Repository navigation
t27b: @log on f64, bit-exact with Zig 0.16 compiler_rt (ExprCall(@log)) #7217
Description
Sub-issue of #6063 (t27b coverage, lane 2: ExprCall(@log)).
specs/port/trinity/src/tri/gen_cross_entropy.t27 passes t27c test-report, but t27b refuses it as ExprCall(@log).
@log on f64 under Zig 0.16 is compiler_rt log (lib/compiler_rt/log.zig).
- It is the table-based routine in the ARM optimized-routines style.
- It reads the float's bits (
@bitCastto i64), then indexes two 128-entry tables (invc/logcandchi/clo) with 7 bits of the mantissa. - It has a separate polynomial path near 1.0.
logfis the musl polynomial.
For t27b to agree bit for bit, it needs two things:
- An f64 <-> u64 bitcast in t27b IR, or
@logas an IR intrinsic whose JIT and interpreter both call a port of that routine. - The two tables copied exactly from Zig 0.16 into t27b.
Measured, not assumed: found while doing #7175 / #7179 (PR #7216). Parked there, because it is larger than one lane slice.
Owner-approved foreign edit in cli/t27b (label owner-approved-foreign).
Boundary
specs/port/trinity/src/tri/gen_cross_entropy.t27
Activity
spec(t27b libm): call the log port's special cases directly (Closes #…
t27b ledger: @log moves gen_cross_entropy to pass, 823 -> 827 with @exp…
gHashTag commentedon Oct 7, 2026
Draft PR #7466 carries the @log port (libm.t27 13/13 in both, gen_cross_entropy 3/3, mismatch 0). It is over the foreign-line budget (#7371): check_budget() on origin/master...HEAD denies, cli/t27b/src/lower.rs +179 and cli/t27b/tests/source.rs +125 against 40 per file and 80 per PR (stacked on #7426, also parked). Auto-merge is off and the work is not split. The l6_nan mutant divergence is explained on the PR and filed as #7445. Refs #6063.
gHashTag commentedon Oct 7, 2026
Reopened 2026-10-07: the bulk close of issues without a
## Boundarysection was reverted by the owner. Nothing to do here; this comment replaces the earlier close notice.t27b ledger: gen_cross_entropy passes (Closes #7217)