This patch, together with a matching clang patch (https://reviews.llvm.org/D38683), implements the lowering of X86 broadcastm intrinsics to IR.
Differential Revision: https://reviews.llvm.org/D38684
Change-Id: I709ac0b34641095397e994c8ff7e15d1315b3540
llvm-svn: 317458
Name=="ssse3.pabs.d.128" || // Added in 6.0
Name.startswith("avx2.pabs.") || // Added in 6.0
Name.startswith("avx512.mask.pabs.") || // Added in 6.0
+ Name.startswith("avx512.broadcastm") || // Added in 6.0
Name.startswith("avx512.mask.pbroadcast") || // Added in 6.0
Name.startswith("sse2.pcmpeq.") || // Added in 3.1
Name.startswith("sse2.pcmpgt.") || // Added in 3.1
Rep = Builder.CreateICmp(CmpEq ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_SGT,
CI->getArgOperand(0), CI->getArgOperand(1));
Rep = Builder.CreateSExt(Rep, CI->getType(), "");
- } else if (IsX86 && (Name.startswith("avx512.mask.pbroadcast"))){
+ } else if (IsX86 && (Name.startswith("avx512.broadcastm"))) {
+ Type *ExtTy = Type::getInt32Ty(C);
+ if (CI->getOperand(0)->getType()->isIntegerTy(8))
+ ExtTy = Type::getInt64Ty(C);
+ unsigned NumElts = CI->getType()->getPrimitiveSizeInBits() /
+ ExtTy->getPrimitiveSizeInBits();
+ Rep = Builder.CreateZExt(CI->getArgOperand(0), ExtTy);
+ Rep = Builder.CreateVectorSplat(NumElts, Rep);
+ } else if (IsX86 && (Name.startswith("avx512.mask.pbroadcast"))) {
unsigned NumElts =
CI->getArgOperand(1)->getType()->getVectorNumElements();
Rep = Builder.CreateVectorSplat(NumElts, CI->getArgOperand(0));
assert((VT.is128BitVector() || VT.is256BitVector() || VT.is512BitVector()) &&
"Unsupported vector type for broadcast.");
+ BitVector UndefElements;
+ SDValue Ld = BVOp->getSplatValue(&UndefElements);
+
// Attempt to use VBROADCASTM
// From this paterrn:
// a. t0 = (zext_i64 (bitcast_i8 v2i1 X))
//
// Create (VBROADCASTM v2i1 X)
if (Subtarget.hasCDI() && (VT.is512BitVector() || Subtarget.hasVLX())) {
- MVT EltType;
- unsigned NumElts;
+ MVT EltType = VT.getScalarType();
+ unsigned NumElts = VT.getVectorNumElements();
+ SDValue BOperand;
SDValue ZeroExtended = isSplatZeroExtended(BVOp, NumElts, EltType);
- if (ZeroExtended && ZeroExtended.getOpcode() == ISD::BITCAST) {
- SDValue BOperand = ZeroExtended.getOperand(0);
+ if ((ZeroExtended && ZeroExtended.getOpcode() == ISD::BITCAST) ||
+ (Ld && Ld.getOpcode() == ISD::ZERO_EXTEND &&
+ Ld.getOperand(0).getOpcode() == ISD::BITCAST)) {
+ if (ZeroExtended)
+ BOperand = ZeroExtended.getOperand(0);
+ else
+ BOperand = Ld.getOperand(0).getOperand(0);
if (BOperand.getValueType().isVector() &&
BOperand.getSimpleValueType().getVectorElementType() == MVT::i1) {
- if ((EltType == MVT::i64 &&
- VT.getVectorElementType() == MVT::i8) || // for broadcastmb2q
- (EltType == MVT::i32 &&
- VT.getVectorElementType() == MVT::i16)) { // for broadcastmw2d
+ if ((EltType == MVT::i64 && (VT.getVectorElementType() == MVT::i8 ||
+ NumElts == 8)) || // for broadcastmb2q
+ (EltType == MVT::i32 && (VT.getVectorElementType() == MVT::i16 ||
+ NumElts == 16))) { // for broadcastmw2d
SDValue Brdcst =
DAG.getNode(X86ISD::VBROADCASTM, dl,
MVT::getVectorVT(EltType, NumElts), BOperand);
}
}
- BitVector UndefElements;
- SDValue Ld = BVOp->getSplatValue(&UndefElements);
-
// We need a splat of a single value to use broadcast, and it doesn't
// make any sense if the value is only in one element of the vector.
if (!Ld || (VT.getVectorNumElements() - UndefElements.count()) <= 1) {
X86_INTRINSIC_DATA(avx2_psubs_w, INTR_TYPE_2OP, X86ISD::SUBS, 0),
X86_INTRINSIC_DATA(avx2_psubus_b, INTR_TYPE_2OP, X86ISD::SUBUS, 0),
X86_INTRINSIC_DATA(avx2_psubus_w, INTR_TYPE_2OP, X86ISD::SUBUS, 0),
- X86_INTRINSIC_DATA(avx512_broadcastmb_128, BROADCASTM, X86ISD::VBROADCASTM, 0),
- X86_INTRINSIC_DATA(avx512_broadcastmb_256, BROADCASTM, X86ISD::VBROADCASTM, 0),
- X86_INTRINSIC_DATA(avx512_broadcastmb_512, BROADCASTM, X86ISD::VBROADCASTM, 0),
- X86_INTRINSIC_DATA(avx512_broadcastmw_128, BROADCASTM, X86ISD::VBROADCASTM, 0),
- X86_INTRINSIC_DATA(avx512_broadcastmw_256, BROADCASTM, X86ISD::VBROADCASTM, 0),
- X86_INTRINSIC_DATA(avx512_broadcastmw_512, BROADCASTM, X86ISD::VBROADCASTM, 0),
X86_INTRINSIC_DATA(avx512_cvtb2mask_128, CONVERT_TO_MASK, X86ISD::CVT2MASK, 0),
X86_INTRINSIC_DATA(avx512_cvtb2mask_256, CONVERT_TO_MASK, X86ISD::CVT2MASK, 0),
X86_INTRINSIC_DATA(avx512_cvtb2mask_512, CONVERT_TO_MASK, X86ISD::CVT2MASK, 0),
--- /dev/null
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc < %s -fast-isel -mtriple=x86_64-unknown-unknown -mattr=+avx512cd | FileCheck %s
+
+define <8 x i64> @test_mm512_broadcastmb_epi64(<8 x i64> %a, <8 x i64> %b) {
+; CHECK-LABEL: test_mm512_broadcastmb_epi64:
+; CHECK: # BB#0: # %entry
+; CHECK-NEXT: vpcmpeqq %zmm1, %zmm0, %k0
+; CHECK-NEXT: vpbroadcastmb2q %k0, %zmm0
+; CHECK-NEXT: retq
+entry:
+ %0 = icmp eq <8 x i64> %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 <8 x i64> @test_mm512_broadcastmw_epi32(<8 x i64> %a, <8 x i64> %b) {
+; CHECK-LABEL: test_mm512_broadcastmw_epi32:
+; CHECK: # BB#0: # %entry
+; CHECK-NEXT: vpcmpeqd %zmm1, %zmm0, %k0
+; CHECK-NEXT: vpbroadcastmw2d %k0, %zmm0
+; CHECK-NEXT: retq
+entry:
+ %0 = bitcast <8 x i64> %a to <16 x i32>
+ %1 = bitcast <8 x i64> %b to <16 x i32>
+ %2 = icmp eq <16 x i32> %0, %1
+ %3 = bitcast <16 x i1> %2 to i16
+ %conv.i = zext i16 %3 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
+ %4 = bitcast <16 x i32> %vecinit15.i.i to <8 x i64>
+ ret <8 x i64> %4
+}
+
+
%res = call <8 x i64> @llvm.x86.avx512.mask.lzcnt.q.512(<8 x i64> %a, <8 x i64> %b, i8 %mask)
ret <8 x i64> %res
}
+
+define <16 x i32> @test_x86_vbroadcastmw_512(i16 %a0) {
+; CHECK-LABEL: test_x86_vbroadcastmw_512:
+; CHECK: ## BB#0:
+; CHECK-NEXT: movzwl %di, %eax
+; CHECK-NEXT: vpbroadcastd %eax, %zmm0
+; CHECK-NEXT: retq
+ %res = call <16 x i32> @llvm.x86.avx512.broadcastmw.512(i16 %a0)
+ ret <16 x i32> %res
+}
+declare <16 x i32> @llvm.x86.avx512.broadcastmw.512(i16)
+
+define <8 x i64> @test_x86_broadcastmb_512(i8 %a0) {
+; CHECK-LABEL: test_x86_broadcastmb_512:
+; CHECK: ## BB#0:
+; CHECK-NEXT: movzbl %dil, %eax
+; CHECK-NEXT: vpbroadcastq %rax, %zmm0
+; CHECK-NEXT: retq
+ %res = call <8 x i64> @llvm.x86.avx512.broadcastmb.512(i8 %a0)
+ ret <8 x i64> %res
+}
+declare <8 x i64> @llvm.x86.avx512.broadcastmb.512(i8)
+
; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
; RUN: llc < %s -mtriple=x86_64-apple-darwin -mcpu=knl -mattr=+avx512cd | FileCheck %s
-define <16 x i32> @test_x86_vbroadcastmw_512(i16 %a0) {
-; CHECK-LABEL: test_x86_vbroadcastmw_512:
-; CHECK: ## BB#0:
-; CHECK-NEXT: kmovw %edi, %k0
-; CHECK-NEXT: vpbroadcastmw2d %k0, %zmm0
-; CHECK-NEXT: retq
- %res = call <16 x i32> @llvm.x86.avx512.broadcastmw.512(i16 %a0)
- ret <16 x i32> %res
-}
-declare <16 x i32> @llvm.x86.avx512.broadcastmw.512(i16)
-
-define <8 x i64> @test_x86_broadcastmb_512(i8 %a0) {
-; CHECK-LABEL: test_x86_broadcastmb_512:
-; CHECK: ## BB#0:
-; CHECK-NEXT: kmovw %edi, %k0
-; CHECK-NEXT: vpbroadcastmb2q %k0, %zmm0
-; CHECK-NEXT: retq
- %res = call <8 x i64> @llvm.x86.avx512.broadcastmb.512(i8 %a0)
- ret <8 x i64> %res
-}
-declare <8 x i64> @llvm.x86.avx512.broadcastmb.512(i8)
-
declare <16 x i32> @llvm.x86.avx512.mask.conflict.d.512(<16 x i32>, <16 x i32>, i16) nounwind readonly
define <8 x i64> @test_conflict_q(<8 x i64> %a) {
ret <4 x i64> %res2
}
+define <8 x i32> @test_x86_vbroadcastmw_256(i16 %a0) {
+; CHECK-LABEL: test_x86_vbroadcastmw_256:
+; CHECK: ## BB#0:
+; CHECK-NEXT: movzwl %di, %eax
+; CHECK-NEXT: vpbroadcastd %eax, %ymm0
+; CHECK-NEXT: retq
+ %res = call <8 x i32> @llvm.x86.avx512.broadcastmw.256(i16 %a0) ;
+ ret <8 x i32> %res
+}
+declare <8 x i32> @llvm.x86.avx512.broadcastmw.256(i16)
+
+define <4 x i32> @test_x86_vbroadcastmw_128(i16 %a0) {
+; CHECK-LABEL: test_x86_vbroadcastmw_128:
+; CHECK: ## BB#0:
+; CHECK-NEXT: movzwl %di, %eax
+; CHECK-NEXT: vpbroadcastd %eax, %xmm0
+; CHECK-NEXT: retq
+ %res = call <4 x i32> @llvm.x86.avx512.broadcastmw.128(i16 %a0) ;
+ ret <4 x i32> %res
+}
+declare <4 x i32> @llvm.x86.avx512.broadcastmw.128(i16)
+
+define <4 x i64> @test_x86_broadcastmb_256(i8 %a0) {
+; CHECK-LABEL: test_x86_broadcastmb_256:
+; CHECK: ## BB#0:
+; CHECK-NEXT: movzbl %dil, %eax
+; CHECK-NEXT: vpbroadcastq %rax, %ymm0
+; CHECK-NEXT: retq
+ %res = call <4 x i64> @llvm.x86.avx512.broadcastmb.256(i8 %a0) ;
+ ret <4 x i64> %res
+}
+declare <4 x i64> @llvm.x86.avx512.broadcastmb.256(i8)
+
+define <2 x i64> @test_x86_broadcastmb_128(i8 %a0) {
+; CHECK-LABEL: test_x86_broadcastmb_128:
+; CHECK: ## BB#0:
+; CHECK-NEXT: movzbl %dil, %eax
+; CHECK-NEXT: vpbroadcastq %rax, %xmm0
+; CHECK-NEXT: retq
+ %res = call <2 x i64> @llvm.x86.avx512.broadcastmb.128(i8 %a0) ;
+ ret <2 x i64> %res
+}
+declare <2 x i64> @llvm.x86.avx512.broadcastmb.128(i8)
+
ret <4 x i64> %res2
}
-define <8 x i32> @test_x86_vbroadcastmw_256(i16 %a0) {
-; CHECK-LABEL: test_x86_vbroadcastmw_256:
-; CHECK: ## BB#0:
-; CHECK-NEXT: kmovw %edi, %k0
-; CHECK-NEXT: vpbroadcastmw2d %k0, %ymm0
-; CHECK-NEXT: retq
- %res = call <8 x i32> @llvm.x86.avx512.broadcastmw.256(i16 %a0) ;
- ret <8 x i32> %res
-}
-declare <8 x i32> @llvm.x86.avx512.broadcastmw.256(i16)
-
-define <4 x i32> @test_x86_vbroadcastmw_128(i16 %a0) {
-; CHECK-LABEL: test_x86_vbroadcastmw_128:
-; CHECK: ## BB#0:
-; CHECK-NEXT: kmovw %edi, %k0
-; CHECK-NEXT: vpbroadcastmw2d %k0, %xmm0
-; CHECK-NEXT: retq
- %res = call <4 x i32> @llvm.x86.avx512.broadcastmw.128(i16 %a0) ;
- ret <4 x i32> %res
-}
-declare <4 x i32> @llvm.x86.avx512.broadcastmw.128(i16)
-
-define <4 x i64> @test_x86_broadcastmb_256(i8 %a0) {
-; CHECK-LABEL: test_x86_broadcastmb_256:
-; CHECK: ## BB#0:
-; CHECK-NEXT: kmovw %edi, %k0
-; CHECK-NEXT: vpbroadcastmb2q %k0, %ymm0
-; CHECK-NEXT: retq
- %res = call <4 x i64> @llvm.x86.avx512.broadcastmb.256(i8 %a0) ;
- ret <4 x i64> %res
-}
-declare <4 x i64> @llvm.x86.avx512.broadcastmb.256(i8)
-
-define <2 x i64> @test_x86_broadcastmb_128(i8 %a0) {
-; CHECK-LABEL: test_x86_broadcastmb_128:
-; CHECK: ## BB#0:
-; CHECK-NEXT: kmovw %edi, %k0
-; CHECK-NEXT: vpbroadcastmb2q %k0, %xmm0
-; CHECK-NEXT: retq
- %res = call <2 x i64> @llvm.x86.avx512.broadcastmb.128(i8 %a0) ;
- ret <2 x i64> %res
-}
-declare <2 x i64> @llvm.x86.avx512.broadcastmb.128(i8)
--- /dev/null
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc < %s -fast-isel -mtriple=x86_64-unknown-unknown -mattr=+avx512cd,+avx512vl | FileCheck %s
+
+define <2 x i64> @test_mm_broadcastmb_epi64(<2 x i64> %a, <2 x i64> %b) {
+; CHECK-LABEL: test_mm_broadcastmb_epi64:
+; CHECK: # BB#0: # %entry
+; CHECK-NEXT: vpcmpeqd %xmm1, %xmm0, %k0
+; CHECK-NEXT: vpbroadcastmb2q %k0, %xmm0
+; CHECK-NEXT: retq
+entry:
+ %0 = bitcast <2 x i64> %a to <4 x i32>
+ %1 = bitcast <2 x i64> %b to <4 x i32>
+ %2 = icmp eq <4 x i32> %0, %1
+ %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
+ %4 = bitcast <8 x i1> %3 to i8
+ %conv.i = zext i8 %4 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 i64> @test_mm256_broadcastmb_epi64(<4 x i64> %a, <4 x i64> %b) {
+; CHECK-LABEL: test_mm256_broadcastmb_epi64:
+; CHECK: # BB#0: # %entry
+; CHECK-NEXT: vpcmpeqq %ymm1, %ymm0, %k0
+; CHECK-NEXT: vpbroadcastmb2q %k0, %ymm0
+; CHECK-NEXT: retq
+entry:
+ %0 = icmp eq <4 x i64> %a, %b
+ %1 = shufflevector <4 x i1> %0, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
+ %2 = bitcast <8 x i1> %1 to i8
+ %conv.i = zext i8 %2 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 <2 x i64> @test_mm_broadcastmw_epi32(<8 x i64> %a, <8 x i64> %b) {
+; CHECK-LABEL: test_mm_broadcastmw_epi32:
+; CHECK: # BB#0: # %entry
+; CHECK-NEXT: vpcmpeqd %zmm1, %zmm0, %k0
+; CHECK-NEXT: vpbroadcastmw2d %k0, %xmm0
+; CHECK-NEXT: vzeroupper
+; CHECK-NEXT: retq
+entry:
+ %0 = bitcast <8 x i64> %a to <16 x i32>
+ %1 = bitcast <8 x i64> %b to <16 x i32>
+ %2 = icmp eq <16 x i32> %0, %1
+ %3 = bitcast <16 x i1> %2 to i16
+ %conv.i = zext i16 %3 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
+ %4 = bitcast <4 x i32> %vecinit3.i.i to <2 x i64>
+ ret <2 x i64> %4
+}
+
+define <4 x i64> @test_mm256_broadcastmw_epi32(<8 x i64> %a, <8 x i64> %b) {
+; CHECK-LABEL: test_mm256_broadcastmw_epi32:
+; CHECK: # BB#0: # %entry
+; CHECK-NEXT: vpcmpeqd %zmm1, %zmm0, %k0
+; CHECK-NEXT: vpbroadcastmw2d %k0, %ymm0
+; CHECK-NEXT: retq
+entry:
+ %0 = bitcast <8 x i64> %a to <16 x i32>
+ %1 = bitcast <8 x i64> %b to <16 x i32>
+ %2 = icmp eq <16 x i32> %0, %1
+ %3 = bitcast <16 x i1> %2 to i16
+ %conv.i = zext i16 %3 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
+ %4 = bitcast <8 x i32> %vecinit7.i.i to <4 x i64>
+ ret <4 x i64> %4
+}
+
+
; 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: vpbroadcastmw2d %k0, %zmm0
; AVX512CD-NEXT: retq
;
; AVX512VLCDBW-LABEL: test_mm512_epi32:
; 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: vpbroadcastmb2q %k0, %zmm0
; AVX512CD-NEXT: retq
;
; AVX512VLCDBW-LABEL: test_mm512_epi64: