[X86] Type legalize v2f32 stores by widening to v4f32, casting to v2f64, extracting f64 and storing.

[craig.topper]

Summary

This is similar to what D52528 did for loads. It should match what generic type legalization does in 64-bit mode where it uses a v2i64 cast and an i64 store.

[craig.topper]

test/CodeGen/X86/2012-01-18-vbitcast.ll

10

Something odd happened here. Suspect something to do with how bitcasts from 2 x float to 2 x i32 are legalized. v2f32 is widened but v2i32 is promoted. Generic legalization probably went through the stack Can probably fix with a custom legalization of this bitcast.

test/CodeGen/X86/fold-load-vec.ll

17

Need to investigate

[craig.topper]

Try to keep combineStore from using i64 when the original type was fp. Prevents a regression in fold-load-vec.ll

[craig.topper]

Rebase

[RKSimon]

test/CodeGen/X86/2012-01-18-vbitcast.ll

10

Are you looking at this regression or should we just add it to bugzilla for now?

[craig.topper]

Rebase.

[craig.topper]

I fixed the bitcast legalization to avoid the stack in r344425. That improved that test and fixed the regression here.

[RKSimon]

LGTM - cheers

lib/Target/X86/X86ISelLowering.cpp

if (!Subtarget.useSoftFloat() && Subtarget.hasSSE2()) {
 
     setOperationAction(ISD::FP_EXTEND,          MVT::v2f32, Custom);
     setOperationAction(ISD::FP_ROUND,           MVT::v2f32, Custom);
 
     for (MVT VT : MVT::fp_vector_valuetypes())
       setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2f32, Legal);
 
     // We want to legalize this to an f64 load rather than an i64 load on
-    // 64-bit targets and two 32-bit loads on a 32-bit target.
+    // 64-bit targets and two 32-bit loads on a 32-bit target. Similar for
+    // store.
     setOperationAction(ISD::LOAD,               MVT::v2f32, Custom);
+    setOperationAction(ISD::STORE,              MVT::v2f32, Custom);
 
     setOperationAction(ISD::BITCAST,            MVT::v2i32, Custom);
     setOperationAction(ISD::BITCAST,            MVT::v4i16, Custom);
     setOperationAction(ISD::BITCAST,            MVT::v8i8,  Custom);
     if (!Subtarget.hasAVX512())
       setOperationAction(ISD::BITCAST, MVT::v16i1, Custom);
 
     setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, MVT::v2i64, Custom);

 
 static SDValue LowerStore(SDValue Op, const X86Subtarget &Subtarget,
                           SelectionDAG &DAG) {
   StoreSDNode *St = cast<StoreSDNode>(Op.getNode());
   SDLoc dl(St);
   SDValue StoredVal = St->getValue();
 
   // Without AVX512DQ, we need to use a scalar type for v2i1/v4i1/v8i1 loads.
-  assert(StoredVal.getValueType().isVector() &&
-         StoredVal.getValueType().getVectorElementType() == MVT::i1 &&
-         StoredVal.getValueType().getVectorNumElements() <= 8 &&
+  if (StoredVal.getValueType().isVector() &&
+            StoredVal.getValueType().getVectorElementType() == MVT::i1) {
+    assert(StoredVal.getValueType().getVectorNumElements() <= 8 &&
            "Unexpected VT");
     assert(!St->isTruncatingStore() && "Expected non-truncating store");
     assert(Subtarget.hasAVX512() && !Subtarget.hasDQI() &&
            "Expected AVX512F without AVX512DQI");
 
     StoredVal = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, MVT::v8i1,
                             DAG.getUNDEF(MVT::v8i1), StoredVal,
                             DAG.getIntPtrConstant(0, dl));
     StoredVal = DAG.getBitcast(MVT::i8, StoredVal);
 
     return DAG.getStore(St->getChain(), dl, StoredVal, St->getBasePtr(),
                         St->getPointerInfo(), St->getAlignment(),
                         St->getMemOperand()->getFlags());
   }
 
+  if (St->isTruncatingStore())
+    return SDValue();
+
+  assert(StoredVal.getValueType() == MVT::v2f32 && "Unexpected VT");
+
+  // Widen the vector, cast to a v2x64 type, extract the single 64-bit
+  // element and store it.
+  StoredVal = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4f32, StoredVal,
+                          DAG.getUNDEF(MVT::v2f32));
+  StoredVal = DAG.getBitcast(MVT::v2f64, StoredVal);
+  StoredVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, StoredVal,
+                          DAG.getIntPtrConstant(0, dl));
+
+  return DAG.getStore(St->getChain(), dl, StoredVal, St->getBasePtr(),
+                      St->getPointerInfo(), St->getAlignment(),
+                      St->getMemOperand()->getFlags());
+}
+
 // Lower vector extended loads using a shuffle. If SSSE3 is not available we
 // may emit an illegal shuffle but the expansion is still better than scalar
 // code. We generate X86ISD::VSEXT for SEXTLOADs if it's available, otherwise
 // we'll emit a shuffle and a arithmetic shift.
 // FIXME: Is the expansion actually better than scalar code? It doesn't seem so.
 // TODO: It is possible to support ZExt by zeroing the undef values during
 // the shuffle phase or after the shuffle.
 static SDValue LowerLoad(SDValue Op, const X86Subtarget &Subtarget,

test/CodeGen/X86/2011-10-19-widen_vselect.ll

 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
 ; RUN: llc < %s -mtriple=i686-unknown-linux-gnu -mcpu=corei7   | FileCheck %s --check-prefix=X32
 ; RUN: llc < %s -mtriple=x86_64-unknown-linux-gnu -mcpu=corei7 | FileCheck %s --check-prefix=X64
 
 ; Make sure that we don't crash when legalizing vselect and vsetcc and that
 ; we are able to generate vector blend instructions.
 
 define void @simple_widen(<2 x float> %a, <2 x float> %b) {
 ; X32-LABEL: simple_widen:
 ; X32:       # %bb.0: # %entry
-; X32-NEXT:    extractps $1, %xmm1, (%eax)
-; X32-NEXT:    movss %xmm1, (%eax)
+; X32-NEXT:    movlps %xmm1, (%eax)
 ; X32-NEXT:    retl
 ;
 ; X64-LABEL: simple_widen:
 ; X64:       # %bb.0: # %entry
 ; X64-NEXT:    movlps %xmm1, (%rax)
 ; X64-NEXT:    retq
 entry:
   %0 = select <2 x i1> undef, <2 x float> %a, <2 x float> %b
   store <2 x float> %0, <2 x float>* undef
   ret void
 }
 
 define void @complex_inreg_work(<2 x float> %a, <2 x float> %b) {
 ; X32-LABEL: complex_inreg_work:
 ; X32:       # %bb.0: # %entry
 ; X32-NEXT:    movaps %xmm0, %xmm2
 ; X32-NEXT:    cmpordps %xmm0, %xmm0
 ; X32-NEXT:    blendvps %xmm0, %xmm2, %xmm1
-; X32-NEXT:    extractps $1, %xmm1, (%eax)
-; X32-NEXT:    movss %xmm1, (%eax)
+; X32-NEXT:    movlps %xmm1, (%eax)
 ; X32-NEXT:    retl
 ;
 ; X64-LABEL: complex_inreg_work:
 ; X64:       # %bb.0: # %entry
 ; X64-NEXT:    movaps %xmm0, %xmm2
 ; X64-NEXT:    cmpordps %xmm0, %xmm0
 ; X64-NEXT:    blendvps %xmm0, %xmm2, %xmm1
 ; X64-NEXT:    movlps %xmm1, (%rax)
 ; X64-NEXT:    retq
 entry:
   %0 = fcmp oeq <2 x float> undef, undef
   %1 = select <2 x i1> %0, <2 x float> %a, <2 x float> %b
   store <2 x float> %1, <2 x float>* undef
   ret void
 }
 
 define void @zero_test() {
 ; X32-LABEL: zero_test:
 ; X32:       # %bb.0: # %entry
 ; X32-NEXT:    xorps %xmm0, %xmm0
-; X32-NEXT:    extractps $1, %xmm0, (%eax)
-; X32-NEXT:    movss %xmm0, (%eax)
+; X32-NEXT:    movlps %xmm0, (%eax)
 ; X32-NEXT:    retl
 ;
 ; X64-LABEL: zero_test:
 ; X64:       # %bb.0: # %entry
 ; X64-NEXT:    xorps %xmm0, %xmm0
 ; X64-NEXT:    movlps %xmm0, (%rax)
 ; X64-NEXT:    retq
 entry:

 ; X32-NEXT:    cmpltps %xmm2, %xmm0
 ; X32-NEXT:    movaps {{.*#+}} xmm3 = <1,1,u,u>
 ; X32-NEXT:    addps %xmm1, %xmm3
 ; X32-NEXT:    movaps %xmm1, %xmm4
 ; X32-NEXT:    blendvps %xmm0, %xmm3, %xmm4
 ; X32-NEXT:    cmpeqps %xmm2, %xmm1
 ; X32-NEXT:    movaps %xmm1, %xmm0
 ; X32-NEXT:    blendvps %xmm0, %xmm2, %xmm4
-; X32-NEXT:    movss %xmm4, {{[0-9]+}}(%esp)
-; X32-NEXT:    movshdup {{.*#+}} xmm0 = xmm4[1,1,3,3]
-; X32-NEXT:    movss %xmm0, {{[0-9]+}}(%esp)
-; X32-NEXT:    movss %xmm4, {{[0-9]+}}(%esp)
-; X32-NEXT:    movss %xmm0, {{[0-9]+}}(%esp)
+; X32-NEXT:    movlps %xmm4, {{[0-9]+}}(%esp)
+; X32-NEXT:    movlps %xmm4, {{[0-9]+}}(%esp)
 ; X32-NEXT:    addl $60, %esp
 ; X32-NEXT:    .cfi_def_cfa_offset 4
 ; X32-NEXT:    retl
 ;
 ; X64-LABEL: full_test:
 ; X64:       # %bb.0: # %entry
 ; X64-NEXT:    movsd {{.*#+}} xmm2 = mem[0],zero
 ; X64-NEXT:    cvttps2dq %xmm2, %xmm0

test/CodeGen/X86/2012-01-18-vbitcast.ll

 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
 ; RUN: llc < %s -mcpu=corei7 -mtriple=x86_64-pc-win32 | FileCheck %s
 
 define <2 x i32> @vcast(<2 x float> %a, <2 x float> %b) {
 ; CHECK-LABEL: vcast:
 ; CHECK:       # %bb.0:
 ; CHECK-NEXT:    subq $16, %rsp
 ; CHECK-NEXT:    .seh_stackalloc 16
 ; CHECK-NEXT:    .seh_endprologue
+; CHECK-NEXT:    movsd {{.*#+}} xmm0 = mem[0],zero
+; CHECK-NEXT:    movsd %xmm0, (%rsp)
+; CHECK-NEXT:    movsd {{.*#+}} xmm0 = mem[0],zero
+; CHECK-NEXT:    movsd %xmm0, {{[0-9]+}}(%rsp)
 ; CHECK-NEXT:    pmovzxdq {{.*#+}} xmm1 = mem[0],zero,mem[1],zero
 ; CHECK-NEXT:    pmovzxdq {{.*#+}} xmm0 = mem[0],zero,mem[1],zero
 ; CHECK-NEXT:    psubq %xmm1, %xmm0
 ; CHECK-NEXT:    addq $16, %rsp
 ; CHECK-NEXT:    retq
 ; CHECK-NEXT:    .seh_handlerdata
 ; CHECK-NEXT:    .text
 ; CHECK-NEXT:    .seh_endproc
   %af = bitcast <2 x float> %a to <2 x i32>
   %bf = bitcast <2 x float> %b to <2 x i32>
   %x = sub <2 x i32> %af, %bf
   ret <2 x i32> %x
 }
 

test/CodeGen/X86/fold-load-vec.ll

 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
 ; RUN: llc < %s -mtriple=x86_64-unknown -mcpu=corei7 -mattr=+sse4.1 | FileCheck %s
 
 ; rdar://12721174
 ; We should not fold movss into pshufd since pshufd expects m128 while movss
 ; loads from m32.
 define void @sample_test(<4 x float>* %source, <2 x float>* %dest) nounwind {
 ; CHECK-LABEL: sample_test:
 ; CHECK:       # %bb.0: # %entry
 ; CHECK-NEXT:    subq $24, %rsp
 ; CHECK-NEXT:    movq %rdi, {{[0-9]+}}(%rsp)
 ; CHECK-NEXT:    movq %rsi, {{[0-9]+}}(%rsp)
 ; CHECK-NEXT:    xorps %xmm0, %xmm0
 ; CHECK-NEXT:    movlps %xmm0, (%rsp)
-; CHECK-NEXT:    insertps {{.*#+}} xmm0 = xmm0[0],mem[0],xmm0[2,3]
+; CHECK-NEXT:    insertps {{.*#+}} xmm0 = zero,mem[0],zero,zero
 ; CHECK-NEXT:    movlps %xmm0, (%rsp)
-; CHECK-NEXT:    movlps %xmm0, (%rsi)
+; CHECK-NEXT:    movq (%rsp), %rax
+; CHECK-NEXT:    movq %rax, (%rsi)
 ; CHECK-NEXT:    movq {{[0-9]+}}(%rsp), %rax
 ; CHECK-NEXT:    movsd {{.*#+}} xmm0 = mem[0],zero
 ; CHECK-NEXT:    movshdup {{.*#+}} xmm1 = xmm0[1,1,3,3]
 ; CHECK-NEXT:    callq ext
 ; CHECK-NEXT:    addq $24, %rsp
 ; CHECK-NEXT:    retq
 entry:
   %source.addr = alloca <4 x float>*, align 8

test/CodeGen/X86/sse-schedule.ll

 
 ; TODO (v)movhps
 
 define <4 x float> @test_movhps(<4 x float> %a0, <4 x float> %a1, x86_mmx *%a2) {
 ; GENERIC-LABEL: test_movhps:
 ; GENERIC:       # %bb.0:
 ; GENERIC-NEXT:    movhpd {{.*#+}} xmm1 = xmm1[0],mem[0] sched: [7:1.00]
 ; GENERIC-NEXT:    addps %xmm1, %xmm0 # sched: [3:1.00]
-; GENERIC-NEXT:    movhlps {{.*#+}} xmm0 = xmm0[1,1] sched: [1:1.00]
-; GENERIC-NEXT:    movlps %xmm0, (%rdi) # sched: [1:1.00]
+; GENERIC-NEXT:    movhps %xmm0, (%rdi) # sched: [1:1.00]
 ; GENERIC-NEXT:    movaps %xmm1, %xmm0 # sched: [1:1.00]
 ; GENERIC-NEXT:    retq # sched: [1:1.00]
 ;
 ; ATOM-LABEL: test_movhps:
 ; ATOM:       # %bb.0:
 ; ATOM-NEXT:    movaps %xmm0, %xmm2 # sched: [1:0.50]
 ; ATOM-NEXT:    movhpd {{.*#+}} xmm1 = xmm1[0],mem[0] sched: [1:1.00]
 ; ATOM-NEXT:    addps %xmm1, %xmm2 # sched: [5:5.00]
 ; ATOM-NEXT:    movaps %xmm1, %xmm0 # sched: [1:0.50]
-; ATOM-NEXT:    movhlps {{.*#+}} xmm2 = xmm2[1,1] sched: [1:1.00]
-; ATOM-NEXT:    movlps %xmm2, (%rdi) # sched: [1:1.00]
+; ATOM-NEXT:    movhps %xmm2, (%rdi) # sched: [1:1.00]
 ; ATOM-NEXT:    retq # sched: [79:39.50]
 ;
 ; SLM-LABEL: test_movhps:
 ; SLM:       # %bb.0:
 ; SLM-NEXT:    movhpd {{.*#+}} xmm1 = xmm1[0],mem[0] sched: [4:1.00]
 ; SLM-NEXT:    addps %xmm1, %xmm0 # sched: [3:1.00]
-; SLM-NEXT:    movhlps {{.*#+}} xmm0 = xmm0[1,1] sched: [1:1.00]
-; SLM-NEXT:    movlps %xmm0, (%rdi) # sched: [1:1.00]
+; SLM-NEXT:    movhps %xmm0, (%rdi) # sched: [1:1.00]
 ; SLM-NEXT:    movaps %xmm1, %xmm0 # sched: [1:0.50]
 ; SLM-NEXT:    retq # sched: [4:1.00]
 ;
 ; SANDY-SSE-LABEL: test_movhps:
 ; SANDY-SSE:       # %bb.0:
 ; SANDY-SSE-NEXT:    movhpd {{.*#+}} xmm1 = xmm1[0],mem[0] sched: [7:1.00]
 ; SANDY-SSE-NEXT:    addps %xmm1, %xmm0 # sched: [3:1.00]
-; SANDY-SSE-NEXT:    movhlps {{.*#+}} xmm0 = xmm0[1,1] sched: [1:1.00]
-; SANDY-SSE-NEXT:    movlps %xmm0, (%rdi) # sched: [1:1.00]
+; SANDY-SSE-NEXT:    movhps %xmm0, (%rdi) # sched: [1:1.00]
 ; SANDY-SSE-NEXT:    movaps %xmm1, %xmm0 # sched: [1:1.00]
 ; SANDY-SSE-NEXT:    retq # sched: [1:1.00]
 ;
 ; SANDY-LABEL: test_movhps:
 ; SANDY:       # %bb.0:
 ; SANDY-NEXT:    vmovhpd {{.*#+}} xmm1 = xmm1[0],mem[0] sched: [7:1.00]
 ; SANDY-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
-; SANDY-NEXT:    vpextrq $1, %xmm0, (%rdi) # sched: [5:1.00]
+; SANDY-NEXT:    vmovhpd %xmm0, (%rdi) # sched: [1:1.00]
 ; SANDY-NEXT:    vmovaps %xmm1, %xmm0 # sched: [1:1.00]
 ; SANDY-NEXT:    retq # sched: [1:1.00]
 ;
 ; HASWELL-SSE-LABEL: test_movhps:
 ; HASWELL-SSE:       # %bb.0:
 ; HASWELL-SSE-NEXT:    movhpd {{.*#+}} xmm1 = xmm1[0],mem[0] sched: [6:1.00]
 ; HASWELL-SSE-NEXT:    addps %xmm1, %xmm0 # sched: [3:1.00]
-; HASWELL-SSE-NEXT:    movhlps {{.*#+}} xmm0 = xmm0[1,1] sched: [1:1.00]
-; HASWELL-SSE-NEXT:    movlps %xmm0, (%rdi) # sched: [1:1.00]
+; HASWELL-SSE-NEXT:    movhps %xmm0, (%rdi) # sched: [1:1.00]
 ; HASWELL-SSE-NEXT:    movaps %xmm1, %xmm0 # sched: [1:1.00]
 ; HASWELL-SSE-NEXT:    retq # sched: [7:1.00]
 ;
 ; HASWELL-LABEL: test_movhps:
 ; HASWELL:       # %bb.0:
 ; HASWELL-NEXT:    vmovhpd {{.*#+}} xmm1 = xmm1[0],mem[0] sched: [6:1.00]
 ; HASWELL-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
-; HASWELL-NEXT:    vpextrq $1, %xmm0, (%rdi) # sched: [2:1.00]
+; HASWELL-NEXT:    vmovhpd %xmm0, (%rdi) # sched: [1:1.00]
 ; HASWELL-NEXT:    vmovaps %xmm1, %xmm0 # sched: [1:1.00]
 ; HASWELL-NEXT:    retq # sched: [7:1.00]
 ;
 ; BROADWELL-SSE-LABEL: test_movhps:
 ; BROADWELL-SSE:       # %bb.0:
 ; BROADWELL-SSE-NEXT:    movhpd {{.*#+}} xmm1 = xmm1[0],mem[0] sched: [6:1.00]
 ; BROADWELL-SSE-NEXT:    addps %xmm1, %xmm0 # sched: [3:1.00]
-; BROADWELL-SSE-NEXT:    movhlps {{.*#+}} xmm0 = xmm0[1,1] sched: [1:1.00]
-; BROADWELL-SSE-NEXT:    movlps %xmm0, (%rdi) # sched: [1:1.00]
+; BROADWELL-SSE-NEXT:    movhps %xmm0, (%rdi) # sched: [1:1.00]
 ; BROADWELL-SSE-NEXT:    movaps %xmm1, %xmm0 # sched: [1:1.00]
 ; BROADWELL-SSE-NEXT:    retq # sched: [7:1.00]
 ;
 ; BROADWELL-LABEL: test_movhps:
 ; BROADWELL:       # %bb.0:
 ; BROADWELL-NEXT:    vmovhpd {{.*#+}} xmm1 = xmm1[0],mem[0] sched: [6:1.00]
 ; BROADWELL-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
-; BROADWELL-NEXT:    vpextrq $1, %xmm0, (%rdi) # sched: [2:1.00]
+; BROADWELL-NEXT:    vmovhpd %xmm0, (%rdi) # sched: [1:1.00]
 ; BROADWELL-NEXT:    vmovaps %xmm1, %xmm0 # sched: [1:1.00]
 ; BROADWELL-NEXT:    retq # sched: [7:1.00]
 ;
 ; SKYLAKE-SSE-LABEL: test_movhps:
 ; SKYLAKE-SSE:       # %bb.0:
 ; SKYLAKE-SSE-NEXT:    movhpd {{.*#+}} xmm1 = xmm1[0],mem[0] sched: [6:1.00]
 ; SKYLAKE-SSE-NEXT:    addps %xmm1, %xmm0 # sched: [4:0.50]
-; SKYLAKE-SSE-NEXT:    movhlps {{.*#+}} xmm0 = xmm0[1,1] sched: [1:1.00]
-; SKYLAKE-SSE-NEXT:    movlps %xmm0, (%rdi) # sched: [1:1.00]
+; SKYLAKE-SSE-NEXT:    movhps %xmm0, (%rdi) # sched: [1:1.00]
 ; SKYLAKE-SSE-NEXT:    movaps %xmm1, %xmm0 # sched: [1:0.33]
 ; SKYLAKE-SSE-NEXT:    retq # sched: [7:1.00]
 ;
 ; SKYLAKE-LABEL: test_movhps:
 ; SKYLAKE:       # %bb.0:
 ; SKYLAKE-NEXT:    vmovhpd {{.*#+}} xmm1 = xmm1[0],mem[0] sched: [6:1.00]
 ; SKYLAKE-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [4:0.50]
-; SKYLAKE-NEXT:    vpextrq $1, %xmm0, (%rdi) # sched: [2:1.00]
+; SKYLAKE-NEXT:    vmovhpd %xmm0, (%rdi) # sched: [1:1.00]
 ; SKYLAKE-NEXT:    vmovaps %xmm1, %xmm0 # sched: [1:0.33]
 ; SKYLAKE-NEXT:    retq # sched: [7:1.00]
 ;
 ; SKX-SSE-LABEL: test_movhps:
 ; SKX-SSE:       # %bb.0:
 ; SKX-SSE-NEXT:    movhpd {{.*#+}} xmm1 = xmm1[0],mem[0] sched: [6:1.00]
 ; SKX-SSE-NEXT:    addps %xmm1, %xmm0 # sched: [4:0.50]
-; SKX-SSE-NEXT:    movhlps {{.*#+}} xmm0 = xmm0[1,1] sched: [1:1.00]
-; SKX-SSE-NEXT:    movlps %xmm0, (%rdi) # sched: [1:1.00]
+; SKX-SSE-NEXT:    movhps %xmm0, (%rdi) # sched: [1:1.00]
 ; SKX-SSE-NEXT:    movaps %xmm1, %xmm0 # sched: [1:0.33]
 ; SKX-SSE-NEXT:    retq # sched: [7:1.00]
 ;
 ; SKX-LABEL: test_movhps:
 ; SKX:       # %bb.0:
 ; SKX-NEXT:    vmovhpd {{.*#+}} xmm1 = xmm1[0],mem[0] sched: [6:1.00]
 ; SKX-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [4:0.50]
-; SKX-NEXT:    vpextrq $1, %xmm0, (%rdi) # sched: [2:1.00]
+; SKX-NEXT:    vmovhpd %xmm0, (%rdi) # sched: [1:1.00]
 ; SKX-NEXT:    vmovaps %xmm1, %xmm0 # sched: [1:0.33]
 ; SKX-NEXT:    retq # sched: [7:1.00]
 ;
 ; BTVER2-SSE-LABEL: test_movhps:
 ; BTVER2-SSE:       # %bb.0:
 ; BTVER2-SSE-NEXT:    movhpd {{.*#+}} xmm1 = xmm1[0],mem[0] sched: [6:1.00]
 ; BTVER2-SSE-NEXT:    addps %xmm1, %xmm0 # sched: [3:1.00]
-; BTVER2-SSE-NEXT:    movhlps {{.*#+}} xmm0 = xmm0[1,1] sched: [1:0.50]
-; BTVER2-SSE-NEXT:    movlps %xmm0, (%rdi) # sched: [2:1.00]
+; BTVER2-SSE-NEXT:    movhps %xmm0, (%rdi) # sched: [2:1.00]
 ; BTVER2-SSE-NEXT:    movaps %xmm1, %xmm0 # sched: [1:0.50]
 ; BTVER2-SSE-NEXT:    retq # sched: [4:1.00]
 ;
 ; BTVER2-LABEL: test_movhps:
 ; BTVER2:       # %bb.0:
 ; BTVER2-NEXT:    vmovhpd {{.*#+}} xmm1 = xmm1[0],mem[0] sched: [6:1.00]
 ; BTVER2-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
-; BTVER2-NEXT:    vpextrq $1, %xmm0, (%rdi) # sched: [3:1.00]
+; BTVER2-NEXT:    vmovhpd %xmm0, (%rdi) # sched: [2:1.00]
 ; BTVER2-NEXT:    vmovaps %xmm1, %xmm0 # sched: [1:0.50]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
 ; ZNVER1-SSE-LABEL: test_movhps:
 ; ZNVER1-SSE:       # %bb.0:
 ; ZNVER1-SSE-NEXT:    movhpd {{.*#+}} xmm1 = xmm1[0],mem[0] sched: [8:0.50]
 ; ZNVER1-SSE-NEXT:    addps %xmm1, %xmm0 # sched: [3:1.00]
-; ZNVER1-SSE-NEXT:    movhlps {{.*#+}} xmm0 = xmm0[1,1] sched: [1:0.50]
-; ZNVER1-SSE-NEXT:    movlps %xmm0, (%rdi) # sched: [1:0.50]
+; ZNVER1-SSE-NEXT:    movhps %xmm0, (%rdi) # sched: [1:0.50]
 ; ZNVER1-SSE-NEXT:    movaps %xmm1, %xmm0 # sched: [1:0.25]
 ; ZNVER1-SSE-NEXT:    retq # sched: [1:0.50]
 ;
 ; ZNVER1-LABEL: test_movhps:
 ; ZNVER1:       # %bb.0:
 ; ZNVER1-NEXT:    vmovhpd {{.*#+}} xmm1 = xmm1[0],mem[0] sched: [8:0.50]
 ; ZNVER1-NEXT:    vaddps %xmm1, %xmm0, %xmm0 # sched: [3:1.00]
-; ZNVER1-NEXT:    vpextrq $1, %xmm0, (%rdi) # sched: [5:3.00]
+; ZNVER1-NEXT:    vmovhpd %xmm0, (%rdi) # sched: [1:0.50]
 ; ZNVER1-NEXT:    vmovaps %xmm1, %xmm0 # sched: [1:0.25]
 ; ZNVER1-NEXT:    retq # sched: [1:0.50]
   %1 = bitcast x86_mmx* %a2 to <2 x float>*
   %2 = load <2 x float>, <2 x float> *%1, align 8
   %3 = shufflevector <2 x float> %2, <2 x float> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
   %4 = shufflevector <4 x float> %a1, <4 x float> %3, <4 x i32> <i32 0, i32 1, i32 4, i32 5>
   %5 = fadd <4 x float> %a0, %4
   %6 = shufflevector <4 x float> %5, <4 x float> undef, <2 x i32> <i32 2, i32 3>

test/CodeGen/X86/vec_fptrunc.ll

 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
 ; RUN: llc < %s -mtriple=i686-unknown -mattr=+sse4.1 | FileCheck %s --check-prefix=X32-SSE
 ; RUN: llc < %s -mtriple=i686-unknown -mattr=+avx | FileCheck %s --check-prefix=X32-AVX
 ; RUN: llc < %s -mtriple=x86_64-unknown -mattr=+sse4.1 | FileCheck %s --check-prefix=X64-SSE
 ; RUN: llc < %s -mtriple=x86_64-unknown -mattr=+avx | FileCheck %s --check-prefix=X64-AVX
 
 define void @fptrunc_frommem2(<2 x double>* %in, <2 x float>* %out) {
 ; X32-SSE-LABEL: fptrunc_frommem2:
 ; X32-SSE:       # %bb.0: # %entry
 ; X32-SSE-NEXT:    movl {{[0-9]+}}(%esp), %eax
 ; X32-SSE-NEXT:    movl {{[0-9]+}}(%esp), %ecx
 ; X32-SSE-NEXT:    cvtpd2ps (%ecx), %xmm0
-; X32-SSE-NEXT:    extractps $1, %xmm0, 4(%eax)
-; X32-SSE-NEXT:    movss %xmm0, (%eax)
+; X32-SSE-NEXT:    movlpd %xmm0, (%eax)
 ; X32-SSE-NEXT:    retl
 ;
 ; X32-AVX-LABEL: fptrunc_frommem2:
 ; X32-AVX:       # %bb.0: # %entry
 ; X32-AVX-NEXT:    movl {{[0-9]+}}(%esp), %eax
 ; X32-AVX-NEXT:    movl {{[0-9]+}}(%esp), %ecx
 ; X32-AVX-NEXT:    vcvtpd2psx (%ecx), %xmm0
-; X32-AVX-NEXT:    vextractps $1, %xmm0, 4(%eax)
-; X32-AVX-NEXT:    vmovss %xmm0, (%eax)
+; X32-AVX-NEXT:    vmovlpd %xmm0, (%eax)
 ; X32-AVX-NEXT:    retl
 ;
 ; X64-SSE-LABEL: fptrunc_frommem2:
 ; X64-SSE:       # %bb.0: # %entry
 ; X64-SSE-NEXT:    cvtpd2ps (%rdi), %xmm0
 ; X64-SSE-NEXT:    movlpd %xmm0, (%rsi)
 ; X64-SSE-NEXT:    retq
 ;

test/CodeGen/X86/widen_conv-3.ll

 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
 ; RUN: llc < %s -mtriple=i686-unknown -mattr=+sse2 | FileCheck %s --check-prefix=X86 --check-prefix=X86-SSE2
 ; RUN: llc < %s -mtriple=i686-unknown -mattr=+sse4.2 | FileCheck %s --check-prefix=X86 --check-prefix=X86-SSE42
 ; RUN: llc < %s -mtriple=x86_64-unknown -mattr=+sse2 | FileCheck %s --check-prefix=X64 --check-prefix=X64-SSE2
 ; RUN: llc < %s -mtriple=x86_64-unknown -mattr=+sse4.2 | FileCheck %s --check-prefix=X64 --check-prefix=X64-SSE42
 
 ; sign to float v2i16 to v2f32
 
 define void @convert_v2i16_to_v2f32(<2 x float>* %dst.addr, <2 x i16> %src) nounwind {
-; X86-SSE2-LABEL: convert_v2i16_to_v2f32:
-; X86-SSE2:       # %bb.0: # %entry
-; X86-SSE2-NEXT:    movl {{[0-9]+}}(%esp), %eax
-; X86-SSE2-NEXT:    psllq $48, %xmm0
-; X86-SSE2-NEXT:    psrad $16, %xmm0
-; X86-SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,3,2,3]
-; X86-SSE2-NEXT:    cvtdq2ps %xmm0, %xmm0
-; X86-SSE2-NEXT:    movss %xmm0, (%eax)
-; X86-SSE2-NEXT:    shufps {{.*#+}} xmm0 = xmm0[1,1,2,3]
-; X86-SSE2-NEXT:    movss %xmm0, 4(%eax)
-; X86-SSE2-NEXT:    retl
-;
-; X86-SSE42-LABEL: convert_v2i16_to_v2f32:
-; X86-SSE42:       # %bb.0: # %entry
-; X86-SSE42-NEXT:    movl {{[0-9]+}}(%esp), %eax
-; X86-SSE42-NEXT:    psllq $48, %xmm0
-; X86-SSE42-NEXT:    psrad $16, %xmm0
-; X86-SSE42-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,3,2,3]
-; X86-SSE42-NEXT:    cvtdq2ps %xmm0, %xmm0
-; X86-SSE42-NEXT:    extractps $1, %xmm0, 4(%eax)
-; X86-SSE42-NEXT:    movss %xmm0, (%eax)
-; X86-SSE42-NEXT:    retl
+; X86-LABEL: convert_v2i16_to_v2f32:
+; X86:       # %bb.0: # %entry
+; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X86-NEXT:    psllq $48, %xmm0
+; X86-NEXT:    psrad $16, %xmm0
+; X86-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,3,2,3]
+; X86-NEXT:    cvtdq2ps %xmm0, %xmm0
+; X86-NEXT:    movlps %xmm0, (%eax)
+; X86-NEXT:    retl
 ;
 ; X64-LABEL: convert_v2i16_to_v2f32:
 ; X64:       # %bb.0: # %entry
 ; X64-NEXT:    psllq $48, %xmm0
 ; X64-NEXT:    psrad $16, %xmm0
 ; X64-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[1,3,2,3]
 ; X64-NEXT:    cvtdq2ps %xmm0, %xmm0
 ; X64-NEXT:    movlps %xmm0, (%rdi)