[ValueTracking][InstCombine] Generalize ignoreSignBitOfZero/NaN to handle more cases

llvm/include/llvm/Analysis/ValueTracking.h

@@ -330,6 +330,13 @@ bool isKnownNeverNaN(const Value *V, unsigned Depth, const SimplifyQuery &SQ);
 std::optional<bool> computeKnownFPSignBit(const Value *V, unsigned Depth,
                                           const SimplifyQuery &SQ);
 
+/// Return true if the sign bit of result can be ignored when the result is
+/// zero.
+bool ignoreSignBitOfZero(Instruction &I);
+
+/// Return true if the sign bit of result can be ignored when the result is NaN.
+bool ignoreSignBitOfNaN(Instruction &I);
+
 /// If the specified value can be set by repeating the same byte in memory,
 /// return the i8 value that it is represented with. This is true for all i8
 /// values obviously, but is also true for i32 0, i32 -1, i16 0xF0F0, double

llvm/lib/Analysis/ValueTracking.cpp

@@ -6357,6 +6357,130 @@ std::optional<bool> llvm::computeKnownFPSignBit(const Value *V, unsigned Depth,
   return Known.SignBit;
 }
 
+/// Return true if the sign bit of result can be ignored when the result is
+/// zero.
+bool llvm::ignoreSignBitOfZero(Instruction &I) {
+  if (auto *FPOp = dyn_cast<FPMathOperator>(&I))
+    if (FPOp->hasNoSignedZeros())
+      return true;
+
+  // Check if the sign bit is ignored by the only user.
+  if (!I.hasOneUse())
+    return false;
+  Instruction *User = I.user_back();
+  if (auto *FPOp = dyn_cast<FPMathOperator>(User)) {
+    if (FPOp->hasNoSignedZeros())
+      return true;
+  }
+
+  switch (User->getOpcode()) {
+  case Instruction::FPToSI:
+  case Instruction::FPToUI:
+    return true;
+  case Instruction::FCmp:
+    // fcmp treats both positive and negative zero as equal.
+    return true;
+  case Instruction::Call:
+    if (auto *II = dyn_cast<IntrinsicInst>(User)) {
+      switch (II->getIntrinsicID()) {
+      case Intrinsic::fabs:
+        return true;
+      case Intrinsic::copysign:
+        return II->getArgOperand(0) == &I;
+      case Intrinsic::is_fpclass:
+      case Intrinsic::vp_is_fpclass: {
+        auto Test =
+            static_cast<FPClassTest>(
+                cast<ConstantInt>(II->getArgOperand(1))->getZExtValue()) &
+            FPClassTest::fcZero;
+        return Test == FPClassTest::fcZero || Test == FPClassTest::fcNone;
+      }
+      default:
+        return false;
+      }
+    }
+    return false;
+  default:
+    return false;
+  }
+}
+
+bool llvm::ignoreSignBitOfNaN(Instruction &I) {
+  if (auto *FPOp = dyn_cast<FPMathOperator>(&I))
+    if (FPOp->hasNoNaNs())
+      return true;
+
+  // Check if the sign bit is ignored by the only user.
+  if (!I.hasOneUse())
+    return false;
+  Instruction *User = I.user_back();
+  if (auto *FPOp = dyn_cast<FPMathOperator>(User)) {
+    if (FPOp->hasNoNaNs())
+      return true;
+  }
+
+  switch (User->getOpcode()) {
+  case Instruction::FPToSI:
+  case Instruction::FPToUI:
+    return true;
+  // Proper FP math operations ignore the sign bit of NaN.
+  case Instruction::FAdd:
+  case Instruction::FSub:
+  case Instruction::FMul:
+  case Instruction::FDiv:
+  case Instruction::FRem:
+  case Instruction::FPTrunc:
+  case Instruction::FPExt:
+  case Instruction::FCmp:
+    return true;
+  // Bitwise FP operations should preserve the sign bit of NaN.
+  case Instruction::FNeg:
+  case Instruction::Select:
+  case Instruction::PHI:
+    return false;
+  case Instruction::Ret:
+    return I.getFunction()->getAttributes().getRetNoFPClass() &
+           FPClassTest::fcNan;
+  case Instruction::Call:
+  case Instruction::Invoke: {
+    if (auto *II = dyn_cast<IntrinsicInst>(User)) {
+      switch (II->getIntrinsicID()) {
+      case Intrinsic::fabs:
+        return true;
+      case Intrinsic::copysign:
+        return II->getArgOperand(0) == &I;
+      // Other proper FP math intrinsics ignore the sign bit of NaN.
+      case Intrinsic::maxnum:
+      case Intrinsic::minnum:
+      case Intrinsic::maximum:
+      case Intrinsic::minimum:
+      case Intrinsic::maximumnum:
+      case Intrinsic::minimumnum:
+      case Intrinsic::canonicalize:
+      case Intrinsic::fma:
+      case Intrinsic::fmuladd:
+      case Intrinsic::sqrt:
+      case Intrinsic::pow:
+      case Intrinsic::powi:
+      case Intrinsic::fptoui_sat:
+      case Intrinsic::fptosi_sat:
+      case Intrinsic::is_fpclass:
+      case Intrinsic::vp_is_fpclass:
+        return true;
+      default:
+        return false;
+      }
+    }
+
+    FPClassTest NoFPClass = cast<CallBase>(User)->getParamNoFPClass(
+        I.uses().begin()->getOperandNo());
+    return NoFPClass & FPClassTest::fcNan;
+  }
+  default:
+    return false;
+  }
+}
+
 Value *llvm::isBytewiseValue(Value *V, const DataLayout &DL) {
 
   // All byte-wide stores are splatable, even of arbitrary variables.

llvm/lib/Transforms/InstCombine/InstCombineSelect.cpp

@@ -2773,47 +2773,6 @@ Instruction *InstCombinerImpl::foldAndOrOfSelectUsingImpliedCond(Value *Op,
   return nullptr;
 }
 
-/// Return true if the sign bit of result can be ignored when the result is
-/// zero.
-static bool ignoreSignBitOfZero(Instruction &I) {
-  if (I.hasNoSignedZeros())
-    return true;
-
-  // Check if the sign bit is ignored by the only user.
-  if (!I.hasOneUse())
-    return false;
-  Instruction *User = I.user_back();
-
-  // fcmp treats both positive and negative zero as equal.
-  if (User->getOpcode() == Instruction::FCmp)
-    return true;
-
-  if (auto *FPOp = dyn_cast<FPMathOperator>(User))
-    return FPOp->hasNoSignedZeros();
-
-  return false;
-}
-
-/// Return true if the sign bit of result can be ignored when the result is NaN.
-static bool ignoreSignBitOfNaN(Instruction &I) {
-  if (I.hasNoNaNs())
-    return true;
-
-  // Check if the sign bit is ignored by the only user.
-  if (!I.hasOneUse())
-    return false;
-  Instruction *User = I.user_back();
-
-  // fcmp ignores the sign bit of NaN.
-  if (User->getOpcode() == Instruction::FCmp)
-    return true;
-
-  if (auto *FPOp = dyn_cast<FPMathOperator>(User))
-    return FPOp->hasNoNaNs();
-
-  return false;
-}
-
 // Canonicalize select with fcmp to fabs(). -0.0 makes this tricky. We need
 // fast-math-flags (nsz) or fsub with +0.0 (not fneg) for this to work.
 static Instruction *foldSelectWithFCmpToFabs(SelectInst &SI,