![[BACK]](/cvsweb/icons/back.gif){kind=icon}
Return to unroll.c CVS log ![[TXT]](/cvsweb/icons/text.gif){kind=icon}
![[DIR]](/cvsweb/icons/dir.gif){kind=icon}
Up to [GCC 2.x] / gcc|
Return to unroll.c CVS log | Up to [GCC 2.x] / gcc |
1.1 root 1: /* Try to unroll loops, and split induction variables.
2: Copyright (C) 1992 Free Software Foundation, Inc.
3: Contributed by James E. Wilson, Cygnus Support/UC Berkeley.
4:
5: This file is part of GNU CC.
6:
7: GNU CC is free software; you can redistribute it and/or modify
8: it under the terms of the GNU General Public License as published by
9: the Free Software Foundation; either version 2, or (at your option)
10: any later version.
11:
12: GNU CC is distributed in the hope that it will be useful,
13: but WITHOUT ANY WARRANTY; without even the implied warranty of
14: MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
15: GNU General Public License for more details.
16:
17: You should have received a copy of the GNU General Public License
18: along with GNU CC; see the file COPYING. If not, write to
19: the Free Software Foundation, 675 Mass Ave, Cambridge, MA 02139, USA. */
20:
21: /* Try to unroll a loop, and split induction variables.
22:
23: Loops for which the number of iterations can be calculated exactly are
24: handled specially. If the number of iterations times the insn_count is
25: less than MAX_UNROLLED_INSNS, then the loop is unrolled completely.
26: Otherwise, we try to unroll the loop a number of times modulo the number
27: of iterations, so that only one exit test will be needed. It is unrolled
28: a number of times approximately equal to MAX_UNROLLED_INSNS divided by
29: the insn count.
30:
31: Otherwise, if the number of iterations can be calculated exactly at
32: run time, and the loop is always entered at the top, then we try to
33: precondition the loop. That is, at run time, calculate how many times
34: the loop will execute, and then execute the loop body a few times so
35: that the remaining iterations will be some multiple of 4 (or 2 if the
36: loop is large). Then fall through to a loop unrolled 4 (or 2) times,
37: with only one exit test needed at the end of the loop.
38:
39: Otherwise, if the number of iterations can not be calculated exactly,
40: not even at run time, then we still unroll the loop a number of times
41: approximately equal to MAX_UNROLLED_INSNS divided by the insn count,
42: but there must be an exit test after each copy of the loop body.
43:
44: For each induction variable, which is dead outside the loop (replaceable)
45: or for which we can easily calculate the final value, if we can easily
46: calculate its value at each place where it is set as a function of the
47: current loop unroll count and the variable's value at loop entry, then
48: the induction variable is split into `N' different variables, one for
49: each copy of the loop body. One variable is live across the backward
50: branch, and the others are all calculated as a function of this variable.
51: This helps eliminate data dependencies, and leads to further opportunities
52: for cse. */
53:
54: /* Possible improvements follow: */
55:
56: /* ??? Add an extra pass somewhere to determine whether unrolling will
57: give any benefit. E.g. after generating all unrolled insns, compute the
58: cost of all insns and compare against cost of insns in rolled loop.
59:
60: - On traditional architectures, unrolling a non-constant bound loop
61: is a win if there is a giv whose only use is in memory addresses, the
62: memory addresses can be split, and hence giv incremenets can be
63: eliminated.
64: - It is also a win if the loop is executed many times, and preconditioning
65: can be performed for the loop.
66: Add code to check for these and similar cases. */
67:
68: /* ??? Improve control of which loops get unrolled. Could use profiling
69: info to only unroll the most commonly executed loops. Perhaps have
70: a user specifyable option to control the amount of code expansion,
71: or the percent of loops to consider for unrolling. Etc. */
72:
73: /* ??? Look at the register copies inside the loop to see if they form a
74: simple permutation. If so, iterate the permutation until it gets back to
75: the start state. This is how many times we should unroll the loop, for
76: best results, because then all register copies can be eliminated.
77: For example, the lisp nreverse function should be unrolled 3 times
78: while (this)
79: {
80: next = this->cdr;
81: this->cdr = prev;
82: prev = this;
83: this = next;
84: }
85:
86: ??? The number of times to unroll the loop may also be based on data
87: references in the loop. For example, if we have a loop that references
88: x[i-1], x[i], and x[i+1], we should unroll it a multiple of 3 times. */
89:
90: /* ??? Add some simple linear equation solving capability so that we can
91: determine the number of loop iterations for more complex loops.
92: For example, consider this loop from gdb
93: #define SWAP_TARGET_AND_HOST(buffer,len)
94: {
95: char tmp;
96: char *p = (char *) buffer;
97: char *q = ((char *) buffer) + len - 1;
98: int iterations = (len + 1) >> 1;
99: int i;
100: for (p; p < q; p++, q--;)
101: {
102: tmp = *q;
103: *q = *p;
104: *p = tmp;
105: }
106: }
107: Note that:
108: start value = p = &buffer + current_iteration
109: end value = q = &buffer + len - 1 - current_iteration
110: Given the loop exit test of "p < q", then there must be "q - p" iterations,
111: set equal to zero and solve for number of iterations:
112: q - p = len - 1 - 2*current_iteration = 0
113: current_iteration = (len - 1) / 2
114: Hence, there are (len - 1) / 2 (rounded up to the nearest integer)
115: iterations of this loop. */
116:
117: /* ??? Currently, no labels are marked as loop invariant when doing loop
118: unrolling. This is because an insn inside the loop, that loads the address
119: of a label inside the loop into a register, could be moved outside the loop
120: by the invariant code motion pass if labels were invariant. If the loop
121: is subsequently unrolled, the code will be wrong because each unrolled
122: body of the loop will use the same address, whereas each actually needs a
123: different address. A case where this happens is when a loop containing
124: a switch statement is unrolled.
125:
126: It would be better to let labels be considered invariant. When we
127: unroll loops here, check to see if any insns using a label local to the
128: loop were moved before the loop. If so, then correct the problem, by
129: moving the insn back into the loop, or perhaps replicate the insn before
130: the loop, one copy for each time the loop is unrolled. */
131:
132: /* The prime factors looked for when trying to unroll a loop by some
133: number which is modulo the total number of iterations. Just checking
134: for these 4 prime factors will find at least one factor for 75% of
135: all numbers theoretically. Practically speaking, this will succeed
136: almost all of the time since loops are generally a multiple of 2
137: and/or 5. */
138:
139: #define NUM_FACTORS 4
140:
141: struct _factor { int factor, count; } factors[NUM_FACTORS]
142: = { {2, 0}, {3, 0}, {5, 0}, {7, 0}};
143:
144: /* Describes the different types of loop unrolling performed. */
145:
146: enum unroll_types { UNROLL_COMPLETELY, UNROLL_MODULO, UNROLL_NAIVE };
147:
148: #include "config.h"
149: #include "rtl.h"
150: #include "insn-config.h"
151: #include "integrate.h"
152: #include "regs.h"
153: #include "flags.h"
154: #include "expr.h"
155: #include <stdio.h>
156: #include "loop.h"
157:
158: /* This controls which loops are unrolled, and by how much we unroll
159: them. */
160:
161: #ifndef MAX_UNROLLED_INSNS
162: #define MAX_UNROLLED_INSNS 100
163: #endif
164:
165: /* Indexed by register number, if non-zero, then it contains a pointer
166: to a struct induction for a DEST_REG giv which has been combined with
167: one of more address givs. This is needed because whenever such a DEST_REG
168: giv is modified, we must modify the value of all split address givs
169: that were combined with this DEST_REG giv. */
170:
171: static struct induction **addr_combined_regs;
172:
173: /* Indexed by register number, if this is a splittable induction variable,
174: then this will hold the current value of the register, which depends on the
175: iteration number. */
176:
177: static rtx *splittable_regs;
178:
179: /* Indexed by register number, if this is a splittable induction variable,
180: then this will hold the number of instructions in the loop that modify
181: the induction variable. Used to ensure that only the last insn modifying
182: a split iv will update the original iv of the dest. */
183:
184: static int *splittable_regs_updates;
185:
186: /* Values describing the current loop's iteration variable. These are set up
187: by loop_iterations, and used by precondition_loop_p. */
188:
189: static rtx loop_iteration_var;
190: static rtx loop_initial_value;
191: static rtx loop_increment;
192: static rtx loop_final_value;
193:
194: /* Forward declarations. */
195:
196: static void init_reg_map ();
197: static int precondition_loop_p ();
198: static void copy_loop_body ();
199: static void iteration_info ();
200: static rtx approx_final_value ();
201: static int find_splittable_regs ();
202: static int find_splittable_givs ();
203: static rtx fold_rtx_mult_add ();
204:
205: /* Try to unroll one loop and split induction variables in the loop.
206:
207: The loop is described by the arguments LOOP_END, INSN_COUNT, and
208: LOOP_START. END_INSERT_BEDFORE indicates where insns should be added
209: which need to be executed when the loop falls through. STRENGTH_REDUCTION_P
210: indicates whether information generated in the strength reduction pass
211: is available.
212:
213: This function is intended to be called from within `strength_reduce'
214: in loop.c. */
215:
216: void
217: unroll_loop (loop_end, insn_count, loop_start, end_insert_before,
218: strength_reduce_p)
219: rtx loop_end;
220: int insn_count;
221: rtx loop_start;
222: rtx end_insert_before;
223: int strength_reduce_p;
224: {
225: int i, j, temp;
226: int unroll_number = 1;
227: rtx copy_start, copy_end;
228: rtx insn, copy, sequence, pattern, tem;
229: int max_labelno, max_insnno;
230: rtx insert_before;
231: struct inline_remap *map;
232: char *local_label;
233: int maxregnum;
234: int new_maxregnum;
235: rtx exit_label = 0;
236: rtx start_label;
237: struct iv_class *bl;
238: struct induction *v;
239: int splitting_not_safe = 0;
240: enum unroll_types unroll_type;
241: int loop_preconditioned = 0;
242: rtx safety_label;
243: /* This points to the last real insn in the loop, which should be either
244: a JUMP_INSN (for conditional jumps) or a BARRIER (for unconditional
245: jumps). */
246: rtx last_loop_insn;
247:
248: /* Don't bother unrolling huge loops. Since the minimum factor is
249: two, loops greater than one half of MAX_UNROLLED_INSNS will never
250: be unrolled. */
251: if (insn_count > MAX_UNROLLED_INSNS / 2)
252: {
253: if (loop_dump_stream)
254: fprintf (loop_dump_stream, "Unrolling failure: Loop too big.\n");
255: return;
256: }
257:
258: /* When emitting debugger info, we can't unroll loops with unequal numbers
259: of block_beg and block_end notes, because that would unbalance the block
260: structure of the function. This can happen as a result of the
261: "if (foo) bar; else break;" optimization in jump.c. */
262:
263: if (write_symbols != NO_DEBUG)
264: {
265: int block_begins = 0;
266: int block_ends = 0;
267:
268: for (insn = loop_start; insn != loop_end; insn = NEXT_INSN (insn))
269: {
270: if (GET_CODE (insn) == NOTE)
271: {
272: if (NOTE_LINE_NUMBER (insn) == NOTE_INSN_BLOCK_BEG)
273: block_begins++;
274: else if (NOTE_LINE_NUMBER (insn) == NOTE_INSN_BLOCK_END)
275: block_ends++;
276: }
277: }
278:
279: if (block_begins != block_ends)
280: {
281: if (loop_dump_stream)
282: fprintf (loop_dump_stream,
283: "Unrolling failure: Unbalanced block notes.\n");
284: return;
285: }
286: }
287:
288: /* Determine type of unroll to perform. Depends on the number of iterations
289: and the size of the loop. */
290:
291: /* If there is no strength reduce info, then set loop_n_iterations to zero.
292: This can happen if strength_reduce can't find any bivs in the loop.
293: A value of zero indicates that the number of iterations could not be
294: calculated. */
295:
296: if (! strength_reduce_p)
297: loop_n_iterations = 0;
298:
299: if (loop_dump_stream && loop_n_iterations > 0)
300: fprintf (loop_dump_stream,
301: "Loop unrolling: %d iterations.\n", loop_n_iterations);
302:
303: /* Find and save a pointer to the last nonnote insn in the loop. */
304:
305: last_loop_insn = prev_nonnote_insn (loop_end);
306:
307: /* Calculate how many times to unroll the loop. Indicate whether or
308: not the loop is being completely unrolled. */
309:
310: if (loop_n_iterations == 1)
311: {
312: /* If number of iterations is exactly 1, then eliminate the compare and
313: branch at the end of the loop since they will never be taken.
314: Then return, since no other action is needed here. */
315:
316: /* If the last instruction is not a BARRIER or a JUMP_INSN, then
317: don't do anything. */
318:
319: if (GET_CODE (last_loop_insn) == BARRIER)
320: {
321: /* Delete the jump insn. This will delete the barrier also. */
322: delete_insn (PREV_INSN (last_loop_insn));
323: }
324: else if (GET_CODE (last_loop_insn) == JUMP_INSN)
325: {
326: #ifdef HAVE_cc0
327: /* The immediately preceeding insn is a compare which must be
328: deleted. */
329: delete_insn (last_loop_insn);
330: delete_insn (PREV_INSN (last_loop_insn));
331: #else
332: /* The immediately preceeding insn may not be the compare, so don't
333: delete it. */
334: delete_insn (last_loop_insn);
335: #endif
336: }
337: return;
338: }
339: else if (loop_n_iterations > 0
340: && loop_n_iterations * insn_count < MAX_UNROLLED_INSNS)
341: {
342: unroll_number = loop_n_iterations;
343: unroll_type = UNROLL_COMPLETELY;
344: }
345: else if (loop_n_iterations > 0)
346: {
347: /* Try to factor the number of iterations. Don't bother with the
348: general case, only using 2, 3, 5, and 7 will get 75% of all
349: numbers theoretically, and almost all in practice. */
350:
351: for (i = 0; i < NUM_FACTORS; i++)
352: factors[i].count = 0;
353:
354: temp = loop_n_iterations;
355: for (i = NUM_FACTORS - 1; i >= 0; i--)
356: while (temp % factors[i].factor == 0)
357: {
358: factors[i].count++;
359: temp = temp / factors[i].factor;
360: }
361:
362: /* Start with the larger factors first so that we generally
363: get lots of unrolling. */
364:
365: unroll_number = 1;
366: temp = insn_count;
367: for (i = 3; i >= 0; i--)
368: while (factors[i].count--)
369: {
370: if (temp * factors[i].factor < MAX_UNROLLED_INSNS)
371: {
372: unroll_number *= factors[i].factor;
373: temp *= factors[i].factor;
374: }
375: else
376: break;
377: }
378:
379: /* If we couldn't find any factors, then unroll as in the normal
380: case. */
381: if (unroll_number == 1)
382: {
383: if (loop_dump_stream)
384: fprintf (loop_dump_stream,
385: "Loop unrolling: No factors found.\n");
386: }
387: else
388: unroll_type = UNROLL_MODULO;
389: }
390:
391:
392: /* Default case, calculate number of times to unroll loop based on its
393: size. */
394: if (unroll_number == 1)
395: {
396: if (8 * insn_count < MAX_UNROLLED_INSNS)
397: unroll_number = 8;
398: else if (4 * insn_count < MAX_UNROLLED_INSNS)
399: unroll_number = 4;
400: else
401: unroll_number = 2;
402:
403: unroll_type = UNROLL_NAIVE;
404: }
405:
406: /* Now we know how many times to unroll the loop. */
407:
408: if (loop_dump_stream)
409: fprintf (loop_dump_stream,
410: "Unrolling loop %d times.\n", unroll_number);
411:
412:
413: if (unroll_type == UNROLL_COMPLETELY || unroll_type == UNROLL_MODULO)
414: {
415: /* Loops of these types should never start with a jump down to
416: the exit condition test. For now, check for this case just to
417: be sure. UNROLL_NAIVE loops can be of this form, this case is
418: handled below. */
419: insn = loop_start;
420: while (GET_CODE (insn) != CODE_LABEL && GET_CODE (insn) != JUMP_INSN)
421: insn = NEXT_INSN (insn);
422: if (GET_CODE (insn) == JUMP_INSN)
423: abort ();
424: }
425:
426: if (unroll_type == UNROLL_COMPLETELY)
427: {
428: /* Completely unrolling the loop: Delete the compare and branch at
429: the end (the last two instructions). This delete must done at the
430: very end of loop unrolling, to avoid problems with calls to
431: back_branch_in_range_p, which is called by find_splittable_regs.
432: All increments of splittable bivs/givs are changed to load constant
433: instructions. */
434:
435: copy_start = loop_start;
436:
437: /* Set insert_before to the instruction immediately after the JUMP_INSN
438: (or BARRIER), so that any NOTEs between the JUMP_INSN and the end of
439: the loop will be correctly handled by copy_loop_body. */
440: insert_before = NEXT_INSN (last_loop_insn);
441:
442: /* Set copy_end to the insn before the jump at the end of the loop. */
443: if (GET_CODE (last_loop_insn) == BARRIER)
444: copy_end = PREV_INSN (PREV_INSN (last_loop_insn));
445: else if (GET_CODE (last_loop_insn) == JUMP_INSN)
446: {
447: #ifdef HAVE_cc0
448: /* The instruction immediately before the JUMP_INSN is a compare
449: instruction which we do not want to copy. */
450: copy_end = PREV_INSN (PREV_INSN (last_loop_insn));
451: #else
452: /* The instruction immediately before the JUMP_INSN may not be the
453: compare, so we must copy it. */
454: copy_end = PREV_INSN (last_loop_insn);
455: #endif
456: }
457: else
458: {
459: /* We currently can't unroll a loop if it doesn't end with a
460: JUMP_INSN. There would need to be a mechanism that recognizes
461: this case, and then inserts a jump after each loop body, which
462: jumps to after the last loop body. */
463: if (loop_dump_stream)
464: fprintf (loop_dump_stream,
465: "Unrolling failure: loop does not end with a JUMP_INSN.\n");
466: return;
467: }
468: }
469: else if (unroll_type == UNROLL_MODULO)
470: {
471: /* Partially unrolling the loop: The compare and branch at the end
472: (the last two instructions) must remain. Don't copy the compare
473: and branch instructions at the end of the loop. Insert the unrolled
474: code immediately before the compare/branch at the end so that the
475: code will fall through to them as before. */
476:
477: copy_start = loop_start;
478:
479: /* Set insert_before to the jump insn at the end of the loop.
480: Set copy_end to before the jump insn at the end of the loop. */
481: if (GET_CODE (last_loop_insn) == BARRIER)
482: {
483: insert_before = PREV_INSN (last_loop_insn);
484: copy_end = PREV_INSN (insert_before);
485: }
486: else if (GET_CODE (last_loop_insn) == JUMP_INSN)
487: {
488: #ifdef HAVE_cc0
489: /* The instruction immediately before the JUMP_INSN is a compare
490: instruction which we do not want to copy or delete. */
491: insert_before = PREV_INSN (last_loop_insn);
492: copy_end = PREV_INSN (insert_before);
493: #else
494: /* The instruction immediately before the JUMP_INSN may not be the
495: compare, so we must copy it. */
496: insert_before = last_loop_insn;
497: copy_end = PREV_INSN (last_loop_insn);
498: #endif
499: }
500: else
501: {
502: /* We currently can't unroll a loop if it doesn't end with a
503: JUMP_INSN. There would need to be a mechanism that recognizes
504: this case, and then inserts a jump after each loop body, which
505: jumps to after the last loop body. */
506: if (loop_dump_stream)
507: fprintf (loop_dump_stream,
508: "Unrolling failure: loop does not end with a JUMP_INSN.\n");
509: return;
510: }
511: }
512: else
513: {
514: /* Normal case: Must copy the compare and branch instructions at the
515: end of the loop. */
516:
517: if (GET_CODE (last_loop_insn) == BARRIER)
518: {
519: /* Loop ends with an unconditional jump and a barrier.
520: Handle this like above, don't copy jump and barrier.
521: This is not strictly necessary, but doing so prevents generating
522: unconditional jumps to an immediately following label.
523:
524: This will be corrected below if the target of this jump is
525: not the start_label. */
526:
527: insert_before = PREV_INSN (last_loop_insn);
528: copy_end = PREV_INSN (insert_before);
529: }
530: else if (GET_CODE (last_loop_insn) == JUMP_INSN)
531: {
532: /* Set insert_before to immediately after the JUMP_INSN, so that
533: NOTEs at the end of the loop will be correctly handled by
534: copy_loop_body. */
535: insert_before = NEXT_INSN (last_loop_insn);
536: copy_end = last_loop_insn;
537: }
538: else
539: {
540: /* We currently can't unroll a loop if it doesn't end with a
541: JUMP_INSN. There would need to be a mechanism that recognizes
542: this case, and then inserts a jump after each loop body, which
543: jumps to after the last loop body. */
544: if (loop_dump_stream)
545: fprintf (loop_dump_stream,
546: "Unrolling failure: loop does not end with a JUMP_INSN.\n");
547: return;
548: }
549:
550: /* If copying exit test branches because they can not be eliminated,
551: then must convert the fall through case of the branch to a jump past
552: the end of the loop. Create a label to emit after the loop and save
553: it for later use. Do not use the label after the loop, if any, since
554: it might be used by insns outside the loop, or there might be insns
555: added before it later by final_[bg]iv_value which must be after
556: the real exit label. */
557: exit_label = gen_label_rtx ();
558:
559: insn = loop_start;
560: while (GET_CODE (insn) != CODE_LABEL && GET_CODE (insn) != JUMP_INSN)
561: insn = NEXT_INSN (insn);
562:
563: if (GET_CODE (insn) == JUMP_INSN)
564: {
565: /* The loop starts with a jump down to the exit condition test.
566: Start copying the loop after the barrier following this
567: jump insn. */
568: copy_start = NEXT_INSN (insn);
569:
570: /* Splitting induction variables doesn't work when the loop is
571: entered via a jump to the bottom, because then we end up doing
572: a comparison against a new register for a split variable, but
573: we did not execute the set insn for the new register because
574: it was skipped over. */
575: splitting_not_safe = 1;
576: if (loop_dump_stream)
577: fprintf (loop_dump_stream,
578: "Splitting not safe, because loop not entered at top.\n");
579: }
580: else
581: copy_start = loop_start;
582: }
583:
584: /* This should always be the first label in the loop. */
585: start_label = NEXT_INSN (copy_start);
586: /* There may be a line number note and/or a loop continue note here. */
587: while (GET_CODE (start_label) == NOTE)
588: start_label = NEXT_INSN (start_label);
589: if (GET_CODE (start_label) != CODE_LABEL)
590: {
591: /* This can happen as a result of jump threading. If the first insns in
592: the loop test the same condition as the loop's backward jump, or the
593: opposite condition, then the backward jump will be modified to point
594: to elsewhere, and the loop's start label is deleted.
595:
596: This case currently can not be handled by the loop unrolling code. */
597:
598: if (loop_dump_stream)
599: fprintf (loop_dump_stream,
600: "Unrolling failure: unknown insns between BEG note and loop label.\n");
601: return;
602: }
603:
604: if (unroll_type == UNROLL_NAIVE
605: && GET_CODE (last_loop_insn) == BARRIER
606: && start_label != JUMP_LABEL (PREV_INSN (last_loop_insn)))
607: {
608: /* In this case, we must copy the jump and barrier, because they will
609: not be converted to jumps to an immediately following label. */
610:
611: insert_before = NEXT_INSN (last_loop_insn);
612: copy_end = last_loop_insn;
613: }
614:
615: /* Allocate a translation table for the labels and insn numbers.
616: They will be filled in as we copy the insns in the loop. */
617:
618: max_labelno = max_label_num ();
619: max_insnno = get_max_uid ();
620:
621: map = (struct inline_remap *) alloca (sizeof (struct inline_remap));
622:
623: /* Allocate the label map. */
624:
625: if (max_labelno > 0)
626: {
627: map->label_map = (rtx *) alloca (max_labelno * sizeof (rtx));
628:
629: local_label = (char *) alloca (max_labelno);
630: bzero (local_label, max_labelno);
631: }
632: else
633: map->label_map = 0;
634:
635: /* Search the loop and mark all local labels, i.e. the ones which have to
636: be distinct labels when copied. For all labels which might be
637: non-local, set their label_map entries to point to themselves.
638: If they happen to be local their label_map entries will be overwritten
639: before the loop body is copied. The label_map entries for local labels
640: will be set to a different value each time the loop body is copied. */
641:
642: for (insn = copy_start; insn != loop_end; insn = NEXT_INSN (insn))
643: {
644: if (GET_CODE (insn) == CODE_LABEL)
645: local_label[CODE_LABEL_NUMBER (insn)] = 1;
646: else if (GET_CODE (insn) == JUMP_INSN)
647: {
648: if (JUMP_LABEL (insn))
649: map->label_map[CODE_LABEL_NUMBER (JUMP_LABEL (insn))]
650: = JUMP_LABEL (insn);
651: else if (GET_CODE (PATTERN (insn)) == ADDR_VEC
652: || GET_CODE (PATTERN (insn)) == ADDR_DIFF_VEC)
653: {
654: rtx pat = PATTERN (insn);
655: int diff_vec_p = GET_CODE (PATTERN (insn)) == ADDR_DIFF_VEC;
656: int len = XVECLEN (pat, diff_vec_p);
657: rtx label;
658:
659: for (i = 0; i < len; i++)
660: {
661: label = XEXP (XVECEXP (pat, diff_vec_p, i), 0);
662: map->label_map[CODE_LABEL_NUMBER (label)] = label;
663: }
664: }
665: }
666: }
667:
668: /* Allocate space for the insn map. */
669:
670: map->insn_map = (rtx *) alloca (max_insnno * sizeof (rtx));
671:
672: /* Set this to zero, to indicate that we are doing loop unrolling,
673: not function inlining. */
674: map->inline_target = 0;
675:
676: /* The register and constant maps depend on the number of registers
677: present, so the final maps can't be created until after
678: find_splittable_regs is called. However, they are needed for
679: preconditioning, so we create temporary maps when preconditioning
680: is performed. */
681:
682: /* The preconditioning code may allocate two new pseudo registers. */
683: maxregnum = max_reg_num ();
684:
685: /* Allocate and zero out the splittable_regs and addr_combined_regs
686: arrays. These must be zeroed here because they will be used if
687: loop preconditioning is performed, and must be zero for that case.
688:
689: It is safe to do this here, since the extra registers created by the
690: preconditioning code and find_splittable_regs will never be used
691: to accees the splittable_regs[] and addr_combined_regs[] arrays. */
692:
693: splittable_regs = (rtx *) alloca (maxregnum * sizeof (rtx));
694: bzero (splittable_regs, maxregnum * sizeof (rtx));
695: splittable_regs_updates = (int *) alloca (maxregnum * sizeof (int));
696: bzero (splittable_regs_updates, maxregnum * sizeof (int));
697: addr_combined_regs
698: = (struct induction **) alloca (maxregnum * sizeof (struct induction *));
699: bzero (addr_combined_regs, maxregnum * sizeof (struct induction *));
700:
701: /* If this loop requires exit tests when unrolled, check to see if we
702: can precondition the loop so as to make the exit tests unnecessary.
703: Just like variable splitting, this is not safe if the loop is entered
704: via a jump to the bottom. Also, can not do this if no strength
705: reduce info, because precondition_loop_p uses this info. */
706:
707: /* Must copy the loop body for preconditioning before the following
708: find_splittable_regs call since that will emit insns which need to
709: be after the preconditioned loop copies, but immediately before the
710: unrolled loop copies. */
711:
712: /* Also, it is not safe to split induction variables for the preconditioned
713: copies of the loop body. If we split induction variables, then the code
714: assumes that each induction variable can be represented as a function
715: of its initial value and the loop iteration number. This is not true
716: in this case, because the last preconditioned copy of the loop body
717: could be any iteration from the first up to the `unroll_number-1'th,
718: depending on the initial value of the iteration variable. Therefore
719: we can not split induction variables here, because we can not calculate
720: their value. Hence, this code must occur before find_splittable_regs
721: is called. */
722:
723: if (unroll_type == UNROLL_NAIVE && ! splitting_not_safe && strength_reduce_p)
724: {
725: rtx initial_value, final_value, increment;
726:
727: if (precondition_loop_p (&initial_value, &final_value, &increment,
728: loop_start, loop_end))
729: {
730: register rtx diff, temp;
731: enum machine_mode mode;
732: rtx *labels;
733: int abs_inc, neg_inc;
734:
735: map->reg_map = (rtx *) alloca (maxregnum * sizeof (rtx));
736:
737: map->const_equiv_map = (rtx *) alloca (maxregnum * sizeof (rtx));
738: map->const_age_map = (unsigned *) alloca (maxregnum
739: * sizeof (unsigned));
740: map->const_equiv_map_size = maxregnum;
741: global_const_equiv_map = map->const_equiv_map;
742:
743: init_reg_map (map, maxregnum);
744:
745: /* Limit loop unrolling to 4, since this will make 7 copies of
746: the loop body. */
747: if (unroll_number > 4)
748: unroll_number = 4;
749:
750: /* Save the absolute value of the increment, and also whether or
751: not it is negative. */
752: neg_inc = 0;
753: abs_inc = INTVAL (increment);
754: if (abs_inc < 0)
755: {
756: abs_inc = - abs_inc;
757: neg_inc = 1;
758: }
759:
760: start_sequence ();
761:
762: /* Decide what mode to do these calculations in. Choose the larger
763: of final_value's mode and initial_value's mode, or a full-word if
764: both are constants. */
765: mode = GET_MODE (final_value);
766: if (mode == VOIDmode)
767: {
768: mode = GET_MODE (initial_value);
769: if (mode == VOIDmode)
770: mode = word_mode;
771: }
772: else if (mode != GET_MODE (initial_value)
773: && (GET_MODE_SIZE (mode)
774: < GET_MODE_SIZE (GET_MODE (initial_value))))
775: mode = GET_MODE (initial_value);
776:
777: /* Calculate the difference between the final and initial values.
778: Final value may be a (plus (reg x) (const_int 1)) rtx.
779: Let the following cse pass simplify this if initial value is
780: a constant.
781:
782: We must copy the final and initial values here to avoid
783: improperly shared rtl. */
784:
785: diff = expand_binop (mode, sub_optab, copy_rtx (final_value),
786: copy_rtx (initial_value), 0, 0,
787: OPTAB_LIB_WIDEN);
788:
789: /* Now calculate (diff % (unroll * abs (increment))) by using an
790: and instruction. */
791: diff = expand_binop (GET_MODE (diff), and_optab, diff,
792: gen_rtx (CONST_INT, VOIDmode,
793: unroll_number * abs_inc - 1),
794: 0, 0, OPTAB_LIB_WIDEN);
795:
796: /* Now emit a sequence of branches to jump to the proper precond
797: loop entry point. */
798:
799: labels = (rtx *) alloca (sizeof (rtx) * unroll_number);
800: for (i = 0; i < unroll_number; i++)
801: labels[i] = gen_label_rtx ();
802:
803: /* Assuming the unroll_number is 4, and the increment is 2, then
804: for a negative increment: for a positive increment:
805: diff = 0,1 precond 0 diff = 0,7 precond 0
806: diff = 2,3 precond 3 diff = 1,2 precond 1
807: diff = 4,5 precond 2 diff = 3,4 precond 2
808: diff = 6,7 precond 1 diff = 5,6 precond 3 */
809:
810: /* We only need to emit (unroll_number - 1) branches here, the
811: last case just falls through to the following code. */
812:
813: /* ??? This would give better code if we emitted a tree of branches
814: instead of the current linear list of branches. */
815:
816: for (i = 0; i < unroll_number - 1; i++)
817: {
818: int cmp_const;
819:
820: /* For negative increments, must invert the constant compared
821: against, except when comparing against zero. */
822: if (i == 0)
823: cmp_const = 0;
824: else if (neg_inc)
825: cmp_const = unroll_number - i;
826: else
827: cmp_const = i;
828:
829: emit_cmp_insn (diff, gen_rtx (CONST_INT, VOIDmode,
830: abs_inc * cmp_const),
831: EQ, 0, mode, 0, 0);
832:
833: if (i == 0)
834: emit_jump_insn (gen_beq (labels[i]));
835: else if (neg_inc)
836: emit_jump_insn (gen_bge (labels[i]));
837: else
838: emit_jump_insn (gen_ble (labels[i]));
839: JUMP_LABEL (get_last_insn ()) = labels[i];
840: LABEL_NUSES (labels[i])++;
841: }
842:
843: /* If the increment is greater than one, then we need another branch,
844: to handle other cases equivalent to 0. */
845:
846: /* ??? This should be merged into the code above somehow to help
847: simplify the code here, and reduce the number of branches emitted.
848: For the negative increment case, the branch here could easily
849: be merged with the `0' case branch above. For the positive
850: increment case, it is not clear how this can be simplified. */
851:
852: if (abs_inc != 1)
853: {
854: int cmp_const;
855:
856: if (neg_inc)
857: cmp_const = abs_inc - 1;
858: else
859: cmp_const = abs_inc * (unroll_number - 1) + 1;
860:
861: emit_cmp_insn (diff, gen_rtx (CONST_INT, VOIDmode, cmp_const),
862: EQ, 0, mode, 0, 0);
863:
864: if (neg_inc)
865: emit_jump_insn (gen_ble (labels[0]));
866: else
867: emit_jump_insn (gen_bge (labels[0]));
868: JUMP_LABEL (get_last_insn ()) = labels[0];
869: LABEL_NUSES (labels[0])++;
870: }
871:
872: sequence = gen_sequence ();
873: end_sequence ();
874: emit_insn_before (sequence, loop_start);
875:
876: /* Only the last copy of the loop body here needs the exit
877: test, so set copy_end to exclude the compare/branch here,
878: and then reset it inside the loop when get to the last
879: copy. */
880:
881: if (GET_CODE (last_loop_insn) == BARRIER)
882: copy_end = PREV_INSN (PREV_INSN (last_loop_insn));
883: else if (GET_CODE (last_loop_insn) == JUMP_INSN)
884: {
885: #ifdef HAVE_cc0
886: /* The immediately preceeding insn is a compare which we do not
887: want to copy. */
888: copy_end = PREV_INSN (PREV_INSN (last_loop_insn));
889: #else
890: /* The immediately preceeding insn may not be a compare, so we
891: must copy it. */
892: copy_end = PREV_INSN (last_loop_insn);
893: #endif
894: }
895: else
896: abort ();
897:
898: for (i = 1; i < unroll_number; i++)
899: {
900: emit_label_after (labels[unroll_number - i],
901: PREV_INSN (loop_start));
902:
903: bzero (map->insn_map, max_insnno * sizeof (rtx));
904: bzero (map->const_equiv_map, maxregnum * sizeof (rtx));
905: bzero (map->const_age_map, maxregnum * sizeof (unsigned));
906: map->const_age = 0;
907:
908: for (j = 0; j < max_labelno; j++)
909: if (local_label[j])
910: map->label_map[j] = gen_label_rtx ();
911:
912: /* The last copy needs the compare/branch insns at the end,
913: so reset copy_end here if the loop ends with a conditional
914: branch. */
915:
916: if (i == unroll_number - 1)
917: {
918: if (GET_CODE (last_loop_insn) == BARRIER)
919: copy_end = PREV_INSN (PREV_INSN (last_loop_insn));
920: else
921: copy_end = last_loop_insn;
922: }
923:
924: /* None of the copies are the `last_iteration', so just
925: pass zero for that parameter. */
926: copy_loop_body (copy_start, copy_end, map, exit_label, 0,
927: unroll_type, start_label, loop_end,
928: loop_start, copy_end);
929: }
930: emit_label_after (labels[0], PREV_INSN (loop_start));
931:
932: if (GET_CODE (last_loop_insn) == BARRIER)
933: {
934: insert_before = PREV_INSN (last_loop_insn);
935: copy_end = PREV_INSN (insert_before);
936: }
937: else
938: {
939: #ifdef HAVE_cc0
940: /* The immediately preceeding insn is a compare which we do not
941: want to copy. */
942: insert_before = PREV_INSN (last_loop_insn);
943: copy_end = PREV_INSN (insert_before);
944: #else
945: /* The immediately preceeding insn may not be a compare, so we
946: must copy it. */
947: insert_before = last_loop_insn;
948: copy_end = PREV_INSN (last_loop_insn);
949: #endif
950: }
951:
952: /* Set unroll type to MODULO now. */
953: unroll_type = UNROLL_MODULO;
954: loop_preconditioned = 1;
955: }
956: }
957:
958: /* If reach here, and the loop type is UNROLL_NAIVE, then don't unroll
959: the loop unless all loops are being unrolled. */
960: if (unroll_type == UNROLL_NAIVE && ! flag_unroll_all_loops)
961: {
962: if (loop_dump_stream)
963: fprintf (loop_dump_stream, "Unrolling failure: Naive unrolling not being done.\n");
964: return;
965: }
966:
967: /* At this point, we are guaranteed to unroll the loop. */
968:
969: /* For each biv and giv, determine whether it can be safely split into
970: a different variable for each unrolled copy of the loop body.
971: We precalculate and save this info here, since computing it is
972: expensive.
973:
974: Do this before deleting any instructions from the loop, so that
975: back_branch_in_range_p will work correctly. */
976:
977: if (splitting_not_safe)
978: temp = 0;
979: else
980: temp = find_splittable_regs (unroll_type, loop_start, loop_end,
981: end_insert_before, unroll_number);
982:
983: /* find_splittable_regs may have created some new registers, so must
984: reallocate the reg_map with the new larger size, and must realloc
985: the constant maps also. */
986:
987: maxregnum = max_reg_num ();
988: map->reg_map = (rtx *) alloca (maxregnum * sizeof (rtx));
989:
990: init_reg_map (map, maxregnum);
991:
992: /* Space is needed in some of the map for new registers, so new_maxregnum
993: is an (over)estimate of how many registers will exist at the end. */
994: new_maxregnum = maxregnum + (temp * unroll_number * 2);
995:
996: /* Must realloc space for the constant maps, because the number of registers
997: may have changed. */
998:
999: map->const_equiv_map = (rtx *) alloca (new_maxregnum * sizeof (rtx));
1000: map->const_age_map = (unsigned *) alloca (new_maxregnum * sizeof (unsigned));
1001:
1002: global_const_equiv_map = map->const_equiv_map;
1003:
1004: /* Search the list of bivs and givs to find ones which need to be remapped
1005: when split, and set their reg_map entry appropriately. */
1006:
1007: for (bl = loop_iv_list; bl; bl = bl->next)
1008: {
1009: if (REGNO (bl->biv->src_reg) != bl->regno)
1010: map->reg_map[bl->regno] = bl->biv->src_reg;
1011: #if 0
1012: /* Currently, non-reduced/final-value givs are never split. */
1013: for (v = bl->giv; v; v = v->next_iv)
1014: if (REGNO (v->src_reg) != bl->regno)
1015: map->reg_map[REGNO (v->dest_reg)] = v->src_reg;
1016: #endif
1017: }
1018:
1019: /* If the loop is being partially unrolled, and the iteration variables
1020: are being split, and are being renamed for the split, then must fix up
1021: the compare instruction at the end of the loop to refer to the new
1022: registers. This compare isn't copied, so the registers used in it
1023: will never be replaced if it isn't done here. */
1024:
1025: if (unroll_type == UNROLL_MODULO)
1026: {
1027: insn = NEXT_INSN (copy_end);
1028: if (GET_CODE (insn) == INSN && GET_CODE (PATTERN (insn)) == SET)
1029: {
1030: #if 0
1031: /* If non-reduced/final-value givs were split, then this would also
1032: have to remap those givs. */
1033: #endif
1034:
1035: tem = SET_SRC (PATTERN (insn));
1036: /* The set source is a register. */
1037: if (GET_CODE (tem) == REG)
1038: {
1039: if (REGNO (tem) < max_reg_before_loop
1040: && reg_iv_type[REGNO (tem)] == BASIC_INDUCT)
1041: SET_SRC (PATTERN (insn))
1042: = reg_biv_class[REGNO (tem)]->biv->src_reg;
1043: }
1044: else
1045: {
1046: /* The set source is a compare of some sort. */
1047: tem = XEXP (SET_SRC (PATTERN (insn)), 0);
1048: if (GET_CODE (tem) == REG
1049: && REGNO (tem) < max_reg_before_loop
1050: && reg_iv_type[REGNO (tem)] == BASIC_INDUCT)
1051: XEXP (SET_SRC (PATTERN (insn)), 0)
1052: = reg_biv_class[REGNO (tem)]->biv->src_reg;
1053:
1054: tem = XEXP (SET_SRC (PATTERN (insn)), 1);
1055: if (GET_CODE (tem) == REG
1056: && REGNO (tem) < max_reg_before_loop
1057: && reg_iv_type[REGNO (tem)] == BASIC_INDUCT)
1058: XEXP (SET_SRC (PATTERN (insn)), 1)
1059: = reg_biv_class[REGNO (tem)]->biv->src_reg;
1060: }
1061: }
1062: }
1063:
1064: /* For unroll_number - 1 times, make a copy of each instruction
1065: between copy_start and copy_end, and insert these new instructions
1066: before the end of the loop. */
1067:
1068: for (i = 0; i < unroll_number; i++)
1069: {
1070: bzero (map->insn_map, max_insnno * sizeof (rtx));
1071: bzero (map->const_equiv_map, new_maxregnum * sizeof (rtx));
1072: bzero (map->const_age_map, new_maxregnum * sizeof (unsigned));
1073: map->const_age = 0;
1074:
1075: for (j = 0; j < max_labelno; j++)
1076: if (local_label[j])
1077: map->label_map[j] = gen_label_rtx ();
1078:
1079: /* If loop starts with a branch to the test, then fix it so that
1080: it points to the test of the first unrolled copy of the loop. */
1081: if (i == 0 && loop_start != copy_start)
1082: {
1083: insn = PREV_INSN (copy_start);
1084: pattern = PATTERN (insn);
1085:
1086: tem = map->label_map[CODE_LABEL_NUMBER
1087: (XEXP (SET_SRC (pattern), 0))];
1088: SET_SRC (pattern) = gen_rtx (LABEL_REF, VOIDmode, tem);
1089:
1090: /* Set the jump label so that it can be used by later loop unrolling
1091: passes. */
1092: JUMP_LABEL (insn) = tem;
1093: LABEL_NUSES (tem)++;
1094: }
1095:
1096: copy_loop_body (copy_start, copy_end, map, exit_label,
1097: i == unroll_number - 1, unroll_type, start_label,
1098: loop_end, insert_before, insert_before);
1099: }
1100:
1101: /* Before deleting any insns, emit a CODE_LABEL immediately after the last
1102: insn to be deleted. This prevents any runaway delete_insn call from
1103: more insns that it should, as it always stops at a CODE_LABEL. */
1104:
1105: /* Delete the compare and branch at the end of the loop if completely
1106: unrolling the loop. Deleting the backward branch at the end also
1107: deletes the code label at the start of the loop. This is done at
1108: the very end to avoid problems with back_branch_in_range_p. */
1109:
1110: if (unroll_type == UNROLL_COMPLETELY)
1111: safety_label = emit_label_after (gen_label_rtx (), last_loop_insn);
1112: else
1113: safety_label = emit_label_after (gen_label_rtx (), copy_end);
1114:
1115: /* Delete all of the original loop instructions. Don't delete the
1116: LOOP_BEG note, or the first code label in the loop. */
1117:
1118: insn = NEXT_INSN (copy_start);
1119: while (insn != safety_label)
1120: {
1121: if (insn != start_label)
1122: insn = delete_insn (insn);
1123: else
1124: insn = NEXT_INSN (insn);
1125: }
1126:
1127: /* Can now delete the 'safety' label emitted to protect us from runaway
1128: delete_insn calls. */
1129: if (INSN_DELETED_P (safety_label))
1130: abort ();
1131: delete_insn (safety_label);
1132:
1133: /* If exit_label exists, emit it after the loop. Doing the emit here
1134: forces it to have a higher INSN_UID than any insn in the unrolled loop.
1135: This is needed so that mostly_true_jump in reorg.c will treat jumps
1136: to this loop end label correctly, i.e. predict that they are usually
1137: not taken. */
1138: if (exit_label)
1139: emit_label_after (exit_label, loop_end);
1140:
1141: /* If debugging, we must replicate the tree nodes corresponsing to the blocks
1142: inside the loop, so that the original one to one mapping will remain. */
1143:
1144: if (write_symbols != NO_DEBUG)
1145: {
1146: int copies = unroll_number;
1147:
1148: if (loop_preconditioned)
1149: copies += unroll_number - 1;
1150:
1151: unroll_block_trees (uid_loop_num[INSN_UID (loop_start)], copies);
1152: }
1153: }
1154:
1155: /* Return true if the loop can be safely, and profitably, preconditioned
1156: so that the unrolled copies of the loop body don't need exit tests.
1157:
1158: This only works if final_value, initial_value and increment can be
1159: determined, and if increment is a constant power of 2.
1160: If increment is not a power of 2, then the preconditioning modulo
1161: operation would require a real modulo instead of a boolean AND, and this
1162: is not considered `profitable'. */
1163:
1164: /* ??? If the loop is known to be executed very many times, or the machine
1165: has a very cheap divide instruction, then preconditioning is a win even
1166: when the increment is not a power of 2. Use RTX_COST to compute
1167: whether divide is cheap. */
1168:
1169: static int
1170: precondition_loop_p (initial_value, final_value, increment, loop_start,
1171: loop_end)
1172: rtx *initial_value, *final_value, *increment;
1173: rtx loop_start, loop_end;
1174: {
1175: int unsigned_compare, compare_dir;
1176:
1177: if (loop_n_iterations > 0)
1178: {
1179: *initial_value = const0_rtx;
1180: *increment = const1_rtx;
1181: *final_value = gen_rtx (CONST_INT, VOIDmode, loop_n_iterations);
1182:
1183: if (loop_dump_stream)
1184: fprintf (loop_dump_stream,
1185: "Preconditioning: Success, number of iterations known, %d.\n",
1186: loop_n_iterations);
1187: return 1;
1188: }
1189:
1190: if (loop_initial_value == 0)
1191: {
1192: if (loop_dump_stream)
1193: fprintf (loop_dump_stream,
1194: "Preconditioning: Could not find initial value.\n");
1195: return 0;
1196: }
1197: else if (loop_increment == 0)
1198: {
1199: if (loop_dump_stream)
1200: fprintf (loop_dump_stream,
1201: "Preconditioning: Could not find increment value.\n");
1202: return 0;
1203: }
1204: else if (GET_CODE (loop_increment) != CONST_INT)
1205: {
1206: if (loop_dump_stream)
1207: fprintf (loop_dump_stream,
1208: "Preconditioning: Increment not a constant.\n");
1209: return 0;
1210: }
1211: else if ((exact_log2 (INTVAL (loop_increment)) < 0)
1212: && (exact_log2 (- INTVAL (loop_increment)) < 0))
1213: {
1214: if (loop_dump_stream)
1215: fprintf (loop_dump_stream,
1216: "Preconditioning: Increment not a constant power of 2.\n");
1217: return 0;
1218: }
1219:
1220: /* Unsigned_compare and compare_dir can be ignored here, since they do
1221: not matter for preconditioning. */
1222:
1223: if (loop_final_value == 0)
1224: {
1225: if (loop_dump_stream)
1226: fprintf (loop_dump_stream,
1227: "Preconditioning: EQ comparison loop.\n");
1228: return 0;
1229: }
1230:
1231: /* Must ensure that final_value is invariant, so call invariant_p to
1232: check. Before doing so, must check regno against max_reg_before_loop
1233: to make sure that the register is in the range convered by invariant_p.
1234: If it isn't, then it is most likely a biv/giv which by definition are
1235: not invariant. */
1236: if ((GET_CODE (loop_final_value) == REG
1237: && REGNO (loop_final_value) >= max_reg_before_loop)
1238: || (GET_CODE (loop_final_value) == PLUS
1239: && REGNO (XEXP (loop_final_value, 0)) >= max_reg_before_loop)
1240: || ! invariant_p (loop_final_value))
1241: {
1242: if (loop_dump_stream)
1243: fprintf (loop_dump_stream,
1244: "Preconditioning: Final value not invariant.\n");
1245: return 0;
1246: }
1247:
1248: /* Fail for floating point values, since the caller of this function
1249: does not have code to deal with them. */
1250: if (GET_MODE_CLASS (GET_MODE (loop_final_value)) == MODE_FLOAT
1251: || GET_MODE_CLASS (GET_MODE (loop_initial_value) == MODE_FLOAT))
1252: {
1253: if (loop_dump_stream)
1254: fprintf (loop_dump_stream,
1255: "Preconditioning: Floating point final or initial value.\n");
1256: return 0;
1257: }
1258:
1259: /* Now set initial_value to be the iteration_var, since that may be a
1260: simpler expression, and is guaranteed to be correct if all of the
1261: above tests succeed.
1262:
1263: We can not use the initial_value as calculated, because it will be
1264: one too small for loops of the form "while (i-- > 0)". We can not
1265: emit code before the loop_skip_over insns to fix this problem as this
1266: will then give a number one too large for loops of the form
1267: "while (--i > 0)".
1268:
1269: Note that all loops that reach here are entered at the top, because
1270: this function is not called if the loop starts with a jump. */
1271:
1272: /* Fail if loop_iteration_var is not live before loop_start, since we need
1273: to test its value in the preconditioning code. */
1274:
1275: if (uid_luid[regno_first_uid[REGNO (loop_iteration_var)]]
1276: > INSN_LUID (loop_start))
1277: {
1278: if (loop_dump_stream)
1279: fprintf (loop_dump_stream,
1280: "Preconditioning: Iteration var not live before loop start.\n");
1281: return 0;
1282: }
1283:
1284: *initial_value = loop_iteration_var;
1285: *increment = loop_increment;
1286: *final_value = loop_final_value;
1287:
1288: /* Success! */
1289: if (loop_dump_stream)
1290: fprintf (loop_dump_stream, "Preconditioning: Successful.\n");
1291: return 1;
1292: }
1293:
1294:
1295: /* All pseudo-registers must be mapped to themselves. Two hard registers
1296: must be mapped, VIRTUAL_STACK_VARS_REGNUM and VIRTUAL_INCOMING_ARGS_
1297: REGNUM, to avoid function-inlining specific conversions of these
1298: registers. All other hard regs can not be mapped because they may be
1299: used with different
1300: modes. */
1301:
1302: static void
1303: init_reg_map (map, maxregnum)
1304: struct inline_remap *map;
1305: int maxregnum;
1306: {
1307: int i;
1308:
1309: for (i = maxregnum - 1; i > LAST_VIRTUAL_REGISTER; i--)
1310: map->reg_map[i] = regno_reg_rtx[i];
1311: /* Just clear the rest of the entries. */
1312: for (i = LAST_VIRTUAL_REGISTER; i >= 0; i--)
1313: map->reg_map[i] = 0;
1314:
1315: map->reg_map[VIRTUAL_STACK_VARS_REGNUM]
1316: = regno_reg_rtx[VIRTUAL_STACK_VARS_REGNUM];
1317: map->reg_map[VIRTUAL_INCOMING_ARGS_REGNUM]
1318: = regno_reg_rtx[VIRTUAL_INCOMING_ARGS_REGNUM];
1319: }
1320:
1321: /* Strength-reduction will often emit code for optimized biv/givs which
1322: calculates their value in a temporary register, and then copies the result
1323: to the iv. This procedure reconstructs the pattern computing the iv;
1324: verifying that all operands are of the proper form.
1325:
1326: The return value is the amount that the giv is incremented by. */
1327:
1328: static rtx
1329: calculate_giv_inc (pattern, src_insn, regno)
1330: rtx pattern, src_insn;
1331: int regno;
1332: {
1333: rtx increment;
1334:
1335: /* Verify that we have an increment insn here. First check for a plus
1336: as the set source. */
1337: if (GET_CODE (SET_SRC (pattern)) != PLUS)
1338: {
1339: /* SR sometimes computes the new giv value in a temp, then copies it
1340: to the new_reg. */
1341: src_insn = PREV_INSN (src_insn);
1342: pattern = PATTERN (src_insn);
1343: if (GET_CODE (SET_SRC (pattern)) != PLUS)
1344: abort ();
1345:
1346: /* The last insn emitted is not needed, so delete it to avoid confusing
1347: the second cse pass. This insn sets the giv unnecessarily. */
1348: delete_insn (get_last_insn ());
1349: }
1350:
1351: /* Verify that we have a constant as the second operand of the plus. */
1352: increment = XEXP (SET_SRC (pattern), 1);
1353: if (GET_CODE (increment) != CONST_INT)
1354: {
1355: /* SR sometimes puts the constant in a register, especially if it is
1356: too big to be an add immed operand. */
1357: increment = SET_SRC (PATTERN (PREV_INSN (src_insn)));
1358:
1359: /* SR may have used LO_SUM to compute the constant if it is too large
1360: for a load immed operand. In this case, the constant is in operand
1361: one of the LO_SUM rtx. */
1362: if (GET_CODE (increment) == LO_SUM)
1363: increment = XEXP (increment, 1);
1364:
1365: if (GET_CODE (increment) != CONST_INT)
1366: abort ();
1367:
1368: /* The insn loading the constant into a register is not longer needed,
1369: so delete it. */
1370: delete_insn (get_last_insn ());
1371: }
1372:
1373: /* Check that the source register is the same as the dest register. */
1374: if (GET_CODE (XEXP (SET_SRC (pattern), 0)) != REG
1375: || REGNO (XEXP (SET_SRC (pattern), 0)) != regno)
1376: abort ();
1377:
1378: return increment;
1379: }
1380:
1381:
1382: /* Copy each instruction in the loop, substituting from map as appropriate.
1383: This is very similar to a loop in expand_inline_function. */
1384:
1385: static void
1386: copy_loop_body (copy_start, copy_end, map, exit_label, last_iteration,
1387: unroll_type, start_label, loop_end, insert_before,
1388: copy_notes_from)
1389: rtx copy_start, copy_end;
1390: struct inline_remap *map;
1391: int last_iteration;
1392: enum unroll_types unroll_type;
1393: rtx start_label, loop_end, insert_before, copy_notes_from;
1394: {
1395: rtx insn, pattern;
1396: rtx tem, copy;
1397: int dest_reg_was_split, i;
1398: rtx cc0_insn = 0;
1399: rtx final_label = 0;
1400: rtx giv_inc, giv_dest_reg, giv_src_reg;
1401:
1402: /* If this isn't the last iteration, then map any references to the
1403: start_label to final_label. Final label will then be emitted immediately
1404: after the end of this loop body if it was ever used.
1405:
1406: If this is the last iteration, then map references to the start_label
1407: to itself. */
1408: if (! last_iteration)
1409: {
1410: final_label = gen_label_rtx ();
1411: map->label_map[CODE_LABEL_NUMBER (start_label)] = final_label;
1412: }
1413: else
1414: map->label_map[CODE_LABEL_NUMBER (start_label)] = start_label;
1415:
1416: start_sequence ();
1417:
1418: insn = copy_start;
1419: do
1420: {
1421: insn = NEXT_INSN (insn);
1422:
1423: map->orig_asm_operands_vector = 0;
1424:
1425: switch (GET_CODE (insn))
1426: {
1427: case INSN:
1428: pattern = PATTERN (insn);
1429: copy = 0;
1430: giv_inc = 0;
1431:
1432: /* Check to see if this is a giv that has been combined with
1433: some split address givs. (Combined in the sense that
1434: `combine_givs' in loop.c has put two givs in the same register.)
1435: In this case, we must search all givs based on the same biv to
1436: find the address givs. Then split the address givs.
1437: Do this before splitting the giv, since that may map the
1438: SET_DEST to a new register. */
1439:
1440: if (GET_CODE (pattern) == SET
1441: && GET_CODE (SET_DEST (pattern)) == REG
1442: && addr_combined_regs[REGNO (SET_DEST (pattern))])
1443: {
1444: struct iv_class *bl;
1445: struct induction *v, *tv;
1446: int regno = REGNO (SET_DEST (pattern));
1447:
1448: v = addr_combined_regs[REGNO (SET_DEST (pattern))];
1449: bl = reg_biv_class[REGNO (v->src_reg)];
1450:
1451: /* Although the giv_inc amount is not needed here, we must call
1452: calculate_giv_inc here since it might try to delete the
1453: last insn emitted. If we wait until later to call it,
1454: we might accidentally delete insns generated immediately
1455: below by emit_unrolled_add. */
1456:
1457: giv_inc = calculate_giv_inc (pattern, insn, regno);
1458:
1459: /* Now find all address giv's that were combined with this
1460: giv 'v'. */
1461: for (tv = bl->giv; tv; tv = tv->next_iv)
1462: if (tv->giv_type == DEST_ADDR && tv->same == v)
1463: {
1464: tv->dest_reg = plus_constant (tv->dest_reg,
1465: INTVAL (giv_inc));
1466: *tv->location = tv->dest_reg;
1467:
1468: if (last_iteration && unroll_type != UNROLL_COMPLETELY)
1469: {
1470: /* Must emit an insn to increment the split address
1471: giv. Add in the const_adjust field in case there
1472: was a constant eliminated from the address. */
1473: rtx value, dest_reg;
1474:
1475: /* tv->dest_reg will be either a bare register,
1476: or else a register plus a constant. */
1477: if (GET_CODE (tv->dest_reg) == REG)
1478: dest_reg = tv->dest_reg;
1479: else
1480: dest_reg = XEXP (tv->dest_reg, 0);
1481:
1482: /* tv->dest_reg may actually be a (PLUS (REG) (CONST))
1483: here, so we must call plus_constant to add
1484: the const_adjust amount before calling
1485: emit_unrolled_add below. */
1486: value = plus_constant (tv->dest_reg, tv->const_adjust);
1487:
1488: /* The constant could be too large for an add
1489: immediate, so can't directly emit an insn here. */
1490: emit_unrolled_add (dest_reg, XEXP (value, 0),
1491: XEXP (value, 1));
1492:
1493: /* Reset the giv to be just the register again, in case
1494: it is used after the set we have just emitted. */
1495: tv->dest_reg = dest_reg;
1496: *tv->location = tv->dest_reg;
1497: }
1498: }
1499: }
1500:
1501: /* If this is a setting of a splittable variable, then determine
1502: how to split the variable, create a new set based on this split,
1503: and set up the reg_map so that later uses of the variable will
1504: use the new split variable. */
1505:
1506: dest_reg_was_split = 0;
1507:
1508: if (GET_CODE (pattern) == SET
1509: && GET_CODE (SET_DEST (pattern)) == REG
1510: && splittable_regs[REGNO (SET_DEST (pattern))])
1511: {
1512: int regno = REGNO (SET_DEST (pattern));
1513:
1514: dest_reg_was_split = 1;
1515:
1516: /* Compute the increment value for the giv, if it wasn't
1517: already computed above. */
1518:
1519: if (giv_inc == 0)
1520: giv_inc = calculate_giv_inc (pattern, insn, regno);
1521: giv_dest_reg = SET_DEST (pattern);
1522: giv_src_reg = SET_DEST (pattern);
1523:
1524: if (unroll_type == UNROLL_COMPLETELY)
1525: {
1526: /* Completely unrolling the loop. Set the induction
1527: variable to a known constant value. */
1528:
1529: /* The value in splittable_regs may be an invariant
1530: value, so we must use plus_constant here. */
1531: splittable_regs[regno]
1532: = plus_constant (splittable_regs[regno], INTVAL (giv_inc));
1533:
1534: if (GET_CODE (splittable_regs[regno]) == PLUS)
1535: {
1536: giv_src_reg = XEXP (splittable_regs[regno], 0);
1537: giv_inc = XEXP (splittable_regs[regno], 1);
1538: }
1539: else
1540: {
1541: /* The splittable_regs value must be a REG or a
1542: CONST_INT, so put the entire value in the giv_src_reg
1543: variable. */
1544: giv_src_reg = splittable_regs[regno];
1545: giv_inc = const0_rtx;
1546: }
1547: }
1548: else
1549: {
1550: /* Partially unrolling loop. Create a new pseudo
1551: register for the iteration variable, and set it to
1552: be a constant plus the original register. Except
1553: on the last iteration, when the result has to
1554: go back into the original iteration var register. */
1555:
1556: /* Handle bivs which must be mapped to a new register
1557: when split. This happens for bivs which need their
1558: final value set before loop entry. The new register
1559: for the biv was stored in the biv's first struct
1560: induction entry by find_splittable_regs. */
1561:
1562: if (regno < max_reg_before_loop
1563: && reg_iv_type[regno] == BASIC_INDUCT)
1564: {
1565: giv_src_reg = reg_biv_class[regno]->biv->src_reg;
1566: giv_dest_reg = giv_src_reg;
1567: }
1568:
1569: #if 0
1570: /* If non-reduced/final-value givs were split, then
1571: this would have to remap those givs also. See
1572: find_splittable_regs. */
1573: #endif
1574:
1575: splittable_regs[regno]
1576: = gen_rtx (CONST_INT, VOIDmode,
1577: INTVAL (giv_inc)
1578: + INTVAL (splittable_regs[regno]));
1579: giv_inc = splittable_regs[regno];
1580:
1581: /* Now split the induction variable by changing the dest
1582: of this insn to a new register, and setting its
1583: reg_map entry to point to this new register.
1584:
1585: If this is the last iteration, and this is the last insn
1586: that will update the iv, then reuse the original dest,
1587: to ensure that the iv will have the proper value when
1588: the loop exits or repeats.
1589:
1590: Using splittable_regs_updates here like this is safe,
1591: because it can only be greater than one if all
1592: instructions modifying the iv are always executed in
1593: order. */
1594:
1595: if (! last_iteration
1596: || (splittable_regs_updates[regno]-- != 1))
1597: {
1598: tem = gen_reg_rtx (GET_MODE (giv_src_reg));
1599: giv_dest_reg = tem;
1600: map->reg_map[regno] = tem;
1601: }
1602: else
1603: map->reg_map[regno] = giv_src_reg;
1604: }
1605:
1606: /* The constant being added could be too large for an add
1607: immediate, so can't directly emit an insn here. */
1608: emit_unrolled_add (giv_dest_reg, giv_src_reg, giv_inc);
1609: copy = get_last_insn ();
1610: pattern = PATTERN (copy);
1611: }
1612: else
1613: {
1614: pattern = copy_rtx_and_substitute (pattern, map);
1615: copy = emit_insn (pattern);
1616: }
1617: /* REG_NOTES will be copied later. */
1618:
1619: #ifdef HAVE_cc0
1620: /* If this insn is setting CC0, it may need to look at
1621: the insn that uses CC0 to see what type of insn it is.
1622: In that case, the call to recog via validate_change will
1623: fail. So don't substitute constants here. Instead,
1624: do it when we emit the following insn.
1625:
1626: For example, see the pyr.md file. That machine has signed and
1627: unsigned compares. The compare patterns must check the
1628: following branch insn to see which what kind of compare to
1629: emit.
1630:
1631: If the previous insn set CC0, substitute constants on it as
1632: well. */
1633: if (sets_cc0_p (copy) != 0)
1634: cc0_insn = copy;
1635: else
1636: {
1637: if (cc0_insn)
1638: try_constants (cc0_insn, map);
1639: cc0_insn = 0;
1640: try_constants (copy, map);
1641: }
1642: #else
1643: try_constants (copy, map);
1644: #endif
1645:
1646: /* Make split induction variable constants `permanent' since we
1647: know there are no backward branches across iteration variable
1648: settings which would invalidate this. */
1649: if (dest_reg_was_split)
1650: {
1651: int regno = REGNO (SET_DEST (pattern));
1652:
1653: if (map->const_age_map[regno] == map->const_age)
1654: map->const_age_map[regno] = -1;
1655: }
1656: break;
1657:
1658: case JUMP_INSN:
1659: if (JUMP_LABEL (insn) == start_label && insn == copy_end
1660: && ! last_iteration)
1661: {
1662: /* This is a branch to the beginning of the loop; this is the
1663: last insn being copied; and this is not the last iteration.
1664: In this case, we want to change the original fall through
1665: case to be a branch past the end of the loop, and the
1666: original jump label case to fall_through. */
1667:
1668: int fall_through;
1669:
1670: /* Never map the label in this case. */
1671: pattern = copy_rtx (PATTERN (insn));
1672:
1673: /* Assume a conditional branch, since the code above
1674: does not let unconditional branches be copied. */
1675: if (! condjump_p (insn))
1676: abort ();
1677: fall_through
1678: = (XEXP (SET_SRC (PATTERN (insn)), 2) == pc_rtx) + 1;
1679:
1680: /* Set the fall through case to the exit label. Must
1681: create a new label_ref since they can't be shared. */
1682: XEXP (SET_SRC (pattern), fall_through)
1683: = gen_rtx (LABEL_REF, VOIDmode, exit_label);
1684:
1685: /* Set the original branch case to fall through. */
1686: XEXP (SET_SRC (pattern), 3 - fall_through)
1687: = pc_rtx;
1688: }
1689: else
1690: pattern = copy_rtx_and_substitute (PATTERN (insn), map);
1691:
1692: copy = emit_jump_insn (pattern);
1693:
1694: #ifdef HAVE_cc0
1695: if (cc0_insn)
1696: try_constants (cc0_insn, map);
1697: cc0_insn = 0;
1698: #endif
1699: try_constants (copy, map);
1700:
1701: /* Set the jump label of COPY correctly to avoid problems with
1702: later passes of unroll_loop, if INSN had jump label set. */
1703: if (JUMP_LABEL (insn))
1704: {
1705: /* Can't use the label_map for every insn, since this may be
1706: the backward branch, and hence the label was not mapped. */
1707: if (GET_CODE (pattern) == SET)
1708: {
1709: tem = SET_SRC (pattern);
1710: if (GET_CODE (tem) == LABEL_REF)
1711: JUMP_LABEL (copy) = XEXP (tem, 0);
1712: else if (GET_CODE (tem) == IF_THEN_ELSE)
1713: {
1714: if (XEXP (tem, 1) != pc_rtx)
1715: JUMP_LABEL (copy) = XEXP (XEXP (tem, 1), 0);
1716: else
1717: JUMP_LABEL (copy) = XEXP (XEXP (tem, 2), 0);
1718: }
1719: else
1720: abort ();
1721: }
1722: else
1723: {
1724: /* An unrecognizable jump insn, probably the entry jump
1725: for a switch statement. This label must have been mapped,
1726: so just use the label_map to get the new jump label. */
1727: JUMP_LABEL (copy) = map->label_map[CODE_LABEL_NUMBER
1728: (JUMP_LABEL (insn))];
1729: }
1730:
1731: /* If this is a non-local jump, then must increase the label
1732: use count so that the label will not be deleted when the
1733: original jump is deleted. */
1734: LABEL_NUSES (JUMP_LABEL (copy))++;
1735: }
1736: else if (GET_CODE (PATTERN (copy)) == ADDR_VEC
1737: || GET_CODE (PATTERN (copy)) == ADDR_DIFF_VEC)
1738: {
1739: rtx pat = PATTERN (copy);
1740: int diff_vec_p = GET_CODE (pat) == ADDR_DIFF_VEC;
1741: int len = XVECLEN (pat, diff_vec_p);
1742: int i;
1743:
1744: for (i = 0; i < len; i++)
1745: LABEL_NUSES (XEXP (XVECEXP (pat, diff_vec_p, i), 0))++;
1746: }
1747:
1748: /* If this used to be a conditional jump insn but whose branch
1749: direction is now known, we must do something special. */
1750: if (condjump_p (insn) && !simplejump_p (insn) && map->last_pc_value)
1751: {
1752: #ifdef HAVE_cc0
1753: /* The previous insn set cc0 for us. So delete it. */
1754: delete_insn (PREV_INSN (copy));
1755: #endif
1756:
1757: /* If this is now a no-op, delete it. */
1758: if (map->last_pc_value == pc_rtx)
1759: {
1760: delete_insn (copy);
1761: copy = 0;
1762: }
1763: else
1764: /* Otherwise, this is unconditional jump so we must put a
1765: BARRIER after it. We could do some dead code elimination
1766: here, but jump.c will do it just as well. */
1767: emit_barrier ();
1768: }
1769: break;
1770:
1771: case CALL_INSN:
1772: pattern = copy_rtx_and_substitute (PATTERN (insn), map);
1773: copy = emit_call_insn (pattern);
1774:
1775: #ifdef HAVE_cc0
1776: if (cc0_insn)
1777: try_constants (cc0_insn, map);
1778: cc0_insn = 0;
1779: #endif
1780: try_constants (copy, map);
1781:
1782: /* Be lazy and assume CALL_INSNs clobber all hard registers. */
1783: for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
1784: map->const_equiv_map[i] = 0;
1785: break;
1786:
1787: case CODE_LABEL:
1788: /* If this is the loop start label, then we don't need to emit a
1789: copy of this label since no one will use it. */
1790:
1791: if (insn != start_label)
1792: {
1793: copy = emit_label (map->label_map[CODE_LABEL_NUMBER (insn)]);
1794: map->const_age++;
1795: }
1796: break;
1797:
1798: case BARRIER:
1799: copy = emit_barrier ();
1800: break;
1801:
1802: case NOTE:
1803: if (NOTE_LINE_NUMBER (insn) != NOTE_INSN_DELETED)
1804: copy = emit_note (NOTE_SOURCE_FILE (insn),
1805: NOTE_LINE_NUMBER (insn));
1806: else
1807: copy = 0;
1808: break;
1809:
1810: default:
1811: abort ();
1812: break;
1813: }
1814:
1815: map->insn_map[INSN_UID (insn)] = copy;
1816: }
1817: while (insn != copy_end);
1818:
1819: /* Now copy the REG_NOTES. */
1820: insn = copy_start;
1821: do
1822: {
1823: insn = NEXT_INSN (insn);
1824: if ((GET_CODE (insn) == INSN || GET_CODE (insn) == JUMP_INSN
1825: || GET_CODE (insn) == CALL_INSN)
1826: && map->insn_map[INSN_UID (insn)])
1827: REG_NOTES (map->insn_map[INSN_UID (insn)])
1828: = copy_rtx_and_substitute (REG_NOTES (insn), map);
1829: }
1830: while (insn != copy_end);
1831:
1832: /* There may be notes between copy_notes_from and loop_end. Emit a copy of
1833: each of these notes here, since there may be some important ones, such as
1834: NOTE_INSN_BLOCK_END notes, in this group. We don't do this on the last
1835: iteration, because the original notes won't be deleted.
1836:
1837: We can't use insert_before here, because when from preconditioning,
1838: insert_before points before the loop. We can't use copy_end, because
1839: there may be insns already inserted after it (which we don't want to
1840: copy) when not from preconditioning code. */
1841:
1842: if (! last_iteration)
1843: {
1844: for (insn = copy_notes_from; insn != loop_end; insn = NEXT_INSN (insn))
1845: {
1846: if (GET_CODE (insn) == NOTE
1847: && NOTE_LINE_NUMBER (insn) != NOTE_INSN_DELETED)
1848: emit_note (NOTE_SOURCE_FILE (insn), NOTE_LINE_NUMBER (insn));
1849: }
1850: }
1851:
1852: if (final_label && LABEL_NUSES (final_label) > 0)
1853: emit_label (final_label);
1854:
1855: tem = gen_sequence ();
1856: end_sequence ();
1857: emit_insn_before (tem, insert_before);
1858: }
1859:
1860: /* Emit an insn, using the expand_binop to ensure that a valid insn is
1861: emitted. This will correctly handle the case where the increment value
1862: won't fit in the immediate field of a PLUS insns. */
1863:
1864: void
1865: emit_unrolled_add (dest_reg, src_reg, increment)
1866: rtx dest_reg, src_reg, increment;
1867: {
1868: rtx result;
1869:
1870: result = expand_binop (GET_MODE (dest_reg), add_optab, src_reg, increment,
1871: dest_reg, 0, OPTAB_LIB_WIDEN);
1872:
1873: if (dest_reg != result)
1874: emit_move_insn (dest_reg, result);
1875: }
1876:
1877: /* Searches the insns between INSN and LOOP_END. Returns 1 if there
1878: is a backward branch in that range that branches to somewhere between
1879: LOOP_START and INSN. Returns 0 otherwise. */
1880:
1881: /* ??? This is quadratic algorithm. Could be rewriten to be linear.
1882: In practice, this is not a problem, because this function is seldom called,
1883: and uses a negligible amount of CPU time on average. */
1884:
1885: static int
1886: back_branch_in_range_p (insn, loop_start, loop_end)
1887: rtx insn;
1888: rtx loop_start, loop_end;
1889: {
1890: rtx p, q, target_insn;
1891:
1892: /* Stop before we get to the backward branch at the end of the loop. */
1893: loop_end = prev_nonnote_insn (loop_end);
1894: if (GET_CODE (loop_end) == BARRIER)
1895: loop_end = PREV_INSN (loop_end);
1896:
1897: /* Check in case insn has been deleted, search forward for first non
1898: deleted insn following it. */
1899: while (INSN_DELETED_P (insn))
1900: insn = NEXT_INSN (insn);
1901:
1902: /* Check for the case where insn is the last insn in the loop. */
1903: if (insn == loop_end)
1904: return 0;
1905:
1906: for (p = NEXT_INSN (insn); p != loop_end; p = NEXT_INSN (p))
1907: {
1908: if (GET_CODE (p) == JUMP_INSN)
1909: {
1910: target_insn = JUMP_LABEL (p);
1911:
1912: /* Search from loop_start to insn, to see if one of them is
1913: the target_insn. We can't use INSN_LUID comparisons here,
1914: since insn may not have an LUID entry. */
1915: for (q = loop_start; q != insn; q = NEXT_INSN (q))
1916: if (q == target_insn)
1917: return 1;
1918: }
1919: }
1920:
1921: return 0;
1922: }
1923:
1924: /* Try to generate the simplest rtx for the expression
1925: (PLUS (MULT mult1 mult2) add1). This is used to calculate the initial
1926: value of giv's. */
1927:
1928: static rtx
1929: fold_rtx_mult_add (mult1, mult2, add1, mode)
1930: rtx mult1, mult2, add1;
1931: enum machine_mode mode;
1932: {
1933: rtx temp, mult_res;
1934: rtx result;
1935:
1936: /* The modes must all be the same. This should always be true. For now,
1937: check to make sure. */
1938: if ((GET_MODE (mult1) != mode && GET_MODE (mult1) != VOIDmode)
1939: || (GET_MODE (mult2) != mode && GET_MODE (mult2) != VOIDmode)
1940: || (GET_MODE (add1) != mode && GET_MODE (add1) != VOIDmode))
1941: abort ();
1942:
1943: /* Ensure that if at least one of mult1/mult2 are constant, then mult2
1944: will be a constant. */
1945: if (GET_CODE (mult1) == CONST_INT)
1946: {
1947: temp = mult2;
1948: mult2 = mult1;
1949: mult1 = temp;
1950: }
1951:
1952: mult_res = simplify_binary_operation (MULT, mode, mult1, mult2);
1953: if (! mult_res)
1954: mult_res = gen_rtx (MULT, mode, mult1, mult2);
1955:
1956: /* Again, put the constant second. */
1957: if (GET_CODE (add1) == CONST_INT)
1958: {
1959: temp = add1;
1960: add1 = mult_res;
1961: mult_res = temp;
1962: }
1963:
1964: result = simplify_binary_operation (PLUS, mode, add1, mult_res);
1965: if (! result)
1966: result = gen_rtx (PLUS, mode, add1, mult_res);
1967:
1968: return result;
1969: }
1970:
1971: /* Searches the list of induction struct's for the biv BL, to try to calculate
1972: the total increment value for one iteration of the loop as a constant.
1973:
1974: Returns the increment value as an rtx, simplified as much as possible,
1975: if it can be calculated. Otherwise, returns 0. */
1976:
1977: rtx
1978: biv_total_increment (bl, loop_start, loop_end)
1979: struct iv_class *bl;
1980: rtx loop_start, loop_end;
1981: {
1982: struct induction *v;
1983: rtx result;
1984:
1985: /* For increment, must check every instruction that sets it. Each
1986: instruction must be executed only once each time through the loop.
1987: To verify this, we check that the the insn is always executed, and that
1988: there are no backward branches after the insn that branch to before it.
1989: Also, the insn must have a mult_val of one (to make sure it really is
1990: an increment). */
1991:
1992: result = const0_rtx;
1993: for (v = bl->biv; v; v = v->next_iv)
1994: {
1995: if (v->always_computable && v->mult_val == const1_rtx
1996: && ! back_branch_in_range_p (v->insn, loop_start, loop_end))
1997: result = fold_rtx_mult_add (result, const1_rtx, v->add_val, v->mode);
1998: else
1999: return 0;
2000: }
2001:
2002: return result;
2003: }
2004:
2005: /* Determine the initial value of the iteration variable, and the amount
2006: that it is incremented each loop. Use the tables constructed by
2007: the strength reduction pass to calculate these values.
2008:
2009: Initial_value and/or increment are set to zero if their values could not
2010: be calculated. */
2011:
2012: static void
2013: iteration_info (iteration_var, initial_value, increment, loop_start, loop_end)
2014: rtx iteration_var, *initial_value, *increment;
2015: rtx loop_start, loop_end;
2016: {
2017: struct iv_class *bl;
2018: struct induction *v, *b;
2019:
2020: /* Clear the result values, in case no answer can be found. */
2021: *initial_value = 0;
2022: *increment = 0;
2023:
2024: /* The iteration variable can be either a giv or a biv. Check to see
2025: which it is, and compute the variable's initial value, and increment
2026: value if possible. */
2027:
2028: /* If this is a new register, can't handle it since we don't have any
2029: reg_iv_type entry for it. */
2030: if (REGNO (iteration_var) >= max_reg_before_loop)
2031: {
2032: if (loop_dump_stream)
2033: fprintf (loop_dump_stream,
2034: "Loop unrolling: No reg_iv_type entry for iteration var.\n");
2035: return;
2036: }
2037: /* Reject iteration variables larger than the host long size, since they
2038: could result in a number of iterations greater than the range of our
2039: `unsigned long' variable loop_n_iterations. */
2040: else if (GET_MODE_BITSIZE (GET_MODE (iteration_var)) > HOST_BITS_PER_LONG)
2041: {
2042: if (loop_dump_stream)
2043: fprintf (loop_dump_stream,
2044: "Loop unrolling: Iteration var rejected because mode larger than host long.\n");
2045: return;
2046: }
2047: else if (GET_MODE_CLASS (GET_MODE (iteration_var)) != MODE_INT)
2048: {
2049: if (loop_dump_stream)
2050: fprintf (loop_dump_stream,
2051: "Loop unrolling: Iteration var not an interger.\n");
2052: return;
2053: }
2054: else if (reg_iv_type[REGNO (iteration_var)] == BASIC_INDUCT)
2055: {
2056: /* Grab initial value, only useful if it is a constant. */
2057: bl = reg_biv_class[REGNO (iteration_var)];
2058: *initial_value = bl->initial_value;
2059:
2060: *increment = biv_total_increment (bl, loop_start, loop_end);
2061: }
2062: else if (reg_iv_type[REGNO (iteration_var)] == GENERAL_INDUCT)
2063: {
2064: #if 1
2065: /* ??? The code below does not work because the incorrect number of
2066: iterations is calculated when the biv is incremented after the giv
2067: is set (which is the usual case). This can probably be accounted
2068: for by biasing the initial_value by subtracting the amount of the
2069: increment that occurs between the giv set and the giv test. However,
2070: a giv as an iterator is very rare, so it does not seem worthwhile
2071: to handle this. */
2072: /* ??? An example failure is: i = 6; do {;} while (i++ < 9). */
2073: if (loop_dump_stream)
2074: fprintf (loop_dump_stream,
2075: "Loop unrolling: Giv iterators are not handled.\n");
2076: return;
2077: #else
2078: /* Initial value is mult_val times the biv's initial value plus
2079: add_val. Only useful if it is a constant. */
2080: v = reg_iv_info[REGNO (iteration_var)];
2081: bl = reg_biv_class[REGNO (v->src_reg)];
2082: *initial_value = fold_rtx_mult_add (v->mult_val, bl->initial_value,
2083: v->add_val, v->mode);
2084:
2085: /* Increment value is mult_val times the increment value of the biv. */
2086:
2087: *increment = biv_total_increment (bl, loop_start, loop_end);
2088: if (*increment)
2089: *increment = fold_rtx_mult_add (v->mult_val, *increment, const0_rtx,
2090: v->mode);
2091: #endif
2092: }
2093: else
2094: {
2095: if (loop_dump_stream)
2096: fprintf (loop_dump_stream,
2097: "Loop unrolling: Not basic or general induction var.\n");
2098: return;
2099: }
2100: }
2101:
2102: /* Calculate the approximate final value of the iteration variable
2103: which has an loop exit test with code COMPARISON_CODE and comparison value
2104: of COMPARISON_VALUE. Also returns an indication of whether the comparison
2105: was signed or unsigned, and the direction of the comparison. This info is
2106: needed to calculate the number of loop iterations. */
2107:
2108: static rtx
2109: approx_final_value (comparison_code, comparison_value, unsigned_p, compare_dir)
2110: enum rtx_code comparison_code;
2111: rtx comparison_value;
2112: int *unsigned_p;
2113: int *compare_dir;
2114: {
2115: /* Calculate the final value of the induction variable.
2116: The exact final value depends on the branch operator, and increment sign.
2117: This is only an approximate value. It will be wrong if the iteration
2118: variable is not incremented by one each time through the loop, and
2119: approx final value - start value % increment != 0. */
2120:
2121: *unsigned_p = 0;
2122: switch (comparison_code)
2123: {
2124: case LEU:
2125: *unsigned_p = 1;
2126: case LE:
2127: *compare_dir = 1;
2128: return plus_constant (comparison_value, 1);
2129: case GEU:
2130: *unsigned_p = 1;
2131: case GE:
2132: *compare_dir = -1;
2133: return plus_constant (comparison_value, -1);
2134: case EQ:
2135: /* Can not calculate a final value for this case. */
2136: *compare_dir = 0;
2137: return 0;
2138: case LTU:
2139: *unsigned_p = 1;
2140: case LT:
2141: *compare_dir = 1;
2142: return comparison_value;
2143: break;
2144: case GTU:
2145: *unsigned_p = 1;
2146: case GT:
2147: *compare_dir = -1;
2148: return comparison_value;
2149: case NE:
2150: *compare_dir = 0;
2151: return comparison_value;
2152: default:
2153: abort ();
2154: }
2155: }
2156:
2157: /* For each biv and giv, determine whether it can be safely split into
2158: a different variable for each unrolled copy of the loop body. If it
2159: is safe to split, then indicate that by saving some useful info
2160: in the splittable_regs array.
2161:
2162: If the loop is being completely unrolled, then splittable_regs will hold
2163: the current value of the induction variable while the loop is unrolled.
2164: It must be set to the initial value of the induction variable here.
2165: Otherwise, splittable_regs will hold the difference between the current
2166: value of the induction variable and the value the induction variable had
2167: at the top of the loop. It must be set to the value 0 here. */
2168:
2169: /* ?? If the loop is only unrolled twice, then most of the restrictions to
2170: constant values are unnecessary, since we can easily calculate increment
2171: values in this case even if nothing is constant. The increment value
2172: should not involve a multiply however. */
2173:
2174: /* ?? Even if the biv/giv increment values aren't constant, it may still
2175: be beneficial to split the variable if the loop is only unrolled a few
2176: times, since multiplies by small integers (1,2,3,4) are very cheap. */
2177:
2178: static int
2179: find_splittable_regs (unroll_type, loop_start, loop_end, end_insert_before,
2180: unroll_number)
2181: enum unroll_types unroll_type;
2182: rtx loop_start, loop_end;
2183: rtx end_insert_before;
2184: int unroll_number;
2185: {
2186: struct iv_class *bl;
2187: rtx increment, tem;
2188: rtx biv_final_value;
2189: int biv_splittable;
2190: int result = 0;
2191:
2192: for (bl = loop_iv_list; bl; bl = bl->next)
2193: {
2194: /* Biv_total_increment must return a constant value,
2195: otherwise we can not calculate the split values. */
2196:
2197: increment = biv_total_increment (bl, loop_start, loop_end);
2198: if (! increment || GET_CODE (increment) != CONST_INT)
2199: continue;
2200:
2201: /* The loop must be unrolled completely, or else have a known number
2202: of iterations and only one exit, or else the biv must be dead
2203: outside the loop, or else the final value must be known. Otherwise,
2204: it is unsafe to split the biv since it may not have the proper
2205: value on loop exit. */
2206:
2207: /* loop_number_exit_labels is non-zero if the loop has an exit other than
2208: a fall through at the end. */
2209:
2210: biv_splittable = 1;
2211: biv_final_value = 0;
2212: if (unroll_type != UNROLL_COMPLETELY
2213: && (loop_number_exit_labels[uid_loop_num[INSN_UID (loop_start)]]
2214: || unroll_type == UNROLL_NAIVE)
2215: && (uid_luid[regno_last_uid[bl->regno]] >= INSN_LUID (loop_end)
2216: || ! bl->init_insn
2217: || INSN_UID (bl->init_insn) >= max_uid_for_loop
2218: || (uid_luid[regno_first_uid[bl->regno]]
2219: < INSN_LUID (bl->init_insn))
2220: || reg_mentioned_p (bl->biv->dest_reg, SET_SRC (bl->init_set)))
2221: && ! (biv_final_value = final_biv_value (bl, loop_start, loop_end)))
2222: biv_splittable = 0;
2223:
2224: /* If final value is non-zero, then must emit an instruction which sets
2225: the value of the biv to the proper value. This is done after
2226: handling all of the givs, since some of them may need to use the
2227: biv's value in their initialization code. */
2228:
2229: /* This biv is splittable. If completely unrolling the loop, save
2230: the biv's initial value. Otherwise, save the constant zero. */
2231:
2232: if (biv_splittable == 1)
2233: {
2234: if (unroll_type == UNROLL_COMPLETELY)
2235: {
2236: /* If the initial value of the biv is itself (i.e. it is too
2237: complicated for strength_reduce to compute), or is a hard
2238: register, then we must create a new psuedo reg to hold the
2239: initial value of the biv. */
2240:
2241: if (GET_CODE (bl->initial_value) == REG
2242: && (REGNO (bl->initial_value) == bl->regno
2243: || REGNO (bl->initial_value) < FIRST_PSEUDO_REGISTER))
2244: {
2245: rtx tem = gen_reg_rtx (bl->biv->mode);
2246:
2247: emit_insn_before (gen_move_insn (tem, bl->biv->src_reg),
2248: loop_start);
2249:
2250: if (loop_dump_stream)
2251: fprintf (loop_dump_stream, "Biv %d initial value remapped to %d.\n",
2252: bl->regno, REGNO (tem));
2253:
2254: splittable_regs[bl->regno] = tem;
2255: }
2256: else
2257: splittable_regs[bl->regno] = bl->initial_value;
2258: }
2259: else
2260: splittable_regs[bl->regno] = const0_rtx;
2261:
2262: /* Save the number of instructions that modify the biv, so that
2263: we can treat the last one specially. */
2264:
2265: splittable_regs_updates[bl->regno] = bl->biv_count;
2266:
2267: result++;
2268:
2269: if (loop_dump_stream)
2270: fprintf (loop_dump_stream,
2271: "Biv %d safe to split.\n", bl->regno);
2272: }
2273:
2274: /* Check every giv that depends on this biv to see whether it is
2275: splittable also. Even if the biv isn't splittable, givs which
2276: depend on it may be splittable if the biv is live outside the
2277: loop, and the givs aren't. */
2278:
2279: result = find_splittable_givs (bl, unroll_type, loop_start, loop_end,
2280: increment, unroll_number, result);
2281:
2282: /* If final value is non-zero, then must emit an instruction which sets
2283: the value of the biv to the proper value. This is done after
2284: handling all of the givs, since some of them may need to use the
2285: biv's value in their initialization code. */
2286: if (biv_final_value)
2287: {
2288: /* If the loop has multiple exits, emit the insns before the
2289: loop to ensure that it will always be executed no matter
2290: how the loop exits. Otherwise emit the insn after the loop,
2291: since this is slightly more efficient. */
2292: if (! loop_number_exit_labels[uid_loop_num[INSN_UID (loop_start)]])
2293: emit_insn_before (gen_move_insn (bl->biv->src_reg,
2294: biv_final_value),
2295: end_insert_before);
2296: else
2297: {
2298: /* Create a new register to hold the value of the biv, and then
2299: set the biv to its final value before the loop start. The biv
2300: is set to its final value before loop start to ensure that
2301: this insn will always be executed, no matter how the loop
2302: exits. */
2303: rtx tem = gen_reg_rtx (bl->biv->mode);
2304: emit_insn_before (gen_move_insn (tem, bl->biv->src_reg),
2305: loop_start);
2306: emit_insn_before (gen_move_insn (bl->biv->src_reg,
2307: biv_final_value),
2308: loop_start);
2309:
2310: if (loop_dump_stream)
2311: fprintf (loop_dump_stream, "Biv %d mapped to %d for split.\n",
2312: REGNO (bl->biv->src_reg), REGNO (tem));
2313:
2314: /* Set up the mapping from the original biv register to the new
2315: register. */
2316: bl->biv->src_reg = tem;
2317: }
2318: }
2319: }
2320: return result;
2321: }
2322:
2323: /* For every giv based on the biv BL, check to determine whether it is
2324: splittable. This is a subroutine to find_splittable_regs (). */
2325:
2326: static int
2327: find_splittable_givs (bl, unroll_type, loop_start, loop_end, increment,
2328: unroll_number, result)
2329: struct iv_class *bl;
2330: enum unroll_types unroll_type;
2331: rtx loop_start, loop_end;
2332: rtx increment;
2333: int unroll_number, result;
2334: {
2335: struct induction *v;
2336: rtx final_value;
2337: rtx tem;
2338:
2339: for (v = bl->giv; v; v = v->next_iv)
2340: {
2341: rtx giv_inc, value;
2342:
2343: /* Only split the giv if it has already been reduced, or if the loop is
2344: being completely unrolled. */
2345: if (unroll_type != UNROLL_COMPLETELY && v->ignore)
2346: continue;
2347:
2348: /* The giv can be split if the insn that sets the giv is executed once
2349: and only once on every iteration of the loop. */
2350: /* An address giv can always be split. v->insn is just a use not a set,
2351: and hence it does not matter whether it is always executed. All that
2352: matters is that all the biv increments are always executed, and we
2353: won't reach here if they aren't. */
2354: if (v->giv_type != DEST_ADDR
2355: && (! v->always_computable
2356: || back_branch_in_range_p (v->insn, loop_start, loop_end)))
2357: continue;
2358:
2359: /* The giv increment value must be a constant. */
2360: giv_inc = fold_rtx_mult_add (v->mult_val, increment, const0_rtx,
2361: v->mode);
2362: if (! giv_inc || GET_CODE (giv_inc) != CONST_INT)
2363: continue;
2364:
2365: /* The loop must be unrolled completely, or else have a known number of
2366: iterations and only one exit, or else the giv must be dead outside
2367: the loop, or else the final value of the giv must be known.
2368: Otherwise, it is not safe to split the giv since it may not have the
2369: proper value on loop exit. */
2370:
2371: /* The used outside loop test will fail for DEST_ADDR givs. They are
2372: never used outside the loop anyways, so it is always safe to split a
2373: DEST_ADDR giv. */
2374:
2375: final_value = 0;
2376: if (unroll_type != UNROLL_COMPLETELY
2377: && (loop_number_exit_labels[uid_loop_num[INSN_UID (loop_start)]]
2378: || unroll_type == UNROLL_NAIVE)
2379: && v->giv_type != DEST_ADDR
2380: && ((regno_first_uid[REGNO (v->dest_reg)] != INSN_UID (v->insn)
2381: /* Check for the case where the pseudo is set by a shift/add
2382: sequence, in which case the first insn setting the pseudo
2383: is the first insn of the shift/add sequence. */
2384: && (! (tem = find_reg_note (v->insn, REG_RETVAL, 0))
2385: || (regno_first_uid[REGNO (v->dest_reg)]
2386: != INSN_UID (XEXP (tem, 0)))))
2387: /* Line above always fails if INSN was moved by loop opt. */
2388: || (uid_luid[regno_last_uid[REGNO (v->dest_reg)]]
2389: >= INSN_LUID (loop_end)))
2390: && ! (final_value = v->final_value))
2391: continue;
2392:
2393: #if 0
2394: /* Currently, non-reduced/final-value givs are never split. */
2395: /* Should emit insns after the loop if possible, as the biv final value
2396: code below does. */
2397:
2398: /* If the final value is non-zero, and the giv has not been reduced,
2399: then must emit an instruction to set the final value. */
2400: if (final_value && !v->new_reg)
2401: {
2402: /* Create a new register to hold the value of the giv, and then set
2403: the giv to its final value before the loop start. The giv is set
2404: to its final value before loop start to ensure that this insn
2405: will always be executed, no matter how we exit. */
2406: tem = gen_reg_rtx (v->mode);
2407: emit_insn_before (gen_move_insn (tem, v->dest_reg), loop_start);
2408: emit_insn_before (gen_move_insn (v->dest_reg, final_value),
2409: loop_start);
2410:
2411: if (loop_dump_stream)
2412: fprintf (loop_dump_stream, "Giv %d mapped to %d for split.\n",
2413: REGNO (v->dest_reg), REGNO (tem));
2414:
2415: v->src_reg = tem;
2416: }
2417: #endif
2418:
2419: /* This giv is splittable. If completely unrolling the loop, save the
2420: giv's initial value. Otherwise, save the constant zero for it. */
2421:
2422: if (unroll_type == UNROLL_COMPLETELY)
2423: /* It is not safe to use bl->initial_value here, because it may not
2424: be invariant. It is safe to use the initial value stored in
2425: the splittable_regs array. */
2426: value = fold_rtx_mult_add (v->mult_val, splittable_regs[bl->regno],
2427: v->add_val, v->mode);
2428: else
2429: value = const0_rtx;
2430:
2431: if (v->new_reg)
2432: {
2433: /* If the giv is an address destination, it could be something other
2434: than a simple register, these have to be treated differently. */
2435: if (v->giv_type == DEST_REG)
2436: splittable_regs[REGNO (v->new_reg)] = value;
2437:
2438: /* If an addr giv was combined with another addr giv, then we
2439: can only split this giv if the addr giv it was combined with
2440: was reduced. This is because the value of v->new_reg is
2441: meaningless in this case. (There is no problem if it was
2442: combined with a dest_reg giv which wasn't reduced, v->new_reg
2443: is still meaningful in this case.) */
2444:
2445: else if (v->same && v->same->giv_type == DEST_ADDR
2446: && ! v->same->new_reg)
2447: {
2448: if (loop_dump_stream)
2449: fprintf (loop_dump_stream,
2450: "DEST_ADDR giv not split, because combined with unreduced DEST_ADDR giv.\n");
2451: }
2452: else
2453: {
2454: /* Splitting address givs is useful since it will often allow us
2455: to eliminate some increment insns for the base giv as
2456: unnecessary. */
2457:
2458: /* If the addr giv is combined with a dest_reg giv, then all
2459: references to that dest reg will be remapped, which is NOT
2460: what we want for split addr regs. We always create a new
2461: register for the split addr giv, just to be safe. */
2462:
2463: /* ??? If there are multiple address givs which have been
2464: combined with the same dest_reg giv, then we may only need
2465: one new register for them. Pulling out constants below will
2466: catch some of the common cases of this. Currently, I leave
2467: the work of simplifying multiple address givs to the
2468: following cse pass. */
2469:
2470: v->const_adjust = 0;
2471: if (unroll_type != UNROLL_COMPLETELY)
2472: {
2473: /* If not completely unrolling the loop, then create a new
2474: register to hold the split value of the DEST_ADDR giv.
2475: Emit insn to initialize its value before loop start. */
2476: tem = gen_reg_rtx (v->mode);
2477:
2478: /* If the address giv has a constant in its new_reg value,
2479: then this constant can be pulled out and put in value,
2480: instead of being part of the initialization code. */
2481:
2482: if (GET_CODE (v->new_reg) == PLUS
2483: && GET_CODE (XEXP (v->new_reg, 1)) == CONST_INT)
2484: {
2485: v->dest_reg
2486: = plus_constant (tem, INTVAL (XEXP (v->new_reg,1)));
2487:
2488: /* Only succeed if this will give valid addresses.
2489: Try to validate both the first and the last
2490: address resulting from loop unrolling, if
2491: one fails, then can't do const elim here. */
2492: if (memory_address_p (v->mode, v->dest_reg)
2493: && memory_address_p (v->mode,
2494: plus_constant (v->dest_reg,
2495: INTVAL (giv_inc)
2496: * (unroll_number - 1))))
2497: {
2498: /* Save the negative of the eliminated const, so
2499: that we can calculate the dest_reg's increment
2500: value later. */
2501: v->const_adjust = - INTVAL (XEXP (v->new_reg, 1));
2502:
2503: v->new_reg = XEXP (v->new_reg, 0);
2504: if (loop_dump_stream)
2505: fprintf (loop_dump_stream,
2506: "Eliminating constant from giv %d\n",
2507: REGNO (tem));
2508: }
2509: else
2510: v->dest_reg = tem;
2511: }
2512: else
2513: v->dest_reg = tem;
2514:
2515: /* If the address hasn't been checked for validity yet, do so
2516: now, and fail completely if either the first or the last
2517: unrolled copy of the address is not a valid address. */
2518: if (v->dest_reg == tem
2519: && (! memory_address_p (v->mode, v->dest_reg)
2520: || ! memory_address_p (v->mode,
2521: plus_constant (v->dest_reg,
2522: INTVAL (giv_inc)
2523: * (unroll_number -1)))))
2524: {
2525: if (loop_dump_stream)
2526: fprintf (loop_dump_stream,
2527: "Illegal address for giv at insn %d\n",
2528: INSN_UID (v->insn));
2529: continue;
2530: }
2531:
2532: /* To initialize the new register, just move the value of
2533: new_reg into it. This is not guaranteed to give a valid
2534: instruction on machines with complex addressing modes.
2535: If we can't recognize it, then delete it and emit insns
2536: to calculate the value from scratch. */
2537: emit_insn_before (gen_rtx (SET, VOIDmode, tem,
2538: copy_rtx (v->new_reg)),
2539: loop_start);
2540: if (! recog_memoized (PREV_INSN (loop_start)))
2541: {
2542: delete_insn (PREV_INSN (loop_start));
2543: emit_iv_add_mult (bl->initial_value, v->mult_val,
2544: v->add_val, tem, loop_start);
2545: if (loop_dump_stream)
2546: fprintf (loop_dump_stream,
2547: "Illegal init insn, rewritten.\n");
2548: }
2549: }
2550: else
2551: {
2552: v->dest_reg = value;
2553:
2554: /* Check the resulting address for validity, and fail
2555: if the resulting address would be illegal. */
2556: if (! memory_address_p (v->mode, v->dest_reg)
2557: || ! memory_address_p (v->mode,
2558: plus_constant (v->dest_reg,
2559: INTVAL (giv_inc) *
2560: (unroll_number -1))))
2561: {
2562: if (loop_dump_stream)
2563: fprintf (loop_dump_stream,
2564: "Illegal address for giv at insn %d\n",
2565: INSN_UID (v->insn));
2566: continue;
2567: }
2568: }
2569:
2570: /* Store the value of dest_reg into the insn. This sharing
2571: will not be a problem as this insn will always be copied
2572: later. */
2573:
2574: *v->location = v->dest_reg;
2575:
2576: /* If this address giv is combined with a dest reg giv, then
2577: save the base giv's induction pointer so that we will be
2578: able to handle this address giv properly. The base giv
2579: itself does not have to be splittable. */
2580:
2581: if (v->same && v->same->giv_type == DEST_REG)
2582: addr_combined_regs[REGNO (v->same->new_reg)] = v->same;
2583:
2584: if (GET_CODE (v->new_reg) == REG)
2585: {
2586: /* This giv maybe hasn't been combined with any others.
2587: Make sure that it's giv is marked as splittable here. */
2588:
2589: splittable_regs[REGNO (v->new_reg)] = value;
2590:
2591: /* Make it appear to depend upon itself, so that the
2592: giv will be properly split in the main loop above. */
2593: if (! v->same)
2594: {
2595: v->same = v;
2596: addr_combined_regs[REGNO (v->new_reg)] = v;
2597: }
2598: }
2599:
2600: if (loop_dump_stream)
2601: fprintf (loop_dump_stream, "DEST_ADDR giv being split.\n");
2602: }
2603: }
2604: else
2605: {
2606: #if 0
2607: /* Currently, unreduced giv's can't be split. This is not too much
2608: of a problem since unreduced giv's are not live across loop
2609: iterations anyways. When unrolling a loop completely though,
2610: it makes sense to reduce&split givs when possible, as this will
2611: result in simpler instructions, and will not require that a reg
2612: be live across loop iterations. */
2613:
2614: splittable_regs[REGNO (v->dest_reg)] = value;
2615: fprintf (stderr, "Giv %d at insn %d not reduced\n",
2616: REGNO (v->dest_reg), INSN_UID (v->insn));
2617: #else
2618: continue;
2619: #endif
2620: }
2621:
2622: /* Givs are only updated once by definition. Mark it so if this is
2623: a splittable register. Don't need to do anything for address givs
2624: where this may not be a register. */
2625:
2626: if (GET_CODE (v->new_reg) == REG)
2627: splittable_regs_updates[REGNO (v->new_reg)] = 1;
2628:
2629: result++;
2630:
2631: if (loop_dump_stream)
2632: {
2633: int regnum;
2634:
2635: if (GET_CODE (v->dest_reg) == CONST_INT)
2636: regnum = -1;
2637: else if (GET_CODE (v->dest_reg) != REG)
2638: regnum = REGNO (XEXP (v->dest_reg, 0));
2639: else
2640: regnum = REGNO (v->dest_reg);
2641: fprintf (loop_dump_stream, "Giv %d at insn %d safe to split.\n",
2642: regnum, INSN_UID (v->insn));
2643: }
2644: }
2645:
2646: return result;
2647: }
2648:
2649: /* Try to prove that the register is dead after the loop exits. Trace every
2650: loop exit looking for an insn that will always be executed, which sets
2651: the register to some value, and appears before the first use of the register
2652: is found. If successful, then return 1, otherwise return 0. */
2653:
2654: /* ?? Could be made more intelligent in the handling of jumps, so that
2655: it can search past if statements and other similar structures. */
2656:
2657: static int
2658: reg_dead_after_loop (reg, loop_start, loop_end)
2659: rtx reg, loop_start, loop_end;
2660: {
2661: rtx insn, label;
2662: enum rtx_code code;
2663:
2664: /* HACK: Must also search the loop fall through exit, create a label_ref
2665: here which points to the loop_end, and append the loop_number_exit_labels
2666: list to it. */
2667: label = gen_rtx (LABEL_REF, VOIDmode, loop_end);
2668: LABEL_NEXTREF (label)
2669: = loop_number_exit_labels[uid_loop_num[INSN_UID (loop_start)]];
2670:
2671: for ( ; label; label = LABEL_NEXTREF (label))
2672: {
2673: /* Succeed if find an insn which sets the biv or if reach end of
2674: function. Fail if find an insn that uses the biv, or if come to
2675: a conditional jump. */
2676:
2677: insn = NEXT_INSN (XEXP (label, 0));
2678: while (1)
2679: {
2680: if (insn == 0)
2681: break;
2682:
2683: if ((code = GET_CODE (insn)) == INSN || code == JUMP_INSN
2684: || code == CALL_INSN)
2685: {
2686: if (GET_CODE (PATTERN (insn)) == SET)
2687: {
2688: if (reg_mentioned_p (reg, SET_SRC (PATTERN (insn))))
2689: return 0;
2690: if (SET_DEST (PATTERN (insn)) == reg)
2691: break;
2692: if (reg_mentioned_p (reg, SET_DEST (PATTERN (insn))))
2693: return 0;
2694: }
2695: else if (reg_mentioned_p (reg, PATTERN (insn)))
2696: return 0;
2697: }
2698: if (code == JUMP_INSN)
2699: {
2700: if (GET_CODE (PATTERN (insn)) == RETURN)
2701: break;
2702: else if (! simplejump_p (insn))
2703: return 0;
2704: else
2705: {
2706: insn = JUMP_LABEL (insn);
2707: /* If this branches to a code label after a LOOP_BEG or
2708: a LOOP_CONT note, then assume it is a loop back edge.
2709: Must fail in that case to prevent going into an infinite
2710: loop trying to trace infinite loops.
2711:
2712: In the presence of syntax errors, this may be a jump to
2713: a CODE_LABEL that was never emitted. Fail in this case
2714: also. */
2715:
2716: if (! PREV_INSN (insn)
2717: || (GET_CODE (PREV_INSN (insn)) == NOTE
2718: && ((NOTE_LINE_NUMBER (PREV_INSN (insn))
2719: == NOTE_INSN_LOOP_BEG)
2720: || (NOTE_LINE_NUMBER (PREV_INSN (insn))
2721: == NOTE_INSN_LOOP_CONT))))
2722: return 0;
2723: }
2724: }
2725:
2726: insn = NEXT_INSN (insn);
2727: }
2728: }
2729:
2730: /* Success, the register is dead on all loop exits. */
2731: return 1;
2732: }
2733:
2734: /* Try to calculate the final value of the biv, the value it will have at
2735: the end of the loop. If we can do it, return that value. */
2736:
2737: rtx
2738: final_biv_value (bl, loop_start, loop_end)
2739: struct iv_class *bl;
2740: rtx loop_start, loop_end;
2741: {
2742: rtx increment, tem;
2743:
2744: /* The final value for reversed bivs must be calculated differently than
2745: for ordinary bivs. In this case, there is already an insn after the
2746: loop which sets this biv's final value (if necessary), and there are
2747: no other loop exits, so we can return any value. */
2748: if (bl->reversed)
2749: {
2750: if (loop_dump_stream)
2751: fprintf (loop_dump_stream,
2752: "Final biv value for %d, reversed biv.\n", bl->regno);
2753:
2754: return const0_rtx;
2755: }
2756:
2757: /* Try to calculate the final value as initial value + (number of iterations
2758: * increment). For this to work, increment must be invariant, the only
2759: exit from the loop must be the fall through at the bottom (otherwise
2760: it may not have its final value when the loop exits), and the initial
2761: value of the biv must be invariant. */
2762:
2763: if (loop_n_iterations != 0
2764: && ! loop_number_exit_labels[uid_loop_num[INSN_UID (loop_start)]]
2765: && invariant_p (bl->initial_value))
2766: {
2767: increment = biv_total_increment (bl, loop_start, loop_end);
2768:
2769: if (increment && invariant_p (increment))
2770: {
2771: /* Can calculate the loop exit value, emit insns after loop
2772: end to calculate this value into a temporary register in
2773: case it is needed later. */
2774:
2775: tem = gen_reg_rtx (bl->biv->mode);
2776: emit_iv_add_mult (increment,
2777: gen_rtx (CONST_INT, VOIDmode, loop_n_iterations),
2778: bl->initial_value, tem, NEXT_INSN (loop_end));
2779:
2780: if (loop_dump_stream)
2781: fprintf (loop_dump_stream,
2782: "Final biv value for %d, calculated.\n", bl->regno);
2783:
2784: return tem;
2785: }
2786: }
2787:
2788: /* Check to see if the biv is dead at all loop exits. */
2789: if (reg_dead_after_loop (bl->biv->src_reg, loop_start, loop_end))
2790: {
2791: if (loop_dump_stream)
2792: fprintf (loop_dump_stream,
2793: "Final biv value for %d, biv dead after loop exit.\n",
2794: bl->regno);
2795:
2796: return const0_rtx;
2797: }
2798:
2799: return 0;
2800: }
2801:
2802: /* Try to calculate the final value of the giv, the value it will have at
2803: the end of the loop. If we can do it, return that value. */
2804:
2805: rtx
2806: final_giv_value (v, loop_start, loop_end)
2807: struct induction *v;
2808: rtx loop_start, loop_end;
2809: {
2810: struct iv_class *bl;
2811: rtx reg, insn, pattern;
2812: rtx increment, tem;
2813: enum rtx_code code;
2814: rtx insert_before;
2815:
2816: bl = reg_biv_class[REGNO (v->src_reg)];
2817:
2818: /* The final value for givs which depend on reversed bivs must be calculated
2819: differently than for ordinary givs. In this case, there is already an
2820: insn after the loop which sets this giv's final value (if necessary),
2821: and there are no other loop exits, so we can return any value. */
2822: if (bl->reversed)
2823: {
2824: if (loop_dump_stream)
2825: fprintf (loop_dump_stream,
2826: "Final giv value for %d, depends on reversed biv\n",
2827: REGNO (v->dest_reg));
2828: return const0_rtx;
2829: }
2830:
2831: /* Try to calculate the final value as a function of the biv it depends
2832: upon. The only exit from the loop must be the fall through at the bottom
2833: (otherwise it may not have its final value when the loop exits). */
2834:
2835: /* ??? Can calculate the final giv value by subtracting off the
2836: extra biv increments times the giv's mult_val. The loop must have
2837: only one exit for this to work, but the loop iterations does not need
2838: to be known. */
2839:
2840: if (loop_n_iterations != 0
2841: && ! loop_number_exit_labels[uid_loop_num[INSN_UID (loop_start)]])
2842: {
2843: /* ?? It is tempting to use the biv's value here since these insns will
2844: be put after the loop, and hence the biv will have its final value
2845: then. However, this fails if the biv is subsequently eliminated.
2846: Perhaps determine whether biv's are eliminable before trying to
2847: determine whether giv's are replaceable so that we can use the
2848: biv value here if it is not eliminable. */
2849:
2850: increment = biv_total_increment (bl, loop_start, loop_end);
2851:
2852: if (increment && invariant_p (increment))
2853: {
2854: /* Can calculate the loop exit value of its biv as
2855: (loop_n_iterations * increment) + initial_value */
2856:
2857: /* The loop exit value of the giv is then
2858: (final_biv_value - extra increments) * mult_val + add_val.
2859: The extra increments are any increments to the biv which
2860: occur in the loop after the giv's value is calculated.
2861: We must search from the insn that sets the giv to the end
2862: of the loop to calculate this value. */
2863:
2864: insert_before = NEXT_INSN (loop_end);
2865:
2866: /* Put the final biv value in tem. */
2867: tem = gen_reg_rtx (bl->biv->mode);
2868: emit_iv_add_mult (increment,
2869: gen_rtx (CONST_INT, VOIDmode, loop_n_iterations),
2870: bl->initial_value, tem, insert_before);
2871:
2872: /* Subtract off extra increments as we find them. */
2873: for (insn = NEXT_INSN (v->insn); insn != loop_end;
2874: insn = NEXT_INSN (insn))
2875: {
2876: if (GET_CODE (insn) == INSN
2877: && GET_CODE (PATTERN (insn)) == SET
2878: && SET_DEST (PATTERN (insn)) == v->src_reg)
2879: {
2880: pattern = PATTERN (insn);
2881: if (GET_CODE (SET_SRC (pattern)) != PLUS)
2882: {
2883: /* Sometimes a biv is computed in a temp reg,
2884: and then copied into the biv reg. */
2885: pattern = PATTERN (PREV_INSN (insn));
2886: if (GET_CODE (SET_SRC (pattern)) != PLUS)
2887: abort ();
2888: }
2889: if (GET_CODE (XEXP (SET_SRC (pattern), 0)) != REG
2890: || REGNO (XEXP (SET_SRC (pattern), 0)) != bl->regno)
2891: abort ();
2892:
2893: tem = expand_binop (GET_MODE (tem), sub_optab, tem,
2894: XEXP (SET_SRC (pattern), 1), 0, 0,
2895: OPTAB_LIB_WIDEN);
2896: }
2897: }
2898:
2899: /* Now calculate the giv's final value. */
2900: emit_iv_add_mult (tem, v->mult_val, v->add_val, tem,
2901: insert_before);
2902:
2903: if (loop_dump_stream)
2904: fprintf (loop_dump_stream,
2905: "Final giv value for %d, calc from biv's value.\n",
2906: REGNO (v->dest_reg));
2907:
2908: return tem;
2909: }
2910: }
2911:
2912: /* Replaceable giv's should never reach here. */
2913: if (v->replaceable)
2914: abort ();
2915:
2916: /* Check to see if the biv is dead at all loop exits. */
2917: if (reg_dead_after_loop (v->dest_reg, loop_start, loop_end))
2918: {
2919: if (loop_dump_stream)
2920: fprintf (loop_dump_stream,
2921: "Final giv value for %d, giv dead after loop exit.\n",
2922: REGNO (v->dest_reg));
2923:
2924: return const0_rtx;
2925: }
2926:
2927: return 0;
2928: }
2929:
2930:
2931: /* Calculate the number of loop iterations. Returns the exact number of loop
2932: iterations if it can be calculated, otherwise retusns zero. */
2933:
2934: unsigned long
2935: loop_iterations (loop_start, loop_end)
2936: rtx loop_start, loop_end;
2937: {
2938: rtx comparison, comparison_value;
2939: rtx iteration_var, initial_value, increment, final_value;
2940: enum rtx_code comparison_code;
2941: int i, increment_dir;
2942: int unsigned_compare, compare_dir, final_larger;
2943: unsigned long tempu;
2944: rtx last_loop_insn;
2945:
2946: /* First find the iteration variable. If the last insn is a conditional
2947: branch, and the insn before tests a register value, make that the
2948: iteration variable. */
2949:
2950: loop_initial_value = 0;
2951: loop_increment = 0;
2952: loop_final_value = 0;
2953: loop_iteration_var = 0;
2954:
2955: last_loop_insn = prev_nonnote_insn (loop_end);
2956:
2957: comparison = get_condition_for_loop (last_loop_insn);
2958: if (comparison == 0)
2959: {
2960: if (loop_dump_stream)
2961: fprintf (loop_dump_stream,
2962: "Loop unrolling: No final conditional branch found.\n");
2963: return 0;
2964: }
2965:
2966: /* ??? Get_condition may switch position of induction variable and
2967: invariant register when it canonicalizes the comparison. */
2968:
2969: comparison_code = GET_CODE (comparison);
2970: iteration_var = XEXP (comparison, 0);
2971: comparison_value = XEXP (comparison, 1);
2972:
2973: if (GET_CODE (iteration_var) != REG)
2974: {
2975: if (loop_dump_stream)
2976: fprintf (loop_dump_stream,
2977: "Loop unrolling: Comparison not against register.\n");
2978: return 0;
2979: }
2980:
2981: /* Loop iterations is always called before any new registers are created
2982: now, so this should never occur. */
2983:
2984: if (REGNO (iteration_var) >= max_reg_before_loop)
2985: abort ();
2986:
2987: iteration_info (iteration_var, &initial_value, &increment,
2988: loop_start, loop_end);
2989: if (initial_value == 0)
2990: /* iteration_info already printed a message. */
2991: return 0;
2992:
2993: if (increment == 0)
2994: {
2995: if (loop_dump_stream)
2996: fprintf (loop_dump_stream,
2997: "Loop unrolling: Increment value can't be calculated.\n");
2998: return 0;
2999: }
3000: if (GET_CODE (increment) != CONST_INT)
3001: {
3002: if (loop_dump_stream)
3003: fprintf (loop_dump_stream,
3004: "Loop unrolling: Increment value not constant.\n");
3005: return 0;
3006: }
3007: if (GET_CODE (initial_value) != CONST_INT)
3008: {
3009: if (loop_dump_stream)
3010: fprintf (loop_dump_stream,
3011: "Loop unrolling: Initial value not constant.\n");
3012: return 0;
3013: }
3014:
3015: /* If the comparison value is an invariant register, then try to find
3016: its value from the insns before the start of the loop. */
3017:
3018: if (GET_CODE (comparison_value) == REG && invariant_p (comparison_value))
3019: {
3020: rtx insn, set;
3021:
3022: for (insn = PREV_INSN (loop_start); insn ; insn = PREV_INSN (insn))
3023: {
3024: if (GET_CODE (insn) == CODE_LABEL)
3025: break;
3026:
3027: else if (GET_RTX_CLASS (GET_CODE (insn)) == 'i'
3028: && (set = single_set (insn))
3029: && (SET_DEST (set) == comparison_value))
3030: {
3031: rtx note = find_reg_note (insn, REG_EQUAL, 0);
3032:
3033: if (note && GET_CODE (XEXP (note, 0)) != EXPR_LIST)
3034: comparison_value = XEXP (note, 0);
3035:
3036: break;
3037: }
3038: }
3039: }
3040:
3041: final_value = approx_final_value (comparison_code, comparison_value,
3042: &unsigned_compare, &compare_dir);
3043:
3044: /* Save the calculated values describing this loop's bounds, in case
3045: precondition_loop_p will need them later. These values can not be
3046: recalculated inside precondition_loop_p because strength reduction
3047: optimizations may obscure the loop's structure. */
3048:
3049: loop_iteration_var = iteration_var;
3050: loop_initial_value = initial_value;
3051: loop_increment = increment;
3052: loop_final_value = final_value;
3053:
3054: if (final_value == 0)
3055: {
3056: if (loop_dump_stream)
3057: fprintf (loop_dump_stream,
3058: "Loop unrolling: EQ comparison loop.\n");
3059: return 0;
3060: }
3061: else if (GET_CODE (final_value) != CONST_INT)
3062: {
3063: if (loop_dump_stream)
3064: fprintf (loop_dump_stream,
3065: "Loop unrolling: Final value not constant.\n");
3066: return 0;
3067: }
3068:
3069: /* ?? Final value and initial value do not have to be constants.
3070: Only their difference has to be constant. When the iteration variable
3071: is an array address, the final value and initial value might both
3072: be addresses with the same base but different constant offsets.
3073: Final value must be invariant for this to work.
3074:
3075: To do this, need someway to find the values of registers which are
3076: invariant. */
3077:
3078: /* Final_larger is 1 if final larger, 0 if they are equal, otherwise -1. */
3079: if (unsigned_compare)
3080: final_larger
3081: = ((unsigned) INTVAL (final_value) > (unsigned) INTVAL (initial_value)) -
3082: ((unsigned) INTVAL (final_value) < (unsigned) INTVAL (initial_value));
3083: else
3084: final_larger = (INTVAL (final_value) > INTVAL (initial_value)) -
3085: (INTVAL (final_value) < INTVAL (initial_value));
3086:
3087: if (INTVAL (increment) > 0)
3088: increment_dir = 1;
3089: else if (INTVAL (increment) == 0)
3090: increment_dir = 0;
3091: else
3092: increment_dir = -1;
3093:
3094: /* There are 27 different cases: compare_dir = -1, 0, 1;
3095: final_larger = -1, 0, 1; increment_dir = -1, 0, 1.
3096: There are 4 normal cases, 4 reverse cases (where the iteration variable
3097: will overflow before the loop exits), 4 infinite loop cases, and 15
3098: immediate exit (0 or 1 iteration depending on loop type) cases.
3099: Only try to optimize the normal cases. */
3100:
3101: /* (compare_dir/final_larger/increment_dir)
3102: Normal cases: (0/-1/-1), (0/1/1), (-1/-1/-1), (1/1/1)
3103: Reverse cases: (0/-1/1), (0/1/-1), (-1/-1/1), (1/1/-1)
3104: Infinite loops: (0/-1/0), (0/1/0), (-1/-1/0), (1/1/0)
3105: Immediate exit: (0/0/X), (-1/0/X), (-1/1/X), (1/0/X), (1/-1/X) */
3106:
3107: /* ?? If the meaning of reverse loops (where the iteration variable
3108: will overflow before the loop exits) is undefined, then could
3109: eliminate all of these special checks, and just always assume
3110: the loops are normal/immediate/infinite. Note that this means
3111: the sign of increment_dir does not have to be known. Also,
3112: since it does not really hurt if immediate exit loops or infinite loops
3113: are optimized, then that case could be ignored also, and hence all
3114: loops can be optimized.
3115:
3116: According to ANSI Spec, the reverse loop case result is undefined,
3117: because the action on overflow is undefined.
3118:
3119: See also the special test for NE loops below. */
3120:
3121: if (final_larger == increment_dir && final_larger != 0
3122: && (final_larger == compare_dir || compare_dir == 0))
3123: /* Normal case. */
3124: ;
3125: else
3126: {
3127: if (loop_dump_stream)
3128: fprintf (loop_dump_stream,
3129: "Loop unrolling: Not normal loop.\n");
3130: return 0;
3131: }
3132:
3133: /* Calculate the number of iterations, final_value is only an approximation,
3134: so correct for that. Note that tempu and loop_n_iterations are
3135: unsigned, because they can be as large as 2^n - 1. */
3136:
3137: i = INTVAL (increment);
3138: if (i > 0)
3139: tempu = INTVAL (final_value) - INTVAL (initial_value);
3140: else if (i < 0)
3141: {
3142: tempu = INTVAL (initial_value) - INTVAL (final_value);
3143: i = -i;
3144: }
3145: else
3146: abort ();
3147:
3148: /* For NE tests, make sure that the iteration variable won't miss the
3149: final value. If tempu mod i is not zero, then the iteration variable
3150: will overflow before the loop exits, and we can not calculate the
3151: number of iterations. */
3152: if (compare_dir == 0 && (tempu % i) != 0)
3153: return 0;
3154:
3155: return tempu / i + ((tempu % i) != 0);
3156: }
This archive runs on limited infrastructure. Preserving old code on modern bandwidth. Automated agents are requested to crawl responsibly.