stillwater-sc / universal / 29619455885

85%

master: 84%

Build Type

push

github

Committed by web-flow

Commit Message

fix(elreal): drop trailing subnormal limbs in zbcl_from_blocks (#1136)

On a narrow host (bfloat16 k=7, fp16 k=11) a deep ZBCL expansion can bottom out
at the denormal floor and leave a trailing subnormal limb whose exponent gap to
its head is below k. zbcl_from_blocks() only stripped trailing zero blocks, so
the subnormal limb was carried into the ZBCL; forcing the tail that holds it
trips the 0-overlap debug assert in ZBCL::tail().

A subnormal block cannot satisfy the k-gap 0-overlap invariant -- block::
is_normalised() documents exactly this ("0-overlap accounting assumes the
leading bit is set"). The streaming producers (mul_scalar, online_divide,
online_multiply) already drop non-normalised blocks; centralise the same guard
in zbcl_from_blocks so every block-list producer (add / multiply / divide / sum)
is covered. Trailing removal is value-preserving: the list is priestRenorm'd
(descending exponent), so subnormal limbs sit strictly below every normalised
limb. is_normalised() is false for zero blocks too, so this subsumes the prior
zero-drop.

Latent, pre-existing fragility masked while bfloat16 conversion truncated;
exposed by the bfloat16 RNE-conversion fix (#1133), which changes the deep sqrt
expansion so the denormal-floor configuration is reached. el_math_sqrt aborted
under Debug/ASan with FpType=bfloat16.

Validated: all 42 elreal regression tests pass on gcc and clang; el_math_sqrt
passes with RNE bfloat16 on both compilers.

Resolves #1135

Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Coverage Stats

1 of 1 new or added line in 1 file covered. (100.0%)

14 existing lines in 2 files now uncovered.

49207 of 58157 relevant lines covered (84.61%)

6680545.96 hits per line