Add __builtin_fmaf16.

[zahiraam]

No description provided.

[llvmbot]

@llvm/pr-subscribers-clang-codegen
@llvm/pr-subscribers-backend-x86

@llvm/pr-subscribers-clang

Author: Zahira Ammarguellat (zahiraam)

Changes

---

Full diff: https://github.com/llvm/llvm-project/pull/97424.diff

4 Files Affected:

• (modified) clang/include/clang/Basic/Builtins.td (+6)
• (modified) clang/lib/CodeGen/CGBuiltin.cpp (+1)
• (modified) clang/test/CodeGen/X86/math-builtins.c (+5-1)
• (modified) clang/test/CodeGen/constrained-math-builtins.c (+2-1)

diff --git a/clang/include/clang/Basic/Builtins.td b/clang/include/clang/Basic/Builtins.td
index f5b15cf90d1f8..3de13d1d22c95 100644
--- a/clang/include/clang/Basic/Builtins.td
+++ b/clang/include/clang/Basic/Builtins.td
@@ -203,6 +203,12 @@ def FmaF16F128 : Builtin, F16F128MathTemplate {
   let Prototype = "T(T, T, T)";
 }
 
+def Fmaf16 : Builtin, F16F128MathTemplate {
+  let Spellings = ["__builtin_fmaf16"];
+  let Attributes = [FunctionWithBuiltinPrefix, NoThrow, ConstIgnoringErrnoAndExceptions];
+  let Prototype = "T(T, T, T)";
+}
+
 def FmaxF16F128 : Builtin, F16F128MathTemplate {
   let Spellings = ["__builtin_fmax"];
   let Attributes = [FunctionWithBuiltinPrefix, NoThrow, Const, Constexpr];
diff --git a/clang/lib/CodeGen/CGBuiltin.cpp b/clang/lib/CodeGen/CGBuiltin.cpp
index ed37267efe715..326a43a33a0ff 100644
--- a/clang/lib/CodeGen/CGBuiltin.cpp
+++ b/clang/lib/CodeGen/CGBuiltin.cpp
@@ -2587,6 +2587,7 @@ RValue CodeGenFunction::EmitBuiltinExpr(const GlobalDecl GD, unsigned BuiltinID,
   case Builtin::BI__builtin_fma:
   case Builtin::BI__builtin_fmaf:
   case Builtin::BI__builtin_fmal:
+  case Builtin::BI__builtin_fmaf16:
   case Builtin::BIfma:
   case Builtin::BIfmaf:
   case Builtin::BIfmal: {
diff --git a/clang/test/CodeGen/X86/math-builtins.c b/clang/test/CodeGen/X86/math-builtins.c
index 1e0f129b98610..d26db19574051 100644
--- a/clang/test/CodeGen/X86/math-builtins.c
+++ b/clang/test/CodeGen/X86/math-builtins.c
@@ -364,27 +364,31 @@ __builtin_floor(f);      __builtin_floorf(f);     __builtin_floorl(f); __builtin
 // HAS_ERRNO: declare x86_fp80 @llvm.floor.f80(x86_fp80) [[READNONE_INTRINSIC]]
 // HAS_ERRNO: declare fp128 @llvm.floor.f128(fp128) [[READNONE_INTRINSIC]]
 
-__builtin_fma(f,f,f);        __builtin_fmaf(f,f,f);       __builtin_fmal(f,f,f); __builtin_fmaf128(f,f,f);
+__builtin_fma(f,f,f);        __builtin_fmaf(f,f,f);       __builtin_fmal(f,f,f); __builtin_fmaf128(f,f,f);  __builtin_fmaf16(f,f,f);
 
 // NO__ERRNO: declare double @llvm.fma.f64(double, double, double) [[READNONE_INTRINSIC]]
 // NO__ERRNO: declare float @llvm.fma.f32(float, float, float) [[READNONE_INTRINSIC]]
 // NO__ERRNO: declare x86_fp80 @llvm.fma.f80(x86_fp80, x86_fp80, x86_fp80) [[READNONE_INTRINSIC]]
 // NO__ERRNO: declare fp128 @llvm.fma.f128(fp128, fp128, fp128) [[READNONE_INTRINSIC]]
+// NO__ERRONO: declare half @llvm.fma.f16(half, half, half) [[READNONE_INTRINSIC]]
 // HAS_ERRNO: declare double @fma(double noundef, double noundef, double noundef) [[NOT_READNONE]]
 // HAS_ERRNO: declare float @fmaf(float noundef, float noundef, float noundef) [[NOT_READNONE]]
 // HAS_ERRNO: declare x86_fp80 @fmal(x86_fp80 noundef, x86_fp80 noundef, x86_fp80 noundef) [[NOT_READNONE]]
 // HAS_ERRNO: declare fp128 @fmaf128(fp128 noundef, fp128 noundef, fp128 noundef) [[NOT_READNONE]]
+// HAS_ERRNO: declare half @fmaf16(half noundef, half noundef, half noundef) [[NOT_READNONE]]
 
 // On GNU or Win, fma never sets errno, so we can convert to the intrinsic.
 
 // HAS_ERRNO_GNU: declare double @llvm.fma.f64(double, double, double) [[READNONE_INTRINSIC:#[0-9]+]]
 // HAS_ERRNO_GNU: declare float @llvm.fma.f32(float, float, float) [[READNONE_INTRINSIC]]
 // HAS_ERRNO_GNU: declare x86_fp80 @llvm.fma.f80(x86_fp80, x86_fp80, x86_fp80) [[READNONE_INTRINSIC]]
+// HAS_ERRNO_GNU: declare half @llvm.fma.f16(half, half, half) [[READNONE_INTRINSIC]]
 
 // HAS_ERRNO_WIN: declare double @llvm.fma.f64(double, double, double) [[READNONE_INTRINSIC:#[0-9]+]]
 // HAS_ERRNO_WIN: declare float @llvm.fma.f32(float, float, float) [[READNONE_INTRINSIC]]
 // Long double is just double on win, so no f80 use/declaration.
 // HAS_ERRNO_WIN-NOT: declare x86_fp80 @llvm.fma.f80(x86_fp80, x86_fp80, x86_fp80)
+// HAS_ERRNO_WIN: declare half @llvm.fma.f16(half, half, half) [[READNONE_INTRINSIC]]
 
 __builtin_fmax(f,f);       __builtin_fmaxf(f,f);      __builtin_fmaxl(f,f); __builtin_fmaxf128(f,f);
 
diff --git a/clang/test/CodeGen/constrained-math-builtins.c b/clang/test/CodeGen/constrained-math-builtins.c
index 6cc3a10a1e794..42c9e3c5008a3 100644
--- a/clang/test/CodeGen/constrained-math-builtins.c
+++ b/clang/test/CodeGen/constrained-math-builtins.c
@@ -78,12 +78,13 @@ void foo(double *d, float f, float *fp, long double *l, int *i, const char *c, _
 // CHECK: call x86_fp80 @llvm.experimental.constrained.floor.f80(x86_fp80 %{{.*}}, metadata !"fpexcept.strict")
 // CHECK: call fp128 @llvm.experimental.constrained.floor.f128(fp128 %{{.*}}, metadata !"fpexcept.strict")
 
-  __builtin_fma(f,f,f);        __builtin_fmaf(f,f,f);       __builtin_fmal(f,f,f); __builtin_fmaf128(f,f,f);
+  __builtin_fma(f,f,f);        __builtin_fmaf(f,f,f);       __builtin_fmal(f,f,f);  __builtin_fmaf128(f,f,f); __builtin_fmaf16(f,f,f);
 
 // CHECK: call double @llvm.experimental.constrained.fma.f64(double %{{.*}}, double %{{.*}}, double %{{.*}}, metadata !"round.tonearest", metadata !"fpexcept.strict")
 // CHECK: call float @llvm.experimental.constrained.fma.f32(float %{{.*}}, float %{{.*}}, float %{{.*}}, metadata !"round.tonearest", metadata !"fpexcept.strict")
 // CHECK: call x86_fp80 @llvm.experimental.constrained.fma.f80(x86_fp80 %{{.*}}, x86_fp80 %{{.*}}, x86_fp80 %{{.*}}, metadata !"round.tonearest", metadata !"fpexcept.strict")
 // CHECK: call fp128 @llvm.experimental.constrained.fma.f128(fp128 %{{.*}}, fp128 %{{.*}}, fp128 %{{.*}}, metadata !"round.tonearest", metadata !"fpexcept.strict")
+// CHECK: call half @llvm.experimental.constrained.fma.f16(half %{{.*}}, half %{{.*}}, half %{{.*}}, metadata !"fpexcept.strict")
 
   __builtin_fmax(f,f);       __builtin_fmaxf(f,f);      __builtin_fmaxl(f,f); __builtin_fmaxf128(f,f);
 

[zahiraam]

@AaronBallman ping?

[AaronBallman]

Why is this needed when we have:

llvm-project/clang/include/clang/Basic/Builtins.td

Line 201 in a3913a1

let Spellings = ["__builtin_fma"];

That seems to already support fp16 for fma?

[AaronBallman]

Changes LGTM but please be sure to add a release note when landing so users know about the improvement.

[zahiraam]

Changes LGTM but please be sure to add a release note when landing so users know about the improvement.

Thanks for the review.