[Comgr][hotswap] Add bare-bones hotswap transpiler scaffolding

[tgymnich]

Lands the minimal scaffolding
needed to build a hotswap-transpiler static library:

• raiser.{hpp,cpp} — raiseToIR() entry point. The implementation creates an llvm::Module with a single ret void AMDGPU kernel function for any input. ELF ingestion, MC disassembly and the per-format raise handlers land in subsequent patches.
• raise_failure.{hpp,cpp} — structured failure values consumed by RaiseResult::failure.
• code_object_utils.hpp — KernelMeta struct used in the raiseToIR signature.
• Minimal CMakeLists.txt linking only LLVMCore, LLVMSupport, LLVMTargetParser. The full library/include surface grows incrementally as later patches need it.
• tests/raiser_empty_module_test.cpp (gtest) verifies the scaffolding contract: an empty input produces a well-formed module with one AMDGPU_KERNEL function whose body is ret void.

[tgymnich]

Warning

Stacked PR

1. [Comgr][hotswap] Add bare-bones hotswap transpiler scaffolding #2438 <----
2. [Comgr][hotswap] Add code-object metadata parsing #2437
3. [Comgr][hotswap] Add AMDGPU MC state setup #2439

[z1-cciauto]

PSDB Link: https://compiler-ci.amd.com/job/compiler-psdb-amd-staging/5772

[chinmaydd]

@tgymnich per today's call, would you be okay with landing the refactor of moving the transpiler into hotswap/transpiler first or do we want to do that as a followup ?

[chinmaydd]

Disclaimer: I've gathered most of these through Claude Opus 4.7, but I believe they are retrieved through the agent looking at the AGENT_CONVENTIONS.md file and the current conventions that we follow in Comgr.

---

Scaffolding isn't wired into the comgr build. amd/comgr/CMakeLists.txt is unchanged, so neither hotswap-transpiler nor transpiler_tests is built under make check-comgr. HOTSWAP_TRANSPILER_BUILD_TOOLS also defaults OFF in the embedded case, so the gtest stays unbuilt even after a future add_subdirectory(). The contract this patch advertises is currently unverifiable in CI.

Duplicates amd/comgr/test-unit/ infrastructure. That tree already has 3-tier gtest discovery (llvm_gtest → find_package(GTest CONFIG) → LLVM-installed gtest), comgr_match_llvm_rtti(), and check-comgr plumbing. The new find_package(GTest REQUIRED) is strictly the 2nd tier and hard-fails in monorepo builds without a system gtest. Per AGENT_CONVENTIONS.md §1, I'd prefer the transpiler refactor land first and this slot into the existing pattern.

License headers missing or incomplete. code_object_utils.hpp, raise_failure.{hpp,cpp}, raiser.hpp, tests/test_main.cpp have no header at all; raiser.cpp and raiser_empty_module_test.cpp have the banner but lack the Apache / SPDX block. Compare test-unit/HotswapElfTest.cpp.

Conventions diverge from the rest of amd/comgr/. .clang-tidy mandates CamelCase for members / params / locals — the new code is lowerCamelCase throughout (kernargSegmentSize, kernelName, funcTy). Headers are .hpp (rest of comgr is .h); namespace is transpiler (rest is COMGR::hotswap). Cheaper to fix now than after #2437 / #2439 stack on top.

Test doesn't verify the advertised contract, and raiseToIR masks failures. Banner says "body is ret void"; the test only checks !fn->empty() and the calling convention — not the terminator, the target triple, the (missing) DataLayout, or any failure path. raiseToIR itself /…/s every parameter except kernelName, never calls parseTargetIdentifier(sourceISA) (AGENT_CONVENTIONS.md §1 reuse list), doesn't reject empty inputs, and unconditionally sets success = true. Both downstream PRs will inherit this.

Smaller stuff (dead fields on RaiseResult, single-valued RaiseFailureReason plus its unreachable fallback in reasonString, the auto / rfind(...,0)==0 / hand-rolled (x+7)&~7 / signed int sizes in KernelMeta::implicitArgsBase())

[z1-cciauto]

PSDB Link: https://compiler-ci.amd.com/job/compiler-psdb-amd-staging/5801

[martin-luecke]

Thanks, this looks much better with the recent changes. The scaffold is small, the COMGR test-unit integration is the right direction, and the tests now cover the advertised contract much better.

I left a few comments. Once those are addressed and CI is green, I’m fine with merging this PR.

• Not approving because I am partly author of this.

[martin-luecke]

CI is currently failing because this file skips include(AddLLVM) when LLVMSupport exists, but llvm_update_compile_flags is still not defined in the COMGR/TheRock configure path. Could we instead check if(NOT COMMAND llvm_update_compile_flags) before including AddLLVM?

[tgymnich]

fixed

[martin-luecke]

This makes the hotswap subproject configure unconditionally in normal COMGR builds, before any public API lands. Is that intentional for test-unit coverage? If yes, the hotswap CMake needs to be robust in all COMGR/TheRock configurations. If not, consider gating or EXCLUDE_FROM_ALL until integration lands.

[tgymnich]

Yes, this is intentional. Better figure this out as soon as possible, and keep CI green.

[lamb-j]

Are we wanting to keep comgr/hotswap or move to comgr/src/hotswap? Having a new top-level directory with source code would be a departure from the existing setup. (Is hotswap a Comgr appendage, or an integrated feature)

[chinmaydd]

I'm okay with either. TBH since we provide an API through Comgr for the hotswap rewrite, I think keeping it under comgr/src/hotswap makes sense to me.

If this was exclusively a Comgr API consumer, I would do comgr/hotswap.

[tgymnich]

I re-worked this with the clang-tidy rules and the new home in comgr/src/hotswap

[chinmaydd]

Mostly LGTM.

However, what can we do to resolve this ? We want to reuse as much as possible from Comgr / LLVM and lay a foundation for doing that from the get-go.

[tgymnich]

I'm at it for PR2, this may be backported to this PR

[chinmaydd]

Awesome, thanks!

[chinmaydd]

LGTM