basic_block bb = stmt ? gimple_bb (stmt) : exit_bb ();
if (stmt && range_defined_in_block (r, name, bb))
{
+ if (TREE_CODE (name) == SSA_NAME)
+ r.intersect (gimple_range_global (name));
+
set_cache (r, name);
return true;
}
/* { dg-do compile } */
-/* { dg-options "-O2 -fdump-tree-thread1-stats -fdump-tree-dom2-stats -fdisable-tree-ethread" } */
+/* { dg-options "-O2 -fdump-tree-thread3-stats -fdump-tree-dom2-stats -fdisable-tree-ethread" } */
void foo();
void bla();
case. And we want to thread through the header as well. These
are both caught by threading in DOM. */
/* { dg-final { scan-tree-dump-not "Jumps threaded" "dom2"} } */
-/* { dg-final { scan-tree-dump-times "Jumps threaded: 1" 1 "thread1"} } */
+/* { dg-final { scan-tree-dump-times "Jumps threaded: 1" 1 "thread3"} } */
/* { dg-do compile } */
-/* { dg-options "-O2 -fdump-tree-thread1-details -fdump-tree-thread2-details" } */
+/* { dg-options "-O2 -fdump-tree-thread1-details -fdump-tree-thread3-details" } */
-/* All the threads in the thread1 dump start on a X->BB12 edge, as can
- be seen in the dump:
-
- Registering FSM jump thread: (x, 12) incoming edge; ...
- etc
- etc
-
- Before the new evrp, we were threading paths that started at the
- following edges:
-
- Registering FSM jump thread: (10, 12) incoming edge
- Registering FSM jump thread: (6, 12) incoming edge
- Registering FSM jump thread: (9, 12) incoming edge
-
- This was because the PHI at BB12 had constant values coming in from
- BB10, BB6, and BB9:
-
- # state_10 = PHI <state_11(7), 0(10), state_11(5), 1(6), state_11(8), 2(9), state_11(11)>
-
- Now with the new evrp, we get:
-
- # state_10 = PHI <0(7), 0(10), state_11(5), 1(6), 0(8), 2(9), 1(11)>
-
- Thus, we have 3 more paths that are known to be constant and can be
- threaded. Which means that by the second threading pass, we can
- only find one profitable path.
-
- For the record, all these extra constants are better paths coming
- out of switches. For example:
-
- SWITCH_BB -> BBx -> BBy -> BBz -> PHI
-
- We now know the value of the switch index at PHI. */
/* { dg-final { scan-tree-dump-times "Registering FSM jump" 6 "thread1" } } */
-/* { dg-final { scan-tree-dump-times "Registering FSM jump" 1 "thread2" } } */
+/* { dg-final { scan-tree-dump-times "Registering FSM jump" 1 "thread3" } } */
int sum0, sum1, sum2, sum3;
int foo (char *s, char **ret)
/* { dg-do compile } */
/* { dg-options "-O2 -fdump-tree-thread1-stats -fdump-tree-thread2-stats -fdump-tree-dom2-stats -fdump-tree-thread3-stats -fdump-tree-dom3-stats -fdump-tree-vrp2-stats -fno-guess-branch-probability" } */
-/* Here we have the same issue as was commented in ssa-dom-thread-6.c.
- The PHI coming into the threader has a lot more constants, so the
- threader can thread more paths.
-
-$ diff clean/a.c.105t.mergephi2 a.c.105t.mergephi2
-252c252
-< # s_50 = PHI <s_49(10), 5(14), s_51(18), s_51(22), 1(26), 1(29), 1(31), s_51(5), 4(12), 1(15), 5(17), 1(19), 3(21), 1(23), 6(25), 7(28), s_51(30)>
----
-> # s_50 = PHI <s_49(10), 5(14), 4(18), 5(22), 1(26), 1(29), 1(31), s_51(5), 4(12), 1(15), 5(17), 1(19), 3(21), 1(23), 6(25), 7(28), 7(30)>
-272a273
-
- I spot checked a few and they all have the same pattern. We are
- basically tracking the switch index better through multiple
- paths. */
-
/* { dg-final { scan-tree-dump "Jumps threaded: 18" "thread1" } } */
-/* { dg-final { scan-tree-dump "Jumps threaded: 8" "thread2" } } */
+/* { dg-final { scan-tree-dump "Jumps threaded: 8" "thread3" } } */
/* { dg-final { scan-tree-dump-not "Jumps threaded" "dom2" } } */
/* aarch64 has the highest CASE_VALUES_THRESHOLD in GCC. It's high enough
been optimized. */
#define PROP_gimple_lomp_dev (1 << 16) /* done omp_device_lower */
#define PROP_rtl_split_insns (1 << 17) /* RTL has insns split. */
+#define PROP_loop_opts_done (1 << 18) /* SSA loop optimizations
+ have completed. */
#define PROP_gimple \
(PROP_gimple_any | PROP_gimple_lcf | PROP_gimple_leh | PROP_gimple_lomp)
OPTGROUP_LOOP, /* optinfo_flags */
TV_NONE, /* tv_id */
PROP_cfg, /* properties_required */
- 0, /* properties_provided */
+ PROP_loop_opts_done, /* properties_provided */
0, /* properties_destroyed */
0, /* todo_flags_start */
TODO_cleanup_cfg, /* todo_flags_finish */
#include "ssa.h"
#include "tree-cfgcleanup.h"
#include "tree-pretty-print.h"
+#include "cfghooks.h"
// Path registry for the backwards threader. After all paths have been
// registered with register_path(), thread_through_all_blocks() is called
TAKEN_EDGE, otherwise it is NULL.
CREATES_IRREDUCIBLE_LOOP, if non-null is set to TRUE if threading this path
- would create an irreducible loop. */
+ would create an irreducible loop.
+
+ ?? It seems we should be able to loosen some of the restrictions in
+ this function after loop optimizations have run. */
bool
back_threader_profitability::profitable_path_p (const vec<basic_block> &m_path,
the last entry in the array when determining if we thread
through the loop latch. */
if (loop->latch == bb)
- threaded_through_latch = true;
+ {
+ threaded_through_latch = true;
+ if (dump_file && (dump_flags & TDF_DETAILS))
+ fprintf (dump_file, " (latch)");
+ }
}
gimple *stmt = get_gimple_control_stmt (m_path[0]);
"a multiway branch.\n");
return false;
}
+
+ /* Threading through an empty latch would cause code to be added to
+ the latch. This could alter the loop form sufficiently to cause
+ loop optimizations to fail. Disable these threads until after
+ loop optimizations have run. */
+ if ((threaded_through_latch
+ || (taken_edge && taken_edge->dest == loop->latch))
+ && !(cfun->curr_properties & PROP_loop_opts_done)
+ && empty_block_p (loop->latch))
+ {
+ if (dump_file && (dump_flags & TDF_DETAILS))
+ fprintf (dump_file,
+ " FAIL: FSM Thread through latch before loop opts would create non-empty latch\n");
+ return false;
+
+ }
return true;
}
projects / platform / upstream / gcc.git / commitdiff