------- Comment #19 from dominiq at lps dot ens dot fr 2010-08-10 12:00 ------- (In reply to comment #18) > Although it is probably better set the stride during resolution.
Probably. A few comments before leaving for ten days without access to the net. (1) I think all the changes done recently to avoid unneeded temporaries can be extended by reversing the loops. (2) The following code generates an unneeded temporary: integer, parameter :: n = 10 integer :: i integer :: a(0:n+1) = (/(i, i =0, n), 0/) a(:6:2) = a(::3) print *, a end It is in the class abs(r_stride)>max(0,abs(l_stride)) for which there is no need for temporary if l_start<r_start+r_stride for r_stride>max(0,l_stride) or l_start>r_start+r_stride for r_stride<min(0,l_stride) If I am not mistaken, this is the last case for which one can conclude for any size of the sections. (I have done the proof on the back of an envelope that I cannot find right now.) For all the remaining cases it exists some set of triplets for which there is a dependency even if for some others there is none (depending upon the dependent indices are inside or not of the actual range). Although this can be implemented, I seriously doubt it is worth the pain. (3) I think the coverage for missing temporaries is rather limited right now and I am planning to write some torture test(s) to improve this situation. -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45159
