http://gcc.gnu.org/bugzilla/show_bug.cgi?id=48636
--- Comment #10 from Thomas Koenig <tkoenig at gcc dot gnu.org> 2011-04-20
16:40:46 UTC ---
(In reply to comment #6)
> Not strictly related to inlining, but in the new descriptor we'll have a field
> specifying whether the array is simply contiguous, so it might make sense to
> generate two loops for each loop over the array in the source, one for the
> contiguous case where it can be vectorized etc. and another loop for the
> general case. This might reduce the profitability of inlining.
Consider the following, hand-crafted matmul:
Here, we have three nested loops. The most interesting one is
the innermost loop of the matmul, which we vectorize by inlining if we omit
the call to my_matmul with non-unity stride for a when compiling with
-fwhole-program -O3.
How many versions of the loop should we generate? Two or eight, depending
on what the caller may do? ;-)
module foo
implicit none
contains
subroutine my_matmul(a,b,c)
implicit none
integer :: count, m, n
real, dimension(:,:), intent(in) :: a,b
real, dimension(:,:), intent(out) :: c
integer :: i,j,k
m = ubound(a,1)
n = ubound(b,2)
count = ubound(a,2)
c = 0
do j=1,n
do k=1, count
do i=1,m
c(i,j) = c(i,j) + a(i,k) * b(k,j)
end do
end do
end do
end subroutine my_matmul
end module foo
program main
use foo
implicit none
integer, parameter :: factor=100
integer, parameter :: n = 2*factor, m = 3*factor, count = 4*factor
real, dimension(m, count) :: a
real, dimension(count, n) :: b
real, dimension(m,n) :: c1, c2
real, dimension(m/2, n) :: ch_1, ch_2
call random_number(a)
call random_number(b)
call my_matmul(a,b,c1)
c2 = matmul(a,b)
if (any(abs(c1 - c2) > 1e-5)) call abort
call my_matmul(a(1:m:2,:),b,ch_1)
ch_2 = matmul(a(1:m:2,:),b)
if (any(abs(ch_1 - ch_2) > 1e-5)) call abort
end program main
