1 /* Swing Modulo Scheduling implementation.
2 Copyright (C) 2004, 2005, 2006, 2007
3 Free Software Foundation, Inc.
4 Contributed by Ayal Zaks and Mustafa Hagog <zaks,mustafa@il.ibm.com>
6 This file is part of GCC.
8 GCC is free software; you can redistribute it and/or modify it under
9 the terms of the GNU General Public License as published by the Free
10 Software Foundation; either version 2, or (at your option) any later
13 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
14 WARRANTY; without even the implied warranty of MERCHANTABILITY or
15 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
18 You should have received a copy of the GNU General Public License
19 along with GCC; see the file COPYING. If not, write to the Free
20 Software Foundation, 51 Franklin Street, Fifth Floor, Boston, MA
26 #include "coretypes.h"
31 #include "hard-reg-set.h"
35 #include "insn-config.h"
36 #include "insn-attr.h"
40 #include "sched-int.h"
42 #include "cfglayout.h"
51 #include "tree-pass.h"
53 #ifdef INSN_SCHEDULING
55 /* This file contains the implementation of the Swing Modulo Scheduler,
56 described in the following references:
57 [1] J. Llosa, A. Gonzalez, E. Ayguade, M. Valero., and J. Eckhardt.
58 Lifetime--sensitive modulo scheduling in a production environment.
59 IEEE Trans. on Comps., 50(3), March 2001
60 [2] J. Llosa, A. Gonzalez, E. Ayguade, and M. Valero.
61 Swing Modulo Scheduling: A Lifetime Sensitive Approach.
62 PACT '96 , pages 80-87, October 1996 (Boston - Massachusetts - USA).
64 The basic structure is:
65 1. Build a data-dependence graph (DDG) for each loop.
66 2. Use the DDG to order the insns of a loop (not in topological order
67 necessarily, but rather) trying to place each insn after all its
68 predecessors _or_ after all its successors.
69 3. Compute MII: a lower bound on the number of cycles to schedule the loop.
70 4. Use the ordering to perform list-scheduling of the loop:
71 1. Set II = MII. We will try to schedule the loop within II cycles.
72 2. Try to schedule the insns one by one according to the ordering.
73 For each insn compute an interval of cycles by considering already-
74 scheduled preds and succs (and associated latencies); try to place
75 the insn in the cycles of this window checking for potential
76 resource conflicts (using the DFA interface).
77 Note: this is different from the cycle-scheduling of schedule_insns;
78 here the insns are not scheduled monotonically top-down (nor bottom-
80 3. If failed in scheduling all insns - bump II++ and try again, unless
81 II reaches an upper bound MaxII, in which case report failure.
82 5. If we succeeded in scheduling the loop within II cycles, we now
83 generate prolog and epilog, decrease the counter of the loop, and
84 perform modulo variable expansion for live ranges that span more than
85 II cycles (i.e. use register copies to prevent a def from overwriting
86 itself before reaching the use).
90 /* This page defines partial-schedule structures and functions for
93 typedef struct partial_schedule
*partial_schedule_ptr
;
94 typedef struct ps_insn
*ps_insn_ptr
;
96 /* The minimum (absolute) cycle that a node of ps was scheduled in. */
97 #define PS_MIN_CYCLE(ps) (((partial_schedule_ptr)(ps))->min_cycle)
99 /* The maximum (absolute) cycle that a node of ps was scheduled in. */
100 #define PS_MAX_CYCLE(ps) (((partial_schedule_ptr)(ps))->max_cycle)
102 /* Perform signed modulo, always returning a non-negative value. */
103 #define SMODULO(x,y) ((x) % (y) < 0 ? ((x) % (y) + (y)) : (x) % (y))
105 /* The number of different iterations the nodes in ps span, assuming
106 the stage boundaries are placed efficiently. */
107 #define PS_STAGE_COUNT(ps) ((PS_MAX_CYCLE (ps) - PS_MIN_CYCLE (ps) \
108 + 1 + (ps)->ii - 1) / (ps)->ii)
110 /* A single instruction in the partial schedule. */
113 /* The corresponding DDG_NODE. */
116 /* The (absolute) cycle in which the PS instruction is scheduled.
117 Same as SCHED_TIME (node). */
120 /* The next/prev PS_INSN in the same row. */
121 ps_insn_ptr next_in_row
,
124 /* The number of nodes in the same row that come after this node. */
128 /* Holds the partial schedule as an array of II rows. Each entry of the
129 array points to a linked list of PS_INSNs, which represents the
130 instructions that are scheduled for that row. */
131 struct partial_schedule
133 int ii
; /* Number of rows in the partial schedule. */
134 int history
; /* Threshold for conflict checking using DFA. */
136 /* rows[i] points to linked list of insns scheduled in row i (0<=i<ii). */
139 /* The earliest absolute cycle of an insn in the partial schedule. */
142 /* The latest absolute cycle of an insn in the partial schedule. */
145 ddg_ptr g
; /* The DDG of the insns in the partial schedule. */
148 /* We use this to record all the register replacements we do in
149 the kernel so we can undo SMS if it is not profitable. */
150 struct undo_replace_buff_elem
155 struct undo_replace_buff_elem
*next
;
160 static partial_schedule_ptr
create_partial_schedule (int ii
, ddg_ptr
, int history
);
161 static void free_partial_schedule (partial_schedule_ptr
);
162 static void reset_partial_schedule (partial_schedule_ptr
, int new_ii
);
163 void print_partial_schedule (partial_schedule_ptr
, FILE *);
164 static int kernel_number_of_cycles (rtx first_insn
, rtx last_insn
);
165 static ps_insn_ptr
ps_add_node_check_conflicts (partial_schedule_ptr
,
166 ddg_node_ptr node
, int cycle
,
167 sbitmap must_precede
,
168 sbitmap must_follow
);
169 static void rotate_partial_schedule (partial_schedule_ptr
, int);
170 void set_row_column_for_ps (partial_schedule_ptr
);
171 static bool ps_unschedule_node (partial_schedule_ptr
, ddg_node_ptr
);
174 /* This page defines constants and structures for the modulo scheduling
177 /* As in haifa-sched.c: */
178 /* issue_rate is the number of insns that can be scheduled in the same
179 machine cycle. It can be defined in the config/mach/mach.h file,
180 otherwise we set it to 1. */
182 static int issue_rate
;
184 static int sms_order_nodes (ddg_ptr
, int, int * result
);
185 static void set_node_sched_params (ddg_ptr
);
186 static partial_schedule_ptr
sms_schedule_by_order (ddg_ptr
, int, int, int *);
187 static void permute_partial_schedule (partial_schedule_ptr ps
, rtx last
);
188 static void generate_prolog_epilog (partial_schedule_ptr
,struct loop
* loop
, rtx
);
189 static void duplicate_insns_of_cycles (partial_schedule_ptr ps
,
190 int from_stage
, int to_stage
,
193 #define SCHED_ASAP(x) (((node_sched_params_ptr)(x)->aux.info)->asap)
194 #define SCHED_TIME(x) (((node_sched_params_ptr)(x)->aux.info)->time)
195 #define SCHED_FIRST_REG_MOVE(x) \
196 (((node_sched_params_ptr)(x)->aux.info)->first_reg_move)
197 #define SCHED_NREG_MOVES(x) \
198 (((node_sched_params_ptr)(x)->aux.info)->nreg_moves)
199 #define SCHED_ROW(x) (((node_sched_params_ptr)(x)->aux.info)->row)
200 #define SCHED_STAGE(x) (((node_sched_params_ptr)(x)->aux.info)->stage)
201 #define SCHED_COLUMN(x) (((node_sched_params_ptr)(x)->aux.info)->column)
203 /* The scheduling parameters held for each node. */
204 typedef struct node_sched_params
206 int asap
; /* A lower-bound on the absolute scheduling cycle. */
207 int time
; /* The absolute scheduling cycle (time >= asap). */
209 /* The following field (first_reg_move) is a pointer to the first
210 register-move instruction added to handle the modulo-variable-expansion
211 of the register defined by this node. This register-move copies the
212 original register defined by the node. */
215 /* The number of register-move instructions added, immediately preceding
219 int row
; /* Holds time % ii. */
220 int stage
; /* Holds time / ii. */
222 /* The column of a node inside the ps. If nodes u, v are on the same row,
223 u will precede v if column (u) < column (v). */
225 } *node_sched_params_ptr
;
228 /* The following three functions are copied from the current scheduler
229 code in order to use sched_analyze() for computing the dependencies.
230 They are used when initializing the sched_info structure. */
232 sms_print_insn (rtx insn
, int aligned ATTRIBUTE_UNUSED
)
236 sprintf (tmp
, "i%4d", INSN_UID (insn
));
241 compute_jump_reg_dependencies (rtx insn ATTRIBUTE_UNUSED
,
242 regset cond_exec ATTRIBUTE_UNUSED
,
243 regset used ATTRIBUTE_UNUSED
,
244 regset set ATTRIBUTE_UNUSED
)
248 static struct sched_info sms_sched_info
=
257 compute_jump_reg_dependencies
,
262 NULL
, NULL
, NULL
, NULL
, NULL
,
263 #ifdef ENABLE_CHECKING
270 /* Return the register decremented and tested in INSN,
271 or zero if it is not a decrement-and-branch insn. */
274 doloop_register_get (rtx insn ATTRIBUTE_UNUSED
)
276 #ifdef HAVE_doloop_end
277 rtx pattern
, reg
, condition
;
282 pattern
= PATTERN (insn
);
283 condition
= doloop_condition_get (pattern
);
287 if (REG_P (XEXP (condition
, 0)))
288 reg
= XEXP (condition
, 0);
289 else if (GET_CODE (XEXP (condition
, 0)) == PLUS
290 && REG_P (XEXP (XEXP (condition
, 0), 0)))
291 reg
= XEXP (XEXP (condition
, 0), 0);
301 /* Check if COUNT_REG is set to a constant in the PRE_HEADER block, so
302 that the number of iterations is a compile-time constant. If so,
303 return the rtx that sets COUNT_REG to a constant, and set COUNT to
304 this constant. Otherwise return 0. */
306 const_iteration_count (rtx count_reg
, basic_block pre_header
,
307 HOST_WIDEST_INT
* count
)
315 get_ebb_head_tail (pre_header
, pre_header
, &head
, &tail
);
317 for (insn
= tail
; insn
!= PREV_INSN (head
); insn
= PREV_INSN (insn
))
318 if (INSN_P (insn
) && single_set (insn
) &&
319 rtx_equal_p (count_reg
, SET_DEST (single_set (insn
))))
321 rtx pat
= single_set (insn
);
323 if (GET_CODE (SET_SRC (pat
)) == CONST_INT
)
325 *count
= INTVAL (SET_SRC (pat
));
335 /* A very simple resource-based lower bound on the initiation interval.
336 ??? Improve the accuracy of this bound by considering the
337 utilization of various units. */
341 return (g
->num_nodes
/ issue_rate
);
345 /* Points to the array that contains the sched data for each node. */
346 static node_sched_params_ptr node_sched_params
;
348 /* Allocate sched_params for each node and initialize it. Assumes that
349 the aux field of each node contain the asap bound (computed earlier),
350 and copies it into the sched_params field. */
352 set_node_sched_params (ddg_ptr g
)
356 /* Allocate for each node in the DDG a place to hold the "sched_data". */
357 /* Initialize ASAP/ALAP/HIGHT to zero. */
358 node_sched_params
= (node_sched_params_ptr
)
359 xcalloc (g
->num_nodes
,
360 sizeof (struct node_sched_params
));
362 /* Set the pointer of the general data of the node to point to the
363 appropriate sched_params structure. */
364 for (i
= 0; i
< g
->num_nodes
; i
++)
366 /* Watch out for aliasing problems? */
367 node_sched_params
[i
].asap
= g
->nodes
[i
].aux
.count
;
368 g
->nodes
[i
].aux
.info
= &node_sched_params
[i
];
373 print_node_sched_params (FILE * file
, int num_nodes
)
379 for (i
= 0; i
< num_nodes
; i
++)
381 node_sched_params_ptr nsp
= &node_sched_params
[i
];
382 rtx reg_move
= nsp
->first_reg_move
;
385 fprintf (file
, "Node %d:\n", i
);
386 fprintf (file
, " asap = %d:\n", nsp
->asap
);
387 fprintf (file
, " time = %d:\n", nsp
->time
);
388 fprintf (file
, " nreg_moves = %d:\n", nsp
->nreg_moves
);
389 for (j
= 0; j
< nsp
->nreg_moves
; j
++)
391 fprintf (file
, " reg_move = ");
392 print_rtl_single (file
, reg_move
);
393 reg_move
= PREV_INSN (reg_move
);
398 /* Calculate an upper bound for II. SMS should not schedule the loop if it
399 requires more cycles than this bound. Currently set to the sum of the
400 longest latency edge for each node. Reset based on experiments. */
402 calculate_maxii (ddg_ptr g
)
407 for (i
= 0; i
< g
->num_nodes
; i
++)
409 ddg_node_ptr u
= &g
->nodes
[i
];
411 int max_edge_latency
= 0;
413 for (e
= u
->out
; e
; e
= e
->next_out
)
414 max_edge_latency
= MAX (max_edge_latency
, e
->latency
);
416 maxii
+= max_edge_latency
;
422 Breaking intra-loop register anti-dependences:
423 Each intra-loop register anti-dependence implies a cross-iteration true
424 dependence of distance 1. Therefore, we can remove such false dependencies
425 and figure out if the partial schedule broke them by checking if (for a
426 true-dependence of distance 1): SCHED_TIME (def) < SCHED_TIME (use) and
427 if so generate a register move. The number of such moves is equal to:
428 SCHED_TIME (use) - SCHED_TIME (def) { 0 broken
429 nreg_moves = ----------------------------------- + 1 - { dependence.
432 static struct undo_replace_buff_elem
*
433 generate_reg_moves (partial_schedule_ptr ps
)
438 struct undo_replace_buff_elem
*reg_move_replaces
= NULL
;
440 for (i
= 0; i
< g
->num_nodes
; i
++)
442 ddg_node_ptr u
= &g
->nodes
[i
];
444 int nreg_moves
= 0, i_reg_move
;
445 sbitmap
*uses_of_defs
;
447 rtx prev_reg
, old_reg
;
449 /* Compute the number of reg_moves needed for u, by looking at life
450 ranges started at u (excluding self-loops). */
451 for (e
= u
->out
; e
; e
= e
->next_out
)
452 if (e
->type
== TRUE_DEP
&& e
->dest
!= e
->src
)
454 int nreg_moves4e
= (SCHED_TIME (e
->dest
) - SCHED_TIME (e
->src
)) / ii
;
456 if (e
->distance
== 1)
457 nreg_moves4e
= (SCHED_TIME (e
->dest
) - SCHED_TIME (e
->src
) + ii
) / ii
;
459 /* If dest precedes src in the schedule of the kernel, then dest
460 will read before src writes and we can save one reg_copy. */
461 if (SCHED_ROW (e
->dest
) == SCHED_ROW (e
->src
)
462 && SCHED_COLUMN (e
->dest
) < SCHED_COLUMN (e
->src
))
465 nreg_moves
= MAX (nreg_moves
, nreg_moves4e
);
471 /* Every use of the register defined by node may require a different
472 copy of this register, depending on the time the use is scheduled.
473 Set a bitmap vector, telling which nodes use each copy of this
475 uses_of_defs
= sbitmap_vector_alloc (nreg_moves
, g
->num_nodes
);
476 sbitmap_vector_zero (uses_of_defs
, nreg_moves
);
477 for (e
= u
->out
; e
; e
= e
->next_out
)
478 if (e
->type
== TRUE_DEP
&& e
->dest
!= e
->src
)
480 int dest_copy
= (SCHED_TIME (e
->dest
) - SCHED_TIME (e
->src
)) / ii
;
482 if (e
->distance
== 1)
483 dest_copy
= (SCHED_TIME (e
->dest
) - SCHED_TIME (e
->src
) + ii
) / ii
;
485 if (SCHED_ROW (e
->dest
) == SCHED_ROW (e
->src
)
486 && SCHED_COLUMN (e
->dest
) < SCHED_COLUMN (e
->src
))
490 SET_BIT (uses_of_defs
[dest_copy
- 1], e
->dest
->cuid
);
493 /* Now generate the reg_moves, attaching relevant uses to them. */
494 SCHED_NREG_MOVES (u
) = nreg_moves
;
495 old_reg
= prev_reg
= copy_rtx (SET_DEST (single_set (u
->insn
)));
496 last_reg_move
= u
->insn
;
498 for (i_reg_move
= 0; i_reg_move
< nreg_moves
; i_reg_move
++)
500 unsigned int i_use
= 0;
501 rtx new_reg
= gen_reg_rtx (GET_MODE (prev_reg
));
502 rtx reg_move
= gen_move_insn (new_reg
, prev_reg
);
503 sbitmap_iterator sbi
;
505 add_insn_before (reg_move
, last_reg_move
);
506 last_reg_move
= reg_move
;
508 if (!SCHED_FIRST_REG_MOVE (u
))
509 SCHED_FIRST_REG_MOVE (u
) = reg_move
;
511 EXECUTE_IF_SET_IN_SBITMAP (uses_of_defs
[i_reg_move
], 0, i_use
, sbi
)
513 struct undo_replace_buff_elem
*rep
;
515 rep
= (struct undo_replace_buff_elem
*)
516 xcalloc (1, sizeof (struct undo_replace_buff_elem
));
517 rep
->insn
= g
->nodes
[i_use
].insn
;
518 rep
->orig_reg
= old_reg
;
519 rep
->new_reg
= new_reg
;
521 if (! reg_move_replaces
)
522 reg_move_replaces
= rep
;
525 rep
->next
= reg_move_replaces
;
526 reg_move_replaces
= rep
;
529 replace_rtx (g
->nodes
[i_use
].insn
, old_reg
, new_reg
);
534 sbitmap_vector_free (uses_of_defs
);
536 return reg_move_replaces
;
539 /* We call this when we want to undo the SMS schedule for a given loop.
540 One of the things that we do is to delete the register moves generated
541 for the sake of SMS; this function deletes the register move instructions
542 recorded in the undo buffer. */
544 undo_generate_reg_moves (partial_schedule_ptr ps
,
545 struct undo_replace_buff_elem
*reg_move_replaces
)
549 for (i
= 0; i
< ps
->g
->num_nodes
; i
++)
551 ddg_node_ptr u
= &ps
->g
->nodes
[i
];
553 rtx crr
= SCHED_FIRST_REG_MOVE (u
);
555 for (j
= 0; j
< SCHED_NREG_MOVES (u
); j
++)
557 prev
= PREV_INSN (crr
);
561 SCHED_FIRST_REG_MOVE (u
) = NULL_RTX
;
564 while (reg_move_replaces
)
566 struct undo_replace_buff_elem
*rep
= reg_move_replaces
;
568 reg_move_replaces
= reg_move_replaces
->next
;
569 replace_rtx (rep
->insn
, rep
->new_reg
, rep
->orig_reg
);
573 /* Free memory allocated for the undo buffer. */
575 free_undo_replace_buff (struct undo_replace_buff_elem
*reg_move_replaces
)
578 while (reg_move_replaces
)
580 struct undo_replace_buff_elem
*rep
= reg_move_replaces
;
582 reg_move_replaces
= reg_move_replaces
->next
;
587 /* Bump the SCHED_TIMEs of all nodes to start from zero. Set the values
588 of SCHED_ROW and SCHED_STAGE. */
590 normalize_sched_times (partial_schedule_ptr ps
)
594 int amount
= PS_MIN_CYCLE (ps
);
597 /* Don't include the closing branch assuming that it is the last node. */
598 for (i
= 0; i
< g
->num_nodes
- 1; i
++)
600 ddg_node_ptr u
= &g
->nodes
[i
];
601 int normalized_time
= SCHED_TIME (u
) - amount
;
603 gcc_assert (normalized_time
>= 0);
605 SCHED_TIME (u
) = normalized_time
;
606 SCHED_ROW (u
) = normalized_time
% ii
;
607 SCHED_STAGE (u
) = normalized_time
/ ii
;
611 /* Set SCHED_COLUMN of each node according to its position in PS. */
613 set_columns_for_ps (partial_schedule_ptr ps
)
617 for (row
= 0; row
< ps
->ii
; row
++)
619 ps_insn_ptr cur_insn
= ps
->rows
[row
];
622 for (; cur_insn
; cur_insn
= cur_insn
->next_in_row
)
623 SCHED_COLUMN (cur_insn
->node
) = column
++;
627 /* Permute the insns according to their order in PS, from row 0 to
628 row ii-1, and position them right before LAST. This schedules
629 the insns of the loop kernel. */
631 permute_partial_schedule (partial_schedule_ptr ps
, rtx last
)
637 for (row
= 0; row
< ii
; row
++)
638 for (ps_ij
= ps
->rows
[row
]; ps_ij
; ps_ij
= ps_ij
->next_in_row
)
639 if (PREV_INSN (last
) != ps_ij
->node
->insn
)
640 reorder_insns_nobb (ps_ij
->node
->first_note
, ps_ij
->node
->insn
,
644 /* As part of undoing SMS we return to the original ordering of the
645 instructions inside the loop kernel. Given the partial schedule PS, this
646 function returns the ordering of the instruction according to their CUID
647 in the DDG (PS->G), which is the original order of the instruction before
650 undo_permute_partial_schedule (partial_schedule_ptr ps
, rtx last
)
654 for (i
= 0 ; i
< ps
->g
->num_nodes
; i
++)
655 if (last
== ps
->g
->nodes
[i
].insn
656 || last
== ps
->g
->nodes
[i
].first_note
)
658 else if (PREV_INSN (last
) != ps
->g
->nodes
[i
].insn
)
659 reorder_insns_nobb (ps
->g
->nodes
[i
].first_note
, ps
->g
->nodes
[i
].insn
,
663 /* Used to generate the prologue & epilogue. Duplicate the subset of
664 nodes whose stages are between FROM_STAGE and TO_STAGE (inclusive
665 of both), together with a prefix/suffix of their reg_moves. */
667 duplicate_insns_of_cycles (partial_schedule_ptr ps
, int from_stage
,
668 int to_stage
, int for_prolog
)
673 for (row
= 0; row
< ps
->ii
; row
++)
674 for (ps_ij
= ps
->rows
[row
]; ps_ij
; ps_ij
= ps_ij
->next_in_row
)
676 ddg_node_ptr u_node
= ps_ij
->node
;
678 rtx reg_move
= NULL_RTX
;
682 /* SCHED_STAGE (u_node) >= from_stage == 0. Generate increasing
683 number of reg_moves starting with the second occurrence of
684 u_node, which is generated if its SCHED_STAGE <= to_stage. */
685 i_reg_moves
= to_stage
- SCHED_STAGE (u_node
) + 1;
686 i_reg_moves
= MAX (i_reg_moves
, 0);
687 i_reg_moves
= MIN (i_reg_moves
, SCHED_NREG_MOVES (u_node
));
689 /* The reg_moves start from the *first* reg_move backwards. */
692 reg_move
= SCHED_FIRST_REG_MOVE (u_node
);
693 for (j
= 1; j
< i_reg_moves
; j
++)
694 reg_move
= PREV_INSN (reg_move
);
697 else /* It's for the epilog. */
699 /* SCHED_STAGE (u_node) <= to_stage. Generate all reg_moves,
700 starting to decrease one stage after u_node no longer occurs;
701 that is, generate all reg_moves until
702 SCHED_STAGE (u_node) == from_stage - 1. */
703 i_reg_moves
= SCHED_NREG_MOVES (u_node
)
704 - (from_stage
- SCHED_STAGE (u_node
) - 1);
705 i_reg_moves
= MAX (i_reg_moves
, 0);
706 i_reg_moves
= MIN (i_reg_moves
, SCHED_NREG_MOVES (u_node
));
708 /* The reg_moves start from the *last* reg_move forwards. */
711 reg_move
= SCHED_FIRST_REG_MOVE (u_node
);
712 for (j
= 1; j
< SCHED_NREG_MOVES (u_node
); j
++)
713 reg_move
= PREV_INSN (reg_move
);
717 for (j
= 0; j
< i_reg_moves
; j
++, reg_move
= NEXT_INSN (reg_move
))
718 emit_insn (copy_rtx (PATTERN (reg_move
)));
719 if (SCHED_STAGE (u_node
) >= from_stage
720 && SCHED_STAGE (u_node
) <= to_stage
)
721 duplicate_insn_chain (u_node
->first_note
, u_node
->insn
);
726 /* Generate the instructions (including reg_moves) for prolog & epilog. */
728 generate_prolog_epilog (partial_schedule_ptr ps
, struct loop
* loop
, rtx count_reg
)
731 int last_stage
= PS_STAGE_COUNT (ps
) - 1;
734 /* Generate the prolog, inserting its insns on the loop-entry edge. */
738 /* Generate a subtract instruction at the beginning of the prolog to
739 adjust the loop count by STAGE_COUNT. */
740 emit_insn (gen_sub2_insn (count_reg
, GEN_INT (last_stage
)));
742 for (i
= 0; i
< last_stage
; i
++)
743 duplicate_insns_of_cycles (ps
, 0, i
, 1);
745 /* Put the prolog on the entry edge. */
746 e
= loop_preheader_edge (loop
);
747 split_edge_and_insert (e
, get_insns ());
751 /* Generate the epilog, inserting its insns on the loop-exit edge. */
754 for (i
= 0; i
< last_stage
; i
++)
755 duplicate_insns_of_cycles (ps
, i
+ 1, last_stage
, 0);
757 /* Put the epilogue on the exit edge. */
758 gcc_assert (single_exit (loop
));
759 e
= single_exit (loop
);
760 split_edge_and_insert (e
, get_insns ());
764 /* Return true if all the BBs of the loop are empty except the
767 loop_single_full_bb_p (struct loop
*loop
)
770 basic_block
*bbs
= get_loop_body (loop
);
772 for (i
= 0; i
< loop
->num_nodes
; i
++)
775 bool empty_bb
= true;
777 if (bbs
[i
] == loop
->header
)
780 /* Make sure that basic blocks other than the header
781 have only notes labels or jumps. */
782 get_ebb_head_tail (bbs
[i
], bbs
[i
], &head
, &tail
);
783 for (; head
!= NEXT_INSN (tail
); head
= NEXT_INSN (head
))
785 if (NOTE_P (head
) || LABEL_P (head
)
786 || (INSN_P (head
) && JUMP_P (head
)))
802 /* A simple loop from SMS point of view; it is a loop that is composed of
803 either a single basic block or two BBs - a header and a latch. */
804 #define SIMPLE_SMS_LOOP_P(loop) ((loop->num_nodes < 3 ) \
805 && (EDGE_COUNT (loop->latch->preds) == 1) \
806 && (EDGE_COUNT (loop->latch->succs) == 1))
808 /* Return true if the loop is in its canonical form and false if not.
809 i.e. SIMPLE_SMS_LOOP_P and have one preheader block, and single exit. */
811 loop_canon_p (struct loop
*loop
)
814 if (loop
->inner
|| !loop_outer (loop
))
817 if (!single_exit (loop
))
821 rtx insn
= BB_END (loop
->header
);
823 fprintf (dump_file
, "SMS loop many exits ");
824 fprintf (dump_file
, " %s %d (file, line)\n",
825 insn_file (insn
), insn_line (insn
));
830 if (! SIMPLE_SMS_LOOP_P (loop
) && ! loop_single_full_bb_p (loop
))
834 rtx insn
= BB_END (loop
->header
);
836 fprintf (dump_file
, "SMS loop many BBs. ");
837 fprintf (dump_file
, " %s %d (file, line)\n",
838 insn_file (insn
), insn_line (insn
));
846 /* If there are more than one entry for the loop,
847 make it one by splitting the first entry edge and
848 redirecting the others to the new BB. */
850 canon_loop (struct loop
*loop
)
855 /* Avoid annoying special cases of edges going to exit
857 FOR_EACH_EDGE (e
, i
, EXIT_BLOCK_PTR
->preds
)
858 if ((e
->flags
& EDGE_FALLTHRU
) && (EDGE_COUNT (e
->src
->succs
) > 1))
861 if (loop
->latch
== loop
->header
862 || EDGE_COUNT (loop
->latch
->succs
) > 1)
864 FOR_EACH_EDGE (e
, i
, loop
->header
->preds
)
865 if (e
->src
== loop
->latch
)
871 /* Probability in % that the sms-ed loop rolls enough so that optimized
872 version may be entered. Just a guess. */
873 #define PROB_SMS_ENOUGH_ITERATIONS 80
875 /* Main entry point, perform SMS scheduling on the loops of the function
876 that consist of single basic blocks. */
880 static int passes
= 0;
886 partial_schedule_ptr ps
;
888 basic_block bb
= NULL
;
890 basic_block condition_bb
= NULL
;
892 gcov_type trip_count
= 0;
894 loop_optimizer_init (LOOPS_HAVE_PREHEADERS
895 | LOOPS_HAVE_RECORDED_EXITS
);
896 if (number_of_loops () <= 1)
898 loop_optimizer_finalize ();
899 return; /* There are no loops to schedule. */
902 /* Initialize issue_rate. */
903 if (targetm
.sched
.issue_rate
)
905 int temp
= reload_completed
;
907 reload_completed
= 1;
908 issue_rate
= targetm
.sched
.issue_rate ();
909 reload_completed
= temp
;
914 /* Initialize the scheduler. */
915 current_sched_info
= &sms_sched_info
;
918 /* Init Data Flow analysis, to be used in interloop dep calculation. */
919 df
= df_init (DF_HARD_REGS
| DF_EQUIV_NOTES
| DF_SUBREGS
);
920 df_rd_add_problem (df
, 0);
921 df_ru_add_problem (df
, 0);
922 df_chain_add_problem (df
, DF_DU_CHAIN
| DF_UD_CHAIN
);
926 df_dump (df
, dump_file
);
928 /* Allocate memory to hold the DDG array one entry for each loop.
929 We use loop->num as index into this array. */
930 g_arr
= XCNEWVEC (ddg_ptr
, number_of_loops ());
932 /* Build DDGs for all the relevant loops and hold them in G_ARR
933 indexed by the loop index. */
934 FOR_EACH_LOOP (li
, loop
, 0)
940 if ((passes
++ > MAX_SMS_LOOP_NUMBER
) && (MAX_SMS_LOOP_NUMBER
!= -1))
943 fprintf (dump_file
, "SMS reached MAX_PASSES... \n");
948 if (! loop_canon_p (loop
))
951 if (! loop_single_full_bb_p (loop
))
956 get_ebb_head_tail (bb
, bb
, &head
, &tail
);
957 latch_edge
= loop_latch_edge (loop
);
958 gcc_assert (single_exit (loop
));
959 if (single_exit (loop
)->count
)
960 trip_count
= latch_edge
->count
/ single_exit (loop
)->count
;
962 /* Perfrom SMS only on loops that their average count is above threshold. */
964 if ( latch_edge
->count
965 && (latch_edge
->count
< single_exit (loop
)->count
* SMS_LOOP_AVERAGE_COUNT_THRESHOLD
))
969 fprintf (dump_file
, " %s %d (file, line)\n",
970 insn_file (tail
), insn_line (tail
));
971 fprintf (dump_file
, "SMS single-bb-loop\n");
972 if (profile_info
&& flag_branch_probabilities
)
974 fprintf (dump_file
, "SMS loop-count ");
975 fprintf (dump_file
, HOST_WIDEST_INT_PRINT_DEC
,
976 (HOST_WIDEST_INT
) bb
->count
);
977 fprintf (dump_file
, "\n");
978 fprintf (dump_file
, "SMS trip-count ");
979 fprintf (dump_file
, HOST_WIDEST_INT_PRINT_DEC
,
980 (HOST_WIDEST_INT
) trip_count
);
981 fprintf (dump_file
, "\n");
982 fprintf (dump_file
, "SMS profile-sum-max ");
983 fprintf (dump_file
, HOST_WIDEST_INT_PRINT_DEC
,
984 (HOST_WIDEST_INT
) profile_info
->sum_max
);
985 fprintf (dump_file
, "\n");
991 /* Make sure this is a doloop. */
992 if ( !(count_reg
= doloop_register_get (tail
)))
995 /* Don't handle BBs with calls or barriers, or !single_set insns. */
996 for (insn
= head
; insn
!= NEXT_INSN (tail
); insn
= NEXT_INSN (insn
))
999 || (INSN_P (insn
) && !JUMP_P (insn
)
1000 && !single_set (insn
) && GET_CODE (PATTERN (insn
)) != USE
))
1003 if (insn
!= NEXT_INSN (tail
))
1008 fprintf (dump_file
, "SMS loop-with-call\n");
1009 else if (BARRIER_P (insn
))
1010 fprintf (dump_file
, "SMS loop-with-barrier\n");
1012 fprintf (dump_file
, "SMS loop-with-not-single-set\n");
1013 print_rtl_single (dump_file
, insn
);
1019 if (! (g
= create_ddg (bb
, df
, 0)))
1022 fprintf (dump_file
, "SMS doloop\n");
1026 g_arr
[loop
->num
] = g
;
1029 /* Release Data Flow analysis data structures. */
1033 /* We don't want to perform SMS on new loops - created by versioning. */
1034 FOR_EACH_LOOP (li
, loop
, 0)
1037 rtx count_reg
, count_init
;
1039 unsigned stage_count
= 0;
1040 HOST_WIDEST_INT loop_count
= 0;
1042 if (! (g
= g_arr
[loop
->num
]))
1046 print_ddg (dump_file
, g
);
1048 get_ebb_head_tail (loop
->header
, loop
->header
, &head
, &tail
);
1050 latch_edge
= loop_latch_edge (loop
);
1051 gcc_assert (single_exit (loop
));
1052 if (single_exit (loop
)->count
)
1053 trip_count
= latch_edge
->count
/ single_exit (loop
)->count
;
1057 fprintf (dump_file
, " %s %d (file, line)\n",
1058 insn_file (tail
), insn_line (tail
));
1059 fprintf (dump_file
, "SMS single-bb-loop\n");
1060 if (profile_info
&& flag_branch_probabilities
)
1062 fprintf (dump_file
, "SMS loop-count ");
1063 fprintf (dump_file
, HOST_WIDEST_INT_PRINT_DEC
,
1064 (HOST_WIDEST_INT
) bb
->count
);
1065 fprintf (dump_file
, "\n");
1066 fprintf (dump_file
, "SMS profile-sum-max ");
1067 fprintf (dump_file
, HOST_WIDEST_INT_PRINT_DEC
,
1068 (HOST_WIDEST_INT
) profile_info
->sum_max
);
1069 fprintf (dump_file
, "\n");
1071 fprintf (dump_file
, "SMS doloop\n");
1072 fprintf (dump_file
, "SMS built-ddg %d\n", g
->num_nodes
);
1073 fprintf (dump_file
, "SMS num-loads %d\n", g
->num_loads
);
1074 fprintf (dump_file
, "SMS num-stores %d\n", g
->num_stores
);
1078 /* In case of th loop have doloop register it gets special
1080 count_init
= NULL_RTX
;
1081 if ((count_reg
= doloop_register_get (tail
)))
1083 basic_block pre_header
;
1085 pre_header
= loop_preheader_edge (loop
)->src
;
1086 count_init
= const_iteration_count (count_reg
, pre_header
,
1089 gcc_assert (count_reg
);
1091 if (dump_file
&& count_init
)
1093 fprintf (dump_file
, "SMS const-doloop ");
1094 fprintf (dump_file
, HOST_WIDEST_INT_PRINT_DEC
,
1096 fprintf (dump_file
, "\n");
1099 node_order
= XNEWVEC (int, g
->num_nodes
);
1101 mii
= 1; /* Need to pass some estimate of mii. */
1102 rec_mii
= sms_order_nodes (g
, mii
, node_order
);
1103 mii
= MAX (res_MII (g
), rec_mii
);
1104 maxii
= (calculate_maxii (g
) * SMS_MAX_II_FACTOR
) / 100;
1107 fprintf (dump_file
, "SMS iis %d %d %d (rec_mii, mii, maxii)\n",
1108 rec_mii
, mii
, maxii
);
1110 /* After sms_order_nodes and before sms_schedule_by_order, to copy over
1112 set_node_sched_params (g
);
1114 ps
= sms_schedule_by_order (g
, mii
, maxii
, node_order
);
1117 stage_count
= PS_STAGE_COUNT (ps
);
1119 /* Stage count of 1 means that there is no interleaving between
1120 iterations, let the scheduling passes do the job. */
1122 || (count_init
&& (loop_count
<= stage_count
))
1123 || (flag_branch_probabilities
&& (trip_count
<= stage_count
)))
1127 fprintf (dump_file
, "SMS failed... \n");
1128 fprintf (dump_file
, "SMS sched-failed (stage-count=%d, loop-count=", stage_count
);
1129 fprintf (dump_file
, HOST_WIDEST_INT_PRINT_DEC
, loop_count
);
1130 fprintf (dump_file
, ", trip-count=");
1131 fprintf (dump_file
, HOST_WIDEST_INT_PRINT_DEC
, trip_count
);
1132 fprintf (dump_file
, ")\n");
1138 int orig_cycles
= kernel_number_of_cycles (BB_HEAD (g
->bb
), BB_END (g
->bb
));
1140 struct undo_replace_buff_elem
*reg_move_replaces
;
1145 "SMS succeeded %d %d (with ii, sc)\n", ps
->ii
,
1147 print_partial_schedule (ps
, dump_file
);
1149 "SMS Branch (%d) will later be scheduled at cycle %d.\n",
1150 g
->closing_branch
->cuid
, PS_MIN_CYCLE (ps
) - 1);
1153 /* Set the stage boundaries. If the DDG is built with closing_branch_deps,
1154 the closing_branch was scheduled and should appear in the last (ii-1)
1155 row. Otherwise, we are free to schedule the branch, and we let nodes
1156 that were scheduled at the first PS_MIN_CYCLE cycle appear in the first
1157 row; this should reduce stage_count to minimum. */
1158 normalize_sched_times (ps
);
1159 rotate_partial_schedule (ps
, PS_MIN_CYCLE (ps
));
1160 set_columns_for_ps (ps
);
1162 /* Generate the kernel just to be able to measure its cycles. */
1163 permute_partial_schedule (ps
, g
->closing_branch
->first_note
);
1164 reg_move_replaces
= generate_reg_moves (ps
);
1166 /* Get the number of cycles the new kernel expect to execute in. */
1167 new_cycles
= kernel_number_of_cycles (BB_HEAD (g
->bb
), BB_END (g
->bb
));
1169 /* Get back to the original loop so we can do loop versioning. */
1170 undo_permute_partial_schedule (ps
, g
->closing_branch
->first_note
);
1171 if (reg_move_replaces
)
1172 undo_generate_reg_moves (ps
, reg_move_replaces
);
1174 if ( new_cycles
>= orig_cycles
)
1176 /* SMS is not profitable so undo the permutation and reg move generation
1177 and return the kernel to its original state. */
1179 fprintf (dump_file
, "Undoing SMS because it is not profitable.\n");
1186 /* case the BCT count is not known , Do loop-versioning */
1187 if (count_reg
&& ! count_init
)
1189 rtx comp_rtx
= gen_rtx_fmt_ee (GT
, VOIDmode
, count_reg
,
1190 GEN_INT(stage_count
));
1191 unsigned prob
= (PROB_SMS_ENOUGH_ITERATIONS
1192 * REG_BR_PROB_BASE
) / 100;
1194 loop_version (loop
, comp_rtx
, &condition_bb
,
1195 prob
, prob
, REG_BR_PROB_BASE
- prob
,
1199 /* Set new iteration count of loop kernel. */
1200 if (count_reg
&& count_init
)
1201 SET_SRC (single_set (count_init
)) = GEN_INT (loop_count
1204 /* Now apply the scheduled kernel to the RTL of the loop. */
1205 permute_partial_schedule (ps
, g
->closing_branch
->first_note
);
1207 /* Mark this loop as software pipelined so the later
1208 scheduling passes doesn't touch it. */
1209 if (! flag_resched_modulo_sched
)
1210 g
->bb
->flags
|= BB_DISABLE_SCHEDULE
;
1211 /* The life-info is not valid any more. */
1212 g
->bb
->flags
|= BB_DIRTY
;
1214 reg_move_replaces
= generate_reg_moves (ps
);
1216 print_node_sched_params (dump_file
, g
->num_nodes
);
1217 /* Generate prolog and epilog. */
1218 if (count_reg
&& !count_init
)
1219 generate_prolog_epilog (ps
, loop
, count_reg
);
1221 generate_prolog_epilog (ps
, loop
, NULL_RTX
);
1223 free_undo_replace_buff (reg_move_replaces
);
1226 free_partial_schedule (ps
);
1227 free (node_sched_params
);
1234 /* Release scheduler data, needed until now because of DFA. */
1236 loop_optimizer_finalize ();
1239 /* The SMS scheduling algorithm itself
1240 -----------------------------------
1241 Input: 'O' an ordered list of insns of a loop.
1242 Output: A scheduling of the loop - kernel, prolog, and epilogue.
1244 'Q' is the empty Set
1245 'PS' is the partial schedule; it holds the currently scheduled nodes with
1247 'PSP' previously scheduled predecessors.
1248 'PSS' previously scheduled successors.
1249 't(u)' the cycle where u is scheduled.
1250 'l(u)' is the latency of u.
1251 'd(v,u)' is the dependence distance from v to u.
1252 'ASAP(u)' the earliest time at which u could be scheduled as computed in
1253 the node ordering phase.
1254 'check_hardware_resources_conflicts(u, PS, c)'
1255 run a trace around cycle/slot through DFA model
1256 to check resource conflicts involving instruction u
1257 at cycle c given the partial schedule PS.
1258 'add_to_partial_schedule_at_time(u, PS, c)'
1259 Add the node/instruction u to the partial schedule
1261 'calculate_register_pressure(PS)'
1262 Given a schedule of instructions, calculate the register
1263 pressure it implies. One implementation could be the
1264 maximum number of overlapping live ranges.
1265 'maxRP' The maximum allowed register pressure, it is usually derived from the number
1266 registers available in the hardware.
1270 3. for each node u in O in pre-computed order
1271 4. if (PSP(u) != Q && PSS(u) == Q) then
1272 5. Early_start(u) = max ( t(v) + l(v) - d(v,u)*II ) over all every v in PSP(u).
1273 6. start = Early_start; end = Early_start + II - 1; step = 1
1274 11. else if (PSP(u) == Q && PSS(u) != Q) then
1275 12. Late_start(u) = min ( t(v) - l(v) + d(v,u)*II ) over all every v in PSS(u).
1276 13. start = Late_start; end = Late_start - II + 1; step = -1
1277 14. else if (PSP(u) != Q && PSS(u) != Q) then
1278 15. Early_start(u) = max ( t(v) + l(v) - d(v,u)*II ) over all every v in PSP(u).
1279 16. Late_start(u) = min ( t(v) - l(v) + d(v,u)*II ) over all every v in PSS(u).
1280 17. start = Early_start;
1281 18. end = min(Early_start + II - 1 , Late_start);
1283 20. else "if (PSP(u) == Q && PSS(u) == Q)"
1284 21. start = ASAP(u); end = start + II - 1; step = 1
1288 24. for (c = start ; c != end ; c += step)
1289 25. if check_hardware_resources_conflicts(u, PS, c) then
1290 26. add_to_partial_schedule_at_time(u, PS, c)
1295 31. if (success == false) then
1297 33. if (II > maxII) then
1298 34. finish - failed to schedule
1303 39. if (calculate_register_pressure(PS) > maxRP) then
1306 42. compute epilogue & prologue
1307 43. finish - succeeded to schedule
1310 /* A limit on the number of cycles that resource conflicts can span. ??? Should
1311 be provided by DFA, and be dependent on the type of insn scheduled. Currently
1312 set to 0 to save compile time. */
1313 #define DFA_HISTORY SMS_DFA_HISTORY
1315 /* Given the partial schedule PS, this function calculates and returns the
1316 cycles in which we can schedule the node with the given index I.
1317 NOTE: Here we do the backtracking in SMS, in some special cases. We have
1318 noticed that there are several cases in which we fail to SMS the loop
1319 because the sched window of a node is empty due to tight data-deps. In
1320 such cases we want to unschedule some of the predecessors/successors
1321 until we get non-empty scheduling window. It returns -1 if the
1322 scheduling window is empty and zero otherwise. */
1325 get_sched_window (partial_schedule_ptr ps
, int *nodes_order
, int i
,
1326 sbitmap sched_nodes
, int ii
, int *start_p
, int *step_p
, int *end_p
)
1328 int start
, step
, end
;
1330 int u
= nodes_order
[i
];
1331 ddg_node_ptr u_node
= &ps
->g
->nodes
[u
];
1332 sbitmap psp
= sbitmap_alloc (ps
->g
->num_nodes
);
1333 sbitmap pss
= sbitmap_alloc (ps
->g
->num_nodes
);
1334 sbitmap u_node_preds
= NODE_PREDECESSORS (u_node
);
1335 sbitmap u_node_succs
= NODE_SUCCESSORS (u_node
);
1339 /* 1. compute sched window for u (start, end, step). */
1342 psp_not_empty
= sbitmap_a_and_b_cg (psp
, u_node_preds
, sched_nodes
);
1343 pss_not_empty
= sbitmap_a_and_b_cg (pss
, u_node_succs
, sched_nodes
);
1345 if (psp_not_empty
&& !pss_not_empty
)
1347 int early_start
= INT_MIN
;
1350 for (e
= u_node
->in
; e
!= 0; e
= e
->next_in
)
1352 ddg_node_ptr v_node
= e
->src
;
1353 if (TEST_BIT (sched_nodes
, v_node
->cuid
))
1355 int node_st
= SCHED_TIME (v_node
)
1356 + e
->latency
- (e
->distance
* ii
);
1358 early_start
= MAX (early_start
, node_st
);
1360 if (e
->data_type
== MEM_DEP
)
1361 end
= MIN (end
, SCHED_TIME (v_node
) + ii
- 1);
1364 start
= early_start
;
1365 end
= MIN (end
, early_start
+ ii
);
1369 else if (!psp_not_empty
&& pss_not_empty
)
1371 int late_start
= INT_MAX
;
1374 for (e
= u_node
->out
; e
!= 0; e
= e
->next_out
)
1376 ddg_node_ptr v_node
= e
->dest
;
1377 if (TEST_BIT (sched_nodes
, v_node
->cuid
))
1379 late_start
= MIN (late_start
,
1380 SCHED_TIME (v_node
) - e
->latency
1381 + (e
->distance
* ii
));
1382 if (e
->data_type
== MEM_DEP
)
1383 end
= MAX (end
, SCHED_TIME (v_node
) - ii
+ 1);
1387 end
= MAX (end
, late_start
- ii
);
1391 else if (psp_not_empty
&& pss_not_empty
)
1393 int early_start
= INT_MIN
;
1394 int late_start
= INT_MAX
;
1398 for (e
= u_node
->in
; e
!= 0; e
= e
->next_in
)
1400 ddg_node_ptr v_node
= e
->src
;
1402 if (TEST_BIT (sched_nodes
, v_node
->cuid
))
1404 early_start
= MAX (early_start
,
1405 SCHED_TIME (v_node
) + e
->latency
1406 - (e
->distance
* ii
));
1407 if (e
->data_type
== MEM_DEP
)
1408 end
= MIN (end
, SCHED_TIME (v_node
) + ii
- 1);
1411 for (e
= u_node
->out
; e
!= 0; e
= e
->next_out
)
1413 ddg_node_ptr v_node
= e
->dest
;
1415 if (TEST_BIT (sched_nodes
, v_node
->cuid
))
1417 late_start
= MIN (late_start
,
1418 SCHED_TIME (v_node
) - e
->latency
1419 + (e
->distance
* ii
));
1420 if (e
->data_type
== MEM_DEP
)
1421 start
= MAX (start
, SCHED_TIME (v_node
) - ii
+ 1);
1424 start
= MAX (start
, early_start
);
1425 end
= MIN (end
, MIN (early_start
+ ii
, late_start
+ 1));
1428 else /* psp is empty && pss is empty. */
1430 start
= SCHED_ASAP (u_node
);
1441 if ((start
>= end
&& step
== 1) || (start
<= end
&& step
== -1))
1447 /* This function implements the scheduling algorithm for SMS according to the
1449 static partial_schedule_ptr
1450 sms_schedule_by_order (ddg_ptr g
, int mii
, int maxii
, int *nodes_order
)
1454 int try_again_with_larger_ii
= true;
1455 int num_nodes
= g
->num_nodes
;
1457 int start
, end
, step
; /* Place together into one struct? */
1458 sbitmap sched_nodes
= sbitmap_alloc (num_nodes
);
1459 sbitmap must_precede
= sbitmap_alloc (num_nodes
);
1460 sbitmap must_follow
= sbitmap_alloc (num_nodes
);
1461 sbitmap tobe_scheduled
= sbitmap_alloc (num_nodes
);
1463 partial_schedule_ptr ps
= create_partial_schedule (ii
, g
, DFA_HISTORY
);
1465 sbitmap_ones (tobe_scheduled
);
1466 sbitmap_zero (sched_nodes
);
1468 while ((! sbitmap_equal (tobe_scheduled
, sched_nodes
)
1469 || try_again_with_larger_ii
) && ii
< maxii
)
1472 bool unscheduled_nodes
= false;
1475 fprintf (dump_file
, "Starting with ii=%d\n", ii
);
1476 if (try_again_with_larger_ii
)
1478 try_again_with_larger_ii
= false;
1479 sbitmap_zero (sched_nodes
);
1482 for (i
= 0; i
< num_nodes
; i
++)
1484 int u
= nodes_order
[i
];
1485 ddg_node_ptr u_node
= &ps
->g
->nodes
[u
];
1486 rtx insn
= u_node
->insn
;
1490 RESET_BIT (tobe_scheduled
, u
);
1494 if (JUMP_P (insn
)) /* Closing branch handled later. */
1496 RESET_BIT (tobe_scheduled
, u
);
1500 if (TEST_BIT (sched_nodes
, u
))
1503 /* Try to get non-empty scheduling window. */
1505 while (get_sched_window (ps
, nodes_order
, i
, sched_nodes
, ii
, &start
, &step
, &end
) < 0
1508 unscheduled_nodes
= true;
1509 if (TEST_BIT (NODE_PREDECESSORS (u_node
), nodes_order
[j
- 1])
1510 || TEST_BIT (NODE_SUCCESSORS (u_node
), nodes_order
[j
- 1]))
1512 ps_unschedule_node (ps
, &ps
->g
->nodes
[nodes_order
[j
- 1]]);
1513 RESET_BIT (sched_nodes
, nodes_order
[j
- 1]);
1519 /* ??? Try backtracking instead of immediately ii++? */
1521 try_again_with_larger_ii
= true;
1522 reset_partial_schedule (ps
, ii
);
1525 /* 2. Try scheduling u in window. */
1528 "Trying to schedule node %d in (%d .. %d) step %d\n",
1529 u
, start
, end
, step
);
1531 /* use must_follow & must_precede bitmaps to determine order
1532 of nodes within the cycle. */
1533 sbitmap_zero (must_precede
);
1534 sbitmap_zero (must_follow
);
1535 for (e
= u_node
->in
; e
!= 0; e
= e
->next_in
)
1536 if (TEST_BIT (sched_nodes
, e
->src
->cuid
)
1537 && e
->latency
== (ii
* e
->distance
)
1538 && start
== SCHED_TIME (e
->src
))
1539 SET_BIT (must_precede
, e
->src
->cuid
);
1541 for (e
= u_node
->out
; e
!= 0; e
= e
->next_out
)
1542 if (TEST_BIT (sched_nodes
, e
->dest
->cuid
)
1543 && e
->latency
== (ii
* e
->distance
)
1544 && end
== SCHED_TIME (e
->dest
))
1545 SET_BIT (must_follow
, e
->dest
->cuid
);
1548 if ((step
> 0 && start
< end
) || (step
< 0 && start
> end
))
1549 for (c
= start
; c
!= end
; c
+= step
)
1553 psi
= ps_add_node_check_conflicts (ps
, u_node
, c
,
1559 SCHED_TIME (u_node
) = c
;
1560 SET_BIT (sched_nodes
, u
);
1563 fprintf (dump_file
, "Schedule in %d\n", c
);
1569 /* ??? Try backtracking instead of immediately ii++? */
1571 try_again_with_larger_ii
= true;
1572 reset_partial_schedule (ps
, ii
);
1575 if (unscheduled_nodes
)
1578 /* ??? If (success), check register pressure estimates. */
1579 } /* Continue with next node. */
1580 } /* While try_again_with_larger_ii. */
1582 sbitmap_free (sched_nodes
);
1583 sbitmap_free (must_precede
);
1584 sbitmap_free (must_follow
);
1585 sbitmap_free (tobe_scheduled
);
1589 free_partial_schedule (ps
);
1596 /* This page implements the algorithm for ordering the nodes of a DDG
1597 for modulo scheduling, activated through the
1598 "int sms_order_nodes (ddg_ptr, int mii, int * result)" API. */
1600 #define ORDER_PARAMS(x) ((struct node_order_params *) (x)->aux.info)
1601 #define ASAP(x) (ORDER_PARAMS ((x))->asap)
1602 #define ALAP(x) (ORDER_PARAMS ((x))->alap)
1603 #define HEIGHT(x) (ORDER_PARAMS ((x))->height)
1604 #define MOB(x) (ALAP ((x)) - ASAP ((x)))
1605 #define DEPTH(x) (ASAP ((x)))
1607 typedef struct node_order_params
* nopa
;
1609 static void order_nodes_of_sccs (ddg_all_sccs_ptr
, int * result
);
1610 static int order_nodes_in_scc (ddg_ptr
, sbitmap
, sbitmap
, int*, int);
1611 static nopa
calculate_order_params (ddg_ptr
, int mii
);
1612 static int find_max_asap (ddg_ptr
, sbitmap
);
1613 static int find_max_hv_min_mob (ddg_ptr
, sbitmap
);
1614 static int find_max_dv_min_mob (ddg_ptr
, sbitmap
);
1616 enum sms_direction
{BOTTOMUP
, TOPDOWN
};
1618 struct node_order_params
1625 /* Check if NODE_ORDER contains a permutation of 0 .. NUM_NODES-1. */
1627 check_nodes_order (int *node_order
, int num_nodes
)
1630 sbitmap tmp
= sbitmap_alloc (num_nodes
);
1634 for (i
= 0; i
< num_nodes
; i
++)
1636 int u
= node_order
[i
];
1638 gcc_assert (u
< num_nodes
&& u
>= 0 && !TEST_BIT (tmp
, u
));
1646 /* Order the nodes of G for scheduling and pass the result in
1647 NODE_ORDER. Also set aux.count of each node to ASAP.
1648 Return the recMII for the given DDG. */
1650 sms_order_nodes (ddg_ptr g
, int mii
, int * node_order
)
1654 ddg_all_sccs_ptr sccs
= create_ddg_all_sccs (g
);
1656 nopa nops
= calculate_order_params (g
, mii
);
1658 order_nodes_of_sccs (sccs
, node_order
);
1660 if (sccs
->num_sccs
> 0)
1661 /* First SCC has the largest recurrence_length. */
1662 rec_mii
= sccs
->sccs
[0]->recurrence_length
;
1664 /* Save ASAP before destroying node_order_params. */
1665 for (i
= 0; i
< g
->num_nodes
; i
++)
1667 ddg_node_ptr v
= &g
->nodes
[i
];
1668 v
->aux
.count
= ASAP (v
);
1672 free_ddg_all_sccs (sccs
);
1673 check_nodes_order (node_order
, g
->num_nodes
);
1679 order_nodes_of_sccs (ddg_all_sccs_ptr all_sccs
, int * node_order
)
1682 ddg_ptr g
= all_sccs
->ddg
;
1683 int num_nodes
= g
->num_nodes
;
1684 sbitmap prev_sccs
= sbitmap_alloc (num_nodes
);
1685 sbitmap on_path
= sbitmap_alloc (num_nodes
);
1686 sbitmap tmp
= sbitmap_alloc (num_nodes
);
1687 sbitmap ones
= sbitmap_alloc (num_nodes
);
1689 sbitmap_zero (prev_sccs
);
1690 sbitmap_ones (ones
);
1692 /* Perfrom the node ordering starting from the SCC with the highest recMII.
1693 For each SCC order the nodes according to their ASAP/ALAP/HEIGHT etc. */
1694 for (i
= 0; i
< all_sccs
->num_sccs
; i
++)
1696 ddg_scc_ptr scc
= all_sccs
->sccs
[i
];
1698 /* Add nodes on paths from previous SCCs to the current SCC. */
1699 find_nodes_on_paths (on_path
, g
, prev_sccs
, scc
->nodes
);
1700 sbitmap_a_or_b (tmp
, scc
->nodes
, on_path
);
1702 /* Add nodes on paths from the current SCC to previous SCCs. */
1703 find_nodes_on_paths (on_path
, g
, scc
->nodes
, prev_sccs
);
1704 sbitmap_a_or_b (tmp
, tmp
, on_path
);
1706 /* Remove nodes of previous SCCs from current extended SCC. */
1707 sbitmap_difference (tmp
, tmp
, prev_sccs
);
1709 pos
= order_nodes_in_scc (g
, prev_sccs
, tmp
, node_order
, pos
);
1710 /* Above call to order_nodes_in_scc updated prev_sccs |= tmp. */
1713 /* Handle the remaining nodes that do not belong to any scc. Each call
1714 to order_nodes_in_scc handles a single connected component. */
1715 while (pos
< g
->num_nodes
)
1717 sbitmap_difference (tmp
, ones
, prev_sccs
);
1718 pos
= order_nodes_in_scc (g
, prev_sccs
, tmp
, node_order
, pos
);
1720 sbitmap_free (prev_sccs
);
1721 sbitmap_free (on_path
);
1723 sbitmap_free (ones
);
1726 /* MII is needed if we consider backarcs (that do not close recursive cycles). */
1727 static struct node_order_params
*
1728 calculate_order_params (ddg_ptr g
, int mii ATTRIBUTE_UNUSED
)
1732 int num_nodes
= g
->num_nodes
;
1734 /* Allocate a place to hold ordering params for each node in the DDG. */
1735 nopa node_order_params_arr
;
1737 /* Initialize of ASAP/ALAP/HEIGHT to zero. */
1738 node_order_params_arr
= (nopa
) xcalloc (num_nodes
,
1739 sizeof (struct node_order_params
));
1741 /* Set the aux pointer of each node to point to its order_params structure. */
1742 for (u
= 0; u
< num_nodes
; u
++)
1743 g
->nodes
[u
].aux
.info
= &node_order_params_arr
[u
];
1745 /* Disregarding a backarc from each recursive cycle to obtain a DAG,
1746 calculate ASAP, ALAP, mobility, distance, and height for each node
1747 in the dependence (direct acyclic) graph. */
1749 /* We assume that the nodes in the array are in topological order. */
1752 for (u
= 0; u
< num_nodes
; u
++)
1754 ddg_node_ptr u_node
= &g
->nodes
[u
];
1757 for (e
= u_node
->in
; e
; e
= e
->next_in
)
1758 if (e
->distance
== 0)
1759 ASAP (u_node
) = MAX (ASAP (u_node
),
1760 ASAP (e
->src
) + e
->latency
);
1761 max_asap
= MAX (max_asap
, ASAP (u_node
));
1764 for (u
= num_nodes
- 1; u
> -1; u
--)
1766 ddg_node_ptr u_node
= &g
->nodes
[u
];
1768 ALAP (u_node
) = max_asap
;
1769 HEIGHT (u_node
) = 0;
1770 for (e
= u_node
->out
; e
; e
= e
->next_out
)
1771 if (e
->distance
== 0)
1773 ALAP (u_node
) = MIN (ALAP (u_node
),
1774 ALAP (e
->dest
) - e
->latency
);
1775 HEIGHT (u_node
) = MAX (HEIGHT (u_node
),
1776 HEIGHT (e
->dest
) + e
->latency
);
1780 return node_order_params_arr
;
1784 find_max_asap (ddg_ptr g
, sbitmap nodes
)
1789 sbitmap_iterator sbi
;
1791 EXECUTE_IF_SET_IN_SBITMAP (nodes
, 0, u
, sbi
)
1793 ddg_node_ptr u_node
= &g
->nodes
[u
];
1795 if (max_asap
< ASAP (u_node
))
1797 max_asap
= ASAP (u_node
);
1805 find_max_hv_min_mob (ddg_ptr g
, sbitmap nodes
)
1809 int min_mob
= INT_MAX
;
1811 sbitmap_iterator sbi
;
1813 EXECUTE_IF_SET_IN_SBITMAP (nodes
, 0, u
, sbi
)
1815 ddg_node_ptr u_node
= &g
->nodes
[u
];
1817 if (max_hv
< HEIGHT (u_node
))
1819 max_hv
= HEIGHT (u_node
);
1820 min_mob
= MOB (u_node
);
1823 else if ((max_hv
== HEIGHT (u_node
))
1824 && (min_mob
> MOB (u_node
)))
1826 min_mob
= MOB (u_node
);
1834 find_max_dv_min_mob (ddg_ptr g
, sbitmap nodes
)
1838 int min_mob
= INT_MAX
;
1840 sbitmap_iterator sbi
;
1842 EXECUTE_IF_SET_IN_SBITMAP (nodes
, 0, u
, sbi
)
1844 ddg_node_ptr u_node
= &g
->nodes
[u
];
1846 if (max_dv
< DEPTH (u_node
))
1848 max_dv
= DEPTH (u_node
);
1849 min_mob
= MOB (u_node
);
1852 else if ((max_dv
== DEPTH (u_node
))
1853 && (min_mob
> MOB (u_node
)))
1855 min_mob
= MOB (u_node
);
1862 /* Places the nodes of SCC into the NODE_ORDER array starting
1863 at position POS, according to the SMS ordering algorithm.
1864 NODES_ORDERED (in&out parameter) holds the bitset of all nodes in
1865 the NODE_ORDER array, starting from position zero. */
1867 order_nodes_in_scc (ddg_ptr g
, sbitmap nodes_ordered
, sbitmap scc
,
1868 int * node_order
, int pos
)
1870 enum sms_direction dir
;
1871 int num_nodes
= g
->num_nodes
;
1872 sbitmap workset
= sbitmap_alloc (num_nodes
);
1873 sbitmap tmp
= sbitmap_alloc (num_nodes
);
1874 sbitmap zero_bitmap
= sbitmap_alloc (num_nodes
);
1875 sbitmap predecessors
= sbitmap_alloc (num_nodes
);
1876 sbitmap successors
= sbitmap_alloc (num_nodes
);
1878 sbitmap_zero (predecessors
);
1879 find_predecessors (predecessors
, g
, nodes_ordered
);
1881 sbitmap_zero (successors
);
1882 find_successors (successors
, g
, nodes_ordered
);
1885 if (sbitmap_a_and_b_cg (tmp
, predecessors
, scc
))
1887 sbitmap_copy (workset
, tmp
);
1890 else if (sbitmap_a_and_b_cg (tmp
, successors
, scc
))
1892 sbitmap_copy (workset
, tmp
);
1899 sbitmap_zero (workset
);
1900 if ((u
= find_max_asap (g
, scc
)) >= 0)
1901 SET_BIT (workset
, u
);
1905 sbitmap_zero (zero_bitmap
);
1906 while (!sbitmap_equal (workset
, zero_bitmap
))
1909 ddg_node_ptr v_node
;
1910 sbitmap v_node_preds
;
1911 sbitmap v_node_succs
;
1915 while (!sbitmap_equal (workset
, zero_bitmap
))
1917 v
= find_max_hv_min_mob (g
, workset
);
1918 v_node
= &g
->nodes
[v
];
1919 node_order
[pos
++] = v
;
1920 v_node_succs
= NODE_SUCCESSORS (v_node
);
1921 sbitmap_a_and_b (tmp
, v_node_succs
, scc
);
1923 /* Don't consider the already ordered successors again. */
1924 sbitmap_difference (tmp
, tmp
, nodes_ordered
);
1925 sbitmap_a_or_b (workset
, workset
, tmp
);
1926 RESET_BIT (workset
, v
);
1927 SET_BIT (nodes_ordered
, v
);
1930 sbitmap_zero (predecessors
);
1931 find_predecessors (predecessors
, g
, nodes_ordered
);
1932 sbitmap_a_and_b (workset
, predecessors
, scc
);
1936 while (!sbitmap_equal (workset
, zero_bitmap
))
1938 v
= find_max_dv_min_mob (g
, workset
);
1939 v_node
= &g
->nodes
[v
];
1940 node_order
[pos
++] = v
;
1941 v_node_preds
= NODE_PREDECESSORS (v_node
);
1942 sbitmap_a_and_b (tmp
, v_node_preds
, scc
);
1944 /* Don't consider the already ordered predecessors again. */
1945 sbitmap_difference (tmp
, tmp
, nodes_ordered
);
1946 sbitmap_a_or_b (workset
, workset
, tmp
);
1947 RESET_BIT (workset
, v
);
1948 SET_BIT (nodes_ordered
, v
);
1951 sbitmap_zero (successors
);
1952 find_successors (successors
, g
, nodes_ordered
);
1953 sbitmap_a_and_b (workset
, successors
, scc
);
1957 sbitmap_free (workset
);
1958 sbitmap_free (zero_bitmap
);
1959 sbitmap_free (predecessors
);
1960 sbitmap_free (successors
);
1965 /* This page contains functions for manipulating partial-schedules during
1966 modulo scheduling. */
1968 /* Create a partial schedule and allocate a memory to hold II rows. */
1970 static partial_schedule_ptr
1971 create_partial_schedule (int ii
, ddg_ptr g
, int history
)
1973 partial_schedule_ptr ps
= XNEW (struct partial_schedule
);
1974 ps
->rows
= (ps_insn_ptr
*) xcalloc (ii
, sizeof (ps_insn_ptr
));
1976 ps
->history
= history
;
1977 ps
->min_cycle
= INT_MAX
;
1978 ps
->max_cycle
= INT_MIN
;
1984 /* Free the PS_INSNs in rows array of the given partial schedule.
1985 ??? Consider caching the PS_INSN's. */
1987 free_ps_insns (partial_schedule_ptr ps
)
1991 for (i
= 0; i
< ps
->ii
; i
++)
1995 ps_insn_ptr ps_insn
= ps
->rows
[i
]->next_in_row
;
1998 ps
->rows
[i
] = ps_insn
;
2004 /* Free all the memory allocated to the partial schedule. */
2007 free_partial_schedule (partial_schedule_ptr ps
)
2016 /* Clear the rows array with its PS_INSNs, and create a new one with
2020 reset_partial_schedule (partial_schedule_ptr ps
, int new_ii
)
2025 if (new_ii
== ps
->ii
)
2027 ps
->rows
= (ps_insn_ptr
*) xrealloc (ps
->rows
, new_ii
2028 * sizeof (ps_insn_ptr
));
2029 memset (ps
->rows
, 0, new_ii
* sizeof (ps_insn_ptr
));
2031 ps
->min_cycle
= INT_MAX
;
2032 ps
->max_cycle
= INT_MIN
;
2035 /* Prints the partial schedule as an ii rows array, for each rows
2036 print the ids of the insns in it. */
2038 print_partial_schedule (partial_schedule_ptr ps
, FILE *dump
)
2042 for (i
= 0; i
< ps
->ii
; i
++)
2044 ps_insn_ptr ps_i
= ps
->rows
[i
];
2046 fprintf (dump
, "\n[CYCLE %d ]: ", i
);
2049 fprintf (dump
, "%d, ",
2050 INSN_UID (ps_i
->node
->insn
));
2051 ps_i
= ps_i
->next_in_row
;
2056 /* Creates an object of PS_INSN and initializes it to the given parameters. */
2058 create_ps_insn (ddg_node_ptr node
, int rest_count
, int cycle
)
2060 ps_insn_ptr ps_i
= XNEW (struct ps_insn
);
2063 ps_i
->next_in_row
= NULL
;
2064 ps_i
->prev_in_row
= NULL
;
2065 ps_i
->row_rest_count
= rest_count
;
2066 ps_i
->cycle
= cycle
;
2072 /* Removes the given PS_INSN from the partial schedule. Returns false if the
2073 node is not found in the partial schedule, else returns true. */
2075 remove_node_from_ps (partial_schedule_ptr ps
, ps_insn_ptr ps_i
)
2082 row
= SMODULO (ps_i
->cycle
, ps
->ii
);
2083 if (! ps_i
->prev_in_row
)
2085 if (ps_i
!= ps
->rows
[row
])
2088 ps
->rows
[row
] = ps_i
->next_in_row
;
2090 ps
->rows
[row
]->prev_in_row
= NULL
;
2094 ps_i
->prev_in_row
->next_in_row
= ps_i
->next_in_row
;
2095 if (ps_i
->next_in_row
)
2096 ps_i
->next_in_row
->prev_in_row
= ps_i
->prev_in_row
;
2102 /* Unlike what literature describes for modulo scheduling (which focuses
2103 on VLIW machines) the order of the instructions inside a cycle is
2104 important. Given the bitmaps MUST_FOLLOW and MUST_PRECEDE we know
2105 where the current instruction should go relative to the already
2106 scheduled instructions in the given cycle. Go over these
2107 instructions and find the first possible column to put it in. */
2109 ps_insn_find_column (partial_schedule_ptr ps
, ps_insn_ptr ps_i
,
2110 sbitmap must_precede
, sbitmap must_follow
)
2112 ps_insn_ptr next_ps_i
;
2113 ps_insn_ptr first_must_follow
= NULL
;
2114 ps_insn_ptr last_must_precede
= NULL
;
2120 row
= SMODULO (ps_i
->cycle
, ps
->ii
);
2122 /* Find the first must follow and the last must precede
2123 and insert the node immediately after the must precede
2124 but make sure that it there is no must follow after it. */
2125 for (next_ps_i
= ps
->rows
[row
];
2127 next_ps_i
= next_ps_i
->next_in_row
)
2129 if (TEST_BIT (must_follow
, next_ps_i
->node
->cuid
)
2130 && ! first_must_follow
)
2131 first_must_follow
= next_ps_i
;
2132 if (TEST_BIT (must_precede
, next_ps_i
->node
->cuid
))
2134 /* If we have already met a node that must follow, then
2135 there is no possible column. */
2136 if (first_must_follow
)
2139 last_must_precede
= next_ps_i
;
2143 /* Now insert the node after INSERT_AFTER_PSI. */
2145 if (! last_must_precede
)
2147 ps_i
->next_in_row
= ps
->rows
[row
];
2148 ps_i
->prev_in_row
= NULL
;
2149 if (ps_i
->next_in_row
)
2150 ps_i
->next_in_row
->prev_in_row
= ps_i
;
2151 ps
->rows
[row
] = ps_i
;
2155 ps_i
->next_in_row
= last_must_precede
->next_in_row
;
2156 last_must_precede
->next_in_row
= ps_i
;
2157 ps_i
->prev_in_row
= last_must_precede
;
2158 if (ps_i
->next_in_row
)
2159 ps_i
->next_in_row
->prev_in_row
= ps_i
;
2165 /* Advances the PS_INSN one column in its current row; returns false
2166 in failure and true in success. Bit N is set in MUST_FOLLOW if
2167 the node with cuid N must be come after the node pointed to by
2168 PS_I when scheduled in the same cycle. */
2170 ps_insn_advance_column (partial_schedule_ptr ps
, ps_insn_ptr ps_i
,
2171 sbitmap must_follow
)
2173 ps_insn_ptr prev
, next
;
2175 ddg_node_ptr next_node
;
2180 row
= SMODULO (ps_i
->cycle
, ps
->ii
);
2182 if (! ps_i
->next_in_row
)
2185 next_node
= ps_i
->next_in_row
->node
;
2187 /* Check if next_in_row is dependent on ps_i, both having same sched
2188 times (typically ANTI_DEP). If so, ps_i cannot skip over it. */
2189 if (TEST_BIT (must_follow
, next_node
->cuid
))
2192 /* Advance PS_I over its next_in_row in the doubly linked list. */
2193 prev
= ps_i
->prev_in_row
;
2194 next
= ps_i
->next_in_row
;
2196 if (ps_i
== ps
->rows
[row
])
2197 ps
->rows
[row
] = next
;
2199 ps_i
->next_in_row
= next
->next_in_row
;
2201 if (next
->next_in_row
)
2202 next
->next_in_row
->prev_in_row
= ps_i
;
2204 next
->next_in_row
= ps_i
;
2205 ps_i
->prev_in_row
= next
;
2207 next
->prev_in_row
= prev
;
2209 prev
->next_in_row
= next
;
2214 /* Inserts a DDG_NODE to the given partial schedule at the given cycle.
2215 Returns 0 if this is not possible and a PS_INSN otherwise. Bit N is
2216 set in MUST_PRECEDE/MUST_FOLLOW if the node with cuid N must be come
2217 before/after (respectively) the node pointed to by PS_I when scheduled
2218 in the same cycle. */
2220 add_node_to_ps (partial_schedule_ptr ps
, ddg_node_ptr node
, int cycle
,
2221 sbitmap must_precede
, sbitmap must_follow
)
2225 int row
= SMODULO (cycle
, ps
->ii
);
2228 && ps
->rows
[row
]->row_rest_count
>= issue_rate
)
2232 rest_count
+= ps
->rows
[row
]->row_rest_count
;
2234 ps_i
= create_ps_insn (node
, rest_count
, cycle
);
2236 /* Finds and inserts PS_I according to MUST_FOLLOW and
2238 if (! ps_insn_find_column (ps
, ps_i
, must_precede
, must_follow
))
2247 /* Advance time one cycle. Assumes DFA is being used. */
2249 advance_one_cycle (void)
2251 if (targetm
.sched
.dfa_pre_cycle_insn
)
2252 state_transition (curr_state
,
2253 targetm
.sched
.dfa_pre_cycle_insn ());
2255 state_transition (curr_state
, NULL
);
2257 if (targetm
.sched
.dfa_post_cycle_insn
)
2258 state_transition (curr_state
,
2259 targetm
.sched
.dfa_post_cycle_insn ());
2262 /* Given the kernel of a loop (from FIRST_INSN to LAST_INSN), finds
2263 the number of cycles according to DFA that the kernel fits in,
2264 we use this to check if we done well with SMS after we add
2265 register moves. In some cases register moves overhead makes
2266 it even worse than the original loop. We want SMS to be performed
2267 when it gives less cycles after register moves are added. */
2269 kernel_number_of_cycles (rtx first_insn
, rtx last_insn
)
2273 int can_issue_more
= issue_rate
;
2275 state_reset (curr_state
);
2277 for (insn
= first_insn
;
2278 insn
!= NULL_RTX
&& insn
!= last_insn
;
2279 insn
= NEXT_INSN (insn
))
2281 if (! INSN_P (insn
) || GET_CODE (PATTERN (insn
)) == USE
)
2284 /* Check if there is room for the current insn. */
2285 if (!can_issue_more
|| state_dead_lock_p (curr_state
))
2288 advance_one_cycle ();
2289 can_issue_more
= issue_rate
;
2292 /* Update the DFA state and return with failure if the DFA found
2293 recource conflicts. */
2294 if (state_transition (curr_state
, insn
) >= 0)
2297 advance_one_cycle ();
2298 can_issue_more
= issue_rate
;
2301 if (targetm
.sched
.variable_issue
)
2303 targetm
.sched
.variable_issue (sched_dump
, sched_verbose
,
2304 insn
, can_issue_more
);
2305 /* A naked CLOBBER or USE generates no instruction, so don't
2306 let them consume issue slots. */
2307 else if (GET_CODE (PATTERN (insn
)) != USE
2308 && GET_CODE (PATTERN (insn
)) != CLOBBER
)
2314 /* Checks if PS has resource conflicts according to DFA, starting from
2315 FROM cycle to TO cycle; returns true if there are conflicts and false
2316 if there are no conflicts. Assumes DFA is being used. */
2318 ps_has_conflicts (partial_schedule_ptr ps
, int from
, int to
)
2322 state_reset (curr_state
);
2324 for (cycle
= from
; cycle
<= to
; cycle
++)
2326 ps_insn_ptr crr_insn
;
2327 /* Holds the remaining issue slots in the current row. */
2328 int can_issue_more
= issue_rate
;
2330 /* Walk through the DFA for the current row. */
2331 for (crr_insn
= ps
->rows
[SMODULO (cycle
, ps
->ii
)];
2333 crr_insn
= crr_insn
->next_in_row
)
2335 rtx insn
= crr_insn
->node
->insn
;
2340 /* Check if there is room for the current insn. */
2341 if (!can_issue_more
|| state_dead_lock_p (curr_state
))
2344 /* Update the DFA state and return with failure if the DFA found
2345 recource conflicts. */
2346 if (state_transition (curr_state
, insn
) >= 0)
2349 if (targetm
.sched
.variable_issue
)
2351 targetm
.sched
.variable_issue (sched_dump
, sched_verbose
,
2352 insn
, can_issue_more
);
2353 /* A naked CLOBBER or USE generates no instruction, so don't
2354 let them consume issue slots. */
2355 else if (GET_CODE (PATTERN (insn
)) != USE
2356 && GET_CODE (PATTERN (insn
)) != CLOBBER
)
2360 /* Advance the DFA to the next cycle. */
2361 advance_one_cycle ();
2366 /* Checks if the given node causes resource conflicts when added to PS at
2367 cycle C. If not the node is added to PS and returned; otherwise zero
2368 is returned. Bit N is set in MUST_PRECEDE/MUST_FOLLOW if the node with
2369 cuid N must be come before/after (respectively) the node pointed to by
2370 PS_I when scheduled in the same cycle. */
2372 ps_add_node_check_conflicts (partial_schedule_ptr ps
, ddg_node_ptr n
,
2373 int c
, sbitmap must_precede
,
2374 sbitmap must_follow
)
2376 int has_conflicts
= 0;
2379 /* First add the node to the PS, if this succeeds check for
2380 conflicts, trying different issue slots in the same row. */
2381 if (! (ps_i
= add_node_to_ps (ps
, n
, c
, must_precede
, must_follow
)))
2382 return NULL
; /* Failed to insert the node at the given cycle. */
2384 has_conflicts
= ps_has_conflicts (ps
, c
, c
)
2386 && ps_has_conflicts (ps
,
2390 /* Try different issue slots to find one that the given node can be
2391 scheduled in without conflicts. */
2392 while (has_conflicts
)
2394 if (! ps_insn_advance_column (ps
, ps_i
, must_follow
))
2396 has_conflicts
= ps_has_conflicts (ps
, c
, c
)
2398 && ps_has_conflicts (ps
,
2405 remove_node_from_ps (ps
, ps_i
);
2409 ps
->min_cycle
= MIN (ps
->min_cycle
, c
);
2410 ps
->max_cycle
= MAX (ps
->max_cycle
, c
);
2414 /* Rotate the rows of PS such that insns scheduled at time
2415 START_CYCLE will appear in row 0. Updates max/min_cycles. */
2417 rotate_partial_schedule (partial_schedule_ptr ps
, int start_cycle
)
2419 int i
, row
, backward_rotates
;
2420 int last_row
= ps
->ii
- 1;
2422 if (start_cycle
== 0)
2425 backward_rotates
= SMODULO (start_cycle
, ps
->ii
);
2427 /* Revisit later and optimize this into a single loop. */
2428 for (i
= 0; i
< backward_rotates
; i
++)
2430 ps_insn_ptr first_row
= ps
->rows
[0];
2432 for (row
= 0; row
< last_row
; row
++)
2433 ps
->rows
[row
] = ps
->rows
[row
+1];
2435 ps
->rows
[last_row
] = first_row
;
2438 ps
->max_cycle
-= start_cycle
;
2439 ps
->min_cycle
-= start_cycle
;
2442 /* Remove the node N from the partial schedule PS; because we restart the DFA
2443 each time we want to check for resource conflicts; this is equivalent to
2444 unscheduling the node N. */
2446 ps_unschedule_node (partial_schedule_ptr ps
, ddg_node_ptr n
)
2449 int row
= SMODULO (SCHED_TIME (n
), ps
->ii
);
2451 if (row
< 0 || row
> ps
->ii
)
2454 for (ps_i
= ps
->rows
[row
];
2455 ps_i
&& ps_i
->node
!= n
;
2456 ps_i
= ps_i
->next_in_row
);
2460 return remove_node_from_ps (ps
, ps_i
);
2462 #endif /* INSN_SCHEDULING */
2465 gate_handle_sms (void)
2467 return (optimize
> 0 && flag_modulo_sched
);
2471 /* Run instruction scheduler. */
2472 /* Perform SMS module scheduling. */
2474 rest_of_handle_sms (void)
2476 #ifdef INSN_SCHEDULING
2479 /* We want to be able to create new pseudos. */
2481 /* Collect loop information to be used in SMS. */
2482 cfg_layout_initialize (CLEANUP_UPDATE_LIFE
);
2485 /* Update the life information, because we add pseudos. */
2486 max_regno
= max_reg_num ();
2487 allocate_reg_info (max_regno
, FALSE
, FALSE
);
2488 update_life_info (NULL
, UPDATE_LIFE_GLOBAL_RM_NOTES
,
2491 | PROP_KILL_DEAD_CODE
2492 | PROP_SCAN_DEAD_CODE
));
2496 /* Finalize layout changes. */
2498 if (bb
->next_bb
!= EXIT_BLOCK_PTR
)
2499 bb
->aux
= bb
->next_bb
;
2500 cfg_layout_finalize ();
2501 free_dominance_info (CDI_DOMINATORS
);
2502 #endif /* INSN_SCHEDULING */
2506 struct tree_opt_pass pass_sms
=
2509 gate_handle_sms
, /* gate */
2510 rest_of_handle_sms
, /* execute */
2513 0, /* static_pass_number */
2515 0, /* properties_required */
2516 0, /* properties_provided */
2517 0, /* properties_destroyed */
2518 TODO_dump_func
, /* todo_flags_start */
2520 TODO_ggc_collect
, /* todo_flags_finish */