The following should fix PR88651 (don't have UBSAN GCC to verify).
We mangle HWI max_stmt_execution result in ways not considering
overflow but max_stmt_execution can return quite large numbers.
The following switches the code over to widest_ints.
Bootstrapped and tested on x86_64-unknown-linux-gnu, applied to trunk.
Richard.
2019-01-02 Richard Biener <rguent...@suse.de>
PR middle-end/88651
* tree-data-ref.c (analyze_subscript_affine_affine): Use
widest_ints when mangling max_stmt_execution results.
diff --git a/gcc/tree-data-ref.c b/gcc/tree-data-ref.c
index d9a8d3a7d9d..7d1f03c66af 100644
--- a/gcc/tree-data-ref.c
+++ b/gcc/tree-data-ref.c
@@ -3761,10 +3761,6 @@ analyze_subscript_affine_affine (tree chrec_a,
if (niter > 0)
{
- HOST_WIDE_INT tau2 = MIN (FLOOR_DIV (niter_a - i0, i1),
- FLOOR_DIV (niter_b - j0, j1));
- HOST_WIDE_INT last_conflict = tau2 - (x1 - i0)/i1;
-
/* If the overlap occurs outside of the bounds of the
loop, there is no dependence. */
if (x1 >= niter_a || y1 >= niter_b)
@@ -3774,8 +3770,20 @@ analyze_subscript_affine_affine (tree chrec_a,
*last_conflicts = integer_zero_node;
goto end_analyze_subs_aa;
}
+
+ /* max stmt executions can get quite large, avoid
+ overflows by using wide ints here. */
+ widest_int tau2
+ = wi::smin (wi::sdiv_floor (wi::sub (niter_a, i0), i1),
+ wi::sdiv_floor (wi::sub (niter_b, j0), j1));
+ widest_int last_conflict = wi::sub (tau2, (x1 - i0)/i1);
+ if (wi::min_precision (last_conflict, SIGNED)
+ <= TYPE_PRECISION (integer_type_node))
+ *last_conflicts
+ = build_int_cst (integer_type_node,
+ last_conflict.to_shwi ());
else
- *last_conflicts = build_int_cst (NULL_TREE, last_conflict);
+ *last_conflicts = chrec_dont_know;
}
else
*last_conflicts = chrec_dont_know;
