[clang] [llvm] [clang][SPIR-V] Add support for AMDGCN flavoured SPIRV (PR #89796)

Alex Voicu via cfe-commits cfe-commits at lists.llvm.org
Sun Apr 28 11:09:00 PDT 2024


================
@@ -54,3 +56,76 @@ void SPIRV64TargetInfo::getTargetDefines(const LangOptions &Opts,
   BaseSPIRVTargetInfo::getTargetDefines(Opts, Builder);
   DefineStd(Builder, "SPIRV64", Opts);
 }
+
+static const AMDGPUTargetInfo AMDGPUTI(llvm::Triple("amdgcn-amd-amdhsa"), {});
+
+ArrayRef<const char *> SPIRV64AMDGCNTargetInfo::getGCCRegNames() const {
+  return AMDGPUTI.getGCCRegNames();
+}
+
+bool SPIRV64AMDGCNTargetInfo::initFeatureMap(
+    llvm::StringMap<bool> &Features, DiagnosticsEngine &Diags, StringRef,
+    const std::vector<std::string> &FeatureVec) const {
+  llvm::AMDGPU::fillAMDGPUFeatureMap({}, getTriple(), Features);
+
+  return TargetInfo::initFeatureMap(Features, Diags, {}, FeatureVec);
+}
+
+bool SPIRV64AMDGCNTargetInfo::validateAsmConstraint(
+    const char *&Name, TargetInfo::ConstraintInfo &Info) const {
+  return AMDGPUTI.validateAsmConstraint(Name, Info);
+}
+
+std::string
+SPIRV64AMDGCNTargetInfo::convertConstraint(const char *&Constraint) const {
+  return AMDGPUTI.convertConstraint(Constraint);
+}
+
+ArrayRef<Builtin::Info> SPIRV64AMDGCNTargetInfo::getTargetBuiltins() const {
+  return AMDGPUTI.getTargetBuiltins();
+}
+
+void SPIRV64AMDGCNTargetInfo::getTargetDefines(const LangOptions &Opts,
+                                               MacroBuilder &Builder) const {
+  BaseSPIRVTargetInfo::getTargetDefines(Opts, Builder);
+  DefineStd(Builder, "SPIRV64", Opts);
+
+  Builder.defineMacro("__AMD__");
+  Builder.defineMacro("__AMDGPU__");
+  Builder.defineMacro("__AMDGCN__");
----------------
AlexVlx wrote:

`__AMDGCN_WAVEFRONT_SIZE` is a pretty terrible error, which should never have been defined on host (not in the least because it's non-uniform across targets). These do (should) end up defined on host too, but they are harmless because they are uniform i.e. all potential consumers of `spirv64-amd-amdhsa` are `__AMD__`, `__AMDGPU__` and `__AMDGCN__`. They're mostly a necessary nuisance for getting library code to work.

https://github.com/llvm/llvm-project/pull/89796


More information about the cfe-commits mailing list