This is an archive of the discontinued LLVM Phabricator instance.

Differential D36536

[X86][AVX512] Add combine for TESTM
ClosedPublic

Authored by guyblank on Aug 9 2017, 10:54 AM.

Details

Reviewers
craig.topper
zvi
RKSimon
Commits
rGde425ae75386: [X86][AVX512] Add combine for TESTM
rL310787: [X86][AVX512] Add combine for TESTM
Summary

The patch adds an X86 combine for TESTM when one of the operands is a BUILD_VECTOR(0,0,...)

Diff Detail

Event Timeline

guyblank created this revision.Aug 9 2017, 10:54 AM
RKSimon added a reviewer: RKSimon.Aug 9 2017, 1:58 PM
RKSimon added a subscriber: RKSimon.
RKSimon added inline comments.
lib/Target/X86/X86ISelLowering.cpp
35498
// TEST BUILD_VECTOR(all_zero), op1 -> BUILD_VECTOR(all_zero)
35502

Worth adding vXi1 support to the getZeroVector() helper?

guyblank added inline comments.Aug 9 2017, 10:57 PM
lib/Target/X86/X86ISelLowering.cpp
35502

It's already supported, I somehow missed that.
Thanks.

guyblank updated this revision to Diff 110520.Aug 9 2017, 11:05 PM

Address Simon's comments.

RKSimon accepted this revision.Aug 11 2017, 5:30 AM

LGTM, thanks

This revision is now accepted and ready to land.Aug 11 2017, 5:30 AM
Closed by commit rL310787: [X86][AVX512] Add combine for TESTM (authored by guyblank). · Explain WhyAug 13 2017, 1:04 AM
This revision was automatically updated to reflect the committed changes.

Revision Contents

PathSize
lib/
Target/
X86/
25 lines
test/
CodeGen/
X86/
2156 lines

Diff 110520

lib/Target/X86/X86ISelLowering.cpp

  • This file is larger than 256 KB, so syntax highlighting is disabled by default.
Show First 20 Lines • Show All 35,475 Lines • ▼ Show 20 Linesstatic SDValue combineLockSub(SDNode *N, SelectionDAG &DAG,
RHS = DAG.getConstant(-1, DL, VT); RHS = DAG.getConstant(-1, DL, VT);
MachineMemOperand *MMO = cast<MemSDNode>(N)->getMemOperand(); MachineMemOperand *MMO = cast<MemSDNode>(N)->getMemOperand();
return DAG.getMemIntrinsicNode(X86ISD::LADD, DL, return DAG.getMemIntrinsicNode(X86ISD::LADD, DL,
DAG.getVTList(MVT::i32, MVT::Other), DAG.getVTList(MVT::i32, MVT::Other),
{Chain, LHS, RHS}, VT, MMO); {Chain, LHS, RHS}, VT, MMO);
} }
// TEST (AND a, b) ,(AND a, b) -> TEST a, b static SDValue combineTestM(SDNode *N, SelectionDAG &DAG,
static SDValue combineTestM(SDNode *N, SelectionDAG &DAG) { const X86Subtarget &Subtarget) {
SDValue Op0 = N->getOperand(0); SDValue Op0 = N->getOperand(0);
SDValue Op1 = N->getOperand(1); SDValue Op1 = N->getOperand(1);
if (Op0 != Op1 || Op1->getOpcode() != ISD::AND) MVT VT = N->getSimpleValueType(0);
return SDValue();
EVT VT = N->getValueType(0);
SDLoc DL(N); SDLoc DL(N);
return DAG.getNode(X86ISD::TESTM, DL, VT, // TEST (AND a, b) ,(AND a, b) -> TEST a, b
Op0->getOperand(0), Op0->getOperand(1)); if (Op0 == Op1 && Op1->getOpcode() == ISD::AND)
return DAG.getNode(X86ISD::TESTM, DL, VT, Op0->getOperand(0),
Op0->getOperand(1));
// TEST op0, BUILD_VECTOR(all_zero) -> BUILD_VECTOR(all_zero)
// TEST BUILD_VECTOR(all_zero), op1 -> BUILD_VECTOR(all_zero)
RKSimonUnsubmitted
Not Done
ReplyInline Actions
// TEST BUILD_VECTOR(all_zero), op1 -> BUILD_VECTOR(all_zero)
RKSimon: ``` // TEST BUILD_VECTOR(all_zero), op1 -> BUILD_VECTOR(all_zero) ```
if (ISD::isBuildVectorAllZeros(Op0.getNode()) ||
ISD::isBuildVectorAllZeros(Op1.getNode()))
return getZeroVector(VT, Subtarget, DAG, DL);
RKSimonUnsubmitted
Not Done
ReplyInline Actions

Worth adding vXi1 support to the getZeroVector() helper?

RKSimon: Worth adding vXi1 support to the getZeroVector() helper?
guyblankAuthorUnsubmitted
Not Done
ReplyInline Actions

It's already supported, I somehow missed that.
Thanks.

guyblank: It's already supported, I somehow missed that. Thanks.
return SDValue();
} }
static SDValue combineVectorCompare(SDNode *N, SelectionDAG &DAG, static SDValue combineVectorCompare(SDNode *N, SelectionDAG &DAG,
const X86Subtarget &Subtarget) { const X86Subtarget &Subtarget) {
MVT VT = N->getSimpleValueType(0); MVT VT = N->getSimpleValueType(0);
SDLoc DL(N); SDLoc DL(N);
if (N->getOperand(0) == N->getOperand(1)) { if (N->getOperand(0) == N->getOperand(1)) {
▲ Show 20 Lines • Show All 200 Lines • ▼ Show 20 LinesSDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
case X86ISD::FMADD: case X86ISD::FMADD:
case X86ISD::FMADD_RND: case X86ISD::FMADD_RND:
case X86ISD::FMADDS1_RND: case X86ISD::FMADDS1_RND:
case X86ISD::FMADDS3_RND: case X86ISD::FMADDS3_RND:
case ISD::FMA: return combineFMA(N, DAG, Subtarget); case ISD::FMA: return combineFMA(N, DAG, Subtarget);
case ISD::MGATHER: case ISD::MGATHER:
case ISD::MSCATTER: return combineGatherScatter(N, DAG); case ISD::MSCATTER: return combineGatherScatter(N, DAG);
case X86ISD::LSUB: return combineLockSub(N, DAG, Subtarget); case X86ISD::LSUB: return combineLockSub(N, DAG, Subtarget);
case X86ISD::TESTM: return combineTestM(N, DAG); case X86ISD::TESTM: return combineTestM(N, DAG, Subtarget);
case X86ISD::PCMPEQ: case X86ISD::PCMPEQ:
case X86ISD::PCMPGT: return combineVectorCompare(N, DAG, Subtarget); case X86ISD::PCMPGT: return combineVectorCompare(N, DAG, Subtarget);
} }
return SDValue(); return SDValue();
} }
/// Return true if the target has native support for the specified value type /// Return true if the target has native support for the specified value type
▲ Show 20 Lines • Show All 996 Lines • Show Last 20 Lines

test/CodeGen/X86/avx512vl-vec-masked-cmp.ll

  • This file is larger than 256 KB, so syntax highlighting is disabled by default.
Show All 34 Lines
; NoVLX-NEXT: .Lcfi6: ; NoVLX-NEXT: .Lcfi6:
; NoVLX-NEXT: .cfi_offset %r14, -32 ; NoVLX-NEXT: .cfi_offset %r14, -32
; NoVLX-NEXT: .Lcfi7: ; NoVLX-NEXT: .Lcfi7:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: vpcmpeqb %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vpcmpeqb %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0 ; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r10d ; NoVLX-NEXT: kmovw %k1, %r10d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 108 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi14: ; NoVLX-NEXT: .Lcfi14:
; NoVLX-NEXT: .cfi_offset %r14, -32 ; NoVLX-NEXT: .cfi_offset %r14, -32
; NoVLX-NEXT: .Lcfi15: ; NoVLX-NEXT: .Lcfi15:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: vpcmpeqb (%rdi), %xmm0, %xmm0 ; NoVLX-NEXT: vpcmpeqb (%rdi), %xmm0, %xmm0
; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0 ; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r10d ; NoVLX-NEXT: kmovw %k1, %r10d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 111 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %r14, -32 ; NoVLX-NEXT: .cfi_offset %r14, -32
; NoVLX-NEXT: .Lcfi23: ; NoVLX-NEXT: .Lcfi23:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: vpcmpeqb %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vpcmpeqb %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0 ; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 {%k1} ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r10d ; NoVLX-NEXT: kmovw %k1, %r10d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 112 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %r14, -32 ; NoVLX-NEXT: .cfi_offset %r14, -32
; NoVLX-NEXT: .Lcfi31: ; NoVLX-NEXT: .Lcfi31:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: vpcmpeqb (%rsi), %xmm0, %xmm0 ; NoVLX-NEXT: vpcmpeqb (%rsi), %xmm0, %xmm0
; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0 ; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 {%k1} ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r10d ; NoVLX-NEXT: kmovw %k1, %r10d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 112 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi38: ; NoVLX-NEXT: .Lcfi38:
; NoVLX-NEXT: .cfi_offset %r14, -32 ; NoVLX-NEXT: .cfi_offset %r14, -32
; NoVLX-NEXT: .Lcfi39: ; NoVLX-NEXT: .Lcfi39:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: vpcmpeqb %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vpcmpeqb %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0 ; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 113 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi46: ; NoVLX-NEXT: .Lcfi46:
; NoVLX-NEXT: .cfi_offset %r14, -32 ; NoVLX-NEXT: .cfi_offset %r14, -32
; NoVLX-NEXT: .Lcfi47: ; NoVLX-NEXT: .Lcfi47:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: vpcmpeqb (%rdi), %xmm0, %xmm0 ; NoVLX-NEXT: vpcmpeqb (%rdi), %xmm0, %xmm0
; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0 ; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 116 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %r14, -32 ; NoVLX-NEXT: .cfi_offset %r14, -32
; NoVLX-NEXT: .Lcfi55: ; NoVLX-NEXT: .Lcfi55:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: vpcmpeqb %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vpcmpeqb %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0 ; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 {%k1} ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 117 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %r14, -32 ; NoVLX-NEXT: .cfi_offset %r14, -32
; NoVLX-NEXT: .Lcfi63: ; NoVLX-NEXT: .Lcfi63:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: vpcmpeqb (%rsi), %xmm0, %xmm0 ; NoVLX-NEXT: vpcmpeqb (%rsi), %xmm0, %xmm0
; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0 ; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 {%k1} ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 110 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0 ; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT: kmovw %k0, (%rsp) ; NoVLX-NEXT: kmovw %k0, (%rsp)
; NoVLX-NEXT: movl (%rsp), %ecx ; NoVLX-NEXT: movl (%rsp), %ecx
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: movl {{[0-9]+}}(%rsp), %eax ; NoVLX-NEXT: movl {{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT: shlq $32, %rax ; NoVLX-NEXT: shlq $32, %rax
; NoVLX-NEXT: orq %rcx, %rax ; NoVLX-NEXT: orq %rcx, %rax
; NoVLX-NEXT: movq %rbp, %rsp ; NoVLX-NEXT: movq %rbp, %rsp
; NoVLX-NEXT: popq %rbp ; NoVLX-NEXT: popq %rbp
; NoVLX-NEXT: retq ; NoVLX-NEXT: retq
Show All 32 Lines
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0 ; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT: kmovw %k0, (%rsp) ; NoVLX-NEXT: kmovw %k0, (%rsp)
; NoVLX-NEXT: movl (%rsp), %ecx ; NoVLX-NEXT: movl (%rsp), %ecx
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: movl {{[0-9]+}}(%rsp), %eax ; NoVLX-NEXT: movl {{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT: shlq $32, %rax ; NoVLX-NEXT: shlq $32, %rax
; NoVLX-NEXT: orq %rcx, %rax ; NoVLX-NEXT: orq %rcx, %rax
; NoVLX-NEXT: movq %rbp, %rsp ; NoVLX-NEXT: movq %rbp, %rsp
; NoVLX-NEXT: popq %rbp ; NoVLX-NEXT: popq %rbp
; NoVLX-NEXT: retq ; NoVLX-NEXT: retq
Show All 30 Lines
; NoVLX-NEXT: subq $96, %rsp ; NoVLX-NEXT: subq $96, %rsp
; NoVLX-NEXT: movl %edi, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: movl %edi, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw {{[0-9]+}}(%rsp), %k1 ; NoVLX-NEXT: kmovw {{[0-9]+}}(%rsp), %k1
; NoVLX-NEXT: kmovw {{[0-9]+}}(%rsp), %k2 ; NoVLX-NEXT: kmovw {{[0-9]+}}(%rsp), %k2
; NoVLX-NEXT: vpternlogd $255, %zmm2, %zmm2, %zmm2 {%k1} {z} ; NoVLX-NEXT: vpternlogd $255, %zmm2, %zmm2, %zmm2 {%k1} {z}
; NoVLX-NEXT: vpmovdb %zmm2, %xmm2 ; NoVLX-NEXT: vpmovdb %zmm2, %xmm2
; NoVLX-NEXT: vpternlogd $255, %zmm3, %zmm3, %zmm3 {%k2} {z} ; NoVLX-NEXT: vpternlogd $255, %zmm3, %zmm3, %zmm3 {%k2} {z}
; NoVLX-NEXT: vpmovdb %zmm3, %xmm3 ; NoVLX-NEXT: vpmovdb %zmm3, %xmm3
; NoVLX-NEXT: vxorps %xmm4, %xmm4, %xmm4
; NoVLX-NEXT: vpcmpeqb %ymm1, %ymm0, %ymm0 ; NoVLX-NEXT: vpcmpeqb %ymm1, %ymm0, %ymm0
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT: vpand %xmm3, %xmm1, %xmm1 ; NoVLX-NEXT: vpand %xmm3, %xmm1, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpand %xmm2, %xmm0, %xmm0 ; NoVLX-NEXT: vpand %xmm2, %xmm0, %xmm0
; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0 ; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT: kmovw %k0, (%rsp) ; NoVLX-NEXT: kmovw %k0, (%rsp)
; NoVLX-NEXT: movl (%rsp), %ecx ; NoVLX-NEXT: movl (%rsp), %ecx
; NoVLX-NEXT: vptestmd %zmm4, %zmm4, %k0 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: movl {{[0-9]+}}(%rsp), %eax ; NoVLX-NEXT: movl {{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT: shlq $32, %rax ; NoVLX-NEXT: shlq $32, %rax
; NoVLX-NEXT: orq %rcx, %rax ; NoVLX-NEXT: orq %rcx, %rax
; NoVLX-NEXT: movq %rbp, %rsp ; NoVLX-NEXT: movq %rbp, %rsp
; NoVLX-NEXT: popq %rbp ; NoVLX-NEXT: popq %rbp
; NoVLX-NEXT: retq ; NoVLX-NEXT: retq
Show All 31 Lines
; NoVLX-NEXT: subq $96, %rsp ; NoVLX-NEXT: subq $96, %rsp
; NoVLX-NEXT: movl %edi, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: movl %edi, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw {{[0-9]+}}(%rsp), %k1 ; NoVLX-NEXT: kmovw {{[0-9]+}}(%rsp), %k1
; NoVLX-NEXT: kmovw {{[0-9]+}}(%rsp), %k2 ; NoVLX-NEXT: kmovw {{[0-9]+}}(%rsp), %k2
; NoVLX-NEXT: vpternlogd $255, %zmm1, %zmm1, %zmm1 {%k1} {z} ; NoVLX-NEXT: vpternlogd $255, %zmm1, %zmm1, %zmm1 {%k1} {z}
; NoVLX-NEXT: vpmovdb %zmm1, %xmm1 ; NoVLX-NEXT: vpmovdb %zmm1, %xmm1
; NoVLX-NEXT: vpternlogd $255, %zmm2, %zmm2, %zmm2 {%k2} {z} ; NoVLX-NEXT: vpternlogd $255, %zmm2, %zmm2, %zmm2 {%k2} {z}
; NoVLX-NEXT: vpmovdb %zmm2, %xmm2 ; NoVLX-NEXT: vpmovdb %zmm2, %xmm2
; NoVLX-NEXT: vxorps %xmm3, %xmm3, %xmm3
; NoVLX-NEXT: vpcmpeqb (%rsi), %ymm0, %ymm0 ; NoVLX-NEXT: vpcmpeqb (%rsi), %ymm0, %ymm0
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm4 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm3
; NoVLX-NEXT: vpand %xmm2, %xmm4, %xmm2 ; NoVLX-NEXT: vpand %xmm2, %xmm3, %xmm2
; NoVLX-NEXT: vpmovsxbd %xmm2, %zmm2 ; NoVLX-NEXT: vpmovsxbd %xmm2, %zmm2
; NoVLX-NEXT: vpslld $31, %zmm2, %zmm2 ; NoVLX-NEXT: vpslld $31, %zmm2, %zmm2
; NoVLX-NEXT: vptestmd %zmm2, %zmm2, %k0 ; NoVLX-NEXT: vptestmd %zmm2, %zmm2, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpand %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0 ; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT: kmovw %k0, (%rsp) ; NoVLX-NEXT: kmovw %k0, (%rsp)
; NoVLX-NEXT: movl (%rsp), %ecx ; NoVLX-NEXT: movl (%rsp), %ecx
; NoVLX-NEXT: vptestmd %zmm3, %zmm3, %k0 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: movl {{[0-9]+}}(%rsp), %eax ; NoVLX-NEXT: movl {{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT: shlq $32, %rax ; NoVLX-NEXT: shlq $32, %rax
; NoVLX-NEXT: orq %rcx, %rax ; NoVLX-NEXT: orq %rcx, %rax
; NoVLX-NEXT: movq %rbp, %rsp ; NoVLX-NEXT: movq %rbp, %rsp
; NoVLX-NEXT: popq %rbp ; NoVLX-NEXT: popq %rbp
; NoVLX-NEXT: retq ; NoVLX-NEXT: retq
▲ Show 20 Lines • Show All 151 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi78: ; NoVLX-NEXT: .Lcfi78:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $32, %rsp ; NoVLX-NEXT: subq $32, %rsp
; NoVLX-NEXT: vpcmpeqw %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vpcmpeqw %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0 ; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0
; NoVLX-NEXT: vpsllq $63, %zmm0, %zmm0 ; NoVLX-NEXT: vpsllq $63, %zmm0, %zmm0
; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0 ; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r9d ; NoVLX-NEXT: kmovw %k1, %r9d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 58 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi81: ; NoVLX-NEXT: .Lcfi81:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $32, %rsp ; NoVLX-NEXT: subq $32, %rsp
; NoVLX-NEXT: vpcmpeqw (%rdi), %xmm0, %xmm0 ; NoVLX-NEXT: vpcmpeqw (%rdi), %xmm0, %xmm0
; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0 ; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0
; NoVLX-NEXT: vpsllq $63, %zmm0, %zmm0 ; NoVLX-NEXT: vpsllq $63, %zmm0, %zmm0
; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0 ; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r9d ; NoVLX-NEXT: kmovw %k1, %r9d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 61 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $32, %rsp ; NoVLX-NEXT: subq $32, %rsp
; NoVLX-NEXT: vpcmpeqw %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vpcmpeqw %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0 ; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0
; NoVLX-NEXT: vpsllq $63, %zmm0, %zmm0 ; NoVLX-NEXT: vpsllq $63, %zmm0, %zmm0
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0 {%k1} ; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r9d ; NoVLX-NEXT: kmovw %k1, %r9d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 62 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $32, %rsp ; NoVLX-NEXT: subq $32, %rsp
; NoVLX-NEXT: vpcmpeqw (%rsi), %xmm0, %xmm0 ; NoVLX-NEXT: vpcmpeqw (%rsi), %xmm0, %xmm0
; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0 ; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0
; NoVLX-NEXT: vpsllq $63, %zmm0, %zmm0 ; NoVLX-NEXT: vpsllq $63, %zmm0, %zmm0
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0 {%k1} ; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r9d ; NoVLX-NEXT: kmovw %k1, %r9d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 62 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi90: ; NoVLX-NEXT: .Lcfi90:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: vpcmpeqw %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vpcmpeqw %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0 ; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0
; NoVLX-NEXT: vpsllq $63, %zmm0, %zmm0 ; NoVLX-NEXT: vpsllq $63, %zmm0, %zmm0
; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0 ; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 63 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi93: ; NoVLX-NEXT: .Lcfi93:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: vpcmpeqw (%rdi), %xmm0, %xmm0 ; NoVLX-NEXT: vpcmpeqw (%rdi), %xmm0, %xmm0
; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0 ; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0
; NoVLX-NEXT: vpsllq $63, %zmm0, %zmm0 ; NoVLX-NEXT: vpsllq $63, %zmm0, %zmm0
; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0 ; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 66 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: vpcmpeqw %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vpcmpeqw %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0 ; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0
; NoVLX-NEXT: vpsllq $63, %zmm0, %zmm0 ; NoVLX-NEXT: vpsllq $63, %zmm0, %zmm0
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0 {%k1} ; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 67 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: vpcmpeqw (%rsi), %xmm0, %xmm0 ; NoVLX-NEXT: vpcmpeqw (%rsi), %xmm0, %xmm0
; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0 ; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0
; NoVLX-NEXT: vpsllq $63, %zmm0, %zmm0 ; NoVLX-NEXT: vpsllq $63, %zmm0, %zmm0
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0 {%k1} ; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 83 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi106: ; NoVLX-NEXT: .Lcfi106:
; NoVLX-NEXT: .cfi_offset %r14, -32 ; NoVLX-NEXT: .cfi_offset %r14, -32
; NoVLX-NEXT: .Lcfi107: ; NoVLX-NEXT: .Lcfi107:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: vpcmpeqw %ymm1, %ymm0, %ymm0 ; NoVLX-NEXT: vpcmpeqw %ymm1, %ymm0, %ymm0
; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0 ; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0 ; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r10d ; NoVLX-NEXT: kmovw %k1, %r10d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 109 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi114: ; NoVLX-NEXT: .Lcfi114:
; NoVLX-NEXT: .cfi_offset %r14, -32 ; NoVLX-NEXT: .cfi_offset %r14, -32
; NoVLX-NEXT: .Lcfi115: ; NoVLX-NEXT: .Lcfi115:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: vpcmpeqw (%rdi), %ymm0, %ymm0 ; NoVLX-NEXT: vpcmpeqw (%rdi), %ymm0, %ymm0
; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0 ; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0 ; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r10d ; NoVLX-NEXT: kmovw %k1, %r10d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 112 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %r14, -32 ; NoVLX-NEXT: .cfi_offset %r14, -32
; NoVLX-NEXT: .Lcfi123: ; NoVLX-NEXT: .Lcfi123:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: vpcmpeqw %ymm1, %ymm0, %ymm0 ; NoVLX-NEXT: vpcmpeqw %ymm1, %ymm0, %ymm0
; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0 ; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0 ; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 {%k1} ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r10d ; NoVLX-NEXT: kmovw %k1, %r10d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 113 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %r14, -32 ; NoVLX-NEXT: .cfi_offset %r14, -32
; NoVLX-NEXT: .Lcfi131: ; NoVLX-NEXT: .Lcfi131:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: vpcmpeqw (%rsi), %ymm0, %ymm0 ; NoVLX-NEXT: vpcmpeqw (%rsi), %ymm0, %ymm0
; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0 ; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0 ; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 {%k1} ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r10d ; NoVLX-NEXT: kmovw %k1, %r10d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 113 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi138: ; NoVLX-NEXT: .Lcfi138:
; NoVLX-NEXT: .cfi_offset %r14, -32 ; NoVLX-NEXT: .cfi_offset %r14, -32
; NoVLX-NEXT: .Lcfi139: ; NoVLX-NEXT: .Lcfi139:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: vpcmpeqw %ymm1, %ymm0, %ymm0 ; NoVLX-NEXT: vpcmpeqw %ymm1, %ymm0, %ymm0
; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0 ; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0 ; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 114 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi146: ; NoVLX-NEXT: .Lcfi146:
; NoVLX-NEXT: .cfi_offset %r14, -32 ; NoVLX-NEXT: .cfi_offset %r14, -32
; NoVLX-NEXT: .Lcfi147: ; NoVLX-NEXT: .Lcfi147:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: vpcmpeqw (%rdi), %ymm0, %ymm0 ; NoVLX-NEXT: vpcmpeqw (%rdi), %ymm0, %ymm0
; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0 ; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0 ; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 117 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %r14, -32 ; NoVLX-NEXT: .cfi_offset %r14, -32
; NoVLX-NEXT: .Lcfi155: ; NoVLX-NEXT: .Lcfi155:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: vpcmpeqw %ymm1, %ymm0, %ymm0 ; NoVLX-NEXT: vpcmpeqw %ymm1, %ymm0, %ymm0
; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0 ; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0 ; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 {%k1} ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 118 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %r14, -32 ; NoVLX-NEXT: .cfi_offset %r14, -32
; NoVLX-NEXT: .Lcfi163: ; NoVLX-NEXT: .Lcfi163:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: vpcmpeqw (%rsi), %ymm0, %ymm0 ; NoVLX-NEXT: vpcmpeqw (%rsi), %ymm0, %ymm0
; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0 ; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0 ; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 {%k1} ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 412 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: kshiftrw $15, %k0, %k0 ; NoVLX-NEXT: kshiftrw $15, %k0, %k0
; NoVLX-NEXT: kmovw %k0, %eax ; NoVLX-NEXT: kmovw %k0, %eax
; NoVLX-NEXT: vpinsrb $15, %eax, %xmm0, %xmm0 ; NoVLX-NEXT: vpinsrb $15, %eax, %xmm0, %xmm0
; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0 ; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT: kmovw %k0, (%rsp) ; NoVLX-NEXT: kmovw %k0, (%rsp)
; NoVLX-NEXT: movl (%rsp), %ecx ; NoVLX-NEXT: movl (%rsp), %ecx
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: movl {{[0-9]+}}(%rsp), %eax ; NoVLX-NEXT: movl {{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT: shlq $32, %rax ; NoVLX-NEXT: shlq $32, %rax
; NoVLX-NEXT: orq %rcx, %rax ; NoVLX-NEXT: orq %rcx, %rax
; NoVLX-NEXT: movq %rbp, %rsp ; NoVLX-NEXT: movq %rbp, %rsp
; NoVLX-NEXT: popq %rbp ; NoVLX-NEXT: popq %rbp
; NoVLX-NEXT: retq ; NoVLX-NEXT: retq
▲ Show 20 Lines • Show All 249 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: kshiftrw $15, %k0, %k0 ; NoVLX-NEXT: kshiftrw $15, %k0, %k0
; NoVLX-NEXT: kmovw %k0, %eax ; NoVLX-NEXT: kmovw %k0, %eax
; NoVLX-NEXT: vpinsrb $15, %eax, %xmm0, %xmm0 ; NoVLX-NEXT: vpinsrb $15, %eax, %xmm0, %xmm0
; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0 ; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT: kmovw %k0, (%rsp) ; NoVLX-NEXT: kmovw %k0, (%rsp)
; NoVLX-NEXT: movl (%rsp), %ecx ; NoVLX-NEXT: movl (%rsp), %ecx
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: movl {{[0-9]+}}(%rsp), %eax ; NoVLX-NEXT: movl {{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT: shlq $32, %rax ; NoVLX-NEXT: shlq $32, %rax
; NoVLX-NEXT: orq %rcx, %rax ; NoVLX-NEXT: orq %rcx, %rax
; NoVLX-NEXT: movq %rbp, %rsp ; NoVLX-NEXT: movq %rbp, %rsp
; NoVLX-NEXT: popq %rbp ; NoVLX-NEXT: popq %rbp
; NoVLX-NEXT: retq ; NoVLX-NEXT: retq
▲ Show 20 Lines • Show All 180 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: shrl $16, %eax ; NoVLX-NEXT: shrl $16, %eax
; NoVLX-NEXT: vmovd %ecx, %xmm2 ; NoVLX-NEXT: vmovd %ecx, %xmm2
; NoVLX-NEXT: vpinsrw $1, %eax, %xmm2, %xmm2 ; NoVLX-NEXT: vpinsrw $1, %eax, %xmm2, %xmm2
; NoVLX-NEXT: movq %rcx, %rax ; NoVLX-NEXT: movq %rcx, %rax
; NoVLX-NEXT: shrq $32, %rax ; NoVLX-NEXT: shrq $32, %rax
; NoVLX-NEXT: vpinsrw $2, %eax, %xmm2, %xmm2 ; NoVLX-NEXT: vpinsrw $2, %eax, %xmm2, %xmm2
; NoVLX-NEXT: vpextrq $1, %xmm1, %rax ; NoVLX-NEXT: vpextrq $1, %xmm1, %rax
; NoVLX-NEXT: vinserti128 $1, %xmm9, %ymm4, %ymm1 ; NoVLX-NEXT: vinserti128 $1, %xmm9, %ymm4, %ymm1
; NoVLX-NEXT: vinserti128 $1, %xmm6, %ymm0, %ymm8 ; NoVLX-NEXT: vinserti128 $1, %xmm6, %ymm0, %ymm4
; NoVLX-NEXT: kmovw {{[0-9]+}}(%rsp), %k1 ; NoVLX-NEXT: kmovw {{[0-9]+}}(%rsp), %k1
; NoVLX-NEXT: kmovw {{[0-9]+}}(%rsp), %k2 ; NoVLX-NEXT: kmovw {{[0-9]+}}(%rsp), %k2
; NoVLX-NEXT: vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT: vinserti128 $1, %xmm7, %ymm3, %ymm3 ; NoVLX-NEXT: vinserti128 $1, %xmm7, %ymm3, %ymm3
; NoVLX-NEXT: vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT: vpmovdb %zmm0, %xmm0 ; NoVLX-NEXT: vpmovdb %zmm0, %xmm0
; NoVLX-NEXT: vpternlogd $255, %zmm6, %zmm6, %zmm6 {%k2} {z} ; NoVLX-NEXT: vpcmpeqw %ymm3, %ymm1, %ymm3
; NoVLX-NEXT: vpcmpeqw %ymm3, %ymm1, %ymm4 ; NoVLX-NEXT: vpternlogd $255, %zmm1, %zmm1, %zmm1 {%k2} {z}
; NoVLX-NEXT: vpmovdb %zmm6, %xmm6 ; NoVLX-NEXT: vpmovdb %zmm1, %xmm1
; NoVLX-NEXT: shrq $48, %rcx ; NoVLX-NEXT: shrq $48, %rcx
; NoVLX-NEXT: vpinsrw $3, %ecx, %xmm2, %xmm1 ; NoVLX-NEXT: vpinsrw $3, %ecx, %xmm2, %xmm2
; NoVLX-NEXT: movl %eax, %ecx ; NoVLX-NEXT: movl %eax, %ecx
; NoVLX-NEXT: shrl $16, %ecx ; NoVLX-NEXT: shrl $16, %ecx
; NoVLX-NEXT: vpinsrw $4, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrw $4, %eax, %xmm2, %xmm2
; NoVLX-NEXT: vpinsrw $5, %ecx, %xmm1, %xmm2 ; NoVLX-NEXT: vpinsrw $5, %ecx, %xmm2, %xmm2
; NoVLX-NEXT: movq %rax, %rcx ; NoVLX-NEXT: movq %rax, %rcx
; NoVLX-NEXT: shrq $48, %rax ; NoVLX-NEXT: shrq $48, %rax
; NoVLX-NEXT: shrq $32, %rcx ; NoVLX-NEXT: shrq $32, %rcx
; NoVLX-NEXT: vpxor %xmm1, %xmm1, %xmm1
; NoVLX-NEXT: vpinsrw $6, %ecx, %xmm2, %xmm2 ; NoVLX-NEXT: vpinsrw $6, %ecx, %xmm2, %xmm2
; NoVLX-NEXT: vpinsrw $7, %eax, %xmm2, %xmm2 ; NoVLX-NEXT: vpinsrw $7, %eax, %xmm2, %xmm2
; NoVLX-NEXT: vinserti128 $1, %xmm5, %ymm2, %ymm2 ; NoVLX-NEXT: vinserti128 $1, %xmm5, %ymm2, %ymm2
; NoVLX-NEXT: vpcmpeqw %ymm2, %ymm8, %ymm2 ; NoVLX-NEXT: vpcmpeqw %ymm2, %ymm4, %ymm2
; NoVLX-NEXT: vpmovsxwd %ymm2, %zmm2 ; NoVLX-NEXT: vpmovsxwd %ymm2, %zmm2
; NoVLX-NEXT: vpslld $31, %zmm2, %zmm2 ; NoVLX-NEXT: vpslld $31, %zmm2, %zmm2
; NoVLX-NEXT: vptestmd %zmm2, %zmm2, %k0 ; NoVLX-NEXT: vptestmd %zmm2, %zmm2, %k0
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 49 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $13, %eax, %xmm2, %xmm2 ; NoVLX-NEXT: vpinsrb $13, %eax, %xmm2, %xmm2
; NoVLX-NEXT: kshiftlw $1, %k0, %k1 ; NoVLX-NEXT: kshiftlw $1, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $14, %eax, %xmm2, %xmm2 ; NoVLX-NEXT: vpinsrb $14, %eax, %xmm2, %xmm2
; NoVLX-NEXT: kshiftrw $15, %k0, %k0 ; NoVLX-NEXT: kshiftrw $15, %k0, %k0
; NoVLX-NEXT: kmovw %k0, %eax ; NoVLX-NEXT: kmovw %k0, %eax
; NoVLX-NEXT: vpinsrb $15, %eax, %xmm2, %xmm3 ; NoVLX-NEXT: vpinsrb $15, %eax, %xmm2, %xmm2
; NoVLX-NEXT: vpmovsxwd %ymm4, %zmm2 ; NoVLX-NEXT: vpmovsxwd %ymm3, %zmm3
; NoVLX-NEXT: vpslld $31, %zmm2, %zmm2 ; NoVLX-NEXT: vpslld $31, %zmm3, %zmm3
; NoVLX-NEXT: vptestmd %zmm2, %zmm2, %k0 ; NoVLX-NEXT: vptestmd %zmm3, %zmm3, %k0
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %ecx ; NoVLX-NEXT: kmovw %k1, %ecx
; NoVLX-NEXT: vmovd %ecx, %xmm2 ; NoVLX-NEXT: vmovd %ecx, %xmm3
; NoVLX-NEXT: vpinsrb $1, %eax, %xmm2, %xmm2 ; NoVLX-NEXT: vpinsrb $1, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $2, %eax, %xmm2, %xmm2 ; NoVLX-NEXT: vpinsrb $2, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftlw $12, %k0, %k1 ; NoVLX-NEXT: kshiftlw $12, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $3, %eax, %xmm2, %xmm2 ; NoVLX-NEXT: vpinsrb $3, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftlw $11, %k0, %k1 ; NoVLX-NEXT: kshiftlw $11, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $4, %eax, %xmm2, %xmm2 ; NoVLX-NEXT: vpinsrb $4, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftlw $10, %k0, %k1 ; NoVLX-NEXT: kshiftlw $10, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $5, %eax, %xmm2, %xmm2 ; NoVLX-NEXT: vpinsrb $5, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftlw $9, %k0, %k1 ; NoVLX-NEXT: kshiftlw $9, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $6, %eax, %xmm2, %xmm2 ; NoVLX-NEXT: vpinsrb $6, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftlw $8, %k0, %k1 ; NoVLX-NEXT: kshiftlw $8, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $7, %eax, %xmm2, %xmm2 ; NoVLX-NEXT: vpinsrb $7, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftlw $7, %k0, %k1 ; NoVLX-NEXT: kshiftlw $7, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $8, %eax, %xmm2, %xmm2 ; NoVLX-NEXT: vpinsrb $8, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftlw $6, %k0, %k1 ; NoVLX-NEXT: kshiftlw $6, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $9, %eax, %xmm2, %xmm2 ; NoVLX-NEXT: vpinsrb $9, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftlw $5, %k0, %k1 ; NoVLX-NEXT: kshiftlw $5, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $10, %eax, %xmm2, %xmm2 ; NoVLX-NEXT: vpinsrb $10, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftlw $4, %k0, %k1 ; NoVLX-NEXT: kshiftlw $4, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $11, %eax, %xmm2, %xmm2 ; NoVLX-NEXT: vpinsrb $11, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftlw $3, %k0, %k1 ; NoVLX-NEXT: kshiftlw $3, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $12, %eax, %xmm2, %xmm2 ; NoVLX-NEXT: vpinsrb $12, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftlw $2, %k0, %k1 ; NoVLX-NEXT: kshiftlw $2, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $13, %eax, %xmm2, %xmm2 ; NoVLX-NEXT: vpinsrb $13, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftlw $1, %k0, %k1 ; NoVLX-NEXT: kshiftlw $1, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $14, %eax, %xmm2, %xmm2 ; NoVLX-NEXT: vpinsrb $14, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftrw $15, %k0, %k0 ; NoVLX-NEXT: kshiftrw $15, %k0, %k0
; NoVLX-NEXT: kmovw %k0, %eax ; NoVLX-NEXT: kmovw %k0, %eax
; NoVLX-NEXT: vpinsrb $15, %eax, %xmm2, %xmm2 ; NoVLX-NEXT: vpinsrb $15, %eax, %xmm3, %xmm3
; NoVLX-NEXT: vpand %xmm6, %xmm2, %xmm2 ; NoVLX-NEXT: vpand %xmm1, %xmm3, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm2, %zmm2 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm2, %zmm2 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm2, %zmm2, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpand %xmm0, %xmm3, %xmm0 ; NoVLX-NEXT: vpand %xmm0, %xmm2, %xmm0
; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0 ; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT: kmovw %k0, (%rsp) ; NoVLX-NEXT: kmovw %k0, (%rsp)
; NoVLX-NEXT: movl (%rsp), %ecx ; NoVLX-NEXT: movl (%rsp), %ecx
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: movl {{[0-9]+}}(%rsp), %eax ; NoVLX-NEXT: movl {{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT: shlq $32, %rax ; NoVLX-NEXT: shlq $32, %rax
; NoVLX-NEXT: orq %rcx, %rax ; NoVLX-NEXT: orq %rcx, %rax
; NoVLX-NEXT: movq %rbp, %rsp ; NoVLX-NEXT: movq %rbp, %rsp
; NoVLX-NEXT: popq %rbp ; NoVLX-NEXT: popq %rbp
; NoVLX-NEXT: retq ; NoVLX-NEXT: retq
▲ Show 20 Lines • Show All 48 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: shrl $16, %ecx ; NoVLX-NEXT: shrl $16, %ecx
; NoVLX-NEXT: vpinsrw $4, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrw $4, %eax, %xmm1, %xmm1
; NoVLX-NEXT: vpinsrw $5, %ecx, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrw $5, %ecx, %xmm1, %xmm1
; NoVLX-NEXT: movq %rax, %rcx ; NoVLX-NEXT: movq %rax, %rcx
; NoVLX-NEXT: shrq $32, %rcx ; NoVLX-NEXT: shrq $32, %rcx
; NoVLX-NEXT: vpinsrw $6, %ecx, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrw $6, %ecx, %xmm1, %xmm1
; NoVLX-NEXT: vmovq %xmm4, %rcx ; NoVLX-NEXT: vmovq %xmm4, %rcx
; NoVLX-NEXT: shrq $48, %rax ; NoVLX-NEXT: shrq $48, %rax
; NoVLX-NEXT: vpinsrw $7, %eax, %xmm1, %xmm2 ; NoVLX-NEXT: vpinsrw $7, %eax, %xmm1, %xmm1
; NoVLX-NEXT: movl %ecx, %eax ; NoVLX-NEXT: movl %ecx, %eax
; NoVLX-NEXT: shrl $16, %eax ; NoVLX-NEXT: shrl $16, %eax
; NoVLX-NEXT: vmovd %ecx, %xmm1 ; NoVLX-NEXT: vmovd %ecx, %xmm2
; NoVLX-NEXT: vpinsrw $1, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrw $1, %eax, %xmm2, %xmm2
; NoVLX-NEXT: movq %rcx, %rax ; NoVLX-NEXT: movq %rcx, %rax
; NoVLX-NEXT: shrq $32, %rax ; NoVLX-NEXT: shrq $32, %rax
; NoVLX-NEXT: vpinsrw $2, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrw $2, %eax, %xmm2, %xmm2
; NoVLX-NEXT: vpextrq $1, %xmm4, %rax ; NoVLX-NEXT: vpextrq $1, %xmm4, %rax
; NoVLX-NEXT: shrq $48, %rcx ; NoVLX-NEXT: shrq $48, %rcx
; NoVLX-NEXT: vpinsrw $3, %ecx, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrw $3, %ecx, %xmm2, %xmm2
; NoVLX-NEXT: movl %eax, %ecx ; NoVLX-NEXT: movl %eax, %ecx
; NoVLX-NEXT: shrl $16, %ecx ; NoVLX-NEXT: shrl $16, %ecx
; NoVLX-NEXT: vpinsrw $4, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrw $4, %eax, %xmm2, %xmm2
; NoVLX-NEXT: vpinsrw $5, %ecx, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrw $5, %ecx, %xmm2, %xmm2
; NoVLX-NEXT: movq %rax, %rcx ; NoVLX-NEXT: movq %rax, %rcx
; NoVLX-NEXT: shrq $32, %rcx ; NoVLX-NEXT: shrq $32, %rcx
; NoVLX-NEXT: vpinsrw $6, %ecx, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrw $6, %ecx, %xmm2, %xmm2
; NoVLX-NEXT: vmovq %xmm3, %rcx ; NoVLX-NEXT: vmovq %xmm3, %rcx
; NoVLX-NEXT: shrq $48, %rax ; NoVLX-NEXT: shrq $48, %rax
; NoVLX-NEXT: vpinsrw $7, %eax, %xmm1, %xmm4 ; NoVLX-NEXT: vpinsrw $7, %eax, %xmm2, %xmm2
; NoVLX-NEXT: movl %ecx, %eax ; NoVLX-NEXT: movl %ecx, %eax
; NoVLX-NEXT: shrl $16, %eax ; NoVLX-NEXT: shrl $16, %eax
; NoVLX-NEXT: vmovd %ecx, %xmm1 ; NoVLX-NEXT: vmovd %ecx, %xmm4
; NoVLX-NEXT: vpinsrw $1, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrw $1, %eax, %xmm4, %xmm4
; NoVLX-NEXT: movq %rcx, %rax ; NoVLX-NEXT: movq %rcx, %rax
; NoVLX-NEXT: shrq $32, %rax ; NoVLX-NEXT: shrq $32, %rax
; NoVLX-NEXT: vpinsrw $2, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrw $2, %eax, %xmm4, %xmm4
; NoVLX-NEXT: vpextrq $1, %xmm3, %rax ; NoVLX-NEXT: vpextrq $1, %xmm3, %rax
; NoVLX-NEXT: shrq $48, %rcx ; NoVLX-NEXT: shrq $48, %rcx
; NoVLX-NEXT: vpinsrw $3, %ecx, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrw $3, %ecx, %xmm4, %xmm3
; NoVLX-NEXT: movl %eax, %ecx ; NoVLX-NEXT: movl %eax, %ecx
; NoVLX-NEXT: shrl $16, %ecx ; NoVLX-NEXT: shrl $16, %ecx
; NoVLX-NEXT: vpinsrw $4, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrw $4, %eax, %xmm3, %xmm3
; NoVLX-NEXT: vpinsrw $5, %ecx, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrw $5, %ecx, %xmm3, %xmm3
; NoVLX-NEXT: movq %rax, %rcx ; NoVLX-NEXT: movq %rax, %rcx
; NoVLX-NEXT: shrq $32, %rcx ; NoVLX-NEXT: shrq $32, %rcx
; NoVLX-NEXT: vpinsrw $6, %ecx, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrw $6, %ecx, %xmm3, %xmm3
; NoVLX-NEXT: vmovq %xmm0, %rcx ; NoVLX-NEXT: vmovq %xmm0, %rcx
; NoVLX-NEXT: shrq $48, %rax ; NoVLX-NEXT: shrq $48, %rax
; NoVLX-NEXT: vpinsrw $7, %eax, %xmm1, %xmm3 ; NoVLX-NEXT: vpinsrw $7, %eax, %xmm3, %xmm4
; NoVLX-NEXT: movl %ecx, %eax ; NoVLX-NEXT: movl %ecx, %eax
; NoVLX-NEXT: shrl $16, %eax ; NoVLX-NEXT: shrl $16, %eax
; NoVLX-NEXT: vmovd %ecx, %xmm1 ; NoVLX-NEXT: vmovd %ecx, %xmm3
; NoVLX-NEXT: vpinsrw $1, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrw $1, %eax, %xmm3, %xmm3
; NoVLX-NEXT: movq %rcx, %rax ; NoVLX-NEXT: movq %rcx, %rax
; NoVLX-NEXT: shrq $32, %rax ; NoVLX-NEXT: shrq $32, %rax
; NoVLX-NEXT: vpinsrw $2, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrw $2, %eax, %xmm3, %xmm3
; NoVLX-NEXT: vpextrq $1, %xmm0, %rax ; NoVLX-NEXT: vpextrq $1, %xmm0, %rax
; NoVLX-NEXT: shrq $48, %rcx ; NoVLX-NEXT: shrq $48, %rcx
; NoVLX-NEXT: vpinsrw $3, %ecx, %xmm1, %xmm0 ; NoVLX-NEXT: vpinsrw $3, %ecx, %xmm3, %xmm0
; NoVLX-NEXT: movl %eax, %ecx ; NoVLX-NEXT: movl %eax, %ecx
; NoVLX-NEXT: shrl $16, %ecx ; NoVLX-NEXT: shrl $16, %ecx
; NoVLX-NEXT: vpinsrw $4, %eax, %xmm0, %xmm0 ; NoVLX-NEXT: vpinsrw $4, %eax, %xmm0, %xmm0
; NoVLX-NEXT: vpinsrw $5, %ecx, %xmm0, %xmm0 ; NoVLX-NEXT: vpinsrw $5, %ecx, %xmm0, %xmm0
; NoVLX-NEXT: movq %rax, %rcx ; NoVLX-NEXT: movq %rax, %rcx
; NoVLX-NEXT: shrq $32, %rcx ; NoVLX-NEXT: shrq $32, %rcx
; NoVLX-NEXT: vpinsrw $6, %ecx, %xmm0, %xmm5 ; NoVLX-NEXT: vpinsrw $6, %ecx, %xmm0, %xmm5
; NoVLX-NEXT: kmovw {{[0-9]+}}(%rsp), %k1 ; NoVLX-NEXT: kmovw {{[0-9]+}}(%rsp), %k1
; NoVLX-NEXT: kmovw {{[0-9]+}}(%rsp), %k2 ; NoVLX-NEXT: kmovw {{[0-9]+}}(%rsp), %k2
; NoVLX-NEXT: vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z} ; NoVLX-NEXT: vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT: vpmovdb %zmm0, %xmm1 ; NoVLX-NEXT: vpmovdb %zmm0, %xmm0
; NoVLX-NEXT: vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k2} {z} ; NoVLX-NEXT: vinserti128 $1, %xmm1, %ymm2, %ymm3
; NoVLX-NEXT: vinserti128 $1, %xmm2, %ymm4, %ymm4 ; NoVLX-NEXT: vpternlogd $255, %zmm1, %zmm1, %zmm1 {%k2} {z}
; NoVLX-NEXT: vpmovdb %zmm0, %xmm2 ; NoVLX-NEXT: vpmovdb %zmm1, %xmm1
; NoVLX-NEXT: shrq $48, %rax ; NoVLX-NEXT: shrq $48, %rax
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: vpinsrw $7, %eax, %xmm5, %xmm2
; NoVLX-NEXT: vpinsrw $7, %eax, %xmm5, %xmm5 ; NoVLX-NEXT: vinserti128 $1, %xmm4, %ymm2, %ymm2
; NoVLX-NEXT: vinserti128 $1, %xmm3, %ymm5, %ymm3 ; NoVLX-NEXT: vpcmpeqw (%rsi), %ymm2, %ymm2
; NoVLX-NEXT: vpcmpeqw (%rsi), %ymm3, %ymm3 ; NoVLX-NEXT: vpmovsxwd %ymm2, %zmm2
; NoVLX-NEXT: vpmovsxwd %ymm3, %zmm3 ; NoVLX-NEXT: vpslld $31, %zmm2, %zmm2
; NoVLX-NEXT: vpslld $31, %zmm3, %zmm3 ; NoVLX-NEXT: vptestmd %zmm2, %zmm2, %k0
; NoVLX-NEXT: vptestmd %zmm3, %zmm3, %k0
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: vmovd %eax, %xmm3 ; NoVLX-NEXT: vmovd %eax, %xmm2
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $1, %eax, %xmm3, %xmm3 ; NoVLX-NEXT: vpinsrb $1, %eax, %xmm2, %xmm2
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $2, %eax, %xmm3, %xmm3 ; NoVLX-NEXT: vpinsrb $2, %eax, %xmm2, %xmm2
; NoVLX-NEXT: kshiftlw $12, %k0, %k1 ; NoVLX-NEXT: kshiftlw $12, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $3, %eax, %xmm3, %xmm3 ; NoVLX-NEXT: vpinsrb $3, %eax, %xmm2, %xmm2
; NoVLX-NEXT: kshiftlw $11, %k0, %k1 ; NoVLX-NEXT: kshiftlw $11, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $4, %eax, %xmm3, %xmm3 ; NoVLX-NEXT: vpinsrb $4, %eax, %xmm2, %xmm2
; NoVLX-NEXT: kshiftlw $10, %k0, %k1 ; NoVLX-NEXT: kshiftlw $10, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $5, %eax, %xmm3, %xmm3 ; NoVLX-NEXT: vpinsrb $5, %eax, %xmm2, %xmm2
; NoVLX-NEXT: kshiftlw $9, %k0, %k1 ; NoVLX-NEXT: kshiftlw $9, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $6, %eax, %xmm3, %xmm3 ; NoVLX-NEXT: vpinsrb $6, %eax, %xmm2, %xmm2
; NoVLX-NEXT: kshiftlw $8, %k0, %k1 ; NoVLX-NEXT: kshiftlw $8, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $7, %eax, %xmm3, %xmm3 ; NoVLX-NEXT: vpinsrb $7, %eax, %xmm2, %xmm2
; NoVLX-NEXT: kshiftlw $7, %k0, %k1 ; NoVLX-NEXT: kshiftlw $7, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $8, %eax, %xmm3, %xmm3 ; NoVLX-NEXT: vpinsrb $8, %eax, %xmm2, %xmm2
; NoVLX-NEXT: kshiftlw $6, %k0, %k1 ; NoVLX-NEXT: kshiftlw $6, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $9, %eax, %xmm3, %xmm3 ; NoVLX-NEXT: vpinsrb $9, %eax, %xmm2, %xmm2
; NoVLX-NEXT: kshiftlw $5, %k0, %k1 ; NoVLX-NEXT: kshiftlw $5, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $10, %eax, %xmm3, %xmm3 ; NoVLX-NEXT: vpinsrb $10, %eax, %xmm2, %xmm2
; NoVLX-NEXT: kshiftlw $4, %k0, %k1 ; NoVLX-NEXT: kshiftlw $4, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $11, %eax, %xmm3, %xmm3 ; NoVLX-NEXT: vpinsrb $11, %eax, %xmm2, %xmm2
; NoVLX-NEXT: kshiftlw $3, %k0, %k1 ; NoVLX-NEXT: kshiftlw $3, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $12, %eax, %xmm3, %xmm3 ; NoVLX-NEXT: vpinsrb $12, %eax, %xmm2, %xmm2
; NoVLX-NEXT: kshiftlw $2, %k0, %k1 ; NoVLX-NEXT: kshiftlw $2, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $13, %eax, %xmm3, %xmm3 ; NoVLX-NEXT: vpinsrb $13, %eax, %xmm2, %xmm2
; NoVLX-NEXT: kshiftlw $1, %k0, %k1 ; NoVLX-NEXT: kshiftlw $1, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $14, %eax, %xmm3, %xmm3 ; NoVLX-NEXT: vpinsrb $14, %eax, %xmm2, %xmm2
; NoVLX-NEXT: kshiftrw $15, %k0, %k0 ; NoVLX-NEXT: kshiftrw $15, %k0, %k0
; NoVLX-NEXT: kmovw %k0, %eax ; NoVLX-NEXT: kmovw %k0, %eax
; NoVLX-NEXT: vpinsrb $15, %eax, %xmm3, %xmm3 ; NoVLX-NEXT: vpinsrb $15, %eax, %xmm2, %xmm2
; NoVLX-NEXT: vpcmpeqw 32(%rsi), %ymm4, %ymm4 ; NoVLX-NEXT: vpcmpeqw 32(%rsi), %ymm3, %ymm3
; NoVLX-NEXT: vpmovsxwd %ymm4, %zmm4 ; NoVLX-NEXT: vpmovsxwd %ymm3, %zmm3
; NoVLX-NEXT: vpslld $31, %zmm4, %zmm4 ; NoVLX-NEXT: vpslld $31, %zmm3, %zmm3
; NoVLX-NEXT: vptestmd %zmm4, %zmm4, %k0 ; NoVLX-NEXT: vptestmd %zmm3, %zmm3, %k0
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %ecx ; NoVLX-NEXT: kmovw %k1, %ecx
; NoVLX-NEXT: vmovd %ecx, %xmm4 ; NoVLX-NEXT: vmovd %ecx, %xmm3
; NoVLX-NEXT: vpinsrb $1, %eax, %xmm4, %xmm4 ; NoVLX-NEXT: vpinsrb $1, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $2, %eax, %xmm4, %xmm4 ; NoVLX-NEXT: vpinsrb $2, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftlw $12, %k0, %k1 ; NoVLX-NEXT: kshiftlw $12, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $3, %eax, %xmm4, %xmm4 ; NoVLX-NEXT: vpinsrb $3, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftlw $11, %k0, %k1 ; NoVLX-NEXT: kshiftlw $11, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $4, %eax, %xmm4, %xmm4 ; NoVLX-NEXT: vpinsrb $4, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftlw $10, %k0, %k1 ; NoVLX-NEXT: kshiftlw $10, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $5, %eax, %xmm4, %xmm4 ; NoVLX-NEXT: vpinsrb $5, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftlw $9, %k0, %k1 ; NoVLX-NEXT: kshiftlw $9, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $6, %eax, %xmm4, %xmm4 ; NoVLX-NEXT: vpinsrb $6, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftlw $8, %k0, %k1 ; NoVLX-NEXT: kshiftlw $8, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $7, %eax, %xmm4, %xmm4 ; NoVLX-NEXT: vpinsrb $7, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftlw $7, %k0, %k1 ; NoVLX-NEXT: kshiftlw $7, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $8, %eax, %xmm4, %xmm4 ; NoVLX-NEXT: vpinsrb $8, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftlw $6, %k0, %k1 ; NoVLX-NEXT: kshiftlw $6, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $9, %eax, %xmm4, %xmm4 ; NoVLX-NEXT: vpinsrb $9, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftlw $5, %k0, %k1 ; NoVLX-NEXT: kshiftlw $5, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $10, %eax, %xmm4, %xmm4 ; NoVLX-NEXT: vpinsrb $10, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftlw $4, %k0, %k1 ; NoVLX-NEXT: kshiftlw $4, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $11, %eax, %xmm4, %xmm4 ; NoVLX-NEXT: vpinsrb $11, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftlw $3, %k0, %k1 ; NoVLX-NEXT: kshiftlw $3, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $12, %eax, %xmm4, %xmm4 ; NoVLX-NEXT: vpinsrb $12, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftlw $2, %k0, %k1 ; NoVLX-NEXT: kshiftlw $2, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $13, %eax, %xmm4, %xmm4 ; NoVLX-NEXT: vpinsrb $13, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftlw $1, %k0, %k1 ; NoVLX-NEXT: kshiftlw $1, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $14, %eax, %xmm4, %xmm4 ; NoVLX-NEXT: vpinsrb $14, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftrw $15, %k0, %k0 ; NoVLX-NEXT: kshiftrw $15, %k0, %k0
; NoVLX-NEXT: kmovw %k0, %eax ; NoVLX-NEXT: kmovw %k0, %eax
; NoVLX-NEXT: vpinsrb $15, %eax, %xmm4, %xmm4 ; NoVLX-NEXT: vpinsrb $15, %eax, %xmm3, %xmm3
; NoVLX-NEXT: vpand %xmm2, %xmm4, %xmm2
; NoVLX-NEXT: vpmovsxbd %xmm2, %zmm2
; NoVLX-NEXT: vpslld $31, %zmm2, %zmm2
; NoVLX-NEXT: vptestmd %zmm2, %zmm2, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpand %xmm1, %xmm3, %xmm1 ; NoVLX-NEXT: vpand %xmm1, %xmm3, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpand %xmm0, %xmm2, %xmm0
; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT: kmovw %k0, (%rsp) ; NoVLX-NEXT: kmovw %k0, (%rsp)
; NoVLX-NEXT: movl (%rsp), %ecx ; NoVLX-NEXT: movl (%rsp), %ecx
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: movl {{[0-9]+}}(%rsp), %eax ; NoVLX-NEXT: movl {{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT: shlq $32, %rax ; NoVLX-NEXT: shlq $32, %rax
; NoVLX-NEXT: orq %rcx, %rax ; NoVLX-NEXT: orq %rcx, %rax
; NoVLX-NEXT: movq %rbp, %rsp ; NoVLX-NEXT: movq %rbp, %rsp
; NoVLX-NEXT: popq %rbp ; NoVLX-NEXT: popq %rbp
; NoVLX-NEXT: retq ; NoVLX-NEXT: retq
▲ Show 20 Lines • Show All 1,165 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi195: ; NoVLX-NEXT: .Lcfi195:
; NoVLX-NEXT: .cfi_offset %rbp, -16 ; NoVLX-NEXT: .cfi_offset %rbp, -16
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi196: ; NoVLX-NEXT: .Lcfi196:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: vpcmpeqd %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vpcmpeqd %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vpxor %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero ; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
Show All 32 Lines
; NoVLX-NEXT: .Lcfi198: ; NoVLX-NEXT: .Lcfi198:
; NoVLX-NEXT: .cfi_offset %rbp, -16 ; NoVLX-NEXT: .cfi_offset %rbp, -16
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi199: ; NoVLX-NEXT: .Lcfi199:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: vpcmpeqd (%rdi), %xmm0, %xmm0 ; NoVLX-NEXT: vpcmpeqd (%rdi), %xmm0, %xmm0
; NoVLX-NEXT: vxorps %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero ; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
▲ Show 20 Lines • Show All 52 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: kmovw %k3, %ecx ; NoVLX-NEXT: kmovw %k3, %ecx
; NoVLX-NEXT: vmovd %ecx, %xmm1 ; NoVLX-NEXT: vmovd %ecx, %xmm1
; NoVLX-NEXT: vpinsrb $4, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrb $4, %eax, %xmm1, %xmm1
; NoVLX-NEXT: kmovw %k2, %eax ; NoVLX-NEXT: kmovw %k2, %eax
; NoVLX-NEXT: vpinsrb $8, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $12, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT: vpand %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vpxor %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero ; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
▲ Show 20 Lines • Show All 54 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: kmovw %k3, %ecx ; NoVLX-NEXT: kmovw %k3, %ecx
; NoVLX-NEXT: vmovd %ecx, %xmm1 ; NoVLX-NEXT: vmovd %ecx, %xmm1
; NoVLX-NEXT: vpinsrb $4, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrb $4, %eax, %xmm1, %xmm1
; NoVLX-NEXT: kmovw %k2, %eax ; NoVLX-NEXT: kmovw %k2, %eax
; NoVLX-NEXT: vpinsrb $8, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $12, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT: vpand %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vpxor %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero ; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
Show All 38 Lines
; NoVLX-NEXT: .cfi_offset %rbp, -16 ; NoVLX-NEXT: .cfi_offset %rbp, -16
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi208: ; NoVLX-NEXT: .Lcfi208:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: vpbroadcastd (%rdi), %xmm1 ; NoVLX-NEXT: vpbroadcastd (%rdi), %xmm1
; NoVLX-NEXT: vpcmpeqd %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vpcmpeqd %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vpxor %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero ; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
▲ Show 20 Lines • Show All 54 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: kmovw %k3, %ecx ; NoVLX-NEXT: kmovw %k3, %ecx
; NoVLX-NEXT: vmovd %ecx, %xmm1 ; NoVLX-NEXT: vmovd %ecx, %xmm1
; NoVLX-NEXT: vpinsrb $4, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrb $4, %eax, %xmm1, %xmm1
; NoVLX-NEXT: kmovw %k2, %eax ; NoVLX-NEXT: kmovw %k2, %eax
; NoVLX-NEXT: vpinsrb $8, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $12, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT: vpand %xmm0, %xmm1, %xmm0 ; NoVLX-NEXT: vpand %xmm0, %xmm1, %xmm0
; NoVLX-NEXT: vpxor %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero ; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
▲ Show 20 Lines • Show All 229 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi214: ; NoVLX-NEXT: .Lcfi214:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $32, %rsp ; NoVLX-NEXT: subq $32, %rsp
; NoVLX-NEXT: # kill: %YMM1<def> %YMM1<kill> %ZMM1<def> ; NoVLX-NEXT: # kill: %YMM1<def> %YMM1<kill> %ZMM1<def>
; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def> ; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def>
; NoVLX-NEXT: vpcmpeqd %zmm1, %zmm0, %k0 ; NoVLX-NEXT: vpcmpeqd %zmm1, %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r9d ; NoVLX-NEXT: kmovw %k1, %r9d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 58 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi217: ; NoVLX-NEXT: .Lcfi217:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $32, %rsp ; NoVLX-NEXT: subq $32, %rsp
; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def> ; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def>
; NoVLX-NEXT: vmovdqa (%rdi), %ymm1 ; NoVLX-NEXT: vmovdqa (%rdi), %ymm1
; NoVLX-NEXT: vpcmpeqd %zmm1, %zmm0, %k0 ; NoVLX-NEXT: vpcmpeqd %zmm1, %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r9d ; NoVLX-NEXT: kmovw %k1, %r9d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 62 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $32, %rsp ; NoVLX-NEXT: subq $32, %rsp
; NoVLX-NEXT: # kill: %YMM1<def> %YMM1<kill> %ZMM1<def> ; NoVLX-NEXT: # kill: %YMM1<def> %YMM1<kill> %ZMM1<def>
; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def> ; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def>
; NoVLX-NEXT: vpcmpeqd %zmm1, %zmm0, %k0 ; NoVLX-NEXT: vpcmpeqd %zmm1, %zmm0, %k0
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: kandw %k1, %k0, %k0 ; NoVLX-NEXT: kandw %k1, %k0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r9d ; NoVLX-NEXT: kmovw %k1, %r9d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 63 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $32, %rsp ; NoVLX-NEXT: subq $32, %rsp
; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def> ; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def>
; NoVLX-NEXT: vmovdqa (%rsi), %ymm1 ; NoVLX-NEXT: vmovdqa (%rsi), %ymm1
; NoVLX-NEXT: vpcmpeqd %zmm1, %zmm0, %k0 ; NoVLX-NEXT: vpcmpeqd %zmm1, %zmm0, %k0
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: kandw %k1, %k0, %k0 ; NoVLX-NEXT: kandw %k1, %k0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r9d ; NoVLX-NEXT: kmovw %k1, %r9d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 62 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi226: ; NoVLX-NEXT: .Lcfi226:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $32, %rsp ; NoVLX-NEXT: subq $32, %rsp
; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def> ; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def>
; NoVLX-NEXT: vpbroadcastd (%rdi), %ymm1 ; NoVLX-NEXT: vpbroadcastd (%rdi), %ymm1
; NoVLX-NEXT: vpcmpeqd %zmm1, %zmm0, %k0 ; NoVLX-NEXT: vpcmpeqd %zmm1, %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r9d ; NoVLX-NEXT: kmovw %k1, %r9d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 63 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $32, %rsp ; NoVLX-NEXT: subq $32, %rsp
; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def> ; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def>
; NoVLX-NEXT: vpbroadcastd (%rsi), %ymm1 ; NoVLX-NEXT: vpbroadcastd (%rsi), %ymm1
; NoVLX-NEXT: vpcmpeqd %zmm1, %zmm0, %k0 ; NoVLX-NEXT: vpcmpeqd %zmm1, %zmm0, %k0
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: kandw %k0, %k1, %k0 ; NoVLX-NEXT: kandw %k0, %k1, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r9d ; NoVLX-NEXT: kmovw %k1, %r9d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 63 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi232: ; NoVLX-NEXT: .Lcfi232:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: # kill: %YMM1<def> %YMM1<kill> %ZMM1<def> ; NoVLX-NEXT: # kill: %YMM1<def> %YMM1<kill> %ZMM1<def>
; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def> ; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def>
; NoVLX-NEXT: vpcmpeqd %zmm1, %zmm0, %k0 ; NoVLX-NEXT: vpcmpeqd %zmm1, %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 63 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi235: ; NoVLX-NEXT: .Lcfi235:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def> ; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def>
; NoVLX-NEXT: vmovdqa (%rdi), %ymm1 ; NoVLX-NEXT: vmovdqa (%rdi), %ymm1
; NoVLX-NEXT: vpcmpeqd %zmm1, %zmm0, %k0 ; NoVLX-NEXT: vpcmpeqd %zmm1, %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 67 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: # kill: %YMM1<def> %YMM1<kill> %ZMM1<def> ; NoVLX-NEXT: # kill: %YMM1<def> %YMM1<kill> %ZMM1<def>
; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def> ; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def>
; NoVLX-NEXT: vpcmpeqd %zmm1, %zmm0, %k0 ; NoVLX-NEXT: vpcmpeqd %zmm1, %zmm0, %k0
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: kandw %k1, %k0, %k0 ; NoVLX-NEXT: kandw %k1, %k0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 68 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def> ; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def>
; NoVLX-NEXT: vmovdqa (%rsi), %ymm1 ; NoVLX-NEXT: vmovdqa (%rsi), %ymm1
; NoVLX-NEXT: vpcmpeqd %zmm1, %zmm0, %k0 ; NoVLX-NEXT: vpcmpeqd %zmm1, %zmm0, %k0
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: kandw %k1, %k0, %k0 ; NoVLX-NEXT: kandw %k1, %k0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 67 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi244: ; NoVLX-NEXT: .Lcfi244:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def> ; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def>
; NoVLX-NEXT: vpbroadcastd (%rdi), %ymm1 ; NoVLX-NEXT: vpbroadcastd (%rdi), %ymm1
; NoVLX-NEXT: vpcmpeqd %zmm1, %zmm0, %k0 ; NoVLX-NEXT: vpcmpeqd %zmm1, %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 68 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def> ; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def>
; NoVLX-NEXT: vpbroadcastd (%rsi), %ymm1 ; NoVLX-NEXT: vpbroadcastd (%rsi), %ymm1
; NoVLX-NEXT: vpcmpeqd %zmm1, %zmm0, %k0 ; NoVLX-NEXT: vpcmpeqd %zmm1, %zmm0, %k0
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: kandw %k0, %k1, %k0 ; NoVLX-NEXT: kandw %k0, %k1, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 81 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %r12, -48 ; NoVLX-NEXT: .cfi_offset %r12, -48
; NoVLX-NEXT: .Lcfi253: ; NoVLX-NEXT: .Lcfi253:
; NoVLX-NEXT: .cfi_offset %r13, -40 ; NoVLX-NEXT: .cfi_offset %r13, -40
; NoVLX-NEXT: .Lcfi254: ; NoVLX-NEXT: .Lcfi254:
; NoVLX-NEXT: .cfi_offset %r14, -32 ; NoVLX-NEXT: .cfi_offset %r14, -32
; NoVLX-NEXT: .Lcfi255: ; NoVLX-NEXT: .Lcfi255:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: vpcmpeqd %zmm1, %zmm0, %k0 ; NoVLX-NEXT: vpcmpeqd %zmm1, %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r10d ; NoVLX-NEXT: kmovw %k1, %r10d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 106 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %r12, -48 ; NoVLX-NEXT: .cfi_offset %r12, -48
; NoVLX-NEXT: .Lcfi261: ; NoVLX-NEXT: .Lcfi261:
; NoVLX-NEXT: .cfi_offset %r13, -40 ; NoVLX-NEXT: .cfi_offset %r13, -40
; NoVLX-NEXT: .Lcfi262: ; NoVLX-NEXT: .Lcfi262:
; NoVLX-NEXT: .cfi_offset %r14, -32 ; NoVLX-NEXT: .cfi_offset %r14, -32
; NoVLX-NEXT: .Lcfi263: ; NoVLX-NEXT: .Lcfi263:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: vpcmpeqd (%rdi), %zmm0, %k0 ; NoVLX-NEXT: vpcmpeqd (%rdi), %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r10d ; NoVLX-NEXT: kmovw %k1, %r10d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 109 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi269: ; NoVLX-NEXT: .Lcfi269:
; NoVLX-NEXT: .cfi_offset %r13, -40 ; NoVLX-NEXT: .cfi_offset %r13, -40
; NoVLX-NEXT: .Lcfi270: ; NoVLX-NEXT: .Lcfi270:
; NoVLX-NEXT: .cfi_offset %r14, -32 ; NoVLX-NEXT: .cfi_offset %r14, -32
; NoVLX-NEXT: .Lcfi271: ; NoVLX-NEXT: .Lcfi271:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vpcmpeqd %zmm1, %zmm0, %k0 {%k1} ; NoVLX-NEXT: vpcmpeqd %zmm1, %zmm0, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r10d ; NoVLX-NEXT: kmovw %k1, %r10d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 110 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi277: ; NoVLX-NEXT: .Lcfi277:
; NoVLX-NEXT: .cfi_offset %r13, -40 ; NoVLX-NEXT: .cfi_offset %r13, -40
; NoVLX-NEXT: .Lcfi278: ; NoVLX-NEXT: .Lcfi278:
; NoVLX-NEXT: .cfi_offset %r14, -32 ; NoVLX-NEXT: .cfi_offset %r14, -32
; NoVLX-NEXT: .Lcfi279: ; NoVLX-NEXT: .Lcfi279:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vpcmpeqd (%rsi), %zmm0, %k0 {%k1} ; NoVLX-NEXT: vpcmpeqd (%rsi), %zmm0, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r10d ; NoVLX-NEXT: kmovw %k1, %r10d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 110 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %r12, -48 ; NoVLX-NEXT: .cfi_offset %r12, -48
; NoVLX-NEXT: .Lcfi285: ; NoVLX-NEXT: .Lcfi285:
; NoVLX-NEXT: .cfi_offset %r13, -40 ; NoVLX-NEXT: .cfi_offset %r13, -40
; NoVLX-NEXT: .Lcfi286: ; NoVLX-NEXT: .Lcfi286:
; NoVLX-NEXT: .cfi_offset %r14, -32 ; NoVLX-NEXT: .cfi_offset %r14, -32
; NoVLX-NEXT: .Lcfi287: ; NoVLX-NEXT: .Lcfi287:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: vpcmpeqd (%rdi){1to16}, %zmm0, %k0 ; NoVLX-NEXT: vpcmpeqd (%rdi){1to16}, %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r10d ; NoVLX-NEXT: kmovw %k1, %r10d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 110 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi293: ; NoVLX-NEXT: .Lcfi293:
; NoVLX-NEXT: .cfi_offset %r13, -40 ; NoVLX-NEXT: .cfi_offset %r13, -40
; NoVLX-NEXT: .Lcfi294: ; NoVLX-NEXT: .Lcfi294:
; NoVLX-NEXT: .cfi_offset %r14, -32 ; NoVLX-NEXT: .cfi_offset %r14, -32
; NoVLX-NEXT: .Lcfi295: ; NoVLX-NEXT: .Lcfi295:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vpcmpeqd (%rsi){1to16}, %zmm0, %k0 {%k1} ; NoVLX-NEXT: vpcmpeqd (%rsi){1to16}, %zmm0, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r10d ; NoVLX-NEXT: kmovw %k1, %r10d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 111 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %r12, -48 ; NoVLX-NEXT: .cfi_offset %r12, -48
; NoVLX-NEXT: .Lcfi301: ; NoVLX-NEXT: .Lcfi301:
; NoVLX-NEXT: .cfi_offset %r13, -40 ; NoVLX-NEXT: .cfi_offset %r13, -40
; NoVLX-NEXT: .Lcfi302: ; NoVLX-NEXT: .Lcfi302:
; NoVLX-NEXT: .cfi_offset %r14, -32 ; NoVLX-NEXT: .cfi_offset %r14, -32
; NoVLX-NEXT: .Lcfi303: ; NoVLX-NEXT: .Lcfi303:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: vpcmpeqd %zmm1, %zmm0, %k0 ; NoVLX-NEXT: vpcmpeqd %zmm1, %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 111 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %r12, -48 ; NoVLX-NEXT: .cfi_offset %r12, -48
; NoVLX-NEXT: .Lcfi309: ; NoVLX-NEXT: .Lcfi309:
; NoVLX-NEXT: .cfi_offset %r13, -40 ; NoVLX-NEXT: .cfi_offset %r13, -40
; NoVLX-NEXT: .Lcfi310: ; NoVLX-NEXT: .Lcfi310:
; NoVLX-NEXT: .cfi_offset %r14, -32 ; NoVLX-NEXT: .cfi_offset %r14, -32
; NoVLX-NEXT: .Lcfi311: ; NoVLX-NEXT: .Lcfi311:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: vpcmpeqd (%rdi), %zmm0, %k0 ; NoVLX-NEXT: vpcmpeqd (%rdi), %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 114 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi317: ; NoVLX-NEXT: .Lcfi317:
; NoVLX-NEXT: .cfi_offset %r13, -40 ; NoVLX-NEXT: .cfi_offset %r13, -40
; NoVLX-NEXT: .Lcfi318: ; NoVLX-NEXT: .Lcfi318:
; NoVLX-NEXT: .cfi_offset %r14, -32 ; NoVLX-NEXT: .cfi_offset %r14, -32
; NoVLX-NEXT: .Lcfi319: ; NoVLX-NEXT: .Lcfi319:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vpcmpeqd %zmm1, %zmm0, %k0 {%k1} ; NoVLX-NEXT: vpcmpeqd %zmm1, %zmm0, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 115 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi325: ; NoVLX-NEXT: .Lcfi325:
; NoVLX-NEXT: .cfi_offset %r13, -40 ; NoVLX-NEXT: .cfi_offset %r13, -40
; NoVLX-NEXT: .Lcfi326: ; NoVLX-NEXT: .Lcfi326:
; NoVLX-NEXT: .cfi_offset %r14, -32 ; NoVLX-NEXT: .cfi_offset %r14, -32
; NoVLX-NEXT: .Lcfi327: ; NoVLX-NEXT: .Lcfi327:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vpcmpeqd (%rsi), %zmm0, %k0 {%k1} ; NoVLX-NEXT: vpcmpeqd (%rsi), %zmm0, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 115 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %r12, -48 ; NoVLX-NEXT: .cfi_offset %r12, -48
; NoVLX-NEXT: .Lcfi333: ; NoVLX-NEXT: .Lcfi333:
; NoVLX-NEXT: .cfi_offset %r13, -40 ; NoVLX-NEXT: .cfi_offset %r13, -40
; NoVLX-NEXT: .Lcfi334: ; NoVLX-NEXT: .Lcfi334:
; NoVLX-NEXT: .cfi_offset %r14, -32 ; NoVLX-NEXT: .cfi_offset %r14, -32
; NoVLX-NEXT: .Lcfi335: ; NoVLX-NEXT: .Lcfi335:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: vpcmpeqd (%rdi){1to16}, %zmm0, %k0 ; NoVLX-NEXT: vpcmpeqd (%rdi){1to16}, %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 115 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi341: ; NoVLX-NEXT: .Lcfi341:
; NoVLX-NEXT: .cfi_offset %r13, -40 ; NoVLX-NEXT: .cfi_offset %r13, -40
; NoVLX-NEXT: .Lcfi342: ; NoVLX-NEXT: .Lcfi342:
; NoVLX-NEXT: .cfi_offset %r14, -32 ; NoVLX-NEXT: .cfi_offset %r14, -32
; NoVLX-NEXT: .Lcfi343: ; NoVLX-NEXT: .Lcfi343:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vpcmpeqd (%rsi){1to16}, %zmm0, %k0 {%k1} ; NoVLX-NEXT: vpcmpeqd (%rsi){1to16}, %zmm0, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 1,190 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi363: ; NoVLX-NEXT: .Lcfi363:
; NoVLX-NEXT: .cfi_offset %rbp, -16 ; NoVLX-NEXT: .cfi_offset %rbp, -16
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi364: ; NoVLX-NEXT: .Lcfi364:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: vpcmpeqq %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vpcmpeqq %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vpxor %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero ; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
Show All 32 Lines
; NoVLX-NEXT: .Lcfi366: ; NoVLX-NEXT: .Lcfi366:
; NoVLX-NEXT: .cfi_offset %rbp, -16 ; NoVLX-NEXT: .cfi_offset %rbp, -16
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi367: ; NoVLX-NEXT: .Lcfi367:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: vpcmpeqq (%rdi), %xmm0, %xmm0 ; NoVLX-NEXT: vpcmpeqq (%rdi), %xmm0, %xmm0
; NoVLX-NEXT: vxorps %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero ; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
▲ Show 20 Lines • Show All 44 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kshiftlw $14, %k0, %k0 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
; NoVLX-NEXT: kshiftrw $15, %k0, %k0 ; NoVLX-NEXT: kshiftrw $15, %k0, %k0
; NoVLX-NEXT: kmovw %k0, %eax ; NoVLX-NEXT: kmovw %k0, %eax
; NoVLX-NEXT: kmovw %k1, %ecx ; NoVLX-NEXT: kmovw %k1, %ecx
; NoVLX-NEXT: vmovd %ecx, %xmm1 ; NoVLX-NEXT: vmovd %ecx, %xmm1
; NoVLX-NEXT: vpinsrb $8, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT: vpand %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vpxor %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero ; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
▲ Show 20 Lines • Show All 46 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kshiftlw $14, %k0, %k0 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
; NoVLX-NEXT: kshiftrw $15, %k0, %k0 ; NoVLX-NEXT: kshiftrw $15, %k0, %k0
; NoVLX-NEXT: kmovw %k0, %eax ; NoVLX-NEXT: kmovw %k0, %eax
; NoVLX-NEXT: kmovw %k1, %ecx ; NoVLX-NEXT: kmovw %k1, %ecx
; NoVLX-NEXT: vmovd %ecx, %xmm1 ; NoVLX-NEXT: vmovd %ecx, %xmm1
; NoVLX-NEXT: vpinsrb $8, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT: vpand %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vpxor %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero ; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
Show All 38 Lines
; NoVLX-NEXT: .cfi_offset %rbp, -16 ; NoVLX-NEXT: .cfi_offset %rbp, -16
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi376: ; NoVLX-NEXT: .Lcfi376:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: vpbroadcastq (%rdi), %xmm1 ; NoVLX-NEXT: vpbroadcastq (%rdi), %xmm1
; NoVLX-NEXT: vpcmpeqq %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vpcmpeqq %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vpxor %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero ; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
▲ Show 20 Lines • Show All 46 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kshiftlw $14, %k0, %k0 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
; NoVLX-NEXT: kshiftrw $15, %k0, %k0 ; NoVLX-NEXT: kshiftrw $15, %k0, %k0
; NoVLX-NEXT: kmovw %k0, %eax ; NoVLX-NEXT: kmovw %k0, %eax
; NoVLX-NEXT: kmovw %k1, %ecx ; NoVLX-NEXT: kmovw %k1, %ecx
; NoVLX-NEXT: vmovd %ecx, %xmm1 ; NoVLX-NEXT: vmovd %ecx, %xmm1
; NoVLX-NEXT: vpinsrb $8, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT: vpand %xmm0, %xmm1, %xmm0 ; NoVLX-NEXT: vpand %xmm0, %xmm1, %xmm0
; NoVLX-NEXT: vpxor %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero ; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
▲ Show 20 Lines • Show All 1,216 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %rbp, -16 ; NoVLX-NEXT: .cfi_offset %rbp, -16
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi400: ; NoVLX-NEXT: .Lcfi400:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: vpcmpeqq %ymm1, %ymm0, %ymm0 ; NoVLX-NEXT: vpcmpeqq %ymm1, %ymm0, %ymm0
; NoVLX-NEXT: vpmovqd %zmm0, %ymm0 ; NoVLX-NEXT: vpmovqd %zmm0, %ymm0
; NoVLX-NEXT: vpxor %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero ; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
Show All 34 Lines
; NoVLX-NEXT: .cfi_offset %rbp, -16 ; NoVLX-NEXT: .cfi_offset %rbp, -16
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi403: ; NoVLX-NEXT: .Lcfi403:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: vpcmpeqq (%rdi), %ymm0, %ymm0 ; NoVLX-NEXT: vpcmpeqq (%rdi), %ymm0, %ymm0
; NoVLX-NEXT: vpmovqd %zmm0, %ymm0 ; NoVLX-NEXT: vpmovqd %zmm0, %ymm0
; NoVLX-NEXT: vxorps %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero ; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
▲ Show 20 Lines • Show All 54 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: kmovw %k3, %ecx ; NoVLX-NEXT: kmovw %k3, %ecx
; NoVLX-NEXT: vmovd %ecx, %xmm1 ; NoVLX-NEXT: vmovd %ecx, %xmm1
; NoVLX-NEXT: vpinsrb $4, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrb $4, %eax, %xmm1, %xmm1
; NoVLX-NEXT: kmovw %k2, %eax ; NoVLX-NEXT: kmovw %k2, %eax
; NoVLX-NEXT: vpinsrb $8, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $12, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT: vpand %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vpxor %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero ; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
▲ Show 20 Lines • Show All 56 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: kmovw %k3, %ecx ; NoVLX-NEXT: kmovw %k3, %ecx
; NoVLX-NEXT: vmovd %ecx, %xmm1 ; NoVLX-NEXT: vmovd %ecx, %xmm1
; NoVLX-NEXT: vpinsrb $4, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrb $4, %eax, %xmm1, %xmm1
; NoVLX-NEXT: kmovw %k2, %eax ; NoVLX-NEXT: kmovw %k2, %eax
; NoVLX-NEXT: vpinsrb $8, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $12, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT: vpand %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vpxor %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero ; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
Show All 40 Lines
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi412: ; NoVLX-NEXT: .Lcfi412:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: vpbroadcastq (%rdi), %ymm1 ; NoVLX-NEXT: vpbroadcastq (%rdi), %ymm1
; NoVLX-NEXT: vpcmpeqq %ymm1, %ymm0, %ymm0 ; NoVLX-NEXT: vpcmpeqq %ymm1, %ymm0, %ymm0
; NoVLX-NEXT: vpmovqd %zmm0, %ymm0 ; NoVLX-NEXT: vpmovqd %zmm0, %ymm0
; NoVLX-NEXT: vpxor %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero ; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
▲ Show 20 Lines • Show All 56 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: kmovw %k3, %ecx ; NoVLX-NEXT: kmovw %k3, %ecx
; NoVLX-NEXT: vmovd %ecx, %xmm1 ; NoVLX-NEXT: vmovd %ecx, %xmm1
; NoVLX-NEXT: vpinsrb $4, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrb $4, %eax, %xmm1, %xmm1
; NoVLX-NEXT: kmovw %k2, %eax ; NoVLX-NEXT: kmovw %k2, %eax
; NoVLX-NEXT: vpinsrb $8, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $12, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT: vpand %xmm0, %xmm1, %xmm0 ; NoVLX-NEXT: vpand %xmm0, %xmm1, %xmm0
; NoVLX-NEXT: vpxor %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero ; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
▲ Show 20 Lines • Show All 203 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi417: ; NoVLX-NEXT: .Lcfi417:
; NoVLX-NEXT: .cfi_offset %rbp, -16 ; NoVLX-NEXT: .cfi_offset %rbp, -16
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi418: ; NoVLX-NEXT: .Lcfi418:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $32, %rsp ; NoVLX-NEXT: subq $32, %rsp
; NoVLX-NEXT: vpcmpeqq %zmm1, %zmm0, %k0 ; NoVLX-NEXT: vpcmpeqq %zmm1, %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r9d ; NoVLX-NEXT: kmovw %k1, %r9d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 56 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi420: ; NoVLX-NEXT: .Lcfi420:
; NoVLX-NEXT: .cfi_offset %rbp, -16 ; NoVLX-NEXT: .cfi_offset %rbp, -16
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi421: ; NoVLX-NEXT: .Lcfi421:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $32, %rsp ; NoVLX-NEXT: subq $32, %rsp
; NoVLX-NEXT: vpcmpeqq (%rdi), %zmm0, %k0 ; NoVLX-NEXT: vpcmpeqq (%rdi), %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r9d ; NoVLX-NEXT: kmovw %k1, %r9d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 59 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %rbp, -16 ; NoVLX-NEXT: .cfi_offset %rbp, -16
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi424: ; NoVLX-NEXT: .Lcfi424:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $32, %rsp ; NoVLX-NEXT: subq $32, %rsp
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vpcmpeqq %zmm1, %zmm0, %k0 {%k1} ; NoVLX-NEXT: vpcmpeqq %zmm1, %zmm0, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r9d ; NoVLX-NEXT: kmovw %k1, %r9d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 60 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %rbp, -16 ; NoVLX-NEXT: .cfi_offset %rbp, -16
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi427: ; NoVLX-NEXT: .Lcfi427:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $32, %rsp ; NoVLX-NEXT: subq $32, %rsp
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vpcmpeqq (%rsi), %zmm0, %k0 {%k1} ; NoVLX-NEXT: vpcmpeqq (%rsi), %zmm0, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r9d ; NoVLX-NEXT: kmovw %k1, %r9d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 60 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi429: ; NoVLX-NEXT: .Lcfi429:
; NoVLX-NEXT: .cfi_offset %rbp, -16 ; NoVLX-NEXT: .cfi_offset %rbp, -16
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi430: ; NoVLX-NEXT: .Lcfi430:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $32, %rsp ; NoVLX-NEXT: subq $32, %rsp
; NoVLX-NEXT: vpcmpeqq (%rdi){1to8}, %zmm0, %k0 ; NoVLX-NEXT: vpcmpeqq (%rdi){1to8}, %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r9d ; NoVLX-NEXT: kmovw %k1, %r9d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 60 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %rbp, -16 ; NoVLX-NEXT: .cfi_offset %rbp, -16
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi433: ; NoVLX-NEXT: .Lcfi433:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $32, %rsp ; NoVLX-NEXT: subq $32, %rsp
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vpcmpeqq (%rsi){1to8}, %zmm0, %k0 {%k1} ; NoVLX-NEXT: vpcmpeqq (%rsi){1to8}, %zmm0, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r9d ; NoVLX-NEXT: kmovw %k1, %r9d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 61 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi435: ; NoVLX-NEXT: .Lcfi435:
; NoVLX-NEXT: .cfi_offset %rbp, -16 ; NoVLX-NEXT: .cfi_offset %rbp, -16
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi436: ; NoVLX-NEXT: .Lcfi436:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: vpcmpeqq %zmm1, %zmm0, %k0 ; NoVLX-NEXT: vpcmpeqq %zmm1, %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 61 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi438: ; NoVLX-NEXT: .Lcfi438:
; NoVLX-NEXT: .cfi_offset %rbp, -16 ; NoVLX-NEXT: .cfi_offset %rbp, -16
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi439: ; NoVLX-NEXT: .Lcfi439:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: vpcmpeqq (%rdi), %zmm0, %k0 ; NoVLX-NEXT: vpcmpeqq (%rdi), %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 64 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %rbp, -16 ; NoVLX-NEXT: .cfi_offset %rbp, -16
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi442: ; NoVLX-NEXT: .Lcfi442:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vpcmpeqq %zmm1, %zmm0, %k0 {%k1} ; NoVLX-NEXT: vpcmpeqq %zmm1, %zmm0, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 65 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %rbp, -16 ; NoVLX-NEXT: .cfi_offset %rbp, -16
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi445: ; NoVLX-NEXT: .Lcfi445:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vpcmpeqq (%rsi), %zmm0, %k0 {%k1} ; NoVLX-NEXT: vpcmpeqq (%rsi), %zmm0, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 65 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi447: ; NoVLX-NEXT: .Lcfi447:
; NoVLX-NEXT: .cfi_offset %rbp, -16 ; NoVLX-NEXT: .cfi_offset %rbp, -16
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi448: ; NoVLX-NEXT: .Lcfi448:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: vpcmpeqq (%rdi){1to8}, %zmm0, %k0 ; NoVLX-NEXT: vpcmpeqq (%rdi){1to8}, %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 65 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %rbp, -16 ; NoVLX-NEXT: .cfi_offset %rbp, -16
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi451: ; NoVLX-NEXT: .Lcfi451:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vpcmpeqq (%rsi){1to8}, %zmm0, %k0 {%k1} ; NoVLX-NEXT: vpcmpeqq (%rsi){1to8}, %zmm0, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 83 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi458: ; NoVLX-NEXT: .Lcfi458:
; NoVLX-NEXT: .cfi_offset %r14, -32 ; NoVLX-NEXT: .cfi_offset %r14, -32
; NoVLX-NEXT: .Lcfi459: ; NoVLX-NEXT: .Lcfi459:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: vpcmpgtb %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vpcmpgtb %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0 ; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r10d ; NoVLX-NEXT: kmovw %k1, %r10d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 108 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi466: ; NoVLX-NEXT: .Lcfi466:
; NoVLX-NEXT: .cfi_offset %r14, -32 ; NoVLX-NEXT: .cfi_offset %r14, -32
; NoVLX-NEXT: .Lcfi467: ; NoVLX-NEXT: .Lcfi467:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: vpcmpgtb (%rdi), %xmm0, %xmm0 ; NoVLX-NEXT: vpcmpgtb (%rdi), %xmm0, %xmm0
; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0 ; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r10d ; NoVLX-NEXT: kmovw %k1, %r10d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 111 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %r14, -32 ; NoVLX-NEXT: .cfi_offset %r14, -32
; NoVLX-NEXT: .Lcfi475: ; NoVLX-NEXT: .Lcfi475:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: vpcmpgtb %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vpcmpgtb %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0 ; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 {%k1} ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r10d ; NoVLX-NEXT: kmovw %k1, %r10d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 112 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %r14, -32 ; NoVLX-NEXT: .cfi_offset %r14, -32
; NoVLX-NEXT: .Lcfi483: ; NoVLX-NEXT: .Lcfi483:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: vpcmpgtb (%rsi), %xmm0, %xmm0 ; NoVLX-NEXT: vpcmpgtb (%rsi), %xmm0, %xmm0
; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0 ; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 {%k1} ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r10d ; NoVLX-NEXT: kmovw %k1, %r10d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 112 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi490: ; NoVLX-NEXT: .Lcfi490:
; NoVLX-NEXT: .cfi_offset %r14, -32 ; NoVLX-NEXT: .cfi_offset %r14, -32
; NoVLX-NEXT: .Lcfi491: ; NoVLX-NEXT: .Lcfi491:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: vpcmpgtb %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vpcmpgtb %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0 ; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 113 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi498: ; NoVLX-NEXT: .Lcfi498:
; NoVLX-NEXT: .cfi_offset %r14, -32 ; NoVLX-NEXT: .cfi_offset %r14, -32
; NoVLX-NEXT: .Lcfi499: ; NoVLX-NEXT: .Lcfi499:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: vpcmpgtb (%rdi), %xmm0, %xmm0 ; NoVLX-NEXT: vpcmpgtb (%rdi), %xmm0, %xmm0
; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0 ; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 116 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %r14, -32 ; NoVLX-NEXT: .cfi_offset %r14, -32
; NoVLX-NEXT: .Lcfi507: ; NoVLX-NEXT: .Lcfi507:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: vpcmpgtb %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vpcmpgtb %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0 ; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 {%k1} ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 117 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %r14, -32 ; NoVLX-NEXT: .cfi_offset %r14, -32
; NoVLX-NEXT: .Lcfi515: ; NoVLX-NEXT: .Lcfi515:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: vpcmpgtb (%rsi), %xmm0, %xmm0 ; NoVLX-NEXT: vpcmpgtb (%rsi), %xmm0, %xmm0
; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0 ; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 {%k1} ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 110 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0 ; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT: kmovw %k0, (%rsp) ; NoVLX-NEXT: kmovw %k0, (%rsp)
; NoVLX-NEXT: movl (%rsp), %ecx ; NoVLX-NEXT: movl (%rsp), %ecx
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: movl {{[0-9]+}}(%rsp), %eax ; NoVLX-NEXT: movl {{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT: shlq $32, %rax ; NoVLX-NEXT: shlq $32, %rax
; NoVLX-NEXT: orq %rcx, %rax ; NoVLX-NEXT: orq %rcx, %rax
; NoVLX-NEXT: movq %rbp, %rsp ; NoVLX-NEXT: movq %rbp, %rsp
; NoVLX-NEXT: popq %rbp ; NoVLX-NEXT: popq %rbp
; NoVLX-NEXT: retq ; NoVLX-NEXT: retq
Show All 32 Lines
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0 ; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT: kmovw %k0, (%rsp) ; NoVLX-NEXT: kmovw %k0, (%rsp)
; NoVLX-NEXT: movl (%rsp), %ecx ; NoVLX-NEXT: movl (%rsp), %ecx
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: movl {{[0-9]+}}(%rsp), %eax ; NoVLX-NEXT: movl {{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT: shlq $32, %rax ; NoVLX-NEXT: shlq $32, %rax
; NoVLX-NEXT: orq %rcx, %rax ; NoVLX-NEXT: orq %rcx, %rax
; NoVLX-NEXT: movq %rbp, %rsp ; NoVLX-NEXT: movq %rbp, %rsp
; NoVLX-NEXT: popq %rbp ; NoVLX-NEXT: popq %rbp
; NoVLX-NEXT: retq ; NoVLX-NEXT: retq
Show All 30 Lines
; NoVLX-NEXT: subq $96, %rsp ; NoVLX-NEXT: subq $96, %rsp
; NoVLX-NEXT: movl %edi, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: movl %edi, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw {{[0-9]+}}(%rsp), %k1 ; NoVLX-NEXT: kmovw {{[0-9]+}}(%rsp), %k1
; NoVLX-NEXT: kmovw {{[0-9]+}}(%rsp), %k2 ; NoVLX-NEXT: kmovw {{[0-9]+}}(%rsp), %k2
; NoVLX-NEXT: vpternlogd $255, %zmm2, %zmm2, %zmm2 {%k1} {z} ; NoVLX-NEXT: vpternlogd $255, %zmm2, %zmm2, %zmm2 {%k1} {z}
; NoVLX-NEXT: vpmovdb %zmm2, %xmm2 ; NoVLX-NEXT: vpmovdb %zmm2, %xmm2
; NoVLX-NEXT: vpternlogd $255, %zmm3, %zmm3, %zmm3 {%k2} {z} ; NoVLX-NEXT: vpternlogd $255, %zmm3, %zmm3, %zmm3 {%k2} {z}
; NoVLX-NEXT: vpmovdb %zmm3, %xmm3 ; NoVLX-NEXT: vpmovdb %zmm3, %xmm3
; NoVLX-NEXT: vxorps %xmm4, %xmm4, %xmm4
; NoVLX-NEXT: vpcmpgtb %ymm1, %ymm0, %ymm0 ; NoVLX-NEXT: vpcmpgtb %ymm1, %ymm0, %ymm0
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT: vpand %xmm3, %xmm1, %xmm1 ; NoVLX-NEXT: vpand %xmm3, %xmm1, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpand %xmm2, %xmm0, %xmm0 ; NoVLX-NEXT: vpand %xmm2, %xmm0, %xmm0
; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0 ; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT: kmovw %k0, (%rsp) ; NoVLX-NEXT: kmovw %k0, (%rsp)
; NoVLX-NEXT: movl (%rsp), %ecx ; NoVLX-NEXT: movl (%rsp), %ecx
; NoVLX-NEXT: vptestmd %zmm4, %zmm4, %k0 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: movl {{[0-9]+}}(%rsp), %eax ; NoVLX-NEXT: movl {{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT: shlq $32, %rax ; NoVLX-NEXT: shlq $32, %rax
; NoVLX-NEXT: orq %rcx, %rax ; NoVLX-NEXT: orq %rcx, %rax
; NoVLX-NEXT: movq %rbp, %rsp ; NoVLX-NEXT: movq %rbp, %rsp
; NoVLX-NEXT: popq %rbp ; NoVLX-NEXT: popq %rbp
; NoVLX-NEXT: retq ; NoVLX-NEXT: retq
Show All 31 Lines
; NoVLX-NEXT: subq $96, %rsp ; NoVLX-NEXT: subq $96, %rsp
; NoVLX-NEXT: movl %edi, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: movl %edi, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw {{[0-9]+}}(%rsp), %k1 ; NoVLX-NEXT: kmovw {{[0-9]+}}(%rsp), %k1
; NoVLX-NEXT: kmovw {{[0-9]+}}(%rsp), %k2 ; NoVLX-NEXT: kmovw {{[0-9]+}}(%rsp), %k2
; NoVLX-NEXT: vpternlogd $255, %zmm1, %zmm1, %zmm1 {%k1} {z} ; NoVLX-NEXT: vpternlogd $255, %zmm1, %zmm1, %zmm1 {%k1} {z}
; NoVLX-NEXT: vpmovdb %zmm1, %xmm1 ; NoVLX-NEXT: vpmovdb %zmm1, %xmm1
; NoVLX-NEXT: vpternlogd $255, %zmm2, %zmm2, %zmm2 {%k2} {z} ; NoVLX-NEXT: vpternlogd $255, %zmm2, %zmm2, %zmm2 {%k2} {z}
; NoVLX-NEXT: vpmovdb %zmm2, %xmm2 ; NoVLX-NEXT: vpmovdb %zmm2, %xmm2
; NoVLX-NEXT: vxorps %xmm3, %xmm3, %xmm3
; NoVLX-NEXT: vpcmpgtb (%rsi), %ymm0, %ymm0 ; NoVLX-NEXT: vpcmpgtb (%rsi), %ymm0, %ymm0
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm4 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm3
; NoVLX-NEXT: vpand %xmm2, %xmm4, %xmm2 ; NoVLX-NEXT: vpand %xmm2, %xmm3, %xmm2
; NoVLX-NEXT: vpmovsxbd %xmm2, %zmm2 ; NoVLX-NEXT: vpmovsxbd %xmm2, %zmm2
; NoVLX-NEXT: vpslld $31, %zmm2, %zmm2 ; NoVLX-NEXT: vpslld $31, %zmm2, %zmm2
; NoVLX-NEXT: vptestmd %zmm2, %zmm2, %k0 ; NoVLX-NEXT: vptestmd %zmm2, %zmm2, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpand %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0 ; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT: kmovw %k0, (%rsp) ; NoVLX-NEXT: kmovw %k0, (%rsp)
; NoVLX-NEXT: movl (%rsp), %ecx ; NoVLX-NEXT: movl (%rsp), %ecx
; NoVLX-NEXT: vptestmd %zmm3, %zmm3, %k0 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: movl {{[0-9]+}}(%rsp), %eax ; NoVLX-NEXT: movl {{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT: shlq $32, %rax ; NoVLX-NEXT: shlq $32, %rax
; NoVLX-NEXT: orq %rcx, %rax ; NoVLX-NEXT: orq %rcx, %rax
; NoVLX-NEXT: movq %rbp, %rsp ; NoVLX-NEXT: movq %rbp, %rsp
; NoVLX-NEXT: popq %rbp ; NoVLX-NEXT: popq %rbp
; NoVLX-NEXT: retq ; NoVLX-NEXT: retq
▲ Show 20 Lines • Show All 151 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi530: ; NoVLX-NEXT: .Lcfi530:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $32, %rsp ; NoVLX-NEXT: subq $32, %rsp
; NoVLX-NEXT: vpcmpgtw %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vpcmpgtw %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0 ; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0
; NoVLX-NEXT: vpsllq $63, %zmm0, %zmm0 ; NoVLX-NEXT: vpsllq $63, %zmm0, %zmm0
; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0 ; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r9d ; NoVLX-NEXT: kmovw %k1, %r9d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 58 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi533: ; NoVLX-NEXT: .Lcfi533:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $32, %rsp ; NoVLX-NEXT: subq $32, %rsp
; NoVLX-NEXT: vpcmpgtw (%rdi), %xmm0, %xmm0 ; NoVLX-NEXT: vpcmpgtw (%rdi), %xmm0, %xmm0
; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0 ; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0
; NoVLX-NEXT: vpsllq $63, %zmm0, %zmm0 ; NoVLX-NEXT: vpsllq $63, %zmm0, %zmm0
; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0 ; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r9d ; NoVLX-NEXT: kmovw %k1, %r9d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 61 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $32, %rsp ; NoVLX-NEXT: subq $32, %rsp
; NoVLX-NEXT: vpcmpgtw %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vpcmpgtw %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0 ; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0
; NoVLX-NEXT: vpsllq $63, %zmm0, %zmm0 ; NoVLX-NEXT: vpsllq $63, %zmm0, %zmm0
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0 {%k1} ; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r9d ; NoVLX-NEXT: kmovw %k1, %r9d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 62 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $32, %rsp ; NoVLX-NEXT: subq $32, %rsp
; NoVLX-NEXT: vpcmpgtw (%rsi), %xmm0, %xmm0 ; NoVLX-NEXT: vpcmpgtw (%rsi), %xmm0, %xmm0
; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0 ; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0
; NoVLX-NEXT: vpsllq $63, %zmm0, %zmm0 ; NoVLX-NEXT: vpsllq $63, %zmm0, %zmm0
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0 {%k1} ; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r9d ; NoVLX-NEXT: kmovw %k1, %r9d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 62 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi542: ; NoVLX-NEXT: .Lcfi542:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: vpcmpgtw %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vpcmpgtw %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0 ; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0
; NoVLX-NEXT: vpsllq $63, %zmm0, %zmm0 ; NoVLX-NEXT: vpsllq $63, %zmm0, %zmm0
; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0 ; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 63 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi545: ; NoVLX-NEXT: .Lcfi545:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: vpcmpgtw (%rdi), %xmm0, %xmm0 ; NoVLX-NEXT: vpcmpgtw (%rdi), %xmm0, %xmm0
; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0 ; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0
; NoVLX-NEXT: vpsllq $63, %zmm0, %zmm0 ; NoVLX-NEXT: vpsllq $63, %zmm0, %zmm0
; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0 ; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 66 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: vpcmpgtw %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vpcmpgtw %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0 ; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0
; NoVLX-NEXT: vpsllq $63, %zmm0, %zmm0 ; NoVLX-NEXT: vpsllq $63, %zmm0, %zmm0
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0 {%k1} ; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 67 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: vpcmpgtw (%rsi), %xmm0, %xmm0 ; NoVLX-NEXT: vpcmpgtw (%rsi), %xmm0, %xmm0
; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0 ; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0
; NoVLX-NEXT: vpsllq $63, %zmm0, %zmm0 ; NoVLX-NEXT: vpsllq $63, %zmm0, %zmm0
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0 {%k1} ; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 83 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi558: ; NoVLX-NEXT: .Lcfi558:
; NoVLX-NEXT: .cfi_offset %r14, -32 ; NoVLX-NEXT: .cfi_offset %r14, -32
; NoVLX-NEXT: .Lcfi559: ; NoVLX-NEXT: .Lcfi559:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: vpcmpgtw %ymm1, %ymm0, %ymm0 ; NoVLX-NEXT: vpcmpgtw %ymm1, %ymm0, %ymm0
; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0 ; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0 ; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r10d ; NoVLX-NEXT: kmovw %k1, %r10d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 109 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi566: ; NoVLX-NEXT: .Lcfi566:
; NoVLX-NEXT: .cfi_offset %r14, -32 ; NoVLX-NEXT: .cfi_offset %r14, -32
; NoVLX-NEXT: .Lcfi567: ; NoVLX-NEXT: .Lcfi567:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: vpcmpgtw (%rdi), %ymm0, %ymm0 ; NoVLX-NEXT: vpcmpgtw (%rdi), %ymm0, %ymm0
; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0 ; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0 ; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r10d ; NoVLX-NEXT: kmovw %k1, %r10d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 112 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %r14, -32 ; NoVLX-NEXT: .cfi_offset %r14, -32
; NoVLX-NEXT: .Lcfi575: ; NoVLX-NEXT: .Lcfi575:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: vpcmpgtw %ymm1, %ymm0, %ymm0 ; NoVLX-NEXT: vpcmpgtw %ymm1, %ymm0, %ymm0
; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0 ; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0 ; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 {%k1} ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r10d ; NoVLX-NEXT: kmovw %k1, %r10d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 113 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %r14, -32 ; NoVLX-NEXT: .cfi_offset %r14, -32
; NoVLX-NEXT: .Lcfi583: ; NoVLX-NEXT: .Lcfi583:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: vpcmpgtw (%rsi), %ymm0, %ymm0 ; NoVLX-NEXT: vpcmpgtw (%rsi), %ymm0, %ymm0
; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0 ; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0 ; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 {%k1} ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r10d ; NoVLX-NEXT: kmovw %k1, %r10d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 113 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi590: ; NoVLX-NEXT: .Lcfi590:
; NoVLX-NEXT: .cfi_offset %r14, -32 ; NoVLX-NEXT: .cfi_offset %r14, -32
; NoVLX-NEXT: .Lcfi591: ; NoVLX-NEXT: .Lcfi591:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: vpcmpgtw %ymm1, %ymm0, %ymm0 ; NoVLX-NEXT: vpcmpgtw %ymm1, %ymm0, %ymm0
; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0 ; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0 ; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 114 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi598: ; NoVLX-NEXT: .Lcfi598:
; NoVLX-NEXT: .cfi_offset %r14, -32 ; NoVLX-NEXT: .cfi_offset %r14, -32
; NoVLX-NEXT: .Lcfi599: ; NoVLX-NEXT: .Lcfi599:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: vpcmpgtw (%rdi), %ymm0, %ymm0 ; NoVLX-NEXT: vpcmpgtw (%rdi), %ymm0, %ymm0
; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0 ; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0 ; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 117 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %r14, -32 ; NoVLX-NEXT: .cfi_offset %r14, -32
; NoVLX-NEXT: .Lcfi607: ; NoVLX-NEXT: .Lcfi607:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: vpcmpgtw %ymm1, %ymm0, %ymm0 ; NoVLX-NEXT: vpcmpgtw %ymm1, %ymm0, %ymm0
; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0 ; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0 ; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 {%k1} ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 118 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %r14, -32 ; NoVLX-NEXT: .cfi_offset %r14, -32
; NoVLX-NEXT: .Lcfi615: ; NoVLX-NEXT: .Lcfi615:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: vpcmpgtw (%rsi), %ymm0, %ymm0 ; NoVLX-NEXT: vpcmpgtw (%rsi), %ymm0, %ymm0
; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0 ; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0 ; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 {%k1} ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 412 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: kshiftrw $15, %k0, %k0 ; NoVLX-NEXT: kshiftrw $15, %k0, %k0
; NoVLX-NEXT: kmovw %k0, %eax ; NoVLX-NEXT: kmovw %k0, %eax
; NoVLX-NEXT: vpinsrb $15, %eax, %xmm0, %xmm0 ; NoVLX-NEXT: vpinsrb $15, %eax, %xmm0, %xmm0
; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0 ; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT: kmovw %k0, (%rsp) ; NoVLX-NEXT: kmovw %k0, (%rsp)
; NoVLX-NEXT: movl (%rsp), %ecx ; NoVLX-NEXT: movl (%rsp), %ecx
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: movl {{[0-9]+}}(%rsp), %eax ; NoVLX-NEXT: movl {{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT: shlq $32, %rax ; NoVLX-NEXT: shlq $32, %rax
; NoVLX-NEXT: orq %rcx, %rax ; NoVLX-NEXT: orq %rcx, %rax
; NoVLX-NEXT: movq %rbp, %rsp ; NoVLX-NEXT: movq %rbp, %rsp
; NoVLX-NEXT: popq %rbp ; NoVLX-NEXT: popq %rbp
; NoVLX-NEXT: retq ; NoVLX-NEXT: retq
▲ Show 20 Lines • Show All 249 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: kshiftrw $15, %k0, %k0 ; NoVLX-NEXT: kshiftrw $15, %k0, %k0
; NoVLX-NEXT: kmovw %k0, %eax ; NoVLX-NEXT: kmovw %k0, %eax
; NoVLX-NEXT: vpinsrb $15, %eax, %xmm0, %xmm0 ; NoVLX-NEXT: vpinsrb $15, %eax, %xmm0, %xmm0
; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0 ; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT: kmovw %k0, (%rsp) ; NoVLX-NEXT: kmovw %k0, (%rsp)
; NoVLX-NEXT: movl (%rsp), %ecx ; NoVLX-NEXT: movl (%rsp), %ecx
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: movl {{[0-9]+}}(%rsp), %eax ; NoVLX-NEXT: movl {{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT: shlq $32, %rax ; NoVLX-NEXT: shlq $32, %rax
; NoVLX-NEXT: orq %rcx, %rax ; NoVLX-NEXT: orq %rcx, %rax
; NoVLX-NEXT: movq %rbp, %rsp ; NoVLX-NEXT: movq %rbp, %rsp
; NoVLX-NEXT: popq %rbp ; NoVLX-NEXT: popq %rbp
; NoVLX-NEXT: retq ; NoVLX-NEXT: retq
▲ Show 20 Lines • Show All 180 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: shrl $16, %eax ; NoVLX-NEXT: shrl $16, %eax
; NoVLX-NEXT: vmovd %ecx, %xmm2 ; NoVLX-NEXT: vmovd %ecx, %xmm2
; NoVLX-NEXT: vpinsrw $1, %eax, %xmm2, %xmm2 ; NoVLX-NEXT: vpinsrw $1, %eax, %xmm2, %xmm2
; NoVLX-NEXT: movq %rcx, %rax ; NoVLX-NEXT: movq %rcx, %rax
; NoVLX-NEXT: shrq $32, %rax ; NoVLX-NEXT: shrq $32, %rax
; NoVLX-NEXT: vpinsrw $2, %eax, %xmm2, %xmm2 ; NoVLX-NEXT: vpinsrw $2, %eax, %xmm2, %xmm2
; NoVLX-NEXT: vpextrq $1, %xmm1, %rax ; NoVLX-NEXT: vpextrq $1, %xmm1, %rax
; NoVLX-NEXT: vinserti128 $1, %xmm9, %ymm4, %ymm1 ; NoVLX-NEXT: vinserti128 $1, %xmm9, %ymm4, %ymm1
; NoVLX-NEXT: vinserti128 $1, %xmm6, %ymm0, %ymm8 ; NoVLX-NEXT: vinserti128 $1, %xmm6, %ymm0, %ymm4
; NoVLX-NEXT: kmovw {{[0-9]+}}(%rsp), %k1 ; NoVLX-NEXT: kmovw {{[0-9]+}}(%rsp), %k1
; NoVLX-NEXT: kmovw {{[0-9]+}}(%rsp), %k2 ; NoVLX-NEXT: kmovw {{[0-9]+}}(%rsp), %k2
; NoVLX-NEXT: vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT: vinserti128 $1, %xmm7, %ymm3, %ymm3 ; NoVLX-NEXT: vinserti128 $1, %xmm7, %ymm3, %ymm3
; NoVLX-NEXT: vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT: vpmovdb %zmm0, %xmm0 ; NoVLX-NEXT: vpmovdb %zmm0, %xmm0
; NoVLX-NEXT: vpternlogd $255, %zmm6, %zmm6, %zmm6 {%k2} {z} ; NoVLX-NEXT: vpcmpgtw %ymm3, %ymm1, %ymm3
; NoVLX-NEXT: vpcmpgtw %ymm3, %ymm1, %ymm4 ; NoVLX-NEXT: vpternlogd $255, %zmm1, %zmm1, %zmm1 {%k2} {z}
; NoVLX-NEXT: vpmovdb %zmm6, %xmm6 ; NoVLX-NEXT: vpmovdb %zmm1, %xmm1
; NoVLX-NEXT: shrq $48, %rcx ; NoVLX-NEXT: shrq $48, %rcx
; NoVLX-NEXT: vpinsrw $3, %ecx, %xmm2, %xmm1 ; NoVLX-NEXT: vpinsrw $3, %ecx, %xmm2, %xmm2
; NoVLX-NEXT: movl %eax, %ecx ; NoVLX-NEXT: movl %eax, %ecx
; NoVLX-NEXT: shrl $16, %ecx ; NoVLX-NEXT: shrl $16, %ecx
; NoVLX-NEXT: vpinsrw $4, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrw $4, %eax, %xmm2, %xmm2
; NoVLX-NEXT: vpinsrw $5, %ecx, %xmm1, %xmm2 ; NoVLX-NEXT: vpinsrw $5, %ecx, %xmm2, %xmm2
; NoVLX-NEXT: movq %rax, %rcx ; NoVLX-NEXT: movq %rax, %rcx
; NoVLX-NEXT: shrq $48, %rax ; NoVLX-NEXT: shrq $48, %rax
; NoVLX-NEXT: shrq $32, %rcx ; NoVLX-NEXT: shrq $32, %rcx
; NoVLX-NEXT: vpxor %xmm1, %xmm1, %xmm1
; NoVLX-NEXT: vpinsrw $6, %ecx, %xmm2, %xmm2 ; NoVLX-NEXT: vpinsrw $6, %ecx, %xmm2, %xmm2
; NoVLX-NEXT: vpinsrw $7, %eax, %xmm2, %xmm2 ; NoVLX-NEXT: vpinsrw $7, %eax, %xmm2, %xmm2
; NoVLX-NEXT: vinserti128 $1, %xmm5, %ymm2, %ymm2 ; NoVLX-NEXT: vinserti128 $1, %xmm5, %ymm2, %ymm2
; NoVLX-NEXT: vpcmpgtw %ymm2, %ymm8, %ymm2 ; NoVLX-NEXT: vpcmpgtw %ymm2, %ymm4, %ymm2
; NoVLX-NEXT: vpmovsxwd %ymm2, %zmm2 ; NoVLX-NEXT: vpmovsxwd %ymm2, %zmm2
; NoVLX-NEXT: vpslld $31, %zmm2, %zmm2 ; NoVLX-NEXT: vpslld $31, %zmm2, %zmm2
; NoVLX-NEXT: vptestmd %zmm2, %zmm2, %k0 ; NoVLX-NEXT: vptestmd %zmm2, %zmm2, %k0
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 49 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $13, %eax, %xmm2, %xmm2 ; NoVLX-NEXT: vpinsrb $13, %eax, %xmm2, %xmm2
; NoVLX-NEXT: kshiftlw $1, %k0, %k1 ; NoVLX-NEXT: kshiftlw $1, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $14, %eax, %xmm2, %xmm2 ; NoVLX-NEXT: vpinsrb $14, %eax, %xmm2, %xmm2
; NoVLX-NEXT: kshiftrw $15, %k0, %k0 ; NoVLX-NEXT: kshiftrw $15, %k0, %k0
; NoVLX-NEXT: kmovw %k0, %eax ; NoVLX-NEXT: kmovw %k0, %eax
; NoVLX-NEXT: vpinsrb $15, %eax, %xmm2, %xmm3 ; NoVLX-NEXT: vpinsrb $15, %eax, %xmm2, %xmm2
; NoVLX-NEXT: vpmovsxwd %ymm4, %zmm2 ; NoVLX-NEXT: vpmovsxwd %ymm3, %zmm3
; NoVLX-NEXT: vpslld $31, %zmm2, %zmm2 ; NoVLX-NEXT: vpslld $31, %zmm3, %zmm3
; NoVLX-NEXT: vptestmd %zmm2, %zmm2, %k0 ; NoVLX-NEXT: vptestmd %zmm3, %zmm3, %k0
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %ecx ; NoVLX-NEXT: kmovw %k1, %ecx
; NoVLX-NEXT: vmovd %ecx, %xmm2 ; NoVLX-NEXT: vmovd %ecx, %xmm3
; NoVLX-NEXT: vpinsrb $1, %eax, %xmm2, %xmm2 ; NoVLX-NEXT: vpinsrb $1, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $2, %eax, %xmm2, %xmm2 ; NoVLX-NEXT: vpinsrb $2, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftlw $12, %k0, %k1 ; NoVLX-NEXT: kshiftlw $12, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $3, %eax, %xmm2, %xmm2 ; NoVLX-NEXT: vpinsrb $3, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftlw $11, %k0, %k1 ; NoVLX-NEXT: kshiftlw $11, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $4, %eax, %xmm2, %xmm2 ; NoVLX-NEXT: vpinsrb $4, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftlw $10, %k0, %k1 ; NoVLX-NEXT: kshiftlw $10, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $5, %eax, %xmm2, %xmm2 ; NoVLX-NEXT: vpinsrb $5, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftlw $9, %k0, %k1 ; NoVLX-NEXT: kshiftlw $9, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $6, %eax, %xmm2, %xmm2 ; NoVLX-NEXT: vpinsrb $6, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftlw $8, %k0, %k1 ; NoVLX-NEXT: kshiftlw $8, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $7, %eax, %xmm2, %xmm2 ; NoVLX-NEXT: vpinsrb $7, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftlw $7, %k0, %k1 ; NoVLX-NEXT: kshiftlw $7, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $8, %eax, %xmm2, %xmm2 ; NoVLX-NEXT: vpinsrb $8, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftlw $6, %k0, %k1 ; NoVLX-NEXT: kshiftlw $6, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $9, %eax, %xmm2, %xmm2 ; NoVLX-NEXT: vpinsrb $9, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftlw $5, %k0, %k1 ; NoVLX-NEXT: kshiftlw $5, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $10, %eax, %xmm2, %xmm2 ; NoVLX-NEXT: vpinsrb $10, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftlw $4, %k0, %k1 ; NoVLX-NEXT: kshiftlw $4, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $11, %eax, %xmm2, %xmm2 ; NoVLX-NEXT: vpinsrb $11, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftlw $3, %k0, %k1 ; NoVLX-NEXT: kshiftlw $3, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $12, %eax, %xmm2, %xmm2 ; NoVLX-NEXT: vpinsrb $12, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftlw $2, %k0, %k1 ; NoVLX-NEXT: kshiftlw $2, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $13, %eax, %xmm2, %xmm2 ; NoVLX-NEXT: vpinsrb $13, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftlw $1, %k0, %k1 ; NoVLX-NEXT: kshiftlw $1, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $14, %eax, %xmm2, %xmm2 ; NoVLX-NEXT: vpinsrb $14, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftrw $15, %k0, %k0 ; NoVLX-NEXT: kshiftrw $15, %k0, %k0
; NoVLX-NEXT: kmovw %k0, %eax ; NoVLX-NEXT: kmovw %k0, %eax
; NoVLX-NEXT: vpinsrb $15, %eax, %xmm2, %xmm2 ; NoVLX-NEXT: vpinsrb $15, %eax, %xmm3, %xmm3
; NoVLX-NEXT: vpand %xmm6, %xmm2, %xmm2 ; NoVLX-NEXT: vpand %xmm1, %xmm3, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm2, %zmm2 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm2, %zmm2 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm2, %zmm2, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpand %xmm0, %xmm3, %xmm0 ; NoVLX-NEXT: vpand %xmm0, %xmm2, %xmm0
; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0 ; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT: kmovw %k0, (%rsp) ; NoVLX-NEXT: kmovw %k0, (%rsp)
; NoVLX-NEXT: movl (%rsp), %ecx ; NoVLX-NEXT: movl (%rsp), %ecx
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: movl {{[0-9]+}}(%rsp), %eax ; NoVLX-NEXT: movl {{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT: shlq $32, %rax ; NoVLX-NEXT: shlq $32, %rax
; NoVLX-NEXT: orq %rcx, %rax ; NoVLX-NEXT: orq %rcx, %rax
; NoVLX-NEXT: movq %rbp, %rsp ; NoVLX-NEXT: movq %rbp, %rsp
; NoVLX-NEXT: popq %rbp ; NoVLX-NEXT: popq %rbp
; NoVLX-NEXT: retq ; NoVLX-NEXT: retq
▲ Show 20 Lines • Show All 48 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: shrl $16, %ecx ; NoVLX-NEXT: shrl $16, %ecx
; NoVLX-NEXT: vpinsrw $4, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrw $4, %eax, %xmm1, %xmm1
; NoVLX-NEXT: vpinsrw $5, %ecx, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrw $5, %ecx, %xmm1, %xmm1
; NoVLX-NEXT: movq %rax, %rcx ; NoVLX-NEXT: movq %rax, %rcx
; NoVLX-NEXT: shrq $32, %rcx ; NoVLX-NEXT: shrq $32, %rcx
; NoVLX-NEXT: vpinsrw $6, %ecx, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrw $6, %ecx, %xmm1, %xmm1
; NoVLX-NEXT: vmovq %xmm4, %rcx ; NoVLX-NEXT: vmovq %xmm4, %rcx
; NoVLX-NEXT: shrq $48, %rax ; NoVLX-NEXT: shrq $48, %rax
; NoVLX-NEXT: vpinsrw $7, %eax, %xmm1, %xmm2 ; NoVLX-NEXT: vpinsrw $7, %eax, %xmm1, %xmm1
; NoVLX-NEXT: movl %ecx, %eax ; NoVLX-NEXT: movl %ecx, %eax
; NoVLX-NEXT: shrl $16, %eax ; NoVLX-NEXT: shrl $16, %eax
; NoVLX-NEXT: vmovd %ecx, %xmm1 ; NoVLX-NEXT: vmovd %ecx, %xmm2
; NoVLX-NEXT: vpinsrw $1, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrw $1, %eax, %xmm2, %xmm2
; NoVLX-NEXT: movq %rcx, %rax ; NoVLX-NEXT: movq %rcx, %rax
; NoVLX-NEXT: shrq $32, %rax ; NoVLX-NEXT: shrq $32, %rax
; NoVLX-NEXT: vpinsrw $2, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrw $2, %eax, %xmm2, %xmm2
; NoVLX-NEXT: vpextrq $1, %xmm4, %rax ; NoVLX-NEXT: vpextrq $1, %xmm4, %rax
; NoVLX-NEXT: shrq $48, %rcx ; NoVLX-NEXT: shrq $48, %rcx
; NoVLX-NEXT: vpinsrw $3, %ecx, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrw $3, %ecx, %xmm2, %xmm2
; NoVLX-NEXT: movl %eax, %ecx ; NoVLX-NEXT: movl %eax, %ecx
; NoVLX-NEXT: shrl $16, %ecx ; NoVLX-NEXT: shrl $16, %ecx
; NoVLX-NEXT: vpinsrw $4, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrw $4, %eax, %xmm2, %xmm2
; NoVLX-NEXT: vpinsrw $5, %ecx, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrw $5, %ecx, %xmm2, %xmm2
; NoVLX-NEXT: movq %rax, %rcx ; NoVLX-NEXT: movq %rax, %rcx
; NoVLX-NEXT: shrq $32, %rcx ; NoVLX-NEXT: shrq $32, %rcx
; NoVLX-NEXT: vpinsrw $6, %ecx, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrw $6, %ecx, %xmm2, %xmm2
; NoVLX-NEXT: vmovq %xmm3, %rcx ; NoVLX-NEXT: vmovq %xmm3, %rcx
; NoVLX-NEXT: shrq $48, %rax ; NoVLX-NEXT: shrq $48, %rax
; NoVLX-NEXT: vpinsrw $7, %eax, %xmm1, %xmm4 ; NoVLX-NEXT: vpinsrw $7, %eax, %xmm2, %xmm2
; NoVLX-NEXT: movl %ecx, %eax ; NoVLX-NEXT: movl %ecx, %eax
; NoVLX-NEXT: shrl $16, %eax ; NoVLX-NEXT: shrl $16, %eax
; NoVLX-NEXT: vmovd %ecx, %xmm1 ; NoVLX-NEXT: vmovd %ecx, %xmm4
; NoVLX-NEXT: vpinsrw $1, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrw $1, %eax, %xmm4, %xmm4
; NoVLX-NEXT: movq %rcx, %rax ; NoVLX-NEXT: movq %rcx, %rax
; NoVLX-NEXT: shrq $32, %rax ; NoVLX-NEXT: shrq $32, %rax
; NoVLX-NEXT: vpinsrw $2, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrw $2, %eax, %xmm4, %xmm4
; NoVLX-NEXT: vpextrq $1, %xmm3, %rax ; NoVLX-NEXT: vpextrq $1, %xmm3, %rax
; NoVLX-NEXT: shrq $48, %rcx ; NoVLX-NEXT: shrq $48, %rcx
; NoVLX-NEXT: vpinsrw $3, %ecx, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrw $3, %ecx, %xmm4, %xmm3
; NoVLX-NEXT: movl %eax, %ecx ; NoVLX-NEXT: movl %eax, %ecx
; NoVLX-NEXT: shrl $16, %ecx ; NoVLX-NEXT: shrl $16, %ecx
; NoVLX-NEXT: vpinsrw $4, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrw $4, %eax, %xmm3, %xmm3
; NoVLX-NEXT: vpinsrw $5, %ecx, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrw $5, %ecx, %xmm3, %xmm3
; NoVLX-NEXT: movq %rax, %rcx ; NoVLX-NEXT: movq %rax, %rcx
; NoVLX-NEXT: shrq $32, %rcx ; NoVLX-NEXT: shrq $32, %rcx
; NoVLX-NEXT: vpinsrw $6, %ecx, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrw $6, %ecx, %xmm3, %xmm3
; NoVLX-NEXT: vmovq %xmm0, %rcx ; NoVLX-NEXT: vmovq %xmm0, %rcx
; NoVLX-NEXT: shrq $48, %rax ; NoVLX-NEXT: shrq $48, %rax
; NoVLX-NEXT: vpinsrw $7, %eax, %xmm1, %xmm3 ; NoVLX-NEXT: vpinsrw $7, %eax, %xmm3, %xmm4
; NoVLX-NEXT: movl %ecx, %eax ; NoVLX-NEXT: movl %ecx, %eax
; NoVLX-NEXT: shrl $16, %eax ; NoVLX-NEXT: shrl $16, %eax
; NoVLX-NEXT: vmovd %ecx, %xmm1 ; NoVLX-NEXT: vmovd %ecx, %xmm3
; NoVLX-NEXT: vpinsrw $1, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrw $1, %eax, %xmm3, %xmm3
; NoVLX-NEXT: movq %rcx, %rax ; NoVLX-NEXT: movq %rcx, %rax
; NoVLX-NEXT: shrq $32, %rax ; NoVLX-NEXT: shrq $32, %rax
; NoVLX-NEXT: vpinsrw $2, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrw $2, %eax, %xmm3, %xmm3
; NoVLX-NEXT: vpextrq $1, %xmm0, %rax ; NoVLX-NEXT: vpextrq $1, %xmm0, %rax
; NoVLX-NEXT: shrq $48, %rcx ; NoVLX-NEXT: shrq $48, %rcx
; NoVLX-NEXT: vpinsrw $3, %ecx, %xmm1, %xmm0 ; NoVLX-NEXT: vpinsrw $3, %ecx, %xmm3, %xmm0
; NoVLX-NEXT: movl %eax, %ecx ; NoVLX-NEXT: movl %eax, %ecx
; NoVLX-NEXT: shrl $16, %ecx ; NoVLX-NEXT: shrl $16, %ecx
; NoVLX-NEXT: vpinsrw $4, %eax, %xmm0, %xmm0 ; NoVLX-NEXT: vpinsrw $4, %eax, %xmm0, %xmm0
; NoVLX-NEXT: vpinsrw $5, %ecx, %xmm0, %xmm0 ; NoVLX-NEXT: vpinsrw $5, %ecx, %xmm0, %xmm0
; NoVLX-NEXT: movq %rax, %rcx ; NoVLX-NEXT: movq %rax, %rcx
; NoVLX-NEXT: shrq $32, %rcx ; NoVLX-NEXT: shrq $32, %rcx
; NoVLX-NEXT: vpinsrw $6, %ecx, %xmm0, %xmm5 ; NoVLX-NEXT: vpinsrw $6, %ecx, %xmm0, %xmm5
; NoVLX-NEXT: kmovw {{[0-9]+}}(%rsp), %k1 ; NoVLX-NEXT: kmovw {{[0-9]+}}(%rsp), %k1
; NoVLX-NEXT: kmovw {{[0-9]+}}(%rsp), %k2 ; NoVLX-NEXT: kmovw {{[0-9]+}}(%rsp), %k2
; NoVLX-NEXT: vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z} ; NoVLX-NEXT: vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT: vpmovdb %zmm0, %xmm1 ; NoVLX-NEXT: vpmovdb %zmm0, %xmm0
; NoVLX-NEXT: vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k2} {z} ; NoVLX-NEXT: vinserti128 $1, %xmm1, %ymm2, %ymm3
; NoVLX-NEXT: vinserti128 $1, %xmm2, %ymm4, %ymm4 ; NoVLX-NEXT: vpternlogd $255, %zmm1, %zmm1, %zmm1 {%k2} {z}
; NoVLX-NEXT: vpmovdb %zmm0, %xmm2 ; NoVLX-NEXT: vpmovdb %zmm1, %xmm1
; NoVLX-NEXT: shrq $48, %rax ; NoVLX-NEXT: shrq $48, %rax
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: vpinsrw $7, %eax, %xmm5, %xmm2
; NoVLX-NEXT: vpinsrw $7, %eax, %xmm5, %xmm5 ; NoVLX-NEXT: vinserti128 $1, %xmm4, %ymm2, %ymm2
; NoVLX-NEXT: vinserti128 $1, %xmm3, %ymm5, %ymm3 ; NoVLX-NEXT: vpcmpgtw (%rsi), %ymm2, %ymm2
; NoVLX-NEXT: vpcmpgtw (%rsi), %ymm3, %ymm3 ; NoVLX-NEXT: vpmovsxwd %ymm2, %zmm2
; NoVLX-NEXT: vpmovsxwd %ymm3, %zmm3 ; NoVLX-NEXT: vpslld $31, %zmm2, %zmm2
; NoVLX-NEXT: vpslld $31, %zmm3, %zmm3 ; NoVLX-NEXT: vptestmd %zmm2, %zmm2, %k0
; NoVLX-NEXT: vptestmd %zmm3, %zmm3, %k0
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: vmovd %eax, %xmm3 ; NoVLX-NEXT: vmovd %eax, %xmm2
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $1, %eax, %xmm3, %xmm3 ; NoVLX-NEXT: vpinsrb $1, %eax, %xmm2, %xmm2
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $2, %eax, %xmm3, %xmm3 ; NoVLX-NEXT: vpinsrb $2, %eax, %xmm2, %xmm2
; NoVLX-NEXT: kshiftlw $12, %k0, %k1 ; NoVLX-NEXT: kshiftlw $12, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $3, %eax, %xmm3, %xmm3 ; NoVLX-NEXT: vpinsrb $3, %eax, %xmm2, %xmm2
; NoVLX-NEXT: kshiftlw $11, %k0, %k1 ; NoVLX-NEXT: kshiftlw $11, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $4, %eax, %xmm3, %xmm3 ; NoVLX-NEXT: vpinsrb $4, %eax, %xmm2, %xmm2
; NoVLX-NEXT: kshiftlw $10, %k0, %k1 ; NoVLX-NEXT: kshiftlw $10, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $5, %eax, %xmm3, %xmm3 ; NoVLX-NEXT: vpinsrb $5, %eax, %xmm2, %xmm2
; NoVLX-NEXT: kshiftlw $9, %k0, %k1 ; NoVLX-NEXT: kshiftlw $9, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $6, %eax, %xmm3, %xmm3 ; NoVLX-NEXT: vpinsrb $6, %eax, %xmm2, %xmm2
; NoVLX-NEXT: kshiftlw $8, %k0, %k1 ; NoVLX-NEXT: kshiftlw $8, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $7, %eax, %xmm3, %xmm3 ; NoVLX-NEXT: vpinsrb $7, %eax, %xmm2, %xmm2
; NoVLX-NEXT: kshiftlw $7, %k0, %k1 ; NoVLX-NEXT: kshiftlw $7, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $8, %eax, %xmm3, %xmm3 ; NoVLX-NEXT: vpinsrb $8, %eax, %xmm2, %xmm2
; NoVLX-NEXT: kshiftlw $6, %k0, %k1 ; NoVLX-NEXT: kshiftlw $6, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $9, %eax, %xmm3, %xmm3 ; NoVLX-NEXT: vpinsrb $9, %eax, %xmm2, %xmm2
; NoVLX-NEXT: kshiftlw $5, %k0, %k1 ; NoVLX-NEXT: kshiftlw $5, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $10, %eax, %xmm3, %xmm3 ; NoVLX-NEXT: vpinsrb $10, %eax, %xmm2, %xmm2
; NoVLX-NEXT: kshiftlw $4, %k0, %k1 ; NoVLX-NEXT: kshiftlw $4, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $11, %eax, %xmm3, %xmm3 ; NoVLX-NEXT: vpinsrb $11, %eax, %xmm2, %xmm2
; NoVLX-NEXT: kshiftlw $3, %k0, %k1 ; NoVLX-NEXT: kshiftlw $3, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $12, %eax, %xmm3, %xmm3 ; NoVLX-NEXT: vpinsrb $12, %eax, %xmm2, %xmm2
; NoVLX-NEXT: kshiftlw $2, %k0, %k1 ; NoVLX-NEXT: kshiftlw $2, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $13, %eax, %xmm3, %xmm3 ; NoVLX-NEXT: vpinsrb $13, %eax, %xmm2, %xmm2
; NoVLX-NEXT: kshiftlw $1, %k0, %k1 ; NoVLX-NEXT: kshiftlw $1, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $14, %eax, %xmm3, %xmm3 ; NoVLX-NEXT: vpinsrb $14, %eax, %xmm2, %xmm2
; NoVLX-NEXT: kshiftrw $15, %k0, %k0 ; NoVLX-NEXT: kshiftrw $15, %k0, %k0
; NoVLX-NEXT: kmovw %k0, %eax ; NoVLX-NEXT: kmovw %k0, %eax
; NoVLX-NEXT: vpinsrb $15, %eax, %xmm3, %xmm3 ; NoVLX-NEXT: vpinsrb $15, %eax, %xmm2, %xmm2
; NoVLX-NEXT: vpcmpgtw 32(%rsi), %ymm4, %ymm4 ; NoVLX-NEXT: vpcmpgtw 32(%rsi), %ymm3, %ymm3
; NoVLX-NEXT: vpmovsxwd %ymm4, %zmm4 ; NoVLX-NEXT: vpmovsxwd %ymm3, %zmm3
; NoVLX-NEXT: vpslld $31, %zmm4, %zmm4 ; NoVLX-NEXT: vpslld $31, %zmm3, %zmm3
; NoVLX-NEXT: vptestmd %zmm4, %zmm4, %k0 ; NoVLX-NEXT: vptestmd %zmm3, %zmm3, %k0
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %ecx ; NoVLX-NEXT: kmovw %k1, %ecx
; NoVLX-NEXT: vmovd %ecx, %xmm4 ; NoVLX-NEXT: vmovd %ecx, %xmm3
; NoVLX-NEXT: vpinsrb $1, %eax, %xmm4, %xmm4 ; NoVLX-NEXT: vpinsrb $1, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $2, %eax, %xmm4, %xmm4 ; NoVLX-NEXT: vpinsrb $2, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftlw $12, %k0, %k1 ; NoVLX-NEXT: kshiftlw $12, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $3, %eax, %xmm4, %xmm4 ; NoVLX-NEXT: vpinsrb $3, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftlw $11, %k0, %k1 ; NoVLX-NEXT: kshiftlw $11, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $4, %eax, %xmm4, %xmm4 ; NoVLX-NEXT: vpinsrb $4, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftlw $10, %k0, %k1 ; NoVLX-NEXT: kshiftlw $10, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $5, %eax, %xmm4, %xmm4 ; NoVLX-NEXT: vpinsrb $5, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftlw $9, %k0, %k1 ; NoVLX-NEXT: kshiftlw $9, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $6, %eax, %xmm4, %xmm4 ; NoVLX-NEXT: vpinsrb $6, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftlw $8, %k0, %k1 ; NoVLX-NEXT: kshiftlw $8, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $7, %eax, %xmm4, %xmm4 ; NoVLX-NEXT: vpinsrb $7, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftlw $7, %k0, %k1 ; NoVLX-NEXT: kshiftlw $7, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $8, %eax, %xmm4, %xmm4 ; NoVLX-NEXT: vpinsrb $8, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftlw $6, %k0, %k1 ; NoVLX-NEXT: kshiftlw $6, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $9, %eax, %xmm4, %xmm4 ; NoVLX-NEXT: vpinsrb $9, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftlw $5, %k0, %k1 ; NoVLX-NEXT: kshiftlw $5, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $10, %eax, %xmm4, %xmm4 ; NoVLX-NEXT: vpinsrb $10, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftlw $4, %k0, %k1 ; NoVLX-NEXT: kshiftlw $4, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $11, %eax, %xmm4, %xmm4 ; NoVLX-NEXT: vpinsrb $11, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftlw $3, %k0, %k1 ; NoVLX-NEXT: kshiftlw $3, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $12, %eax, %xmm4, %xmm4 ; NoVLX-NEXT: vpinsrb $12, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftlw $2, %k0, %k1 ; NoVLX-NEXT: kshiftlw $2, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $13, %eax, %xmm4, %xmm4 ; NoVLX-NEXT: vpinsrb $13, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftlw $1, %k0, %k1 ; NoVLX-NEXT: kshiftlw $1, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $14, %eax, %xmm4, %xmm4 ; NoVLX-NEXT: vpinsrb $14, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftrw $15, %k0, %k0 ; NoVLX-NEXT: kshiftrw $15, %k0, %k0
; NoVLX-NEXT: kmovw %k0, %eax ; NoVLX-NEXT: kmovw %k0, %eax
; NoVLX-NEXT: vpinsrb $15, %eax, %xmm4, %xmm4 ; NoVLX-NEXT: vpinsrb $15, %eax, %xmm3, %xmm3
; NoVLX-NEXT: vpand %xmm2, %xmm4, %xmm2
; NoVLX-NEXT: vpmovsxbd %xmm2, %zmm2
; NoVLX-NEXT: vpslld $31, %zmm2, %zmm2
; NoVLX-NEXT: vptestmd %zmm2, %zmm2, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpand %xmm1, %xmm3, %xmm1 ; NoVLX-NEXT: vpand %xmm1, %xmm3, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpand %xmm0, %xmm2, %xmm0
; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT: kmovw %k0, (%rsp) ; NoVLX-NEXT: kmovw %k0, (%rsp)
; NoVLX-NEXT: movl (%rsp), %ecx ; NoVLX-NEXT: movl (%rsp), %ecx
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: movl {{[0-9]+}}(%rsp), %eax ; NoVLX-NEXT: movl {{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT: shlq $32, %rax ; NoVLX-NEXT: shlq $32, %rax
; NoVLX-NEXT: orq %rcx, %rax ; NoVLX-NEXT: orq %rcx, %rax
; NoVLX-NEXT: movq %rbp, %rsp ; NoVLX-NEXT: movq %rbp, %rsp
; NoVLX-NEXT: popq %rbp ; NoVLX-NEXT: popq %rbp
; NoVLX-NEXT: retq ; NoVLX-NEXT: retq
▲ Show 20 Lines • Show All 1,165 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi647: ; NoVLX-NEXT: .Lcfi647:
; NoVLX-NEXT: .cfi_offset %rbp, -16 ; NoVLX-NEXT: .cfi_offset %rbp, -16
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi648: ; NoVLX-NEXT: .Lcfi648:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: vpcmpgtd %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vpcmpgtd %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vpxor %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero ; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
Show All 32 Lines
; NoVLX-NEXT: .Lcfi650: ; NoVLX-NEXT: .Lcfi650:
; NoVLX-NEXT: .cfi_offset %rbp, -16 ; NoVLX-NEXT: .cfi_offset %rbp, -16
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi651: ; NoVLX-NEXT: .Lcfi651:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: vpcmpgtd (%rdi), %xmm0, %xmm0 ; NoVLX-NEXT: vpcmpgtd (%rdi), %xmm0, %xmm0
; NoVLX-NEXT: vxorps %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero ; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
▲ Show 20 Lines • Show All 52 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: kmovw %k3, %ecx ; NoVLX-NEXT: kmovw %k3, %ecx
; NoVLX-NEXT: vmovd %ecx, %xmm1 ; NoVLX-NEXT: vmovd %ecx, %xmm1
; NoVLX-NEXT: vpinsrb $4, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrb $4, %eax, %xmm1, %xmm1
; NoVLX-NEXT: kmovw %k2, %eax ; NoVLX-NEXT: kmovw %k2, %eax
; NoVLX-NEXT: vpinsrb $8, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $12, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT: vpand %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vpxor %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero ; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
▲ Show 20 Lines • Show All 54 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: kmovw %k3, %ecx ; NoVLX-NEXT: kmovw %k3, %ecx
; NoVLX-NEXT: vmovd %ecx, %xmm1 ; NoVLX-NEXT: vmovd %ecx, %xmm1
; NoVLX-NEXT: vpinsrb $4, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrb $4, %eax, %xmm1, %xmm1
; NoVLX-NEXT: kmovw %k2, %eax ; NoVLX-NEXT: kmovw %k2, %eax
; NoVLX-NEXT: vpinsrb $8, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $12, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT: vpand %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vpxor %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero ; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
Show All 38 Lines
; NoVLX-NEXT: .cfi_offset %rbp, -16 ; NoVLX-NEXT: .cfi_offset %rbp, -16
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi660: ; NoVLX-NEXT: .Lcfi660:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: vpbroadcastd (%rdi), %xmm1 ; NoVLX-NEXT: vpbroadcastd (%rdi), %xmm1
; NoVLX-NEXT: vpcmpgtd %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vpcmpgtd %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vpxor %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero ; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
▲ Show 20 Lines • Show All 54 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: kmovw %k3, %ecx ; NoVLX-NEXT: kmovw %k3, %ecx
; NoVLX-NEXT: vmovd %ecx, %xmm1 ; NoVLX-NEXT: vmovd %ecx, %xmm1
; NoVLX-NEXT: vpinsrb $4, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrb $4, %eax, %xmm1, %xmm1
; NoVLX-NEXT: kmovw %k2, %eax ; NoVLX-NEXT: kmovw %k2, %eax
; NoVLX-NEXT: vpinsrb $8, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $12, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT: vpand %xmm0, %xmm1, %xmm0 ; NoVLX-NEXT: vpand %xmm0, %xmm1, %xmm0
; NoVLX-NEXT: vpxor %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero ; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
▲ Show 20 Lines • Show All 229 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi666: ; NoVLX-NEXT: .Lcfi666:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $32, %rsp ; NoVLX-NEXT: subq $32, %rsp
; NoVLX-NEXT: # kill: %YMM1<def> %YMM1<kill> %ZMM1<def> ; NoVLX-NEXT: # kill: %YMM1<def> %YMM1<kill> %ZMM1<def>
; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def> ; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def>
; NoVLX-NEXT: vpcmpgtd %zmm1, %zmm0, %k0 ; NoVLX-NEXT: vpcmpgtd %zmm1, %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r9d ; NoVLX-NEXT: kmovw %k1, %r9d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 58 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi669: ; NoVLX-NEXT: .Lcfi669:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $32, %rsp ; NoVLX-NEXT: subq $32, %rsp
; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def> ; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def>
; NoVLX-NEXT: vmovdqa (%rdi), %ymm1 ; NoVLX-NEXT: vmovdqa (%rdi), %ymm1
; NoVLX-NEXT: vpcmpgtd %zmm1, %zmm0, %k0 ; NoVLX-NEXT: vpcmpgtd %zmm1, %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r9d ; NoVLX-NEXT: kmovw %k1, %r9d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 62 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $32, %rsp ; NoVLX-NEXT: subq $32, %rsp
; NoVLX-NEXT: # kill: %YMM1<def> %YMM1<kill> %ZMM1<def> ; NoVLX-NEXT: # kill: %YMM1<def> %YMM1<kill> %ZMM1<def>
; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def> ; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def>
; NoVLX-NEXT: vpcmpgtd %zmm1, %zmm0, %k0 ; NoVLX-NEXT: vpcmpgtd %zmm1, %zmm0, %k0
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: kandw %k1, %k0, %k0 ; NoVLX-NEXT: kandw %k1, %k0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r9d ; NoVLX-NEXT: kmovw %k1, %r9d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 63 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $32, %rsp ; NoVLX-NEXT: subq $32, %rsp
; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def> ; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def>
; NoVLX-NEXT: vmovdqa (%rsi), %ymm1 ; NoVLX-NEXT: vmovdqa (%rsi), %ymm1
; NoVLX-NEXT: vpcmpgtd %zmm1, %zmm0, %k0 ; NoVLX-NEXT: vpcmpgtd %zmm1, %zmm0, %k0
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: kandw %k1, %k0, %k0 ; NoVLX-NEXT: kandw %k1, %k0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r9d ; NoVLX-NEXT: kmovw %k1, %r9d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 62 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi678: ; NoVLX-NEXT: .Lcfi678:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $32, %rsp ; NoVLX-NEXT: subq $32, %rsp
; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def> ; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def>
; NoVLX-NEXT: vpbroadcastd (%rdi), %ymm1 ; NoVLX-NEXT: vpbroadcastd (%rdi), %ymm1
; NoVLX-NEXT: vpcmpgtd %zmm1, %zmm0, %k0 ; NoVLX-NEXT: vpcmpgtd %zmm1, %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r9d ; NoVLX-NEXT: kmovw %k1, %r9d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 63 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $32, %rsp ; NoVLX-NEXT: subq $32, %rsp
; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def> ; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def>
; NoVLX-NEXT: vpbroadcastd (%rsi), %ymm1 ; NoVLX-NEXT: vpbroadcastd (%rsi), %ymm1
; NoVLX-NEXT: vpcmpgtd %zmm1, %zmm0, %k0 ; NoVLX-NEXT: vpcmpgtd %zmm1, %zmm0, %k0
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: kandw %k0, %k1, %k0 ; NoVLX-NEXT: kandw %k0, %k1, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r9d ; NoVLX-NEXT: kmovw %k1, %r9d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 63 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi684: ; NoVLX-NEXT: .Lcfi684:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: # kill: %YMM1<def> %YMM1<kill> %ZMM1<def> ; NoVLX-NEXT: # kill: %YMM1<def> %YMM1<kill> %ZMM1<def>
; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def> ; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def>
; NoVLX-NEXT: vpcmpgtd %zmm1, %zmm0, %k0 ; NoVLX-NEXT: vpcmpgtd %zmm1, %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 63 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi687: ; NoVLX-NEXT: .Lcfi687:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def> ; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def>
; NoVLX-NEXT: vmovdqa (%rdi), %ymm1 ; NoVLX-NEXT: vmovdqa (%rdi), %ymm1
; NoVLX-NEXT: vpcmpgtd %zmm1, %zmm0, %k0 ; NoVLX-NEXT: vpcmpgtd %zmm1, %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 67 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: # kill: %YMM1<def> %YMM1<kill> %ZMM1<def> ; NoVLX-NEXT: # kill: %YMM1<def> %YMM1<kill> %ZMM1<def>
; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def> ; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def>
; NoVLX-NEXT: vpcmpgtd %zmm1, %zmm0, %k0 ; NoVLX-NEXT: vpcmpgtd %zmm1, %zmm0, %k0
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: kandw %k1, %k0, %k0 ; NoVLX-NEXT: kandw %k1, %k0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 68 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def> ; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def>
; NoVLX-NEXT: vmovdqa (%rsi), %ymm1 ; NoVLX-NEXT: vmovdqa (%rsi), %ymm1
; NoVLX-NEXT: vpcmpgtd %zmm1, %zmm0, %k0 ; NoVLX-NEXT: vpcmpgtd %zmm1, %zmm0, %k0
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: kandw %k1, %k0, %k0 ; NoVLX-NEXT: kandw %k1, %k0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 67 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi696: ; NoVLX-NEXT: .Lcfi696:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def> ; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def>
; NoVLX-NEXT: vpbroadcastd (%rdi), %ymm1 ; NoVLX-NEXT: vpbroadcastd (%rdi), %ymm1
; NoVLX-NEXT: vpcmpgtd %zmm1, %zmm0, %k0 ; NoVLX-NEXT: vpcmpgtd %zmm1, %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 68 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def> ; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def>
; NoVLX-NEXT: vpbroadcastd (%rsi), %ymm1 ; NoVLX-NEXT: vpbroadcastd (%rsi), %ymm1
; NoVLX-NEXT: vpcmpgtd %zmm1, %zmm0, %k0 ; NoVLX-NEXT: vpcmpgtd %zmm1, %zmm0, %k0
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: kandw %k0, %k1, %k0 ; NoVLX-NEXT: kandw %k0, %k1, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 81 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %r12, -48 ; NoVLX-NEXT: .cfi_offset %r12, -48
; NoVLX-NEXT: .Lcfi705: ; NoVLX-NEXT: .Lcfi705:
; NoVLX-NEXT: .cfi_offset %r13, -40 ; NoVLX-NEXT: .cfi_offset %r13, -40
; NoVLX-NEXT: .Lcfi706: ; NoVLX-NEXT: .Lcfi706:
; NoVLX-NEXT: .cfi_offset %r14, -32 ; NoVLX-NEXT: .cfi_offset %r14, -32
; NoVLX-NEXT: .Lcfi707: ; NoVLX-NEXT: .Lcfi707:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: vpcmpgtd %zmm1, %zmm0, %k0 ; NoVLX-NEXT: vpcmpgtd %zmm1, %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r10d ; NoVLX-NEXT: kmovw %k1, %r10d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 106 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %r12, -48 ; NoVLX-NEXT: .cfi_offset %r12, -48
; NoVLX-NEXT: .Lcfi713: ; NoVLX-NEXT: .Lcfi713:
; NoVLX-NEXT: .cfi_offset %r13, -40 ; NoVLX-NEXT: .cfi_offset %r13, -40
; NoVLX-NEXT: .Lcfi714: ; NoVLX-NEXT: .Lcfi714:
; NoVLX-NEXT: .cfi_offset %r14, -32 ; NoVLX-NEXT: .cfi_offset %r14, -32
; NoVLX-NEXT: .Lcfi715: ; NoVLX-NEXT: .Lcfi715:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: vpcmpgtd (%rdi), %zmm0, %k0 ; NoVLX-NEXT: vpcmpgtd (%rdi), %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r10d ; NoVLX-NEXT: kmovw %k1, %r10d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 109 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi721: ; NoVLX-NEXT: .Lcfi721:
; NoVLX-NEXT: .cfi_offset %r13, -40 ; NoVLX-NEXT: .cfi_offset %r13, -40
; NoVLX-NEXT: .Lcfi722: ; NoVLX-NEXT: .Lcfi722:
; NoVLX-NEXT: .cfi_offset %r14, -32 ; NoVLX-NEXT: .cfi_offset %r14, -32
; NoVLX-NEXT: .Lcfi723: ; NoVLX-NEXT: .Lcfi723:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vpcmpgtd %zmm1, %zmm0, %k0 {%k1} ; NoVLX-NEXT: vpcmpgtd %zmm1, %zmm0, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r10d ; NoVLX-NEXT: kmovw %k1, %r10d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 110 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi729: ; NoVLX-NEXT: .Lcfi729:
; NoVLX-NEXT: .cfi_offset %r13, -40 ; NoVLX-NEXT: .cfi_offset %r13, -40
; NoVLX-NEXT: .Lcfi730: ; NoVLX-NEXT: .Lcfi730:
; NoVLX-NEXT: .cfi_offset %r14, -32 ; NoVLX-NEXT: .cfi_offset %r14, -32
; NoVLX-NEXT: .Lcfi731: ; NoVLX-NEXT: .Lcfi731:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vpcmpgtd (%rsi), %zmm0, %k0 {%k1} ; NoVLX-NEXT: vpcmpgtd (%rsi), %zmm0, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r10d ; NoVLX-NEXT: kmovw %k1, %r10d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 110 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %r12, -48 ; NoVLX-NEXT: .cfi_offset %r12, -48
; NoVLX-NEXT: .Lcfi737: ; NoVLX-NEXT: .Lcfi737:
; NoVLX-NEXT: .cfi_offset %r13, -40 ; NoVLX-NEXT: .cfi_offset %r13, -40
; NoVLX-NEXT: .Lcfi738: ; NoVLX-NEXT: .Lcfi738:
; NoVLX-NEXT: .cfi_offset %r14, -32 ; NoVLX-NEXT: .cfi_offset %r14, -32
; NoVLX-NEXT: .Lcfi739: ; NoVLX-NEXT: .Lcfi739:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: vpcmpgtd (%rdi){1to16}, %zmm0, %k0 ; NoVLX-NEXT: vpcmpgtd (%rdi){1to16}, %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r10d ; NoVLX-NEXT: kmovw %k1, %r10d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 110 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi745: ; NoVLX-NEXT: .Lcfi745:
; NoVLX-NEXT: .cfi_offset %r13, -40 ; NoVLX-NEXT: .cfi_offset %r13, -40
; NoVLX-NEXT: .Lcfi746: ; NoVLX-NEXT: .Lcfi746:
; NoVLX-NEXT: .cfi_offset %r14, -32 ; NoVLX-NEXT: .cfi_offset %r14, -32
; NoVLX-NEXT: .Lcfi747: ; NoVLX-NEXT: .Lcfi747:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vpcmpgtd (%rsi){1to16}, %zmm0, %k0 {%k1} ; NoVLX-NEXT: vpcmpgtd (%rsi){1to16}, %zmm0, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r10d ; NoVLX-NEXT: kmovw %k1, %r10d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 111 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %r12, -48 ; NoVLX-NEXT: .cfi_offset %r12, -48
; NoVLX-NEXT: .Lcfi753: ; NoVLX-NEXT: .Lcfi753:
; NoVLX-NEXT: .cfi_offset %r13, -40 ; NoVLX-NEXT: .cfi_offset %r13, -40
; NoVLX-NEXT: .Lcfi754: ; NoVLX-NEXT: .Lcfi754:
; NoVLX-NEXT: .cfi_offset %r14, -32 ; NoVLX-NEXT: .cfi_offset %r14, -32
; NoVLX-NEXT: .Lcfi755: ; NoVLX-NEXT: .Lcfi755:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: vpcmpgtd %zmm1, %zmm0, %k0 ; NoVLX-NEXT: vpcmpgtd %zmm1, %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 111 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %r12, -48 ; NoVLX-NEXT: .cfi_offset %r12, -48
; NoVLX-NEXT: .Lcfi761: ; NoVLX-NEXT: .Lcfi761:
; NoVLX-NEXT: .cfi_offset %r13, -40 ; NoVLX-NEXT: .cfi_offset %r13, -40
; NoVLX-NEXT: .Lcfi762: ; NoVLX-NEXT: .Lcfi762:
; NoVLX-NEXT: .cfi_offset %r14, -32 ; NoVLX-NEXT: .cfi_offset %r14, -32
; NoVLX-NEXT: .Lcfi763: ; NoVLX-NEXT: .Lcfi763:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: vpcmpgtd (%rdi), %zmm0, %k0 ; NoVLX-NEXT: vpcmpgtd (%rdi), %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 114 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi769: ; NoVLX-NEXT: .Lcfi769:
; NoVLX-NEXT: .cfi_offset %r13, -40 ; NoVLX-NEXT: .cfi_offset %r13, -40
; NoVLX-NEXT: .Lcfi770: ; NoVLX-NEXT: .Lcfi770:
; NoVLX-NEXT: .cfi_offset %r14, -32 ; NoVLX-NEXT: .cfi_offset %r14, -32
; NoVLX-NEXT: .Lcfi771: ; NoVLX-NEXT: .Lcfi771:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vpcmpgtd %zmm1, %zmm0, %k0 {%k1} ; NoVLX-NEXT: vpcmpgtd %zmm1, %zmm0, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 115 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi777: ; NoVLX-NEXT: .Lcfi777:
; NoVLX-NEXT: .cfi_offset %r13, -40 ; NoVLX-NEXT: .cfi_offset %r13, -40
; NoVLX-NEXT: .Lcfi778: ; NoVLX-NEXT: .Lcfi778:
; NoVLX-NEXT: .cfi_offset %r14, -32 ; NoVLX-NEXT: .cfi_offset %r14, -32
; NoVLX-NEXT: .Lcfi779: ; NoVLX-NEXT: .Lcfi779:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vpcmpgtd (%rsi), %zmm0, %k0 {%k1} ; NoVLX-NEXT: vpcmpgtd (%rsi), %zmm0, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 115 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %r12, -48 ; NoVLX-NEXT: .cfi_offset %r12, -48
; NoVLX-NEXT: .Lcfi785: ; NoVLX-NEXT: .Lcfi785:
; NoVLX-NEXT: .cfi_offset %r13, -40 ; NoVLX-NEXT: .cfi_offset %r13, -40
; NoVLX-NEXT: .Lcfi786: ; NoVLX-NEXT: .Lcfi786:
; NoVLX-NEXT: .cfi_offset %r14, -32 ; NoVLX-NEXT: .cfi_offset %r14, -32
; NoVLX-NEXT: .Lcfi787: ; NoVLX-NEXT: .Lcfi787:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: vpcmpgtd (%rdi){1to16}, %zmm0, %k0 ; NoVLX-NEXT: vpcmpgtd (%rdi){1to16}, %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 115 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi793: ; NoVLX-NEXT: .Lcfi793:
; NoVLX-NEXT: .cfi_offset %r13, -40 ; NoVLX-NEXT: .cfi_offset %r13, -40
; NoVLX-NEXT: .Lcfi794: ; NoVLX-NEXT: .Lcfi794:
; NoVLX-NEXT: .cfi_offset %r14, -32 ; NoVLX-NEXT: .cfi_offset %r14, -32
; NoVLX-NEXT: .Lcfi795: ; NoVLX-NEXT: .Lcfi795:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vpcmpgtd (%rsi){1to16}, %zmm0, %k0 {%k1} ; NoVLX-NEXT: vpcmpgtd (%rsi){1to16}, %zmm0, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 1,190 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi815: ; NoVLX-NEXT: .Lcfi815:
; NoVLX-NEXT: .cfi_offset %rbp, -16 ; NoVLX-NEXT: .cfi_offset %rbp, -16
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi816: ; NoVLX-NEXT: .Lcfi816:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: vpcmpgtq %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vpcmpgtq %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vpxor %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero ; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
Show All 32 Lines
; NoVLX-NEXT: .Lcfi818: ; NoVLX-NEXT: .Lcfi818:
; NoVLX-NEXT: .cfi_offset %rbp, -16 ; NoVLX-NEXT: .cfi_offset %rbp, -16
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi819: ; NoVLX-NEXT: .Lcfi819:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: vpcmpgtq (%rdi), %xmm0, %xmm0 ; NoVLX-NEXT: vpcmpgtq (%rdi), %xmm0, %xmm0
; NoVLX-NEXT: vxorps %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero ; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
▲ Show 20 Lines • Show All 44 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kshiftlw $14, %k0, %k0 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
; NoVLX-NEXT: kshiftrw $15, %k0, %k0 ; NoVLX-NEXT: kshiftrw $15, %k0, %k0
; NoVLX-NEXT: kmovw %k0, %eax ; NoVLX-NEXT: kmovw %k0, %eax
; NoVLX-NEXT: kmovw %k1, %ecx ; NoVLX-NEXT: kmovw %k1, %ecx
; NoVLX-NEXT: vmovd %ecx, %xmm1 ; NoVLX-NEXT: vmovd %ecx, %xmm1
; NoVLX-NEXT: vpinsrb $8, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT: vpand %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vpxor %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero ; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
▲ Show 20 Lines • Show All 46 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kshiftlw $14, %k0, %k0 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
; NoVLX-NEXT: kshiftrw $15, %k0, %k0 ; NoVLX-NEXT: kshiftrw $15, %k0, %k0
; NoVLX-NEXT: kmovw %k0, %eax ; NoVLX-NEXT: kmovw %k0, %eax
; NoVLX-NEXT: kmovw %k1, %ecx ; NoVLX-NEXT: kmovw %k1, %ecx
; NoVLX-NEXT: vmovd %ecx, %xmm1 ; NoVLX-NEXT: vmovd %ecx, %xmm1
; NoVLX-NEXT: vpinsrb $8, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT: vpand %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vpxor %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero ; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
Show All 38 Lines
; NoVLX-NEXT: .cfi_offset %rbp, -16 ; NoVLX-NEXT: .cfi_offset %rbp, -16
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi828: ; NoVLX-NEXT: .Lcfi828:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: vpbroadcastq (%rdi), %xmm1 ; NoVLX-NEXT: vpbroadcastq (%rdi), %xmm1
; NoVLX-NEXT: vpcmpgtq %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vpcmpgtq %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vpxor %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero ; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
▲ Show 20 Lines • Show All 46 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kshiftlw $14, %k0, %k0 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
; NoVLX-NEXT: kshiftrw $15, %k0, %k0 ; NoVLX-NEXT: kshiftrw $15, %k0, %k0
; NoVLX-NEXT: kmovw %k0, %eax ; NoVLX-NEXT: kmovw %k0, %eax
; NoVLX-NEXT: kmovw %k1, %ecx ; NoVLX-NEXT: kmovw %k1, %ecx
; NoVLX-NEXT: vmovd %ecx, %xmm1 ; NoVLX-NEXT: vmovd %ecx, %xmm1
; NoVLX-NEXT: vpinsrb $8, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT: vpand %xmm0, %xmm1, %xmm0 ; NoVLX-NEXT: vpand %xmm0, %xmm1, %xmm0
; NoVLX-NEXT: vpxor %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero ; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
▲ Show 20 Lines • Show All 1,216 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %rbp, -16 ; NoVLX-NEXT: .cfi_offset %rbp, -16
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi852: ; NoVLX-NEXT: .Lcfi852:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: vpcmpgtq %ymm1, %ymm0, %ymm0 ; NoVLX-NEXT: vpcmpgtq %ymm1, %ymm0, %ymm0
; NoVLX-NEXT: vpmovqd %zmm0, %ymm0 ; NoVLX-NEXT: vpmovqd %zmm0, %ymm0
; NoVLX-NEXT: vpxor %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero ; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
Show All 34 Lines
; NoVLX-NEXT: .cfi_offset %rbp, -16 ; NoVLX-NEXT: .cfi_offset %rbp, -16
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi855: ; NoVLX-NEXT: .Lcfi855:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: vpcmpgtq (%rdi), %ymm0, %ymm0 ; NoVLX-NEXT: vpcmpgtq (%rdi), %ymm0, %ymm0
; NoVLX-NEXT: vpmovqd %zmm0, %ymm0 ; NoVLX-NEXT: vpmovqd %zmm0, %ymm0
; NoVLX-NEXT: vxorps %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero ; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
▲ Show 20 Lines • Show All 54 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: kmovw %k3, %ecx ; NoVLX-NEXT: kmovw %k3, %ecx
; NoVLX-NEXT: vmovd %ecx, %xmm1 ; NoVLX-NEXT: vmovd %ecx, %xmm1
; NoVLX-NEXT: vpinsrb $4, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrb $4, %eax, %xmm1, %xmm1
; NoVLX-NEXT: kmovw %k2, %eax ; NoVLX-NEXT: kmovw %k2, %eax
; NoVLX-NEXT: vpinsrb $8, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $12, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT: vpand %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vpxor %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero ; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
▲ Show 20 Lines • Show All 56 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: kmovw %k3, %ecx ; NoVLX-NEXT: kmovw %k3, %ecx
; NoVLX-NEXT: vmovd %ecx, %xmm1 ; NoVLX-NEXT: vmovd %ecx, %xmm1
; NoVLX-NEXT: vpinsrb $4, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrb $4, %eax, %xmm1, %xmm1
; NoVLX-NEXT: kmovw %k2, %eax ; NoVLX-NEXT: kmovw %k2, %eax
; NoVLX-NEXT: vpinsrb $8, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $12, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT: vpand %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vpxor %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero ; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
Show All 40 Lines
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi864: ; NoVLX-NEXT: .Lcfi864:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: vpbroadcastq (%rdi), %ymm1 ; NoVLX-NEXT: vpbroadcastq (%rdi), %ymm1
; NoVLX-NEXT: vpcmpgtq %ymm1, %ymm0, %ymm0 ; NoVLX-NEXT: vpcmpgtq %ymm1, %ymm0, %ymm0
; NoVLX-NEXT: vpmovqd %zmm0, %ymm0 ; NoVLX-NEXT: vpmovqd %zmm0, %ymm0
; NoVLX-NEXT: vpxor %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero ; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
▲ Show 20 Lines • Show All 56 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: kmovw %k3, %ecx ; NoVLX-NEXT: kmovw %k3, %ecx
; NoVLX-NEXT: vmovd %ecx, %xmm1 ; NoVLX-NEXT: vmovd %ecx, %xmm1
; NoVLX-NEXT: vpinsrb $4, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrb $4, %eax, %xmm1, %xmm1
; NoVLX-NEXT: kmovw %k2, %eax ; NoVLX-NEXT: kmovw %k2, %eax
; NoVLX-NEXT: vpinsrb $8, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $12, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT: vpand %xmm0, %xmm1, %xmm0 ; NoVLX-NEXT: vpand %xmm0, %xmm1, %xmm0
; NoVLX-NEXT: vpxor %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero ; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
▲ Show 20 Lines • Show All 203 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi869: ; NoVLX-NEXT: .Lcfi869:
; NoVLX-NEXT: .cfi_offset %rbp, -16 ; NoVLX-NEXT: .cfi_offset %rbp, -16
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi870: ; NoVLX-NEXT: .Lcfi870:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $32, %rsp ; NoVLX-NEXT: subq $32, %rsp
; NoVLX-NEXT: vpcmpgtq %zmm1, %zmm0, %k0 ; NoVLX-NEXT: vpcmpgtq %zmm1, %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r9d ; NoVLX-NEXT: kmovw %k1, %r9d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 56 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi872: ; NoVLX-NEXT: .Lcfi872:
; NoVLX-NEXT: .cfi_offset %rbp, -16 ; NoVLX-NEXT: .cfi_offset %rbp, -16
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi873: ; NoVLX-NEXT: .Lcfi873:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $32, %rsp ; NoVLX-NEXT: subq $32, %rsp
; NoVLX-NEXT: vpcmpgtq (%rdi), %zmm0, %k0 ; NoVLX-NEXT: vpcmpgtq (%rdi), %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r9d ; NoVLX-NEXT: kmovw %k1, %r9d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 59 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %rbp, -16 ; NoVLX-NEXT: .cfi_offset %rbp, -16
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi876: ; NoVLX-NEXT: .Lcfi876:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $32, %rsp ; NoVLX-NEXT: subq $32, %rsp
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vpcmpgtq %zmm1, %zmm0, %k0 {%k1} ; NoVLX-NEXT: vpcmpgtq %zmm1, %zmm0, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r9d ; NoVLX-NEXT: kmovw %k1, %r9d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 60 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %rbp, -16 ; NoVLX-NEXT: .cfi_offset %rbp, -16
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi879: ; NoVLX-NEXT: .Lcfi879:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $32, %rsp ; NoVLX-NEXT: subq $32, %rsp
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vpcmpgtq (%rsi), %zmm0, %k0 {%k1} ; NoVLX-NEXT: vpcmpgtq (%rsi), %zmm0, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r9d ; NoVLX-NEXT: kmovw %k1, %r9d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 60 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi881: ; NoVLX-NEXT: .Lcfi881:
; NoVLX-NEXT: .cfi_offset %rbp, -16 ; NoVLX-NEXT: .cfi_offset %rbp, -16
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi882: ; NoVLX-NEXT: .Lcfi882:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $32, %rsp ; NoVLX-NEXT: subq $32, %rsp
; NoVLX-NEXT: vpcmpgtq (%rdi){1to8}, %zmm0, %k0 ; NoVLX-NEXT: vpcmpgtq (%rdi){1to8}, %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r9d ; NoVLX-NEXT: kmovw %k1, %r9d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 60 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %rbp, -16 ; NoVLX-NEXT: .cfi_offset %rbp, -16
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi885: ; NoVLX-NEXT: .Lcfi885:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $32, %rsp ; NoVLX-NEXT: subq $32, %rsp
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vpcmpgtq (%rsi){1to8}, %zmm0, %k0 {%k1} ; NoVLX-NEXT: vpcmpgtq (%rsi){1to8}, %zmm0, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r9d ; NoVLX-NEXT: kmovw %k1, %r9d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 61 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi887: ; NoVLX-NEXT: .Lcfi887:
; NoVLX-NEXT: .cfi_offset %rbp, -16 ; NoVLX-NEXT: .cfi_offset %rbp, -16
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi888: ; NoVLX-NEXT: .Lcfi888:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: vpcmpgtq %zmm1, %zmm0, %k0 ; NoVLX-NEXT: vpcmpgtq %zmm1, %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 61 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi890: ; NoVLX-NEXT: .Lcfi890:
; NoVLX-NEXT: .cfi_offset %rbp, -16 ; NoVLX-NEXT: .cfi_offset %rbp, -16
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi891: ; NoVLX-NEXT: .Lcfi891:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: vpcmpgtq (%rdi), %zmm0, %k0 ; NoVLX-NEXT: vpcmpgtq (%rdi), %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 64 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %rbp, -16 ; NoVLX-NEXT: .cfi_offset %rbp, -16
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi894: ; NoVLX-NEXT: .Lcfi894:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vpcmpgtq %zmm1, %zmm0, %k0 {%k1} ; NoVLX-NEXT: vpcmpgtq %zmm1, %zmm0, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 65 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %rbp, -16 ; NoVLX-NEXT: .cfi_offset %rbp, -16
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi897: ; NoVLX-NEXT: .Lcfi897:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vpcmpgtq (%rsi), %zmm0, %k0 {%k1} ; NoVLX-NEXT: vpcmpgtq (%rsi), %zmm0, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 65 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi899: ; NoVLX-NEXT: .Lcfi899:
; NoVLX-NEXT: .cfi_offset %rbp, -16 ; NoVLX-NEXT: .cfi_offset %rbp, -16
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi900: ; NoVLX-NEXT: .Lcfi900:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: vpcmpgtq (%rdi){1to8}, %zmm0, %k0 ; NoVLX-NEXT: vpcmpgtq (%rdi){1to8}, %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 65 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %rbp, -16 ; NoVLX-NEXT: .cfi_offset %rbp, -16
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi903: ; NoVLX-NEXT: .Lcfi903:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vpcmpgtq (%rsi){1to8}, %zmm0, %k0 {%k1} ; NoVLX-NEXT: vpcmpgtq (%rsi){1to8}, %zmm0, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 85 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi911: ; NoVLX-NEXT: .Lcfi911:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: vpcmpgtb %xmm0, %xmm1, %xmm0 ; NoVLX-NEXT: vpcmpgtb %xmm0, %xmm1, %xmm0
; NoVLX-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1
; NoVLX-NEXT: vpxor %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0 ; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r10d ; NoVLX-NEXT: kmovw %k1, %r10d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 111 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: vmovdqa (%rdi), %xmm1 ; NoVLX-NEXT: vmovdqa (%rdi), %xmm1
; NoVLX-NEXT: vpcmpgtb %xmm0, %xmm1, %xmm0 ; NoVLX-NEXT: vpcmpgtb %xmm0, %xmm1, %xmm0
; NoVLX-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1
; NoVLX-NEXT: vpxor %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0 ; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r10d ; NoVLX-NEXT: kmovw %k1, %r10d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 113 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: vpcmpgtb %xmm0, %xmm1, %xmm0 ; NoVLX-NEXT: vpcmpgtb %xmm0, %xmm1, %xmm0
; NoVLX-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1
; NoVLX-NEXT: vpxor %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0 ; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 {%k1} ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r10d ; NoVLX-NEXT: kmovw %k1, %r10d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 115 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: vmovdqa (%rsi), %xmm1 ; NoVLX-NEXT: vmovdqa (%rsi), %xmm1
; NoVLX-NEXT: vpcmpgtb %xmm0, %xmm1, %xmm0 ; NoVLX-NEXT: vpcmpgtb %xmm0, %xmm1, %xmm0
; NoVLX-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1
; NoVLX-NEXT: vpxor %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0 ; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 {%k1} ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r10d ; NoVLX-NEXT: kmovw %k1, %r10d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 114 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi943: ; NoVLX-NEXT: .Lcfi943:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: vpcmpgtb %xmm0, %xmm1, %xmm0 ; NoVLX-NEXT: vpcmpgtb %xmm0, %xmm1, %xmm0
; NoVLX-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1
; NoVLX-NEXT: vpxor %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0 ; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 116 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: vmovdqa (%rdi), %xmm1 ; NoVLX-NEXT: vmovdqa (%rdi), %xmm1
; NoVLX-NEXT: vpcmpgtb %xmm0, %xmm1, %xmm0 ; NoVLX-NEXT: vpcmpgtb %xmm0, %xmm1, %xmm0
; NoVLX-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1
; NoVLX-NEXT: vpxor %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0 ; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 118 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: vpcmpgtb %xmm0, %xmm1, %xmm0 ; NoVLX-NEXT: vpcmpgtb %xmm0, %xmm1, %xmm0
; NoVLX-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1
; NoVLX-NEXT: vpxor %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0 ; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 {%k1} ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 120 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: vmovdqa (%rsi), %xmm1 ; NoVLX-NEXT: vmovdqa (%rsi), %xmm1
; NoVLX-NEXT: vpcmpgtb %xmm0, %xmm1, %xmm0 ; NoVLX-NEXT: vpcmpgtb %xmm0, %xmm1, %xmm0
; NoVLX-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1
; NoVLX-NEXT: vpxor %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0 ; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 {%k1} ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 112 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0 ; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT: kmovw %k0, (%rsp) ; NoVLX-NEXT: kmovw %k0, (%rsp)
; NoVLX-NEXT: movl (%rsp), %ecx ; NoVLX-NEXT: movl (%rsp), %ecx
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: movl {{[0-9]+}}(%rsp), %eax ; NoVLX-NEXT: movl {{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT: shlq $32, %rax ; NoVLX-NEXT: shlq $32, %rax
; NoVLX-NEXT: orq %rcx, %rax ; NoVLX-NEXT: orq %rcx, %rax
; NoVLX-NEXT: movq %rbp, %rsp ; NoVLX-NEXT: movq %rbp, %rsp
; NoVLX-NEXT: popq %rbp ; NoVLX-NEXT: popq %rbp
; NoVLX-NEXT: retq ; NoVLX-NEXT: retq
Show All 35 Lines
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0 ; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT: kmovw %k0, (%rsp) ; NoVLX-NEXT: kmovw %k0, (%rsp)
; NoVLX-NEXT: movl (%rsp), %ecx ; NoVLX-NEXT: movl (%rsp), %ecx
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: movl {{[0-9]+}}(%rsp), %eax ; NoVLX-NEXT: movl {{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT: shlq $32, %rax ; NoVLX-NEXT: shlq $32, %rax
; NoVLX-NEXT: orq %rcx, %rax ; NoVLX-NEXT: orq %rcx, %rax
; NoVLX-NEXT: movq %rbp, %rsp ; NoVLX-NEXT: movq %rbp, %rsp
; NoVLX-NEXT: popq %rbp ; NoVLX-NEXT: popq %rbp
; NoVLX-NEXT: retq ; NoVLX-NEXT: retq
Show All 30 Lines
; NoVLX-NEXT: subq $96, %rsp ; NoVLX-NEXT: subq $96, %rsp
; NoVLX-NEXT: movl %edi, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: movl %edi, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw {{[0-9]+}}(%rsp), %k1 ; NoVLX-NEXT: kmovw {{[0-9]+}}(%rsp), %k1
; NoVLX-NEXT: kmovw {{[0-9]+}}(%rsp), %k2 ; NoVLX-NEXT: kmovw {{[0-9]+}}(%rsp), %k2
; NoVLX-NEXT: vpternlogd $255, %zmm2, %zmm2, %zmm2 {%k1} {z} ; NoVLX-NEXT: vpternlogd $255, %zmm2, %zmm2, %zmm2 {%k1} {z}
; NoVLX-NEXT: vpmovdb %zmm2, %xmm2 ; NoVLX-NEXT: vpmovdb %zmm2, %xmm2
; NoVLX-NEXT: vpternlogd $255, %zmm3, %zmm3, %zmm3 {%k2} {z} ; NoVLX-NEXT: vpternlogd $255, %zmm3, %zmm3, %zmm3 {%k2} {z}
; NoVLX-NEXT: vpmovdb %zmm3, %xmm3 ; NoVLX-NEXT: vpmovdb %zmm3, %xmm3
; NoVLX-NEXT: vxorps %xmm4, %xmm4, %xmm4
; NoVLX-NEXT: vpcmpgtb %ymm0, %ymm1, %ymm0 ; NoVLX-NEXT: vpcmpgtb %ymm0, %ymm1, %ymm0
; NoVLX-NEXT: vpcmpeqd %ymm1, %ymm1, %ymm1 ; NoVLX-NEXT: vpcmpeqd %ymm1, %ymm1, %ymm1
; NoVLX-NEXT: vpxor %ymm1, %ymm0, %ymm0 ; NoVLX-NEXT: vpxor %ymm1, %ymm0, %ymm0
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT: vpand %xmm3, %xmm1, %xmm1 ; NoVLX-NEXT: vpand %xmm3, %xmm1, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpand %xmm2, %xmm0, %xmm0 ; NoVLX-NEXT: vpand %xmm2, %xmm0, %xmm0
; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0 ; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT: kmovw %k0, (%rsp) ; NoVLX-NEXT: kmovw %k0, (%rsp)
; NoVLX-NEXT: movl (%rsp), %ecx ; NoVLX-NEXT: movl (%rsp), %ecx
; NoVLX-NEXT: vptestmd %zmm4, %zmm4, %k0 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: movl {{[0-9]+}}(%rsp), %eax ; NoVLX-NEXT: movl {{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT: shlq $32, %rax ; NoVLX-NEXT: shlq $32, %rax
; NoVLX-NEXT: orq %rcx, %rax ; NoVLX-NEXT: orq %rcx, %rax
; NoVLX-NEXT: movq %rbp, %rsp ; NoVLX-NEXT: movq %rbp, %rsp
; NoVLX-NEXT: popq %rbp ; NoVLX-NEXT: popq %rbp
; NoVLX-NEXT: retq ; NoVLX-NEXT: retq
Show All 31 Lines
; NoVLX-NEXT: subq $96, %rsp ; NoVLX-NEXT: subq $96, %rsp
; NoVLX-NEXT: movl %edi, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: movl %edi, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw {{[0-9]+}}(%rsp), %k1 ; NoVLX-NEXT: kmovw {{[0-9]+}}(%rsp), %k1
; NoVLX-NEXT: kmovw {{[0-9]+}}(%rsp), %k2 ; NoVLX-NEXT: kmovw {{[0-9]+}}(%rsp), %k2
; NoVLX-NEXT: vpternlogd $255, %zmm1, %zmm1, %zmm1 {%k1} {z} ; NoVLX-NEXT: vpternlogd $255, %zmm1, %zmm1, %zmm1 {%k1} {z}
; NoVLX-NEXT: vpmovdb %zmm1, %xmm1 ; NoVLX-NEXT: vpmovdb %zmm1, %xmm1
; NoVLX-NEXT: vpternlogd $255, %zmm2, %zmm2, %zmm2 {%k2} {z} ; NoVLX-NEXT: vpternlogd $255, %zmm2, %zmm2, %zmm2 {%k2} {z}
; NoVLX-NEXT: vpmovdb %zmm2, %xmm2 ; NoVLX-NEXT: vpmovdb %zmm2, %xmm2
; NoVLX-NEXT: vxorps %xmm3, %xmm3, %xmm3 ; NoVLX-NEXT: vmovdqa (%rsi), %ymm3
; NoVLX-NEXT: vmovdqa (%rsi), %ymm4 ; NoVLX-NEXT: vpcmpgtb %ymm0, %ymm3, %ymm0
; NoVLX-NEXT: vpcmpgtb %ymm0, %ymm4, %ymm0 ; NoVLX-NEXT: vpcmpeqd %ymm3, %ymm3, %ymm3
; NoVLX-NEXT: vpcmpeqd %ymm4, %ymm4, %ymm4 ; NoVLX-NEXT: vpxor %ymm3, %ymm0, %ymm0
; NoVLX-NEXT: vpxor %ymm4, %ymm0, %ymm0 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm3
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm4 ; NoVLX-NEXT: vpand %xmm2, %xmm3, %xmm2
; NoVLX-NEXT: vpand %xmm2, %xmm4, %xmm2
; NoVLX-NEXT: vpmovsxbd %xmm2, %zmm2 ; NoVLX-NEXT: vpmovsxbd %xmm2, %zmm2
; NoVLX-NEXT: vpslld $31, %zmm2, %zmm2 ; NoVLX-NEXT: vpslld $31, %zmm2, %zmm2
; NoVLX-NEXT: vptestmd %zmm2, %zmm2, %k0 ; NoVLX-NEXT: vptestmd %zmm2, %zmm2, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpand %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0 ; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT: kmovw %k0, (%rsp) ; NoVLX-NEXT: kmovw %k0, (%rsp)
; NoVLX-NEXT: movl (%rsp), %ecx ; NoVLX-NEXT: movl (%rsp), %ecx
; NoVLX-NEXT: vptestmd %zmm3, %zmm3, %k0 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: movl {{[0-9]+}}(%rsp), %eax ; NoVLX-NEXT: movl {{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT: shlq $32, %rax ; NoVLX-NEXT: shlq $32, %rax
; NoVLX-NEXT: orq %rcx, %rax ; NoVLX-NEXT: orq %rcx, %rax
; NoVLX-NEXT: movq %rbp, %rsp ; NoVLX-NEXT: movq %rbp, %rsp
; NoVLX-NEXT: popq %rbp ; NoVLX-NEXT: popq %rbp
; NoVLX-NEXT: retq ; NoVLX-NEXT: retq
▲ Show 20 Lines • Show All 163 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $32, %rsp ; NoVLX-NEXT: subq $32, %rsp
; NoVLX-NEXT: vpcmpgtw %xmm0, %xmm1, %xmm0 ; NoVLX-NEXT: vpcmpgtw %xmm0, %xmm1, %xmm0
; NoVLX-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1
; NoVLX-NEXT: vpxor %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0 ; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0
; NoVLX-NEXT: vpsllq $63, %zmm0, %zmm0 ; NoVLX-NEXT: vpsllq $63, %zmm0, %zmm0
; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0 ; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r9d ; NoVLX-NEXT: kmovw %k1, %r9d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 61 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: subq $32, %rsp ; NoVLX-NEXT: subq $32, %rsp
; NoVLX-NEXT: vmovdqa (%rdi), %xmm1 ; NoVLX-NEXT: vmovdqa (%rdi), %xmm1
; NoVLX-NEXT: vpcmpgtw %xmm0, %xmm1, %xmm0 ; NoVLX-NEXT: vpcmpgtw %xmm0, %xmm1, %xmm0
; NoVLX-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1
; NoVLX-NEXT: vpxor %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0 ; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0
; NoVLX-NEXT: vpsllq $63, %zmm0, %zmm0 ; NoVLX-NEXT: vpsllq $63, %zmm0, %zmm0
; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0 ; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r9d ; NoVLX-NEXT: kmovw %k1, %r9d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 63 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: subq $32, %rsp ; NoVLX-NEXT: subq $32, %rsp
; NoVLX-NEXT: vpcmpgtw %xmm0, %xmm1, %xmm0 ; NoVLX-NEXT: vpcmpgtw %xmm0, %xmm1, %xmm0
; NoVLX-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1
; NoVLX-NEXT: vpxor %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0 ; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0
; NoVLX-NEXT: vpsllq $63, %zmm0, %zmm0 ; NoVLX-NEXT: vpsllq $63, %zmm0, %zmm0
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0 {%k1} ; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r9d ; NoVLX-NEXT: kmovw %k1, %r9d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 65 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: vmovdqa (%rsi), %xmm1 ; NoVLX-NEXT: vmovdqa (%rsi), %xmm1
; NoVLX-NEXT: vpcmpgtw %xmm0, %xmm1, %xmm0 ; NoVLX-NEXT: vpcmpgtw %xmm0, %xmm1, %xmm0
; NoVLX-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1
; NoVLX-NEXT: vpxor %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0 ; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0
; NoVLX-NEXT: vpsllq $63, %zmm0, %zmm0 ; NoVLX-NEXT: vpsllq $63, %zmm0, %zmm0
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0 {%k1} ; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r9d ; NoVLX-NEXT: kmovw %k1, %r9d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 64 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: vpcmpgtw %xmm0, %xmm1, %xmm0 ; NoVLX-NEXT: vpcmpgtw %xmm0, %xmm1, %xmm0
; NoVLX-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1
; NoVLX-NEXT: vpxor %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0 ; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0
; NoVLX-NEXT: vpsllq $63, %zmm0, %zmm0 ; NoVLX-NEXT: vpsllq $63, %zmm0, %zmm0
; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0 ; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 66 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: vmovdqa (%rdi), %xmm1 ; NoVLX-NEXT: vmovdqa (%rdi), %xmm1
; NoVLX-NEXT: vpcmpgtw %xmm0, %xmm1, %xmm0 ; NoVLX-NEXT: vpcmpgtw %xmm0, %xmm1, %xmm0
; NoVLX-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1
; NoVLX-NEXT: vpxor %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0 ; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0
; NoVLX-NEXT: vpsllq $63, %zmm0, %zmm0 ; NoVLX-NEXT: vpsllq $63, %zmm0, %zmm0
; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0 ; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 68 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: vpcmpgtw %xmm0, %xmm1, %xmm0 ; NoVLX-NEXT: vpcmpgtw %xmm0, %xmm1, %xmm0
; NoVLX-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1
; NoVLX-NEXT: vpxor %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0 ; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0
; NoVLX-NEXT: vpsllq $63, %zmm0, %zmm0 ; NoVLX-NEXT: vpsllq $63, %zmm0, %zmm0
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0 {%k1} ; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 70 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: vmovdqa (%rsi), %xmm1 ; NoVLX-NEXT: vmovdqa (%rsi), %xmm1
; NoVLX-NEXT: vpcmpgtw %xmm0, %xmm1, %xmm0 ; NoVLX-NEXT: vpcmpgtw %xmm0, %xmm1, %xmm0
; NoVLX-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1
; NoVLX-NEXT: vpxor %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0 ; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0
; NoVLX-NEXT: vpsllq $63, %zmm0, %zmm0 ; NoVLX-NEXT: vpsllq $63, %zmm0, %zmm0
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0 {%k1} ; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 85 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi1011: ; NoVLX-NEXT: .Lcfi1011:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: vpcmpgtw %ymm0, %ymm1, %ymm0 ; NoVLX-NEXT: vpcmpgtw %ymm0, %ymm1, %ymm0
; NoVLX-NEXT: vpcmpeqd %ymm1, %ymm1, %ymm1 ; NoVLX-NEXT: vpcmpeqd %ymm1, %ymm1, %ymm1
; NoVLX-NEXT: vpxor %ymm1, %ymm0, %ymm0 ; NoVLX-NEXT: vpxor %ymm1, %ymm0, %ymm0
; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0 ; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0 ; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r10d ; NoVLX-NEXT: kmovw %k1, %r10d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 112 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: vmovdqa (%rdi), %ymm1 ; NoVLX-NEXT: vmovdqa (%rdi), %ymm1
; NoVLX-NEXT: vpcmpgtw %ymm0, %ymm1, %ymm0 ; NoVLX-NEXT: vpcmpgtw %ymm0, %ymm1, %ymm0
; NoVLX-NEXT: vpcmpeqd %ymm1, %ymm1, %ymm1 ; NoVLX-NEXT: vpcmpeqd %ymm1, %ymm1, %ymm1
; NoVLX-NEXT: vpxor %ymm1, %ymm0, %ymm0 ; NoVLX-NEXT: vpxor %ymm1, %ymm0, %ymm0
; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0 ; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0 ; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r10d ; NoVLX-NEXT: kmovw %k1, %r10d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 114 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: vpcmpgtw %ymm0, %ymm1, %ymm0 ; NoVLX-NEXT: vpcmpgtw %ymm0, %ymm1, %ymm0
; NoVLX-NEXT: vpcmpeqd %ymm1, %ymm1, %ymm1 ; NoVLX-NEXT: vpcmpeqd %ymm1, %ymm1, %ymm1
; NoVLX-NEXT: vpxor %ymm1, %ymm0, %ymm0 ; NoVLX-NEXT: vpxor %ymm1, %ymm0, %ymm0
; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0 ; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0 ; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 {%k1} ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r10d ; NoVLX-NEXT: kmovw %k1, %r10d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 116 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: vmovdqa (%rsi), %ymm1 ; NoVLX-NEXT: vmovdqa (%rsi), %ymm1
; NoVLX-NEXT: vpcmpgtw %ymm0, %ymm1, %ymm0 ; NoVLX-NEXT: vpcmpgtw %ymm0, %ymm1, %ymm0
; NoVLX-NEXT: vpcmpeqd %ymm1, %ymm1, %ymm1 ; NoVLX-NEXT: vpcmpeqd %ymm1, %ymm1, %ymm1
; NoVLX-NEXT: vpxor %ymm1, %ymm0, %ymm0 ; NoVLX-NEXT: vpxor %ymm1, %ymm0, %ymm0
; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0 ; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0 ; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 {%k1} ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r10d ; NoVLX-NEXT: kmovw %k1, %r10d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 115 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi1043: ; NoVLX-NEXT: .Lcfi1043:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: vpcmpgtw %ymm0, %ymm1, %ymm0 ; NoVLX-NEXT: vpcmpgtw %ymm0, %ymm1, %ymm0
; NoVLX-NEXT: vpcmpeqd %ymm1, %ymm1, %ymm1 ; NoVLX-NEXT: vpcmpeqd %ymm1, %ymm1, %ymm1
; NoVLX-NEXT: vpxor %ymm1, %ymm0, %ymm0 ; NoVLX-NEXT: vpxor %ymm1, %ymm0, %ymm0
; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0 ; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0 ; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 117 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: vmovdqa (%rdi), %ymm1 ; NoVLX-NEXT: vmovdqa (%rdi), %ymm1
; NoVLX-NEXT: vpcmpgtw %ymm0, %ymm1, %ymm0 ; NoVLX-NEXT: vpcmpgtw %ymm0, %ymm1, %ymm0
; NoVLX-NEXT: vpcmpeqd %ymm1, %ymm1, %ymm1 ; NoVLX-NEXT: vpcmpeqd %ymm1, %ymm1, %ymm1
; NoVLX-NEXT: vpxor %ymm1, %ymm0, %ymm0 ; NoVLX-NEXT: vpxor %ymm1, %ymm0, %ymm0
; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0 ; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0 ; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 119 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: vpcmpgtw %ymm0, %ymm1, %ymm0 ; NoVLX-NEXT: vpcmpgtw %ymm0, %ymm1, %ymm0
; NoVLX-NEXT: vpcmpeqd %ymm1, %ymm1, %ymm1 ; NoVLX-NEXT: vpcmpeqd %ymm1, %ymm1, %ymm1
; NoVLX-NEXT: vpxor %ymm1, %ymm0, %ymm0 ; NoVLX-NEXT: vpxor %ymm1, %ymm0, %ymm0
; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0 ; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0 ; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 {%k1} ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 121 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: vmovdqa (%rsi), %ymm1 ; NoVLX-NEXT: vmovdqa (%rsi), %ymm1
; NoVLX-NEXT: vpcmpgtw %ymm0, %ymm1, %ymm0 ; NoVLX-NEXT: vpcmpgtw %ymm0, %ymm1, %ymm0
; NoVLX-NEXT: vpcmpeqd %ymm1, %ymm1, %ymm1 ; NoVLX-NEXT: vpcmpeqd %ymm1, %ymm1, %ymm1
; NoVLX-NEXT: vpxor %ymm1, %ymm0, %ymm0 ; NoVLX-NEXT: vpxor %ymm1, %ymm0, %ymm0
; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0 ; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0 ; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 {%k1} ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 415 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: kshiftrw $15, %k0, %k0 ; NoVLX-NEXT: kshiftrw $15, %k0, %k0
; NoVLX-NEXT: kmovw %k0, %eax ; NoVLX-NEXT: kmovw %k0, %eax
; NoVLX-NEXT: vpinsrb $15, %eax, %xmm0, %xmm0 ; NoVLX-NEXT: vpinsrb $15, %eax, %xmm0, %xmm0
; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0 ; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT: kmovw %k0, (%rsp) ; NoVLX-NEXT: kmovw %k0, (%rsp)
; NoVLX-NEXT: movl (%rsp), %ecx ; NoVLX-NEXT: movl (%rsp), %ecx
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: movl {{[0-9]+}}(%rsp), %eax ; NoVLX-NEXT: movl {{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT: shlq $32, %rax ; NoVLX-NEXT: shlq $32, %rax
; NoVLX-NEXT: orq %rcx, %rax ; NoVLX-NEXT: orq %rcx, %rax
; NoVLX-NEXT: movq %rbp, %rsp ; NoVLX-NEXT: movq %rbp, %rsp
; NoVLX-NEXT: popq %rbp ; NoVLX-NEXT: popq %rbp
; NoVLX-NEXT: retq ; NoVLX-NEXT: retq
▲ Show 20 Lines • Show All 254 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: kshiftrw $15, %k0, %k0 ; NoVLX-NEXT: kshiftrw $15, %k0, %k0
; NoVLX-NEXT: kmovw %k0, %eax ; NoVLX-NEXT: kmovw %k0, %eax
; NoVLX-NEXT: vpinsrb $15, %eax, %xmm0, %xmm0 ; NoVLX-NEXT: vpinsrb $15, %eax, %xmm0, %xmm0
; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0 ; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT: kmovw %k0, (%rsp) ; NoVLX-NEXT: kmovw %k0, (%rsp)
; NoVLX-NEXT: movl (%rsp), %ecx ; NoVLX-NEXT: movl (%rsp), %ecx
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: movl {{[0-9]+}}(%rsp), %eax ; NoVLX-NEXT: movl {{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT: shlq $32, %rax ; NoVLX-NEXT: shlq $32, %rax
; NoVLX-NEXT: orq %rcx, %rax ; NoVLX-NEXT: orq %rcx, %rax
; NoVLX-NEXT: movq %rbp, %rsp ; NoVLX-NEXT: movq %rbp, %rsp
; NoVLX-NEXT: popq %rbp ; NoVLX-NEXT: popq %rbp
; NoVLX-NEXT: retq ; NoVLX-NEXT: retq
▲ Show 20 Lines • Show All 180 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: shrl $16, %eax ; NoVLX-NEXT: shrl $16, %eax
; NoVLX-NEXT: vmovd %ecx, %xmm2 ; NoVLX-NEXT: vmovd %ecx, %xmm2
; NoVLX-NEXT: vpinsrw $1, %eax, %xmm2, %xmm2 ; NoVLX-NEXT: vpinsrw $1, %eax, %xmm2, %xmm2
; NoVLX-NEXT: movq %rcx, %rax ; NoVLX-NEXT: movq %rcx, %rax
; NoVLX-NEXT: shrq $32, %rax ; NoVLX-NEXT: shrq $32, %rax
; NoVLX-NEXT: vpinsrw $2, %eax, %xmm2, %xmm2 ; NoVLX-NEXT: vpinsrw $2, %eax, %xmm2, %xmm2
; NoVLX-NEXT: vpextrq $1, %xmm1, %rax ; NoVLX-NEXT: vpextrq $1, %xmm1, %rax
; NoVLX-NEXT: vinserti128 $1, %xmm9, %ymm4, %ymm1 ; NoVLX-NEXT: vinserti128 $1, %xmm9, %ymm4, %ymm1
; NoVLX-NEXT: vinserti128 $1, %xmm6, %ymm0, %ymm8 ; NoVLX-NEXT: vinserti128 $1, %xmm6, %ymm0, %ymm4
; NoVLX-NEXT: kmovw {{[0-9]+}}(%rsp), %k1 ; NoVLX-NEXT: kmovw {{[0-9]+}}(%rsp), %k1
; NoVLX-NEXT: kmovw {{[0-9]+}}(%rsp), %k2 ; NoVLX-NEXT: kmovw {{[0-9]+}}(%rsp), %k2
; NoVLX-NEXT: vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT: vinserti128 $1, %xmm7, %ymm3, %ymm3 ; NoVLX-NEXT: vinserti128 $1, %xmm7, %ymm3, %ymm3
; NoVLX-NEXT: vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT: vpmovdb %zmm0, %xmm0 ; NoVLX-NEXT: vpmovdb %zmm0, %xmm0
; NoVLX-NEXT: vpternlogd $255, %zmm6, %zmm6, %zmm6 {%k2} {z} ; NoVLX-NEXT: vpcmpgtw %ymm1, %ymm3, %ymm3
; NoVLX-NEXT: vpcmpgtw %ymm1, %ymm3, %ymm4 ; NoVLX-NEXT: vpternlogd $255, %zmm1, %zmm1, %zmm1 {%k2} {z}
; NoVLX-NEXT: vpmovdb %zmm6, %xmm6 ; NoVLX-NEXT: vpmovdb %zmm1, %xmm1
; NoVLX-NEXT: shrq $48, %rcx ; NoVLX-NEXT: shrq $48, %rcx
; NoVLX-NEXT: vpinsrw $3, %ecx, %xmm2, %xmm1 ; NoVLX-NEXT: vpinsrw $3, %ecx, %xmm2, %xmm2
; NoVLX-NEXT: movl %eax, %ecx ; NoVLX-NEXT: movl %eax, %ecx
; NoVLX-NEXT: shrl $16, %ecx ; NoVLX-NEXT: shrl $16, %ecx
; NoVLX-NEXT: vpinsrw $4, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrw $4, %eax, %xmm2, %xmm2
; NoVLX-NEXT: vpinsrw $5, %ecx, %xmm1, %xmm2 ; NoVLX-NEXT: vpinsrw $5, %ecx, %xmm2, %xmm2
; NoVLX-NEXT: movq %rax, %rcx ; NoVLX-NEXT: movq %rax, %rcx
; NoVLX-NEXT: shrq $48, %rax ; NoVLX-NEXT: shrq $48, %rax
; NoVLX-NEXT: shrq $32, %rcx ; NoVLX-NEXT: shrq $32, %rcx
; NoVLX-NEXT: vpxor %xmm1, %xmm1, %xmm1
; NoVLX-NEXT: vpinsrw $6, %ecx, %xmm2, %xmm2 ; NoVLX-NEXT: vpinsrw $6, %ecx, %xmm2, %xmm2
; NoVLX-NEXT: vpinsrw $7, %eax, %xmm2, %xmm2 ; NoVLX-NEXT: vpinsrw $7, %eax, %xmm2, %xmm2
; NoVLX-NEXT: vinserti128 $1, %xmm5, %ymm2, %ymm2 ; NoVLX-NEXT: vinserti128 $1, %xmm5, %ymm2, %ymm2
; NoVLX-NEXT: vpcmpgtw %ymm8, %ymm2, %ymm2 ; NoVLX-NEXT: vpcmpgtw %ymm4, %ymm2, %ymm2
; NoVLX-NEXT: vpcmpeqd %ymm5, %ymm5, %ymm5 ; NoVLX-NEXT: vpcmpeqd %ymm4, %ymm4, %ymm4
; NoVLX-NEXT: vpxor %ymm5, %ymm2, %ymm2 ; NoVLX-NEXT: vpxor %ymm4, %ymm2, %ymm2
; NoVLX-NEXT: vpmovsxwd %ymm2, %zmm2 ; NoVLX-NEXT: vpmovsxwd %ymm2, %zmm2
; NoVLX-NEXT: vpslld $31, %zmm2, %zmm2 ; NoVLX-NEXT: vpslld $31, %zmm2, %zmm2
; NoVLX-NEXT: vptestmd %zmm2, %zmm2, %k0 ; NoVLX-NEXT: vptestmd %zmm2, %zmm2, %k0
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 49 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $13, %eax, %xmm2, %xmm2 ; NoVLX-NEXT: vpinsrb $13, %eax, %xmm2, %xmm2
; NoVLX-NEXT: kshiftlw $1, %k0, %k1 ; NoVLX-NEXT: kshiftlw $1, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $14, %eax, %xmm2, %xmm2 ; NoVLX-NEXT: vpinsrb $14, %eax, %xmm2, %xmm2
; NoVLX-NEXT: kshiftrw $15, %k0, %k0 ; NoVLX-NEXT: kshiftrw $15, %k0, %k0
; NoVLX-NEXT: kmovw %k0, %eax ; NoVLX-NEXT: kmovw %k0, %eax
; NoVLX-NEXT: vpinsrb $15, %eax, %xmm2, %xmm3 ; NoVLX-NEXT: vpinsrb $15, %eax, %xmm2, %xmm2
; NoVLX-NEXT: vpxor %ymm5, %ymm4, %ymm2 ; NoVLX-NEXT: vpxor %ymm4, %ymm3, %ymm3
; NoVLX-NEXT: vpmovsxwd %ymm2, %zmm2 ; NoVLX-NEXT: vpmovsxwd %ymm3, %zmm3
; NoVLX-NEXT: vpslld $31, %zmm2, %zmm2 ; NoVLX-NEXT: vpslld $31, %zmm3, %zmm3
; NoVLX-NEXT: vptestmd %zmm2, %zmm2, %k0 ; NoVLX-NEXT: vptestmd %zmm3, %zmm3, %k0
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %ecx ; NoVLX-NEXT: kmovw %k1, %ecx
; NoVLX-NEXT: vmovd %ecx, %xmm2 ; NoVLX-NEXT: vmovd %ecx, %xmm3
; NoVLX-NEXT: vpinsrb $1, %eax, %xmm2, %xmm2 ; NoVLX-NEXT: vpinsrb $1, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $2, %eax, %xmm2, %xmm2 ; NoVLX-NEXT: vpinsrb $2, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftlw $12, %k0, %k1 ; NoVLX-NEXT: kshiftlw $12, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $3, %eax, %xmm2, %xmm2 ; NoVLX-NEXT: vpinsrb $3, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftlw $11, %k0, %k1 ; NoVLX-NEXT: kshiftlw $11, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $4, %eax, %xmm2, %xmm2 ; NoVLX-NEXT: vpinsrb $4, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftlw $10, %k0, %k1 ; NoVLX-NEXT: kshiftlw $10, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $5, %eax, %xmm2, %xmm2 ; NoVLX-NEXT: vpinsrb $5, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftlw $9, %k0, %k1 ; NoVLX-NEXT: kshiftlw $9, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $6, %eax, %xmm2, %xmm2 ; NoVLX-NEXT: vpinsrb $6, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftlw $8, %k0, %k1 ; NoVLX-NEXT: kshiftlw $8, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $7, %eax, %xmm2, %xmm2 ; NoVLX-NEXT: vpinsrb $7, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftlw $7, %k0, %k1 ; NoVLX-NEXT: kshiftlw $7, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $8, %eax, %xmm2, %xmm2 ; NoVLX-NEXT: vpinsrb $8, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftlw $6, %k0, %k1 ; NoVLX-NEXT: kshiftlw $6, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $9, %eax, %xmm2, %xmm2 ; NoVLX-NEXT: vpinsrb $9, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftlw $5, %k0, %k1 ; NoVLX-NEXT: kshiftlw $5, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $10, %eax, %xmm2, %xmm2 ; NoVLX-NEXT: vpinsrb $10, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftlw $4, %k0, %k1 ; NoVLX-NEXT: kshiftlw $4, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $11, %eax, %xmm2, %xmm2 ; NoVLX-NEXT: vpinsrb $11, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftlw $3, %k0, %k1 ; NoVLX-NEXT: kshiftlw $3, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $12, %eax, %xmm2, %xmm2 ; NoVLX-NEXT: vpinsrb $12, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftlw $2, %k0, %k1 ; NoVLX-NEXT: kshiftlw $2, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $13, %eax, %xmm2, %xmm2 ; NoVLX-NEXT: vpinsrb $13, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftlw $1, %k0, %k1 ; NoVLX-NEXT: kshiftlw $1, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $14, %eax, %xmm2, %xmm2 ; NoVLX-NEXT: vpinsrb $14, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftrw $15, %k0, %k0 ; NoVLX-NEXT: kshiftrw $15, %k0, %k0
; NoVLX-NEXT: kmovw %k0, %eax ; NoVLX-NEXT: kmovw %k0, %eax
; NoVLX-NEXT: vpinsrb $15, %eax, %xmm2, %xmm2 ; NoVLX-NEXT: vpinsrb $15, %eax, %xmm3, %xmm3
; NoVLX-NEXT: vpand %xmm6, %xmm2, %xmm2 ; NoVLX-NEXT: vpand %xmm1, %xmm3, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm2, %zmm2 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm2, %zmm2 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm2, %zmm2, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpand %xmm0, %xmm3, %xmm0 ; NoVLX-NEXT: vpand %xmm0, %xmm2, %xmm0
; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0 ; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT: kmovw %k0, (%rsp) ; NoVLX-NEXT: kmovw %k0, (%rsp)
; NoVLX-NEXT: movl (%rsp), %ecx ; NoVLX-NEXT: movl (%rsp), %ecx
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: movl {{[0-9]+}}(%rsp), %eax ; NoVLX-NEXT: movl {{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT: shlq $32, %rax ; NoVLX-NEXT: shlq $32, %rax
; NoVLX-NEXT: orq %rcx, %rax ; NoVLX-NEXT: orq %rcx, %rax
; NoVLX-NEXT: movq %rbp, %rsp ; NoVLX-NEXT: movq %rbp, %rsp
; NoVLX-NEXT: popq %rbp ; NoVLX-NEXT: popq %rbp
; NoVLX-NEXT: retq ; NoVLX-NEXT: retq
▲ Show 20 Lines • Show All 48 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: shrl $16, %ecx ; NoVLX-NEXT: shrl $16, %ecx
; NoVLX-NEXT: vpinsrw $4, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrw $4, %eax, %xmm1, %xmm1
; NoVLX-NEXT: vpinsrw $5, %ecx, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrw $5, %ecx, %xmm1, %xmm1
; NoVLX-NEXT: movq %rax, %rcx ; NoVLX-NEXT: movq %rax, %rcx
; NoVLX-NEXT: shrq $32, %rcx ; NoVLX-NEXT: shrq $32, %rcx
; NoVLX-NEXT: vpinsrw $6, %ecx, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrw $6, %ecx, %xmm1, %xmm1
; NoVLX-NEXT: vmovq %xmm4, %rcx ; NoVLX-NEXT: vmovq %xmm4, %rcx
; NoVLX-NEXT: shrq $48, %rax ; NoVLX-NEXT: shrq $48, %rax
; NoVLX-NEXT: vpinsrw $7, %eax, %xmm1, %xmm2 ; NoVLX-NEXT: vpinsrw $7, %eax, %xmm1, %xmm1
; NoVLX-NEXT: movl %ecx, %eax ; NoVLX-NEXT: movl %ecx, %eax
; NoVLX-NEXT: shrl $16, %eax ; NoVLX-NEXT: shrl $16, %eax
; NoVLX-NEXT: vmovd %ecx, %xmm1 ; NoVLX-NEXT: vmovd %ecx, %xmm2
; NoVLX-NEXT: vpinsrw $1, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrw $1, %eax, %xmm2, %xmm2
; NoVLX-NEXT: movq %rcx, %rax ; NoVLX-NEXT: movq %rcx, %rax
; NoVLX-NEXT: shrq $32, %rax ; NoVLX-NEXT: shrq $32, %rax
; NoVLX-NEXT: vpinsrw $2, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrw $2, %eax, %xmm2, %xmm2
; NoVLX-NEXT: vpextrq $1, %xmm4, %rax ; NoVLX-NEXT: vpextrq $1, %xmm4, %rax
; NoVLX-NEXT: shrq $48, %rcx ; NoVLX-NEXT: shrq $48, %rcx
; NoVLX-NEXT: vpinsrw $3, %ecx, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrw $3, %ecx, %xmm2, %xmm2
; NoVLX-NEXT: movl %eax, %ecx ; NoVLX-NEXT: movl %eax, %ecx
; NoVLX-NEXT: shrl $16, %ecx ; NoVLX-NEXT: shrl $16, %ecx
; NoVLX-NEXT: vpinsrw $4, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrw $4, %eax, %xmm2, %xmm2
; NoVLX-NEXT: vpinsrw $5, %ecx, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrw $5, %ecx, %xmm2, %xmm2
; NoVLX-NEXT: movq %rax, %rcx ; NoVLX-NEXT: movq %rax, %rcx
; NoVLX-NEXT: shrq $32, %rcx ; NoVLX-NEXT: shrq $32, %rcx
; NoVLX-NEXT: vpinsrw $6, %ecx, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrw $6, %ecx, %xmm2, %xmm2
; NoVLX-NEXT: vmovq %xmm3, %rcx ; NoVLX-NEXT: vmovq %xmm3, %rcx
; NoVLX-NEXT: shrq $48, %rax ; NoVLX-NEXT: shrq $48, %rax
; NoVLX-NEXT: vpinsrw $7, %eax, %xmm1, %xmm4 ; NoVLX-NEXT: vpinsrw $7, %eax, %xmm2, %xmm2
; NoVLX-NEXT: movl %ecx, %eax ; NoVLX-NEXT: movl %ecx, %eax
; NoVLX-NEXT: shrl $16, %eax ; NoVLX-NEXT: shrl $16, %eax
; NoVLX-NEXT: vmovd %ecx, %xmm1 ; NoVLX-NEXT: vmovd %ecx, %xmm4
; NoVLX-NEXT: vpinsrw $1, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrw $1, %eax, %xmm4, %xmm4
; NoVLX-NEXT: movq %rcx, %rax ; NoVLX-NEXT: movq %rcx, %rax
; NoVLX-NEXT: shrq $32, %rax ; NoVLX-NEXT: shrq $32, %rax
; NoVLX-NEXT: vpinsrw $2, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrw $2, %eax, %xmm4, %xmm4
; NoVLX-NEXT: vpextrq $1, %xmm3, %rax ; NoVLX-NEXT: vpextrq $1, %xmm3, %rax
; NoVLX-NEXT: shrq $48, %rcx ; NoVLX-NEXT: shrq $48, %rcx
; NoVLX-NEXT: vpinsrw $3, %ecx, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrw $3, %ecx, %xmm4, %xmm3
; NoVLX-NEXT: movl %eax, %ecx ; NoVLX-NEXT: movl %eax, %ecx
; NoVLX-NEXT: shrl $16, %ecx ; NoVLX-NEXT: shrl $16, %ecx
; NoVLX-NEXT: vpinsrw $4, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrw $4, %eax, %xmm3, %xmm3
; NoVLX-NEXT: vpinsrw $5, %ecx, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrw $5, %ecx, %xmm3, %xmm3
; NoVLX-NEXT: movq %rax, %rcx ; NoVLX-NEXT: movq %rax, %rcx
; NoVLX-NEXT: shrq $32, %rcx ; NoVLX-NEXT: shrq $32, %rcx
; NoVLX-NEXT: vpinsrw $6, %ecx, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrw $6, %ecx, %xmm3, %xmm3
; NoVLX-NEXT: vmovq %xmm0, %rcx ; NoVLX-NEXT: vmovq %xmm0, %rcx
; NoVLX-NEXT: shrq $48, %rax ; NoVLX-NEXT: shrq $48, %rax
; NoVLX-NEXT: vpinsrw $7, %eax, %xmm1, %xmm3 ; NoVLX-NEXT: vpinsrw $7, %eax, %xmm3, %xmm3
; NoVLX-NEXT: movl %ecx, %eax ; NoVLX-NEXT: movl %ecx, %eax
; NoVLX-NEXT: shrl $16, %eax ; NoVLX-NEXT: shrl $16, %eax
; NoVLX-NEXT: vmovd %ecx, %xmm1 ; NoVLX-NEXT: vmovd %ecx, %xmm4
; NoVLX-NEXT: vpinsrw $1, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrw $1, %eax, %xmm4, %xmm4
; NoVLX-NEXT: movq %rcx, %rax ; NoVLX-NEXT: movq %rcx, %rax
; NoVLX-NEXT: shrq $32, %rax ; NoVLX-NEXT: shrq $32, %rax
; NoVLX-NEXT: vpinsrw $2, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrw $2, %eax, %xmm4, %xmm4
; NoVLX-NEXT: vpextrq $1, %xmm0, %rax ; NoVLX-NEXT: vpextrq $1, %xmm0, %rax
; NoVLX-NEXT: shrq $48, %rcx ; NoVLX-NEXT: shrq $48, %rcx
; NoVLX-NEXT: vpinsrw $3, %ecx, %xmm1, %xmm0 ; NoVLX-NEXT: vpinsrw $3, %ecx, %xmm4, %xmm0
; NoVLX-NEXT: movl %eax, %ecx ; NoVLX-NEXT: movl %eax, %ecx
; NoVLX-NEXT: shrl $16, %ecx ; NoVLX-NEXT: shrl $16, %ecx
; NoVLX-NEXT: vpinsrw $4, %eax, %xmm0, %xmm0 ; NoVLX-NEXT: vpinsrw $4, %eax, %xmm0, %xmm0
; NoVLX-NEXT: vpinsrw $5, %ecx, %xmm0, %xmm0 ; NoVLX-NEXT: vpinsrw $5, %ecx, %xmm0, %xmm0
; NoVLX-NEXT: movq %rax, %rcx ; NoVLX-NEXT: movq %rax, %rcx
; NoVLX-NEXT: shrq $32, %rcx ; NoVLX-NEXT: shrq $32, %rcx
; NoVLX-NEXT: vpinsrw $6, %ecx, %xmm0, %xmm5 ; NoVLX-NEXT: vpinsrw $6, %ecx, %xmm0, %xmm4
; NoVLX-NEXT: kmovw {{[0-9]+}}(%rsp), %k1 ; NoVLX-NEXT: kmovw {{[0-9]+}}(%rsp), %k1
; NoVLX-NEXT: kmovw {{[0-9]+}}(%rsp), %k2 ; NoVLX-NEXT: kmovw {{[0-9]+}}(%rsp), %k2
; NoVLX-NEXT: vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z} ; NoVLX-NEXT: vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT: vpmovdb %zmm0, %xmm1 ; NoVLX-NEXT: vpmovdb %zmm0, %xmm0
; NoVLX-NEXT: vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k2} {z} ; NoVLX-NEXT: vinserti128 $1, %xmm1, %ymm2, %ymm2
; NoVLX-NEXT: vinserti128 $1, %xmm2, %ymm4, %ymm4 ; NoVLX-NEXT: vpternlogd $255, %zmm1, %zmm1, %zmm1 {%k2} {z}
; NoVLX-NEXT: vpmovdb %zmm0, %xmm2 ; NoVLX-NEXT: vpmovdb %zmm1, %xmm1
; NoVLX-NEXT: shrq $48, %rax ; NoVLX-NEXT: shrq $48, %rax
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: vpinsrw $7, %eax, %xmm4, %xmm4
; NoVLX-NEXT: vpinsrw $7, %eax, %xmm5, %xmm5 ; NoVLX-NEXT: vinserti128 $1, %xmm3, %ymm4, %ymm3
; NoVLX-NEXT: vinserti128 $1, %xmm3, %ymm5, %ymm3 ; NoVLX-NEXT: vmovdqa (%rsi), %ymm4
; NoVLX-NEXT: vmovdqa (%rsi), %ymm5 ; NoVLX-NEXT: vpcmpgtw %ymm3, %ymm4, %ymm5
; NoVLX-NEXT: vpcmpgtw %ymm3, %ymm5, %ymm3 ; NoVLX-NEXT: vmovdqa 32(%rsi), %ymm3
; NoVLX-NEXT: vmovdqa 32(%rsi), %ymm5 ; NoVLX-NEXT: vpcmpgtw %ymm2, %ymm3, %ymm3
; NoVLX-NEXT: vpcmpgtw %ymm4, %ymm5, %ymm4 ; NoVLX-NEXT: vpcmpeqd %ymm4, %ymm4, %ymm4
; NoVLX-NEXT: vpcmpeqd %ymm5, %ymm5, %ymm5 ; NoVLX-NEXT: vpxor %ymm4, %ymm5, %ymm2
; NoVLX-NEXT: vpxor %ymm5, %ymm3, %ymm3 ; NoVLX-NEXT: vpmovsxwd %ymm2, %zmm2
; NoVLX-NEXT: vpmovsxwd %ymm3, %zmm3 ; NoVLX-NEXT: vpslld $31, %zmm2, %zmm2
; NoVLX-NEXT: vpslld $31, %zmm3, %zmm3 ; NoVLX-NEXT: vptestmd %zmm2, %zmm2, %k0
; NoVLX-NEXT: vptestmd %zmm3, %zmm3, %k0
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %ecx ; NoVLX-NEXT: kmovw %k1, %ecx
; NoVLX-NEXT: vmovd %ecx, %xmm3 ; NoVLX-NEXT: vmovd %ecx, %xmm2
; NoVLX-NEXT: vpinsrb $1, %eax, %xmm3, %xmm3 ; NoVLX-NEXT: vpinsrb $1, %eax, %xmm2, %xmm2
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $2, %eax, %xmm3, %xmm3 ; NoVLX-NEXT: vpinsrb $2, %eax, %xmm2, %xmm2
; NoVLX-NEXT: kshiftlw $12, %k0, %k1 ; NoVLX-NEXT: kshiftlw $12, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $3, %eax, %xmm3, %xmm3 ; NoVLX-NEXT: vpinsrb $3, %eax, %xmm2, %xmm2
; NoVLX-NEXT: kshiftlw $11, %k0, %k1 ; NoVLX-NEXT: kshiftlw $11, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $4, %eax, %xmm3, %xmm3 ; NoVLX-NEXT: vpinsrb $4, %eax, %xmm2, %xmm2
; NoVLX-NEXT: kshiftlw $10, %k0, %k1 ; NoVLX-NEXT: kshiftlw $10, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $5, %eax, %xmm3, %xmm3 ; NoVLX-NEXT: vpinsrb $5, %eax, %xmm2, %xmm2
; NoVLX-NEXT: kshiftlw $9, %k0, %k1 ; NoVLX-NEXT: kshiftlw $9, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $6, %eax, %xmm3, %xmm3 ; NoVLX-NEXT: vpinsrb $6, %eax, %xmm2, %xmm2
; NoVLX-NEXT: kshiftlw $8, %k0, %k1 ; NoVLX-NEXT: kshiftlw $8, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $7, %eax, %xmm3, %xmm3 ; NoVLX-NEXT: vpinsrb $7, %eax, %xmm2, %xmm2
; NoVLX-NEXT: kshiftlw $7, %k0, %k1 ; NoVLX-NEXT: kshiftlw $7, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $8, %eax, %xmm3, %xmm3 ; NoVLX-NEXT: vpinsrb $8, %eax, %xmm2, %xmm2
; NoVLX-NEXT: kshiftlw $6, %k0, %k1 ; NoVLX-NEXT: kshiftlw $6, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $9, %eax, %xmm3, %xmm3 ; NoVLX-NEXT: vpinsrb $9, %eax, %xmm2, %xmm2
; NoVLX-NEXT: kshiftlw $5, %k0, %k1 ; NoVLX-NEXT: kshiftlw $5, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $10, %eax, %xmm3, %xmm3 ; NoVLX-NEXT: vpinsrb $10, %eax, %xmm2, %xmm2
; NoVLX-NEXT: kshiftlw $4, %k0, %k1 ; NoVLX-NEXT: kshiftlw $4, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $11, %eax, %xmm3, %xmm3 ; NoVLX-NEXT: vpinsrb $11, %eax, %xmm2, %xmm2
; NoVLX-NEXT: kshiftlw $3, %k0, %k1 ; NoVLX-NEXT: kshiftlw $3, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $12, %eax, %xmm3, %xmm3 ; NoVLX-NEXT: vpinsrb $12, %eax, %xmm2, %xmm2
; NoVLX-NEXT: kshiftlw $2, %k0, %k1 ; NoVLX-NEXT: kshiftlw $2, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $13, %eax, %xmm3, %xmm3 ; NoVLX-NEXT: vpinsrb $13, %eax, %xmm2, %xmm2
; NoVLX-NEXT: kshiftlw $1, %k0, %k1 ; NoVLX-NEXT: kshiftlw $1, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $14, %eax, %xmm3, %xmm3 ; NoVLX-NEXT: vpinsrb $14, %eax, %xmm2, %xmm2
; NoVLX-NEXT: kshiftrw $15, %k0, %k0 ; NoVLX-NEXT: kshiftrw $15, %k0, %k0
; NoVLX-NEXT: kmovw %k0, %eax ; NoVLX-NEXT: kmovw %k0, %eax
; NoVLX-NEXT: vpinsrb $15, %eax, %xmm3, %xmm3 ; NoVLX-NEXT: vpinsrb $15, %eax, %xmm2, %xmm2
; NoVLX-NEXT: vpxor %ymm5, %ymm4, %ymm4 ; NoVLX-NEXT: vpxor %ymm4, %ymm3, %ymm3
; NoVLX-NEXT: vpmovsxwd %ymm4, %zmm4 ; NoVLX-NEXT: vpmovsxwd %ymm3, %zmm3
; NoVLX-NEXT: vpslld $31, %zmm4, %zmm4 ; NoVLX-NEXT: vpslld $31, %zmm3, %zmm3
; NoVLX-NEXT: vptestmd %zmm4, %zmm4, %k0 ; NoVLX-NEXT: vptestmd %zmm3, %zmm3, %k0
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %ecx ; NoVLX-NEXT: kmovw %k1, %ecx
; NoVLX-NEXT: vmovd %ecx, %xmm4 ; NoVLX-NEXT: vmovd %ecx, %xmm3
; NoVLX-NEXT: vpinsrb $1, %eax, %xmm4, %xmm4 ; NoVLX-NEXT: vpinsrb $1, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $2, %eax, %xmm4, %xmm4 ; NoVLX-NEXT: vpinsrb $2, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftlw $12, %k0, %k1 ; NoVLX-NEXT: kshiftlw $12, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $3, %eax, %xmm4, %xmm4 ; NoVLX-NEXT: vpinsrb $3, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftlw $11, %k0, %k1 ; NoVLX-NEXT: kshiftlw $11, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $4, %eax, %xmm4, %xmm4 ; NoVLX-NEXT: vpinsrb $4, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftlw $10, %k0, %k1 ; NoVLX-NEXT: kshiftlw $10, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $5, %eax, %xmm4, %xmm4 ; NoVLX-NEXT: vpinsrb $5, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftlw $9, %k0, %k1 ; NoVLX-NEXT: kshiftlw $9, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $6, %eax, %xmm4, %xmm4 ; NoVLX-NEXT: vpinsrb $6, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftlw $8, %k0, %k1 ; NoVLX-NEXT: kshiftlw $8, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $7, %eax, %xmm4, %xmm4 ; NoVLX-NEXT: vpinsrb $7, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftlw $7, %k0, %k1 ; NoVLX-NEXT: kshiftlw $7, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $8, %eax, %xmm4, %xmm4 ; NoVLX-NEXT: vpinsrb $8, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftlw $6, %k0, %k1 ; NoVLX-NEXT: kshiftlw $6, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $9, %eax, %xmm4, %xmm4 ; NoVLX-NEXT: vpinsrb $9, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftlw $5, %k0, %k1 ; NoVLX-NEXT: kshiftlw $5, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $10, %eax, %xmm4, %xmm4 ; NoVLX-NEXT: vpinsrb $10, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftlw $4, %k0, %k1 ; NoVLX-NEXT: kshiftlw $4, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $11, %eax, %xmm4, %xmm4 ; NoVLX-NEXT: vpinsrb $11, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftlw $3, %k0, %k1 ; NoVLX-NEXT: kshiftlw $3, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $12, %eax, %xmm4, %xmm4 ; NoVLX-NEXT: vpinsrb $12, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftlw $2, %k0, %k1 ; NoVLX-NEXT: kshiftlw $2, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $13, %eax, %xmm4, %xmm4 ; NoVLX-NEXT: vpinsrb $13, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftlw $1, %k0, %k1 ; NoVLX-NEXT: kshiftlw $1, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $14, %eax, %xmm4, %xmm4 ; NoVLX-NEXT: vpinsrb $14, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftrw $15, %k0, %k0 ; NoVLX-NEXT: kshiftrw $15, %k0, %k0
; NoVLX-NEXT: kmovw %k0, %eax ; NoVLX-NEXT: kmovw %k0, %eax
; NoVLX-NEXT: vpinsrb $15, %eax, %xmm4, %xmm4 ; NoVLX-NEXT: vpinsrb $15, %eax, %xmm3, %xmm3
; NoVLX-NEXT: vpand %xmm2, %xmm4, %xmm2
; NoVLX-NEXT: vpmovsxbd %xmm2, %zmm2
; NoVLX-NEXT: vpslld $31, %zmm2, %zmm2
; NoVLX-NEXT: vptestmd %zmm2, %zmm2, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpand %xmm1, %xmm3, %xmm1 ; NoVLX-NEXT: vpand %xmm1, %xmm3, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpand %xmm0, %xmm2, %xmm0
; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT: kmovw %k0, (%rsp) ; NoVLX-NEXT: kmovw %k0, (%rsp)
; NoVLX-NEXT: movl (%rsp), %ecx ; NoVLX-NEXT: movl (%rsp), %ecx
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: movl {{[0-9]+}}(%rsp), %eax ; NoVLX-NEXT: movl {{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT: shlq $32, %rax ; NoVLX-NEXT: shlq $32, %rax
; NoVLX-NEXT: orq %rcx, %rax ; NoVLX-NEXT: orq %rcx, %rax
; NoVLX-NEXT: movq %rbp, %rsp ; NoVLX-NEXT: movq %rbp, %rsp
; NoVLX-NEXT: popq %rbp ; NoVLX-NEXT: popq %rbp
; NoVLX-NEXT: retq ; NoVLX-NEXT: retq
▲ Show 20 Lines • Show All 1,197 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi1100: ; NoVLX-NEXT: .Lcfi1100:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: vpcmpgtd %xmm0, %xmm1, %xmm0 ; NoVLX-NEXT: vpcmpgtd %xmm0, %xmm1, %xmm0
; NoVLX-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1
; NoVLX-NEXT: vpxor %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vpxor %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero ; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
Show All 35 Lines
; NoVLX-NEXT: .Lcfi1103: ; NoVLX-NEXT: .Lcfi1103:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: vmovdqa (%rdi), %xmm1 ; NoVLX-NEXT: vmovdqa (%rdi), %xmm1
; NoVLX-NEXT: vpcmpgtd %xmm0, %xmm1, %xmm0 ; NoVLX-NEXT: vpcmpgtd %xmm0, %xmm1, %xmm0
; NoVLX-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1
; NoVLX-NEXT: vpxor %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vpxor %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero ; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
▲ Show 20 Lines • Show All 52 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: kmovw %k3, %ecx ; NoVLX-NEXT: kmovw %k3, %ecx
; NoVLX-NEXT: vmovd %ecx, %xmm1 ; NoVLX-NEXT: vmovd %ecx, %xmm1
; NoVLX-NEXT: vpinsrb $4, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrb $4, %eax, %xmm1, %xmm1
; NoVLX-NEXT: kmovw %k2, %eax ; NoVLX-NEXT: kmovw %k2, %eax
; NoVLX-NEXT: vpinsrb $8, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $12, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT: vpandn %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vpandn %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vpxor %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero ; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
▲ Show 20 Lines • Show All 55 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: kmovw %k3, %ecx ; NoVLX-NEXT: kmovw %k3, %ecx
; NoVLX-NEXT: vmovd %ecx, %xmm1 ; NoVLX-NEXT: vmovd %ecx, %xmm1
; NoVLX-NEXT: vpinsrb $4, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrb $4, %eax, %xmm1, %xmm1
; NoVLX-NEXT: kmovw %k2, %eax ; NoVLX-NEXT: kmovw %k2, %eax
; NoVLX-NEXT: vpinsrb $8, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $12, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT: vpandn %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vpandn %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vpxor %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero ; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
▲ Show 20 Lines • Show All 41 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi1112: ; NoVLX-NEXT: .Lcfi1112:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: vpbroadcastd (%rdi), %xmm1 ; NoVLX-NEXT: vpbroadcastd (%rdi), %xmm1
; NoVLX-NEXT: vpcmpgtd %xmm0, %xmm1, %xmm0 ; NoVLX-NEXT: vpcmpgtd %xmm0, %xmm1, %xmm0
; NoVLX-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1
; NoVLX-NEXT: vpxor %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vpxor %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero ; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
▲ Show 20 Lines • Show All 55 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: kmovw %k3, %ecx ; NoVLX-NEXT: kmovw %k3, %ecx
; NoVLX-NEXT: vmovd %ecx, %xmm1 ; NoVLX-NEXT: vmovd %ecx, %xmm1
; NoVLX-NEXT: vpinsrb $4, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrb $4, %eax, %xmm1, %xmm1
; NoVLX-NEXT: kmovw %k2, %eax ; NoVLX-NEXT: kmovw %k2, %eax
; NoVLX-NEXT: vpinsrb $8, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $12, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT: vpandn %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vpandn %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vpxor %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero ; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
▲ Show 20 Lines • Show All 231 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi1118: ; NoVLX-NEXT: .Lcfi1118:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $32, %rsp ; NoVLX-NEXT: subq $32, %rsp
; NoVLX-NEXT: # kill: %YMM1<def> %YMM1<kill> %ZMM1<def> ; NoVLX-NEXT: # kill: %YMM1<def> %YMM1<kill> %ZMM1<def>
; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def> ; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def>
; NoVLX-NEXT: vpcmpled %zmm0, %zmm1, %k0 ; NoVLX-NEXT: vpcmpled %zmm0, %zmm1, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r9d ; NoVLX-NEXT: kmovw %k1, %r9d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 58 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi1121: ; NoVLX-NEXT: .Lcfi1121:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $32, %rsp ; NoVLX-NEXT: subq $32, %rsp
; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def> ; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def>
; NoVLX-NEXT: vmovdqa (%rdi), %ymm1 ; NoVLX-NEXT: vmovdqa (%rdi), %ymm1
; NoVLX-NEXT: vpcmpled %zmm0, %zmm1, %k0 ; NoVLX-NEXT: vpcmpled %zmm0, %zmm1, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r9d ; NoVLX-NEXT: kmovw %k1, %r9d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 62 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $32, %rsp ; NoVLX-NEXT: subq $32, %rsp
; NoVLX-NEXT: # kill: %YMM1<def> %YMM1<kill> %ZMM1<def> ; NoVLX-NEXT: # kill: %YMM1<def> %YMM1<kill> %ZMM1<def>
; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def> ; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def>
; NoVLX-NEXT: vpcmpled %zmm0, %zmm1, %k0 ; NoVLX-NEXT: vpcmpled %zmm0, %zmm1, %k0
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: kandw %k1, %k0, %k0 ; NoVLX-NEXT: kandw %k1, %k0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r9d ; NoVLX-NEXT: kmovw %k1, %r9d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 63 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $32, %rsp ; NoVLX-NEXT: subq $32, %rsp
; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def> ; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def>
; NoVLX-NEXT: vmovdqa (%rsi), %ymm1 ; NoVLX-NEXT: vmovdqa (%rsi), %ymm1
; NoVLX-NEXT: vpcmpled %zmm0, %zmm1, %k0 ; NoVLX-NEXT: vpcmpled %zmm0, %zmm1, %k0
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: kandw %k1, %k0, %k0 ; NoVLX-NEXT: kandw %k1, %k0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r9d ; NoVLX-NEXT: kmovw %k1, %r9d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 63 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi1130: ; NoVLX-NEXT: .Lcfi1130:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $32, %rsp ; NoVLX-NEXT: subq $32, %rsp
; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def> ; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def>
; NoVLX-NEXT: vpbroadcastd (%rdi), %ymm1 ; NoVLX-NEXT: vpbroadcastd (%rdi), %ymm1
; NoVLX-NEXT: vpcmpled %zmm0, %zmm1, %k0 ; NoVLX-NEXT: vpcmpled %zmm0, %zmm1, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r9d ; NoVLX-NEXT: kmovw %k1, %r9d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 64 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $32, %rsp ; NoVLX-NEXT: subq $32, %rsp
; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def> ; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def>
; NoVLX-NEXT: vpbroadcastd (%rsi), %ymm1 ; NoVLX-NEXT: vpbroadcastd (%rsi), %ymm1
; NoVLX-NEXT: vpcmpled %zmm0, %zmm1, %k0 ; NoVLX-NEXT: vpcmpled %zmm0, %zmm1, %k0
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: kandw %k0, %k1, %k0 ; NoVLX-NEXT: kandw %k0, %k1, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r9d ; NoVLX-NEXT: kmovw %k1, %r9d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 63 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi1136: ; NoVLX-NEXT: .Lcfi1136:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: # kill: %YMM1<def> %YMM1<kill> %ZMM1<def> ; NoVLX-NEXT: # kill: %YMM1<def> %YMM1<kill> %ZMM1<def>
; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def> ; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def>
; NoVLX-NEXT: vpcmpled %zmm0, %zmm1, %k0 ; NoVLX-NEXT: vpcmpled %zmm0, %zmm1, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 63 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi1139: ; NoVLX-NEXT: .Lcfi1139:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def> ; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def>
; NoVLX-NEXT: vmovdqa (%rdi), %ymm1 ; NoVLX-NEXT: vmovdqa (%rdi), %ymm1
; NoVLX-NEXT: vpcmpled %zmm0, %zmm1, %k0 ; NoVLX-NEXT: vpcmpled %zmm0, %zmm1, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 67 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: # kill: %YMM1<def> %YMM1<kill> %ZMM1<def> ; NoVLX-NEXT: # kill: %YMM1<def> %YMM1<kill> %ZMM1<def>
; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def> ; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def>
; NoVLX-NEXT: vpcmpled %zmm0, %zmm1, %k0 ; NoVLX-NEXT: vpcmpled %zmm0, %zmm1, %k0
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: kandw %k1, %k0, %k0 ; NoVLX-NEXT: kandw %k1, %k0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 68 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def> ; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def>
; NoVLX-NEXT: vmovdqa (%rsi), %ymm1 ; NoVLX-NEXT: vmovdqa (%rsi), %ymm1
; NoVLX-NEXT: vpcmpled %zmm0, %zmm1, %k0 ; NoVLX-NEXT: vpcmpled %zmm0, %zmm1, %k0
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: kandw %k1, %k0, %k0 ; NoVLX-NEXT: kandw %k1, %k0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 68 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi1148: ; NoVLX-NEXT: .Lcfi1148:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def> ; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def>
; NoVLX-NEXT: vpbroadcastd (%rdi), %ymm1 ; NoVLX-NEXT: vpbroadcastd (%rdi), %ymm1
; NoVLX-NEXT: vpcmpled %zmm0, %zmm1, %k0 ; NoVLX-NEXT: vpcmpled %zmm0, %zmm1, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 69 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def> ; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def>
; NoVLX-NEXT: vpbroadcastd (%rsi), %ymm1 ; NoVLX-NEXT: vpbroadcastd (%rsi), %ymm1
; NoVLX-NEXT: vpcmpled %zmm0, %zmm1, %k0 ; NoVLX-NEXT: vpcmpled %zmm0, %zmm1, %k0
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: kandw %k0, %k1, %k0 ; NoVLX-NEXT: kandw %k0, %k1, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 81 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %r12, -48 ; NoVLX-NEXT: .cfi_offset %r12, -48
; NoVLX-NEXT: .Lcfi1157: ; NoVLX-NEXT: .Lcfi1157:
; NoVLX-NEXT: .cfi_offset %r13, -40 ; NoVLX-NEXT: .cfi_offset %r13, -40
; NoVLX-NEXT: .Lcfi1158: ; NoVLX-NEXT: .Lcfi1158:
; NoVLX-NEXT: .cfi_offset %r14, -32 ; NoVLX-NEXT: .cfi_offset %r14, -32
; NoVLX-NEXT: .Lcfi1159: ; NoVLX-NEXT: .Lcfi1159:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: vpcmpled %zmm0, %zmm1, %k0 ; NoVLX-NEXT: vpcmpled %zmm0, %zmm1, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r10d ; NoVLX-NEXT: kmovw %k1, %r10d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 106 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %r12, -48 ; NoVLX-NEXT: .cfi_offset %r12, -48
; NoVLX-NEXT: .Lcfi1165: ; NoVLX-NEXT: .Lcfi1165:
; NoVLX-NEXT: .cfi_offset %r13, -40 ; NoVLX-NEXT: .cfi_offset %r13, -40
; NoVLX-NEXT: .Lcfi1166: ; NoVLX-NEXT: .Lcfi1166:
; NoVLX-NEXT: .cfi_offset %r14, -32 ; NoVLX-NEXT: .cfi_offset %r14, -32
; NoVLX-NEXT: .Lcfi1167: ; NoVLX-NEXT: .Lcfi1167:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: vpcmpnltd (%rdi), %zmm0, %k0 ; NoVLX-NEXT: vpcmpnltd (%rdi), %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r10d ; NoVLX-NEXT: kmovw %k1, %r10d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 109 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi1173: ; NoVLX-NEXT: .Lcfi1173:
; NoVLX-NEXT: .cfi_offset %r13, -40 ; NoVLX-NEXT: .cfi_offset %r13, -40
; NoVLX-NEXT: .Lcfi1174: ; NoVLX-NEXT: .Lcfi1174:
; NoVLX-NEXT: .cfi_offset %r14, -32 ; NoVLX-NEXT: .cfi_offset %r14, -32
; NoVLX-NEXT: .Lcfi1175: ; NoVLX-NEXT: .Lcfi1175:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vpcmpled %zmm0, %zmm1, %k0 {%k1} ; NoVLX-NEXT: vpcmpled %zmm0, %zmm1, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r10d ; NoVLX-NEXT: kmovw %k1, %r10d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 110 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi1181: ; NoVLX-NEXT: .Lcfi1181:
; NoVLX-NEXT: .cfi_offset %r13, -40 ; NoVLX-NEXT: .cfi_offset %r13, -40
; NoVLX-NEXT: .Lcfi1182: ; NoVLX-NEXT: .Lcfi1182:
; NoVLX-NEXT: .cfi_offset %r14, -32 ; NoVLX-NEXT: .cfi_offset %r14, -32
; NoVLX-NEXT: .Lcfi1183: ; NoVLX-NEXT: .Lcfi1183:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vpcmpnltd (%rsi), %zmm0, %k0 {%k1} ; NoVLX-NEXT: vpcmpnltd (%rsi), %zmm0, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r10d ; NoVLX-NEXT: kmovw %k1, %r10d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 112 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi1189: ; NoVLX-NEXT: .Lcfi1189:
; NoVLX-NEXT: .cfi_offset %r13, -40 ; NoVLX-NEXT: .cfi_offset %r13, -40
; NoVLX-NEXT: .Lcfi1190: ; NoVLX-NEXT: .Lcfi1190:
; NoVLX-NEXT: .cfi_offset %r14, -32 ; NoVLX-NEXT: .cfi_offset %r14, -32
; NoVLX-NEXT: .Lcfi1191: ; NoVLX-NEXT: .Lcfi1191:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: vpbroadcastd (%rdi), %zmm1 ; NoVLX-NEXT: vpbroadcastd (%rdi), %zmm1
; NoVLX-NEXT: vpcmpled %zmm0, %zmm1, %k0 ; NoVLX-NEXT: vpcmpled %zmm0, %zmm1, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r10d ; NoVLX-NEXT: kmovw %k1, %r10d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 112 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %r13, -40 ; NoVLX-NEXT: .cfi_offset %r13, -40
; NoVLX-NEXT: .Lcfi1198: ; NoVLX-NEXT: .Lcfi1198:
; NoVLX-NEXT: .cfi_offset %r14, -32 ; NoVLX-NEXT: .cfi_offset %r14, -32
; NoVLX-NEXT: .Lcfi1199: ; NoVLX-NEXT: .Lcfi1199:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: vpbroadcastd (%rsi), %zmm1 ; NoVLX-NEXT: vpbroadcastd (%rsi), %zmm1
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vpcmpled %zmm0, %zmm1, %k0 {%k1} ; NoVLX-NEXT: vpcmpled %zmm0, %zmm1, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r10d ; NoVLX-NEXT: kmovw %k1, %r10d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 111 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %r12, -48 ; NoVLX-NEXT: .cfi_offset %r12, -48
; NoVLX-NEXT: .Lcfi1205: ; NoVLX-NEXT: .Lcfi1205:
; NoVLX-NEXT: .cfi_offset %r13, -40 ; NoVLX-NEXT: .cfi_offset %r13, -40
; NoVLX-NEXT: .Lcfi1206: ; NoVLX-NEXT: .Lcfi1206:
; NoVLX-NEXT: .cfi_offset %r14, -32 ; NoVLX-NEXT: .cfi_offset %r14, -32
; NoVLX-NEXT: .Lcfi1207: ; NoVLX-NEXT: .Lcfi1207:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: vpcmpled %zmm0, %zmm1, %k0 ; NoVLX-NEXT: vpcmpled %zmm0, %zmm1, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 111 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %r12, -48 ; NoVLX-NEXT: .cfi_offset %r12, -48
; NoVLX-NEXT: .Lcfi1213: ; NoVLX-NEXT: .Lcfi1213:
; NoVLX-NEXT: .cfi_offset %r13, -40 ; NoVLX-NEXT: .cfi_offset %r13, -40
; NoVLX-NEXT: .Lcfi1214: ; NoVLX-NEXT: .Lcfi1214:
; NoVLX-NEXT: .cfi_offset %r14, -32 ; NoVLX-NEXT: .cfi_offset %r14, -32
; NoVLX-NEXT: .Lcfi1215: ; NoVLX-NEXT: .Lcfi1215:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: vpcmpnltd (%rdi), %zmm0, %k0 ; NoVLX-NEXT: vpcmpnltd (%rdi), %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 114 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi1221: ; NoVLX-NEXT: .Lcfi1221:
; NoVLX-NEXT: .cfi_offset %r13, -40 ; NoVLX-NEXT: .cfi_offset %r13, -40
; NoVLX-NEXT: .Lcfi1222: ; NoVLX-NEXT: .Lcfi1222:
; NoVLX-NEXT: .cfi_offset %r14, -32 ; NoVLX-NEXT: .cfi_offset %r14, -32
; NoVLX-NEXT: .Lcfi1223: ; NoVLX-NEXT: .Lcfi1223:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vpcmpled %zmm0, %zmm1, %k0 {%k1} ; NoVLX-NEXT: vpcmpled %zmm0, %zmm1, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 115 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi1229: ; NoVLX-NEXT: .Lcfi1229:
; NoVLX-NEXT: .cfi_offset %r13, -40 ; NoVLX-NEXT: .cfi_offset %r13, -40
; NoVLX-NEXT: .Lcfi1230: ; NoVLX-NEXT: .Lcfi1230:
; NoVLX-NEXT: .cfi_offset %r14, -32 ; NoVLX-NEXT: .cfi_offset %r14, -32
; NoVLX-NEXT: .Lcfi1231: ; NoVLX-NEXT: .Lcfi1231:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vpcmpnltd (%rsi), %zmm0, %k0 {%k1} ; NoVLX-NEXT: vpcmpnltd (%rsi), %zmm0, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 117 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi1237: ; NoVLX-NEXT: .Lcfi1237:
; NoVLX-NEXT: .cfi_offset %r13, -40 ; NoVLX-NEXT: .cfi_offset %r13, -40
; NoVLX-NEXT: .Lcfi1238: ; NoVLX-NEXT: .Lcfi1238:
; NoVLX-NEXT: .cfi_offset %r14, -32 ; NoVLX-NEXT: .cfi_offset %r14, -32
; NoVLX-NEXT: .Lcfi1239: ; NoVLX-NEXT: .Lcfi1239:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: vpbroadcastd (%rdi), %zmm1 ; NoVLX-NEXT: vpbroadcastd (%rdi), %zmm1
; NoVLX-NEXT: vpcmpled %zmm0, %zmm1, %k0 ; NoVLX-NEXT: vpcmpled %zmm0, %zmm1, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 117 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %r13, -40 ; NoVLX-NEXT: .cfi_offset %r13, -40
; NoVLX-NEXT: .Lcfi1246: ; NoVLX-NEXT: .Lcfi1246:
; NoVLX-NEXT: .cfi_offset %r14, -32 ; NoVLX-NEXT: .cfi_offset %r14, -32
; NoVLX-NEXT: .Lcfi1247: ; NoVLX-NEXT: .Lcfi1247:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: vpbroadcastd (%rsi), %zmm1 ; NoVLX-NEXT: vpbroadcastd (%rsi), %zmm1
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vpcmpled %zmm0, %zmm1, %k0 {%k1} ; NoVLX-NEXT: vpcmpled %zmm0, %zmm1, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 1,232 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi1268: ; NoVLX-NEXT: .Lcfi1268:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: vpcmpgtq %xmm0, %xmm1, %xmm0 ; NoVLX-NEXT: vpcmpgtq %xmm0, %xmm1, %xmm0
; NoVLX-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1
; NoVLX-NEXT: vpxor %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vpxor %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero ; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
Show All 35 Lines
; NoVLX-NEXT: .Lcfi1271: ; NoVLX-NEXT: .Lcfi1271:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: vmovdqa (%rdi), %xmm1 ; NoVLX-NEXT: vmovdqa (%rdi), %xmm1
; NoVLX-NEXT: vpcmpgtq %xmm0, %xmm1, %xmm0 ; NoVLX-NEXT: vpcmpgtq %xmm0, %xmm1, %xmm0
; NoVLX-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1
; NoVLX-NEXT: vpxor %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vpxor %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero ; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
▲ Show 20 Lines • Show All 44 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kshiftlw $14, %k0, %k0 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
; NoVLX-NEXT: kshiftrw $15, %k0, %k0 ; NoVLX-NEXT: kshiftrw $15, %k0, %k0
; NoVLX-NEXT: kmovw %k0, %eax ; NoVLX-NEXT: kmovw %k0, %eax
; NoVLX-NEXT: kmovw %k1, %ecx ; NoVLX-NEXT: kmovw %k1, %ecx
; NoVLX-NEXT: vmovd %ecx, %xmm1 ; NoVLX-NEXT: vmovd %ecx, %xmm1
; NoVLX-NEXT: vpinsrb $8, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT: vpandn %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vpandn %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vpxor %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero ; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
▲ Show 20 Lines • Show All 47 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kshiftlw $14, %k0, %k0 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
; NoVLX-NEXT: kshiftrw $15, %k0, %k0 ; NoVLX-NEXT: kshiftrw $15, %k0, %k0
; NoVLX-NEXT: kmovw %k0, %eax ; NoVLX-NEXT: kmovw %k0, %eax
; NoVLX-NEXT: kmovw %k1, %ecx ; NoVLX-NEXT: kmovw %k1, %ecx
; NoVLX-NEXT: vmovd %ecx, %xmm1 ; NoVLX-NEXT: vmovd %ecx, %xmm1
; NoVLX-NEXT: vpinsrb $8, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT: vpandn %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vpandn %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vpxor %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero ; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
▲ Show 20 Lines • Show All 41 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi1280: ; NoVLX-NEXT: .Lcfi1280:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: vpbroadcastq (%rdi), %xmm1 ; NoVLX-NEXT: vpbroadcastq (%rdi), %xmm1
; NoVLX-NEXT: vpcmpgtq %xmm0, %xmm1, %xmm0 ; NoVLX-NEXT: vpcmpgtq %xmm0, %xmm1, %xmm0
; NoVLX-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: vpcmpeqd %xmm1, %xmm1, %xmm1
; NoVLX-NEXT: vpxor %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vpxor %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero ; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
▲ Show 20 Lines • Show All 47 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kshiftlw $14, %k0, %k0 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
; NoVLX-NEXT: kshiftrw $15, %k0, %k0 ; NoVLX-NEXT: kshiftrw $15, %k0, %k0
; NoVLX-NEXT: kmovw %k0, %eax ; NoVLX-NEXT: kmovw %k0, %eax
; NoVLX-NEXT: kmovw %k1, %ecx ; NoVLX-NEXT: kmovw %k1, %ecx
; NoVLX-NEXT: vmovd %ecx, %xmm1 ; NoVLX-NEXT: vmovd %ecx, %xmm1
; NoVLX-NEXT: vpinsrb $8, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT: vpandn %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vpandn %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vpxor %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero ; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
▲ Show 20 Lines • Show All 1,266 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi1304: ; NoVLX-NEXT: .Lcfi1304:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: vpcmpgtq %ymm0, %ymm1, %ymm0 ; NoVLX-NEXT: vpcmpgtq %ymm0, %ymm1, %ymm0
; NoVLX-NEXT: vpcmpeqd %ymm1, %ymm1, %ymm1 ; NoVLX-NEXT: vpcmpeqd %ymm1, %ymm1, %ymm1
; NoVLX-NEXT: vpxor %ymm1, %ymm0, %ymm0 ; NoVLX-NEXT: vpxor %ymm1, %ymm0, %ymm0
; NoVLX-NEXT: vpmovqd %zmm0, %ymm0 ; NoVLX-NEXT: vpmovqd %zmm0, %ymm0
; NoVLX-NEXT: vpxor %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero ; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
Show All 37 Lines
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: vmovdqa (%rdi), %ymm1 ; NoVLX-NEXT: vmovdqa (%rdi), %ymm1
; NoVLX-NEXT: vpcmpgtq %ymm0, %ymm1, %ymm0 ; NoVLX-NEXT: vpcmpgtq %ymm0, %ymm1, %ymm0
; NoVLX-NEXT: vpcmpeqd %ymm1, %ymm1, %ymm1 ; NoVLX-NEXT: vpcmpeqd %ymm1, %ymm1, %ymm1
; NoVLX-NEXT: vpxor %ymm1, %ymm0, %ymm0 ; NoVLX-NEXT: vpxor %ymm1, %ymm0, %ymm0
; NoVLX-NEXT: vpmovqd %zmm0, %ymm0 ; NoVLX-NEXT: vpmovqd %zmm0, %ymm0
; NoVLX-NEXT: vpxor %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero ; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
▲ Show 20 Lines • Show All 56 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: kmovw %k3, %ecx ; NoVLX-NEXT: kmovw %k3, %ecx
; NoVLX-NEXT: vmovd %ecx, %xmm1 ; NoVLX-NEXT: vmovd %ecx, %xmm1
; NoVLX-NEXT: vpinsrb $4, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrb $4, %eax, %xmm1, %xmm1
; NoVLX-NEXT: kmovw %k2, %eax ; NoVLX-NEXT: kmovw %k2, %eax
; NoVLX-NEXT: vpinsrb $8, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $12, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT: vpand %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vpxor %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero ; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
▲ Show 20 Lines • Show All 59 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: kmovw %k3, %ecx ; NoVLX-NEXT: kmovw %k3, %ecx
; NoVLX-NEXT: vmovd %ecx, %xmm1 ; NoVLX-NEXT: vmovd %ecx, %xmm1
; NoVLX-NEXT: vpinsrb $4, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrb $4, %eax, %xmm1, %xmm1
; NoVLX-NEXT: kmovw %k2, %eax ; NoVLX-NEXT: kmovw %k2, %eax
; NoVLX-NEXT: vpinsrb $8, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $12, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT: vpand %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vpxor %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero ; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
▲ Show 20 Lines • Show All 43 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: vpbroadcastq (%rdi), %ymm1 ; NoVLX-NEXT: vpbroadcastq (%rdi), %ymm1
; NoVLX-NEXT: vpcmpgtq %ymm0, %ymm1, %ymm0 ; NoVLX-NEXT: vpcmpgtq %ymm0, %ymm1, %ymm0
; NoVLX-NEXT: vpcmpeqd %ymm1, %ymm1, %ymm1 ; NoVLX-NEXT: vpcmpeqd %ymm1, %ymm1, %ymm1
; NoVLX-NEXT: vpxor %ymm1, %ymm0, %ymm0 ; NoVLX-NEXT: vpxor %ymm1, %ymm0, %ymm0
; NoVLX-NEXT: vpmovqd %zmm0, %ymm0 ; NoVLX-NEXT: vpmovqd %zmm0, %ymm0
; NoVLX-NEXT: vpxor %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero ; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
▲ Show 20 Lines • Show All 59 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: kmovw %k3, %ecx ; NoVLX-NEXT: kmovw %k3, %ecx
; NoVLX-NEXT: vmovd %ecx, %xmm1 ; NoVLX-NEXT: vmovd %ecx, %xmm1
; NoVLX-NEXT: vpinsrb $4, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrb $4, %eax, %xmm1, %xmm1
; NoVLX-NEXT: kmovw %k2, %eax ; NoVLX-NEXT: kmovw %k2, %eax
; NoVLX-NEXT: vpinsrb $8, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $12, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT: vpand %xmm0, %xmm1, %xmm0 ; NoVLX-NEXT: vpand %xmm0, %xmm1, %xmm0
; NoVLX-NEXT: vpxor %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero ; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
▲ Show 20 Lines • Show All 207 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi1321: ; NoVLX-NEXT: .Lcfi1321:
; NoVLX-NEXT: .cfi_offset %rbp, -16 ; NoVLX-NEXT: .cfi_offset %rbp, -16
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi1322: ; NoVLX-NEXT: .Lcfi1322:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $32, %rsp ; NoVLX-NEXT: subq $32, %rsp
; NoVLX-NEXT: vpcmpleq %zmm0, %zmm1, %k0 ; NoVLX-NEXT: vpcmpleq %zmm0, %zmm1, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r9d ; NoVLX-NEXT: kmovw %k1, %r9d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 56 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi1324: ; NoVLX-NEXT: .Lcfi1324:
; NoVLX-NEXT: .cfi_offset %rbp, -16 ; NoVLX-NEXT: .cfi_offset %rbp, -16
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi1325: ; NoVLX-NEXT: .Lcfi1325:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $32, %rsp ; NoVLX-NEXT: subq $32, %rsp
; NoVLX-NEXT: vpcmpnltq (%rdi), %zmm0, %k0 ; NoVLX-NEXT: vpcmpnltq (%rdi), %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r9d ; NoVLX-NEXT: kmovw %k1, %r9d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 59 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %rbp, -16 ; NoVLX-NEXT: .cfi_offset %rbp, -16
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi1328: ; NoVLX-NEXT: .Lcfi1328:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $32, %rsp ; NoVLX-NEXT: subq $32, %rsp
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vpcmpleq %zmm0, %zmm1, %k0 {%k1} ; NoVLX-NEXT: vpcmpleq %zmm0, %zmm1, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r9d ; NoVLX-NEXT: kmovw %k1, %r9d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 60 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %rbp, -16 ; NoVLX-NEXT: .cfi_offset %rbp, -16
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi1331: ; NoVLX-NEXT: .Lcfi1331:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $32, %rsp ; NoVLX-NEXT: subq $32, %rsp
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vpcmpnltq (%rsi), %zmm0, %k0 {%k1} ; NoVLX-NEXT: vpcmpnltq (%rsi), %zmm0, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r9d ; NoVLX-NEXT: kmovw %k1, %r9d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 62 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %rbp, -16 ; NoVLX-NEXT: .cfi_offset %rbp, -16
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi1334: ; NoVLX-NEXT: .Lcfi1334:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $32, %rsp ; NoVLX-NEXT: subq $32, %rsp
; NoVLX-NEXT: vpbroadcastq (%rdi), %zmm1 ; NoVLX-NEXT: vpbroadcastq (%rdi), %zmm1
; NoVLX-NEXT: vpcmpleq %zmm0, %zmm1, %k0 ; NoVLX-NEXT: vpcmpleq %zmm0, %zmm1, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r9d ; NoVLX-NEXT: kmovw %k1, %r9d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 62 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi1337: ; NoVLX-NEXT: .Lcfi1337:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $32, %rsp ; NoVLX-NEXT: subq $32, %rsp
; NoVLX-NEXT: vpbroadcastq (%rsi), %zmm1 ; NoVLX-NEXT: vpbroadcastq (%rsi), %zmm1
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vpcmpleq %zmm0, %zmm1, %k0 {%k1} ; NoVLX-NEXT: vpcmpleq %zmm0, %zmm1, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r9d ; NoVLX-NEXT: kmovw %k1, %r9d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 61 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi1339: ; NoVLX-NEXT: .Lcfi1339:
; NoVLX-NEXT: .cfi_offset %rbp, -16 ; NoVLX-NEXT: .cfi_offset %rbp, -16
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi1340: ; NoVLX-NEXT: .Lcfi1340:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: vpcmpleq %zmm0, %zmm1, %k0 ; NoVLX-NEXT: vpcmpleq %zmm0, %zmm1, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 61 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi1342: ; NoVLX-NEXT: .Lcfi1342:
; NoVLX-NEXT: .cfi_offset %rbp, -16 ; NoVLX-NEXT: .cfi_offset %rbp, -16
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi1343: ; NoVLX-NEXT: .Lcfi1343:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: vpcmpnltq (%rdi), %zmm0, %k0 ; NoVLX-NEXT: vpcmpnltq (%rdi), %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 64 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %rbp, -16 ; NoVLX-NEXT: .cfi_offset %rbp, -16
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi1346: ; NoVLX-NEXT: .Lcfi1346:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vpcmpleq %zmm0, %zmm1, %k0 {%k1} ; NoVLX-NEXT: vpcmpleq %zmm0, %zmm1, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 65 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %rbp, -16 ; NoVLX-NEXT: .cfi_offset %rbp, -16
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi1349: ; NoVLX-NEXT: .Lcfi1349:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vpcmpnltq (%rsi), %zmm0, %k0 {%k1} ; NoVLX-NEXT: vpcmpnltq (%rsi), %zmm0, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 67 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %rbp, -16 ; NoVLX-NEXT: .cfi_offset %rbp, -16
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi1352: ; NoVLX-NEXT: .Lcfi1352:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: vpbroadcastq (%rdi), %zmm1 ; NoVLX-NEXT: vpbroadcastq (%rdi), %zmm1
; NoVLX-NEXT: vpcmpleq %zmm0, %zmm1, %k0 ; NoVLX-NEXT: vpcmpleq %zmm0, %zmm1, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 67 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi1355: ; NoVLX-NEXT: .Lcfi1355:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: vpbroadcastq (%rsi), %zmm1 ; NoVLX-NEXT: vpbroadcastq (%rsi), %zmm1
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vpcmpleq %zmm0, %zmm1, %k0 {%k1} ; NoVLX-NEXT: vpcmpleq %zmm0, %zmm1, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 86 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: vmovdqa {{.*#+}} xmm2 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128] ; NoVLX-NEXT: vmovdqa {{.*#+}} xmm2 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128]
; NoVLX-NEXT: vpxor %xmm2, %xmm0, %xmm0 ; NoVLX-NEXT: vpxor %xmm2, %xmm0, %xmm0
; NoVLX-NEXT: vpxor %xmm2, %xmm1, %xmm1 ; NoVLX-NEXT: vpxor %xmm2, %xmm1, %xmm1
; NoVLX-NEXT: vpcmpgtb %xmm0, %xmm1, %xmm0 ; NoVLX-NEXT: vpcmpgtb %xmm0, %xmm1, %xmm0
; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0 ; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r10d ; NoVLX-NEXT: kmovw %k1, %r10d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 111 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: vmovdqa {{.*#+}} xmm1 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128] ; NoVLX-NEXT: vmovdqa {{.*#+}} xmm1 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128]
; NoVLX-NEXT: vpxor %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vpxor (%rdi), %xmm1, %xmm1 ; NoVLX-NEXT: vpxor (%rdi), %xmm1, %xmm1
; NoVLX-NEXT: vpcmpgtb %xmm0, %xmm1, %xmm0 ; NoVLX-NEXT: vpcmpgtb %xmm0, %xmm1, %xmm0
; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0 ; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r10d ; NoVLX-NEXT: kmovw %k1, %r10d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 114 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: vmovdqa {{.*#+}} xmm2 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128] ; NoVLX-NEXT: vmovdqa {{.*#+}} xmm2 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128]
; NoVLX-NEXT: vpxor %xmm2, %xmm0, %xmm0 ; NoVLX-NEXT: vpxor %xmm2, %xmm0, %xmm0
; NoVLX-NEXT: vpxor %xmm2, %xmm1, %xmm1 ; NoVLX-NEXT: vpxor %xmm2, %xmm1, %xmm1
; NoVLX-NEXT: vpcmpgtb %xmm0, %xmm1, %xmm0 ; NoVLX-NEXT: vpcmpgtb %xmm0, %xmm1, %xmm0
; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0 ; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 {%k1} ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r10d ; NoVLX-NEXT: kmovw %k1, %r10d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 115 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: vmovdqa {{.*#+}} xmm1 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128] ; NoVLX-NEXT: vmovdqa {{.*#+}} xmm1 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128]
; NoVLX-NEXT: vpxor %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vpxor (%rsi), %xmm1, %xmm1 ; NoVLX-NEXT: vpxor (%rsi), %xmm1, %xmm1
; NoVLX-NEXT: vpcmpgtb %xmm0, %xmm1, %xmm0 ; NoVLX-NEXT: vpcmpgtb %xmm0, %xmm1, %xmm0
; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0 ; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 {%k1} ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r10d ; NoVLX-NEXT: kmovw %k1, %r10d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 115 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: vmovdqa {{.*#+}} xmm2 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128] ; NoVLX-NEXT: vmovdqa {{.*#+}} xmm2 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128]
; NoVLX-NEXT: vpxor %xmm2, %xmm0, %xmm0 ; NoVLX-NEXT: vpxor %xmm2, %xmm0, %xmm0
; NoVLX-NEXT: vpxor %xmm2, %xmm1, %xmm1 ; NoVLX-NEXT: vpxor %xmm2, %xmm1, %xmm1
; NoVLX-NEXT: vpcmpgtb %xmm0, %xmm1, %xmm0 ; NoVLX-NEXT: vpcmpgtb %xmm0, %xmm1, %xmm0
; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0 ; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 116 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: vmovdqa {{.*#+}} xmm1 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128] ; NoVLX-NEXT: vmovdqa {{.*#+}} xmm1 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128]
; NoVLX-NEXT: vpxor %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vpxor (%rdi), %xmm1, %xmm1 ; NoVLX-NEXT: vpxor (%rdi), %xmm1, %xmm1
; NoVLX-NEXT: vpcmpgtb %xmm0, %xmm1, %xmm0 ; NoVLX-NEXT: vpcmpgtb %xmm0, %xmm1, %xmm0
; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0 ; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 119 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: vmovdqa {{.*#+}} xmm2 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128] ; NoVLX-NEXT: vmovdqa {{.*#+}} xmm2 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128]
; NoVLX-NEXT: vpxor %xmm2, %xmm0, %xmm0 ; NoVLX-NEXT: vpxor %xmm2, %xmm0, %xmm0
; NoVLX-NEXT: vpxor %xmm2, %xmm1, %xmm1 ; NoVLX-NEXT: vpxor %xmm2, %xmm1, %xmm1
; NoVLX-NEXT: vpcmpgtb %xmm0, %xmm1, %xmm0 ; NoVLX-NEXT: vpcmpgtb %xmm0, %xmm1, %xmm0
; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0 ; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 {%k1} ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 120 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: vmovdqa {{.*#+}} xmm1 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128] ; NoVLX-NEXT: vmovdqa {{.*#+}} xmm1 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128]
; NoVLX-NEXT: vpxor %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vpxor (%rsi), %xmm1, %xmm1 ; NoVLX-NEXT: vpxor (%rsi), %xmm1, %xmm1
; NoVLX-NEXT: vpcmpgtb %xmm0, %xmm1, %xmm0 ; NoVLX-NEXT: vpcmpgtb %xmm0, %xmm1, %xmm0
; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0 ; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 {%k1} ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 113 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0 ; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT: kmovw %k0, (%rsp) ; NoVLX-NEXT: kmovw %k0, (%rsp)
; NoVLX-NEXT: movl (%rsp), %ecx ; NoVLX-NEXT: movl (%rsp), %ecx
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: movl {{[0-9]+}}(%rsp), %eax ; NoVLX-NEXT: movl {{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT: shlq $32, %rax ; NoVLX-NEXT: shlq $32, %rax
; NoVLX-NEXT: orq %rcx, %rax ; NoVLX-NEXT: orq %rcx, %rax
; NoVLX-NEXT: movq %rbp, %rsp ; NoVLX-NEXT: movq %rbp, %rsp
; NoVLX-NEXT: popq %rbp ; NoVLX-NEXT: popq %rbp
; NoVLX-NEXT: retq ; NoVLX-NEXT: retq
Show All 35 Lines
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0 ; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT: kmovw %k0, (%rsp) ; NoVLX-NEXT: kmovw %k0, (%rsp)
; NoVLX-NEXT: movl (%rsp), %ecx ; NoVLX-NEXT: movl (%rsp), %ecx
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: movl {{[0-9]+}}(%rsp), %eax ; NoVLX-NEXT: movl {{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT: shlq $32, %rax ; NoVLX-NEXT: shlq $32, %rax
; NoVLX-NEXT: orq %rcx, %rax ; NoVLX-NEXT: orq %rcx, %rax
; NoVLX-NEXT: movq %rbp, %rsp ; NoVLX-NEXT: movq %rbp, %rsp
; NoVLX-NEXT: popq %rbp ; NoVLX-NEXT: popq %rbp
; NoVLX-NEXT: retq ; NoVLX-NEXT: retq
Show All 30 Lines
; NoVLX-NEXT: subq $96, %rsp ; NoVLX-NEXT: subq $96, %rsp
; NoVLX-NEXT: movl %edi, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: movl %edi, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw {{[0-9]+}}(%rsp), %k1 ; NoVLX-NEXT: kmovw {{[0-9]+}}(%rsp), %k1
; NoVLX-NEXT: kmovw {{[0-9]+}}(%rsp), %k2 ; NoVLX-NEXT: kmovw {{[0-9]+}}(%rsp), %k2
; NoVLX-NEXT: vpternlogd $255, %zmm2, %zmm2, %zmm2 {%k1} {z} ; NoVLX-NEXT: vpternlogd $255, %zmm2, %zmm2, %zmm2 {%k1} {z}
; NoVLX-NEXT: vpmovdb %zmm2, %xmm2 ; NoVLX-NEXT: vpmovdb %zmm2, %xmm2
; NoVLX-NEXT: vpternlogd $255, %zmm3, %zmm3, %zmm3 {%k2} {z} ; NoVLX-NEXT: vpternlogd $255, %zmm3, %zmm3, %zmm3 {%k2} {z}
; NoVLX-NEXT: vpmovdb %zmm3, %xmm3 ; NoVLX-NEXT: vpmovdb %zmm3, %xmm3
; NoVLX-NEXT: vxorps %xmm4, %xmm4, %xmm4 ; NoVLX-NEXT: vmovdqa {{.*#+}} ymm4 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128]
; NoVLX-NEXT: vmovdqa {{.*#+}} ymm5 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128] ; NoVLX-NEXT: vpxor %ymm4, %ymm0, %ymm0
; NoVLX-NEXT: vpxor %ymm5, %ymm0, %ymm0 ; NoVLX-NEXT: vpxor %ymm4, %ymm1, %ymm1
; NoVLX-NEXT: vpxor %ymm5, %ymm1, %ymm1
; NoVLX-NEXT: vpcmpgtb %ymm0, %ymm1, %ymm0 ; NoVLX-NEXT: vpcmpgtb %ymm0, %ymm1, %ymm0
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT: vpand %xmm3, %xmm1, %xmm1 ; NoVLX-NEXT: vpand %xmm3, %xmm1, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpand %xmm2, %xmm0, %xmm0 ; NoVLX-NEXT: vpand %xmm2, %xmm0, %xmm0
; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0 ; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT: kmovw %k0, (%rsp) ; NoVLX-NEXT: kmovw %k0, (%rsp)
; NoVLX-NEXT: movl (%rsp), %ecx ; NoVLX-NEXT: movl (%rsp), %ecx
; NoVLX-NEXT: vptestmd %zmm4, %zmm4, %k0 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: movl {{[0-9]+}}(%rsp), %eax ; NoVLX-NEXT: movl {{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT: shlq $32, %rax ; NoVLX-NEXT: shlq $32, %rax
; NoVLX-NEXT: orq %rcx, %rax ; NoVLX-NEXT: orq %rcx, %rax
; NoVLX-NEXT: movq %rbp, %rsp ; NoVLX-NEXT: movq %rbp, %rsp
; NoVLX-NEXT: popq %rbp ; NoVLX-NEXT: popq %rbp
; NoVLX-NEXT: retq ; NoVLX-NEXT: retq
Show All 31 Lines
; NoVLX-NEXT: subq $96, %rsp ; NoVLX-NEXT: subq $96, %rsp
; NoVLX-NEXT: movl %edi, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: movl %edi, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw {{[0-9]+}}(%rsp), %k1 ; NoVLX-NEXT: kmovw {{[0-9]+}}(%rsp), %k1
; NoVLX-NEXT: kmovw {{[0-9]+}}(%rsp), %k2 ; NoVLX-NEXT: kmovw {{[0-9]+}}(%rsp), %k2
; NoVLX-NEXT: vpternlogd $255, %zmm1, %zmm1, %zmm1 {%k1} {z} ; NoVLX-NEXT: vpternlogd $255, %zmm1, %zmm1, %zmm1 {%k1} {z}
; NoVLX-NEXT: vpmovdb %zmm1, %xmm1 ; NoVLX-NEXT: vpmovdb %zmm1, %xmm1
; NoVLX-NEXT: vpternlogd $255, %zmm2, %zmm2, %zmm2 {%k2} {z} ; NoVLX-NEXT: vpternlogd $255, %zmm2, %zmm2, %zmm2 {%k2} {z}
; NoVLX-NEXT: vpmovdb %zmm2, %xmm2 ; NoVLX-NEXT: vpmovdb %zmm2, %xmm2
; NoVLX-NEXT: vxorps %xmm3, %xmm3, %xmm3 ; NoVLX-NEXT: vmovdqa {{.*#+}} ymm3 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128]
; NoVLX-NEXT: vmovdqa {{.*#+}} ymm4 = [128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128,128] ; NoVLX-NEXT: vpxor %ymm3, %ymm0, %ymm0
; NoVLX-NEXT: vpxor %ymm4, %ymm0, %ymm0 ; NoVLX-NEXT: vpxor (%rsi), %ymm3, %ymm3
; NoVLX-NEXT: vpxor (%rsi), %ymm4, %ymm4 ; NoVLX-NEXT: vpcmpgtb %ymm0, %ymm3, %ymm0
; NoVLX-NEXT: vpcmpgtb %ymm0, %ymm4, %ymm0 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm3
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm4 ; NoVLX-NEXT: vpand %xmm2, %xmm3, %xmm2
; NoVLX-NEXT: vpand %xmm2, %xmm4, %xmm2
; NoVLX-NEXT: vpmovsxbd %xmm2, %zmm2 ; NoVLX-NEXT: vpmovsxbd %xmm2, %zmm2
; NoVLX-NEXT: vpslld $31, %zmm2, %zmm2 ; NoVLX-NEXT: vpslld $31, %zmm2, %zmm2
; NoVLX-NEXT: vptestmd %zmm2, %zmm2, %k0 ; NoVLX-NEXT: vptestmd %zmm2, %zmm2, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpand %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0 ; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT: kmovw %k0, (%rsp) ; NoVLX-NEXT: kmovw %k0, (%rsp)
; NoVLX-NEXT: movl (%rsp), %ecx ; NoVLX-NEXT: movl (%rsp), %ecx
; NoVLX-NEXT: vptestmd %zmm3, %zmm3, %k0 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: movl {{[0-9]+}}(%rsp), %eax ; NoVLX-NEXT: movl {{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT: shlq $32, %rax ; NoVLX-NEXT: shlq $32, %rax
; NoVLX-NEXT: orq %rcx, %rax ; NoVLX-NEXT: orq %rcx, %rax
; NoVLX-NEXT: movq %rbp, %rsp ; NoVLX-NEXT: movq %rbp, %rsp
; NoVLX-NEXT: popq %rbp ; NoVLX-NEXT: popq %rbp
; NoVLX-NEXT: retq ; NoVLX-NEXT: retq
▲ Show 20 Lines • Show All 166 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: subq $32, %rsp ; NoVLX-NEXT: subq $32, %rsp
; NoVLX-NEXT: vmovdqa {{.*#+}} xmm2 = [32768,32768,32768,32768,32768,32768,32768,32768] ; NoVLX-NEXT: vmovdqa {{.*#+}} xmm2 = [32768,32768,32768,32768,32768,32768,32768,32768]
; NoVLX-NEXT: vpxor %xmm2, %xmm0, %xmm0 ; NoVLX-NEXT: vpxor %xmm2, %xmm0, %xmm0
; NoVLX-NEXT: vpxor %xmm2, %xmm1, %xmm1 ; NoVLX-NEXT: vpxor %xmm2, %xmm1, %xmm1
; NoVLX-NEXT: vpcmpgtw %xmm0, %xmm1, %xmm0 ; NoVLX-NEXT: vpcmpgtw %xmm0, %xmm1, %xmm0
; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0 ; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0
; NoVLX-NEXT: vpsllq $63, %zmm0, %zmm0 ; NoVLX-NEXT: vpsllq $63, %zmm0, %zmm0
; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0 ; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r9d ; NoVLX-NEXT: kmovw %k1, %r9d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 61 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: subq $32, %rsp ; NoVLX-NEXT: subq $32, %rsp
; NoVLX-NEXT: vmovdqa {{.*#+}} xmm1 = [32768,32768,32768,32768,32768,32768,32768,32768] ; NoVLX-NEXT: vmovdqa {{.*#+}} xmm1 = [32768,32768,32768,32768,32768,32768,32768,32768]
; NoVLX-NEXT: vpxor %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vpxor (%rdi), %xmm1, %xmm1 ; NoVLX-NEXT: vpxor (%rdi), %xmm1, %xmm1
; NoVLX-NEXT: vpcmpgtw %xmm0, %xmm1, %xmm0 ; NoVLX-NEXT: vpcmpgtw %xmm0, %xmm1, %xmm0
; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0 ; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0
; NoVLX-NEXT: vpsllq $63, %zmm0, %zmm0 ; NoVLX-NEXT: vpsllq $63, %zmm0, %zmm0
; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0 ; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r9d ; NoVLX-NEXT: kmovw %k1, %r9d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 64 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: vmovdqa {{.*#+}} xmm2 = [32768,32768,32768,32768,32768,32768,32768,32768] ; NoVLX-NEXT: vmovdqa {{.*#+}} xmm2 = [32768,32768,32768,32768,32768,32768,32768,32768]
; NoVLX-NEXT: vpxor %xmm2, %xmm0, %xmm0 ; NoVLX-NEXT: vpxor %xmm2, %xmm0, %xmm0
; NoVLX-NEXT: vpxor %xmm2, %xmm1, %xmm1 ; NoVLX-NEXT: vpxor %xmm2, %xmm1, %xmm1
; NoVLX-NEXT: vpcmpgtw %xmm0, %xmm1, %xmm0 ; NoVLX-NEXT: vpcmpgtw %xmm0, %xmm1, %xmm0
; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0 ; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0
; NoVLX-NEXT: vpsllq $63, %zmm0, %zmm0 ; NoVLX-NEXT: vpsllq $63, %zmm0, %zmm0
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0 {%k1} ; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r9d ; NoVLX-NEXT: kmovw %k1, %r9d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 65 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: vmovdqa {{.*#+}} xmm1 = [32768,32768,32768,32768,32768,32768,32768,32768] ; NoVLX-NEXT: vmovdqa {{.*#+}} xmm1 = [32768,32768,32768,32768,32768,32768,32768,32768]
; NoVLX-NEXT: vpxor %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vpxor (%rsi), %xmm1, %xmm1 ; NoVLX-NEXT: vpxor (%rsi), %xmm1, %xmm1
; NoVLX-NEXT: vpcmpgtw %xmm0, %xmm1, %xmm0 ; NoVLX-NEXT: vpcmpgtw %xmm0, %xmm1, %xmm0
; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0 ; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0
; NoVLX-NEXT: vpsllq $63, %zmm0, %zmm0 ; NoVLX-NEXT: vpsllq $63, %zmm0, %zmm0
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0 {%k1} ; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r9d ; NoVLX-NEXT: kmovw %k1, %r9d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 65 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: vmovdqa {{.*#+}} xmm2 = [32768,32768,32768,32768,32768,32768,32768,32768] ; NoVLX-NEXT: vmovdqa {{.*#+}} xmm2 = [32768,32768,32768,32768,32768,32768,32768,32768]
; NoVLX-NEXT: vpxor %xmm2, %xmm0, %xmm0 ; NoVLX-NEXT: vpxor %xmm2, %xmm0, %xmm0
; NoVLX-NEXT: vpxor %xmm2, %xmm1, %xmm1 ; NoVLX-NEXT: vpxor %xmm2, %xmm1, %xmm1
; NoVLX-NEXT: vpcmpgtw %xmm0, %xmm1, %xmm0 ; NoVLX-NEXT: vpcmpgtw %xmm0, %xmm1, %xmm0
; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0 ; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0
; NoVLX-NEXT: vpsllq $63, %zmm0, %zmm0 ; NoVLX-NEXT: vpsllq $63, %zmm0, %zmm0
; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0 ; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 66 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: vmovdqa {{.*#+}} xmm1 = [32768,32768,32768,32768,32768,32768,32768,32768] ; NoVLX-NEXT: vmovdqa {{.*#+}} xmm1 = [32768,32768,32768,32768,32768,32768,32768,32768]
; NoVLX-NEXT: vpxor %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vpxor (%rdi), %xmm1, %xmm1 ; NoVLX-NEXT: vpxor (%rdi), %xmm1, %xmm1
; NoVLX-NEXT: vpcmpgtw %xmm0, %xmm1, %xmm0 ; NoVLX-NEXT: vpcmpgtw %xmm0, %xmm1, %xmm0
; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0 ; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0
; NoVLX-NEXT: vpsllq $63, %zmm0, %zmm0 ; NoVLX-NEXT: vpsllq $63, %zmm0, %zmm0
; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0 ; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 69 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: vmovdqa {{.*#+}} xmm2 = [32768,32768,32768,32768,32768,32768,32768,32768] ; NoVLX-NEXT: vmovdqa {{.*#+}} xmm2 = [32768,32768,32768,32768,32768,32768,32768,32768]
; NoVLX-NEXT: vpxor %xmm2, %xmm0, %xmm0 ; NoVLX-NEXT: vpxor %xmm2, %xmm0, %xmm0
; NoVLX-NEXT: vpxor %xmm2, %xmm1, %xmm1 ; NoVLX-NEXT: vpxor %xmm2, %xmm1, %xmm1
; NoVLX-NEXT: vpcmpgtw %xmm0, %xmm1, %xmm0 ; NoVLX-NEXT: vpcmpgtw %xmm0, %xmm1, %xmm0
; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0 ; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0
; NoVLX-NEXT: vpsllq $63, %zmm0, %zmm0 ; NoVLX-NEXT: vpsllq $63, %zmm0, %zmm0
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0 {%k1} ; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 70 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: vmovdqa {{.*#+}} xmm1 = [32768,32768,32768,32768,32768,32768,32768,32768] ; NoVLX-NEXT: vmovdqa {{.*#+}} xmm1 = [32768,32768,32768,32768,32768,32768,32768,32768]
; NoVLX-NEXT: vpxor %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vpxor (%rsi), %xmm1, %xmm1 ; NoVLX-NEXT: vpxor (%rsi), %xmm1, %xmm1
; NoVLX-NEXT: vpcmpgtw %xmm0, %xmm1, %xmm0 ; NoVLX-NEXT: vpcmpgtw %xmm0, %xmm1, %xmm0
; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0 ; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0
; NoVLX-NEXT: vpsllq $63, %zmm0, %zmm0 ; NoVLX-NEXT: vpsllq $63, %zmm0, %zmm0
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0 {%k1} ; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 86 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: vmovdqa {{.*#+}} ymm2 = [32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768] ; NoVLX-NEXT: vmovdqa {{.*#+}} ymm2 = [32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768]
; NoVLX-NEXT: vpxor %ymm2, %ymm0, %ymm0 ; NoVLX-NEXT: vpxor %ymm2, %ymm0, %ymm0
; NoVLX-NEXT: vpxor %ymm2, %ymm1, %ymm1 ; NoVLX-NEXT: vpxor %ymm2, %ymm1, %ymm1
; NoVLX-NEXT: vpcmpgtw %ymm0, %ymm1, %ymm0 ; NoVLX-NEXT: vpcmpgtw %ymm0, %ymm1, %ymm0
; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0 ; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0 ; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r10d ; NoVLX-NEXT: kmovw %k1, %r10d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 112 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: vmovdqa {{.*#+}} ymm1 = [32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768] ; NoVLX-NEXT: vmovdqa {{.*#+}} ymm1 = [32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768]
; NoVLX-NEXT: vpxor %ymm1, %ymm0, %ymm0 ; NoVLX-NEXT: vpxor %ymm1, %ymm0, %ymm0
; NoVLX-NEXT: vpxor (%rdi), %ymm1, %ymm1 ; NoVLX-NEXT: vpxor (%rdi), %ymm1, %ymm1
; NoVLX-NEXT: vpcmpgtw %ymm0, %ymm1, %ymm0 ; NoVLX-NEXT: vpcmpgtw %ymm0, %ymm1, %ymm0
; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0 ; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0 ; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r10d ; NoVLX-NEXT: kmovw %k1, %r10d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 115 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: vmovdqa {{.*#+}} ymm2 = [32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768] ; NoVLX-NEXT: vmovdqa {{.*#+}} ymm2 = [32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768]
; NoVLX-NEXT: vpxor %ymm2, %ymm0, %ymm0 ; NoVLX-NEXT: vpxor %ymm2, %ymm0, %ymm0
; NoVLX-NEXT: vpxor %ymm2, %ymm1, %ymm1 ; NoVLX-NEXT: vpxor %ymm2, %ymm1, %ymm1
; NoVLX-NEXT: vpcmpgtw %ymm0, %ymm1, %ymm0 ; NoVLX-NEXT: vpcmpgtw %ymm0, %ymm1, %ymm0
; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0 ; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0 ; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 {%k1} ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r10d ; NoVLX-NEXT: kmovw %k1, %r10d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 116 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: vmovdqa {{.*#+}} ymm1 = [32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768] ; NoVLX-NEXT: vmovdqa {{.*#+}} ymm1 = [32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768]
; NoVLX-NEXT: vpxor %ymm1, %ymm0, %ymm0 ; NoVLX-NEXT: vpxor %ymm1, %ymm0, %ymm0
; NoVLX-NEXT: vpxor (%rsi), %ymm1, %ymm1 ; NoVLX-NEXT: vpxor (%rsi), %ymm1, %ymm1
; NoVLX-NEXT: vpcmpgtw %ymm0, %ymm1, %ymm0 ; NoVLX-NEXT: vpcmpgtw %ymm0, %ymm1, %ymm0
; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0 ; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0 ; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 {%k1} ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r10d ; NoVLX-NEXT: kmovw %k1, %r10d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 116 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: vmovdqa {{.*#+}} ymm2 = [32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768] ; NoVLX-NEXT: vmovdqa {{.*#+}} ymm2 = [32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768]
; NoVLX-NEXT: vpxor %ymm2, %ymm0, %ymm0 ; NoVLX-NEXT: vpxor %ymm2, %ymm0, %ymm0
; NoVLX-NEXT: vpxor %ymm2, %ymm1, %ymm1 ; NoVLX-NEXT: vpxor %ymm2, %ymm1, %ymm1
; NoVLX-NEXT: vpcmpgtw %ymm0, %ymm1, %ymm0 ; NoVLX-NEXT: vpcmpgtw %ymm0, %ymm1, %ymm0
; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0 ; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0 ; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 117 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: vmovdqa {{.*#+}} ymm1 = [32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768] ; NoVLX-NEXT: vmovdqa {{.*#+}} ymm1 = [32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768]
; NoVLX-NEXT: vpxor %ymm1, %ymm0, %ymm0 ; NoVLX-NEXT: vpxor %ymm1, %ymm0, %ymm0
; NoVLX-NEXT: vpxor (%rdi), %ymm1, %ymm1 ; NoVLX-NEXT: vpxor (%rdi), %ymm1, %ymm1
; NoVLX-NEXT: vpcmpgtw %ymm0, %ymm1, %ymm0 ; NoVLX-NEXT: vpcmpgtw %ymm0, %ymm1, %ymm0
; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0 ; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0 ; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 120 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: vmovdqa {{.*#+}} ymm2 = [32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768] ; NoVLX-NEXT: vmovdqa {{.*#+}} ymm2 = [32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768]
; NoVLX-NEXT: vpxor %ymm2, %ymm0, %ymm0 ; NoVLX-NEXT: vpxor %ymm2, %ymm0, %ymm0
; NoVLX-NEXT: vpxor %ymm2, %ymm1, %ymm1 ; NoVLX-NEXT: vpxor %ymm2, %ymm1, %ymm1
; NoVLX-NEXT: vpcmpgtw %ymm0, %ymm1, %ymm0 ; NoVLX-NEXT: vpcmpgtw %ymm0, %ymm1, %ymm0
; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0 ; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0 ; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 {%k1} ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 121 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: vmovdqa {{.*#+}} ymm1 = [32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768] ; NoVLX-NEXT: vmovdqa {{.*#+}} ymm1 = [32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768]
; NoVLX-NEXT: vpxor %ymm1, %ymm0, %ymm0 ; NoVLX-NEXT: vpxor %ymm1, %ymm0, %ymm0
; NoVLX-NEXT: vpxor (%rsi), %ymm1, %ymm1 ; NoVLX-NEXT: vpxor (%rsi), %ymm1, %ymm1
; NoVLX-NEXT: vpcmpgtw %ymm0, %ymm1, %ymm0 ; NoVLX-NEXT: vpcmpgtw %ymm0, %ymm1, %ymm0
; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0 ; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0 ; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 {%k1} ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 417 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: kshiftrw $15, %k0, %k0 ; NoVLX-NEXT: kshiftrw $15, %k0, %k0
; NoVLX-NEXT: kmovw %k0, %eax ; NoVLX-NEXT: kmovw %k0, %eax
; NoVLX-NEXT: vpinsrb $15, %eax, %xmm0, %xmm0 ; NoVLX-NEXT: vpinsrb $15, %eax, %xmm0, %xmm0
; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0 ; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT: kmovw %k0, (%rsp) ; NoVLX-NEXT: kmovw %k0, (%rsp)
; NoVLX-NEXT: movl (%rsp), %ecx ; NoVLX-NEXT: movl (%rsp), %ecx
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: movl {{[0-9]+}}(%rsp), %eax ; NoVLX-NEXT: movl {{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT: shlq $32, %rax ; NoVLX-NEXT: shlq $32, %rax
; NoVLX-NEXT: orq %rcx, %rax ; NoVLX-NEXT: orq %rcx, %rax
; NoVLX-NEXT: movq %rbp, %rsp ; NoVLX-NEXT: movq %rbp, %rsp
; NoVLX-NEXT: popq %rbp ; NoVLX-NEXT: popq %rbp
; NoVLX-NEXT: retq ; NoVLX-NEXT: retq
▲ Show 20 Lines • Show All 254 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: kshiftrw $15, %k0, %k0 ; NoVLX-NEXT: kshiftrw $15, %k0, %k0
; NoVLX-NEXT: kmovw %k0, %eax ; NoVLX-NEXT: kmovw %k0, %eax
; NoVLX-NEXT: vpinsrb $15, %eax, %xmm0, %xmm0 ; NoVLX-NEXT: vpinsrb $15, %eax, %xmm0, %xmm0
; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0 ; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT: kmovw %k0, (%rsp) ; NoVLX-NEXT: kmovw %k0, (%rsp)
; NoVLX-NEXT: movl (%rsp), %ecx ; NoVLX-NEXT: movl (%rsp), %ecx
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: movl {{[0-9]+}}(%rsp), %eax ; NoVLX-NEXT: movl {{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT: shlq $32, %rax ; NoVLX-NEXT: shlq $32, %rax
; NoVLX-NEXT: orq %rcx, %rax ; NoVLX-NEXT: orq %rcx, %rax
; NoVLX-NEXT: movq %rbp, %rsp ; NoVLX-NEXT: movq %rbp, %rsp
; NoVLX-NEXT: popq %rbp ; NoVLX-NEXT: popq %rbp
; NoVLX-NEXT: retq ; NoVLX-NEXT: retq
▲ Show 20 Lines • Show All 150 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: shrl $16, %ecx ; NoVLX-NEXT: shrl $16, %ecx
; NoVLX-NEXT: vpinsrw $4, %eax, %xmm3, %xmm3 ; NoVLX-NEXT: vpinsrw $4, %eax, %xmm3, %xmm3
; NoVLX-NEXT: vpinsrw $5, %ecx, %xmm3, %xmm3 ; NoVLX-NEXT: vpinsrw $5, %ecx, %xmm3, %xmm3
; NoVLX-NEXT: movq %rax, %rcx ; NoVLX-NEXT: movq %rax, %rcx
; NoVLX-NEXT: shrq $32, %rcx ; NoVLX-NEXT: shrq $32, %rcx
; NoVLX-NEXT: vpinsrw $6, %ecx, %xmm3, %xmm3 ; NoVLX-NEXT: vpinsrw $6, %ecx, %xmm3, %xmm3
; NoVLX-NEXT: vmovq %xmm8, %rcx ; NoVLX-NEXT: vmovq %xmm8, %rcx
; NoVLX-NEXT: shrq $48, %rax ; NoVLX-NEXT: shrq $48, %rax
; NoVLX-NEXT: vpinsrw $7, %eax, %xmm3, %xmm5 ; NoVLX-NEXT: vpinsrw $7, %eax, %xmm3, %xmm3
; NoVLX-NEXT: movl %ecx, %eax ; NoVLX-NEXT: movl %ecx, %eax
; NoVLX-NEXT: shrl $16, %eax ; NoVLX-NEXT: shrl $16, %eax
; NoVLX-NEXT: vmovd %ecx, %xmm3 ; NoVLX-NEXT: vmovd %ecx, %xmm5
; NoVLX-NEXT: vpinsrw $1, %eax, %xmm3, %xmm3 ; NoVLX-NEXT: vpinsrw $1, %eax, %xmm5, %xmm5
; NoVLX-NEXT: movq %rcx, %rax ; NoVLX-NEXT: movq %rcx, %rax
; NoVLX-NEXT: shrq $32, %rax ; NoVLX-NEXT: shrq $32, %rax
; NoVLX-NEXT: vpinsrw $2, %eax, %xmm3, %xmm3 ; NoVLX-NEXT: vpinsrw $2, %eax, %xmm5, %xmm5
; NoVLX-NEXT: vpextrq $1, %xmm8, %rax ; NoVLX-NEXT: vpextrq $1, %xmm8, %rax
; NoVLX-NEXT: shrq $48, %rcx ; NoVLX-NEXT: shrq $48, %rcx
; NoVLX-NEXT: vpinsrw $3, %ecx, %xmm3, %xmm3 ; NoVLX-NEXT: vpinsrw $3, %ecx, %xmm5, %xmm5
; NoVLX-NEXT: movl %eax, %ecx ; NoVLX-NEXT: movl %eax, %ecx
; NoVLX-NEXT: shrl $16, %ecx ; NoVLX-NEXT: shrl $16, %ecx
; NoVLX-NEXT: vpinsrw $4, %eax, %xmm3, %xmm3 ; NoVLX-NEXT: vpinsrw $4, %eax, %xmm5, %xmm5
; NoVLX-NEXT: vpinsrw $5, %ecx, %xmm3, %xmm3 ; NoVLX-NEXT: vpinsrw $5, %ecx, %xmm5, %xmm5
; NoVLX-NEXT: movq %rax, %rcx ; NoVLX-NEXT: movq %rax, %rcx
; NoVLX-NEXT: shrq $32, %rcx ; NoVLX-NEXT: shrq $32, %rcx
; NoVLX-NEXT: vpinsrw $6, %ecx, %xmm3, %xmm3 ; NoVLX-NEXT: vpinsrw $6, %ecx, %xmm5, %xmm5
; NoVLX-NEXT: vmovq %xmm1, %rcx ; NoVLX-NEXT: vmovq %xmm1, %rcx
; NoVLX-NEXT: shrq $48, %rax ; NoVLX-NEXT: shrq $48, %rax
; NoVLX-NEXT: vpinsrw $7, %eax, %xmm3, %xmm2 ; NoVLX-NEXT: vpinsrw $7, %eax, %xmm5, %xmm5
; NoVLX-NEXT: movl %ecx, %eax ; NoVLX-NEXT: movl %ecx, %eax
; NoVLX-NEXT: shrl $16, %eax ; NoVLX-NEXT: shrl $16, %eax
; NoVLX-NEXT: vmovd %ecx, %xmm3 ; NoVLX-NEXT: vmovd %ecx, %xmm2
; NoVLX-NEXT: vpinsrw $1, %eax, %xmm3, %xmm3 ; NoVLX-NEXT: vpinsrw $1, %eax, %xmm2, %xmm2
; NoVLX-NEXT: movq %rcx, %rax ; NoVLX-NEXT: movq %rcx, %rax
; NoVLX-NEXT: shrq $32, %rax ; NoVLX-NEXT: shrq $32, %rax
; NoVLX-NEXT: vpinsrw $2, %eax, %xmm3, %xmm3 ; NoVLX-NEXT: vpinsrw $2, %eax, %xmm2, %xmm2
; NoVLX-NEXT: vpextrq $1, %xmm1, %rax ; NoVLX-NEXT: vpextrq $1, %xmm1, %rax
; NoVLX-NEXT: vinserti128 $1, %xmm9, %ymm4, %ymm8 ; NoVLX-NEXT: vinserti128 $1, %xmm9, %ymm4, %ymm8
; NoVLX-NEXT: kmovw {{[0-9]+}}(%rsp), %k1 ; NoVLX-NEXT: kmovw {{[0-9]+}}(%rsp), %k1
; NoVLX-NEXT: kmovw {{[0-9]+}}(%rsp), %k2 ; NoVLX-NEXT: kmovw {{[0-9]+}}(%rsp), %k2
; NoVLX-NEXT: vpternlogd $255, %zmm1, %zmm1, %zmm1 {%k1} {z} ; NoVLX-NEXT: vinserti128 $1, %xmm6, %ymm0, %ymm6
; NoVLX-NEXT: vinserti128 $1, %xmm6, %ymm0, %ymm4 ; NoVLX-NEXT: vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT: vpmovdb %zmm1, %xmm0 ; NoVLX-NEXT: vpmovdb %zmm0, %xmm0
; NoVLX-NEXT: vinserti128 $1, %xmm7, %ymm3, %ymm4
; NoVLX-NEXT: vpternlogd $255, %zmm1, %zmm1, %zmm1 {%k2} {z} ; NoVLX-NEXT: vpternlogd $255, %zmm1, %zmm1, %zmm1 {%k2} {z}
; NoVLX-NEXT: vinserti128 $1, %xmm7, %ymm5, %ymm5 ; NoVLX-NEXT: vpmovdb %zmm1, %xmm1
; NoVLX-NEXT: vpmovdb %zmm1, %xmm7
; NoVLX-NEXT: shrq $48, %rcx ; NoVLX-NEXT: shrq $48, %rcx
; NoVLX-NEXT: vpinsrw $3, %ecx, %xmm3, %xmm1 ; NoVLX-NEXT: vpinsrw $3, %ecx, %xmm2, %xmm2
; NoVLX-NEXT: movl %eax, %ecx ; NoVLX-NEXT: movl %eax, %ecx
; NoVLX-NEXT: shrl $16, %ecx ; NoVLX-NEXT: shrl $16, %ecx
; NoVLX-NEXT: vpinsrw $4, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrw $4, %eax, %xmm2, %xmm2
; NoVLX-NEXT: vpinsrw $5, %ecx, %xmm1, %xmm3 ; NoVLX-NEXT: vpinsrw $5, %ecx, %xmm2, %xmm2
; NoVLX-NEXT: movq %rax, %rcx ; NoVLX-NEXT: movq %rax, %rcx
; NoVLX-NEXT: shrq $48, %rax ; NoVLX-NEXT: shrq $48, %rax
; NoVLX-NEXT: shrq $32, %rcx ; NoVLX-NEXT: shrq $32, %rcx
; NoVLX-NEXT: vpxor %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrw $6, %ecx, %xmm2, %xmm2
; NoVLX-NEXT: vpinsrw $6, %ecx, %xmm3, %xmm3 ; NoVLX-NEXT: vpinsrw $7, %eax, %xmm2, %xmm2
; NoVLX-NEXT: vpinsrw $7, %eax, %xmm3, %xmm3 ; NoVLX-NEXT: vinserti128 $1, %xmm5, %ymm2, %ymm2
; NoVLX-NEXT: vinserti128 $1, %xmm2, %ymm3, %ymm2 ; NoVLX-NEXT: vmovdqa {{.*#+}} ymm5 = [32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768]
; NoVLX-NEXT: vmovdqa {{.*#+}} ymm6 = [32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768] ; NoVLX-NEXT: vpxor %ymm5, %ymm6, %ymm3
; NoVLX-NEXT: vpxor %ymm6, %ymm4, %ymm3 ; NoVLX-NEXT: vpxor %ymm5, %ymm2, %ymm2
; NoVLX-NEXT: vpxor %ymm6, %ymm2, %ymm2
; NoVLX-NEXT: vpcmpgtw %ymm3, %ymm2, %ymm2 ; NoVLX-NEXT: vpcmpgtw %ymm3, %ymm2, %ymm2
; NoVLX-NEXT: vpmovsxwd %ymm2, %zmm2 ; NoVLX-NEXT: vpmovsxwd %ymm2, %zmm2
; NoVLX-NEXT: vpslld $31, %zmm2, %zmm2 ; NoVLX-NEXT: vpslld $31, %zmm2, %zmm2
; NoVLX-NEXT: vptestmd %zmm2, %zmm2, %k0 ; NoVLX-NEXT: vptestmd %zmm2, %zmm2, %k0
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
▲ Show 20 Lines • Show All 50 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $13, %eax, %xmm2, %xmm2 ; NoVLX-NEXT: vpinsrb $13, %eax, %xmm2, %xmm2
; NoVLX-NEXT: kshiftlw $1, %k0, %k1 ; NoVLX-NEXT: kshiftlw $1, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $14, %eax, %xmm2, %xmm2 ; NoVLX-NEXT: vpinsrb $14, %eax, %xmm2, %xmm2
; NoVLX-NEXT: kshiftrw $15, %k0, %k0 ; NoVLX-NEXT: kshiftrw $15, %k0, %k0
; NoVLX-NEXT: kmovw %k0, %eax ; NoVLX-NEXT: kmovw %k0, %eax
; NoVLX-NEXT: vpinsrb $15, %eax, %xmm2, %xmm4 ; NoVLX-NEXT: vpinsrb $15, %eax, %xmm2, %xmm3
; NoVLX-NEXT: vpxor %ymm6, %ymm8, %ymm2 ; NoVLX-NEXT: vpxor %ymm5, %ymm8, %ymm2
; NoVLX-NEXT: vpxor %ymm6, %ymm5, %ymm3 ; NoVLX-NEXT: vpxor %ymm5, %ymm4, %ymm4
; NoVLX-NEXT: vpcmpgtw %ymm2, %ymm3, %ymm2 ; NoVLX-NEXT: vpcmpgtw %ymm2, %ymm4, %ymm2
; NoVLX-NEXT: vpmovsxwd %ymm2, %zmm2 ; NoVLX-NEXT: vpmovsxwd %ymm2, %zmm2
; NoVLX-NEXT: vpslld $31, %zmm2, %zmm2 ; NoVLX-NEXT: vpslld $31, %zmm2, %zmm2
; NoVLX-NEXT: vptestmd %zmm2, %zmm2, %k0 ; NoVLX-NEXT: vptestmd %zmm2, %zmm2, %k0
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 50 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: vpinsrb $13, %eax, %xmm2, %xmm2 ; NoVLX-NEXT: vpinsrb $13, %eax, %xmm2, %xmm2
; NoVLX-NEXT: kshiftlw $1, %k0, %k1 ; NoVLX-NEXT: kshiftlw $1, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $14, %eax, %xmm2, %xmm2 ; NoVLX-NEXT: vpinsrb $14, %eax, %xmm2, %xmm2
; NoVLX-NEXT: kshiftrw $15, %k0, %k0 ; NoVLX-NEXT: kshiftrw $15, %k0, %k0
; NoVLX-NEXT: kmovw %k0, %eax ; NoVLX-NEXT: kmovw %k0, %eax
; NoVLX-NEXT: vpinsrb $15, %eax, %xmm2, %xmm2 ; NoVLX-NEXT: vpinsrb $15, %eax, %xmm2, %xmm2
; NoVLX-NEXT: vpand %xmm7, %xmm2, %xmm2 ; NoVLX-NEXT: vpand %xmm1, %xmm2, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm2, %zmm2 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm2, %zmm2 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm2, %zmm2, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpand %xmm0, %xmm4, %xmm0 ; NoVLX-NEXT: vpand %xmm0, %xmm3, %xmm0
; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0 ; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT: kmovw %k0, (%rsp) ; NoVLX-NEXT: kmovw %k0, (%rsp)
; NoVLX-NEXT: movl (%rsp), %ecx ; NoVLX-NEXT: movl (%rsp), %ecx
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: movl {{[0-9]+}}(%rsp), %eax ; NoVLX-NEXT: movl {{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT: shlq $32, %rax ; NoVLX-NEXT: shlq $32, %rax
; NoVLX-NEXT: orq %rcx, %rax ; NoVLX-NEXT: orq %rcx, %rax
; NoVLX-NEXT: movq %rbp, %rsp ; NoVLX-NEXT: movq %rbp, %rsp
; NoVLX-NEXT: popq %rbp ; NoVLX-NEXT: popq %rbp
; NoVLX-NEXT: retq ; NoVLX-NEXT: retq
Show All 32 Lines
; NoVLX-NEXT: movl %edi, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: movl %edi, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vextracti32x4 $3, %zmm0, %xmm1 ; NoVLX-NEXT: vextracti32x4 $3, %zmm0, %xmm1
; NoVLX-NEXT: vmovq %xmm1, %rax ; NoVLX-NEXT: vmovq %xmm1, %rax
; NoVLX-NEXT: movq %rax, %rcx ; NoVLX-NEXT: movq %rax, %rcx
; NoVLX-NEXT: movq %rax, %rdx ; NoVLX-NEXT: movq %rax, %rdx
; NoVLX-NEXT: vmovd %eax, %xmm2 ; NoVLX-NEXT: vmovd %eax, %xmm2
; NoVLX-NEXT: shrl $16, %eax ; NoVLX-NEXT: shrl $16, %eax
; NoVLX-NEXT: vpinsrw $1, %eax, %xmm2, %xmm2 ; NoVLX-NEXT: vpinsrw $1, %eax, %xmm2, %xmm2
; NoVLX-NEXT: vextracti32x4 $1, %zmm0, %xmm4 ; NoVLX-NEXT: vextracti32x4 $1, %zmm0, %xmm3
; NoVLX-NEXT: vextracti32x4 $2, %zmm0, %xmm5 ; NoVLX-NEXT: vextracti32x4 $2, %zmm0, %xmm4
; NoVLX-NEXT: shrq $32, %rdx ; NoVLX-NEXT: shrq $32, %rdx
; NoVLX-NEXT: vpinsrw $2, %edx, %xmm2, %xmm2 ; NoVLX-NEXT: vpinsrw $2, %edx, %xmm2, %xmm2
; NoVLX-NEXT: vpextrq $1, %xmm1, %rax ; NoVLX-NEXT: vpextrq $1, %xmm1, %rax
; NoVLX-NEXT: shrq $48, %rcx ; NoVLX-NEXT: shrq $48, %rcx
; NoVLX-NEXT: vpinsrw $3, %ecx, %xmm2, %xmm1 ; NoVLX-NEXT: vpinsrw $3, %ecx, %xmm2, %xmm1
; NoVLX-NEXT: movl %eax, %ecx ; NoVLX-NEXT: movl %eax, %ecx
; NoVLX-NEXT: shrl $16, %ecx ; NoVLX-NEXT: shrl $16, %ecx
; NoVLX-NEXT: vpinsrw $4, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrw $4, %eax, %xmm1, %xmm1
; NoVLX-NEXT: vpinsrw $5, %ecx, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrw $5, %ecx, %xmm1, %xmm1
; NoVLX-NEXT: movq %rax, %rcx ; NoVLX-NEXT: movq %rax, %rcx
; NoVLX-NEXT: shrq $32, %rcx ; NoVLX-NEXT: shrq $32, %rcx
; NoVLX-NEXT: vpinsrw $6, %ecx, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrw $6, %ecx, %xmm1, %xmm1
; NoVLX-NEXT: vmovq %xmm5, %rcx ; NoVLX-NEXT: vmovq %xmm4, %rcx
; NoVLX-NEXT: shrq $48, %rax ; NoVLX-NEXT: shrq $48, %rax
; NoVLX-NEXT: vpinsrw $7, %eax, %xmm1, %xmm3 ; NoVLX-NEXT: vpinsrw $7, %eax, %xmm1, %xmm2
; NoVLX-NEXT: movl %ecx, %eax ; NoVLX-NEXT: movl %ecx, %eax
; NoVLX-NEXT: shrl $16, %eax ; NoVLX-NEXT: shrl $16, %eax
; NoVLX-NEXT: vmovd %ecx, %xmm1 ; NoVLX-NEXT: vmovd %ecx, %xmm1
; NoVLX-NEXT: vpinsrw $1, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrw $1, %eax, %xmm1, %xmm1
; NoVLX-NEXT: movq %rcx, %rax ; NoVLX-NEXT: movq %rcx, %rax
; NoVLX-NEXT: shrq $32, %rax ; NoVLX-NEXT: shrq $32, %rax
; NoVLX-NEXT: vpinsrw $2, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrw $2, %eax, %xmm1, %xmm1
; NoVLX-NEXT: vpextrq $1, %xmm5, %rax ; NoVLX-NEXT: vpextrq $1, %xmm4, %rax
; NoVLX-NEXT: shrq $48, %rcx ; NoVLX-NEXT: shrq $48, %rcx
; NoVLX-NEXT: vpinsrw $3, %ecx, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrw $3, %ecx, %xmm1, %xmm1
; NoVLX-NEXT: movl %eax, %ecx ; NoVLX-NEXT: movl %eax, %ecx
; NoVLX-NEXT: shrl $16, %ecx ; NoVLX-NEXT: shrl $16, %ecx
; NoVLX-NEXT: vpinsrw $4, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrw $4, %eax, %xmm1, %xmm1
; NoVLX-NEXT: vpinsrw $5, %ecx, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrw $5, %ecx, %xmm1, %xmm1
; NoVLX-NEXT: movq %rax, %rcx ; NoVLX-NEXT: movq %rax, %rcx
; NoVLX-NEXT: shrq $32, %rcx ; NoVLX-NEXT: shrq $32, %rcx
; NoVLX-NEXT: vpinsrw $6, %ecx, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrw $6, %ecx, %xmm1, %xmm1
; NoVLX-NEXT: vmovq %xmm4, %rcx ; NoVLX-NEXT: vmovq %xmm3, %rcx
; NoVLX-NEXT: shrq $48, %rax ; NoVLX-NEXT: shrq $48, %rax
; NoVLX-NEXT: vpinsrw $7, %eax, %xmm1, %xmm5 ; NoVLX-NEXT: vpinsrw $7, %eax, %xmm1, %xmm4
; NoVLX-NEXT: movl %ecx, %eax ; NoVLX-NEXT: movl %ecx, %eax
; NoVLX-NEXT: shrl $16, %eax ; NoVLX-NEXT: shrl $16, %eax
; NoVLX-NEXT: vmovd %ecx, %xmm1 ; NoVLX-NEXT: vmovd %ecx, %xmm1
; NoVLX-NEXT: vpinsrw $1, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrw $1, %eax, %xmm1, %xmm1
; NoVLX-NEXT: movq %rcx, %rax ; NoVLX-NEXT: movq %rcx, %rax
; NoVLX-NEXT: shrq $32, %rax ; NoVLX-NEXT: shrq $32, %rax
; NoVLX-NEXT: vpinsrw $2, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrw $2, %eax, %xmm1, %xmm1
; NoVLX-NEXT: vpextrq $1, %xmm4, %rax ; NoVLX-NEXT: vpextrq $1, %xmm3, %rax
; NoVLX-NEXT: shrq $48, %rcx ; NoVLX-NEXT: shrq $48, %rcx
; NoVLX-NEXT: vpinsrw $3, %ecx, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrw $3, %ecx, %xmm1, %xmm1
; NoVLX-NEXT: movl %eax, %ecx ; NoVLX-NEXT: movl %eax, %ecx
; NoVLX-NEXT: shrl $16, %ecx ; NoVLX-NEXT: shrl $16, %ecx
; NoVLX-NEXT: vpinsrw $4, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrw $4, %eax, %xmm1, %xmm1
; NoVLX-NEXT: vpinsrw $5, %ecx, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrw $5, %ecx, %xmm1, %xmm1
; NoVLX-NEXT: movq %rax, %rcx ; NoVLX-NEXT: movq %rax, %rcx
; NoVLX-NEXT: shrq $32, %rcx ; NoVLX-NEXT: shrq $32, %rcx
; NoVLX-NEXT: vpinsrw $6, %ecx, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrw $6, %ecx, %xmm1, %xmm1
; NoVLX-NEXT: vmovq %xmm0, %rcx ; NoVLX-NEXT: vmovq %xmm0, %rcx
; NoVLX-NEXT: shrq $48, %rax ; NoVLX-NEXT: shrq $48, %rax
; NoVLX-NEXT: vpinsrw $7, %eax, %xmm1, %xmm6 ; NoVLX-NEXT: vpinsrw $7, %eax, %xmm1, %xmm5
; NoVLX-NEXT: movl %ecx, %eax ; NoVLX-NEXT: movl %ecx, %eax
; NoVLX-NEXT: shrl $16, %eax ; NoVLX-NEXT: shrl $16, %eax
; NoVLX-NEXT: vmovd %ecx, %xmm1 ; NoVLX-NEXT: vmovd %ecx, %xmm1
; NoVLX-NEXT: vpinsrw $1, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrw $1, %eax, %xmm1, %xmm1
; NoVLX-NEXT: movq %rcx, %rax ; NoVLX-NEXT: movq %rcx, %rax
; NoVLX-NEXT: shrq $32, %rax ; NoVLX-NEXT: shrq $32, %rax
; NoVLX-NEXT: vpinsrw $2, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrw $2, %eax, %xmm1, %xmm1
; NoVLX-NEXT: vpextrq $1, %xmm0, %rax ; NoVLX-NEXT: vpextrq $1, %xmm0, %rax
; NoVLX-NEXT: shrq $48, %rcx ; NoVLX-NEXT: shrq $48, %rcx
; NoVLX-NEXT: vpinsrw $3, %ecx, %xmm1, %xmm0 ; NoVLX-NEXT: vpinsrw $3, %ecx, %xmm1, %xmm0
; NoVLX-NEXT: movl %eax, %ecx ; NoVLX-NEXT: movl %eax, %ecx
; NoVLX-NEXT: shrl $16, %ecx ; NoVLX-NEXT: shrl $16, %ecx
; NoVLX-NEXT: vpinsrw $4, %eax, %xmm0, %xmm0 ; NoVLX-NEXT: vpinsrw $4, %eax, %xmm0, %xmm0
; NoVLX-NEXT: vpinsrw $5, %ecx, %xmm0, %xmm0 ; NoVLX-NEXT: vpinsrw $5, %ecx, %xmm0, %xmm0
; NoVLX-NEXT: movq %rax, %rcx ; NoVLX-NEXT: movq %rax, %rcx
; NoVLX-NEXT: shrq $32, %rcx ; NoVLX-NEXT: shrq $32, %rcx
; NoVLX-NEXT: vpinsrw $6, %ecx, %xmm0, %xmm7 ; NoVLX-NEXT: vpinsrw $6, %ecx, %xmm0, %xmm6
; NoVLX-NEXT: kmovw {{[0-9]+}}(%rsp), %k1 ; NoVLX-NEXT: kmovw {{[0-9]+}}(%rsp), %k1
; NoVLX-NEXT: kmovw {{[0-9]+}}(%rsp), %k2 ; NoVLX-NEXT: kmovw {{[0-9]+}}(%rsp), %k2
; NoVLX-NEXT: vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z} ; NoVLX-NEXT: vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
; NoVLX-NEXT: vpmovdb %zmm0, %xmm1 ; NoVLX-NEXT: vpmovdb %zmm0, %xmm0
; NoVLX-NEXT: vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k2} {z} ; NoVLX-NEXT: vpternlogd $255, %zmm1, %zmm1, %zmm1 {%k2} {z}
; NoVLX-NEXT: vpmovdb %zmm0, %xmm2 ; NoVLX-NEXT: vpmovdb %zmm1, %xmm1
; NoVLX-NEXT: shrq $48, %rax ; NoVLX-NEXT: shrq $48, %rax
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: vinserti128 $1, %xmm2, %ymm4, %ymm3
; NoVLX-NEXT: vinserti128 $1, %xmm3, %ymm5, %ymm4 ; NoVLX-NEXT: vpinsrw $7, %eax, %xmm6, %xmm2
; NoVLX-NEXT: vpinsrw $7, %eax, %xmm7, %xmm3 ; NoVLX-NEXT: vinserti128 $1, %xmm5, %ymm2, %ymm2
; NoVLX-NEXT: vinserti128 $1, %xmm6, %ymm3, %ymm3 ; NoVLX-NEXT: vmovdqa {{.*#+}} ymm4 = [32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768]
; NoVLX-NEXT: vmovdqa {{.*#+}} ymm5 = [32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768,32768] ; NoVLX-NEXT: vpxor %ymm4, %ymm2, %ymm2
; NoVLX-NEXT: vpxor %ymm5, %ymm3, %ymm3 ; NoVLX-NEXT: vpxor (%rsi), %ymm4, %ymm5
; NoVLX-NEXT: vpxor (%rsi), %ymm5, %ymm6 ; NoVLX-NEXT: vpcmpgtw %ymm2, %ymm5, %ymm2
; NoVLX-NEXT: vpcmpgtw %ymm3, %ymm6, %ymm3 ; NoVLX-NEXT: vpmovsxwd %ymm2, %zmm2
; NoVLX-NEXT: vpmovsxwd %ymm3, %zmm3 ; NoVLX-NEXT: vpslld $31, %zmm2, %zmm2
; NoVLX-NEXT: vpslld $31, %zmm3, %zmm3 ; NoVLX-NEXT: vptestmd %zmm2, %zmm2, %k0
; NoVLX-NEXT: vptestmd %zmm3, %zmm3, %k0
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: vmovd %eax, %xmm3 ; NoVLX-NEXT: vmovd %eax, %xmm2
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $1, %eax, %xmm3, %xmm3 ; NoVLX-NEXT: vpinsrb $1, %eax, %xmm2, %xmm2
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $2, %eax, %xmm3, %xmm3 ; NoVLX-NEXT: vpinsrb $2, %eax, %xmm2, %xmm2
; NoVLX-NEXT: kshiftlw $12, %k0, %k1 ; NoVLX-NEXT: kshiftlw $12, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $3, %eax, %xmm3, %xmm3 ; NoVLX-NEXT: vpinsrb $3, %eax, %xmm2, %xmm2
; NoVLX-NEXT: kshiftlw $11, %k0, %k1 ; NoVLX-NEXT: kshiftlw $11, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $4, %eax, %xmm3, %xmm3 ; NoVLX-NEXT: vpinsrb $4, %eax, %xmm2, %xmm2
; NoVLX-NEXT: kshiftlw $10, %k0, %k1 ; NoVLX-NEXT: kshiftlw $10, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $5, %eax, %xmm3, %xmm3 ; NoVLX-NEXT: vpinsrb $5, %eax, %xmm2, %xmm2
; NoVLX-NEXT: kshiftlw $9, %k0, %k1 ; NoVLX-NEXT: kshiftlw $9, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $6, %eax, %xmm3, %xmm3 ; NoVLX-NEXT: vpinsrb $6, %eax, %xmm2, %xmm2
; NoVLX-NEXT: kshiftlw $8, %k0, %k1 ; NoVLX-NEXT: kshiftlw $8, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $7, %eax, %xmm3, %xmm3 ; NoVLX-NEXT: vpinsrb $7, %eax, %xmm2, %xmm2
; NoVLX-NEXT: kshiftlw $7, %k0, %k1 ; NoVLX-NEXT: kshiftlw $7, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $8, %eax, %xmm3, %xmm3 ; NoVLX-NEXT: vpinsrb $8, %eax, %xmm2, %xmm2
; NoVLX-NEXT: kshiftlw $6, %k0, %k1 ; NoVLX-NEXT: kshiftlw $6, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $9, %eax, %xmm3, %xmm3 ; NoVLX-NEXT: vpinsrb $9, %eax, %xmm2, %xmm2
; NoVLX-NEXT: kshiftlw $5, %k0, %k1 ; NoVLX-NEXT: kshiftlw $5, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $10, %eax, %xmm3, %xmm3 ; NoVLX-NEXT: vpinsrb $10, %eax, %xmm2, %xmm2
; NoVLX-NEXT: kshiftlw $4, %k0, %k1 ; NoVLX-NEXT: kshiftlw $4, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $11, %eax, %xmm3, %xmm3 ; NoVLX-NEXT: vpinsrb $11, %eax, %xmm2, %xmm2
; NoVLX-NEXT: kshiftlw $3, %k0, %k1 ; NoVLX-NEXT: kshiftlw $3, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $12, %eax, %xmm3, %xmm3 ; NoVLX-NEXT: vpinsrb $12, %eax, %xmm2, %xmm2
; NoVLX-NEXT: kshiftlw $2, %k0, %k1 ; NoVLX-NEXT: kshiftlw $2, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $13, %eax, %xmm3, %xmm3 ; NoVLX-NEXT: vpinsrb $13, %eax, %xmm2, %xmm2
; NoVLX-NEXT: kshiftlw $1, %k0, %k1 ; NoVLX-NEXT: kshiftlw $1, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $14, %eax, %xmm3, %xmm3 ; NoVLX-NEXT: vpinsrb $14, %eax, %xmm2, %xmm2
; NoVLX-NEXT: kshiftrw $15, %k0, %k0 ; NoVLX-NEXT: kshiftrw $15, %k0, %k0
; NoVLX-NEXT: kmovw %k0, %eax ; NoVLX-NEXT: kmovw %k0, %eax
; NoVLX-NEXT: vpinsrb $15, %eax, %xmm3, %xmm3 ; NoVLX-NEXT: vpinsrb $15, %eax, %xmm2, %xmm2
; NoVLX-NEXT: vpxor %ymm5, %ymm4, %ymm4 ; NoVLX-NEXT: vpxor %ymm4, %ymm3, %ymm3
; NoVLX-NEXT: vpxor 32(%rsi), %ymm5, %ymm5 ; NoVLX-NEXT: vpxor 32(%rsi), %ymm4, %ymm4
; NoVLX-NEXT: vpcmpgtw %ymm4, %ymm5, %ymm4 ; NoVLX-NEXT: vpcmpgtw %ymm3, %ymm4, %ymm3
; NoVLX-NEXT: vpmovsxwd %ymm4, %zmm4 ; NoVLX-NEXT: vpmovsxwd %ymm3, %zmm3
; NoVLX-NEXT: vpslld $31, %zmm4, %zmm4 ; NoVLX-NEXT: vpslld $31, %zmm3, %zmm3
; NoVLX-NEXT: vptestmd %zmm4, %zmm4, %k0 ; NoVLX-NEXT: vptestmd %zmm3, %zmm3, %k0
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %ecx ; NoVLX-NEXT: kmovw %k1, %ecx
; NoVLX-NEXT: vmovd %ecx, %xmm4 ; NoVLX-NEXT: vmovd %ecx, %xmm3
; NoVLX-NEXT: vpinsrb $1, %eax, %xmm4, %xmm4 ; NoVLX-NEXT: vpinsrb $1, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $2, %eax, %xmm4, %xmm4 ; NoVLX-NEXT: vpinsrb $2, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftlw $12, %k0, %k1 ; NoVLX-NEXT: kshiftlw $12, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $3, %eax, %xmm4, %xmm4 ; NoVLX-NEXT: vpinsrb $3, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftlw $11, %k0, %k1 ; NoVLX-NEXT: kshiftlw $11, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $4, %eax, %xmm4, %xmm4 ; NoVLX-NEXT: vpinsrb $4, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftlw $10, %k0, %k1 ; NoVLX-NEXT: kshiftlw $10, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $5, %eax, %xmm4, %xmm4 ; NoVLX-NEXT: vpinsrb $5, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftlw $9, %k0, %k1 ; NoVLX-NEXT: kshiftlw $9, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $6, %eax, %xmm4, %xmm4 ; NoVLX-NEXT: vpinsrb $6, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftlw $8, %k0, %k1 ; NoVLX-NEXT: kshiftlw $8, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $7, %eax, %xmm4, %xmm4 ; NoVLX-NEXT: vpinsrb $7, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftlw $7, %k0, %k1 ; NoVLX-NEXT: kshiftlw $7, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $8, %eax, %xmm4, %xmm4 ; NoVLX-NEXT: vpinsrb $8, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftlw $6, %k0, %k1 ; NoVLX-NEXT: kshiftlw $6, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $9, %eax, %xmm4, %xmm4 ; NoVLX-NEXT: vpinsrb $9, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftlw $5, %k0, %k1 ; NoVLX-NEXT: kshiftlw $5, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $10, %eax, %xmm4, %xmm4 ; NoVLX-NEXT: vpinsrb $10, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftlw $4, %k0, %k1 ; NoVLX-NEXT: kshiftlw $4, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $11, %eax, %xmm4, %xmm4 ; NoVLX-NEXT: vpinsrb $11, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftlw $3, %k0, %k1 ; NoVLX-NEXT: kshiftlw $3, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $12, %eax, %xmm4, %xmm4 ; NoVLX-NEXT: vpinsrb $12, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftlw $2, %k0, %k1 ; NoVLX-NEXT: kshiftlw $2, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $13, %eax, %xmm4, %xmm4 ; NoVLX-NEXT: vpinsrb $13, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftlw $1, %k0, %k1 ; NoVLX-NEXT: kshiftlw $1, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $14, %eax, %xmm4, %xmm4 ; NoVLX-NEXT: vpinsrb $14, %eax, %xmm3, %xmm3
; NoVLX-NEXT: kshiftrw $15, %k0, %k0 ; NoVLX-NEXT: kshiftrw $15, %k0, %k0
; NoVLX-NEXT: kmovw %k0, %eax ; NoVLX-NEXT: kmovw %k0, %eax
; NoVLX-NEXT: vpinsrb $15, %eax, %xmm4, %xmm4 ; NoVLX-NEXT: vpinsrb $15, %eax, %xmm3, %xmm3
; NoVLX-NEXT: vpand %xmm2, %xmm4, %xmm2
; NoVLX-NEXT: vpmovsxbd %xmm2, %zmm2
; NoVLX-NEXT: vpslld $31, %zmm2, %zmm2
; NoVLX-NEXT: vptestmd %zmm2, %zmm2, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpand %xmm1, %xmm3, %xmm1 ; NoVLX-NEXT: vpand %xmm1, %xmm3, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpand %xmm0, %xmm2, %xmm0
; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
; NoVLX-NEXT: vpslld $31, %zmm0, %zmm0
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
; NoVLX-NEXT: kmovw %k0, (%rsp) ; NoVLX-NEXT: kmovw %k0, (%rsp)
; NoVLX-NEXT: movl (%rsp), %ecx ; NoVLX-NEXT: movl (%rsp), %ecx
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: movl {{[0-9]+}}(%rsp), %eax ; NoVLX-NEXT: movl {{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT: shlq $32, %rax ; NoVLX-NEXT: shlq $32, %rax
; NoVLX-NEXT: orq %rcx, %rax ; NoVLX-NEXT: orq %rcx, %rax
; NoVLX-NEXT: movq %rbp, %rsp ; NoVLX-NEXT: movq %rbp, %rsp
; NoVLX-NEXT: popq %rbp ; NoVLX-NEXT: popq %rbp
; NoVLX-NEXT: retq ; NoVLX-NEXT: retq
▲ Show 20 Lines • Show All 1,222 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi1552: ; NoVLX-NEXT: .Lcfi1552:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: vpbroadcastd {{.*#+}} xmm2 = [2147483648,2147483648,2147483648,2147483648] ; NoVLX-NEXT: vpbroadcastd {{.*#+}} xmm2 = [2147483648,2147483648,2147483648,2147483648]
; NoVLX-NEXT: vpxor %xmm2, %xmm0, %xmm0 ; NoVLX-NEXT: vpxor %xmm2, %xmm0, %xmm0
; NoVLX-NEXT: vpxor %xmm2, %xmm1, %xmm1 ; NoVLX-NEXT: vpxor %xmm2, %xmm1, %xmm1
; NoVLX-NEXT: vpcmpgtd %xmm0, %xmm1, %xmm0 ; NoVLX-NEXT: vpcmpgtd %xmm0, %xmm1, %xmm0
; NoVLX-NEXT: vpxor %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero ; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
Show All 35 Lines
; NoVLX-NEXT: .Lcfi1555: ; NoVLX-NEXT: .Lcfi1555:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: vpbroadcastd {{.*#+}} xmm1 = [2147483648,2147483648,2147483648,2147483648] ; NoVLX-NEXT: vpbroadcastd {{.*#+}} xmm1 = [2147483648,2147483648,2147483648,2147483648]
; NoVLX-NEXT: vpxor %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vpxor (%rdi), %xmm1, %xmm1 ; NoVLX-NEXT: vpxor (%rdi), %xmm1, %xmm1
; NoVLX-NEXT: vpcmpgtd %xmm0, %xmm1, %xmm0 ; NoVLX-NEXT: vpcmpgtd %xmm0, %xmm1, %xmm0
; NoVLX-NEXT: vpxor %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero ; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
▲ Show 20 Lines • Show All 55 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: kmovw %k3, %ecx ; NoVLX-NEXT: kmovw %k3, %ecx
; NoVLX-NEXT: vmovd %ecx, %xmm1 ; NoVLX-NEXT: vmovd %ecx, %xmm1
; NoVLX-NEXT: vpinsrb $4, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrb $4, %eax, %xmm1, %xmm1
; NoVLX-NEXT: kmovw %k2, %eax ; NoVLX-NEXT: kmovw %k2, %eax
; NoVLX-NEXT: vpinsrb $8, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $12, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT: vpand %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vpxor %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero ; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
▲ Show 20 Lines • Show All 57 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: kmovw %k3, %ecx ; NoVLX-NEXT: kmovw %k3, %ecx
; NoVLX-NEXT: vmovd %ecx, %xmm1 ; NoVLX-NEXT: vmovd %ecx, %xmm1
; NoVLX-NEXT: vpinsrb $4, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrb $4, %eax, %xmm1, %xmm1
; NoVLX-NEXT: kmovw %k2, %eax ; NoVLX-NEXT: kmovw %k2, %eax
; NoVLX-NEXT: vpinsrb $8, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $12, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT: vpand %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vpxor %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero ; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
▲ Show 20 Lines • Show All 41 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: vpbroadcastd (%rdi), %xmm1 ; NoVLX-NEXT: vpbroadcastd (%rdi), %xmm1
; NoVLX-NEXT: vpbroadcastd {{.*#+}} xmm2 = [2147483648,2147483648,2147483648,2147483648] ; NoVLX-NEXT: vpbroadcastd {{.*#+}} xmm2 = [2147483648,2147483648,2147483648,2147483648]
; NoVLX-NEXT: vpxor %xmm2, %xmm0, %xmm0 ; NoVLX-NEXT: vpxor %xmm2, %xmm0, %xmm0
; NoVLX-NEXT: vpxor %xmm2, %xmm1, %xmm1 ; NoVLX-NEXT: vpxor %xmm2, %xmm1, %xmm1
; NoVLX-NEXT: vpcmpgtd %xmm0, %xmm1, %xmm0 ; NoVLX-NEXT: vpcmpgtd %xmm0, %xmm1, %xmm0
; NoVLX-NEXT: vpxor %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero ; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
▲ Show 20 Lines • Show All 57 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: kmovw %k3, %ecx ; NoVLX-NEXT: kmovw %k3, %ecx
; NoVLX-NEXT: vmovd %ecx, %xmm1 ; NoVLX-NEXT: vmovd %ecx, %xmm1
; NoVLX-NEXT: vpinsrb $4, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrb $4, %eax, %xmm1, %xmm1
; NoVLX-NEXT: kmovw %k2, %eax ; NoVLX-NEXT: kmovw %k2, %eax
; NoVLX-NEXT: vpinsrb $8, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $12, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT: vpand %xmm0, %xmm1, %xmm0 ; NoVLX-NEXT: vpand %xmm0, %xmm1, %xmm0
; NoVLX-NEXT: vpxor %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero ; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
▲ Show 20 Lines • Show All 229 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi1570: ; NoVLX-NEXT: .Lcfi1570:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $32, %rsp ; NoVLX-NEXT: subq $32, %rsp
; NoVLX-NEXT: # kill: %YMM1<def> %YMM1<kill> %ZMM1<def> ; NoVLX-NEXT: # kill: %YMM1<def> %YMM1<kill> %ZMM1<def>
; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def> ; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def>
; NoVLX-NEXT: vpcmpltud %zmm1, %zmm0, %k0 ; NoVLX-NEXT: vpcmpltud %zmm1, %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r9d ; NoVLX-NEXT: kmovw %k1, %r9d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 58 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi1573: ; NoVLX-NEXT: .Lcfi1573:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $32, %rsp ; NoVLX-NEXT: subq $32, %rsp
; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def> ; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def>
; NoVLX-NEXT: vmovdqa (%rdi), %ymm1 ; NoVLX-NEXT: vmovdqa (%rdi), %ymm1
; NoVLX-NEXT: vpcmpltud %zmm1, %zmm0, %k0 ; NoVLX-NEXT: vpcmpltud %zmm1, %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r9d ; NoVLX-NEXT: kmovw %k1, %r9d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 62 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $32, %rsp ; NoVLX-NEXT: subq $32, %rsp
; NoVLX-NEXT: # kill: %YMM1<def> %YMM1<kill> %ZMM1<def> ; NoVLX-NEXT: # kill: %YMM1<def> %YMM1<kill> %ZMM1<def>
; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def> ; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def>
; NoVLX-NEXT: vpcmpltud %zmm1, %zmm0, %k0 ; NoVLX-NEXT: vpcmpltud %zmm1, %zmm0, %k0
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: kandw %k1, %k0, %k0 ; NoVLX-NEXT: kandw %k1, %k0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r9d ; NoVLX-NEXT: kmovw %k1, %r9d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 63 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $32, %rsp ; NoVLX-NEXT: subq $32, %rsp
; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def> ; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def>
; NoVLX-NEXT: vmovdqa (%rsi), %ymm1 ; NoVLX-NEXT: vmovdqa (%rsi), %ymm1
; NoVLX-NEXT: vpcmpltud %zmm1, %zmm0, %k0 ; NoVLX-NEXT: vpcmpltud %zmm1, %zmm0, %k0
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: kandw %k1, %k0, %k0 ; NoVLX-NEXT: kandw %k1, %k0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r9d ; NoVLX-NEXT: kmovw %k1, %r9d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 62 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi1582: ; NoVLX-NEXT: .Lcfi1582:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $32, %rsp ; NoVLX-NEXT: subq $32, %rsp
; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def> ; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def>
; NoVLX-NEXT: vpbroadcastd (%rdi), %ymm1 ; NoVLX-NEXT: vpbroadcastd (%rdi), %ymm1
; NoVLX-NEXT: vpcmpltud %zmm1, %zmm0, %k0 ; NoVLX-NEXT: vpcmpltud %zmm1, %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r9d ; NoVLX-NEXT: kmovw %k1, %r9d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 63 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $32, %rsp ; NoVLX-NEXT: subq $32, %rsp
; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def> ; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def>
; NoVLX-NEXT: vpbroadcastd (%rsi), %ymm1 ; NoVLX-NEXT: vpbroadcastd (%rsi), %ymm1
; NoVLX-NEXT: vpcmpltud %zmm1, %zmm0, %k0 ; NoVLX-NEXT: vpcmpltud %zmm1, %zmm0, %k0
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: kandw %k0, %k1, %k0 ; NoVLX-NEXT: kandw %k0, %k1, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r9d ; NoVLX-NEXT: kmovw %k1, %r9d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 63 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi1588: ; NoVLX-NEXT: .Lcfi1588:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: # kill: %YMM1<def> %YMM1<kill> %ZMM1<def> ; NoVLX-NEXT: # kill: %YMM1<def> %YMM1<kill> %ZMM1<def>
; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def> ; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def>
; NoVLX-NEXT: vpcmpltud %zmm1, %zmm0, %k0 ; NoVLX-NEXT: vpcmpltud %zmm1, %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 63 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi1591: ; NoVLX-NEXT: .Lcfi1591:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def> ; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def>
; NoVLX-NEXT: vmovdqa (%rdi), %ymm1 ; NoVLX-NEXT: vmovdqa (%rdi), %ymm1
; NoVLX-NEXT: vpcmpltud %zmm1, %zmm0, %k0 ; NoVLX-NEXT: vpcmpltud %zmm1, %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 67 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: # kill: %YMM1<def> %YMM1<kill> %ZMM1<def> ; NoVLX-NEXT: # kill: %YMM1<def> %YMM1<kill> %ZMM1<def>
; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def> ; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def>
; NoVLX-NEXT: vpcmpltud %zmm1, %zmm0, %k0 ; NoVLX-NEXT: vpcmpltud %zmm1, %zmm0, %k0
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: kandw %k1, %k0, %k0 ; NoVLX-NEXT: kandw %k1, %k0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 68 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def> ; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def>
; NoVLX-NEXT: vmovdqa (%rsi), %ymm1 ; NoVLX-NEXT: vmovdqa (%rsi), %ymm1
; NoVLX-NEXT: vpcmpltud %zmm1, %zmm0, %k0 ; NoVLX-NEXT: vpcmpltud %zmm1, %zmm0, %k0
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: kandw %k1, %k0, %k0 ; NoVLX-NEXT: kandw %k1, %k0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 67 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi1600: ; NoVLX-NEXT: .Lcfi1600:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def> ; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def>
; NoVLX-NEXT: vpbroadcastd (%rdi), %ymm1 ; NoVLX-NEXT: vpbroadcastd (%rdi), %ymm1
; NoVLX-NEXT: vpcmpltud %zmm1, %zmm0, %k0 ; NoVLX-NEXT: vpcmpltud %zmm1, %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 68 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def> ; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def>
; NoVLX-NEXT: vpbroadcastd (%rsi), %ymm1 ; NoVLX-NEXT: vpbroadcastd (%rsi), %ymm1
; NoVLX-NEXT: vpcmpltud %zmm1, %zmm0, %k0 ; NoVLX-NEXT: vpcmpltud %zmm1, %zmm0, %k0
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: kandw %k0, %k1, %k0 ; NoVLX-NEXT: kandw %k0, %k1, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 81 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %r12, -48 ; NoVLX-NEXT: .cfi_offset %r12, -48
; NoVLX-NEXT: .Lcfi1609: ; NoVLX-NEXT: .Lcfi1609:
; NoVLX-NEXT: .cfi_offset %r13, -40 ; NoVLX-NEXT: .cfi_offset %r13, -40
; NoVLX-NEXT: .Lcfi1610: ; NoVLX-NEXT: .Lcfi1610:
; NoVLX-NEXT: .cfi_offset %r14, -32 ; NoVLX-NEXT: .cfi_offset %r14, -32
; NoVLX-NEXT: .Lcfi1611: ; NoVLX-NEXT: .Lcfi1611:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: vpcmpltud %zmm1, %zmm0, %k0 ; NoVLX-NEXT: vpcmpltud %zmm1, %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r10d ; NoVLX-NEXT: kmovw %k1, %r10d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 106 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %r12, -48 ; NoVLX-NEXT: .cfi_offset %r12, -48
; NoVLX-NEXT: .Lcfi1617: ; NoVLX-NEXT: .Lcfi1617:
; NoVLX-NEXT: .cfi_offset %r13, -40 ; NoVLX-NEXT: .cfi_offset %r13, -40
; NoVLX-NEXT: .Lcfi1618: ; NoVLX-NEXT: .Lcfi1618:
; NoVLX-NEXT: .cfi_offset %r14, -32 ; NoVLX-NEXT: .cfi_offset %r14, -32
; NoVLX-NEXT: .Lcfi1619: ; NoVLX-NEXT: .Lcfi1619:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: vpcmpltud (%rdi), %zmm0, %k0 ; NoVLX-NEXT: vpcmpltud (%rdi), %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r10d ; NoVLX-NEXT: kmovw %k1, %r10d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 109 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi1625: ; NoVLX-NEXT: .Lcfi1625:
; NoVLX-NEXT: .cfi_offset %r13, -40 ; NoVLX-NEXT: .cfi_offset %r13, -40
; NoVLX-NEXT: .Lcfi1626: ; NoVLX-NEXT: .Lcfi1626:
; NoVLX-NEXT: .cfi_offset %r14, -32 ; NoVLX-NEXT: .cfi_offset %r14, -32
; NoVLX-NEXT: .Lcfi1627: ; NoVLX-NEXT: .Lcfi1627:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vpcmpltud %zmm1, %zmm0, %k0 {%k1} ; NoVLX-NEXT: vpcmpltud %zmm1, %zmm0, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r10d ; NoVLX-NEXT: kmovw %k1, %r10d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 110 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi1633: ; NoVLX-NEXT: .Lcfi1633:
; NoVLX-NEXT: .cfi_offset %r13, -40 ; NoVLX-NEXT: .cfi_offset %r13, -40
; NoVLX-NEXT: .Lcfi1634: ; NoVLX-NEXT: .Lcfi1634:
; NoVLX-NEXT: .cfi_offset %r14, -32 ; NoVLX-NEXT: .cfi_offset %r14, -32
; NoVLX-NEXT: .Lcfi1635: ; NoVLX-NEXT: .Lcfi1635:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vpcmpltud (%rsi), %zmm0, %k0 {%k1} ; NoVLX-NEXT: vpcmpltud (%rsi), %zmm0, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r10d ; NoVLX-NEXT: kmovw %k1, %r10d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 110 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %r12, -48 ; NoVLX-NEXT: .cfi_offset %r12, -48
; NoVLX-NEXT: .Lcfi1641: ; NoVLX-NEXT: .Lcfi1641:
; NoVLX-NEXT: .cfi_offset %r13, -40 ; NoVLX-NEXT: .cfi_offset %r13, -40
; NoVLX-NEXT: .Lcfi1642: ; NoVLX-NEXT: .Lcfi1642:
; NoVLX-NEXT: .cfi_offset %r14, -32 ; NoVLX-NEXT: .cfi_offset %r14, -32
; NoVLX-NEXT: .Lcfi1643: ; NoVLX-NEXT: .Lcfi1643:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: vpcmpltud (%rdi){1to16}, %zmm0, %k0 ; NoVLX-NEXT: vpcmpltud (%rdi){1to16}, %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r10d ; NoVLX-NEXT: kmovw %k1, %r10d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 110 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi1649: ; NoVLX-NEXT: .Lcfi1649:
; NoVLX-NEXT: .cfi_offset %r13, -40 ; NoVLX-NEXT: .cfi_offset %r13, -40
; NoVLX-NEXT: .Lcfi1650: ; NoVLX-NEXT: .Lcfi1650:
; NoVLX-NEXT: .cfi_offset %r14, -32 ; NoVLX-NEXT: .cfi_offset %r14, -32
; NoVLX-NEXT: .Lcfi1651: ; NoVLX-NEXT: .Lcfi1651:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vpcmpltud (%rsi){1to16}, %zmm0, %k0 {%k1} ; NoVLX-NEXT: vpcmpltud (%rsi){1to16}, %zmm0, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r10d ; NoVLX-NEXT: kmovw %k1, %r10d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 111 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %r12, -48 ; NoVLX-NEXT: .cfi_offset %r12, -48
; NoVLX-NEXT: .Lcfi1657: ; NoVLX-NEXT: .Lcfi1657:
; NoVLX-NEXT: .cfi_offset %r13, -40 ; NoVLX-NEXT: .cfi_offset %r13, -40
; NoVLX-NEXT: .Lcfi1658: ; NoVLX-NEXT: .Lcfi1658:
; NoVLX-NEXT: .cfi_offset %r14, -32 ; NoVLX-NEXT: .cfi_offset %r14, -32
; NoVLX-NEXT: .Lcfi1659: ; NoVLX-NEXT: .Lcfi1659:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: vpcmpltud %zmm1, %zmm0, %k0 ; NoVLX-NEXT: vpcmpltud %zmm1, %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 111 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %r12, -48 ; NoVLX-NEXT: .cfi_offset %r12, -48
; NoVLX-NEXT: .Lcfi1665: ; NoVLX-NEXT: .Lcfi1665:
; NoVLX-NEXT: .cfi_offset %r13, -40 ; NoVLX-NEXT: .cfi_offset %r13, -40
; NoVLX-NEXT: .Lcfi1666: ; NoVLX-NEXT: .Lcfi1666:
; NoVLX-NEXT: .cfi_offset %r14, -32 ; NoVLX-NEXT: .cfi_offset %r14, -32
; NoVLX-NEXT: .Lcfi1667: ; NoVLX-NEXT: .Lcfi1667:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: vpcmpltud (%rdi), %zmm0, %k0 ; NoVLX-NEXT: vpcmpltud (%rdi), %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 114 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi1673: ; NoVLX-NEXT: .Lcfi1673:
; NoVLX-NEXT: .cfi_offset %r13, -40 ; NoVLX-NEXT: .cfi_offset %r13, -40
; NoVLX-NEXT: .Lcfi1674: ; NoVLX-NEXT: .Lcfi1674:
; NoVLX-NEXT: .cfi_offset %r14, -32 ; NoVLX-NEXT: .cfi_offset %r14, -32
; NoVLX-NEXT: .Lcfi1675: ; NoVLX-NEXT: .Lcfi1675:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vpcmpltud %zmm1, %zmm0, %k0 {%k1} ; NoVLX-NEXT: vpcmpltud %zmm1, %zmm0, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 115 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi1681: ; NoVLX-NEXT: .Lcfi1681:
; NoVLX-NEXT: .cfi_offset %r13, -40 ; NoVLX-NEXT: .cfi_offset %r13, -40
; NoVLX-NEXT: .Lcfi1682: ; NoVLX-NEXT: .Lcfi1682:
; NoVLX-NEXT: .cfi_offset %r14, -32 ; NoVLX-NEXT: .cfi_offset %r14, -32
; NoVLX-NEXT: .Lcfi1683: ; NoVLX-NEXT: .Lcfi1683:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vpcmpltud (%rsi), %zmm0, %k0 {%k1} ; NoVLX-NEXT: vpcmpltud (%rsi), %zmm0, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 115 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %r12, -48 ; NoVLX-NEXT: .cfi_offset %r12, -48
; NoVLX-NEXT: .Lcfi1689: ; NoVLX-NEXT: .Lcfi1689:
; NoVLX-NEXT: .cfi_offset %r13, -40 ; NoVLX-NEXT: .cfi_offset %r13, -40
; NoVLX-NEXT: .Lcfi1690: ; NoVLX-NEXT: .Lcfi1690:
; NoVLX-NEXT: .cfi_offset %r14, -32 ; NoVLX-NEXT: .cfi_offset %r14, -32
; NoVLX-NEXT: .Lcfi1691: ; NoVLX-NEXT: .Lcfi1691:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: vpcmpltud (%rdi){1to16}, %zmm0, %k0 ; NoVLX-NEXT: vpcmpltud (%rdi){1to16}, %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 115 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi1697: ; NoVLX-NEXT: .Lcfi1697:
; NoVLX-NEXT: .cfi_offset %r13, -40 ; NoVLX-NEXT: .cfi_offset %r13, -40
; NoVLX-NEXT: .Lcfi1698: ; NoVLX-NEXT: .Lcfi1698:
; NoVLX-NEXT: .cfi_offset %r14, -32 ; NoVLX-NEXT: .cfi_offset %r14, -32
; NoVLX-NEXT: .Lcfi1699: ; NoVLX-NEXT: .Lcfi1699:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vpcmpltud (%rsi){1to16}, %zmm0, %k0 {%k1} ; NoVLX-NEXT: vpcmpltud (%rsi){1to16}, %zmm0, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 1,265 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi1720: ; NoVLX-NEXT: .Lcfi1720:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: vmovdqa {{.*#+}} xmm2 = [9223372036854775808,9223372036854775808] ; NoVLX-NEXT: vmovdqa {{.*#+}} xmm2 = [9223372036854775808,9223372036854775808]
; NoVLX-NEXT: vpxor %xmm2, %xmm0, %xmm0 ; NoVLX-NEXT: vpxor %xmm2, %xmm0, %xmm0
; NoVLX-NEXT: vpxor %xmm2, %xmm1, %xmm1 ; NoVLX-NEXT: vpxor %xmm2, %xmm1, %xmm1
; NoVLX-NEXT: vpcmpgtq %xmm0, %xmm1, %xmm0 ; NoVLX-NEXT: vpcmpgtq %xmm0, %xmm1, %xmm0
; NoVLX-NEXT: vpxor %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero ; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
Show All 35 Lines
; NoVLX-NEXT: .Lcfi1723: ; NoVLX-NEXT: .Lcfi1723:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: vmovdqa {{.*#+}} xmm1 = [9223372036854775808,9223372036854775808] ; NoVLX-NEXT: vmovdqa {{.*#+}} xmm1 = [9223372036854775808,9223372036854775808]
; NoVLX-NEXT: vpxor %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vpxor %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vpxor (%rdi), %xmm1, %xmm1 ; NoVLX-NEXT: vpxor (%rdi), %xmm1, %xmm1
; NoVLX-NEXT: vpcmpgtq %xmm0, %xmm1, %xmm0 ; NoVLX-NEXT: vpcmpgtq %xmm0, %xmm1, %xmm0
; NoVLX-NEXT: vpxor %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero ; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
▲ Show 20 Lines • Show All 47 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kshiftlw $14, %k0, %k0 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
; NoVLX-NEXT: kshiftrw $15, %k0, %k0 ; NoVLX-NEXT: kshiftrw $15, %k0, %k0
; NoVLX-NEXT: kmovw %k0, %eax ; NoVLX-NEXT: kmovw %k0, %eax
; NoVLX-NEXT: kmovw %k1, %ecx ; NoVLX-NEXT: kmovw %k1, %ecx
; NoVLX-NEXT: vmovd %ecx, %xmm1 ; NoVLX-NEXT: vmovd %ecx, %xmm1
; NoVLX-NEXT: vpinsrb $8, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT: vpand %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vpxor %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero ; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
▲ Show 20 Lines • Show All 49 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kshiftlw $14, %k0, %k0 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
; NoVLX-NEXT: kshiftrw $15, %k0, %k0 ; NoVLX-NEXT: kshiftrw $15, %k0, %k0
; NoVLX-NEXT: kmovw %k0, %eax ; NoVLX-NEXT: kmovw %k0, %eax
; NoVLX-NEXT: kmovw %k1, %ecx ; NoVLX-NEXT: kmovw %k1, %ecx
; NoVLX-NEXT: vmovd %ecx, %xmm1 ; NoVLX-NEXT: vmovd %ecx, %xmm1
; NoVLX-NEXT: vpinsrb $8, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT: vpand %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vpxor %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero ; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
▲ Show 20 Lines • Show All 41 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: vpbroadcastq (%rdi), %xmm1 ; NoVLX-NEXT: vpbroadcastq (%rdi), %xmm1
; NoVLX-NEXT: vmovdqa {{.*#+}} xmm2 = [9223372036854775808,9223372036854775808] ; NoVLX-NEXT: vmovdqa {{.*#+}} xmm2 = [9223372036854775808,9223372036854775808]
; NoVLX-NEXT: vpxor %xmm2, %xmm0, %xmm0 ; NoVLX-NEXT: vpxor %xmm2, %xmm0, %xmm0
; NoVLX-NEXT: vpxor %xmm2, %xmm1, %xmm1 ; NoVLX-NEXT: vpxor %xmm2, %xmm1, %xmm1
; NoVLX-NEXT: vpcmpgtq %xmm0, %xmm1, %xmm0 ; NoVLX-NEXT: vpcmpgtq %xmm0, %xmm1, %xmm0
; NoVLX-NEXT: vpxor %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero ; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
▲ Show 20 Lines • Show All 49 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kshiftlw $14, %k0, %k0 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
; NoVLX-NEXT: kshiftrw $15, %k0, %k0 ; NoVLX-NEXT: kshiftrw $15, %k0, %k0
; NoVLX-NEXT: kmovw %k0, %eax ; NoVLX-NEXT: kmovw %k0, %eax
; NoVLX-NEXT: kmovw %k1, %ecx ; NoVLX-NEXT: kmovw %k1, %ecx
; NoVLX-NEXT: vmovd %ecx, %xmm1 ; NoVLX-NEXT: vmovd %ecx, %xmm1
; NoVLX-NEXT: vpinsrb $8, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT: vpand %xmm0, %xmm1, %xmm0 ; NoVLX-NEXT: vpand %xmm0, %xmm1, %xmm0
; NoVLX-NEXT: vpxor %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero ; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
▲ Show 20 Lines • Show All 1,273 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: vpbroadcastq {{.*#+}} ymm2 = [9223372036854775808,9223372036854775808,9223372036854775808,9223372036854775808] ; NoVLX-NEXT: vpbroadcastq {{.*#+}} ymm2 = [9223372036854775808,9223372036854775808,9223372036854775808,9223372036854775808]
; NoVLX-NEXT: vpxor %ymm2, %ymm0, %ymm0 ; NoVLX-NEXT: vpxor %ymm2, %ymm0, %ymm0
; NoVLX-NEXT: vpxor %ymm2, %ymm1, %ymm1 ; NoVLX-NEXT: vpxor %ymm2, %ymm1, %ymm1
; NoVLX-NEXT: vpcmpgtq %ymm0, %ymm1, %ymm0 ; NoVLX-NEXT: vpcmpgtq %ymm0, %ymm1, %ymm0
; NoVLX-NEXT: vpmovqd %zmm0, %ymm0 ; NoVLX-NEXT: vpmovqd %zmm0, %ymm0
; NoVLX-NEXT: vpxor %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero ; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
Show All 37 Lines
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: vpbroadcastq {{.*#+}} ymm1 = [9223372036854775808,9223372036854775808,9223372036854775808,9223372036854775808] ; NoVLX-NEXT: vpbroadcastq {{.*#+}} ymm1 = [9223372036854775808,9223372036854775808,9223372036854775808,9223372036854775808]
; NoVLX-NEXT: vpxor %ymm1, %ymm0, %ymm0 ; NoVLX-NEXT: vpxor %ymm1, %ymm0, %ymm0
; NoVLX-NEXT: vpxor (%rdi), %ymm1, %ymm1 ; NoVLX-NEXT: vpxor (%rdi), %ymm1, %ymm1
; NoVLX-NEXT: vpcmpgtq %ymm0, %ymm1, %ymm0 ; NoVLX-NEXT: vpcmpgtq %ymm0, %ymm1, %ymm0
; NoVLX-NEXT: vpmovqd %zmm0, %ymm0 ; NoVLX-NEXT: vpmovqd %zmm0, %ymm0
; NoVLX-NEXT: vpxor %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero ; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
▲ Show 20 Lines • Show All 57 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: kmovw %k3, %ecx ; NoVLX-NEXT: kmovw %k3, %ecx
; NoVLX-NEXT: vmovd %ecx, %xmm1 ; NoVLX-NEXT: vmovd %ecx, %xmm1
; NoVLX-NEXT: vpinsrb $4, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrb $4, %eax, %xmm1, %xmm1
; NoVLX-NEXT: kmovw %k2, %eax ; NoVLX-NEXT: kmovw %k2, %eax
; NoVLX-NEXT: vpinsrb $8, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $12, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT: vpand %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vpxor %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero ; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
▲ Show 20 Lines • Show All 59 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: kmovw %k3, %ecx ; NoVLX-NEXT: kmovw %k3, %ecx
; NoVLX-NEXT: vmovd %ecx, %xmm1 ; NoVLX-NEXT: vmovd %ecx, %xmm1
; NoVLX-NEXT: vpinsrb $4, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrb $4, %eax, %xmm1, %xmm1
; NoVLX-NEXT: kmovw %k2, %eax ; NoVLX-NEXT: kmovw %k2, %eax
; NoVLX-NEXT: vpinsrb $8, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $12, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT: vpand %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vpxor %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero ; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
▲ Show 20 Lines • Show All 43 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: vpbroadcastq (%rdi), %ymm1 ; NoVLX-NEXT: vpbroadcastq (%rdi), %ymm1
; NoVLX-NEXT: vpbroadcastq {{.*#+}} ymm2 = [9223372036854775808,9223372036854775808,9223372036854775808,9223372036854775808] ; NoVLX-NEXT: vpbroadcastq {{.*#+}} ymm2 = [9223372036854775808,9223372036854775808,9223372036854775808,9223372036854775808]
; NoVLX-NEXT: vpxor %ymm2, %ymm0, %ymm0 ; NoVLX-NEXT: vpxor %ymm2, %ymm0, %ymm0
; NoVLX-NEXT: vpxor %ymm2, %ymm1, %ymm1 ; NoVLX-NEXT: vpxor %ymm2, %ymm1, %ymm1
; NoVLX-NEXT: vpcmpgtq %ymm0, %ymm1, %ymm0 ; NoVLX-NEXT: vpcmpgtq %ymm0, %ymm1, %ymm0
; NoVLX-NEXT: vpmovqd %zmm0, %ymm0 ; NoVLX-NEXT: vpmovqd %zmm0, %ymm0
; NoVLX-NEXT: vpxor %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero ; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
▲ Show 20 Lines • Show All 59 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: kmovw %k3, %ecx ; NoVLX-NEXT: kmovw %k3, %ecx
; NoVLX-NEXT: vmovd %ecx, %xmm1 ; NoVLX-NEXT: vmovd %ecx, %xmm1
; NoVLX-NEXT: vpinsrb $4, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrb $4, %eax, %xmm1, %xmm1
; NoVLX-NEXT: kmovw %k2, %eax ; NoVLX-NEXT: kmovw %k2, %eax
; NoVLX-NEXT: vpinsrb $8, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrb $8, %eax, %xmm1, %xmm1
; NoVLX-NEXT: kmovw %k1, %eax ; NoVLX-NEXT: kmovw %k1, %eax
; NoVLX-NEXT: vpinsrb $12, %eax, %xmm1, %xmm1 ; NoVLX-NEXT: vpinsrb $12, %eax, %xmm1, %xmm1
; NoVLX-NEXT: vpand %xmm0, %xmm1, %xmm0 ; NoVLX-NEXT: vpand %xmm0, %xmm1, %xmm0
; NoVLX-NEXT: vpxor %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero ; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
▲ Show 20 Lines • Show All 203 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi1773: ; NoVLX-NEXT: .Lcfi1773:
; NoVLX-NEXT: .cfi_offset %rbp, -16 ; NoVLX-NEXT: .cfi_offset %rbp, -16
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi1774: ; NoVLX-NEXT: .Lcfi1774:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $32, %rsp ; NoVLX-NEXT: subq $32, %rsp
; NoVLX-NEXT: vpcmpltuq %zmm1, %zmm0, %k0 ; NoVLX-NEXT: vpcmpltuq %zmm1, %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r9d ; NoVLX-NEXT: kmovw %k1, %r9d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 56 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi1776: ; NoVLX-NEXT: .Lcfi1776:
; NoVLX-NEXT: .cfi_offset %rbp, -16 ; NoVLX-NEXT: .cfi_offset %rbp, -16
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi1777: ; NoVLX-NEXT: .Lcfi1777:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $32, %rsp ; NoVLX-NEXT: subq $32, %rsp
; NoVLX-NEXT: vpcmpltuq (%rdi), %zmm0, %k0 ; NoVLX-NEXT: vpcmpltuq (%rdi), %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r9d ; NoVLX-NEXT: kmovw %k1, %r9d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 59 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %rbp, -16 ; NoVLX-NEXT: .cfi_offset %rbp, -16
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi1780: ; NoVLX-NEXT: .Lcfi1780:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $32, %rsp ; NoVLX-NEXT: subq $32, %rsp
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vpcmpltuq %zmm1, %zmm0, %k0 {%k1} ; NoVLX-NEXT: vpcmpltuq %zmm1, %zmm0, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r9d ; NoVLX-NEXT: kmovw %k1, %r9d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 60 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %rbp, -16 ; NoVLX-NEXT: .cfi_offset %rbp, -16
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi1783: ; NoVLX-NEXT: .Lcfi1783:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $32, %rsp ; NoVLX-NEXT: subq $32, %rsp
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vpcmpltuq (%rsi), %zmm0, %k0 {%k1} ; NoVLX-NEXT: vpcmpltuq (%rsi), %zmm0, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r9d ; NoVLX-NEXT: kmovw %k1, %r9d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 60 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi1785: ; NoVLX-NEXT: .Lcfi1785:
; NoVLX-NEXT: .cfi_offset %rbp, -16 ; NoVLX-NEXT: .cfi_offset %rbp, -16
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi1786: ; NoVLX-NEXT: .Lcfi1786:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $32, %rsp ; NoVLX-NEXT: subq $32, %rsp
; NoVLX-NEXT: vpcmpltuq (%rdi){1to8}, %zmm0, %k0 ; NoVLX-NEXT: vpcmpltuq (%rdi){1to8}, %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r9d ; NoVLX-NEXT: kmovw %k1, %r9d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 60 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %rbp, -16 ; NoVLX-NEXT: .cfi_offset %rbp, -16
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi1789: ; NoVLX-NEXT: .Lcfi1789:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $32, %rsp ; NoVLX-NEXT: subq $32, %rsp
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vpcmpltuq (%rsi){1to8}, %zmm0, %k0 {%k1} ; NoVLX-NEXT: vpcmpltuq (%rsi){1to8}, %zmm0, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r9d ; NoVLX-NEXT: kmovw %k1, %r9d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 61 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi1791: ; NoVLX-NEXT: .Lcfi1791:
; NoVLX-NEXT: .cfi_offset %rbp, -16 ; NoVLX-NEXT: .cfi_offset %rbp, -16
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi1792: ; NoVLX-NEXT: .Lcfi1792:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: vpcmpltuq %zmm1, %zmm0, %k0 ; NoVLX-NEXT: vpcmpltuq %zmm1, %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 61 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi1794: ; NoVLX-NEXT: .Lcfi1794:
; NoVLX-NEXT: .cfi_offset %rbp, -16 ; NoVLX-NEXT: .cfi_offset %rbp, -16
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi1795: ; NoVLX-NEXT: .Lcfi1795:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: vpcmpltuq (%rdi), %zmm0, %k0 ; NoVLX-NEXT: vpcmpltuq (%rdi), %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 64 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %rbp, -16 ; NoVLX-NEXT: .cfi_offset %rbp, -16
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi1798: ; NoVLX-NEXT: .Lcfi1798:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vpcmpltuq %zmm1, %zmm0, %k0 {%k1} ; NoVLX-NEXT: vpcmpltuq %zmm1, %zmm0, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 65 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %rbp, -16 ; NoVLX-NEXT: .cfi_offset %rbp, -16
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi1801: ; NoVLX-NEXT: .Lcfi1801:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vpcmpltuq (%rsi), %zmm0, %k0 {%k1} ; NoVLX-NEXT: vpcmpltuq (%rsi), %zmm0, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 65 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi1803: ; NoVLX-NEXT: .Lcfi1803:
; NoVLX-NEXT: .cfi_offset %rbp, -16 ; NoVLX-NEXT: .cfi_offset %rbp, -16
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi1804: ; NoVLX-NEXT: .Lcfi1804:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: vpcmpltuq (%rdi){1to8}, %zmm0, %k0 ; NoVLX-NEXT: vpcmpltuq (%rdi){1to8}, %zmm0, %k0
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 65 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %rbp, -16 ; NoVLX-NEXT: .cfi_offset %rbp, -16
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi1807: ; NoVLX-NEXT: .Lcfi1807:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vpcmpltuq (%rsi){1to8}, %zmm0, %k0 {%k1} ; NoVLX-NEXT: vpcmpltuq (%rsi){1to8}, %zmm0, %k0 {%k1}
; NoVLX-NEXT: vpxor %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 1,098 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi1827: ; NoVLX-NEXT: .Lcfi1827:
; NoVLX-NEXT: .cfi_offset %rbp, -16 ; NoVLX-NEXT: .cfi_offset %rbp, -16
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi1828: ; NoVLX-NEXT: .Lcfi1828:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: vcmpeqps %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vcmpeqps %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vxorps %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero ; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
Show All 32 Lines
; NoVLX-NEXT: .Lcfi1830: ; NoVLX-NEXT: .Lcfi1830:
; NoVLX-NEXT: .cfi_offset %rbp, -16 ; NoVLX-NEXT: .cfi_offset %rbp, -16
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi1831: ; NoVLX-NEXT: .Lcfi1831:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: vcmpeqps (%rdi), %xmm0, %xmm0 ; NoVLX-NEXT: vcmpeqps (%rdi), %xmm0, %xmm0
; NoVLX-NEXT: vxorps %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero ; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
Show All 34 Lines
; NoVLX-NEXT: .cfi_offset %rbp, -16 ; NoVLX-NEXT: .cfi_offset %rbp, -16
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi1834: ; NoVLX-NEXT: .Lcfi1834:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: vbroadcastss (%rdi), %xmm1 ; NoVLX-NEXT: vbroadcastss (%rdi), %xmm1
; NoVLX-NEXT: vcmpeqps %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vcmpeqps %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vxorps %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero ; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
Show All 40 Lines
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $96, %rsp ; NoVLX-NEXT: subq $96, %rsp
; NoVLX-NEXT: movb %dil, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: movb %dil, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: movzbl {{[0-9]+}}(%rsp), %eax ; NoVLX-NEXT: movzbl {{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT: kmovw %eax, %k1 ; NoVLX-NEXT: kmovw %eax, %k1
; NoVLX-NEXT: vpternlogq $255, %zmm2, %zmm2, %zmm2 {%k1} {z} ; NoVLX-NEXT: vpternlogq $255, %zmm2, %zmm2, %zmm2 {%k1} {z}
; NoVLX-NEXT: vpmovqd %zmm2, %ymm2 ; NoVLX-NEXT: vpmovqd %zmm2, %ymm2
; NoVLX-NEXT: vxorps %xmm3, %xmm3, %xmm3
; NoVLX-NEXT: vcmpeqps %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vcmpeqps %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vandps %xmm2, %xmm0, %xmm0 ; NoVLX-NEXT: vandps %xmm2, %xmm0, %xmm0
; NoVLX-NEXT: vptestmd %zmm3, %zmm3, %k0 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero ; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
Show All 40 Lines
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $96, %rsp ; NoVLX-NEXT: subq $96, %rsp
; NoVLX-NEXT: movb %dil, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: movb %dil, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: movzbl {{[0-9]+}}(%rsp), %eax ; NoVLX-NEXT: movzbl {{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT: kmovw %eax, %k1 ; NoVLX-NEXT: kmovw %eax, %k1
; NoVLX-NEXT: vpternlogq $255, %zmm1, %zmm1, %zmm1 {%k1} {z} ; NoVLX-NEXT: vpternlogq $255, %zmm1, %zmm1, %zmm1 {%k1} {z}
; NoVLX-NEXT: vpmovqd %zmm1, %ymm1 ; NoVLX-NEXT: vpmovqd %zmm1, %ymm1
; NoVLX-NEXT: vxorps %xmm2, %xmm2, %xmm2
; NoVLX-NEXT: vcmpeqps (%rsi), %xmm0, %xmm0 ; NoVLX-NEXT: vcmpeqps (%rsi), %xmm0, %xmm0
; NoVLX-NEXT: vandps %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vandps %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vptestmd %zmm2, %zmm2, %k0 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero ; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
▲ Show 20 Lines • Show All 41 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $96, %rsp ; NoVLX-NEXT: subq $96, %rsp
; NoVLX-NEXT: movb %dil, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: movb %dil, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: movzbl {{[0-9]+}}(%rsp), %eax ; NoVLX-NEXT: movzbl {{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT: kmovw %eax, %k1 ; NoVLX-NEXT: kmovw %eax, %k1
; NoVLX-NEXT: vpternlogq $255, %zmm1, %zmm1, %zmm1 {%k1} {z} ; NoVLX-NEXT: vpternlogq $255, %zmm1, %zmm1, %zmm1 {%k1} {z}
; NoVLX-NEXT: vpmovqd %zmm1, %ymm1 ; NoVLX-NEXT: vpmovqd %zmm1, %ymm1
; NoVLX-NEXT: vxorps %xmm2, %xmm2, %xmm2 ; NoVLX-NEXT: vbroadcastss (%rsi), %xmm2
; NoVLX-NEXT: vbroadcastss (%rsi), %xmm3 ; NoVLX-NEXT: vcmpeqps %xmm2, %xmm0, %xmm0
; NoVLX-NEXT: vcmpeqps %xmm3, %xmm0, %xmm0
; NoVLX-NEXT: vandps %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vandps %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vptestmd %zmm2, %zmm2, %k0 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero ; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
▲ Show 20 Lines • Show All 229 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi1846: ; NoVLX-NEXT: .Lcfi1846:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $32, %rsp ; NoVLX-NEXT: subq $32, %rsp
; NoVLX-NEXT: # kill: %YMM1<def> %YMM1<kill> %ZMM1<def> ; NoVLX-NEXT: # kill: %YMM1<def> %YMM1<kill> %ZMM1<def>
; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def> ; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def>
; NoVLX-NEXT: vcmpeqps %zmm1, %zmm0, %k0 ; NoVLX-NEXT: vcmpeqps %zmm1, %zmm0, %k0
; NoVLX-NEXT: vxorps %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r9d ; NoVLX-NEXT: kmovw %k1, %r9d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 58 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi1849: ; NoVLX-NEXT: .Lcfi1849:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $32, %rsp ; NoVLX-NEXT: subq $32, %rsp
; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def> ; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def>
; NoVLX-NEXT: vmovaps (%rdi), %ymm1 ; NoVLX-NEXT: vmovaps (%rdi), %ymm1
; NoVLX-NEXT: vcmpeqps %zmm1, %zmm0, %k0 ; NoVLX-NEXT: vcmpeqps %zmm1, %zmm0, %k0
; NoVLX-NEXT: vxorps %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r9d ; NoVLX-NEXT: kmovw %k1, %r9d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 59 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi1852: ; NoVLX-NEXT: .Lcfi1852:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $32, %rsp ; NoVLX-NEXT: subq $32, %rsp
; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def> ; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def>
; NoVLX-NEXT: vbroadcastss (%rdi), %ymm1 ; NoVLX-NEXT: vbroadcastss (%rdi), %ymm1
; NoVLX-NEXT: vcmpeqps %zmm1, %zmm0, %k0 ; NoVLX-NEXT: vcmpeqps %zmm1, %zmm0, %k0
; NoVLX-NEXT: vxorps %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r9d ; NoVLX-NEXT: kmovw %k1, %r9d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 63 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $32, %rsp ; NoVLX-NEXT: subq $32, %rsp
; NoVLX-NEXT: # kill: %YMM1<def> %YMM1<kill> %ZMM1<def> ; NoVLX-NEXT: # kill: %YMM1<def> %YMM1<kill> %ZMM1<def>
; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def> ; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def>
; NoVLX-NEXT: vcmpeqps %zmm1, %zmm0, %k0 ; NoVLX-NEXT: vcmpeqps %zmm1, %zmm0, %k0
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: kandw %k1, %k0, %k0 ; NoVLX-NEXT: kandw %k1, %k0, %k0
; NoVLX-NEXT: vxorps %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r9d ; NoVLX-NEXT: kmovw %k1, %r9d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 63 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $32, %rsp ; NoVLX-NEXT: subq $32, %rsp
; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def> ; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def>
; NoVLX-NEXT: vmovaps (%rsi), %ymm1 ; NoVLX-NEXT: vmovaps (%rsi), %ymm1
; NoVLX-NEXT: vcmpeqps %zmm1, %zmm0, %k0 ; NoVLX-NEXT: vcmpeqps %zmm1, %zmm0, %k0
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: kandw %k1, %k0, %k0 ; NoVLX-NEXT: kandw %k1, %k0, %k0
; NoVLX-NEXT: vxorps %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r9d ; NoVLX-NEXT: kmovw %k1, %r9d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 64 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $32, %rsp ; NoVLX-NEXT: subq $32, %rsp
; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def> ; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def>
; NoVLX-NEXT: vbroadcastss (%rsi), %ymm1 ; NoVLX-NEXT: vbroadcastss (%rsi), %ymm1
; NoVLX-NEXT: vcmpeqps %zmm1, %zmm0, %k0 ; NoVLX-NEXT: vcmpeqps %zmm1, %zmm0, %k0
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: kandw %k1, %k0, %k0 ; NoVLX-NEXT: kandw %k1, %k0, %k0
; NoVLX-NEXT: vxorps %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r9d ; NoVLX-NEXT: kmovw %k1, %r9d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 64 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi1864: ; NoVLX-NEXT: .Lcfi1864:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: # kill: %YMM1<def> %YMM1<kill> %ZMM1<def> ; NoVLX-NEXT: # kill: %YMM1<def> %YMM1<kill> %ZMM1<def>
; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def> ; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def>
; NoVLX-NEXT: vcmpeqps %zmm1, %zmm0, %k0 ; NoVLX-NEXT: vcmpeqps %zmm1, %zmm0, %k0
; NoVLX-NEXT: vxorps %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 63 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi1867: ; NoVLX-NEXT: .Lcfi1867:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def> ; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def>
; NoVLX-NEXT: vmovaps (%rdi), %ymm1 ; NoVLX-NEXT: vmovaps (%rdi), %ymm1
; NoVLX-NEXT: vcmpeqps %zmm1, %zmm0, %k0 ; NoVLX-NEXT: vcmpeqps %zmm1, %zmm0, %k0
; NoVLX-NEXT: vxorps %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 64 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi1870: ; NoVLX-NEXT: .Lcfi1870:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def> ; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def>
; NoVLX-NEXT: vbroadcastss (%rdi), %ymm1 ; NoVLX-NEXT: vbroadcastss (%rdi), %ymm1
; NoVLX-NEXT: vcmpeqps %zmm1, %zmm0, %k0 ; NoVLX-NEXT: vcmpeqps %zmm1, %zmm0, %k0
; NoVLX-NEXT: vxorps %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 68 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: # kill: %YMM1<def> %YMM1<kill> %ZMM1<def> ; NoVLX-NEXT: # kill: %YMM1<def> %YMM1<kill> %ZMM1<def>
; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def> ; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def>
; NoVLX-NEXT: vcmpeqps %zmm1, %zmm0, %k0 ; NoVLX-NEXT: vcmpeqps %zmm1, %zmm0, %k0
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: kandw %k1, %k0, %k0 ; NoVLX-NEXT: kandw %k1, %k0, %k0
; NoVLX-NEXT: vxorps %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 68 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def> ; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def>
; NoVLX-NEXT: vmovaps (%rsi), %ymm1 ; NoVLX-NEXT: vmovaps (%rsi), %ymm1
; NoVLX-NEXT: vcmpeqps %zmm1, %zmm0, %k0 ; NoVLX-NEXT: vcmpeqps %zmm1, %zmm0, %k0
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: kandw %k1, %k0, %k0 ; NoVLX-NEXT: kandw %k1, %k0, %k0
; NoVLX-NEXT: vxorps %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 69 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def> ; NoVLX-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def>
; NoVLX-NEXT: vbroadcastss (%rsi), %ymm1 ; NoVLX-NEXT: vbroadcastss (%rsi), %ymm1
; NoVLX-NEXT: vcmpeqps %zmm1, %zmm0, %k0 ; NoVLX-NEXT: vcmpeqps %zmm1, %zmm0, %k0
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: kandw %k1, %k0, %k0 ; NoVLX-NEXT: kandw %k1, %k0, %k0
; NoVLX-NEXT: vxorps %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 82 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %r12, -48 ; NoVLX-NEXT: .cfi_offset %r12, -48
; NoVLX-NEXT: .Lcfi1885: ; NoVLX-NEXT: .Lcfi1885:
; NoVLX-NEXT: .cfi_offset %r13, -40 ; NoVLX-NEXT: .cfi_offset %r13, -40
; NoVLX-NEXT: .Lcfi1886: ; NoVLX-NEXT: .Lcfi1886:
; NoVLX-NEXT: .cfi_offset %r14, -32 ; NoVLX-NEXT: .cfi_offset %r14, -32
; NoVLX-NEXT: .Lcfi1887: ; NoVLX-NEXT: .Lcfi1887:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: vcmpeqps %zmm1, %zmm0, %k0 ; NoVLX-NEXT: vcmpeqps %zmm1, %zmm0, %k0
; NoVLX-NEXT: vxorps %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r10d ; NoVLX-NEXT: kmovw %k1, %r10d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 106 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %r12, -48 ; NoVLX-NEXT: .cfi_offset %r12, -48
; NoVLX-NEXT: .Lcfi1893: ; NoVLX-NEXT: .Lcfi1893:
; NoVLX-NEXT: .cfi_offset %r13, -40 ; NoVLX-NEXT: .cfi_offset %r13, -40
; NoVLX-NEXT: .Lcfi1894: ; NoVLX-NEXT: .Lcfi1894:
; NoVLX-NEXT: .cfi_offset %r14, -32 ; NoVLX-NEXT: .cfi_offset %r14, -32
; NoVLX-NEXT: .Lcfi1895: ; NoVLX-NEXT: .Lcfi1895:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: vcmpeqps (%rdi), %zmm0, %k0 ; NoVLX-NEXT: vcmpeqps (%rdi), %zmm0, %k0
; NoVLX-NEXT: vxorps %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r10d ; NoVLX-NEXT: kmovw %k1, %r10d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 107 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %r12, -48 ; NoVLX-NEXT: .cfi_offset %r12, -48
; NoVLX-NEXT: .Lcfi1901: ; NoVLX-NEXT: .Lcfi1901:
; NoVLX-NEXT: .cfi_offset %r13, -40 ; NoVLX-NEXT: .cfi_offset %r13, -40
; NoVLX-NEXT: .Lcfi1902: ; NoVLX-NEXT: .Lcfi1902:
; NoVLX-NEXT: .cfi_offset %r14, -32 ; NoVLX-NEXT: .cfi_offset %r14, -32
; NoVLX-NEXT: .Lcfi1903: ; NoVLX-NEXT: .Lcfi1903:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: vcmpeqps (%rdi){1to16}, %zmm0, %k0 ; NoVLX-NEXT: vcmpeqps (%rdi){1to16}, %zmm0, %k0
; NoVLX-NEXT: vxorps %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r10d ; NoVLX-NEXT: kmovw %k1, %r10d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 110 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi1909: ; NoVLX-NEXT: .Lcfi1909:
; NoVLX-NEXT: .cfi_offset %r13, -40 ; NoVLX-NEXT: .cfi_offset %r13, -40
; NoVLX-NEXT: .Lcfi1910: ; NoVLX-NEXT: .Lcfi1910:
; NoVLX-NEXT: .cfi_offset %r14, -32 ; NoVLX-NEXT: .cfi_offset %r14, -32
; NoVLX-NEXT: .Lcfi1911: ; NoVLX-NEXT: .Lcfi1911:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vcmpeqps %zmm1, %zmm0, %k0 {%k1} ; NoVLX-NEXT: vcmpeqps %zmm1, %zmm0, %k0 {%k1}
; NoVLX-NEXT: vxorps %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r10d ; NoVLX-NEXT: kmovw %k1, %r10d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 110 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi1917: ; NoVLX-NEXT: .Lcfi1917:
; NoVLX-NEXT: .cfi_offset %r13, -40 ; NoVLX-NEXT: .cfi_offset %r13, -40
; NoVLX-NEXT: .Lcfi1918: ; NoVLX-NEXT: .Lcfi1918:
; NoVLX-NEXT: .cfi_offset %r14, -32 ; NoVLX-NEXT: .cfi_offset %r14, -32
; NoVLX-NEXT: .Lcfi1919: ; NoVLX-NEXT: .Lcfi1919:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vcmpeqps (%rsi), %zmm0, %k0 {%k1} ; NoVLX-NEXT: vcmpeqps (%rsi), %zmm0, %k0 {%k1}
; NoVLX-NEXT: vxorps %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r10d ; NoVLX-NEXT: kmovw %k1, %r10d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 111 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi1925: ; NoVLX-NEXT: .Lcfi1925:
; NoVLX-NEXT: .cfi_offset %r13, -40 ; NoVLX-NEXT: .cfi_offset %r13, -40
; NoVLX-NEXT: .Lcfi1926: ; NoVLX-NEXT: .Lcfi1926:
; NoVLX-NEXT: .cfi_offset %r14, -32 ; NoVLX-NEXT: .cfi_offset %r14, -32
; NoVLX-NEXT: .Lcfi1927: ; NoVLX-NEXT: .Lcfi1927:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vcmpeqps (%rsi){1to16}, %zmm0, %k0 {%k1} ; NoVLX-NEXT: vcmpeqps (%rsi){1to16}, %zmm0, %k0 {%k1}
; NoVLX-NEXT: vxorps %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r10d ; NoVLX-NEXT: kmovw %k1, %r10d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 158 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %r12, -48 ; NoVLX-NEXT: .cfi_offset %r12, -48
; NoVLX-NEXT: .Lcfi1933: ; NoVLX-NEXT: .Lcfi1933:
; NoVLX-NEXT: .cfi_offset %r13, -40 ; NoVLX-NEXT: .cfi_offset %r13, -40
; NoVLX-NEXT: .Lcfi1934: ; NoVLX-NEXT: .Lcfi1934:
; NoVLX-NEXT: .cfi_offset %r14, -32 ; NoVLX-NEXT: .cfi_offset %r14, -32
; NoVLX-NEXT: .Lcfi1935: ; NoVLX-NEXT: .Lcfi1935:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: vcmpeqps %zmm1, %zmm0, %k0 ; NoVLX-NEXT: vcmpeqps %zmm1, %zmm0, %k0
; NoVLX-NEXT: vxorps %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 111 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %r12, -48 ; NoVLX-NEXT: .cfi_offset %r12, -48
; NoVLX-NEXT: .Lcfi1941: ; NoVLX-NEXT: .Lcfi1941:
; NoVLX-NEXT: .cfi_offset %r13, -40 ; NoVLX-NEXT: .cfi_offset %r13, -40
; NoVLX-NEXT: .Lcfi1942: ; NoVLX-NEXT: .Lcfi1942:
; NoVLX-NEXT: .cfi_offset %r14, -32 ; NoVLX-NEXT: .cfi_offset %r14, -32
; NoVLX-NEXT: .Lcfi1943: ; NoVLX-NEXT: .Lcfi1943:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: vcmpeqps (%rdi), %zmm0, %k0 ; NoVLX-NEXT: vcmpeqps (%rdi), %zmm0, %k0
; NoVLX-NEXT: vxorps %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 112 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %r12, -48 ; NoVLX-NEXT: .cfi_offset %r12, -48
; NoVLX-NEXT: .Lcfi1949: ; NoVLX-NEXT: .Lcfi1949:
; NoVLX-NEXT: .cfi_offset %r13, -40 ; NoVLX-NEXT: .cfi_offset %r13, -40
; NoVLX-NEXT: .Lcfi1950: ; NoVLX-NEXT: .Lcfi1950:
; NoVLX-NEXT: .cfi_offset %r14, -32 ; NoVLX-NEXT: .cfi_offset %r14, -32
; NoVLX-NEXT: .Lcfi1951: ; NoVLX-NEXT: .Lcfi1951:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: vcmpeqps (%rdi){1to16}, %zmm0, %k0 ; NoVLX-NEXT: vcmpeqps (%rdi){1to16}, %zmm0, %k0
; NoVLX-NEXT: vxorps %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 115 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi1957: ; NoVLX-NEXT: .Lcfi1957:
; NoVLX-NEXT: .cfi_offset %r13, -40 ; NoVLX-NEXT: .cfi_offset %r13, -40
; NoVLX-NEXT: .Lcfi1958: ; NoVLX-NEXT: .Lcfi1958:
; NoVLX-NEXT: .cfi_offset %r14, -32 ; NoVLX-NEXT: .cfi_offset %r14, -32
; NoVLX-NEXT: .Lcfi1959: ; NoVLX-NEXT: .Lcfi1959:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vcmpeqps %zmm1, %zmm0, %k0 {%k1} ; NoVLX-NEXT: vcmpeqps %zmm1, %zmm0, %k0 {%k1}
; NoVLX-NEXT: vxorps %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 115 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi1965: ; NoVLX-NEXT: .Lcfi1965:
; NoVLX-NEXT: .cfi_offset %r13, -40 ; NoVLX-NEXT: .cfi_offset %r13, -40
; NoVLX-NEXT: .Lcfi1966: ; NoVLX-NEXT: .Lcfi1966:
; NoVLX-NEXT: .cfi_offset %r14, -32 ; NoVLX-NEXT: .cfi_offset %r14, -32
; NoVLX-NEXT: .Lcfi1967: ; NoVLX-NEXT: .Lcfi1967:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vcmpeqps (%rsi), %zmm0, %k0 {%k1} ; NoVLX-NEXT: vcmpeqps (%rsi), %zmm0, %k0 {%k1}
; NoVLX-NEXT: vxorps %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 116 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi1973: ; NoVLX-NEXT: .Lcfi1973:
; NoVLX-NEXT: .cfi_offset %r13, -40 ; NoVLX-NEXT: .cfi_offset %r13, -40
; NoVLX-NEXT: .Lcfi1974: ; NoVLX-NEXT: .Lcfi1974:
; NoVLX-NEXT: .cfi_offset %r14, -32 ; NoVLX-NEXT: .cfi_offset %r14, -32
; NoVLX-NEXT: .Lcfi1975: ; NoVLX-NEXT: .Lcfi1975:
; NoVLX-NEXT: .cfi_offset %r15, -24 ; NoVLX-NEXT: .cfi_offset %r15, -24
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vcmpeqps (%rsi){1to16}, %zmm0, %k0 {%k1} ; NoVLX-NEXT: vcmpeqps (%rsi){1to16}, %zmm0, %k0 {%k1}
; NoVLX-NEXT: vxorps %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 1,182 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi1995: ; NoVLX-NEXT: .Lcfi1995:
; NoVLX-NEXT: .cfi_offset %rbp, -16 ; NoVLX-NEXT: .cfi_offset %rbp, -16
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi1996: ; NoVLX-NEXT: .Lcfi1996:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: vcmpeqpd %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vcmpeqpd %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vxorpd %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero ; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
Show All 32 Lines
; NoVLX-NEXT: .Lcfi1998: ; NoVLX-NEXT: .Lcfi1998:
; NoVLX-NEXT: .cfi_offset %rbp, -16 ; NoVLX-NEXT: .cfi_offset %rbp, -16
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi1999: ; NoVLX-NEXT: .Lcfi1999:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: vcmpeqpd (%rdi), %xmm0, %xmm0 ; NoVLX-NEXT: vcmpeqpd (%rdi), %xmm0, %xmm0
; NoVLX-NEXT: vxorps %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero ; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
Show All 34 Lines
; NoVLX-NEXT: .cfi_offset %rbp, -16 ; NoVLX-NEXT: .cfi_offset %rbp, -16
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi2002: ; NoVLX-NEXT: .Lcfi2002:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: vmovddup {{.*#+}} xmm1 = mem[0,0] ; NoVLX-NEXT: vmovddup {{.*#+}} xmm1 = mem[0,0]
; NoVLX-NEXT: vcmpeqpd %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vcmpeqpd %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vxorpd %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero ; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
Show All 39 Lines
; NoVLX-NEXT: .Lcfi2005: ; NoVLX-NEXT: .Lcfi2005:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $96, %rsp ; NoVLX-NEXT: subq $96, %rsp
; NoVLX-NEXT: movb %dil, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: movb %dil, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: movzbl {{[0-9]+}}(%rsp), %eax ; NoVLX-NEXT: movzbl {{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT: kmovw %eax, %k1 ; NoVLX-NEXT: kmovw %eax, %k1
; NoVLX-NEXT: vpternlogq $255, %zmm2, %zmm2, %zmm2 {%k1} {z} ; NoVLX-NEXT: vpternlogq $255, %zmm2, %zmm2, %zmm2 {%k1} {z}
; NoVLX-NEXT: vxorps %xmm3, %xmm3, %xmm3
; NoVLX-NEXT: vcmpeqpd %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vcmpeqpd %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vandpd %xmm2, %xmm0, %xmm0 ; NoVLX-NEXT: vandpd %xmm2, %xmm0, %xmm0
; NoVLX-NEXT: vptestmd %zmm3, %zmm3, %k0 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero ; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
Show All 39 Lines
; NoVLX-NEXT: .Lcfi2008: ; NoVLX-NEXT: .Lcfi2008:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $96, %rsp ; NoVLX-NEXT: subq $96, %rsp
; NoVLX-NEXT: movb %dil, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: movb %dil, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: movzbl {{[0-9]+}}(%rsp), %eax ; NoVLX-NEXT: movzbl {{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT: kmovw %eax, %k1 ; NoVLX-NEXT: kmovw %eax, %k1
; NoVLX-NEXT: vpternlogq $255, %zmm1, %zmm1, %zmm1 {%k1} {z} ; NoVLX-NEXT: vpternlogq $255, %zmm1, %zmm1, %zmm1 {%k1} {z}
; NoVLX-NEXT: vxorps %xmm2, %xmm2, %xmm2
; NoVLX-NEXT: vcmpeqpd (%rsi), %xmm0, %xmm0 ; NoVLX-NEXT: vcmpeqpd (%rsi), %xmm0, %xmm0
; NoVLX-NEXT: vandpd %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vandpd %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vptestmd %zmm2, %zmm2, %k0 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero ; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
Show All 40 Lines
; NoVLX-NEXT: .Lcfi2011: ; NoVLX-NEXT: .Lcfi2011:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $96, %rsp ; NoVLX-NEXT: subq $96, %rsp
; NoVLX-NEXT: movb %dil, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: movb %dil, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: movzbl {{[0-9]+}}(%rsp), %eax ; NoVLX-NEXT: movzbl {{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT: kmovw %eax, %k1 ; NoVLX-NEXT: kmovw %eax, %k1
; NoVLX-NEXT: vpternlogq $255, %zmm1, %zmm1, %zmm1 {%k1} {z} ; NoVLX-NEXT: vpternlogq $255, %zmm1, %zmm1, %zmm1 {%k1} {z}
; NoVLX-NEXT: vxorps %xmm2, %xmm2, %xmm2 ; NoVLX-NEXT: vmovddup {{.*#+}} xmm2 = mem[0,0]
; NoVLX-NEXT: vmovddup {{.*#+}} xmm3 = mem[0,0] ; NoVLX-NEXT: vcmpeqpd %xmm2, %xmm0, %xmm0
; NoVLX-NEXT: vcmpeqpd %xmm3, %xmm0, %xmm0
; NoVLX-NEXT: vandpd %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vandpd %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vptestmd %zmm2, %zmm2, %k0 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero ; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,8],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
▲ Show 20 Lines • Show All 1,108 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %rbp, -16 ; NoVLX-NEXT: .cfi_offset %rbp, -16
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi2032: ; NoVLX-NEXT: .Lcfi2032:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: vcmpeqpd %ymm1, %ymm0, %ymm0 ; NoVLX-NEXT: vcmpeqpd %ymm1, %ymm0, %ymm0
; NoVLX-NEXT: vpmovqd %zmm0, %ymm0 ; NoVLX-NEXT: vpmovqd %zmm0, %ymm0
; NoVLX-NEXT: vxorpd %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero ; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
Show All 34 Lines
; NoVLX-NEXT: .cfi_offset %rbp, -16 ; NoVLX-NEXT: .cfi_offset %rbp, -16
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi2035: ; NoVLX-NEXT: .Lcfi2035:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: vcmpeqpd (%rdi), %ymm0, %ymm0 ; NoVLX-NEXT: vcmpeqpd (%rdi), %ymm0, %ymm0
; NoVLX-NEXT: vpmovqd %zmm0, %ymm0 ; NoVLX-NEXT: vpmovqd %zmm0, %ymm0
; NoVLX-NEXT: vxorps %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero ; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
Show All 36 Lines
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi2038: ; NoVLX-NEXT: .Lcfi2038:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: vbroadcastsd (%rdi), %ymm1 ; NoVLX-NEXT: vbroadcastsd (%rdi), %ymm1
; NoVLX-NEXT: vcmpeqpd %ymm1, %ymm0, %ymm0 ; NoVLX-NEXT: vcmpeqpd %ymm1, %ymm0, %ymm0
; NoVLX-NEXT: vpmovqd %zmm0, %ymm0 ; NoVLX-NEXT: vpmovqd %zmm0, %ymm0
; NoVLX-NEXT: vxorpd %xmm1, %xmm1, %xmm1 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero ; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
▲ Show 20 Lines • Show All 41 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $96, %rsp ; NoVLX-NEXT: subq $96, %rsp
; NoVLX-NEXT: movb %dil, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: movb %dil, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: movzbl {{[0-9]+}}(%rsp), %eax ; NoVLX-NEXT: movzbl {{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT: kmovw %eax, %k1 ; NoVLX-NEXT: kmovw %eax, %k1
; NoVLX-NEXT: vpternlogq $255, %zmm2, %zmm2, %zmm2 {%k1} {z} ; NoVLX-NEXT: vpternlogq $255, %zmm2, %zmm2, %zmm2 {%k1} {z}
; NoVLX-NEXT: vpmovqd %zmm2, %ymm2 ; NoVLX-NEXT: vpmovqd %zmm2, %ymm2
; NoVLX-NEXT: vxorps %xmm3, %xmm3, %xmm3
; NoVLX-NEXT: vcmpeqpd %ymm1, %ymm0, %ymm0 ; NoVLX-NEXT: vcmpeqpd %ymm1, %ymm0, %ymm0
; NoVLX-NEXT: vpmovqd %zmm0, %ymm0 ; NoVLX-NEXT: vpmovqd %zmm0, %ymm0
; NoVLX-NEXT: vpand %xmm2, %xmm0, %xmm0 ; NoVLX-NEXT: vpand %xmm2, %xmm0, %xmm0
; NoVLX-NEXT: vptestmd %zmm3, %zmm3, %k0 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero ; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
▲ Show 20 Lines • Show All 41 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $96, %rsp ; NoVLX-NEXT: subq $96, %rsp
; NoVLX-NEXT: movb %dil, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: movb %dil, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: movzbl {{[0-9]+}}(%rsp), %eax ; NoVLX-NEXT: movzbl {{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT: kmovw %eax, %k1 ; NoVLX-NEXT: kmovw %eax, %k1
; NoVLX-NEXT: vpternlogq $255, %zmm1, %zmm1, %zmm1 {%k1} {z} ; NoVLX-NEXT: vpternlogq $255, %zmm1, %zmm1, %zmm1 {%k1} {z}
; NoVLX-NEXT: vpmovqd %zmm1, %ymm1 ; NoVLX-NEXT: vpmovqd %zmm1, %ymm1
; NoVLX-NEXT: vxorps %xmm2, %xmm2, %xmm2
; NoVLX-NEXT: vcmpeqpd (%rsi), %ymm0, %ymm0 ; NoVLX-NEXT: vcmpeqpd (%rsi), %ymm0, %ymm0
; NoVLX-NEXT: vpmovqd %zmm0, %ymm0 ; NoVLX-NEXT: vpmovqd %zmm0, %ymm0
; NoVLX-NEXT: vpand %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vptestmd %zmm2, %zmm2, %k0 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero ; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
▲ Show 20 Lines • Show All 42 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $96, %rsp ; NoVLX-NEXT: subq $96, %rsp
; NoVLX-NEXT: movb %dil, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: movb %dil, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: movzbl {{[0-9]+}}(%rsp), %eax ; NoVLX-NEXT: movzbl {{[0-9]+}}(%rsp), %eax
; NoVLX-NEXT: kmovw %eax, %k1 ; NoVLX-NEXT: kmovw %eax, %k1
; NoVLX-NEXT: vpternlogq $255, %zmm1, %zmm1, %zmm1 {%k1} {z} ; NoVLX-NEXT: vpternlogq $255, %zmm1, %zmm1, %zmm1 {%k1} {z}
; NoVLX-NEXT: vpmovqd %zmm1, %ymm1 ; NoVLX-NEXT: vpmovqd %zmm1, %ymm1
; NoVLX-NEXT: vxorps %xmm2, %xmm2, %xmm2 ; NoVLX-NEXT: vbroadcastsd (%rsi), %ymm2
; NoVLX-NEXT: vbroadcastsd (%rsi), %ymm3 ; NoVLX-NEXT: vcmpeqpd %ymm2, %ymm0, %ymm0
; NoVLX-NEXT: vcmpeqpd %ymm3, %ymm0, %ymm0
; NoVLX-NEXT: vpmovqd %zmm0, %ymm0 ; NoVLX-NEXT: vpmovqd %zmm0, %ymm0
; NoVLX-NEXT: vpand %xmm1, %xmm0, %xmm0 ; NoVLX-NEXT: vpand %xmm1, %xmm0, %xmm0
; NoVLX-NEXT: vptestmd %zmm2, %zmm2, %k0 ; NoVLX-NEXT: kxorw %k0, %k0, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero ; NoVLX-NEXT: vpshufb {{.*#+}} ymm0 = ymm0[0,4,8,12],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm1
; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1 ; NoVLX-NEXT: vpmovsxbd %xmm1, %zmm1
; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1 ; NoVLX-NEXT: vpslld $31, %zmm1, %zmm1
; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k0, {{[0-9]+}}(%rsp)
▲ Show 20 Lines • Show All 257 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi2049: ; NoVLX-NEXT: .Lcfi2049:
; NoVLX-NEXT: .cfi_offset %rbp, -16 ; NoVLX-NEXT: .cfi_offset %rbp, -16
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi2050: ; NoVLX-NEXT: .Lcfi2050:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $32, %rsp ; NoVLX-NEXT: subq $32, %rsp
; NoVLX-NEXT: vcmpeqpd %zmm1, %zmm0, %k0 ; NoVLX-NEXT: vcmpeqpd %zmm1, %zmm0, %k0
; NoVLX-NEXT: vxorpd %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r9d ; NoVLX-NEXT: kmovw %k1, %r9d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 56 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi2052: ; NoVLX-NEXT: .Lcfi2052:
; NoVLX-NEXT: .cfi_offset %rbp, -16 ; NoVLX-NEXT: .cfi_offset %rbp, -16
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi2053: ; NoVLX-NEXT: .Lcfi2053:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $32, %rsp ; NoVLX-NEXT: subq $32, %rsp
; NoVLX-NEXT: vcmpeqpd (%rdi), %zmm0, %k0 ; NoVLX-NEXT: vcmpeqpd (%rdi), %zmm0, %k0
; NoVLX-NEXT: vxorpd %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r9d ; NoVLX-NEXT: kmovw %k1, %r9d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 57 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi2055: ; NoVLX-NEXT: .Lcfi2055:
; NoVLX-NEXT: .cfi_offset %rbp, -16 ; NoVLX-NEXT: .cfi_offset %rbp, -16
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi2056: ; NoVLX-NEXT: .Lcfi2056:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $32, %rsp ; NoVLX-NEXT: subq $32, %rsp
; NoVLX-NEXT: vcmpeqpd (%rdi){1to8}, %zmm0, %k0 ; NoVLX-NEXT: vcmpeqpd (%rdi){1to8}, %zmm0, %k0
; NoVLX-NEXT: vxorpd %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r9d ; NoVLX-NEXT: kmovw %k1, %r9d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 60 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %rbp, -16 ; NoVLX-NEXT: .cfi_offset %rbp, -16
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi2059: ; NoVLX-NEXT: .Lcfi2059:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $32, %rsp ; NoVLX-NEXT: subq $32, %rsp
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vcmpeqpd %zmm1, %zmm0, %k0 {%k1} ; NoVLX-NEXT: vcmpeqpd %zmm1, %zmm0, %k0 {%k1}
; NoVLX-NEXT: vxorpd %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r9d ; NoVLX-NEXT: kmovw %k1, %r9d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 60 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %rbp, -16 ; NoVLX-NEXT: .cfi_offset %rbp, -16
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi2062: ; NoVLX-NEXT: .Lcfi2062:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $32, %rsp ; NoVLX-NEXT: subq $32, %rsp
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vcmpeqpd (%rsi), %zmm0, %k0 {%k1} ; NoVLX-NEXT: vcmpeqpd (%rsi), %zmm0, %k0 {%k1}
; NoVLX-NEXT: vxorpd %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r9d ; NoVLX-NEXT: kmovw %k1, %r9d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 61 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %rbp, -16 ; NoVLX-NEXT: .cfi_offset %rbp, -16
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi2065: ; NoVLX-NEXT: .Lcfi2065:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $32, %rsp ; NoVLX-NEXT: subq $32, %rsp
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vcmpeqpd (%rsi){1to8}, %zmm0, %k0 {%k1} ; NoVLX-NEXT: vcmpeqpd (%rsi){1to8}, %zmm0, %k0 {%k1}
; NoVLX-NEXT: vxorpd %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r9d ; NoVLX-NEXT: kmovw %k1, %r9d
; NoVLX-NEXT: kshiftlw $13, %k0, %k1 ; NoVLX-NEXT: kshiftlw $13, %k0, %k1
▲ Show 20 Lines • Show All 110 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi2067: ; NoVLX-NEXT: .Lcfi2067:
; NoVLX-NEXT: .cfi_offset %rbp, -16 ; NoVLX-NEXT: .cfi_offset %rbp, -16
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi2068: ; NoVLX-NEXT: .Lcfi2068:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: vcmpeqpd %zmm1, %zmm0, %k0 ; NoVLX-NEXT: vcmpeqpd %zmm1, %zmm0, %k0
; NoVLX-NEXT: vxorpd %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 61 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi2070: ; NoVLX-NEXT: .Lcfi2070:
; NoVLX-NEXT: .cfi_offset %rbp, -16 ; NoVLX-NEXT: .cfi_offset %rbp, -16
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi2071: ; NoVLX-NEXT: .Lcfi2071:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: vcmpeqpd (%rdi), %zmm0, %k0 ; NoVLX-NEXT: vcmpeqpd (%rdi), %zmm0, %k0
; NoVLX-NEXT: vxorpd %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 62 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .Lcfi2073: ; NoVLX-NEXT: .Lcfi2073:
; NoVLX-NEXT: .cfi_offset %rbp, -16 ; NoVLX-NEXT: .cfi_offset %rbp, -16
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi2074: ; NoVLX-NEXT: .Lcfi2074:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: vcmpeqpd (%rdi){1to8}, %zmm0, %k0 ; NoVLX-NEXT: vcmpeqpd (%rdi){1to8}, %zmm0, %k0
; NoVLX-NEXT: vxorpd %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 65 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %rbp, -16 ; NoVLX-NEXT: .cfi_offset %rbp, -16
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi2077: ; NoVLX-NEXT: .Lcfi2077:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vcmpeqpd %zmm1, %zmm0, %k0 {%k1} ; NoVLX-NEXT: vcmpeqpd %zmm1, %zmm0, %k0 {%k1}
; NoVLX-NEXT: vxorpd %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 65 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %rbp, -16 ; NoVLX-NEXT: .cfi_offset %rbp, -16
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi2080: ; NoVLX-NEXT: .Lcfi2080:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vcmpeqpd (%rsi), %zmm0, %k0 {%k1} ; NoVLX-NEXT: vcmpeqpd (%rsi), %zmm0, %k0 {%k1}
; NoVLX-NEXT: vxorpd %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 66 Lines • ▼ Show 20 Lines
; NoVLX-NEXT: .cfi_offset %rbp, -16 ; NoVLX-NEXT: .cfi_offset %rbp, -16
; NoVLX-NEXT: movq %rsp, %rbp ; NoVLX-NEXT: movq %rsp, %rbp
; NoVLX-NEXT: .Lcfi2083: ; NoVLX-NEXT: .Lcfi2083:
; NoVLX-NEXT: .cfi_def_cfa_register %rbp ; NoVLX-NEXT: .cfi_def_cfa_register %rbp
; NoVLX-NEXT: andq $-32, %rsp ; NoVLX-NEXT: andq $-32, %rsp
; NoVLX-NEXT: subq $64, %rsp ; NoVLX-NEXT: subq $64, %rsp
; NoVLX-NEXT: kmovw %edi, %k1 ; NoVLX-NEXT: kmovw %edi, %k1
; NoVLX-NEXT: vcmpeqpd (%rsi){1to8}, %zmm0, %k0 {%k1} ; NoVLX-NEXT: vcmpeqpd (%rsi){1to8}, %zmm0, %k0 {%k1}
; NoVLX-NEXT: vxorpd %xmm0, %xmm0, %xmm0 ; NoVLX-NEXT: kxorw %k0, %k0, %k1
; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k1
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp) ; NoVLX-NEXT: kmovw %k1, {{[0-9]+}}(%rsp)
; NoVLX-NEXT: kshiftlw $15, %k0, %k1 ; NoVLX-NEXT: kshiftlw $15, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
; NoVLX-NEXT: kmovw %k1, %r8d ; NoVLX-NEXT: kmovw %k1, %r8d
; NoVLX-NEXT: kshiftlw $14, %k0, %k1 ; NoVLX-NEXT: kshiftlw $14, %k0, %k1
; NoVLX-NEXT: kshiftrw $15, %k1, %k1 ; NoVLX-NEXT: kshiftrw $15, %k1, %k1
▲ Show 20 Lines • Show All 104 Lines • Show Last 20 Lines