https://gcc.gnu.org/bugzilla/show_bug.cgi?id=126208
Bug ID: 126208
Summary: [17 Regression] Missed beneficial path splitting in
libquantum
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: ---
After removing -fsplit-paths in GCC 17 (PR120892)
we have a 3% regression on aarch64 in libquantum from SPEC2006:
https://gcc.gnu.org/pipermail/gcc-patches/2026-July/723522.html
The reduced tescase is:
void
update (long **nodes, int *size)
{
for (int i = 0; i < *size; ++i)
if ((*nodes)[i])
(*nodes)[i] = 0;
}
compiled with -O3 -fno-strict-aliasing .
The before (good) code is:
update(long**, int*):
ldr w6, [x1]
cmp w6, 0
ble .L1
ldr x4, [x0]
mov x2, 0
.L5:
ubfiz x3, x2, 3, 32
ldr x5, [x4, x3]
cbz x5, .L3
.L8:
str xzr, [x4, x3]
add x2, x2, 1
ldr w6, [x1]
cmp w6, w2
ble .L1
ldr x4, [x0]
ubfiz x3, x2, 3, 32
ldr x5, [x4, x3]
cbnz x5, .L8
.L3:
add x2, x2, 1
cmp w6, w2
bgt .L5
.L1:
ret
and the bad (after, GCC 17) code is:
update(long**, int*):
ldr w6, [x1]
cmp w6, 0
ble .L1
mov w2, 0
.L4:
ldr x4, [x0]
ubfiz x3, x2, 3, 32
add x2, x2, 1
ldr x5, [x4, x3]
cbz x5, .L3
str xzr, [x4, x3]
ldr w6, [x1]
.L3:
cmp w6, w2
bgt .L4
.L1:
ret
GCC 16 with -fsplit-paths managed to avoid a pointer-chasing load of x4 in
every iteration of the hot loop.
