[all-commits] [llvm/llvm-project] bacc5b: [AMDGPU] Track allocator-reserved register pressur...

[Dhruva Chakrabarti via All-commits]

  Branch: refs/heads/users/dhruvachak/fix_tuple_handling
  Home:   https://github.com/llvm/llvm-project
  Commit: bacc5bd4b0f58d2d8f9cb2ded9eb5dd390b08676
      https://github.com/llvm/llvm-project/commit/bacc5bd4b0f58d2d8f9cb2ded9eb5dd390b08676
  Author: Dhruva Chakrabarti <Dhruva.Chakrabarti at amd.com>
  Date:   2026-09-25 (Fri, 25 Sep 2026)

  Changed paths:
    M llvm/lib/Target/AMDGPU/GCNRegPressure.cpp
    M llvm/lib/Target/AMDGPU/GCNRegPressure.h
    M llvm/lib/Target/AMDGPU/GCNSchedStrategy.cpp

  Log Message:
  -----------
  [AMDGPU] Track allocator-reserved register pressure separately

GCNRegPressure counts a tuple at its live lane count, but the excess and
critical limits the scheduler compares against are expressed in
allocatable registers. A tuple is reserved at its full class width for
its whole live range, so lanes that are dead at a given point are still
unavailable to any other value, and the live lane count underestimates
what the allocator has to set aside.

Track a third set of per-kind values recording the reserved width, and
use it for the pressure the GCN trackers hand to the scheduling
heuristics.

Also add -amdgpu-trackers-compare-rp to print the GCN tracker and the
generic tracker pressure side by side for every candidate, which is what
the discrepancy was found with.

WIP

AI-Assisted



To unsubscribe from these emails, change your notification settings at https://github.com/llvm/llvm-project/settings/notifications