[ValueTracking] Compute known FPClass from dominating condition

Author: dtcxzyw

llvm/lib/Analysis/DomConditionCache.cpp

@@ -51,6 +51,9 @@ static void findAffectedValues(Value *Cond,
  // Handle (A + C1) u< C2, which is the canonical form of A > C3 && A < C4.
  if (match(A, m_Add(m_Value(X), m_ConstantInt())))
  AddAffected(X);
+ // Handle icmp slt/sgt (bitcast X to int) 0/-1
+ if (match(A, m_BitCast(m_Value(X))))
+ Affected.push_back(X);
  }
  }
 }

llvm/lib/Analysis/ValueTracking.cpp

@@ -4213,9 +4213,82 @@ llvm::fcmpImpliesClass(CmpInst::Predicate Pred, const Function &F, Value *LHS,
  return fcmpImpliesClass(Pred, F, LHS, *ConstRHS, LookThroughSrc);
 }
 
-static FPClassTest computeKnownFPClassFromAssumes(const Value *V,
- const SimplifyQuery &Q) {
- FPClassTest KnownFromAssume = fcAllFlags;
+static void computeKnownFPClassFromCond(const Value *V, Value *Cond,
+ bool CondIsTrue,
+ const Instruction *CxtI,
+ KnownFPClass &KnownFromContext) {
+ CmpInst::Predicate Pred;
+ Value *LHS;
+ Value *RHS;
+ uint64_t ClassVal = 0;
+ if (match(Cond, m_Cmp(Pred, m_Value(LHS), m_Value(RHS)))) {
+ if (CmpInst::isIntPredicate(Pred)) {
+ if (!match(LHS, m_BitCast(m_Specific(V))))
+ return;
+ Type *SrcType = V->getType();
+ Type *DstType = LHS->getType();
+
+ // Make sure the bitcast doesn't change between scalar and vector and
+ // doesn't change the number of vector elements.
+ if (SrcType->isVectorTy() == DstType->isVectorTy() &&
+ SrcType->getScalarSizeInBits() == DstType->getScalarSizeInBits()) {
+ // TODO: move IsSignBitCheck to ValueTracking
+ bool TrueIfSigned;
+ if ((Pred == ICmpInst::ICMP_SLT && match(RHS, m_Zero())) ||
+ (Pred == ICmpInst::ICMP_SLE && match(RHS, m_AllOnes())))
+ TrueIfSigned = true;
+ else if (Pred == ICmpInst::ICMP_SGT && match(RHS, m_AllOnes()) ||
+ (Pred == ICmpInst::ICMP_SGE && match(RHS, m_Zero())))
+ TrueIfSigned = false;
+ else
+ return;
+ if (TrueIfSigned == CondIsTrue)
+ KnownFromContext.signBitMustBeOne();
+ else
+ KnownFromContext.signBitMustBeZero();
+ }
+ } else {
+ const APFloat *CRHS;
+ if (match(RHS, m_APFloat(CRHS))) {
+ auto [CmpVal, MaskIfTrue, MaskIfFalse] = fcmpImpliesClass(
+ Pred, *CxtI->getParent()->getParent(), LHS, *CRHS, LHS != V);
+ if (CmpVal == V)
+ KnownFromContext.knownNot(~(CondIsTrue ? MaskIfTrue : MaskIfFalse));
+ }
+ }
+ } else if (match(Cond, m_Intrinsic<Intrinsic::is_fpclass>(
+ m_Value(LHS), m_ConstantInt(ClassVal)))) {
+ FPClassTest Mask = static_cast<FPClassTest>(ClassVal);
+ KnownFromContext.knownNot(CondIsTrue ? ~Mask : Mask);
+ }
+}
+
+static KnownFPClass computeKnownFPClassFromContext(const Value *V,
+ const SimplifyQuery &Q) {
+ KnownFPClass KnownFromContext;
+
+ if (!Q.CxtI)
+ return KnownFromContext;
+
+ if (Q.DC && Q.DT) {
+ // Handle dominating conditions.
+ for (BranchInst *BI : Q.DC->conditionsFor(V)) {
+ Value *Cond = BI->getCondition();
+
+ BasicBlockEdge Edge0(BI->getParent(), BI->getSuccessor(0));
+ if (Q.DT->dominates(Edge0, Q.CxtI->getParent()))
+ computeKnownFPClassFromCond(V, Cond, /*CondIsTrue=*/true, Q.CxtI,
+ KnownFromContext);
+
+ BasicBlockEdge Edge1(BI->getParent(), BI->getSuccessor(1));
+ if (Q.DT->dominates(Edge1, Q.CxtI->getParent()))
+ computeKnownFPClassFromCond(V, Cond, /*CondIsTrue=*/false, Q.CxtI,
+ KnownFromContext);
+ }
+ }
+
+ if (!Q.AC)
+ return KnownFromContext;
 
  // Try to restrict the floating-point classes based on information from
  // assumptions.
@@ -4233,25 +4306,11 @@ static FPClassTest computeKnownFPClassFromAssumes(const Value *V,
  if (!isValidAssumeForContext(I, Q.CxtI, Q.DT))
  continue;
 
- CmpInst::Predicate Pred;
- Value *LHS, *RHS;
- uint64_t ClassVal = 0;
- if (match(I->getArgOperand(0), m_FCmp(Pred, m_Value(LHS), m_Value(RHS)))) {
- const APFloat *CRHS;
- if (match(RHS, m_APFloat(CRHS))) {
- auto [CmpVal, MaskIfTrue, MaskIfFalse] =
- fcmpImpliesClass(Pred, *F, LHS, *CRHS, LHS != V);
- if (CmpVal == V)
- KnownFromAssume &= MaskIfTrue;
- }
- } else if (match(I->getArgOperand(0),
- m_Intrinsic<Intrinsic::is_fpclass>(
- m_Value(LHS), m_ConstantInt(ClassVal)))) {
- KnownFromAssume &= static_cast<FPClassTest>(ClassVal);
- }
+ computeKnownFPClassFromCond(V, I->getArgOperand(0), /*CondIsTrue=*/true,
+ Q.CxtI, KnownFromContext);
  }
 
- return KnownFromAssume;
+ return KnownFromContext;
 }
 
 void computeKnownFPClass(const Value *V, const APInt &DemandedElts,
@@ -4359,17 +4418,21 @@ void computeKnownFPClass(const Value *V, const APInt &DemandedElts,
  KnownNotFromFlags |= fcInf;
  }
 
- if (Q.AC) {
- FPClassTest AssumedClasses = computeKnownFPClassFromAssumes(V, Q);
- KnownNotFromFlags |= ~AssumedClasses;
- }
+ KnownFPClass AssumedClasses = computeKnownFPClassFromContext(V, Q);
+ KnownNotFromFlags |= ~AssumedClasses.KnownFPClasses;
 
  // We no longer need to find out about these bits from inputs if we can
  // assume this from flags/attributes.
  InterestedClasses &= ~KnownNotFromFlags;
 
  auto ClearClassesFromFlags = make_scope_exit([=, &Known] {
  Known.knownNot(KnownNotFromFlags);
+ if (!Known.SignBit && AssumedClasses.SignBit) {
+ if (*AssumedClasses.SignBit)
+ Known.signBitMustBeOne();
+ else
+ Known.signBitMustBeZero();
+ }
  });
 
  if (!Op)
@@ -5271,7 +5334,8 @@ void computeKnownFPClass(const Value *V, const APInt &DemandedElts,
 
  bool First = true;
 
- for (Value *IncValue : P->incoming_values()) {
+ for (const Use &U : P->operands()) {
+ Value *IncValue = U.get();
  // Skip direct self references.
  if (IncValue == P)
  continue;
@@ -5280,8 +5344,10 @@ void computeKnownFPClass(const Value *V, const APInt &DemandedElts,
  // Recurse, but cap the recursion to two levels, because we don't want
  // to waste time spinning around in loops. We need at least depth 2 to
  // detect known sign bits.
- computeKnownFPClass(IncValue, DemandedElts, InterestedClasses, KnownSrc,
- PhiRecursionLimit, Q);
+ computeKnownFPClass(
+ IncValue, DemandedElts, InterestedClasses, KnownSrc,
+ PhiRecursionLimit,
+ Q.getWithInstruction(P->getIncomingBlock(U)->getTerminator()));
 
  if (First) {
  Known = KnownSrc;

llvm/test/Transforms/InstCombine/fpclass-from-dom-cond.ll

@@ -16,8 +16,7 @@ define float @test_signbit_check(float %x, i1 %cond) {
 ; CHECK-NEXT: br label [[IF_END]]
 ; CHECK: if.end:
 ; CHECK-NEXT: [[VALUE:%.*]] = phi float [ [[FNEG]], [[IF_THEN1]] ], [ [[X]], [[IF_THEN2]] ], [ [[X]], [[IF_ELSE]] ]
-; CHECK-NEXT: [[RET:%.*]] = call float @llvm.fabs.f32(float [[VALUE]])
-; CHECK-NEXT: ret float [[RET]]
+; CHECK-NEXT: ret float [[VALUE]]
 ;
  %i32 = bitcast float %x to i32
  %cmp = icmp slt i32 %i32, 0