As can be seen on the testcase in PR91526 the vectorizer can end
up marking things addressable because it creates very many
pointer IV. Then unrolling can completely elide a loop or IVOPTs
can choose other IVs. Thus it's certainly beneficial for
other passes to have an updated TREE_ADDRESSABLE, esp. NRV
which bails out on TREE_ADDRESSABLE marked vars. Similar
tail-call wants to know this.
Not enough to fix the PR, but anyways.
Bootstrap / regtest running on x86_64-unknown-linux-gnu.
Richard.
2019-08-23 Richard Biener <rguent...@suse.de>
PR tree-optimization/91526
* passes.def: Note that after late FRE we do TODO_update_address_taken.
* tree-ssa-sccvn.c (pass_fre::execute): In late mode schedule
TODO_update_address_taken.
Index: gcc/passes.def
===================================================================
--- gcc/passes.def (revision 274843)
+++ gcc/passes.def (working copy)
@@ -313,6 +313,8 @@ along with GCC; see the file COPYING3.
NEXT_PASS (pass_split_paths);
NEXT_PASS (pass_tracer);
NEXT_PASS (pass_fre, false /* may_iterate */);
+ /* After late FRE we rewrite no longer addressed locals into SSA
+ form if possible. */
NEXT_PASS (pass_thread_jumps);
NEXT_PASS (pass_dominator, false /* may_peel_loop_headers_p */);
NEXT_PASS (pass_strlen);
Index: gcc/tree-ssa-sccvn.c
===================================================================
--- gcc/tree-ssa-sccvn.c (revision 274843)
+++ gcc/tree-ssa-sccvn.c (working copy)
@@ -7312,6 +7415,11 @@ pass_fre::execute (function *fun)
if (iterate_p)
loop_optimizer_finalize ();
+ /* For late FRE after IVOPTs and unrolling, see if we can
+ remove some TREE_ADDRESSABLE and rewrite stuff into SSA. */
+ if (!may_iterate)
+ todo |= TODO_update_address_taken;
+
return todo;
}
