The vectorizer is too restricted in the way it decides by how many iterations
to peel a loop in order to align a certain memory reference in a loop. It
considers only the first (potentially) misaligned store it encounters in the
loop. For this reason the testcases vect-multitypes-1.c, vect-multitypes-4.c
and vect-iv-4.c don't get vectorized. For example (using Vector Size of 16
bytes), in vect-multitypes-1.c we have:
short sa[N], sb[N];
int ia[N], ib[N];
for (i = 0; i < n; i++) {
ia[i+3] = ib[i];
sa[i+3] = sb[i];
}
The current peeling-for-alignment scheme will consider the 'ia[i+3]' access for
peeling, and therefore will examine the option of using a peeling factor =
(4-3)%4 = 1. This will not align the access 'sa[i+3]', for which we need to
peel 5 iterations. As a result the loop doesn't get vectorized (cause we
currently can't handle misaligned stores unless we align them by peeling).
However, if we had considered the 'sa[i+3]' access as well for peeling, we
would have examined the option of using a peeling factor = (8-3)%8 = 5, which
would align both accesses, and would allow us to vectorize the loop. So the
vectorizer needs to be extended to consider more peeling factors, and not just
one.
--
Summary: missed vectorization due to too strict peeling-for-
alignment policy
Product: gcc
Version: 4.3.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: tree-optimization
AssignedTo: unassigned at gcc dot gnu dot org
ReportedBy: dorit at il dot ibm dot com
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=31946
