https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116312
Bug ID: 116312
Summary: Use LDP instead of LD2 on for Advanced SIMD when
possible
Product: gcc
Version: 15.0
Status: UNCONFIRMED
Keywords: missed-optimization
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: ktkachov at gcc dot gnu.org
Target Milestone: ---
Target: aarch64
The testcase:
#define N 32000
float a[N],b[N],c[N];
void foo(struct args_t * func_args)
{
for (int i = 0; i < N; i+=2) {
c[i] = a[i] + b[i];
}
}
Compiled, for example with -O3 -mcpu=neoverse-v2 will generate the loop:
.L2:
ld2 {v29.4s - v30.4s}, [x4], 32
mov x3, x0
add x2, x0, 16
add x1, x0, 24
add x0, x0, 32
ld2 {v30.4s - v31.4s}, [x5], 32
fadd v30.4s, v30.4s, v29.4s
str s30, [x3], 8
st1 {v30.s}[1], [x3]
st1 {v30.s}[2], [x2]
st1 {v30.s}[3], [x1]
cmp x4, x6
bne .L2
LLVM generates:
.LBB0_1:
ldp q0, q1, [x9], #32
ldp q2, q3, [x11], #32
sub x12, x8, #8
add x13, x8, #8
subs x10, x10, #4
fadd v1.4s, v1.4s, v3.4s
fadd v0.4s, v0.4s, v2.4s
st1 { v1.s }[2], [x13]
stur s0, [x8, #-16]
str s1, [x8], #32
st1 { v0.s }[2], [x12]
b.ne .LBB0_1
LDP is known to be more efficient than LD2. For example from the Neoverse V2
SWOG the throughput/latency of thhe LD2 here is 8 and 3/2.
For the equivalent LDP it's 6 and 3/2.
It's not a huge improvement to be honest, but we could implement it as a simple
final assembly output template change for minimal invasion.
Though if we wanted to take advantage of wider post-index immediates available
to LDP we could make it more elaborate
