[PATCH] [PR middle-end/69347] Don't do useless work for abnormal SSA_NAMEs

[Jeff Law]

There's no point in walking through the PHIs to trace values for 
SSA_NAMEs that appear in abnormal PHIs -- we're not supposed to be 
threading those paths to begin with.

This happens to avoid a lot of useless work for BZ 69347.

Bootstrapped and regression tested on x86.  Committed to the trunk.


Jeff

ps.  There's still a significant size regression with the FSM threader, 
this patch shouldn't impact that BZ.
diff --git a/gcc/ChangeLog b/gcc/ChangeLog
index f41b2c6..2930a8f 100644
--- a/gcc/ChangeLog
+++ b/gcc/ChangeLog
@@ -1,3 +1,10 @@
+2016-01-19  Jeff Law  <l...@redhat.com>
+
+       PR middle-end/69347
+       * tree-ssa-threadbackwards.c
+       (fsm_find_control_statement_thread_paths): Do not try to lookup
+       FSM paths for SSA_NAMEs appearing in abnormal PHIs.
+
 2016-01-20  Kugan Vivekanandarajah  <kug...@linaro.org>
 
        * doc/lto.texi: Remove text that says only Gold has linker plugin
diff --git a/gcc/tree-ssa-threadbackward.c b/gcc/tree-ssa-threadbackward.c
index 8c99f6f..8d8aa30 100644
--- a/gcc/tree-ssa-threadbackward.c
+++ b/gcc/tree-ssa-threadbackward.c
@@ -99,6 +99,11 @@ fsm_find_control_statement_thread_paths (tree name,
                                         vec<basic_block, va_gc> *&path,
                                         bool seen_loop_phi)
 {
+  /* If NAME appears in an abnormal PHI, then don't try to trace its
+     value back through PHI nodes.  */
+  if (SSA_NAME_OCCURS_IN_ABNORMAL_PHI (name))
+    return;
+
   gimple *def_stmt = SSA_NAME_DEF_STMT (name);
   basic_block var_bb = gimple_bb (def_stmt);