stillwater-sc / universal / 26584402568

84%

master: 84%

Build Type

push

github

Committed by web-flow

Commit Message

fix(elreal): quad-precision-capable two_prod_host (fused-fma odd-p path) (#1025)

block_two_mult was not exact for quad hosts: two_prod_host computed the residual
of an odd-precision product in a `double` intermediate (#942), which silently
dropped the residual of a 113-bit product (cfloat<128,15>). Fixed by selecting
the residual path by width: even p -> Veltkamp/Dekker (exact at any width); odd p
with 2p <= 53 -> double intermediate (half, cfloat<24,5>, unchanged); odd p with
2p > 53 -> the host's correctly-rounded fused fma, r = fma(a,b,-p) = a*b - p
exactly. Verified exact at 113 bits: 0/20000 inexact vs the exact dyadic oracle.

Also makes the #1022 oracle genuinely quad-capable, which was the precondition
for validating the above. The merged exact_real extracted wide values via cfloat
frexp/floor, both of which are unreliable at wide precision (frexp normalises to
[1,2) not std's [0.5,1); floor mis-handles large integers -- filed separately).
#1023's quad sweeps passed only because they fed double-derived (<=53-bit) q128
values; genuine 113-bit values were mis-extracted. exact_real now reads the
cfloat encoding directly (sign + scale + fraction bits ->
(-1)^sign * (2^fbits + F) * 2^(scale-fbits)), no frexp/floor. Added a full-width
quad sweep (113-bit operands) and re-enabled quad two_mult in the oracle.

Full elreal suite passes, gcc + clang.

Resolves #1024
Relates to #1022, #923

Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Coverage Stats

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

4 existing lines in 2 files now uncovered.

46896 of 55829 relevant lines covered (84.0%)

5789662.09 hits per line