[medvednikov]

Speeds up the self-host transform and trims serial work in cgen, and restores transform memory, which had regressed recently.

-prod v3 from vlib/v, ./v3 -nocache -building-v -o v4 -v v.v, M5 Max, best of 9 alternating runs (load average 7–15):

	master	this PR
transform	222–225 ms	143–145 ms
transform stage peak	~1273 MB	~868 MB
cgen	154–157 ms	152–161 ms
cgen stage peak	~1086 MB	~977 MB

On this shared machine cgen moves by ±15 ms between sessions, so its wall-clock change is within noise. The cgen subphases do shrink: the serial postamble goes from 5.3 to 1.5 ms, and the type-declaration task uses ~8 ms less CPU.

Commit 1: restore transform speed and memory

• Copy-on-write detach per batch (fn.v). call_param_types_from_decl recorded every unindexed name as a miss. On a helper batch, that detached (cloned) the shared prepared index, cache and miss maps; transform: bound comptime monomorphization memory #29334 added the index to that clone. Unindexed names are no longer recorded, since the index lookup costs the same as a recorded miss. Transform allocations drop from 4.1 GB to 2.0 GB.

• Escape precheck matched almost everything. Since v3: fix the bugs that stopped C translated by c2v (SQLite, libpq) from compiling and running #29167 and transform: preserve fixed-array references and owned borrowed storage #29197, any call with a pointer parameter or a pointer receiver forced the full escape walk. The precheck now mirrors the walk's own conditions:

  • a receiver rooted at a mut/& parameter can't move a local of this frame;
  • a non-void reference parameter only matters when an argument holds inline fixed-array storage.

  The walk also checks those cheap argument conditions before resolving callee signatures.

• driver.run lowered serially. v3: constraints on type parameters, editor answers in generic code, faster checks for editors #29254 added function literals to driver.run, the compiler's largest function, so it was transformed alone before the workers started. The callbacks are now small named functions.

• Check scratch retained into transform. Since checker: report hard errors in reachable dependency functions #29214, collect_selected_file_called_fns walks every reachable body, the whole compiler for a self build, and left ~120 MB of scratch in the compilation arena. Scoped checks now run the walk on a disposable fork.

• Worker balance and scratch. Large helper chunks use smaller scratch batches with a 1 MB arena floor. The master takes 85% of a helper's share. Scope save/restore, alias-chain normalization and escape alias lookups avoid needless map clones and allocations.

Commit 2: trim serial work around the parallel stages

transform:

• Builds the scope-owned base node list only when the driver won't keep the whole arena (the self-host path keeps it).
• Finds ignored $for flags and helper-scope text flags with memchr instead of a loop over one flag per AST node.
• Scans anonymous struct literal candidates on the pool.
• Replaces the remaining comparator literals in checker functions with named functions.

cgen:

• Records $embed_file payload nodes in the collect scan instead of rescanning the AST three times.
• Finds C extern declarations among the recorded declaration nodes instead of every transform-appended node.
• Searches runtime const initializers once per global instead of building four patterns per (global, initializer) pair.
• Skips the generic shared-type scan for declarations without shared fields.
• Per-declaration and per-function timing under -v now only runs with -d v3_ttime.

Verification

• The generated C is unchanged. Self-build C for a fixed source tree is identical before and after, with and without V3_NO_ESCAPE_PRECHECK=1. The master-share change only renumbers _str_N constants.
• An 859-program corpus (vlib/v/tests/*.v, examples/*.v, vlib/v/compiler_tests/*_test.v) produces the same C. The only difference is the embedded compiler commit hash.
• v test vlib/v/transform/ passes 33/33, vlib/v/types/ 97/97 and vlib/v/gen/c/ 51/51. fn_test.v is updated: a miss on an unindexed name no longer detaches the worker's shared maps.
• Self-host fixed point: v3 → v4 → v5 with v4 == v5 byte-identical.

Not addressed here: the eager called-function walk still costs ~190 ms serially in the check stage of a self build. It only matters when a dependency function has an error, so it could be deferred like the IError gating.

[medvednikov]

@codex review