[HIP] Do not include the CUID module hash with the new driver

[jhuber6]

Summary:
The new driver does not need this hash and it can lead to redefined
symbol errors when the CUID hash isn't set.

[llvmbot]

@llvm/pr-subscribers-clang

@llvm/pr-subscribers-clang-codegen

Author: Joseph Huber (jhuber6)

Changes

Summary:
The new driver does not need this hash and it can lead to redefined
symbol errors when the CUID hash isn't set.

---

Full diff: https://github.com/llvm/llvm-project/pull/84332.diff

1 Files Affected:

• (modified) clang/lib/CodeGen/CodeGenModule.cpp (+1-1)

diff --git a/clang/lib/CodeGen/CodeGenModule.cpp b/clang/lib/CodeGen/CodeGenModule.cpp
index d02875c6a86d77..967319bdfc4571 100644
--- a/clang/lib/CodeGen/CodeGenModule.cpp
+++ b/clang/lib/CodeGen/CodeGenModule.cpp
@@ -916,7 +916,7 @@ void CodeGenModule::Release() {
         llvm::ConstantArray::get(ATy, UsedArray), "__clang_gpu_used_external");
     addCompilerUsedGlobal(GV);
   }
-  if (LangOpts.HIP) {
+  if (LangOpts.HIP && !getLangOpts().OffloadingNewDriver) {
     // Emit a unique ID so that host and device binaries from the same
     // compilation unit can be associated.
     auto *GV = new llvm::GlobalVariable(

[yxsamliu]

CUID is needed for device static variable to be accessible on host side. Since the driver does not know whether device static variables are accessed on host side, it should always enable CUID for HIP.

[jhuber6]

CUID is needed for device static variable to be accessible on host side. Since the driver does not know whether device static variables are accessed on host side, it should always enable CUID for HIP.

Yeah that's what I was wondering about when I noticed that this CUID thing wasn't being set when @Meinersbur was testing on Windows. However, I don't think we need this module tag regardless.

[jhuber6]

CUID is needed for device static variable to be accessible on host side. Since the driver does not know whether device static variables are accessed on host side, it should always enable CUID for HIP.

Oh! I think I remember what I did. I made the CUID hash generator thing make a fallback that just uses the file's inode because it was likely good enough for the use-case. So the new driver doesn't pass -cuid but it does do some scrambling on the globals.

[yxsamliu]

CUID is needed for device static variable to be accessible on host side. Since the driver does not know whether device static variables are accessed on host side, it should always enable CUID for HIP.

Oh! I think I remember what I did. I made the CUID hash generator thing make a fallback that just uses the file's inode because it was likely good enough for the use-case. So the new driver doesn't pass -cuid but it does do some scrambling on the globals.

As long as each generated .o has a unique CUID that is fine. Using inode as CUID has issue since users may compile the same file twice with different options to generate two different .o file (the case of rccl), then the CUID is not unique.