[clang] [llvm] [X86] Enhance kCFI type IDs with a 3-bit arity indicator. (PR #117121)

[Scott Constable via cfe-commits]

================
@@ -208,10 +209,34 @@ void llvm::setKCFIType(Module &M, Function &F, StringRef MangledType) {
   std::string Type = MangledType.str();
   if (M.getModuleFlag("cfi-normalize-integers"))
     Type += ".normalized";
+
+  uint32_t OutHash = static_cast<uint32_t>(llvm::xxHash64(Type));
+  auto T = Triple(Twine(M.getTargetTriple()));
+  if (T.isX86() && T.isArch64Bit() && T.isOSLinux()) {
+    // Estimate the function's arity (i.e., the number of arguments) at the ABI
+    // level by counting the number of parameters that are likely to be passed
+    // as registers, such as pointers and 64-bit (or smaller) integers. The
+    // Linux x86-64 ABI allows up to 6 parameters to be passed in GPRs.
+    // Additional parameters or parameters larger than 64 bits may be passed on
+    // the stack, in which case the arity is denoted as 7.
+    size_t NumParams = F.arg_size();
+    bool MayHaveStackArgs = NumParams > 6;
+
+    for (unsigned int i = 0; !MayHaveStackArgs && i < NumParams; ++i) {
+      const llvm::Type *PT = F.getArg(i)->getType();
+      if (!(PT->isPointerTy() || PT->getIntegerBitWidth() <= 64))
----------------
scottconstable wrote:

Thank you for the suggestion. I looked at `llvm::Argument::hasPassPointeeByValueCopyAttr()`, but it looks like it is only available where a function is being defined. It does not appear to be available where a call is made through a function pointer. Therefore, I'm not sure that `llvm::Argument::hasPassPointeeByValueCopyAttr()` will be helpful since KCFI requires the ID to be computed identically at both the call site and the call target.

Or, do you think I am overlooking something, and that there is a way to use `llvm::Argument::hasPassPointeeByValueCopyAttr()` or something similar at an indirect call site? As far as I can tell, the only information that is available at an indirect call site is the function pointer type, which does contain the number of arguments and their types, but does not appear to contain an indication as to whether an argument may be passed on the stack.

https://github.com/llvm/llvm-project/pull/117121