https://gcc.gnu.org/bugzilla/show_bug.cgi?id=62173
Bug ID: 62173
Summary: [AArch64] Performance regression due to r213488
Product: gcc
Version: 5.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: spop at gcc dot gnu.org
void bar(int i) {
char A[10];
int d = 0;
while (i > 0)
A[d++] = i--;
while (d > 0)
foo(A[d--]);
}
Compile this function at -O3 with the last good revision r213487 and with the
first bad revision r213488 "[AArch64] Improve TARGET_LEGITIMIZE_ADDRESS_P
hook", and diffing:
--- good.s 2014-08-18 10:56:42.542867000 -0500
+++ bad.s 2014-08-18 10:56:42.504090000 -0500
@@ -8,34 +8,33 @@
stp x29, x30, [sp, -48]!
cmp w0, wzr
add x29, sp, 0
- stp x19, x20, [sp, 16]
+ str x19, [sp, 16]
ble .L1
strb w0, [x29, 32]
subs w19, w0, #1
- add x20, x29, 32
beq .L4
strb w19, [x29, 33]
subs w1, w0, #2
beq .L4
- strb w1, [x20, 2]
+ strb w1, [x29, 34]
subs w1, w0, #3
beq .L4
- strb w1, [x20, 3]
+ strb w1, [x29, 35]
subs w1, w0, #4
beq .L4
- strb w1, [x20, 4]
+ strb w1, [x29, 36]
subs w1, w0, #5
beq .L4
- strb w1, [x20, 5]
+ strb w1, [x29, 37]
subs w1, w0, #6
beq .L4
- strb w1, [x20, 6]
+ strb w1, [x29, 38]
subs w1, w0, #7
beq .L4
- strb w1, [x20, 7]
+ strb w1, [x29, 39]
subs w1, w0, #8
beq .L4
- strb w1, [x20, 8]
+ strb w1, [x29, 40]
subs w1, w0, #9
beq .L4
strb w1, [x29, 41]
@@ -43,14 +42,16 @@
.L35:
sub w19, w19, #1
.L4:
- ldrb w0, [x20, w0, sxtw]
+ add x1, x29, 48
+ add x0, x1, x0, sxtw
+ ldrb w0, [x0, -16]
bl foo
mov w0, w19
cbnz w19, .L35
.L1:
- ldp x19, x20, [sp, 16]
+ ldr x19, [sp, 16]
ldp x29, x30, [sp], 48
ret
.size bar, .-bar
The problem is that gcc now generates an addressing mode that requires two more
add instructions in the second loop:
- ldrb w0, [x20, w0, sxtw]
+ add x1, x29, 48
+ add x0, x1, x0, sxtw
+ ldrb w0, [x0, -16]
