[AMDGPU] comgr: read SGPR count from msgpack metadata instead of reserved KD field

[suryajasper]

Summary

On GFX10+, GRANULATED_WAVEFRONT_SGPR_COUNT in the kernel descriptor is architecturally reserved and must be zero — the hardware ignores it. The scratch SGPR allocator introduced in #2328 was reading this field to determine where to start allocating scratch SGPRs, which is unreliable: a code object that correctly zeroes the field would cause the allocator to start at s8, potentially clobbering live registers.

This PR replaces the KD-field approach with a metadata-first strategy: read .sgpr_count from the amdhsa.kernels msgpack metadata note (NT_AMDGPU_METADATA), falling back to the KD field for minimal ELFs that lack a metadata note (e.g. lit test objects assembled with -nostdlib).

Also fixes Gfx1250MaxVgprs from 256 to 1024 — GFX1250 has Feature1024AddressableVGPRs and with wave32, getAddressableNumVGPRs returns 1024.

Follow-up to #2328 (specifically @jmmartinez's review comment noting the SGPR field is reserved on GFX10+, and comment noting gfx1250 has 1024 VGPRs).

Implementation Details

comgr-hotswap-elf.cpp — getKernelSgprCount() rewritten:

• Primary path: iterate PT_NOTE program headers → find NT_AMDGPU_METADATA note → parse msgpack blob via llvm::msgpack::Document::readFromBlob() → walk amdhsa.kernels array → match .name to the kernel → return .sgpr_count.
• Fallback path: if no metadata note is present, read GRANULATED_WAVEFRONT_SGPR_COUNT from the kernel descriptor with the correct encoding granule (8), matching getSGPREncodingGranule() in AMDGPUBaseInfo.cpp.
• Returns std::nullopt on failure; the caller falls back to MaxSgprs (106), which is conservative (no headroom → patch refused rather than silent clobber).

comgr-hotswap-elf.cpp — updateKernelDescriptor() simplified:

• Removed ExtraSgprs and SgprGranuleSize parameters.
• Deleted the SGPR write block — bumping a reserved field has no effect on hardware.
• VGPR update path unchanged (that field is not reserved).

comgr-hotswap-internal.h:

• getKernelSgprCount() signature: removed SgprGranuleSize parameter (metadata provides the raw count directly).
• updateKernelDescriptor() signature: removed ExtraSgprs and SgprGranuleSize.
• RewriteConfig: removed SgprGranuleSize field.

comgr-hotswap-patch-f32-to-e5m3.cpp:

• allocateScratch(): updated call to getKernelSgprCount() (no granule argument).

comgr-hotswap-b0a0.cpp:

• Removed Gfx1250SgprGranuleSize constant.
• Fixed Gfx1250MaxVgprs from 256 to 1024 (Feature1024AddressableVGPRs + wave32).
• Updated updateKernelDescriptor() call site to match new signature.

No build system changes — LLVMBinaryFormat (which includes msgpack) is already linked by COMGR.

Test Coverage

• 49/49 hotswap lit tests pass, including all 7 FP8 E5M3 tests and the new WMMA split test. The CHECK patterns use s_mov_b32 without specifying SGPR numbers, so they are agnostic to which specific SGPRs the allocator selects.
• 4/4 HotswapElfTests and 44/44 HotswapMCTests pass.
• The KD fallback path is exercised by every existing lit test (assembled with -nostdlib, which produces ELFs without metadata notes). The metadata path will be exercised by production code objects compiled with hipcc/clang which always include NT_AMDGPU_METADATA.

[jmmartinez]

The getKernelSgprCount related changes look good to me. Thanks!

[chinmaydd]

@suryajasper could you take a look at the merge conflicts here ? Thanks ! This one should be good to go.

[suryajasper]

Rebased and addressed merge conflicts. Also fixed Gfx1250MaxVgprs from 256 to 1024 per @jmmartinez's comment on #2328.