Skip to content

t27b: @intCast with an integer result type lowers from a t27 plan (Closes #7412) - #7627

Merged
gHashTag merged 2 commits into
masterfrom
claude/t27b-intcast-plan
Oct 7, 2026
Merged

t27b: @intCast with an integer result type lowers from a t27 plan (Closes #7412)#7627
gHashTag merged 2 commits into
masterfrom
claude/t27b-intcast-plan

Conversation

@gHashTag

@gHashTag gHashTag commented Oct 7, 2026

Copy link
Copy Markdown
Owner

Closes #7412. Part of #6063 (t27b coverage, lane 4), which is item C10 of #6488; written in t27 under the port epic #6198. Supersedes draft #7417, which did the same in 58 hand-written lines in lower.rs and a 120-line Rust test file.

What

t27b now lowers @intCast(x) when the context gives it an integer result type: @as(T, ..), a typed binding, a parameter, or a return. gen-zig prints the call as written, so t27b follows Zig 0.16 Debug. The semantics are #7417's:

  • A lossless conversion widens: same signedness and not narrower, or unsigned into a strictly wider signed type. A constant operand whose value is kept folds.
  • Any other conversion is a checked Cast that traps when the value is outside the result type. That includes unsigned to unsigned, where as truncates.
  • In wrap mode (--overflow wrap) such a conversion truncates, as as does there.
  • Refused:
    • a comptime_int literal that does not fit (Zig refuses it at compile time, in either mode);
    • an operand that is not an integer;
    • a result type that is not an integer;
    • any operand count but one.
  • @intCast with no result type stays refused, as before.

What t27b decides is a t27 spec. specs/tri/t27b/int_cast_plan.t27 holds:

  • shape: the operand count and the result type, asked before the operand is lowered;
  • widens: Zig's lossless rule over widths and signedness, odd widths such as u4 and i21 included;
  • plan: the decision table. It maps the operand's kind, the two types, constness, fit and the overflow mode to FOLD, FOLD_WRAP, KEEP, WIDEN, CHECK, TRUNCATE or a refusal;
  • which answers build a Cast, and which of those trap;
  • the construct a refusal names, its words, and the word a trap site starts with.

t27c gen-rust turns it into gen/rust/tri/t27b/int_cast_plan.rs. cli/t27b/src/lower.rs mounts that file with #[path], as it mounts builtin_plan.rs.

What the hand-written glue does (36 lines in lower.rs):

  • routes @intCast with an integer result type from expr_as;
  • lowers the operand and classifies it: typed integer, comptime_int, or other;
  • reads the widths and signedness of the two types, and whether a constant fits (Ty::fits);
  • asks plan, and builds the IR node the answer names.

Trap cases, written in t27

#7417 kept its trap cases in a 120-line Rust file. Here they are a t27 file, cli/t27b/tests/int_cast_traps.t27. Each test's name ends in its verdict, _expect_pass or _expect_trap, the way specs/tri/t27b/fuzz_oracle.t27 reads a generated test. Every assert of a trap test expects the truncated value, so a cast that truncated instead of trapping would pass it.

cli/t27b/tests/tail.rs runs the file in 19 lines:

  • Every _expect_trap test must trap with TrapKind::Cast, on a line that holds @intCast, in the interpreter and the JIT alike.
  • Every other test must pass, and exactly 7 must trap.
  • Its last two lines check the two shapes Zig refuses at compile time: a literal that does not fit, and no result type.

The file is not under specs/. Its tests fail by design, and the corpus counts a spec with a failing test as not passing.

The reference agrees with it, test by test:

  • t27c test-report fails exactly the 7 _expect_trap tests and passes the other one.
  • zig test --test-filter on its gen-zig output prints panic: integer does not fit in destination type at the @intCast line.
  • On the t27b lab, t27b corpus <dir> --reference <t27c> over that one file gives t27b fail and reference fail, with reference_disagree 0 files and 0 tests.

The passing cases are #7417's conformance spec, specs/tri/t27b/conformance/int_cast.t27. Only its header comment changed, to point at the trap file and the plan.

Lines

file kind added deleted
cli/t27b/src/lower.rs hand-written glue 36 0
cli/t27b/tests/tail.rs hand-written test 19 0
gen/rust/tri/t27b/int_cast_plan.rs t27c gen-rust, not hand-edited 131 0
specs/tri/t27b/int_cast_plan.t27 spec, 11 tests 265 0
specs/tri/t27b/conformance/int_cast.t27 #7417's conformance spec, 6 tests 99 0
cli/t27b/tests/int_cast_traps.t27 trap cases, 8 tests 81 0
.trinity/seals/*.json (2) t27c seal --save 40 0
docs/reports/t27b_expectations.json ledger 7 5
AGENTS.md t27b remainder line 4 1

Hand-written foreign code is 55 added lines: 36 in lower.rs and 19 in tail.rs. That is under the #7371 cap of 40 per file and 80 per PR. Both files were already in master's tools/policy/foreign-exceptions.txt. This PR does not change that file and needs no label.

Hand-written Rust in cli/t27b, measured with wc -l: src/*.rs goes from 14812 to 14848, and tests/*.rs from 7814 to 7833. The AGENTS.md line now says so.

Gates (t27c Railway lab, t27c built from master 1120d30, zig 0.16.0)

  • check_budget(), master's gen/c/policy/own_language.c with the one-line C main from lefthook.yml, over git diff --numstat --no-renames origin/master...HEAD: exit 0.
  • check_all(), with the exception list from origin/master:tools/policy/foreign-exceptions.txt, in both the -- (pre-push) and --ci forms: exit 0.
  • Negative controls: each of these is denied:
  • tools/l2_regen_check.py --base 1120d30ae: t27c gen-rust specs/tri/t27b/int_cast_plan.t27 reproduces the generated file byte for byte. gen-rust gave the same bytes over 3 runs.
  • Seals: tools/check_seal_coverage.py and tools/check_seal_currency.py exit 0. Both new seals pass t27c seal --verify with all hashes MATCH.
  • tools/check_assertionless_spec_tests.py: ok, 3761 in 30, as on the baseline.
  • tools/dupe_scan.py: it reports only the hex_nibble_of group that its base 1120d30 already reports. --like finds nothing from either new spec.

Spec checks

t27c test-report:

  • int_cast_plan.t27: 11/11 pass, 0 of 11 vacuous, 153 runtime asserts.
  • the conformance spec: 6/6 pass, 0 of 6 vacuous.
  • the trap file: 1 pass and 7 FAIL, exactly the _expect_trap tests.

Mutants. I ran 18 mutants of the plan spec, each on a copy, under t27c test-report. All 18 were killed:

  • same signedness widens only to a strictly wider type;
  • unsigned widens into a signed type of the same width;
  • signed widens into a wider unsigned type;
  • a literal that does not fit folds in wrap mode;
  • a literal that fits is refused in trap mode;
  • a widening constant is not folded;
  • a narrowing constant that fits is checked, not folded;
  • a fitting constant folds before wrap mode is asked;
  • wrap mode is ignored;
  • unsigned to unsigned truncates, as as does;
  • a non-integer operand is not refused;
  • the result type is asked before the operand count;
  • the same type is widened, not kept;
  • a truncating cast traps;
  • TRUNCATE builds no Cast;
  • REFUSE_LITERAL is not a refusal;
  • the literal refusal says something else;
  • the refusal names another construct.

t27b lab: full corpus with the reference

Master base 1120d30. These are the lab's own scheduled run numbers: https://t27b-lab-production.up.railway.app/runs/1120d30aed7f5508c44d1d2a62537d04eb0c330b.json. The run used qemu-aarch64 t27b corpus specs --json --runner qemu-aarch64 --timeout-ms 60000 --jobs 24 and 12 reference workers.

Branch d62ac63.

  • A driver in a private work dir in the same container imported the deployed /opt/t27b-lab/lab.py (blob 070419d2a) and ran its lab_run.
  • T27_SRV and T27_WORK pointed at that dir, so the lab's /srv and /work were not touched.
  • Same image and toolchain, --jobs 12, 3 reference workers, an empty reference cache, fuzz off.
  • The run JSON is not published at the lab URL. Its sha256 is a426b79a2e84d3ff....
  • The head e9583a1 adds only the ledger rows below, so it changes nothing the run measures.
master 1120d30 branch d62ac63
files 1568 1570 (+2: the new specs)
pass 872 876
pass_vacuous 151 151
blocked 526 524
fail 16 16
timeout 2 2
mismatch (= jit_interp_mismatch) 1 1
reference pass 1082 1084
reference disagree 0 0
cargo test --release -p t27b (aarch64, qemu) 128 passed, 0 failed 129 passed, 0 failed

Per file:

  • These two move from blocked (ExprCall(@intCast)) to pass, as t27b: @intCast with an integer result type is a checked cast; gen_fuzz passes (Closes #7412) #7417 said they would:
    • specs/port/trinity/src/tri/gen_fuzz.t27: 8 tests, 545 runtime asserts;
    • specs/port/trinity/src/server_main.t27: 4 tests, 22 runtime asserts.
  • Both new specs pass: int_cast_plan.t27 with 11 tests and 153 asserts, the conformance spec with 6 tests and 24 asserts.
  • Eight stay blocked with a later first blocker:
    • base/ternary_memory: TestBlock;
    • isa/ternary_arithmetic: ExprCall(@rem);
    • isa/ternary_shift: InvariantBlock;
    • ml/recurrent/attention_mechanism and tri/utils/random: ExprCall(undeclared fn);
    • numeric/gf16 and numeric/tf3: type mismatch;
    • queen/brain_summaries: StmtAssign.
  • numeric/ocp_mx and tri/search/rabin_karp still have an @intCast blocker. There the call is an array index or an arithmetic operand, which gives it no integer result type.
  • Nothing leaves pass. Every other file has the same verdict, first blocker, test count and assert count on both runs.
  • The new tail.rs test, int_cast_traps_outside_its_result_type, passed under qemu-aarch64 with the JIT.

The one mismatch is specs/port/trinity/src/tri/gen_canvas.t27, on master as well. It is inherited and filed as #7550.

Ledger

tri t27b ratchet --bless of the branch run gave the rows of the 4 files whose verdict this PR changes. Every other row is master's.

  • gen_fuzz.t27 and server_main.t27: blocked -> pass.
  • int_cast_plan.t27 and conformance/int_cast.t27: new rows, pass.
  • Counts: pass 854 -> 858, not_pass 61 -> 59, max_not_pass 61 -> 59.
  • Size: 1067 rows, no duplicate path.
  • Ratchet result: against the new ledger, the branch run shows exactly master's existing findings: UNEXPECTED PASS 2, UNLISTED 18, STALE 1. It adds one informational MOVED: the gf16.t27 blocker goes from ExprCall(@intCast) to type mismatch. Its verdict stays blocked, so its row is master's.

gen-rust gaps

None. gen-rust expressed the whole plan. The plan takes "does this constant fit" as a bool from the existing Ty::fits, so no value wider than 64 bits crosses into it.

For the owner

  • A typed constant that does not fit: in trap mode it becomes a Cast that traps at run time, as in t27b: @intCast with an integer result type is a checked cast; gen_fuzz passes (Closes #7412) #7417. Zig refuses such a cast at compile time when the constant is comptime-known, so the reference blocks those files and they stay outside the denominator.
  • FOLD and FOLD_WRAP: the glue builds both as ty.wrap(c). A FOLD value always fits ty, so the wrap keeps it unchanged. The plan keeps the two answers apart, and its tests pin when each one applies.
  • Where the trap cases live: they sit outside specs/, as a t27 file whose test names carry their verdicts. A trap test can only fail under the reference, and a conformance spec must pass.

Generated with Claude Code

claude and others added 2 commits October 8, 2026 05:45
@claude
t27b: @intcast with an integer result type lowers from a t27 plan (Cl…
d62ac63 
…oses #7412)

`@intCast(x)` where the context gives an integer result type (`@as`, a
typed binding, a parameter, a return) now lowers the way Zig 0.16 Debug
runs it: a lossless conversion widens, a constant whose value is kept
folds, any other conversion is a checked Cast that traps outside the
result type (unsigned to unsigned included, unlike `as`), and wrap mode
truncates. A literal that does not fit, a non-integer operand or result
type, and a wrong operand count stay refused.

What t27b decides is specs/tri/t27b/int_cast_plan.t27 (11 tests), and
`t27c gen-rust` turns it into gen/rust/tri/t27b/int_cast_plan.rs, which
cli/t27b/src/lower.rs mounts with #[path]. The glue there is 36 lines.

The trap cases are a t27 file, cli/t27b/tests/int_cast_traps.t27, whose
test names carry their verdicts (`_expect_trap` / `_expect_pass`, as
fuzz_oracle.t27 reads them); a 19-line test in cli/t27b/tests/tail.rs
runs it. The passing cases are #7417's conformance spec,
specs/tri/t27b/conformance/int_cast.t27. Supersedes draft #7417.

Closes #7412
Part of #6063

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@claude
ledger: gen_fuzz and server_main pass once @intcast lowers (Closes #7412
e9583a1 
)

The rows come from `tri t27b ratchet --bless` of the t27b lab's run of
d62ac63 (the deployed lab.py's lab_run, private work dir): the two
specs whose first blocker was ExprCall(@intcast) move from blocked to
pass, and the two new specs (int_cast_plan.t27 and
conformance/int_cast.t27) get their first rows, pass. Every other row is
master's. Counts: pass 854 -> 858, not_pass 61 -> 59, max_not_pass
61 -> 59; 1067 rows, no duplicate path.

Closes #7412
Part of #6063

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@gHashTag
gHashTag enabled auto-merge (squash) October 7, 2026 23:05
@gHashTag gHashTag mentioned this pull request Oct 7, 2026
Closed
@github-actions

github-actions Bot commented Oct 7, 2026

Copy link
Copy Markdown
Contributor

PR Dashboard

Generated at: 2026-10-07 23:06:07 UTC

Summary

Status Count
Total Open PRs 50
PRs with Failing Checks 47
PRs with All Checks Green 3
READY 1
FAILING 47
PENDING 0
NO CHECKS YET 0

These columns do not partition: 1 + 47 + 0 + 0 = 48, and there are 50 open PRs. A PR is being counted twice or not at all.

Seal Status

  • ⚠️ STALE -- sha256(compiler.rs)=39fa3908f9fd != 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).

@gHashTag
gHashTag merged commit 4dd14bf into master Oct 7, 2026
35 of 37 checks passed
This was referenced Oct 8, 2026
Merged
Closed
Merged
Open
Open
Open
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

None yet

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

t27b: @intCast with a result type (gen_fuzz, lane 4)

2 participants

@gHashTag @claude