[PATCH] D50332: AMDGPU: Turn class x, p_zero|n_zero into fcmp oeq x, 0

[Matt Arsenault via Phabricator via llvm-commits]

arsenm created this revision.
arsenm added reviewers: rampitec, b-sumner.
Herald added subscribers: t-tye, tpr, dstuttard, yaxunl, nhaehnle, wdng, kzhuravl.

The library does use this for some reason


https://reviews.llvm.org/D50332

Files:
  lib/Transforms/InstCombine/InstCombineCalls.cpp
  test/Transforms/InstCombine/AMDGPU/amdgcn-intrinsics.ll


Index: test/Transforms/InstCombine/AMDGPU/amdgcn-intrinsics.ll
===================================================================
--- test/Transforms/InstCombine/AMDGPU/amdgcn-intrinsics.ll
+++ test/Transforms/InstCombine/AMDGPU/amdgcn-intrinsics.ll
@@ -457,6 +457,13 @@
   ret i1 %val
 }
 
+; CHECK-LABEL: @test_class_is_p0_n0_f32(
+; CHECK: %val = fcmp oeq float %x, 0.000000e+00
+define i1 @test_class_is_p0_n0_f32(float %x) nounwind {
+  %val = call i1 @llvm.amdgcn.class.f32(float %x, i32 96)
+  ret i1 %val
+}
+
 ; CHECK-LABEL: @test_constant_class_snan_test_snan_f64(
 ; CHECK: ret i1 true
 define i1 @test_constant_class_snan_test_snan_f64() nounwind {
Index: lib/Transforms/InstCombine/InstCombineCalls.cpp
===================================================================
--- lib/Transforms/InstCombine/InstCombineCalls.cpp
+++ lib/Transforms/InstCombine/InstCombineCalls.cpp
@@ -3260,6 +3260,15 @@
       return replaceInstUsesWith(*II, FCmp);
     }
 
+    if (Mask == (N_ZERO | P_ZERO)) {
+      // Equivalent of == 0.
+      Value *FCmp = Builder.CreateFCmpOEQ(
+        Src0, ConstantFP::get(Src0->getType(), 0.0));
+
+      FCmp->takeName(II);
+      return replaceInstUsesWith(*II, FCmp);
+    }
+
     const ConstantFP *CVal = dyn_cast<ConstantFP>(Src0);
     if (!CVal) {
       if (isa<UndefValue>(Src0))


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D50332.159290.patch
Type: text/x-patch
Size: 1328 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20180806/5058f16c/attachment.bin>