[comgr][hotswap] B0 -> A0 Support Tracker

[xintin]

Umbrella tracker for the comgr HotSwap work that rewrites GFX1250 code objects for B0-to-A0 cross-stepping execution. Collects the open infra/cleanup issues spun out of the HotSwap PR stack, the high-priority in-flight rewrite PRs, foundational landed work, and reference docs. Scope is AMD-fork-specific (amd/comgr); labels: comgr, hotswap.

Foundational PRs (HotSwap infrastructure stack)

• [AMDGPU] comgr: add HotSwap ELF infrastructure (1/3) #2201 [AMDGPU] comgr: add HotSwap ELF infrastructure (1/3)
• [AMDGPU] comgr: add HotSwap MC/LLVM infrastructure (2/3) #2202 HotSwap infrastructure (2/3)
• [AMDGPU] comgr: add HotSwap B0-to-A0 policy and public API (3/3) #2203 HotSwap infrastructure (3/3)
• [AMDGPU] comgr: HotSwap in-place patches for B0-to-A0 rewriting #2222 [AMDGPU] comgr: HotSwap in-place patches for B0-to-A0 rewriting
• [AMDGPU] comgr: add WMMA instruction splitting for GFX1250 B0-to-A0 compatibility #2379 [AMDGPU] comgr: add WMMA instruction splitting for GFX1250 B0-to-A0 compatibility

High-priority pending PRs

• [AMDGPU] comgr: Scratch Patches for B0-to-A0 Hotswap 2/2 -- Scaled WMMA #2731 [AMDGPU] comgr: Scratch Patches for B0-to-A0 Hotswap 2/2 — Scaled WMMA (@jammm)
• [AMDGPU] comgr: read SGPR count from msgpack metadata instead of reserved KD field #2873 [AMDGPU] comgr: read SGPR count from msgpack metadata instead of reserved KD field (@suryajasper)

Open infra / cleanup issues

Each carries granular background, a concrete proposal, and the source files / review thread it came from.

• Comgr hotswap: replace hardcoded AMDGPU VGPR/SGPR granule constants with MC-derived values #2285 replace hardcoded AMDGPU VGPR/SGPR granule constants with MC-derived values. comgr-hotswap-b0a0.cpp hardcodes Gfx1250VgprGranuleSize = 8 / Gfx1250SgprGranuleSize = 16; these already exist in LLVM (AMDGPU::IsaInfo::get{VGPR,SGPR}AllocGranule) but only in the target-internal AMDGPUBaseInfo.h. Promote the getters to a public header (or add AMDHSA helpers), then drive RewriteConfig::{Vgpr,Sgpr}GranuleSize from the parsed ISA. (from PR [AMDGPU] comgr: add HotSwap B0-to-A0 policy and public API (3/3) #2203 review; cc @chinmaydd @lamb-j)
• Comgr: share MC toolchain state between DisassemblyInfo and hotswap LLVMState #2253 share MC toolchain state between DisassemblyInfo and hotswap LLVMState. comgr-disassembly.cpp and comgr-hotswap-llvm.cpp build near-identical per-ISA MC* object bundles. Extract a common ComgrMCToolchain struct both embed (DisassemblyInfo adds MIA+callbacks; LLVMState adds MOFI+MCE). (from PR [AMDGPU] comgr: add HotSwap MC/LLVM infrastructure (2/3) #2202 review; @lamb-j to help on disassembly side)
• comgr: upstream AMDGPU SOPP s_branch offset encoding constants to public LLVM headers #2235 upstream AMDGPU SOPP s_branch offset encoding constants to public LLVM headers. HotSwap locally defines BranchOffsetMin/Max/Mask in comgr-hotswap-internal.h; expose via an inline llvm::AMDGPU::getSBranchOffsetRange() + SBranchOffsetMask (TargetParser.h or a new BinaryFormat/AMDGPU.h), then drop the locals. (from PR [AMDGPU] comgr: add HotSwap ELF infrastructure (1/3) #2201 review; can combine with comgr: upstream AMDGPU note-owner / ISA-version-note constants to public LLVM headers #2225)
• comgr: extract hotswap-agnostic helpers into a comgr-utils module #2227 extract hotswap-agnostic helpers into a comgr-utils module. Move extractCpu() and encodeSBranch() out of COMGR::hotswap into a new COMGR::utils (comgr-utils.{h,cpp}, public-LLVM-only deps), update call sites in comgr-hotswap-llvm.cpp/comgr-hotswap-b0a0.cpp, relocate gtests. (from PR [AMDGPU] comgr: add HotSwap ELF infrastructure (1/3) #2201 review)
• comgr: upstream AMDGPU note-owner / ISA-version-note constants to public LLVM headers #2225 upstream AMDGPU note-owner / ISA-version-note constants to public LLVM headers. HotSwap locally redefines kAmdgpuNoteOwner = "AMDGPU" and kNoteTypeIsaVersion = 27; upstream NoteNameV3 into a public BinaryFormat header and add NT_AMDGPU_ISA_VERSION (27) to ELF.h, then drop the duplicates. (from PR [AMDGPU] comgr: add HotSwap ELF infrastructure (1/3) #2201 review; sibling of comgr: upstream AMDGPU SOPP s_branch offset encoding constants to public LLVM headers #2235)

Landed work

Merged comgr/hotswap PRs: is:pr is:merged label:comgr label:hotswap

Done when

• Both high-priority PRs (Backport cstdint patch for building with Clang #273 Scaled WMMA 2/2, msgpack SGPR-count) are merged.
• B0-to-A0 rewrite validated end-to-end (scaled-WMMA scratch patches + correct SGPR-count sourcing + WMMA splitting) with pinned lit/gtest coverage.
• Extensively tested using kernels for transpiler testing; currently tested on ~20 gemm kernels and lit tests for each merged feature.