------- Comment #2 from bonzini at gnu dot org 2006-09-07 08:11 -------
It's not really CSE's fault, though I agree that doing addressing mode
selection there is wrong. fwprop does not fix the bug for example, though
teaching it about this case could be easier than teaching CSE.
Unlike CSE, fwprop fixes this case:
char a[10], b[10];
int f1(log i)
{
return a[i+1] + b[i+1];
}
which generates
incq %rdi
movsbl a(%rdi),%eax
movsbl b(%rdi),%edx
addl %edx, %eax
with current GCC. So one way to fix this bug (assuming fwprop is merged in
4.3) is to teach something (expand?) that
(set (reg:DI 58 [ D.1872 ])
(sign_extend:DI (reg:SI 60 [ D.1872 ])))
(parallel [
(set (reg:DI 62 [ D.1872 ])
(plus:DI (reg/v:DI 58 [ i ])
(const_int 1 [0x1])))
(clobber (reg:CC 17 flags))
])
at least when we can disregard overflow, might be better than
(parallel [
(set (reg:SI 58 [ D.1872 ])
(plus:SI (reg/v:SI 60 [ i ])
(const_int 1 [0x1])))
(clobber (reg:CC 17 flags))
])
(set (reg:DI 62 [ D.1872 ])
(sign_extend:DI (reg:SI 58 [ D.1872 ])))
--
bonzini at gnu dot org changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |bonzini at gnu dot org
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=28940
