https://gcc.gnu.org/bugzilla/show_bug.cgi?id=117607
Bug ID: 117607
Summary: unnecessary scev optimization for popcnt
Product: gcc
Version: 15.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: fxue at os dot amperecomputing.com
Target Milestone: ---
Look at the case:
int *foo(long v, int *p)
{
while (v) {
long t = v;
v &= (v - 1);
*p++ = __builtin_ctzl(t);
}
return p;
}
On arch with popcnt supported, such as aarch64, with "-O2 -ftree-scev-cprop",
it generates:
foo:
.LFB0:
.cfi_startproc
cbz x0, .L4
mov x4, x1
mov x2, x0
.p2align 5,,15
.L3:
rbit x3, x2
clz x3, x3
str w3, [x4], 4
sub x3, x2, #1
ands x2, x2, x3
bne .L3
fmov d31, x0
add x1, x1, 4
cnt v31.8b, v31.8b
addv b31, v31.8b
fmov w0, s31
sub w2, w0, #1
add x0, x1, w2, uxtw 2
ret
.p2align 2,,3
.L4:
mov x0, x1
ret
For ""-O2 -fno-tree-scev-cprop", it gets simpler codegen:
.LFB0:
.cfi_startproc
mov x2, x0
mov x0, x1
cbz x2, .L2
.p2align 5,,15
.L3:
rbit x1, x2
sub x3, x2, #1
clz x1, x1
str w1, [x0], 4
ands x2, x2, x3
bne .L3
.L2:
ret
The cause is that scev would compute exit value of "p" using POPCNT in
one-shot. However, since "p" value is used and has to be evaluated at every
iteration, so the computation at exit is unneeded.
final value replacement:
p_15 = PHI <p_11(3)>
with expr: (int *) (((sizetype) (unsigned int) (.POPCOUNT ((unsigned long)
v_7(D)) + -1) * 4 + (sizetype) p_8(D)) + 4)
final stmt:
p_15 = (int *) _25
