[llvm] 6ca9676 - [X86] Disable commuting for the first source operand of zero masked scalar fma intrinsic instructions.
Craig Topper via llvm-commits
llvm-commits at lists.llvm.org
Wed Mar 4 14:36:14 PST 2020
Author: Craig Topper
Date: 2020-03-04T14:35:53-08:00
New Revision: 6ca96765c7e6f63b45e6c311918a648ef684ea20
URL: https://github.com/llvm/llvm-project/commit/6ca96765c7e6f63b45e6c311918a648ef684ea20
DIFF: https://github.com/llvm/llvm-project/commit/6ca96765c7e6f63b45e6c311918a648ef684ea20.diff
LOG: [X86] Disable commuting for the first source operand of zero masked scalar fma intrinsic instructions.
I believe this is the correct fix for D75506 rather than disabling all commuting. We can still commute the remaining two sources.
Differential Revision:m https://reviews.llvm.org/D75526
Added:
Modified:
llvm/lib/Target/X86/X86InstrInfo.cpp
llvm/test/CodeGen/X86/avx512-intrinsics.ll
Removed:
################################################################################
diff --git a/llvm/lib/Target/X86/X86InstrInfo.cpp b/llvm/lib/Target/X86/X86InstrInfo.cpp
index a133641dc6a1..7a81198fc9ba 100644
--- a/llvm/lib/Target/X86/X86InstrInfo.cpp
+++ b/llvm/lib/Target/X86/X86InstrInfo.cpp
@@ -1883,7 +1883,7 @@ X86InstrInfo::findThreeSrcCommutedOpIndices(const MachineInstr &MI,
unsigned KMaskOp = -1U;
if (X86II::isKMasked(TSFlags)) {
// For k-zero-masked operations it is Ok to commute the first vector
- // operand.
+ // operand. Unless this is an intrinsic instruction.
// For regular k-masked operations a conservative choice is done as the
// elements of the first vector operand, for which the corresponding bit
// in the k-mask operand is set to 0, are copied to the result of the
@@ -1902,7 +1902,7 @@ X86InstrInfo::findThreeSrcCommutedOpIndices(const MachineInstr &MI,
// The operand with index = 1 is used as a source for those elements for
// which the corresponding bit in the k-mask is set to 0.
- if (X86II::isKMergeMasked(TSFlags))
+ if (X86II::isKMergeMasked(TSFlags) || IsIntrinsic)
FirstCommutableVecOp = 3;
LastCommutableVecOp++;
diff --git a/llvm/test/CodeGen/X86/avx512-intrinsics.ll b/llvm/test/CodeGen/X86/avx512-intrinsics.ll
index a089e1ee99af..8844c0e5e95f 100644
--- a/llvm/test/CodeGen/X86/avx512-intrinsics.ll
+++ b/llvm/test/CodeGen/X86/avx512-intrinsics.ll
@@ -5818,6 +5818,35 @@ define <4 x float>@test_int_x86_avx512_maskz_vfmadd_ss(<4 x float> %x0, <4 x flo
ret <4 x float> %res2
}
+; Make sure we don't commute this to fold the load as that source isn't commutable.
+define <4 x float> @test_int_x86_avx512_maskz_vfmadd_ss_load0(i8 zeroext %0, <4 x float>* nocapture readonly %1, float %2, float %3) {
+; X64-LABEL: test_int_x86_avx512_maskz_vfmadd_ss_load0:
+; X64: # %bb.0:
+; X64-NEXT: vmovaps (%rsi), %xmm2
+; X64-NEXT: kmovw %edi, %k1
+; X64-NEXT: vfmadd213ss {{.*#+}} xmm2 = (xmm0 * xmm2) + xmm1
+; X64-NEXT: vmovaps %xmm2, %xmm0
+; X64-NEXT: retq
+;
+; X86-LABEL: test_int_x86_avx512_maskz_vfmadd_ss_load0:
+; X86: # %bb.0:
+; X86-NEXT: vmovss {{.*#+}} xmm1 = mem[0],zero,zero,zero
+; X86-NEXT: movb {{[0-9]+}}(%esp), %al
+; X86-NEXT: movl {{[0-9]+}}(%esp), %ecx
+; X86-NEXT: vmovaps (%ecx), %xmm0
+; X86-NEXT: kmovw %eax, %k1
+; X86-NEXT: vfmadd132ss {{.*#+}} xmm0 = (xmm0 * mem) + xmm1
+; X86-NEXT: retl
+ %5 = load <4 x float>, <4 x float>* %1, align 16
+ %6 = extractelement <4 x float> %5, i64 0
+ %7 = tail call float @llvm.fma.f32(float %6, float %2, float %3) #2
+ %8 = bitcast i8 %0 to <8 x i1>
+ %9 = extractelement <8 x i1> %8, i64 0
+ %10 = select i1 %9, float %7, float 0.000000e+00
+ %11 = insertelement <4 x float> %5, float %10, i64 0
+ ret <4 x float> %11
+}
+
define <2 x double>@test_int_x86_avx512_mask3_vfmadd_sd(<2 x double> %x0, <2 x double> %x1, <2 x double> %x2, i8 %x3,i32 %x4 ){
; X64-LABEL: test_int_x86_avx512_mask3_vfmadd_sd:
; X64: # %bb.0:
More information about the llvm-commits
mailing list