[clang] [compiler-rt] [llvm] [PGO][AMDGPU] Add offload profiling with uniformity-aware optimization (PR #177665)

[Yaxun Liu via cfe-commits]

================
@@ -303,6 +303,21 @@ HIPAMDToolChain::TranslateArgs(const llvm::opt::DerivedArgList &Args,
   const OptTable &Opts = getDriver().getOpts();
 
   for (Arg *A : Args) {
+    // Handle device-side profile data file for PGO
+    if (A->getOption().matches(options::OPT_fprofile_use_EQ)) {
+      StringRef ProfileFile = A->getValue();
+      auto [Base, Ext] = ProfileFile.rsplit('.');
+      std::string DeviceProfileFile;
+      if (!Ext.empty()) {
+        DeviceProfileFile = (Base + "." + getTriple().str() + "." + Ext).str();
+      } else {
+        DeviceProfileFile = (ProfileFile + "." + getTriple().str()).str();
+      }
+      DAL->AddJoinedArg(A, Opts.getOption(options::OPT_fprofile_instr_use_EQ),
+                        DeviceProfileFile);
+      A->claim();
+      continue;
+    }
----------------
yxsamliu wrote:

Good point. I've removed the auto-suffix logic from HIPAMD.cpp. HIP now uses the same explicit `-Xarch_host`/`-Xarch_device` approach as OpenMP:

```bash
clang++ -x hip app.hip -o app_optimized \
    -Xarch_host -fprofile-use=host.profdata \
    -Xarch_device -fprofile-use=device.profdata
```

This is more flexible since users control the profile file naming, and it's consistent with OpenMP behavior.

I've also added documentation in `clang/docs/HIPSupport.rst` covering the PGO workflow.


https://github.com/llvm/llvm-project/pull/177665