https://gcc.gnu.org/bugzilla/show_bug.cgi?id=69557
Bug ID: 69557
Summary: [ARM] revsh instruction not being conditionalised for
Thumb2
Product: gcc
Version: 6.0
Status: UNCONFIRMED
Keywords: missed-optimization
Severity: normal
Priority: P3
Component: rtl-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: ktkachov at gcc dot gnu.org
Target Milestone: ---
Target: arm*
While doing some of the testcase splitting work for PR 65578
I saw that we don't use conditional execution as much as we should for Thumb2.
The testcase:
extern short foos16 (short);
/* revshne */
short swaps16_cond (short x, int y)
{
short z = x;
if (y)
z = __builtin_bswap16 (x);
return foos16 (z);
}
for -march=armv6t2 -mthumb -O2 I get:
swaps16_cond:
@ args = 0, pretend = 0, frame = 0
@ frame_needed = 0, uses_anonymous_args = 0
@ link register save eliminated.
cbz r1, .L2
revsh r0, r0
.L2:
b
whereas for -march=armv6t2 -mthumb -O2 we get conditional execution:
swaps16_cond:
@ args = 0, pretend = 0, frame = 0
@ frame_needed = 0, uses_anonymous_args = 0
@ link register save eliminated.
cmp r1, #0
revshne r0, r0
.L2:
b
Note that the -marm version gets conditionalised at the very end in
arm_final_prescan_insn which we don't perform for Thumb2.
The ce3 pass doesn't create the COND_EXEC forms like I'd expect.
For the testcase above it refuses to perform any transformations in
cond_exec_find_if_block (from ifcvt.c) where it fails the check:
/* The edges of the THEN and ELSE blocks cannot have complex edges. */
FOR_EACH_EDGE (cur_edge, ei, then_bb->preds)
{
if (cur_edge->flags & EDGE_COMPLEX)
return FALSE;
}
Seems the edge is deemed complex so ifconversion bails out.
