Clang: emit llvm.minnum and llvm.maxnum with nsz always

See: #112852

We will define llvm.minnum and llvm.maxnum with +0.0>-0.0, by default,
while libc doesn't require it.

Author: wzssyqa

clang/lib/CodeGen/CGBuiltin.cpp

@@ -689,19 +689,20 @@ static Value *emitUnaryMaybeConstrainedFPBuiltin(CodeGenFunction &CGF,
 
 // Emit an intrinsic that has 2 operands of the same type as its result.
 // Depending on mode, this may be a constrained floating-point intrinsic.
-static Value *emitBinaryMaybeConstrainedFPBuiltin(CodeGenFunction &CGF,
-                                const CallExpr *E, unsigned IntrinsicID,
-                                unsigned ConstrainedIntrinsicID) {
+static Value *emitBinaryMaybeConstrainedFPBuiltin(
+    CodeGenFunction &CGF, const CallExpr *E, unsigned IntrinsicID,
+    unsigned ConstrainedIntrinsicID, llvm::FastMathFlags *FMF = nullptr) {
   llvm::Value *Src0 = CGF.EmitScalarExpr(E->getArg(0));
   llvm::Value *Src1 = CGF.EmitScalarExpr(E->getArg(1));
 
   CodeGenFunction::CGFPOptionsRAII FPOptsRAII(CGF, E);
   if (CGF.Builder.getIsFPConstrained()) {
     Function *F = CGF.CGM.getIntrinsic(ConstrainedIntrinsicID, Src0->getType());
-    return CGF.Builder.CreateConstrainedFPCall(F, { Src0, Src1 });
+    return CGF.Builder.CreateConstrainedFPCall(F, {Src0, Src1}, "",
+                                               std::nullopt, std::nullopt, FMF);
   } else {
     Function *F = CGF.CGM.getIntrinsic(IntrinsicID, Src0->getType());
-    return CGF.Builder.CreateCall(F, { Src0, Src1 });
+    return CGF.Builder.CreateCall(F, {Src0, Src1}, "", nullptr, FMF);
   }
 }
 
@@ -3203,10 +3204,13 @@ RValue CodeGenFunction::EmitBuiltinExpr(const GlobalDecl GD, unsigned BuiltinID,
     case Builtin::BI__builtin_fmaxf:
     case Builtin::BI__builtin_fmaxf16:
     case Builtin::BI__builtin_fmaxl:
-    case Builtin::BI__builtin_fmaxf128:
-      return RValue::get(emitBinaryMaybeConstrainedFPBuiltin(*this, E,
-                                   Intrinsic::maxnum,
-                                   Intrinsic::experimental_constrained_maxnum));
+    case Builtin::BI__builtin_fmaxf128: {
+      llvm::FastMathFlags FMF;
+      FMF.setNoSignedZeros();
+      return RValue::get(emitBinaryMaybeConstrainedFPBuiltin(
+          *this, E, Intrinsic::maxnum,
+          Intrinsic::experimental_constrained_maxnum, &FMF));
+    }
 
     case Builtin::BIfmin:
     case Builtin::BIfminf:
@@ -3215,10 +3219,13 @@ RValue CodeGenFunction::EmitBuiltinExpr(const GlobalDecl GD, unsigned BuiltinID,
     case Builtin::BI__builtin_fminf:
     case Builtin::BI__builtin_fminf16:
     case Builtin::BI__builtin_fminl:
-    case Builtin::BI__builtin_fminf128:
-      return RValue::get(emitBinaryMaybeConstrainedFPBuiltin(*this, E,
-                                   Intrinsic::minnum,
-                                   Intrinsic::experimental_constrained_minnum));
+    case Builtin::BI__builtin_fminf128: {
+      llvm::FastMathFlags FMF;
+      FMF.setNoSignedZeros();
+      return RValue::get(emitBinaryMaybeConstrainedFPBuiltin(
+          *this, E, Intrinsic::minnum,
+          Intrinsic::experimental_constrained_minnum, &FMF));
+    }
 
     case Builtin::BIfmaximum_num:
     case Builtin::BIfmaximum_numf:

clang/test/CodeGen/fmaxnum_fminnum_use_nsz.c

@@ -0,0 +1,33 @@
+// RUN: %clang_cc1 -triple x86_64 %s -emit-llvm -o - 2>&1 | FileCheck %s --check-prefix=CHECK
+
+float fminf (float, float);
+double fmin (double, double);
+long double fminl (long double, long double);
+float fmaxf (float, float);
+double fmax (double, double);
+long double fmaxl (long double, long double);
+
+// CHECK: call nsz float @llvm.minnum.f32
+float fmin1(float a, float b) {
+        return fminf(a, b);
+}
+// CHECK: call nsz double @llvm.minnum.f64
+float fmin2(double a, double b) {
+        return fmin(a, b);
+}
+// CHECK: call nsz x86_fp80 @llvm.minnum.f80
+float fmin3(long double a, long double b) {
+        return fminl(a, b);
+}
+// CHECK: call nsz float @llvm.maxnum.f32
+float fmax1(float a, float b) {
+        return fmaxf(a, b);
+}
+// CHECK: call nsz double @llvm.maxnum.f64
+float fmax2(double a, double b) {
+        return fmax(a, b);
+}
+// CHECK: call nsz x86_fp80 @llvm.maxnum.f80
+float fmax3(long double a, long double b) {
+        return fmaxl(a, b);
+}

llvm/include/llvm/IR/IRBuilder.h

@@ -2448,12 +2448,14 @@ class IRBuilderBase {
 public:
   CallInst *CreateCall(FunctionType *FTy, Value *Callee,
                        ArrayRef<Value *> Args = {}, const Twine &Name = "",
-                       MDNode *FPMathTag = nullptr) {
+                       MDNode *FPMathTag = nullptr,
+                       FastMathFlags *uFMF = nullptr) {
     CallInst *CI = CallInst::Create(FTy, Callee, Args, DefaultOperandBundles);
     if (IsFPConstrained)
       setConstrainedFPCallAttr(CI);
-    if (isa<FPMathOperator>(CI))
-      setFPAttrs(CI, FPMathTag, FMF);
+    if (isa<FPMathOperator>(CI)) {
+      setFPAttrs(CI, FPMathTag, uFMF ? (FMF | *uFMF) : FMF);
+    }
     return Insert(CI, Name);
   }
 
@@ -2469,9 +2471,10 @@ class IRBuilderBase {
   }
 
   CallInst *CreateCall(FunctionCallee Callee, ArrayRef<Value *> Args = {},
-                       const Twine &Name = "", MDNode *FPMathTag = nullptr) {
+                       const Twine &Name = "", MDNode *FPMathTag = nullptr,
+                       FastMathFlags *uFMF = nullptr) {
     return CreateCall(Callee.getFunctionType(), Callee.getCallee(), Args, Name,
-                      FPMathTag);
+                      FPMathTag, uFMF);
   }
 
   CallInst *CreateCall(FunctionCallee Callee, ArrayRef<Value *> Args,
@@ -2484,7 +2487,8 @@ class IRBuilderBase {
   CallInst *CreateConstrainedFPCall(
       Function *Callee, ArrayRef<Value *> Args, const Twine &Name = "",
       std::optional<RoundingMode> Rounding = std::nullopt,
-      std::optional<fp::ExceptionBehavior> Except = std::nullopt);
+      std::optional<fp::ExceptionBehavior> Except = std::nullopt,
+      FastMathFlags *FMF = nullptr);
 
   Value *CreateSelect(Value *C, Value *True, Value *False,
                       const Twine &Name = "", Instruction *MDFrom = nullptr);

llvm/lib/IR/IRBuilder.cpp

@@ -1036,7 +1036,7 @@ CallInst *IRBuilderBase::CreateConstrainedFPCmp(
 CallInst *IRBuilderBase::CreateConstrainedFPCall(
     Function *Callee, ArrayRef<Value *> Args, const Twine &Name,
     std::optional<RoundingMode> Rounding,
-    std::optional<fp::ExceptionBehavior> Except) {
+    std::optional<fp::ExceptionBehavior> Except, FastMathFlags *FMF) {
   llvm::SmallVector<Value *, 6> UseArgs;
 
   append_range(UseArgs, Args);
@@ -1045,7 +1045,7 @@ CallInst *IRBuilderBase::CreateConstrainedFPCall(
     UseArgs.push_back(getConstrainedFPRounding(Rounding));
   UseArgs.push_back(getConstrainedFPExcept(Except));
 
-  CallInst *C = CreateCall(Callee, UseArgs, Name);
+  CallInst *C = CreateCall(Callee, UseArgs, Name, nullptr, FMF);
   setConstrainedFPCallAttr(C);
   return C;
 }