[AMDGCNSPIRV][NFC] Match AMDGPU's __builtin_va_list type

[AlexVlx]

AMDGCN flavoured SPIRV should math AMDGPU TI as much as possible, and the va_list difference was spurious.

[llvmbot]

@llvm/pr-subscribers-clang

Author: Alex Voicu (AlexVlx)

Changes

AMDGCN flavoured SPIRV should math AMDGPU TI as much as possible, and the va_list difference was spurious.

---

Full diff: https://github.com/llvm/llvm-project/pull/152044.diff

2 Files Affected:

• (modified) clang/lib/Basic/Targets/SPIR.h (+4)
• (added) clang/test/Sema/amdgcn-va-list-type.c (+13)

diff --git a/clang/lib/Basic/Targets/SPIR.h b/clang/lib/Basic/Targets/SPIR.h
index c13b286cd7916..9d0ced2afdbc7 100644
--- a/clang/lib/Basic/Targets/SPIR.h
+++ b/clang/lib/Basic/Targets/SPIR.h
@@ -438,6 +438,10 @@ class LLVM_LIBRARY_VISIBILITY SPIRV64AMDGCNTargetInfo final
 
   ArrayRef<const char *> getGCCRegNames() const override;
 
+  BuiltinVaListKind getBuiltinVaListKind() const override {
+    return TargetInfo::CharPtrBuiltinVaList;
+  }
+
   bool initFeatureMap(llvm::StringMap<bool> &Features, DiagnosticsEngine &Diags,
                       StringRef,
                       const std::vector<std::string> &) const override;
diff --git a/clang/test/Sema/amdgcn-va-list-type.c b/clang/test/Sema/amdgcn-va-list-type.c
new file mode 100644
index 0000000000000..1c0e93adabae4
--- /dev/null
+++ b/clang/test/Sema/amdgcn-va-list-type.c
@@ -0,0 +1,13 @@
+// RUN: %clang_cc1 %s -triple amdgcn-amd-amdhsa -fsyntax-only -verify
+// RUN: %clang_cc1 %s -triple spirv64-amd-amdhsa -fsyntax-only -verify
+
+// expected-no-diagnostics
+
+typedef char* va_list;
+
+void foo(const char* f, ...) {
+    int r;
+    va_list args;
+    __builtin_va_start(args, f);
+    __builtin_va_end(args);
+}
\ No newline at end of file

[chinmaydd]

LGTM.

Moreover, I believe this change is relevant to #119661 i.e representation of the va_list type for the NVPTX backend on Windows.