stillwater-sc / universal / 23597700668

85%

master: 84%

Build Type

push

github

Committed by web-flow

Commit Message

fix(cfloat): add rounding and overflow handling to wide-significand convert path (#608)

* fix(cfloat): handle ALL_ONES_ES round-overflow explicitly in wide path

When rounding carries into the hidden-bit position and the biased
exponent is already ALL_ONES_ES, the wide-significand path left the
result as a no-op, relying on the isnan() remap below. This fails
for two cases:

1. hasMaxExpValues == true: all-ones exponent with zero fraction is a
   valid finite encoding, not NaN, so the remap never fires — producing
   an incorrect mid-range value instead of maxpos/maxneg or inf.

2. isSaturating == true && hasMaxExpValues == false: all-ones exponent
   with zero fraction is infinity, not NaN, so the saturating remap to
   maxpos/maxneg never fires — producing inf instead of maxpos/maxneg.

Fix: explicitly saturate to maxpos/maxneg or set inf and return early,
matching the early overflow handling already in convert().

* test(cfloat): add from_blocktriple regression tests for maxexpvalue conversion

Add missing from_blocktriple.cpp test files to both saturating/maxexpvalue
and nonsaturating/maxexpvalue conversion directories. These exercise the
convert() code path fixed in PR #608, where rounding overflow into the
max-exponent binade was not handled correctly when hasMaxExpValues=true.

- saturating: verifies overflow saturates to maxpos/maxneg (ftt config)
- nonsaturating: verifies overflow maps to +/-inf (ftf config)

Covers es=2 through es=8 across small-to-medium cfloat sizes. The es=8
cases are included unconditionally since the hasMaxExpValues path is
what the fix specifically addressed.

* fix(test): use += to accumulate nrOfFailedTestCases in maxexpvalue from_blocktriple tests

Coverage Stats

0 of 3 new or added lines in 1 file covered. (0.0%)

9 existing lines in 2 files now uncovered.

44346 of 52563 relevant lines covered (84.37%)

6132421.69 hits per line