[clang] [ClangLinkerWrapper] Use discrete steps in verbose mode (PR #204186)

[Joseph Huber via cfe-commits]

================
@@ -790,8 +855,52 @@ wrapDeviceImages(ArrayRef<std::unique_ptr<MemoryBuffer>> Buffers,
   return *FileOrErr;
 }
 
+/// Perform the OpenMP bundling with 'llvm-offload-binary' in verbose mode.
 Expected<SmallVector<std::unique_ptr<MemoryBuffer>>>
-bundleOpenMP(ArrayRef<OffloadingImage> Images) {
+bundleOpenMPVerbose(ArrayRef<OffloadingImage> Images, const ArgList &Args) {
+  Expected<std::string> OffloadBinaryPath = findProgram(
+      "llvm-offload-binary", {getExecutableDir("llvm-offload-binary")});
+  if (!OffloadBinaryPath)
+    return OffloadBinaryPath.takeError();
+
+  BumpPtrAllocator Alloc;
+  StringSaver Saver(Alloc);
+  SmallVector<std::unique_ptr<MemoryBuffer>> Buffers;
+  for (const OffloadingImage &Image : Images) {
+    StringRef ImageFile = Image.Image->getBufferIdentifier();
+    auto BinaryOrErr =
+        createOutputFile(sys::path::stem(ImageFile) + "." +
+                             getOffloadKindName(Image.TheOffloadKind),
+                         "offload");
+    if (!BinaryOrErr)
+      return BinaryOrErr.takeError();
+
+    std::string ImageArg = ("--image=file=" + ImageFile +
----------------
jhuber6 wrote:

Honestly, the `IMG_xyz` metadata is functionally unused. Maybe some placed try to infer some filename on it? But all the real users just use the magic bytes. I think the only case that it would make a difference is like `PTX`, but all PTX used in practice is already in a CUDA fatbinary which has magic. I suppose the one thing people would notice is that the `llvm-objdump --offloading` wouldn't display anything, half considering just making that print based off of the magic instead as well.

I could maybe make it work, but I don't think it's a major priority.

https://github.com/llvm/llvm-project/pull/204186