WIP: [Mips] Partially unroll small loops on the R5900 #12

Closed

techwritescode wants to merge 6 commits from r5900-unroll into r5900-ee

pull from: r5900-unroll

merge into: ps2:r5900-ee

ps2:r5900-ee

ps2:r3000a-delay-slots

ps2:r3000a-load-delay

ps2:ps2-mc-fill-warning

ps2:main

Conversation 1 Commits 6 Files changed 12 +159 -2

techwritescode commented 2026-10-14 22:46:49 +00:00

Owner

Copy link

Adds an R5900 getUnrollingPreferences enabling partial and runtime unrolling (PartialThreshold 64), with a lit test. Also renames the split-hi16 pass argument, which collided with the -mips-split-hi16 option and made opt abort.

Console cycles (-O2): packet_build 61,559 to 55,734; raster 7.36M to 7.50M (+2%); decompress, physics flat. At -O3: packet_build flat, raster +3%.

Mips CodeGen and MC lit tests pass.

🤖 Generated with Claude Code

Adds an R5900 getUnrollingPreferences enabling partial and runtime unrolling (PartialThreshold 64), with a lit test. Also renames the split-hi16 pass argument, which collided with the -mips-split-hi16 option and made opt abort. Console cycles (-O2): packet_build 61,559 to 55,734; raster 7.36M to 7.50M (+2%); decompress, physics flat. At -O3: packet_build flat, raster +3%. Mips CodeGen and MC lit tests pass. 🤖 Generated with [Claude Code](https://claude.com/claude-code)

techwritescode added 2 commits 2026-10-14 22:46:49 +00:00

[Mips] Give the split-hi16 pass its own name so opt can start ... 6610a8a0a1

The pass argument matched the -mips-split-hi16 option, so opt aborted with a duplicate option registration.

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>

[Mips] Partially unroll small loops on the R5900 ... 7bb528e86e

Enable partial and runtime unrolling with a threshold of 64. On the console this cuts packet_build 9% at -O2 and leaves other kernels within 3%.

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>

techwritescode changed title from [Mips] Partially unroll small loops on the R5900 to WIP: [Mips] Partially unroll small loops on the R5900 2026-10-14 22:48:46 +00:00

techwritescode added 3 commits 2026-10-15 05:51:21 +00:00

[Mips] Unroll and jam loop nests on the R5900 ... 371ce32206

Mips never enabled unroll-and-jam, so the pass did nothing. For the R5900:
set UP.UnrollAndJam, allow multi-block inner loops, and prefer a jam count
of 6. The 16 KB data cache thrashes when an outer loop sweeps a larger array
once per iteration; jamming several outer iterations per sweep cuts the
misses (a 1024-particle physics step: 14.7M -> 6.6M cycles on hardware).

The jam clones the inner induction variable, so the R5900 pipeline runs
instcombine, indvars, early-cse and GVN after it to forward the copies'
loads and stores.

Adds TTI::UnrollingPreferences::UnrollAndJamMultiBlock and
UnrollAndJamPreferredCount (defaults off).

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>

Enable non-trivial loop unswitch at -O2 for R5900 ... 9302ab0c61

Unswitching the loop-invariant triangle-area test out of the raster pixel loop
cuts bench_raster at -O2 from 7.49M to 6.84M cycles (GCC: 6.87M).

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>

R5900: unroll and jam at -O1 ... 734ca8cdd6

Enable loop unrolling by default at -O1 for r5900, and run
AggressiveInstCombine and GVN before unroll-and-jam there so the sqrtf call
and redundant loads no longer fail the jam dependence check. bench physics
-O1 goes from 14.9M to 7.59M cycles, raster from 8.03M to 7.49M.

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>

techwritescode added 1 commit 2026-10-15 05:57:23 +00:00

[Mips] Move R5900 pass pipeline tuning into the Mips target ... 455c9cd951

Register the R5900 unswitch, unroll-and-jam and cleanup passes through
MipsTargetMachine::registerPassBuilderCallbacks instead of branching on
the CPU in PassBuilderPipelines.cpp, and set the -O1 unroll default in
the MIPS driver args instead of CompilerInvocation.

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>

techwritescode referenced this pull request 2026-10-15 16:56:18 +00:00

Keep the split-hi16 rename, state what belongs in the fork #13

techwritescode commented 2026-10-15 16:58:28 +00:00

Author

Owner

Copy link

Closing: superseded by #13. The split-hi16 rename went in there. The rest changes generic LLVM passes (LoopUnrollAndJam, LoopUnrollPass, TargetTransformInfo.h), adds an R5900 pass pipeline, adds -funroll-loops at -O1 in the clang driver, and uses unroll thresholds tuned to the benchmark, which UPDATING.md now keeps out of the fork. Partial unroll alone made -O2 bench code 1.70x GCC's size on the console. The branch stays on the remote for reference.

Closing: superseded by #13. The split-hi16 rename went in there. The rest changes generic LLVM passes (LoopUnrollAndJam, LoopUnrollPass, TargetTransformInfo.h), adds an R5900 pass pipeline, adds -funroll-loops at -O1 in the clang driver, and uses unroll thresholds tuned to the benchmark, which UPDATING.md now keeps out of the fork. Partial unroll alone made -O2 bench code 1.70x GCC's size on the console. The branch stays on the remote for reference.

techwritescode closed this pull request 2026-10-15 16:58:29 +00:00

Some checks failed

Hide checks

action.yml / [Mips] Move R5900 pass pipeline tuning into the Mips target (push) Failing after 0s

Details

release-binaries-all.yml / [Mips] Move R5900 pass pipeline tuning into the Mips target (push) Failing after 0s

Details

release-binaries.yml / [Mips] Move R5900 pass pipeline tuning into the Mips target (push) Failing after 0s

Details

release-documentation.yml / [Mips] Move R5900 pass pipeline tuning into the Mips target (push) Failing after 0s

Details

release-sources.yml / [Mips] Move R5900 pass pipeline tuning into the Mips target (push) Failing after 0s

Details

Labelling new pull requests / greeter (pull_request_target) Has been skipped

Details

Labelling new pull requests / automate-prs-labels (pull_request_target) Has been skipped

Details

action.yml / [Mips] Move R5900 pass pipeline tuning into the Mips target (pull_request) Failing after 0s

Details

Check LLVM ABI annotations / Check LLVM_ABI annotations with ids (pull_request) Has been skipped

Details

CI Checks / Build and Test Linux (pull_request) Has been skipped

Details

CI Checks / Build and Test Windows (pull_request) Has been skipped

Details

CI Checks / Compute macOS Projects (pull_request) Has been skipped

Details

CI Checks / Build and Test macOS arm64 (pull_request) Has been skipped

Details

release-binaries-all.yml / [Mips] Move R5900 pass pipeline tuning into the Mips target (pull_request) Failing after 0s

Details

release-binaries.yml / [Mips] Move R5900 pass pipeline tuning into the Mips target (pull_request) Failing after 0s

Details

release-documentation.yml / [Mips] Move R5900 pass pipeline tuning into the Mips target (pull_request) Failing after 0s

Details

release-sources.yml / [Mips] Move R5900 pass pipeline tuning into the Mips target (pull_request) Failing after 0s

Details

Add buildbot information to first PRs from new contributors / buildbot_comment (pull_request_target) Has been skipped

Details

Pull request closed

Please reopen this pull request to perform a merge.

Sign in to join this conversation.

Reviewers

No reviewers

Labels

Clear labels

No items

No labels

Milestone

Clear milestone

No items

No milestone

Projects

Clear projects

No items

No project

Assignees

Clear assignees

No assignees

1 participant

Notifications

Subscribe

Due date

The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference

ps2/llvm!12

No description provided.