[PATCH] D74835: [AMDGPU][ConstantFolding] Fold llvm.amdgcn.fmul.legacy intrinsic

[Jay Foad via Phabricator via llvm-commits]

foad created this revision.
foad added reviewers: arsenm, rampitec, nhaehnle.
Herald added subscribers: llvm-commits, kerbowa, hiraditya, t-tye, tpr, dstuttard, yaxunl, wdng, jvesely, kzhuravl.
Herald added a project: LLVM.

Repository:
  rG LLVM Github Monorepo

https://reviews.llvm.org/D74835

Files:
  llvm/lib/Analysis/ConstantFolding.cpp
  llvm/test/Transforms/ConstProp/amdgpu.ll


Index: llvm/test/Transforms/ConstProp/amdgpu.ll
===================================================================
--- /dev/null
+++ llvm/test/Transforms/ConstProp/amdgpu.ll
@@ -0,0 +1,45 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; RUN: opt < %s -constprop -S | FileCheck %s
+; REQUIRES: amdgpu-registered-target
+
+declare float @llvm.amdgcn.fmul.legacy(float, float)
+
+define void @test(float* %p) {
+; CHECK-LABEL: @test(
+; CHECK-NEXT:    store volatile float 6.000000e+00, float* [[P:%.*]]
+; CHECK-NEXT:    store volatile float 0.000000e+00, float* [[P]]
+; CHECK-NEXT:    store volatile float 0.000000e+00, float* [[P]]
+; CHECK-NEXT:    store volatile float 0.000000e+00, float* [[P]]
+; CHECK-NEXT:    store volatile float 0.000000e+00, float* [[P]]
+; CHECK-NEXT:    store volatile float 0.000000e+00, float* [[P]]
+; CHECK-NEXT:    store volatile float 0.000000e+00, float* [[P]]
+; CHECK-NEXT:    store volatile float 0.000000e+00, float* [[P]]
+; CHECK-NEXT:    store volatile float 0.000000e+00, float* [[P]]
+; CHECK-NEXT:    store volatile float 0.000000e+00, float* [[P]]
+; CHECK-NEXT:    store volatile float 0.000000e+00, float* [[P]]
+; CHECK-NEXT:    ret void
+;
+  %a = call float @llvm.amdgcn.fmul.legacy(float +2.0, float +3.0)
+  store volatile float %a, float* %p
+  %b = call float @llvm.amdgcn.fmul.legacy(float +2.0, float +0.0)
+  store volatile float %b, float* %p
+  %c = call float @llvm.amdgcn.fmul.legacy(float +2.0, float -0.0)
+  store volatile float %c, float* %p
+  %d = call float @llvm.amdgcn.fmul.legacy(float +0.0, float +0.0)
+  store volatile float %d, float* %p
+  %e = call float @llvm.amdgcn.fmul.legacy(float +0.0, float -0.0)
+  store volatile float %e, float* %p
+  %f = call float @llvm.amdgcn.fmul.legacy(float -0.0, float +0.0)
+  store volatile float %f, float* %p
+  %g = call float @llvm.amdgcn.fmul.legacy(float -0.0, float -0.0)
+  store volatile float %g, float* %p
+  %h = call float @llvm.amdgcn.fmul.legacy(float +0.0, float 0x7ff0000000000000) ; +inf
+  store volatile float %h, float* %p
+  %i = call float @llvm.amdgcn.fmul.legacy(float 0xfff0000000000000, float +0.0) ; -inf
+  store volatile float %i, float* %p
+  %j = call float @llvm.amdgcn.fmul.legacy(float 0x7ff0001000000000, float -0.0) ; +nan
+  store volatile float %j, float* %p
+  %k = call float @llvm.amdgcn.fmul.legacy(float -0.0, float 0xfff0000100000000) ; -nan
+  store volatile float %k, float* %p
+  ret void
+}
Index: llvm/lib/Analysis/ConstantFolding.cpp
===================================================================
--- llvm/lib/Analysis/ConstantFolding.cpp
+++ llvm/lib/Analysis/ConstantFolding.cpp
@@ -39,6 +39,7 @@
 #include "llvm/IR/Instruction.h"
 #include "llvm/IR/Instructions.h"
 #include "llvm/IR/Intrinsics.h"
+#include "llvm/IR/IntrinsicsAMDGPU.h"
 #include "llvm/IR/IntrinsicsX86.h"
 #include "llvm/IR/Operator.h"
 #include "llvm/IR/Type.h"
@@ -1458,6 +1459,7 @@
   case Intrinsic::convert_from_fp16:
   case Intrinsic::convert_to_fp16:
   case Intrinsic::bitreverse:
+  case Intrinsic::amdgcn_fmul_legacy:
   case Intrinsic::x86_sse_cvtss2si:
   case Intrinsic::x86_sse_cvtss2si64:
   case Intrinsic::x86_sse_cvttss2si:
@@ -2082,6 +2084,16 @@
         return ConstantFP::get(Ty->getContext(), maximum(C1, C2));
       }
 
+      if (IntrinsicID == Intrinsic::amdgcn_fmul_legacy) {
+        const APFloat &C1 = Op1->getValueAPF();
+        const APFloat &C2 = Op2->getValueAPF();
+        // The legacy behaviour is that multiplying zero by anything, even NaN
+        // or infinity, gives +0.0.
+        if (C1.isZero() || C2.isZero())
+          return ConstantFP::getNullValue(Ty);
+        return ConstantFP::get(Ty->getContext(), C1 * C2);
+      }
+
       if (!TLI)
         return nullptr;
 


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D74835.245402.patch
Type: text/x-patch
Size: 3799 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20200219/629642b5/attachment.bin>