/* Natural loop analysis code for GNU compiler.
- Copyright (C) 2002, 2003, 2004, 2005, 2006 Free Software Foundation, Inc.
+ Copyright (C) 2002, 2003, 2004, 2005, 2006, 2007, 2008, 2009
+ 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 2, or (at your option) any later
+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
for more details.
You should have received a copy of the GNU General Public License
-along with GCC; see the file COPYING. If not, write to the Free
-Software Foundation, 51 Franklin Street, Fifth Floor, Boston, MA
-02110-1301, USA. */
+along with GCC; see the file COPYING3. If not see
+<http://www.gnu.org/licenses/>. */
#include "config.h"
#include "system.h"
#include "expr.h"
#include "output.h"
#include "graphds.h"
+#include "params.h"
+
+struct target_cfgloop default_target_cfgloop;
+#if SWITCHABLE_TARGET
+struct target_cfgloop *this_target_cfgloop = &default_target_cfgloop;
+#endif
/* Checks whether BB is executed exactly once in each LOOP iteration. */
bool
-just_once_each_iteration_p (const struct loop *loop, basic_block bb)
+just_once_each_iteration_p (const struct loop *loop, const_basic_block bb)
{
/* It must be executed at least once each iteration. */
if (!dominated_by_p (CDI_DOMINATORS, loop->latch, bb))
return true;
}
-/* Marks the edge E in graph G irreducible if it connects two vertices in the
- same scc. */
-
-static void
-check_irred (struct graph *g, struct graph_edge *e)
-{
- edge real = (edge) e->data;
-
- /* All edges should lead from a component with higher number to the
- one with lower one. */
- gcc_assert (g->vertices[e->src].component >= g->vertices[e->dest].component);
-
- if (g->vertices[e->src].component != g->vertices[e->dest].component)
- return;
-
- real->flags |= EDGE_IRREDUCIBLE_LOOP;
- if (flow_bb_inside_loop_p (real->src->loop_father, real->dest))
- real->src->flags |= BB_IRREDUCIBLE_LOOP;
-}
-
/* Marks blocks and edges that are part of non-recognized loops; i.e. we
throw away all latch edges and mark blocks inside any remaining cycle.
Everything is a bit complicated due to fact we do not want to do this
#define LOOP_REPR(LOOP) ((LOOP)->num + last_basic_block)
#define BB_REPR(BB) ((BB)->index + 1)
-void
+bool
mark_irreducible_loops (void)
{
basic_block act;
+ struct graph_edge *ge;
edge e;
edge_iterator ei;
int src, dest;
struct graph *g;
int num = number_of_loops ();
struct loop *cloop;
+ bool irred_loop_found = false;
+ int i;
gcc_assert (current_loops != NULL);
graphds_scc (g, NULL);
/* Mark the irreducible loops. */
- for_each_edge (g, check_irred);
+ for (i = 0; i < g->n_vertices; i++)
+ for (ge = g->vertices[i].succ; ge; ge = ge->succ_next)
+ {
+ edge real = (edge) ge->data;
+ /* edge E in graph G is irreducible if it connects two vertices in the
+ same scc. */
+
+ /* All edges should lead from a component with higher number to the
+ one with lower one. */
+ gcc_assert (g->vertices[ge->src].component >= g->vertices[ge->dest].component);
+
+ if (g->vertices[ge->src].component != g->vertices[ge->dest].component)
+ continue;
+
+ real->flags |= EDGE_IRREDUCIBLE_LOOP;
+ irred_loop_found = true;
+ if (flow_bb_inside_loop_p (real->src->loop_father, real->dest))
+ real->src->flags |= BB_IRREDUCIBLE_LOOP;
+ }
free_graph (g);
- current_loops->state |= LOOPS_HAVE_MARKED_IRREDUCIBLE_REGIONS;
+ loops_state_set (LOOPS_HAVE_MARKED_IRREDUCIBLE_REGIONS);
+ return irred_loop_found;
}
/* Counts number of insns inside LOOP. */
int
-num_loop_insns (struct loop *loop)
+num_loop_insns (const struct loop *loop)
{
basic_block *bbs, bb;
unsigned i, ninsns = 0;
for (i = 0; i < loop->num_nodes; i++)
{
bb = bbs[i];
- ninsns++;
- for (insn = BB_HEAD (bb); insn != BB_END (bb); insn = NEXT_INSN (insn))
- if (INSN_P (insn))
+ FOR_BB_INSNS (bb, insn)
+ if (NONDEBUG_INSN_P (insn))
ninsns++;
}
- free(bbs);
+ free (bbs);
+
+ if (!ninsns)
+ ninsns = 1; /* To avoid division by zero. */
return ninsns;
}
/* Counts number of insns executed on average per iteration LOOP. */
int
-average_num_loop_insns (struct loop *loop)
+average_num_loop_insns (const struct loop *loop)
{
basic_block *bbs, bb;
unsigned i, binsns, ninsns, ratio;
{
bb = bbs[i];
- binsns = 1;
- for (insn = BB_HEAD (bb); insn != BB_END (bb); insn = NEXT_INSN (insn))
- if (INSN_P (insn))
+ binsns = 0;
+ FOR_BB_INSNS (bb, insn)
+ if (NONDEBUG_INSN_P (insn))
binsns++;
ratio = loop->header->frequency == 0
: (bb->frequency * BB_FREQ_MAX) / loop->header->frequency;
ninsns += binsns * ratio;
}
- free(bbs);
+ free (bbs);
ninsns /= BB_FREQ_MAX;
if (!ninsns)
/* Returns estimate on cost of computing SEQ. */
static unsigned
-seq_cost (rtx seq)
+seq_cost (const_rtx seq, bool speed)
{
unsigned cost = 0;
rtx set;
{
set = single_set (seq);
if (set)
- cost += rtx_cost (set, SET);
+ cost += rtx_cost (set, SET, speed);
else
cost++;
}
return cost;
}
-/* The properties of the target. */
-
-unsigned target_avail_regs; /* Number of available registers. */
-unsigned target_res_regs; /* Number of registers reserved for temporary
- expressions. */
-unsigned target_reg_cost; /* The cost for register when there still
- is some reserve, but we are approaching
- the number of available registers. */
-unsigned target_spill_cost; /* The cost for register when we need
- to spill. */
-
/* Initialize the constants for computing set costs. */
void
init_set_costs (void)
{
+ int speed;
rtx seq;
rtx reg1 = gen_raw_REG (SImode, FIRST_PSEUDO_REGISTER);
rtx reg2 = gen_raw_REG (SImode, FIRST_PSEUDO_REGISTER + 1);
rtx mem = validize_mem (gen_rtx_MEM (SImode, addr));
unsigned i;
+ target_avail_regs = 0;
+ target_clobbered_regs = 0;
for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
if (TEST_HARD_REG_BIT (reg_class_contents[GENERAL_REGS], i)
&& !fixed_regs[i])
- target_avail_regs++;
+ {
+ target_avail_regs++;
+ if (call_used_regs[i])
+ target_clobbered_regs++;
+ }
target_res_regs = 3;
- /* Set up the costs for using extra registers:
-
- 1) If not many free registers remain, we should prefer having an
- additional move to decreasing the number of available registers.
- (TARGET_REG_COST).
- 2) If no registers are available, we need to spill, which may require
- storing the old value to memory and loading it back
- (TARGET_SPILL_COST). */
-
- start_sequence ();
- emit_move_insn (reg1, reg2);
- seq = get_insns ();
- end_sequence ();
- target_reg_cost = seq_cost (seq);
-
- start_sequence ();
- emit_move_insn (mem, reg1);
- emit_move_insn (reg2, mem);
- seq = get_insns ();
- end_sequence ();
- target_spill_cost = seq_cost (seq);
+ for (speed = 0; speed < 2; speed++)
+ {
+ crtl->maybe_hot_insn_p = speed;
+ /* Set up the costs for using extra registers:
+
+ 1) If not many free registers remain, we should prefer having an
+ additional move to decreasing the number of available registers.
+ (TARGET_REG_COST).
+ 2) If no registers are available, we need to spill, which may require
+ storing the old value to memory and loading it back
+ (TARGET_SPILL_COST). */
+
+ start_sequence ();
+ emit_move_insn (reg1, reg2);
+ seq = get_insns ();
+ end_sequence ();
+ target_reg_cost [speed] = seq_cost (seq, speed);
+
+ start_sequence ();
+ emit_move_insn (mem, reg1);
+ emit_move_insn (reg2, mem);
+ seq = get_insns ();
+ end_sequence ();
+ target_spill_cost [speed] = seq_cost (seq, speed);
+ }
+ default_rtl_profile ();
}
/* Estimates cost of increased register pressure caused by making N_NEW new
registers live around the loop. N_OLD is the number of registers live
- around the loop. */
+ around the loop. If CALL_P is true, also take into account that
+ call-used registers may be clobbered in the loop body, reducing the
+ number of available registers before we spill. */
unsigned
-estimate_reg_pressure_cost (unsigned n_new, unsigned n_old)
+estimate_reg_pressure_cost (unsigned n_new, unsigned n_old, bool speed,
+ bool call_p)
{
+ unsigned cost;
unsigned regs_needed = n_new + n_old;
+ unsigned available_regs = target_avail_regs;
+
+ /* If there is a call in the loop body, the call-clobbered registers
+ are not available for loop invariants. */
+ if (call_p)
+ available_regs = available_regs - target_clobbered_regs;
/* If we have enough registers, we should use them and not restrict
the transformations unnecessarily. */
- if (regs_needed + target_res_regs <= target_avail_regs)
+ if (regs_needed + target_res_regs <= available_regs)
return 0;
- /* If we are close to running out of registers, try to preserve them. */
- if (regs_needed <= target_avail_regs)
- return target_reg_cost * n_new;
-
- /* If we run out of registers, it is very expensive to add another one. */
- return target_spill_cost * n_new;
+ if (regs_needed <= available_regs)
+ /* If we are close to running out of registers, try to preserve
+ them. */
+ cost = target_reg_cost [speed] * n_new;
+ else
+ /* If we run out of registers, it is very expensive to add another
+ one. */
+ cost = target_spill_cost [speed] * n_new;
+
+ if (optimize && (flag_ira_region == IRA_REGION_ALL
+ || flag_ira_region == IRA_REGION_MIXED)
+ && number_of_loops () <= (unsigned) IRA_MAX_LOOPS_NUM)
+ /* IRA regional allocation deals with high register pressure
+ better. So decrease the cost (to do more accurate the cost
+ calculation for IRA, we need to know how many registers lives
+ through the loop transparently). */
+ cost /= 2;
+
+ return cost;
}
/* Sets EDGE_LOOP_EXIT flag for all loop exits. */