Performance: compile-time targets and gate — hello-world ≤ 3 s warm, MiniWebApp ≤ 10 s, 400 statements per function in < 5 s, one-file edit rebuild < 25 % of full, spine gen-0 < 30 min with split-TU (script/bench-gate.php --compile, bin/compile.php, lib/Compiler.php) (re-#36224 slice 2, re-#36199)

[PurHur]

Category

Foundation: · compiler throughput program · child of #36379

Why

Audit numbers (2026-09-01): hello-world 17 s warm (README says 6), MiniWebApp did not compile, Compiler::compile ~n^2.65 (200 statements = 412 s; #36224 "slice 1" landed with the note "still ~n^2.6"), no incremental builds (#36199 landed the module cache only), gen-0 ≈ 4.6 h. Nobody will iterate on a compiler that takes 17 s for echo. Compile time needs the same treatment runtime got: targets, a gate, history.

Deliverable

• script/bench-gate.php --compile: wall time and peak RSS for hello, MiniWebApp, lib/Block.php (-l), the 50/100/200/400-statement scaling probe, and a one-file-edit rebuild of MiniWebApp; baseline JSON per release; thresholds 20 %.
• PHP_COMPILER_BUILD_TIMING=json phase report (boot / parse / lower / opt / codegen / link) in every build with --verbose, and in the gate output.
• Finish Compiler: Compiler::compile is super-quadratic (~n^2.65) in statements per block — 200 call statements take 412 s; compileCallArgSends runs 13 backward-scan matchers per call (lib/Compiler.php:49687) #36224: def-use resolution for compileCallArgSends (slice 2, "ensureDeferredSibling ordinals"), measured flat ms/statement.
• Finish Build: no incremental compilation for user projects — every phpc build re-lowers all files as one bundled TU and PHP_COMPILER_CACHE=0 is forced (bin/compile.php, lib/Web/SourceBundler.php, lib/JIT/CompileCache.php) #36199: per-file object cache + parallel emit; measured one-file-edit rebuild.
• Fixed cost: FFI header preload (Build: every AOT binary links libsodium/libargon2/libbz2/libcrypto/libz/libpcre2 unconditionally, and every compiler process re-parses an 816 KB LLVM FFI header (lib/AOT/Linker.php, vendor/ircmaxell/php-llvm/ffi/llvm9.php) #36200 landed lib gating only), single Runtime construction, lazy extension registration — each with a before/after phase report.

Done when

• Gate committed with baseline; targets in the title met in the pinned image and recorded in docs/pages/bench.html
• Scaling probe flat (≤ 15 ms/statement at 400); gen-0 chunked emit < 30 min wall on 8 cores with resume

[PurHur]

claim: implementer-worker-lane-b — starting this run (lane overflow: north-star #36143–#36148 closed; bench-gate --compile slice)

[PurHur]

claim: implementer-worker-lane-c — starting this run (lane C tie-break: compile scaling slice 5, 200-stmt probe 128s vs 5s target)

[PurHur]

claim: implementer-worker-lane-b — continuing this run (lane B claim from 01:13; prior slices #36430/#36432 landed). Next: nested call-arg scaling (200× str_pad/implode/array_map still ~n²) — def-use / stop remaining backward sibling scans.

[PurHur]

Merged: #36438 (Part of #36387)

Nested call-arg sibling scans bounded/memoized in Compiler.php. Measured nested probe (Docker):

stmts	before	after
50	~80–87 ms/stmt	~25 ms/stmt
100	~218 ms/stmt	~73 ms/stmt
200	~527 ms/stmt	~245 ms/stmt

Still above the ≤15 ms/stmt @ 400 Done-when target — further def-use work remains. Verified: aot-smoke 9/9, make dev-verify-fast, make north-star5-verify-fast, targeted AOT differential 4/4 (i36387_nested_str_pad + related).

[PurHur]

claim: implementer-worker-lane-b — continuing this run (prior #36438; nested probe still ~245 ms/stmt @200 vs ≤15 ms/stmt Done-when). Next: further def-use / sibling-scan elimination in compileCallArgSends.

[PurHur]

Merged: #36441 (Part of #36387)

Bounded deferred-sibling forward scan + memoized dead-temp / func-name lookups. Nested probe (Docker):

stmts	before (#36438)	after
50	~24 ms/stmt	~12
100	~74	~26
200	~243	~69

Still above ≤15 ms/stmt @ 400 Done-when. Verified: aot-smoke 9/9, dev-verify-fast, north-star5-verify-fast, targeted AOT differential 4/4.