https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116022
Bug ID: 116022
Summary: complete (early) unrolling foils vectorizer for vector
initialization
Product: gcc
Version: 15.0
Status: UNCONFIRMED
Keywords: missed-optimization
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: amylaar at gcc dot gnu.org
Target Milestone: ---
#define LENGTH 4
typedef unsigned uint32v_t __attribute ((vector_size (LENGTH * 4)));
uint32v_t vdup_u32(uint32v_t a, unsigned b)
{
uint32v_t r;
int i;
for (i = 0; i < LENGTH; i++)
r[i] = b;
return r;
}
For x86_64-pc-linux-gnu, with -O1 -ftree-vectorize, we get:
vdup_u32:
.LFB0:
.cfi_startproc
movd %edi, %xmm1
pshufd $0, %xmm1, %xmm0
ret
which is fine.
However, with -O3, the complete unroller is run before the vectorizer, and
instead we get:
vdup_u32:
.LFB0:
.cfi_startproc
movd %edi, %xmm0
movd %edi, %xmm1
pshufd $225, %xmm0, %xmm0
movss %xmm1, %xmm0
pshufd $225, %xmm0, %xmm0
pshufd $198, %xmm0, %xmm0
movss %xmm1, %xmm0
pshufd $198, %xmm0, %xmm0
pshufd $39, %xmm0, %xmm0
movss %xmm1, %xmm0
pshufd $39, %xmm0, %xmm0
ret
making the code both larger and slower.
According to https://gcc.gnu.org/projects/tree-ssa/vectorization.htm , this was
supposed
to be handled by SLP, but apparently that is not happening.
See dump files produced by -fdump-tree-rebuild_frequencies -fdump-tree-cunrolli
-fdump-tree-vect
for details.
