------- Comment #1 from mikael dot morin at tele2 dot fr 2008-10-29 20:59
-------
Reduced case:
subroutine subr (m, n, a, b, c, d, p)
implicit none
integer m, n
real a(m,n), b(m,n), c(n,n), d(m,n)
integer p(n)
d = a(:,p) - matmul(b, c)
end subroutine
The problem is with a(:,p) - matmul(b,c)
It arises because:
(1) gfc_conv_loop_setup chooses matmul's ss to setup the loop bounds.
This explains why -matmul(b,c) + a(:,p) works.
(2) The loop->to are asserted to be NULL in gfc_conv_loop_setup
(GFC_SS_FUNCTION case)
(3) gfc_add_ss_loop_code for the second dimension of a(:,p) calls
gfc_set_loop_bounds_from_array_spec which sets the second dimension of
the loop.to
(4) in the loop setting the descriptor in gfc_trans_create_temp_array,
in the first iteration, (n = 0), loop.to is NULL, and the size is set to
NULL.
(5) In the next iteration, (n = 1), loop.to != NULL, and the loop follows
the normal path. The size however was set to NULL (condition (4)).
(4) and (5) explain why it works with a(q,:) - matmul(b,c)
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37749
