AMDGPU/GFX9: Set +fast-fmaf for >=gfx900 unless -cl-denorms-are-zero is set

[kzhuravl]

Add \n to test file.

[kzhuravl]

Ping.

[kzhuravl]

Ping.

[arsenm]

lib/Basic/Targets.cpp

2208–2210

We glue fast-fmaf to the specific subtargets, so I don't think you can actually turn it and off. It has this nonsensical behavior where it assumes if you disable a subtarget feature for the processors you aren't actually targeting the processor at all. I think a dedicated new feature might be needed

[kzhuravl]

Abandoned in favor of D32363.

lib/Basic/Targets.cpp

void adjustTargetOptions(const CodeGenOptions &CGOpts,
     bool hasFP32Denormals = false;
     bool hasFP64Denormals = false;
     for (auto &I : TargetOpts.FeaturesAsWritten) {
       if (I == "+fp32-denormals" || I == "-fp32-denormals")
         hasFP32Denormals = true;
       if (I == "+fp64-fp16-denormals" || I == "-fp64-fp16-denormals")
         hasFP64Denormals = true;
     }
-    if (!hasFP32Denormals)
+    if (!hasFP32Denormals) {
       TargetOpts.Features.push_back(
           (Twine(hasFullSpeedFMAF32(TargetOpts.CPU) &&
           !CGOpts.FlushDenorm ? '+' : '-') + Twine("fp32-denormals")).str());
+      TargetOpts.Features.push_back(
+          (Twine(hasFullSpeedFMAF32(TargetOpts.CPU) &&
+          !CGOpts.FlushDenorm ? '+' : '-') + Twine("fast-fmaf")).str());
+    }
     // Always do not flush fp64 or fp16 denorms.
     if (!hasFP64Denormals && hasFP64)
       TargetOpts.Features.push_back("+fp64-fp16-denormals");
   }
 
   ArrayRef<Builtin::Info> getTargetBuiltins() const override {
     return llvm::makeArrayRef(BuiltinInfo,
                         clang::AMDGPU::LastTSBuiltin - Builtin::FirstTSBuiltin);

test/CodeGenOpenCL/gfx9-fast-fmaf.cl

+// REQUIRES: amdgpu-registered-target
+
+// RUN: %clang_cc1 -triple amdgcn-unknown-unknown -target-cpu gfx900 -S -emit-llvm -o - %s | FileCheck --check-prefix=DEFAULT %s
+// RUN: %clang_cc1 -triple amdgcn-unknown-unknown -target-cpu gfx900 -S -emit-llvm -o - -target-feature +fast-fmaf %s | FileCheck --check-prefix=FEATURE_FP32_DENORMALS_ON %s
+// RUN: %clang_cc1 -triple amdgcn-unknown-unknown -target-cpu gfx900 -S -emit-llvm -o - -target-feature -fast-fmaf %s | FileCheck --check-prefix=FEATURE_FP32_DENORMALS_OFF %s
+// RUN: %clang_cc1 -triple amdgcn-unknown-unknown -target-cpu gfx900 -S -emit-llvm -o - -cl-denorms-are-zero %s | FileCheck --check-prefix=OPT_DENORMS_ARE_ZERO %s
+
+// DEFAULT: +fast-fmaf
+// FEATURE_FP32_DENORMALS_ON: +fast-fmaf
+// FEATURE_FP32_DENORMALS_OFF: -fast-fmaf
+// OPT_DENORMS_ARE_ZERO: -fast-fmaf
+
+kernel void gfx9_fast_fmaf() {}