diff options
Diffstat (limited to 'gcc')
| -rw-r--r-- | gcc/gimple-range-path.cc | 3 | ||||
| -rw-r--r-- | gcc/testsuite/gcc.dg/tree-ssa/ssa-dom-thread-2b.c | 4 | ||||
| -rw-r--r-- | gcc/testsuite/gcc.dg/tree-ssa/ssa-dom-thread-6.c | 37 | ||||
| -rw-r--r-- | gcc/testsuite/gcc.dg/tree-ssa/ssa-dom-thread-7.c | 17 | ||||
| -rw-r--r-- | gcc/tree-pass.h | 2 | ||||
| -rw-r--r-- | gcc/tree-ssa-loop.c | 2 | ||||
| -rw-r--r-- | gcc/tree-ssa-threadbackward.c | 28 |
7 files changed, 37 insertions, 56 deletions
diff --git a/gcc/gimple-range-path.cc b/gcc/gimple-range-path.cc index a4fa3b2..c616b65 100644 --- a/gcc/gimple-range-path.cc +++ b/gcc/gimple-range-path.cc @@ -127,6 +127,9 @@ path_range_query::internal_range_of_expr (irange &r, tree name, gimple *stmt) 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; } diff --git a/gcc/testsuite/gcc.dg/tree-ssa/ssa-dom-thread-2b.c b/gcc/testsuite/gcc.dg/tree-ssa/ssa-dom-thread-2b.c index e1c33e8..823ada9 100644 --- a/gcc/testsuite/gcc.dg/tree-ssa/ssa-dom-thread-2b.c +++ b/gcc/testsuite/gcc.dg/tree-ssa/ssa-dom-thread-2b.c @@ -1,5 +1,5 @@ /* { 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(); @@ -26,4 +26,4 @@ void thread_latch_through_header (void) 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"} } */ diff --git a/gcc/testsuite/gcc.dg/tree-ssa/ssa-dom-thread-6.c b/gcc/testsuite/gcc.dg/tree-ssa/ssa-dom-thread-6.c index c7bf867..ee46759 100644 --- a/gcc/testsuite/gcc.dg/tree-ssa/ssa-dom-thread-6.c +++ b/gcc/testsuite/gcc.dg/tree-ssa/ssa-dom-thread-6.c @@ -1,41 +1,8 @@ /* { 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) diff --git a/gcc/testsuite/gcc.dg/tree-ssa/ssa-dom-thread-7.c b/gcc/testsuite/gcc.dg/tree-ssa/ssa-dom-thread-7.c index 5fc2145..ba07942 100644 --- a/gcc/testsuite/gcc.dg/tree-ssa/ssa-dom-thread-7.c +++ b/gcc/testsuite/gcc.dg/tree-ssa/ssa-dom-thread-7.c @@ -1,23 +1,8 @@ /* { 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 diff --git a/gcc/tree-pass.h b/gcc/tree-pass.h index 83941bc..eb75eb1 100644 --- a/gcc/tree-pass.h +++ b/gcc/tree-pass.h @@ -225,6 +225,8 @@ protected: 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) diff --git a/gcc/tree-ssa-loop.c b/gcc/tree-ssa-loop.c index 0cc4b3b..1bbf2f1 100644 --- a/gcc/tree-ssa-loop.c +++ b/gcc/tree-ssa-loop.c @@ -540,7 +540,7 @@ const pass_data pass_data_tree_loop_done = 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 */ diff --git a/gcc/tree-ssa-threadbackward.c b/gcc/tree-ssa-threadbackward.c index 449232c..e729923 100644 --- a/gcc/tree-ssa-threadbackward.c +++ b/gcc/tree-ssa-threadbackward.c @@ -43,6 +43,7 @@ along with GCC; see the file COPYING3. If not see #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 @@ -564,7 +565,10 @@ back_threader_registry::thread_through_all_blocks (bool may_peel_loop_headers) 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, @@ -725,7 +729,11 @@ 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]); @@ -845,6 +853,22 @@ back_threader_profitability::profitable_path_p (const vec<basic_block> &m_path, "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; } |