Skip to content

AMDGPU: expose legacy stepping instruction classification through installed MC APIs #3638

Description

@harsh-amd

COMGR HotSwap must inspect a gfx1250 B0 legacy VOP3 instruction while using the gfx1250 A0 MC decoder. The A0 decoder advances one dword and reports the B0-only opcode as unknown, so COMGR currently has no installed MC API that can establish that the complete 64-bit pair is a vector ALU instruction which cannot affect scalar control flow or MODE.\n\nPlease expose an installed AMDGPU MC/TargetParser API that can classify legacy stepping encodings (including instruction length and conservative register/control-flow effects) without duplicating backend-private TableGen fields in COMGR. AMDGPUBaseInfo.h and named operand helpers are backend-private and unavailable to installed COMGR builds.\n\nUntil this exists, the HotSwap M32 Scale16 work will keep one exact, fail-closed classifier for the observed B0 pair, reject any known entry into its continuation dword, and cover positive and near-miss encodings with focused tests.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions