AMDGPU: Perform basic folds on llvm.amdgcn.exp2

[arsenm]

c6aaa0b14f2ab7ed764cd494a605f30a9e02c40b

llvm/lib/Target/AMDGPU/AMDGPUInstCombineIntrinsic.cpp

case Intrinsic::amdgcn_rsq: {
     if (isa<UndefValue>(Src)) {
       Type *Ty = II.getType();
       auto *QNaN = ConstantFP::get(Ty, APFloat::getQNaN(Ty->getFltSemantics()));
       return IC.replaceInstUsesWith(II, QNaN);
     }
 
     break;
   }
-  case Intrinsic::amdgcn_log: {
+  case Intrinsic::amdgcn_log:
+  case Intrinsic::amdgcn_exp2: {
+    const bool IsLog = IID == Intrinsic::amdgcn_log;
+    const bool IsExp = IID == Intrinsic::amdgcn_exp2;
     Value *Src = II.getArgOperand(0);
     Type *Ty = II.getType();
 
     if (isa<PoisonValue>(Src))
       return IC.replaceInstUsesWith(II, Src);
 
     if (IC.getSimplifyQuery().isUndefValue(Src))
       return IC.replaceInstUsesWith(II, ConstantFP::getNaN(Ty));
 
     if (ConstantFP *C = dyn_cast<ConstantFP>(Src)) {
-      if (C->isInfinity() && !C->isNegative())
+      if (C->isInfinity()) {
+        // exp2(+inf) -> +inf
+        // log2(+inf) -> +inf
+        if (!C->isNegative())
           return IC.replaceInstUsesWith(II, C);
 
+        // exp2(-inf) -> 0
+        if (IsExp && C->isNegative())
+          return IC.replaceInstUsesWith(II, ConstantFP::getZero(Ty));
+      }
+
       if (II.isStrictFP())
         break;
 
       if (C->isNaN()) {
         Constant *Quieted = ConstantFP::get(Ty, C->getValue().makeQuiet());
         return IC.replaceInstUsesWith(II, Quieted);
       }
 
       // f32 instruction doesn't handle denormals, f16 does.
-      if (C->isZero() || (C->getValue().isDenormal() && Ty->isFloatTy()))
-        return IC.replaceInstUsesWith(II, ConstantFP::getInfinity(Ty, true));
+      if (C->isZero() || (C->getValue().isDenormal() && Ty->isFloatTy())) {
+        Constant *FoldedValue = IsLog ? ConstantFP::getInfinity(Ty, true)
+                                      : ConstantFP::get(Ty, 1.0);
+        return IC.replaceInstUsesWith(II, FoldedValue);
+      }
 
-      if (C->isNegative())
+      if (IsLog && C->isNegative())
         return IC.replaceInstUsesWith(II, ConstantFP::getNaN(Ty));
 
       // TODO: Full constant folding matching hardware behavior.
     }
 
     break;
   }
   case Intrinsic::amdgcn_frexp_mant:

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

 ; llvm.amdgcn.exp2
 ; --------------------------------------------------------------------
 
 declare float @llvm.amdgcn.exp2.f32(float) nounwind readnone
 declare half @llvm.amdgcn.exp2.f16(half) nounwind readnone
 
 define float @test_constant_fold_exp2_f32_undef() {
 ; CHECK-LABEL: @test_constant_fold_exp2_f32_undef(
-; CHECK-NEXT:    [[VAL:%.*]] = call float @llvm.amdgcn.exp2.f32(float undef)
-; CHECK-NEXT:    ret float [[VAL]]
+; CHECK-NEXT:    ret float 0x7FF8000000000000
 ;
   %val = call float @llvm.amdgcn.exp2.f32(float undef)
   ret float %val
 }
 
 define float @test_constant_fold_exp2_f32_poison() {
 ; CHECK-LABEL: @test_constant_fold_exp2_f32_poison(
-; CHECK-NEXT:    [[VAL:%.*]] = call float @llvm.amdgcn.exp2.f32(float poison)
-; CHECK-NEXT:    ret float [[VAL]]
+; CHECK-NEXT:    ret float poison
 ;
   %val = call float @llvm.amdgcn.exp2.f32(float poison)
   ret float %val
 }
 
 define float @test_constant_fold_exp2_f32_p0() {
 ; CHECK-LABEL: @test_constant_fold_exp2_f32_p0(
-; CHECK-NEXT:    [[VAL:%.*]] = call float @llvm.amdgcn.exp2.f32(float 0.000000e+00)
-; CHECK-NEXT:    ret float [[VAL]]
+; CHECK-NEXT:    ret float 1.000000e+00
 ;
   %val = call float @llvm.amdgcn.exp2.f32(float 0.0)
   ret float %val
 }
 
 define float @test_constant_fold_exp2_f32_n0() {
 ; CHECK-LABEL: @test_constant_fold_exp2_f32_n0(
-; CHECK-NEXT:    [[VAL:%.*]] = call float @llvm.amdgcn.exp2.f32(float -0.000000e+00)
-; CHECK-NEXT:    ret float [[VAL]]
+; CHECK-NEXT:    ret float 1.000000e+00
 ;
   %val = call float @llvm.amdgcn.exp2.f32(float -0.0)
   ret float %val
 }
 
 define float @test_constant_fold_exp2_f32_p1() {
 ; CHECK-LABEL: @test_constant_fold_exp2_f32_p1(
 ; CHECK-NEXT:    [[VAL:%.*]] = call float @llvm.amdgcn.exp2.f32(float 1.000000e+00)

 ; CHECK-NEXT:    ret float [[VAL]]
 ;
   %val = call float @llvm.amdgcn.exp2.f32(float -4.0)
   ret float %val
 }
 
 define float @test_constant_fold_exp2_f32_subnormal() {
 ; CHECK-LABEL: @test_constant_fold_exp2_f32_subnormal(
-; CHECK-NEXT:    [[VAL:%.*]] = call float @llvm.amdgcn.exp2.f32(float 0x380FFFFFC0000000)
-; CHECK-NEXT:    ret float [[VAL]]
+; CHECK-NEXT:    ret float 1.000000e+00
 ;
   %val = call float @llvm.amdgcn.exp2.f32(float 0x380FFFFFC0000000)
   ret float %val
 }
 
 define float @test_constant_fold_exp2_f32_negsubnormal() {
 ; CHECK-LABEL: @test_constant_fold_exp2_f32_negsubnormal(
-; CHECK-NEXT:    [[VAL:%.*]] = call float @llvm.amdgcn.exp2.f32(float 0xB80FFFFFC0000000)
-; CHECK-NEXT:    ret float [[VAL]]
+; CHECK-NEXT:    ret float 1.000000e+00
 ;
   %val = call float @llvm.amdgcn.exp2.f32(float 0xB80FFFFFC0000000)
   ret float %val
 }
 
 define float @test_constant_fold_exp2_f32_pinf() {
 ; CHECK-LABEL: @test_constant_fold_exp2_f32_pinf(
-; CHECK-NEXT:    [[VAL:%.*]] = call float @llvm.amdgcn.exp2.f32(float 0x7FF0000000000000)
-; CHECK-NEXT:    ret float [[VAL]]
+; CHECK-NEXT:    ret float 0x7FF0000000000000
 ;
   %val = call float @llvm.amdgcn.exp2.f32(float 0x7FF0000000000000)
   ret float %val
 }
 
 define float @test_constant_fold_exp2_f32_ninf() {
 ; CHECK-LABEL: @test_constant_fold_exp2_f32_ninf(
-; CHECK-NEXT:    [[VAL:%.*]] = call float @llvm.amdgcn.exp2.f32(float 0xFFF0000000000000)
-; CHECK-NEXT:    ret float [[VAL]]
+; CHECK-NEXT:    ret float 0.000000e+00
 ;
   %val = call float @llvm.amdgcn.exp2.f32(float 0xFFF0000000000000)
   ret float %val
 }
 
 define float @test_constant_fold_exp2_f32_p10() {
 ; CHECK-LABEL: @test_constant_fold_exp2_f32_p10(
 ; CHECK-NEXT:    [[VAL:%.*]] = call float @llvm.amdgcn.exp2.f32(float 1.000000e+01)

 ; CHECK-NEXT:    ret float [[VAL]]
 ;
   %val = call float @llvm.amdgcn.exp2.f32(float -10.0)
   ret float %val
 }
 
 define float @test_constant_fold_exp2_f32_qnan() {
 ; CHECK-LABEL: @test_constant_fold_exp2_f32_qnan(
-; CHECK-NEXT:    [[VAL:%.*]] = call float @llvm.amdgcn.exp2.f32(float 0x7FF8000000000000)
-; CHECK-NEXT:    ret float [[VAL]]
+; CHECK-NEXT:    ret float 0x7FF8000000000000
 ;
   %val = call float @llvm.amdgcn.exp2.f32(float 0x7FF8000000000000)
   ret float %val
 }
 
 define float @test_constant_fold_exp2_f32_snan() {
 ; CHECK-LABEL: @test_constant_fold_exp2_f32_snan(
-; CHECK-NEXT:    [[VAL:%.*]] = call float @llvm.amdgcn.exp2.f32(float 0x7FF0000020000000)
-; CHECK-NEXT:    ret float [[VAL]]
+; CHECK-NEXT:    ret float 0x7FF8000020000000
 ;
   %val = call float @llvm.amdgcn.exp2.f32(float 0x7FF0000020000000)
   ret float %val
 }
 
 define half @test_constant_fold_exp2_f16_p0() {
 ; CHECK-LABEL: @test_constant_fold_exp2_f16_p0(
-; CHECK-NEXT:    [[VAL:%.*]] = call half @llvm.amdgcn.exp2.f16(half 0xH0000)
-; CHECK-NEXT:    ret half [[VAL]]
+; CHECK-NEXT:    ret half 0xH3C00
 ;
   %val = call half @llvm.amdgcn.exp2.f16(half 0.0)
   ret half %val
 }
 
 define half @test_constant_fold_exp2_f16_neg10() {
 ; CHECK-LABEL: @test_constant_fold_exp2_f16_neg10(
 ; CHECK-NEXT:    [[VAL:%.*]] = call half @llvm.amdgcn.exp2.f16(half 0xHC900)

 ; CHECK-NEXT:    ret float [[VAL]]
 ;
   %val = call float @llvm.amdgcn.exp2.f32(float -10.0) strictfp
   ret float %val
 }
 
 define float @test_constant_fold_exp2_f32_pinf_strictfp() strictfp {
 ; CHECK-LABEL: @test_constant_fold_exp2_f32_pinf_strictfp(
-; CHECK-NEXT:    [[VAL:%.*]] = call float @llvm.amdgcn.exp2.f32(float 0x7FF0000000000000) #[[ATTR16]]
-; CHECK-NEXT:    ret float [[VAL]]
+; CHECK-NEXT:    ret float 0x7FF0000000000000
 ;
   %val = call float @llvm.amdgcn.exp2.f32(float 0x7FF0000000000000) strictfp
   ret float %val
 }
 
 define float @test_constant_fold_exp2_f32_ninf_strictfp() strictfp {
 ; CHECK-LABEL: @test_constant_fold_exp2_f32_ninf_strictfp(
-; CHECK-NEXT:    [[VAL:%.*]] = call float @llvm.amdgcn.exp2.f32(float 0xFFF0000000000000) #[[ATTR16]]
-; CHECK-NEXT:    ret float [[VAL]]
+; CHECK-NEXT:    ret float 0.000000e+00
 ;
   %val = call float @llvm.amdgcn.exp2.f32(float 0xFFF0000000000000) strictfp
   ret float %val
 }
 
 define half @test_constant_fold_exp2_f16_denorm() {
 ; CHECK-LABEL: @test_constant_fold_exp2_f16_denorm(
 ; CHECK-NEXT:    [[VAL:%.*]] = call half @llvm.amdgcn.exp2.f16(half 0xH03FF)