[llvm] [clang] [AMDGPU] Enable OpenCL hostcall printf (WIP) (PR #72556)

[Sameer Sahasrabuddhe via llvm-commits]

================
@@ -4742,6 +4742,16 @@ void Clang::ConstructJob(Compilation &C, const JobAction &JA,
     Args.ClaimAllArgs(options::OPT_gen_cdb_fragment_path);
   }
 
+  if (TC.getTriple().isAMDGPU() && types::isOpenCL(Input.getType())) {
+    if (Args.getLastArg(options::OPT_mprintf_kind_EQ)) {
+      CmdArgs.push_back(Args.MakeArgString(
+          "-mprintf-kind=" +
+          Args.getLastArgValue(options::OPT_mprintf_kind_EQ)));
+      // Force compiler error on invalid conversion specifiers
+      CmdArgs.push_back(Args.MakeArgString("-Werror=format-invalid-specifier"));
----------------
ssahasra wrote:

Actually this is important. Converting a warning into an error can cause surprises when users build existing applications. It is upto the user to decide if they want this to be an error, which is why it is a command-line option. Even if the spec says undefined behaviour, we should just do something reasonable and accept whatever the user wrote without forcing an error. There is a lot of freedom in deciding what is "something reasonable" ... we could choose to print nothing, or a default value, or a placeholder instead of the actual printf format string, etc.

https://github.com/llvm/llvm-project/pull/72556