On Wed, 9 Nov 2022, Martin Storsjö wrote:
On Wed, 9 Nov 2022, Oleg Oshmyan via Mingw-w64-public wrote:
When int x == INT_MIN, the behaviour of `-x` is undefined.
This causes a crash when isnan is inlined into code instrumented
for Clang's or GCC's -ftrapv, -fsanitize=signed-integer-overflow,
-fsanitize=undefined or similar.
Instead of mucking about with `int` and `unsigned`,
just use `unsigned` throughout the body of isnan.
For unsigned types, all the same arithmetic can be used,
and it is always well-defined and uses two's complement
for negation, which is exactly the same as these platforms'
signed arithmetic that this code already assumes.
Note that there are some inconsistencies in the code of the various
isnan variants, such as whether integer literals are suffixed with `u`
when promoted to `unsigned int` anyway and whether the "true" return
value is 1 or -1. This patch leaves these inconsistencies intact.
Signed-off-by: Oleg Oshmyan <[email protected]>
---
mingw-w64-headers/crt/math.h | 18 +++++++++---------
1 file changed, 9 insertions(+), 9 deletions(-)
Thanks, this looks reasonable to me. I'll include it in some rounds of
testing, and I can push it tomorrow or so, if it doesn't turn up any issues.
No issues encountered so far at least, so I pushed the patch. Thanks for
your contribution!
// Martin
_______________________________________________
Mingw-w64-public mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/mingw-w64-public
