https://gcc.gnu.org/bugzilla/show_bug.cgi?id=128768
Bug ID: 128768
Summary: Vectorised loop keeps one base register per data
reference when their common base is computed in an
outer loop
Product: gcc
Version: 16.2.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: freddie at witherden dot org
Target Milestone: ---
Consider the following C snippet:
#define S(c, i) s[r][i][j] += e*b[((k + c)*5 + i)*8 + j];
#define C(c) { double e = op[(r*8 + c)*8 + j]; \
S(c, 0) S(c, 1) S(c, 2) S(c, 3) S(c, 4) }
void tile(long n, long k, const double *restrict op,
const double *restrict b, double (*restrict s)[5][8])
{
for (long r = 0; r < n; r++)
#pragma omp simd
for (long j = 0; j < 8; j++)
{
C(0) C(1) C(2) C(3) C(4) C(5) C(6) C(7)
}
}
which compiled as gcc -O3 -fopenmp-simd -mcpu=apple-m1 results in inefficient
code being generated for the inner loop of the form:
L4: ldr x8, [sp, 304] ; base reloaded
ldr q31, [x2, x0]
ldr q3, [x6, x0]
ldr q4, [x8, x0]
ldr x8, [sp, 240] ; base reloaded
ldr q26, [x26, x0]
ldr q1, [x5, x0]
fmla v3.2d, v4.2d, v31.2d
ldr q28, [x8, x0]
ldr x8, [sp, 192] ; base reloaded
... ; 23 more reloads
add x0, x0, 16
cmp x0, 64
bne L4
where I count (hopefully correctly) 40 fmla, 53 ldr q, 5 str q, 26 ldr from sp.
The sp loads are the suspect ones: the inner loop of tile does 53 vector loads
per pass (8 of op, 40 of b, 5 of s) and their addresses have three bases, op +
512*r, b + 320*k and s + 320*r. Yet, one of the optimization passes seems to
expand this into 53 base pointers which leads to a bunch of them being spilled
to the stack.
This hurts performance quite a bit on my M4 (hacking the ASM to use 3 base
pointers gives me a ~1.5x speedup). No reason to think this is an ARM specific
issue, however, seems to lie upstream.