Skip to content

gen-c: run defer at every exit of its block (Closes #7352) - #7406

Draft
gHashTag wants to merge 4 commits into
masterfrom
genc-defer
Draft

gen-c: run defer at every exit of its block (Closes #7352)#7406
gHashTag wants to merge 4 commits into
masterfrom
genc-defer

Conversation

@gHashTag

@gHashTag gHashTag commented Oct 7, 2026 •
edited
Loading

Copy link
Copy Markdown
Owner

Closes #7352
Part of #5980

What was wrong

gen-c wrote defer S; as the comment /* unsupported: StmtExpr */; and exited 0. The deferred statement never ran: a test that depended on it trapped at run time, and one that did not passed while the cleanup it claimed never happened.

What gen-c does now

The deferred statement is written at every exit of the block that declared it, the way Zig runs it:

  • at the fall-through end of the block (skipped when the block already ends in return, break or continue);
  • before a return, after the returned value is read: { T __t27_ret = <value>; <defers>; return __t27_ret; } (__auto_type where the fn's return type is not known);
  • before a break or continue, for the blocks up to and including the loop body that declared it;
  • innermost block first, each block in reverse order of declaration.

Blocks covered: fn bodies, invariant and bench bodies, test bodies, if/else branches, while and for bodies.

errdefer is not lowered: gen-c now refuses it by name and line (gen-c: \errdefer` at line N is not lowered to C (#7352), non-zero exit) instead of dropping it. A defer` outside any block is refused the same way.

Tests

bootstrap/tests/genc_defer.rs, 9 tests. Each one compiles the generated C with -DT27_TEST_MAIN and runs it, because the defect compiled cleanly and only running shows the call happened: the issue repro, reverse order, inner-before-outer on return, early return, value read before the defer runs, break and continue running the loop body's defer, a loop defer not running again on the fn's return, and the errdefer refusal.

Lab (t27c-lab), cargo test --test genc_defer: 9 passed on this branch; with compiler.rs and FROZEN_HASH reverted to master, 9 failed.

C output over the corpus

gen-c over all 1527 tracked specs, master binary vs this branch: 13 outputs change, all of them specs that use defer:

  • specs/api/c_api_contract.t27
  • specs/port/tools/jtag/link_relay.t27
  • specs/port/tools/jtag/read_user1.t27
  • specs/port/tools/jtag/read_verdict.t27
  • specs/port/trinity/src/phi-engine/vibeec_original/bogatyrs_registry.t27
  • specs/port/trinity/src/tri/gen_gradient.t27
  • specs/tri/encoding/json.t27
  • specs/tri/io/compress.t27
  • specs/tri/io/filesystem.t27
  • specs/tri/net/url.t27
  • specs/tri/t27b/conformance/scope_exit.t27
  • specs/tri/utils/config.t27
  • specs/tri/utils/text.t27

Every diff is the same three moves and nothing else: the /* unsupported: StmtExpr */; lines are gone, the deferred calls are written at block ends, and returns in a fn with a pending defer are wrapped as above. (read_user1 gets one unreachable mpsse_close after an if/else whose branches both return; harmless.)

cc -fsyntax-only error counts are identical before and after for 12 of the 13. compress.t27 goes 36 -> 37: a return of the Zig-only error.OutOfMemory now lands in an __auto_type temporary; the file did not compile before either. scope_exit.t27 (the t27b conformance spec for defer) now compiles and runs: "All 2 tests passed."

Seals

7 of the 13 have seals; each was fresh on master and stale with the new C, and is resealed here with its twin (14 files): c_api_contract, json, compress, filesystem, url, config, text. Only gen_hash_c (plus sealed_at, built_by, tests) moves; spec, Rust, Verilog and Zig hashes are unchanged. t27c seal --verify exits 0 for all 7 on the lab. The other 6 have no seal on master either. One more seal sits outside specs/: contrib/backend/zig/legacy/main_zig_handwritten.t27 (27 defers), which the lab's seal-currency gate named on 5ffb207 (gen_hash_c sealed 88c4ba2b8d2f, current 2eabfadaa47c). It is resealed in d8e71dd; again only gen_hash_c, built_by and sealed_at move. A gen-c comparison of all 28 sealed specs outside specs/ (master 18689bc vs this fix) shows no other change, so no other seal goes stale.

Foreign code

This edits Rust (bootstrap/src/compiler.rs + its FROZEN_HASH seal, and a new test file), so it carries owner-approved-foreign and a tools/policy/foreign-exceptions.txt entry naming #7352.

🤖 Generated with Claude Code

gHashTag and others added 2 commits October 7, 2026 15:46
@gHashTag @claude
fix(gen-c): lower defer at every exit of its block (Closes #7352)
c6f1eec 
gen-c wrote `defer S;` as `/* unsupported: StmtExpr */;` and exited 0, so
the deferred statement never ran. It is now recorded on the innermost open
block and written at each exit of that block, in Zig's order: at the
fall-through end, before `return` (the returned value goes into a temp
first, so a defer that changes what it reads does not change it), before
`break`/`continue` of the loop whose body declared it; innermost block
first, each block in reverse order. `errdefer`, and a `defer` with no open
block, are refused by name and line instead of dropped.

Regression test: bootstrap/tests/genc_defer.rs compiles and runs the C.

Part of #5980.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@gHashTag @claude
seals: reseal the 7 specs whose C output now runs their defers (#7352)
d422294 
gen_hash_c moves for c_api_contract, json, compress, filesystem, url,
config and text (and each twin); no other hash moves.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@gHashTag gHashTag added the owner-approved-foreign Owner-approved exception to the only-t27 rule: hand-written foreign code allowed in this PR label Oct 7, 2026
@gHashTag
gHashTag enabled auto-merge (squash) October 7, 2026 08:56
This was referenced Oct 7, 2026
Merged
Merged
Merged
@gHashTag
Merge remote-tracking branch 'origin/master' into genc-defer
5ffb207 
# Conflicts:
#	bootstrap/stage0/FROZEN_HASH
This was referenced Oct 7, 2026
Merged
Merged
Merged
Merged
Closed
@github-actions

github-actions Bot commented Oct 7, 2026

Copy link
Copy Markdown
Contributor

PR Dashboard

Generated at: 2026-10-07 09:46:09 UTC

Summary

Status Count
Total Open PRs 50
PRs with Failing Checks 41
PRs with All Checks Green 9
READY 1
FAILING 41
PENDING 0
NO CHECKS YET 0

These columns do not partition: 1 + 41 + 0 + 0 = 42, and there are 50 open PRs. A PR is being counted twice or not at all.

Seal Status

  • ⚠️ STALE -- sha256(compiler.rs)=c0f9065938c4 != manifest seal=87e5cbd3ad94.
    The committed NMSE numbers were certified against an older compiler.rs.
    Run scripts/reseal-check.sh locally for the two-step reseal command (advisory; not a merge gate).

This was referenced Oct 7, 2026
Merged
Merged
Open
@gHashTag @claude
seal: reseal main_zig_handwritten for the defer lowering (#7352)
d8e71dd 
contrib/backend/zig/legacy/main_zig_handwritten.t27 has 27 defers, so its C
moves; only gen_hash_c, built_by and sealed_at change. The lab's seal-currency
gate named it (current=2eabfadaa47c).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This was referenced Oct 7, 2026
Merged
Merged
@gHashTag
gHashTag disabled auto-merge October 7, 2026 09:59
@gHashTag
gHashTag marked this pull request as draft October 7, 2026 09:59
@gHashTag

gHashTag commented Oct 7, 2026

Copy link
Copy Markdown
Owner Author

Parked: this PR is over the foreign-line budget that #7399 (owner rule 2026-10-07, #7371) put on master while it waited for CI. It is now a draft, with auto-merge off.

check_budget() from master's gen/c/policy/own_language.c, run over git diff --numstat --no-renames origin/master...origin/genc-defer, denies it:

  • bootstrap/src/compiler.rs: 189 added lines (limit 40 per file)
  • bootstrap/tests/genc_defer.rs: 207 added lines (limit 40 per file)
  • all files: 396 added hand-written Rust lines (limit 80 per diff)

No exception and no label lifts this limit. Splitting the same Rust into 80-line PRs would get around the rule rather than follow it, so I have not done that. The fix itself is still correct and tested on the t27c lab: the regression tests fail on master's compiler and pass with the fix, and legacy_main_zig_handwritten is resealed. Landing it needs the defer lowering written in .t27 (gen-rust into bootstrap/gen/rust), or an owner decision.

This was referenced Oct 7, 2026
Open
Open
Open
This was referenced Oct 7, 2026
Merged
Draft
Open
Draft
This was referenced Oct 8, 2026
Open
Merged
Open
Merged
Merged
Closed
Merged
Open
Merged
Merged
Merged
Merged
Merged
Merged
Merged
Merged
Merged
Merged
Merged
Merged
Open
Open
Merged
Merged
Merged
Merged
Merged
Merged
Open
Open

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

No reviews

Assignees

No one assigned

Labels

owner-approved-foreign Owner-approved exception to the only-t27 rule: hand-written foreign code allowed in this PR

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

Self-host: gen-c writes defer as an unsupported comment and the deferred call never runs

1 participant

@gHashTag