https://gcc.gnu.org/bugzilla/show_bug.cgi?id=128834
Jakub Jelinek <jakub at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |sayle at gcc dot gnu.org,
| |uros at gcc dot gnu.org
--- Comment #2 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
The difference is between the two revisions in using
(insn 22 20 23 4 (set (reg:CCZ 17 flags)
(compare:CCZ (reg:DI 103 [ _18+16 ])
(const_int 0 [0]))) "pr128834.c":7:16 8 {*cmpdi_ccno_1}
(nil))
before and
(insn 21 20 22 4 (set (reg:CCZ 17 flags)
(compare:CCZ (reg:TI 108 [ _17+16 ])
(const_int 0 [0]))) "pr128834.c":7:16 38 {*cmpti_doubleword}
(expr_list:REG_DEAD (reg:TI 108 [ _17+16 ])
(nil)))
after.
With the one before, noce_try_store_flag during ce1 is successful, attempts to
match
(set (reg:DI 103 [ _18+16 ])
(ltu:DI (reg:CCC 17 flags)
(const_int 0 [0])))
insn and that matches.
But after that change it attempts to match
(set (reg:TI 108 [ _17+16 ])
(ltu:TI (reg:CCC 17 flags)
(const_int 0 [0])))
instead and that doesn't match.
So, wonder if we don't need setcc_ti_1 patterns (and similarly for 32-bit mode
*setcc_di_2) which at some point is lowered to setcc_[ds]i_1 on the lower half
and clearing of the upper half.