[all-commits] [llvm/llvm-project] 6c4819: llvm-reduce: Error on -j with MIR inputs

[Matt Arsenault via All-commits]

  Branch: refs/heads/users/arsenm/llvm-reduce/no-parallel-mir-reduction
  Home:   https://github.com/llvm/llvm-project
  Commit: 6c4819239f477539ad5ceb9f4cdee40be3809e10
      https://github.com/llvm/llvm-project/commit/6c4819239f477539ad5ceb9f4cdee40be3809e10
  Author: Matt Arsenault <Matthew.Arsenault at amd.com>
  Date:   2026-09-24 (Thu, 24 Sep 2026)

  Changed paths:
    A llvm/test/tools/llvm-reduce/mir/parallel-jobs-unsupported.mir
    M llvm/tools/llvm-reduce/deltas/Delta.cpp
    M llvm/tools/llvm-reduce/deltas/Delta.h
    M llvm/tools/llvm-reduce/llvm-reduce.cpp

  Log Message:
  -----------
  llvm-reduce: Error on -j with MIR inputs

Parallel chunk processing distributes work by serializing the program to
bitcode and reparsing it in each worker thread. Bitcode cannot represent
MachineFunctions, so readBitcode leaves ReducerWorkItem::MMI null and the
reparsed item reports isMIR() == false. The delta pass list is still the
MIR one, chosen from the original program, so the worker runs a MIR pass
over an item with no MachineModuleInfo and dereferences null.

This broke every MIR reduction using -j > 1, though not immediately: the
parallel path is only taken once there is more than one chunk left to
process, so the crash appeared after the first granularity increase.

Reject the combination up front instead of crashing partway through a
reduction. Supporting it would mean exchanging MIR text rather than
bitcode, and giving each worker its own TargetMachine.

Co-Authored-By: Claude Opus 5 <noreply at anthropic.com>



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