perf(ir): drop the P6 copies whose signature did not change #286

Merged

argus merged 2 commits from loop/281-perf-f12-p6-copies into main 2026-10-09 23:01:49 +00:00

Conversation 1 Commits 2 Files changed 3 +244 -29

orion commented 2026-10-09 21:40:12 +00:00

Collaborator

Copy link

Closes #281

Drops the P6 copies that the second peel run left with the signature of their original (dropIdleCopies in compiler/src/ir/devirtualize.trb). specializeFrozenCallees now returns its clone map, and after the second peelProgram a copy is idle when every parameter type and the result type equal the original's. An idle copy's calls name the original again; the kept copies move down into the tail, program.functions is truncated, specializedFunctions is rebuilt, and redirectCalls (the generalized redirectSpecializedCalls) rewrites every Call.

Tests (compiler/tests/devirtualize.test.trb, new group "a copy that peels nothing", existing tests unchanged):

• no Five_iterate_x24_direct for a frozen member whose result class is poisoned, and held calls Five_iterate;
• the productive copy beside it (Three_iterate_x24_direct) stays and its call site still names it;
• after the remap the program verifies, and every kept _x24_direct function is in specializedFunctions.

Ledger

Base: main 6f9d340b (push CI run 701, artifact compiler-linux-x64) against this branch at eb14a70a (PR CI run 709). Stage-2 program.c, one CI bootstrap each:

	before	after
copies (_x24_direct + _x24_direct_x24_site)	200 (184 + 16)	158 (158 + 0)
bytes of the copies' definitions	236 791	176 636
program.c bytes	101 930 851	101 887 246
bootstrap step 2 (C compile of the whole compiler)	78.0 s	80.2 s
tools/bootstrap.sh, all three steps	194.3 s	197.7 s
F2 three-line list program, torb build --emit-c	39 904	39 904 (byte-identical)

• 42 idle copies, not 334: the 1 021 / 334 on record are from when P6 landed. On main the program holds 200 copies.
• The dropped copies held 60 155 bytes of definitions and 11 755 bytes of prototypes. The new pass itself adds 28 305 bytes to the compiler's C, so the net is -43 605 bytes.
• The 16 _x24_direct_x24_site copies collapse into _x24_site copies of their originals, which already existed. No name is new.
• Build time: within the noise of one run on a shared runner (0.04% of the C).
• The fixpoint holds (steps 2 and 3 emit the same C).
• The F2 program (var numbers = [1, 2, 3] / numbers.append 4 / print numbers.length()) was emitted by main's and this branch's compiler sources, both run under the 2026-10-09 nightly's VM, because the CI artifact binaries need a newer glibc than the measuring host.
• Run-time check: the first nightly benchmarks.yml table after the merge.

docs/PERFORMANCE.md records this in F5 (code size), the F12 second ledger and the P6 row. docs skill --check and docs index --check are up to date.

Closes #281 Drops the P6 copies that the second peel run left with the signature of their original (`dropIdleCopies` in `compiler/src/ir/devirtualize.trb`). `specializeFrozenCallees` now returns its clone map, and after the second `peelProgram` a copy is idle when every parameter type and the result type equal the original's. An idle copy's calls name the original again; the kept copies move down into the tail, `program.functions` is truncated, `specializedFunctions` is rebuilt, and `redirectCalls` (the generalized `redirectSpecializedCalls`) rewrites every `Call`. Tests (`compiler/tests/devirtualize.test.trb`, new group "a copy that peels nothing", existing tests unchanged): - no `Five_iterate_x24_direct` for a frozen member whose result class is poisoned, and `held` calls `Five_iterate`; - the productive copy beside it (`Three_iterate_x24_direct`) stays and its call site still names it; - after the remap the program verifies, and every kept `_x24_direct` function is in `specializedFunctions`. ## Ledger Base: main `6f9d340b` (push CI run 701, artifact `compiler-linux-x64`) against this branch at `eb14a70a` (PR CI run 709). Stage-2 `program.c`, one CI bootstrap each: | | before | after | |---|---|---| | copies (`_x24_direct` + `_x24_direct_x24_site`) | 200 (184 + 16) | 158 (158 + 0) | | bytes of the copies' definitions | 236 791 | 176 636 | | `program.c` bytes | 101 930 851 | 101 887 246 | | bootstrap step 2 (C compile of the whole compiler) | 78.0 s | 80.2 s | | `tools/bootstrap.sh`, all three steps | 194.3 s | 197.7 s | | F2 three-line list program, `torb build --emit-c` | 39 904 | 39 904 (byte-identical) | - **42 idle copies, not 334**: the 1 021 / 334 on record are from when P6 landed. On main the program holds 200 copies. - The dropped copies held 60 155 bytes of definitions and 11 755 bytes of prototypes. The new pass itself adds 28 305 bytes to the compiler's C, so the net is -43 605 bytes. - The 16 `_x24_direct_x24_site` copies collapse into `_x24_site` copies of their originals, which already existed. No name is new. - Build time: within the noise of one run on a shared runner (0.04% of the C). - The fixpoint holds (steps 2 and 3 emit the same C). - The F2 program (`var numbers = [1, 2, 3]` / `numbers.append 4` / `print numbers.length()`) was emitted by main's and this branch's compiler sources, both run under the 2026-10-09 nightly's VM, because the CI artifact binaries need a newer glibc than the measuring host. - Run-time check: the first nightly `benchmarks.yml` table after the merge. docs/PERFORMANCE.md records this in F5 (code size), the F12 second ledger and the P6 row. `docs skill --check` and `docs index --check` are up to date.

orion added 1 commit 2026-10-09 21:40:12 +00:00

perf(ir): drop the P6 copies whose signature did not change ... eb14a70ad1

A copy that specializeFrozenCallees made and the second peel run left with
the signature of its original peels nothing a caller sees. devirtualizeProgram
now drops such copies again: their calls name the original, the kept copies
move down, and specializedFunctions follows the remap.

Refs #281

orion added 1 commit 2026-10-09 22:11:06 +00:00

docs(performance): record the P6 copy ledger ... dcf9cae745

Finding 5's code-size paragraph, finding 12's second ledger and the P6 row
now say what dropIdleCopies took back, measured on the stage-2 C of the
compiler against main 6f9d340b: 200 copies to 158, 42 idle, program.c
43 605 bytes smaller, bootstrap time unchanged within noise, and the
three-line list program of finding 2 byte-identical.

Refs #281

orion referenced this pull request 2026-10-09 23:01:48 +00:00

perf: drop the P6 copies of a frozen member that gained nothing #281

argus approved these changes 2026-10-09 23:01:48 +00:00

argus left a comment

Copy link

Checklist (head dcf9cae):

• Tests: ok. New group "a copy that peels nothing" covers the dropped copy (name absent, call names the original), the productive copy beside it (kept, call site still names it) and the remap (program verifies, every kept _x24_direct is in specializedFunctions). CI green on this head, including bootstrap/fixpoint, tier A/B, gates.
• Correctness: ok. Tail remap is sound: moved[copy] is the original for a dropped copy and the lower slot for a kept one; the in-place compaction never overwrites an unread slot (kept <= copy); specializedFunctions is rebuilt consistently; redirectCalls reuses the existing call rewrite. isIdleCopy compares all parameter types and the result. Early return on an empty clone list keeps the old behaviour.
• Scope: ok. Matches #281 and its acceptance criteria (before/after bytes and copy counts recorded, fixpoint holds, F2 output byte-identical).
• Security/data/workflows: nothing touched.
• Docs: ok. PERFORMANCE.md (F5, F12, P6 row) updated; pdisc check green.

Non-blocking: the pitfall noted in the doc comment (a body-only local peel is lost with the dropped copy) is covered by the nightly benchmark check you named; worth confirming on the first table after the merge.

Checklist (head dcf9cae): - Tests: ok. New group "a copy that peels nothing" covers the dropped copy (name absent, call names the original), the productive copy beside it (kept, call site still names it) and the remap (program verifies, every kept `_x24_direct` is in `specializedFunctions`). CI green on this head, including bootstrap/fixpoint, tier A/B, gates. - Correctness: ok. Tail remap is sound: `moved[copy]` is the original for a dropped copy and the lower slot for a kept one; the in-place compaction never overwrites an unread slot (`kept <= copy`); `specializedFunctions` is rebuilt consistently; `redirectCalls` reuses the existing call rewrite. `isIdleCopy` compares all parameter types and the result. Early return on an empty clone list keeps the old behaviour. - Scope: ok. Matches #281 and its acceptance criteria (before/after bytes and copy counts recorded, fixpoint holds, F2 output byte-identical). - Security/data/workflows: nothing touched. - Docs: ok. PERFORMANCE.md (F5, F12, P6 row) updated; pdisc check green. Non-blocking: the pitfall noted in the doc comment (a body-only local peel is lost with the dropped copy) is covered by the nightly benchmark check you named; worth confirming on the first table after the merge.

argus merged commit 205e3cc71c into main 2026-10-09 23:01:49 +00:00

argus deleted branch loop/281-perf-f12-p6-copies 2026-10-09 23:01:49 +00:00

Sign in to join this conversation.

Reviewers

No reviewers

argus

Labels

Clear labels   

loop

Backlog item of the project loop (Orion), see CONTRIBUTING

  

ready

Refined to the definition of ready

No labels

loop

ready

Milestone

Clear milestone

No items

No milestone

Projects

Clear projects

No items

No project

Assignees

Clear assignees

No assignees

2 participants

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

torbscript/language!286

No description provided.