[ValueTracking][InstCombine] Generalize ignoreSignBitOfZero/NaN to handle more cases (#141015)

dtcxzyw · web-flow · commit 6c86b7d7d8bc · 2025-05-28T19:17:51.000+08:00
This patch was originally part of llvm/llvm-project#139861. It generalizes `ignoreSignBitOfZero/NaN` to handle more instructions/intrinsics. BTW, I find it mitigates performance regressions caused by llvm/llvm-project#141010 (IR diff https://github.com/dtcxzyw/llvm-opt-benchmark/pull/2365/files). We don't need to propagate FMF from fcmp into select, since we can infer demanded properties from the user of select.
diff --git a/llvm/include/llvm/Analysis/ValueTracking.h b/llvm/include/llvm/Analysis/ValueTracking.h
@@ -300,6 +300,14 @@ LLVM_ABI bool isKnownNeverNaN(const Value *V, unsigned Depth,
 LLVM_ABI std::optional<bool>
 computeKnownFPSignBit(const Value *V, unsigned Depth, const SimplifyQuery &SQ);
 
+/// Return true if the sign bit of the FP value can be ignored by the user when
+/// the value is zero.
+bool canIgnoreSignBitOfZero(const Use &U);
+
+/// Return true if the sign bit of the FP value can be ignored by the user when
+/// the value is NaN.
+bool canIgnoreSignBitOfNaN(const Use &U);
+
 /// 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
diff --git a/llvm/lib/Analysis/ValueTracking.cpp b/llvm/lib/Analysis/ValueTracking.cpp
@@ -5937,6 +5937,114 @@ std::optional<bool> llvm::computeKnownFPSignBit(const Value *V, unsigned Depth,
   return Known.SignBit;
 }
 
+bool llvm::canIgnoreSignBitOfZero(const Use &U) {
+  auto *User = cast<Instruction>(U.getUser());
+  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 U.getOperandNo() == 0;
+      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::canIgnoreSignBitOfNaN(const Use &U) {
+  auto *User = cast<Instruction>(U.getUser());
+  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 User->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 U.getOperandNo() == 0;
+      // 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(U.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.
diff --git a/llvm/lib/Transforms/InstCombine/InstCombineSelect.cpp b/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,
@@ -2838,7 +2797,8 @@ static Instruction *foldSelectWithFCmpToFabs(SelectInst &SI,
     //       of NAN, but IEEE-754 specifies the signbit of NAN values with
     //       fneg/fabs operations.
     if (match(TrueVal, m_FSub(m_PosZeroFP(), m_Specific(X))) &&
-        (cast<FPMathOperator>(CondVal)->hasNoNaNs() || ignoreSignBitOfNaN(SI) ||
+        (cast<FPMathOperator>(CondVal)->hasNoNaNs() || SI.hasNoNaNs() ||
+         (SI.hasOneUse() && canIgnoreSignBitOfNaN(*SI.use_begin())) ||
          isKnownNeverNaN(X, /*Depth=*/0,
                          IC.getSimplifyQuery().getWithInstruction(
                              cast<Instruction>(CondVal))))) {
@@ -2885,7 +2845,11 @@ static Instruction *foldSelectWithFCmpToFabs(SelectInst &SI,
     // Note: We require "nnan" for this fold because fcmp ignores the signbit
     //       of NAN, but IEEE-754 specifies the signbit of NAN values with
     //       fneg/fabs operations.
-    if (!ignoreSignBitOfZero(SI) || !ignoreSignBitOfNaN(SI))
+    if (!SI.hasNoSignedZeros() &&
+        (!SI.hasOneUse() || !canIgnoreSignBitOfZero(*SI.use_begin())))
+      return nullptr;
+    if (!SI.hasNoNaNs() &&
+        (!SI.hasOneUse() || !canIgnoreSignBitOfNaN(*SI.use_begin())))
       return nullptr;
 
     if (Swap)
diff --git a/llvm/test/Transforms/InstCombine/fabs.ll b/llvm/test/Transforms/InstCombine/fabs.ll
