tree-if-conv.c: fix ICE seen with -fno-tree-forwprop (PR tree-optimization/84178)
PR tree-optimization/84178 reports a couple of source files that ICE inside
ifcvt when compiled with -03 -fno-tree-forwprop (trunk and gcc 7).
Both cases involve problems with ifcvt's per-BB gimplified predicates.
Testcase 1 fails this assertion within release_bb_predicate during cleanup:
283 if (flag_checking)
284 for (gimple_stmt_iterator i = gsi_start (stmts);
285 !gsi_end_p (i); gsi_next (&i))
286 gcc_assert (! gimple_use_ops (gsi_stmt (i)));
The testcase contains a division in the loop, which leads to
if_convertible_loop_p returning false (due to gimple_could_trap_p being true
for the division). This happens *after* the per-BB gimplified predicates
have been created in predicate_bbs (loop).
Hence tree_if_conversion bails out to "cleanup", but the gimplified predicates
exist and make use of SSA names; for example this conjunction for two BB
conditions:
_4 = h4.1_112 != 0;
_175 = (signed char) _117;
_176 = _175 >= 0;
_174 = _4 & _176;
is using SSA names.
This assertion was added in r236498 (aka
c3deca2519d97c55876869c57cf11ae1e5c6cf8b):
2016-05-20 Richard Biener <rguenther@suse.de>
* tree-if-conv.c (add_bb_predicate_gimplified_stmts): Use
gimple_seq_add_seq_without_update.
(release_bb_predicate): Assert we have no operands to free.
(if_convertible_loop_p_1): Calculate post dominators later.
Do not free BB predicates here.
(combine_blocks): Do not recompute BB predicates.
(version_loop_for_if_conversion): Save BB predicates around
loop versioning.
* gcc.dg/tree-ssa/ifc-cd.c: Adjust.
The following patch fixes this by adding a call to gimple_seq_discard
to release_bb_predicate. It also updates the assertion, so that
instead of asserting the stmts have no imm uses, instead assert that
they weren't added to a bb before discarding them (otherwise discarding
them would be a bug). We know this is the case because
insert_gimplified_predicates has:
/* Once the sequence is code generated, set it to NULL. */
set_bb_predicate_gimplified_stmts (bb, NULL);
but asserting it seems appropriate as a double-check.
The patch doesn't address the 2nd issue within PR tree-optimization/84178.
gcc/ChangeLog:
PR tree-optimization/84178
* tree-if-conv.c (release_bb_predicate): Remove the
the assertion that the stmts have NULL use_ops.
Discard the statements, asserting that they haven't
yet been added to a BB.
gcc/testsuite/ChangeLog:
PR tree-optimization/84178
* gcc.c-torture/compile/pr84178-1.c: New test.
git-svn-id: svn+ssh://gcc.gnu.org/svn/gcc/trunk@258363 138bc75d-0d04-0410-961f-82ee72b054a4