https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116541
Bug ID: 116541
Summary: [14/15 Regression] Inefficient missing use of reg+reg
addressing modes
Product: gcc
Version: 15.0
Status: UNCONFIRMED
Keywords: missed-optimization
Severity: normal
Priority: P3
Component: rtl-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: ktkachov at gcc dot gnu.org
CC: tnfchris at gcc dot gnu.org
Target Milestone: ---
Target: aarch64
The loop from TSVC:
__attribute__((aligned(64))) float
a[32000],b[32000],c[32000],d[32000],e[32000],
aa[256][256],bb[256][256],cc[256][256],tt[256][256];
int dummy(float[32000], float[32000], float[32000], float[32000], float[32000],
float[256][256], float[256][256], float[256][256], float);
#define real_t float
#define iterations 100000
#define LEN_1D 32000
real_t s242(real_t s1, real_t s2)
{
for (int nl = 0; nl < iterations/5; nl++) {
//#pragma unroll(1)
for (int i = 1; i < LEN_1D; ++i) {
a[i] = a[i - 1] + s1 + s2 + b[i] + c[i] + d[i];
}
}
return a[0];
}
Compiled with -O3 -mcpu=neoverse-v2 generates the following loop:
.L3:
fadd s31, s0, s31
add x4, x8, x0
add x3, x7, x0
add x2, x6, x0
add x1, x5, x0
ldr s3, [x4, 4]
add x0, x0, 4
ldr s2, [x3, 4]
fadd s31, s31, s1
ldr s30, [x2, 4]
fadd s31, s31, s3
fadd s31, s31, s2
fadd s31, s31, s30
str s31, [x1, 4]
cmp x0, x9
bne .L3
With GCC 13 it used better addressing modes:
.L3:
fadd s0, s5, s0
ldr s3, [x4, x0]
ldr s2, [x3, x0]
ldr s1, [x2, x0]
fadd s0, s0, s4
fadd s0, s0, s3
fadd s0, s0, s2
fadd s0, s0, s1
str s0, [x1, x0]
add x0, x0, 4
cmp x0, x5
bne .L3
I don't know if Tamar's pending IVOPTs fix this but filing it here just in case