https://gcc.gnu.org/bugzilla/show_bug.cgi?id=118662
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Priority|P3 |P2
Known to fail| |14.1.0
Target|X86_64 |x86_64-*-*
Known to work| |13.3.0
--- Comment #12 from Richard Biener <rguenth at gcc dot gnu.org> ---
We combine
Trying 7, 9 -> 10:
7: r106:V4QI=[r119:DI]
REG_DEAD r119:DI
9: r108:V4HI=sign_extend(vec_select(r106:V4QI#0,parallel))
10: r109:V4SI=sign_extend(vec_select(r108:V4HI#0,parallel))
REG_DEAD r108:V4HI
to
modifying insn i2 9: r109:V4SI=sign_extend([r119:DI])
modifying insn i3 10: r106:V4QI=r109:V4SI#0
where i3 is problematic. r106 is no longer computed correctly (while
a QImode subreg of a SImode reg that's sign-extended from a QImode val
is a no-op round-trip the same is of course not true for vector modes).
