Index: lib/Target/X86/X86ISelLowering.cpp
===================================================================
--- lib/Target/X86/X86ISelLowering.cpp
+++ lib/Target/X86/X86ISelLowering.cpp
@@ -6602,6 +6602,40 @@
return false;
}
+// Check if the current node of build vector is a zero extended vector.
+// If so, return the value extended.
+static SDValue isSplatZeroExtended(const BuildVectorSDNode *Op,
+ unsigned &NumElt, MVT &EltType) {
+ SDValue ExtValue = Op->getOperand(0);
+ unsigned Delta = Op->getNumOperands();
+
+ for (unsigned i = 1; i < Op->getNumOperands(); i++) {
+ if (Op->getOperand(i) == ExtValue) {
+ Delta = i;
+ break;
+ }
+ if (!(Op->getOperand(i).isUndef() || isNullConstant(Op->getOperand(i))))
+ return SDValue();
+ }
+ if (!isPowerOf2_32(Delta) || Delta == 1)
+ return SDValue();
+
+ for (unsigned i = Delta; i < Op->getNumOperands(); i++) {
+ if (i % Delta == 0) {
+ if (Op->getOperand(i) != ExtValue)
+ return SDValue();
+ } else if (!(isNullConstant(Op->getOperand(i)) ||
+ Op->getOperand(i).isUndef()))
+ return SDValue();
+ }
+ unsigned EltSize =
+ Op->getSimpleValueType(0).getVectorElementType().getSizeInBits();
+ unsigned ExtVTSize = EltSize * Delta;
+ EltType = MVT::getIntegerVT(ExtVTSize);
+ NumElt = Op->getNumOperands() / Delta;
+ return ExtValue;
+}
+
/// Attempt to use the vbroadcast instruction to generate a splat value
/// from a splat BUILD_VECTOR which uses:
/// a. A single scalar load, or a constant.
@@ -6624,6 +6658,32 @@
assert((VT.is128BitVector() || VT.is256BitVector() || VT.is512BitVector()) &&
"Unsupported vector type for broadcast.");
+ // Attempt to use VBROADCASTM
+ // From this paterrn:
+ // a. t0 = (zext_i64 (bitcast_i8 v2i1 X))
+ // b. t1 = (build_vector t0 t0)
+ //
+ // Create (VBROADCASTM v2i1 X)
+ if (Subtarget.hasCDI() && (VT.is512BitVector() || Subtarget.hasVLX())) {
+ MVT EltType;
+ unsigned NumElts = 1;
+ SDValue ZeroExtended = isSplatZeroExtended(BVOp, NumElts, EltType);
+ if (ZeroExtended && ZeroExtended.getOpcode() == ISD::BITCAST) {
+ SDValue BOperand = ZeroExtended.getOperand(0);
+ if (BOperand.getSimpleValueType().getVectorElementType() == MVT::i1) {
+ if ((EltType == MVT::i64 &&
+ VT.getVectorElementType() == MVT::i8) || // for broadcastmb2q
+ (EltType == MVT::i32 &&
+ VT.getVectorElementType() == MVT::i16)) { // for broadcastmw2d
+ SDValue Brdcst =
+ DAG.getNode(X86ISD::VBROADCASTM, dl,
+ MVT::getVectorVT(EltType, NumElts), BOperand);
+ return DAG.getBitcast(VT, Brdcst);
+ }
+ }
+ }
+ }
+
BitVector UndefElements;
SDValue Ld = BVOp->getSplatValue(&UndefElements);
Index: test/CodeGen/X86/broadcastm-lowering.ll
===================================================================
--- /dev/null
+++ test/CodeGen/X86/broadcastm-lowering.ll
@@ -0,0 +1,215 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512cd | FileCheck %s --check-prefix=ALL --check-prefix=AVX512CD
+; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512vl,avx512cd,+avx512bw| FileCheck %s --check-prefix=ALL --check-prefix=AVX512VLCDBW
+; RUN: llc < %s -mtriple=i686-unknown-unknown -mattr=+avx512vl,avx512cd,+avx512bw| FileCheck %s --check-prefix=ALL --check-prefix=X86-AVX512VLCDBW
+
+define <2 x i64> @test_mm_epi64(<8 x i16> %a, <8 x i16> %b) {
+; AVX512CD-LABEL: test_mm_epi64:
+; AVX512CD: # BB#0: # %entry
+; AVX512CD-NEXT: vpcmpeqw %xmm1, %xmm0, %xmm0
+; AVX512CD-NEXT: vpmovsxwq %xmm0, %zmm0
+; AVX512CD-NEXT: vpsllq $63, %zmm0, %zmm0
+; AVX512CD-NEXT: vptestmq %zmm0, %zmm0, %k0
+; AVX512CD-NEXT: kmovw %k0, %eax
+; AVX512CD-NEXT: vpxor %xmm0, %xmm0, %xmm0
+; AVX512CD-NEXT: vpinsrb $0, %eax, %xmm0, %xmm0
+; AVX512CD-NEXT: vpinsrb $8, %eax, %xmm0, %xmm0
+; AVX512CD-NEXT: vzeroupper
+; AVX512CD-NEXT: retq
+;
+; AVX512VLCDBW-LABEL: test_mm_epi64:
+; AVX512VLCDBW: # BB#0: # %entry
+; AVX512VLCDBW-NEXT: vpcmpeqw %xmm1, %xmm0, %k0
+; AVX512VLCDBW-NEXT: vpbroadcastmb2q %k0, %xmm0
+; AVX512VLCDBW-NEXT: retq
+;
+; X86-AVX512VLCDBW-LABEL: test_mm_epi64:
+; X86-AVX512VLCDBW: # BB#0: # %entry
+; X86-AVX512VLCDBW-NEXT: vpcmpeqw %xmm1, %xmm0, %k0
+; X86-AVX512VLCDBW-NEXT: kmovd %k0, %eax
+; X86-AVX512VLCDBW-NEXT: movzbl %al, %eax
+; X86-AVX512VLCDBW-NEXT: vmovd %eax, %xmm0
+; X86-AVX512VLCDBW-NEXT: vpbroadcastq %xmm0, %xmm0
+; X86-AVX512VLCDBW-NEXT: retl
+entry:
+ %0 = icmp eq <8 x i16> %a, %b
+ %1 = bitcast <8 x i1> %0 to i8
+ %conv.i = zext i8 %1 to i64
+ %vecinit.i.i = insertelement <2 x i64> undef, i64 %conv.i, i32 0
+ %vecinit1.i.i = shufflevector <2 x i64> %vecinit.i.i, <2 x i64> undef, <2 x i32> zeroinitializer
+ ret <2 x i64> %vecinit1.i.i
+}
+
+define <4 x i32> @test_mm_epi32(<16 x i8> %a, <16 x i8> %b) {
+; AVX512CD-LABEL: test_mm_epi32:
+; AVX512CD: # BB#0: # %entry
+; AVX512CD-NEXT: vpcmpeqb %xmm1, %xmm0, %xmm0
+; AVX512CD-NEXT: vpmovsxbd %xmm0, %zmm0
+; AVX512CD-NEXT: vpslld $31, %zmm0, %zmm0
+; AVX512CD-NEXT: vptestmd %zmm0, %zmm0, %k0
+; AVX512CD-NEXT: kmovw %k0, %eax
+; AVX512CD-NEXT: vpxor %xmm0, %xmm0, %xmm0
+; AVX512CD-NEXT: vpinsrw $0, %eax, %xmm0, %xmm0
+; AVX512CD-NEXT: vpinsrw $2, %eax, %xmm0, %xmm0
+; AVX512CD-NEXT: vpinsrw $4, %eax, %xmm0, %xmm0
+; AVX512CD-NEXT: vpinsrw $6, %eax, %xmm0, %xmm0
+; AVX512CD-NEXT: vzeroupper
+; AVX512CD-NEXT: retq
+;
+; AVX512VLCDBW-LABEL: test_mm_epi32:
+; AVX512VLCDBW: # BB#0: # %entry
+; AVX512VLCDBW-NEXT: vpcmpeqb %xmm1, %xmm0, %k0
+; AVX512VLCDBW-NEXT: vpbroadcastmw2d %k0, %xmm0
+; AVX512VLCDBW-NEXT: retq
+;
+; X86-AVX512VLCDBW-LABEL: test_mm_epi32:
+; X86-AVX512VLCDBW: # BB#0: # %entry
+; X86-AVX512VLCDBW-NEXT: vpcmpeqb %xmm1, %xmm0, %k0
+; X86-AVX512VLCDBW-NEXT: vpbroadcastmw2d %k0, %xmm0
+; X86-AVX512VLCDBW-NEXT: retl
+entry:
+ %0 = icmp eq <16 x i8> %a, %b
+ %1 = bitcast <16 x i1> %0 to i16
+ %conv.i = zext i16 %1 to i32
+ %vecinit.i.i = insertelement <4 x i32> undef, i32 %conv.i, i32 0
+ %vecinit3.i.i = shufflevector <4 x i32> %vecinit.i.i, <4 x i32> undef, <4 x i32> zeroinitializer
+ ret <4 x i32> %vecinit3.i.i
+}
+
+define <16 x i32> @test_mm512_epi32(<16 x i32> %a, <16 x i32> %b) {
+; AVX512CD-LABEL: test_mm512_epi32:
+; AVX512CD: # BB#0: # %entry
+; AVX512CD-NEXT: vpcmpeqd %zmm1, %zmm0, %k0
+; AVX512CD-NEXT: kmovw %k0, %eax
+; AVX512CD-NEXT: vpbroadcastd %eax, %zmm0
+; AVX512CD-NEXT: retq
+;
+; AVX512VLCDBW-LABEL: test_mm512_epi32:
+; AVX512VLCDBW: # BB#0: # %entry
+; AVX512VLCDBW-NEXT: vpcmpeqd %zmm1, %zmm0, %k0
+; AVX512VLCDBW-NEXT: vpbroadcastmw2d %k0, %zmm0
+; AVX512VLCDBW-NEXT: retq
+;
+; X86-AVX512VLCDBW-LABEL: test_mm512_epi32:
+; X86-AVX512VLCDBW: # BB#0: # %entry
+; X86-AVX512VLCDBW-NEXT: vpcmpeqd %zmm1, %zmm0, %k0
+; X86-AVX512VLCDBW-NEXT: vpbroadcastmw2d %k0, %zmm0
+; X86-AVX512VLCDBW-NEXT: retl
+entry:
+ %0 = icmp eq <16 x i32> %a, %b
+ %1 = bitcast <16 x i1> %0 to i16
+ %conv.i = zext i16 %1 to i32
+ %vecinit.i.i = insertelement <16 x i32> undef, i32 %conv.i, i32 0
+ %vecinit15.i.i = shufflevector <16 x i32> %vecinit.i.i, <16 x i32> undef, <16 x i32> zeroinitializer
+ ret <16 x i32> %vecinit15.i.i
+}
+
+define <8 x i64> @test_mm512_epi64(<8 x i32> %a, <8 x i32> %b) {
+; AVX512CD-LABEL: test_mm512_epi64:
+; AVX512CD: # BB#0: # %entry
+; AVX512CD-NEXT: # kill: %YMM1<def> %YMM1<kill> %ZMM1<def>
+; AVX512CD-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def>
+; AVX512CD-NEXT: vpcmpeqd %zmm1, %zmm0, %k0
+; AVX512CD-NEXT: kmovw %k0, %eax
+; AVX512CD-NEXT: movzbl %al, %eax
+; AVX512CD-NEXT: vpbroadcastq %rax, %zmm0
+; AVX512CD-NEXT: retq
+;
+; AVX512VLCDBW-LABEL: test_mm512_epi64:
+; AVX512VLCDBW: # BB#0: # %entry
+; AVX512VLCDBW-NEXT: vpcmpeqd %ymm1, %ymm0, %k0
+; AVX512VLCDBW-NEXT: vpbroadcastmb2q %k0, %zmm0
+; AVX512VLCDBW-NEXT: retq
+;
+; X86-AVX512VLCDBW-LABEL: test_mm512_epi64:
+; X86-AVX512VLCDBW: # BB#0: # %entry
+; X86-AVX512VLCDBW-NEXT: vpcmpeqd %ymm1, %ymm0, %k0
+; X86-AVX512VLCDBW-NEXT: kmovd %k0, %eax
+; X86-AVX512VLCDBW-NEXT: movzbl %al, %eax
+; X86-AVX512VLCDBW-NEXT: vmovd %eax, %xmm0
+; X86-AVX512VLCDBW-NEXT: vpbroadcastq %xmm0, %xmm0
+; X86-AVX512VLCDBW-NEXT: vinserti128 $1, %xmm0, %ymm0, %ymm0
+; X86-AVX512VLCDBW-NEXT: vinserti64x4 $1, %ymm0, %zmm0, %zmm0
+; X86-AVX512VLCDBW-NEXT: retl
+entry:
+ %0 = icmp eq <8 x i32> %a, %b
+ %1 = bitcast <8 x i1> %0 to i8
+ %conv.i = zext i8 %1 to i64
+ %vecinit.i.i = insertelement <8 x i64> undef, i64 %conv.i, i32 0
+ %vecinit7.i.i = shufflevector <8 x i64> %vecinit.i.i, <8 x i64> undef, <8 x i32> zeroinitializer
+ ret <8 x i64> %vecinit7.i.i
+}
+
+define <4 x i64> @test_mm256_epi64(<8 x i32> %a, <8 x i32> %b) {
+; AVX512CD-LABEL: test_mm256_epi64:
+; AVX512CD: # BB#0: # %entry
+; AVX512CD-NEXT: # kill: %YMM1<def> %YMM1<kill> %ZMM1<def>
+; AVX512CD-NEXT: # kill: %YMM0<def> %YMM0<kill> %ZMM0<def>
+; AVX512CD-NEXT: vpcmpeqd %zmm1, %zmm0, %k0
+; AVX512CD-NEXT: kmovw %k0, %eax
+; AVX512CD-NEXT: vpxor %xmm0, %xmm0, %xmm0
+; AVX512CD-NEXT: vpinsrb $0, %eax, %xmm0, %xmm0
+; AVX512CD-NEXT: vpinsrb $8, %eax, %xmm0, %xmm0
+; AVX512CD-NEXT: vinserti128 $1, %xmm0, %ymm0, %ymm0
+; AVX512CD-NEXT: retq
+;
+; AVX512VLCDBW-LABEL: test_mm256_epi64:
+; AVX512VLCDBW: # BB#0: # %entry
+; AVX512VLCDBW-NEXT: vpcmpeqd %ymm1, %ymm0, %k0
+; AVX512VLCDBW-NEXT: vpbroadcastmb2q %k0, %ymm0
+; AVX512VLCDBW-NEXT: retq
+;
+; X86-AVX512VLCDBW-LABEL: test_mm256_epi64:
+; X86-AVX512VLCDBW: # BB#0: # %entry
+; X86-AVX512VLCDBW-NEXT: vpcmpeqd %ymm1, %ymm0, %k0
+; X86-AVX512VLCDBW-NEXT: kmovd %k0, %eax
+; X86-AVX512VLCDBW-NEXT: movzbl %al, %eax
+; X86-AVX512VLCDBW-NEXT: vmovd %eax, %xmm0
+; X86-AVX512VLCDBW-NEXT: vpbroadcastq %xmm0, %xmm0
+; X86-AVX512VLCDBW-NEXT: vinserti128 $1, %xmm0, %ymm0, %ymm0
+; X86-AVX512VLCDBW-NEXT: retl
+entry:
+ %0 = icmp eq <8 x i32> %a, %b
+ %1 = bitcast <8 x i1> %0 to i8
+ %conv.i = zext i8 %1 to i64
+ %vecinit.i.i = insertelement <4 x i64> undef, i64 %conv.i, i32 0
+ %vecinit3.i.i = shufflevector <4 x i64> %vecinit.i.i, <4 x i64> undef, <4 x i32> zeroinitializer
+ ret <4 x i64> %vecinit3.i.i
+}
+
+define <8 x i32> @test_mm256_epi32(<16 x i16> %a, <16 x i16> %b) {
+; AVX512CD-LABEL: test_mm256_epi32:
+; AVX512CD: # BB#0: # %entry
+; AVX512CD-NEXT: vpcmpeqw %ymm1, %ymm0, %ymm0
+; AVX512CD-NEXT: vpmovsxwd %ymm0, %zmm0
+; AVX512CD-NEXT: vpslld $31, %zmm0, %zmm0
+; AVX512CD-NEXT: vptestmd %zmm0, %zmm0, %k0
+; AVX512CD-NEXT: kmovw %k0, %eax
+; AVX512CD-NEXT: vpxor %xmm0, %xmm0, %xmm0
+; AVX512CD-NEXT: vpinsrw $0, %eax, %xmm0, %xmm0
+; AVX512CD-NEXT: vpinsrw $2, %eax, %xmm0, %xmm0
+; AVX512CD-NEXT: vpinsrw $4, %eax, %xmm0, %xmm0
+; AVX512CD-NEXT: vpinsrw $6, %eax, %xmm0, %xmm0
+; AVX512CD-NEXT: vinserti128 $1, %xmm0, %ymm0, %ymm0
+; AVX512CD-NEXT: retq
+;
+; AVX512VLCDBW-LABEL: test_mm256_epi32:
+; AVX512VLCDBW: # BB#0: # %entry
+; AVX512VLCDBW-NEXT: vpcmpeqw %ymm1, %ymm0, %k0
+; AVX512VLCDBW-NEXT: vpbroadcastmw2d %k0, %ymm0
+; AVX512VLCDBW-NEXT: retq
+;
+; X86-AVX512VLCDBW-LABEL: test_mm256_epi32:
+; X86-AVX512VLCDBW: # BB#0: # %entry
+; X86-AVX512VLCDBW-NEXT: vpcmpeqw %ymm1, %ymm0, %k0
+; X86-AVX512VLCDBW-NEXT: vpbroadcastmw2d %k0, %ymm0
+; X86-AVX512VLCDBW-NEXT: retl
+entry:
+ %0 = icmp eq <16 x i16> %a, %b
+ %1 = bitcast <16 x i1> %0 to i16
+ %conv.i = zext i16 %1 to i32
+ %vecinit.i.i = insertelement <8 x i32> undef, i32 %conv.i, i32 0
+ %vecinit7.i.i = shufflevector <8 x i32> %vecinit.i.i, <8 x i32> undef, <8 x i32> zeroinitializer
+ ret <8 x i32> %vecinit7.i.i
+}
+