1 /* Code to analyze doloop loops in order for targets to perform late
2 optimizations converting doloops to other forms of hardware loops.
3 Copyright (C) 2011 Free Software Foundation, Inc.
5 This file is part of GCC.
7 GCC is free software; you can redistribute it and/or modify it under
8 the terms of the GNU General Public License as published by the Free
9 Software Foundation; either version 3, or (at your option) any later
12 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
13 WARRANTY; without even the implied warranty of MERCHANTABILITY or
14 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
17 You should have received a copy of the GNU General Public License
18 along with GCC; see the file COPYING3. If not see
19 <http://www.gnu.org/licenses/>. */
23 #include "coretypes.h"
28 #include "hard-reg-set.h"
30 #include "basic-block.h"
36 #include "hw-doloop.h"
38 #ifdef HAVE_doloop_end
40 /* Dump information collected in LOOPS. */
42 dump_hwloops (hwloop_info loops
)
46 for (loop
= loops
; loop
; loop
= loop
->next
)
52 fprintf (dump_file
, ";; loop %d: ", loop
->loop_no
);
54 fprintf (dump_file
, "(bad) ");
55 fprintf (dump_file
, "{head:%d, depth:%d, reg:%u}",
56 loop
->head
== NULL
? -1 : loop
->head
->index
,
57 loop
->depth
, REGNO (loop
->iter_reg
));
59 fprintf (dump_file
, " blocks: [ ");
60 for (ix
= 0; VEC_iterate (basic_block
, loop
->blocks
, ix
, b
); ix
++)
61 fprintf (dump_file
, "%d ", b
->index
);
62 fprintf (dump_file
, "] ");
64 fprintf (dump_file
, " inner loops: [ ");
65 for (ix
= 0; VEC_iterate (hwloop_info
, loop
->loops
, ix
, i
); ix
++)
66 fprintf (dump_file
, "%d ", i
->loop_no
);
67 fprintf (dump_file
, "]\n");
69 fprintf (dump_file
, "\n");
72 /* Return true if BB is part of LOOP. */
74 bb_in_loop_p (hwloop_info loop
, basic_block bb
)
76 return bitmap_bit_p (loop
->block_bitmap
, bb
->index
);
79 /* Scan the blocks of LOOP (and its inferiors), and record things such as
80 hard registers set, jumps out of and within the loop. */
82 scan_loop (hwloop_info loop
)
90 if (REGNO_REG_SET_P (df_get_live_in (loop
->successor
),
91 REGNO (loop
->iter_reg
)))
92 loop
->iter_reg_used_outside
= true;
94 for (ix
= 0; VEC_iterate (basic_block
, loop
->blocks
, ix
, bb
); ix
++)
100 if (bb
!= loop
->tail
)
101 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
103 if (bb_in_loop_p (loop
, e
->dest
))
105 if (!(e
->flags
& EDGE_FALLTHRU
))
106 loop
->jumps_within
= true;
110 loop
->jumps_outof
= true;
112 gcc_assert (!REGNO_REG_SET_P (df_get_live_in (e
->dest
),
113 REGNO (loop
->iter_reg
)));
117 for (insn
= BB_HEAD (bb
);
118 insn
!= NEXT_INSN (BB_END (bb
));
119 insn
= NEXT_INSN (insn
))
122 HARD_REG_SET set_this_insn
;
124 if (!NONDEBUG_INSN_P (insn
))
127 if (recog_memoized (insn
) < 0
128 && (GET_CODE (PATTERN (insn
)) == ASM_INPUT
129 || asm_noperands (PATTERN (insn
)) >= 0))
130 loop
->has_asm
= true;
132 CLEAR_HARD_REG_SET (set_this_insn
);
133 for (def_rec
= DF_INSN_DEFS (insn
); *def_rec
; def_rec
++)
135 rtx dreg
= DF_REF_REG (*def_rec
);
140 add_to_hard_reg_set (&set_this_insn
, GET_MODE (dreg
),
144 if (insn
== loop
->loop_end
)
145 CLEAR_HARD_REG_BIT (set_this_insn
, REGNO (loop
->iter_reg
));
146 else if (reg_mentioned_p (loop
->iter_reg
, PATTERN (insn
)))
147 loop
->iter_reg_used
= true;
148 IOR_HARD_REG_SET (loop
->regs_set_in_loop
, set_this_insn
);
153 /* Compute the incoming_dest and incoming_src members of LOOP by
154 identifying the edges that jump into the loop. If there is more
155 than one block that jumps into the loop, incoming_src will be set
156 to NULL; likewise, if there is more than one block in the loop that
157 is the destination of an incoming edge, incoming_dest will be NULL.
159 Return true if either of these two fields is nonnull, false
162 process_incoming_edges (hwloop_info loop
)
168 FOR_EACH_EDGE (e
, ei
, loop
->incoming
)
172 loop
->incoming_src
= e
->src
;
173 loop
->incoming_dest
= e
->dest
;
178 if (e
->dest
!= loop
->incoming_dest
)
179 loop
->incoming_dest
= NULL
;
180 if (e
->src
!= loop
->incoming_src
)
181 loop
->incoming_src
= NULL
;
184 if (loop
->incoming_src
== NULL
&& loop
->incoming_dest
== NULL
)
190 /* Try to identify a forwarder block that jump into LOOP, and add it to
191 the set of blocks in the loop, updating the vector of incoming blocks as
192 well. This transformation gives a second chance to loops we couldn't
193 otherwise handle by increasing the chance that we'll end up with one
195 Return true if we made a change, false otherwise. */
197 add_forwarder_blocks (hwloop_info loop
)
202 FOR_EACH_EDGE (e
, ei
, loop
->incoming
)
204 if (forwarder_block_p (e
->src
))
211 ";; Adding forwarder block %d to loop %d and retrying\n",
212 e
->src
->index
, loop
->loop_no
);
213 VEC_safe_push (basic_block
, heap
, loop
->blocks
, e
->src
);
214 bitmap_set_bit (loop
->block_bitmap
, e
->src
->index
);
215 FOR_EACH_EDGE (e2
, ei2
, e
->src
->preds
)
216 VEC_safe_push (edge
, gc
, loop
->incoming
, e2
);
217 VEC_unordered_remove (edge
, loop
->incoming
, ei
.index
);
224 /* Called from reorg_loops when a potential loop end is found. LOOP is
225 a newly set up structure describing the loop, it is this function's
226 responsibility to fill most of it. TAIL_BB and TAIL_INSN point to the
227 loop_end insn and its enclosing basic block. REG is the loop counter
229 For our purposes, a loop is defined by the set of blocks reachable from
230 the loop head in which the loop counter register is live. This matches
231 the expected use; targets that call into this code usually replace the
232 loop counter with a different special register. */
234 discover_loop (hwloop_info loop
, basic_block tail_bb
, rtx tail_insn
, rtx reg
)
239 VEC (basic_block
,heap
) *works
;
241 loop
->tail
= tail_bb
;
242 loop
->loop_end
= tail_insn
;
243 loop
->iter_reg
= reg
;
244 loop
->incoming
= VEC_alloc (edge
, gc
, 2);
245 loop
->start_label
= JUMP_LABEL (tail_insn
);
247 if (EDGE_COUNT (tail_bb
->succs
) != 2)
252 loop
->head
= BRANCH_EDGE (tail_bb
)->dest
;
253 loop
->successor
= FALLTHRU_EDGE (tail_bb
)->dest
;
255 works
= VEC_alloc (basic_block
, heap
, 20);
256 VEC_safe_push (basic_block
, heap
, works
, loop
->head
);
259 for (dwork
= 0; VEC_iterate (basic_block
, works
, dwork
, bb
); dwork
++)
263 if (bb
== EXIT_BLOCK_PTR
)
265 /* We've reached the exit block. The loop must be bad. */
268 ";; Loop is bad - reached exit block while scanning\n");
273 if (bitmap_bit_p (loop
->block_bitmap
, bb
->index
))
276 /* We've not seen this block before. Add it to the loop's
277 list and then add each successor to the work list. */
279 VEC_safe_push (basic_block
, heap
, loop
->blocks
, bb
);
280 bitmap_set_bit (loop
->block_bitmap
, bb
->index
);
286 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
288 basic_block succ
= EDGE_SUCC (bb
, ei
.index
)->dest
;
289 if (REGNO_REG_SET_P (df_get_live_in (succ
),
290 REGNO (loop
->iter_reg
)))
291 VEC_safe_push (basic_block
, heap
, works
, succ
);
299 /* Find the predecessor, and make sure nothing else jumps into this loop. */
302 FOR_EACH_VEC_ELT (basic_block
, loop
->blocks
, dwork
, bb
)
306 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
308 basic_block pred
= e
->src
;
310 if (!bb_in_loop_p (loop
, pred
))
313 fprintf (dump_file
, ";; Loop %d: incoming edge %d -> %d\n",
314 loop
->loop_no
, pred
->index
,
316 VEC_safe_push (edge
, gc
, loop
->incoming
, e
);
321 if (!process_incoming_edges (loop
))
325 ";; retrying loop %d with forwarder blocks\n",
327 if (!add_forwarder_blocks (loop
))
330 fprintf (dump_file
, ";; No forwarder blocks found\n");
333 else if (!process_incoming_edges (loop
))
337 ";; can't find suitable entry for loop %d\n",
343 VEC_free (basic_block
, heap
, works
);
346 /* Analyze the structure of the loops in the current function. Use
347 STACK for bitmap allocations. Returns all the valid candidates for
348 hardware loops found in this function. HOOKS is the argument
349 passed to reorg_loops, used here to find the iteration registers
350 from a loop_end pattern. */
352 discover_loops (bitmap_obstack
*stack
, struct hw_doloop_hooks
*hooks
)
354 hwloop_info loops
= NULL
;
359 /* Find all the possible loop tails. This means searching for every
360 loop_end instruction. For each one found, create a hwloop_info
361 structure and add the head block to the work list. */
364 rtx tail
= BB_END (bb
);
367 while (tail
&& GET_CODE (tail
) == NOTE
&& tail
!= BB_HEAD (bb
))
368 tail
= PREV_INSN (tail
);
370 if (tail
== NULL_RTX
)
375 reg
= hooks
->end_pattern_reg (tail
);
379 /* A possible loop end */
381 /* There's a degenerate case we can handle - an empty loop consisting
382 of only a back branch. Handle that by deleting the branch. */
383 insn
= JUMP_LABEL (tail
);
384 while (insn
&& !NONDEBUG_INSN_P (insn
))
385 insn
= NEXT_INSN (insn
);
388 basic_block succ
= FALLTHRU_EDGE (bb
)->dest
;
391 fprintf (dump_file
, ";; degenerate loop ending at\n");
392 print_rtl_single (dump_file
, tail
);
394 if (!REGNO_REG_SET_P (df_get_live_in (succ
), REGNO (reg
)))
397 fprintf (dump_file
, ";; deleting it\n");
398 delete_insn_and_edges (tail
);
403 loop
= XCNEW (struct hwloop_info_d
);
406 loop
->loop_no
= nloops
++;
407 loop
->blocks
= VEC_alloc (basic_block
, heap
, 20);
408 loop
->block_bitmap
= BITMAP_ALLOC (stack
);
412 fprintf (dump_file
, ";; potential loop %d ending at\n",
414 print_rtl_single (dump_file
, tail
);
417 discover_loop (loop
, bb
, tail
, reg
);
420 /* Compute loop nestings. Given two loops A and B, either the sets
421 of their blocks don't intersect at all, or one is the subset of
422 the other, or the blocks don't form a good nesting structure. */
423 for (loop
= loops
; loop
; loop
= loop
->next
)
430 for (other
= loops
; other
; other
= other
->next
)
435 if (!bitmap_intersect_p (other
->block_bitmap
, loop
->block_bitmap
))
437 if (!bitmap_intersect_compl_p (other
->block_bitmap
,
439 VEC_safe_push (hwloop_info
, heap
, loop
->loops
, other
);
440 else if (!bitmap_intersect_compl_p (loop
->block_bitmap
,
441 other
->block_bitmap
))
442 VEC_safe_push (hwloop_info
, heap
, other
->loops
, loop
);
447 ";; can't find suitable nesting for loops %d and %d\n",
448 loop
->loop_no
, other
->loop_no
);
449 loop
->bad
= other
->bad
= true;
455 dump_hwloops (loops
);
460 /* Free up the loop structures in LOOPS. */
462 free_loops (hwloop_info loops
)
466 hwloop_info loop
= loops
;
468 VEC_free (hwloop_info
, heap
, loop
->loops
);
469 VEC_free (basic_block
, heap
, loop
->blocks
);
470 BITMAP_FREE (loop
->block_bitmap
);
475 #define BB_AUX_INDEX(BB) ((intptr_t) (BB)->aux)
477 /* Initialize the aux fields to give ascending indices to basic blocks. */
479 set_bb_indices (void)
486 bb
->aux
= (void *) index
++;
489 /* The taken-branch edge from the loop end can actually go forward.
490 If the target's hardware loop support requires that the loop end be
491 after the loop start, try to reorder a loop's basic blocks when we
493 This is not very aggressive; it only moves at most one block. It
494 does not introduce new branches into loops; it may remove them, or
495 it may switch fallthru/jump edges. */
497 reorder_loops (hwloop_info loops
)
502 cfg_layout_initialize (0);
506 for (loop
= loops
; loop
; loop
= loop
->next
)
514 if (BB_AUX_INDEX (loop
->head
) <= BB_AUX_INDEX (loop
->tail
))
517 FOR_EACH_EDGE (e
, ei
, loop
->head
->succs
)
519 if (bitmap_bit_p (loop
->block_bitmap
, e
->dest
->index
)
520 && BB_AUX_INDEX (e
->dest
) < BB_AUX_INDEX (loop
->tail
))
522 basic_block start_bb
= e
->dest
;
523 basic_block start_prev_bb
= start_bb
->prev_bb
;
526 fprintf (dump_file
, ";; Moving block %d before block %d\n",
527 loop
->head
->index
, start_bb
->index
);
528 loop
->head
->prev_bb
->next_bb
= loop
->head
->next_bb
;
529 loop
->head
->next_bb
->prev_bb
= loop
->head
->prev_bb
;
531 loop
->head
->prev_bb
= start_prev_bb
;
532 loop
->head
->next_bb
= start_bb
;
533 start_prev_bb
->next_bb
= start_bb
->prev_bb
= loop
->head
;
544 if (bb
->next_bb
!= EXIT_BLOCK_PTR
)
545 bb
->aux
= bb
->next_bb
;
549 cfg_layout_finalize ();
550 clear_aux_for_blocks ();
554 /* Call the OPT function for LOOP and all of its sub-loops. This is
555 done in a depth-first search; innermost loops are visited first.
556 OPTIMIZE and FAIL are the functions passed to reorg_loops by the
557 target's reorg pass. */
559 optimize_loop (hwloop_info loop
, struct hw_doloop_hooks
*hooks
)
573 fprintf (dump_file
, ";; loop %d bad when found\n", loop
->loop_no
);
577 /* Every loop contains in its list of inner loops every loop nested inside
578 it, even if there are intermediate loops. This works because we're doing
579 a depth-first search here and never visit a loop more than once.
580 Recursion depth is effectively limited by the number of available
581 hardware registers. */
582 for (ix
= 0; VEC_iterate (hwloop_info
, loop
->loops
, ix
, inner
); ix
++)
584 optimize_loop (inner
, hooks
);
586 if (!inner
->bad
&& inner_depth
< inner
->depth
)
587 inner_depth
= inner
->depth
;
588 /* The set of registers may be changed while optimizing the inner
590 IOR_HARD_REG_SET (loop
->regs_set_in_loop
, inner
->regs_set_in_loop
);
593 loop
->depth
= inner_depth
+ 1;
595 if (hooks
->opt (loop
))
600 fprintf (dump_file
, ";; loop %d is bad\n", loop
->loop_no
);
606 /* This function can be used from a port's machine_dependent_reorg to
607 find and analyze loops that end in loop_end instructions. It uses
608 a set of function pointers in HOOKS to call back into the
609 target-specific functions to perform the actual machine-specific
612 Such transformations typically involve additional set-up
613 instructions before the loop, to define loop bounds or set up a
614 special loop counter register.
616 DO_REORDER should be set to true if we should try to use the
617 reorder_loops function to ensure the loop end occurs after the loop
618 start. This is for use by targets where the loop hardware requires
621 HOOKS is used to pass in target specific hooks; see
624 reorg_loops (bool do_reorder
, struct hw_doloop_hooks
*hooks
)
626 hwloop_info loops
= NULL
;
628 bitmap_obstack stack
;
630 df_live_add_problem ();
631 df_live_set_all_dirty ();
634 bitmap_obstack_initialize (&stack
);
637 fprintf (dump_file
, ";; Find loops, first pass\n\n");
639 loops
= discover_loops (&stack
, hooks
);
643 reorder_loops (loops
);
647 fprintf (dump_file
, ";; Find loops, second pass\n\n");
649 loops
= discover_loops (&stack
, hooks
);
652 for (loop
= loops
; loop
; loop
= loop
->next
)
655 /* Now apply the optimizations. */
656 for (loop
= loops
; loop
; loop
= loop
->next
)
657 optimize_loop (loop
, hooks
);
661 fprintf (dump_file
, ";; After hardware loops optimization:\n\n");
662 dump_hwloops (loops
);
668 print_rtl (dump_file
, get_insns ());