http://gcc.gnu.org/bugzilla/show_bug.cgi?id=55858
Bug #: 55858
Summary: When scalarizing contiguous whole-arrays, consider
folding into a single loop
Classification: Unclassified
Product: gcc
Version: 4.8.0
Status: UNCONFIRMED
Keywords: missed-optimization
Severity: normal
Priority: P3
Component: fortran
AssignedTo: unassig...@gcc.gnu.org
ReportedBy: bur...@gcc.gnu.org
Motivated by http://gcc.gnu.org/ml/fortran/2013-01/msg00015.html
Currently, the scalarizer generates (rank) loops:
subroutine g(x)
integer, pointer, intent(in), contiguous :: x(:,:)
x = 8
end subroutine g
gives
S.0 = Dx->dim[1].lbound
while (1)
{
if (S.0 > x->dim[1].ubound) goto L.2;
{
integer(kind=8) D.1903;
integer(kind=8) S.1;
D.1903 = x->dim[1].stride * S.0 + x->offset;
S.1 = x->dim[0].lbound;
while (1)
{
if (S.1 > x->dim[0].ubound) goto L.1;
(*D.1895)[S.1 + D.1903] = 8;
S.1 = S.1 + 1;
}
L.1:;
}
S.0 = S.0 + 1;
}
If one knows that the memory is contiguous (i.e. it is simply contiguous), one
can use a single loop.
Or at least something which allows the ME to fold the two loops into a single
loop, e.g. "stride[1]" could be replaced by the extent
("ubound[0]-lbound[0]+1"), which allows the ME to deduce that one can fold it
into a single loop. Using "stride", the contiguity information is lost.
