]>
Commit | Line | Data |
---|---|---|
67f2de41 | 1 | /* Try to unroll loops, and split induction variables. |
e9a25f70 | 2 | Copyright (C) 1992, 1993, 1994, 1995, 1997 Free Software Foundation, Inc. |
67f2de41 RK |
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 | |
e9fa0c7c RK |
19 | the Free Software Foundation, 59 Temple Place - Suite 330, |
20 | Boston, MA 02111-1307, USA. */ | |
67f2de41 RK |
21 | |
22 | /* Try to unroll a loop, and split induction variables. | |
23 | ||
24 | Loops for which the number of iterations can be calculated exactly are | |
25 | handled specially. If the number of iterations times the insn_count is | |
26 | less than MAX_UNROLLED_INSNS, then the loop is unrolled completely. | |
27 | Otherwise, we try to unroll the loop a number of times modulo the number | |
28 | of iterations, so that only one exit test will be needed. It is unrolled | |
29 | a number of times approximately equal to MAX_UNROLLED_INSNS divided by | |
30 | the insn count. | |
31 | ||
32 | Otherwise, if the number of iterations can be calculated exactly at | |
33 | run time, and the loop is always entered at the top, then we try to | |
34 | precondition the loop. That is, at run time, calculate how many times | |
35 | the loop will execute, and then execute the loop body a few times so | |
36 | that the remaining iterations will be some multiple of 4 (or 2 if the | |
37 | loop is large). Then fall through to a loop unrolled 4 (or 2) times, | |
38 | with only one exit test needed at the end of the loop. | |
39 | ||
40 | Otherwise, if the number of iterations can not be calculated exactly, | |
41 | not even at run time, then we still unroll the loop a number of times | |
42 | approximately equal to MAX_UNROLLED_INSNS divided by the insn count, | |
43 | but there must be an exit test after each copy of the loop body. | |
44 | ||
45 | For each induction variable, which is dead outside the loop (replaceable) | |
46 | or for which we can easily calculate the final value, if we can easily | |
47 | calculate its value at each place where it is set as a function of the | |
48 | current loop unroll count and the variable's value at loop entry, then | |
49 | the induction variable is split into `N' different variables, one for | |
50 | each copy of the loop body. One variable is live across the backward | |
51 | branch, and the others are all calculated as a function of this variable. | |
52 | This helps eliminate data dependencies, and leads to further opportunities | |
53 | for cse. */ | |
54 | ||
55 | /* Possible improvements follow: */ | |
56 | ||
57 | /* ??? Add an extra pass somewhere to determine whether unrolling will | |
58 | give any benefit. E.g. after generating all unrolled insns, compute the | |
59 | cost of all insns and compare against cost of insns in rolled loop. | |
60 | ||
61 | - On traditional architectures, unrolling a non-constant bound loop | |
62 | is a win if there is a giv whose only use is in memory addresses, the | |
6dc42e49 | 63 | memory addresses can be split, and hence giv increments can be |
67f2de41 RK |
64 | eliminated. |
65 | - It is also a win if the loop is executed many times, and preconditioning | |
66 | can be performed for the loop. | |
67 | Add code to check for these and similar cases. */ | |
68 | ||
69 | /* ??? Improve control of which loops get unrolled. Could use profiling | |
70 | info to only unroll the most commonly executed loops. Perhaps have | |
71 | a user specifyable option to control the amount of code expansion, | |
72 | or the percent of loops to consider for unrolling. Etc. */ | |
73 | ||
74 | /* ??? Look at the register copies inside the loop to see if they form a | |
75 | simple permutation. If so, iterate the permutation until it gets back to | |
76 | the start state. This is how many times we should unroll the loop, for | |
77 | best results, because then all register copies can be eliminated. | |
78 | For example, the lisp nreverse function should be unrolled 3 times | |
79 | while (this) | |
80 | { | |
81 | next = this->cdr; | |
82 | this->cdr = prev; | |
83 | prev = this; | |
84 | this = next; | |
85 | } | |
86 | ||
87 | ??? The number of times to unroll the loop may also be based on data | |
88 | references in the loop. For example, if we have a loop that references | |
89 | x[i-1], x[i], and x[i+1], we should unroll it a multiple of 3 times. */ | |
90 | ||
91 | /* ??? Add some simple linear equation solving capability so that we can | |
92 | determine the number of loop iterations for more complex loops. | |
93 | For example, consider this loop from gdb | |
94 | #define SWAP_TARGET_AND_HOST(buffer,len) | |
95 | { | |
96 | char tmp; | |
97 | char *p = (char *) buffer; | |
98 | char *q = ((char *) buffer) + len - 1; | |
99 | int iterations = (len + 1) >> 1; | |
100 | int i; | |
101 | for (p; p < q; p++, q--;) | |
102 | { | |
103 | tmp = *q; | |
104 | *q = *p; | |
105 | *p = tmp; | |
106 | } | |
107 | } | |
108 | Note that: | |
109 | start value = p = &buffer + current_iteration | |
110 | end value = q = &buffer + len - 1 - current_iteration | |
111 | Given the loop exit test of "p < q", then there must be "q - p" iterations, | |
112 | set equal to zero and solve for number of iterations: | |
113 | q - p = len - 1 - 2*current_iteration = 0 | |
114 | current_iteration = (len - 1) / 2 | |
115 | Hence, there are (len - 1) / 2 (rounded up to the nearest integer) | |
116 | iterations of this loop. */ | |
117 | ||
118 | /* ??? Currently, no labels are marked as loop invariant when doing loop | |
119 | unrolling. This is because an insn inside the loop, that loads the address | |
120 | of a label inside the loop into a register, could be moved outside the loop | |
121 | by the invariant code motion pass if labels were invariant. If the loop | |
122 | is subsequently unrolled, the code will be wrong because each unrolled | |
123 | body of the loop will use the same address, whereas each actually needs a | |
124 | different address. A case where this happens is when a loop containing | |
125 | a switch statement is unrolled. | |
126 | ||
127 | It would be better to let labels be considered invariant. When we | |
128 | unroll loops here, check to see if any insns using a label local to the | |
129 | loop were moved before the loop. If so, then correct the problem, by | |
130 | moving the insn back into the loop, or perhaps replicate the insn before | |
131 | the loop, one copy for each time the loop is unrolled. */ | |
132 | ||
133 | /* The prime factors looked for when trying to unroll a loop by some | |
134 | number which is modulo the total number of iterations. Just checking | |
135 | for these 4 prime factors will find at least one factor for 75% of | |
136 | all numbers theoretically. Practically speaking, this will succeed | |
137 | almost all of the time since loops are generally a multiple of 2 | |
138 | and/or 5. */ | |
139 | ||
140 | #define NUM_FACTORS 4 | |
141 | ||
142 | struct _factor { int factor, count; } factors[NUM_FACTORS] | |
143 | = { {2, 0}, {3, 0}, {5, 0}, {7, 0}}; | |
144 | ||
145 | /* Describes the different types of loop unrolling performed. */ | |
146 | ||
147 | enum unroll_types { UNROLL_COMPLETELY, UNROLL_MODULO, UNROLL_NAIVE }; | |
148 | ||
149 | #include "config.h" | |
e9a25f70 | 150 | #include <stdio.h> |
67f2de41 RK |
151 | #include "rtl.h" |
152 | #include "insn-config.h" | |
153 | #include "integrate.h" | |
154 | #include "regs.h" | |
e9a25f70 | 155 | #include "recog.h" |
67f2de41 RK |
156 | #include "flags.h" |
157 | #include "expr.h" | |
67f2de41 RK |
158 | #include "loop.h" |
159 | ||
160 | /* This controls which loops are unrolled, and by how much we unroll | |
161 | them. */ | |
162 | ||
163 | #ifndef MAX_UNROLLED_INSNS | |
164 | #define MAX_UNROLLED_INSNS 100 | |
165 | #endif | |
166 | ||
167 | /* Indexed by register number, if non-zero, then it contains a pointer | |
168 | to a struct induction for a DEST_REG giv which has been combined with | |
169 | one of more address givs. This is needed because whenever such a DEST_REG | |
170 | giv is modified, we must modify the value of all split address givs | |
171 | that were combined with this DEST_REG giv. */ | |
172 | ||
173 | static struct induction **addr_combined_regs; | |
174 | ||
175 | /* Indexed by register number, if this is a splittable induction variable, | |
176 | then this will hold the current value of the register, which depends on the | |
177 | iteration number. */ | |
178 | ||
179 | static rtx *splittable_regs; | |
180 | ||
181 | /* Indexed by register number, if this is a splittable induction variable, | |
182 | then this will hold the number of instructions in the loop that modify | |
183 | the induction variable. Used to ensure that only the last insn modifying | |
184 | a split iv will update the original iv of the dest. */ | |
185 | ||
186 | static int *splittable_regs_updates; | |
187 | ||
188 | /* Values describing the current loop's iteration variable. These are set up | |
189 | by loop_iterations, and used by precondition_loop_p. */ | |
190 | ||
191 | static rtx loop_iteration_var; | |
192 | static rtx loop_initial_value; | |
193 | static rtx loop_increment; | |
194 | static rtx loop_final_value; | |
1dcfa896 | 195 | static enum rtx_code loop_comparison_code; |
67f2de41 RK |
196 | |
197 | /* Forward declarations. */ | |
198 | ||
7f5b8ca7 RK |
199 | static void init_reg_map PROTO((struct inline_remap *, int)); |
200 | static int precondition_loop_p PROTO((rtx *, rtx *, rtx *, rtx, rtx)); | |
201 | static rtx calculate_giv_inc PROTO((rtx, rtx, int)); | |
202 | static rtx initial_reg_note_copy PROTO((rtx, struct inline_remap *)); | |
203 | static void final_reg_note_copy PROTO((rtx, struct inline_remap *)); | |
204 | static void copy_loop_body PROTO((rtx, rtx, struct inline_remap *, rtx, int, | |
205 | enum unroll_types, rtx, rtx, rtx, rtx)); | |
8c660648 | 206 | void iteration_info PROTO((rtx, rtx *, rtx *, rtx, rtx)); |
7f5b8ca7 RK |
207 | static rtx approx_final_value PROTO((enum rtx_code, rtx, int *, int *)); |
208 | static int find_splittable_regs PROTO((enum unroll_types, rtx, rtx, rtx, int)); | |
209 | static int find_splittable_givs PROTO((struct iv_class *,enum unroll_types, | |
210 | rtx, rtx, rtx, int)); | |
211 | static int reg_dead_after_loop PROTO((rtx, rtx, rtx)); | |
212 | static rtx fold_rtx_mult_add PROTO((rtx, rtx, rtx, enum machine_mode)); | |
213 | static rtx remap_split_bivs PROTO((rtx)); | |
67f2de41 RK |
214 | |
215 | /* Try to unroll one loop and split induction variables in the loop. | |
216 | ||
217 | The loop is described by the arguments LOOP_END, INSN_COUNT, and | |
6dc42e49 | 218 | LOOP_START. END_INSERT_BEFORE indicates where insns should be added |
67f2de41 RK |
219 | which need to be executed when the loop falls through. STRENGTH_REDUCTION_P |
220 | indicates whether information generated in the strength reduction pass | |
221 | is available. | |
222 | ||
223 | This function is intended to be called from within `strength_reduce' | |
224 | in loop.c. */ | |
225 | ||
226 | void | |
227 | unroll_loop (loop_end, insn_count, loop_start, end_insert_before, | |
228 | strength_reduce_p) | |
229 | rtx loop_end; | |
230 | int insn_count; | |
231 | rtx loop_start; | |
232 | rtx end_insert_before; | |
233 | int strength_reduce_p; | |
234 | { | |
235 | int i, j, temp; | |
236 | int unroll_number = 1; | |
237 | rtx copy_start, copy_end; | |
238 | rtx insn, copy, sequence, pattern, tem; | |
239 | int max_labelno, max_insnno; | |
240 | rtx insert_before; | |
241 | struct inline_remap *map; | |
242 | char *local_label; | |
7f5b8ca7 | 243 | char *local_regno; |
67f2de41 RK |
244 | int maxregnum; |
245 | int new_maxregnum; | |
246 | rtx exit_label = 0; | |
247 | rtx start_label; | |
248 | struct iv_class *bl; | |
67f2de41 RK |
249 | int splitting_not_safe = 0; |
250 | enum unroll_types unroll_type; | |
251 | int loop_preconditioned = 0; | |
252 | rtx safety_label; | |
253 | /* This points to the last real insn in the loop, which should be either | |
254 | a JUMP_INSN (for conditional jumps) or a BARRIER (for unconditional | |
255 | jumps). */ | |
256 | rtx last_loop_insn; | |
257 | ||
258 | /* Don't bother unrolling huge loops. Since the minimum factor is | |
259 | two, loops greater than one half of MAX_UNROLLED_INSNS will never | |
260 | be unrolled. */ | |
261 | if (insn_count > MAX_UNROLLED_INSNS / 2) | |
262 | { | |
263 | if (loop_dump_stream) | |
264 | fprintf (loop_dump_stream, "Unrolling failure: Loop too big.\n"); | |
265 | return; | |
266 | } | |
267 | ||
268 | /* When emitting debugger info, we can't unroll loops with unequal numbers | |
269 | of block_beg and block_end notes, because that would unbalance the block | |
270 | structure of the function. This can happen as a result of the | |
271 | "if (foo) bar; else break;" optimization in jump.c. */ | |
d1b9c52c JW |
272 | /* ??? Gcc has a general policy that -g is never supposed to change the code |
273 | that the compiler emits, so we must disable this optimization always, | |
274 | even if debug info is not being output. This is rare, so this should | |
275 | not be a significant performance problem. */ | |
67f2de41 | 276 | |
d1b9c52c | 277 | if (1 /* write_symbols != NO_DEBUG */) |
67f2de41 RK |
278 | { |
279 | int block_begins = 0; | |
280 | int block_ends = 0; | |
281 | ||
282 | for (insn = loop_start; insn != loop_end; insn = NEXT_INSN (insn)) | |
283 | { | |
284 | if (GET_CODE (insn) == NOTE) | |
285 | { | |
286 | if (NOTE_LINE_NUMBER (insn) == NOTE_INSN_BLOCK_BEG) | |
287 | block_begins++; | |
288 | else if (NOTE_LINE_NUMBER (insn) == NOTE_INSN_BLOCK_END) | |
289 | block_ends++; | |
290 | } | |
291 | } | |
292 | ||
293 | if (block_begins != block_ends) | |
294 | { | |
295 | if (loop_dump_stream) | |
296 | fprintf (loop_dump_stream, | |
297 | "Unrolling failure: Unbalanced block notes.\n"); | |
298 | return; | |
299 | } | |
300 | } | |
301 | ||
302 | /* Determine type of unroll to perform. Depends on the number of iterations | |
303 | and the size of the loop. */ | |
304 | ||
305 | /* If there is no strength reduce info, then set loop_n_iterations to zero. | |
306 | This can happen if strength_reduce can't find any bivs in the loop. | |
307 | A value of zero indicates that the number of iterations could not be | |
308 | calculated. */ | |
309 | ||
310 | if (! strength_reduce_p) | |
311 | loop_n_iterations = 0; | |
312 | ||
313 | if (loop_dump_stream && loop_n_iterations > 0) | |
314 | fprintf (loop_dump_stream, | |
315 | "Loop unrolling: %d iterations.\n", loop_n_iterations); | |
316 | ||
317 | /* Find and save a pointer to the last nonnote insn in the loop. */ | |
318 | ||
319 | last_loop_insn = prev_nonnote_insn (loop_end); | |
320 | ||
321 | /* Calculate how many times to unroll the loop. Indicate whether or | |
322 | not the loop is being completely unrolled. */ | |
323 | ||
324 | if (loop_n_iterations == 1) | |
325 | { | |
326 | /* If number of iterations is exactly 1, then eliminate the compare and | |
327 | branch at the end of the loop since they will never be taken. | |
328 | Then return, since no other action is needed here. */ | |
329 | ||
330 | /* If the last instruction is not a BARRIER or a JUMP_INSN, then | |
331 | don't do anything. */ | |
332 | ||
333 | if (GET_CODE (last_loop_insn) == BARRIER) | |
334 | { | |
335 | /* Delete the jump insn. This will delete the barrier also. */ | |
336 | delete_insn (PREV_INSN (last_loop_insn)); | |
337 | } | |
338 | else if (GET_CODE (last_loop_insn) == JUMP_INSN) | |
339 | { | |
340 | #ifdef HAVE_cc0 | |
b4ac57ab | 341 | /* The immediately preceding insn is a compare which must be |
67f2de41 RK |
342 | deleted. */ |
343 | delete_insn (last_loop_insn); | |
344 | delete_insn (PREV_INSN (last_loop_insn)); | |
345 | #else | |
b4ac57ab | 346 | /* The immediately preceding insn may not be the compare, so don't |
67f2de41 RK |
347 | delete it. */ |
348 | delete_insn (last_loop_insn); | |
349 | #endif | |
350 | } | |
351 | return; | |
352 | } | |
353 | else if (loop_n_iterations > 0 | |
354 | && loop_n_iterations * insn_count < MAX_UNROLLED_INSNS) | |
355 | { | |
356 | unroll_number = loop_n_iterations; | |
357 | unroll_type = UNROLL_COMPLETELY; | |
358 | } | |
359 | else if (loop_n_iterations > 0) | |
360 | { | |
361 | /* Try to factor the number of iterations. Don't bother with the | |
362 | general case, only using 2, 3, 5, and 7 will get 75% of all | |
363 | numbers theoretically, and almost all in practice. */ | |
364 | ||
365 | for (i = 0; i < NUM_FACTORS; i++) | |
366 | factors[i].count = 0; | |
367 | ||
368 | temp = loop_n_iterations; | |
369 | for (i = NUM_FACTORS - 1; i >= 0; i--) | |
370 | while (temp % factors[i].factor == 0) | |
371 | { | |
372 | factors[i].count++; | |
373 | temp = temp / factors[i].factor; | |
374 | } | |
375 | ||
376 | /* Start with the larger factors first so that we generally | |
377 | get lots of unrolling. */ | |
378 | ||
379 | unroll_number = 1; | |
380 | temp = insn_count; | |
381 | for (i = 3; i >= 0; i--) | |
382 | while (factors[i].count--) | |
383 | { | |
384 | if (temp * factors[i].factor < MAX_UNROLLED_INSNS) | |
385 | { | |
386 | unroll_number *= factors[i].factor; | |
387 | temp *= factors[i].factor; | |
388 | } | |
389 | else | |
390 | break; | |
391 | } | |
392 | ||
393 | /* If we couldn't find any factors, then unroll as in the normal | |
394 | case. */ | |
395 | if (unroll_number == 1) | |
396 | { | |
397 | if (loop_dump_stream) | |
398 | fprintf (loop_dump_stream, | |
399 | "Loop unrolling: No factors found.\n"); | |
400 | } | |
401 | else | |
402 | unroll_type = UNROLL_MODULO; | |
403 | } | |
404 | ||
405 | ||
406 | /* Default case, calculate number of times to unroll loop based on its | |
407 | size. */ | |
408 | if (unroll_number == 1) | |
409 | { | |
410 | if (8 * insn_count < MAX_UNROLLED_INSNS) | |
411 | unroll_number = 8; | |
412 | else if (4 * insn_count < MAX_UNROLLED_INSNS) | |
413 | unroll_number = 4; | |
414 | else | |
415 | unroll_number = 2; | |
416 | ||
417 | unroll_type = UNROLL_NAIVE; | |
418 | } | |
419 | ||
420 | /* Now we know how many times to unroll the loop. */ | |
421 | ||
422 | if (loop_dump_stream) | |
423 | fprintf (loop_dump_stream, | |
424 | "Unrolling loop %d times.\n", unroll_number); | |
425 | ||
426 | ||
427 | if (unroll_type == UNROLL_COMPLETELY || unroll_type == UNROLL_MODULO) | |
428 | { | |
429 | /* Loops of these types should never start with a jump down to | |
430 | the exit condition test. For now, check for this case just to | |
431 | be sure. UNROLL_NAIVE loops can be of this form, this case is | |
432 | handled below. */ | |
433 | insn = loop_start; | |
434 | while (GET_CODE (insn) != CODE_LABEL && GET_CODE (insn) != JUMP_INSN) | |
435 | insn = NEXT_INSN (insn); | |
436 | if (GET_CODE (insn) == JUMP_INSN) | |
437 | abort (); | |
438 | } | |
439 | ||
440 | if (unroll_type == UNROLL_COMPLETELY) | |
441 | { | |
442 | /* Completely unrolling the loop: Delete the compare and branch at | |
443 | the end (the last two instructions). This delete must done at the | |
444 | very end of loop unrolling, to avoid problems with calls to | |
445 | back_branch_in_range_p, which is called by find_splittable_regs. | |
446 | All increments of splittable bivs/givs are changed to load constant | |
447 | instructions. */ | |
448 | ||
449 | copy_start = loop_start; | |
450 | ||
451 | /* Set insert_before to the instruction immediately after the JUMP_INSN | |
452 | (or BARRIER), so that any NOTEs between the JUMP_INSN and the end of | |
453 | the loop will be correctly handled by copy_loop_body. */ | |
454 | insert_before = NEXT_INSN (last_loop_insn); | |
455 | ||
456 | /* Set copy_end to the insn before the jump at the end of the loop. */ | |
457 | if (GET_CODE (last_loop_insn) == BARRIER) | |
458 | copy_end = PREV_INSN (PREV_INSN (last_loop_insn)); | |
459 | else if (GET_CODE (last_loop_insn) == JUMP_INSN) | |
460 | { | |
461 | #ifdef HAVE_cc0 | |
462 | /* The instruction immediately before the JUMP_INSN is a compare | |
463 | instruction which we do not want to copy. */ | |
464 | copy_end = PREV_INSN (PREV_INSN (last_loop_insn)); | |
465 | #else | |
466 | /* The instruction immediately before the JUMP_INSN may not be the | |
467 | compare, so we must copy it. */ | |
468 | copy_end = PREV_INSN (last_loop_insn); | |
469 | #endif | |
470 | } | |
471 | else | |
472 | { | |
473 | /* We currently can't unroll a loop if it doesn't end with a | |
474 | JUMP_INSN. There would need to be a mechanism that recognizes | |
475 | this case, and then inserts a jump after each loop body, which | |
476 | jumps to after the last loop body. */ | |
477 | if (loop_dump_stream) | |
478 | fprintf (loop_dump_stream, | |
479 | "Unrolling failure: loop does not end with a JUMP_INSN.\n"); | |
480 | return; | |
481 | } | |
482 | } | |
483 | else if (unroll_type == UNROLL_MODULO) | |
484 | { | |
485 | /* Partially unrolling the loop: The compare and branch at the end | |
486 | (the last two instructions) must remain. Don't copy the compare | |
487 | and branch instructions at the end of the loop. Insert the unrolled | |
488 | code immediately before the compare/branch at the end so that the | |
489 | code will fall through to them as before. */ | |
490 | ||
491 | copy_start = loop_start; | |
492 | ||
493 | /* Set insert_before to the jump insn at the end of the loop. | |
494 | Set copy_end to before the jump insn at the end of the loop. */ | |
495 | if (GET_CODE (last_loop_insn) == BARRIER) | |
496 | { | |
497 | insert_before = PREV_INSN (last_loop_insn); | |
498 | copy_end = PREV_INSN (insert_before); | |
499 | } | |
500 | else if (GET_CODE (last_loop_insn) == JUMP_INSN) | |
501 | { | |
502 | #ifdef HAVE_cc0 | |
503 | /* The instruction immediately before the JUMP_INSN is a compare | |
504 | instruction which we do not want to copy or delete. */ | |
505 | insert_before = PREV_INSN (last_loop_insn); | |
506 | copy_end = PREV_INSN (insert_before); | |
507 | #else | |
508 | /* The instruction immediately before the JUMP_INSN may not be the | |
509 | compare, so we must copy it. */ | |
510 | insert_before = last_loop_insn; | |
511 | copy_end = PREV_INSN (last_loop_insn); | |
512 | #endif | |
513 | } | |
514 | else | |
515 | { | |
516 | /* We currently can't unroll a loop if it doesn't end with a | |
517 | JUMP_INSN. There would need to be a mechanism that recognizes | |
518 | this case, and then inserts a jump after each loop body, which | |
519 | jumps to after the last loop body. */ | |
520 | if (loop_dump_stream) | |
521 | fprintf (loop_dump_stream, | |
522 | "Unrolling failure: loop does not end with a JUMP_INSN.\n"); | |
523 | return; | |
524 | } | |
525 | } | |
526 | else | |
527 | { | |
528 | /* Normal case: Must copy the compare and branch instructions at the | |
529 | end of the loop. */ | |
530 | ||
531 | if (GET_CODE (last_loop_insn) == BARRIER) | |
532 | { | |
533 | /* Loop ends with an unconditional jump and a barrier. | |
534 | Handle this like above, don't copy jump and barrier. | |
535 | This is not strictly necessary, but doing so prevents generating | |
536 | unconditional jumps to an immediately following label. | |
537 | ||
538 | This will be corrected below if the target of this jump is | |
539 | not the start_label. */ | |
540 | ||
541 | insert_before = PREV_INSN (last_loop_insn); | |
542 | copy_end = PREV_INSN (insert_before); | |
543 | } | |
544 | else if (GET_CODE (last_loop_insn) == JUMP_INSN) | |
545 | { | |
546 | /* Set insert_before to immediately after the JUMP_INSN, so that | |
547 | NOTEs at the end of the loop will be correctly handled by | |
548 | copy_loop_body. */ | |
549 | insert_before = NEXT_INSN (last_loop_insn); | |
550 | copy_end = last_loop_insn; | |
551 | } | |
552 | else | |
553 | { | |
554 | /* We currently can't unroll a loop if it doesn't end with a | |
555 | JUMP_INSN. There would need to be a mechanism that recognizes | |
556 | this case, and then inserts a jump after each loop body, which | |
557 | jumps to after the last loop body. */ | |
558 | if (loop_dump_stream) | |
559 | fprintf (loop_dump_stream, | |
560 | "Unrolling failure: loop does not end with a JUMP_INSN.\n"); | |
561 | return; | |
562 | } | |
563 | ||
564 | /* If copying exit test branches because they can not be eliminated, | |
565 | then must convert the fall through case of the branch to a jump past | |
566 | the end of the loop. Create a label to emit after the loop and save | |
567 | it for later use. Do not use the label after the loop, if any, since | |
568 | it might be used by insns outside the loop, or there might be insns | |
569 | added before it later by final_[bg]iv_value which must be after | |
570 | the real exit label. */ | |
571 | exit_label = gen_label_rtx (); | |
572 | ||
573 | insn = loop_start; | |
574 | while (GET_CODE (insn) != CODE_LABEL && GET_CODE (insn) != JUMP_INSN) | |
575 | insn = NEXT_INSN (insn); | |
576 | ||
577 | if (GET_CODE (insn) == JUMP_INSN) | |
578 | { | |
579 | /* The loop starts with a jump down to the exit condition test. | |
580 | Start copying the loop after the barrier following this | |
581 | jump insn. */ | |
582 | copy_start = NEXT_INSN (insn); | |
583 | ||
584 | /* Splitting induction variables doesn't work when the loop is | |
585 | entered via a jump to the bottom, because then we end up doing | |
586 | a comparison against a new register for a split variable, but | |
587 | we did not execute the set insn for the new register because | |
588 | it was skipped over. */ | |
589 | splitting_not_safe = 1; | |
590 | if (loop_dump_stream) | |
591 | fprintf (loop_dump_stream, | |
592 | "Splitting not safe, because loop not entered at top.\n"); | |
593 | } | |
594 | else | |
595 | copy_start = loop_start; | |
596 | } | |
597 | ||
598 | /* This should always be the first label in the loop. */ | |
599 | start_label = NEXT_INSN (copy_start); | |
600 | /* There may be a line number note and/or a loop continue note here. */ | |
601 | while (GET_CODE (start_label) == NOTE) | |
602 | start_label = NEXT_INSN (start_label); | |
603 | if (GET_CODE (start_label) != CODE_LABEL) | |
604 | { | |
605 | /* This can happen as a result of jump threading. If the first insns in | |
606 | the loop test the same condition as the loop's backward jump, or the | |
607 | opposite condition, then the backward jump will be modified to point | |
608 | to elsewhere, and the loop's start label is deleted. | |
609 | ||
610 | This case currently can not be handled by the loop unrolling code. */ | |
611 | ||
612 | if (loop_dump_stream) | |
613 | fprintf (loop_dump_stream, | |
614 | "Unrolling failure: unknown insns between BEG note and loop label.\n"); | |
615 | return; | |
616 | } | |
b9b817f0 RS |
617 | if (LABEL_NAME (start_label)) |
618 | { | |
619 | /* The jump optimization pass must have combined the original start label | |
620 | with a named label for a goto. We can't unroll this case because | |
621 | jumps which go to the named label must be handled differently than | |
622 | jumps to the loop start, and it is impossible to differentiate them | |
623 | in this case. */ | |
624 | if (loop_dump_stream) | |
625 | fprintf (loop_dump_stream, | |
626 | "Unrolling failure: loop start label is gone\n"); | |
627 | return; | |
628 | } | |
67f2de41 RK |
629 | |
630 | if (unroll_type == UNROLL_NAIVE | |
631 | && GET_CODE (last_loop_insn) == BARRIER | |
632 | && start_label != JUMP_LABEL (PREV_INSN (last_loop_insn))) | |
633 | { | |
634 | /* In this case, we must copy the jump and barrier, because they will | |
635 | not be converted to jumps to an immediately following label. */ | |
636 | ||
637 | insert_before = NEXT_INSN (last_loop_insn); | |
638 | copy_end = last_loop_insn; | |
639 | } | |
640 | ||
92905582 JW |
641 | if (unroll_type == UNROLL_NAIVE |
642 | && GET_CODE (last_loop_insn) == JUMP_INSN | |
643 | && start_label != JUMP_LABEL (last_loop_insn)) | |
644 | { | |
645 | /* ??? The loop ends with a conditional branch that does not branch back | |
646 | to the loop start label. In this case, we must emit an unconditional | |
647 | branch to the loop exit after emitting the final branch. | |
648 | copy_loop_body does not have support for this currently, so we | |
649 | give up. It doesn't seem worthwhile to unroll anyways since | |
650 | unrolling would increase the number of branch instructions | |
651 | executed. */ | |
652 | if (loop_dump_stream) | |
653 | fprintf (loop_dump_stream, | |
654 | "Unrolling failure: final conditional branch not to loop start\n"); | |
655 | return; | |
656 | } | |
657 | ||
67f2de41 RK |
658 | /* Allocate a translation table for the labels and insn numbers. |
659 | They will be filled in as we copy the insns in the loop. */ | |
660 | ||
661 | max_labelno = max_label_num (); | |
662 | max_insnno = get_max_uid (); | |
663 | ||
664 | map = (struct inline_remap *) alloca (sizeof (struct inline_remap)); | |
665 | ||
bae12186 JW |
666 | map->integrating = 0; |
667 | ||
67f2de41 RK |
668 | /* Allocate the label map. */ |
669 | ||
670 | if (max_labelno > 0) | |
671 | { | |
672 | map->label_map = (rtx *) alloca (max_labelno * sizeof (rtx)); | |
673 | ||
674 | local_label = (char *) alloca (max_labelno); | |
675 | bzero (local_label, max_labelno); | |
676 | } | |
677 | else | |
678 | map->label_map = 0; | |
679 | ||
680 | /* Search the loop and mark all local labels, i.e. the ones which have to | |
681 | be distinct labels when copied. For all labels which might be | |
682 | non-local, set their label_map entries to point to themselves. | |
683 | If they happen to be local their label_map entries will be overwritten | |
684 | before the loop body is copied. The label_map entries for local labels | |
685 | will be set to a different value each time the loop body is copied. */ | |
686 | ||
687 | for (insn = copy_start; insn != loop_end; insn = NEXT_INSN (insn)) | |
688 | { | |
689 | if (GET_CODE (insn) == CODE_LABEL) | |
690 | local_label[CODE_LABEL_NUMBER (insn)] = 1; | |
691 | else if (GET_CODE (insn) == JUMP_INSN) | |
692 | { | |
693 | if (JUMP_LABEL (insn)) | |
1f3d3a31 JL |
694 | set_label_in_map (map, |
695 | CODE_LABEL_NUMBER (JUMP_LABEL (insn)), | |
696 | JUMP_LABEL (insn)); | |
67f2de41 RK |
697 | else if (GET_CODE (PATTERN (insn)) == ADDR_VEC |
698 | || GET_CODE (PATTERN (insn)) == ADDR_DIFF_VEC) | |
699 | { | |
700 | rtx pat = PATTERN (insn); | |
701 | int diff_vec_p = GET_CODE (PATTERN (insn)) == ADDR_DIFF_VEC; | |
702 | int len = XVECLEN (pat, diff_vec_p); | |
703 | rtx label; | |
704 | ||
705 | for (i = 0; i < len; i++) | |
706 | { | |
707 | label = XEXP (XVECEXP (pat, diff_vec_p, i), 0); | |
1f3d3a31 JL |
708 | set_label_in_map (map, |
709 | CODE_LABEL_NUMBER (label), | |
710 | label); | |
67f2de41 RK |
711 | } |
712 | } | |
713 | } | |
714 | } | |
715 | ||
716 | /* Allocate space for the insn map. */ | |
717 | ||
718 | map->insn_map = (rtx *) alloca (max_insnno * sizeof (rtx)); | |
719 | ||
720 | /* Set this to zero, to indicate that we are doing loop unrolling, | |
721 | not function inlining. */ | |
722 | map->inline_target = 0; | |
723 | ||
724 | /* The register and constant maps depend on the number of registers | |
725 | present, so the final maps can't be created until after | |
726 | find_splittable_regs is called. However, they are needed for | |
727 | preconditioning, so we create temporary maps when preconditioning | |
728 | is performed. */ | |
729 | ||
730 | /* The preconditioning code may allocate two new pseudo registers. */ | |
731 | maxregnum = max_reg_num (); | |
732 | ||
733 | /* Allocate and zero out the splittable_regs and addr_combined_regs | |
734 | arrays. These must be zeroed here because they will be used if | |
735 | loop preconditioning is performed, and must be zero for that case. | |
736 | ||
737 | It is safe to do this here, since the extra registers created by the | |
738 | preconditioning code and find_splittable_regs will never be used | |
6dc42e49 | 739 | to access the splittable_regs[] and addr_combined_regs[] arrays. */ |
67f2de41 RK |
740 | |
741 | splittable_regs = (rtx *) alloca (maxregnum * sizeof (rtx)); | |
4c9a05bc | 742 | bzero ((char *) splittable_regs, maxregnum * sizeof (rtx)); |
67f2de41 | 743 | splittable_regs_updates = (int *) alloca (maxregnum * sizeof (int)); |
4c9a05bc | 744 | bzero ((char *) splittable_regs_updates, maxregnum * sizeof (int)); |
67f2de41 RK |
745 | addr_combined_regs |
746 | = (struct induction **) alloca (maxregnum * sizeof (struct induction *)); | |
4c9a05bc | 747 | bzero ((char *) addr_combined_regs, maxregnum * sizeof (struct induction *)); |
ba68fc32 JW |
748 | /* We must limit it to max_reg_before_loop, because only these pseudo |
749 | registers have valid regno_first_uid info. Any register created after | |
750 | that is unlikely to be local to the loop anyways. */ | |
751 | local_regno = (char *) alloca (max_reg_before_loop); | |
752 | bzero (local_regno, max_reg_before_loop); | |
7f5b8ca7 RK |
753 | |
754 | /* Mark all local registers, i.e. the ones which are referenced only | |
ba68fc32 JW |
755 | inside the loop. */ |
756 | if (INSN_UID (copy_end) < max_uid_for_loop) | |
7f5b8ca7 RK |
757 | { |
758 | int copy_start_luid = INSN_LUID (copy_start); | |
759 | int copy_end_luid = INSN_LUID (copy_end); | |
760 | ||
b5f89788 RK |
761 | /* If a register is used in the jump insn, we must not duplicate it |
762 | since it will also be used outside the loop. */ | |
763 | if (GET_CODE (copy_end) == JUMP_INSN) | |
764 | copy_end_luid--; | |
01e7bd9a JW |
765 | /* If copy_start points to the NOTE that starts the loop, then we must |
766 | use the next luid, because invariant pseudo-regs moved out of the loop | |
767 | have their lifetimes modified to start here, but they are not safe | |
768 | to duplicate. */ | |
769 | if (copy_start == loop_start) | |
770 | copy_start_luid++; | |
b5f89788 | 771 | |
1fe33d17 JW |
772 | /* If a pseudo's lifetime is entirely contained within this loop, then we |
773 | can use a different pseudo in each unrolled copy of the loop. This | |
774 | results in better code. */ | |
ba68fc32 | 775 | for (j = FIRST_PSEUDO_REGISTER; j < max_reg_before_loop; ++j) |
b1f21e0a MM |
776 | if (REGNO_FIRST_UID (j) > 0 && REGNO_FIRST_UID (j) <= max_uid_for_loop |
777 | && uid_luid[REGNO_FIRST_UID (j)] >= copy_start_luid | |
778 | && REGNO_LAST_UID (j) > 0 && REGNO_LAST_UID (j) <= max_uid_for_loop | |
779 | && uid_luid[REGNO_LAST_UID (j)] <= copy_end_luid) | |
1fe33d17 JW |
780 | { |
781 | /* However, we must also check for loop-carried dependencies. | |
782 | If the value the pseudo has at the end of iteration X is | |
783 | used by iteration X+1, then we can not use a different pseudo | |
784 | for each unrolled copy of the loop. */ | |
785 | /* A pseudo is safe if regno_first_uid is a set, and this | |
786 | set dominates all instructions from regno_first_uid to | |
787 | regno_last_uid. */ | |
788 | /* ??? This check is simplistic. We would get better code if | |
789 | this check was more sophisticated. */ | |
b1f21e0a | 790 | if (set_dominates_use (j, REGNO_FIRST_UID (j), REGNO_LAST_UID (j), |
1fe33d17 JW |
791 | copy_start, copy_end)) |
792 | local_regno[j] = 1; | |
793 | ||
794 | if (loop_dump_stream) | |
795 | { | |
796 | if (local_regno[j]) | |
797 | fprintf (loop_dump_stream, "Marked reg %d as local\n", j); | |
798 | else | |
799 | fprintf (loop_dump_stream, "Did not mark reg %d as local\n", | |
800 | j); | |
801 | } | |
802 | } | |
7f5b8ca7 | 803 | } |
67f2de41 RK |
804 | |
805 | /* If this loop requires exit tests when unrolled, check to see if we | |
806 | can precondition the loop so as to make the exit tests unnecessary. | |
807 | Just like variable splitting, this is not safe if the loop is entered | |
808 | via a jump to the bottom. Also, can not do this if no strength | |
809 | reduce info, because precondition_loop_p uses this info. */ | |
810 | ||
811 | /* Must copy the loop body for preconditioning before the following | |
812 | find_splittable_regs call since that will emit insns which need to | |
813 | be after the preconditioned loop copies, but immediately before the | |
814 | unrolled loop copies. */ | |
815 | ||
816 | /* Also, it is not safe to split induction variables for the preconditioned | |
817 | copies of the loop body. If we split induction variables, then the code | |
818 | assumes that each induction variable can be represented as a function | |
819 | of its initial value and the loop iteration number. This is not true | |
820 | in this case, because the last preconditioned copy of the loop body | |
821 | could be any iteration from the first up to the `unroll_number-1'th, | |
822 | depending on the initial value of the iteration variable. Therefore | |
823 | we can not split induction variables here, because we can not calculate | |
824 | their value. Hence, this code must occur before find_splittable_regs | |
825 | is called. */ | |
826 | ||
827 | if (unroll_type == UNROLL_NAIVE && ! splitting_not_safe && strength_reduce_p) | |
828 | { | |
829 | rtx initial_value, final_value, increment; | |
830 | ||
831 | if (precondition_loop_p (&initial_value, &final_value, &increment, | |
832 | loop_start, loop_end)) | |
833 | { | |
ccd043a9 | 834 | register rtx diff ; |
67f2de41 RK |
835 | enum machine_mode mode; |
836 | rtx *labels; | |
837 | int abs_inc, neg_inc; | |
838 | ||
839 | map->reg_map = (rtx *) alloca (maxregnum * sizeof (rtx)); | |
840 | ||
841 | map->const_equiv_map = (rtx *) alloca (maxregnum * sizeof (rtx)); | |
842 | map->const_age_map = (unsigned *) alloca (maxregnum | |
843 | * sizeof (unsigned)); | |
844 | map->const_equiv_map_size = maxregnum; | |
845 | global_const_equiv_map = map->const_equiv_map; | |
e39bf2d9 | 846 | global_const_equiv_map_size = maxregnum; |
67f2de41 RK |
847 | |
848 | init_reg_map (map, maxregnum); | |
849 | ||
850 | /* Limit loop unrolling to 4, since this will make 7 copies of | |
851 | the loop body. */ | |
852 | if (unroll_number > 4) | |
853 | unroll_number = 4; | |
854 | ||
855 | /* Save the absolute value of the increment, and also whether or | |
856 | not it is negative. */ | |
857 | neg_inc = 0; | |
858 | abs_inc = INTVAL (increment); | |
859 | if (abs_inc < 0) | |
860 | { | |
861 | abs_inc = - abs_inc; | |
862 | neg_inc = 1; | |
863 | } | |
864 | ||
865 | start_sequence (); | |
866 | ||
867 | /* Decide what mode to do these calculations in. Choose the larger | |
868 | of final_value's mode and initial_value's mode, or a full-word if | |
869 | both are constants. */ | |
870 | mode = GET_MODE (final_value); | |
871 | if (mode == VOIDmode) | |
872 | { | |
873 | mode = GET_MODE (initial_value); | |
874 | if (mode == VOIDmode) | |
875 | mode = word_mode; | |
876 | } | |
877 | else if (mode != GET_MODE (initial_value) | |
878 | && (GET_MODE_SIZE (mode) | |
879 | < GET_MODE_SIZE (GET_MODE (initial_value)))) | |
880 | mode = GET_MODE (initial_value); | |
881 | ||
882 | /* Calculate the difference between the final and initial values. | |
883 | Final value may be a (plus (reg x) (const_int 1)) rtx. | |
884 | Let the following cse pass simplify this if initial value is | |
885 | a constant. | |
886 | ||
887 | We must copy the final and initial values here to avoid | |
888 | improperly shared rtl. */ | |
889 | ||
890 | diff = expand_binop (mode, sub_optab, copy_rtx (final_value), | |
c166a311 | 891 | copy_rtx (initial_value), NULL_RTX, 0, |
67f2de41 RK |
892 | OPTAB_LIB_WIDEN); |
893 | ||
894 | /* Now calculate (diff % (unroll * abs (increment))) by using an | |
895 | and instruction. */ | |
896 | diff = expand_binop (GET_MODE (diff), and_optab, diff, | |
c166a311 CH |
897 | GEN_INT (unroll_number * abs_inc - 1), |
898 | NULL_RTX, 0, OPTAB_LIB_WIDEN); | |
67f2de41 RK |
899 | |
900 | /* Now emit a sequence of branches to jump to the proper precond | |
901 | loop entry point. */ | |
902 | ||
903 | labels = (rtx *) alloca (sizeof (rtx) * unroll_number); | |
904 | for (i = 0; i < unroll_number; i++) | |
905 | labels[i] = gen_label_rtx (); | |
906 | ||
1dcfa896 JW |
907 | /* Check for the case where the initial value is greater than or |
908 | equal to the final value. In that case, we want to execute | |
909 | exactly one loop iteration. The code below will fail for this | |
910 | case. This check does not apply if the loop has a NE | |
911 | comparison at the end. */ | |
7edd39eb | 912 | |
1dcfa896 JW |
913 | if (loop_comparison_code != NE) |
914 | { | |
915 | emit_cmp_insn (initial_value, final_value, neg_inc ? LE : GE, | |
916 | NULL_RTX, mode, 0, 0); | |
917 | if (neg_inc) | |
918 | emit_jump_insn (gen_ble (labels[1])); | |
919 | else | |
920 | emit_jump_insn (gen_bge (labels[1])); | |
921 | JUMP_LABEL (get_last_insn ()) = labels[1]; | |
922 | LABEL_NUSES (labels[1])++; | |
923 | } | |
7edd39eb | 924 | |
67f2de41 RK |
925 | /* Assuming the unroll_number is 4, and the increment is 2, then |
926 | for a negative increment: for a positive increment: | |
927 | diff = 0,1 precond 0 diff = 0,7 precond 0 | |
928 | diff = 2,3 precond 3 diff = 1,2 precond 1 | |
929 | diff = 4,5 precond 2 diff = 3,4 precond 2 | |
930 | diff = 6,7 precond 1 diff = 5,6 precond 3 */ | |
931 | ||
932 | /* We only need to emit (unroll_number - 1) branches here, the | |
933 | last case just falls through to the following code. */ | |
934 | ||
935 | /* ??? This would give better code if we emitted a tree of branches | |
936 | instead of the current linear list of branches. */ | |
937 | ||
938 | for (i = 0; i < unroll_number - 1; i++) | |
939 | { | |
940 | int cmp_const; | |
7edd39eb | 941 | enum rtx_code cmp_code; |
67f2de41 RK |
942 | |
943 | /* For negative increments, must invert the constant compared | |
944 | against, except when comparing against zero. */ | |
945 | if (i == 0) | |
7edd39eb RK |
946 | { |
947 | cmp_const = 0; | |
948 | cmp_code = EQ; | |
949 | } | |
67f2de41 | 950 | else if (neg_inc) |
7edd39eb RK |
951 | { |
952 | cmp_const = unroll_number - i; | |
953 | cmp_code = GE; | |
954 | } | |
67f2de41 | 955 | else |
7edd39eb RK |
956 | { |
957 | cmp_const = i; | |
958 | cmp_code = LE; | |
959 | } | |
67f2de41 | 960 | |
c166a311 | 961 | emit_cmp_insn (diff, GEN_INT (abs_inc * cmp_const), |
7edd39eb | 962 | cmp_code, NULL_RTX, mode, 0, 0); |
67f2de41 RK |
963 | |
964 | if (i == 0) | |
965 | emit_jump_insn (gen_beq (labels[i])); | |
966 | else if (neg_inc) | |
967 | emit_jump_insn (gen_bge (labels[i])); | |
968 | else | |
969 | emit_jump_insn (gen_ble (labels[i])); | |
970 | JUMP_LABEL (get_last_insn ()) = labels[i]; | |
971 | LABEL_NUSES (labels[i])++; | |
972 | } | |
973 | ||
974 | /* If the increment is greater than one, then we need another branch, | |
975 | to handle other cases equivalent to 0. */ | |
976 | ||
977 | /* ??? This should be merged into the code above somehow to help | |
978 | simplify the code here, and reduce the number of branches emitted. | |
979 | For the negative increment case, the branch here could easily | |
980 | be merged with the `0' case branch above. For the positive | |
981 | increment case, it is not clear how this can be simplified. */ | |
982 | ||
983 | if (abs_inc != 1) | |
984 | { | |
985 | int cmp_const; | |
7edd39eb | 986 | enum rtx_code cmp_code; |
67f2de41 RK |
987 | |
988 | if (neg_inc) | |
7edd39eb RK |
989 | { |
990 | cmp_const = abs_inc - 1; | |
991 | cmp_code = LE; | |
992 | } | |
67f2de41 | 993 | else |
7edd39eb RK |
994 | { |
995 | cmp_const = abs_inc * (unroll_number - 1) + 1; | |
996 | cmp_code = GE; | |
997 | } | |
67f2de41 | 998 | |
7edd39eb | 999 | emit_cmp_insn (diff, GEN_INT (cmp_const), cmp_code, NULL_RTX, |
c166a311 | 1000 | mode, 0, 0); |
67f2de41 RK |
1001 | |
1002 | if (neg_inc) | |
1003 | emit_jump_insn (gen_ble (labels[0])); | |
1004 | else | |
1005 | emit_jump_insn (gen_bge (labels[0])); | |
1006 | JUMP_LABEL (get_last_insn ()) = labels[0]; | |
1007 | LABEL_NUSES (labels[0])++; | |
1008 | } | |
1009 | ||
1010 | sequence = gen_sequence (); | |
1011 | end_sequence (); | |
1012 | emit_insn_before (sequence, loop_start); | |
1013 | ||
1014 | /* Only the last copy of the loop body here needs the exit | |
1015 | test, so set copy_end to exclude the compare/branch here, | |
1016 | and then reset it inside the loop when get to the last | |
1017 | copy. */ | |
1018 | ||
1019 | if (GET_CODE (last_loop_insn) == BARRIER) | |
1020 | copy_end = PREV_INSN (PREV_INSN (last_loop_insn)); | |
1021 | else if (GET_CODE (last_loop_insn) == JUMP_INSN) | |
1022 | { | |
1023 | #ifdef HAVE_cc0 | |
b4ac57ab | 1024 | /* The immediately preceding insn is a compare which we do not |
67f2de41 RK |
1025 | want to copy. */ |
1026 | copy_end = PREV_INSN (PREV_INSN (last_loop_insn)); | |
1027 | #else | |
b4ac57ab | 1028 | /* The immediately preceding insn may not be a compare, so we |
67f2de41 RK |
1029 | must copy it. */ |
1030 | copy_end = PREV_INSN (last_loop_insn); | |
1031 | #endif | |
1032 | } | |
1033 | else | |
1034 | abort (); | |
1035 | ||
1036 | for (i = 1; i < unroll_number; i++) | |
1037 | { | |
1038 | emit_label_after (labels[unroll_number - i], | |
1039 | PREV_INSN (loop_start)); | |
1040 | ||
4c9a05bc RK |
1041 | bzero ((char *) map->insn_map, max_insnno * sizeof (rtx)); |
1042 | bzero ((char *) map->const_equiv_map, maxregnum * sizeof (rtx)); | |
1043 | bzero ((char *) map->const_age_map, | |
1044 | maxregnum * sizeof (unsigned)); | |
67f2de41 RK |
1045 | map->const_age = 0; |
1046 | ||
1047 | for (j = 0; j < max_labelno; j++) | |
1048 | if (local_label[j]) | |
1f3d3a31 | 1049 | set_label_in_map (map, j, gen_label_rtx ()); |
67f2de41 | 1050 | |
ba68fc32 | 1051 | for (j = FIRST_PSEUDO_REGISTER; j < max_reg_before_loop; j++) |
7f5b8ca7 | 1052 | if (local_regno[j]) |
9ae8ffe7 JL |
1053 | { |
1054 | map->reg_map[j] = gen_reg_rtx (GET_MODE (regno_reg_rtx[j])); | |
1055 | record_base_value (REGNO (map->reg_map[j]), | |
1056 | regno_reg_rtx[j]); | |
1057 | } | |
67f2de41 RK |
1058 | /* The last copy needs the compare/branch insns at the end, |
1059 | so reset copy_end here if the loop ends with a conditional | |
1060 | branch. */ | |
1061 | ||
1062 | if (i == unroll_number - 1) | |
1063 | { | |
1064 | if (GET_CODE (last_loop_insn) == BARRIER) | |
1065 | copy_end = PREV_INSN (PREV_INSN (last_loop_insn)); | |
1066 | else | |
1067 | copy_end = last_loop_insn; | |
1068 | } | |
1069 | ||
1070 | /* None of the copies are the `last_iteration', so just | |
1071 | pass zero for that parameter. */ | |
1072 | copy_loop_body (copy_start, copy_end, map, exit_label, 0, | |
1073 | unroll_type, start_label, loop_end, | |
1074 | loop_start, copy_end); | |
1075 | } | |
1076 | emit_label_after (labels[0], PREV_INSN (loop_start)); | |
1077 | ||
1078 | if (GET_CODE (last_loop_insn) == BARRIER) | |
1079 | { | |
1080 | insert_before = PREV_INSN (last_loop_insn); | |
1081 | copy_end = PREV_INSN (insert_before); | |
1082 | } | |
1083 | else | |
1084 | { | |
1085 | #ifdef HAVE_cc0 | |
b4ac57ab | 1086 | /* The immediately preceding insn is a compare which we do not |
67f2de41 RK |
1087 | want to copy. */ |
1088 | insert_before = PREV_INSN (last_loop_insn); | |
1089 | copy_end = PREV_INSN (insert_before); | |
1090 | #else | |
b4ac57ab | 1091 | /* The immediately preceding insn may not be a compare, so we |
67f2de41 RK |
1092 | must copy it. */ |
1093 | insert_before = last_loop_insn; | |
1094 | copy_end = PREV_INSN (last_loop_insn); | |
1095 | #endif | |
1096 | } | |
1097 | ||
1098 | /* Set unroll type to MODULO now. */ | |
1099 | unroll_type = UNROLL_MODULO; | |
1100 | loop_preconditioned = 1; | |
237a9795 | 1101 | |
8c660648 | 1102 | #ifdef HAIFA |
237a9795 JL |
1103 | /* Fix the initial value for the loop as needed. */ |
1104 | if (loop_n_iterations <= 0) | |
37aa45a2 | 1105 | loop_start_value [uid_loop_num [INSN_UID (loop_start)]] |
237a9795 | 1106 | = initial_value; |
8c660648 | 1107 | #endif |
67f2de41 RK |
1108 | } |
1109 | } | |
1110 | ||
1111 | /* If reach here, and the loop type is UNROLL_NAIVE, then don't unroll | |
1112 | the loop unless all loops are being unrolled. */ | |
1113 | if (unroll_type == UNROLL_NAIVE && ! flag_unroll_all_loops) | |
1114 | { | |
1115 | if (loop_dump_stream) | |
1116 | fprintf (loop_dump_stream, "Unrolling failure: Naive unrolling not being done.\n"); | |
1117 | return; | |
1118 | } | |
1119 | ||
1120 | /* At this point, we are guaranteed to unroll the loop. */ | |
1121 | ||
237a9795 | 1122 | /* Keep track of the unroll factor for each loop. */ |
8c660648 | 1123 | if (unroll_type == UNROLL_COMPLETELY) |
37aa45a2 | 1124 | loop_unroll_factor [uid_loop_num [INSN_UID (loop_start)]] = -1; |
8c660648 | 1125 | else |
37aa45a2 | 1126 | loop_unroll_factor [uid_loop_num [INSN_UID (loop_start)]] = unroll_number; |
8c660648 JL |
1127 | |
1128 | ||
67f2de41 RK |
1129 | /* For each biv and giv, determine whether it can be safely split into |
1130 | a different variable for each unrolled copy of the loop body. | |
1131 | We precalculate and save this info here, since computing it is | |
1132 | expensive. | |
1133 | ||
1134 | Do this before deleting any instructions from the loop, so that | |
1135 | back_branch_in_range_p will work correctly. */ | |
1136 | ||
1137 | if (splitting_not_safe) | |
1138 | temp = 0; | |
1139 | else | |
1140 | temp = find_splittable_regs (unroll_type, loop_start, loop_end, | |
1141 | end_insert_before, unroll_number); | |
1142 | ||
1143 | /* find_splittable_regs may have created some new registers, so must | |
1144 | reallocate the reg_map with the new larger size, and must realloc | |
1145 | the constant maps also. */ | |
1146 | ||
1147 | maxregnum = max_reg_num (); | |
1148 | map->reg_map = (rtx *) alloca (maxregnum * sizeof (rtx)); | |
1149 | ||
1150 | init_reg_map (map, maxregnum); | |
1151 | ||
1152 | /* Space is needed in some of the map for new registers, so new_maxregnum | |
1153 | is an (over)estimate of how many registers will exist at the end. */ | |
1154 | new_maxregnum = maxregnum + (temp * unroll_number * 2); | |
1155 | ||
1156 | /* Must realloc space for the constant maps, because the number of registers | |
1157 | may have changed. */ | |
1158 | ||
1159 | map->const_equiv_map = (rtx *) alloca (new_maxregnum * sizeof (rtx)); | |
1160 | map->const_age_map = (unsigned *) alloca (new_maxregnum * sizeof (unsigned)); | |
1161 | ||
8dc0b179 | 1162 | map->const_equiv_map_size = new_maxregnum; |
67f2de41 | 1163 | global_const_equiv_map = map->const_equiv_map; |
e39bf2d9 | 1164 | global_const_equiv_map_size = new_maxregnum; |
67f2de41 RK |
1165 | |
1166 | /* Search the list of bivs and givs to find ones which need to be remapped | |
1167 | when split, and set their reg_map entry appropriately. */ | |
1168 | ||
1169 | for (bl = loop_iv_list; bl; bl = bl->next) | |
1170 | { | |
1171 | if (REGNO (bl->biv->src_reg) != bl->regno) | |
1172 | map->reg_map[bl->regno] = bl->biv->src_reg; | |
1173 | #if 0 | |
1174 | /* Currently, non-reduced/final-value givs are never split. */ | |
1175 | for (v = bl->giv; v; v = v->next_iv) | |
1176 | if (REGNO (v->src_reg) != bl->regno) | |
1177 | map->reg_map[REGNO (v->dest_reg)] = v->src_reg; | |
1178 | #endif | |
1179 | } | |
1180 | ||
8b4904a3 RK |
1181 | /* Use our current register alignment and pointer flags. */ |
1182 | map->regno_pointer_flag = regno_pointer_flag; | |
1183 | map->regno_pointer_align = regno_pointer_align; | |
1184 | ||
67f2de41 RK |
1185 | /* If the loop is being partially unrolled, and the iteration variables |
1186 | are being split, and are being renamed for the split, then must fix up | |
2b59419a | 1187 | the compare/jump instruction at the end of the loop to refer to the new |
67f2de41 RK |
1188 | registers. This compare isn't copied, so the registers used in it |
1189 | will never be replaced if it isn't done here. */ | |
1190 | ||
1191 | if (unroll_type == UNROLL_MODULO) | |
1192 | { | |
1193 | insn = NEXT_INSN (copy_end); | |
2b59419a JW |
1194 | if (GET_CODE (insn) == INSN || GET_CODE (insn) == JUMP_INSN) |
1195 | PATTERN (insn) = remap_split_bivs (PATTERN (insn)); | |
67f2de41 RK |
1196 | } |
1197 | ||
1198 | /* For unroll_number - 1 times, make a copy of each instruction | |
1199 | between copy_start and copy_end, and insert these new instructions | |
1200 | before the end of the loop. */ | |
1201 | ||
1202 | for (i = 0; i < unroll_number; i++) | |
1203 | { | |
4c9a05bc RK |
1204 | bzero ((char *) map->insn_map, max_insnno * sizeof (rtx)); |
1205 | bzero ((char *) map->const_equiv_map, new_maxregnum * sizeof (rtx)); | |
1206 | bzero ((char *) map->const_age_map, new_maxregnum * sizeof (unsigned)); | |
67f2de41 RK |
1207 | map->const_age = 0; |
1208 | ||
1209 | for (j = 0; j < max_labelno; j++) | |
1210 | if (local_label[j]) | |
1f3d3a31 | 1211 | set_label_in_map (map, j, gen_label_rtx ()); |
67f2de41 | 1212 | |
ba68fc32 | 1213 | for (j = FIRST_PSEUDO_REGISTER; j < max_reg_before_loop; j++) |
7f5b8ca7 | 1214 | if (local_regno[j]) |
9ae8ffe7 JL |
1215 | { |
1216 | map->reg_map[j] = gen_reg_rtx (GET_MODE (regno_reg_rtx[j])); | |
1217 | record_base_value (REGNO (map->reg_map[j]), | |
1218 | regno_reg_rtx[j]); | |
1219 | } | |
7f5b8ca7 | 1220 | |
67f2de41 RK |
1221 | /* If loop starts with a branch to the test, then fix it so that |
1222 | it points to the test of the first unrolled copy of the loop. */ | |
1223 | if (i == 0 && loop_start != copy_start) | |
1224 | { | |
1225 | insn = PREV_INSN (copy_start); | |
1226 | pattern = PATTERN (insn); | |
1227 | ||
1f3d3a31 JL |
1228 | tem = get_label_from_map (map, |
1229 | CODE_LABEL_NUMBER | |
1230 | (XEXP (SET_SRC (pattern), 0))); | |
38a448ca | 1231 | SET_SRC (pattern) = gen_rtx_LABEL_REF (VOIDmode, tem); |
67f2de41 RK |
1232 | |
1233 | /* Set the jump label so that it can be used by later loop unrolling | |
1234 | passes. */ | |
1235 | JUMP_LABEL (insn) = tem; | |
1236 | LABEL_NUSES (tem)++; | |
1237 | } | |
1238 | ||
1239 | copy_loop_body (copy_start, copy_end, map, exit_label, | |
1240 | i == unroll_number - 1, unroll_type, start_label, | |
1241 | loop_end, insert_before, insert_before); | |
1242 | } | |
1243 | ||
1244 | /* Before deleting any insns, emit a CODE_LABEL immediately after the last | |
1245 | insn to be deleted. This prevents any runaway delete_insn call from | |
1246 | more insns that it should, as it always stops at a CODE_LABEL. */ | |
1247 | ||
1248 | /* Delete the compare and branch at the end of the loop if completely | |
1249 | unrolling the loop. Deleting the backward branch at the end also | |
1250 | deletes the code label at the start of the loop. This is done at | |
1251 | the very end to avoid problems with back_branch_in_range_p. */ | |
1252 | ||
1253 | if (unroll_type == UNROLL_COMPLETELY) | |
1254 | safety_label = emit_label_after (gen_label_rtx (), last_loop_insn); | |
1255 | else | |
1256 | safety_label = emit_label_after (gen_label_rtx (), copy_end); | |
1257 | ||
1258 | /* Delete all of the original loop instructions. Don't delete the | |
1259 | LOOP_BEG note, or the first code label in the loop. */ | |
1260 | ||
1261 | insn = NEXT_INSN (copy_start); | |
1262 | while (insn != safety_label) | |
1263 | { | |
1264 | if (insn != start_label) | |
1265 | insn = delete_insn (insn); | |
1266 | else | |
1267 | insn = NEXT_INSN (insn); | |
1268 | } | |
1269 | ||
1270 | /* Can now delete the 'safety' label emitted to protect us from runaway | |
1271 | delete_insn calls. */ | |
1272 | if (INSN_DELETED_P (safety_label)) | |
1273 | abort (); | |
1274 | delete_insn (safety_label); | |
1275 | ||
1276 | /* If exit_label exists, emit it after the loop. Doing the emit here | |
1277 | forces it to have a higher INSN_UID than any insn in the unrolled loop. | |
1278 | This is needed so that mostly_true_jump in reorg.c will treat jumps | |
1279 | to this loop end label correctly, i.e. predict that they are usually | |
1280 | not taken. */ | |
1281 | if (exit_label) | |
1282 | emit_label_after (exit_label, loop_end); | |
67f2de41 RK |
1283 | } |
1284 | \f | |
1285 | /* Return true if the loop can be safely, and profitably, preconditioned | |
1286 | so that the unrolled copies of the loop body don't need exit tests. | |
1287 | ||
1288 | This only works if final_value, initial_value and increment can be | |
1289 | determined, and if increment is a constant power of 2. | |
1290 | If increment is not a power of 2, then the preconditioning modulo | |
1291 | operation would require a real modulo instead of a boolean AND, and this | |
1292 | is not considered `profitable'. */ | |
1293 | ||
1294 | /* ??? If the loop is known to be executed very many times, or the machine | |
1295 | has a very cheap divide instruction, then preconditioning is a win even | |
1296 | when the increment is not a power of 2. Use RTX_COST to compute | |
1297 | whether divide is cheap. */ | |
1298 | ||
1299 | static int | |
1300 | precondition_loop_p (initial_value, final_value, increment, loop_start, | |
1301 | loop_end) | |
1302 | rtx *initial_value, *final_value, *increment; | |
1303 | rtx loop_start, loop_end; | |
1304 | { | |
67f2de41 RK |
1305 | |
1306 | if (loop_n_iterations > 0) | |
1307 | { | |
1308 | *initial_value = const0_rtx; | |
1309 | *increment = const1_rtx; | |
c166a311 | 1310 | *final_value = GEN_INT (loop_n_iterations); |
67f2de41 RK |
1311 | |
1312 | if (loop_dump_stream) | |
1313 | fprintf (loop_dump_stream, | |
1314 | "Preconditioning: Success, number of iterations known, %d.\n", | |
1315 | loop_n_iterations); | |
1316 | return 1; | |
1317 | } | |
1318 | ||
1319 | if (loop_initial_value == 0) | |
1320 | { | |
1321 | if (loop_dump_stream) | |
1322 | fprintf (loop_dump_stream, | |
1323 | "Preconditioning: Could not find initial value.\n"); | |
1324 | return 0; | |
1325 | } | |
1326 | else if (loop_increment == 0) | |
1327 | { | |
1328 | if (loop_dump_stream) | |
1329 | fprintf (loop_dump_stream, | |
1330 | "Preconditioning: Could not find increment value.\n"); | |
1331 | return 0; | |
1332 | } | |
1333 | else if (GET_CODE (loop_increment) != CONST_INT) | |
1334 | { | |
1335 | if (loop_dump_stream) | |
1336 | fprintf (loop_dump_stream, | |
1337 | "Preconditioning: Increment not a constant.\n"); | |
1338 | return 0; | |
1339 | } | |
1340 | else if ((exact_log2 (INTVAL (loop_increment)) < 0) | |
1341 | && (exact_log2 (- INTVAL (loop_increment)) < 0)) | |
1342 | { | |
1343 | if (loop_dump_stream) | |
1344 | fprintf (loop_dump_stream, | |
1345 | "Preconditioning: Increment not a constant power of 2.\n"); | |
1346 | return 0; | |
1347 | } | |
1348 | ||
1349 | /* Unsigned_compare and compare_dir can be ignored here, since they do | |
1350 | not matter for preconditioning. */ | |
1351 | ||
1352 | if (loop_final_value == 0) | |
1353 | { | |
1354 | if (loop_dump_stream) | |
1355 | fprintf (loop_dump_stream, | |
1356 | "Preconditioning: EQ comparison loop.\n"); | |
1357 | return 0; | |
1358 | } | |
1359 | ||
1360 | /* Must ensure that final_value is invariant, so call invariant_p to | |
1361 | check. Before doing so, must check regno against max_reg_before_loop | |
6dc42e49 | 1362 | to make sure that the register is in the range covered by invariant_p. |
67f2de41 RK |
1363 | If it isn't, then it is most likely a biv/giv which by definition are |
1364 | not invariant. */ | |
1365 | if ((GET_CODE (loop_final_value) == REG | |
1366 | && REGNO (loop_final_value) >= max_reg_before_loop) | |
1367 | || (GET_CODE (loop_final_value) == PLUS | |
1368 | && REGNO (XEXP (loop_final_value, 0)) >= max_reg_before_loop) | |
1369 | || ! invariant_p (loop_final_value)) | |
1370 | { | |
1371 | if (loop_dump_stream) | |
1372 | fprintf (loop_dump_stream, | |
1373 | "Preconditioning: Final value not invariant.\n"); | |
1374 | return 0; | |
1375 | } | |
1376 | ||
1377 | /* Fail for floating point values, since the caller of this function | |
1378 | does not have code to deal with them. */ | |
1379 | if (GET_MODE_CLASS (GET_MODE (loop_final_value)) == MODE_FLOAT | |
412dc348 | 1380 | || GET_MODE_CLASS (GET_MODE (loop_initial_value)) == MODE_FLOAT) |
67f2de41 RK |
1381 | { |
1382 | if (loop_dump_stream) | |
1383 | fprintf (loop_dump_stream, | |
1384 | "Preconditioning: Floating point final or initial value.\n"); | |
1385 | return 0; | |
1386 | } | |
1387 | ||
1388 | /* Now set initial_value to be the iteration_var, since that may be a | |
1389 | simpler expression, and is guaranteed to be correct if all of the | |
1390 | above tests succeed. | |
1391 | ||
1392 | We can not use the initial_value as calculated, because it will be | |
1393 | one too small for loops of the form "while (i-- > 0)". We can not | |
1394 | emit code before the loop_skip_over insns to fix this problem as this | |
1395 | will then give a number one too large for loops of the form | |
1396 | "while (--i > 0)". | |
1397 | ||
1398 | Note that all loops that reach here are entered at the top, because | |
1399 | this function is not called if the loop starts with a jump. */ | |
1400 | ||
1401 | /* Fail if loop_iteration_var is not live before loop_start, since we need | |
1402 | to test its value in the preconditioning code. */ | |
1403 | ||
b1f21e0a | 1404 | if (uid_luid[REGNO_FIRST_UID (REGNO (loop_iteration_var))] |
67f2de41 RK |
1405 | > INSN_LUID (loop_start)) |
1406 | { | |
1407 | if (loop_dump_stream) | |
1408 | fprintf (loop_dump_stream, | |
1409 | "Preconditioning: Iteration var not live before loop start.\n"); | |
1410 | return 0; | |
1411 | } | |
1412 | ||
1413 | *initial_value = loop_iteration_var; | |
1414 | *increment = loop_increment; | |
1415 | *final_value = loop_final_value; | |
1416 | ||
1417 | /* Success! */ | |
1418 | if (loop_dump_stream) | |
1419 | fprintf (loop_dump_stream, "Preconditioning: Successful.\n"); | |
1420 | return 1; | |
1421 | } | |
1422 | ||
1423 | ||
1424 | /* All pseudo-registers must be mapped to themselves. Two hard registers | |
1425 | must be mapped, VIRTUAL_STACK_VARS_REGNUM and VIRTUAL_INCOMING_ARGS_ | |
1426 | REGNUM, to avoid function-inlining specific conversions of these | |
1427 | registers. All other hard regs can not be mapped because they may be | |
1428 | used with different | |
1429 | modes. */ | |
1430 | ||
1431 | static void | |
1432 | init_reg_map (map, maxregnum) | |
1433 | struct inline_remap *map; | |
1434 | int maxregnum; | |
1435 | { | |
1436 | int i; | |
1437 | ||
1438 | for (i = maxregnum - 1; i > LAST_VIRTUAL_REGISTER; i--) | |
1439 | map->reg_map[i] = regno_reg_rtx[i]; | |
1440 | /* Just clear the rest of the entries. */ | |
1441 | for (i = LAST_VIRTUAL_REGISTER; i >= 0; i--) | |
1442 | map->reg_map[i] = 0; | |
1443 | ||
1444 | map->reg_map[VIRTUAL_STACK_VARS_REGNUM] | |
1445 | = regno_reg_rtx[VIRTUAL_STACK_VARS_REGNUM]; | |
1446 | map->reg_map[VIRTUAL_INCOMING_ARGS_REGNUM] | |
1447 | = regno_reg_rtx[VIRTUAL_INCOMING_ARGS_REGNUM]; | |
1448 | } | |
1449 | \f | |
1450 | /* Strength-reduction will often emit code for optimized biv/givs which | |
1451 | calculates their value in a temporary register, and then copies the result | |
1452 | to the iv. This procedure reconstructs the pattern computing the iv; | |
1453 | verifying that all operands are of the proper form. | |
1454 | ||
1455 | The return value is the amount that the giv is incremented by. */ | |
1456 | ||
1457 | static rtx | |
1458 | calculate_giv_inc (pattern, src_insn, regno) | |
1459 | rtx pattern, src_insn; | |
1460 | int regno; | |
1461 | { | |
1462 | rtx increment; | |
6511b4f8 RS |
1463 | rtx increment_total = 0; |
1464 | int tries = 0; | |
67f2de41 | 1465 | |
6511b4f8 | 1466 | retry: |
67f2de41 RK |
1467 | /* Verify that we have an increment insn here. First check for a plus |
1468 | as the set source. */ | |
1469 | if (GET_CODE (SET_SRC (pattern)) != PLUS) | |
1470 | { | |
1471 | /* SR sometimes computes the new giv value in a temp, then copies it | |
1472 | to the new_reg. */ | |
1473 | src_insn = PREV_INSN (src_insn); | |
1474 | pattern = PATTERN (src_insn); | |
1475 | if (GET_CODE (SET_SRC (pattern)) != PLUS) | |
1476 | abort (); | |
1477 | ||
1478 | /* The last insn emitted is not needed, so delete it to avoid confusing | |
1479 | the second cse pass. This insn sets the giv unnecessarily. */ | |
1480 | delete_insn (get_last_insn ()); | |
1481 | } | |
1482 | ||
1483 | /* Verify that we have a constant as the second operand of the plus. */ | |
1484 | increment = XEXP (SET_SRC (pattern), 1); | |
1485 | if (GET_CODE (increment) != CONST_INT) | |
1486 | { | |
1487 | /* SR sometimes puts the constant in a register, especially if it is | |
1488 | too big to be an add immed operand. */ | |
6511b4f8 RS |
1489 | src_insn = PREV_INSN (src_insn); |
1490 | increment = SET_SRC (PATTERN (src_insn)); | |
67f2de41 RK |
1491 | |
1492 | /* SR may have used LO_SUM to compute the constant if it is too large | |
1493 | for a load immed operand. In this case, the constant is in operand | |
1494 | one of the LO_SUM rtx. */ | |
1495 | if (GET_CODE (increment) == LO_SUM) | |
1496 | increment = XEXP (increment, 1); | |
efb84aa5 GK |
1497 | |
1498 | /* Some ports store large constants in memory and add a REG_EQUAL | |
1499 | note to the store insn. */ | |
1500 | else if (GET_CODE (increment) == MEM) | |
1501 | { | |
1502 | rtx note = find_reg_note (src_insn, REG_EQUAL, 0); | |
1503 | if (note) | |
1504 | increment = XEXP (note, 0); | |
1505 | } | |
1506 | ||
92065e1f | 1507 | else if (GET_CODE (increment) == IOR |
2599dcc7 JW |
1508 | || GET_CODE (increment) == ASHIFT |
1509 | || GET_CODE (increment) == PLUS) | |
8700c494 | 1510 | { |
92065e1f | 1511 | /* The rs6000 port loads some constants with IOR. |
2599dcc7 | 1512 | The alpha port loads some constants with ASHIFT and PLUS. */ |
8700c494 | 1513 | rtx second_part = XEXP (increment, 1); |
92065e1f | 1514 | enum rtx_code code = GET_CODE (increment); |
8700c494 JW |
1515 | |
1516 | src_insn = PREV_INSN (src_insn); | |
1517 | increment = SET_SRC (PATTERN (src_insn)); | |
1518 | /* Don't need the last insn anymore. */ | |
1519 | delete_insn (get_last_insn ()); | |
1520 | ||
1521 | if (GET_CODE (second_part) != CONST_INT | |
1522 | || GET_CODE (increment) != CONST_INT) | |
1523 | abort (); | |
1524 | ||
92065e1f RK |
1525 | if (code == IOR) |
1526 | increment = GEN_INT (INTVAL (increment) | INTVAL (second_part)); | |
2599dcc7 JW |
1527 | else if (code == PLUS) |
1528 | increment = GEN_INT (INTVAL (increment) + INTVAL (second_part)); | |
92065e1f RK |
1529 | else |
1530 | increment = GEN_INT (INTVAL (increment) << INTVAL (second_part)); | |
8700c494 | 1531 | } |
67f2de41 RK |
1532 | |
1533 | if (GET_CODE (increment) != CONST_INT) | |
1534 | abort (); | |
1535 | ||
8700c494 | 1536 | /* The insn loading the constant into a register is no longer needed, |
67f2de41 RK |
1537 | so delete it. */ |
1538 | delete_insn (get_last_insn ()); | |
1539 | } | |
1540 | ||
6511b4f8 RS |
1541 | if (increment_total) |
1542 | increment_total = GEN_INT (INTVAL (increment_total) + INTVAL (increment)); | |
1543 | else | |
1544 | increment_total = increment; | |
1545 | ||
1546 | /* Check that the source register is the same as the register we expected | |
1547 | to see as the source. If not, something is seriously wrong. */ | |
67f2de41 RK |
1548 | if (GET_CODE (XEXP (SET_SRC (pattern), 0)) != REG |
1549 | || REGNO (XEXP (SET_SRC (pattern), 0)) != regno) | |
6511b4f8 RS |
1550 | { |
1551 | /* Some machines (e.g. the romp), may emit two add instructions for | |
1552 | certain constants, so lets try looking for another add immediately | |
1553 | before this one if we have only seen one add insn so far. */ | |
1554 | ||
1555 | if (tries == 0) | |
1556 | { | |
1557 | tries++; | |
1558 | ||
1559 | src_insn = PREV_INSN (src_insn); | |
1560 | pattern = PATTERN (src_insn); | |
1561 | ||
1562 | delete_insn (get_last_insn ()); | |
1563 | ||
1564 | goto retry; | |
1565 | } | |
1566 | ||
1567 | abort (); | |
1568 | } | |
67f2de41 | 1569 | |
6511b4f8 | 1570 | return increment_total; |
67f2de41 RK |
1571 | } |
1572 | ||
6bd43524 RS |
1573 | /* Copy REG_NOTES, except for insn references, because not all insn_map |
1574 | entries are valid yet. We do need to copy registers now though, because | |
1575 | the reg_map entries can change during copying. */ | |
1576 | ||
1577 | static rtx | |
1578 | initial_reg_note_copy (notes, map) | |
1579 | rtx notes; | |
1580 | struct inline_remap *map; | |
1581 | { | |
1582 | rtx copy; | |
1583 | ||
1584 | if (notes == 0) | |
1585 | return 0; | |
1586 | ||
1587 | copy = rtx_alloc (GET_CODE (notes)); | |
1588 | PUT_MODE (copy, GET_MODE (notes)); | |
1589 | ||
1590 | if (GET_CODE (notes) == EXPR_LIST) | |
1591 | XEXP (copy, 0) = copy_rtx_and_substitute (XEXP (notes, 0), map); | |
1592 | else if (GET_CODE (notes) == INSN_LIST) | |
1593 | /* Don't substitute for these yet. */ | |
1594 | XEXP (copy, 0) = XEXP (notes, 0); | |
1595 | else | |
1596 | abort (); | |
1597 | ||
1598 | XEXP (copy, 1) = initial_reg_note_copy (XEXP (notes, 1), map); | |
1599 | ||
1600 | return copy; | |
1601 | } | |
1602 | ||
1603 | /* Fixup insn references in copied REG_NOTES. */ | |
1604 | ||
1605 | static void | |
1606 | final_reg_note_copy (notes, map) | |
1607 | rtx notes; | |
1608 | struct inline_remap *map; | |
1609 | { | |
1610 | rtx note; | |
1611 | ||
1612 | for (note = notes; note; note = XEXP (note, 1)) | |
1613 | if (GET_CODE (note) == INSN_LIST) | |
1614 | XEXP (note, 0) = map->insn_map[INSN_UID (XEXP (note, 0))]; | |
1615 | } | |
67f2de41 RK |
1616 | |
1617 | /* Copy each instruction in the loop, substituting from map as appropriate. | |
1618 | This is very similar to a loop in expand_inline_function. */ | |
1619 | ||
1620 | static void | |
1621 | copy_loop_body (copy_start, copy_end, map, exit_label, last_iteration, | |
1622 | unroll_type, start_label, loop_end, insert_before, | |
1623 | copy_notes_from) | |
1624 | rtx copy_start, copy_end; | |
1625 | struct inline_remap *map; | |
87f60669 | 1626 | rtx exit_label; |
67f2de41 RK |
1627 | int last_iteration; |
1628 | enum unroll_types unroll_type; | |
1629 | rtx start_label, loop_end, insert_before, copy_notes_from; | |
1630 | { | |
1631 | rtx insn, pattern; | |
1632 | rtx tem, copy; | |
1633 | int dest_reg_was_split, i; | |
1634 | rtx cc0_insn = 0; | |
1635 | rtx final_label = 0; | |
1636 | rtx giv_inc, giv_dest_reg, giv_src_reg; | |
1637 | ||
1638 | /* If this isn't the last iteration, then map any references to the | |
1639 | start_label to final_label. Final label will then be emitted immediately | |
1640 | after the end of this loop body if it was ever used. | |
1641 | ||
1642 | If this is the last iteration, then map references to the start_label | |
1643 | to itself. */ | |
1644 | if (! last_iteration) | |
1645 | { | |
1646 | final_label = gen_label_rtx (); | |
1f3d3a31 JL |
1647 | set_label_in_map (map, CODE_LABEL_NUMBER (start_label), |
1648 | final_label); | |
67f2de41 RK |
1649 | } |
1650 | else | |
1f3d3a31 | 1651 | set_label_in_map (map, CODE_LABEL_NUMBER (start_label), start_label); |
67f2de41 RK |
1652 | |
1653 | start_sequence (); | |
1654 | ||
1655 | insn = copy_start; | |
1656 | do | |
1657 | { | |
1658 | insn = NEXT_INSN (insn); | |
1659 | ||
1660 | map->orig_asm_operands_vector = 0; | |
1661 | ||
1662 | switch (GET_CODE (insn)) | |
1663 | { | |
1664 | case INSN: | |
1665 | pattern = PATTERN (insn); | |
1666 | copy = 0; | |
1667 | giv_inc = 0; | |
1668 | ||
1669 | /* Check to see if this is a giv that has been combined with | |
1670 | some split address givs. (Combined in the sense that | |
1671 | `combine_givs' in loop.c has put two givs in the same register.) | |
1672 | In this case, we must search all givs based on the same biv to | |
1673 | find the address givs. Then split the address givs. | |
1674 | Do this before splitting the giv, since that may map the | |
1675 | SET_DEST to a new register. */ | |
1676 | ||
1677 | if (GET_CODE (pattern) == SET | |
1678 | && GET_CODE (SET_DEST (pattern)) == REG | |
1679 | && addr_combined_regs[REGNO (SET_DEST (pattern))]) | |
1680 | { | |
1681 | struct iv_class *bl; | |
1682 | struct induction *v, *tv; | |
1683 | int regno = REGNO (SET_DEST (pattern)); | |
1684 | ||
1685 | v = addr_combined_regs[REGNO (SET_DEST (pattern))]; | |
1686 | bl = reg_biv_class[REGNO (v->src_reg)]; | |
1687 | ||
1688 | /* Although the giv_inc amount is not needed here, we must call | |
1689 | calculate_giv_inc here since it might try to delete the | |
1690 | last insn emitted. If we wait until later to call it, | |
1691 | we might accidentally delete insns generated immediately | |
1692 | below by emit_unrolled_add. */ | |
1693 | ||
1694 | giv_inc = calculate_giv_inc (pattern, insn, regno); | |
1695 | ||
1696 | /* Now find all address giv's that were combined with this | |
1697 | giv 'v'. */ | |
1698 | for (tv = bl->giv; tv; tv = tv->next_iv) | |
1699 | if (tv->giv_type == DEST_ADDR && tv->same == v) | |
1700 | { | |
7085bad3 JW |
1701 | int this_giv_inc; |
1702 | ||
1703 | /* If this DEST_ADDR giv was not split, then ignore it. */ | |
1704 | if (*tv->location != tv->dest_reg) | |
1705 | continue; | |
2b4bd1bc JW |
1706 | |
1707 | /* Scale this_giv_inc if the multiplicative factors of | |
1708 | the two givs are different. */ | |
7085bad3 | 1709 | this_giv_inc = INTVAL (giv_inc); |
2b4bd1bc JW |
1710 | if (tv->mult_val != v->mult_val) |
1711 | this_giv_inc = (this_giv_inc / INTVAL (v->mult_val) | |
1712 | * INTVAL (tv->mult_val)); | |
1713 | ||
1714 | tv->dest_reg = plus_constant (tv->dest_reg, this_giv_inc); | |
67f2de41 RK |
1715 | *tv->location = tv->dest_reg; |
1716 | ||
1717 | if (last_iteration && unroll_type != UNROLL_COMPLETELY) | |
1718 | { | |
1719 | /* Must emit an insn to increment the split address | |
1720 | giv. Add in the const_adjust field in case there | |
1721 | was a constant eliminated from the address. */ | |
1722 | rtx value, dest_reg; | |
1723 | ||
1724 | /* tv->dest_reg will be either a bare register, | |
1725 | or else a register plus a constant. */ | |
1726 | if (GET_CODE (tv->dest_reg) == REG) | |
1727 | dest_reg = tv->dest_reg; | |
1728 | else | |
1729 | dest_reg = XEXP (tv->dest_reg, 0); | |
1730 | ||
8d092274 | 1731 | /* Check for shared address givs, and avoid |
9faa82d8 | 1732 | incrementing the shared pseudo reg more than |
8d092274 | 1733 | once. */ |
9ae8ffe7 | 1734 | if (! tv->same_insn && ! tv->shared) |
8d092274 JW |
1735 | { |
1736 | /* tv->dest_reg may actually be a (PLUS (REG) | |
1737 | (CONST)) here, so we must call plus_constant | |
1738 | to add the const_adjust amount before calling | |
1739 | emit_unrolled_add below. */ | |
1740 | value = plus_constant (tv->dest_reg, | |
1741 | tv->const_adjust); | |
1742 | ||
1743 | /* The constant could be too large for an add | |
1744 | immediate, so can't directly emit an insn | |
1745 | here. */ | |
1746 | emit_unrolled_add (dest_reg, XEXP (value, 0), | |
1747 | XEXP (value, 1)); | |
1748 | } | |
67f2de41 RK |
1749 | |
1750 | /* Reset the giv to be just the register again, in case | |
412dc348 RK |
1751 | it is used after the set we have just emitted. |
1752 | We must subtract the const_adjust factor added in | |
1753 | above. */ | |
1754 | tv->dest_reg = plus_constant (dest_reg, | |
1755 | - tv->const_adjust); | |
67f2de41 RK |
1756 | *tv->location = tv->dest_reg; |
1757 | } | |
1758 | } | |
1759 | } | |
1760 | ||
1761 | /* If this is a setting of a splittable variable, then determine | |
1762 | how to split the variable, create a new set based on this split, | |
1763 | and set up the reg_map so that later uses of the variable will | |
1764 | use the new split variable. */ | |
1765 | ||
1766 | dest_reg_was_split = 0; | |
1767 | ||
1768 | if (GET_CODE (pattern) == SET | |
1769 | && GET_CODE (SET_DEST (pattern)) == REG | |
1770 | && splittable_regs[REGNO (SET_DEST (pattern))]) | |
1771 | { | |
1772 | int regno = REGNO (SET_DEST (pattern)); | |
1773 | ||
1774 | dest_reg_was_split = 1; | |
1775 | ||
1776 | /* Compute the increment value for the giv, if it wasn't | |
1777 | already computed above. */ | |
1778 | ||
1779 | if (giv_inc == 0) | |
1780 | giv_inc = calculate_giv_inc (pattern, insn, regno); | |
1781 | giv_dest_reg = SET_DEST (pattern); | |
1782 | giv_src_reg = SET_DEST (pattern); | |
1783 | ||
1784 | if (unroll_type == UNROLL_COMPLETELY) | |
1785 | { | |
1786 | /* Completely unrolling the loop. Set the induction | |
1787 | variable to a known constant value. */ | |
1788 | ||
1789 | /* The value in splittable_regs may be an invariant | |
1790 | value, so we must use plus_constant here. */ | |
1791 | splittable_regs[regno] | |
1792 | = plus_constant (splittable_regs[regno], INTVAL (giv_inc)); | |
1793 | ||
1794 | if (GET_CODE (splittable_regs[regno]) == PLUS) | |
1795 | { | |
1796 | giv_src_reg = XEXP (splittable_regs[regno], 0); | |
1797 | giv_inc = XEXP (splittable_regs[regno], 1); | |
1798 | } | |
1799 | else | |
1800 | { | |
1801 | /* The splittable_regs value must be a REG or a | |
1802 | CONST_INT, so put the entire value in the giv_src_reg | |
1803 | variable. */ | |
1804 | giv_src_reg = splittable_regs[regno]; | |
1805 | giv_inc = const0_rtx; | |
1806 | } | |
1807 | } | |
1808 | else | |
1809 | { | |
1810 | /* Partially unrolling loop. Create a new pseudo | |
1811 | register for the iteration variable, and set it to | |
1812 | be a constant plus the original register. Except | |
1813 | on the last iteration, when the result has to | |
1814 | go back into the original iteration var register. */ | |
1815 | ||
1816 | /* Handle bivs which must be mapped to a new register | |
1817 | when split. This happens for bivs which need their | |
1818 | final value set before loop entry. The new register | |
1819 | for the biv was stored in the biv's first struct | |
1820 | induction entry by find_splittable_regs. */ | |
1821 | ||
1822 | if (regno < max_reg_before_loop | |
1823 | && reg_iv_type[regno] == BASIC_INDUCT) | |
1824 | { | |
1825 | giv_src_reg = reg_biv_class[regno]->biv->src_reg; | |
1826 | giv_dest_reg = giv_src_reg; | |
1827 | } | |
1828 | ||
1829 | #if 0 | |
1830 | /* If non-reduced/final-value givs were split, then | |
1831 | this would have to remap those givs also. See | |
1832 | find_splittable_regs. */ | |
1833 | #endif | |
1834 | ||
1835 | splittable_regs[regno] | |
c166a311 | 1836 | = GEN_INT (INTVAL (giv_inc) |
67f2de41 RK |
1837 | + INTVAL (splittable_regs[regno])); |
1838 | giv_inc = splittable_regs[regno]; | |
1839 | ||
1840 | /* Now split the induction variable by changing the dest | |
1841 | of this insn to a new register, and setting its | |
1842 | reg_map entry to point to this new register. | |
1843 | ||
1844 | If this is the last iteration, and this is the last insn | |
1845 | that will update the iv, then reuse the original dest, | |
1846 | to ensure that the iv will have the proper value when | |
1847 | the loop exits or repeats. | |
1848 | ||
1849 | Using splittable_regs_updates here like this is safe, | |
1850 | because it can only be greater than one if all | |
1851 | instructions modifying the iv are always executed in | |
1852 | order. */ | |
1853 | ||
1854 | if (! last_iteration | |
1855 | || (splittable_regs_updates[regno]-- != 1)) | |
1856 | { | |
1857 | tem = gen_reg_rtx (GET_MODE (giv_src_reg)); | |
1858 | giv_dest_reg = tem; | |
1859 | map->reg_map[regno] = tem; | |
9ae8ffe7 | 1860 | record_base_value (REGNO (tem), giv_src_reg); |
67f2de41 RK |
1861 | } |
1862 | else | |
1863 | map->reg_map[regno] = giv_src_reg; | |
1864 | } | |
1865 | ||
1866 | /* The constant being added could be too large for an add | |
1867 | immediate, so can't directly emit an insn here. */ | |
1868 | emit_unrolled_add (giv_dest_reg, giv_src_reg, giv_inc); | |
1869 | copy = get_last_insn (); | |
1870 | pattern = PATTERN (copy); | |
1871 | } | |
1872 | else | |
1873 | { | |
1874 | pattern = copy_rtx_and_substitute (pattern, map); | |
1875 | copy = emit_insn (pattern); | |
1876 | } | |
6bd43524 | 1877 | REG_NOTES (copy) = initial_reg_note_copy (REG_NOTES (insn), map); |
67f2de41 RK |
1878 | |
1879 | #ifdef HAVE_cc0 | |
1880 | /* If this insn is setting CC0, it may need to look at | |
1881 | the insn that uses CC0 to see what type of insn it is. | |
1882 | In that case, the call to recog via validate_change will | |
1883 | fail. So don't substitute constants here. Instead, | |
1884 | do it when we emit the following insn. | |
1885 | ||
1886 | For example, see the pyr.md file. That machine has signed and | |
1887 | unsigned compares. The compare patterns must check the | |
1888 | following branch insn to see which what kind of compare to | |
1889 | emit. | |
1890 | ||
1891 | If the previous insn set CC0, substitute constants on it as | |
1892 | well. */ | |
e65886db | 1893 | if (sets_cc0_p (PATTERN (copy)) != 0) |
67f2de41 RK |
1894 | cc0_insn = copy; |
1895 | else | |
1896 | { | |
1897 | if (cc0_insn) | |
1898 | try_constants (cc0_insn, map); | |
1899 | cc0_insn = 0; | |
1900 | try_constants (copy, map); | |
1901 | } | |
1902 | #else | |
1903 | try_constants (copy, map); | |
1904 | #endif | |
1905 | ||
1906 | /* Make split induction variable constants `permanent' since we | |
1907 | know there are no backward branches across iteration variable | |
1908 | settings which would invalidate this. */ | |
1909 | if (dest_reg_was_split) | |
1910 | { | |
1911 | int regno = REGNO (SET_DEST (pattern)); | |
1912 | ||
8dc0b179 RS |
1913 | if (regno < map->const_equiv_map_size |
1914 | && map->const_age_map[regno] == map->const_age) | |
67f2de41 RK |
1915 | map->const_age_map[regno] = -1; |
1916 | } | |
1917 | break; | |
1918 | ||
1919 | case JUMP_INSN: | |
483404b6 JW |
1920 | pattern = copy_rtx_and_substitute (PATTERN (insn), map); |
1921 | copy = emit_jump_insn (pattern); | |
6bd43524 | 1922 | REG_NOTES (copy) = initial_reg_note_copy (REG_NOTES (insn), map); |
483404b6 | 1923 | |
67f2de41 RK |
1924 | if (JUMP_LABEL (insn) == start_label && insn == copy_end |
1925 | && ! last_iteration) | |
1926 | { | |
1927 | /* This is a branch to the beginning of the loop; this is the | |
1928 | last insn being copied; and this is not the last iteration. | |
1929 | In this case, we want to change the original fall through | |
1930 | case to be a branch past the end of the loop, and the | |
1931 | original jump label case to fall_through. */ | |
a9d27cb2 | 1932 | |
ca72f752 RE |
1933 | if (invert_exp (pattern, copy)) |
1934 | { | |
1935 | if (! redirect_exp (&pattern, | |
1f3d3a31 JL |
1936 | get_label_from_map (map, |
1937 | CODE_LABEL_NUMBER | |
1938 | (JUMP_LABEL (insn))), | |
ca72f752 RE |
1939 | exit_label, copy)) |
1940 | abort (); | |
1941 | } | |
1942 | else | |
1943 | { | |
1944 | rtx jmp; | |
1945 | rtx lab = gen_label_rtx (); | |
9faa82d8 RK |
1946 | /* Can't do it by reversing the jump (probably because we |
1947 | couldn't reverse the conditions), so emit a new | |
ca72f752 RE |
1948 | jump_insn after COPY, and redirect the jump around |
1949 | that. */ | |
1950 | jmp = emit_jump_insn_after (gen_jump (exit_label), copy); | |
1951 | jmp = emit_barrier_after (jmp); | |
1952 | emit_label_after (lab, jmp); | |
1953 | LABEL_NUSES (lab) = 0; | |
1954 | if (! redirect_exp (&pattern, | |
1f3d3a31 JL |
1955 | get_label_from_map (map, |
1956 | CODE_LABEL_NUMBER | |
1957 | (JUMP_LABEL (insn))), | |
ca72f752 RE |
1958 | lab, copy)) |
1959 | abort (); | |
1960 | } | |
67f2de41 | 1961 | } |
67f2de41 RK |
1962 | |
1963 | #ifdef HAVE_cc0 | |
1964 | if (cc0_insn) | |
1965 | try_constants (cc0_insn, map); | |
1966 | cc0_insn = 0; | |
1967 | #endif | |
1968 | try_constants (copy, map); | |
1969 | ||
1970 | /* Set the jump label of COPY correctly to avoid problems with | |
1971 | later passes of unroll_loop, if INSN had jump label set. */ | |
1972 | if (JUMP_LABEL (insn)) | |
1973 | { | |
57467646 JW |
1974 | rtx label = 0; |
1975 | ||
67f2de41 RK |
1976 | /* Can't use the label_map for every insn, since this may be |
1977 | the backward branch, and hence the label was not mapped. */ | |
1978 | if (GET_CODE (pattern) == SET) | |
1979 | { | |
1980 | tem = SET_SRC (pattern); | |
1981 | if (GET_CODE (tem) == LABEL_REF) | |
57467646 | 1982 | label = XEXP (tem, 0); |
67f2de41 RK |
1983 | else if (GET_CODE (tem) == IF_THEN_ELSE) |
1984 | { | |
1985 | if (XEXP (tem, 1) != pc_rtx) | |
57467646 | 1986 | label = XEXP (XEXP (tem, 1), 0); |
67f2de41 | 1987 | else |
57467646 | 1988 | label = XEXP (XEXP (tem, 2), 0); |
67f2de41 | 1989 | } |
67f2de41 | 1990 | } |
57467646 JW |
1991 | |
1992 | if (label && GET_CODE (label) == CODE_LABEL) | |
1993 | JUMP_LABEL (copy) = label; | |
67f2de41 RK |
1994 | else |
1995 | { | |
1996 | /* An unrecognizable jump insn, probably the entry jump | |
1997 | for a switch statement. This label must have been mapped, | |
1998 | so just use the label_map to get the new jump label. */ | |
4ac75a4e | 1999 | JUMP_LABEL (copy) |
1f3d3a31 JL |
2000 | = get_label_from_map (map, |
2001 | CODE_LABEL_NUMBER (JUMP_LABEL (insn))); | |
67f2de41 RK |
2002 | } |
2003 | ||
2004 | /* If this is a non-local jump, then must increase the label | |
2005 | use count so that the label will not be deleted when the | |
2006 | original jump is deleted. */ | |
2007 | LABEL_NUSES (JUMP_LABEL (copy))++; | |
2008 | } | |
2009 | else if (GET_CODE (PATTERN (copy)) == ADDR_VEC | |
2010 | || GET_CODE (PATTERN (copy)) == ADDR_DIFF_VEC) | |
2011 | { | |
2012 | rtx pat = PATTERN (copy); | |
2013 | int diff_vec_p = GET_CODE (pat) == ADDR_DIFF_VEC; | |
2014 | int len = XVECLEN (pat, diff_vec_p); | |
2015 | int i; | |
2016 | ||
2017 | for (i = 0; i < len; i++) | |
2018 | LABEL_NUSES (XEXP (XVECEXP (pat, diff_vec_p, i), 0))++; | |
2019 | } | |
2020 | ||
2021 | /* If this used to be a conditional jump insn but whose branch | |
2022 | direction is now known, we must do something special. */ | |
2023 | if (condjump_p (insn) && !simplejump_p (insn) && map->last_pc_value) | |
2024 | { | |
2025 | #ifdef HAVE_cc0 | |
2026 | /* The previous insn set cc0 for us. So delete it. */ | |
2027 | delete_insn (PREV_INSN (copy)); | |
2028 | #endif | |
2029 | ||
2030 | /* If this is now a no-op, delete it. */ | |
2031 | if (map->last_pc_value == pc_rtx) | |
2032 | { | |
648a09ee JW |
2033 | /* Don't let delete_insn delete the label referenced here, |
2034 | because we might possibly need it later for some other | |
2035 | instruction in the loop. */ | |
2036 | if (JUMP_LABEL (copy)) | |
2037 | LABEL_NUSES (JUMP_LABEL (copy))++; | |
67f2de41 | 2038 | delete_insn (copy); |
648a09ee JW |
2039 | if (JUMP_LABEL (copy)) |
2040 | LABEL_NUSES (JUMP_LABEL (copy))--; | |
67f2de41 RK |
2041 | copy = 0; |
2042 | } | |
2043 | else | |
2044 | /* Otherwise, this is unconditional jump so we must put a | |
2045 | BARRIER after it. We could do some dead code elimination | |
2046 | here, but jump.c will do it just as well. */ | |
2047 | emit_barrier (); | |
2048 | } | |
2049 | break; | |
2050 | ||
2051 | case CALL_INSN: | |
2052 | pattern = copy_rtx_and_substitute (PATTERN (insn), map); | |
2053 | copy = emit_call_insn (pattern); | |
6bd43524 | 2054 | REG_NOTES (copy) = initial_reg_note_copy (REG_NOTES (insn), map); |
67f2de41 | 2055 | |
2d98fe23 JW |
2056 | /* Because the USAGE information potentially contains objects other |
2057 | than hard registers, we need to copy it. */ | |
db3cf6fb MS |
2058 | CALL_INSN_FUNCTION_USAGE (copy) |
2059 | = copy_rtx_and_substitute (CALL_INSN_FUNCTION_USAGE (insn), map); | |
2d98fe23 | 2060 | |
67f2de41 RK |
2061 | #ifdef HAVE_cc0 |
2062 | if (cc0_insn) | |
2063 | try_constants (cc0_insn, map); | |
2064 | cc0_insn = 0; | |
2065 | #endif | |
2066 | try_constants (copy, map); | |
2067 | ||
2068 | /* Be lazy and assume CALL_INSNs clobber all hard registers. */ | |
2069 | for (i = 0; i < FIRST_PSEUDO_REGISTER; i++) | |
2070 | map->const_equiv_map[i] = 0; | |
2071 | break; | |
2072 | ||
2073 | case CODE_LABEL: | |
2074 | /* If this is the loop start label, then we don't need to emit a | |
2075 | copy of this label since no one will use it. */ | |
2076 | ||
2077 | if (insn != start_label) | |
2078 | { | |
1f3d3a31 JL |
2079 | copy = emit_label (get_label_from_map (map, |
2080 | CODE_LABEL_NUMBER (insn))); | |
67f2de41 RK |
2081 | map->const_age++; |
2082 | } | |
2083 | break; | |
2084 | ||
2085 | case BARRIER: | |
2086 | copy = emit_barrier (); | |
2087 | break; | |
2088 | ||
2089 | case NOTE: | |
715e6efb JW |
2090 | /* VTOP notes are valid only before the loop exit test. If placed |
2091 | anywhere else, loop may generate bad code. */ | |
2092 | ||
2093 | if (NOTE_LINE_NUMBER (insn) != NOTE_INSN_DELETED | |
2094 | && (NOTE_LINE_NUMBER (insn) != NOTE_INSN_LOOP_VTOP | |
2095 | || (last_iteration && unroll_type != UNROLL_COMPLETELY))) | |
67f2de41 RK |
2096 | copy = emit_note (NOTE_SOURCE_FILE (insn), |
2097 | NOTE_LINE_NUMBER (insn)); | |
2098 | else | |
2099 | copy = 0; | |
2100 | break; | |
2101 | ||
2102 | default: | |
2103 | abort (); | |
2104 | break; | |
2105 | } | |
2106 | ||
2107 | map->insn_map[INSN_UID (insn)] = copy; | |
2108 | } | |
2109 | while (insn != copy_end); | |
2110 | ||
6bd43524 | 2111 | /* Now finish coping the REG_NOTES. */ |
ef39bb95 JW |
2112 | insn = copy_start; |
2113 | do | |
2114 | { | |
2115 | insn = NEXT_INSN (insn); | |
2116 | if ((GET_CODE (insn) == INSN || GET_CODE (insn) == JUMP_INSN | |
2117 | || GET_CODE (insn) == CALL_INSN) | |
2118 | && map->insn_map[INSN_UID (insn)]) | |
6bd43524 | 2119 | final_reg_note_copy (REG_NOTES (map->insn_map[INSN_UID (insn)]), map); |
ef39bb95 JW |
2120 | } |
2121 | while (insn != copy_end); | |
2122 | ||
67f2de41 RK |
2123 | /* There may be notes between copy_notes_from and loop_end. Emit a copy of |
2124 | each of these notes here, since there may be some important ones, such as | |
2125 | NOTE_INSN_BLOCK_END notes, in this group. We don't do this on the last | |
2126 | iteration, because the original notes won't be deleted. | |
2127 | ||
2128 | We can't use insert_before here, because when from preconditioning, | |
2129 | insert_before points before the loop. We can't use copy_end, because | |
2130 | there may be insns already inserted after it (which we don't want to | |
2131 | copy) when not from preconditioning code. */ | |
2132 | ||
2133 | if (! last_iteration) | |
2134 | { | |
2135 | for (insn = copy_notes_from; insn != loop_end; insn = NEXT_INSN (insn)) | |
2136 | { | |
2137 | if (GET_CODE (insn) == NOTE | |
2138 | && NOTE_LINE_NUMBER (insn) != NOTE_INSN_DELETED) | |
2139 | emit_note (NOTE_SOURCE_FILE (insn), NOTE_LINE_NUMBER (insn)); | |
2140 | } | |
2141 | } | |
2142 | ||
2143 | if (final_label && LABEL_NUSES (final_label) > 0) | |
2144 | emit_label (final_label); | |
2145 | ||
2146 | tem = gen_sequence (); | |
2147 | end_sequence (); | |
2148 | emit_insn_before (tem, insert_before); | |
2149 | } | |
2150 | \f | |
2151 | /* Emit an insn, using the expand_binop to ensure that a valid insn is | |
2152 | emitted. This will correctly handle the case where the increment value | |
2153 | won't fit in the immediate field of a PLUS insns. */ | |
2154 | ||
2155 | void | |
2156 | emit_unrolled_add (dest_reg, src_reg, increment) | |
2157 | rtx dest_reg, src_reg, increment; | |
2158 | { | |
2159 | rtx result; | |
2160 | ||
2161 | result = expand_binop (GET_MODE (dest_reg), add_optab, src_reg, increment, | |
2162 | dest_reg, 0, OPTAB_LIB_WIDEN); | |
2163 | ||
2164 | if (dest_reg != result) | |
2165 | emit_move_insn (dest_reg, result); | |
2166 | } | |
2167 | \f | |
2168 | /* Searches the insns between INSN and LOOP_END. Returns 1 if there | |
2169 | is a backward branch in that range that branches to somewhere between | |
2170 | LOOP_START and INSN. Returns 0 otherwise. */ | |
2171 | ||
6dc42e49 | 2172 | /* ??? This is quadratic algorithm. Could be rewritten to be linear. |
67f2de41 RK |
2173 | In practice, this is not a problem, because this function is seldom called, |
2174 | and uses a negligible amount of CPU time on average. */ | |
2175 | ||
9923a30d | 2176 | int |
67f2de41 RK |
2177 | back_branch_in_range_p (insn, loop_start, loop_end) |
2178 | rtx insn; | |
2179 | rtx loop_start, loop_end; | |
2180 | { | |
2181 | rtx p, q, target_insn; | |
956d6950 | 2182 | rtx orig_loop_end = loop_end; |
67f2de41 RK |
2183 | |
2184 | /* Stop before we get to the backward branch at the end of the loop. */ | |
2185 | loop_end = prev_nonnote_insn (loop_end); | |
2186 | if (GET_CODE (loop_end) == BARRIER) | |
2187 | loop_end = PREV_INSN (loop_end); | |
2188 | ||
2189 | /* Check in case insn has been deleted, search forward for first non | |
2190 | deleted insn following it. */ | |
2191 | while (INSN_DELETED_P (insn)) | |
2192 | insn = NEXT_INSN (insn); | |
2193 | ||
956d6950 JL |
2194 | /* Check for the case where insn is the last insn in the loop. Deal |
2195 | with the case where INSN was a deleted loop test insn, in which case | |
2196 | it will now be the NOTE_LOOP_END. */ | |
2197 | if (insn == loop_end || insn == orig_loop_end) | |
67f2de41 RK |
2198 | return 0; |
2199 | ||
2200 | for (p = NEXT_INSN (insn); p != loop_end; p = NEXT_INSN (p)) | |
2201 | { | |
2202 | if (GET_CODE (p) == JUMP_INSN) | |
2203 | { | |
2204 | target_insn = JUMP_LABEL (p); | |
2205 | ||
2206 | /* Search from loop_start to insn, to see if one of them is | |
2207 | the target_insn. We can't use INSN_LUID comparisons here, | |
2208 | since insn may not have an LUID entry. */ | |
2209 | for (q = loop_start; q != insn; q = NEXT_INSN (q)) | |
2210 | if (q == target_insn) | |
2211 | return 1; | |
2212 | } | |
2213 | } | |
2214 | ||
2215 | return 0; | |
2216 | } | |
2217 | ||
2218 | /* Try to generate the simplest rtx for the expression | |
2219 | (PLUS (MULT mult1 mult2) add1). This is used to calculate the initial | |
2220 | value of giv's. */ | |
2221 | ||
2222 | static rtx | |
2223 | fold_rtx_mult_add (mult1, mult2, add1, mode) | |
2224 | rtx mult1, mult2, add1; | |
2225 | enum machine_mode mode; | |
2226 | { | |
2227 | rtx temp, mult_res; | |
2228 | rtx result; | |
2229 | ||
2230 | /* The modes must all be the same. This should always be true. For now, | |
2231 | check to make sure. */ | |
2232 | if ((GET_MODE (mult1) != mode && GET_MODE (mult1) != VOIDmode) | |
2233 | || (GET_MODE (mult2) != mode && GET_MODE (mult2) != VOIDmode) | |
2234 | || (GET_MODE (add1) != mode && GET_MODE (add1) != VOIDmode)) | |
2235 | abort (); | |
2236 | ||
2237 | /* Ensure that if at least one of mult1/mult2 are constant, then mult2 | |
2238 | will be a constant. */ | |
2239 | if (GET_CODE (mult1) == CONST_INT) | |
2240 | { | |
2241 | temp = mult2; | |
2242 | mult2 = mult1; | |
2243 | mult1 = temp; | |
2244 | } | |
2245 | ||
2246 | mult_res = simplify_binary_operation (MULT, mode, mult1, mult2); | |
2247 | if (! mult_res) | |
38a448ca | 2248 | mult_res = gen_rtx_MULT (mode, mult1, mult2); |
67f2de41 RK |
2249 | |
2250 | /* Again, put the constant second. */ | |
2251 | if (GET_CODE (add1) == CONST_INT) | |
2252 | { | |
2253 | temp = add1; | |
2254 | add1 = mult_res; | |
2255 | mult_res = temp; | |
2256 | } | |
2257 | ||
2258 | result = simplify_binary_operation (PLUS, mode, add1, mult_res); | |
2259 | if (! result) | |
38a448ca | 2260 | result = gen_rtx_PLUS (mode, add1, mult_res); |
67f2de41 RK |
2261 | |
2262 | return result; | |
2263 | } | |
2264 | ||
2265 | /* Searches the list of induction struct's for the biv BL, to try to calculate | |
2266 | the total increment value for one iteration of the loop as a constant. | |
2267 | ||
2268 | Returns the increment value as an rtx, simplified as much as possible, | |
2269 | if it can be calculated. Otherwise, returns 0. */ | |
2270 | ||
2271 | rtx | |
2272 | biv_total_increment (bl, loop_start, loop_end) | |
2273 | struct iv_class *bl; | |
2274 | rtx loop_start, loop_end; | |
2275 | { | |
2276 | struct induction *v; | |
2277 | rtx result; | |
2278 | ||
2279 | /* For increment, must check every instruction that sets it. Each | |
2280 | instruction must be executed only once each time through the loop. | |
2281 | To verify this, we check that the the insn is always executed, and that | |
2282 | there are no backward branches after the insn that branch to before it. | |
2283 | Also, the insn must have a mult_val of one (to make sure it really is | |
2284 | an increment). */ | |
2285 | ||
2286 | result = const0_rtx; | |
2287 | for (v = bl->biv; v; v = v->next_iv) | |
2288 | { | |
2289 | if (v->always_computable && v->mult_val == const1_rtx | |
2290 | && ! back_branch_in_range_p (v->insn, loop_start, loop_end)) | |
2291 | result = fold_rtx_mult_add (result, const1_rtx, v->add_val, v->mode); | |
2292 | else | |
2293 | return 0; | |
2294 | } | |
2295 | ||
2296 | return result; | |
2297 | } | |
2298 | ||
2299 | /* Determine the initial value of the iteration variable, and the amount | |
2300 | that it is incremented each loop. Use the tables constructed by | |
2301 | the strength reduction pass to calculate these values. | |
2302 | ||
2303 | Initial_value and/or increment are set to zero if their values could not | |
2304 | be calculated. */ | |
2305 | ||
8c660648 | 2306 | void |
67f2de41 RK |
2307 | iteration_info (iteration_var, initial_value, increment, loop_start, loop_end) |
2308 | rtx iteration_var, *initial_value, *increment; | |
2309 | rtx loop_start, loop_end; | |
2310 | { | |
2311 | struct iv_class *bl; | |
2312 | struct induction *v, *b; | |
2313 | ||
2314 | /* Clear the result values, in case no answer can be found. */ | |
2315 | *initial_value = 0; | |
2316 | *increment = 0; | |
2317 | ||
2318 | /* The iteration variable can be either a giv or a biv. Check to see | |
2319 | which it is, and compute the variable's initial value, and increment | |
2320 | value if possible. */ | |
2321 | ||
2322 | /* If this is a new register, can't handle it since we don't have any | |
2323 | reg_iv_type entry for it. */ | |
2324 | if (REGNO (iteration_var) >= max_reg_before_loop) | |
2325 | { | |
2326 | if (loop_dump_stream) | |
2327 | fprintf (loop_dump_stream, | |
2328 | "Loop unrolling: No reg_iv_type entry for iteration var.\n"); | |
2329 | return; | |
2330 | } | |
f6beed21 RK |
2331 | |
2332 | /* Reject iteration variables larger than the host wide int size, since they | |
67f2de41 | 2333 | could result in a number of iterations greater than the range of our |
f6beed21 RK |
2334 | `unsigned HOST_WIDE_INT' variable loop_n_iterations. */ |
2335 | else if ((GET_MODE_BITSIZE (GET_MODE (iteration_var)) | |
2336 | > HOST_BITS_PER_WIDE_INT)) | |
67f2de41 RK |
2337 | { |
2338 | if (loop_dump_stream) | |
2339 | fprintf (loop_dump_stream, | |
f6beed21 | 2340 | "Loop unrolling: Iteration var rejected because mode too large.\n"); |
67f2de41 RK |
2341 | return; |
2342 | } | |
2343 | else if (GET_MODE_CLASS (GET_MODE (iteration_var)) != MODE_INT) | |
2344 | { | |
2345 | if (loop_dump_stream) | |
2346 | fprintf (loop_dump_stream, | |
b4ac57ab | 2347 | "Loop unrolling: Iteration var not an integer.\n"); |
67f2de41 RK |
2348 | return; |
2349 | } | |
2350 | else if (reg_iv_type[REGNO (iteration_var)] == BASIC_INDUCT) | |
2351 | { | |
2352 | /* Grab initial value, only useful if it is a constant. */ | |
2353 | bl = reg_biv_class[REGNO (iteration_var)]; | |
2354 | *initial_value = bl->initial_value; | |
2355 | ||
2356 | *increment = biv_total_increment (bl, loop_start, loop_end); | |
2357 | } | |
2358 | else if (reg_iv_type[REGNO (iteration_var)] == GENERAL_INDUCT) | |
2359 | { | |
2360 | #if 1 | |
2361 | /* ??? The code below does not work because the incorrect number of | |
2362 | iterations is calculated when the biv is incremented after the giv | |
2363 | is set (which is the usual case). This can probably be accounted | |
2364 | for by biasing the initial_value by subtracting the amount of the | |
2365 | increment that occurs between the giv set and the giv test. However, | |
2366 | a giv as an iterator is very rare, so it does not seem worthwhile | |
2367 | to handle this. */ | |
2368 | /* ??? An example failure is: i = 6; do {;} while (i++ < 9). */ | |
2369 | if (loop_dump_stream) | |
2370 | fprintf (loop_dump_stream, | |
2371 | "Loop unrolling: Giv iterators are not handled.\n"); | |
2372 | return; | |
2373 | #else | |
2374 | /* Initial value is mult_val times the biv's initial value plus | |
2375 | add_val. Only useful if it is a constant. */ | |
2376 | v = reg_iv_info[REGNO (iteration_var)]; | |
2377 | bl = reg_biv_class[REGNO (v->src_reg)]; | |
2378 | *initial_value = fold_rtx_mult_add (v->mult_val, bl->initial_value, | |
2379 | v->add_val, v->mode); | |
2380 | ||
2381 | /* Increment value is mult_val times the increment value of the biv. */ | |
2382 | ||
2383 | *increment = biv_total_increment (bl, loop_start, loop_end); | |
2384 | if (*increment) | |
2385 | *increment = fold_rtx_mult_add (v->mult_val, *increment, const0_rtx, | |
2386 | v->mode); | |
2387 | #endif | |
2388 | } | |
2389 | else | |
2390 | { | |
2391 | if (loop_dump_stream) | |
2392 | fprintf (loop_dump_stream, | |
2393 | "Loop unrolling: Not basic or general induction var.\n"); | |
2394 | return; | |
2395 | } | |
2396 | } | |
2397 | ||
2398 | /* Calculate the approximate final value of the iteration variable | |
2399 | which has an loop exit test with code COMPARISON_CODE and comparison value | |
2400 | of COMPARISON_VALUE. Also returns an indication of whether the comparison | |
2401 | was signed or unsigned, and the direction of the comparison. This info is | |
2402 | needed to calculate the number of loop iterations. */ | |
2403 | ||
2404 | static rtx | |
2405 | approx_final_value (comparison_code, comparison_value, unsigned_p, compare_dir) | |
2406 | enum rtx_code comparison_code; | |
2407 | rtx comparison_value; | |
2408 | int *unsigned_p; | |
2409 | int *compare_dir; | |
2410 | { | |
2411 | /* Calculate the final value of the induction variable. | |
2412 | The exact final value depends on the branch operator, and increment sign. | |
2413 | This is only an approximate value. It will be wrong if the iteration | |
2414 | variable is not incremented by one each time through the loop, and | |
2415 | approx final value - start value % increment != 0. */ | |
2416 | ||
2417 | *unsigned_p = 0; | |
2418 | switch (comparison_code) | |
2419 | { | |
2420 | case LEU: | |
2421 | *unsigned_p = 1; | |
2422 | case LE: | |
2423 | *compare_dir = 1; | |
2424 | return plus_constant (comparison_value, 1); | |
2425 | case GEU: | |
2426 | *unsigned_p = 1; | |
2427 | case GE: | |
2428 | *compare_dir = -1; | |
2429 | return plus_constant (comparison_value, -1); | |
2430 | case EQ: | |
2431 | /* Can not calculate a final value for this case. */ | |
2432 | *compare_dir = 0; | |
2433 | return 0; | |
2434 | case LTU: | |
2435 | *unsigned_p = 1; | |
2436 | case LT: | |
2437 | *compare_dir = 1; | |
2438 | return comparison_value; | |
2439 | break; | |
2440 | case GTU: | |
2441 | *unsigned_p = 1; | |
2442 | case GT: | |
2443 | *compare_dir = -1; | |
2444 | return comparison_value; | |
2445 | case NE: | |
2446 | *compare_dir = 0; | |
2447 | return comparison_value; | |
2448 | default: | |
2449 | abort (); | |
2450 | } | |
2451 | } | |
2452 | ||
2453 | /* For each biv and giv, determine whether it can be safely split into | |
2454 | a different variable for each unrolled copy of the loop body. If it | |
2455 | is safe to split, then indicate that by saving some useful info | |
2456 | in the splittable_regs array. | |
2457 | ||
2458 | If the loop is being completely unrolled, then splittable_regs will hold | |
2459 | the current value of the induction variable while the loop is unrolled. | |
2460 | It must be set to the initial value of the induction variable here. | |
2461 | Otherwise, splittable_regs will hold the difference between the current | |
2462 | value of the induction variable and the value the induction variable had | |
8dc0b179 RS |
2463 | at the top of the loop. It must be set to the value 0 here. |
2464 | ||
2465 | Returns the total number of instructions that set registers that are | |
2466 | splittable. */ | |
67f2de41 RK |
2467 | |
2468 | /* ?? If the loop is only unrolled twice, then most of the restrictions to | |
2469 | constant values are unnecessary, since we can easily calculate increment | |
2470 | values in this case even if nothing is constant. The increment value | |
2471 | should not involve a multiply however. */ | |
2472 | ||
2473 | /* ?? Even if the biv/giv increment values aren't constant, it may still | |
2474 | be beneficial to split the variable if the loop is only unrolled a few | |
2475 | times, since multiplies by small integers (1,2,3,4) are very cheap. */ | |
2476 | ||
2477 | static int | |
2478 | find_splittable_regs (unroll_type, loop_start, loop_end, end_insert_before, | |
2479 | unroll_number) | |
2480 | enum unroll_types unroll_type; | |
2481 | rtx loop_start, loop_end; | |
2482 | rtx end_insert_before; | |
2483 | int unroll_number; | |
2484 | { | |
2485 | struct iv_class *bl; | |
0e91429a | 2486 | struct induction *v; |
67f2de41 RK |
2487 | rtx increment, tem; |
2488 | rtx biv_final_value; | |
2489 | int biv_splittable; | |
2490 | int result = 0; | |
2491 | ||
2492 | for (bl = loop_iv_list; bl; bl = bl->next) | |
2493 | { | |
2494 | /* Biv_total_increment must return a constant value, | |
2495 | otherwise we can not calculate the split values. */ | |
2496 | ||
2497 | increment = biv_total_increment (bl, loop_start, loop_end); | |
2498 | if (! increment || GET_CODE (increment) != CONST_INT) | |
2499 | continue; | |
2500 | ||
2501 | /* The loop must be unrolled completely, or else have a known number | |
2502 | of iterations and only one exit, or else the biv must be dead | |
2503 | outside the loop, or else the final value must be known. Otherwise, | |
2504 | it is unsafe to split the biv since it may not have the proper | |
2505 | value on loop exit. */ | |
2506 | ||
3669e646 | 2507 | /* loop_number_exit_count is non-zero if the loop has an exit other than |
67f2de41 RK |
2508 | a fall through at the end. */ |
2509 | ||
2510 | biv_splittable = 1; | |
2511 | biv_final_value = 0; | |
2512 | if (unroll_type != UNROLL_COMPLETELY | |
3669e646 | 2513 | && (loop_number_exit_count[uid_loop_num[INSN_UID (loop_start)]] |
67f2de41 | 2514 | || unroll_type == UNROLL_NAIVE) |
b1f21e0a | 2515 | && (uid_luid[REGNO_LAST_UID (bl->regno)] >= INSN_LUID (loop_end) |
67f2de41 RK |
2516 | || ! bl->init_insn |
2517 | || INSN_UID (bl->init_insn) >= max_uid_for_loop | |
b1f21e0a | 2518 | || (uid_luid[REGNO_FIRST_UID (bl->regno)] |
67f2de41 RK |
2519 | < INSN_LUID (bl->init_insn)) |
2520 | || reg_mentioned_p (bl->biv->dest_reg, SET_SRC (bl->init_set))) | |
2521 | && ! (biv_final_value = final_biv_value (bl, loop_start, loop_end))) | |
2522 | biv_splittable = 0; | |
2523 | ||
0e91429a RK |
2524 | /* If any of the insns setting the BIV don't do so with a simple |
2525 | PLUS, we don't know how to split it. */ | |
2526 | for (v = bl->biv; biv_splittable && v; v = v->next_iv) | |
2527 | if ((tem = single_set (v->insn)) == 0 | |
2528 | || GET_CODE (SET_DEST (tem)) != REG | |
2529 | || REGNO (SET_DEST (tem)) != bl->regno | |
2530 | || GET_CODE (SET_SRC (tem)) != PLUS) | |
2531 | biv_splittable = 0; | |
2532 | ||
67f2de41 RK |
2533 | /* If final value is non-zero, then must emit an instruction which sets |
2534 | the value of the biv to the proper value. This is done after | |
2535 | handling all of the givs, since some of them may need to use the | |
2536 | biv's value in their initialization code. */ | |
2537 | ||
2538 | /* This biv is splittable. If completely unrolling the loop, save | |
2539 | the biv's initial value. Otherwise, save the constant zero. */ | |
2540 | ||
2541 | if (biv_splittable == 1) | |
2542 | { | |
2543 | if (unroll_type == UNROLL_COMPLETELY) | |
2544 | { | |
2545 | /* If the initial value of the biv is itself (i.e. it is too | |
2546 | complicated for strength_reduce to compute), or is a hard | |
f47f2c17 RK |
2547 | register, or it isn't invariant, then we must create a new |
2548 | pseudo reg to hold the initial value of the biv. */ | |
67f2de41 RK |
2549 | |
2550 | if (GET_CODE (bl->initial_value) == REG | |
2551 | && (REGNO (bl->initial_value) == bl->regno | |
f47f2c17 RK |
2552 | || REGNO (bl->initial_value) < FIRST_PSEUDO_REGISTER |
2553 | || ! invariant_p (bl->initial_value))) | |
67f2de41 RK |
2554 | { |
2555 | rtx tem = gen_reg_rtx (bl->biv->mode); | |
9ae8ffe7 JL |
2556 | |
2557 | record_base_value (REGNO (tem), bl->biv->add_val); | |
67f2de41 RK |
2558 | emit_insn_before (gen_move_insn (tem, bl->biv->src_reg), |
2559 | loop_start); | |
2560 | ||
2561 | if (loop_dump_stream) | |
2562 | fprintf (loop_dump_stream, "Biv %d initial value remapped to %d.\n", | |
2563 | bl->regno, REGNO (tem)); | |
2564 | ||
2565 | splittable_regs[bl->regno] = tem; | |
2566 | } | |
2567 | else | |
2568 | splittable_regs[bl->regno] = bl->initial_value; | |
2569 | } | |
2570 | else | |
2571 | splittable_regs[bl->regno] = const0_rtx; | |
2572 | ||
2573 | /* Save the number of instructions that modify the biv, so that | |
2574 | we can treat the last one specially. */ | |
2575 | ||
2576 | splittable_regs_updates[bl->regno] = bl->biv_count; | |
8dc0b179 | 2577 | result += bl->biv_count; |
67f2de41 RK |
2578 | |
2579 | if (loop_dump_stream) | |
2580 | fprintf (loop_dump_stream, | |
2581 | "Biv %d safe to split.\n", bl->regno); | |
2582 | } | |
2583 | ||
2584 | /* Check every giv that depends on this biv to see whether it is | |
2585 | splittable also. Even if the biv isn't splittable, givs which | |
2586 | depend on it may be splittable if the biv is live outside the | |
2587 | loop, and the givs aren't. */ | |
2588 | ||
8dc0b179 RS |
2589 | result += find_splittable_givs (bl, unroll_type, loop_start, loop_end, |
2590 | increment, unroll_number); | |
67f2de41 RK |
2591 | |
2592 | /* If final value is non-zero, then must emit an instruction which sets | |
2593 | the value of the biv to the proper value. This is done after | |
2594 | handling all of the givs, since some of them may need to use the | |
2595 | biv's value in their initialization code. */ | |
2596 | if (biv_final_value) | |
2597 | { | |
2598 | /* If the loop has multiple exits, emit the insns before the | |
2599 | loop to ensure that it will always be executed no matter | |
2600 | how the loop exits. Otherwise emit the insn after the loop, | |
2601 | since this is slightly more efficient. */ | |
3669e646 | 2602 | if (! loop_number_exit_count[uid_loop_num[INSN_UID (loop_start)]]) |
67f2de41 RK |
2603 | emit_insn_before (gen_move_insn (bl->biv->src_reg, |
2604 | biv_final_value), | |
2605 | end_insert_before); | |
2606 | else | |
2607 | { | |
2608 | /* Create a new register to hold the value of the biv, and then | |
2609 | set the biv to its final value before the loop start. The biv | |
2610 | is set to its final value before loop start to ensure that | |
2611 | this insn will always be executed, no matter how the loop | |
2612 | exits. */ | |
2613 | rtx tem = gen_reg_rtx (bl->biv->mode); | |
9ae8ffe7 JL |
2614 | record_base_value (REGNO (tem), bl->biv->add_val); |
2615 | ||
67f2de41 RK |
2616 | emit_insn_before (gen_move_insn (tem, bl->biv->src_reg), |
2617 | loop_start); | |
2618 | emit_insn_before (gen_move_insn (bl->biv->src_reg, | |
2619 | biv_final_value), | |
2620 | loop_start); | |
2621 | ||
2622 | if (loop_dump_stream) | |
2623 | fprintf (loop_dump_stream, "Biv %d mapped to %d for split.\n", | |
2624 | REGNO (bl->biv->src_reg), REGNO (tem)); | |
2625 | ||
2626 | /* Set up the mapping from the original biv register to the new | |
2627 | register. */ | |
2628 | bl->biv->src_reg = tem; | |
2629 | } | |
2630 | } | |
2631 | } | |
2632 | return result; | |
2633 | } | |
2634 | ||
d8191d86 RK |
2635 | /* Return 1 if the first and last unrolled copy of the address giv V is valid |
2636 | for the instruction that is using it. Do not make any changes to that | |
2637 | instruction. */ | |
2638 | ||
2639 | static int | |
2640 | verify_addresses (v, giv_inc, unroll_number) | |
2641 | struct induction *v; | |
2642 | rtx giv_inc; | |
2643 | int unroll_number; | |
2644 | { | |
2645 | int ret = 1; | |
2646 | rtx orig_addr = *v->location; | |
2647 | rtx last_addr = plus_constant (v->dest_reg, | |
2648 | INTVAL (giv_inc) * (unroll_number - 1)); | |
2649 | ||
2650 | /* First check to see if either address would fail. */ | |
2651 | if (! validate_change (v->insn, v->location, v->dest_reg, 0) | |
2652 | || ! validate_change (v->insn, v->location, last_addr, 0)) | |
2653 | ret = 0; | |
2654 | ||
2655 | /* Now put things back the way they were before. This will always | |
2656 | succeed. */ | |
2657 | validate_change (v->insn, v->location, orig_addr, 0); | |
2658 | ||
2659 | return ret; | |
2660 | } | |
2661 | ||
67f2de41 | 2662 | /* For every giv based on the biv BL, check to determine whether it is |
8dc0b179 RS |
2663 | splittable. This is a subroutine to find_splittable_regs (). |
2664 | ||
2665 | Return the number of instructions that set splittable registers. */ | |
67f2de41 RK |
2666 | |
2667 | static int | |
2668 | find_splittable_givs (bl, unroll_type, loop_start, loop_end, increment, | |
8dc0b179 | 2669 | unroll_number) |
67f2de41 RK |
2670 | struct iv_class *bl; |
2671 | enum unroll_types unroll_type; | |
2672 | rtx loop_start, loop_end; | |
2673 | rtx increment; | |
8dc0b179 | 2674 | int unroll_number; |
67f2de41 | 2675 | { |
3fc347fa | 2676 | struct induction *v, *v2; |
67f2de41 RK |
2677 | rtx final_value; |
2678 | rtx tem; | |
8dc0b179 | 2679 | int result = 0; |
67f2de41 | 2680 | |
3fc347fa JW |
2681 | /* Scan the list of givs, and set the same_insn field when there are |
2682 | multiple identical givs in the same insn. */ | |
2683 | for (v = bl->giv; v; v = v->next_iv) | |
2684 | for (v2 = v->next_iv; v2; v2 = v2->next_iv) | |
2685 | if (v->insn == v2->insn && rtx_equal_p (v->new_reg, v2->new_reg) | |
2686 | && ! v2->same_insn) | |
2687 | v2->same_insn = v; | |
2688 | ||
67f2de41 RK |
2689 | for (v = bl->giv; v; v = v->next_iv) |
2690 | { | |
2691 | rtx giv_inc, value; | |
2692 | ||
2693 | /* Only split the giv if it has already been reduced, or if the loop is | |
2694 | being completely unrolled. */ | |
2695 | if (unroll_type != UNROLL_COMPLETELY && v->ignore) | |
2696 | continue; | |
2697 | ||
2698 | /* The giv can be split if the insn that sets the giv is executed once | |
2699 | and only once on every iteration of the loop. */ | |
2700 | /* An address giv can always be split. v->insn is just a use not a set, | |
2701 | and hence it does not matter whether it is always executed. All that | |
2702 | matters is that all the biv increments are always executed, and we | |
2703 | won't reach here if they aren't. */ | |
2704 | if (v->giv_type != DEST_ADDR | |
2705 | && (! v->always_computable | |
2706 | || back_branch_in_range_p (v->insn, loop_start, loop_end))) | |
2707 | continue; | |
2708 | ||
2709 | /* The giv increment value must be a constant. */ | |
2710 | giv_inc = fold_rtx_mult_add (v->mult_val, increment, const0_rtx, | |
2711 | v->mode); | |
2712 | if (! giv_inc || GET_CODE (giv_inc) != CONST_INT) | |
2713 | continue; | |
2714 | ||
2715 | /* The loop must be unrolled completely, or else have a known number of | |
2716 | iterations and only one exit, or else the giv must be dead outside | |
2717 | the loop, or else the final value of the giv must be known. | |
2718 | Otherwise, it is not safe to split the giv since it may not have the | |
2719 | proper value on loop exit. */ | |
2720 | ||
2721 | /* The used outside loop test will fail for DEST_ADDR givs. They are | |
2722 | never used outside the loop anyways, so it is always safe to split a | |
2723 | DEST_ADDR giv. */ | |
2724 | ||
2725 | final_value = 0; | |
2726 | if (unroll_type != UNROLL_COMPLETELY | |
3669e646 | 2727 | && (loop_number_exit_count[uid_loop_num[INSN_UID (loop_start)]] |
67f2de41 RK |
2728 | || unroll_type == UNROLL_NAIVE) |
2729 | && v->giv_type != DEST_ADDR | |
9fb82071 JW |
2730 | /* The next part is true if the pseudo is used outside the loop. |
2731 | We assume that this is true for any pseudo created after loop | |
2732 | starts, because we don't have a reg_n_info entry for them. */ | |
2733 | && (REGNO (v->dest_reg) >= max_reg_before_loop | |
2734 | || (REGNO_FIRST_UID (REGNO (v->dest_reg)) != INSN_UID (v->insn) | |
2735 | /* Check for the case where the pseudo is set by a shift/add | |
2736 | sequence, in which case the first insn setting the pseudo | |
2737 | is the first insn of the shift/add sequence. */ | |
2738 | && (! (tem = find_reg_note (v->insn, REG_RETVAL, NULL_RTX)) | |
2739 | || (REGNO_FIRST_UID (REGNO (v->dest_reg)) | |
2740 | != INSN_UID (XEXP (tem, 0))))) | |
67f2de41 | 2741 | /* Line above always fails if INSN was moved by loop opt. */ |
b1f21e0a | 2742 | || (uid_luid[REGNO_LAST_UID (REGNO (v->dest_reg))] |
67f2de41 RK |
2743 | >= INSN_LUID (loop_end))) |
2744 | && ! (final_value = v->final_value)) | |
2745 | continue; | |
2746 | ||
2747 | #if 0 | |
2748 | /* Currently, non-reduced/final-value givs are never split. */ | |
2749 | /* Should emit insns after the loop if possible, as the biv final value | |
2750 | code below does. */ | |
2751 | ||
2752 | /* If the final value is non-zero, and the giv has not been reduced, | |
2753 | then must emit an instruction to set the final value. */ | |
2754 | if (final_value && !v->new_reg) | |
2755 | { | |
2756 | /* Create a new register to hold the value of the giv, and then set | |
2757 | the giv to its final value before the loop start. The giv is set | |
2758 | to its final value before loop start to ensure that this insn | |
2759 | will always be executed, no matter how we exit. */ | |
2760 | tem = gen_reg_rtx (v->mode); | |
2761 | emit_insn_before (gen_move_insn (tem, v->dest_reg), loop_start); | |
2762 | emit_insn_before (gen_move_insn (v->dest_reg, final_value), | |
2763 | loop_start); | |
2764 | ||
2765 | if (loop_dump_stream) | |
2766 | fprintf (loop_dump_stream, "Giv %d mapped to %d for split.\n", | |
2767 | REGNO (v->dest_reg), REGNO (tem)); | |
2768 | ||
2769 | v->src_reg = tem; | |
2770 | } | |
2771 | #endif | |
2772 | ||
2773 | /* This giv is splittable. If completely unrolling the loop, save the | |
2774 | giv's initial value. Otherwise, save the constant zero for it. */ | |
2775 | ||
2776 | if (unroll_type == UNROLL_COMPLETELY) | |
0e91429a RK |
2777 | { |
2778 | /* It is not safe to use bl->initial_value here, because it may not | |
2779 | be invariant. It is safe to use the initial value stored in | |
2780 | the splittable_regs array if it is set. In rare cases, it won't | |
2781 | be set, so then we do exactly the same thing as | |
2782 | find_splittable_regs does to get a safe value. */ | |
2783 | rtx biv_initial_value; | |
2784 | ||
2785 | if (splittable_regs[bl->regno]) | |
2786 | biv_initial_value = splittable_regs[bl->regno]; | |
2787 | else if (GET_CODE (bl->initial_value) != REG | |
2788 | || (REGNO (bl->initial_value) != bl->regno | |
2789 | && REGNO (bl->initial_value) >= FIRST_PSEUDO_REGISTER)) | |
2790 | biv_initial_value = bl->initial_value; | |
2791 | else | |
2792 | { | |
2793 | rtx tem = gen_reg_rtx (bl->biv->mode); | |
2794 | ||
9ae8ffe7 | 2795 | record_base_value (REGNO (tem), bl->biv->add_val); |
0e91429a RK |
2796 | emit_insn_before (gen_move_insn (tem, bl->biv->src_reg), |
2797 | loop_start); | |
2798 | biv_initial_value = tem; | |
2799 | } | |
2800 | value = fold_rtx_mult_add (v->mult_val, biv_initial_value, | |
2801 | v->add_val, v->mode); | |
2802 | } | |
67f2de41 RK |
2803 | else |
2804 | value = const0_rtx; | |
2805 | ||
2806 | if (v->new_reg) | |
2807 | { | |
bf1c6940 JW |
2808 | /* If a giv was combined with another giv, then we can only split |
2809 | this giv if the giv it was combined with was reduced. This | |
2810 | is because the value of v->new_reg is meaningless in this | |
2811 | case. */ | |
2812 | if (v->same && ! v->same->new_reg) | |
67f2de41 RK |
2813 | { |
2814 | if (loop_dump_stream) | |
2815 | fprintf (loop_dump_stream, | |
bf1c6940 JW |
2816 | "giv combined with unreduced giv not split.\n"); |
2817 | continue; | |
67f2de41 | 2818 | } |
bf1c6940 JW |
2819 | /* If the giv is an address destination, it could be something other |
2820 | than a simple register, these have to be treated differently. */ | |
2821 | else if (v->giv_type == DEST_REG) | |
2b9a9aea JW |
2822 | { |
2823 | /* If value is not a constant, register, or register plus | |
2824 | constant, then compute its value into a register before | |
181c6568 | 2825 | loop start. This prevents invalid rtx sharing, and should |
2b9a9aea JW |
2826 | generate better code. We can use bl->initial_value here |
2827 | instead of splittable_regs[bl->regno] because this code | |
2828 | is going before the loop start. */ | |
2829 | if (unroll_type == UNROLL_COMPLETELY | |
2830 | && GET_CODE (value) != CONST_INT | |
2831 | && GET_CODE (value) != REG | |
2832 | && (GET_CODE (value) != PLUS | |
2833 | || GET_CODE (XEXP (value, 0)) != REG | |
2834 | || GET_CODE (XEXP (value, 1)) != CONST_INT)) | |
2835 | { | |
2836 | rtx tem = gen_reg_rtx (v->mode); | |
9ae8ffe7 | 2837 | record_base_value (REGNO (tem), v->add_val); |
2b9a9aea JW |
2838 | emit_iv_add_mult (bl->initial_value, v->mult_val, |
2839 | v->add_val, tem, loop_start); | |
2840 | value = tem; | |
2841 | } | |
2842 | ||
2843 | splittable_regs[REGNO (v->new_reg)] = value; | |
2844 | } | |
67f2de41 RK |
2845 | else |
2846 | { | |
2847 | /* Splitting address givs is useful since it will often allow us | |
2848 | to eliminate some increment insns for the base giv as | |
2849 | unnecessary. */ | |
2850 | ||
2851 | /* If the addr giv is combined with a dest_reg giv, then all | |
2852 | references to that dest reg will be remapped, which is NOT | |
2853 | what we want for split addr regs. We always create a new | |
2854 | register for the split addr giv, just to be safe. */ | |
2855 | ||
9ae8ffe7 JL |
2856 | /* If we have multiple identical address givs within a |
2857 | single instruction, then use a single pseudo reg for | |
2858 | both. This is necessary in case one is a match_dup | |
8d092274 JW |
2859 | of the other. */ |
2860 | ||
67f2de41 | 2861 | v->const_adjust = 0; |
8d092274 | 2862 | |
3fc347fa | 2863 | if (v->same_insn) |
8d092274 | 2864 | { |
3fc347fa | 2865 | v->dest_reg = v->same_insn->dest_reg; |
8d092274 JW |
2866 | if (loop_dump_stream) |
2867 | fprintf (loop_dump_stream, | |
3fc347fa JW |
2868 | "Sharing address givs in insn %d\n", |
2869 | INSN_UID (v->insn)); | |
8d092274 | 2870 | } |
9ae8ffe7 JL |
2871 | /* If multiple address GIVs have been combined with the |
2872 | same dest_reg GIV, do not create a new register for | |
2873 | each. */ | |
2874 | else if (unroll_type != UNROLL_COMPLETELY | |
2875 | && v->giv_type == DEST_ADDR | |
2876 | && v->same && v->same->giv_type == DEST_ADDR | |
b6ffe602 JC |
2877 | && v->same->unrolled |
2878 | /* combine_givs_p may return true for some cases | |
2879 | where the add and mult values are not equal. | |
2880 | To share a register here, the values must be | |
2881 | equal. */ | |
2882 | && rtx_equal_p (v->same->mult_val, v->mult_val) | |
2883 | && rtx_equal_p (v->same->add_val, v->add_val)) | |
2884 | ||
9ae8ffe7 JL |
2885 | { |
2886 | v->dest_reg = v->same->dest_reg; | |
2887 | v->shared = 1; | |
2888 | } | |
8d092274 | 2889 | else if (unroll_type != UNROLL_COMPLETELY) |
67f2de41 RK |
2890 | { |
2891 | /* If not completely unrolling the loop, then create a new | |
2892 | register to hold the split value of the DEST_ADDR giv. | |
2893 | Emit insn to initialize its value before loop start. */ | |
9ae8ffe7 JL |
2894 | |
2895 | rtx tem = gen_reg_rtx (v->mode); | |
2896 | record_base_value (REGNO (tem), v->add_val); | |
2897 | v->unrolled = 1; | |
67f2de41 RK |
2898 | |
2899 | /* If the address giv has a constant in its new_reg value, | |
2900 | then this constant can be pulled out and put in value, | |
2901 | instead of being part of the initialization code. */ | |
2902 | ||
2903 | if (GET_CODE (v->new_reg) == PLUS | |
2904 | && GET_CODE (XEXP (v->new_reg, 1)) == CONST_INT) | |
2905 | { | |
2906 | v->dest_reg | |
2907 | = plus_constant (tem, INTVAL (XEXP (v->new_reg,1))); | |
9ae8ffe7 | 2908 | |
67f2de41 RK |
2909 | /* Only succeed if this will give valid addresses. |
2910 | Try to validate both the first and the last | |
2911 | address resulting from loop unrolling, if | |
2912 | one fails, then can't do const elim here. */ | |
7085bad3 | 2913 | if (verify_addresses (v, giv_inc, unroll_number)) |
67f2de41 RK |
2914 | { |
2915 | /* Save the negative of the eliminated const, so | |
2916 | that we can calculate the dest_reg's increment | |
2917 | value later. */ | |
2918 | v->const_adjust = - INTVAL (XEXP (v->new_reg, 1)); | |
2919 | ||
2920 | v->new_reg = XEXP (v->new_reg, 0); | |
2921 | if (loop_dump_stream) | |
2922 | fprintf (loop_dump_stream, | |
2923 | "Eliminating constant from giv %d\n", | |
2924 | REGNO (tem)); | |
2925 | } | |
2926 | else | |
2927 | v->dest_reg = tem; | |
2928 | } | |
2929 | else | |
2930 | v->dest_reg = tem; | |
2931 | ||
2932 | /* If the address hasn't been checked for validity yet, do so | |
2933 | now, and fail completely if either the first or the last | |
d8191d86 RK |
2934 | unrolled copy of the address is not a valid address |
2935 | for the instruction that uses it. */ | |
67f2de41 | 2936 | if (v->dest_reg == tem |
d8191d86 | 2937 | && ! verify_addresses (v, giv_inc, unroll_number)) |
67f2de41 RK |
2938 | { |
2939 | if (loop_dump_stream) | |
2940 | fprintf (loop_dump_stream, | |
181c6568 | 2941 | "Invalid address for giv at insn %d\n", |
67f2de41 RK |
2942 | INSN_UID (v->insn)); |
2943 | continue; | |
2944 | } | |
2945 | ||
2946 | /* To initialize the new register, just move the value of | |
2947 | new_reg into it. This is not guaranteed to give a valid | |
2948 | instruction on machines with complex addressing modes. | |
2949 | If we can't recognize it, then delete it and emit insns | |
2950 | to calculate the value from scratch. */ | |
38a448ca RH |
2951 | emit_insn_before (gen_rtx_SET (VOIDmode, tem, |
2952 | copy_rtx (v->new_reg)), | |
67f2de41 | 2953 | loop_start); |
9fdc1dca | 2954 | if (recog_memoized (PREV_INSN (loop_start)) < 0) |
67f2de41 | 2955 | { |
75540fea JW |
2956 | rtx sequence, ret; |
2957 | ||
2958 | /* We can't use bl->initial_value to compute the initial | |
2959 | value, because the loop may have been preconditioned. | |
2960 | We must calculate it from NEW_REG. Try using | |
2961 | force_operand instead of emit_iv_add_mult. */ | |
67f2de41 | 2962 | delete_insn (PREV_INSN (loop_start)); |
75540fea JW |
2963 | |
2964 | start_sequence (); | |
2965 | ret = force_operand (v->new_reg, tem); | |
2966 | if (ret != tem) | |
2967 | emit_move_insn (tem, ret); | |
2968 | sequence = gen_sequence (); | |
2969 | end_sequence (); | |
2970 | emit_insn_before (sequence, loop_start); | |
2971 | ||
67f2de41 RK |
2972 | if (loop_dump_stream) |
2973 | fprintf (loop_dump_stream, | |
181c6568 | 2974 | "Invalid init insn, rewritten.\n"); |
67f2de41 RK |
2975 | } |
2976 | } | |
2977 | else | |
2978 | { | |
2979 | v->dest_reg = value; | |
2980 | ||
2981 | /* Check the resulting address for validity, and fail | |
181c6568 | 2982 | if the resulting address would be invalid. */ |
d8191d86 | 2983 | if (! verify_addresses (v, giv_inc, unroll_number)) |
67f2de41 RK |
2984 | { |
2985 | if (loop_dump_stream) | |
2986 | fprintf (loop_dump_stream, | |
181c6568 | 2987 | "Invalid address for giv at insn %d\n", |
67f2de41 RK |
2988 | INSN_UID (v->insn)); |
2989 | continue; | |
2990 | } | |
2991 | } | |
2992 | ||
2993 | /* Store the value of dest_reg into the insn. This sharing | |
2994 | will not be a problem as this insn will always be copied | |
2995 | later. */ | |
2996 | ||
2997 | *v->location = v->dest_reg; | |
2998 | ||
2999 | /* If this address giv is combined with a dest reg giv, then | |
3000 | save the base giv's induction pointer so that we will be | |
3001 | able to handle this address giv properly. The base giv | |
3002 | itself does not have to be splittable. */ | |
3003 | ||
3004 | if (v->same && v->same->giv_type == DEST_REG) | |
3005 | addr_combined_regs[REGNO (v->same->new_reg)] = v->same; | |
3006 | ||
3007 | if (GET_CODE (v->new_reg) == REG) | |
3008 | { | |
3009 | /* This giv maybe hasn't been combined with any others. | |
3010 | Make sure that it's giv is marked as splittable here. */ | |
3011 | ||
3012 | splittable_regs[REGNO (v->new_reg)] = value; | |
3013 | ||
3014 | /* Make it appear to depend upon itself, so that the | |
3015 | giv will be properly split in the main loop above. */ | |
3016 | if (! v->same) | |
3017 | { | |
3018 | v->same = v; | |
3019 | addr_combined_regs[REGNO (v->new_reg)] = v; | |
3020 | } | |
3021 | } | |
3f07e47a | 3022 | |
67f2de41 RK |
3023 | if (loop_dump_stream) |
3024 | fprintf (loop_dump_stream, "DEST_ADDR giv being split.\n"); | |
3025 | } | |
3026 | } | |
3027 | else | |
3028 | { | |
3029 | #if 0 | |
3030 | /* Currently, unreduced giv's can't be split. This is not too much | |
3031 | of a problem since unreduced giv's are not live across loop | |
3032 | iterations anyways. When unrolling a loop completely though, | |
3033 | it makes sense to reduce&split givs when possible, as this will | |
3034 | result in simpler instructions, and will not require that a reg | |
3035 | be live across loop iterations. */ | |
3036 | ||
3037 | splittable_regs[REGNO (v->dest_reg)] = value; | |
3038 | fprintf (stderr, "Giv %d at insn %d not reduced\n", | |
3039 | REGNO (v->dest_reg), INSN_UID (v->insn)); | |
3040 | #else | |
3041 | continue; | |
3042 | #endif | |
3043 | } | |
3044 | ||
3bb24246 JW |
3045 | /* Unreduced givs are only updated once by definition. Reduced givs |
3046 | are updated as many times as their biv is. Mark it so if this is | |
67f2de41 RK |
3047 | a splittable register. Don't need to do anything for address givs |
3048 | where this may not be a register. */ | |
3049 | ||
3050 | if (GET_CODE (v->new_reg) == REG) | |
3bb24246 JW |
3051 | { |
3052 | int count = 1; | |
3053 | if (! v->ignore) | |
3054 | count = reg_biv_class[REGNO (v->src_reg)]->biv_count; | |
3055 | ||
3056 | splittable_regs_updates[REGNO (v->new_reg)] = count; | |
3057 | } | |
67f2de41 RK |
3058 | |
3059 | result++; | |
3060 | ||
3061 | if (loop_dump_stream) | |
3062 | { | |
3063 | int regnum; | |
3064 | ||
3065 | if (GET_CODE (v->dest_reg) == CONST_INT) | |
3066 | regnum = -1; | |
3067 | else if (GET_CODE (v->dest_reg) != REG) | |
3068 | regnum = REGNO (XEXP (v->dest_reg, 0)); | |
3069 | else | |
3070 | regnum = REGNO (v->dest_reg); | |
3071 | fprintf (loop_dump_stream, "Giv %d at insn %d safe to split.\n", | |
3072 | regnum, INSN_UID (v->insn)); | |
3073 | } | |
3074 | } | |
3075 | ||
3076 | return result; | |
3077 | } | |
3078 | \f | |
3079 | /* Try to prove that the register is dead after the loop exits. Trace every | |
3080 | loop exit looking for an insn that will always be executed, which sets | |
3081 | the register to some value, and appears before the first use of the register | |
3082 | is found. If successful, then return 1, otherwise return 0. */ | |
3083 | ||
3084 | /* ?? Could be made more intelligent in the handling of jumps, so that | |
3085 | it can search past if statements and other similar structures. */ | |
3086 | ||
3087 | static int | |
3088 | reg_dead_after_loop (reg, loop_start, loop_end) | |
3089 | rtx reg, loop_start, loop_end; | |
3090 | { | |
3091 | rtx insn, label; | |
3092 | enum rtx_code code; | |
412dc348 | 3093 | int jump_count = 0; |
3669e646 RK |
3094 | int label_count = 0; |
3095 | int this_loop_num = uid_loop_num[INSN_UID (loop_start)]; | |
3096 | ||
3097 | /* In addition to checking all exits of this loop, we must also check | |
3098 | all exits of inner nested loops that would exit this loop. We don't | |
3099 | have any way to identify those, so we just give up if there are any | |
3100 | such inner loop exits. */ | |
3101 | ||
3102 | for (label = loop_number_exit_labels[this_loop_num]; label; | |
3103 | label = LABEL_NEXTREF (label)) | |
3104 | label_count++; | |
3105 | ||
3106 | if (label_count != loop_number_exit_count[this_loop_num]) | |
3107 | return 0; | |
67f2de41 RK |
3108 | |
3109 | /* HACK: Must also search the loop fall through exit, create a label_ref | |
3110 | here which points to the loop_end, and append the loop_number_exit_labels | |
3111 | list to it. */ | |
38a448ca | 3112 | label = gen_rtx_LABEL_REF (VOIDmode, loop_end); |
3669e646 | 3113 | LABEL_NEXTREF (label) = loop_number_exit_labels[this_loop_num]; |
67f2de41 RK |
3114 | |
3115 | for ( ; label; label = LABEL_NEXTREF (label)) | |
3116 | { | |
3117 | /* Succeed if find an insn which sets the biv or if reach end of | |
3118 | function. Fail if find an insn that uses the biv, or if come to | |
3119 | a conditional jump. */ | |
3120 | ||
3121 | insn = NEXT_INSN (XEXP (label, 0)); | |
412dc348 | 3122 | while (insn) |
67f2de41 | 3123 | { |
412dc348 RK |
3124 | code = GET_CODE (insn); |
3125 | if (GET_RTX_CLASS (code) == 'i') | |
67f2de41 | 3126 | { |
412dc348 RK |
3127 | rtx set; |
3128 | ||
3129 | if (reg_referenced_p (reg, PATTERN (insn))) | |
67f2de41 | 3130 | return 0; |
412dc348 RK |
3131 | |
3132 | set = single_set (insn); | |
3133 | if (set && rtx_equal_p (SET_DEST (set), reg)) | |
3134 | break; | |
67f2de41 | 3135 | } |
412dc348 | 3136 | |
67f2de41 RK |
3137 | if (code == JUMP_INSN) |
3138 | { | |
3139 | if (GET_CODE (PATTERN (insn)) == RETURN) | |
3140 | break; | |
412dc348 | 3141 | else if (! simplejump_p (insn) |
0f41302f | 3142 | /* Prevent infinite loop following infinite loops. */ |
412dc348 | 3143 | || jump_count++ > 20) |
67f2de41 RK |
3144 | return 0; |
3145 | else | |
412dc348 | 3146 | insn = JUMP_LABEL (insn); |
67f2de41 | 3147 | } |
412dc348 | 3148 | |
67f2de41 RK |
3149 | insn = NEXT_INSN (insn); |
3150 | } | |
3151 | } | |
3152 | ||
3153 | /* Success, the register is dead on all loop exits. */ | |
3154 | return 1; | |
3155 | } | |
3156 | ||
3157 | /* Try to calculate the final value of the biv, the value it will have at | |
3158 | the end of the loop. If we can do it, return that value. */ | |
3159 | ||
3160 | rtx | |
3161 | final_biv_value (bl, loop_start, loop_end) | |
3162 | struct iv_class *bl; | |
3163 | rtx loop_start, loop_end; | |
3164 | { | |
3165 | rtx increment, tem; | |
3166 | ||
b4ac57ab RS |
3167 | /* ??? This only works for MODE_INT biv's. Reject all others for now. */ |
3168 | ||
3169 | if (GET_MODE_CLASS (bl->biv->mode) != MODE_INT) | |
3170 | return 0; | |
3171 | ||
67f2de41 RK |
3172 | /* The final value for reversed bivs must be calculated differently than |
3173 | for ordinary bivs. In this case, there is already an insn after the | |
3174 | loop which sets this biv's final value (if necessary), and there are | |
3175 | no other loop exits, so we can return any value. */ | |
3176 | if (bl->reversed) | |
3177 | { | |
3178 | if (loop_dump_stream) | |
3179 | fprintf (loop_dump_stream, | |
3180 | "Final biv value for %d, reversed biv.\n", bl->regno); | |
3181 | ||
3182 | return const0_rtx; | |
3183 | } | |
3184 | ||
3185 | /* Try to calculate the final value as initial value + (number of iterations | |
3186 | * increment). For this to work, increment must be invariant, the only | |
3187 | exit from the loop must be the fall through at the bottom (otherwise | |
3188 | it may not have its final value when the loop exits), and the initial | |
3189 | value of the biv must be invariant. */ | |
3190 | ||
3191 | if (loop_n_iterations != 0 | |
3669e646 | 3192 | && ! loop_number_exit_count[uid_loop_num[INSN_UID (loop_start)]] |
67f2de41 RK |
3193 | && invariant_p (bl->initial_value)) |
3194 | { | |
3195 | increment = biv_total_increment (bl, loop_start, loop_end); | |
3196 | ||
3197 | if (increment && invariant_p (increment)) | |
3198 | { | |
3199 | /* Can calculate the loop exit value, emit insns after loop | |
3200 | end to calculate this value into a temporary register in | |
3201 | case it is needed later. */ | |
3202 | ||
3203 | tem = gen_reg_rtx (bl->biv->mode); | |
9ae8ffe7 | 3204 | record_base_value (REGNO (tem), bl->biv->add_val); |
885cde01 RS |
3205 | /* Make sure loop_end is not the last insn. */ |
3206 | if (NEXT_INSN (loop_end) == 0) | |
3207 | emit_note_after (NOTE_INSN_DELETED, loop_end); | |
c166a311 | 3208 | emit_iv_add_mult (increment, GEN_INT (loop_n_iterations), |
67f2de41 RK |
3209 | bl->initial_value, tem, NEXT_INSN (loop_end)); |
3210 | ||
3211 | if (loop_dump_stream) | |
3212 | fprintf (loop_dump_stream, | |
3213 | "Final biv value for %d, calculated.\n", bl->regno); | |
3214 | ||
3215 | return tem; | |
3216 | } | |
3217 | } | |
3218 | ||
3219 | /* Check to see if the biv is dead at all loop exits. */ | |
3220 | if (reg_dead_after_loop (bl->biv->src_reg, loop_start, loop_end)) | |
3221 | { | |
3222 | if (loop_dump_stream) | |
3223 | fprintf (loop_dump_stream, | |
3224 | "Final biv value for %d, biv dead after loop exit.\n", | |
3225 | bl->regno); | |
3226 | ||
3227 | return const0_rtx; | |
3228 | } | |
3229 | ||
3230 | return 0; | |
3231 | } | |
3232 | ||
3233 | /* Try to calculate the final value of the giv, the value it will have at | |
3234 | the end of the loop. If we can do it, return that value. */ | |
3235 | ||
3236 | rtx | |
3237 | final_giv_value (v, loop_start, loop_end) | |
3238 | struct induction *v; | |
3239 | rtx loop_start, loop_end; | |
3240 | { | |
3241 | struct iv_class *bl; | |
0e91429a | 3242 | rtx insn; |
67f2de41 | 3243 | rtx increment, tem; |
b82fb263 | 3244 | rtx insert_before, seq; |
67f2de41 RK |
3245 | |
3246 | bl = reg_biv_class[REGNO (v->src_reg)]; | |
3247 | ||
3248 | /* The final value for givs which depend on reversed bivs must be calculated | |
3249 | differently than for ordinary givs. In this case, there is already an | |
3250 | insn after the loop which sets this giv's final value (if necessary), | |
3251 | and there are no other loop exits, so we can return any value. */ | |
3252 | if (bl->reversed) | |
3253 | { | |
3254 | if (loop_dump_stream) | |
3255 | fprintf (loop_dump_stream, | |
3256 | "Final giv value for %d, depends on reversed biv\n", | |
3257 | REGNO (v->dest_reg)); | |
3258 | return const0_rtx; | |
3259 | } | |
3260 | ||
3261 | /* Try to calculate the final value as a function of the biv it depends | |
3262 | upon. The only exit from the loop must be the fall through at the bottom | |
3263 | (otherwise it may not have its final value when the loop exits). */ | |
3264 | ||
3265 | /* ??? Can calculate the final giv value by subtracting off the | |
3266 | extra biv increments times the giv's mult_val. The loop must have | |
3267 | only one exit for this to work, but the loop iterations does not need | |
3268 | to be known. */ | |
3269 | ||
3270 | if (loop_n_iterations != 0 | |
3669e646 | 3271 | && ! loop_number_exit_count[uid_loop_num[INSN_UID (loop_start)]]) |
67f2de41 RK |
3272 | { |
3273 | /* ?? It is tempting to use the biv's value here since these insns will | |
3274 | be put after the loop, and hence the biv will have its final value | |
3275 | then. However, this fails if the biv is subsequently eliminated. | |
3276 | Perhaps determine whether biv's are eliminable before trying to | |
3277 | determine whether giv's are replaceable so that we can use the | |
3278 | biv value here if it is not eliminable. */ | |
3279 | ||
db2f7559 JW |
3280 | /* We are emitting code after the end of the loop, so we must make |
3281 | sure that bl->initial_value is still valid then. It will still | |
3282 | be valid if it is invariant. */ | |
3283 | ||
67f2de41 RK |
3284 | increment = biv_total_increment (bl, loop_start, loop_end); |
3285 | ||
db2f7559 JW |
3286 | if (increment && invariant_p (increment) |
3287 | && invariant_p (bl->initial_value)) | |
67f2de41 RK |
3288 | { |
3289 | /* Can calculate the loop exit value of its biv as | |
3290 | (loop_n_iterations * increment) + initial_value */ | |
3291 | ||
3292 | /* The loop exit value of the giv is then | |
3293 | (final_biv_value - extra increments) * mult_val + add_val. | |
3294 | The extra increments are any increments to the biv which | |
3295 | occur in the loop after the giv's value is calculated. | |
3296 | We must search from the insn that sets the giv to the end | |
3297 | of the loop to calculate this value. */ | |
3298 | ||
3299 | insert_before = NEXT_INSN (loop_end); | |
3300 | ||
3301 | /* Put the final biv value in tem. */ | |
3302 | tem = gen_reg_rtx (bl->biv->mode); | |
9ae8ffe7 | 3303 | record_base_value (REGNO (tem), bl->biv->add_val); |
c166a311 | 3304 | emit_iv_add_mult (increment, GEN_INT (loop_n_iterations), |
67f2de41 RK |
3305 | bl->initial_value, tem, insert_before); |
3306 | ||
3307 | /* Subtract off extra increments as we find them. */ | |
3308 | for (insn = NEXT_INSN (v->insn); insn != loop_end; | |
3309 | insn = NEXT_INSN (insn)) | |
3310 | { | |
0e91429a RK |
3311 | struct induction *biv; |
3312 | ||
3313 | for (biv = bl->biv; biv; biv = biv->next_iv) | |
3314 | if (biv->insn == insn) | |
3315 | { | |
3316 | start_sequence (); | |
3317 | tem = expand_binop (GET_MODE (tem), sub_optab, tem, | |
3318 | biv->add_val, NULL_RTX, 0, | |
3319 | OPTAB_LIB_WIDEN); | |
3320 | seq = gen_sequence (); | |
3321 | end_sequence (); | |
3322 | emit_insn_before (seq, insert_before); | |
3323 | } | |
67f2de41 RK |
3324 | } |
3325 | ||
3326 | /* Now calculate the giv's final value. */ | |
3327 | emit_iv_add_mult (tem, v->mult_val, v->add_val, tem, | |
3328 | insert_before); | |
3329 | ||
3330 | if (loop_dump_stream) | |
3331 | fprintf (loop_dump_stream, | |
3332 | "Final giv value for %d, calc from biv's value.\n", | |
3333 | REGNO (v->dest_reg)); | |
3334 | ||
3335 | return tem; | |
3336 | } | |
3337 | } | |
3338 | ||
3339 | /* Replaceable giv's should never reach here. */ | |
3340 | if (v->replaceable) | |
3341 | abort (); | |
3342 | ||
3343 | /* Check to see if the biv is dead at all loop exits. */ | |
3344 | if (reg_dead_after_loop (v->dest_reg, loop_start, loop_end)) | |
3345 | { | |
3346 | if (loop_dump_stream) | |
3347 | fprintf (loop_dump_stream, | |
3348 | "Final giv value for %d, giv dead after loop exit.\n", | |
3349 | REGNO (v->dest_reg)); | |
3350 | ||
3351 | return const0_rtx; | |
3352 | } | |
3353 | ||
3354 | return 0; | |
3355 | } | |
3356 | ||
3357 | ||
3358 | /* Calculate the number of loop iterations. Returns the exact number of loop | |
6dc42e49 | 3359 | iterations if it can be calculated, otherwise returns zero. */ |
67f2de41 | 3360 | |
c166a311 | 3361 | unsigned HOST_WIDE_INT |
67f2de41 RK |
3362 | loop_iterations (loop_start, loop_end) |
3363 | rtx loop_start, loop_end; | |
3364 | { | |
3365 | rtx comparison, comparison_value; | |
3366 | rtx iteration_var, initial_value, increment, final_value; | |
3367 | enum rtx_code comparison_code; | |
c166a311 CH |
3368 | HOST_WIDE_INT i; |
3369 | int increment_dir; | |
67f2de41 RK |
3370 | int unsigned_compare, compare_dir, final_larger; |
3371 | unsigned long tempu; | |
3372 | rtx last_loop_insn; | |
3373 | ||
3374 | /* First find the iteration variable. If the last insn is a conditional | |
3375 | branch, and the insn before tests a register value, make that the | |
3376 | iteration variable. */ | |
3377 | ||
3378 | loop_initial_value = 0; | |
3379 | loop_increment = 0; | |
3380 | loop_final_value = 0; | |
3381 | loop_iteration_var = 0; | |
3382 | ||
40e81af5 JW |
3383 | /* We used to use pren_nonnote_insn here, but that fails because it might |
3384 | accidentally get the branch for a contained loop if the branch for this | |
3385 | loop was deleted. We can only trust branches immediately before the | |
3386 | loop_end. */ | |
3387 | last_loop_insn = PREV_INSN (loop_end); | |
67f2de41 RK |
3388 | |
3389 | comparison = get_condition_for_loop (last_loop_insn); | |
3390 | if (comparison == 0) | |
3391 | { | |
3392 | if (loop_dump_stream) | |
3393 | fprintf (loop_dump_stream, | |
3394 | "Loop unrolling: No final conditional branch found.\n"); | |
3395 | return 0; | |
3396 | } | |
3397 | ||
3398 | /* ??? Get_condition may switch position of induction variable and | |
3399 | invariant register when it canonicalizes the comparison. */ | |
3400 | ||
3401 | comparison_code = GET_CODE (comparison); | |
3402 | iteration_var = XEXP (comparison, 0); | |
3403 | comparison_value = XEXP (comparison, 1); | |
3404 | ||
3405 | if (GET_CODE (iteration_var) != REG) | |
3406 | { | |
3407 | if (loop_dump_stream) | |
3408 | fprintf (loop_dump_stream, | |
3409 | "Loop unrolling: Comparison not against register.\n"); | |
3410 | return 0; | |
3411 | } | |
3412 | ||
3413 | /* Loop iterations is always called before any new registers are created | |
3414 | now, so this should never occur. */ | |
3415 | ||
3416 | if (REGNO (iteration_var) >= max_reg_before_loop) | |
3417 | abort (); | |
3418 | ||
3419 | iteration_info (iteration_var, &initial_value, &increment, | |
3420 | loop_start, loop_end); | |
3421 | if (initial_value == 0) | |
3422 | /* iteration_info already printed a message. */ | |
3423 | return 0; | |
3424 | ||
67f2de41 RK |
3425 | /* If the comparison value is an invariant register, then try to find |
3426 | its value from the insns before the start of the loop. */ | |
3427 | ||
3428 | if (GET_CODE (comparison_value) == REG && invariant_p (comparison_value)) | |
3429 | { | |
3430 | rtx insn, set; | |
3431 | ||
3432 | for (insn = PREV_INSN (loop_start); insn ; insn = PREV_INSN (insn)) | |
3433 | { | |
3434 | if (GET_CODE (insn) == CODE_LABEL) | |
3435 | break; | |
3436 | ||
3437 | else if (GET_RTX_CLASS (GET_CODE (insn)) == 'i' | |
1f1e8172 | 3438 | && reg_set_p (comparison_value, insn)) |
67f2de41 | 3439 | { |
1f1e8172 JW |
3440 | /* We found the last insn before the loop that sets the register. |
3441 | If it sets the entire register, and has a REG_EQUAL note, | |
3442 | then use the value of the REG_EQUAL note. */ | |
3443 | if ((set = single_set (insn)) | |
3444 | && (SET_DEST (set) == comparison_value)) | |
3445 | { | |
3446 | rtx note = find_reg_note (insn, REG_EQUAL, NULL_RTX); | |
67f2de41 | 3447 | |
f0c94be0 JW |
3448 | /* Only use the REG_EQUAL note if it is a constant. |
3449 | Other things, divide in particular, will cause | |
3450 | problems later if we use them. */ | |
3451 | if (note && GET_CODE (XEXP (note, 0)) != EXPR_LIST | |
3452 | && CONSTANT_P (XEXP (note, 0))) | |
1f1e8172 JW |
3453 | comparison_value = XEXP (note, 0); |
3454 | } | |
67f2de41 RK |
3455 | break; |
3456 | } | |
3457 | } | |
3458 | } | |
3459 | ||
3460 | final_value = approx_final_value (comparison_code, comparison_value, | |
3461 | &unsigned_compare, &compare_dir); | |
3462 | ||
3463 | /* Save the calculated values describing this loop's bounds, in case | |
3464 | precondition_loop_p will need them later. These values can not be | |
3465 | recalculated inside precondition_loop_p because strength reduction | |
3466 | optimizations may obscure the loop's structure. */ | |
3467 | ||
3468 | loop_iteration_var = iteration_var; | |
3469 | loop_initial_value = initial_value; | |
3470 | loop_increment = increment; | |
3471 | loop_final_value = final_value; | |
1dcfa896 | 3472 | loop_comparison_code = comparison_code; |
67f2de41 | 3473 | |
00c0c63c JW |
3474 | if (increment == 0) |
3475 | { | |
3476 | if (loop_dump_stream) | |
3477 | fprintf (loop_dump_stream, | |
3478 | "Loop unrolling: Increment value can't be calculated.\n"); | |
3479 | return 0; | |
3480 | } | |
3481 | else if (GET_CODE (increment) != CONST_INT) | |
3482 | { | |
3483 | if (loop_dump_stream) | |
3484 | fprintf (loop_dump_stream, | |
3485 | "Loop unrolling: Increment value not constant.\n"); | |
3486 | return 0; | |
3487 | } | |
3488 | else if (GET_CODE (initial_value) != CONST_INT) | |
3489 | { | |
3490 | if (loop_dump_stream) | |
3491 | fprintf (loop_dump_stream, | |
3492 | "Loop unrolling: Initial value not constant.\n"); | |
3493 | return 0; | |
3494 | } | |
3495 | else if (final_value == 0) | |
67f2de41 RK |
3496 | { |
3497 | if (loop_dump_stream) | |
3498 | fprintf (loop_dump_stream, | |
3499 | "Loop unrolling: EQ comparison loop.\n"); | |
3500 | return 0; | |
3501 | } | |
3502 | else if (GET_CODE (final_value) != CONST_INT) | |
3503 | { | |
3504 | if (loop_dump_stream) | |
3505 | fprintf (loop_dump_stream, | |
3506 | "Loop unrolling: Final value not constant.\n"); | |
3507 | return 0; | |
3508 | } | |
3509 | ||
3510 | /* ?? Final value and initial value do not have to be constants. | |
3511 | Only their difference has to be constant. When the iteration variable | |
3512 | is an array address, the final value and initial value might both | |
3513 | be addresses with the same base but different constant offsets. | |
3514 | Final value must be invariant for this to work. | |
3515 | ||
6dc42e49 | 3516 | To do this, need some way to find the values of registers which are |
67f2de41 RK |
3517 | invariant. */ |
3518 | ||
3519 | /* Final_larger is 1 if final larger, 0 if they are equal, otherwise -1. */ | |
3520 | if (unsigned_compare) | |
3521 | final_larger | |
c166a311 CH |
3522 | = ((unsigned HOST_WIDE_INT) INTVAL (final_value) |
3523 | > (unsigned HOST_WIDE_INT) INTVAL (initial_value)) | |
3524 | - ((unsigned HOST_WIDE_INT) INTVAL (final_value) | |
3525 | < (unsigned HOST_WIDE_INT) INTVAL (initial_value)); | |
67f2de41 | 3526 | else |
c166a311 CH |
3527 | final_larger = (INTVAL (final_value) > INTVAL (initial_value)) |
3528 | - (INTVAL (final_value) < INTVAL (initial_value)); | |
67f2de41 RK |
3529 | |
3530 | if (INTVAL (increment) > 0) | |
3531 | increment_dir = 1; | |
3532 | else if (INTVAL (increment) == 0) | |
3533 | increment_dir = 0; | |
3534 | else | |
3535 | increment_dir = -1; | |
3536 | ||
3537 | /* There are 27 different cases: compare_dir = -1, 0, 1; | |
3538 | final_larger = -1, 0, 1; increment_dir = -1, 0, 1. | |
3539 | There are 4 normal cases, 4 reverse cases (where the iteration variable | |
3540 | will overflow before the loop exits), 4 infinite loop cases, and 15 | |
3541 | immediate exit (0 or 1 iteration depending on loop type) cases. | |
3542 | Only try to optimize the normal cases. */ | |
3543 | ||
3544 | /* (compare_dir/final_larger/increment_dir) | |
3545 | Normal cases: (0/-1/-1), (0/1/1), (-1/-1/-1), (1/1/1) | |
3546 | Reverse cases: (0/-1/1), (0/1/-1), (-1/-1/1), (1/1/-1) | |
3547 | Infinite loops: (0/-1/0), (0/1/0), (-1/-1/0), (1/1/0) | |
3548 | Immediate exit: (0/0/X), (-1/0/X), (-1/1/X), (1/0/X), (1/-1/X) */ | |
3549 | ||
3550 | /* ?? If the meaning of reverse loops (where the iteration variable | |
3551 | will overflow before the loop exits) is undefined, then could | |
3552 | eliminate all of these special checks, and just always assume | |
3553 | the loops are normal/immediate/infinite. Note that this means | |
3554 | the sign of increment_dir does not have to be known. Also, | |
3555 | since it does not really hurt if immediate exit loops or infinite loops | |
3556 | are optimized, then that case could be ignored also, and hence all | |
3557 | loops can be optimized. | |
3558 | ||
3559 | According to ANSI Spec, the reverse loop case result is undefined, | |
3560 | because the action on overflow is undefined. | |
3561 | ||
3562 | See also the special test for NE loops below. */ | |
3563 | ||
3564 | if (final_larger == increment_dir && final_larger != 0 | |
3565 | && (final_larger == compare_dir || compare_dir == 0)) | |
3566 | /* Normal case. */ | |
3567 | ; | |
3568 | else | |
3569 | { | |
3570 | if (loop_dump_stream) | |
3571 | fprintf (loop_dump_stream, | |
3572 | "Loop unrolling: Not normal loop.\n"); | |
3573 | return 0; | |
3574 | } | |
3575 | ||
3576 | /* Calculate the number of iterations, final_value is only an approximation, | |
3577 | so correct for that. Note that tempu and loop_n_iterations are | |
3578 | unsigned, because they can be as large as 2^n - 1. */ | |
3579 | ||
3580 | i = INTVAL (increment); | |
3581 | if (i > 0) | |
3582 | tempu = INTVAL (final_value) - INTVAL (initial_value); | |
3583 | else if (i < 0) | |
3584 | { | |
3585 | tempu = INTVAL (initial_value) - INTVAL (final_value); | |
3586 | i = -i; | |
3587 | } | |
3588 | else | |
3589 | abort (); | |
3590 | ||
3591 | /* For NE tests, make sure that the iteration variable won't miss the | |
3592 | final value. If tempu mod i is not zero, then the iteration variable | |
3593 | will overflow before the loop exits, and we can not calculate the | |
3594 | number of iterations. */ | |
3595 | if (compare_dir == 0 && (tempu % i) != 0) | |
3596 | return 0; | |
3597 | ||
3598 | return tempu / i + ((tempu % i) != 0); | |
3599 | } | |
2b59419a | 3600 | |
9faa82d8 | 3601 | /* Replace uses of split bivs with their split pseudo register. This is |
2b59419a JW |
3602 | for original instructions which remain after loop unrolling without |
3603 | copying. */ | |
3604 | ||
3605 | static rtx | |
3606 | remap_split_bivs (x) | |
3607 | rtx x; | |
3608 | { | |
3609 | register enum rtx_code code; | |
3610 | register int i; | |
3611 | register char *fmt; | |
3612 | ||
3613 | if (x == 0) | |
3614 | return x; | |
3615 | ||
3616 | code = GET_CODE (x); | |
3617 | switch (code) | |
3618 | { | |
3619 | case SCRATCH: | |
3620 | case PC: | |
3621 | case CC0: | |
3622 | case CONST_INT: | |
3623 | case CONST_DOUBLE: | |
3624 | case CONST: | |
3625 | case SYMBOL_REF: | |
3626 | case LABEL_REF: | |
3627 | return x; | |
3628 | ||
3629 | case REG: | |
3630 | #if 0 | |
3631 | /* If non-reduced/final-value givs were split, then this would also | |
3632 | have to remap those givs also. */ | |
3633 | #endif | |
3634 | if (REGNO (x) < max_reg_before_loop | |
3635 | && reg_iv_type[REGNO (x)] == BASIC_INDUCT) | |
3636 | return reg_biv_class[REGNO (x)]->biv->src_reg; | |
e9a25f70 JL |
3637 | break; |
3638 | ||
3639 | default: | |
3640 | break; | |
2b59419a JW |
3641 | } |
3642 | ||
3643 | fmt = GET_RTX_FORMAT (code); | |
3644 | for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) | |
3645 | { | |
3646 | if (fmt[i] == 'e') | |
3647 | XEXP (x, i) = remap_split_bivs (XEXP (x, i)); | |
3648 | if (fmt[i] == 'E') | |
3649 | { | |
3650 | register int j; | |
3651 | for (j = 0; j < XVECLEN (x, i); j++) | |
3652 | XVECEXP (x, i, j) = remap_split_bivs (XVECEXP (x, i, j)); | |
3653 | } | |
3654 | } | |
3655 | return x; | |
3656 | } | |
1fe33d17 JW |
3657 | |
3658 | /* If FIRST_UID is a set of REGNO, and FIRST_UID dominates LAST_UID (e.g. | |
3659 | FIST_UID is always executed if LAST_UID is), then return 1. Otherwise | |
3660 | return 0. COPY_START is where we can start looking for the insns | |
3661 | FIRST_UID and LAST_UID. COPY_END is where we stop looking for these | |
3662 | insns. | |
3663 | ||
3664 | If there is no JUMP_INSN between LOOP_START and FIRST_UID, then FIRST_UID | |
3665 | must dominate LAST_UID. | |
3666 | ||
3667 | If there is a CODE_LABEL between FIRST_UID and LAST_UID, then FIRST_UID | |
3668 | may not dominate LAST_UID. | |
3669 | ||
3670 | If there is no CODE_LABEL between FIRST_UID and LAST_UID, then FIRST_UID | |
3671 | must dominate LAST_UID. */ | |
3672 | ||
3673 | int | |
3674 | set_dominates_use (regno, first_uid, last_uid, copy_start, copy_end) | |
3675 | int regno; | |
3676 | int first_uid; | |
3677 | int last_uid; | |
3678 | rtx copy_start; | |
3679 | rtx copy_end; | |
3680 | { | |
3681 | int passed_jump = 0; | |
3682 | rtx p = NEXT_INSN (copy_start); | |
3683 | ||
3684 | while (INSN_UID (p) != first_uid) | |
3685 | { | |
3686 | if (GET_CODE (p) == JUMP_INSN) | |
3687 | passed_jump= 1; | |
3688 | /* Could not find FIRST_UID. */ | |
3689 | if (p == copy_end) | |
3690 | return 0; | |
3691 | p = NEXT_INSN (p); | |
3692 | } | |
3693 | ||
3694 | /* Verify that FIRST_UID is an insn that entirely sets REGNO. */ | |
3695 | if (GET_RTX_CLASS (GET_CODE (p)) != 'i' | |
3696 | || ! dead_or_set_regno_p (p, regno)) | |
3697 | return 0; | |
3698 | ||
3699 | /* FIRST_UID is always executed. */ | |
3700 | if (passed_jump == 0) | |
3701 | return 1; | |
3702 | ||
3703 | while (INSN_UID (p) != last_uid) | |
3704 | { | |
3705 | /* If we see a CODE_LABEL between FIRST_UID and LAST_UID, then we | |
3706 | can not be sure that FIRST_UID dominates LAST_UID. */ | |
3707 | if (GET_CODE (p) == CODE_LABEL) | |
3708 | return 0; | |
6b857f04 JW |
3709 | /* Could not find LAST_UID, but we reached the end of the loop, so |
3710 | it must be safe. */ | |
3711 | else if (p == copy_end) | |
3712 | return 1; | |
1fe33d17 JW |
3713 | p = NEXT_INSN (p); |
3714 | } | |
3715 | ||
3716 | /* FIRST_UID is always executed if LAST_UID is executed. */ | |
3717 | return 1; | |
3718 | } |