On Tue, Jan 09, 2024 at 09:42:30PM +0100, Theo Buehler wrote: > With the below diff regress on aarch64 passes. It only changes things > there and maybe on arm, so there is no new round of testing needed. > > 119 failed, 33819 passed, 220 skipped, 1308 deselected, 31 xfailed, 4 > xpassed, 461 warnings in 737.96s (0:12:17) > > This is a bit worse than 1.24 which had > > 32 failed, 25330 passed, 225 skipped, 1306 deselected, 33 xfailed, 4 xpassed, > 100 warnings in 365.47s (0:06:05) > > The breakage came from this PR switching numpy 1.25 to use SIMD > for isnan, isfinite, etc: https://github.com/numpy/numpy/pull/22165 > > It would result in out-of-bounds heap reads which would sometimes crash > after hitting a guard page. Fortunately, the linalg test would do this > quite reliably. I managed to narrow the issue down to a specific > instruction and then it was easy to find very recent upstream fixes. > > It causes at least one new failure in the linalg test, but I'm not going > to debug this.
I forgot to attach the full test.log for posterity and to say ok tb
test.log.xz
Description: Binary data
