https://gcc.gnu.org/bugzilla/show_bug.cgi?id=77729
Segher Boessenkool <segher at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |segher at gcc dot gnu.org
--- Comment #11 from Segher Boessenkool <segher at gcc dot gnu.org> ---
> Basically combine is going funny.
Hey, don't blame combine, it's all simplify-rtx's doing.
If in simplify_binary_operation_1 you disable the blocks after
/* Minimize the number of bits set in C1, i.e. C1 := C1 & ~C2. */
and after
/* If we have (ior (and (X C1) C2)), simplify this by making
C1 as small as possible if C1 actually changes. */
all works as you want (this is on rs6000, pretty much the same thing):
===
Trying 10, 11 -> 12:
Failed to match this instruction:
(set (reg:SI 131)
(ior:SI (zero_extend:SI (mem:QI (reg/v/f:DI 128 [ string ]) [0
*string_9(D)+0 S1 A8]))
(const_int 96 [0x60])))
Successfully matched this instruction:
(set (reg:SI 130)
(zero_extend:SI (mem:QI (reg/v/f:DI 128 [ string ]) [0 *string_9(D)+0 S1
A8])))
Successfully matched this instruction:
(set (reg:SI 131)
(ior:SI (reg:SI 130)
(const_int 96 [0x60])))
allowing combination of insns 10, 11 and 12
original costs 8 + 4 + 4 = 16
replacement costs 8 + 4 = 12
deferring deletion of insn with uid = 10.
modifying insn i2 11: r130:SI=zero_extend([r128:DI])
deferring rescan insn with uid = 11.
modifying insn i3 12: r131:SI=r130:SI|0x60
REG_DEAD r130:SI
deferring rescan insn with uid = 12.
===
giving
lbz 9,0(3)
ori 9,9,0x60
cmpwi 7,9,116
bne 7,.L5
etc.
