Reapply "InstCombine: Introduce SimplifyDemandedUseFPClass"" (#74056)

This reverts commit ef38833.

The referenced issue violates the spec for finite-only math only by
using a return value for a constant infinity. If the interpretation
is results and arguments cannot violate nofpclass, then any
std::numeric_limits<T>::infinity() result is invalid under
-ffinite-math-only. Without this interpretation the utility of
nofpclass is slashed.

llvm/include/llvm/Analysis/ValueTracking.h

@@ -248,6 +248,10 @@ struct KnownFPClass {
   /// definitely set or false if the sign bit is definitely unset.
   std::optional<bool> SignBit;
 
+  bool operator==(KnownFPClass Other) const {
+    return KnownFPClasses == Other.KnownFPClasses && SignBit == Other.SignBit;
+  }
+
   /// Return true if it's known this can never be one of the mask entries.
   bool isKnownNever(FPClassTest Mask) const {
     return (KnownFPClasses & Mask) == fcNone;

llvm/lib/Transforms/InstCombine/InstCombineInternal.h

@@ -566,6 +566,15 @@ class LLVM_LIBRARY_VISIBILITY InstCombinerImpl final
                                     APInt &PoisonElts, unsigned Depth = 0,
                                     bool AllowMultipleUsers = false) override;
 
+  /// Attempts to replace V with a simpler value based on the demanded
+  /// floating-point classes
+  Value *SimplifyDemandedUseFPClass(Value *V, FPClassTest DemandedMask,
+                                    KnownFPClass &Known, unsigned Depth,
+                                    Instruction *CxtI);
+  bool SimplifyDemandedFPClass(Instruction *I, unsigned Op,
+                               FPClassTest DemandedMask, KnownFPClass &Known,
+                               unsigned Depth = 0);
+
   /// Canonicalize the position of binops relative to shufflevector.
   Instruction *foldVectorBinop(BinaryOperator &Inst);
   Instruction *foldVectorSelect(SelectInst &Sel);

llvm/lib/Transforms/InstCombine/InstCombineSimplifyDemanded.cpp

@@ -1877,3 +1877,139 @@ Value *InstCombinerImpl::SimplifyDemandedVectorElts(Value *V,
 
   return MadeChange ? I : nullptr;
 }
+
+/// For floating-point classes that resolve to a single bit pattern, return that
+/// value.
+static Constant *getFPClassConstant(Type *Ty, FPClassTest Mask) {
+  switch (Mask) {
+  case fcPosZero:
+    return ConstantFP::getZero(Ty);
+  case fcNegZero:
+    return ConstantFP::getZero(Ty, true);
+  case fcPosInf:
+    return ConstantFP::getInfinity(Ty);
+  case fcNegInf:
+    return ConstantFP::getInfinity(Ty, true);
+  case fcNone:
+    return PoisonValue::get(Ty);
+  default:
+    return nullptr;
+  }
+}
+
+Value *InstCombinerImpl::SimplifyDemandedUseFPClass(
+    Value *V, const FPClassTest DemandedMask, KnownFPClass &Known,
+    unsigned Depth, Instruction *CxtI) {
+  assert(Depth <= MaxAnalysisRecursionDepth && "Limit Search Depth");
+  Type *VTy = V->getType();
+
+  assert(Known == KnownFPClass() && "expected uninitialized state");
+
+  if (DemandedMask == fcNone)
+    return isa<UndefValue>(V) ? nullptr : PoisonValue::get(VTy);
+
+  if (Depth == MaxAnalysisRecursionDepth)
+    return nullptr;
+
+  Instruction *I = dyn_cast<Instruction>(V);
+  if (!I) {
+    // Handle constants and arguments
+    Known = computeKnownFPClass(V, fcAllFlags, CxtI, Depth + 1);
+    Value *FoldedToConst =
+        getFPClassConstant(VTy, DemandedMask & Known.KnownFPClasses);
+    return FoldedToConst == V ? nullptr : FoldedToConst;
+  }
+
+  if (!I->hasOneUse())
+    return nullptr;
+
+  // TODO: Should account for nofpclass/FastMathFlags on current instruction
+  switch (I->getOpcode()) {
+  case Instruction::FNeg: {
+    if (SimplifyDemandedFPClass(I, 0, llvm::fneg(DemandedMask), Known,
+                                Depth + 1))
+      return I;
+    Known.fneg();
+    break;
+  }
+  case Instruction::Call: {
+    CallInst *CI = cast<CallInst>(I);
+    switch (CI->getIntrinsicID()) {
+    case Intrinsic::fabs:
+      if (SimplifyDemandedFPClass(I, 0, llvm::inverse_fabs(DemandedMask), Known,
+                                  Depth + 1))
+        return I;
+      Known.fabs();
+      break;
+    case Intrinsic::arithmetic_fence:
+      if (SimplifyDemandedFPClass(I, 0, DemandedMask, Known, Depth + 1))
+        return I;
+      break;
+    case Intrinsic::copysign: {
+      // Flip on more potentially demanded classes
+      const FPClassTest DemandedMaskAnySign = llvm::unknown_sign(DemandedMask);
+      if (SimplifyDemandedFPClass(I, 0, DemandedMaskAnySign, Known, Depth + 1))
+        return I;
+
+      if ((DemandedMask & fcPositive) == fcNone) {
+        // Roundabout way of replacing with fneg(fabs)
+        I->setOperand(1, ConstantFP::get(VTy, -1.0));
+        return I;
+      }
+
+      if ((DemandedMask & fcNegative) == fcNone) {
+        // Roundabout way of replacing with fabs
+        I->setOperand(1, ConstantFP::getZero(VTy));
+        return I;
+      }
+
+      KnownFPClass KnownSign =
+          computeKnownFPClass(I->getOperand(1), fcAllFlags, CxtI, Depth + 1);
+      Known.copysign(KnownSign);
+      break;
+    }
+    default:
+      Known = computeKnownFPClass(I, ~DemandedMask, CxtI, Depth + 1);
+      break;
+    }
+
+    break;
+  }
+  case Instruction::Select: {
+    KnownFPClass KnownLHS, KnownRHS;
+    if (SimplifyDemandedFPClass(I, 2, DemandedMask, KnownRHS, Depth + 1) ||
+        SimplifyDemandedFPClass(I, 1, DemandedMask, KnownLHS, Depth + 1))
+      return I;
+
+    if (KnownLHS.isKnownNever(DemandedMask))
+      return I->getOperand(2);
+    if (KnownRHS.isKnownNever(DemandedMask))
+      return I->getOperand(1);
+
+    // TODO: Recognize clamping patterns
+    Known = KnownLHS | KnownRHS;
+    break;
+  }
+  default:
+    Known = computeKnownFPClass(I, ~DemandedMask, CxtI, Depth + 1);
+    break;
+  }
+
+  return getFPClassConstant(VTy, DemandedMask & Known.KnownFPClasses);
+}
+
+bool InstCombinerImpl::SimplifyDemandedFPClass(Instruction *I, unsigned OpNo,
+                                               FPClassTest DemandedMask,
+                                               KnownFPClass &Known,
+                                               unsigned Depth) {
+  Use &U = I->getOperandUse(OpNo);
+  Value *NewVal =
+      SimplifyDemandedUseFPClass(U.get(), DemandedMask, Known, Depth, I);
+  if (!NewVal)
+    return false;
+  if (Instruction *OpInst = dyn_cast<Instruction>(U))
+    salvageDebugInfo(*OpInst);
+
+  replaceUse(U, NewVal);
+  return true;
+}

llvm/lib/Transforms/InstCombine/InstructionCombining.cpp

@@ -142,6 +142,12 @@ static cl::opt<unsigned>
 MaxArraySize("instcombine-maxarray-size", cl::init(1024),
              cl::desc("Maximum array size considered when doing a combine"));
 
+// TODO: Remove this option
+static cl::opt<bool> EnableSimplifyDemandedUseFPClass(
+    "instcombine-simplify-demanded-fp-class",
+    cl::desc("Enable demanded floating-point class optimizations"),
+    cl::init(false));
+
 // FIXME: Remove this flag when it is no longer necessary to convert
 // llvm.dbg.declare to avoid inaccurate debug info. Setting this to false
 // increases variable availability at the cost of accuracy. Variables that
@@ -3105,8 +3111,25 @@ Instruction *InstCombinerImpl::visitFree(CallInst &FI, Value *Op) {
 }
 
 Instruction *InstCombinerImpl::visitReturnInst(ReturnInst &RI) {
-  // Nothing for now.
-  return nullptr;
+  if (!EnableSimplifyDemandedUseFPClass)
+    return nullptr;
+
+  Value *RetVal = RI.getReturnValue();
+  if (!RetVal || !AttributeFuncs::isNoFPClassCompatibleType(RetVal->getType()))
+    return nullptr;
+
+  Function *F = RI.getFunction();
+  FPClassTest ReturnClass = F->getAttributes().getRetNoFPClass();
+  if (ReturnClass == fcNone)
+    return nullptr;
+
+  KnownFPClass KnownClass;
+  Value *Simplified =
+      SimplifyDemandedUseFPClass(RetVal, ~ReturnClass, KnownClass, 0, &RI);
+  if (!Simplified)
+    return nullptr;
+
+  return ReturnInst::Create(RI.getContext(), Simplified);
 }
 
 // WARNING: keep in sync with SimplifyCFGOpt::simplifyUnreachable()!