Hi!
On the testcase in the PR we ICE because we create non-canonical CONST_INTs
(e.g. for HImode niter 0xfffe) and the compare_and_jump_seq code then ICEs
on that.
Fixed thusly, bootstrapped/regtested on x86_64-linux and i686-linux, ok for
trunk?
Testcase not included, because it creates 480000 basic blocks and we handle
that with O(n^2) complexity.
2019-03-19 Jakub Jelinek <ja...@redhat.com>
PR rtl-optimization/89768
* loop-unroll.c (unroll_loop_constant_iterations): Use gen_int_mode
instead of GEN_INT.
(unroll_loop_runtime_iterations): Likewise.
--- gcc/loop-unroll.c.jj 2019-03-19 09:09:32.686006683 +0100
+++ gcc/loop-unroll.c 2019-03-19 10:15:50.319343904 +0100
@@ -652,7 +652,7 @@ unroll_loop_constant_iterations (struct
if (loop->any_likely_upper_bound)
loop->nb_iterations_likely_upper_bound
= wi::udiv_trunc (loop->nb_iterations_likely_upper_bound, max_unroll +
1);
- desc->niter_expr = GEN_INT (desc->niter);
+ desc->niter_expr = gen_int_mode (desc->niter, desc->mode);
/* Remove the edges. */
FOR_EACH_VEC_ELT (remove_edges, i, e)
@@ -1020,9 +1020,9 @@ unroll_loop_runtime_iterations (struct l
preheader = split_edge (loop_preheader_edge (loop));
/* Add in count of edge from switch block. */
preheader->count += iter_count;
- branch_code = compare_and_jump_seq (copy_rtx (niter), GEN_INT (j), EQ,
- block_label (preheader), p,
- NULL);
+ branch_code = compare_and_jump_seq (copy_rtx (niter),
+ gen_int_mode (j, desc->mode), EQ,
+ block_label (preheader), p, NULL);
/* We rely on the fact that the compare and jump cannot be optimized out,
and hence the cfg we create is correct. */
Jakub
