On 7/29/24 2:13 PM, Richard Sandiford wrote:
Sorry for the slow reply!
Slow?!? I think I sent that Thu/Fri last week. Nothing to be sorry
about at all in my mind!
No, I agree it looks like a bug. Seems like:
if (outermode == innermostmode
&& known_eq (byte, 0U)
&& known_eq (SUBREG_BYTE (op), 0))
return SUBREG_REG (op);
should be testing for subreg_lowpart_offset in both cases.
I'd largely concluded over the weekend that the simplify-rtx code was
just wrong and should be using subreg_lowpart.
Essentially I was trying to write a comment in ext-dce.cc around a check
to avoid this problem and I kept coming back to the phrase "avoid bug in
SUBREG simplification". And that little hack in ext-dce that I have
locally uses subreg_lowpart_p as one of its components.
jeff
