https://gcc.gnu.org/bugzilla/show_bug.cgi?id=65363
Bug ID: 65363
Summary: trivial redundant code reordering makes code less
optimal
Product: gcc
Version: 5.0
Status: UNCONFIRMED
Severity: minor
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: vries at gcc dot gnu.org
Consider this test-case test.c (based on PR65270 comment 27/28):
...
#define N 100000
struct a
{
int a[N];
};
typedef struct a misaligned_t __attribute__ ((aligned (8)));
typedef struct a aligned_t __attribute__ ((aligned (32)));
static void
__attribute__ ((noinline))
__attribute__ ((noclone))
__attribute__ ((used))
t (void *a, aligned_t *d)
{
int v, v2;
int i;
for (i=0; i < N; i++)
{
#if REORDER
v2 = ((misaligned_t *)a)->a[i];
v = ((aligned_t *)a)->a[i];
#else
v = ((aligned_t *)a)->a[i];
v2 = ((misaligned_t *)a)->a[i];
#endif
d->a[i] += v + v2;
}
}
aligned_t aa;
aligned_t d;
int
main (void)
{
t (&aa, &d);
return 0;
}
...
Changing the order of loads in the loop body results in different instructions
(and I assume the unaligned one (movdqu) is more expensive than the aligned one
(movdqa)):
...
$ n=0; gcc -O2 -ftree-vectorize test.c -DREORDER=$n -S -o $n
$ n=1; gcc -O2 -ftree-vectorize test.c -DREORDER=$n -S -o $n
$ diff -u 0 1
--- 0 2015-03-09 15:46:41.395919753 +0100
+++ 1 2015-03-09 15:46:43.747919840 +0100
@@ -19,7 +19,7 @@
.p2align 4,,10
.p2align 3
.L4:
- movdqa (%rdi,%rax), %xmm0
+ movdqu (%rdi,%rax), %xmm0
paddd %xmm0, %xmm0
paddd (%rsi,%rax), %xmm0
movaps %xmm0, (%rsi,%rax)
...
The two loads are redundant, and fre is the pass that picks the first one and
eliminates the second one. I'm not sure though whether you want to fix this
particular example in fre. Perhaps you want to propagate alignment before doing
fre.
OTOH, fre does not take the cost of the value producing statements into account
when determining which to choose as representative and which to eliminate.
