1 /* Natural loop analysis code for GNU compiler.
2 Copyright (C) 2002, 2003, 2004, 2005, 2006, 2007 Free Software Foundation, Inc.
4 This file is part of GCC.
6 GCC is free software; you can redistribute it and/or modify it under
7 the terms of the GNU General Public License as published by the Free
8 Software Foundation; either version 3, or (at your option) any later
11 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
12 WARRANTY; without even the implied warranty of MERCHANTABILITY or
13 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
16 You should have received a copy of the GNU General Public License
17 along with GCC; see the file COPYING3. If not see
18 <http://www.gnu.org/licenses/>. */
22 #include "coretypes.h"
25 #include "hard-reg-set.h"
27 #include "basic-block.h"
32 /* Checks whether BB is executed exactly once in each LOOP iteration. */
35 just_once_each_iteration_p (const struct loop
*loop
, basic_block bb
)
37 /* It must be executed at least once each iteration. */
38 if (!dominated_by_p (CDI_DOMINATORS
, loop
->latch
, bb
))
42 if (bb
->loop_father
!= loop
)
45 /* But this was not enough. We might have some irreducible loop here. */
46 if (bb
->flags
& BB_IRREDUCIBLE_LOOP
)
52 /* Structure representing edge of a graph. */
56 int src
, dest
; /* Source and destination. */
57 struct edge
*pred_next
, *succ_next
;
58 /* Next edge in predecessor and successor lists. */
59 void *data
; /* Data attached to the edge. */
62 /* Structure representing vertex of a graph. */
66 struct edge
*pred
, *succ
;
67 /* Lists of predecessors and successors. */
68 int component
; /* Number of dfs restarts before reaching the
70 int post
; /* Postorder number. */
73 /* Structure representing a graph. */
77 int n_vertices
; /* Number of vertices. */
78 struct vertex
*vertices
;
82 /* Dumps graph G into F. */
84 extern void dump_graph (FILE *, struct graph
*);
87 dump_graph (FILE *f
, struct graph
*g
)
92 for (i
= 0; i
< g
->n_vertices
; i
++)
94 if (!g
->vertices
[i
].pred
95 && !g
->vertices
[i
].succ
)
98 fprintf (f
, "%d (%d)\t<-", i
, g
->vertices
[i
].component
);
99 for (e
= g
->vertices
[i
].pred
; e
; e
= e
->pred_next
)
100 fprintf (f
, " %d", e
->src
);
104 for (e
= g
->vertices
[i
].succ
; e
; e
= e
->succ_next
)
105 fprintf (f
, " %d", e
->dest
);
110 /* Creates a new graph with N_VERTICES vertices. */
112 static struct graph
*
113 new_graph (int n_vertices
)
115 struct graph
*g
= XNEW (struct graph
);
117 g
->n_vertices
= n_vertices
;
118 g
->vertices
= XCNEWVEC (struct vertex
, n_vertices
);
123 /* Adds an edge from F to T to graph G, with DATA attached. */
126 add_edge (struct graph
*g
, int f
, int t
, void *data
)
128 struct edge
*e
= xmalloc (sizeof (struct edge
));
134 e
->pred_next
= g
->vertices
[t
].pred
;
135 g
->vertices
[t
].pred
= e
;
137 e
->succ_next
= g
->vertices
[f
].succ
;
138 g
->vertices
[f
].succ
= e
;
141 /* Runs dfs search over vertices of G, from NQ vertices in queue QS.
142 The vertices in postorder are stored into QT. If FORWARD is false,
143 backward dfs is run. */
146 dfs (struct graph
*g
, int *qs
, int nq
, int *qt
, bool forward
)
148 int i
, tick
= 0, v
, comp
= 0, top
;
150 struct edge
**stack
= xmalloc (sizeof (struct edge
*) * g
->n_vertices
);
152 for (i
= 0; i
< g
->n_vertices
; i
++)
154 g
->vertices
[i
].component
= -1;
155 g
->vertices
[i
].post
= -1;
158 #define FST_EDGE(V) (forward ? g->vertices[(V)].succ : g->vertices[(V)].pred)
159 #define NEXT_EDGE(E) (forward ? (E)->succ_next : (E)->pred_next)
160 #define EDGE_SRC(E) (forward ? (E)->src : (E)->dest)
161 #define EDGE_DEST(E) (forward ? (E)->dest : (E)->src)
163 for (i
= 0; i
< nq
; i
++)
166 if (g
->vertices
[v
].post
!= -1)
169 g
->vertices
[v
].component
= comp
++;
175 while (e
&& g
->vertices
[EDGE_DEST (e
)].component
!= -1)
182 g
->vertices
[v
].post
= tick
++;
196 g
->vertices
[v
].component
= comp
- 1;
203 /* Marks the edge E in graph G irreducible if it connects two vertices in the
207 check_irred (struct graph
*g
, struct edge
*e
)
211 /* All edges should lead from a component with higher number to the
212 one with lower one. */
213 gcc_assert (g
->vertices
[e
->src
].component
>= g
->vertices
[e
->dest
].component
);
215 if (g
->vertices
[e
->src
].component
!= g
->vertices
[e
->dest
].component
)
218 real
->flags
|= EDGE_IRREDUCIBLE_LOOP
;
219 if (flow_bb_inside_loop_p (real
->src
->loop_father
, real
->dest
))
220 real
->src
->flags
|= BB_IRREDUCIBLE_LOOP
;
223 /* Runs CALLBACK for all edges in G. */
226 for_each_edge (struct graph
*g
,
227 void (callback
) (struct graph
*, struct edge
*))
232 for (i
= 0; i
< g
->n_vertices
; i
++)
233 for (e
= g
->vertices
[i
].succ
; e
; e
= e
->succ_next
)
237 /* Releases the memory occupied by G. */
240 free_graph (struct graph
*g
)
245 for (i
= 0; i
< g
->n_vertices
; i
++)
246 for (e
= g
->vertices
[i
].succ
; e
; e
= n
)
255 /* Marks blocks and edges that are part of non-recognized loops; i.e. we
256 throw away all latch edges and mark blocks inside any remaining cycle.
257 Everything is a bit complicated due to fact we do not want to do this
258 for parts of cycles that only "pass" through some loop -- i.e. for
259 each cycle, we want to mark blocks that belong directly to innermost
260 loop containing the whole cycle.
262 LOOPS is the loop tree. */
264 #define LOOP_REPR(LOOP) ((LOOP)->num + last_basic_block)
265 #define BB_REPR(BB) ((BB)->index + 1)
268 mark_irreducible_loops (struct loops
*loops
)
275 int *queue1
= XNEWVEC (int, last_basic_block
+ loops
->num
);
276 int *queue2
= XNEWVEC (int, last_basic_block
+ loops
->num
);
280 /* Reset the flags. */
281 FOR_BB_BETWEEN (act
, ENTRY_BLOCK_PTR
, EXIT_BLOCK_PTR
, next_bb
)
283 act
->flags
&= ~BB_IRREDUCIBLE_LOOP
;
284 FOR_EACH_EDGE (e
, ei
, act
->succs
)
285 e
->flags
&= ~EDGE_IRREDUCIBLE_LOOP
;
288 /* Create the edge lists. */
289 g
= new_graph (last_basic_block
+ loops
->num
);
291 FOR_BB_BETWEEN (act
, ENTRY_BLOCK_PTR
, EXIT_BLOCK_PTR
, next_bb
)
292 FOR_EACH_EDGE (e
, ei
, act
->succs
)
294 /* Ignore edges to exit. */
295 if (e
->dest
== EXIT_BLOCK_PTR
)
298 /* And latch edges. */
299 if (e
->dest
->loop_father
->header
== e
->dest
300 && e
->dest
->loop_father
->latch
== act
)
303 /* Edges inside a single loop should be left where they are. Edges
304 to subloop headers should lead to representative of the subloop,
305 but from the same place.
307 Edges exiting loops should lead from representative
308 of the son of nearest common ancestor of the loops in that
312 dest
= BB_REPR (e
->dest
);
314 if (e
->dest
->loop_father
->header
== e
->dest
)
315 dest
= LOOP_REPR (e
->dest
->loop_father
);
317 if (!flow_bb_inside_loop_p (act
->loop_father
, e
->dest
))
319 depth
= find_common_loop (act
->loop_father
,
320 e
->dest
->loop_father
)->depth
+ 1;
321 if (depth
== act
->loop_father
->depth
)
322 cloop
= act
->loop_father
;
324 cloop
= act
->loop_father
->pred
[depth
];
326 src
= LOOP_REPR (cloop
);
329 add_edge (g
, src
, dest
, e
);
332 /* Find the strongly connected components. Use the algorithm of Tarjan --
333 first determine the postorder dfs numbering in reversed graph, then
334 run the dfs on the original graph in the order given by decreasing
335 numbers assigned by the previous pass. */
337 FOR_BB_BETWEEN (act
, ENTRY_BLOCK_PTR
, EXIT_BLOCK_PTR
, next_bb
)
339 queue1
[nq
++] = BB_REPR (act
);
341 for (i
= 1; i
< (int) loops
->num
; i
++)
342 if (loops
->parray
[i
])
343 queue1
[nq
++] = LOOP_REPR (loops
->parray
[i
]);
344 dfs (g
, queue1
, nq
, queue2
, false);
345 for (i
= 0; i
< nq
; i
++)
346 queue1
[i
] = queue2
[nq
- i
- 1];
347 dfs (g
, queue1
, nq
, NULL
, true);
349 /* Mark the irreducible loops. */
350 for_each_edge (g
, check_irred
);
356 loops
->state
|= LOOPS_HAVE_MARKED_IRREDUCIBLE_REGIONS
;
359 /* Counts number of insns inside LOOP. */
361 num_loop_insns (struct loop
*loop
)
363 basic_block
*bbs
, bb
;
364 unsigned i
, ninsns
= 0;
367 bbs
= get_loop_body (loop
);
368 for (i
= 0; i
< loop
->num_nodes
; i
++)
372 for (insn
= BB_HEAD (bb
); insn
!= BB_END (bb
); insn
= NEXT_INSN (insn
))
381 /* Counts number of insns executed on average per iteration LOOP. */
383 average_num_loop_insns (struct loop
*loop
)
385 basic_block
*bbs
, bb
;
386 unsigned i
, binsns
, ninsns
, ratio
;
390 bbs
= get_loop_body (loop
);
391 for (i
= 0; i
< loop
->num_nodes
; i
++)
396 for (insn
= BB_HEAD (bb
); insn
!= BB_END (bb
); insn
= NEXT_INSN (insn
))
400 ratio
= loop
->header
->frequency
== 0
402 : (bb
->frequency
* BB_FREQ_MAX
) / loop
->header
->frequency
;
403 ninsns
+= binsns
* ratio
;
407 ninsns
/= BB_FREQ_MAX
;
409 ninsns
= 1; /* To avoid division by zero. */
414 /* Returns expected number of LOOP iterations.
415 Compute upper bound on number of iterations in case they do not fit integer
416 to help loop peeling heuristics. Use exact counts if at all possible. */
418 expected_loop_iterations (const struct loop
*loop
)
423 if (loop
->latch
->count
|| loop
->header
->count
)
425 gcov_type count_in
, count_latch
, expected
;
430 FOR_EACH_EDGE (e
, ei
, loop
->header
->preds
)
431 if (e
->src
== loop
->latch
)
432 count_latch
= e
->count
;
434 count_in
+= e
->count
;
437 expected
= count_latch
* 2;
439 expected
= (count_latch
+ count_in
- 1) / count_in
;
441 /* Avoid overflows. */
442 return (expected
> REG_BR_PROB_BASE
? REG_BR_PROB_BASE
: expected
);
446 int freq_in
, freq_latch
;
451 FOR_EACH_EDGE (e
, ei
, loop
->header
->preds
)
452 if (e
->src
== loop
->latch
)
453 freq_latch
= EDGE_FREQUENCY (e
);
455 freq_in
+= EDGE_FREQUENCY (e
);
458 return freq_latch
* 2;
460 return (freq_latch
+ freq_in
- 1) / freq_in
;
464 /* Returns the maximum level of nesting of subloops of LOOP. */
467 get_loop_level (const struct loop
*loop
)
469 const struct loop
*ploop
;
472 for (ploop
= loop
->inner
; ploop
; ploop
= ploop
->next
)
474 l
= get_loop_level (ploop
);
481 /* Returns estimate on cost of computing SEQ. */
489 for (; seq
; seq
= NEXT_INSN (seq
))
491 set
= single_set (seq
);
493 cost
+= rtx_cost (set
, SET
);
501 /* The properties of the target. */
503 unsigned target_avail_regs
; /* Number of available registers. */
504 unsigned target_res_regs
; /* Number of reserved registers. */
505 unsigned target_small_cost
; /* The cost for register when there is a free one. */
506 unsigned target_pres_cost
; /* The cost for register when there are not too many
508 unsigned target_spill_cost
; /* The cost for register when we need to spill. */
510 /* Initialize the constants for computing set costs. */
513 init_set_costs (void)
516 rtx reg1
= gen_raw_REG (SImode
, FIRST_PSEUDO_REGISTER
);
517 rtx reg2
= gen_raw_REG (SImode
, FIRST_PSEUDO_REGISTER
+ 1);
518 rtx addr
= gen_raw_REG (Pmode
, FIRST_PSEUDO_REGISTER
+ 2);
519 rtx mem
= validize_mem (gen_rtx_MEM (SImode
, addr
));
522 for (i
= 0; i
< FIRST_PSEUDO_REGISTER
; i
++)
523 if (TEST_HARD_REG_BIT (reg_class_contents
[GENERAL_REGS
], i
)
529 /* These are really just heuristic values. */
532 emit_move_insn (reg1
, reg2
);
535 target_small_cost
= seq_cost (seq
);
536 target_pres_cost
= 2 * target_small_cost
;
539 emit_move_insn (mem
, reg1
);
540 emit_move_insn (reg2
, mem
);
543 target_spill_cost
= seq_cost (seq
);
546 /* Calculates cost for having SIZE new loop global variables. REGS_USED is the
547 number of global registers used in loop. N_USES is the number of relevant
551 global_cost_for_size (unsigned size
, unsigned regs_used
, unsigned n_uses
)
553 unsigned regs_needed
= regs_used
+ size
;
556 if (regs_needed
+ target_res_regs
<= target_avail_regs
)
557 cost
+= target_small_cost
* size
;
558 else if (regs_needed
<= target_avail_regs
)
559 cost
+= target_pres_cost
* size
;
562 cost
+= target_pres_cost
* size
;
563 cost
+= target_spill_cost
* n_uses
* (regs_needed
- target_avail_regs
) / regs_needed
;
569 /* Sets EDGE_LOOP_EXIT flag for all exits of LOOPS. */
572 mark_loop_exit_edges (struct loops
*loops
)
584 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
586 if (bb
->loop_father
->outer
587 && loop_exit_edge_p (bb
->loop_father
, e
))
588 e
->flags
|= EDGE_LOOP_EXIT
;
590 e
->flags
&= ~EDGE_LOOP_EXIT
;