https://gcc.gnu.org/bugzilla/show_bug.cgi?id=126332
Bug ID: 126332
Summary: -ftracer hurts phiopt load commoning
Product: gcc
Version: 16.0
Status: UNCONFIRMED
Keywords: missed-optimization
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: ktkachov at gcc dot gnu.org
Target Milestone: ---
The hot loop from Snappy decompression gets worse codegen when compiled with
PGO because -ftracer messes up the load commoning that was introduced in
phiopt:
#include <stddef.h>
#include <stdint.h>
__attribute__ ((noinline, noclone))
const uint8_t *
advance (const uint8_t *ip, size_t tag, const uint8_t *end)
{
size_t sum = 0;
while (ip < end)
{
size_t type = tag & 3;
if (type == 0)
{
size_t n = (tag >> 2) + 1;
tag = ip[n];
ip += n + 1;
}
else
{
tag = ip[type];
ip += type + 1;
}
sum += tag;
}
return ip - (sum & 1);
}
Compiled for aarch64 with -O3 gives:
advance(unsigned char const*, unsigned long, unsigned char const*):
cmp x0, x2
bcs .L1
mov w5, 0
.L5:
ands x3, x1, 3
lsr x1, x1, 2
add w4, w3, 1
csinc x3, x3, x1, ne
add x1, x1, 2
csel x4, x4, x1, ne
ldrb w1, [x0, x3]
add x0, x0, x4
add x5, x5, x1
cmp x0, x2
bcc .L5
and x5, x5, 1
sub x0, x0, x5
.L1:
ret
with -O3 -ftracer it gives:
advance(unsigned char const*, unsigned long, unsigned char const*):
mov w4, 0
cmp x0, x2
bcs .L1
.L2:
lsr x3, x1, 2
ands x1, x1, 3
add x6, x0, x3
add w5, w1, 1
add x3, x3, 2
bne .L4
ldrb w1, [x6, 1]
add x0, x0, x3
add x4, x4, x1
cmp x0, x2
bcc .L2
and x4, x4, 1
sub x0, x0, x4
.L1:
ret
.L4:
ldrb w1, [x0, x1]
add x0, x0, x5
add x4, x4, x1
cmp x2, x0
bhi .L2
and x4, x4, 1
sub x0, x0, x4
b .L1
