https://gcc.gnu.org/bugzilla/show_bug.cgi?id=121248
Bug ID: 121248
Summary: -ftree-slp-vectorize with ARM/NEON results in
different runtime outcome
Product: gcc
Version: 14.3.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: jpalus+gcc at fastmail dot com
Target Milestone: ---
Created attachment 61971
--> https://gcc.gnu.org/bugzilla/attachment.cgi?id=61971&action=edit
reproducer
FreeType's FT_Matrix_Check [0] gives different results than expected when
compiled for ARMv7 with NEON extensions.
Attached reproducer is supposed to return "1" but if compiled with
"-march=armv7-a -mfpu=neon-vfpv3 -O2" returns "0". Further narrowed flags down
to "-march=armv7-a -mfpu=neon-vfpv3 -O1 -ftree-slp-vectorize". Taking away
either one of "neon-", "-O1" or "-ftree-slp-vectorize" no longer reproduces the
issue.
Works fine when compiled for aarch64 though.
[0]
https://gitlab.freedesktop.org/freetype/freetype/-/blob/VER-2-13-3/src/base/ftcalc.c?ref_type=tags#L727
