------- Comment #4 from fxcoudert at gcc dot gnu dot org 2008-05-12 23:15
-------
Well, that one is subtle, but I think I found it: it's a TRUNC_DIV_EXPR that
should be a FLOOR_DIV_EXPR in gfc_conv_loop_setup(). That simple fix makes all
testcases in this PR run, except when -fbounds-check is added. To fix
-fbounds-check, we need to correctly make all zero-sized section zero-sized,
and not any-negative-sized, which is done by two appropriate MAX_EXPR. The
complete patch is thus:
Index: trans-array.c
===================================================================
--- trans-array.c (revision 135088)
+++ trans-array.c (working copy)
@@ -3083,6 +3083,8 @@ gfc_conv_ss_startstride (gfc_loopinfo *
info->start[n]);
tmp = fold_build2 (FLOOR_DIV_EXPR, gfc_array_index_type, tmp,
info->stride[n]);
+ tmp = fold_build2 (MAX_EXPR, gfc_array_index_type, tmp,
+ build_int_cst (gfc_array_index_type, 0));
/* We remember the size of the first section, and check all the
others against this. */
if (size[n])
@@ -3435,8 +3437,10 @@ gfc_conv_loop_setup (gfc_loopinfo * loop
for (i = 0; i<=last; i++){...}; */
tmp = fold_build2 (MINUS_EXPR, gfc_array_index_type,
loop->to[n], loop->from[n]);
- tmp = fold_build2 (TRUNC_DIV_EXPR, gfc_array_index_type,
+ tmp = fold_build2 (FLOOR_DIV_EXPR, gfc_array_index_type,
tmp, info->stride[n]);
+ tmp = fold_build2 (MAX_EXPR, gfc_array_index_type, tmp,
+ build_int_cst (gfc_array_index_type, -1));
loop->to[n] = gfc_evaluate_now (tmp, &loop->pre);
/* Make the loop variable start at 0. */
loop->from[n] = gfc_index_zero_node;
Bootstraps OK, I've started regtesting.
--
fxcoudert at gcc dot gnu dot org changed:
What |Removed |Added
----------------------------------------------------------------------------
AssignedTo|unassigned at gcc dot gnu |fxcoudert at gcc dot gnu dot
|dot org |org
Status|NEW |ASSIGNED
Last reconfirmed|2008-03-25 16:37:47 |2008-05-12 23:15:08
date| |
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=35682
