1 | /* Try to unroll loops, and split induction variables. |
---|
2 | Copyright (C) 1992, 93, 94, 95, 97, 1998 Free Software Foundation, Inc. |
---|
3 | Contributed by James E. Wilson, Cygnus Support/UC Berkeley. |
---|
4 | |
---|
5 | This file is part of GNU CC. |
---|
6 | |
---|
7 | GNU CC is free software; you can redistribute it and/or modify |
---|
8 | it under the terms of the GNU General Public License as published by |
---|
9 | the Free Software Foundation; either version 2, or (at your option) |
---|
10 | any later version. |
---|
11 | |
---|
12 | GNU CC is distributed in the hope that it will be useful, |
---|
13 | but WITHOUT ANY WARRANTY; without even the implied warranty of |
---|
14 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
---|
15 | GNU General Public License for more details. |
---|
16 | |
---|
17 | You should have received a copy of the GNU General Public License |
---|
18 | along with GNU CC; see the file COPYING. If not, write to |
---|
19 | the Free Software Foundation, 59 Temple Place - Suite 330, |
---|
20 | Boston, MA 02111-1307, USA. */ |
---|
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 |
---|
63 | memory addresses can be split, and hence giv increments can be |
---|
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" |
---|
150 | #include <stdio.h> |
---|
151 | #include "rtl.h" |
---|
152 | #include "insn-config.h" |
---|
153 | #include "integrate.h" |
---|
154 | #include "regs.h" |
---|
155 | #include "recog.h" |
---|
156 | #include "flags.h" |
---|
157 | #include "expr.h" |
---|
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; |
---|
195 | static enum rtx_code loop_comparison_code; |
---|
196 | |
---|
197 | /* Forward declarations. */ |
---|
198 | |
---|
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)); |
---|
206 | static void iteration_info PROTO((rtx, rtx *, rtx *, rtx, rtx)); |
---|
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)); |
---|
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 |
---|
218 | LOOP_START. END_INSERT_BEFORE indicates where insns should be added |
---|
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; |
---|
243 | char *local_regno; |
---|
244 | int maxregnum; |
---|
245 | int new_maxregnum; |
---|
246 | rtx exit_label = 0; |
---|
247 | rtx start_label; |
---|
248 | struct iv_class *bl; |
---|
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. */ |
---|
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. */ |
---|
276 | |
---|
277 | if (1 /* write_symbols != NO_DEBUG */) |
---|
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 |
---|
341 | /* The immediately preceding insn is a compare which must be |
---|
342 | deleted. */ |
---|
343 | delete_insn (last_loop_insn); |
---|
344 | delete_insn (PREV_INSN (last_loop_insn)); |
---|
345 | #else |
---|
346 | /* The immediately preceding insn may not be the compare, so don't |
---|
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 | } |
---|
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 | } |
---|
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 | |
---|
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 | |
---|
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 | |
---|
666 | map->integrating = 0; |
---|
667 | |
---|
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)) |
---|
694 | set_label_in_map (map, |
---|
695 | CODE_LABEL_NUMBER (JUMP_LABEL (insn)), |
---|
696 | JUMP_LABEL (insn)); |
---|
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); |
---|
708 | set_label_in_map (map, |
---|
709 | CODE_LABEL_NUMBER (label), |
---|
710 | label); |
---|
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 |
---|
739 | to access the splittable_regs[] and addr_combined_regs[] arrays. */ |
---|
740 | |
---|
741 | splittable_regs = (rtx *) alloca (maxregnum * sizeof (rtx)); |
---|
742 | bzero ((char *) splittable_regs, maxregnum * sizeof (rtx)); |
---|
743 | splittable_regs_updates = (int *) alloca (maxregnum * sizeof (int)); |
---|
744 | bzero ((char *) splittable_regs_updates, maxregnum * sizeof (int)); |
---|
745 | addr_combined_regs |
---|
746 | = (struct induction **) alloca (maxregnum * sizeof (struct induction *)); |
---|
747 | bzero ((char *) addr_combined_regs, maxregnum * sizeof (struct induction *)); |
---|
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); |
---|
753 | |
---|
754 | /* Mark all local registers, i.e. the ones which are referenced only |
---|
755 | inside the loop. */ |
---|
756 | if (INSN_UID (copy_end) < max_uid_for_loop) |
---|
757 | { |
---|
758 | int copy_start_luid = INSN_LUID (copy_start); |
---|
759 | int copy_end_luid = INSN_LUID (copy_end); |
---|
760 | |
---|
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--; |
---|
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++; |
---|
771 | |
---|
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. */ |
---|
775 | for (j = FIRST_PSEUDO_REGISTER; j < max_reg_before_loop; ++j) |
---|
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) |
---|
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. */ |
---|
790 | if (set_dominates_use (j, REGNO_FIRST_UID (j), REGNO_LAST_UID (j), |
---|
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 | } |
---|
803 | } |
---|
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 | { |
---|
834 | register rtx diff, temp; |
---|
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; |
---|
846 | global_const_equiv_map_size = maxregnum; |
---|
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), |
---|
891 | copy_rtx (initial_value), NULL_RTX, 0, |
---|
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, |
---|
897 | GEN_INT (unroll_number * abs_inc - 1), |
---|
898 | NULL_RTX, 0, OPTAB_LIB_WIDEN); |
---|
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 | |
---|
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. */ |
---|
912 | |
---|
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 | } |
---|
924 | |
---|
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; |
---|
941 | enum rtx_code cmp_code; |
---|
942 | |
---|
943 | /* For negative increments, must invert the constant compared |
---|
944 | against, except when comparing against zero. */ |
---|
945 | if (i == 0) |
---|
946 | { |
---|
947 | cmp_const = 0; |
---|
948 | cmp_code = EQ; |
---|
949 | } |
---|
950 | else if (neg_inc) |
---|
951 | { |
---|
952 | cmp_const = unroll_number - i; |
---|
953 | cmp_code = GE; |
---|
954 | } |
---|
955 | else |
---|
956 | { |
---|
957 | cmp_const = i; |
---|
958 | cmp_code = LE; |
---|
959 | } |
---|
960 | |
---|
961 | emit_cmp_insn (diff, GEN_INT (abs_inc * cmp_const), |
---|
962 | cmp_code, NULL_RTX, mode, 0, 0); |
---|
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; |
---|
986 | enum rtx_code cmp_code; |
---|
987 | |
---|
988 | if (neg_inc) |
---|
989 | { |
---|
990 | cmp_const = abs_inc - 1; |
---|
991 | cmp_code = LE; |
---|
992 | } |
---|
993 | else |
---|
994 | { |
---|
995 | cmp_const = abs_inc * (unroll_number - 1) + 1; |
---|
996 | cmp_code = GE; |
---|
997 | } |
---|
998 | |
---|
999 | emit_cmp_insn (diff, GEN_INT (cmp_const), cmp_code, NULL_RTX, |
---|
1000 | mode, 0, 0); |
---|
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 |
---|
1024 | /* The immediately preceding insn is a compare which we do not |
---|
1025 | want to copy. */ |
---|
1026 | copy_end = PREV_INSN (PREV_INSN (last_loop_insn)); |
---|
1027 | #else |
---|
1028 | /* The immediately preceding insn may not be a compare, so we |
---|
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 | |
---|
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)); |
---|
1045 | map->const_age = 0; |
---|
1046 | |
---|
1047 | for (j = 0; j < max_labelno; j++) |
---|
1048 | if (local_label[j]) |
---|
1049 | set_label_in_map (map, j, gen_label_rtx ()); |
---|
1050 | |
---|
1051 | for (j = FIRST_PSEUDO_REGISTER; j < max_reg_before_loop; j++) |
---|
1052 | if (local_regno[j]) |
---|
1053 | map->reg_map[j] = gen_reg_rtx (GET_MODE (regno_reg_rtx[j])); |
---|
1054 | |
---|
1055 | /* The last copy needs the compare/branch insns at the end, |
---|
1056 | so reset copy_end here if the loop ends with a conditional |
---|
1057 | branch. */ |
---|
1058 | |
---|
1059 | if (i == unroll_number - 1) |
---|
1060 | { |
---|
1061 | if (GET_CODE (last_loop_insn) == BARRIER) |
---|
1062 | copy_end = PREV_INSN (PREV_INSN (last_loop_insn)); |
---|
1063 | else |
---|
1064 | copy_end = last_loop_insn; |
---|
1065 | } |
---|
1066 | |
---|
1067 | /* None of the copies are the `last_iteration', so just |
---|
1068 | pass zero for that parameter. */ |
---|
1069 | copy_loop_body (copy_start, copy_end, map, exit_label, 0, |
---|
1070 | unroll_type, start_label, loop_end, |
---|
1071 | loop_start, copy_end); |
---|
1072 | } |
---|
1073 | emit_label_after (labels[0], PREV_INSN (loop_start)); |
---|
1074 | |
---|
1075 | if (GET_CODE (last_loop_insn) == BARRIER) |
---|
1076 | { |
---|
1077 | insert_before = PREV_INSN (last_loop_insn); |
---|
1078 | copy_end = PREV_INSN (insert_before); |
---|
1079 | } |
---|
1080 | else |
---|
1081 | { |
---|
1082 | #ifdef HAVE_cc0 |
---|
1083 | /* The immediately preceding insn is a compare which we do not |
---|
1084 | want to copy. */ |
---|
1085 | insert_before = PREV_INSN (last_loop_insn); |
---|
1086 | copy_end = PREV_INSN (insert_before); |
---|
1087 | #else |
---|
1088 | /* The immediately preceding insn may not be a compare, so we |
---|
1089 | must copy it. */ |
---|
1090 | insert_before = last_loop_insn; |
---|
1091 | copy_end = PREV_INSN (last_loop_insn); |
---|
1092 | #endif |
---|
1093 | } |
---|
1094 | |
---|
1095 | /* Set unroll type to MODULO now. */ |
---|
1096 | unroll_type = UNROLL_MODULO; |
---|
1097 | loop_preconditioned = 1; |
---|
1098 | } |
---|
1099 | } |
---|
1100 | |
---|
1101 | /* If reach here, and the loop type is UNROLL_NAIVE, then don't unroll |
---|
1102 | the loop unless all loops are being unrolled. */ |
---|
1103 | if (unroll_type == UNROLL_NAIVE && ! flag_unroll_all_loops) |
---|
1104 | { |
---|
1105 | if (loop_dump_stream) |
---|
1106 | fprintf (loop_dump_stream, "Unrolling failure: Naive unrolling not being done.\n"); |
---|
1107 | return; |
---|
1108 | } |
---|
1109 | |
---|
1110 | /* At this point, we are guaranteed to unroll the loop. */ |
---|
1111 | |
---|
1112 | /* For each biv and giv, determine whether it can be safely split into |
---|
1113 | a different variable for each unrolled copy of the loop body. |
---|
1114 | We precalculate and save this info here, since computing it is |
---|
1115 | expensive. |
---|
1116 | |
---|
1117 | Do this before deleting any instructions from the loop, so that |
---|
1118 | back_branch_in_range_p will work correctly. */ |
---|
1119 | |
---|
1120 | if (splitting_not_safe) |
---|
1121 | temp = 0; |
---|
1122 | else |
---|
1123 | temp = find_splittable_regs (unroll_type, loop_start, loop_end, |
---|
1124 | end_insert_before, unroll_number); |
---|
1125 | |
---|
1126 | /* find_splittable_regs may have created some new registers, so must |
---|
1127 | reallocate the reg_map with the new larger size, and must realloc |
---|
1128 | the constant maps also. */ |
---|
1129 | |
---|
1130 | maxregnum = max_reg_num (); |
---|
1131 | map->reg_map = (rtx *) alloca (maxregnum * sizeof (rtx)); |
---|
1132 | |
---|
1133 | init_reg_map (map, maxregnum); |
---|
1134 | |
---|
1135 | /* Space is needed in some of the map for new registers, so new_maxregnum |
---|
1136 | is an (over)estimate of how many registers will exist at the end. */ |
---|
1137 | new_maxregnum = maxregnum + (temp * unroll_number * 2); |
---|
1138 | |
---|
1139 | /* Must realloc space for the constant maps, because the number of registers |
---|
1140 | may have changed. */ |
---|
1141 | |
---|
1142 | map->const_equiv_map = (rtx *) alloca (new_maxregnum * sizeof (rtx)); |
---|
1143 | map->const_age_map = (unsigned *) alloca (new_maxregnum * sizeof (unsigned)); |
---|
1144 | |
---|
1145 | map->const_equiv_map_size = new_maxregnum; |
---|
1146 | global_const_equiv_map = map->const_equiv_map; |
---|
1147 | global_const_equiv_map_size = new_maxregnum; |
---|
1148 | |
---|
1149 | /* Search the list of bivs and givs to find ones which need to be remapped |
---|
1150 | when split, and set their reg_map entry appropriately. */ |
---|
1151 | |
---|
1152 | for (bl = loop_iv_list; bl; bl = bl->next) |
---|
1153 | { |
---|
1154 | if (REGNO (bl->biv->src_reg) != bl->regno) |
---|
1155 | map->reg_map[bl->regno] = bl->biv->src_reg; |
---|
1156 | #if 0 |
---|
1157 | /* Currently, non-reduced/final-value givs are never split. */ |
---|
1158 | for (v = bl->giv; v; v = v->next_iv) |
---|
1159 | if (REGNO (v->src_reg) != bl->regno) |
---|
1160 | map->reg_map[REGNO (v->dest_reg)] = v->src_reg; |
---|
1161 | #endif |
---|
1162 | } |
---|
1163 | |
---|
1164 | /* Use our current register alignment and pointer flags. */ |
---|
1165 | map->regno_pointer_flag = regno_pointer_flag; |
---|
1166 | map->regno_pointer_align = regno_pointer_align; |
---|
1167 | |
---|
1168 | /* If the loop is being partially unrolled, and the iteration variables |
---|
1169 | are being split, and are being renamed for the split, then must fix up |
---|
1170 | the compare/jump instruction at the end of the loop to refer to the new |
---|
1171 | registers. This compare isn't copied, so the registers used in it |
---|
1172 | will never be replaced if it isn't done here. */ |
---|
1173 | |
---|
1174 | if (unroll_type == UNROLL_MODULO) |
---|
1175 | { |
---|
1176 | insn = NEXT_INSN (copy_end); |
---|
1177 | if (GET_CODE (insn) == INSN || GET_CODE (insn) == JUMP_INSN) |
---|
1178 | PATTERN (insn) = remap_split_bivs (PATTERN (insn)); |
---|
1179 | } |
---|
1180 | |
---|
1181 | /* For unroll_number - 1 times, make a copy of each instruction |
---|
1182 | between copy_start and copy_end, and insert these new instructions |
---|
1183 | before the end of the loop. */ |
---|
1184 | |
---|
1185 | for (i = 0; i < unroll_number; i++) |
---|
1186 | { |
---|
1187 | bzero ((char *) map->insn_map, max_insnno * sizeof (rtx)); |
---|
1188 | bzero ((char *) map->const_equiv_map, new_maxregnum * sizeof (rtx)); |
---|
1189 | bzero ((char *) map->const_age_map, new_maxregnum * sizeof (unsigned)); |
---|
1190 | map->const_age = 0; |
---|
1191 | |
---|
1192 | for (j = 0; j < max_labelno; j++) |
---|
1193 | if (local_label[j]) |
---|
1194 | set_label_in_map (map, j, gen_label_rtx ()); |
---|
1195 | |
---|
1196 | for (j = FIRST_PSEUDO_REGISTER; j < max_reg_before_loop; j++) |
---|
1197 | if (local_regno[j]) |
---|
1198 | map->reg_map[j] = gen_reg_rtx (GET_MODE (regno_reg_rtx[j])); |
---|
1199 | |
---|
1200 | /* If loop starts with a branch to the test, then fix it so that |
---|
1201 | it points to the test of the first unrolled copy of the loop. */ |
---|
1202 | if (i == 0 && loop_start != copy_start) |
---|
1203 | { |
---|
1204 | insn = PREV_INSN (copy_start); |
---|
1205 | pattern = PATTERN (insn); |
---|
1206 | |
---|
1207 | tem = get_label_from_map (map, |
---|
1208 | CODE_LABEL_NUMBER |
---|
1209 | (XEXP (SET_SRC (pattern), 0))); |
---|
1210 | SET_SRC (pattern) = gen_rtx (LABEL_REF, VOIDmode, tem); |
---|
1211 | |
---|
1212 | /* Set the jump label so that it can be used by later loop unrolling |
---|
1213 | passes. */ |
---|
1214 | JUMP_LABEL (insn) = tem; |
---|
1215 | LABEL_NUSES (tem)++; |
---|
1216 | } |
---|
1217 | |
---|
1218 | copy_loop_body (copy_start, copy_end, map, exit_label, |
---|
1219 | i == unroll_number - 1, unroll_type, start_label, |
---|
1220 | loop_end, insert_before, insert_before); |
---|
1221 | } |
---|
1222 | |
---|
1223 | /* Before deleting any insns, emit a CODE_LABEL immediately after the last |
---|
1224 | insn to be deleted. This prevents any runaway delete_insn call from |
---|
1225 | more insns that it should, as it always stops at a CODE_LABEL. */ |
---|
1226 | |
---|
1227 | /* Delete the compare and branch at the end of the loop if completely |
---|
1228 | unrolling the loop. Deleting the backward branch at the end also |
---|
1229 | deletes the code label at the start of the loop. This is done at |
---|
1230 | the very end to avoid problems with back_branch_in_range_p. */ |
---|
1231 | |
---|
1232 | if (unroll_type == UNROLL_COMPLETELY) |
---|
1233 | safety_label = emit_label_after (gen_label_rtx (), last_loop_insn); |
---|
1234 | else |
---|
1235 | safety_label = emit_label_after (gen_label_rtx (), copy_end); |
---|
1236 | |
---|
1237 | /* Delete all of the original loop instructions. Don't delete the |
---|
1238 | LOOP_BEG note, or the first code label in the loop. */ |
---|
1239 | |
---|
1240 | insn = NEXT_INSN (copy_start); |
---|
1241 | while (insn != safety_label) |
---|
1242 | { |
---|
1243 | if (insn != start_label) |
---|
1244 | insn = delete_insn (insn); |
---|
1245 | else |
---|
1246 | insn = NEXT_INSN (insn); |
---|
1247 | } |
---|
1248 | |
---|
1249 | /* Can now delete the 'safety' label emitted to protect us from runaway |
---|
1250 | delete_insn calls. */ |
---|
1251 | if (INSN_DELETED_P (safety_label)) |
---|
1252 | abort (); |
---|
1253 | delete_insn (safety_label); |
---|
1254 | |
---|
1255 | /* If exit_label exists, emit it after the loop. Doing the emit here |
---|
1256 | forces it to have a higher INSN_UID than any insn in the unrolled loop. |
---|
1257 | This is needed so that mostly_true_jump in reorg.c will treat jumps |
---|
1258 | to this loop end label correctly, i.e. predict that they are usually |
---|
1259 | not taken. */ |
---|
1260 | if (exit_label) |
---|
1261 | emit_label_after (exit_label, loop_end); |
---|
1262 | } |
---|
1263 | |
---|
1264 | /* Return true if the loop can be safely, and profitably, preconditioned |
---|
1265 | so that the unrolled copies of the loop body don't need exit tests. |
---|
1266 | |
---|
1267 | This only works if final_value, initial_value and increment can be |
---|
1268 | determined, and if increment is a constant power of 2. |
---|
1269 | If increment is not a power of 2, then the preconditioning modulo |
---|
1270 | operation would require a real modulo instead of a boolean AND, and this |
---|
1271 | is not considered `profitable'. */ |
---|
1272 | |
---|
1273 | /* ??? If the loop is known to be executed very many times, or the machine |
---|
1274 | has a very cheap divide instruction, then preconditioning is a win even |
---|
1275 | when the increment is not a power of 2. Use RTX_COST to compute |
---|
1276 | whether divide is cheap. */ |
---|
1277 | |
---|
1278 | static int |
---|
1279 | precondition_loop_p (initial_value, final_value, increment, loop_start, |
---|
1280 | loop_end) |
---|
1281 | rtx *initial_value, *final_value, *increment; |
---|
1282 | rtx loop_start, loop_end; |
---|
1283 | { |
---|
1284 | |
---|
1285 | if (loop_n_iterations > 0) |
---|
1286 | { |
---|
1287 | *initial_value = const0_rtx; |
---|
1288 | *increment = const1_rtx; |
---|
1289 | *final_value = GEN_INT (loop_n_iterations); |
---|
1290 | |
---|
1291 | if (loop_dump_stream) |
---|
1292 | fprintf (loop_dump_stream, |
---|
1293 | "Preconditioning: Success, number of iterations known, %d.\n", |
---|
1294 | loop_n_iterations); |
---|
1295 | return 1; |
---|
1296 | } |
---|
1297 | |
---|
1298 | if (loop_initial_value == 0) |
---|
1299 | { |
---|
1300 | if (loop_dump_stream) |
---|
1301 | fprintf (loop_dump_stream, |
---|
1302 | "Preconditioning: Could not find initial value.\n"); |
---|
1303 | return 0; |
---|
1304 | } |
---|
1305 | else if (loop_increment == 0) |
---|
1306 | { |
---|
1307 | if (loop_dump_stream) |
---|
1308 | fprintf (loop_dump_stream, |
---|
1309 | "Preconditioning: Could not find increment value.\n"); |
---|
1310 | return 0; |
---|
1311 | } |
---|
1312 | else if (GET_CODE (loop_increment) != CONST_INT) |
---|
1313 | { |
---|
1314 | if (loop_dump_stream) |
---|
1315 | fprintf (loop_dump_stream, |
---|
1316 | "Preconditioning: Increment not a constant.\n"); |
---|
1317 | return 0; |
---|
1318 | } |
---|
1319 | else if ((exact_log2 (INTVAL (loop_increment)) < 0) |
---|
1320 | && (exact_log2 (- INTVAL (loop_increment)) < 0)) |
---|
1321 | { |
---|
1322 | if (loop_dump_stream) |
---|
1323 | fprintf (loop_dump_stream, |
---|
1324 | "Preconditioning: Increment not a constant power of 2.\n"); |
---|
1325 | return 0; |
---|
1326 | } |
---|
1327 | |
---|
1328 | /* Unsigned_compare and compare_dir can be ignored here, since they do |
---|
1329 | not matter for preconditioning. */ |
---|
1330 | |
---|
1331 | if (loop_final_value == 0) |
---|
1332 | { |
---|
1333 | if (loop_dump_stream) |
---|
1334 | fprintf (loop_dump_stream, |
---|
1335 | "Preconditioning: EQ comparison loop.\n"); |
---|
1336 | return 0; |
---|
1337 | } |
---|
1338 | |
---|
1339 | /* Must ensure that final_value is invariant, so call invariant_p to |
---|
1340 | check. Before doing so, must check regno against max_reg_before_loop |
---|
1341 | to make sure that the register is in the range covered by invariant_p. |
---|
1342 | If it isn't, then it is most likely a biv/giv which by definition are |
---|
1343 | not invariant. */ |
---|
1344 | if ((GET_CODE (loop_final_value) == REG |
---|
1345 | && REGNO (loop_final_value) >= max_reg_before_loop) |
---|
1346 | || (GET_CODE (loop_final_value) == PLUS |
---|
1347 | && REGNO (XEXP (loop_final_value, 0)) >= max_reg_before_loop) |
---|
1348 | || ! invariant_p (loop_final_value)) |
---|
1349 | { |
---|
1350 | if (loop_dump_stream) |
---|
1351 | fprintf (loop_dump_stream, |
---|
1352 | "Preconditioning: Final value not invariant.\n"); |
---|
1353 | return 0; |
---|
1354 | } |
---|
1355 | |
---|
1356 | /* Fail for floating point values, since the caller of this function |
---|
1357 | does not have code to deal with them. */ |
---|
1358 | if (GET_MODE_CLASS (GET_MODE (loop_final_value)) == MODE_FLOAT |
---|
1359 | || GET_MODE_CLASS (GET_MODE (loop_initial_value)) == MODE_FLOAT) |
---|
1360 | { |
---|
1361 | if (loop_dump_stream) |
---|
1362 | fprintf (loop_dump_stream, |
---|
1363 | "Preconditioning: Floating point final or initial value.\n"); |
---|
1364 | return 0; |
---|
1365 | } |
---|
1366 | |
---|
1367 | /* Now set initial_value to be the iteration_var, since that may be a |
---|
1368 | simpler expression, and is guaranteed to be correct if all of the |
---|
1369 | above tests succeed. |
---|
1370 | |
---|
1371 | We can not use the initial_value as calculated, because it will be |
---|
1372 | one too small for loops of the form "while (i-- > 0)". We can not |
---|
1373 | emit code before the loop_skip_over insns to fix this problem as this |
---|
1374 | will then give a number one too large for loops of the form |
---|
1375 | "while (--i > 0)". |
---|
1376 | |
---|
1377 | Note that all loops that reach here are entered at the top, because |
---|
1378 | this function is not called if the loop starts with a jump. */ |
---|
1379 | |
---|
1380 | /* Fail if loop_iteration_var is not live before loop_start, since we need |
---|
1381 | to test its value in the preconditioning code. */ |
---|
1382 | |
---|
1383 | if (uid_luid[REGNO_FIRST_UID (REGNO (loop_iteration_var))] |
---|
1384 | > INSN_LUID (loop_start)) |
---|
1385 | { |
---|
1386 | if (loop_dump_stream) |
---|
1387 | fprintf (loop_dump_stream, |
---|
1388 | "Preconditioning: Iteration var not live before loop start.\n"); |
---|
1389 | return 0; |
---|
1390 | } |
---|
1391 | |
---|
1392 | *initial_value = loop_iteration_var; |
---|
1393 | *increment = loop_increment; |
---|
1394 | *final_value = loop_final_value; |
---|
1395 | |
---|
1396 | /* Success! */ |
---|
1397 | if (loop_dump_stream) |
---|
1398 | fprintf (loop_dump_stream, "Preconditioning: Successful.\n"); |
---|
1399 | return 1; |
---|
1400 | } |
---|
1401 | |
---|
1402 | |
---|
1403 | /* All pseudo-registers must be mapped to themselves. Two hard registers |
---|
1404 | must be mapped, VIRTUAL_STACK_VARS_REGNUM and VIRTUAL_INCOMING_ARGS_ |
---|
1405 | REGNUM, to avoid function-inlining specific conversions of these |
---|
1406 | registers. All other hard regs can not be mapped because they may be |
---|
1407 | used with different |
---|
1408 | modes. */ |
---|
1409 | |
---|
1410 | static void |
---|
1411 | init_reg_map (map, maxregnum) |
---|
1412 | struct inline_remap *map; |
---|
1413 | int maxregnum; |
---|
1414 | { |
---|
1415 | int i; |
---|
1416 | |
---|
1417 | for (i = maxregnum - 1; i > LAST_VIRTUAL_REGISTER; i--) |
---|
1418 | map->reg_map[i] = regno_reg_rtx[i]; |
---|
1419 | /* Just clear the rest of the entries. */ |
---|
1420 | for (i = LAST_VIRTUAL_REGISTER; i >= 0; i--) |
---|
1421 | map->reg_map[i] = 0; |
---|
1422 | |
---|
1423 | map->reg_map[VIRTUAL_STACK_VARS_REGNUM] |
---|
1424 | = regno_reg_rtx[VIRTUAL_STACK_VARS_REGNUM]; |
---|
1425 | map->reg_map[VIRTUAL_INCOMING_ARGS_REGNUM] |
---|
1426 | = regno_reg_rtx[VIRTUAL_INCOMING_ARGS_REGNUM]; |
---|
1427 | } |
---|
1428 | |
---|
1429 | /* Strength-reduction will often emit code for optimized biv/givs which |
---|
1430 | calculates their value in a temporary register, and then copies the result |
---|
1431 | to the iv. This procedure reconstructs the pattern computing the iv; |
---|
1432 | verifying that all operands are of the proper form. |
---|
1433 | |
---|
1434 | PATTERN must be the result of single_set. |
---|
1435 | The return value is the amount that the giv is incremented by. */ |
---|
1436 | |
---|
1437 | static rtx |
---|
1438 | calculate_giv_inc (pattern, src_insn, regno) |
---|
1439 | rtx pattern, src_insn; |
---|
1440 | int regno; |
---|
1441 | { |
---|
1442 | rtx increment; |
---|
1443 | rtx increment_total = 0; |
---|
1444 | int tries = 0; |
---|
1445 | |
---|
1446 | retry: |
---|
1447 | /* Verify that we have an increment insn here. First check for a plus |
---|
1448 | as the set source. */ |
---|
1449 | if (GET_CODE (SET_SRC (pattern)) != PLUS) |
---|
1450 | { |
---|
1451 | /* SR sometimes computes the new giv value in a temp, then copies it |
---|
1452 | to the new_reg. */ |
---|
1453 | src_insn = PREV_INSN (src_insn); |
---|
1454 | pattern = PATTERN (src_insn); |
---|
1455 | if (GET_CODE (SET_SRC (pattern)) != PLUS) |
---|
1456 | abort (); |
---|
1457 | |
---|
1458 | /* The last insn emitted is not needed, so delete it to avoid confusing |
---|
1459 | the second cse pass. This insn sets the giv unnecessarily. */ |
---|
1460 | delete_insn (get_last_insn ()); |
---|
1461 | } |
---|
1462 | |
---|
1463 | /* Verify that we have a constant as the second operand of the plus. */ |
---|
1464 | increment = XEXP (SET_SRC (pattern), 1); |
---|
1465 | if (GET_CODE (increment) != CONST_INT) |
---|
1466 | { |
---|
1467 | /* SR sometimes puts the constant in a register, especially if it is |
---|
1468 | too big to be an add immed operand. */ |
---|
1469 | src_insn = PREV_INSN (src_insn); |
---|
1470 | increment = SET_SRC (PATTERN (src_insn)); |
---|
1471 | |
---|
1472 | /* SR may have used LO_SUM to compute the constant if it is too large |
---|
1473 | for a load immed operand. In this case, the constant is in operand |
---|
1474 | one of the LO_SUM rtx. */ |
---|
1475 | if (GET_CODE (increment) == LO_SUM) |
---|
1476 | increment = XEXP (increment, 1); |
---|
1477 | else if (GET_CODE (increment) == IOR |
---|
1478 | || GET_CODE (increment) == ASHIFT |
---|
1479 | || GET_CODE (increment) == PLUS) |
---|
1480 | { |
---|
1481 | /* The rs6000 port loads some constants with IOR. |
---|
1482 | The alpha port loads some constants with ASHIFT and PLUS. */ |
---|
1483 | rtx second_part = XEXP (increment, 1); |
---|
1484 | enum rtx_code code = GET_CODE (increment); |
---|
1485 | |
---|
1486 | src_insn = PREV_INSN (src_insn); |
---|
1487 | increment = SET_SRC (PATTERN (src_insn)); |
---|
1488 | /* Don't need the last insn anymore. */ |
---|
1489 | delete_insn (get_last_insn ()); |
---|
1490 | |
---|
1491 | if (GET_CODE (second_part) != CONST_INT |
---|
1492 | || GET_CODE (increment) != CONST_INT) |
---|
1493 | abort (); |
---|
1494 | |
---|
1495 | if (code == IOR) |
---|
1496 | increment = GEN_INT (INTVAL (increment) | INTVAL (second_part)); |
---|
1497 | else if (code == PLUS) |
---|
1498 | increment = GEN_INT (INTVAL (increment) + INTVAL (second_part)); |
---|
1499 | else |
---|
1500 | increment = GEN_INT (INTVAL (increment) << INTVAL (second_part)); |
---|
1501 | } |
---|
1502 | |
---|
1503 | if (GET_CODE (increment) != CONST_INT) |
---|
1504 | abort (); |
---|
1505 | |
---|
1506 | /* The insn loading the constant into a register is no longer needed, |
---|
1507 | so delete it. */ |
---|
1508 | delete_insn (get_last_insn ()); |
---|
1509 | } |
---|
1510 | |
---|
1511 | if (increment_total) |
---|
1512 | increment_total = GEN_INT (INTVAL (increment_total) + INTVAL (increment)); |
---|
1513 | else |
---|
1514 | increment_total = increment; |
---|
1515 | |
---|
1516 | /* Check that the source register is the same as the register we expected |
---|
1517 | to see as the source. If not, something is seriously wrong. */ |
---|
1518 | if (GET_CODE (XEXP (SET_SRC (pattern), 0)) != REG |
---|
1519 | || REGNO (XEXP (SET_SRC (pattern), 0)) != regno) |
---|
1520 | { |
---|
1521 | /* Some machines (e.g. the romp), may emit two add instructions for |
---|
1522 | certain constants, so lets try looking for another add immediately |
---|
1523 | before this one if we have only seen one add insn so far. */ |
---|
1524 | |
---|
1525 | if (tries == 0) |
---|
1526 | { |
---|
1527 | tries++; |
---|
1528 | |
---|
1529 | src_insn = PREV_INSN (src_insn); |
---|
1530 | pattern = PATTERN (src_insn); |
---|
1531 | |
---|
1532 | delete_insn (get_last_insn ()); |
---|
1533 | |
---|
1534 | goto retry; |
---|
1535 | } |
---|
1536 | |
---|
1537 | abort (); |
---|
1538 | } |
---|
1539 | |
---|
1540 | return increment_total; |
---|
1541 | } |
---|
1542 | |
---|
1543 | /* Copy REG_NOTES, except for insn references, because not all insn_map |
---|
1544 | entries are valid yet. We do need to copy registers now though, because |
---|
1545 | the reg_map entries can change during copying. */ |
---|
1546 | |
---|
1547 | static rtx |
---|
1548 | initial_reg_note_copy (notes, map) |
---|
1549 | rtx notes; |
---|
1550 | struct inline_remap *map; |
---|
1551 | { |
---|
1552 | rtx copy; |
---|
1553 | |
---|
1554 | if (notes == 0) |
---|
1555 | return 0; |
---|
1556 | |
---|
1557 | copy = rtx_alloc (GET_CODE (notes)); |
---|
1558 | PUT_MODE (copy, GET_MODE (notes)); |
---|
1559 | |
---|
1560 | if (GET_CODE (notes) == EXPR_LIST) |
---|
1561 | XEXP (copy, 0) = copy_rtx_and_substitute (XEXP (notes, 0), map); |
---|
1562 | else if (GET_CODE (notes) == INSN_LIST) |
---|
1563 | /* Don't substitute for these yet. */ |
---|
1564 | XEXP (copy, 0) = XEXP (notes, 0); |
---|
1565 | else |
---|
1566 | abort (); |
---|
1567 | |
---|
1568 | XEXP (copy, 1) = initial_reg_note_copy (XEXP (notes, 1), map); |
---|
1569 | |
---|
1570 | return copy; |
---|
1571 | } |
---|
1572 | |
---|
1573 | /* Fixup insn references in copied REG_NOTES. */ |
---|
1574 | |
---|
1575 | static void |
---|
1576 | final_reg_note_copy (notes, map) |
---|
1577 | rtx notes; |
---|
1578 | struct inline_remap *map; |
---|
1579 | { |
---|
1580 | rtx note; |
---|
1581 | |
---|
1582 | for (note = notes; note; note = XEXP (note, 1)) |
---|
1583 | if (GET_CODE (note) == INSN_LIST) |
---|
1584 | XEXP (note, 0) = map->insn_map[INSN_UID (XEXP (note, 0))]; |
---|
1585 | } |
---|
1586 | |
---|
1587 | /* Copy each instruction in the loop, substituting from map as appropriate. |
---|
1588 | This is very similar to a loop in expand_inline_function. */ |
---|
1589 | |
---|
1590 | static void |
---|
1591 | copy_loop_body (copy_start, copy_end, map, exit_label, last_iteration, |
---|
1592 | unroll_type, start_label, loop_end, insert_before, |
---|
1593 | copy_notes_from) |
---|
1594 | rtx copy_start, copy_end; |
---|
1595 | struct inline_remap *map; |
---|
1596 | rtx exit_label; |
---|
1597 | int last_iteration; |
---|
1598 | enum unroll_types unroll_type; |
---|
1599 | rtx start_label, loop_end, insert_before, copy_notes_from; |
---|
1600 | { |
---|
1601 | rtx insn, pattern; |
---|
1602 | rtx set, tem, copy; |
---|
1603 | int dest_reg_was_split, i; |
---|
1604 | rtx cc0_insn = 0; |
---|
1605 | rtx final_label = 0; |
---|
1606 | rtx giv_inc, giv_dest_reg, giv_src_reg; |
---|
1607 | |
---|
1608 | /* If this isn't the last iteration, then map any references to the |
---|
1609 | start_label to final_label. Final label will then be emitted immediately |
---|
1610 | after the end of this loop body if it was ever used. |
---|
1611 | |
---|
1612 | If this is the last iteration, then map references to the start_label |
---|
1613 | to itself. */ |
---|
1614 | if (! last_iteration) |
---|
1615 | { |
---|
1616 | final_label = gen_label_rtx (); |
---|
1617 | set_label_in_map (map, CODE_LABEL_NUMBER (start_label), |
---|
1618 | final_label); |
---|
1619 | } |
---|
1620 | else |
---|
1621 | set_label_in_map (map, CODE_LABEL_NUMBER (start_label), start_label); |
---|
1622 | |
---|
1623 | start_sequence (); |
---|
1624 | |
---|
1625 | insn = copy_start; |
---|
1626 | do |
---|
1627 | { |
---|
1628 | insn = NEXT_INSN (insn); |
---|
1629 | |
---|
1630 | map->orig_asm_operands_vector = 0; |
---|
1631 | |
---|
1632 | switch (GET_CODE (insn)) |
---|
1633 | { |
---|
1634 | case INSN: |
---|
1635 | pattern = PATTERN (insn); |
---|
1636 | copy = 0; |
---|
1637 | giv_inc = 0; |
---|
1638 | |
---|
1639 | /* Check to see if this is a giv that has been combined with |
---|
1640 | some split address givs. (Combined in the sense that |
---|
1641 | `combine_givs' in loop.c has put two givs in the same register.) |
---|
1642 | In this case, we must search all givs based on the same biv to |
---|
1643 | find the address givs. Then split the address givs. |
---|
1644 | Do this before splitting the giv, since that may map the |
---|
1645 | SET_DEST to a new register. */ |
---|
1646 | |
---|
1647 | if ((set = single_set (insn)) |
---|
1648 | && GET_CODE (SET_DEST (set)) == REG |
---|
1649 | && addr_combined_regs[REGNO (SET_DEST (set))]) |
---|
1650 | { |
---|
1651 | struct iv_class *bl; |
---|
1652 | struct induction *v, *tv; |
---|
1653 | int regno = REGNO (SET_DEST (set)); |
---|
1654 | |
---|
1655 | v = addr_combined_regs[REGNO (SET_DEST (set))]; |
---|
1656 | bl = reg_biv_class[REGNO (v->src_reg)]; |
---|
1657 | |
---|
1658 | /* Although the giv_inc amount is not needed here, we must call |
---|
1659 | calculate_giv_inc here since it might try to delete the |
---|
1660 | last insn emitted. If we wait until later to call it, |
---|
1661 | we might accidentally delete insns generated immediately |
---|
1662 | below by emit_unrolled_add. */ |
---|
1663 | |
---|
1664 | giv_inc = calculate_giv_inc (set, insn, regno); |
---|
1665 | |
---|
1666 | /* Now find all address giv's that were combined with this |
---|
1667 | giv 'v'. */ |
---|
1668 | for (tv = bl->giv; tv; tv = tv->next_iv) |
---|
1669 | if (tv->giv_type == DEST_ADDR && tv->same == v) |
---|
1670 | { |
---|
1671 | int this_giv_inc; |
---|
1672 | |
---|
1673 | /* If this DEST_ADDR giv was not split, then ignore it. */ |
---|
1674 | if (*tv->location != tv->dest_reg) |
---|
1675 | continue; |
---|
1676 | |
---|
1677 | /* Scale this_giv_inc if the multiplicative factors of |
---|
1678 | the two givs are different. */ |
---|
1679 | this_giv_inc = INTVAL (giv_inc); |
---|
1680 | if (tv->mult_val != v->mult_val) |
---|
1681 | this_giv_inc = (this_giv_inc / INTVAL (v->mult_val) |
---|
1682 | * INTVAL (tv->mult_val)); |
---|
1683 | |
---|
1684 | tv->dest_reg = plus_constant (tv->dest_reg, this_giv_inc); |
---|
1685 | *tv->location = tv->dest_reg; |
---|
1686 | |
---|
1687 | if (last_iteration && unroll_type != UNROLL_COMPLETELY) |
---|
1688 | { |
---|
1689 | /* Must emit an insn to increment the split address |
---|
1690 | giv. Add in the const_adjust field in case there |
---|
1691 | was a constant eliminated from the address. */ |
---|
1692 | rtx value, dest_reg; |
---|
1693 | |
---|
1694 | /* tv->dest_reg will be either a bare register, |
---|
1695 | or else a register plus a constant. */ |
---|
1696 | if (GET_CODE (tv->dest_reg) == REG) |
---|
1697 | dest_reg = tv->dest_reg; |
---|
1698 | else |
---|
1699 | dest_reg = XEXP (tv->dest_reg, 0); |
---|
1700 | |
---|
1701 | /* Check for shared address givs, and avoid |
---|
1702 | incrementing the shared pseudo reg more than |
---|
1703 | once. */ |
---|
1704 | if (! tv->same_insn) |
---|
1705 | { |
---|
1706 | /* tv->dest_reg may actually be a (PLUS (REG) |
---|
1707 | (CONST)) here, so we must call plus_constant |
---|
1708 | to add the const_adjust amount before calling |
---|
1709 | emit_unrolled_add below. */ |
---|
1710 | value = plus_constant (tv->dest_reg, |
---|
1711 | tv->const_adjust); |
---|
1712 | |
---|
1713 | /* The constant could be too large for an add |
---|
1714 | immediate, so can't directly emit an insn |
---|
1715 | here. */ |
---|
1716 | emit_unrolled_add (dest_reg, XEXP (value, 0), |
---|
1717 | XEXP (value, 1)); |
---|
1718 | } |
---|
1719 | |
---|
1720 | /* Reset the giv to be just the register again, in case |
---|
1721 | it is used after the set we have just emitted. |
---|
1722 | We must subtract the const_adjust factor added in |
---|
1723 | above. */ |
---|
1724 | tv->dest_reg = plus_constant (dest_reg, |
---|
1725 | - tv->const_adjust); |
---|
1726 | *tv->location = tv->dest_reg; |
---|
1727 | } |
---|
1728 | } |
---|
1729 | } |
---|
1730 | |
---|
1731 | /* If this is a setting of a splittable variable, then determine |
---|
1732 | how to split the variable, create a new set based on this split, |
---|
1733 | and set up the reg_map so that later uses of the variable will |
---|
1734 | use the new split variable. */ |
---|
1735 | |
---|
1736 | dest_reg_was_split = 0; |
---|
1737 | |
---|
1738 | if ((set = single_set (insn)) |
---|
1739 | && GET_CODE (SET_DEST (set)) == REG |
---|
1740 | && splittable_regs[REGNO (SET_DEST (set))]) |
---|
1741 | { |
---|
1742 | int regno = REGNO (SET_DEST (set)); |
---|
1743 | |
---|
1744 | dest_reg_was_split = 1; |
---|
1745 | |
---|
1746 | /* Compute the increment value for the giv, if it wasn't |
---|
1747 | already computed above. */ |
---|
1748 | |
---|
1749 | if (giv_inc == 0) |
---|
1750 | giv_inc = calculate_giv_inc (set, insn, regno); |
---|
1751 | giv_dest_reg = SET_DEST (set); |
---|
1752 | giv_src_reg = SET_DEST (set); |
---|
1753 | |
---|
1754 | if (unroll_type == UNROLL_COMPLETELY) |
---|
1755 | { |
---|
1756 | /* Completely unrolling the loop. Set the induction |
---|
1757 | variable to a known constant value. */ |
---|
1758 | |
---|
1759 | /* The value in splittable_regs may be an invariant |
---|
1760 | value, so we must use plus_constant here. */ |
---|
1761 | splittable_regs[regno] |
---|
1762 | = plus_constant (splittable_regs[regno], INTVAL (giv_inc)); |
---|
1763 | |
---|
1764 | if (GET_CODE (splittable_regs[regno]) == PLUS) |
---|
1765 | { |
---|
1766 | giv_src_reg = XEXP (splittable_regs[regno], 0); |
---|
1767 | giv_inc = XEXP (splittable_regs[regno], 1); |
---|
1768 | } |
---|
1769 | else |
---|
1770 | { |
---|
1771 | /* The splittable_regs value must be a REG or a |
---|
1772 | CONST_INT, so put the entire value in the giv_src_reg |
---|
1773 | variable. */ |
---|
1774 | giv_src_reg = splittable_regs[regno]; |
---|
1775 | giv_inc = const0_rtx; |
---|
1776 | } |
---|
1777 | } |
---|
1778 | else |
---|
1779 | { |
---|
1780 | /* Partially unrolling loop. Create a new pseudo |
---|
1781 | register for the iteration variable, and set it to |
---|
1782 | be a constant plus the original register. Except |
---|
1783 | on the last iteration, when the result has to |
---|
1784 | go back into the original iteration var register. */ |
---|
1785 | |
---|
1786 | /* Handle bivs which must be mapped to a new register |
---|
1787 | when split. This happens for bivs which need their |
---|
1788 | final value set before loop entry. The new register |
---|
1789 | for the biv was stored in the biv's first struct |
---|
1790 | induction entry by find_splittable_regs. */ |
---|
1791 | |
---|
1792 | if (regno < max_reg_before_loop |
---|
1793 | && reg_iv_type[regno] == BASIC_INDUCT) |
---|
1794 | { |
---|
1795 | giv_src_reg = reg_biv_class[regno]->biv->src_reg; |
---|
1796 | giv_dest_reg = giv_src_reg; |
---|
1797 | } |
---|
1798 | |
---|
1799 | #if 0 |
---|
1800 | /* If non-reduced/final-value givs were split, then |
---|
1801 | this would have to remap those givs also. See |
---|
1802 | find_splittable_regs. */ |
---|
1803 | #endif |
---|
1804 | |
---|
1805 | splittable_regs[regno] |
---|
1806 | = GEN_INT (INTVAL (giv_inc) |
---|
1807 | + INTVAL (splittable_regs[regno])); |
---|
1808 | giv_inc = splittable_regs[regno]; |
---|
1809 | |
---|
1810 | /* Now split the induction variable by changing the dest |
---|
1811 | of this insn to a new register, and setting its |
---|
1812 | reg_map entry to point to this new register. |
---|
1813 | |
---|
1814 | If this is the last iteration, and this is the last insn |
---|
1815 | that will update the iv, then reuse the original dest, |
---|
1816 | to ensure that the iv will have the proper value when |
---|
1817 | the loop exits or repeats. |
---|
1818 | |
---|
1819 | Using splittable_regs_updates here like this is safe, |
---|
1820 | because it can only be greater than one if all |
---|
1821 | instructions modifying the iv are always executed in |
---|
1822 | order. */ |
---|
1823 | |
---|
1824 | if (! last_iteration |
---|
1825 | || (splittable_regs_updates[regno]-- != 1)) |
---|
1826 | { |
---|
1827 | tem = gen_reg_rtx (GET_MODE (giv_src_reg)); |
---|
1828 | giv_dest_reg = tem; |
---|
1829 | map->reg_map[regno] = tem; |
---|
1830 | } |
---|
1831 | else |
---|
1832 | map->reg_map[regno] = giv_src_reg; |
---|
1833 | } |
---|
1834 | |
---|
1835 | /* The constant being added could be too large for an add |
---|
1836 | immediate, so can't directly emit an insn here. */ |
---|
1837 | emit_unrolled_add (giv_dest_reg, giv_src_reg, giv_inc); |
---|
1838 | copy = get_last_insn (); |
---|
1839 | pattern = PATTERN (copy); |
---|
1840 | } |
---|
1841 | else |
---|
1842 | { |
---|
1843 | pattern = copy_rtx_and_substitute (pattern, map); |
---|
1844 | copy = emit_insn (pattern); |
---|
1845 | } |
---|
1846 | REG_NOTES (copy) = initial_reg_note_copy (REG_NOTES (insn), map); |
---|
1847 | |
---|
1848 | #ifdef HAVE_cc0 |
---|
1849 | /* If this insn is setting CC0, it may need to look at |
---|
1850 | the insn that uses CC0 to see what type of insn it is. |
---|
1851 | In that case, the call to recog via validate_change will |
---|
1852 | fail. So don't substitute constants here. Instead, |
---|
1853 | do it when we emit the following insn. |
---|
1854 | |
---|
1855 | For example, see the pyr.md file. That machine has signed and |
---|
1856 | unsigned compares. The compare patterns must check the |
---|
1857 | following branch insn to see which what kind of compare to |
---|
1858 | emit. |
---|
1859 | |
---|
1860 | If the previous insn set CC0, substitute constants on it as |
---|
1861 | well. */ |
---|
1862 | if (sets_cc0_p (PATTERN (copy)) != 0) |
---|
1863 | cc0_insn = copy; |
---|
1864 | else |
---|
1865 | { |
---|
1866 | if (cc0_insn) |
---|
1867 | try_constants (cc0_insn, map); |
---|
1868 | cc0_insn = 0; |
---|
1869 | try_constants (copy, map); |
---|
1870 | } |
---|
1871 | #else |
---|
1872 | try_constants (copy, map); |
---|
1873 | #endif |
---|
1874 | |
---|
1875 | /* Make split induction variable constants `permanent' since we |
---|
1876 | know there are no backward branches across iteration variable |
---|
1877 | settings which would invalidate this. */ |
---|
1878 | if (dest_reg_was_split) |
---|
1879 | { |
---|
1880 | int regno = REGNO (SET_DEST (pattern)); |
---|
1881 | |
---|
1882 | if (regno < map->const_equiv_map_size |
---|
1883 | && map->const_age_map[regno] == map->const_age) |
---|
1884 | map->const_age_map[regno] = -1; |
---|
1885 | } |
---|
1886 | break; |
---|
1887 | |
---|
1888 | case JUMP_INSN: |
---|
1889 | pattern = copy_rtx_and_substitute (PATTERN (insn), map); |
---|
1890 | copy = emit_jump_insn (pattern); |
---|
1891 | REG_NOTES (copy) = initial_reg_note_copy (REG_NOTES (insn), map); |
---|
1892 | |
---|
1893 | if (JUMP_LABEL (insn) == start_label && insn == copy_end |
---|
1894 | && ! last_iteration) |
---|
1895 | { |
---|
1896 | /* This is a branch to the beginning of the loop; this is the |
---|
1897 | last insn being copied; and this is not the last iteration. |
---|
1898 | In this case, we want to change the original fall through |
---|
1899 | case to be a branch past the end of the loop, and the |
---|
1900 | original jump label case to fall_through. */ |
---|
1901 | |
---|
1902 | if (invert_exp (pattern, copy)) |
---|
1903 | { |
---|
1904 | if (! redirect_exp (&pattern, |
---|
1905 | get_label_from_map (map, |
---|
1906 | CODE_LABEL_NUMBER |
---|
1907 | (JUMP_LABEL (insn))), |
---|
1908 | exit_label, copy)) |
---|
1909 | abort (); |
---|
1910 | } |
---|
1911 | else |
---|
1912 | { |
---|
1913 | rtx jmp; |
---|
1914 | rtx lab = gen_label_rtx (); |
---|
1915 | /* Can't do it by reversing the jump (probably because we |
---|
1916 | couldn't reverse the conditions), so emit a new |
---|
1917 | jump_insn after COPY, and redirect the jump around |
---|
1918 | that. */ |
---|
1919 | jmp = emit_jump_insn_after (gen_jump (exit_label), copy); |
---|
1920 | jmp = emit_barrier_after (jmp); |
---|
1921 | emit_label_after (lab, jmp); |
---|
1922 | LABEL_NUSES (lab) = 0; |
---|
1923 | if (! redirect_exp (&pattern, |
---|
1924 | get_label_from_map (map, |
---|
1925 | CODE_LABEL_NUMBER |
---|
1926 | (JUMP_LABEL (insn))), |
---|
1927 | lab, copy)) |
---|
1928 | abort (); |
---|
1929 | } |
---|
1930 | } |
---|
1931 | |
---|
1932 | #ifdef HAVE_cc0 |
---|
1933 | if (cc0_insn) |
---|
1934 | try_constants (cc0_insn, map); |
---|
1935 | cc0_insn = 0; |
---|
1936 | #endif |
---|
1937 | try_constants (copy, map); |
---|
1938 | |
---|
1939 | /* Set the jump label of COPY correctly to avoid problems with |
---|
1940 | later passes of unroll_loop, if INSN had jump label set. */ |
---|
1941 | if (JUMP_LABEL (insn)) |
---|
1942 | { |
---|
1943 | rtx label = 0; |
---|
1944 | |
---|
1945 | /* Can't use the label_map for every insn, since this may be |
---|
1946 | the backward branch, and hence the label was not mapped. */ |
---|
1947 | if ((set = single_set (copy))) |
---|
1948 | { |
---|
1949 | tem = SET_SRC (set); |
---|
1950 | if (GET_CODE (tem) == LABEL_REF) |
---|
1951 | label = XEXP (tem, 0); |
---|
1952 | else if (GET_CODE (tem) == IF_THEN_ELSE) |
---|
1953 | { |
---|
1954 | if (XEXP (tem, 1) != pc_rtx) |
---|
1955 | label = XEXP (XEXP (tem, 1), 0); |
---|
1956 | else |
---|
1957 | label = XEXP (XEXP (tem, 2), 0); |
---|
1958 | } |
---|
1959 | } |
---|
1960 | |
---|
1961 | if (label && GET_CODE (label) == CODE_LABEL) |
---|
1962 | JUMP_LABEL (copy) = label; |
---|
1963 | else |
---|
1964 | { |
---|
1965 | /* An unrecognizable jump insn, probably the entry jump |
---|
1966 | for a switch statement. This label must have been mapped, |
---|
1967 | so just use the label_map to get the new jump label. */ |
---|
1968 | JUMP_LABEL (copy) |
---|
1969 | = get_label_from_map (map, |
---|
1970 | CODE_LABEL_NUMBER (JUMP_LABEL (insn))); |
---|
1971 | } |
---|
1972 | |
---|
1973 | /* If this is a non-local jump, then must increase the label |
---|
1974 | use count so that the label will not be deleted when the |
---|
1975 | original jump is deleted. */ |
---|
1976 | LABEL_NUSES (JUMP_LABEL (copy))++; |
---|
1977 | } |
---|
1978 | else if (GET_CODE (PATTERN (copy)) == ADDR_VEC |
---|
1979 | || GET_CODE (PATTERN (copy)) == ADDR_DIFF_VEC) |
---|
1980 | { |
---|
1981 | rtx pat = PATTERN (copy); |
---|
1982 | int diff_vec_p = GET_CODE (pat) == ADDR_DIFF_VEC; |
---|
1983 | int len = XVECLEN (pat, diff_vec_p); |
---|
1984 | int i; |
---|
1985 | |
---|
1986 | for (i = 0; i < len; i++) |
---|
1987 | LABEL_NUSES (XEXP (XVECEXP (pat, diff_vec_p, i), 0))++; |
---|
1988 | } |
---|
1989 | |
---|
1990 | /* If this used to be a conditional jump insn but whose branch |
---|
1991 | direction is now known, we must do something special. */ |
---|
1992 | if (condjump_p (insn) && !simplejump_p (insn) && map->last_pc_value) |
---|
1993 | { |
---|
1994 | #ifdef HAVE_cc0 |
---|
1995 | /* The previous insn set cc0 for us. So delete it. */ |
---|
1996 | delete_insn (PREV_INSN (copy)); |
---|
1997 | #endif |
---|
1998 | |
---|
1999 | /* If this is now a no-op, delete it. */ |
---|
2000 | if (map->last_pc_value == pc_rtx) |
---|
2001 | { |
---|
2002 | /* Don't let delete_insn delete the label referenced here, |
---|
2003 | because we might possibly need it later for some other |
---|
2004 | instruction in the loop. */ |
---|
2005 | if (JUMP_LABEL (copy)) |
---|
2006 | LABEL_NUSES (JUMP_LABEL (copy))++; |
---|
2007 | delete_insn (copy); |
---|
2008 | if (JUMP_LABEL (copy)) |
---|
2009 | LABEL_NUSES (JUMP_LABEL (copy))--; |
---|
2010 | copy = 0; |
---|
2011 | } |
---|
2012 | else |
---|
2013 | /* Otherwise, this is unconditional jump so we must put a |
---|
2014 | BARRIER after it. We could do some dead code elimination |
---|
2015 | here, but jump.c will do it just as well. */ |
---|
2016 | emit_barrier (); |
---|
2017 | } |
---|
2018 | break; |
---|
2019 | |
---|
2020 | case CALL_INSN: |
---|
2021 | pattern = copy_rtx_and_substitute (PATTERN (insn), map); |
---|
2022 | copy = emit_call_insn (pattern); |
---|
2023 | REG_NOTES (copy) = initial_reg_note_copy (REG_NOTES (insn), map); |
---|
2024 | |
---|
2025 | /* Because the USAGE information potentially contains objects other |
---|
2026 | than hard registers, we need to copy it. */ |
---|
2027 | CALL_INSN_FUNCTION_USAGE (copy) |
---|
2028 | = copy_rtx_and_substitute (CALL_INSN_FUNCTION_USAGE (insn), map); |
---|
2029 | |
---|
2030 | #ifdef HAVE_cc0 |
---|
2031 | if (cc0_insn) |
---|
2032 | try_constants (cc0_insn, map); |
---|
2033 | cc0_insn = 0; |
---|
2034 | #endif |
---|
2035 | try_constants (copy, map); |
---|
2036 | |
---|
2037 | /* Be lazy and assume CALL_INSNs clobber all hard registers. */ |
---|
2038 | for (i = 0; i < FIRST_PSEUDO_REGISTER; i++) |
---|
2039 | map->const_equiv_map[i] = 0; |
---|
2040 | break; |
---|
2041 | |
---|
2042 | case CODE_LABEL: |
---|
2043 | /* If this is the loop start label, then we don't need to emit a |
---|
2044 | copy of this label since no one will use it. */ |
---|
2045 | |
---|
2046 | if (insn != start_label) |
---|
2047 | { |
---|
2048 | copy = emit_label (get_label_from_map (map, |
---|
2049 | CODE_LABEL_NUMBER (insn))); |
---|
2050 | map->const_age++; |
---|
2051 | } |
---|
2052 | break; |
---|
2053 | |
---|
2054 | case BARRIER: |
---|
2055 | copy = emit_barrier (); |
---|
2056 | break; |
---|
2057 | |
---|
2058 | case NOTE: |
---|
2059 | /* VTOP notes are valid only before the loop exit test. If placed |
---|
2060 | anywhere else, loop may generate bad code. */ |
---|
2061 | |
---|
2062 | if (NOTE_LINE_NUMBER (insn) != NOTE_INSN_DELETED |
---|
2063 | && (NOTE_LINE_NUMBER (insn) != NOTE_INSN_LOOP_VTOP |
---|
2064 | || (last_iteration && unroll_type != UNROLL_COMPLETELY))) |
---|
2065 | copy = emit_note (NOTE_SOURCE_FILE (insn), |
---|
2066 | NOTE_LINE_NUMBER (insn)); |
---|
2067 | else |
---|
2068 | copy = 0; |
---|
2069 | break; |
---|
2070 | |
---|
2071 | default: |
---|
2072 | abort (); |
---|
2073 | break; |
---|
2074 | } |
---|
2075 | |
---|
2076 | map->insn_map[INSN_UID (insn)] = copy; |
---|
2077 | } |
---|
2078 | while (insn != copy_end); |
---|
2079 | |
---|
2080 | /* Now finish coping the REG_NOTES. */ |
---|
2081 | insn = copy_start; |
---|
2082 | do |
---|
2083 | { |
---|
2084 | insn = NEXT_INSN (insn); |
---|
2085 | if ((GET_CODE (insn) == INSN || GET_CODE (insn) == JUMP_INSN |
---|
2086 | || GET_CODE (insn) == CALL_INSN) |
---|
2087 | && map->insn_map[INSN_UID (insn)]) |
---|
2088 | final_reg_note_copy (REG_NOTES (map->insn_map[INSN_UID (insn)]), map); |
---|
2089 | } |
---|
2090 | while (insn != copy_end); |
---|
2091 | |
---|
2092 | /* There may be notes between copy_notes_from and loop_end. Emit a copy of |
---|
2093 | each of these notes here, since there may be some important ones, such as |
---|
2094 | NOTE_INSN_BLOCK_END notes, in this group. We don't do this on the last |
---|
2095 | iteration, because the original notes won't be deleted. |
---|
2096 | |
---|
2097 | We can't use insert_before here, because when from preconditioning, |
---|
2098 | insert_before points before the loop. We can't use copy_end, because |
---|
2099 | there may be insns already inserted after it (which we don't want to |
---|
2100 | copy) when not from preconditioning code. */ |
---|
2101 | |
---|
2102 | if (! last_iteration) |
---|
2103 | { |
---|
2104 | for (insn = copy_notes_from; insn != loop_end; insn = NEXT_INSN (insn)) |
---|
2105 | { |
---|
2106 | if (GET_CODE (insn) == NOTE |
---|
2107 | && NOTE_LINE_NUMBER (insn) != NOTE_INSN_DELETED) |
---|
2108 | emit_note (NOTE_SOURCE_FILE (insn), NOTE_LINE_NUMBER (insn)); |
---|
2109 | } |
---|
2110 | } |
---|
2111 | |
---|
2112 | if (final_label && LABEL_NUSES (final_label) > 0) |
---|
2113 | emit_label (final_label); |
---|
2114 | |
---|
2115 | tem = gen_sequence (); |
---|
2116 | end_sequence (); |
---|
2117 | emit_insn_before (tem, insert_before); |
---|
2118 | } |
---|
2119 | |
---|
2120 | /* Emit an insn, using the expand_binop to ensure that a valid insn is |
---|
2121 | emitted. This will correctly handle the case where the increment value |
---|
2122 | won't fit in the immediate field of a PLUS insns. */ |
---|
2123 | |
---|
2124 | void |
---|
2125 | emit_unrolled_add (dest_reg, src_reg, increment) |
---|
2126 | rtx dest_reg, src_reg, increment; |
---|
2127 | { |
---|
2128 | rtx result; |
---|
2129 | |
---|
2130 | result = expand_binop (GET_MODE (dest_reg), add_optab, src_reg, increment, |
---|
2131 | dest_reg, 0, OPTAB_LIB_WIDEN); |
---|
2132 | |
---|
2133 | if (dest_reg != result) |
---|
2134 | emit_move_insn (dest_reg, result); |
---|
2135 | } |
---|
2136 | |
---|
2137 | /* Searches the insns between INSN and LOOP_END. Returns 1 if there |
---|
2138 | is a backward branch in that range that branches to somewhere between |
---|
2139 | LOOP_START and INSN. Returns 0 otherwise. */ |
---|
2140 | |
---|
2141 | /* ??? This is quadratic algorithm. Could be rewritten to be linear. |
---|
2142 | In practice, this is not a problem, because this function is seldom called, |
---|
2143 | and uses a negligible amount of CPU time on average. */ |
---|
2144 | |
---|
2145 | int |
---|
2146 | back_branch_in_range_p (insn, loop_start, loop_end) |
---|
2147 | rtx insn; |
---|
2148 | rtx loop_start, loop_end; |
---|
2149 | { |
---|
2150 | rtx p, q, target_insn; |
---|
2151 | rtx orig_loop_end = loop_end; |
---|
2152 | |
---|
2153 | /* Stop before we get to the backward branch at the end of the loop. */ |
---|
2154 | loop_end = prev_nonnote_insn (loop_end); |
---|
2155 | if (GET_CODE (loop_end) == BARRIER) |
---|
2156 | loop_end = PREV_INSN (loop_end); |
---|
2157 | |
---|
2158 | /* Check in case insn has been deleted, search forward for first non |
---|
2159 | deleted insn following it. */ |
---|
2160 | while (INSN_DELETED_P (insn)) |
---|
2161 | insn = NEXT_INSN (insn); |
---|
2162 | |
---|
2163 | /* Check for the case where insn is the last insn in the loop. Deal |
---|
2164 | with the case where INSN was a deleted loop test insn, in which case |
---|
2165 | it will now be the NOTE_LOOP_END. */ |
---|
2166 | if (insn == loop_end || insn == orig_loop_end) |
---|
2167 | return 0; |
---|
2168 | |
---|
2169 | for (p = NEXT_INSN (insn); p != loop_end; p = NEXT_INSN (p)) |
---|
2170 | { |
---|
2171 | if (GET_CODE (p) == JUMP_INSN) |
---|
2172 | { |
---|
2173 | target_insn = JUMP_LABEL (p); |
---|
2174 | |
---|
2175 | /* Search from loop_start to insn, to see if one of them is |
---|
2176 | the target_insn. We can't use INSN_LUID comparisons here, |
---|
2177 | since insn may not have an LUID entry. */ |
---|
2178 | for (q = loop_start; q != insn; q = NEXT_INSN (q)) |
---|
2179 | if (q == target_insn) |
---|
2180 | return 1; |
---|
2181 | } |
---|
2182 | } |
---|
2183 | |
---|
2184 | return 0; |
---|
2185 | } |
---|
2186 | |
---|
2187 | /* Try to generate the simplest rtx for the expression |
---|
2188 | (PLUS (MULT mult1 mult2) add1). This is used to calculate the initial |
---|
2189 | value of giv's. */ |
---|
2190 | |
---|
2191 | static rtx |
---|
2192 | fold_rtx_mult_add (mult1, mult2, add1, mode) |
---|
2193 | rtx mult1, mult2, add1; |
---|
2194 | enum machine_mode mode; |
---|
2195 | { |
---|
2196 | rtx temp, mult_res; |
---|
2197 | rtx result; |
---|
2198 | |
---|
2199 | /* The modes must all be the same. This should always be true. For now, |
---|
2200 | check to make sure. */ |
---|
2201 | if ((GET_MODE (mult1) != mode && GET_MODE (mult1) != VOIDmode) |
---|
2202 | || (GET_MODE (mult2) != mode && GET_MODE (mult2) != VOIDmode) |
---|
2203 | || (GET_MODE (add1) != mode && GET_MODE (add1) != VOIDmode)) |
---|
2204 | abort (); |
---|
2205 | |
---|
2206 | /* Ensure that if at least one of mult1/mult2 are constant, then mult2 |
---|
2207 | will be a constant. */ |
---|
2208 | if (GET_CODE (mult1) == CONST_INT) |
---|
2209 | { |
---|
2210 | temp = mult2; |
---|
2211 | mult2 = mult1; |
---|
2212 | mult1 = temp; |
---|
2213 | } |
---|
2214 | |
---|
2215 | mult_res = simplify_binary_operation (MULT, mode, mult1, mult2); |
---|
2216 | if (! mult_res) |
---|
2217 | mult_res = gen_rtx (MULT, mode, mult1, mult2); |
---|
2218 | |
---|
2219 | /* Again, put the constant second. */ |
---|
2220 | if (GET_CODE (add1) == CONST_INT) |
---|
2221 | { |
---|
2222 | temp = add1; |
---|
2223 | add1 = mult_res; |
---|
2224 | mult_res = temp; |
---|
2225 | } |
---|
2226 | |
---|
2227 | result = simplify_binary_operation (PLUS, mode, add1, mult_res); |
---|
2228 | if (! result) |
---|
2229 | result = gen_rtx (PLUS, mode, add1, mult_res); |
---|
2230 | |
---|
2231 | return result; |
---|
2232 | } |
---|
2233 | |
---|
2234 | /* Searches the list of induction struct's for the biv BL, to try to calculate |
---|
2235 | the total increment value for one iteration of the loop as a constant. |
---|
2236 | |
---|
2237 | Returns the increment value as an rtx, simplified as much as possible, |
---|
2238 | if it can be calculated. Otherwise, returns 0. */ |
---|
2239 | |
---|
2240 | rtx |
---|
2241 | biv_total_increment (bl, loop_start, loop_end) |
---|
2242 | struct iv_class *bl; |
---|
2243 | rtx loop_start, loop_end; |
---|
2244 | { |
---|
2245 | struct induction *v; |
---|
2246 | rtx result; |
---|
2247 | |
---|
2248 | /* For increment, must check every instruction that sets it. Each |
---|
2249 | instruction must be executed only once each time through the loop. |
---|
2250 | To verify this, we check that the the insn is always executed, and that |
---|
2251 | there are no backward branches after the insn that branch to before it. |
---|
2252 | Also, the insn must have a mult_val of one (to make sure it really is |
---|
2253 | an increment). */ |
---|
2254 | |
---|
2255 | result = const0_rtx; |
---|
2256 | for (v = bl->biv; v; v = v->next_iv) |
---|
2257 | { |
---|
2258 | if (v->always_computable && v->mult_val == const1_rtx |
---|
2259 | && ! back_branch_in_range_p (v->insn, loop_start, loop_end)) |
---|
2260 | result = fold_rtx_mult_add (result, const1_rtx, v->add_val, v->mode); |
---|
2261 | else |
---|
2262 | return 0; |
---|
2263 | } |
---|
2264 | |
---|
2265 | return result; |
---|
2266 | } |
---|
2267 | |
---|
2268 | /* Determine the initial value of the iteration variable, and the amount |
---|
2269 | that it is incremented each loop. Use the tables constructed by |
---|
2270 | the strength reduction pass to calculate these values. |
---|
2271 | |
---|
2272 | Initial_value and/or increment are set to zero if their values could not |
---|
2273 | be calculated. */ |
---|
2274 | |
---|
2275 | static void |
---|
2276 | iteration_info (iteration_var, initial_value, increment, loop_start, loop_end) |
---|
2277 | rtx iteration_var, *initial_value, *increment; |
---|
2278 | rtx loop_start, loop_end; |
---|
2279 | { |
---|
2280 | struct iv_class *bl; |
---|
2281 | struct induction *v, *b; |
---|
2282 | |
---|
2283 | /* Clear the result values, in case no answer can be found. */ |
---|
2284 | *initial_value = 0; |
---|
2285 | *increment = 0; |
---|
2286 | |
---|
2287 | /* The iteration variable can be either a giv or a biv. Check to see |
---|
2288 | which it is, and compute the variable's initial value, and increment |
---|
2289 | value if possible. */ |
---|
2290 | |
---|
2291 | /* If this is a new register, can't handle it since we don't have any |
---|
2292 | reg_iv_type entry for it. */ |
---|
2293 | if (REGNO (iteration_var) >= max_reg_before_loop) |
---|
2294 | { |
---|
2295 | if (loop_dump_stream) |
---|
2296 | fprintf (loop_dump_stream, |
---|
2297 | "Loop unrolling: No reg_iv_type entry for iteration var.\n"); |
---|
2298 | return; |
---|
2299 | } |
---|
2300 | |
---|
2301 | /* Reject iteration variables larger than the host wide int size, since they |
---|
2302 | could result in a number of iterations greater than the range of our |
---|
2303 | `unsigned HOST_WIDE_INT' variable loop_n_iterations. */ |
---|
2304 | else if ((GET_MODE_BITSIZE (GET_MODE (iteration_var)) |
---|
2305 | > HOST_BITS_PER_WIDE_INT)) |
---|
2306 | { |
---|
2307 | if (loop_dump_stream) |
---|
2308 | fprintf (loop_dump_stream, |
---|
2309 | "Loop unrolling: Iteration var rejected because mode too large.\n"); |
---|
2310 | return; |
---|
2311 | } |
---|
2312 | else if (GET_MODE_CLASS (GET_MODE (iteration_var)) != MODE_INT) |
---|
2313 | { |
---|
2314 | if (loop_dump_stream) |
---|
2315 | fprintf (loop_dump_stream, |
---|
2316 | "Loop unrolling: Iteration var not an integer.\n"); |
---|
2317 | return; |
---|
2318 | } |
---|
2319 | else if (reg_iv_type[REGNO (iteration_var)] == BASIC_INDUCT) |
---|
2320 | { |
---|
2321 | /* Grab initial value, only useful if it is a constant. */ |
---|
2322 | bl = reg_biv_class[REGNO (iteration_var)]; |
---|
2323 | *initial_value = bl->initial_value; |
---|
2324 | |
---|
2325 | *increment = biv_total_increment (bl, loop_start, loop_end); |
---|
2326 | } |
---|
2327 | else if (reg_iv_type[REGNO (iteration_var)] == GENERAL_INDUCT) |
---|
2328 | { |
---|
2329 | #if 1 |
---|
2330 | /* ??? The code below does not work because the incorrect number of |
---|
2331 | iterations is calculated when the biv is incremented after the giv |
---|
2332 | is set (which is the usual case). This can probably be accounted |
---|
2333 | for by biasing the initial_value by subtracting the amount of the |
---|
2334 | increment that occurs between the giv set and the giv test. However, |
---|
2335 | a giv as an iterator is very rare, so it does not seem worthwhile |
---|
2336 | to handle this. */ |
---|
2337 | /* ??? An example failure is: i = 6; do {;} while (i++ < 9). */ |
---|
2338 | if (loop_dump_stream) |
---|
2339 | fprintf (loop_dump_stream, |
---|
2340 | "Loop unrolling: Giv iterators are not handled.\n"); |
---|
2341 | return; |
---|
2342 | #else |
---|
2343 | /* Initial value is mult_val times the biv's initial value plus |
---|
2344 | add_val. Only useful if it is a constant. */ |
---|
2345 | v = reg_iv_info[REGNO (iteration_var)]; |
---|
2346 | bl = reg_biv_class[REGNO (v->src_reg)]; |
---|
2347 | *initial_value = fold_rtx_mult_add (v->mult_val, bl->initial_value, |
---|
2348 | v->add_val, v->mode); |
---|
2349 | |
---|
2350 | /* Increment value is mult_val times the increment value of the biv. */ |
---|
2351 | |
---|
2352 | *increment = biv_total_increment (bl, loop_start, loop_end); |
---|
2353 | if (*increment) |
---|
2354 | *increment = fold_rtx_mult_add (v->mult_val, *increment, const0_rtx, |
---|
2355 | v->mode); |
---|
2356 | #endif |
---|
2357 | } |
---|
2358 | else |
---|
2359 | { |
---|
2360 | if (loop_dump_stream) |
---|
2361 | fprintf (loop_dump_stream, |
---|
2362 | "Loop unrolling: Not basic or general induction var.\n"); |
---|
2363 | return; |
---|
2364 | } |
---|
2365 | } |
---|
2366 | |
---|
2367 | /* Calculate the approximate final value of the iteration variable |
---|
2368 | which has an loop exit test with code COMPARISON_CODE and comparison value |
---|
2369 | of COMPARISON_VALUE. Also returns an indication of whether the comparison |
---|
2370 | was signed or unsigned, and the direction of the comparison. This info is |
---|
2371 | needed to calculate the number of loop iterations. */ |
---|
2372 | |
---|
2373 | static rtx |
---|
2374 | approx_final_value (comparison_code, comparison_value, unsigned_p, compare_dir) |
---|
2375 | enum rtx_code comparison_code; |
---|
2376 | rtx comparison_value; |
---|
2377 | int *unsigned_p; |
---|
2378 | int *compare_dir; |
---|
2379 | { |
---|
2380 | /* Calculate the final value of the induction variable. |
---|
2381 | The exact final value depends on the branch operator, and increment sign. |
---|
2382 | This is only an approximate value. It will be wrong if the iteration |
---|
2383 | variable is not incremented by one each time through the loop, and |
---|
2384 | approx final value - start value % increment != 0. */ |
---|
2385 | |
---|
2386 | *unsigned_p = 0; |
---|
2387 | switch (comparison_code) |
---|
2388 | { |
---|
2389 | case LEU: |
---|
2390 | *unsigned_p = 1; |
---|
2391 | case LE: |
---|
2392 | *compare_dir = 1; |
---|
2393 | return plus_constant (comparison_value, 1); |
---|
2394 | case GEU: |
---|
2395 | *unsigned_p = 1; |
---|
2396 | case GE: |
---|
2397 | *compare_dir = -1; |
---|
2398 | return plus_constant (comparison_value, -1); |
---|
2399 | case EQ: |
---|
2400 | /* Can not calculate a final value for this case. */ |
---|
2401 | *compare_dir = 0; |
---|
2402 | return 0; |
---|
2403 | case LTU: |
---|
2404 | *unsigned_p = 1; |
---|
2405 | case LT: |
---|
2406 | *compare_dir = 1; |
---|
2407 | return comparison_value; |
---|
2408 | break; |
---|
2409 | case GTU: |
---|
2410 | *unsigned_p = 1; |
---|
2411 | case GT: |
---|
2412 | *compare_dir = -1; |
---|
2413 | return comparison_value; |
---|
2414 | case NE: |
---|
2415 | *compare_dir = 0; |
---|
2416 | return comparison_value; |
---|
2417 | default: |
---|
2418 | abort (); |
---|
2419 | } |
---|
2420 | } |
---|
2421 | |
---|
2422 | /* For each biv and giv, determine whether it can be safely split into |
---|
2423 | a different variable for each unrolled copy of the loop body. If it |
---|
2424 | is safe to split, then indicate that by saving some useful info |
---|
2425 | in the splittable_regs array. |
---|
2426 | |
---|
2427 | If the loop is being completely unrolled, then splittable_regs will hold |
---|
2428 | the current value of the induction variable while the loop is unrolled. |
---|
2429 | It must be set to the initial value of the induction variable here. |
---|
2430 | Otherwise, splittable_regs will hold the difference between the current |
---|
2431 | value of the induction variable and the value the induction variable had |
---|
2432 | at the top of the loop. It must be set to the value 0 here. |
---|
2433 | |
---|
2434 | Returns the total number of instructions that set registers that are |
---|
2435 | splittable. */ |
---|
2436 | |
---|
2437 | /* ?? If the loop is only unrolled twice, then most of the restrictions to |
---|
2438 | constant values are unnecessary, since we can easily calculate increment |
---|
2439 | values in this case even if nothing is constant. The increment value |
---|
2440 | should not involve a multiply however. */ |
---|
2441 | |
---|
2442 | /* ?? Even if the biv/giv increment values aren't constant, it may still |
---|
2443 | be beneficial to split the variable if the loop is only unrolled a few |
---|
2444 | times, since multiplies by small integers (1,2,3,4) are very cheap. */ |
---|
2445 | |
---|
2446 | static int |
---|
2447 | find_splittable_regs (unroll_type, loop_start, loop_end, end_insert_before, |
---|
2448 | unroll_number) |
---|
2449 | enum unroll_types unroll_type; |
---|
2450 | rtx loop_start, loop_end; |
---|
2451 | rtx end_insert_before; |
---|
2452 | int unroll_number; |
---|
2453 | { |
---|
2454 | struct iv_class *bl; |
---|
2455 | struct induction *v; |
---|
2456 | rtx increment, tem; |
---|
2457 | rtx biv_final_value; |
---|
2458 | int biv_splittable; |
---|
2459 | int result = 0; |
---|
2460 | |
---|
2461 | for (bl = loop_iv_list; bl; bl = bl->next) |
---|
2462 | { |
---|
2463 | /* Biv_total_increment must return a constant value, |
---|
2464 | otherwise we can not calculate the split values. */ |
---|
2465 | |
---|
2466 | increment = biv_total_increment (bl, loop_start, loop_end); |
---|
2467 | if (! increment || GET_CODE (increment) != CONST_INT) |
---|
2468 | continue; |
---|
2469 | |
---|
2470 | /* The loop must be unrolled completely, or else have a known number |
---|
2471 | of iterations and only one exit, or else the biv must be dead |
---|
2472 | outside the loop, or else the final value must be known. Otherwise, |
---|
2473 | it is unsafe to split the biv since it may not have the proper |
---|
2474 | value on loop exit. */ |
---|
2475 | |
---|
2476 | /* loop_number_exit_count is non-zero if the loop has an exit other than |
---|
2477 | a fall through at the end. */ |
---|
2478 | |
---|
2479 | biv_splittable = 1; |
---|
2480 | biv_final_value = 0; |
---|
2481 | if (unroll_type != UNROLL_COMPLETELY |
---|
2482 | && (loop_number_exit_count[uid_loop_num[INSN_UID (loop_start)]] |
---|
2483 | || unroll_type == UNROLL_NAIVE) |
---|
2484 | && (uid_luid[REGNO_LAST_UID (bl->regno)] >= INSN_LUID (loop_end) |
---|
2485 | || ! bl->init_insn |
---|
2486 | || INSN_UID (bl->init_insn) >= max_uid_for_loop |
---|
2487 | || (uid_luid[REGNO_FIRST_UID (bl->regno)] |
---|
2488 | < INSN_LUID (bl->init_insn)) |
---|
2489 | || reg_mentioned_p (bl->biv->dest_reg, SET_SRC (bl->init_set))) |
---|
2490 | && ! (biv_final_value = final_biv_value (bl, loop_start, loop_end))) |
---|
2491 | biv_splittable = 0; |
---|
2492 | |
---|
2493 | /* If any of the insns setting the BIV don't do so with a simple |
---|
2494 | PLUS, we don't know how to split it. */ |
---|
2495 | for (v = bl->biv; biv_splittable && v; v = v->next_iv) |
---|
2496 | if ((tem = single_set (v->insn)) == 0 |
---|
2497 | || GET_CODE (SET_DEST (tem)) != REG |
---|
2498 | || REGNO (SET_DEST (tem)) != bl->regno |
---|
2499 | || GET_CODE (SET_SRC (tem)) != PLUS) |
---|
2500 | biv_splittable = 0; |
---|
2501 | |
---|
2502 | /* If final value is non-zero, then must emit an instruction which sets |
---|
2503 | the value of the biv to the proper value. This is done after |
---|
2504 | handling all of the givs, since some of them may need to use the |
---|
2505 | biv's value in their initialization code. */ |
---|
2506 | |
---|
2507 | /* This biv is splittable. If completely unrolling the loop, save |
---|
2508 | the biv's initial value. Otherwise, save the constant zero. */ |
---|
2509 | |
---|
2510 | if (biv_splittable == 1) |
---|
2511 | { |
---|
2512 | if (unroll_type == UNROLL_COMPLETELY) |
---|
2513 | { |
---|
2514 | /* If the initial value of the biv is itself (i.e. it is too |
---|
2515 | complicated for strength_reduce to compute), or is a hard |
---|
2516 | register, or it isn't invariant, then we must create a new |
---|
2517 | pseudo reg to hold the initial value of the biv. */ |
---|
2518 | |
---|
2519 | if (GET_CODE (bl->initial_value) == REG |
---|
2520 | && (REGNO (bl->initial_value) == bl->regno |
---|
2521 | || REGNO (bl->initial_value) < FIRST_PSEUDO_REGISTER |
---|
2522 | || ! invariant_p (bl->initial_value))) |
---|
2523 | { |
---|
2524 | rtx tem = gen_reg_rtx (bl->biv->mode); |
---|
2525 | |
---|
2526 | emit_insn_before (gen_move_insn (tem, bl->biv->src_reg), |
---|
2527 | loop_start); |
---|
2528 | |
---|
2529 | if (loop_dump_stream) |
---|
2530 | fprintf (loop_dump_stream, "Biv %d initial value remapped to %d.\n", |
---|
2531 | bl->regno, REGNO (tem)); |
---|
2532 | |
---|
2533 | splittable_regs[bl->regno] = tem; |
---|
2534 | } |
---|
2535 | else |
---|
2536 | splittable_regs[bl->regno] = bl->initial_value; |
---|
2537 | } |
---|
2538 | else |
---|
2539 | splittable_regs[bl->regno] = const0_rtx; |
---|
2540 | |
---|
2541 | /* Save the number of instructions that modify the biv, so that |
---|
2542 | we can treat the last one specially. */ |
---|
2543 | |
---|
2544 | splittable_regs_updates[bl->regno] = bl->biv_count; |
---|
2545 | result += bl->biv_count; |
---|
2546 | |
---|
2547 | if (loop_dump_stream) |
---|
2548 | fprintf (loop_dump_stream, |
---|
2549 | "Biv %d safe to split.\n", bl->regno); |
---|
2550 | } |
---|
2551 | |
---|
2552 | /* Check every giv that depends on this biv to see whether it is |
---|
2553 | splittable also. Even if the biv isn't splittable, givs which |
---|
2554 | depend on it may be splittable if the biv is live outside the |
---|
2555 | loop, and the givs aren't. */ |
---|
2556 | |
---|
2557 | result += find_splittable_givs (bl, unroll_type, loop_start, loop_end, |
---|
2558 | increment, unroll_number); |
---|
2559 | |
---|
2560 | /* If final value is non-zero, then must emit an instruction which sets |
---|
2561 | the value of the biv to the proper value. This is done after |
---|
2562 | handling all of the givs, since some of them may need to use the |
---|
2563 | biv's value in their initialization code. */ |
---|
2564 | if (biv_final_value) |
---|
2565 | { |
---|
2566 | /* If the loop has multiple exits, emit the insns before the |
---|
2567 | loop to ensure that it will always be executed no matter |
---|
2568 | how the loop exits. Otherwise emit the insn after the loop, |
---|
2569 | since this is slightly more efficient. */ |
---|
2570 | if (! loop_number_exit_count[uid_loop_num[INSN_UID (loop_start)]]) |
---|
2571 | emit_insn_before (gen_move_insn (bl->biv->src_reg, |
---|
2572 | biv_final_value), |
---|
2573 | end_insert_before); |
---|
2574 | else |
---|
2575 | { |
---|
2576 | /* Create a new register to hold the value of the biv, and then |
---|
2577 | set the biv to its final value before the loop start. The biv |
---|
2578 | is set to its final value before loop start to ensure that |
---|
2579 | this insn will always be executed, no matter how the loop |
---|
2580 | exits. */ |
---|
2581 | rtx tem = gen_reg_rtx (bl->biv->mode); |
---|
2582 | emit_insn_before (gen_move_insn (tem, bl->biv->src_reg), |
---|
2583 | loop_start); |
---|
2584 | emit_insn_before (gen_move_insn (bl->biv->src_reg, |
---|
2585 | biv_final_value), |
---|
2586 | loop_start); |
---|
2587 | |
---|
2588 | if (loop_dump_stream) |
---|
2589 | fprintf (loop_dump_stream, "Biv %d mapped to %d for split.\n", |
---|
2590 | REGNO (bl->biv->src_reg), REGNO (tem)); |
---|
2591 | |
---|
2592 | /* Set up the mapping from the original biv register to the new |
---|
2593 | register. */ |
---|
2594 | bl->biv->src_reg = tem; |
---|
2595 | } |
---|
2596 | } |
---|
2597 | } |
---|
2598 | return result; |
---|
2599 | } |
---|
2600 | |
---|
2601 | /* Return 1 if the first and last unrolled copy of the address giv V is valid |
---|
2602 | for the instruction that is using it. Do not make any changes to that |
---|
2603 | instruction. */ |
---|
2604 | |
---|
2605 | static int |
---|
2606 | verify_addresses (v, giv_inc, unroll_number) |
---|
2607 | struct induction *v; |
---|
2608 | rtx giv_inc; |
---|
2609 | int unroll_number; |
---|
2610 | { |
---|
2611 | int ret = 1; |
---|
2612 | rtx orig_addr = *v->location; |
---|
2613 | rtx last_addr = plus_constant (v->dest_reg, |
---|
2614 | INTVAL (giv_inc) * (unroll_number - 1)); |
---|
2615 | |
---|
2616 | /* First check to see if either address would fail. */ |
---|
2617 | if (! validate_change (v->insn, v->location, v->dest_reg, 0) |
---|
2618 | || ! validate_change (v->insn, v->location, last_addr, 0)) |
---|
2619 | ret = 0; |
---|
2620 | |
---|
2621 | /* Now put things back the way they were before. This will always |
---|
2622 | succeed. */ |
---|
2623 | validate_change (v->insn, v->location, orig_addr, 0); |
---|
2624 | |
---|
2625 | return ret; |
---|
2626 | } |
---|
2627 | |
---|
2628 | /* For every giv based on the biv BL, check to determine whether it is |
---|
2629 | splittable. This is a subroutine to find_splittable_regs (). |
---|
2630 | |
---|
2631 | Return the number of instructions that set splittable registers. */ |
---|
2632 | |
---|
2633 | static int |
---|
2634 | find_splittable_givs (bl, unroll_type, loop_start, loop_end, increment, |
---|
2635 | unroll_number) |
---|
2636 | struct iv_class *bl; |
---|
2637 | enum unroll_types unroll_type; |
---|
2638 | rtx loop_start, loop_end; |
---|
2639 | rtx increment; |
---|
2640 | int unroll_number; |
---|
2641 | { |
---|
2642 | struct induction *v, *v2; |
---|
2643 | rtx final_value; |
---|
2644 | rtx tem; |
---|
2645 | int result = 0; |
---|
2646 | |
---|
2647 | /* Scan the list of givs, and set the same_insn field when there are |
---|
2648 | multiple identical givs in the same insn. */ |
---|
2649 | for (v = bl->giv; v; v = v->next_iv) |
---|
2650 | for (v2 = v->next_iv; v2; v2 = v2->next_iv) |
---|
2651 | if (v->insn == v2->insn && rtx_equal_p (v->new_reg, v2->new_reg) |
---|
2652 | && ! v2->same_insn) |
---|
2653 | v2->same_insn = v; |
---|
2654 | |
---|
2655 | for (v = bl->giv; v; v = v->next_iv) |
---|
2656 | { |
---|
2657 | rtx giv_inc, value; |
---|
2658 | |
---|
2659 | /* Only split the giv if it has already been reduced, or if the loop is |
---|
2660 | being completely unrolled. */ |
---|
2661 | if (unroll_type != UNROLL_COMPLETELY && v->ignore) |
---|
2662 | continue; |
---|
2663 | |
---|
2664 | /* The giv can be split if the insn that sets the giv is executed once |
---|
2665 | and only once on every iteration of the loop. */ |
---|
2666 | /* An address giv can always be split. v->insn is just a use not a set, |
---|
2667 | and hence it does not matter whether it is always executed. All that |
---|
2668 | matters is that all the biv increments are always executed, and we |
---|
2669 | won't reach here if they aren't. */ |
---|
2670 | if (v->giv_type != DEST_ADDR |
---|
2671 | && (! v->always_computable |
---|
2672 | || back_branch_in_range_p (v->insn, loop_start, loop_end))) |
---|
2673 | continue; |
---|
2674 | |
---|
2675 | /* The giv increment value must be a constant. */ |
---|
2676 | giv_inc = fold_rtx_mult_add (v->mult_val, increment, const0_rtx, |
---|
2677 | v->mode); |
---|
2678 | if (! giv_inc || GET_CODE (giv_inc) != CONST_INT) |
---|
2679 | continue; |
---|
2680 | |
---|
2681 | /* The loop must be unrolled completely, or else have a known number of |
---|
2682 | iterations and only one exit, or else the giv must be dead outside |
---|
2683 | the loop, or else the final value of the giv must be known. |
---|
2684 | Otherwise, it is not safe to split the giv since it may not have the |
---|
2685 | proper value on loop exit. */ |
---|
2686 | |
---|
2687 | /* The used outside loop test will fail for DEST_ADDR givs. They are |
---|
2688 | never used outside the loop anyways, so it is always safe to split a |
---|
2689 | DEST_ADDR giv. */ |
---|
2690 | |
---|
2691 | final_value = 0; |
---|
2692 | if (unroll_type != UNROLL_COMPLETELY |
---|
2693 | && (loop_number_exit_count[uid_loop_num[INSN_UID (loop_start)]] |
---|
2694 | || unroll_type == UNROLL_NAIVE) |
---|
2695 | && v->giv_type != DEST_ADDR |
---|
2696 | /* The next part is true if the pseudo is used outside the loop. |
---|
2697 | We assume that this is true for any pseudo created after loop |
---|
2698 | starts, because we don't have a reg_n_info entry for them. */ |
---|
2699 | && (REGNO (v->dest_reg) >= max_reg_before_loop |
---|
2700 | || (REGNO_FIRST_UID (REGNO (v->dest_reg)) != INSN_UID (v->insn) |
---|
2701 | /* Check for the case where the pseudo is set by a shift/add |
---|
2702 | sequence, in which case the first insn setting the pseudo |
---|
2703 | is the first insn of the shift/add sequence. */ |
---|
2704 | && (! (tem = find_reg_note (v->insn, REG_RETVAL, NULL_RTX)) |
---|
2705 | || (REGNO_FIRST_UID (REGNO (v->dest_reg)) |
---|
2706 | != INSN_UID (XEXP (tem, 0))))) |
---|
2707 | /* Line above always fails if INSN was moved by loop opt. */ |
---|
2708 | || (uid_luid[REGNO_LAST_UID (REGNO (v->dest_reg))] |
---|
2709 | >= INSN_LUID (loop_end))) |
---|
2710 | && ! (final_value = v->final_value)) |
---|
2711 | continue; |
---|
2712 | |
---|
2713 | #if 0 |
---|
2714 | /* Currently, non-reduced/final-value givs are never split. */ |
---|
2715 | /* Should emit insns after the loop if possible, as the biv final value |
---|
2716 | code below does. */ |
---|
2717 | |
---|
2718 | /* If the final value is non-zero, and the giv has not been reduced, |
---|
2719 | then must emit an instruction to set the final value. */ |
---|
2720 | if (final_value && !v->new_reg) |
---|
2721 | { |
---|
2722 | /* Create a new register to hold the value of the giv, and then set |
---|
2723 | the giv to its final value before the loop start. The giv is set |
---|
2724 | to its final value before loop start to ensure that this insn |
---|
2725 | will always be executed, no matter how we exit. */ |
---|
2726 | tem = gen_reg_rtx (v->mode); |
---|
2727 | emit_insn_before (gen_move_insn (tem, v->dest_reg), loop_start); |
---|
2728 | emit_insn_before (gen_move_insn (v->dest_reg, final_value), |
---|
2729 | loop_start); |
---|
2730 | |
---|
2731 | if (loop_dump_stream) |
---|
2732 | fprintf (loop_dump_stream, "Giv %d mapped to %d for split.\n", |
---|
2733 | REGNO (v->dest_reg), REGNO (tem)); |
---|
2734 | |
---|
2735 | v->src_reg = tem; |
---|
2736 | } |
---|
2737 | #endif |
---|
2738 | |
---|
2739 | /* This giv is splittable. If completely unrolling the loop, save the |
---|
2740 | giv's initial value. Otherwise, save the constant zero for it. */ |
---|
2741 | |
---|
2742 | if (unroll_type == UNROLL_COMPLETELY) |
---|
2743 | { |
---|
2744 | /* It is not safe to use bl->initial_value here, because it may not |
---|
2745 | be invariant. It is safe to use the initial value stored in |
---|
2746 | the splittable_regs array if it is set. In rare cases, it won't |
---|
2747 | be set, so then we do exactly the same thing as |
---|
2748 | find_splittable_regs does to get a safe value. */ |
---|
2749 | rtx biv_initial_value; |
---|
2750 | |
---|
2751 | if (splittable_regs[bl->regno]) |
---|
2752 | biv_initial_value = splittable_regs[bl->regno]; |
---|
2753 | else if (GET_CODE (bl->initial_value) != REG |
---|
2754 | || (REGNO (bl->initial_value) != bl->regno |
---|
2755 | && REGNO (bl->initial_value) >= FIRST_PSEUDO_REGISTER)) |
---|
2756 | biv_initial_value = bl->initial_value; |
---|
2757 | else |
---|
2758 | { |
---|
2759 | rtx tem = gen_reg_rtx (bl->biv->mode); |
---|
2760 | |
---|
2761 | emit_insn_before (gen_move_insn (tem, bl->biv->src_reg), |
---|
2762 | loop_start); |
---|
2763 | biv_initial_value = tem; |
---|
2764 | } |
---|
2765 | value = fold_rtx_mult_add (v->mult_val, biv_initial_value, |
---|
2766 | v->add_val, v->mode); |
---|
2767 | } |
---|
2768 | else |
---|
2769 | value = const0_rtx; |
---|
2770 | |
---|
2771 | if (v->new_reg) |
---|
2772 | { |
---|
2773 | /* If a giv was combined with another giv, then we can only split |
---|
2774 | this giv if the giv it was combined with was reduced. This |
---|
2775 | is because the value of v->new_reg is meaningless in this |
---|
2776 | case. */ |
---|
2777 | if (v->same && ! v->same->new_reg) |
---|
2778 | { |
---|
2779 | if (loop_dump_stream) |
---|
2780 | fprintf (loop_dump_stream, |
---|
2781 | "giv combined with unreduced giv not split.\n"); |
---|
2782 | continue; |
---|
2783 | } |
---|
2784 | /* If the giv is an address destination, it could be something other |
---|
2785 | than a simple register, these have to be treated differently. */ |
---|
2786 | else if (v->giv_type == DEST_REG) |
---|
2787 | { |
---|
2788 | /* If value is not a constant, register, or register plus |
---|
2789 | constant, then compute its value into a register before |
---|
2790 | loop start. This prevents invalid rtx sharing, and should |
---|
2791 | generate better code. We can use bl->initial_value here |
---|
2792 | instead of splittable_regs[bl->regno] because this code |
---|
2793 | is going before the loop start. */ |
---|
2794 | if (unroll_type == UNROLL_COMPLETELY |
---|
2795 | && GET_CODE (value) != CONST_INT |
---|
2796 | && GET_CODE (value) != REG |
---|
2797 | && (GET_CODE (value) != PLUS |
---|
2798 | || GET_CODE (XEXP (value, 0)) != REG |
---|
2799 | || GET_CODE (XEXP (value, 1)) != CONST_INT)) |
---|
2800 | { |
---|
2801 | rtx tem = gen_reg_rtx (v->mode); |
---|
2802 | emit_iv_add_mult (bl->initial_value, v->mult_val, |
---|
2803 | v->add_val, tem, loop_start); |
---|
2804 | value = tem; |
---|
2805 | } |
---|
2806 | |
---|
2807 | splittable_regs[REGNO (v->new_reg)] = value; |
---|
2808 | } |
---|
2809 | else |
---|
2810 | { |
---|
2811 | /* Splitting address givs is useful since it will often allow us |
---|
2812 | to eliminate some increment insns for the base giv as |
---|
2813 | unnecessary. */ |
---|
2814 | |
---|
2815 | /* If the addr giv is combined with a dest_reg giv, then all |
---|
2816 | references to that dest reg will be remapped, which is NOT |
---|
2817 | what we want for split addr regs. We always create a new |
---|
2818 | register for the split addr giv, just to be safe. */ |
---|
2819 | |
---|
2820 | /* ??? If there are multiple address givs which have been |
---|
2821 | combined with the same dest_reg giv, then we may only need |
---|
2822 | one new register for them. Pulling out constants below will |
---|
2823 | catch some of the common cases of this. Currently, I leave |
---|
2824 | the work of simplifying multiple address givs to the |
---|
2825 | following cse pass. */ |
---|
2826 | |
---|
2827 | /* As a special case, if we have multiple identical address givs |
---|
2828 | within a single instruction, then we do use a single pseudo |
---|
2829 | reg for both. This is necessary in case one is a match_dup |
---|
2830 | of the other. */ |
---|
2831 | |
---|
2832 | v->const_adjust = 0; |
---|
2833 | |
---|
2834 | if (v->same_insn) |
---|
2835 | { |
---|
2836 | v->dest_reg = v->same_insn->dest_reg; |
---|
2837 | if (loop_dump_stream) |
---|
2838 | fprintf (loop_dump_stream, |
---|
2839 | "Sharing address givs in insn %d\n", |
---|
2840 | INSN_UID (v->insn)); |
---|
2841 | } |
---|
2842 | else if (unroll_type != UNROLL_COMPLETELY) |
---|
2843 | { |
---|
2844 | /* If not completely unrolling the loop, then create a new |
---|
2845 | register to hold the split value of the DEST_ADDR giv. |
---|
2846 | Emit insn to initialize its value before loop start. */ |
---|
2847 | tem = gen_reg_rtx (v->mode); |
---|
2848 | |
---|
2849 | /* If the address giv has a constant in its new_reg value, |
---|
2850 | then this constant can be pulled out and put in value, |
---|
2851 | instead of being part of the initialization code. */ |
---|
2852 | |
---|
2853 | if (GET_CODE (v->new_reg) == PLUS |
---|
2854 | && GET_CODE (XEXP (v->new_reg, 1)) == CONST_INT) |
---|
2855 | { |
---|
2856 | v->dest_reg |
---|
2857 | = plus_constant (tem, INTVAL (XEXP (v->new_reg,1))); |
---|
2858 | |
---|
2859 | /* Only succeed if this will give valid addresses. |
---|
2860 | Try to validate both the first and the last |
---|
2861 | address resulting from loop unrolling, if |
---|
2862 | one fails, then can't do const elim here. */ |
---|
2863 | if (verify_addresses (v, giv_inc, unroll_number)) |
---|
2864 | { |
---|
2865 | /* Save the negative of the eliminated const, so |
---|
2866 | that we can calculate the dest_reg's increment |
---|
2867 | value later. */ |
---|
2868 | v->const_adjust = - INTVAL (XEXP (v->new_reg, 1)); |
---|
2869 | |
---|
2870 | v->new_reg = XEXP (v->new_reg, 0); |
---|
2871 | if (loop_dump_stream) |
---|
2872 | fprintf (loop_dump_stream, |
---|
2873 | "Eliminating constant from giv %d\n", |
---|
2874 | REGNO (tem)); |
---|
2875 | } |
---|
2876 | else |
---|
2877 | v->dest_reg = tem; |
---|
2878 | } |
---|
2879 | else |
---|
2880 | v->dest_reg = tem; |
---|
2881 | |
---|
2882 | /* If the address hasn't been checked for validity yet, do so |
---|
2883 | now, and fail completely if either the first or the last |
---|
2884 | unrolled copy of the address is not a valid address |
---|
2885 | for the instruction that uses it. */ |
---|
2886 | if (v->dest_reg == tem |
---|
2887 | && ! verify_addresses (v, giv_inc, unroll_number)) |
---|
2888 | { |
---|
2889 | if (loop_dump_stream) |
---|
2890 | fprintf (loop_dump_stream, |
---|
2891 | "Invalid address for giv at insn %d\n", |
---|
2892 | INSN_UID (v->insn)); |
---|
2893 | continue; |
---|
2894 | } |
---|
2895 | |
---|
2896 | /* To initialize the new register, just move the value of |
---|
2897 | new_reg into it. This is not guaranteed to give a valid |
---|
2898 | instruction on machines with complex addressing modes. |
---|
2899 | If we can't recognize it, then delete it and emit insns |
---|
2900 | to calculate the value from scratch. */ |
---|
2901 | emit_insn_before (gen_rtx (SET, VOIDmode, tem, |
---|
2902 | copy_rtx (v->new_reg)), |
---|
2903 | loop_start); |
---|
2904 | if (recog_memoized (PREV_INSN (loop_start)) < 0) |
---|
2905 | { |
---|
2906 | rtx sequence, ret; |
---|
2907 | |
---|
2908 | /* We can't use bl->initial_value to compute the initial |
---|
2909 | value, because the loop may have been preconditioned. |
---|
2910 | We must calculate it from NEW_REG. Try using |
---|
2911 | force_operand instead of emit_iv_add_mult. */ |
---|
2912 | delete_insn (PREV_INSN (loop_start)); |
---|
2913 | |
---|
2914 | start_sequence (); |
---|
2915 | ret = force_operand (v->new_reg, tem); |
---|
2916 | if (ret != tem) |
---|
2917 | emit_move_insn (tem, ret); |
---|
2918 | sequence = gen_sequence (); |
---|
2919 | end_sequence (); |
---|
2920 | emit_insn_before (sequence, loop_start); |
---|
2921 | |
---|
2922 | if (loop_dump_stream) |
---|
2923 | fprintf (loop_dump_stream, |
---|
2924 | "Invalid init insn, rewritten.\n"); |
---|
2925 | } |
---|
2926 | } |
---|
2927 | else |
---|
2928 | { |
---|
2929 | v->dest_reg = value; |
---|
2930 | |
---|
2931 | /* Check the resulting address for validity, and fail |
---|
2932 | if the resulting address would be invalid. */ |
---|
2933 | if (! verify_addresses (v, giv_inc, unroll_number)) |
---|
2934 | { |
---|
2935 | if (loop_dump_stream) |
---|
2936 | fprintf (loop_dump_stream, |
---|
2937 | "Invalid address for giv at insn %d\n", |
---|
2938 | INSN_UID (v->insn)); |
---|
2939 | continue; |
---|
2940 | } |
---|
2941 | } |
---|
2942 | |
---|
2943 | /* Store the value of dest_reg into the insn. This sharing |
---|
2944 | will not be a problem as this insn will always be copied |
---|
2945 | later. */ |
---|
2946 | |
---|
2947 | *v->location = v->dest_reg; |
---|
2948 | |
---|
2949 | /* If this address giv is combined with a dest reg giv, then |
---|
2950 | save the base giv's induction pointer so that we will be |
---|
2951 | able to handle this address giv properly. The base giv |
---|
2952 | itself does not have to be splittable. */ |
---|
2953 | |
---|
2954 | if (v->same && v->same->giv_type == DEST_REG) |
---|
2955 | addr_combined_regs[REGNO (v->same->new_reg)] = v->same; |
---|
2956 | |
---|
2957 | if (GET_CODE (v->new_reg) == REG) |
---|
2958 | { |
---|
2959 | /* This giv maybe hasn't been combined with any others. |
---|
2960 | Make sure that it's giv is marked as splittable here. */ |
---|
2961 | |
---|
2962 | splittable_regs[REGNO (v->new_reg)] = value; |
---|
2963 | |
---|
2964 | /* Make it appear to depend upon itself, so that the |
---|
2965 | giv will be properly split in the main loop above. */ |
---|
2966 | if (! v->same) |
---|
2967 | { |
---|
2968 | v->same = v; |
---|
2969 | addr_combined_regs[REGNO (v->new_reg)] = v; |
---|
2970 | } |
---|
2971 | } |
---|
2972 | |
---|
2973 | if (loop_dump_stream) |
---|
2974 | fprintf (loop_dump_stream, "DEST_ADDR giv being split.\n"); |
---|
2975 | } |
---|
2976 | } |
---|
2977 | else |
---|
2978 | { |
---|
2979 | #if 0 |
---|
2980 | /* Currently, unreduced giv's can't be split. This is not too much |
---|
2981 | of a problem since unreduced giv's are not live across loop |
---|
2982 | iterations anyways. When unrolling a loop completely though, |
---|
2983 | it makes sense to reduce&split givs when possible, as this will |
---|
2984 | result in simpler instructions, and will not require that a reg |
---|
2985 | be live across loop iterations. */ |
---|
2986 | |
---|
2987 | splittable_regs[REGNO (v->dest_reg)] = value; |
---|
2988 | fprintf (stderr, "Giv %d at insn %d not reduced\n", |
---|
2989 | REGNO (v->dest_reg), INSN_UID (v->insn)); |
---|
2990 | #else |
---|
2991 | continue; |
---|
2992 | #endif |
---|
2993 | } |
---|
2994 | |
---|
2995 | /* Unreduced givs are only updated once by definition. Reduced givs |
---|
2996 | are updated as many times as their biv is. Mark it so if this is |
---|
2997 | a splittable register. Don't need to do anything for address givs |
---|
2998 | where this may not be a register. */ |
---|
2999 | |
---|
3000 | if (GET_CODE (v->new_reg) == REG) |
---|
3001 | { |
---|
3002 | int count = 1; |
---|
3003 | if (! v->ignore) |
---|
3004 | count = reg_biv_class[REGNO (v->src_reg)]->biv_count; |
---|
3005 | |
---|
3006 | splittable_regs_updates[REGNO (v->new_reg)] = count; |
---|
3007 | } |
---|
3008 | |
---|
3009 | result++; |
---|
3010 | |
---|
3011 | if (loop_dump_stream) |
---|
3012 | { |
---|
3013 | int regnum; |
---|
3014 | |
---|
3015 | if (GET_CODE (v->dest_reg) == CONST_INT) |
---|
3016 | regnum = -1; |
---|
3017 | else if (GET_CODE (v->dest_reg) != REG) |
---|
3018 | regnum = REGNO (XEXP (v->dest_reg, 0)); |
---|
3019 | else |
---|
3020 | regnum = REGNO (v->dest_reg); |
---|
3021 | fprintf (loop_dump_stream, "Giv %d at insn %d safe to split.\n", |
---|
3022 | regnum, INSN_UID (v->insn)); |
---|
3023 | } |
---|
3024 | } |
---|
3025 | |
---|
3026 | return result; |
---|
3027 | } |
---|
3028 | |
---|
3029 | /* Try to prove that the register is dead after the loop exits. Trace every |
---|
3030 | loop exit looking for an insn that will always be executed, which sets |
---|
3031 | the register to some value, and appears before the first use of the register |
---|
3032 | is found. If successful, then return 1, otherwise return 0. */ |
---|
3033 | |
---|
3034 | /* ?? Could be made more intelligent in the handling of jumps, so that |
---|
3035 | it can search past if statements and other similar structures. */ |
---|
3036 | |
---|
3037 | static int |
---|
3038 | reg_dead_after_loop (reg, loop_start, loop_end) |
---|
3039 | rtx reg, loop_start, loop_end; |
---|
3040 | { |
---|
3041 | rtx insn, label; |
---|
3042 | enum rtx_code code; |
---|
3043 | int jump_count = 0; |
---|
3044 | int label_count = 0; |
---|
3045 | int this_loop_num = uid_loop_num[INSN_UID (loop_start)]; |
---|
3046 | |
---|
3047 | /* In addition to checking all exits of this loop, we must also check |
---|
3048 | all exits of inner nested loops that would exit this loop. We don't |
---|
3049 | have any way to identify those, so we just give up if there are any |
---|
3050 | such inner loop exits. */ |
---|
3051 | |
---|
3052 | for (label = loop_number_exit_labels[this_loop_num]; label; |
---|
3053 | label = LABEL_NEXTREF (label)) |
---|
3054 | label_count++; |
---|
3055 | |
---|
3056 | if (label_count != loop_number_exit_count[this_loop_num]) |
---|
3057 | return 0; |
---|
3058 | |
---|
3059 | /* HACK: Must also search the loop fall through exit, create a label_ref |
---|
3060 | here which points to the loop_end, and append the loop_number_exit_labels |
---|
3061 | list to it. */ |
---|
3062 | label = gen_rtx (LABEL_REF, VOIDmode, loop_end); |
---|
3063 | LABEL_NEXTREF (label) = loop_number_exit_labels[this_loop_num]; |
---|
3064 | |
---|
3065 | for ( ; label; label = LABEL_NEXTREF (label)) |
---|
3066 | { |
---|
3067 | /* Succeed if find an insn which sets the biv or if reach end of |
---|
3068 | function. Fail if find an insn that uses the biv, or if come to |
---|
3069 | a conditional jump. */ |
---|
3070 | |
---|
3071 | insn = NEXT_INSN (XEXP (label, 0)); |
---|
3072 | while (insn) |
---|
3073 | { |
---|
3074 | code = GET_CODE (insn); |
---|
3075 | if (GET_RTX_CLASS (code) == 'i') |
---|
3076 | { |
---|
3077 | rtx set; |
---|
3078 | |
---|
3079 | if (reg_referenced_p (reg, PATTERN (insn))) |
---|
3080 | return 0; |
---|
3081 | |
---|
3082 | set = single_set (insn); |
---|
3083 | if (set && rtx_equal_p (SET_DEST (set), reg)) |
---|
3084 | break; |
---|
3085 | } |
---|
3086 | |
---|
3087 | if (code == JUMP_INSN) |
---|
3088 | { |
---|
3089 | if (GET_CODE (PATTERN (insn)) == RETURN) |
---|
3090 | break; |
---|
3091 | else if (! simplejump_p (insn) |
---|
3092 | /* Prevent infinite loop following infinite loops. */ |
---|
3093 | || jump_count++ > 20) |
---|
3094 | return 0; |
---|
3095 | else |
---|
3096 | insn = JUMP_LABEL (insn); |
---|
3097 | } |
---|
3098 | |
---|
3099 | insn = NEXT_INSN (insn); |
---|
3100 | } |
---|
3101 | } |
---|
3102 | |
---|
3103 | /* Success, the register is dead on all loop exits. */ |
---|
3104 | return 1; |
---|
3105 | } |
---|
3106 | |
---|
3107 | /* Try to calculate the final value of the biv, the value it will have at |
---|
3108 | the end of the loop. If we can do it, return that value. */ |
---|
3109 | |
---|
3110 | rtx |
---|
3111 | final_biv_value (bl, loop_start, loop_end) |
---|
3112 | struct iv_class *bl; |
---|
3113 | rtx loop_start, loop_end; |
---|
3114 | { |
---|
3115 | rtx increment, tem; |
---|
3116 | |
---|
3117 | /* ??? This only works for MODE_INT biv's. Reject all others for now. */ |
---|
3118 | |
---|
3119 | if (GET_MODE_CLASS (bl->biv->mode) != MODE_INT) |
---|
3120 | return 0; |
---|
3121 | |
---|
3122 | /* The final value for reversed bivs must be calculated differently than |
---|
3123 | for ordinary bivs. In this case, there is already an insn after the |
---|
3124 | loop which sets this biv's final value (if necessary), and there are |
---|
3125 | no other loop exits, so we can return any value. */ |
---|
3126 | if (bl->reversed) |
---|
3127 | { |
---|
3128 | if (loop_dump_stream) |
---|
3129 | fprintf (loop_dump_stream, |
---|
3130 | "Final biv value for %d, reversed biv.\n", bl->regno); |
---|
3131 | |
---|
3132 | return const0_rtx; |
---|
3133 | } |
---|
3134 | |
---|
3135 | /* Try to calculate the final value as initial value + (number of iterations |
---|
3136 | * increment). For this to work, increment must be invariant, the only |
---|
3137 | exit from the loop must be the fall through at the bottom (otherwise |
---|
3138 | it may not have its final value when the loop exits), and the initial |
---|
3139 | value of the biv must be invariant. */ |
---|
3140 | |
---|
3141 | if (loop_n_iterations != 0 |
---|
3142 | && ! loop_number_exit_count[uid_loop_num[INSN_UID (loop_start)]] |
---|
3143 | && invariant_p (bl->initial_value)) |
---|
3144 | { |
---|
3145 | increment = biv_total_increment (bl, loop_start, loop_end); |
---|
3146 | |
---|
3147 | if (increment && invariant_p (increment)) |
---|
3148 | { |
---|
3149 | /* Can calculate the loop exit value, emit insns after loop |
---|
3150 | end to calculate this value into a temporary register in |
---|
3151 | case it is needed later. */ |
---|
3152 | |
---|
3153 | tem = gen_reg_rtx (bl->biv->mode); |
---|
3154 | /* Make sure loop_end is not the last insn. */ |
---|
3155 | if (NEXT_INSN (loop_end) == 0) |
---|
3156 | emit_note_after (NOTE_INSN_DELETED, loop_end); |
---|
3157 | emit_iv_add_mult (increment, GEN_INT (loop_n_iterations), |
---|
3158 | bl->initial_value, tem, NEXT_INSN (loop_end)); |
---|
3159 | |
---|
3160 | if (loop_dump_stream) |
---|
3161 | fprintf (loop_dump_stream, |
---|
3162 | "Final biv value for %d, calculated.\n", bl->regno); |
---|
3163 | |
---|
3164 | return tem; |
---|
3165 | } |
---|
3166 | } |
---|
3167 | |
---|
3168 | /* Check to see if the biv is dead at all loop exits. */ |
---|
3169 | if (reg_dead_after_loop (bl->biv->src_reg, loop_start, loop_end)) |
---|
3170 | { |
---|
3171 | if (loop_dump_stream) |
---|
3172 | fprintf (loop_dump_stream, |
---|
3173 | "Final biv value for %d, biv dead after loop exit.\n", |
---|
3174 | bl->regno); |
---|
3175 | |
---|
3176 | return const0_rtx; |
---|
3177 | } |
---|
3178 | |
---|
3179 | return 0; |
---|
3180 | } |
---|
3181 | |
---|
3182 | /* Try to calculate the final value of the giv, the value it will have at |
---|
3183 | the end of the loop. If we can do it, return that value. */ |
---|
3184 | |
---|
3185 | rtx |
---|
3186 | final_giv_value (v, loop_start, loop_end) |
---|
3187 | struct induction *v; |
---|
3188 | rtx loop_start, loop_end; |
---|
3189 | { |
---|
3190 | struct iv_class *bl; |
---|
3191 | rtx insn; |
---|
3192 | rtx increment, tem; |
---|
3193 | rtx insert_before, seq; |
---|
3194 | |
---|
3195 | bl = reg_biv_class[REGNO (v->src_reg)]; |
---|
3196 | |
---|
3197 | /* The final value for givs which depend on reversed bivs must be calculated |
---|
3198 | differently than for ordinary givs. In this case, there is already an |
---|
3199 | insn after the loop which sets this giv's final value (if necessary), |
---|
3200 | and there are no other loop exits, so we can return any value. */ |
---|
3201 | if (bl->reversed) |
---|
3202 | { |
---|
3203 | if (loop_dump_stream) |
---|
3204 | fprintf (loop_dump_stream, |
---|
3205 | "Final giv value for %d, depends on reversed biv\n", |
---|
3206 | REGNO (v->dest_reg)); |
---|
3207 | return const0_rtx; |
---|
3208 | } |
---|
3209 | |
---|
3210 | /* Try to calculate the final value as a function of the biv it depends |
---|
3211 | upon. The only exit from the loop must be the fall through at the bottom |
---|
3212 | (otherwise it may not have its final value when the loop exits). */ |
---|
3213 | |
---|
3214 | /* ??? Can calculate the final giv value by subtracting off the |
---|
3215 | extra biv increments times the giv's mult_val. The loop must have |
---|
3216 | only one exit for this to work, but the loop iterations does not need |
---|
3217 | to be known. */ |
---|
3218 | |
---|
3219 | if (loop_n_iterations != 0 |
---|
3220 | && ! loop_number_exit_count[uid_loop_num[INSN_UID (loop_start)]]) |
---|
3221 | { |
---|
3222 | /* ?? It is tempting to use the biv's value here since these insns will |
---|
3223 | be put after the loop, and hence the biv will have its final value |
---|
3224 | then. However, this fails if the biv is subsequently eliminated. |
---|
3225 | Perhaps determine whether biv's are eliminable before trying to |
---|
3226 | determine whether giv's are replaceable so that we can use the |
---|
3227 | biv value here if it is not eliminable. */ |
---|
3228 | |
---|
3229 | /* We are emitting code after the end of the loop, so we must make |
---|
3230 | sure that bl->initial_value is still valid then. It will still |
---|
3231 | be valid if it is invariant. */ |
---|
3232 | |
---|
3233 | increment = biv_total_increment (bl, loop_start, loop_end); |
---|
3234 | |
---|
3235 | if (increment && invariant_p (increment) |
---|
3236 | && invariant_p (bl->initial_value)) |
---|
3237 | { |
---|
3238 | /* Can calculate the loop exit value of its biv as |
---|
3239 | (loop_n_iterations * increment) + initial_value */ |
---|
3240 | |
---|
3241 | /* The loop exit value of the giv is then |
---|
3242 | (final_biv_value - extra increments) * mult_val + add_val. |
---|
3243 | The extra increments are any increments to the biv which |
---|
3244 | occur in the loop after the giv's value is calculated. |
---|
3245 | We must search from the insn that sets the giv to the end |
---|
3246 | of the loop to calculate this value. */ |
---|
3247 | |
---|
3248 | insert_before = NEXT_INSN (loop_end); |
---|
3249 | |
---|
3250 | /* Put the final biv value in tem. */ |
---|
3251 | tem = gen_reg_rtx (bl->biv->mode); |
---|
3252 | emit_iv_add_mult (increment, GEN_INT (loop_n_iterations), |
---|
3253 | bl->initial_value, tem, insert_before); |
---|
3254 | |
---|
3255 | /* Subtract off extra increments as we find them. */ |
---|
3256 | for (insn = NEXT_INSN (v->insn); insn != loop_end; |
---|
3257 | insn = NEXT_INSN (insn)) |
---|
3258 | { |
---|
3259 | struct induction *biv; |
---|
3260 | |
---|
3261 | for (biv = bl->biv; biv; biv = biv->next_iv) |
---|
3262 | if (biv->insn == insn) |
---|
3263 | { |
---|
3264 | start_sequence (); |
---|
3265 | tem = expand_binop (GET_MODE (tem), sub_optab, tem, |
---|
3266 | biv->add_val, NULL_RTX, 0, |
---|
3267 | OPTAB_LIB_WIDEN); |
---|
3268 | seq = gen_sequence (); |
---|
3269 | end_sequence (); |
---|
3270 | emit_insn_before (seq, insert_before); |
---|
3271 | } |
---|
3272 | } |
---|
3273 | |
---|
3274 | /* Now calculate the giv's final value. */ |
---|
3275 | emit_iv_add_mult (tem, v->mult_val, v->add_val, tem, |
---|
3276 | insert_before); |
---|
3277 | |
---|
3278 | if (loop_dump_stream) |
---|
3279 | fprintf (loop_dump_stream, |
---|
3280 | "Final giv value for %d, calc from biv's value.\n", |
---|
3281 | REGNO (v->dest_reg)); |
---|
3282 | |
---|
3283 | return tem; |
---|
3284 | } |
---|
3285 | } |
---|
3286 | |
---|
3287 | /* Replaceable giv's should never reach here. */ |
---|
3288 | if (v->replaceable) |
---|
3289 | abort (); |
---|
3290 | |
---|
3291 | /* Check to see if the biv is dead at all loop exits. */ |
---|
3292 | if (reg_dead_after_loop (v->dest_reg, loop_start, loop_end)) |
---|
3293 | { |
---|
3294 | if (loop_dump_stream) |
---|
3295 | fprintf (loop_dump_stream, |
---|
3296 | "Final giv value for %d, giv dead after loop exit.\n", |
---|
3297 | REGNO (v->dest_reg)); |
---|
3298 | |
---|
3299 | return const0_rtx; |
---|
3300 | } |
---|
3301 | |
---|
3302 | return 0; |
---|
3303 | } |
---|
3304 | |
---|
3305 | |
---|
3306 | /* Calculate the number of loop iterations. Returns the exact number of loop |
---|
3307 | iterations if it can be calculated, otherwise returns zero. */ |
---|
3308 | |
---|
3309 | unsigned HOST_WIDE_INT |
---|
3310 | loop_iterations (loop_start, loop_end) |
---|
3311 | rtx loop_start, loop_end; |
---|
3312 | { |
---|
3313 | rtx comparison, comparison_value; |
---|
3314 | rtx iteration_var, initial_value, increment, final_value; |
---|
3315 | enum rtx_code comparison_code; |
---|
3316 | HOST_WIDE_INT i; |
---|
3317 | int increment_dir; |
---|
3318 | int unsigned_compare, compare_dir, final_larger; |
---|
3319 | unsigned long tempu; |
---|
3320 | rtx last_loop_insn; |
---|
3321 | |
---|
3322 | /* First find the iteration variable. If the last insn is a conditional |
---|
3323 | branch, and the insn before tests a register value, make that the |
---|
3324 | iteration variable. */ |
---|
3325 | |
---|
3326 | loop_initial_value = 0; |
---|
3327 | loop_increment = 0; |
---|
3328 | loop_final_value = 0; |
---|
3329 | loop_iteration_var = 0; |
---|
3330 | |
---|
3331 | /* We used to use pren_nonnote_insn here, but that fails because it might |
---|
3332 | accidentally get the branch for a contained loop if the branch for this |
---|
3333 | loop was deleted. We can only trust branches immediately before the |
---|
3334 | loop_end. */ |
---|
3335 | last_loop_insn = PREV_INSN (loop_end); |
---|
3336 | |
---|
3337 | comparison = get_condition_for_loop (last_loop_insn); |
---|
3338 | if (comparison == 0) |
---|
3339 | { |
---|
3340 | if (loop_dump_stream) |
---|
3341 | fprintf (loop_dump_stream, |
---|
3342 | "Loop unrolling: No final conditional branch found.\n"); |
---|
3343 | return 0; |
---|
3344 | } |
---|
3345 | |
---|
3346 | /* ??? Get_condition may switch position of induction variable and |
---|
3347 | invariant register when it canonicalizes the comparison. */ |
---|
3348 | |
---|
3349 | comparison_code = GET_CODE (comparison); |
---|
3350 | iteration_var = XEXP (comparison, 0); |
---|
3351 | comparison_value = XEXP (comparison, 1); |
---|
3352 | |
---|
3353 | if (GET_CODE (iteration_var) != REG) |
---|
3354 | { |
---|
3355 | if (loop_dump_stream) |
---|
3356 | fprintf (loop_dump_stream, |
---|
3357 | "Loop unrolling: Comparison not against register.\n"); |
---|
3358 | return 0; |
---|
3359 | } |
---|
3360 | |
---|
3361 | /* Loop iterations is always called before any new registers are created |
---|
3362 | now, so this should never occur. */ |
---|
3363 | |
---|
3364 | if (REGNO (iteration_var) >= max_reg_before_loop) |
---|
3365 | abort (); |
---|
3366 | |
---|
3367 | iteration_info (iteration_var, &initial_value, &increment, |
---|
3368 | loop_start, loop_end); |
---|
3369 | if (initial_value == 0) |
---|
3370 | /* iteration_info already printed a message. */ |
---|
3371 | return 0; |
---|
3372 | |
---|
3373 | /* If the comparison value is an invariant register, then try to find |
---|
3374 | its value from the insns before the start of the loop. */ |
---|
3375 | |
---|
3376 | if (GET_CODE (comparison_value) == REG && invariant_p (comparison_value)) |
---|
3377 | { |
---|
3378 | rtx insn, set; |
---|
3379 | |
---|
3380 | for (insn = PREV_INSN (loop_start); insn ; insn = PREV_INSN (insn)) |
---|
3381 | { |
---|
3382 | if (GET_CODE (insn) == CODE_LABEL) |
---|
3383 | break; |
---|
3384 | |
---|
3385 | else if (GET_RTX_CLASS (GET_CODE (insn)) == 'i' |
---|
3386 | && reg_set_p (comparison_value, insn)) |
---|
3387 | { |
---|
3388 | /* We found the last insn before the loop that sets the register. |
---|
3389 | If it sets the entire register, and has a REG_EQUAL note, |
---|
3390 | then use the value of the REG_EQUAL note. */ |
---|
3391 | if ((set = single_set (insn)) |
---|
3392 | && (SET_DEST (set) == comparison_value)) |
---|
3393 | { |
---|
3394 | rtx note = find_reg_note (insn, REG_EQUAL, NULL_RTX); |
---|
3395 | |
---|
3396 | /* Only use the REG_EQUAL note if it is a constant. |
---|
3397 | Other things, divide in particular, will cause |
---|
3398 | problems later if we use them. */ |
---|
3399 | if (note && GET_CODE (XEXP (note, 0)) != EXPR_LIST |
---|
3400 | && CONSTANT_P (XEXP (note, 0))) |
---|
3401 | comparison_value = XEXP (note, 0); |
---|
3402 | } |
---|
3403 | break; |
---|
3404 | } |
---|
3405 | } |
---|
3406 | } |
---|
3407 | |
---|
3408 | final_value = approx_final_value (comparison_code, comparison_value, |
---|
3409 | &unsigned_compare, &compare_dir); |
---|
3410 | |
---|
3411 | /* Save the calculated values describing this loop's bounds, in case |
---|
3412 | precondition_loop_p will need them later. These values can not be |
---|
3413 | recalculated inside precondition_loop_p because strength reduction |
---|
3414 | optimizations may obscure the loop's structure. */ |
---|
3415 | |
---|
3416 | loop_iteration_var = iteration_var; |
---|
3417 | loop_initial_value = initial_value; |
---|
3418 | loop_increment = increment; |
---|
3419 | loop_final_value = final_value; |
---|
3420 | loop_comparison_code = comparison_code; |
---|
3421 | |
---|
3422 | if (increment == 0) |
---|
3423 | { |
---|
3424 | if (loop_dump_stream) |
---|
3425 | fprintf (loop_dump_stream, |
---|
3426 | "Loop unrolling: Increment value can't be calculated.\n"); |
---|
3427 | return 0; |
---|
3428 | } |
---|
3429 | else if (GET_CODE (increment) != CONST_INT) |
---|
3430 | { |
---|
3431 | if (loop_dump_stream) |
---|
3432 | fprintf (loop_dump_stream, |
---|
3433 | "Loop unrolling: Increment value not constant.\n"); |
---|
3434 | return 0; |
---|
3435 | } |
---|
3436 | else if (GET_CODE (initial_value) != CONST_INT) |
---|
3437 | { |
---|
3438 | if (loop_dump_stream) |
---|
3439 | fprintf (loop_dump_stream, |
---|
3440 | "Loop unrolling: Initial value not constant.\n"); |
---|
3441 | return 0; |
---|
3442 | } |
---|
3443 | else if (final_value == 0) |
---|
3444 | { |
---|
3445 | if (loop_dump_stream) |
---|
3446 | fprintf (loop_dump_stream, |
---|
3447 | "Loop unrolling: EQ comparison loop.\n"); |
---|
3448 | return 0; |
---|
3449 | } |
---|
3450 | else if (GET_CODE (final_value) != CONST_INT) |
---|
3451 | { |
---|
3452 | if (loop_dump_stream) |
---|
3453 | fprintf (loop_dump_stream, |
---|
3454 | "Loop unrolling: Final value not constant.\n"); |
---|
3455 | return 0; |
---|
3456 | } |
---|
3457 | |
---|
3458 | /* ?? Final value and initial value do not have to be constants. |
---|
3459 | Only their difference has to be constant. When the iteration variable |
---|
3460 | is an array address, the final value and initial value might both |
---|
3461 | be addresses with the same base but different constant offsets. |
---|
3462 | Final value must be invariant for this to work. |
---|
3463 | |
---|
3464 | To do this, need some way to find the values of registers which are |
---|
3465 | invariant. */ |
---|
3466 | |
---|
3467 | /* Final_larger is 1 if final larger, 0 if they are equal, otherwise -1. */ |
---|
3468 | if (unsigned_compare) |
---|
3469 | final_larger |
---|
3470 | = ((unsigned HOST_WIDE_INT) INTVAL (final_value) |
---|
3471 | > (unsigned HOST_WIDE_INT) INTVAL (initial_value)) |
---|
3472 | - ((unsigned HOST_WIDE_INT) INTVAL (final_value) |
---|
3473 | < (unsigned HOST_WIDE_INT) INTVAL (initial_value)); |
---|
3474 | else |
---|
3475 | final_larger = (INTVAL (final_value) > INTVAL (initial_value)) |
---|
3476 | - (INTVAL (final_value) < INTVAL (initial_value)); |
---|
3477 | |
---|
3478 | if (INTVAL (increment) > 0) |
---|
3479 | increment_dir = 1; |
---|
3480 | else if (INTVAL (increment) == 0) |
---|
3481 | increment_dir = 0; |
---|
3482 | else |
---|
3483 | increment_dir = -1; |
---|
3484 | |
---|
3485 | /* There are 27 different cases: compare_dir = -1, 0, 1; |
---|
3486 | final_larger = -1, 0, 1; increment_dir = -1, 0, 1. |
---|
3487 | There are 4 normal cases, 4 reverse cases (where the iteration variable |
---|
3488 | will overflow before the loop exits), 4 infinite loop cases, and 15 |
---|
3489 | immediate exit (0 or 1 iteration depending on loop type) cases. |
---|
3490 | Only try to optimize the normal cases. */ |
---|
3491 | |
---|
3492 | /* (compare_dir/final_larger/increment_dir) |
---|
3493 | Normal cases: (0/-1/-1), (0/1/1), (-1/-1/-1), (1/1/1) |
---|
3494 | Reverse cases: (0/-1/1), (0/1/-1), (-1/-1/1), (1/1/-1) |
---|
3495 | Infinite loops: (0/-1/0), (0/1/0), (-1/-1/0), (1/1/0) |
---|
3496 | Immediate exit: (0/0/X), (-1/0/X), (-1/1/X), (1/0/X), (1/-1/X) */ |
---|
3497 | |
---|
3498 | /* ?? If the meaning of reverse loops (where the iteration variable |
---|
3499 | will overflow before the loop exits) is undefined, then could |
---|
3500 | eliminate all of these special checks, and just always assume |
---|
3501 | the loops are normal/immediate/infinite. Note that this means |
---|
3502 | the sign of increment_dir does not have to be known. Also, |
---|
3503 | since it does not really hurt if immediate exit loops or infinite loops |
---|
3504 | are optimized, then that case could be ignored also, and hence all |
---|
3505 | loops can be optimized. |
---|
3506 | |
---|
3507 | According to ANSI Spec, the reverse loop case result is undefined, |
---|
3508 | because the action on overflow is undefined. |
---|
3509 | |
---|
3510 | See also the special test for NE loops below. */ |
---|
3511 | |
---|
3512 | if (final_larger == increment_dir && final_larger != 0 |
---|
3513 | && (final_larger == compare_dir || compare_dir == 0)) |
---|
3514 | /* Normal case. */ |
---|
3515 | ; |
---|
3516 | else |
---|
3517 | { |
---|
3518 | if (loop_dump_stream) |
---|
3519 | fprintf (loop_dump_stream, |
---|
3520 | "Loop unrolling: Not normal loop.\n"); |
---|
3521 | return 0; |
---|
3522 | } |
---|
3523 | |
---|
3524 | /* Calculate the number of iterations, final_value is only an approximation, |
---|
3525 | so correct for that. Note that tempu and loop_n_iterations are |
---|
3526 | unsigned, because they can be as large as 2^n - 1. */ |
---|
3527 | |
---|
3528 | i = INTVAL (increment); |
---|
3529 | if (i > 0) |
---|
3530 | tempu = INTVAL (final_value) - INTVAL (initial_value); |
---|
3531 | else if (i < 0) |
---|
3532 | { |
---|
3533 | tempu = INTVAL (initial_value) - INTVAL (final_value); |
---|
3534 | i = -i; |
---|
3535 | } |
---|
3536 | else |
---|
3537 | abort (); |
---|
3538 | |
---|
3539 | /* For NE tests, make sure that the iteration variable won't miss the |
---|
3540 | final value. If tempu mod i is not zero, then the iteration variable |
---|
3541 | will overflow before the loop exits, and we can not calculate the |
---|
3542 | number of iterations. */ |
---|
3543 | if (compare_dir == 0 && (tempu % i) != 0) |
---|
3544 | return 0; |
---|
3545 | |
---|
3546 | return tempu / i + ((tempu % i) != 0); |
---|
3547 | } |
---|
3548 | |
---|
3549 | /* Replace uses of split bivs with their split pseudo register. This is |
---|
3550 | for original instructions which remain after loop unrolling without |
---|
3551 | copying. */ |
---|
3552 | |
---|
3553 | static rtx |
---|
3554 | remap_split_bivs (x) |
---|
3555 | rtx x; |
---|
3556 | { |
---|
3557 | register enum rtx_code code; |
---|
3558 | register int i; |
---|
3559 | register char *fmt; |
---|
3560 | |
---|
3561 | if (x == 0) |
---|
3562 | return x; |
---|
3563 | |
---|
3564 | code = GET_CODE (x); |
---|
3565 | switch (code) |
---|
3566 | { |
---|
3567 | case SCRATCH: |
---|
3568 | case PC: |
---|
3569 | case CC0: |
---|
3570 | case CONST_INT: |
---|
3571 | case CONST_DOUBLE: |
---|
3572 | case CONST: |
---|
3573 | case SYMBOL_REF: |
---|
3574 | case LABEL_REF: |
---|
3575 | return x; |
---|
3576 | |
---|
3577 | case REG: |
---|
3578 | #if 0 |
---|
3579 | /* If non-reduced/final-value givs were split, then this would also |
---|
3580 | have to remap those givs also. */ |
---|
3581 | #endif |
---|
3582 | if (REGNO (x) < max_reg_before_loop |
---|
3583 | && reg_iv_type[REGNO (x)] == BASIC_INDUCT) |
---|
3584 | return reg_biv_class[REGNO (x)]->biv->src_reg; |
---|
3585 | break; |
---|
3586 | |
---|
3587 | default: |
---|
3588 | break; |
---|
3589 | } |
---|
3590 | |
---|
3591 | fmt = GET_RTX_FORMAT (code); |
---|
3592 | for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) |
---|
3593 | { |
---|
3594 | if (fmt[i] == 'e') |
---|
3595 | XEXP (x, i) = remap_split_bivs (XEXP (x, i)); |
---|
3596 | if (fmt[i] == 'E') |
---|
3597 | { |
---|
3598 | register int j; |
---|
3599 | for (j = 0; j < XVECLEN (x, i); j++) |
---|
3600 | XVECEXP (x, i, j) = remap_split_bivs (XVECEXP (x, i, j)); |
---|
3601 | } |
---|
3602 | } |
---|
3603 | return x; |
---|
3604 | } |
---|
3605 | |
---|
3606 | /* If FIRST_UID is a set of REGNO, and FIRST_UID dominates LAST_UID (e.g. |
---|
3607 | FIST_UID is always executed if LAST_UID is), then return 1. Otherwise |
---|
3608 | return 0. COPY_START is where we can start looking for the insns |
---|
3609 | FIRST_UID and LAST_UID. COPY_END is where we stop looking for these |
---|
3610 | insns. |
---|
3611 | |
---|
3612 | If there is no JUMP_INSN between LOOP_START and FIRST_UID, then FIRST_UID |
---|
3613 | must dominate LAST_UID. |
---|
3614 | |
---|
3615 | If there is a CODE_LABEL between FIRST_UID and LAST_UID, then FIRST_UID |
---|
3616 | may not dominate LAST_UID. |
---|
3617 | |
---|
3618 | If there is no CODE_LABEL between FIRST_UID and LAST_UID, then FIRST_UID |
---|
3619 | must dominate LAST_UID. */ |
---|
3620 | |
---|
3621 | int |
---|
3622 | set_dominates_use (regno, first_uid, last_uid, copy_start, copy_end) |
---|
3623 | int regno; |
---|
3624 | int first_uid; |
---|
3625 | int last_uid; |
---|
3626 | rtx copy_start; |
---|
3627 | rtx copy_end; |
---|
3628 | { |
---|
3629 | int passed_jump = 0; |
---|
3630 | rtx p = NEXT_INSN (copy_start); |
---|
3631 | |
---|
3632 | while (INSN_UID (p) != first_uid) |
---|
3633 | { |
---|
3634 | if (GET_CODE (p) == JUMP_INSN) |
---|
3635 | passed_jump= 1; |
---|
3636 | /* Could not find FIRST_UID. */ |
---|
3637 | if (p == copy_end) |
---|
3638 | return 0; |
---|
3639 | p = NEXT_INSN (p); |
---|
3640 | } |
---|
3641 | |
---|
3642 | /* Verify that FIRST_UID is an insn that entirely sets REGNO. */ |
---|
3643 | if (GET_RTX_CLASS (GET_CODE (p)) != 'i' |
---|
3644 | || ! dead_or_set_regno_p (p, regno)) |
---|
3645 | return 0; |
---|
3646 | |
---|
3647 | /* FIRST_UID is always executed. */ |
---|
3648 | if (passed_jump == 0) |
---|
3649 | return 1; |
---|
3650 | |
---|
3651 | while (INSN_UID (p) != last_uid) |
---|
3652 | { |
---|
3653 | /* If we see a CODE_LABEL between FIRST_UID and LAST_UID, then we |
---|
3654 | can not be sure that FIRST_UID dominates LAST_UID. */ |
---|
3655 | if (GET_CODE (p) == CODE_LABEL) |
---|
3656 | return 0; |
---|
3657 | /* Could not find LAST_UID, but we reached the end of the loop, so |
---|
3658 | it must be safe. */ |
---|
3659 | else if (p == copy_end) |
---|
3660 | return 1; |
---|
3661 | p = NEXT_INSN (p); |
---|
3662 | } |
---|
3663 | |
---|
3664 | /* FIRST_UID is always executed if LAST_UID is executed. */ |
---|
3665 | return 1; |
---|
3666 | } |
---|