[PATCH] D75526: [X86] Disable commuting for the first source operand of zero masked scalar fma intrinsic instructions.

[Craig Topper via Phabricator via llvm-commits]

craig.topper created this revision.
craig.topper added reviewers: RKSimon, spatel, LiuChen3.
Herald added a subscriber: hiraditya.
Herald added a project: LLVM.

I believe this is the correct fix for D75506 <https://reviews.llvm.org/D75506> rather than disabling all commuting. We can still commute the remaining two sources.


Repository:
  rG LLVM Github Monorepo

https://reviews.llvm.org/D75526

Files:
  llvm/lib/Target/X86/X86InstrInfo.cpp
  llvm/test/CodeGen/X86/avx512-intrinsics.ll


Index: llvm/test/CodeGen/X86/avx512-intrinsics.ll
===================================================================
--- llvm/test/CodeGen/X86/avx512-intrinsics.ll
+++ llvm/test/CodeGen/X86/avx512-intrinsics.ll
@@ -5812,6 +5812,37 @@
   ret <4 x float> %8
 }
 
+; 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, <4 x float> %2, <4 x 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:    movb {{[0-9]+}}(%esp), %al
+; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
+; X86-NEXT:    vmovaps (%ecx), %xmm2
+; X86-NEXT:    kmovw %eax, %k1
+; X86-NEXT:    vfmadd213ss {{.*#+}} xmm2 = (xmm0 * xmm2) + xmm1
+; X86-NEXT:    vmovaps %xmm2, %xmm0
+; X86-NEXT:    retl
+  %5 = load <4 x float>, <4 x float>* %1, align 16
+  %6 = extractelement <4 x float> %5, i64 0
+  %7 = extractelement <4 x float> %2, i64 0
+  %8 = extractelement <4 x float> %3, i64 0
+  %9 = tail call float @llvm.fma.f32(float %6, float %7, float %8) #2
+  %10 = bitcast i8 %0 to <8 x i1>
+  %11 = extractelement <8 x i1> %10, i64 0
+  %12 = select i1 %11, float %9, float 0.000000e+00
+  %13 = insertelement <4 x float> %5, float %12, i64 0
+  ret <4 x float> %13
+}
+
 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:
Index: llvm/lib/Target/X86/X86InstrInfo.cpp
===================================================================
--- llvm/lib/Target/X86/X86InstrInfo.cpp
+++ llvm/lib/Target/X86/X86InstrInfo.cpp
@@ -1883,7 +1883,7 @@
   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 @@
 
     // 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++;


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D75526.247902.patch
Type: text/x-patch
Size: 2791 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20200303/cd5484dc/attachment.bin>