------- Comment #5 from bonzini at gnu dot org 2010-06-30 11:51 -------
I agree that the problem is a wrong pattern.
Here you have non-matching mode between the udiv/umod and the first argument:
> (ior:HI (ashift:HI (zero_extend:HI (umod:QI (reg:HI 68)
> (reg:QI 61 [ D.2750 ])))
> (const_int 8 [0x8]))
> (zero_extend:HI (udiv:QI (reg:HI 68)
> (reg:QI 61 [ D.2750 ]))))
Instead you need to have this before reload:
(set (match_operand:HI "register_operand" "=a")
(ior:HI (ashift:HI
(zero_extend:HI (umod:QI (match_operand:QI "register_operand" "0")
(match_operand:QI "register_operand" "q")))
(const_int 8 [0x8]))
(zero_extend:HI (udiv:QI (match_dup 1) (match_dup 2))))
and after reload split it into a zero/sign extension of al into ax followed by
(set (match_dup 0)
(ior:HI (ashift:HI (umod:HI (match_dup 0)
(zero_extend:HI (match_dup 2)))
(const_int 8 [0x8]))
(udiv:HI (match_dup 0)
(zero_extend:HI (match_dup 2))))
(or maybe sign extend into eax? QImode->SImode is definitely cheaper for zero
extension, I don't know about the cost of cbw because of partial register
stalls).
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=44695
