Clang: Add nsz to llvm.minnum and llvm.maxnum emitted from fmin and fmax #113133

wzssyqa · 2024-10-21T07:24:33Z

See: #112852

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

llvmbot · 2024-10-21T07:25:06Z

@llvm/pr-subscribers-backend-risc-v
@llvm/pr-subscribers-llvm-ir

@llvm/pr-subscribers-clang-codegen

Author: YunQiang Su (wzssyqa)

Changes

See: #112852

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

Full diff: https://github.com/llvm/llvm-project/pull/113133.diff

4 Files Affected:

(modified) clang/lib/CodeGen/CGBuiltin.cpp (+20-13)
(added) clang/test/CodeGen/fmaxnum_fminnum_use_nsz.c (+33)
(modified) llvm/include/llvm/IR/IRBuilder.h (+10-6)
(modified) llvm/lib/IR/IRBuilder.cpp (+2-2)

diff --git a/clang/lib/CodeGen/CGBuiltin.cpp b/clang/lib/CodeGen/CGBuiltin.cpp
index 28f28c70b5ae52..f2d6049908720b 100644
--- a/clang/lib/CodeGen/CGBuiltin.cpp
+++ b/clang/lib/CodeGen/CGBuiltin.cpp
@@ -510,19 +510,20 @@ 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);
   }
 }
 
@@ -2846,10 +2847,13 @@ 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:
@@ -2858,10 +2862,13 @@ 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:
diff --git a/clang/test/CodeGen/fmaxnum_fminnum_use_nsz.c b/clang/test/CodeGen/fmaxnum_fminnum_use_nsz.c
new file mode 100644
index 00000000000000..9798baf0432fea
--- /dev/null
+++ b/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);
+}
diff --git a/llvm/include/llvm/IR/IRBuilder.h b/llvm/include/llvm/IR/IRBuilder.h
index 23fd8350a29b3d..1baca4f003cad6 100644
--- a/llvm/include/llvm/IR/IRBuilder.h
+++ b/llvm/include/llvm/IR/IRBuilder.h
@@ -2438,12 +2438,14 @@ 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);
   }
 
@@ -2459,9 +2461,10 @@ 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,
@@ -2474,7 +2477,8 @@ 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);
diff --git a/llvm/lib/IR/IRBuilder.cpp b/llvm/lib/IR/IRBuilder.cpp
index f340f7aafdc76f..5feaf956b45a97 100644
--- a/llvm/lib/IR/IRBuilder.cpp
+++ b/llvm/lib/IR/IRBuilder.cpp
@@ -1031,7 +1031,7 @@ 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);
@@ -1040,7 +1040,7 @@ 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;
 }

llvmbot · 2024-10-21T07:25:07Z

@llvm/pr-subscribers-clang

Author: YunQiang Su (wzssyqa)

Changes

See: #112852

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

Full diff: https://github.com/llvm/llvm-project/pull/113133.diff

4 Files Affected:

(modified) clang/lib/CodeGen/CGBuiltin.cpp (+20-13)
(added) clang/test/CodeGen/fmaxnum_fminnum_use_nsz.c (+33)
(modified) llvm/include/llvm/IR/IRBuilder.h (+10-6)
(modified) llvm/lib/IR/IRBuilder.cpp (+2-2)

diff --git a/clang/lib/CodeGen/CGBuiltin.cpp b/clang/lib/CodeGen/CGBuiltin.cpp
index 28f28c70b5ae52..f2d6049908720b 100644
--- a/clang/lib/CodeGen/CGBuiltin.cpp
+++ b/clang/lib/CodeGen/CGBuiltin.cpp
@@ -510,19 +510,20 @@ 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);
   }
 }
 
@@ -2846,10 +2847,13 @@ 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:
@@ -2858,10 +2862,13 @@ 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:
diff --git a/clang/test/CodeGen/fmaxnum_fminnum_use_nsz.c b/clang/test/CodeGen/fmaxnum_fminnum_use_nsz.c
new file mode 100644
index 00000000000000..9798baf0432fea
--- /dev/null
+++ b/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);
+}
diff --git a/llvm/include/llvm/IR/IRBuilder.h b/llvm/include/llvm/IR/IRBuilder.h
index 23fd8350a29b3d..1baca4f003cad6 100644
--- a/llvm/include/llvm/IR/IRBuilder.h
+++ b/llvm/include/llvm/IR/IRBuilder.h
@@ -2438,12 +2438,14 @@ 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);
   }
 
@@ -2459,9 +2461,10 @@ 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,
@@ -2474,7 +2477,8 @@ 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);
diff --git a/llvm/lib/IR/IRBuilder.cpp b/llvm/lib/IR/IRBuilder.cpp
index f340f7aafdc76f..5feaf956b45a97 100644
--- a/llvm/lib/IR/IRBuilder.cpp
+++ b/llvm/lib/IR/IRBuilder.cpp
@@ -1031,7 +1031,7 @@ 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);
@@ -1040,7 +1040,7 @@ 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;
 }

arsenm

We should also add new builtins (with the minnum/maxnum names) to get the stronger 2019 behavior

clang/test/CodeGen/fmaxnum_fminnum_use_nsz.c

clang/lib/CodeGen/CGBuiltin.cpp

wzssyqa · 2024-11-25T06:26:05Z

@arsenm ping

jcranmer-intel

You'll want to merge the fast-math flags, so that compiling with -ffinite-math-only gets you nnan ninf nsz on the maxnum/minnum, rather than just nsz.

clang/test/CodeGen/fmaxnum_fminnum_use_nsz.c

wzssyqa · 2025-02-27T10:15:14Z

nsz is missing for __builtin_elementwise_min.
I am working on it.

arsenm · 2025-02-27T10:24:54Z

We should also have new builtins for the raw minnum / maxnum intrinsics (plus elementwise) to not get the nsz

arsenm · 2025-02-27T11:34:53Z

nsz is missing for __builtin_elementwise_min. I am working on it.

I'm not sure that's a good idea. For fmin/fmax it seems OK, but this doesn't take the libm name

wzssyqa · 2025-03-04T06:20:19Z

You'll want to merge the fast-math flags, so that compiling with -ffinite-math-only gets you nnan ninf nsz on the maxnum/minnum, rather than just nsz.

Done, for cc1, we use -menable-no-nans and -menable-no-infs.

clang/lib/CodeGen/CGBuiltin.cpp

clang/test/CodeGen/builtins-elementwise-math.c

efriedma-quic · 2025-04-27T20:31:14Z

Please propose a fix for the definition of nsz itself itself in LangRef; burying an exception to the general nsz rules in the middle of the definition of min/max is, at best, confusing.

wzssyqa · 2025-09-14T03:27:28Z

It takes too long to take this decision.
Since we cannot agree on the wording of nsz now, and in fact we had some words in maxnum/minnum about nsz.
I prefer we can continue with this PR.

wzssyqa · 2025-09-14T03:29:10Z

It takes too long to take this decision. Since we cannot agree on the wording of nsz now, and in fact we had some words in maxnum/minnum about nsz. I prefer we can continue with this PR.

We have got some complains: #138303

See: llvm#112852 We will define llvm.minnum and llvm.maxnum with +0.0>-0.0, by default, while libc doesn't require it. fix testcases -ffp-exception-behavior=strict add missing builtin test test auto vectorize fix test cases update testcase disable-llvm-passes fix elementswise fix some tests

wzssyqa · 2025-12-02T02:20:10Z

#168838

OK, now we decides to define llvm.minnum and llvm.maxnum as same as libm.
So we don't need it now. Let's close it.

arsenm · 2025-12-02T06:20:51Z

#168838

OK, now we decides to define llvm.minnum and llvm.maxnum as same as libm. So we don't need it now. Let's close it.

No, this should still be done

github-actions · 2025-12-02T09:13:41Z

🐧 Linux x64 Test Results

193631 tests passed
6242 tests skipped

✅ The build succeeded and all tests passed.

wzssyqa · 2025-12-04T09:12:23Z

@arsenm ping

fhahn

Do you have any data on the impact of the change?

wzssyqa · 2025-12-04T10:32:15Z

Do you have any data on the impact of the change?

It is used in TargetLowering::expandFMINNUM_FMAXNUM:

  if ((Node->getFlags().hasNoNaNs() ||
       (DAG.isKnownNeverNaN(Node->getOperand(0)) &&
        DAG.isKnownNeverNaN(Node->getOperand(1)))) &&
      (Node->getFlags().hasNoSignedZeros() ||
       DAG.isKnownNeverZeroFloat(Node->getOperand(0)) ||
       DAG.isKnownNeverZeroFloat(Node->getOperand(1)))) {
    unsigned IEEE2018Op =
        Node->getOpcode() == ISD::FMINNUM ? ISD::FMINIMUM : ISD::FMAXIMUM;
    if (isOperationLegalOrCustom(IEEE2018Op, VT))
      return DAG.getNode(IEEE2018Op, dl, VT, Node->getOperand(0),
                         Node->getOperand(1), Node->getFlags());
  }

We will add more expand rules there.

arsenm · 2025-12-04T10:59:42Z

Do you have any data on the impact of the change?

This change in isolation shouldn't change anything. In the absence of changing the minnum definition to have strict zero handling, adding nsz is a no-op. With strict signed zero handling, this would help avoid regressions in the lowering

clang/lib/CodeGen/CGBuiltin.cpp

clang/test/CodeGen/fmaxnum_fminnum_use_nsz.c

github-actions · 2025-12-05T01:58:14Z

✅ With the latest revision this PR passed the C/C++ code formatter.

arsenm · 2025-12-05T08:19:25Z

clang/lib/CodeGen/CGBuiltin.cpp

-      Result = Builder.CreateMinNum(Op0, Op1, /*FMFSource=*/nullptr, "elt.min");
+    } else {
+      FastMathFlags FMF;
+      FMF.setNoSignedZeros(true);


I don't think this should apply in the elementwise case.

I also think it was a mistake to allow floating point in elementwise min/max

See: #129207

In that PR, we planned to use the same naming scheme
__builtin_elementwise_max -> max (fmax)
__builtin_elementwise_maxnum -> maxnum

If we are planning to drop float support of __builtin_elementwise, it should be in another patchset.

Yes this should be discussed separately. Feedback from library authors has been that different builtins for floats/ints are a bit of a pain

But they are in fact, different operations. And you have many choices for which FP min/max.

Given this name doesn't have the historic fmin/fmax in it, I don't think this should take the fuzzy signed zero handling

In fact I am planning to add __builtin_elementwise_maximumnum after this PR is merged.

I agree this should not continue to use minnum/maxnum, though that change should not be part of this PR.

minnum/maxnum behavior has never been consistent across targets (or across scenarios). But we could also add __builtin_elementwise_maximumnum/__builtin_elementwise_minimnumnum

I agree this should not continue to use minnum/maxnum, though that change should not be part of this PR.

So can we merge this PR?

minnum/maxnum behavior has never been consistent across targets (or across scenarios).

In fact, all of the architectures that claims implement IEEE754-2008, has the same behavior:
AArch64, MIPSr6, LoongArch, PowerPC/VSX
That's why I'd plan to define minnum/maxnum as the same as these architectures.

But we could also add __builtin_elementwise_maximumnum/__builtin_elementwise_minimnumnum

I will do it.

So can we merge this PR?

The discussion in #137567 is still unresolved.

#137567 is about all opcodes, and in fact for min/max we have done with
#112852

llvmbot added clang Clang issues not falling into any other category clang:codegen IR generation bugs: mangling, exceptions, etc. llvm:ir labels Oct 21, 2024

wzssyqa requested a review from arsenm October 21, 2024 07:25

arsenm added the floating-point Floating-point math label Oct 21, 2024

arsenm reviewed Oct 21, 2024

View reviewed changes

clang/test/CodeGen/fmaxnum_fminnum_use_nsz.c Outdated Show resolved Hide resolved

arsenm reviewed Oct 21, 2024

View reviewed changes

clang/test/CodeGen/fmaxnum_fminnum_use_nsz.c Outdated Show resolved Hide resolved

arsenm reviewed Oct 21, 2024

View reviewed changes

clang/test/CodeGen/fmaxnum_fminnum_use_nsz.c Outdated Show resolved Hide resolved

clang/lib/CodeGen/CGBuiltin.cpp Outdated Show resolved Hide resolved

wzssyqa force-pushed the clang-fmin-with-nsz branch from e3bff67 to 8655cb0 Compare October 23, 2024 03:11

wzssyqa requested a review from arsenm October 23, 2024 04:48

jcranmer-intel reviewed Dec 2, 2024

View reviewed changes

arsenm reviewed Dec 3, 2024

View reviewed changes

wzssyqa force-pushed the clang-fmin-with-nsz branch from 105c9a7 to a3f3567 Compare February 27, 2025 10:06

wzssyqa marked this pull request as draft February 27, 2025 10:13

wzssyqa force-pushed the clang-fmin-with-nsz branch from 8b63af7 to ce675ee Compare March 4, 2025 06:15

wzssyqa marked this pull request as ready for review March 4, 2025 06:16

wzssyqa requested review from arsenm and jcranmer-intel March 4, 2025 06:16

arsenm changed the title ~~Clang: emit llvm.minnum and llvm.maxnum with nsz always~~ Clang: Add nsz to llvm.minnum and llvm.maxnum emitted from fmin and fmax Mar 4, 2025

arsenm reviewed Mar 4, 2025

View reviewed changes

clang/lib/CodeGen/CGBuiltin.cpp Outdated Show resolved Hide resolved

wzssyqa requested a review from arsenm March 4, 2025 07:49

arsenm reviewed Mar 4, 2025

View reviewed changes

clang/test/CodeGen/builtins-elementwise-math.c Outdated Show resolved Hide resolved

wzssyqa mentioned this pull request Apr 28, 2025

LangRef: Clarify behaviors of nsz in fast math flag #137567

Open

wzssyqa force-pushed the clang-fmin-with-nsz branch from bb4a5fd to 2170e67 Compare September 14, 2025 07:12

llvmbot added the backend:RISC-V label Sep 14, 2025

wzssyqa closed this Dec 2, 2025

arsenm reopened this Dec 2, 2025

Merge branch 'main' into clang-fmin-with-nsz

d65e845

wzssyqa added 2 commits December 2, 2025 17:18

Update clang/test/CodeGen/fmaxnum_fminnum_use_nsz.c

b754adc

Merge branch 'main' into clang-fmin-with-nsz

67b7f54

Merge branch 'main' into clang-fmin-with-nsz

883919f

fhahn reviewed Dec 4, 2025

View reviewed changes

arsenm reviewed Dec 4, 2025

View reviewed changes

clang/lib/CodeGen/CGBuiltin.cpp Outdated Show resolved Hide resolved

clang/lib/CodeGen/CGBuiltin.cpp Outdated Show resolved Hide resolved

clang/lib/CodeGen/CGBuiltin.cpp Outdated Show resolved Hide resolved

clang/test/CodeGen/fmaxnum_fminnum_use_nsz.c Outdated Show resolved Hide resolved

wzssyqa added 2 commits December 5, 2025 08:52

Merge branch 'main' into clang-fmin-with-nsz

d170c4d

Not use pointer

48cce5f

wzssyqa requested review from arsenm and fhahn December 5, 2025 01:56

Fix code format

9398186

arsenm reviewed Dec 5, 2025

View reviewed changes

wzssyqa requested review from arsenm and nikic December 9, 2025 08:37

Clang: Add nsz to llvm.minnum and llvm.maxnum emitted from fmin and fmax #113133

Are you sure you want to change the base?

Clang: Add nsz to llvm.minnum and llvm.maxnum emitted from fmin and fmax #113133

Uh oh!

Conversation

wzssyqa commented Oct 21, 2024

Uh oh!

llvmbot commented Oct 21, 2024 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

llvmbot commented Oct 21, 2024

Uh oh!

arsenm left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

wzssyqa commented Nov 25, 2024

Uh oh!

jcranmer-intel left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

wzssyqa commented Feb 27, 2025

Uh oh!

arsenm commented Feb 27, 2025

Uh oh!

arsenm commented Feb 27, 2025

Uh oh!

wzssyqa commented Mar 4, 2025

Uh oh!

Uh oh!

Uh oh!

efriedma-quic commented Apr 27, 2025

Uh oh!

wzssyqa commented Sep 14, 2025

Uh oh!

wzssyqa commented Sep 14, 2025

Uh oh!

wzssyqa commented Dec 2, 2025

Uh oh!

arsenm commented Dec 2, 2025

Uh oh!

github-actions bot commented Dec 2, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

🐧 Linux x64 Test Results

Uh oh!

wzssyqa commented Dec 4, 2025

Uh oh!

fhahn left a comment

Choose a reason for hiding this comment

Uh oh!

wzssyqa commented Dec 4, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

arsenm commented Dec 4, 2025

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

github-actions bot commented Dec 5, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

wzssyqa Dec 8, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

llvmbot commented Oct 21, 2024 •

edited

Loading

github-actions bot commented Dec 2, 2025 •

edited

Loading

wzssyqa commented Dec 4, 2025 •

edited

Loading

github-actions bot commented Dec 5, 2025 •

edited

Loading

wzssyqa Dec 8, 2025 •

edited

Loading