Support TheRock LLVM layout in Windows HIP build script

[jamesweiym-ops]

What changed

• detect Windows ROCm clang in both bin and lib/llvm/bin
• recognize the TheRock layout when resolving HIP_PATH or auto-discovering C:\TheRock\build
• prepend the selected compiler directory to PATH while keeping the ROCm runtime bin directory

Why

Current TheRock-based Windows ROCm packages place hipcc.exe under bin, but place clang.exe and clang++.exe under lib/llvm/bin. The Windows HIP build script previously required <ROCm>\bin\clang++.exe, so an otherwise valid TheRock install failed immediately with:

ROCm clang not found under C:\TheRock\build\bin

The normal AMD HIP SDK layout remains preferred and unchanged.

Validation

Tested on Windows 11 with Ryzen AI MAX+ 395 / Radeon 8060S (gfx1151):

• ROCm/TheRock 7.14.60850
• AMD Clang 23.0.0
• Visual Studio 2022 toolset 14.44
• explicit -RocmPath C:\TheRock\build configure: passed
• automatic HIP_PATH=C:\TheRock\build detection: passed
• distribution-style HIP build with -GpuTargets gfx1151 -NoNativeCpu -DeploymentBuild -Target audiocpp_cli: passed (708 build steps)
• resulting audiocpp_cli.exe --help: passed
• resulting audiocpp_cli.exe --list-loaders --json: passed

[IIIIIllllIIIIIlllll]

Sorry! I've never used TheRock, so that didn't occur to me. I'm not familiar with it. I'll take a look at how it works tomorrow. Thanks for your PR!

[IIIIIllllIIIIIlllll]

Validated this on Windows 11 with a TheRock nightly tarball instead of a source build, targeting a different GPU family:

Environment

Ryzen 7 7840U iGPU (gfx1103)
TheRock nightly therock-dist-windows-gfx110X-all-7.15.0a20260728 (extracted to C:\TheRock\7.15)
AMD Clang 23.0.0 (from the tarball), Visual Studio 18 toolset 14.5x, CMake 4.3
What passed

Explicit -RocmPath C:\TheRock\7.15 configure: passed — clang is correctly picked up from lib\llvm\bin (fails on main with ROCm clang not found under C:\TheRock\7.15\bin, as expected)
Distribution-style build -GpuTargets gfx1103 -NoNativeCpu -DeploymentBuild -Target audiocpp_cli: passed (684 build steps)
audiocpp_cli.exe --help and --list-loaders --json: passed
So the change itself works as advertised for the tarball layout too. However, TheRock 7.15 + a newer MSVC toolset surfaced three issues outside this PR's scope that are worth documenting for the next person:

No resource compiler in TheRock tarballs. They ship neither llvm-rc nor windres, and CMake 4.x fails configure with No CMAKE_RC_COMPILER could be found. Workaround: run from a VS developer prompt, or set RC to the Windows SDK's rc.exe.

MSVC ≥14.40 vs clang 23 HIP wrapper headers. Compiling any HIP TU that pulls fails with error: device function 'isgreater' cannot overload host device function 'isgreater' (and siblings). This is the known collision fixed upstream by llvm/llvm-project#201563, which TheRock's clang 23 does not yet carry. Workaround: apply the same one-file reorder to lib/llvm/lib/clang/23/include/__clang_hip_runtime_wrapper.h (move #include <__clang_cuda_math_forward_declares.h> above #include ).

ROCm device libraries not found. The tarball keeps amdgcn/bitcode under lib\llvm, where clang 23 does not look; every HIP compile fails with cannot find ROCm device library. Workaround: a directory junction mklink /J C:\TheRock\7.15\amdgcn C:\TheRock\7.15\lib\llvm\amdgcn so $ROCM_PATH\amdgcn\bitcode resolves (same class of workaround as ROCm/TheRock#1979).

One more observation in this PR's direct scope: amdgpu-arch.exe also lives in lib\llvm\bin in the TheRock layout, not bin, so the script's GPU auto-detection can't find it either — I had to pass -GpuTargets explicitly. Might be worth resolving it from $compilerDir instead of $RocmPath\bin, since you already compute that.

[0xShug0]

@jamesweiym-ops Let me know when the PR is ready.

[0xShug0]

@IIIIIllllIIIIIlllll I'd like to merge the PR. Is it safe to merge as is based on your tests?

[IIIIIllllIIIIIlllll]

@0xShug0 No problem, safe, but I think we need to make it clear—users need to know how to handle the operations themselves, given that there can be differences between various versions of ROCm. But since they're using TheRock, I'm sure he can handle it.

[0xShug0]

@IIIIIllllIIIIIlllll Thank! I'll update the docs later (let AI handle it) based on this PR and your feedback.

[0xShug0]

@jamesweiym-ops @IIIIIllllIIIIIlllll PR merged! Thank you both!