https://gcc.gnu.org/bugzilla/show_bug.cgi?id=88837
Bug ID: 88837
Summary: [SVE] Poor vector construction code in VL-specific
mode
Product: gcc
Version: unknown
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: rsandifo at gcc dot gnu.org
Target Milestone: ---
The reduction testcases in gcc.target/aarch64/sve/slp_5.c require an initial
vector in which all elements except the first two are zero. For the default
VL-agnostic mode we generate reasonable code, e.g.:
vec_slp_int32_t:
.LFB4:
.cfi_startproc
ldp s2, s1, [x1]
cmp w2, 0
ble .L19
mov x3, 0
sbfiz x2, x2, 1, 32
mov z0.b, #0
whilelo p0.s, xzr, x2
insr z0.s, s1
ptrue p1.s, all
insr z0.s, s2
.p2align 3,,7
.L20:
ld1w z1.s, p0/z, [x0, x3, lsl 2]
incw x3
add z0.s, p0/m, z0.s, z1.s
whilelo p0.s, x3, x2
bne .L20
But with -msve-vector-bits=256 the code is much worse:
vec_slp_int32_t:
.LFB4:
.cfi_startproc
ldp w5, w4, [x1]
cmp w2, 0
ble .L31
sub sp, sp, #32
.cfi_def_cfa_offset 32
mov z0.b, #0
str z0, [sp]
mov x3, 0
sbfiz x2, x2, 1, 32
whilelo p0.s, xzr, x2
ldr x6, [sp]
bfi x6, x5, 0, 32
mov x5, x6
bfi x5, x4, 32, 32
str x5, [sp]
ldr z0, [sp]
.p2align 3,,7
.L28:
ld1w z1.s, p0/z, [x0, x3, lsl 2]
add x3, x3, 8
add z0.s, p0/m, z0.s, z1.s
whilelo p0.s, x3, x2
bne .L28
We should try to optimise this, probably by implementing vec_init_optab for
fixed-length vectors.
