![[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
1.1.1.3 root 62: memory addresses can be split, and hence giv increments can be
1.1 root 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
1.1.1.3 root 208: LOOP_START. END_INSERT_BEFORE indicates where insns should be added
1.1 root 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
1.1.1.2 root 327: /* The immediately preceding insn is a compare which must be
1.1 root 328: deleted. */
329: delete_insn (last_loop_insn);
330: delete_insn (PREV_INSN (last_loop_insn));
331: #else
1.1.1.2 root 332: /* The immediately preceding insn may not be the compare, so don't
1.1 root 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
1.1.1.3 root 691: to access the splittable_regs[] and addr_combined_regs[] arrays. */
1.1 root 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),
1.1.1.4 ! root 786: copy_rtx (initial_value), NULL_RTX, 0,
1.1 root 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,
1.1.1.4 ! root 792: GEN_INT (unroll_number * abs_inc - 1),
! 793: NULL_RTX, 0, OPTAB_LIB_WIDEN);
1.1 root 794:
795: /* Now emit a sequence of branches to jump to the proper precond
796: loop entry point. */
797:
798: labels = (rtx *) alloca (sizeof (rtx) * unroll_number);
799: for (i = 0; i < unroll_number; i++)
800: labels[i] = gen_label_rtx ();
801:
802: /* Assuming the unroll_number is 4, and the increment is 2, then
803: for a negative increment: for a positive increment:
804: diff = 0,1 precond 0 diff = 0,7 precond 0
805: diff = 2,3 precond 3 diff = 1,2 precond 1
806: diff = 4,5 precond 2 diff = 3,4 precond 2
807: diff = 6,7 precond 1 diff = 5,6 precond 3 */
808:
809: /* We only need to emit (unroll_number - 1) branches here, the
810: last case just falls through to the following code. */
811:
812: /* ??? This would give better code if we emitted a tree of branches
813: instead of the current linear list of branches. */
814:
815: for (i = 0; i < unroll_number - 1; i++)
816: {
817: int cmp_const;
818:
819: /* For negative increments, must invert the constant compared
820: against, except when comparing against zero. */
821: if (i == 0)
822: cmp_const = 0;
823: else if (neg_inc)
824: cmp_const = unroll_number - i;
825: else
826: cmp_const = i;
827:
1.1.1.4 ! root 828: emit_cmp_insn (diff, GEN_INT (abs_inc * cmp_const),
! 829: EQ, NULL_RTX, mode, 0, 0);
1.1 root 830:
831: if (i == 0)
832: emit_jump_insn (gen_beq (labels[i]));
833: else if (neg_inc)
834: emit_jump_insn (gen_bge (labels[i]));
835: else
836: emit_jump_insn (gen_ble (labels[i]));
837: JUMP_LABEL (get_last_insn ()) = labels[i];
838: LABEL_NUSES (labels[i])++;
839: }
840:
841: /* If the increment is greater than one, then we need another branch,
842: to handle other cases equivalent to 0. */
843:
844: /* ??? This should be merged into the code above somehow to help
845: simplify the code here, and reduce the number of branches emitted.
846: For the negative increment case, the branch here could easily
847: be merged with the `0' case branch above. For the positive
848: increment case, it is not clear how this can be simplified. */
849:
850: if (abs_inc != 1)
851: {
852: int cmp_const;
853:
854: if (neg_inc)
855: cmp_const = abs_inc - 1;
856: else
857: cmp_const = abs_inc * (unroll_number - 1) + 1;
858:
1.1.1.4 ! root 859: emit_cmp_insn (diff, GEN_INT (cmp_const), EQ, NULL_RTX,
! 860: mode, 0, 0);
1.1 root 861:
862: if (neg_inc)
863: emit_jump_insn (gen_ble (labels[0]));
864: else
865: emit_jump_insn (gen_bge (labels[0]));
866: JUMP_LABEL (get_last_insn ()) = labels[0];
867: LABEL_NUSES (labels[0])++;
868: }
869:
870: sequence = gen_sequence ();
871: end_sequence ();
872: emit_insn_before (sequence, loop_start);
873:
874: /* Only the last copy of the loop body here needs the exit
875: test, so set copy_end to exclude the compare/branch here,
876: and then reset it inside the loop when get to the last
877: copy. */
878:
879: if (GET_CODE (last_loop_insn) == BARRIER)
880: copy_end = PREV_INSN (PREV_INSN (last_loop_insn));
881: else if (GET_CODE (last_loop_insn) == JUMP_INSN)
882: {
883: #ifdef HAVE_cc0
1.1.1.2 root 884: /* The immediately preceding insn is a compare which we do not
1.1 root 885: want to copy. */
886: copy_end = PREV_INSN (PREV_INSN (last_loop_insn));
887: #else
1.1.1.2 root 888: /* The immediately preceding insn may not be a compare, so we
1.1 root 889: must copy it. */
890: copy_end = PREV_INSN (last_loop_insn);
891: #endif
892: }
893: else
894: abort ();
895:
896: for (i = 1; i < unroll_number; i++)
897: {
898: emit_label_after (labels[unroll_number - i],
899: PREV_INSN (loop_start));
900:
901: bzero (map->insn_map, max_insnno * sizeof (rtx));
902: bzero (map->const_equiv_map, maxregnum * sizeof (rtx));
903: bzero (map->const_age_map, maxregnum * sizeof (unsigned));
904: map->const_age = 0;
905:
906: for (j = 0; j < max_labelno; j++)
907: if (local_label[j])
908: map->label_map[j] = gen_label_rtx ();
909:
910: /* The last copy needs the compare/branch insns at the end,
911: so reset copy_end here if the loop ends with a conditional
912: branch. */
913:
914: if (i == unroll_number - 1)
915: {
916: if (GET_CODE (last_loop_insn) == BARRIER)
917: copy_end = PREV_INSN (PREV_INSN (last_loop_insn));
918: else
919: copy_end = last_loop_insn;
920: }
921:
922: /* None of the copies are the `last_iteration', so just
923: pass zero for that parameter. */
924: copy_loop_body (copy_start, copy_end, map, exit_label, 0,
925: unroll_type, start_label, loop_end,
926: loop_start, copy_end);
927: }
928: emit_label_after (labels[0], PREV_INSN (loop_start));
929:
930: if (GET_CODE (last_loop_insn) == BARRIER)
931: {
932: insert_before = PREV_INSN (last_loop_insn);
933: copy_end = PREV_INSN (insert_before);
934: }
935: else
936: {
937: #ifdef HAVE_cc0
1.1.1.2 root 938: /* The immediately preceding insn is a compare which we do not
1.1 root 939: want to copy. */
940: insert_before = PREV_INSN (last_loop_insn);
941: copy_end = PREV_INSN (insert_before);
942: #else
1.1.1.2 root 943: /* The immediately preceding insn may not be a compare, so we
1.1 root 944: must copy it. */
945: insert_before = last_loop_insn;
946: copy_end = PREV_INSN (last_loop_insn);
947: #endif
948: }
949:
950: /* Set unroll type to MODULO now. */
951: unroll_type = UNROLL_MODULO;
952: loop_preconditioned = 1;
953: }
954: }
955:
956: /* If reach here, and the loop type is UNROLL_NAIVE, then don't unroll
957: the loop unless all loops are being unrolled. */
958: if (unroll_type == UNROLL_NAIVE && ! flag_unroll_all_loops)
959: {
960: if (loop_dump_stream)
961: fprintf (loop_dump_stream, "Unrolling failure: Naive unrolling not being done.\n");
962: return;
963: }
964:
965: /* At this point, we are guaranteed to unroll the loop. */
966:
967: /* For each biv and giv, determine whether it can be safely split into
968: a different variable for each unrolled copy of the loop body.
969: We precalculate and save this info here, since computing it is
970: expensive.
971:
972: Do this before deleting any instructions from the loop, so that
973: back_branch_in_range_p will work correctly. */
974:
975: if (splitting_not_safe)
976: temp = 0;
977: else
978: temp = find_splittable_regs (unroll_type, loop_start, loop_end,
979: end_insert_before, unroll_number);
980:
981: /* find_splittable_regs may have created some new registers, so must
982: reallocate the reg_map with the new larger size, and must realloc
983: the constant maps also. */
984:
985: maxregnum = max_reg_num ();
986: map->reg_map = (rtx *) alloca (maxregnum * sizeof (rtx));
987:
988: init_reg_map (map, maxregnum);
989:
990: /* Space is needed in some of the map for new registers, so new_maxregnum
991: is an (over)estimate of how many registers will exist at the end. */
992: new_maxregnum = maxregnum + (temp * unroll_number * 2);
993:
994: /* Must realloc space for the constant maps, because the number of registers
995: may have changed. */
996:
997: map->const_equiv_map = (rtx *) alloca (new_maxregnum * sizeof (rtx));
998: map->const_age_map = (unsigned *) alloca (new_maxregnum * sizeof (unsigned));
999:
1000: global_const_equiv_map = map->const_equiv_map;
1001:
1002: /* Search the list of bivs and givs to find ones which need to be remapped
1003: when split, and set their reg_map entry appropriately. */
1004:
1005: for (bl = loop_iv_list; bl; bl = bl->next)
1006: {
1007: if (REGNO (bl->biv->src_reg) != bl->regno)
1008: map->reg_map[bl->regno] = bl->biv->src_reg;
1009: #if 0
1010: /* Currently, non-reduced/final-value givs are never split. */
1011: for (v = bl->giv; v; v = v->next_iv)
1012: if (REGNO (v->src_reg) != bl->regno)
1013: map->reg_map[REGNO (v->dest_reg)] = v->src_reg;
1014: #endif
1015: }
1016:
1017: /* If the loop is being partially unrolled, and the iteration variables
1018: are being split, and are being renamed for the split, then must fix up
1019: the compare instruction at the end of the loop to refer to the new
1020: registers. This compare isn't copied, so the registers used in it
1021: will never be replaced if it isn't done here. */
1022:
1023: if (unroll_type == UNROLL_MODULO)
1024: {
1025: insn = NEXT_INSN (copy_end);
1026: if (GET_CODE (insn) == INSN && GET_CODE (PATTERN (insn)) == SET)
1027: {
1028: #if 0
1029: /* If non-reduced/final-value givs were split, then this would also
1030: have to remap those givs. */
1031: #endif
1032:
1033: tem = SET_SRC (PATTERN (insn));
1034: /* The set source is a register. */
1035: if (GET_CODE (tem) == REG)
1036: {
1037: if (REGNO (tem) < max_reg_before_loop
1038: && reg_iv_type[REGNO (tem)] == BASIC_INDUCT)
1039: SET_SRC (PATTERN (insn))
1040: = reg_biv_class[REGNO (tem)]->biv->src_reg;
1041: }
1042: else
1043: {
1044: /* The set source is a compare of some sort. */
1045: tem = XEXP (SET_SRC (PATTERN (insn)), 0);
1046: if (GET_CODE (tem) == REG
1047: && REGNO (tem) < max_reg_before_loop
1048: && reg_iv_type[REGNO (tem)] == BASIC_INDUCT)
1049: XEXP (SET_SRC (PATTERN (insn)), 0)
1050: = reg_biv_class[REGNO (tem)]->biv->src_reg;
1051:
1052: tem = XEXP (SET_SRC (PATTERN (insn)), 1);
1053: if (GET_CODE (tem) == REG
1054: && REGNO (tem) < max_reg_before_loop
1055: && reg_iv_type[REGNO (tem)] == BASIC_INDUCT)
1056: XEXP (SET_SRC (PATTERN (insn)), 1)
1057: = reg_biv_class[REGNO (tem)]->biv->src_reg;
1058: }
1059: }
1060: }
1061:
1062: /* For unroll_number - 1 times, make a copy of each instruction
1063: between copy_start and copy_end, and insert these new instructions
1064: before the end of the loop. */
1065:
1066: for (i = 0; i < unroll_number; i++)
1067: {
1068: bzero (map->insn_map, max_insnno * sizeof (rtx));
1069: bzero (map->const_equiv_map, new_maxregnum * sizeof (rtx));
1070: bzero (map->const_age_map, new_maxregnum * sizeof (unsigned));
1071: map->const_age = 0;
1072:
1073: for (j = 0; j < max_labelno; j++)
1074: if (local_label[j])
1075: map->label_map[j] = gen_label_rtx ();
1076:
1077: /* If loop starts with a branch to the test, then fix it so that
1078: it points to the test of the first unrolled copy of the loop. */
1079: if (i == 0 && loop_start != copy_start)
1080: {
1081: insn = PREV_INSN (copy_start);
1082: pattern = PATTERN (insn);
1083:
1084: tem = map->label_map[CODE_LABEL_NUMBER
1085: (XEXP (SET_SRC (pattern), 0))];
1086: SET_SRC (pattern) = gen_rtx (LABEL_REF, VOIDmode, tem);
1087:
1088: /* Set the jump label so that it can be used by later loop unrolling
1089: passes. */
1090: JUMP_LABEL (insn) = tem;
1091: LABEL_NUSES (tem)++;
1092: }
1093:
1094: copy_loop_body (copy_start, copy_end, map, exit_label,
1095: i == unroll_number - 1, unroll_type, start_label,
1096: loop_end, insert_before, insert_before);
1097: }
1098:
1099: /* Before deleting any insns, emit a CODE_LABEL immediately after the last
1100: insn to be deleted. This prevents any runaway delete_insn call from
1101: more insns that it should, as it always stops at a CODE_LABEL. */
1102:
1103: /* Delete the compare and branch at the end of the loop if completely
1104: unrolling the loop. Deleting the backward branch at the end also
1105: deletes the code label at the start of the loop. This is done at
1106: the very end to avoid problems with back_branch_in_range_p. */
1107:
1108: if (unroll_type == UNROLL_COMPLETELY)
1109: safety_label = emit_label_after (gen_label_rtx (), last_loop_insn);
1110: else
1111: safety_label = emit_label_after (gen_label_rtx (), copy_end);
1112:
1113: /* Delete all of the original loop instructions. Don't delete the
1114: LOOP_BEG note, or the first code label in the loop. */
1115:
1116: insn = NEXT_INSN (copy_start);
1117: while (insn != safety_label)
1118: {
1119: if (insn != start_label)
1120: insn = delete_insn (insn);
1121: else
1122: insn = NEXT_INSN (insn);
1123: }
1124:
1125: /* Can now delete the 'safety' label emitted to protect us from runaway
1126: delete_insn calls. */
1127: if (INSN_DELETED_P (safety_label))
1128: abort ();
1129: delete_insn (safety_label);
1130:
1131: /* If exit_label exists, emit it after the loop. Doing the emit here
1132: forces it to have a higher INSN_UID than any insn in the unrolled loop.
1133: This is needed so that mostly_true_jump in reorg.c will treat jumps
1134: to this loop end label correctly, i.e. predict that they are usually
1135: not taken. */
1136: if (exit_label)
1137: emit_label_after (exit_label, loop_end);
1138: }
1139:
1140: /* Return true if the loop can be safely, and profitably, preconditioned
1141: so that the unrolled copies of the loop body don't need exit tests.
1142:
1143: This only works if final_value, initial_value and increment can be
1144: determined, and if increment is a constant power of 2.
1145: If increment is not a power of 2, then the preconditioning modulo
1146: operation would require a real modulo instead of a boolean AND, and this
1147: is not considered `profitable'. */
1148:
1149: /* ??? If the loop is known to be executed very many times, or the machine
1150: has a very cheap divide instruction, then preconditioning is a win even
1151: when the increment is not a power of 2. Use RTX_COST to compute
1152: whether divide is cheap. */
1153:
1154: static int
1155: precondition_loop_p (initial_value, final_value, increment, loop_start,
1156: loop_end)
1157: rtx *initial_value, *final_value, *increment;
1158: rtx loop_start, loop_end;
1159: {
1160: int unsigned_compare, compare_dir;
1161:
1162: if (loop_n_iterations > 0)
1163: {
1164: *initial_value = const0_rtx;
1165: *increment = const1_rtx;
1.1.1.4 ! root 1166: *final_value = GEN_INT (loop_n_iterations);
1.1 root 1167:
1168: if (loop_dump_stream)
1169: fprintf (loop_dump_stream,
1170: "Preconditioning: Success, number of iterations known, %d.\n",
1171: loop_n_iterations);
1172: return 1;
1173: }
1174:
1175: if (loop_initial_value == 0)
1176: {
1177: if (loop_dump_stream)
1178: fprintf (loop_dump_stream,
1179: "Preconditioning: Could not find initial value.\n");
1180: return 0;
1181: }
1182: else if (loop_increment == 0)
1183: {
1184: if (loop_dump_stream)
1185: fprintf (loop_dump_stream,
1186: "Preconditioning: Could not find increment value.\n");
1187: return 0;
1188: }
1189: else if (GET_CODE (loop_increment) != CONST_INT)
1190: {
1191: if (loop_dump_stream)
1192: fprintf (loop_dump_stream,
1193: "Preconditioning: Increment not a constant.\n");
1194: return 0;
1195: }
1196: else if ((exact_log2 (INTVAL (loop_increment)) < 0)
1197: && (exact_log2 (- INTVAL (loop_increment)) < 0))
1198: {
1199: if (loop_dump_stream)
1200: fprintf (loop_dump_stream,
1201: "Preconditioning: Increment not a constant power of 2.\n");
1202: return 0;
1203: }
1204:
1205: /* Unsigned_compare and compare_dir can be ignored here, since they do
1206: not matter for preconditioning. */
1207:
1208: if (loop_final_value == 0)
1209: {
1210: if (loop_dump_stream)
1211: fprintf (loop_dump_stream,
1212: "Preconditioning: EQ comparison loop.\n");
1213: return 0;
1214: }
1215:
1216: /* Must ensure that final_value is invariant, so call invariant_p to
1217: check. Before doing so, must check regno against max_reg_before_loop
1.1.1.3 root 1218: to make sure that the register is in the range covered by invariant_p.
1.1 root 1219: If it isn't, then it is most likely a biv/giv which by definition are
1220: not invariant. */
1221: if ((GET_CODE (loop_final_value) == REG
1222: && REGNO (loop_final_value) >= max_reg_before_loop)
1223: || (GET_CODE (loop_final_value) == PLUS
1224: && REGNO (XEXP (loop_final_value, 0)) >= max_reg_before_loop)
1225: || ! invariant_p (loop_final_value))
1226: {
1227: if (loop_dump_stream)
1228: fprintf (loop_dump_stream,
1229: "Preconditioning: Final value not invariant.\n");
1230: return 0;
1231: }
1232:
1233: /* Fail for floating point values, since the caller of this function
1234: does not have code to deal with them. */
1235: if (GET_MODE_CLASS (GET_MODE (loop_final_value)) == MODE_FLOAT
1.1.1.2 root 1236: || GET_MODE_CLASS (GET_MODE (loop_initial_value)) == MODE_FLOAT)
1.1 root 1237: {
1238: if (loop_dump_stream)
1239: fprintf (loop_dump_stream,
1240: "Preconditioning: Floating point final or initial value.\n");
1241: return 0;
1242: }
1243:
1244: /* Now set initial_value to be the iteration_var, since that may be a
1245: simpler expression, and is guaranteed to be correct if all of the
1246: above tests succeed.
1247:
1248: We can not use the initial_value as calculated, because it will be
1249: one too small for loops of the form "while (i-- > 0)". We can not
1250: emit code before the loop_skip_over insns to fix this problem as this
1251: will then give a number one too large for loops of the form
1252: "while (--i > 0)".
1253:
1254: Note that all loops that reach here are entered at the top, because
1255: this function is not called if the loop starts with a jump. */
1256:
1257: /* Fail if loop_iteration_var is not live before loop_start, since we need
1258: to test its value in the preconditioning code. */
1259:
1260: if (uid_luid[regno_first_uid[REGNO (loop_iteration_var)]]
1261: > INSN_LUID (loop_start))
1262: {
1263: if (loop_dump_stream)
1264: fprintf (loop_dump_stream,
1265: "Preconditioning: Iteration var not live before loop start.\n");
1266: return 0;
1267: }
1268:
1269: *initial_value = loop_iteration_var;
1270: *increment = loop_increment;
1271: *final_value = loop_final_value;
1272:
1273: /* Success! */
1274: if (loop_dump_stream)
1275: fprintf (loop_dump_stream, "Preconditioning: Successful.\n");
1276: return 1;
1277: }
1278:
1279:
1280: /* All pseudo-registers must be mapped to themselves. Two hard registers
1281: must be mapped, VIRTUAL_STACK_VARS_REGNUM and VIRTUAL_INCOMING_ARGS_
1282: REGNUM, to avoid function-inlining specific conversions of these
1283: registers. All other hard regs can not be mapped because they may be
1284: used with different
1285: modes. */
1286:
1287: static void
1288: init_reg_map (map, maxregnum)
1289: struct inline_remap *map;
1290: int maxregnum;
1291: {
1292: int i;
1293:
1294: for (i = maxregnum - 1; i > LAST_VIRTUAL_REGISTER; i--)
1295: map->reg_map[i] = regno_reg_rtx[i];
1296: /* Just clear the rest of the entries. */
1297: for (i = LAST_VIRTUAL_REGISTER; i >= 0; i--)
1298: map->reg_map[i] = 0;
1299:
1300: map->reg_map[VIRTUAL_STACK_VARS_REGNUM]
1301: = regno_reg_rtx[VIRTUAL_STACK_VARS_REGNUM];
1302: map->reg_map[VIRTUAL_INCOMING_ARGS_REGNUM]
1303: = regno_reg_rtx[VIRTUAL_INCOMING_ARGS_REGNUM];
1304: }
1305:
1306: /* Strength-reduction will often emit code for optimized biv/givs which
1307: calculates their value in a temporary register, and then copies the result
1308: to the iv. This procedure reconstructs the pattern computing the iv;
1309: verifying that all operands are of the proper form.
1310:
1311: The return value is the amount that the giv is incremented by. */
1312:
1313: static rtx
1314: calculate_giv_inc (pattern, src_insn, regno)
1315: rtx pattern, src_insn;
1316: int regno;
1317: {
1318: rtx increment;
1319:
1320: /* Verify that we have an increment insn here. First check for a plus
1321: as the set source. */
1322: if (GET_CODE (SET_SRC (pattern)) != PLUS)
1323: {
1324: /* SR sometimes computes the new giv value in a temp, then copies it
1325: to the new_reg. */
1326: src_insn = PREV_INSN (src_insn);
1327: pattern = PATTERN (src_insn);
1328: if (GET_CODE (SET_SRC (pattern)) != PLUS)
1329: abort ();
1330:
1331: /* The last insn emitted is not needed, so delete it to avoid confusing
1332: the second cse pass. This insn sets the giv unnecessarily. */
1333: delete_insn (get_last_insn ());
1334: }
1335:
1336: /* Verify that we have a constant as the second operand of the plus. */
1337: increment = XEXP (SET_SRC (pattern), 1);
1338: if (GET_CODE (increment) != CONST_INT)
1339: {
1340: /* SR sometimes puts the constant in a register, especially if it is
1341: too big to be an add immed operand. */
1342: increment = SET_SRC (PATTERN (PREV_INSN (src_insn)));
1343:
1344: /* SR may have used LO_SUM to compute the constant if it is too large
1345: for a load immed operand. In this case, the constant is in operand
1346: one of the LO_SUM rtx. */
1347: if (GET_CODE (increment) == LO_SUM)
1348: increment = XEXP (increment, 1);
1349:
1350: if (GET_CODE (increment) != CONST_INT)
1351: abort ();
1352:
1353: /* The insn loading the constant into a register is not longer needed,
1354: so delete it. */
1355: delete_insn (get_last_insn ());
1356: }
1357:
1358: /* Check that the source register is the same as the dest register. */
1359: if (GET_CODE (XEXP (SET_SRC (pattern), 0)) != REG
1360: || REGNO (XEXP (SET_SRC (pattern), 0)) != regno)
1361: abort ();
1362:
1363: return increment;
1364: }
1365:
1366:
1367: /* Copy each instruction in the loop, substituting from map as appropriate.
1368: This is very similar to a loop in expand_inline_function. */
1369:
1370: static void
1371: copy_loop_body (copy_start, copy_end, map, exit_label, last_iteration,
1372: unroll_type, start_label, loop_end, insert_before,
1373: copy_notes_from)
1374: rtx copy_start, copy_end;
1375: struct inline_remap *map;
1.1.1.4 ! root 1376: rtx exit_label;
1.1 root 1377: int last_iteration;
1378: enum unroll_types unroll_type;
1379: rtx start_label, loop_end, insert_before, copy_notes_from;
1380: {
1381: rtx insn, pattern;
1382: rtx tem, copy;
1383: int dest_reg_was_split, i;
1384: rtx cc0_insn = 0;
1385: rtx final_label = 0;
1386: rtx giv_inc, giv_dest_reg, giv_src_reg;
1387:
1388: /* If this isn't the last iteration, then map any references to the
1389: start_label to final_label. Final label will then be emitted immediately
1390: after the end of this loop body if it was ever used.
1391:
1392: If this is the last iteration, then map references to the start_label
1393: to itself. */
1394: if (! last_iteration)
1395: {
1396: final_label = gen_label_rtx ();
1397: map->label_map[CODE_LABEL_NUMBER (start_label)] = final_label;
1398: }
1399: else
1400: map->label_map[CODE_LABEL_NUMBER (start_label)] = start_label;
1401:
1402: start_sequence ();
1403:
1404: insn = copy_start;
1405: do
1406: {
1407: insn = NEXT_INSN (insn);
1408:
1409: map->orig_asm_operands_vector = 0;
1410:
1411: switch (GET_CODE (insn))
1412: {
1413: case INSN:
1414: pattern = PATTERN (insn);
1415: copy = 0;
1416: giv_inc = 0;
1417:
1418: /* Check to see if this is a giv that has been combined with
1419: some split address givs. (Combined in the sense that
1420: `combine_givs' in loop.c has put two givs in the same register.)
1421: In this case, we must search all givs based on the same biv to
1422: find the address givs. Then split the address givs.
1423: Do this before splitting the giv, since that may map the
1424: SET_DEST to a new register. */
1425:
1426: if (GET_CODE (pattern) == SET
1427: && GET_CODE (SET_DEST (pattern)) == REG
1428: && addr_combined_regs[REGNO (SET_DEST (pattern))])
1429: {
1430: struct iv_class *bl;
1431: struct induction *v, *tv;
1432: int regno = REGNO (SET_DEST (pattern));
1433:
1434: v = addr_combined_regs[REGNO (SET_DEST (pattern))];
1435: bl = reg_biv_class[REGNO (v->src_reg)];
1436:
1437: /* Although the giv_inc amount is not needed here, we must call
1438: calculate_giv_inc here since it might try to delete the
1439: last insn emitted. If we wait until later to call it,
1440: we might accidentally delete insns generated immediately
1441: below by emit_unrolled_add. */
1442:
1443: giv_inc = calculate_giv_inc (pattern, insn, regno);
1444:
1445: /* Now find all address giv's that were combined with this
1446: giv 'v'. */
1447: for (tv = bl->giv; tv; tv = tv->next_iv)
1448: if (tv->giv_type == DEST_ADDR && tv->same == v)
1449: {
1.1.1.3 root 1450: int this_giv_inc = INTVAL (giv_inc);
1451:
1452: /* Scale this_giv_inc if the multiplicative factors of
1453: the two givs are different. */
1454: if (tv->mult_val != v->mult_val)
1455: this_giv_inc = (this_giv_inc / INTVAL (v->mult_val)
1456: * INTVAL (tv->mult_val));
1457:
1458: tv->dest_reg = plus_constant (tv->dest_reg, this_giv_inc);
1.1 root 1459: *tv->location = tv->dest_reg;
1460:
1461: if (last_iteration && unroll_type != UNROLL_COMPLETELY)
1462: {
1463: /* Must emit an insn to increment the split address
1464: giv. Add in the const_adjust field in case there
1465: was a constant eliminated from the address. */
1466: rtx value, dest_reg;
1467:
1468: /* tv->dest_reg will be either a bare register,
1469: or else a register plus a constant. */
1470: if (GET_CODE (tv->dest_reg) == REG)
1471: dest_reg = tv->dest_reg;
1472: else
1473: dest_reg = XEXP (tv->dest_reg, 0);
1474:
1475: /* tv->dest_reg may actually be a (PLUS (REG) (CONST))
1476: here, so we must call plus_constant to add
1477: the const_adjust amount before calling
1478: emit_unrolled_add below. */
1479: value = plus_constant (tv->dest_reg, tv->const_adjust);
1480:
1481: /* The constant could be too large for an add
1482: immediate, so can't directly emit an insn here. */
1483: emit_unrolled_add (dest_reg, XEXP (value, 0),
1484: XEXP (value, 1));
1485:
1486: /* Reset the giv to be just the register again, in case
1.1.1.2 root 1487: it is used after the set we have just emitted.
1488: We must subtract the const_adjust factor added in
1489: above. */
1490: tv->dest_reg = plus_constant (dest_reg,
1491: - tv->const_adjust);
1.1 root 1492: *tv->location = tv->dest_reg;
1493: }
1494: }
1495: }
1496:
1497: /* If this is a setting of a splittable variable, then determine
1498: how to split the variable, create a new set based on this split,
1499: and set up the reg_map so that later uses of the variable will
1500: use the new split variable. */
1501:
1502: dest_reg_was_split = 0;
1503:
1504: if (GET_CODE (pattern) == SET
1505: && GET_CODE (SET_DEST (pattern)) == REG
1506: && splittable_regs[REGNO (SET_DEST (pattern))])
1507: {
1508: int regno = REGNO (SET_DEST (pattern));
1509:
1510: dest_reg_was_split = 1;
1511:
1512: /* Compute the increment value for the giv, if it wasn't
1513: already computed above. */
1514:
1515: if (giv_inc == 0)
1516: giv_inc = calculate_giv_inc (pattern, insn, regno);
1517: giv_dest_reg = SET_DEST (pattern);
1518: giv_src_reg = SET_DEST (pattern);
1519:
1520: if (unroll_type == UNROLL_COMPLETELY)
1521: {
1522: /* Completely unrolling the loop. Set the induction
1523: variable to a known constant value. */
1524:
1525: /* The value in splittable_regs may be an invariant
1526: value, so we must use plus_constant here. */
1527: splittable_regs[regno]
1528: = plus_constant (splittable_regs[regno], INTVAL (giv_inc));
1529:
1530: if (GET_CODE (splittable_regs[regno]) == PLUS)
1531: {
1532: giv_src_reg = XEXP (splittable_regs[regno], 0);
1533: giv_inc = XEXP (splittable_regs[regno], 1);
1534: }
1535: else
1536: {
1537: /* The splittable_regs value must be a REG or a
1538: CONST_INT, so put the entire value in the giv_src_reg
1539: variable. */
1540: giv_src_reg = splittable_regs[regno];
1541: giv_inc = const0_rtx;
1542: }
1543: }
1544: else
1545: {
1546: /* Partially unrolling loop. Create a new pseudo
1547: register for the iteration variable, and set it to
1548: be a constant plus the original register. Except
1549: on the last iteration, when the result has to
1550: go back into the original iteration var register. */
1551:
1552: /* Handle bivs which must be mapped to a new register
1553: when split. This happens for bivs which need their
1554: final value set before loop entry. The new register
1555: for the biv was stored in the biv's first struct
1556: induction entry by find_splittable_regs. */
1557:
1558: if (regno < max_reg_before_loop
1559: && reg_iv_type[regno] == BASIC_INDUCT)
1560: {
1561: giv_src_reg = reg_biv_class[regno]->biv->src_reg;
1562: giv_dest_reg = giv_src_reg;
1563: }
1564:
1565: #if 0
1566: /* If non-reduced/final-value givs were split, then
1567: this would have to remap those givs also. See
1568: find_splittable_regs. */
1569: #endif
1570:
1571: splittable_regs[regno]
1.1.1.4 ! root 1572: = GEN_INT (INTVAL (giv_inc)
1.1 root 1573: + INTVAL (splittable_regs[regno]));
1574: giv_inc = splittable_regs[regno];
1575:
1576: /* Now split the induction variable by changing the dest
1577: of this insn to a new register, and setting its
1578: reg_map entry to point to this new register.
1579:
1580: If this is the last iteration, and this is the last insn
1581: that will update the iv, then reuse the original dest,
1582: to ensure that the iv will have the proper value when
1583: the loop exits or repeats.
1584:
1585: Using splittable_regs_updates here like this is safe,
1586: because it can only be greater than one if all
1587: instructions modifying the iv are always executed in
1588: order. */
1589:
1590: if (! last_iteration
1591: || (splittable_regs_updates[regno]-- != 1))
1592: {
1593: tem = gen_reg_rtx (GET_MODE (giv_src_reg));
1594: giv_dest_reg = tem;
1595: map->reg_map[regno] = tem;
1596: }
1597: else
1598: map->reg_map[regno] = giv_src_reg;
1599: }
1600:
1601: /* The constant being added could be too large for an add
1602: immediate, so can't directly emit an insn here. */
1603: emit_unrolled_add (giv_dest_reg, giv_src_reg, giv_inc);
1604: copy = get_last_insn ();
1605: pattern = PATTERN (copy);
1606: }
1607: else
1608: {
1609: pattern = copy_rtx_and_substitute (pattern, map);
1610: copy = emit_insn (pattern);
1611: }
1612: /* REG_NOTES will be copied later. */
1613:
1614: #ifdef HAVE_cc0
1615: /* If this insn is setting CC0, it may need to look at
1616: the insn that uses CC0 to see what type of insn it is.
1617: In that case, the call to recog via validate_change will
1618: fail. So don't substitute constants here. Instead,
1619: do it when we emit the following insn.
1620:
1621: For example, see the pyr.md file. That machine has signed and
1622: unsigned compares. The compare patterns must check the
1623: following branch insn to see which what kind of compare to
1624: emit.
1625:
1626: If the previous insn set CC0, substitute constants on it as
1627: well. */
1628: if (sets_cc0_p (copy) != 0)
1629: cc0_insn = copy;
1630: else
1631: {
1632: if (cc0_insn)
1633: try_constants (cc0_insn, map);
1634: cc0_insn = 0;
1635: try_constants (copy, map);
1636: }
1637: #else
1638: try_constants (copy, map);
1639: #endif
1640:
1641: /* Make split induction variable constants `permanent' since we
1642: know there are no backward branches across iteration variable
1643: settings which would invalidate this. */
1644: if (dest_reg_was_split)
1645: {
1646: int regno = REGNO (SET_DEST (pattern));
1647:
1648: if (map->const_age_map[regno] == map->const_age)
1649: map->const_age_map[regno] = -1;
1650: }
1651: break;
1652:
1653: case JUMP_INSN:
1654: if (JUMP_LABEL (insn) == start_label && insn == copy_end
1655: && ! last_iteration)
1656: {
1657: /* This is a branch to the beginning of the loop; this is the
1658: last insn being copied; and this is not the last iteration.
1659: In this case, we want to change the original fall through
1660: case to be a branch past the end of the loop, and the
1661: original jump label case to fall_through. */
1662: /* Never map the label in this case. */
1.1.1.4 ! root 1663:
1.1 root 1664: pattern = copy_rtx (PATTERN (insn));
1.1.1.4 ! root 1665: copy = emit_jump_insn (pattern);
1.1 root 1666:
1.1.1.4 ! root 1667: if (! invert_exp (pattern, copy)
! 1668: || ! redirect_exp (&pattern, JUMP_LABEL (insn),
! 1669: exit_label, copy))
! 1670: abort ();
1.1 root 1671: }
1672: else
1.1.1.4 ! root 1673: {
! 1674: pattern = copy_rtx_and_substitute (PATTERN (insn), map);
! 1675: copy = emit_jump_insn (pattern);
! 1676: }
1.1 root 1677:
1678: #ifdef HAVE_cc0
1679: if (cc0_insn)
1680: try_constants (cc0_insn, map);
1681: cc0_insn = 0;
1682: #endif
1683: try_constants (copy, map);
1684:
1685: /* Set the jump label of COPY correctly to avoid problems with
1686: later passes of unroll_loop, if INSN had jump label set. */
1687: if (JUMP_LABEL (insn))
1688: {
1.1.1.4 ! root 1689: rtx label = 0;
! 1690:
1.1 root 1691: /* Can't use the label_map for every insn, since this may be
1692: the backward branch, and hence the label was not mapped. */
1693: if (GET_CODE (pattern) == SET)
1694: {
1695: tem = SET_SRC (pattern);
1696: if (GET_CODE (tem) == LABEL_REF)
1.1.1.4 ! root 1697: label = XEXP (tem, 0);
1.1 root 1698: else if (GET_CODE (tem) == IF_THEN_ELSE)
1699: {
1700: if (XEXP (tem, 1) != pc_rtx)
1.1.1.4 ! root 1701: label = XEXP (XEXP (tem, 1), 0);
1.1 root 1702: else
1.1.1.4 ! root 1703: label = XEXP (XEXP (tem, 2), 0);
1.1 root 1704: }
1705: }
1.1.1.4 ! root 1706:
! 1707: if (label && GET_CODE (label) == CODE_LABEL)
! 1708: JUMP_LABEL (copy) = label;
1.1 root 1709: else
1710: {
1711: /* An unrecognizable jump insn, probably the entry jump
1712: for a switch statement. This label must have been mapped,
1713: so just use the label_map to get the new jump label. */
1714: JUMP_LABEL (copy) = map->label_map[CODE_LABEL_NUMBER
1715: (JUMP_LABEL (insn))];
1716: }
1717:
1718: /* If this is a non-local jump, then must increase the label
1719: use count so that the label will not be deleted when the
1720: original jump is deleted. */
1721: LABEL_NUSES (JUMP_LABEL (copy))++;
1722: }
1723: else if (GET_CODE (PATTERN (copy)) == ADDR_VEC
1724: || GET_CODE (PATTERN (copy)) == ADDR_DIFF_VEC)
1725: {
1726: rtx pat = PATTERN (copy);
1727: int diff_vec_p = GET_CODE (pat) == ADDR_DIFF_VEC;
1728: int len = XVECLEN (pat, diff_vec_p);
1729: int i;
1730:
1731: for (i = 0; i < len; i++)
1732: LABEL_NUSES (XEXP (XVECEXP (pat, diff_vec_p, i), 0))++;
1733: }
1734:
1735: /* If this used to be a conditional jump insn but whose branch
1736: direction is now known, we must do something special. */
1737: if (condjump_p (insn) && !simplejump_p (insn) && map->last_pc_value)
1738: {
1739: #ifdef HAVE_cc0
1740: /* The previous insn set cc0 for us. So delete it. */
1741: delete_insn (PREV_INSN (copy));
1742: #endif
1743:
1744: /* If this is now a no-op, delete it. */
1745: if (map->last_pc_value == pc_rtx)
1746: {
1747: delete_insn (copy);
1748: copy = 0;
1749: }
1750: else
1751: /* Otherwise, this is unconditional jump so we must put a
1752: BARRIER after it. We could do some dead code elimination
1753: here, but jump.c will do it just as well. */
1754: emit_barrier ();
1755: }
1756: break;
1757:
1758: case CALL_INSN:
1759: pattern = copy_rtx_and_substitute (PATTERN (insn), map);
1760: copy = emit_call_insn (pattern);
1761:
1762: #ifdef HAVE_cc0
1763: if (cc0_insn)
1764: try_constants (cc0_insn, map);
1765: cc0_insn = 0;
1766: #endif
1767: try_constants (copy, map);
1768:
1769: /* Be lazy and assume CALL_INSNs clobber all hard registers. */
1770: for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
1771: map->const_equiv_map[i] = 0;
1772: break;
1773:
1774: case CODE_LABEL:
1775: /* If this is the loop start label, then we don't need to emit a
1776: copy of this label since no one will use it. */
1777:
1778: if (insn != start_label)
1779: {
1780: copy = emit_label (map->label_map[CODE_LABEL_NUMBER (insn)]);
1781: map->const_age++;
1782: }
1783: break;
1784:
1785: case BARRIER:
1786: copy = emit_barrier ();
1787: break;
1788:
1789: case NOTE:
1.1.1.4 ! root 1790: /* VTOP notes are valid only before the loop exit test. If placed
! 1791: anywhere else, loop may generate bad code. */
! 1792:
! 1793: if (NOTE_LINE_NUMBER (insn) != NOTE_INSN_DELETED
! 1794: && (NOTE_LINE_NUMBER (insn) != NOTE_INSN_LOOP_VTOP
! 1795: || (last_iteration && unroll_type != UNROLL_COMPLETELY)))
1.1 root 1796: copy = emit_note (NOTE_SOURCE_FILE (insn),
1797: NOTE_LINE_NUMBER (insn));
1798: else
1799: copy = 0;
1800: break;
1801:
1802: default:
1803: abort ();
1804: break;
1805: }
1806:
1807: map->insn_map[INSN_UID (insn)] = copy;
1808: }
1809: while (insn != copy_end);
1810:
1811: /* Now copy the REG_NOTES. */
1812: insn = copy_start;
1813: do
1814: {
1815: insn = NEXT_INSN (insn);
1816: if ((GET_CODE (insn) == INSN || GET_CODE (insn) == JUMP_INSN
1817: || GET_CODE (insn) == CALL_INSN)
1818: && map->insn_map[INSN_UID (insn)])
1819: REG_NOTES (map->insn_map[INSN_UID (insn)])
1820: = copy_rtx_and_substitute (REG_NOTES (insn), map);
1821: }
1822: while (insn != copy_end);
1823:
1824: /* There may be notes between copy_notes_from and loop_end. Emit a copy of
1825: each of these notes here, since there may be some important ones, such as
1826: NOTE_INSN_BLOCK_END notes, in this group. We don't do this on the last
1827: iteration, because the original notes won't be deleted.
1828:
1829: We can't use insert_before here, because when from preconditioning,
1830: insert_before points before the loop. We can't use copy_end, because
1831: there may be insns already inserted after it (which we don't want to
1832: copy) when not from preconditioning code. */
1833:
1834: if (! last_iteration)
1835: {
1836: for (insn = copy_notes_from; insn != loop_end; insn = NEXT_INSN (insn))
1837: {
1838: if (GET_CODE (insn) == NOTE
1839: && NOTE_LINE_NUMBER (insn) != NOTE_INSN_DELETED)
1840: emit_note (NOTE_SOURCE_FILE (insn), NOTE_LINE_NUMBER (insn));
1841: }
1842: }
1843:
1844: if (final_label && LABEL_NUSES (final_label) > 0)
1845: emit_label (final_label);
1846:
1847: tem = gen_sequence ();
1848: end_sequence ();
1849: emit_insn_before (tem, insert_before);
1850: }
1851:
1852: /* Emit an insn, using the expand_binop to ensure that a valid insn is
1853: emitted. This will correctly handle the case where the increment value
1854: won't fit in the immediate field of a PLUS insns. */
1855:
1856: void
1857: emit_unrolled_add (dest_reg, src_reg, increment)
1858: rtx dest_reg, src_reg, increment;
1859: {
1860: rtx result;
1861:
1862: result = expand_binop (GET_MODE (dest_reg), add_optab, src_reg, increment,
1863: dest_reg, 0, OPTAB_LIB_WIDEN);
1864:
1865: if (dest_reg != result)
1866: emit_move_insn (dest_reg, result);
1867: }
1868:
1869: /* Searches the insns between INSN and LOOP_END. Returns 1 if there
1870: is a backward branch in that range that branches to somewhere between
1871: LOOP_START and INSN. Returns 0 otherwise. */
1872:
1.1.1.3 root 1873: /* ??? This is quadratic algorithm. Could be rewritten to be linear.
1.1 root 1874: In practice, this is not a problem, because this function is seldom called,
1875: and uses a negligible amount of CPU time on average. */
1876:
1877: static int
1878: back_branch_in_range_p (insn, loop_start, loop_end)
1879: rtx insn;
1880: rtx loop_start, loop_end;
1881: {
1882: rtx p, q, target_insn;
1883:
1884: /* Stop before we get to the backward branch at the end of the loop. */
1885: loop_end = prev_nonnote_insn (loop_end);
1886: if (GET_CODE (loop_end) == BARRIER)
1887: loop_end = PREV_INSN (loop_end);
1888:
1889: /* Check in case insn has been deleted, search forward for first non
1890: deleted insn following it. */
1891: while (INSN_DELETED_P (insn))
1892: insn = NEXT_INSN (insn);
1893:
1894: /* Check for the case where insn is the last insn in the loop. */
1895: if (insn == loop_end)
1896: return 0;
1897:
1898: for (p = NEXT_INSN (insn); p != loop_end; p = NEXT_INSN (p))
1899: {
1900: if (GET_CODE (p) == JUMP_INSN)
1901: {
1902: target_insn = JUMP_LABEL (p);
1903:
1904: /* Search from loop_start to insn, to see if one of them is
1905: the target_insn. We can't use INSN_LUID comparisons here,
1906: since insn may not have an LUID entry. */
1907: for (q = loop_start; q != insn; q = NEXT_INSN (q))
1908: if (q == target_insn)
1909: return 1;
1910: }
1911: }
1912:
1913: return 0;
1914: }
1915:
1916: /* Try to generate the simplest rtx for the expression
1917: (PLUS (MULT mult1 mult2) add1). This is used to calculate the initial
1918: value of giv's. */
1919:
1920: static rtx
1921: fold_rtx_mult_add (mult1, mult2, add1, mode)
1922: rtx mult1, mult2, add1;
1923: enum machine_mode mode;
1924: {
1925: rtx temp, mult_res;
1926: rtx result;
1927:
1928: /* The modes must all be the same. This should always be true. For now,
1929: check to make sure. */
1930: if ((GET_MODE (mult1) != mode && GET_MODE (mult1) != VOIDmode)
1931: || (GET_MODE (mult2) != mode && GET_MODE (mult2) != VOIDmode)
1932: || (GET_MODE (add1) != mode && GET_MODE (add1) != VOIDmode))
1933: abort ();
1934:
1935: /* Ensure that if at least one of mult1/mult2 are constant, then mult2
1936: will be a constant. */
1937: if (GET_CODE (mult1) == CONST_INT)
1938: {
1939: temp = mult2;
1940: mult2 = mult1;
1941: mult1 = temp;
1942: }
1943:
1944: mult_res = simplify_binary_operation (MULT, mode, mult1, mult2);
1945: if (! mult_res)
1946: mult_res = gen_rtx (MULT, mode, mult1, mult2);
1947:
1948: /* Again, put the constant second. */
1949: if (GET_CODE (add1) == CONST_INT)
1950: {
1951: temp = add1;
1952: add1 = mult_res;
1953: mult_res = temp;
1954: }
1955:
1956: result = simplify_binary_operation (PLUS, mode, add1, mult_res);
1957: if (! result)
1958: result = gen_rtx (PLUS, mode, add1, mult_res);
1959:
1960: return result;
1961: }
1962:
1963: /* Searches the list of induction struct's for the biv BL, to try to calculate
1964: the total increment value for one iteration of the loop as a constant.
1965:
1966: Returns the increment value as an rtx, simplified as much as possible,
1967: if it can be calculated. Otherwise, returns 0. */
1968:
1969: rtx
1970: biv_total_increment (bl, loop_start, loop_end)
1971: struct iv_class *bl;
1972: rtx loop_start, loop_end;
1973: {
1974: struct induction *v;
1975: rtx result;
1976:
1977: /* For increment, must check every instruction that sets it. Each
1978: instruction must be executed only once each time through the loop.
1979: To verify this, we check that the the insn is always executed, and that
1980: there are no backward branches after the insn that branch to before it.
1981: Also, the insn must have a mult_val of one (to make sure it really is
1982: an increment). */
1983:
1984: result = const0_rtx;
1985: for (v = bl->biv; v; v = v->next_iv)
1986: {
1987: if (v->always_computable && v->mult_val == const1_rtx
1988: && ! back_branch_in_range_p (v->insn, loop_start, loop_end))
1989: result = fold_rtx_mult_add (result, const1_rtx, v->add_val, v->mode);
1990: else
1991: return 0;
1992: }
1993:
1994: return result;
1995: }
1996:
1997: /* Determine the initial value of the iteration variable, and the amount
1998: that it is incremented each loop. Use the tables constructed by
1999: the strength reduction pass to calculate these values.
2000:
2001: Initial_value and/or increment are set to zero if their values could not
2002: be calculated. */
2003:
2004: static void
2005: iteration_info (iteration_var, initial_value, increment, loop_start, loop_end)
2006: rtx iteration_var, *initial_value, *increment;
2007: rtx loop_start, loop_end;
2008: {
2009: struct iv_class *bl;
2010: struct induction *v, *b;
2011:
2012: /* Clear the result values, in case no answer can be found. */
2013: *initial_value = 0;
2014: *increment = 0;
2015:
2016: /* The iteration variable can be either a giv or a biv. Check to see
2017: which it is, and compute the variable's initial value, and increment
2018: value if possible. */
2019:
2020: /* If this is a new register, can't handle it since we don't have any
2021: reg_iv_type entry for it. */
2022: if (REGNO (iteration_var) >= max_reg_before_loop)
2023: {
2024: if (loop_dump_stream)
2025: fprintf (loop_dump_stream,
2026: "Loop unrolling: No reg_iv_type entry for iteration var.\n");
2027: return;
2028: }
2029: /* Reject iteration variables larger than the host long size, since they
2030: could result in a number of iterations greater than the range of our
2031: `unsigned long' variable loop_n_iterations. */
2032: else if (GET_MODE_BITSIZE (GET_MODE (iteration_var)) > HOST_BITS_PER_LONG)
2033: {
2034: if (loop_dump_stream)
2035: fprintf (loop_dump_stream,
2036: "Loop unrolling: Iteration var rejected because mode larger than host long.\n");
2037: return;
2038: }
2039: else if (GET_MODE_CLASS (GET_MODE (iteration_var)) != MODE_INT)
2040: {
2041: if (loop_dump_stream)
2042: fprintf (loop_dump_stream,
1.1.1.2 root 2043: "Loop unrolling: Iteration var not an integer.\n");
1.1 root 2044: return;
2045: }
2046: else if (reg_iv_type[REGNO (iteration_var)] == BASIC_INDUCT)
2047: {
2048: /* Grab initial value, only useful if it is a constant. */
2049: bl = reg_biv_class[REGNO (iteration_var)];
2050: *initial_value = bl->initial_value;
2051:
2052: *increment = biv_total_increment (bl, loop_start, loop_end);
2053: }
2054: else if (reg_iv_type[REGNO (iteration_var)] == GENERAL_INDUCT)
2055: {
2056: #if 1
2057: /* ??? The code below does not work because the incorrect number of
2058: iterations is calculated when the biv is incremented after the giv
2059: is set (which is the usual case). This can probably be accounted
2060: for by biasing the initial_value by subtracting the amount of the
2061: increment that occurs between the giv set and the giv test. However,
2062: a giv as an iterator is very rare, so it does not seem worthwhile
2063: to handle this. */
2064: /* ??? An example failure is: i = 6; do {;} while (i++ < 9). */
2065: if (loop_dump_stream)
2066: fprintf (loop_dump_stream,
2067: "Loop unrolling: Giv iterators are not handled.\n");
2068: return;
2069: #else
2070: /* Initial value is mult_val times the biv's initial value plus
2071: add_val. Only useful if it is a constant. */
2072: v = reg_iv_info[REGNO (iteration_var)];
2073: bl = reg_biv_class[REGNO (v->src_reg)];
2074: *initial_value = fold_rtx_mult_add (v->mult_val, bl->initial_value,
2075: v->add_val, v->mode);
2076:
2077: /* Increment value is mult_val times the increment value of the biv. */
2078:
2079: *increment = biv_total_increment (bl, loop_start, loop_end);
2080: if (*increment)
2081: *increment = fold_rtx_mult_add (v->mult_val, *increment, const0_rtx,
2082: v->mode);
2083: #endif
2084: }
2085: else
2086: {
2087: if (loop_dump_stream)
2088: fprintf (loop_dump_stream,
2089: "Loop unrolling: Not basic or general induction var.\n");
2090: return;
2091: }
2092: }
2093:
2094: /* Calculate the approximate final value of the iteration variable
2095: which has an loop exit test with code COMPARISON_CODE and comparison value
2096: of COMPARISON_VALUE. Also returns an indication of whether the comparison
2097: was signed or unsigned, and the direction of the comparison. This info is
2098: needed to calculate the number of loop iterations. */
2099:
2100: static rtx
2101: approx_final_value (comparison_code, comparison_value, unsigned_p, compare_dir)
2102: enum rtx_code comparison_code;
2103: rtx comparison_value;
2104: int *unsigned_p;
2105: int *compare_dir;
2106: {
2107: /* Calculate the final value of the induction variable.
2108: The exact final value depends on the branch operator, and increment sign.
2109: This is only an approximate value. It will be wrong if the iteration
2110: variable is not incremented by one each time through the loop, and
2111: approx final value - start value % increment != 0. */
2112:
2113: *unsigned_p = 0;
2114: switch (comparison_code)
2115: {
2116: case LEU:
2117: *unsigned_p = 1;
2118: case LE:
2119: *compare_dir = 1;
2120: return plus_constant (comparison_value, 1);
2121: case GEU:
2122: *unsigned_p = 1;
2123: case GE:
2124: *compare_dir = -1;
2125: return plus_constant (comparison_value, -1);
2126: case EQ:
2127: /* Can not calculate a final value for this case. */
2128: *compare_dir = 0;
2129: return 0;
2130: case LTU:
2131: *unsigned_p = 1;
2132: case LT:
2133: *compare_dir = 1;
2134: return comparison_value;
2135: break;
2136: case GTU:
2137: *unsigned_p = 1;
2138: case GT:
2139: *compare_dir = -1;
2140: return comparison_value;
2141: case NE:
2142: *compare_dir = 0;
2143: return comparison_value;
2144: default:
2145: abort ();
2146: }
2147: }
2148:
2149: /* For each biv and giv, determine whether it can be safely split into
2150: a different variable for each unrolled copy of the loop body. If it
2151: is safe to split, then indicate that by saving some useful info
2152: in the splittable_regs array.
2153:
2154: If the loop is being completely unrolled, then splittable_regs will hold
2155: the current value of the induction variable while the loop is unrolled.
2156: It must be set to the initial value of the induction variable here.
2157: Otherwise, splittable_regs will hold the difference between the current
2158: value of the induction variable and the value the induction variable had
2159: at the top of the loop. It must be set to the value 0 here. */
2160:
2161: /* ?? If the loop is only unrolled twice, then most of the restrictions to
2162: constant values are unnecessary, since we can easily calculate increment
2163: values in this case even if nothing is constant. The increment value
2164: should not involve a multiply however. */
2165:
2166: /* ?? Even if the biv/giv increment values aren't constant, it may still
2167: be beneficial to split the variable if the loop is only unrolled a few
2168: times, since multiplies by small integers (1,2,3,4) are very cheap. */
2169:
2170: static int
2171: find_splittable_regs (unroll_type, loop_start, loop_end, end_insert_before,
2172: unroll_number)
2173: enum unroll_types unroll_type;
2174: rtx loop_start, loop_end;
2175: rtx end_insert_before;
2176: int unroll_number;
2177: {
2178: struct iv_class *bl;
1.1.1.4 ! root 2179: struct induction *v;
1.1 root 2180: rtx increment, tem;
2181: rtx biv_final_value;
2182: int biv_splittable;
2183: int result = 0;
2184:
2185: for (bl = loop_iv_list; bl; bl = bl->next)
2186: {
2187: /* Biv_total_increment must return a constant value,
2188: otherwise we can not calculate the split values. */
2189:
2190: increment = biv_total_increment (bl, loop_start, loop_end);
2191: if (! increment || GET_CODE (increment) != CONST_INT)
2192: continue;
2193:
2194: /* The loop must be unrolled completely, or else have a known number
2195: of iterations and only one exit, or else the biv must be dead
2196: outside the loop, or else the final value must be known. Otherwise,
2197: it is unsafe to split the biv since it may not have the proper
2198: value on loop exit. */
2199:
2200: /* loop_number_exit_labels is non-zero if the loop has an exit other than
2201: a fall through at the end. */
2202:
2203: biv_splittable = 1;
2204: biv_final_value = 0;
2205: if (unroll_type != UNROLL_COMPLETELY
2206: && (loop_number_exit_labels[uid_loop_num[INSN_UID (loop_start)]]
2207: || unroll_type == UNROLL_NAIVE)
2208: && (uid_luid[regno_last_uid[bl->regno]] >= INSN_LUID (loop_end)
2209: || ! bl->init_insn
2210: || INSN_UID (bl->init_insn) >= max_uid_for_loop
2211: || (uid_luid[regno_first_uid[bl->regno]]
2212: < INSN_LUID (bl->init_insn))
2213: || reg_mentioned_p (bl->biv->dest_reg, SET_SRC (bl->init_set)))
2214: && ! (biv_final_value = final_biv_value (bl, loop_start, loop_end)))
2215: biv_splittable = 0;
2216:
1.1.1.4 ! root 2217: /* If any of the insns setting the BIV don't do so with a simple
! 2218: PLUS, we don't know how to split it. */
! 2219: for (v = bl->biv; biv_splittable && v; v = v->next_iv)
! 2220: if ((tem = single_set (v->insn)) == 0
! 2221: || GET_CODE (SET_DEST (tem)) != REG
! 2222: || REGNO (SET_DEST (tem)) != bl->regno
! 2223: || GET_CODE (SET_SRC (tem)) != PLUS)
! 2224: biv_splittable = 0;
! 2225:
1.1 root 2226: /* If final value is non-zero, then must emit an instruction which sets
2227: the value of the biv to the proper value. This is done after
2228: handling all of the givs, since some of them may need to use the
2229: biv's value in their initialization code. */
2230:
2231: /* This biv is splittable. If completely unrolling the loop, save
2232: the biv's initial value. Otherwise, save the constant zero. */
2233:
2234: if (biv_splittable == 1)
2235: {
2236: if (unroll_type == UNROLL_COMPLETELY)
2237: {
2238: /* If the initial value of the biv is itself (i.e. it is too
2239: complicated for strength_reduce to compute), or is a hard
1.1.1.2 root 2240: register, then we must create a new pseudo reg to hold the
1.1 root 2241: initial value of the biv. */
2242:
2243: if (GET_CODE (bl->initial_value) == REG
2244: && (REGNO (bl->initial_value) == bl->regno
2245: || REGNO (bl->initial_value) < FIRST_PSEUDO_REGISTER))
2246: {
2247: rtx tem = gen_reg_rtx (bl->biv->mode);
2248:
2249: emit_insn_before (gen_move_insn (tem, bl->biv->src_reg),
2250: loop_start);
2251:
2252: if (loop_dump_stream)
2253: fprintf (loop_dump_stream, "Biv %d initial value remapped to %d.\n",
2254: bl->regno, REGNO (tem));
2255:
2256: splittable_regs[bl->regno] = tem;
2257: }
2258: else
2259: splittable_regs[bl->regno] = bl->initial_value;
2260: }
2261: else
2262: splittable_regs[bl->regno] = const0_rtx;
2263:
2264: /* Save the number of instructions that modify the biv, so that
2265: we can treat the last one specially. */
2266:
2267: splittable_regs_updates[bl->regno] = bl->biv_count;
2268:
2269: result++;
2270:
2271: if (loop_dump_stream)
2272: fprintf (loop_dump_stream,
2273: "Biv %d safe to split.\n", bl->regno);
2274: }
2275:
2276: /* Check every giv that depends on this biv to see whether it is
2277: splittable also. Even if the biv isn't splittable, givs which
2278: depend on it may be splittable if the biv is live outside the
2279: loop, and the givs aren't. */
2280:
2281: result = find_splittable_givs (bl, unroll_type, loop_start, loop_end,
2282: increment, unroll_number, result);
2283:
2284: /* If final value is non-zero, then must emit an instruction which sets
2285: the value of the biv to the proper value. This is done after
2286: handling all of the givs, since some of them may need to use the
2287: biv's value in their initialization code. */
2288: if (biv_final_value)
2289: {
2290: /* If the loop has multiple exits, emit the insns before the
2291: loop to ensure that it will always be executed no matter
2292: how the loop exits. Otherwise emit the insn after the loop,
2293: since this is slightly more efficient. */
2294: if (! loop_number_exit_labels[uid_loop_num[INSN_UID (loop_start)]])
2295: emit_insn_before (gen_move_insn (bl->biv->src_reg,
2296: biv_final_value),
2297: end_insert_before);
2298: else
2299: {
2300: /* Create a new register to hold the value of the biv, and then
2301: set the biv to its final value before the loop start. The biv
2302: is set to its final value before loop start to ensure that
2303: this insn will always be executed, no matter how the loop
2304: exits. */
2305: rtx tem = gen_reg_rtx (bl->biv->mode);
2306: emit_insn_before (gen_move_insn (tem, bl->biv->src_reg),
2307: loop_start);
2308: emit_insn_before (gen_move_insn (bl->biv->src_reg,
2309: biv_final_value),
2310: loop_start);
2311:
2312: if (loop_dump_stream)
2313: fprintf (loop_dump_stream, "Biv %d mapped to %d for split.\n",
2314: REGNO (bl->biv->src_reg), REGNO (tem));
2315:
2316: /* Set up the mapping from the original biv register to the new
2317: register. */
2318: bl->biv->src_reg = tem;
2319: }
2320: }
2321: }
2322: return result;
2323: }
2324:
2325: /* For every giv based on the biv BL, check to determine whether it is
2326: splittable. This is a subroutine to find_splittable_regs (). */
2327:
2328: static int
2329: find_splittable_givs (bl, unroll_type, loop_start, loop_end, increment,
2330: unroll_number, result)
2331: struct iv_class *bl;
2332: enum unroll_types unroll_type;
2333: rtx loop_start, loop_end;
2334: rtx increment;
2335: int unroll_number, result;
2336: {
2337: struct induction *v;
2338: rtx final_value;
2339: rtx tem;
2340:
2341: for (v = bl->giv; v; v = v->next_iv)
2342: {
2343: rtx giv_inc, value;
2344:
2345: /* Only split the giv if it has already been reduced, or if the loop is
2346: being completely unrolled. */
2347: if (unroll_type != UNROLL_COMPLETELY && v->ignore)
2348: continue;
2349:
2350: /* The giv can be split if the insn that sets the giv is executed once
2351: and only once on every iteration of the loop. */
2352: /* An address giv can always be split. v->insn is just a use not a set,
2353: and hence it does not matter whether it is always executed. All that
2354: matters is that all the biv increments are always executed, and we
2355: won't reach here if they aren't. */
2356: if (v->giv_type != DEST_ADDR
2357: && (! v->always_computable
2358: || back_branch_in_range_p (v->insn, loop_start, loop_end)))
2359: continue;
2360:
2361: /* The giv increment value must be a constant. */
2362: giv_inc = fold_rtx_mult_add (v->mult_val, increment, const0_rtx,
2363: v->mode);
2364: if (! giv_inc || GET_CODE (giv_inc) != CONST_INT)
2365: continue;
2366:
2367: /* The loop must be unrolled completely, or else have a known number of
2368: iterations and only one exit, or else the giv must be dead outside
2369: the loop, or else the final value of the giv must be known.
2370: Otherwise, it is not safe to split the giv since it may not have the
2371: proper value on loop exit. */
2372:
2373: /* The used outside loop test will fail for DEST_ADDR givs. They are
2374: never used outside the loop anyways, so it is always safe to split a
2375: DEST_ADDR giv. */
2376:
2377: final_value = 0;
2378: if (unroll_type != UNROLL_COMPLETELY
2379: && (loop_number_exit_labels[uid_loop_num[INSN_UID (loop_start)]]
2380: || unroll_type == UNROLL_NAIVE)
2381: && v->giv_type != DEST_ADDR
2382: && ((regno_first_uid[REGNO (v->dest_reg)] != INSN_UID (v->insn)
2383: /* Check for the case where the pseudo is set by a shift/add
2384: sequence, in which case the first insn setting the pseudo
2385: is the first insn of the shift/add sequence. */
1.1.1.4 ! root 2386: && (! (tem = find_reg_note (v->insn, REG_RETVAL, NULL_RTX))
1.1 root 2387: || (regno_first_uid[REGNO (v->dest_reg)]
2388: != INSN_UID (XEXP (tem, 0)))))
2389: /* Line above always fails if INSN was moved by loop opt. */
2390: || (uid_luid[regno_last_uid[REGNO (v->dest_reg)]]
2391: >= INSN_LUID (loop_end)))
2392: && ! (final_value = v->final_value))
2393: continue;
2394:
2395: #if 0
2396: /* Currently, non-reduced/final-value givs are never split. */
2397: /* Should emit insns after the loop if possible, as the biv final value
2398: code below does. */
2399:
2400: /* If the final value is non-zero, and the giv has not been reduced,
2401: then must emit an instruction to set the final value. */
2402: if (final_value && !v->new_reg)
2403: {
2404: /* Create a new register to hold the value of the giv, and then set
2405: the giv to its final value before the loop start. The giv is set
2406: to its final value before loop start to ensure that this insn
2407: will always be executed, no matter how we exit. */
2408: tem = gen_reg_rtx (v->mode);
2409: emit_insn_before (gen_move_insn (tem, v->dest_reg), loop_start);
2410: emit_insn_before (gen_move_insn (v->dest_reg, final_value),
2411: loop_start);
2412:
2413: if (loop_dump_stream)
2414: fprintf (loop_dump_stream, "Giv %d mapped to %d for split.\n",
2415: REGNO (v->dest_reg), REGNO (tem));
2416:
2417: v->src_reg = tem;
2418: }
2419: #endif
2420:
2421: /* This giv is splittable. If completely unrolling the loop, save the
2422: giv's initial value. Otherwise, save the constant zero for it. */
2423:
2424: if (unroll_type == UNROLL_COMPLETELY)
1.1.1.4 ! root 2425: {
! 2426: /* It is not safe to use bl->initial_value here, because it may not
! 2427: be invariant. It is safe to use the initial value stored in
! 2428: the splittable_regs array if it is set. In rare cases, it won't
! 2429: be set, so then we do exactly the same thing as
! 2430: find_splittable_regs does to get a safe value. */
! 2431: rtx biv_initial_value;
! 2432:
! 2433: if (splittable_regs[bl->regno])
! 2434: biv_initial_value = splittable_regs[bl->regno];
! 2435: else if (GET_CODE (bl->initial_value) != REG
! 2436: || (REGNO (bl->initial_value) != bl->regno
! 2437: && REGNO (bl->initial_value) >= FIRST_PSEUDO_REGISTER))
! 2438: biv_initial_value = bl->initial_value;
! 2439: else
! 2440: {
! 2441: rtx tem = gen_reg_rtx (bl->biv->mode);
! 2442:
! 2443: emit_insn_before (gen_move_insn (tem, bl->biv->src_reg),
! 2444: loop_start);
! 2445: biv_initial_value = tem;
! 2446: }
! 2447: value = fold_rtx_mult_add (v->mult_val, biv_initial_value,
! 2448: v->add_val, v->mode);
! 2449: }
1.1 root 2450: else
2451: value = const0_rtx;
2452:
2453: if (v->new_reg)
2454: {
1.1.1.3 root 2455: /* If a giv was combined with another giv, then we can only split
2456: this giv if the giv it was combined with was reduced. This
2457: is because the value of v->new_reg is meaningless in this
2458: case. */
2459: if (v->same && ! v->same->new_reg)
1.1 root 2460: {
2461: if (loop_dump_stream)
2462: fprintf (loop_dump_stream,
1.1.1.3 root 2463: "giv combined with unreduced giv not split.\n");
2464: continue;
2465: }
2466: /* If the giv is an address destination, it could be something other
2467: than a simple register, these have to be treated differently. */
2468: else if (v->giv_type == DEST_REG)
2469: {
2470: /* If value is not a constant, register, or register plus
2471: constant, then compute its value into a register before
2472: loop start. This prevents illegal rtx sharing, and should
2473: generate better code. We can use bl->initial_value here
2474: instead of splittable_regs[bl->regno] because this code
2475: is going before the loop start. */
2476: if (unroll_type == UNROLL_COMPLETELY
2477: && GET_CODE (value) != CONST_INT
2478: && GET_CODE (value) != REG
2479: && (GET_CODE (value) != PLUS
2480: || GET_CODE (XEXP (value, 0)) != REG
2481: || GET_CODE (XEXP (value, 1)) != CONST_INT))
2482: {
2483: rtx tem = gen_reg_rtx (v->mode);
2484: emit_iv_add_mult (bl->initial_value, v->mult_val,
2485: v->add_val, tem, loop_start);
2486: value = tem;
2487: }
2488:
2489: splittable_regs[REGNO (v->new_reg)] = value;
1.1 root 2490: }
2491: else
2492: {
2493: /* Splitting address givs is useful since it will often allow us
2494: to eliminate some increment insns for the base giv as
2495: unnecessary. */
2496:
2497: /* If the addr giv is combined with a dest_reg giv, then all
2498: references to that dest reg will be remapped, which is NOT
2499: what we want for split addr regs. We always create a new
2500: register for the split addr giv, just to be safe. */
2501:
2502: /* ??? If there are multiple address givs which have been
2503: combined with the same dest_reg giv, then we may only need
2504: one new register for them. Pulling out constants below will
2505: catch some of the common cases of this. Currently, I leave
2506: the work of simplifying multiple address givs to the
2507: following cse pass. */
2508:
2509: v->const_adjust = 0;
2510: if (unroll_type != UNROLL_COMPLETELY)
2511: {
2512: /* If not completely unrolling the loop, then create a new
2513: register to hold the split value of the DEST_ADDR giv.
2514: Emit insn to initialize its value before loop start. */
2515: tem = gen_reg_rtx (v->mode);
2516:
2517: /* If the address giv has a constant in its new_reg value,
2518: then this constant can be pulled out and put in value,
2519: instead of being part of the initialization code. */
2520:
2521: if (GET_CODE (v->new_reg) == PLUS
2522: && GET_CODE (XEXP (v->new_reg, 1)) == CONST_INT)
2523: {
2524: v->dest_reg
2525: = plus_constant (tem, INTVAL (XEXP (v->new_reg,1)));
2526:
2527: /* Only succeed if this will give valid addresses.
2528: Try to validate both the first and the last
2529: address resulting from loop unrolling, if
2530: one fails, then can't do const elim here. */
1.1.1.4 ! root 2531: if (memory_address_p (v->mem_mode, v->dest_reg)
! 2532: && memory_address_p (v->mem_mode,
1.1 root 2533: plus_constant (v->dest_reg,
2534: INTVAL (giv_inc)
2535: * (unroll_number - 1))))
2536: {
2537: /* Save the negative of the eliminated const, so
2538: that we can calculate the dest_reg's increment
2539: value later. */
2540: v->const_adjust = - INTVAL (XEXP (v->new_reg, 1));
2541:
2542: v->new_reg = XEXP (v->new_reg, 0);
2543: if (loop_dump_stream)
2544: fprintf (loop_dump_stream,
2545: "Eliminating constant from giv %d\n",
2546: REGNO (tem));
2547: }
2548: else
2549: v->dest_reg = tem;
2550: }
2551: else
2552: v->dest_reg = tem;
2553:
2554: /* If the address hasn't been checked for validity yet, do so
2555: now, and fail completely if either the first or the last
2556: unrolled copy of the address is not a valid address. */
2557: if (v->dest_reg == tem
1.1.1.4 ! root 2558: && (! memory_address_p (v->mem_mode, v->dest_reg)
! 2559: || ! memory_address_p (v->mem_mode,
1.1 root 2560: plus_constant (v->dest_reg,
2561: INTVAL (giv_inc)
2562: * (unroll_number -1)))))
2563: {
2564: if (loop_dump_stream)
2565: fprintf (loop_dump_stream,
2566: "Illegal address for giv at insn %d\n",
2567: INSN_UID (v->insn));
2568: continue;
2569: }
2570:
2571: /* To initialize the new register, just move the value of
2572: new_reg into it. This is not guaranteed to give a valid
2573: instruction on machines with complex addressing modes.
2574: If we can't recognize it, then delete it and emit insns
2575: to calculate the value from scratch. */
2576: emit_insn_before (gen_rtx (SET, VOIDmode, tem,
2577: copy_rtx (v->new_reg)),
2578: loop_start);
1.1.1.4 ! root 2579: if (recog_memoized (PREV_INSN (loop_start)) < 0)
1.1 root 2580: {
2581: delete_insn (PREV_INSN (loop_start));
2582: emit_iv_add_mult (bl->initial_value, v->mult_val,
2583: v->add_val, tem, loop_start);
2584: if (loop_dump_stream)
2585: fprintf (loop_dump_stream,
2586: "Illegal init insn, rewritten.\n");
2587: }
2588: }
2589: else
2590: {
2591: v->dest_reg = value;
2592:
2593: /* Check the resulting address for validity, and fail
2594: if the resulting address would be illegal. */
1.1.1.4 ! root 2595: if (! memory_address_p (v->mem_mode, v->dest_reg)
! 2596: || ! memory_address_p (v->mem_mode,
1.1 root 2597: plus_constant (v->dest_reg,
2598: INTVAL (giv_inc) *
2599: (unroll_number -1))))
2600: {
2601: if (loop_dump_stream)
2602: fprintf (loop_dump_stream,
2603: "Illegal address for giv at insn %d\n",
2604: INSN_UID (v->insn));
2605: continue;
2606: }
2607: }
2608:
2609: /* Store the value of dest_reg into the insn. This sharing
2610: will not be a problem as this insn will always be copied
2611: later. */
2612:
2613: *v->location = v->dest_reg;
2614:
2615: /* If this address giv is combined with a dest reg giv, then
2616: save the base giv's induction pointer so that we will be
2617: able to handle this address giv properly. The base giv
2618: itself does not have to be splittable. */
2619:
2620: if (v->same && v->same->giv_type == DEST_REG)
2621: addr_combined_regs[REGNO (v->same->new_reg)] = v->same;
2622:
2623: if (GET_CODE (v->new_reg) == REG)
2624: {
2625: /* This giv maybe hasn't been combined with any others.
2626: Make sure that it's giv is marked as splittable here. */
2627:
2628: splittable_regs[REGNO (v->new_reg)] = value;
2629:
2630: /* Make it appear to depend upon itself, so that the
2631: giv will be properly split in the main loop above. */
2632: if (! v->same)
2633: {
2634: v->same = v;
2635: addr_combined_regs[REGNO (v->new_reg)] = v;
2636: }
2637: }
1.1.1.2 root 2638:
1.1 root 2639: if (loop_dump_stream)
2640: fprintf (loop_dump_stream, "DEST_ADDR giv being split.\n");
2641: }
2642: }
2643: else
2644: {
2645: #if 0
2646: /* Currently, unreduced giv's can't be split. This is not too much
2647: of a problem since unreduced giv's are not live across loop
2648: iterations anyways. When unrolling a loop completely though,
2649: it makes sense to reduce&split givs when possible, as this will
2650: result in simpler instructions, and will not require that a reg
2651: be live across loop iterations. */
2652:
2653: splittable_regs[REGNO (v->dest_reg)] = value;
2654: fprintf (stderr, "Giv %d at insn %d not reduced\n",
2655: REGNO (v->dest_reg), INSN_UID (v->insn));
2656: #else
2657: continue;
2658: #endif
2659: }
2660:
2661: /* Givs are only updated once by definition. Mark it so if this is
2662: a splittable register. Don't need to do anything for address givs
2663: where this may not be a register. */
2664:
2665: if (GET_CODE (v->new_reg) == REG)
2666: splittable_regs_updates[REGNO (v->new_reg)] = 1;
2667:
2668: result++;
2669:
2670: if (loop_dump_stream)
2671: {
2672: int regnum;
2673:
2674: if (GET_CODE (v->dest_reg) == CONST_INT)
2675: regnum = -1;
2676: else if (GET_CODE (v->dest_reg) != REG)
2677: regnum = REGNO (XEXP (v->dest_reg, 0));
2678: else
2679: regnum = REGNO (v->dest_reg);
2680: fprintf (loop_dump_stream, "Giv %d at insn %d safe to split.\n",
2681: regnum, INSN_UID (v->insn));
2682: }
2683: }
2684:
2685: return result;
2686: }
2687:
2688: /* Try to prove that the register is dead after the loop exits. Trace every
2689: loop exit looking for an insn that will always be executed, which sets
2690: the register to some value, and appears before the first use of the register
2691: is found. If successful, then return 1, otherwise return 0. */
2692:
2693: /* ?? Could be made more intelligent in the handling of jumps, so that
2694: it can search past if statements and other similar structures. */
2695:
2696: static int
2697: reg_dead_after_loop (reg, loop_start, loop_end)
2698: rtx reg, loop_start, loop_end;
2699: {
2700: rtx insn, label;
2701: enum rtx_code code;
1.1.1.2 root 2702: int jump_count = 0;
1.1 root 2703:
2704: /* HACK: Must also search the loop fall through exit, create a label_ref
2705: here which points to the loop_end, and append the loop_number_exit_labels
2706: list to it. */
2707: label = gen_rtx (LABEL_REF, VOIDmode, loop_end);
2708: LABEL_NEXTREF (label)
2709: = loop_number_exit_labels[uid_loop_num[INSN_UID (loop_start)]];
2710:
2711: for ( ; label; label = LABEL_NEXTREF (label))
2712: {
2713: /* Succeed if find an insn which sets the biv or if reach end of
2714: function. Fail if find an insn that uses the biv, or if come to
2715: a conditional jump. */
2716:
2717: insn = NEXT_INSN (XEXP (label, 0));
1.1.1.2 root 2718: while (insn)
1.1 root 2719: {
1.1.1.2 root 2720: code = GET_CODE (insn);
2721: if (GET_RTX_CLASS (code) == 'i')
1.1 root 2722: {
1.1.1.2 root 2723: rtx set;
2724:
2725: if (reg_referenced_p (reg, PATTERN (insn)))
1.1 root 2726: return 0;
1.1.1.2 root 2727:
2728: set = single_set (insn);
2729: if (set && rtx_equal_p (SET_DEST (set), reg))
2730: break;
1.1 root 2731: }
1.1.1.2 root 2732:
1.1 root 2733: if (code == JUMP_INSN)
2734: {
2735: if (GET_CODE (PATTERN (insn)) == RETURN)
2736: break;
1.1.1.2 root 2737: else if (! simplejump_p (insn)
2738: /* Prevent infinite loop following infinite loops. */
2739: || jump_count++ > 20)
1.1 root 2740: return 0;
2741: else
1.1.1.2 root 2742: insn = JUMP_LABEL (insn);
1.1 root 2743: }
1.1.1.2 root 2744:
1.1 root 2745: insn = NEXT_INSN (insn);
2746: }
2747: }
2748:
2749: /* Success, the register is dead on all loop exits. */
2750: return 1;
2751: }
2752:
2753: /* Try to calculate the final value of the biv, the value it will have at
2754: the end of the loop. If we can do it, return that value. */
2755:
2756: rtx
2757: final_biv_value (bl, loop_start, loop_end)
2758: struct iv_class *bl;
2759: rtx loop_start, loop_end;
2760: {
2761: rtx increment, tem;
2762:
1.1.1.2 root 2763: /* ??? This only works for MODE_INT biv's. Reject all others for now. */
2764:
2765: if (GET_MODE_CLASS (bl->biv->mode) != MODE_INT)
2766: return 0;
2767:
1.1 root 2768: /* The final value for reversed bivs must be calculated differently than
2769: for ordinary bivs. In this case, there is already an insn after the
2770: loop which sets this biv's final value (if necessary), and there are
2771: no other loop exits, so we can return any value. */
2772: if (bl->reversed)
2773: {
2774: if (loop_dump_stream)
2775: fprintf (loop_dump_stream,
2776: "Final biv value for %d, reversed biv.\n", bl->regno);
2777:
2778: return const0_rtx;
2779: }
2780:
2781: /* Try to calculate the final value as initial value + (number of iterations
2782: * increment). For this to work, increment must be invariant, the only
2783: exit from the loop must be the fall through at the bottom (otherwise
2784: it may not have its final value when the loop exits), and the initial
2785: value of the biv must be invariant. */
2786:
2787: if (loop_n_iterations != 0
2788: && ! loop_number_exit_labels[uid_loop_num[INSN_UID (loop_start)]]
2789: && invariant_p (bl->initial_value))
2790: {
2791: increment = biv_total_increment (bl, loop_start, loop_end);
2792:
2793: if (increment && invariant_p (increment))
2794: {
2795: /* Can calculate the loop exit value, emit insns after loop
2796: end to calculate this value into a temporary register in
2797: case it is needed later. */
2798:
2799: tem = gen_reg_rtx (bl->biv->mode);
1.1.1.4 ! root 2800: /* Make sure loop_end is not the last insn. */
! 2801: if (NEXT_INSN (loop_end) == 0)
! 2802: emit_note_after (NOTE_INSN_DELETED, loop_end);
! 2803: emit_iv_add_mult (increment, GEN_INT (loop_n_iterations),
1.1 root 2804: bl->initial_value, tem, NEXT_INSN (loop_end));
2805:
2806: if (loop_dump_stream)
2807: fprintf (loop_dump_stream,
2808: "Final biv value for %d, calculated.\n", bl->regno);
2809:
2810: return tem;
2811: }
2812: }
2813:
2814: /* Check to see if the biv is dead at all loop exits. */
2815: if (reg_dead_after_loop (bl->biv->src_reg, loop_start, loop_end))
2816: {
2817: if (loop_dump_stream)
2818: fprintf (loop_dump_stream,
2819: "Final biv value for %d, biv dead after loop exit.\n",
2820: bl->regno);
2821:
2822: return const0_rtx;
2823: }
2824:
2825: return 0;
2826: }
2827:
2828: /* Try to calculate the final value of the giv, the value it will have at
2829: the end of the loop. If we can do it, return that value. */
2830:
2831: rtx
2832: final_giv_value (v, loop_start, loop_end)
2833: struct induction *v;
2834: rtx loop_start, loop_end;
2835: {
2836: struct iv_class *bl;
1.1.1.4 ! root 2837: rtx insn;
1.1 root 2838: rtx increment, tem;
2839: enum rtx_code code;
1.1.1.3 root 2840: rtx insert_before, seq;
1.1 root 2841:
2842: bl = reg_biv_class[REGNO (v->src_reg)];
2843:
2844: /* The final value for givs which depend on reversed bivs must be calculated
2845: differently than for ordinary givs. In this case, there is already an
2846: insn after the loop which sets this giv's final value (if necessary),
2847: and there are no other loop exits, so we can return any value. */
2848: if (bl->reversed)
2849: {
2850: if (loop_dump_stream)
2851: fprintf (loop_dump_stream,
2852: "Final giv value for %d, depends on reversed biv\n",
2853: REGNO (v->dest_reg));
2854: return const0_rtx;
2855: }
2856:
2857: /* Try to calculate the final value as a function of the biv it depends
2858: upon. The only exit from the loop must be the fall through at the bottom
2859: (otherwise it may not have its final value when the loop exits). */
2860:
2861: /* ??? Can calculate the final giv value by subtracting off the
2862: extra biv increments times the giv's mult_val. The loop must have
2863: only one exit for this to work, but the loop iterations does not need
2864: to be known. */
2865:
2866: if (loop_n_iterations != 0
2867: && ! loop_number_exit_labels[uid_loop_num[INSN_UID (loop_start)]])
2868: {
2869: /* ?? It is tempting to use the biv's value here since these insns will
2870: be put after the loop, and hence the biv will have its final value
2871: then. However, this fails if the biv is subsequently eliminated.
2872: Perhaps determine whether biv's are eliminable before trying to
2873: determine whether giv's are replaceable so that we can use the
2874: biv value here if it is not eliminable. */
2875:
2876: increment = biv_total_increment (bl, loop_start, loop_end);
2877:
2878: if (increment && invariant_p (increment))
2879: {
2880: /* Can calculate the loop exit value of its biv as
2881: (loop_n_iterations * increment) + initial_value */
2882:
2883: /* The loop exit value of the giv is then
2884: (final_biv_value - extra increments) * mult_val + add_val.
2885: The extra increments are any increments to the biv which
2886: occur in the loop after the giv's value is calculated.
2887: We must search from the insn that sets the giv to the end
2888: of the loop to calculate this value. */
2889:
2890: insert_before = NEXT_INSN (loop_end);
2891:
2892: /* Put the final biv value in tem. */
2893: tem = gen_reg_rtx (bl->biv->mode);
1.1.1.4 ! root 2894: emit_iv_add_mult (increment, GEN_INT (loop_n_iterations),
1.1 root 2895: bl->initial_value, tem, insert_before);
2896:
2897: /* Subtract off extra increments as we find them. */
2898: for (insn = NEXT_INSN (v->insn); insn != loop_end;
2899: insn = NEXT_INSN (insn))
2900: {
1.1.1.4 ! root 2901: struct induction *biv;
! 2902:
! 2903: for (biv = bl->biv; biv; biv = biv->next_iv)
! 2904: if (biv->insn == insn)
! 2905: {
! 2906: start_sequence ();
! 2907: tem = expand_binop (GET_MODE (tem), sub_optab, tem,
! 2908: biv->add_val, NULL_RTX, 0,
! 2909: OPTAB_LIB_WIDEN);
! 2910: seq = gen_sequence ();
! 2911: end_sequence ();
! 2912: emit_insn_before (seq, insert_before);
! 2913: }
1.1 root 2914: }
2915:
2916: /* Now calculate the giv's final value. */
2917: emit_iv_add_mult (tem, v->mult_val, v->add_val, tem,
2918: insert_before);
2919:
2920: if (loop_dump_stream)
2921: fprintf (loop_dump_stream,
2922: "Final giv value for %d, calc from biv's value.\n",
2923: REGNO (v->dest_reg));
2924:
2925: return tem;
2926: }
2927: }
2928:
2929: /* Replaceable giv's should never reach here. */
2930: if (v->replaceable)
2931: abort ();
2932:
2933: /* Check to see if the biv is dead at all loop exits. */
2934: if (reg_dead_after_loop (v->dest_reg, loop_start, loop_end))
2935: {
2936: if (loop_dump_stream)
2937: fprintf (loop_dump_stream,
2938: "Final giv value for %d, giv dead after loop exit.\n",
2939: REGNO (v->dest_reg));
2940:
2941: return const0_rtx;
2942: }
2943:
2944: return 0;
2945: }
2946:
2947:
2948: /* Calculate the number of loop iterations. Returns the exact number of loop
1.1.1.3 root 2949: iterations if it can be calculated, otherwise returns zero. */
1.1 root 2950:
1.1.1.4 ! root 2951: unsigned HOST_WIDE_INT
1.1 root 2952: loop_iterations (loop_start, loop_end)
2953: rtx loop_start, loop_end;
2954: {
2955: rtx comparison, comparison_value;
2956: rtx iteration_var, initial_value, increment, final_value;
2957: enum rtx_code comparison_code;
1.1.1.4 ! root 2958: HOST_WIDE_INT i;
! 2959: int increment_dir;
1.1 root 2960: int unsigned_compare, compare_dir, final_larger;
2961: unsigned long tempu;
2962: rtx last_loop_insn;
2963:
2964: /* First find the iteration variable. If the last insn is a conditional
2965: branch, and the insn before tests a register value, make that the
2966: iteration variable. */
2967:
2968: loop_initial_value = 0;
2969: loop_increment = 0;
2970: loop_final_value = 0;
2971: loop_iteration_var = 0;
2972:
2973: last_loop_insn = prev_nonnote_insn (loop_end);
2974:
2975: comparison = get_condition_for_loop (last_loop_insn);
2976: if (comparison == 0)
2977: {
2978: if (loop_dump_stream)
2979: fprintf (loop_dump_stream,
2980: "Loop unrolling: No final conditional branch found.\n");
2981: return 0;
2982: }
2983:
2984: /* ??? Get_condition may switch position of induction variable and
2985: invariant register when it canonicalizes the comparison. */
2986:
2987: comparison_code = GET_CODE (comparison);
2988: iteration_var = XEXP (comparison, 0);
2989: comparison_value = XEXP (comparison, 1);
2990:
2991: if (GET_CODE (iteration_var) != REG)
2992: {
2993: if (loop_dump_stream)
2994: fprintf (loop_dump_stream,
2995: "Loop unrolling: Comparison not against register.\n");
2996: return 0;
2997: }
2998:
2999: /* Loop iterations is always called before any new registers are created
3000: now, so this should never occur. */
3001:
3002: if (REGNO (iteration_var) >= max_reg_before_loop)
3003: abort ();
3004:
3005: iteration_info (iteration_var, &initial_value, &increment,
3006: loop_start, loop_end);
3007: if (initial_value == 0)
3008: /* iteration_info already printed a message. */
3009: return 0;
3010:
3011: if (increment == 0)
3012: {
3013: if (loop_dump_stream)
3014: fprintf (loop_dump_stream,
3015: "Loop unrolling: Increment value can't be calculated.\n");
3016: return 0;
3017: }
3018: if (GET_CODE (increment) != CONST_INT)
3019: {
3020: if (loop_dump_stream)
3021: fprintf (loop_dump_stream,
3022: "Loop unrolling: Increment value not constant.\n");
3023: return 0;
3024: }
3025: if (GET_CODE (initial_value) != CONST_INT)
3026: {
3027: if (loop_dump_stream)
3028: fprintf (loop_dump_stream,
3029: "Loop unrolling: Initial value not constant.\n");
3030: return 0;
3031: }
3032:
3033: /* If the comparison value is an invariant register, then try to find
3034: its value from the insns before the start of the loop. */
3035:
3036: if (GET_CODE (comparison_value) == REG && invariant_p (comparison_value))
3037: {
3038: rtx insn, set;
3039:
3040: for (insn = PREV_INSN (loop_start); insn ; insn = PREV_INSN (insn))
3041: {
3042: if (GET_CODE (insn) == CODE_LABEL)
3043: break;
3044:
3045: else if (GET_RTX_CLASS (GET_CODE (insn)) == 'i'
3046: && (set = single_set (insn))
3047: && (SET_DEST (set) == comparison_value))
3048: {
1.1.1.4 ! root 3049: rtx note = find_reg_note (insn, REG_EQUAL, NULL_RTX);
1.1 root 3050:
3051: if (note && GET_CODE (XEXP (note, 0)) != EXPR_LIST)
3052: comparison_value = XEXP (note, 0);
3053:
3054: break;
3055: }
3056: }
3057: }
3058:
3059: final_value = approx_final_value (comparison_code, comparison_value,
3060: &unsigned_compare, &compare_dir);
3061:
3062: /* Save the calculated values describing this loop's bounds, in case
3063: precondition_loop_p will need them later. These values can not be
3064: recalculated inside precondition_loop_p because strength reduction
3065: optimizations may obscure the loop's structure. */
3066:
3067: loop_iteration_var = iteration_var;
3068: loop_initial_value = initial_value;
3069: loop_increment = increment;
3070: loop_final_value = final_value;
3071:
3072: if (final_value == 0)
3073: {
3074: if (loop_dump_stream)
3075: fprintf (loop_dump_stream,
3076: "Loop unrolling: EQ comparison loop.\n");
3077: return 0;
3078: }
3079: else if (GET_CODE (final_value) != CONST_INT)
3080: {
3081: if (loop_dump_stream)
3082: fprintf (loop_dump_stream,
3083: "Loop unrolling: Final value not constant.\n");
3084: return 0;
3085: }
3086:
3087: /* ?? Final value and initial value do not have to be constants.
3088: Only their difference has to be constant. When the iteration variable
3089: is an array address, the final value and initial value might both
3090: be addresses with the same base but different constant offsets.
3091: Final value must be invariant for this to work.
3092:
1.1.1.3 root 3093: To do this, need some way to find the values of registers which are
1.1 root 3094: invariant. */
3095:
3096: /* Final_larger is 1 if final larger, 0 if they are equal, otherwise -1. */
3097: if (unsigned_compare)
3098: final_larger
1.1.1.4 ! root 3099: = ((unsigned HOST_WIDE_INT) INTVAL (final_value)
! 3100: > (unsigned HOST_WIDE_INT) INTVAL (initial_value))
! 3101: - ((unsigned HOST_WIDE_INT) INTVAL (final_value)
! 3102: < (unsigned HOST_WIDE_INT) INTVAL (initial_value));
1.1 root 3103: else
1.1.1.4 ! root 3104: final_larger = (INTVAL (final_value) > INTVAL (initial_value))
! 3105: - (INTVAL (final_value) < INTVAL (initial_value));
1.1 root 3106:
3107: if (INTVAL (increment) > 0)
3108: increment_dir = 1;
3109: else if (INTVAL (increment) == 0)
3110: increment_dir = 0;
3111: else
3112: increment_dir = -1;
3113:
3114: /* There are 27 different cases: compare_dir = -1, 0, 1;
3115: final_larger = -1, 0, 1; increment_dir = -1, 0, 1.
3116: There are 4 normal cases, 4 reverse cases (where the iteration variable
3117: will overflow before the loop exits), 4 infinite loop cases, and 15
3118: immediate exit (0 or 1 iteration depending on loop type) cases.
3119: Only try to optimize the normal cases. */
3120:
3121: /* (compare_dir/final_larger/increment_dir)
3122: Normal cases: (0/-1/-1), (0/1/1), (-1/-1/-1), (1/1/1)
3123: Reverse cases: (0/-1/1), (0/1/-1), (-1/-1/1), (1/1/-1)
3124: Infinite loops: (0/-1/0), (0/1/0), (-1/-1/0), (1/1/0)
3125: Immediate exit: (0/0/X), (-1/0/X), (-1/1/X), (1/0/X), (1/-1/X) */
3126:
3127: /* ?? If the meaning of reverse loops (where the iteration variable
3128: will overflow before the loop exits) is undefined, then could
3129: eliminate all of these special checks, and just always assume
3130: the loops are normal/immediate/infinite. Note that this means
3131: the sign of increment_dir does not have to be known. Also,
3132: since it does not really hurt if immediate exit loops or infinite loops
3133: are optimized, then that case could be ignored also, and hence all
3134: loops can be optimized.
3135:
3136: According to ANSI Spec, the reverse loop case result is undefined,
3137: because the action on overflow is undefined.
3138:
3139: See also the special test for NE loops below. */
3140:
3141: if (final_larger == increment_dir && final_larger != 0
3142: && (final_larger == compare_dir || compare_dir == 0))
3143: /* Normal case. */
3144: ;
3145: else
3146: {
3147: if (loop_dump_stream)
3148: fprintf (loop_dump_stream,
3149: "Loop unrolling: Not normal loop.\n");
3150: return 0;
3151: }
3152:
3153: /* Calculate the number of iterations, final_value is only an approximation,
3154: so correct for that. Note that tempu and loop_n_iterations are
3155: unsigned, because they can be as large as 2^n - 1. */
3156:
3157: i = INTVAL (increment);
3158: if (i > 0)
3159: tempu = INTVAL (final_value) - INTVAL (initial_value);
3160: else if (i < 0)
3161: {
3162: tempu = INTVAL (initial_value) - INTVAL (final_value);
3163: i = -i;
3164: }
3165: else
3166: abort ();
3167:
3168: /* For NE tests, make sure that the iteration variable won't miss the
3169: final value. If tempu mod i is not zero, then the iteration variable
3170: will overflow before the loop exits, and we can not calculate the
3171: number of iterations. */
3172: if (compare_dir == 0 && (tempu % i) != 0)
3173: return 0;
3174:
3175: return tempu / i + ((tempu % i) != 0);
3176: }
This archive runs on limited infrastructure. Preserving old code on modern bandwidth. Automated agents are requested to crawl responsibly.