https://gcc.gnu.org/bugzilla/show_bug.cgi?id=113618
Bug ID: 113618
Summary: [14 Regression] AArch64: memmove idiom regression
Product: gcc
Version: 14.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: wilco at gcc dot gnu.org
Target Milestone: ---
The following is often used as an idiom for memmove since GCC mid-end and most
back-ends have no support for inlining memmove:
void move64 (char *a, char *b)
{
char temp[64];
memcpy (temp, a, 64);
memcpy (b, temp, 64);
}
On trunk this generates:
ldp q30, q31, [x0]
sub sp, sp, #64
ldp q28, q29, [x0, 32]
stp q30, q31, [sp]
ldp q30, q31, [sp]
stp q28, q29, [sp, 32]
ldp q28, q29, [sp, 32]
stp q30, q31, [x1]
stp q28, q29, [x1, 32]
add sp, sp, 64
ret
This is a significant regression from GCC13 which has redundant stores but
avoids load-after-store forwarding penalties:
ldp q2, q3, [x0]
sub sp, sp, #64
ldp q0, q1, [x0, 32]
stp q2, q3, [sp]
stp q2, q3, [x1]
stp q0, q1, [sp, 32]
stp q0, q1, [x1, 32]
add sp, sp, 64
ret
LLVM avoids writing to the temporary and removes the stackframe altogether:
ldp q1, q0, [x0, #32]
ldp q2, q3, [x0]
stp q1, q0, [x1, #32]
stp q2, q3, [x1]
ret
The reason for the regression appears to be the changed RTL representation of
LDP/STP. The RTL optimizer does not understand LDP/STP, so emitting LDP/STP
early in memcpy expansion means it cannot remove the redundant stack stores.
A possible fix would be to avoid emitting LDP/STP in memcpy/memmove/memset
expansions.
