libm: remainder: make sure x is zero
ClosedPublic
Actions

Authored by vexeduxr on Oct 10 2025, 12:57 PM.

Details

Reviewers

imp
kib
emaste
dim
siva

Commits

rG4c87cff1c8c5: libm: remainder: make sure x is zero
rG583e976d6d09: libm: remainder: make sure x is zero
rG439909c48b0e: libm: remainder: make sure x is zero
rGd06e31d4a762: libm: remainder: make sure x is zero
rG25cca51ed294: libm: remainder: make sure x is zero

Summary

Make sure the entirety of x is zero before flipping the sign bit.
Otherwise the sign would be wrong for small values of x when it is
negative and |n*y| > |x|

PR: 251091
MFC after: 3 days

Test Plan

This fixes the "rem3" test on powerpc64. It should also fix it on the rest of the architectures, but I haven't tested them.

Diff Detail

Repository

rG FreeBSD src repository

Lint

Lint Not Applicable

Unit

Tests Not Applicable

Event Timeline

vexeduxr requested review of this revision.Oct 10 2025, 12:57 PM

vexeduxr created this revision.

Harbormaster completed remote builds in B67697: Diff 163926.Oct 10 2025, 12:57 PM

vexeduxr edited the test plan for this revision. (Show Details)Oct 10 2025, 1:11 PM

vexeduxr edited the summary of this revision. (Show Details)Oct 10 2025, 2:40 PM

Ask kargl@ for explicit feedback and approval of the patch.

Will do. I tried to add him to the review before uploading the patch but couldn't find an account.

Will do. I tried to add him to the review before uploading the patch but couldn't find an account.

I don't think he has one; I'd suggest mailing him with the patch attached, but you can include a link to this review as well.

On Sat Oct 11, 2025 at 2:26 AM +0300, Steve Kargl wrote:

I do not have a phabricator account. That method to
review patches came after I stopped committing.

I did look at your patch, and do believe that it
is the correct fix. If you compare to remainderf(),
you'll see that the float version considers the
entire significand. I also checked remainderl().
It calls remquol() and it's implementation is
much different than the float and double versions.