https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109153
Bug ID: 109153
Summary: missed vector constructor optimizations
Product: gcc
Version: 13.0
Status: UNCONFIRMED
Keywords: missed-optimization
Severity: normal
Priority: P3
Component: middle-end
Assignee: unassigned at gcc dot gnu.org
Reporter: tnfchris at gcc dot gnu.org
Blocks: 47562
Target Milestone: ---
Target: aarch64*
The following example:
#include <arm_neon.h>
#include <string.h>
uint32x2_t foo(const uint8_t *buf, int stride) {
uint32_t a0, a1;
memcpy(&a0, buf, 4);
memcpy(&a1, buf + stride, 4);
uint32x2_t a_u32 = vdup_n_u32(a0);
return vset_lane_u32(a1, a_u32, 1);
}
generates
foo:
add x1, x0, w1, sxtw
ld1r {v0.2s}, [x0]
ld1 {v0.s}[1], [x1]
ret
where the initial value is replicated to then be overwritten.
At the gimple level we have:
_9 = {_4, _4};
__vec_10 = BIT_INSERT_EXPR <_9, _7, 32 (32 bits)>;
which should have been optimized to:
_9 = {_4, _7 }
but this cannot be blindly done as the resulting vector needs to be
cheaper.
For instance if it was
_9 = {_4, _4, _4, _4};
__vec_10 = BIT_INSERT_EXPR <_9, _7, 32 (32 bits)>;
this wouldn't have been cheaper.
Similarly another testcase gives
<bb 2> [local count: 1073741824]:
_4 = VEC_PERM_EXPR <a_2(D), b_3(D), { 0, 8, 1, 9, 2, 10, 3, 11 }>;
_5 = VEC_PERM_EXPR <a_2(D), b_3(D), { 4, 12, 5, 13, 6, 14, 7, 15 }>;
_6 = {_4, _5};
return _6;
which should have been just a singe VEC_PERM_EXPR.
Referenced Bugs:
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=47562
[Bug 47562] [meta-bug] keep track of Neon Intrinsics enhancements
