[all-commits] [llvm/llvm-project] 09f850: [AMDGPU] Fix speculative register pressure queries...

[ilia-cher via All-commits]

  Branch: refs/heads/main
  Home:   https://github.com/llvm/llvm-project
  Commit: 09f8509bec1fec4636c8e592dcd55a80f39f2667
      https://github.com/llvm/llvm-project/commit/09f8509bec1fec4636c8e592dcd55a80f39f2667
  Author: ilia-cher <30845429+ilia-cher at users.noreply.github.com>
  Date:   2026-07-16 (Thu, 16 Jul 2026)

  Changed paths:
    M llvm/lib/Target/AMDGPU/GCNRegPressure.cpp
    M llvm/unittests/Target/AMDGPU/GCNRegPressureTest.cpp

  Log Message:
  -----------
  [AMDGPU] Fix speculative register pressure queries (#208574)

There are two issues with the way we currently speculate register
pressure:
1. GCNDownwardRPTracker::advance(with UseInternalIterator=false), which
is called by the scheduler in schedNode, resets LastTrackedMI, so the
tracker (bumpDownwardPressure) doesn't know where the last scheduled
instruction is and falls back to the beginning of the basic block. As a
result, when we estimate RP impact for a given MI, we tend to find uses
that are often already scheduled and should be skipped.
2. When looking for the remaining uses between LastTrackedMI and the
candidate MI we should skip already scheduled instruction.



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