AMDGPU: Refactor treatment of denormal mode

[arsenm]

Summary

Start moving towards treating this as a property of the calling
convention, and not the subtarget. The default denormal mode should
not be part of the subtarget, and be moved into a separate function
attribute.

This patch is still NFC. The denormal mode remains as a subtarget
feature for now, but make the necessary changes to switch to using an
attribute.

[rampitec]

LGTM

[arsenm]

Rebase and fix bug in inline compatibility. Also fix test which will be temporarily broken in GlobalISel after previous commit, and fixed here

[arsenm]

Drop accidental changes

[rampitec]

LGTM

[arsenm]

db0ed3e429b55d1730d1ecc253b0643de7fca099

llvm/lib/Target/AMDGPU/AMDGPUAsmPrinter.cpp

 //
 // FIXME: It seems some instructions do not support single precision denormals
 // regardless of the mode (exp_*_f32, rcp_*_f32, rsq_*_f32, rsq_*f32, sqrt_f32,
 // and sin_f32, cos_f32 on most parts).
 
 // 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 &F) {
-  const GCNSubtarget& ST = F.getSubtarget<GCNSubtarget>();
-  // TODO: Is there any real use for the flush in only / flush out only modes?
+static uint32_t getFPMode(AMDGPU::SIModeRegisterDefaults Mode) {
 
+  // 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;
+    Mode.FP32Denormals ? 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;
+    Mode.FP64FP16Denormals ? 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(FP32Denormals) |
          FP_DENORM_MODE_DP(FP64Denormals);
 }
 
 static AsmPrinter *
if (MFI->getLDSSize() > static_cast<unsigned>(STM.getLocalMemorySize())) {
     Ctx.diagnose(Diag);
   }
 
   ProgInfo.SGPRBlocks = IsaInfo::getNumSGPRBlocks(
       &STM, ProgInfo.NumSGPRsForWavesPerEU);
   ProgInfo.VGPRBlocks = IsaInfo::getNumVGPRBlocks(
       &STM, ProgInfo.NumVGPRsForWavesPerEU);
 
+  const SIModeRegisterDefaults Mode = MFI->getMode();
+
   // Set the value to initialize FP_ROUND and FP_DENORM parts of the mode
   // register.
-  ProgInfo.FloatMode = getFPMode(MF);
+  ProgInfo.FloatMode = getFPMode(Mode);
 
-  const SIModeRegisterDefaults Mode = MFI->getMode();
   ProgInfo.IEEEMode = Mode.IEEE;
 
   // Make clamp modifier on NaN input returns 0.
   ProgInfo.DX10Clamp = Mode.DX10Clamp;
 
   unsigned LDSAlignShift;
   if (STM.getGeneration() < AMDGPUSubtarget::SEA_ISLANDS) {
     // LDS is allocated in 64 dword blocks.

llvm/lib/Target/AMDGPU/AMDGPUCodeGenPrepare.cpp

 class AMDGPUCodeGenPrepare : public FunctionPass,
                              public InstVisitor<AMDGPUCodeGenPrepare, bool> {
   const GCNSubtarget *ST = nullptr;
   AssumptionCache *AC = nullptr;
   LegacyDivergenceAnalysis *DA = nullptr;
   Module *Mod = nullptr;
   const DataLayout *DL = nullptr;
   bool HasUnsafeFPMath = false;
+  bool HasFP32Denormals = false;
 
   /// Copies exact/nsw/nuw flags (if any) from binary operation \p I to
   /// binary operation \p V.
   ///
   /// \returns Binary operation \p V.
   /// \returns \p T's base element bit width.
   unsigned getBaseElementBitWidth(const Type *T) const;
 
bool AMDGPUCodeGenPrepare::visitFDiv(BinaryOperator &FDiv) {
 
   Function *Decl = Intrinsic::getDeclaration(Mod, Intrinsic::amdgcn_fdiv_fast);
 
   Value *Num = FDiv.getOperand(0);
   Value *Den = FDiv.getOperand(1);
 
   Value *NewFDiv = nullptr;
 
-  bool HasDenormals = ST->hasFP32Denormals();
   if (VectorType *VT = dyn_cast<VectorType>(Ty)) {
     NewFDiv = UndefValue::get(VT);
 
     // FIXME: Doesn't do the right thing for cases where the vector is partially
     // constant. This works when the scalarizer pass is run first.
     for (unsigned I = 0, E = VT->getNumElements(); I != E; ++I) {
       Value *NumEltI = Builder.CreateExtractElement(Num, I);
       Value *DenEltI = Builder.CreateExtractElement(Den, I);
       Value *NewElt;
 
-      if (shouldKeepFDivF32(NumEltI, UnsafeDiv, HasDenormals)) {
+      if (shouldKeepFDivF32(NumEltI, UnsafeDiv, HasFP32Denormals)) {
         NewElt = Builder.CreateFDiv(NumEltI, DenEltI);
       } else {
         NewElt = Builder.CreateCall(Decl, { NumEltI, DenEltI });
       }
 
       NewFDiv = Builder.CreateInsertElement(NewFDiv, NewElt, I);
     }
   } else {
-    if (!shouldKeepFDivF32(Num, UnsafeDiv, HasDenormals))
+    if (!shouldKeepFDivF32(Num, UnsafeDiv, HasFP32Denormals))
       NewFDiv = Builder.CreateCall(Decl, { Num, Den });
   }
 
   if (NewFDiv) {
     FDiv.replaceAllUsesWith(NewFDiv);
     NewFDiv->takeName(&FDiv);
     FDiv.eraseFromParent();
   }
bool AMDGPUCodeGenPrepare::runOnFunction(Function &F) {
   if (!TPC)
     return false;
 
   const AMDGPUTargetMachine &TM = TPC->getTM<AMDGPUTargetMachine>();
   ST = &TM.getSubtarget<GCNSubtarget>(F);
   AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
   DA = &getAnalysis<LegacyDivergenceAnalysis>();
   HasUnsafeFPMath = hasUnsafeFPMath(F);
+  HasFP32Denormals = ST->hasFP32Denormals(F);
 
   bool MadeChange = false;
 
   for (BasicBlock &BB : F) {
     BasicBlock::iterator Next;
     for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; I = Next) {
       Next = std::next(I);
       MadeChange |= visit(*I);

llvm/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp

 }
 
 /// AMDGPU specific code to select AMDGPU machine instructions for
 /// SelectionDAG operations.
 class AMDGPUDAGToDAGISel : public SelectionDAGISel {
   // Subtarget - Keep a pointer to the AMDGPU Subtarget around so that we can
   // make the right decision when generating code for different targets.
   const GCNSubtarget *Subtarget;
+
+  // Default FP mode for the current function.
+  AMDGPU::SIModeRegisterDefaults Mode;
+
   bool EnableLateStructurizeCFG;
 
 public:
   explicit AMDGPUDAGToDAGISel(TargetMachine *TM = nullptr,
                               CodeGenOpt::Level OptLevel = CodeGenOpt::Default)
     : SelectionDAGISel(*TM, OptLevel) {
     EnableLateStructurizeCFG = AMDGPUTargetMachine::EnableLateStructurizeCFG;
   }

 #ifdef EXPENSIVE_CHECKS
   DominatorTree & DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
   LoopInfo * LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
   for (auto &L : LI->getLoopsInPreorder()) {
     assert(L->isLCSSAForm(DT));
   }
 #endif
   Subtarget = &MF.getSubtarget<GCNSubtarget>();
+  Mode = AMDGPU::SIModeRegisterDefaults(MF.getFunction(), *Subtarget);
   return SelectionDAGISel::runOnMachineFunction(MF);
 }
 
 bool AMDGPUDAGToDAGISel::matchLoadD16FromBuildVector(SDNode *N) const {
   assert(Subtarget->d16PreservesUnusedBits());
   MVT VT = N->getValueType(0).getSimpleVT();
   if (VT != MVT::v2i16 && VT != MVT::v2f16)
     return false;
void AMDGPUDAGToDAGISel::SelectFMAD_FMA(SDNode *N) {
   unsigned Src0Mods, Src1Mods, Src2Mods;
 
   // Avoid using v_mad_mix_f32/v_fma_mix_f32 unless there is actually an operand
   // using the conversion from f16.
   bool Sel0 = SelectVOP3PMadMixModsImpl(Src0, Src0, Src0Mods);
   bool Sel1 = SelectVOP3PMadMixModsImpl(Src1, Src1, Src1Mods);
   bool Sel2 = SelectVOP3PMadMixModsImpl(Src2, Src2, Src2Mods);
 
-  assert((IsFMA || !Subtarget->hasFP32Denormals()) &&
+  assert((IsFMA || !Mode.FP32Denormals) &&
          "fmad selected with denormals enabled");
   // TODO: We can select this with f32 denormals enabled if all the sources are
   // converted from f16 (in which case fmad isn't legal).
 
   if (Sel0 || Sel1 || Sel2) {
     // For dummy operands.
     SDValue Zero = CurDAG->getTargetConstant(0, SDLoc(), MVT::i32);
     SDValue Ops[] = {

llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp

SDValue fq = DAG.getNode(ISD::FMUL, DL, FltVT,
                            fa, DAG.getNode(AMDGPUISD::RCP, DL, FltVT, fb));
 
   // fq = trunc(fq);
   fq = DAG.getNode(ISD::FTRUNC, DL, FltVT, fq);
 
   // float fqneg = -fq;
   SDValue fqneg = DAG.getNode(ISD::FNEG, DL, FltVT, fq);
 
+  MachineFunction &MF = DAG.getMachineFunction();
+  const AMDGPUMachineFunction *MFI = MF.getInfo<AMDGPUMachineFunction>();
+
   // float fr = mad(fqneg, fb, fa);
-  unsigned OpCode = Subtarget->hasFP32Denormals() ?
+  unsigned OpCode = MFI->getMode().FP32Denormals ?
                     (unsigned)AMDGPUISD::FMAD_FTZ :
                     (unsigned)ISD::FMAD;
   SDValue fr = DAG.getNode(OpCode, DL, FltVT, fqneg, fb, fa);
 
   // int iq = (int)fq;
   SDValue iq = DAG.getNode(ToInt, DL, IntVT, fq);
 
   // fr = fabs(fr);
if (DAG.MaskedValueIsZero(RHS, APInt::getHighBitsSet(64, 32)) &&
     SDValue REM = DAG.getBuildVector(MVT::v2i32, DL, {Res.getValue(1), Zero});
 
     Results.push_back(DAG.getNode(ISD::BITCAST, DL, MVT::i64, DIV));
     Results.push_back(DAG.getNode(ISD::BITCAST, DL, MVT::i64, REM));
     return;
   }
 
   if (isTypeLegal(MVT::i64)) {
+    MachineFunction &MF = DAG.getMachineFunction();
+    const SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
+
     // Compute denominator reciprocal.
-    unsigned FMAD = Subtarget->hasFP32Denormals() ?
+    unsigned FMAD = MFI->getMode().FP32Denormals ?
                     (unsigned)AMDGPUISD::FMAD_FTZ :
                     (unsigned)ISD::FMAD;
 
     SDValue Cvt_Lo = DAG.getNode(ISD::UINT_TO_FP, DL, MVT::f32, RHS_Lo);
     SDValue Cvt_Hi = DAG.getNode(ISD::UINT_TO_FP, DL, MVT::f32, RHS_Hi);
     SDValue Mad1 = DAG.getNode(FMAD, DL, MVT::f32, Cvt_Hi,
       DAG.getConstantFP(APInt(32, 0x4f800000).bitsToFloat(), DL, MVT::f32),
       Cvt_Lo);

llvm/lib/Target/AMDGPU/AMDGPUInstructions.td

list<Predicate> Predicates = PredConcat<
                                                        SubtargetPredicate>.ret,
                                             AssemblerPredicate>.ret,
                                  WaveSizePredicate>.ret;
 }
 
 class AMDGPUPat<dag pattern, dag result> : Pat<pattern, result>,
       PredicateControl;
 
-def FP16Denormals : Predicate<"Subtarget->hasFP16Denormals()">;
-def FP32Denormals : Predicate<"Subtarget->hasFP32Denormals()">;
-def FP64Denormals : Predicate<"Subtarget->hasFP64Denormals()">;
-def NoFP16Denormals : Predicate<"!Subtarget->hasFP16Denormals()">;
-def NoFP32Denormals : Predicate<"!Subtarget->hasFP32Denormals()">;
-def NoFP64Denormals : Predicate<"!Subtarget->hasFP64Denormals()">;
+let RecomputePerFunction = 1 in {
+def FP16Denormals : Predicate<"MF->getInfo<SIMachineFunctionInfo>()->getMode().FP64FP16Denormals">;
+def FP32Denormals : Predicate<"MF->getInfo<SIMachineFunctionInfo>()->getMode().FP32Denormals">;
+def FP64Denormals : Predicate<"MF->getInfo<SIMachineFunctionInfo>()->getMode().FP64FP16Denormals">;
+def NoFP16Denormals : Predicate<"!MF->getInfo<SIMachineFunctionInfo>()->getMode().FP64FP16Denormals">;
+def NoFP32Denormals : Predicate<"!MF->getInfo<SIMachineFunctionInfo>()->getMode().FP32Denormals">;
+def NoFP64Denormals : Predicate<"!MF->getInfo<SIMachineFunctionInfo>()->getMode().FP64FP16Denormals">;
 def UnsafeFPMath : Predicate<"TM.Options.UnsafeFPMath">;
+}
+
 def FMA : Predicate<"Subtarget->hasFMA()">;
 
 def InstFlag : OperandWithDefaultOps <i32, (ops (i32 0))>;
 
 def u16ImmTarget : AsmOperandClass {
   let Name = "U16Imm";
   let RenderMethod = "addImmOperands";
 }

llvm/lib/Target/AMDGPU/AMDGPUMachineFunction.h

 //===-- AMDGPUMachineFunctionInfo.h -------------------------------*- C++ -*-=//
 //
 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
 // See https://llvm.org/LICENSE.txt for license information.
 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 //
 //===----------------------------------------------------------------------===//
 
 #ifndef LLVM_LIB_TARGET_AMDGPU_AMDGPUMACHINEFUNCTION_H
 #define LLVM_LIB_TARGET_AMDGPU_AMDGPUMACHINEFUNCTION_H
 
 #include "llvm/ADT/DenseMap.h"
 #include "llvm/CodeGen/MachineFunction.h"
+#include "Utils/AMDGPUBaseInfo.h"
 
 namespace llvm {
 
 class GCNSubtarget;
 
 class AMDGPUMachineFunction : public MachineFunctionInfo {
   /// A map to keep track of local memory objects and their offsets within the
   /// local memory space.
   SmallDenseMap<const GlobalValue *, unsigned, 4> LocalMemoryObjects;
 
 protected:
   uint64_t ExplicitKernArgSize; // Cache for this.
   Align MaxKernArgAlign;        // Cache for this.
 
   /// Number of bytes in the LDS that are being used.
   unsigned LDSSize;
 
+  // State of MODE register, assumed FP mode.
+  AMDGPU::SIModeRegisterDefaults Mode;
+
   // Kernels + shaders. i.e. functions called by the driver and not called
   // by other functions.
   bool IsEntryFunction;
 
   bool NoSignedZerosFPMath;
 
   // Function may be memory bound.
   bool MemoryBound;
public:
   }
 
   unsigned getMaxKernArgAlign() const { return MaxKernArgAlign.value(); }
 
   unsigned getLDSSize() const {
     return LDSSize;
   }
 
+  AMDGPU::SIModeRegisterDefaults getMode() const {
+    return Mode;
+  }
+
   bool isEntryFunction() const {
     return IsEntryFunction;
   }
 
   bool hasNoSignedZerosFPMath() const {
     return NoSignedZerosFPMath;
   }
 

llvm/lib/Target/AMDGPU/AMDGPUMachineFunction.cpp

 
 using namespace llvm;
 
 AMDGPUMachineFunction::AMDGPUMachineFunction(const MachineFunction &MF) :
   MachineFunctionInfo(),
   LocalMemoryObjects(),
   ExplicitKernArgSize(0),
   LDSSize(0),
+  Mode(MF.getFunction(), MF.getSubtarget<GCNSubtarget>()),
   IsEntryFunction(AMDGPU::isEntryFunctionCC(MF.getFunction().getCallingConv())),
   NoSignedZerosFPMath(MF.getTarget().Options.NoSignedZerosFPMath),
   MemoryBound(false),
   WaveLimiter(false) {
   const AMDGPUSubtarget &ST = AMDGPUSubtarget::get(MF);
 
   // FIXME: Should initialize KernArgSize based on ExplicitKernelArgOffset,
   // except reserved size is not correctly aligned.

llvm/lib/Target/AMDGPU/AMDGPUSubtarget.h

public:
   bool has16BitInsts() const {
     return Has16BitInsts;
   }
 
   bool hasMadMixInsts() const {
     return HasMadMixInsts;
   }
 
-  bool hasFP32Denormals() const {
+  bool hasFP32Denormals(const Function &F) const {
+    // FIXME: This should not be a property of the subtarget. This should be a
+    // property with a default set by the calling convention which can be
+    // overridden by attributes. For now, use the subtarget feature as a
+    // placeholder attribute. The function arguments only purpose is to
+    // discourage use without a function context until this is removed.
     return FP32Denormals;
   }
 
   bool hasFPExceptions() const {
     return FPExceptions;
   }
 
   bool hasSDWA() const {
bool dumpCode() const {
     return DumpCode;
   }
 
   /// Return the amount of LDS that can be used that will not restrict the
   /// occupancy lower than WaveCount.
   unsigned getMaxLocalMemSizeWithWaveCount(unsigned WaveCount,
                                            const Function &) const;
 
-  bool hasFP16Denormals() const {
+  /// Alias for hasFP64FP16Denormals
+  bool hasFP16Denormals(const Function &F) const {
     return FP64FP16Denormals;
   }
 
-  bool hasFP64Denormals() const {
+  /// Alias for hasFP64FP16Denormals
+  bool hasFP64Denormals(const Function &F) const {
+    return FP64FP16Denormals;
+  }
+
+  bool hasFP64FP16Denormals(const Function &F) const {
     return FP64FP16Denormals;
   }
 
   bool supportsMinMaxDenormModes() const {
     return getGeneration() >= AMDGPUSubtarget::GFX9;
   }
 
   /// \returns If target supports S_DENORM_MODE.

llvm/lib/Target/AMDGPU/AMDGPUTargetTransformInfo.h

 class AMDGPUTTIImpl final : public BasicTTIImplBase<AMDGPUTTIImpl> {
   using BaseT = BasicTTIImplBase<AMDGPUTTIImpl>;
   using TTI = TargetTransformInfo;
 
   friend BaseT;
 
   Triple TargetTriple;
 
-  const TargetSubtargetInfo *ST;
+  const GCNSubtarget *ST;
   const TargetLoweringBase *TLI;
 
   const TargetSubtargetInfo *getST() const { return ST; }
   const TargetLoweringBase *getTLI() const { return TLI; }
 
 public:
   explicit AMDGPUTTIImpl(const AMDGPUTargetMachine *TM, const Function &F)
       : BaseT(TM, F.getParent()->getDataLayout()),
class GCNTTIImpl final : public BasicTTIImplBase<GCNTTIImpl> {
   using TTI = TargetTransformInfo;
 
   friend BaseT;
 
   const GCNSubtarget *ST;
   const AMDGPUTargetLowering *TLI;
   AMDGPUTTIImpl CommonTTI;
   bool IsGraphicsShader;
+  bool HasFP32Denormals;
 
   const FeatureBitset InlineFeatureIgnoreList = {
     // Codegen control options which don't matter.
     AMDGPU::FeatureEnableLoadStoreOpt,
     AMDGPU::FeatureEnableSIScheduler,
     AMDGPU::FeatureEnableUnsafeDSOffsetFolding,
     AMDGPU::FeatureFlatForGlobal,
     AMDGPU::FeaturePromoteAlloca,
class GCNTTIImpl final : public BasicTTIImplBase<GCNTTIImpl> {
   }
 
 public:
   explicit GCNTTIImpl(const AMDGPUTargetMachine *TM, const Function &F)
     : BaseT(TM, F.getParent()->getDataLayout()),
       ST(static_cast<const GCNSubtarget*>(TM->getSubtargetImpl(F))),
       TLI(ST->getTargetLowering()),
       CommonTTI(TM, F),
-      IsGraphicsShader(AMDGPU::isShader(F.getCallingConv())) {}
+      IsGraphicsShader(AMDGPU::isShader(F.getCallingConv())),
+      HasFP32Denormals(ST->hasFP32Denormals(F)) { }
 
   bool hasBranchDivergence() { return true; }
 
   void getUnrollingPreferences(Loop *L, ScalarEvolution &SE,
                                TTI::UnrollingPreferences &UP);
 
   TTI::PopcntSupportKind getPopcntSupport(unsigned TyWidth) {
     assert(isPowerOf2_32(TyWidth) && "Ty width must be power of 2");

llvm/lib/Target/AMDGPU/AMDGPUTargetTransformInfo.cpp

if (SLT == MVT::f64) {
       if (!ST->hasUsableDivScaleConditionOutput())
         Cost += 3 * getFullRateInstrCost();
 
       return LT.first * Cost * NElts;
     }
 
     if (!Args.empty() && match(Args[0], PatternMatch::m_FPOne())) {
       // TODO: This is more complicated, unsafe flags etc.
-      if ((SLT == MVT::f32 && !ST->hasFP32Denormals()) ||
+      if ((SLT == MVT::f32 && !HasFP32Denormals) ||
           (SLT == MVT::f16 && ST->has16BitInsts())) {
         return LT.first * getQuarterRateInstrCost() * NElts;
       }
     }
 
     if (SLT == MVT::f16 && ST->has16BitInsts()) {
       // 2 x v_cvt_f32_f16
       // f32 rcp
       // f32 fmul
       // v_cvt_f16_f32
       // f16 div_fixup
       int Cost = 4 * getFullRateInstrCost() + 2 * getQuarterRateInstrCost();
       return LT.first * Cost * NElts;
     }
 
     if (SLT == MVT::f32 || SLT == MVT::f16) {
       int Cost = 7 * getFullRateInstrCost() + 1 * getQuarterRateInstrCost();
 
-      if (!ST->hasFP32Denormals()) {
+      if (!HasFP32Denormals) {
         // FP mode switches.
         Cost += 2 * getFullRateInstrCost();
       }
 
       return LT.first * NElts * Cost;
     }
     break;
   default:
unsigned GCNTTIImpl::getShuffleCost(TTI::ShuffleKind Kind, Type *Tp, int Index,
   }
 
   return BaseT::getShuffleCost(Kind, Tp, Index, SubTp);
 }
 
 bool GCNTTIImpl::areInlineCompatible(const Function *Caller,
                                      const Function *Callee) const {
   const TargetMachine &TM = getTLI()->getTargetMachine();
-  const FeatureBitset &CallerBits =
-    TM.getSubtargetImpl(*Caller)->getFeatureBits();
-  const FeatureBitset &CalleeBits =
-    TM.getSubtargetImpl(*Callee)->getFeatureBits();
+  const GCNSubtarget *CallerST
+    = static_cast<const GCNSubtarget *>(TM.getSubtargetImpl(*Caller));
+  const GCNSubtarget *CalleeST
+    = static_cast<const GCNSubtarget *>(TM.getSubtargetImpl(*Callee));
+
+  const FeatureBitset &CallerBits = CallerST->getFeatureBits();
+  const FeatureBitset &CalleeBits = CalleeST->getFeatureBits();
 
   FeatureBitset RealCallerBits = CallerBits & ~InlineFeatureIgnoreList;
   FeatureBitset RealCalleeBits = CalleeBits & ~InlineFeatureIgnoreList;
   if ((RealCallerBits & RealCalleeBits) != RealCalleeBits)
     return false;
 
   // FIXME: dx10_clamp can just take the caller setting, but there seems to be
   // no way to support merge for backend defined attributes.
-  AMDGPU::SIModeRegisterDefaults CallerMode(*Caller);
-  AMDGPU::SIModeRegisterDefaults CalleeMode(*Callee);
+  AMDGPU::SIModeRegisterDefaults CallerMode(*Caller, *CallerST);
+  AMDGPU::SIModeRegisterDefaults CalleeMode(*Callee, *CalleeST);
   return CallerMode.isInlineCompatible(CalleeMode);
 }
 
 void GCNTTIImpl::getUnrollingPreferences(Loop *L, ScalarEvolution &SE,
                                          TTI::UnrollingPreferences &UP) {
   CommonTTI.getUnrollingPreferences(L, SE, UP);
 }
 

llvm/lib/Target/AMDGPU/R600ISelLowering.cpp

R600TargetLowering::R600TargetLowering(const TargetMachine &TM,
   setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
   setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
 
   if (!Subtarget->hasFMA()) {
     setOperationAction(ISD::FMA, MVT::f32, Expand);
     setOperationAction(ISD::FMA, MVT::f64, Expand);
   }
 
-  // FIXME: This was moved from AMDGPUTargetLowering, I'm not sure if we
-  // need it for R600.
-  if (!Subtarget->hasFP32Denormals())
-    setOperationAction(ISD::FMAD, MVT::f32, Legal);
+  // FIXME: May need no denormals check
 
   if (!Subtarget->hasBFI()) {
     // fcopysign can be done in a single instruction with BFI.
     setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
     setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
   }
 
   if (!Subtarget->hasBCNT(32))

llvm/lib/Target/AMDGPU/SIFoldOperands.cpp

 // handled, so will anything other than that break?
 std::pair<const MachineOperand *, int>
 SIFoldOperands::isOMod(const MachineInstr &MI) const {
   unsigned Op = MI.getOpcode();
   switch (Op) {
   case AMDGPU::V_MUL_F32_e64:
   case AMDGPU::V_MUL_F16_e64: {
     // If output denormals are enabled, omod is ignored.
-    if ((Op == AMDGPU::V_MUL_F32_e64 && ST->hasFP32Denormals()) ||
-        (Op == AMDGPU::V_MUL_F16_e64 && ST->hasFP16Denormals()))
+    if ((Op == AMDGPU::V_MUL_F32_e64 && MFI->getMode().FP32Denormals) ||
+        (Op == AMDGPU::V_MUL_F16_e64 && MFI->getMode().FP64FP16Denormals))
       return std::make_pair(nullptr, SIOutMods::NONE);
 
     const MachineOperand *RegOp = nullptr;
     const MachineOperand *ImmOp = nullptr;
     const MachineOperand *Src0 = TII->getNamedOperand(MI, AMDGPU::OpName::src0);
     const MachineOperand *Src1 = TII->getNamedOperand(MI, AMDGPU::OpName::src1);
     if (Src0->isImm()) {
       ImmOp = Src0;
if (OMod == SIOutMods::NONE ||
         TII->hasModifiersSet(MI, AMDGPU::OpName::clamp))
       return std::make_pair(nullptr, SIOutMods::NONE);
 
     return std::make_pair(RegOp, OMod);
   }
   case AMDGPU::V_ADD_F32_e64:
   case AMDGPU::V_ADD_F16_e64: {
     // If output denormals are enabled, omod is ignored.
-    if ((Op == AMDGPU::V_ADD_F32_e64 && ST->hasFP32Denormals()) ||
-        (Op == AMDGPU::V_ADD_F16_e64 && ST->hasFP16Denormals()))
+    if ((Op == AMDGPU::V_ADD_F32_e64 && MFI->getMode().FP32Denormals) ||
+        (Op == AMDGPU::V_ADD_F16_e64 && MFI->getMode().FP64FP16Denormals))
       return std::make_pair(nullptr, SIOutMods::NONE);
 
     // Look through the DAGCombiner canonicalization fmul x, 2 -> fadd x, x
     const MachineOperand *Src0 = TII->getNamedOperand(MI, AMDGPU::OpName::src0);
     const MachineOperand *Src1 = TII->getNamedOperand(MI, AMDGPU::OpName::src1);
 
     if (Src0->isReg() && Src1->isReg() && Src0->getReg() == Src1->getReg() &&
         Src0->getSubReg() == Src1->getSubReg() &&

llvm/lib/Target/AMDGPU/SIISelLowering.h

void computeKnownBitsForFrameIndex(const SDValue Op,
                                      const SelectionDAG &DAG,
                                      unsigned Depth = 0) const override;
 
   bool isSDNodeSourceOfDivergence(const SDNode *N,
     FunctionLoweringInfo *FLI, LegacyDivergenceAnalysis *DA) const override;
 
   bool isCanonicalized(SelectionDAG &DAG, SDValue Op,
                        unsigned MaxDepth = 5) const;
-  bool denormalsEnabledForType(EVT VT) const;
+  bool denormalsEnabledForType(const SelectionDAG &DAG, EVT VT) const;
 
   bool isKnownNeverNaNForTargetNode(SDValue Op,
                                     const SelectionDAG &DAG,
                                     bool SNaN = false,
                                     unsigned Depth = 0) const override;
   AtomicExpansionKind shouldExpandAtomicRMWInIR(AtomicRMWInst *) const override;
 
   virtual const TargetRegisterClass *

llvm/lib/Target/AMDGPU/SIISelLowering.cpp

static cl::opt<bool> EnableVGPRIndexMode(
   cl::desc("Use GPR indexing mode instead of movrel for vector indexing"),
   cl::init(false));
 
 static cl::opt<bool> DisableLoopAlignment(
   "amdgpu-disable-loop-alignment",
   cl::desc("Do not align and prefetch loops"),
   cl::init(false));
 
+static bool hasFP32Denormals(const MachineFunction &MF) {
+  const SIMachineFunctionInfo *Info = MF.getInfo<SIMachineFunctionInfo>();
+  return Info->getMode().FP32Denormals;
+}
+
+static bool hasFP64FP16Denormals(const MachineFunction &MF) {
+  const SIMachineFunctionInfo *Info = MF.getInfo<SIMachineFunctionInfo>();
+  return Info->getMode().FP64FP16Denormals;
+}
+
 static unsigned findFirstFreeSGPR(CCState &CCInfo) {
   unsigned NumSGPRs = AMDGPU::SGPR_32RegClass.getNumRegs();
   for (unsigned Reg = 0; Reg < NumSGPRs; ++Reg) {
     if (!CCInfo.isAllocated(AMDGPU::SGPR0 + Reg)) {
       return AMDGPU::SGPR0 + Reg;
     }
   }
   llvm_unreachable("Cannot allocate sgpr");
#endif
   setOperationAction(ISD::DEBUGTRAP, MVT::Other, Custom);
 
   if (Subtarget->has16BitInsts()) {
     setOperationAction(ISD::FLOG, MVT::f16, Custom);
     setOperationAction(ISD::FEXP, MVT::f16, Custom);
     setOperationAction(ISD::FLOG10, MVT::f16, Custom);
   }
 
-  // v_mad_f32 does not support denormals according to some sources.
-  if (!Subtarget->hasFP32Denormals())
-    setOperationAction(ISD::FMAD, MVT::f32, Legal);
+  // v_mad_f32 does not support denormals. We report it as unconditionally
+  // legal, and the context where it is formed will disallow it when fp32
+  // denormals are enabled.
 
   if (!Subtarget->hasBFI()) {
     // fcopysign can be done in a single instruction with BFI.
     setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
     setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
   }
 
   if (!Subtarget->hasBCNT(32))
if (Subtarget->has16BitInsts()) {
     // F16 - VOP2 Actions.
     setOperationAction(ISD::BR_CC, MVT::f16, Expand);
     setOperationAction(ISD::SELECT_CC, MVT::f16, Expand);
 
     setOperationAction(ISD::FDIV, MVT::f16, Custom);
 
     // F16 - VOP3 Actions.
     setOperationAction(ISD::FMA, MVT::f16, Legal);
-    if (!Subtarget->hasFP16Denormals() && STI.hasMadF16())
+    if (STI.hasMadF16())
       setOperationAction(ISD::FMAD, MVT::f16, Legal);
 
     for (MVT VT : {MVT::v2i16, MVT::v2f16, MVT::v4i16, MVT::v4f16}) {
       for (unsigned Op = 0; Op < ISD::BUILTIN_OP_END; ++Op) {
         switch (Op) {
         case ISD::LOAD:
         case ISD::STORE:
         case ISD::BUILD_VECTOR:

 // v_mad_mix* support a conversion from f16 to f32.
 //
 // There is only one special case when denormals are enabled we don't currently,
 // where this is OK to use.
 bool SITargetLowering::isFPExtFoldable(const SelectionDAG &DAG, unsigned Opcode,
                                        EVT DestVT, EVT SrcVT) const {
   return ((Opcode == ISD::FMAD && Subtarget->hasMadMixInsts()) ||
           (Opcode == ISD::FMA && Subtarget->hasFmaMixInsts())) &&
-         DestVT.getScalarType() == MVT::f32 && !Subtarget->hasFP32Denormals() &&
-         SrcVT.getScalarType() == MVT::f16;
+    DestVT.getScalarType() == MVT::f32 &&
+    SrcVT.getScalarType() == MVT::f16 &&
+    !hasFP32Denormals(DAG.getMachineFunction());
 }
 
 bool SITargetLowering::isShuffleMaskLegal(ArrayRef<int>, EVT) const {
   // SI has some legal vector types, but no legal vector operations. Say no
   // shuffles are legal in order to prefer scalarizing some vector operations.
   return false;
 }
 
bool SITargetLowering::isFMAFasterThanFMulAndFAdd(const MachineFunction &MF,
   VT = VT.getScalarType();
 
   switch (VT.getSimpleVT().SimpleTy) {
   case MVT::f32: {
     // This is as fast on some subtargets. However, we always have full rate f32
     // mad available which returns the same result as the separate operations
     // which we should prefer over fma. We can't use this if we want to support
     // denormals, so only report this in these cases.
-    if (Subtarget->hasFP32Denormals())
+    if (hasFP32Denormals(MF))
       return Subtarget->hasFastFMAF32() || Subtarget->hasDLInsts();
 
     // If the subtarget has v_fmac_f32, that's just as good as v_mac_f32.
     return Subtarget->hasFastFMAF32() && Subtarget->hasDLInsts();
   }
   case MVT::f64:
     return true;
   case MVT::f16:
-    return Subtarget->has16BitInsts() && Subtarget->hasFP16Denormals();
+    return Subtarget->has16BitInsts() && hasFP64FP16Denormals(MF);
   default:
     break;
   }
 
   return false;
 }
 
 bool SITargetLowering::isFMADLegalForFAddFSub(const SelectionDAG &DAG,
                                               const SDNode *N) const {
   // TODO: Check future ftz flag
   // v_mad_f32/v_mac_f32 do not support denormals.
   EVT VT = N->getValueType(0);
   if (VT == MVT::f32)
-    return !Subtarget->hasFP32Denormals();
-  if (VT == MVT::f16)
-    return !Subtarget->hasFP16Denormals() && Subtarget->hasMadF16();
+    return !hasFP32Denormals(DAG.getMachineFunction());
+  if (VT == MVT::f16) {
+    return Subtarget->hasMadF16() &&
+           !hasFP64FP16Denormals(DAG.getMachineFunction());
+  }
 
   return false;
 }
 
 //===----------------------------------------------------------------------===//
 // Custom DAG Lowering Operations
 //===----------------------------------------------------------------------===//
 
SDValue SITargetLowering::lowerFastUnsafeFDIV(SDValue Op,
                                               SelectionDAG &DAG) const {
   SDLoc SL(Op);
   SDValue LHS = Op.getOperand(0);
   SDValue RHS = Op.getOperand(1);
   EVT VT = Op.getValueType();
   const SDNodeFlags Flags = Op->getFlags();
   bool Unsafe = DAG.getTarget().Options.UnsafeFPMath || Flags.hasAllowReciprocal();
 
-  if (!Unsafe && VT == MVT::f32 && Subtarget->hasFP32Denormals())
+  if (!Unsafe && VT == MVT::f32 && hasFP32Denormals(DAG.getMachineFunction()))
     return SDValue();
 
   if (const ConstantFPSDNode *CLHS = dyn_cast<ConstantFPSDNode>(LHS)) {
     if (Unsafe || VT == MVT::f32 || VT == MVT::f16) {
       if (CLHS->isExactlyValue(1.0)) {
         // v_rcp_f32 and v_rsq_f32 do not support denormals, and according to
         // the CI documentation has a worst case error of 1 ulp.
         // OpenCL requires <= 2.5 ulp for 1.0 / x, so it should always be OK to
SDValue SITargetLowering::lowerFDIV_FAST(SDValue Op, SelectionDAG &DAG) const {
   return DAG.getNode(ISD::FMUL, SL, MVT::f32, r3, Mul);
 }
 
 // Returns immediate value for setting the F32 denorm mode when using the
 // S_DENORM_MODE instruction.
 static const SDValue getSPDenormModeValue(int SPDenormMode, SelectionDAG &DAG,
                                           const SDLoc &SL, const GCNSubtarget *ST) {
   assert(ST->hasDenormModeInst() && "Requires S_DENORM_MODE");
-  int DPDenormModeDefault = ST->hasFP64Denormals()
+  int DPDenormModeDefault = hasFP64FP16Denormals(DAG.getMachineFunction())
                                 ? FP_DENORM_FLUSH_NONE
                                 : FP_DENORM_FLUSH_IN_FLUSH_OUT;
 
   int Mode = SPDenormMode | (DPDenormModeDefault << 2);
   return DAG.getTargetConstant(Mode, SL, MVT::i32);
 }
 
 SDValue SITargetLowering::LowerFDIV32(SDValue Op, SelectionDAG &DAG) const {
SDValue SITargetLowering::LowerFDIV32(SDValue Op, SelectionDAG &DAG) const {
   SDValue NegDivScale0 = DAG.getNode(ISD::FNEG, SL, MVT::f32,
                                      DenominatorScaled);
 
   const unsigned Denorm32Reg = AMDGPU::Hwreg::ID_MODE |
                                (4 << AMDGPU::Hwreg::OFFSET_SHIFT_) |
                                (1 << AMDGPU::Hwreg::WIDTH_M1_SHIFT_);
   const SDValue BitField = DAG.getTargetConstant(Denorm32Reg, SL, MVT::i16);
 
-  if (!Subtarget->hasFP32Denormals()) {
+  const bool HasFP32Denormals = hasFP32Denormals(DAG.getMachineFunction());
+
+  if (!HasFP32Denormals) {
     SDVTList BindParamVTs = DAG.getVTList(MVT::Other, MVT::Glue);
 
     SDValue EnableDenorm;
     if (Subtarget->hasDenormModeInst()) {
       const SDValue EnableDenormValue =
           getSPDenormModeValue(FP_DENORM_FLUSH_NONE, DAG, SL, Subtarget);
 
       EnableDenorm = DAG.getNode(AMDGPUISD::DENORM_MODE, SL, BindParamVTs,
SDValue SITargetLowering::LowerFDIV32(SDValue Op, SelectionDAG &DAG) const {
   SDValue Fma2 = getFPTernOp(DAG, ISD::FMA, SL, MVT::f32, NegDivScale0, Mul,
                              NumeratorScaled, Mul);
 
   SDValue Fma3 = getFPTernOp(DAG, ISD::FMA, SL, MVT::f32, Fma2, Fma1, Mul, Fma2);
 
   SDValue Fma4 = getFPTernOp(DAG, ISD::FMA, SL, MVT::f32, NegDivScale0, Fma3,
                              NumeratorScaled, Fma3);
 
-  if (!Subtarget->hasFP32Denormals()) {
-
+  if (!HasFP32Denormals) {
     SDValue DisableDenorm;
     if (Subtarget->hasDenormModeInst()) {
       const SDValue DisableDenormValue =
           getSPDenormModeValue(FP_DENORM_FLUSH_IN_FLUSH_OUT, DAG, SL, Subtarget);
 
       DisableDenorm = DAG.getNode(AMDGPUISD::DENORM_MODE, SL, MVT::Other,
                                   Fma4.getValue(1), DisableDenormValue,
                                   Fma4.getValue(2));
bool SITargetLowering::isCanonicalized(SelectionDAG &DAG, SDValue Op,
   unsigned Opcode = Op.getOpcode();
   if (Opcode == ISD::FCANONICALIZE)
     return true;
 
   if (auto *CFP = dyn_cast<ConstantFPSDNode>(Op)) {
     auto F = CFP->getValueAPF();
     if (F.isNaN() && F.isSignaling())
       return false;
-    return !F.isDenormal() || denormalsEnabledForType(Op.getValueType());
+    return !F.isDenormal() || denormalsEnabledForType(DAG, Op.getValueType());
   }
 
   // If source is a result of another standard FP operation it is already in
   // canonical form.
   if (MaxDepth == 0)
     return false;
 
   switch (Opcode) {
bool SITargetLowering::isCanonicalized(SelectionDAG &DAG, SDValue Op,
   case AMDGPUISD::FMAX3:
   case AMDGPUISD::FMIN3: {
     // FIXME: Shouldn't treat the generic operations different based these.
     // However, we aren't really required to flush the result from
     // minnum/maxnum..
 
     // snans will be quieted, so we only need to worry about denormals.
     if (Subtarget->supportsMinMaxDenormModes() ||
-        denormalsEnabledForType(Op.getValueType()))
+        denormalsEnabledForType(DAG, Op.getValueType()))
       return true;
 
     // Flushing may be required.
     // In pre-GFX9 targets V_MIN_F32 and others do not flush denorms. For such
     // targets need to check their input recursively.
 
     // FIXME: Does this apply with clamp? It's implemented with max.
     for (unsigned I = 0, E = Op.getNumOperands(); I != E; ++I) {
case Intrinsic::amdgcn_fdot2:
       return true;
     default:
       break;
     }
 
     LLVM_FALLTHROUGH;
   }
   default:
-    return denormalsEnabledForType(Op.getValueType()) &&
+    return denormalsEnabledForType(DAG, Op.getValueType()) &&
            DAG.isKnownNeverSNaN(Op);
   }
 
   llvm_unreachable("invalid operation");
 }
 
 // Constant fold canonicalize.
 SDValue SITargetLowering::getCanonicalConstantFP(
   SelectionDAG &DAG, const SDLoc &SL, EVT VT, const APFloat &C) const {
   // Flush denormals to 0 if not enabled.
-  if (C.isDenormal() && !denormalsEnabledForType(VT))
+  if (C.isDenormal() && !denormalsEnabledForType(DAG, VT))
     return DAG.getConstantFP(0.0, SL, VT);
 
   if (C.isNaN()) {
     APFloat CanonicalQNaN = APFloat::getQNaN(C.getSemantics());
     if (C.isSignaling()) {
       // Quiet a signaling NaN.
       // FIXME: Is this supposed to preserve payload bits?
       return DAG.getConstantFP(CanonicalQNaN, SL, VT);

 
 unsigned SITargetLowering::getFusedOpcode(const SelectionDAG &DAG,
                                           const SDNode *N0,
                                           const SDNode *N1) const {
   EVT VT = N0->getValueType(0);
 
   // Only do this if we are not trying to support denormals. v_mad_f32 does not
   // support denormals ever.
-  if (((VT == MVT::f32 && !Subtarget->hasFP32Denormals()) ||
-       (VT == MVT::f16 && !Subtarget->hasFP16Denormals() &&
+  if (((VT == MVT::f32 && !hasFP32Denormals(DAG.getMachineFunction())) ||
+       (VT == MVT::f16 && !hasFP64FP16Denormals(DAG.getMachineFunction()) &&
         getSubtarget()->hasMadF16())) &&
        isOperationLegal(ISD::FMAD, VT))
     return ISD::FMAD;
 
   const TargetOptions &Options = DAG.getTarget().Options;
   if ((Options.AllowFPOpFusion == FPOpFusion::Fast || Options.UnsafeFPMath ||
        (N0->getFlags().hasAllowContract() &&
         N1->getFlags().hasAllowContract())) &&
case ISD::INTRINSIC_WO_CHAIN:
       cast<ConstantSDNode>(N->getOperand(0))->getZExtValue());
     case ISD::INTRINSIC_W_CHAIN:
       return AMDGPU::isIntrinsicSourceOfDivergence(
       cast<ConstantSDNode>(N->getOperand(1))->getZExtValue());
   }
   return false;
 }
 
-bool SITargetLowering::denormalsEnabledForType(EVT VT) const {
+bool SITargetLowering::denormalsEnabledForType(const SelectionDAG &DAG,
+                                               EVT VT) const {
   switch (VT.getScalarType().getSimpleVT().SimpleTy) {
   case MVT::f32:
-    return Subtarget->hasFP32Denormals();
+    return hasFP32Denormals(DAG.getMachineFunction());
   case MVT::f64:
-    return Subtarget->hasFP64Denormals();
   case MVT::f16:
-    return Subtarget->hasFP16Denormals();
+    return hasFP64FP16Denormals(DAG.getMachineFunction());
   default:
     return false;
   }
 }
 
 bool SITargetLowering::isKnownNeverNaNForTargetNode(SDValue Op,
                                                     const SelectionDAG &DAG,
                                                     bool SNaN,

llvm/lib/Target/AMDGPU/SIMachineFunctionInfo.h

class SIMachineFunctionInfo final : public AMDGPUMachineFunction {
   // the ScratchWaveOffsetReg.
   unsigned FrameOffsetReg = AMDGPU::FP_REG;
 
   // Top of the stack SGPR offset derived from the ScratchWaveOffsetReg.
   unsigned StackPtrOffsetReg = AMDGPU::SP_REG;
 
   AMDGPUFunctionArgInfo ArgInfo;
 
-  // State of MODE register, assumed FP mode.
-  AMDGPU::SIModeRegisterDefaults Mode;
-
   // Graphics info.
   unsigned PSInputAddr = 0;
   unsigned PSInputEnable = 0;
 
   /// Number of bytes of arguments this function has on the stack. If the callee
   /// is expected to restore the argument stack this should be a multiple of 16,
   /// all usable during a tail call.
   ///
public:
   }
 
   MCPhysReg getVGPRToAGPRSpill(int FrameIndex, unsigned Lane) const {
     auto I = VGPRToAGPRSpills.find(FrameIndex);
     return (I == VGPRToAGPRSpills.end()) ? (MCPhysReg)AMDGPU::NoRegister
                                          : I->second.Lanes[Lane];
   }
 
-  AMDGPU::SIModeRegisterDefaults getMode() const {
-    return Mode;
-  }
-
   bool haveFreeLanesForSGPRSpill(const MachineFunction &MF,
                                  unsigned NumLane) const;
   bool allocateSGPRSpillToVGPR(MachineFunction &MF, int FI);
   bool allocateVGPRSpillToAGPR(MachineFunction &MF, int FI, bool isAGPRtoVGPR);
   void removeDeadFrameIndices(MachineFrameInfo &MFI);
 
   bool hasCalculatedTID() const { return TIDReg != 0; };
   unsigned getTIDReg() const { return TIDReg; };

llvm/lib/Target/AMDGPU/SIMachineFunctionInfo.cpp

 #include <vector>
 
 #define MAX_LANES 64
 
 using namespace llvm;
 
 SIMachineFunctionInfo::SIMachineFunctionInfo(const MachineFunction &MF)
   : AMDGPUMachineFunction(MF),
-    Mode(MF.getFunction()),
     PrivateSegmentBuffer(false),
     DispatchPtr(false),
     QueuePtr(false),
     KernargSegmentPtr(false),
     DispatchID(false),
     FlatScratchInit(false),
     WorkGroupIDX(false),
     WorkGroupIDY(false),

llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.h

struct SIModeRegisterDefaults {
   bool FP64FP16Denormals : 1;
 
   SIModeRegisterDefaults() :
     IEEE(true),
     DX10Clamp(true),
     FP32Denormals(true),
     FP64FP16Denormals(true) {}
 
-  SIModeRegisterDefaults(const Function &F);
+  // FIXME: Should not depend on the subtarget
+  SIModeRegisterDefaults(const Function &F, const GCNSubtarget &ST);
 
   static SIModeRegisterDefaults getDefaultForCallingConv(CallingConv::ID CC) {
     const bool IsCompute = AMDGPU::isCompute(CC);
 
     SIModeRegisterDefaults Mode;
     Mode.DX10Clamp = true;
     Mode.IEEE = IsCompute;
     Mode.FP32Denormals = false; // FIXME: Should be on by default.
     Mode.FP64FP16Denormals = true;
     return Mode;
   }
 
   bool operator ==(const SIModeRegisterDefaults Other) const {
     return IEEE == Other.IEEE && DX10Clamp == Other.DX10Clamp &&
            FP32Denormals == Other.FP32Denormals &&
            FP64FP16Denormals == Other.FP64FP16Denormals;
   }
 
+  /// Returns true if a flag is compatible if it's enabled in the callee, but
+  /// disabled in the caller.
+  static bool oneWayCompatible(bool CallerMode, bool CalleeMode) {
+    return CallerMode == CalleeMode || (CallerMode && !CalleeMode);
+  }
+
   // FIXME: Inlining should be OK for dx10-clamp, since the caller's mode should
   // be able to override.
   bool isInlineCompatible(SIModeRegisterDefaults CalleeMode) const {
-    return *this == CalleeMode;
+    if (DX10Clamp != CalleeMode.DX10Clamp)
+      return false;
+    if (IEEE != CalleeMode.IEEE)
+      return false;
+
+    // Allow inlining denormals enabled into denormals flushed functions.
+    return oneWayCompatible(FP64FP16Denormals, CalleeMode.FP64FP16Denormals) &&
+           oneWayCompatible(FP32Denormals, CalleeMode.FP32Denormals);
   }
 };
 
 } // end namespace AMDGPU
 } // end namespace llvm
 
 #endif // LLVM_LIB_TARGET_AMDGPU_UTILS_AMDGPUBASEINFO_H

llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.cpp

if (Overflow > 0 &&
       Subtarget->getGeneration() <= AMDGPUSubtarget::SEA_ISLANDS)
     return false;
 
   ImmOffset = Imm;
   SOffset = Overflow;
   return true;
 }
 
-SIModeRegisterDefaults::SIModeRegisterDefaults(const Function &F) {
+SIModeRegisterDefaults::SIModeRegisterDefaults(const Function &F,
+                                               const GCNSubtarget &ST) {
   *this = getDefaultForCallingConv(F.getCallingConv());
 
   StringRef IEEEAttr = F.getFnAttribute("amdgpu-ieee").getValueAsString();
   if (!IEEEAttr.empty())
     IEEE = IEEEAttr == "true";
 
   StringRef DX10ClampAttr
     = F.getFnAttribute("amdgpu-dx10-clamp").getValueAsString();
   if (!DX10ClampAttr.empty())
     DX10Clamp = DX10ClampAttr == "true";
+
+  FP32Denormals = ST.hasFP32Denormals(F);
+  FP64FP16Denormals = ST.hasFP64FP16Denormals(F);
 }
 
 namespace {
 
 struct SourceOfDivergence {
   unsigned Intr;
 };
 const SourceOfDivergence *lookupSourceOfDivergence(unsigned Intr);

llvm/test/CodeGen/AMDGPU/GlobalISel/inst-select-fcanonicalize.mir

machineFunctionInfo:
   mode:
     fp64-fp16-denormals: false
 
 body: |
   bb.0:
     liveins: $vgpr0
     ; GFX9-LABEL: name: fcanonicalize_f16_flush
     ; GFX9: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
-    ; GFX9: [[V_MAX_F16_e64_:%[0-9]+]]:vgpr_32 = V_MAX_F16_e64 0, [[COPY]], 0, [[COPY]], 0, 0, implicit $exec
-    ; GFX9: S_ENDPGM 0, implicit [[V_MAX_F16_e64_]]
+    ; GFX9: [[V_MUL_F16_e64_:%[0-9]+]]:vgpr_32 = V_MUL_F16_e64 0, 15360, 0, [[COPY]], 0, 0, implicit $exec
+    ; GFX9: S_ENDPGM 0, implicit [[V_MUL_F16_e64_]]
     %0:vgpr(s32) = COPY $vgpr0
     %1:vgpr(s16) = G_TRUNC %0
     %2:vgpr(s16) = G_FCANONICALIZE %1
     S_ENDPGM 0, implicit %2
 ...
 
 ---
 
 name:            fcanonicalize_f32_denorm
 legalized:       true
 regBankSelected: true
 machineFunctionInfo:
   mode:
     fp32-denormals: true
 
 body: |
   bb.0:
     liveins: $vgpr0
 
     ; GFX9-LABEL: name: fcanonicalize_f32_denorm
     ; GFX9: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
-    ; GFX9: [[V_MUL_F32_e64_:%[0-9]+]]:vgpr_32 = V_MUL_F32_e64 0, 1065353216, 0, [[COPY]], 0, 0, implicit $exec
-    ; GFX9: S_ENDPGM 0, implicit [[V_MUL_F32_e64_]]
+    ; GFX9: [[V_MAX_F32_e64_:%[0-9]+]]:vgpr_32 = V_MAX_F32_e64 0, [[COPY]], 0, [[COPY]], 0, 0, implicit $exec
+    ; GFX9: S_ENDPGM 0, implicit [[V_MAX_F32_e64_]]
     %0:vgpr(s32) = COPY $vgpr0
     %1:vgpr(s32) = G_FCANONICALIZE %0
     S_ENDPGM 0, implicit %1
 ...
 
 ---
 
 name:            fcanonicalize_f32_flush
mode:
     fp64-fp16-denormals: false
 
 body: |
   bb.0:
     liveins: $vgpr0_vgpr1
 
     ; GFX9-LABEL: name: fcanonicalize_f64_flush
     ; GFX9: [[COPY:%[0-9]+]]:vreg_64 = COPY $vgpr0_vgpr1
-    ; GFX9: [[V_MAX_F64_:%[0-9]+]]:vreg_64 = V_MAX_F64 0, [[COPY]], 0, [[COPY]], 0, 0, implicit $exec
-    ; GFX9: S_ENDPGM 0, implicit [[V_MAX_F64_]]
+    ; GFX9: [[V_MUL_F64_:%[0-9]+]]:vreg_64 = V_MUL_F64 0, 4607182418800017408, 0, [[COPY]], 0, 0, implicit $exec
+    ; GFX9: S_ENDPGM 0, implicit [[V_MUL_F64_]]
     %0:vgpr(s64) = COPY $vgpr0_vgpr1
     %1:vgpr(s64) = G_FCANONICALIZE %0
     S_ENDPGM 0, implicit %1
 ...
 
 ---
 
 name:            fcanonicalize_fabs_f32_denorm
 legalized:       true
 regBankSelected: true
 machineFunctionInfo:
   mode:
     fp32-denormals: true
 
 body: |
   bb.0:
     liveins: $vgpr0
     ; GFX9-LABEL: name: fcanonicalize_fabs_f32_denorm
     ; GFX9: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
-    ; GFX9: [[V_MUL_F32_e64_:%[0-9]+]]:vgpr_32 = V_MUL_F32_e64 0, 1065353216, 2, [[COPY]], 0, 0, implicit $exec
-    ; GFX9: S_ENDPGM 0, implicit [[V_MUL_F32_e64_]]
+    ; GFX9: [[V_MAX_F32_e64_:%[0-9]+]]:vgpr_32 = V_MAX_F32_e64 2, [[COPY]], 2, [[COPY]], 0, 0, implicit $exec
+    ; GFX9: S_ENDPGM 0, implicit [[V_MAX_F32_e64_]]
     %0:vgpr(s32) = COPY $vgpr0
     %1:vgpr(s32) = G_FABS %0
     %2:vgpr(s32) = G_FCANONICALIZE %1
     S_ENDPGM 0, implicit %2
 ...
 
 
 ---
machineFunctionInfo:
   mode:
     fp32-denormals: true
 
 body: |
   bb.0:
     liveins: $vgpr0
     ; GFX9-LABEL: name: fcanonicalize_fneg_f32_denorm
     ; GFX9: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
-    ; GFX9: [[V_MUL_F32_e64_:%[0-9]+]]:vgpr_32 = V_MUL_F32_e64 0, 3212836864, 0, [[COPY]], 0, 0, implicit $exec
-    ; GFX9: S_ENDPGM 0, implicit [[V_MUL_F32_e64_]]
+    ; GFX9: [[V_MAX_F32_e64_:%[0-9]+]]:vgpr_32 = V_MAX_F32_e64 1, [[COPY]], 1, [[COPY]], 0, 0, implicit $exec
+    ; GFX9: S_ENDPGM 0, implicit [[V_MAX_F32_e64_]]
     %0:vgpr(s32) = COPY $vgpr0
     %1:vgpr(s32) = G_FNEG %0
     %2:vgpr(s32) = G_FCANONICALIZE %1
     S_ENDPGM 0, implicit %2
 ...
 
 ---
 

 
 body: |
   bb.0:
     liveins: $vgpr0
     ; GFX9-LABEL: name: fcanonicalize_fneg_fabs_f32_denorm
     ; GFX9: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
     ; GFX9: [[S_MOV_B32_:%[0-9]+]]:sreg_32 = S_MOV_B32 2147483648
     ; GFX9: [[V_XOR_B32_e32_:%[0-9]+]]:vgpr_32 = V_XOR_B32_e32 [[S_MOV_B32_]], [[COPY]], implicit $exec
-    ; GFX9: [[V_MUL_F32_e64_:%[0-9]+]]:vgpr_32 = V_MUL_F32_e64 0, 1065353216, 2, [[V_XOR_B32_e32_]], 0, 0, implicit $exec
-    ; GFX9: S_ENDPGM 0, implicit [[V_MUL_F32_e64_]]
+    ; GFX9: [[V_MAX_F32_e64_:%[0-9]+]]:vgpr_32 = V_MAX_F32_e64 2, [[V_XOR_B32_e32_]], 2, [[V_XOR_B32_e32_]], 0, 0, implicit $exec
+    ; GFX9: S_ENDPGM 0, implicit [[V_MAX_F32_e64_]]
     %0:vgpr(s32) = COPY $vgpr0
     %1:vgpr(s32) = G_FNEG %0
     %2:vgpr(s32) = G_FABS %1
     %3:vgpr(s32) = G_FCANONICALIZE %2
     S_ENDPGM 0, implicit %3
 ...
 
 ---