7ee2468b92
gcc/ * system.h (dump_file): Do not define. * tree-pass.h: Include dumpfile.h, which is a new file containing... (enum tree_dump_index, TDF_*, get_dump_file_name, dump_enabled_p, dump_initialized_p, dump_begin, dump_end, dump_node, dump_switch_p, dump_flag_name, dump_file, dump_flags, dump_file_name, get_dump_file_info, struct dump_file_info): all of this, moved to... * dumpfile.h: Here, new file. * tree-dump.h: Include dumpfile.h, but not tree-pass.h. (dump_stmt): Remove prototype for C++ specific function. (dump_enumerated_decls): Move prototype from here... * tree-flow.h (dump_enumerated_decls): ... to here. (get_ref_base_and_extent) Move prototype from here ... * tree.h (get_ref_base_and_extent) ... to here. * tree-ssa-live.c: Do not inclde tree-pretty-print.h, because gimple-pretty-print.h is enough. Do not include tree-dump.h, include timevar.h and dumpfile.h instead. (struct numbered_tree_d, compare_decls_by_uid, dump_enumerated_decls_push, dump_enumerated_decls): Move from here ... * tree-dfa.c:(struct numbered_tree_d, compare_decls_by_uid, dump_enumerated_decls_push, dump_enumerated_decls):... to here. Do not include timevar.h. * tree.c: Do not include timevar.h. * tree-cfg.c: Do not include langhooks.h, tree-pretty-print.h, and timevar.h. (dump_cfg_stats): Use current_function_name. (gimple_cfg2vcg): Likewise. (dump_function_to_file): Likewise. * df-scan.c: Do not include tree-pass.h and timevar.h. Include dumpfile.h. (df_entry_block_bitmap_verify, df_exit_block_bitmap_verify): Do not use print_current_pass. * df-problems.c: Include dumpfile.h. Always define REG_DEAD_DEBUGGING, avoid #ifdef code, because it leads to errors in the code not selected. (df_note_compute): Do not print_rtl_with_bb here. Fix compilation bug if REG_DEAD_DEBUGGING is not 0, get_insns is not available here. * lcm.c: Include dumpfile.h. Remove obsolete include of insn-attr.h. * dojump.c (do_compare_rtx_and_jump): Remove failure printing for missing probability notes. * stmt.c: Include dumpfile.h. (emit_case_decision_tree): Re-enable printing expand details only if TDF_DETAILS. * alias.c, auto-inc-dec.c, bb-reorder.c, caller-save.c, cfg.c, cfgcleanup.c, cfgexpand.c, cfgloop.c, cfgloopmanip.c, cgraph.c, cgraphclones.c, cgraphunit.c, combine.c, combine-stack-adj.c, coverage.c, cprop.c, cse.c, cselib.c, dbgcnt.c, dce.c, df-core.c, dse.c, dwarf2out.c, emit-rtl.c, except.c, expr.c, final.c, function.c, fwprop.c, gcse.c, gimple-fold.c, gimple-pretty-print.c, gimple-ssa-strength-reduction.c, gimplify.c, graphite-blocking.c, graphite-clast-to-gimple.c, graphite-dependences.c, graphite-interchange.c, graphite-optimize-isl.c, graphite-poly.c, graphite-sese-to-poly.c, haifa-sched.c, hw-doloop.c, ifcvt.c, ipa.c, ipa-cp.c, ipa-inline-analysis.c, ipa-inline.c, ipa-inline-transform.c, ipa-prop.c, ipa-pure-const.c, ipa-reference.c, ipa-split.c, ipa-utils.c, ira.c, ira-emit.c, jump.c, loop-doloop.c, loop-init.c, loop-invariant.c, loop-iv.c, loop-unroll.c, loop-unswitch.c, lower-subreg.c, lto-section-out.c, lto-streamer-in.c, matrix-reorg.c, mcf.c, mode-switching.c, modulo-sched.c, omega.c, omp-low.c, passes.c, plugin.c, postreload.c, postreload-gcse.c, predict.c, print-rtl.c, print-tree.c, profile.c, recog.c, ree.c, regcprop.c, reginfo.c, regmove.c, regrename.c, reg-stack.c, reload1.c, reorg.c, sched-rgn.c, sched-vis.c, sel-sched.c, sel-sched-ir.c, store-motion.c, tracer.c, trans-mem.c, tree-affine.c, tree-call-cdce.c, tree-cfgcleanup.c, tree-chrec.c, tree-data-ref.c, tree-diagnostic.c, tree-dump.c, tree-eh.c, tree-flow-inline.h, tree-if-conv.c, tree-into-ssa.c, tree-mudflap.c, tree-nrv.c, tree-object-size.c, tree-optimize.c, tree-outof-ssa.c, tree-predcom.c, tree-pretty-print.c, tree-profile.c, tree-scalar-evolution.c, tree-sra.c, tree-ssa-address.c, tree-ssa-alias.c, tree-ssa.c, tree-ssa-ccp.c, tree-ssa-coalesce.c, tree-ssa-copy.c, tree-ssa-copyrename.c,, tree-ssa-dce.c, tree-ssa-dom.c, tree-ssa-dse.c, tree-ssa-forwprop.c, tree-ssa-ifcombine.c, tree-ssa-loop.c, tree-ssa-loop-ch.c, tree-ssa-loop-im.c, tree-ssa-loop-ivcanon.c, tree-ssa-loop-ivopts.c, tree-ssa-loop-manip.c, tree-ssa-loop-niter.c, tree-ssa-loop-prefetch.c, tree-ssa-loop-unswitch.c, tree-ssa-math-opts.c, tree-ssa-operands.c, tree-ssa-phiopt.c, tree-ssa-phiprop.c, tree-ssa-pre.c, tree-ssa-propagate.c, tree-ssa-reassoc.c, tree-ssa-sccvn.c, tree-ssa-sink.c, tree-ssa-structalias.c, tree-ssa-tail-merge.c, tree-ssa-ter.c, tree-ssa-threadedge.c, tree-ssa-threadupdate.c, tree-ssa-uncprop.c, tree-ssa-uninit.c, tree-switch-conversion.c, tree-tailcall.c, tree-vect-data-refs.c, tree-vect-loop.c, tree-vect-loop-manip.c, tree-vectorizer.c, tree-vect-patterns.c, tree-vect-slp.c, tree-vect-stmts.c, tree-vrp.c, value-prof.c, var-tracking.c, web.c: Include tree-pass.h only if needed. If tree-pass.h is included, do not include timevar.h and dumpfile.h. If tree-pass.h is not included but dump_file, or dump_flags, or the TDF_* flags are used, include dumpfile.h. If gimple-pretty-print.h is included, don't include tree-pretty-print.h. Remove assorted unnecessary includes. * config/mn10300/mn10300.c, config/c6x/c6x.c, config/ia64/ia64.c, config/arm/arm.c, config/bfin/bfin.c, config/frv/frv.c, config/spu/spu.c, config/mep/mep.c, config/i386/i386.c: Include dumpfile.h. * config/rl78/rl78.c: Include dumpfile.h instead of tree-pass.h. * arm/t-arm, avr/t-avr, i386/t-i386, ia64/t-ia64, mep/t-mep, spu/t-spu-elf: Fix dependencies. c-family/ * c-gimplify.c: Include dumpfile.h instead of tree-dump.h. * c-ada-spec.c: Likewise. * c-dump.c (dump_stmt): Move to cp/dump.c, the only user. c/ * c-decl.c: Include dumpfile.h instead of tree-dump.h. * Make-lang.in: Fix dependencies. cp/ * dump.c (dump_stmt): Moved here from c-dump.c. * optimize.c: Include dumpfile.h instead of tree-dump.h. * class.c: Likewise. * decl2.c: Likewise. * Make-lang.in: Fix dependencies. fortran/ * f95-lang.c: Include dumpfile.h instead of tree-dump.h. * Make-lang.in: Fix dependencies. java/ * java-gimplify.c Include dumpfile.h instead of tree-dump.h * Make-lang.in: Fix dependencies. lto/ * lto.c: Do not include timevar.h. * Make-lang.in: Fix dependencies. ada/ * gcc-interface/utils.c: Include timevar.h. * Make-lang.in: Fix dependencies. From-SVN: r189519
380 lines
9.6 KiB
C
380 lines
9.6 KiB
C
/* Loop unswitching.
|
|
Copyright (C) 2004, 2005, 2007, 2008, 2010 Free Software Foundation, Inc.
|
|
|
|
This file is part of GCC.
|
|
|
|
GCC is free software; you can redistribute it and/or modify it
|
|
under the terms of the GNU General Public License as published by the
|
|
Free Software Foundation; either version 3, or (at your option) any
|
|
later version.
|
|
|
|
GCC is distributed in the hope that it will be useful, but WITHOUT
|
|
ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
|
|
FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
|
|
for more details.
|
|
|
|
You should have received a copy of the GNU General Public License
|
|
along with GCC; see the file COPYING3. If not see
|
|
<http://www.gnu.org/licenses/>. */
|
|
|
|
#include "config.h"
|
|
#include "system.h"
|
|
#include "coretypes.h"
|
|
#include "tm.h"
|
|
#include "tree.h"
|
|
#include "tm_p.h"
|
|
#include "basic-block.h"
|
|
#include "tree-flow.h"
|
|
#include "cfgloop.h"
|
|
#include "params.h"
|
|
#include "tree-pass.h"
|
|
#include "tree-inline.h"
|
|
|
|
/* This file implements the loop unswitching, i.e. transformation of loops like
|
|
|
|
while (A)
|
|
{
|
|
if (inv)
|
|
B;
|
|
|
|
X;
|
|
|
|
if (!inv)
|
|
C;
|
|
}
|
|
|
|
where inv is the loop invariant, into
|
|
|
|
if (inv)
|
|
{
|
|
while (A)
|
|
{
|
|
B;
|
|
X;
|
|
}
|
|
}
|
|
else
|
|
{
|
|
while (A)
|
|
{
|
|
X;
|
|
C;
|
|
}
|
|
}
|
|
|
|
Inv is considered invariant iff the values it compares are both invariant;
|
|
tree-ssa-loop-im.c ensures that all the suitable conditions are in this
|
|
shape. */
|
|
|
|
static struct loop *tree_unswitch_loop (struct loop *, basic_block, tree);
|
|
static bool tree_unswitch_single_loop (struct loop *, int);
|
|
static tree tree_may_unswitch_on (basic_block, struct loop *);
|
|
|
|
/* Main entry point. Perform loop unswitching on all suitable loops. */
|
|
|
|
unsigned int
|
|
tree_ssa_unswitch_loops (void)
|
|
{
|
|
loop_iterator li;
|
|
struct loop *loop;
|
|
bool changed = false;
|
|
|
|
/* Go through inner loops (only original ones). */
|
|
FOR_EACH_LOOP (li, loop, LI_ONLY_INNERMOST)
|
|
{
|
|
if (dump_file && (dump_flags & TDF_DETAILS))
|
|
fprintf (dump_file, ";; Considering loop %d\n", loop->num);
|
|
|
|
/* Do not unswitch in cold regions. */
|
|
if (optimize_loop_for_size_p (loop))
|
|
{
|
|
if (dump_file && (dump_flags & TDF_DETAILS))
|
|
fprintf (dump_file, ";; Not unswitching cold loops\n");
|
|
continue;
|
|
}
|
|
|
|
/* The loop should not be too large, to limit code growth. */
|
|
if (tree_num_loop_insns (loop, &eni_size_weights)
|
|
> (unsigned) PARAM_VALUE (PARAM_MAX_UNSWITCH_INSNS))
|
|
{
|
|
if (dump_file && (dump_flags & TDF_DETAILS))
|
|
fprintf (dump_file, ";; Not unswitching, loop too big\n");
|
|
continue;
|
|
}
|
|
|
|
changed |= tree_unswitch_single_loop (loop, 0);
|
|
}
|
|
|
|
if (changed)
|
|
return TODO_cleanup_cfg;
|
|
return 0;
|
|
}
|
|
|
|
/* Checks whether we can unswitch LOOP on condition at end of BB -- one of its
|
|
basic blocks (for what it means see comments below). */
|
|
|
|
static tree
|
|
tree_may_unswitch_on (basic_block bb, struct loop *loop)
|
|
{
|
|
gimple stmt, def;
|
|
tree cond, use;
|
|
basic_block def_bb;
|
|
ssa_op_iter iter;
|
|
|
|
/* BB must end in a simple conditional jump. */
|
|
stmt = last_stmt (bb);
|
|
if (!stmt || gimple_code (stmt) != GIMPLE_COND)
|
|
return NULL_TREE;
|
|
|
|
/* To keep the things simple, we do not directly remove the conditions,
|
|
but just replace tests with 0 != 0 resp. 1 != 0. Prevent the infinite
|
|
loop where we would unswitch again on such a condition. */
|
|
if (gimple_cond_true_p (stmt) || gimple_cond_false_p (stmt))
|
|
return NULL_TREE;
|
|
|
|
/* Condition must be invariant. */
|
|
FOR_EACH_SSA_TREE_OPERAND (use, stmt, iter, SSA_OP_USE)
|
|
{
|
|
def = SSA_NAME_DEF_STMT (use);
|
|
def_bb = gimple_bb (def);
|
|
if (def_bb
|
|
&& flow_bb_inside_loop_p (loop, def_bb))
|
|
return NULL_TREE;
|
|
}
|
|
|
|
cond = build2 (gimple_cond_code (stmt), boolean_type_node,
|
|
gimple_cond_lhs (stmt), gimple_cond_rhs (stmt));
|
|
|
|
return cond;
|
|
}
|
|
|
|
/* Simplifies COND using checks in front of the entry of the LOOP. Just very
|
|
simplish (sufficient to prevent us from duplicating loop in unswitching
|
|
unnecessarily). */
|
|
|
|
static tree
|
|
simplify_using_entry_checks (struct loop *loop, tree cond)
|
|
{
|
|
edge e = loop_preheader_edge (loop);
|
|
gimple stmt;
|
|
|
|
while (1)
|
|
{
|
|
stmt = last_stmt (e->src);
|
|
if (stmt
|
|
&& gimple_code (stmt) == GIMPLE_COND
|
|
&& gimple_cond_code (stmt) == TREE_CODE (cond)
|
|
&& operand_equal_p (gimple_cond_lhs (stmt),
|
|
TREE_OPERAND (cond, 0), 0)
|
|
&& operand_equal_p (gimple_cond_rhs (stmt),
|
|
TREE_OPERAND (cond, 1), 0))
|
|
return (e->flags & EDGE_TRUE_VALUE
|
|
? boolean_true_node
|
|
: boolean_false_node);
|
|
|
|
if (!single_pred_p (e->src))
|
|
return cond;
|
|
|
|
e = single_pred_edge (e->src);
|
|
if (e->src == ENTRY_BLOCK_PTR)
|
|
return cond;
|
|
}
|
|
}
|
|
|
|
/* Unswitch single LOOP. NUM is number of unswitchings done; we do not allow
|
|
it to grow too much, it is too easy to create example on that the code would
|
|
grow exponentially. */
|
|
|
|
static bool
|
|
tree_unswitch_single_loop (struct loop *loop, int num)
|
|
{
|
|
basic_block *bbs;
|
|
struct loop *nloop;
|
|
unsigned i, found;
|
|
tree cond = NULL_TREE;
|
|
gimple stmt;
|
|
bool changed = false;
|
|
|
|
i = 0;
|
|
bbs = get_loop_body (loop);
|
|
found = loop->num_nodes;
|
|
|
|
while (1)
|
|
{
|
|
/* Find a bb to unswitch on. */
|
|
for (; i < loop->num_nodes; i++)
|
|
if ((cond = tree_may_unswitch_on (bbs[i], loop)))
|
|
break;
|
|
|
|
if (i == loop->num_nodes)
|
|
{
|
|
if (dump_file
|
|
&& num > PARAM_VALUE (PARAM_MAX_UNSWITCH_LEVEL)
|
|
&& (dump_flags & TDF_DETAILS))
|
|
fprintf (dump_file, ";; Not unswitching anymore, hit max level\n");
|
|
|
|
if (found == loop->num_nodes)
|
|
{
|
|
free (bbs);
|
|
return changed;
|
|
}
|
|
break;
|
|
}
|
|
|
|
cond = simplify_using_entry_checks (loop, cond);
|
|
stmt = last_stmt (bbs[i]);
|
|
if (integer_nonzerop (cond))
|
|
{
|
|
/* Remove false path. */
|
|
gimple_cond_set_condition_from_tree (stmt, boolean_true_node);
|
|
changed = true;
|
|
}
|
|
else if (integer_zerop (cond))
|
|
{
|
|
/* Remove true path. */
|
|
gimple_cond_set_condition_from_tree (stmt, boolean_false_node);
|
|
changed = true;
|
|
}
|
|
/* Do not unswitch too much. */
|
|
else if (num > PARAM_VALUE (PARAM_MAX_UNSWITCH_LEVEL))
|
|
{
|
|
i++;
|
|
continue;
|
|
}
|
|
/* In nested tree_unswitch_single_loop first optimize all conditions
|
|
using entry checks, then discover still reachable blocks in the
|
|
loop and find the condition only among those still reachable bbs. */
|
|
else if (num != 0)
|
|
{
|
|
if (found == loop->num_nodes)
|
|
found = i;
|
|
i++;
|
|
continue;
|
|
}
|
|
else
|
|
{
|
|
found = i;
|
|
break;
|
|
}
|
|
|
|
update_stmt (stmt);
|
|
i++;
|
|
}
|
|
|
|
if (num != 0)
|
|
{
|
|
basic_block *tos, *worklist;
|
|
|
|
/* When called recursively, first do a quick discovery
|
|
of reachable bbs after the above changes and only
|
|
consider conditions in still reachable bbs. */
|
|
tos = worklist = XNEWVEC (basic_block, loop->num_nodes);
|
|
|
|
for (i = 0; i < loop->num_nodes; i++)
|
|
bbs[i]->flags &= ~BB_REACHABLE;
|
|
|
|
/* Start with marking header. */
|
|
*tos++ = bbs[0];
|
|
bbs[0]->flags |= BB_REACHABLE;
|
|
|
|
/* Iterate: find everything reachable from what we've already seen
|
|
within the same innermost loop. Don't look through false edges
|
|
if condition is always true or true edges if condition is
|
|
always false. */
|
|
while (tos != worklist)
|
|
{
|
|
basic_block b = *--tos;
|
|
edge e;
|
|
edge_iterator ei;
|
|
int flags = 0;
|
|
|
|
if (EDGE_COUNT (b->succs) == 2)
|
|
{
|
|
gimple stmt = last_stmt (b);
|
|
if (stmt
|
|
&& gimple_code (stmt) == GIMPLE_COND)
|
|
{
|
|
if (gimple_cond_true_p (stmt))
|
|
flags = EDGE_FALSE_VALUE;
|
|
else if (gimple_cond_false_p (stmt))
|
|
flags = EDGE_TRUE_VALUE;
|
|
}
|
|
}
|
|
|
|
FOR_EACH_EDGE (e, ei, b->succs)
|
|
{
|
|
basic_block dest = e->dest;
|
|
|
|
if (dest->loop_father == loop
|
|
&& !(dest->flags & BB_REACHABLE)
|
|
&& !(e->flags & flags))
|
|
{
|
|
*tos++ = dest;
|
|
dest->flags |= BB_REACHABLE;
|
|
}
|
|
}
|
|
}
|
|
|
|
free (worklist);
|
|
|
|
/* Find a bb to unswitch on. */
|
|
for (; found < loop->num_nodes; found++)
|
|
if ((bbs[found]->flags & BB_REACHABLE)
|
|
&& (cond = tree_may_unswitch_on (bbs[found], loop)))
|
|
break;
|
|
|
|
if (found == loop->num_nodes)
|
|
{
|
|
free (bbs);
|
|
return changed;
|
|
}
|
|
}
|
|
|
|
if (dump_file && (dump_flags & TDF_DETAILS))
|
|
fprintf (dump_file, ";; Unswitching loop\n");
|
|
|
|
initialize_original_copy_tables ();
|
|
/* Unswitch the loop on this condition. */
|
|
nloop = tree_unswitch_loop (loop, bbs[found], cond);
|
|
if (!nloop)
|
|
{
|
|
free_original_copy_tables ();
|
|
free (bbs);
|
|
return changed;
|
|
}
|
|
|
|
/* Update the SSA form after unswitching. */
|
|
update_ssa (TODO_update_ssa);
|
|
free_original_copy_tables ();
|
|
|
|
/* Invoke itself on modified loops. */
|
|
tree_unswitch_single_loop (nloop, num + 1);
|
|
tree_unswitch_single_loop (loop, num + 1);
|
|
free (bbs);
|
|
return true;
|
|
}
|
|
|
|
/* Unswitch a LOOP w.r. to given basic block UNSWITCH_ON. We only support
|
|
unswitching of innermost loops. COND is the condition determining which
|
|
loop is entered -- the new loop is entered if COND is true. Returns NULL
|
|
if impossible, new loop otherwise. */
|
|
|
|
static struct loop *
|
|
tree_unswitch_loop (struct loop *loop,
|
|
basic_block unswitch_on, tree cond)
|
|
{
|
|
unsigned prob_true;
|
|
edge edge_true, edge_false;
|
|
|
|
/* Some sanity checking. */
|
|
gcc_assert (flow_bb_inside_loop_p (loop, unswitch_on));
|
|
gcc_assert (EDGE_COUNT (unswitch_on->succs) == 2);
|
|
gcc_assert (loop->inner == NULL);
|
|
|
|
extract_true_false_edges_from_block (unswitch_on, &edge_true, &edge_false);
|
|
prob_true = edge_true->probability;
|
|
return loop_version (loop, unshare_expr (cond),
|
|
NULL, prob_true, prob_true,
|
|
REG_BR_PROB_BASE - prob_true, false);
|
|
}
|