bemanproject / big_int / 32258094846

92%

push

github

Fix the non-canonical magnitude left by decrement across a limb boundary (#332)

unchecked_decrement_magnitude trimmed nothing when the borrow stopped inside the
magnitude, so a decrement that zeroed the top limb left limb_count() too high
and broke the trimmed-top-limb invariant. The increment sibling already handled
its mirror case by extending the count on carry-out, so growth was covered and
shrinking was not. ++ on a negative and both operator~ overloads reach the same
path.

Comparison still reported the right value, so the pre-existing
PrefixDecrementAllocatedBorrowChain passed, but width_mag() read 0 and
everything downstream of it was wrong: the free operator>> discarded the value
because it exits early when the shift is at or beyond width_mag(), to_string and
to_chars printed 0, size() and representation_size() read 0, std::hash disagreed
for equal values, and shrink_to_fit() could not return the value to the in-place
buffer.

The tests assert the representation rather than the value, since asserting the
value is what let this through. Prefix and postfix decrement, increment on
negatives, both operator~ overloads and the in-place path via
basic_big_int<256> are covered at 2^64, 2^128 and 2^192, with the magnitude
observers at 2^128 and shrink_to_fit at 2^64. is_normalized moved to testing.hpp
because hash.test.cpp needs it too, and the consteval cases catch the bug at
compile time.

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

5985 of 6202 relevant lines covered (96.5%)

3161089.41 hits per line