On 21/06/13 09:49, Kyrylo Tkachov wrote:
Hi all,
With r200197 the test pr46975.c now fails because the code:
/* { dg-options "-mthumb -Os" } */
int foo (int s)
{
return s == 1;
}
now generates:
0: f1a0 0001 sub.w r0, r0, #1
4: fab0 f080 clz r0, r0
8: 0940 lsrs r0, r0, #5
a: 4770 bx lr
instead of:
0: 1e43 subs r3, r0, #1
2: 4258 negs r0, r3
4: 4158 adcs r0, r3
6: 4770 bx lr
This hurts code size.
Therefore I've disabled the new peephole2 for optimize_insn_for_size_p so that
the original peephole before r200197 is used when optimising for size.
I've also added a test to confirm that the new peephole2 for the non-CC
setting variants is being used when optimising for speed, since the large code
should be more efficient because the scheduler won't have to deal with
condition-code setting.
Ok for trunk?
Thanks,
Kyrill
gcc/
2013-06-21 Kyrylo Tkachov <kyrylo.tkac...@arm.com>
* config/arm/arm.md (peepholes for eq (reg1) (reg2/imm)):
Use the flag setting variants when optimising for size.
gcc/testsuite/
2013-06-21 Kyrylo Tkachov <kyrylo.tkac...@arm.com>
* gcc.target/arm/pr46975-2.c: New test.
Not ok.
This re-introduces the bug that Meador was trying to fix.
You need to ensure that the condition code register is dead before
applying this peephole.
