[clang][CUDA] Disable float128 diagnostics for device compilation

[pranavk]

No description provided.

[llvmbot]

@llvm/pr-subscribers-clang

Author: Pranav Kant (pranavk)

Changes

---

Full diff: https://github.com/llvm/llvm-project/pull/83918.diff

2 Files Affected:

• (modified) clang/lib/Sema/SemaDeclAttr.cpp (+3-1)
• (modified) clang/lib/Sema/SemaType.cpp (+1)

diff --git a/clang/lib/Sema/SemaDeclAttr.cpp b/clang/lib/Sema/SemaDeclAttr.cpp
index 397b5db0dc0669..e6943efb345ce0 100644
--- a/clang/lib/Sema/SemaDeclAttr.cpp
+++ b/clang/lib/Sema/SemaDeclAttr.cpp
@@ -4877,7 +4877,9 @@ void Sema::AddModeAttr(Decl *D, const AttributeCommonInfo &CI,
     NewElemTy = Context.getRealTypeForBitwidth(DestWidth, ExplicitType);
 
   if (NewElemTy.isNull()) {
-    Diag(AttrLoc, diag::err_machine_mode) << 1 /*Unsupported*/ << Name;
+    // Only emit diagnostic on host for 128-bit mode attribute
+    if (!(DestWidth == 128 && getLangOpts().CUDAIsDevice))
+      Diag(AttrLoc, diag::err_machine_mode) << 1 /*Unsupported*/ << Name;
     return;
   }
 
diff --git a/clang/lib/Sema/SemaType.cpp b/clang/lib/Sema/SemaType.cpp
index 1e43e36016a66f..4a4e6f80d0d049 100644
--- a/clang/lib/Sema/SemaType.cpp
+++ b/clang/lib/Sema/SemaType.cpp
@@ -1562,6 +1562,7 @@ static QualType ConvertDeclSpecToType(TypeProcessingState &state) {
   case DeclSpec::TST_float128:
     if (!S.Context.getTargetInfo().hasFloat128Type() &&
         !S.getLangOpts().SYCLIsDevice &&
+        !S.getLangOpts().CUDAIsDevice &&
         !(S.getLangOpts().OpenMP && S.getLangOpts().OpenMPIsTargetDevice))
       S.Diag(DS.getTypeSpecTypeLoc(), diag::err_type_unsupported)
         << "__float128";

[github-actions]

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

[Artem-B]

What do you expect to see if __float128 is used from a GPU function.

Can you check on a toy example.

__attribute__((device)) __float128 f(__float128 a, float b) {
  __float128 c = b + 1.0;
  return a + c;
}

[pranavk]

This is going to error out like this:

error: 'a' requires 128 bit size '__float128' type support, but target 'nvptx64-nvidia-cuda' does not support it

[pranavk]

On second thought, it may be a good idea to add that in the diagnostic test.

[Artem-B]

This is going to error out like this:

error: 'a' requires 128 bit size '__float128' type support, but target 'nvptx64-nvidia-cuda' does not support it

Something does not add up. How would we get target 'nvptx64-nvidia-cuda' if the diag below only fires if we're compiling for the host?

[pranavk]

There are two diags that I disabled as part of this PR:

1. %0 is not supported on this target
2. unsupported machine mode __TC__

The diag that got emitted as part of building sample snippet that you pasted in this comment thread is neither 1 nor 2. This diag comes from part of clang/LLVM when it starts to generate code for nvptx specifically.

[Artem-B]

OK. As long as you're sure that the remaining diag covers all possible uses of fp128 on the GPU, it should be fine.

[Artem-B]

I'd add a test verifying that we do emit diagnostics if fp128 is used in the GPU code.
It would probably need to be done somewhere in the codegen tests as it will not fire in the syntax-only checks.