The gfortran.dg/lto/pr87689 test fails as a result of
a -Wstringop-overflow warning. The warning is only enabled for
the C family of languages and LTO but it looks like LTO obviates
this distinction. Last week's enhancement to compute_objsize
seems like the most likely culprit, although the root cause of
the problem -- assuming array indices begin with zero, was
probably latent before then. In r279445 I have committed
the attached patch to remove the assumption that array indices
are zero-based to let the test pass again.
Martin
PR middle-end/92952 - gfortran.dg/lto/pr87689 FAILs
gcc/ChangeLog:
PR middle-end/92952
* builtins.c (compute_objsize): Adjust offset by the array low bound.
Index: gcc/builtins.c
===================================================================
--- gcc/builtins.c (revision 279443)
+++ gcc/builtins.c (working copy)
@@ -3999,6 +3999,16 @@ compute_objsize (tree dest, int ostype, tree *pdec
above. */
if (TREE_CODE (dest) == ARRAY_REF)
{
+ tree lowbnd = array_ref_low_bound (dest);
+ if (!integer_zerop (lowbnd) && tree_fits_uhwi_p (lowbnd))
+ {
+ /* Adjust the offset by the low bound of the array
+ domain (normally zero but 1 in Fortran). */
+ unsigned HOST_WIDE_INT lb = tree_to_uhwi (lowbnd);
+ offrng[0] -= lb;
+ offrng[1] -= lb;
+ }
+
/* Convert the array index into a byte offset. */
tree eltype = TREE_TYPE (dest);
tree tpsize = TYPE_SIZE_UNIT (eltype);
