[clang] [llvm] Stop using spir_kernel calling convention on non-SPIR targets. (PR #191090)

[Nick Sarnie via cfe-commits]

================
@@ -866,7 +866,8 @@ const CGFunctionInfo &CodeGenTypes::arrangeLLVMFunctionInfo(
   assert(inserted && "Recursively being processed?");
 
   // Compute ABI information.
-  if (CC == llvm::CallingConv::SPIR_KERNEL) {
+  if (info.getCC() == CC_DeviceKernel &&
+      (CC == llvm::CallingConv::SPIR_KERNEL || CC == llvm::CallingConv::C)) {
----------------
sarnex wrote:

Would an assert like `assert(CC != llvm::CallingConv::CC || getContext().getLangOpts().OpenCL)` before the call to `computeSPIRKernelABIInfo` work? 

Basically I want to lock down that we only have `info.getCC() == CC_DeviceKernel` but LLVM CC `CallingConv::CC` for OpenCL, because in the general case if we see that that means the Clang CC to LLVM CC mapping did something crazy.

https://github.com/llvm/llvm-project/pull/191090