AMDGPU: Turn class x, p_zero|n_zero into fcmp oeq x, 0

[arsenm]

Summary

The library does use this for some reason

[rampitec]

lib/Transforms/InstCombine/InstCombineCalls.cpp

3266

Is it really true for -0?

[arsenm]

lib/Transforms/InstCombine/InstCombineCalls.cpp

3266

Yes, -0.0 and 0.0 compare equal

[rampitec]

LGTM

[arsenm]

r339461

lib/Transforms/InstCombine/InstCombineCalls.cpp

case Intrinsic::amdgcn_class: {
 
     if (Mask == (S_NAN | Q_NAN)) {
       // Equivalent of isnan. Replace with standard fcmp.
       Value *FCmp = Builder.CreateFCmpUNO(Src0, Src0);
       FCmp->takeName(II);
       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))
         return replaceInstUsesWith(*II, UndefValue::get(II->getType()));
 
       // Clamp mask to used bits
       if ((Mask & FullMask) != Mask) {
         CallInst *NewCall = Builder.CreateCall(II->getCalledFunction(),

test/Transforms/InstCombine/AMDGPU/amdgcn-intrinsics.ll

 
 ; CHECK-LABEL: @test_class_isnan_f32(
 ; CHECK: %val = fcmp uno float %x, 0.000000e+00
 define i1 @test_class_isnan_f32(float %x) nounwind {
   %val = call i1 @llvm.amdgcn.class.f32(float %x, i32 3)
   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 {
   %val = call i1 @llvm.amdgcn.class.f64(double 0x7FF0000000000001, i32 1)
   ret i1 %val
 }
 
 ; CHECK-LABEL: @test_constant_class_qnan_test_qnan_f64(