[clang] [clang][X86] Support __attribute__((model("small"/"large"))) (PR #124834)

[Reid Kleckner via cfe-commits]

================
@@ -2949,15 +2950,34 @@ static void handleSectionAttr(Sema &S, Decl *D, const ParsedAttr &AL) {
   }
 }
 
+static bool isValidCodeModelAttr(Sema &S, StringRef Str) {
+  if (S.Context.getTargetInfo().getTriple().isLoongArch()) {
+    return Str == "normal" || Str == "medium" || Str == "extreme";
+  } else {
+    assert(S.Context.getTargetInfo().getTriple().getArch() ==
+               llvm::Triple::x86_64 &&
+           "only loongarch/x86-64 supported");
+    return Str == "small" || Str == "large";
+  }
+}
+
 static void handleCodeModelAttr(Sema &S, Decl *D, const ParsedAttr &AL) {
   StringRef Str;
   SourceLocation LiteralLoc;
   // Check that it is a string.
   if (!S.checkStringLiteralArgumentAttr(AL, 0, Str, &LiteralLoc))
     return;
 
+  // Ignore the attribute for GPU device compiles since it only applies to host
+  // globals.
+  if (S.Context.getTargetInfo().getTriple().isNVPTX() ||
----------------
rnk wrote:

Aren't we? I mean, I'm sure nvidia and vendors everywhere do crazy stuff, but I would like to clarify what the support scope of the Clang project is. My understanding of all the CUDA functionality we've ever added to Clang was that it assumes we're using the Clang compiler on both sides, not the janky nvcc split pre-processing model.

https://github.com/llvm/llvm-project/pull/124834