R600: Add denormal handling subtarget features.

llvm-svn: 213018

llvm/lib/Target/R600/AMDGPU.td

@@ -42,6 +42,20 @@ def FeatureFP64 : SubtargetFeature<"fp64",
         "true",
         "Enable double precision operations">;
 
+def FeatureFP64Denormals : SubtargetFeature<"fp64-denormals",
+        "FP64Denormals",
+        "true",
+        "Enable double precision denormal handling",
+        [FeatureFP64]>;
+
+// Some instructions do not support denormals despite this flag. Using
+// fp32 denormals also causes instructions to run at the double
+// precision rate for the device.
+def FeatureFP32Denormals : SubtargetFeature<"fp32-denormals",
+        "FP32Denormals",
+        "true",
+        "Enable single precision denormal handling">;
+
 def Feature64BitPtr : SubtargetFeature<"64BitPtr",
         "Is64bit",
         "true",

llvm/lib/Target/R600/AMDGPUAsmPrinter.cpp

@@ -55,11 +55,20 @@ using namespace llvm;
 // We want to use these instructions, and using fp32 denormals also causes
 // instructions to run at the double precision rate for the device so it's
 // probably best to just report no single precision denormals.
-static uint32_t getFPMode(const MachineFunction &) {
+static uint32_t getFPMode(const MachineFunction &F) {
+  const AMDGPUSubtarget& ST = F.getTarget().getSubtarget<AMDGPUSubtarget>();
+  // TODO: Is there any real use for the flush in only / flush out only modes?
+
+  uint32_t FP32Denormals =
+    ST.hasFP32Denormals() ? FP_DENORM_FLUSH_NONE : FP_DENORM_FLUSH_IN_FLUSH_OUT;
+
+  uint32_t FP64Denormals =
+    ST.hasFP64Denormals() ? FP_DENORM_FLUSH_NONE : FP_DENORM_FLUSH_IN_FLUSH_OUT;
+
   return FP_ROUND_MODE_SP(FP_ROUND_ROUND_TO_NEAREST) |
          FP_ROUND_MODE_DP(FP_ROUND_ROUND_TO_NEAREST) |
-         FP_DENORM_MODE_SP(FP_DENORM_FLUSH_IN_FLUSH_OUT) |
-         FP_DENORM_MODE_DP(FP_DENORM_FLUSH_NONE);
+         FP_DENORM_MODE_SP(FP32Denormals) |
+         FP_DENORM_MODE_DP(FP64Denormals);
 }
 
 static AsmPrinter *createAMDGPUAsmPrinterPass(TargetMachine &tm,

llvm/lib/Target/R600/AMDGPUInstructions.td

@@ -34,6 +34,9 @@ class AMDGPUShaderInst <dag outs, dag ins, string asm, list<dag> pattern>
 
 }
 
+def FP32Denormals : Predicate<"Subtarget.hasFP32Denormals()">;
+def FP64Denormals : Predicate<"Subtarget.hasFP64Denormals()">;
+
 def InstFlag : OperandWithDefaultOps <i32, (ops (i32 0))>;
 def ADDRIndirect : ComplexPattern<iPTR, 2, "SelectADDRIndirect", [], []>;
 

llvm/lib/Target/R600/AMDGPUSubtarget.cpp

@@ -15,6 +15,7 @@
 #include "AMDGPUSubtarget.h"
 #include "R600InstrInfo.h"
 #include "SIInstrInfo.h"
+#include "llvm/ADT/SmallString.h"
 
 #include "llvm/ADT/SmallString.h"
 
@@ -37,6 +38,8 @@ AMDGPUSubtarget::AMDGPUSubtarget(StringRef TT, StringRef GPU, StringRef FS) :
   TexVTXClauseSize(0),
   Gen(AMDGPUSubtarget::R600),
   FP64(false),
+  FP64Denormals(false),
+  FP32Denormals(false),
   CaymanISA(false),
   EnableIRStructurizer(true),
   EnablePromoteAlloca(false),
@@ -45,14 +48,27 @@ AMDGPUSubtarget::AMDGPUSubtarget(StringRef TT, StringRef GPU, StringRef FS) :
   CFALUBug(false),
   LocalMemorySize(0),
   InstrItins(getInstrItineraryForCPU(GPU)) {
+  // On SI+, we want FP64 denormals to be on by default. FP32 denormals can be
+  // enabled, but some instructions do not respect them and they run at the
+  // double precision rate, so don't enable by default.
+  //
+  // We want to be able to turn these off, but making this a subtarget feature
+  // for SI has the unhelpful behavior that it unsets everything else if you
+  // disable it.
 
-  SmallString<256> FullFS("+promote-alloca,");
+  SmallString<256> FullFS("+promote-alloca,+fp64-denormals,");
   FullFS += FS;
 
   ParseSubtargetFeatures(GPU, FullFS);
 
   if (getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS) {
     InstrInfo.reset(new R600InstrInfo(*this));
+
+    // FIXME: I don't think think Evergreen has any useful support for
+    // denormals, but should be checked. Should we issue a warning somewhere if
+    // someone tries to enable these?
+    FP32Denormals = false;
+    FP64Denormals = false;
   } else {
     InstrInfo.reset(new SIInstrInfo(*this));
   }

llvm/lib/Target/R600/AMDGPUSubtarget.h

@@ -50,6 +50,8 @@ private:
   short TexVTXClauseSize;
   Generation Gen;
   bool FP64;
+  bool FP64Denormals;
+  bool FP32Denormals;
   bool CaymanISA;
   bool EnableIRStructurizer;
   bool EnablePromoteAlloca;
@@ -97,6 +99,14 @@ public:
     return CaymanISA;
   }
 
+  bool hasFP32Denormals() const {
+    return FP32Denormals;
+  }
+
+  bool hasFP64Denormals() const {
+    return FP64Denormals;
+  }
+
   bool hasBFE() const {
     return (getGeneration() >= EVERGREEN);
   }

llvm/test/CodeGen/R600/default-fp-mode.ll

@@ -1,8 +1,27 @@
-; RUN: llc -march=r600 -mcpu=SI < %s | FileCheck -check-prefix=SI %s
+; RUN: llc -march=r600 -mcpu=SI -mattr=-fp32-denormals,+fp64-denormals < %s | FileCheck -check-prefix=FP64-DENORMAL -check-prefix=FUNC %s
+; RUN: llc -march=r600 -mcpu=SI -mattr=+fp32-denormals,-fp64-denormals < %s | FileCheck -check-prefix=FP32-DENORMAL -check-prefix=FUNC %s
+; RUN: llc -march=r600 -mcpu=SI -mattr=+fp32-denormals,+fp64-denormals < %s | FileCheck -check-prefix=BOTH-DENORMAL -check-prefix=FUNC %s
+; RUN: llc -march=r600 -mcpu=SI -mattr=-fp32-denormals,-fp64-denormals < %s | FileCheck -check-prefix=NO-DENORMAL -check-prefix=FUNC %s
+; RUN: llc -march=r600 -mcpu=SI < %s | FileCheck -check-prefix=DEFAULT -check-prefix=FUNC %s
+; RUN: llc -march=r600 -mcpu=SI -mattr=-fp32-denormals < %s | FileCheck -check-prefix=DEFAULT -check-prefix=FUNC %s
+; RUN: llc -march=r600 -mcpu=SI -mattr=+fp64-denormals < %s | FileCheck -check-prefix=DEFAULT -check-prefix=FUNC %s
 
-; SI-LABEL: @test_kernel
-; SI: FloatMode: 192
-; SI: IeeeMode: 0
+; FUNC-LABEL: @test_kernel
+
+; DEFAULT: FloatMode: 192
+; DEFAULT: IeeeMode: 0
+
+; FP64-DENORMAL: FloatMode: 192
+; FP64-DENORMAL: IeeeMode: 0
+
+; FP32-DENORMAL: FloatMode: 48
+; FP32-DENORMAL: IeeeMode: 0
+
+; BOTH-DENORMAL: FloatMode: 240
+; BOTH-DENORMAL: IeeeMode: 0
+
+; NO-DENORMAL: FloatMode: 0
+; NO-DENORMAL: IeeeMode: 0
 define void @test_kernel(float addrspace(1)* %out0, double addrspace(1)* %out1) nounwind {
   store float 0.0, float addrspace(1)* %out0
   store double 0.0, double addrspace(1)* %out1