1 /* Natural loop analysis code for GNU compiler.
2 Copyright (C) 2002-2021 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"
35 #include "function-abi.h"
37 struct target_cfgloop default_target_cfgloop
;
39 struct target_cfgloop
*this_target_cfgloop
= &default_target_cfgloop
;
42 /* Checks whether BB is executed exactly once in each LOOP iteration. */
45 just_once_each_iteration_p (const class loop
*loop
, const_basic_block bb
)
47 /* It must be executed at least once each iteration. */
48 if (!dominated_by_p (CDI_DOMINATORS
, loop
->latch
, bb
))
52 if (bb
->loop_father
!= loop
)
55 /* But this was not enough. We might have some irreducible loop here. */
56 if (bb
->flags
& BB_IRREDUCIBLE_LOOP
)
62 /* Marks blocks and edges that are part of non-recognized loops; i.e. we
63 throw away all latch edges and mark blocks inside any remaining cycle.
64 Everything is a bit complicated due to fact we do not want to do this
65 for parts of cycles that only "pass" through some loop -- i.e. for
66 each cycle, we want to mark blocks that belong directly to innermost
67 loop containing the whole cycle.
69 LOOPS is the loop tree. */
71 #define LOOP_REPR(LOOP) ((LOOP)->num + last_basic_block_for_fn (cfun))
72 #define BB_REPR(BB) ((BB)->index + 1)
75 mark_irreducible_loops (void)
78 struct graph_edge
*ge
;
84 int num
= number_of_loops (cfun
);
86 bool irred_loop_found
= false;
89 gcc_assert (current_loops
!= NULL
);
91 /* Reset the flags. */
92 FOR_BB_BETWEEN (act
, ENTRY_BLOCK_PTR_FOR_FN (cfun
),
93 EXIT_BLOCK_PTR_FOR_FN (cfun
), next_bb
)
95 act
->flags
&= ~BB_IRREDUCIBLE_LOOP
;
96 FOR_EACH_EDGE (e
, ei
, act
->succs
)
97 e
->flags
&= ~EDGE_IRREDUCIBLE_LOOP
;
100 /* Create the edge lists. */
101 g
= new_graph (last_basic_block_for_fn (cfun
) + num
);
103 FOR_BB_BETWEEN (act
, ENTRY_BLOCK_PTR_FOR_FN (cfun
),
104 EXIT_BLOCK_PTR_FOR_FN (cfun
), next_bb
)
105 FOR_EACH_EDGE (e
, ei
, act
->succs
)
107 /* Ignore edges to exit. */
108 if (e
->dest
== EXIT_BLOCK_PTR_FOR_FN (cfun
))
112 dest
= BB_REPR (e
->dest
);
114 /* Ignore latch edges. */
115 if (e
->dest
->loop_father
->header
== e
->dest
116 && e
->dest
->loop_father
->latch
== act
)
119 /* Edges inside a single loop should be left where they are. Edges
120 to subloop headers should lead to representative of the subloop,
121 but from the same place.
123 Edges exiting loops should lead from representative
124 of the son of nearest common ancestor of the loops in that
127 if (e
->dest
->loop_father
->header
== e
->dest
)
128 dest
= LOOP_REPR (e
->dest
->loop_father
);
130 if (!flow_bb_inside_loop_p (act
->loop_father
, e
->dest
))
132 depth
= 1 + loop_depth (find_common_loop (act
->loop_father
,
133 e
->dest
->loop_father
));
134 if (depth
== loop_depth (act
->loop_father
))
135 cloop
= act
->loop_father
;
137 cloop
= (*act
->loop_father
->superloops
)[depth
];
139 src
= LOOP_REPR (cloop
);
142 add_edge (g
, src
, dest
)->data
= e
;
145 /* Find the strongly connected components. */
146 graphds_scc (g
, NULL
);
148 /* Mark the irreducible loops. */
149 for (i
= 0; i
< g
->n_vertices
; i
++)
150 for (ge
= g
->vertices
[i
].succ
; ge
; ge
= ge
->succ_next
)
152 edge real
= (edge
) ge
->data
;
153 /* edge E in graph G is irreducible if it connects two vertices in the
156 /* All edges should lead from a component with higher number to the
157 one with lower one. */
158 gcc_assert (g
->vertices
[ge
->src
].component
>= g
->vertices
[ge
->dest
].component
);
160 if (g
->vertices
[ge
->src
].component
!= g
->vertices
[ge
->dest
].component
)
163 real
->flags
|= EDGE_IRREDUCIBLE_LOOP
;
164 irred_loop_found
= true;
165 if (flow_bb_inside_loop_p (real
->src
->loop_father
, real
->dest
))
166 real
->src
->flags
|= BB_IRREDUCIBLE_LOOP
;
171 loops_state_set (LOOPS_HAVE_MARKED_IRREDUCIBLE_REGIONS
);
172 return irred_loop_found
;
175 /* Counts number of insns inside LOOP. */
177 num_loop_insns (const class loop
*loop
)
179 basic_block
*bbs
, bb
;
180 unsigned i
, ninsns
= 0;
183 bbs
= get_loop_body (loop
);
184 for (i
= 0; i
< loop
->num_nodes
; i
++)
187 FOR_BB_INSNS (bb
, insn
)
188 if (NONDEBUG_INSN_P (insn
))
194 ninsns
= 1; /* To avoid division by zero. */
199 /* Counts number of insns executed on average per iteration LOOP. */
201 average_num_loop_insns (const class loop
*loop
)
203 basic_block
*bbs
, bb
;
209 bbs
= get_loop_body (loop
);
210 for (i
= 0; i
< loop
->num_nodes
; i
++)
215 FOR_BB_INSNS (bb
, insn
)
216 if (NONDEBUG_INSN_P (insn
))
219 ninsns
+= (sreal
)binsns
* bb
->count
.to_sreal_scale (loop
->header
->count
);
220 /* Avoid overflows. */
221 if (ninsns
> 1000000)
229 int64_t ret
= ninsns
.to_int ();
231 ret
= 1; /* To avoid division by zero. */
236 /* Returns expected number of iterations of LOOP, according to
237 measured or guessed profile.
239 This functions attempts to return "sane" value even if profile
240 information is not good enough to derive osmething.
241 If BY_PROFILE_ONLY is set, this logic is bypassed and function
242 return -1 in those scenarios. */
245 expected_loop_iterations_unbounded (const class loop
*loop
,
246 bool *read_profile_p
,
247 bool by_profile_only
)
251 gcov_type expected
= -1;
254 *read_profile_p
= false;
256 /* If we have no profile at all, use AVG_LOOP_NITER. */
257 if (profile_status_for_fn (cfun
) == PROFILE_ABSENT
)
261 expected
= param_avg_loop_niter
;
263 else if (loop
->latch
&& (loop
->latch
->count
.initialized_p ()
264 || loop
->header
->count
.initialized_p ()))
266 profile_count count_in
= profile_count::zero (),
267 count_latch
= profile_count::zero ();
269 FOR_EACH_EDGE (e
, ei
, loop
->header
->preds
)
270 if (e
->src
== loop
->latch
)
271 count_latch
= e
->count ();
273 count_in
+= e
->count ();
275 if (!count_latch
.initialized_p ())
279 expected
= param_avg_loop_niter
;
281 else if (!count_in
.nonzero_p ())
285 expected
= count_latch
.to_gcov_type () * 2;
289 expected
= (count_latch
.to_gcov_type () + count_in
.to_gcov_type ()
290 - 1) / count_in
.to_gcov_type ();
292 && count_latch
.reliable_p () && count_in
.reliable_p ())
293 *read_profile_p
= true;
300 expected
= param_avg_loop_niter
;
303 if (!by_profile_only
)
305 HOST_WIDE_INT max
= get_max_loop_iterations_int (loop
);
306 if (max
!= -1 && max
< expected
)
313 /* Returns expected number of LOOP iterations. The returned value is bounded
314 by REG_BR_PROB_BASE. */
317 expected_loop_iterations (class loop
*loop
)
319 gcov_type expected
= expected_loop_iterations_unbounded (loop
);
320 return (expected
> REG_BR_PROB_BASE
? REG_BR_PROB_BASE
: expected
);
323 /* Returns the maximum level of nesting of subloops of LOOP. */
326 get_loop_level (const class loop
*loop
)
328 const class loop
*ploop
;
331 for (ploop
= loop
->inner
; ploop
; ploop
= ploop
->next
)
333 l
= get_loop_level (ploop
);
340 /* Initialize the constants for computing set costs. */
343 init_set_costs (void)
347 rtx reg1
= gen_raw_REG (SImode
, LAST_VIRTUAL_REGISTER
+ 1);
348 rtx reg2
= gen_raw_REG (SImode
, LAST_VIRTUAL_REGISTER
+ 2);
349 rtx addr
= gen_raw_REG (Pmode
, LAST_VIRTUAL_REGISTER
+ 3);
350 rtx mem
= validize_mem (gen_rtx_MEM (SImode
, addr
));
353 target_avail_regs
= 0;
354 target_clobbered_regs
= 0;
355 for (i
= 0; i
< FIRST_PSEUDO_REGISTER
; i
++)
356 if (TEST_HARD_REG_BIT (reg_class_contents
[GENERAL_REGS
], i
)
360 /* ??? This is only a rough heuristic. It doesn't cope well
361 with alternative ABIs, but that's an optimization rather than
362 correctness issue. */
363 if (default_function_abi
.clobbers_full_reg_p (i
))
364 target_clobbered_regs
++;
369 for (speed
= 0; speed
< 2; speed
++)
371 crtl
->maybe_hot_insn_p
= speed
;
372 /* Set up the costs for using extra registers:
374 1) If not many free registers remain, we should prefer having an
375 additional move to decreasing the number of available registers.
377 2) If no registers are available, we need to spill, which may require
378 storing the old value to memory and loading it back
379 (TARGET_SPILL_COST). */
382 emit_move_insn (reg1
, reg2
);
385 target_reg_cost
[speed
] = seq_cost (seq
, speed
);
388 emit_move_insn (mem
, reg1
);
389 emit_move_insn (reg2
, mem
);
392 target_spill_cost
[speed
] = seq_cost (seq
, speed
);
394 default_rtl_profile ();
397 /* Estimates cost of increased register pressure caused by making N_NEW new
398 registers live around the loop. N_OLD is the number of registers live
399 around the loop. If CALL_P is true, also take into account that
400 call-used registers may be clobbered in the loop body, reducing the
401 number of available registers before we spill. */
404 estimate_reg_pressure_cost (unsigned n_new
, unsigned n_old
, bool speed
,
408 unsigned regs_needed
= n_new
+ n_old
;
409 unsigned available_regs
= target_avail_regs
;
411 /* If there is a call in the loop body, the call-clobbered registers
412 are not available for loop invariants. */
414 available_regs
= available_regs
- target_clobbered_regs
;
416 /* If we have enough registers, we should use them and not restrict
417 the transformations unnecessarily. */
418 if (regs_needed
+ target_res_regs
<= available_regs
)
421 if (regs_needed
<= available_regs
)
422 /* If we are close to running out of registers, try to preserve
424 cost
= target_reg_cost
[speed
] * n_new
;
426 /* If we run out of registers, it is very expensive to add another
428 cost
= target_spill_cost
[speed
] * n_new
;
430 if (optimize
&& (flag_ira_region
== IRA_REGION_ALL
431 || flag_ira_region
== IRA_REGION_MIXED
)
432 && number_of_loops (cfun
) <= (unsigned) param_ira_max_loops_num
)
433 /* IRA regional allocation deals with high register pressure
434 better. So decrease the cost (to do more accurate the cost
435 calculation for IRA, we need to know how many registers lives
436 through the loop transparently). */
442 /* Sets EDGE_LOOP_EXIT flag for all loop exits. */
445 mark_loop_exit_edges (void)
450 if (number_of_loops (cfun
) <= 1)
453 FOR_EACH_BB_FN (bb
, cfun
)
457 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
459 if (loop_outer (bb
->loop_father
)
460 && loop_exit_edge_p (bb
->loop_father
, e
))
461 e
->flags
|= EDGE_LOOP_EXIT
;
463 e
->flags
&= ~EDGE_LOOP_EXIT
;
468 /* Return exit edge if loop has only one exit that is likely
469 to be executed on runtime (i.e. it is not EH or leading
473 single_likely_exit (class loop
*loop
, vec
<edge
> exits
)
475 edge found
= single_exit (loop
);
481 FOR_EACH_VEC_ELT (exits
, i
, ex
)
483 if (probably_never_executed_edge_p (cfun
, ex
)
484 /* We want to rule out paths to noreturns but not low probabilities
485 resulting from adjustments or combining.
486 FIXME: once we have better quality tracking, make this more
488 || ex
->probability
<= profile_probability::very_unlikely ())
499 /* Gets basic blocks of a LOOP. Header is the 0-th block, rest is in dfs
500 order against direction of edges from latch. Specially, if
501 header != latch, latch is the 1-st block. */
504 get_loop_hot_path (const class loop
*loop
)
506 basic_block bb
= loop
->header
;
507 vec
<basic_block
> path
= vNULL
;
508 bitmap visited
= BITMAP_ALLOC (NULL
);
517 bitmap_set_bit (visited
, bb
->index
);
518 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
519 if ((!best
|| e
->probability
> best
->probability
)
520 && !loop_exit_edge_p (loop
, e
)
521 && !bitmap_bit_p (visited
, e
->dest
->index
))
523 if (!best
|| best
->dest
== loop
->header
)
527 BITMAP_FREE (visited
);