https://gcc.gnu.org/bugzilla/show_bug.cgi?id=128836
--- Comment #7 from GCC Commits <cvs-commit at gcc dot gnu.org> --- The master branch has been updated by Uros Bizjak <[email protected]>: https://gcc.gnu.org/g:44f26668c6b35ef7f6d71fdc4c89e667bec9a80f commit r17-6432-g44f26668c6b35ef7f6d71fdc4c89e667bec9a80f Author: Uros Bizjak <[email protected]> Date: Wed Nov 18 20:19:35 2026 +0100 i386: Require reload_completed for the AND -65536 split [PR target/128836] The define_split that converts "and $-65536, reg" to a strict_low_part HImode store of zero can fire before reload. When operands[0] is a (subreg:SI (reg:SF N) 0), gen_lowpart (HImode, operands[0]) cannot be folded into a valid subreg of the SFmode pseudo, so it silently copies the value to a fresh SImode pseudo and returns the lowpart of that. The strict_low_part store then updates a pseudo nobody reads and the original register keeps its old value, which is a miscompile. Fix this by requiring reload_completed, where operands[0] is a hard register and gen_lowpart yields a plain REG that aliases the original. PR target/128836 gcc/ChangeLog: * config/i386/i386.md (split and with -65536 to strict_low_part movhi): Require reload_completed. gcc/testsuite/ChangeLog: * gcc.target/i386/pr128836.c: New test.
