[SYCL] Enable FTZ operations for CUDA/PTX backend via -fcuda-flush-denormals-to-zero (#6411)

Gives a small performance benefit for Gromacs on NVidia targets

Author: P-Jensen

clang/lib/CodeGen/CodeGenModule.cpp

@@ -840,13 +840,15 @@ void CodeGenModule::Release() {
         llvm::MDString::get(Ctx, CodeGenOpts.MemoryProfileOutput));
   }
 
-  if ((LangOpts.CUDAIsDevice || LangOpts.isSYCL()) && getTriple().isNVPTX()) {
+  if ((LangOpts.CUDAIsDevice || LangOpts.SYCLIsDevice) && getTriple().isNVPTX()) {
     // Indicate whether __nvvm_reflect should be configured to flush denormal
     // floating point values to 0.  (This corresponds to its "__CUDA_FTZ"
     // property.)
     getModule().addModuleFlag(llvm::Module::Override, "nvvm-reflect-ftz",
-                              CodeGenOpts.FP32DenormalMode.Output !=
-                                  llvm::DenormalMode::IEEE);
+                              (CodeGenOpts.FP32DenormalMode.Output !=
+                                  llvm::DenormalMode::IEEE) ||
+                              (CodeGenOpts.FPDenormalMode.Output !=
+                                  llvm::DenormalMode::IEEE));
     getModule().addModuleFlag(llvm::Module::Override, "nvvm-reflect-prec-sqrt",
                               getTarget().getTargetOpts().NVVMCudaPrecSqrt);
   }

clang/lib/Driver/ToolChains/Clang.cpp

@@ -3014,6 +3014,11 @@ static void RenderFloatingPointOptions(const ToolChain &TC, const Driver &D,
       RoundingMathPresent = false;
       break;
 
+    case options::OPT_fcuda_flush_denormals_to_zero:
+    case options::OPT_fgpu_flush_denormals_to_zero:
+      DenormalFP32Math = llvm::DenormalMode::getPreserveSign();
+      break;
+
     case options::OPT_fdenormal_fp_math_EQ:
       DenormalFPMath = llvm::parseDenormalFPAttribute(A->getValue());
       DenormalFP32Math = DenormalFPMath;

clang/test/CodeGenSYCL/flush-denormals.cpp

@@ -0,0 +1,18 @@
+// RUN: %clang_cc1 -fcuda-is-device -fdenormal-fp-math-f32=preserve-sign \
+// RUN:   -triple nvptx-nvidia-cuda -emit-llvm -o - %s | \
+// RUN:   FileCheck -check-prefixes=FTZ32,PTXFTZ32 %s
+
+// RUN: %clang_cc1 -fcuda-is-device -fdenormal-fp-math=preserve-sign \
+// RUN:   -triple nvptx-nvidia-cuda -emit-llvm -o - %s | \
+// RUN:   FileCheck -check-prefixes=FTZ,PTXFTZ %s
+
+// CHECK-LABEL: define void @_Z3foov() #0
+void foo() {}
+
+// FTZ32: attributes #0 = {{.*}} "denormal-fp-math-f32"="preserve-sign,preserve-sign"
+// PTXFTZ32:!llvm.module.flags = !{{{.*}}, [[MODFLAG:![0-9]+]], {{.*}}}
+// PTXFTZ32:[[MODFLAG]] = !{i32 4, !"nvvm-reflect-ftz", i32 1}
+
+// FTZ: attributes #0 = {{.*}} "denormal-fp-math"="preserve-sign,preserve-sign"
+// PTXFTZ:!llvm.module.flags = !{{{.*}}, [[MODFLAG:![0-9]+]], {{.*}}}
+// PTXFTZ:[[MODFLAG]] = !{i32 4, !"nvvm-reflect-ftz", i32 1}

clang/test/Driver/ftz-cuda.c

@@ -0,0 +1,4 @@
+// RUN: %clang -### -fcuda-flush-denormals-to-zero -c %s 2>&1 \
+// RUN:   | FileCheck --check-prefix=CHECK-FTZ %s
+// CHECK-FTZ: "-cc1"
+// CHECK-FTZ: "-fdenormal-fp-math-f32=preserve-sign,preserve-sign"