stillwater-sc / universal / 34923927380

86%

master: 84%

Build Type

push

github

Committed by web-flow

Commit Message

fix(areal): an exact zero sum is -0 only when both zeros are -0 (#1511)

+0 + -0 and +0 - +0 gave -0. IEEE-754 gives +0 for both in every
rounding mode but toward negative, and areal truncates toward zero.
operator+= returned rhs whenever *this was zero, so +0 + -0 copied the
-0; operator-= is *this += -rhs, so +0 - +0 reached the same case. Two
exact zeros now sum to -0 only when both are -0. x - x for a non-zero
x, and -0 - +0 = -0, were already right; so is the ubit interval
(0, minpos), which iszero() does not match.

saturation.cpp's exact-reference sweep of + - * no longer accepts
either sign on an exact zero, and its operands now include -0 as well
as +0: the double reference carries IEEE's zero signs, and areal's ==
compares encodings, so the check is sign-exact. Without this change
the sweep fails exactly these two cases in every configuration; with
it, everything passes at every level on gcc and clang. The areal suite
passes 26/26 and the areal consumers print identical output.

BEHAVIOUR-CHANGE: areal +0 + -0 and +0 - +0 now give +0 instead of -0, as IEEE-754 does

Closes #1507


Claude-Session: https://claude.ai/code/session_01NazYbxaii2XXVzQeQKMDTs

Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>

Coverage Stats

3 of 3 new or added lines in 1 file covered. (100.0%)

2 existing lines in 1 file now uncovered.

44081 of 51315 relevant lines covered (85.9%)

8238319.66 hits per line