]>
Commit | Line | Data |
---|---|---|
e9eb809d | 1 | /* Functions to determine/estimate number of iterations of a loop. |
cbe34bb5 | 2 | Copyright (C) 2004-2017 Free Software Foundation, Inc. |
b8698a0f | 3 | |
e9eb809d | 4 | This file is part of GCC. |
b8698a0f | 5 | |
e9eb809d ZD |
6 | GCC is free software; you can redistribute it and/or modify it |
7 | under the terms of the GNU General Public License as published by the | |
9dcd6f09 | 8 | Free Software Foundation; either version 3, or (at your option) any |
e9eb809d | 9 | later version. |
b8698a0f | 10 | |
e9eb809d ZD |
11 | GCC is distributed in the hope that it will be useful, but WITHOUT |
12 | ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
13 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
14 | for more details. | |
b8698a0f | 15 | |
e9eb809d | 16 | You should have received a copy of the GNU General Public License |
9dcd6f09 NC |
17 | along with GCC; see the file COPYING3. If not see |
18 | <http://www.gnu.org/licenses/>. */ | |
e9eb809d ZD |
19 | |
20 | #include "config.h" | |
21 | #include "system.h" | |
22 | #include "coretypes.h" | |
c7131fb2 | 23 | #include "backend.h" |
957060b5 | 24 | #include "rtl.h" |
e9eb809d | 25 | #include "tree.h" |
c7131fb2 | 26 | #include "gimple.h" |
957060b5 | 27 | #include "tree-pass.h" |
c7131fb2 | 28 | #include "ssa.h" |
957060b5 AM |
29 | #include "gimple-pretty-print.h" |
30 | #include "diagnostic-core.h" | |
2f07b722 | 31 | #include "stor-layout.h" |
40e23961 | 32 | #include "fold-const.h" |
d8a2d370 | 33 | #include "calls.h" |
f9cc1a70 | 34 | #include "intl.h" |
45b0be94 | 35 | #include "gimplify.h" |
5be5c238 | 36 | #include "gimple-iterator.h" |
442b4905 | 37 | #include "tree-cfg.h" |
e28030cf AM |
38 | #include "tree-ssa-loop-ivopts.h" |
39 | #include "tree-ssa-loop-niter.h" | |
442b4905 | 40 | #include "tree-ssa-loop.h" |
e9eb809d | 41 | #include "cfgloop.h" |
e9eb809d ZD |
42 | #include "tree-chrec.h" |
43 | #include "tree-scalar-evolution.h" | |
44 | #include "params.h" | |
e9eb809d | 45 | |
71343877 | 46 | |
b3ce5b6e ZD |
47 | /* The maximum number of dominator BBs we search for conditions |
48 | of loop header copies we use for simplifying a conditional | |
49 | expression. */ | |
50 | #define MAX_DOMINATORS_TO_WALK 8 | |
e9eb809d ZD |
51 | |
52 | /* | |
53 | ||
54 | Analysis of number of iterations of an affine exit test. | |
55 | ||
56 | */ | |
57 | ||
b3ce5b6e ZD |
58 | /* Bounds on some value, BELOW <= X <= UP. */ |
59 | ||
a79683d5 | 60 | struct bounds |
b3ce5b6e ZD |
61 | { |
62 | mpz_t below, up; | |
a79683d5 | 63 | }; |
b3ce5b6e | 64 | |
b3ce5b6e ZD |
65 | |
66 | /* Splits expression EXPR to a variable part VAR and constant OFFSET. */ | |
67 | ||
68 | static void | |
69 | split_to_var_and_offset (tree expr, tree *var, mpz_t offset) | |
70 | { | |
71 | tree type = TREE_TYPE (expr); | |
72 | tree op0, op1; | |
b3ce5b6e ZD |
73 | bool negate = false; |
74 | ||
75 | *var = expr; | |
76 | mpz_set_ui (offset, 0); | |
77 | ||
78 | switch (TREE_CODE (expr)) | |
79 | { | |
80 | case MINUS_EXPR: | |
81 | negate = true; | |
82 | /* Fallthru. */ | |
83 | ||
84 | case PLUS_EXPR: | |
5be014d5 | 85 | case POINTER_PLUS_EXPR: |
b3ce5b6e ZD |
86 | op0 = TREE_OPERAND (expr, 0); |
87 | op1 = TREE_OPERAND (expr, 1); | |
88 | ||
89 | if (TREE_CODE (op1) != INTEGER_CST) | |
90 | break; | |
91 | ||
92 | *var = op0; | |
93 | /* Always sign extend the offset. */ | |
807e902e | 94 | wi::to_mpz (op1, offset, SIGNED); |
eab1da69 UB |
95 | if (negate) |
96 | mpz_neg (offset, offset); | |
b3ce5b6e ZD |
97 | break; |
98 | ||
99 | case INTEGER_CST: | |
100 | *var = build_int_cst_type (type, 0); | |
807e902e | 101 | wi::to_mpz (expr, offset, TYPE_SIGN (type)); |
b3ce5b6e ZD |
102 | break; |
103 | ||
104 | default: | |
105 | break; | |
106 | } | |
107 | } | |
108 | ||
7b008bbc BC |
109 | /* From condition C0 CMP C1 derives information regarding the value range |
110 | of VAR, which is of TYPE. Results are stored in to BELOW and UP. */ | |
111 | ||
112 | static void | |
113 | refine_value_range_using_guard (tree type, tree var, | |
114 | tree c0, enum tree_code cmp, tree c1, | |
115 | mpz_t below, mpz_t up) | |
116 | { | |
117 | tree varc0, varc1, ctype; | |
118 | mpz_t offc0, offc1; | |
119 | mpz_t mint, maxt, minc1, maxc1; | |
120 | wide_int minv, maxv; | |
121 | bool no_wrap = nowrap_type_p (type); | |
122 | bool c0_ok, c1_ok; | |
123 | signop sgn = TYPE_SIGN (type); | |
124 | ||
125 | switch (cmp) | |
126 | { | |
127 | case LT_EXPR: | |
128 | case LE_EXPR: | |
129 | case GT_EXPR: | |
130 | case GE_EXPR: | |
131 | STRIP_SIGN_NOPS (c0); | |
132 | STRIP_SIGN_NOPS (c1); | |
133 | ctype = TREE_TYPE (c0); | |
134 | if (!useless_type_conversion_p (ctype, type)) | |
135 | return; | |
136 | ||
137 | break; | |
138 | ||
139 | case EQ_EXPR: | |
140 | /* We could derive quite precise information from EQ_EXPR, however, | |
141 | such a guard is unlikely to appear, so we do not bother with | |
142 | handling it. */ | |
143 | return; | |
144 | ||
145 | case NE_EXPR: | |
146 | /* NE_EXPR comparisons do not contain much of useful information, | |
147 | except for cases of comparing with bounds. */ | |
148 | if (TREE_CODE (c1) != INTEGER_CST | |
149 | || !INTEGRAL_TYPE_P (type)) | |
150 | return; | |
151 | ||
152 | /* Ensure that the condition speaks about an expression in the same | |
153 | type as X and Y. */ | |
154 | ctype = TREE_TYPE (c0); | |
155 | if (TYPE_PRECISION (ctype) != TYPE_PRECISION (type)) | |
156 | return; | |
157 | c0 = fold_convert (type, c0); | |
158 | c1 = fold_convert (type, c1); | |
159 | ||
160 | if (operand_equal_p (var, c0, 0)) | |
161 | { | |
162 | mpz_t valc1; | |
163 | ||
164 | /* Case of comparing VAR with its below/up bounds. */ | |
165 | mpz_init (valc1); | |
166 | wi::to_mpz (c1, valc1, TYPE_SIGN (type)); | |
167 | if (mpz_cmp (valc1, below) == 0) | |
168 | cmp = GT_EXPR; | |
169 | if (mpz_cmp (valc1, up) == 0) | |
170 | cmp = LT_EXPR; | |
171 | ||
172 | mpz_clear (valc1); | |
173 | } | |
174 | else | |
175 | { | |
176 | /* Case of comparing with the bounds of the type. */ | |
177 | wide_int min = wi::min_value (type); | |
178 | wide_int max = wi::max_value (type); | |
179 | ||
180 | if (wi::eq_p (c1, min)) | |
181 | cmp = GT_EXPR; | |
182 | if (wi::eq_p (c1, max)) | |
183 | cmp = LT_EXPR; | |
184 | } | |
185 | ||
186 | /* Quick return if no useful information. */ | |
187 | if (cmp == NE_EXPR) | |
188 | return; | |
189 | ||
190 | break; | |
191 | ||
192 | default: | |
193 | return; | |
194 | } | |
195 | ||
196 | mpz_init (offc0); | |
197 | mpz_init (offc1); | |
198 | split_to_var_and_offset (expand_simple_operations (c0), &varc0, offc0); | |
199 | split_to_var_and_offset (expand_simple_operations (c1), &varc1, offc1); | |
200 | ||
201 | /* We are only interested in comparisons of expressions based on VAR. */ | |
202 | if (operand_equal_p (var, varc1, 0)) | |
203 | { | |
204 | std::swap (varc0, varc1); | |
205 | mpz_swap (offc0, offc1); | |
206 | cmp = swap_tree_comparison (cmp); | |
207 | } | |
208 | else if (!operand_equal_p (var, varc0, 0)) | |
209 | { | |
210 | mpz_clear (offc0); | |
211 | mpz_clear (offc1); | |
212 | return; | |
213 | } | |
214 | ||
215 | mpz_init (mint); | |
216 | mpz_init (maxt); | |
217 | get_type_static_bounds (type, mint, maxt); | |
218 | mpz_init (minc1); | |
219 | mpz_init (maxc1); | |
220 | /* Setup range information for varc1. */ | |
221 | if (integer_zerop (varc1)) | |
222 | { | |
223 | wi::to_mpz (integer_zero_node, minc1, TYPE_SIGN (type)); | |
224 | wi::to_mpz (integer_zero_node, maxc1, TYPE_SIGN (type)); | |
225 | } | |
226 | else if (TREE_CODE (varc1) == SSA_NAME | |
227 | && INTEGRAL_TYPE_P (type) | |
228 | && get_range_info (varc1, &minv, &maxv) == VR_RANGE) | |
229 | { | |
230 | gcc_assert (wi::le_p (minv, maxv, sgn)); | |
231 | wi::to_mpz (minv, minc1, sgn); | |
232 | wi::to_mpz (maxv, maxc1, sgn); | |
233 | } | |
234 | else | |
235 | { | |
236 | mpz_set (minc1, mint); | |
237 | mpz_set (maxc1, maxt); | |
238 | } | |
239 | ||
240 | /* Compute valid range information for varc1 + offc1. Note nothing | |
241 | useful can be derived if it overflows or underflows. Overflow or | |
242 | underflow could happen when: | |
243 | ||
244 | offc1 > 0 && varc1 + offc1 > MAX_VAL (type) | |
245 | offc1 < 0 && varc1 + offc1 < MIN_VAL (type). */ | |
246 | mpz_add (minc1, minc1, offc1); | |
247 | mpz_add (maxc1, maxc1, offc1); | |
248 | c1_ok = (no_wrap | |
249 | || mpz_sgn (offc1) == 0 | |
250 | || (mpz_sgn (offc1) < 0 && mpz_cmp (minc1, mint) >= 0) | |
251 | || (mpz_sgn (offc1) > 0 && mpz_cmp (maxc1, maxt) <= 0)); | |
252 | if (!c1_ok) | |
253 | goto end; | |
254 | ||
255 | if (mpz_cmp (minc1, mint) < 0) | |
256 | mpz_set (minc1, mint); | |
257 | if (mpz_cmp (maxc1, maxt) > 0) | |
258 | mpz_set (maxc1, maxt); | |
259 | ||
260 | if (cmp == LT_EXPR) | |
261 | { | |
262 | cmp = LE_EXPR; | |
263 | mpz_sub_ui (maxc1, maxc1, 1); | |
264 | } | |
265 | if (cmp == GT_EXPR) | |
266 | { | |
267 | cmp = GE_EXPR; | |
268 | mpz_add_ui (minc1, minc1, 1); | |
269 | } | |
270 | ||
271 | /* Compute range information for varc0. If there is no overflow, | |
272 | the condition implied that | |
273 | ||
274 | (varc0) cmp (varc1 + offc1 - offc0) | |
275 | ||
276 | We can possibly improve the upper bound of varc0 if cmp is LE_EXPR, | |
277 | or the below bound if cmp is GE_EXPR. | |
278 | ||
279 | To prove there is no overflow/underflow, we need to check below | |
280 | four cases: | |
281 | 1) cmp == LE_EXPR && offc0 > 0 | |
282 | ||
283 | (varc0 + offc0) doesn't overflow | |
284 | && (varc1 + offc1 - offc0) doesn't underflow | |
285 | ||
286 | 2) cmp == LE_EXPR && offc0 < 0 | |
287 | ||
288 | (varc0 + offc0) doesn't underflow | |
289 | && (varc1 + offc1 - offc0) doesn't overfloe | |
290 | ||
291 | In this case, (varc0 + offc0) will never underflow if we can | |
292 | prove (varc1 + offc1 - offc0) doesn't overflow. | |
293 | ||
294 | 3) cmp == GE_EXPR && offc0 < 0 | |
295 | ||
296 | (varc0 + offc0) doesn't underflow | |
297 | && (varc1 + offc1 - offc0) doesn't overflow | |
298 | ||
299 | 4) cmp == GE_EXPR && offc0 > 0 | |
300 | ||
301 | (varc0 + offc0) doesn't overflow | |
302 | && (varc1 + offc1 - offc0) doesn't underflow | |
303 | ||
304 | In this case, (varc0 + offc0) will never overflow if we can | |
305 | prove (varc1 + offc1 - offc0) doesn't underflow. | |
306 | ||
307 | Note we only handle case 2 and 4 in below code. */ | |
308 | ||
309 | mpz_sub (minc1, minc1, offc0); | |
310 | mpz_sub (maxc1, maxc1, offc0); | |
311 | c0_ok = (no_wrap | |
312 | || mpz_sgn (offc0) == 0 | |
313 | || (cmp == LE_EXPR | |
314 | && mpz_sgn (offc0) < 0 && mpz_cmp (maxc1, maxt) <= 0) | |
315 | || (cmp == GE_EXPR | |
316 | && mpz_sgn (offc0) > 0 && mpz_cmp (minc1, mint) >= 0)); | |
317 | if (!c0_ok) | |
318 | goto end; | |
319 | ||
320 | if (cmp == LE_EXPR) | |
321 | { | |
322 | if (mpz_cmp (up, maxc1) > 0) | |
323 | mpz_set (up, maxc1); | |
324 | } | |
325 | else | |
326 | { | |
327 | if (mpz_cmp (below, minc1) < 0) | |
328 | mpz_set (below, minc1); | |
329 | } | |
330 | ||
331 | end: | |
332 | mpz_clear (mint); | |
333 | mpz_clear (maxt); | |
334 | mpz_clear (minc1); | |
335 | mpz_clear (maxc1); | |
336 | mpz_clear (offc0); | |
337 | mpz_clear (offc1); | |
338 | } | |
339 | ||
b3ce5b6e ZD |
340 | /* Stores estimate on the minimum/maximum value of the expression VAR + OFF |
341 | in TYPE to MIN and MAX. */ | |
342 | ||
343 | static void | |
7190fdc1 | 344 | determine_value_range (struct loop *loop, tree type, tree var, mpz_t off, |
b3ce5b6e ZD |
345 | mpz_t min, mpz_t max) |
346 | { | |
7b008bbc BC |
347 | int cnt = 0; |
348 | mpz_t minm, maxm; | |
349 | basic_block bb; | |
807e902e | 350 | wide_int minv, maxv; |
7190fdc1 JJ |
351 | enum value_range_type rtype = VR_VARYING; |
352 | ||
b3ce5b6e ZD |
353 | /* If the expression is a constant, we know its value exactly. */ |
354 | if (integer_zerop (var)) | |
355 | { | |
356 | mpz_set (min, off); | |
357 | mpz_set (max, off); | |
358 | return; | |
359 | } | |
360 | ||
7190fdc1 JJ |
361 | get_type_static_bounds (type, min, max); |
362 | ||
363 | /* See if we have some range info from VRP. */ | |
364 | if (TREE_CODE (var) == SSA_NAME && INTEGRAL_TYPE_P (type)) | |
365 | { | |
366 | edge e = loop_preheader_edge (loop); | |
807e902e | 367 | signop sgn = TYPE_SIGN (type); |
538dd0b7 | 368 | gphi_iterator gsi; |
7190fdc1 JJ |
369 | |
370 | /* Either for VAR itself... */ | |
371 | rtype = get_range_info (var, &minv, &maxv); | |
372 | /* Or for PHI results in loop->header where VAR is used as | |
373 | PHI argument from the loop preheader edge. */ | |
374 | for (gsi = gsi_start_phis (loop->header); !gsi_end_p (gsi); gsi_next (&gsi)) | |
375 | { | |
538dd0b7 | 376 | gphi *phi = gsi.phi (); |
807e902e | 377 | wide_int minc, maxc; |
7190fdc1 JJ |
378 | if (PHI_ARG_DEF_FROM_EDGE (phi, e) == var |
379 | && (get_range_info (gimple_phi_result (phi), &minc, &maxc) | |
380 | == VR_RANGE)) | |
381 | { | |
382 | if (rtype != VR_RANGE) | |
383 | { | |
384 | rtype = VR_RANGE; | |
385 | minv = minc; | |
386 | maxv = maxc; | |
387 | } | |
388 | else | |
389 | { | |
807e902e KZ |
390 | minv = wi::max (minv, minc, sgn); |
391 | maxv = wi::min (maxv, maxc, sgn); | |
20adc5b1 JJ |
392 | /* If the PHI result range are inconsistent with |
393 | the VAR range, give up on looking at the PHI | |
394 | results. This can happen if VR_UNDEFINED is | |
395 | involved. */ | |
807e902e | 396 | if (wi::gt_p (minv, maxv, sgn)) |
20adc5b1 JJ |
397 | { |
398 | rtype = get_range_info (var, &minv, &maxv); | |
399 | break; | |
400 | } | |
7190fdc1 JJ |
401 | } |
402 | } | |
403 | } | |
7b008bbc BC |
404 | mpz_init (minm); |
405 | mpz_init (maxm); | |
406 | if (rtype != VR_RANGE) | |
407 | { | |
408 | mpz_set (minm, min); | |
409 | mpz_set (maxm, max); | |
410 | } | |
411 | else | |
7190fdc1 | 412 | { |
807e902e | 413 | gcc_assert (wi::le_p (minv, maxv, sgn)); |
807e902e KZ |
414 | wi::to_mpz (minv, minm, sgn); |
415 | wi::to_mpz (maxv, maxm, sgn); | |
7b008bbc BC |
416 | } |
417 | /* Now walk the dominators of the loop header and use the entry | |
418 | guards to refine the estimates. */ | |
419 | for (bb = loop->header; | |
420 | bb != ENTRY_BLOCK_PTR_FOR_FN (cfun) && cnt < MAX_DOMINATORS_TO_WALK; | |
421 | bb = get_immediate_dominator (CDI_DOMINATORS, bb)) | |
422 | { | |
423 | edge e; | |
424 | tree c0, c1; | |
355fe088 | 425 | gimple *cond; |
7b008bbc BC |
426 | enum tree_code cmp; |
427 | ||
428 | if (!single_pred_p (bb)) | |
429 | continue; | |
430 | e = single_pred_edge (bb); | |
431 | ||
432 | if (!(e->flags & (EDGE_TRUE_VALUE | EDGE_FALSE_VALUE))) | |
433 | continue; | |
434 | ||
435 | cond = last_stmt (e->src); | |
436 | c0 = gimple_cond_lhs (cond); | |
437 | cmp = gimple_cond_code (cond); | |
438 | c1 = gimple_cond_rhs (cond); | |
439 | ||
440 | if (e->flags & EDGE_FALSE_VALUE) | |
441 | cmp = invert_tree_comparison (cmp, false); | |
442 | ||
443 | refine_value_range_using_guard (type, var, c0, cmp, c1, minm, maxm); | |
444 | ++cnt; | |
445 | } | |
446 | ||
447 | mpz_add (minm, minm, off); | |
448 | mpz_add (maxm, maxm, off); | |
449 | /* If the computation may not wrap or off is zero, then this | |
450 | is always fine. If off is negative and minv + off isn't | |
451 | smaller than type's minimum, or off is positive and | |
452 | maxv + off isn't bigger than type's maximum, use the more | |
453 | precise range too. */ | |
454 | if (nowrap_type_p (type) | |
455 | || mpz_sgn (off) == 0 | |
456 | || (mpz_sgn (off) < 0 && mpz_cmp (minm, min) >= 0) | |
457 | || (mpz_sgn (off) > 0 && mpz_cmp (maxm, max) <= 0)) | |
458 | { | |
459 | mpz_set (min, minm); | |
460 | mpz_set (max, maxm); | |
7190fdc1 JJ |
461 | mpz_clear (minm); |
462 | mpz_clear (maxm); | |
7b008bbc | 463 | return; |
7190fdc1 | 464 | } |
7b008bbc BC |
465 | mpz_clear (minm); |
466 | mpz_clear (maxm); | |
7190fdc1 JJ |
467 | } |
468 | ||
b3ce5b6e ZD |
469 | /* If the computation may wrap, we know nothing about the value, except for |
470 | the range of the type. */ | |
b3ce5b6e ZD |
471 | if (!nowrap_type_p (type)) |
472 | return; | |
473 | ||
474 | /* Since the addition of OFF does not wrap, if OFF is positive, then we may | |
475 | add it to MIN, otherwise to MAX. */ | |
476 | if (mpz_sgn (off) < 0) | |
477 | mpz_add (max, max, off); | |
478 | else | |
479 | mpz_add (min, min, off); | |
480 | } | |
481 | ||
482 | /* Stores the bounds on the difference of the values of the expressions | |
483 | (var + X) and (var + Y), computed in TYPE, to BNDS. */ | |
484 | ||
485 | static void | |
486 | bound_difference_of_offsetted_base (tree type, mpz_t x, mpz_t y, | |
487 | bounds *bnds) | |
488 | { | |
489 | int rel = mpz_cmp (x, y); | |
490 | bool may_wrap = !nowrap_type_p (type); | |
491 | mpz_t m; | |
492 | ||
493 | /* If X == Y, then the expressions are always equal. | |
494 | If X > Y, there are the following possibilities: | |
495 | a) neither of var + X and var + Y overflow or underflow, or both of | |
496 | them do. Then their difference is X - Y. | |
497 | b) var + X overflows, and var + Y does not. Then the values of the | |
498 | expressions are var + X - M and var + Y, where M is the range of | |
499 | the type, and their difference is X - Y - M. | |
500 | c) var + Y underflows and var + X does not. Their difference again | |
501 | is M - X + Y. | |
502 | Therefore, if the arithmetics in type does not overflow, then the | |
503 | bounds are (X - Y, X - Y), otherwise they are (X - Y - M, X - Y) | |
504 | Similarly, if X < Y, the bounds are either (X - Y, X - Y) or | |
505 | (X - Y, X - Y + M). */ | |
506 | ||
507 | if (rel == 0) | |
508 | { | |
509 | mpz_set_ui (bnds->below, 0); | |
510 | mpz_set_ui (bnds->up, 0); | |
511 | return; | |
512 | } | |
513 | ||
514 | mpz_init (m); | |
807e902e | 515 | wi::to_mpz (wi::minus_one (TYPE_PRECISION (type)), m, UNSIGNED); |
b3ce5b6e ZD |
516 | mpz_add_ui (m, m, 1); |
517 | mpz_sub (bnds->up, x, y); | |
518 | mpz_set (bnds->below, bnds->up); | |
519 | ||
520 | if (may_wrap) | |
521 | { | |
522 | if (rel > 0) | |
523 | mpz_sub (bnds->below, bnds->below, m); | |
524 | else | |
525 | mpz_add (bnds->up, bnds->up, m); | |
526 | } | |
527 | ||
528 | mpz_clear (m); | |
529 | } | |
530 | ||
531 | /* From condition C0 CMP C1 derives information regarding the | |
532 | difference of values of VARX + OFFX and VARY + OFFY, computed in TYPE, | |
533 | and stores it to BNDS. */ | |
534 | ||
535 | static void | |
536 | refine_bounds_using_guard (tree type, tree varx, mpz_t offx, | |
537 | tree vary, mpz_t offy, | |
538 | tree c0, enum tree_code cmp, tree c1, | |
539 | bounds *bnds) | |
540 | { | |
6b4db501 | 541 | tree varc0, varc1, ctype; |
b3ce5b6e ZD |
542 | mpz_t offc0, offc1, loffx, loffy, bnd; |
543 | bool lbound = false; | |
544 | bool no_wrap = nowrap_type_p (type); | |
545 | bool x_ok, y_ok; | |
546 | ||
547 | switch (cmp) | |
548 | { | |
549 | case LT_EXPR: | |
550 | case LE_EXPR: | |
551 | case GT_EXPR: | |
552 | case GE_EXPR: | |
17b236ed ZD |
553 | STRIP_SIGN_NOPS (c0); |
554 | STRIP_SIGN_NOPS (c1); | |
555 | ctype = TREE_TYPE (c0); | |
36618b93 | 556 | if (!useless_type_conversion_p (ctype, type)) |
17b236ed ZD |
557 | return; |
558 | ||
b3ce5b6e ZD |
559 | break; |
560 | ||
561 | case EQ_EXPR: | |
562 | /* We could derive quite precise information from EQ_EXPR, however, such | |
17b236ed ZD |
563 | a guard is unlikely to appear, so we do not bother with handling |
564 | it. */ | |
b3ce5b6e ZD |
565 | return; |
566 | ||
567 | case NE_EXPR: | |
17b236ed ZD |
568 | /* NE_EXPR comparisons do not contain much of useful information, except for |
569 | special case of comparing with the bounds of the type. */ | |
570 | if (TREE_CODE (c1) != INTEGER_CST | |
571 | || !INTEGRAL_TYPE_P (type)) | |
572 | return; | |
573 | ||
574 | /* Ensure that the condition speaks about an expression in the same type | |
575 | as X and Y. */ | |
576 | ctype = TREE_TYPE (c0); | |
577 | if (TYPE_PRECISION (ctype) != TYPE_PRECISION (type)) | |
578 | return; | |
579 | c0 = fold_convert (type, c0); | |
580 | c1 = fold_convert (type, c1); | |
581 | ||
582 | if (TYPE_MIN_VALUE (type) | |
583 | && operand_equal_p (c1, TYPE_MIN_VALUE (type), 0)) | |
584 | { | |
585 | cmp = GT_EXPR; | |
586 | break; | |
587 | } | |
588 | if (TYPE_MAX_VALUE (type) | |
589 | && operand_equal_p (c1, TYPE_MAX_VALUE (type), 0)) | |
590 | { | |
591 | cmp = LT_EXPR; | |
592 | break; | |
593 | } | |
594 | ||
b3ce5b6e ZD |
595 | return; |
596 | default: | |
597 | return; | |
b8698a0f | 598 | } |
b3ce5b6e ZD |
599 | |
600 | mpz_init (offc0); | |
601 | mpz_init (offc1); | |
602 | split_to_var_and_offset (expand_simple_operations (c0), &varc0, offc0); | |
603 | split_to_var_and_offset (expand_simple_operations (c1), &varc1, offc1); | |
604 | ||
605 | /* We are only interested in comparisons of expressions based on VARX and | |
606 | VARY. TODO -- we might also be able to derive some bounds from | |
607 | expressions containing just one of the variables. */ | |
608 | ||
609 | if (operand_equal_p (varx, varc1, 0)) | |
610 | { | |
6b4db501 | 611 | std::swap (varc0, varc1); |
b3ce5b6e ZD |
612 | mpz_swap (offc0, offc1); |
613 | cmp = swap_tree_comparison (cmp); | |
614 | } | |
615 | ||
616 | if (!operand_equal_p (varx, varc0, 0) | |
617 | || !operand_equal_p (vary, varc1, 0)) | |
618 | goto end; | |
619 | ||
620 | mpz_init_set (loffx, offx); | |
621 | mpz_init_set (loffy, offy); | |
622 | ||
623 | if (cmp == GT_EXPR || cmp == GE_EXPR) | |
624 | { | |
6b4db501 | 625 | std::swap (varx, vary); |
b3ce5b6e ZD |
626 | mpz_swap (offc0, offc1); |
627 | mpz_swap (loffx, loffy); | |
628 | cmp = swap_tree_comparison (cmp); | |
629 | lbound = true; | |
630 | } | |
631 | ||
632 | /* If there is no overflow, the condition implies that | |
633 | ||
634 | (VARX + OFFX) cmp (VARY + OFFY) + (OFFX - OFFY + OFFC1 - OFFC0). | |
635 | ||
636 | The overflows and underflows may complicate things a bit; each | |
637 | overflow decreases the appropriate offset by M, and underflow | |
638 | increases it by M. The above inequality would not necessarily be | |
639 | true if | |
b8698a0f | 640 | |
b3ce5b6e ZD |
641 | -- VARX + OFFX underflows and VARX + OFFC0 does not, or |
642 | VARX + OFFC0 overflows, but VARX + OFFX does not. | |
643 | This may only happen if OFFX < OFFC0. | |
644 | -- VARY + OFFY overflows and VARY + OFFC1 does not, or | |
645 | VARY + OFFC1 underflows and VARY + OFFY does not. | |
646 | This may only happen if OFFY > OFFC1. */ | |
647 | ||
648 | if (no_wrap) | |
649 | { | |
650 | x_ok = true; | |
651 | y_ok = true; | |
652 | } | |
653 | else | |
654 | { | |
655 | x_ok = (integer_zerop (varx) | |
656 | || mpz_cmp (loffx, offc0) >= 0); | |
657 | y_ok = (integer_zerop (vary) | |
658 | || mpz_cmp (loffy, offc1) <= 0); | |
659 | } | |
660 | ||
661 | if (x_ok && y_ok) | |
662 | { | |
663 | mpz_init (bnd); | |
664 | mpz_sub (bnd, loffx, loffy); | |
665 | mpz_add (bnd, bnd, offc1); | |
666 | mpz_sub (bnd, bnd, offc0); | |
667 | ||
668 | if (cmp == LT_EXPR) | |
669 | mpz_sub_ui (bnd, bnd, 1); | |
670 | ||
671 | if (lbound) | |
672 | { | |
673 | mpz_neg (bnd, bnd); | |
674 | if (mpz_cmp (bnds->below, bnd) < 0) | |
675 | mpz_set (bnds->below, bnd); | |
676 | } | |
677 | else | |
678 | { | |
679 | if (mpz_cmp (bnd, bnds->up) < 0) | |
680 | mpz_set (bnds->up, bnd); | |
681 | } | |
682 | mpz_clear (bnd); | |
683 | } | |
684 | ||
685 | mpz_clear (loffx); | |
686 | mpz_clear (loffy); | |
687 | end: | |
688 | mpz_clear (offc0); | |
689 | mpz_clear (offc1); | |
690 | } | |
691 | ||
692 | /* Stores the bounds on the value of the expression X - Y in LOOP to BNDS. | |
693 | The subtraction is considered to be performed in arbitrary precision, | |
694 | without overflows. | |
b8698a0f | 695 | |
b3ce5b6e ZD |
696 | We do not attempt to be too clever regarding the value ranges of X and |
697 | Y; most of the time, they are just integers or ssa names offsetted by | |
698 | integer. However, we try to use the information contained in the | |
699 | comparisons before the loop (usually created by loop header copying). */ | |
700 | ||
701 | static void | |
702 | bound_difference (struct loop *loop, tree x, tree y, bounds *bnds) | |
703 | { | |
704 | tree type = TREE_TYPE (x); | |
705 | tree varx, vary; | |
706 | mpz_t offx, offy; | |
707 | mpz_t minx, maxx, miny, maxy; | |
708 | int cnt = 0; | |
709 | edge e; | |
710 | basic_block bb; | |
726a989a | 711 | tree c0, c1; |
355fe088 | 712 | gimple *cond; |
b3ce5b6e ZD |
713 | enum tree_code cmp; |
714 | ||
17b236ed ZD |
715 | /* Get rid of unnecessary casts, but preserve the value of |
716 | the expressions. */ | |
717 | STRIP_SIGN_NOPS (x); | |
718 | STRIP_SIGN_NOPS (y); | |
719 | ||
b3ce5b6e ZD |
720 | mpz_init (bnds->below); |
721 | mpz_init (bnds->up); | |
722 | mpz_init (offx); | |
723 | mpz_init (offy); | |
724 | split_to_var_and_offset (x, &varx, offx); | |
725 | split_to_var_and_offset (y, &vary, offy); | |
726 | ||
727 | if (!integer_zerop (varx) | |
728 | && operand_equal_p (varx, vary, 0)) | |
729 | { | |
730 | /* Special case VARX == VARY -- we just need to compare the | |
731 | offsets. The matters are a bit more complicated in the | |
732 | case addition of offsets may wrap. */ | |
733 | bound_difference_of_offsetted_base (type, offx, offy, bnds); | |
734 | } | |
735 | else | |
736 | { | |
737 | /* Otherwise, use the value ranges to determine the initial | |
738 | estimates on below and up. */ | |
739 | mpz_init (minx); | |
740 | mpz_init (maxx); | |
741 | mpz_init (miny); | |
742 | mpz_init (maxy); | |
7190fdc1 JJ |
743 | determine_value_range (loop, type, varx, offx, minx, maxx); |
744 | determine_value_range (loop, type, vary, offy, miny, maxy); | |
b3ce5b6e ZD |
745 | |
746 | mpz_sub (bnds->below, minx, maxy); | |
747 | mpz_sub (bnds->up, maxx, miny); | |
748 | mpz_clear (minx); | |
749 | mpz_clear (maxx); | |
750 | mpz_clear (miny); | |
751 | mpz_clear (maxy); | |
752 | } | |
753 | ||
754 | /* If both X and Y are constants, we cannot get any more precise. */ | |
755 | if (integer_zerop (varx) && integer_zerop (vary)) | |
756 | goto end; | |
757 | ||
758 | /* Now walk the dominators of the loop header and use the entry | |
759 | guards to refine the estimates. */ | |
760 | for (bb = loop->header; | |
fefa31b5 | 761 | bb != ENTRY_BLOCK_PTR_FOR_FN (cfun) && cnt < MAX_DOMINATORS_TO_WALK; |
b3ce5b6e ZD |
762 | bb = get_immediate_dominator (CDI_DOMINATORS, bb)) |
763 | { | |
764 | if (!single_pred_p (bb)) | |
765 | continue; | |
766 | e = single_pred_edge (bb); | |
767 | ||
768 | if (!(e->flags & (EDGE_TRUE_VALUE | EDGE_FALSE_VALUE))) | |
769 | continue; | |
770 | ||
726a989a RB |
771 | cond = last_stmt (e->src); |
772 | c0 = gimple_cond_lhs (cond); | |
773 | cmp = gimple_cond_code (cond); | |
774 | c1 = gimple_cond_rhs (cond); | |
b3ce5b6e ZD |
775 | |
776 | if (e->flags & EDGE_FALSE_VALUE) | |
777 | cmp = invert_tree_comparison (cmp, false); | |
778 | ||
779 | refine_bounds_using_guard (type, varx, offx, vary, offy, | |
780 | c0, cmp, c1, bnds); | |
781 | ++cnt; | |
782 | } | |
783 | ||
784 | end: | |
785 | mpz_clear (offx); | |
786 | mpz_clear (offy); | |
787 | } | |
788 | ||
789 | /* Update the bounds in BNDS that restrict the value of X to the bounds | |
790 | that restrict the value of X + DELTA. X can be obtained as a | |
791 | difference of two values in TYPE. */ | |
792 | ||
793 | static void | |
807e902e | 794 | bounds_add (bounds *bnds, const widest_int &delta, tree type) |
b3ce5b6e ZD |
795 | { |
796 | mpz_t mdelta, max; | |
797 | ||
798 | mpz_init (mdelta); | |
807e902e | 799 | wi::to_mpz (delta, mdelta, SIGNED); |
b3ce5b6e ZD |
800 | |
801 | mpz_init (max); | |
807e902e | 802 | wi::to_mpz (wi::minus_one (TYPE_PRECISION (type)), max, UNSIGNED); |
b3ce5b6e ZD |
803 | |
804 | mpz_add (bnds->up, bnds->up, mdelta); | |
805 | mpz_add (bnds->below, bnds->below, mdelta); | |
806 | ||
807 | if (mpz_cmp (bnds->up, max) > 0) | |
808 | mpz_set (bnds->up, max); | |
809 | ||
810 | mpz_neg (max, max); | |
811 | if (mpz_cmp (bnds->below, max) < 0) | |
812 | mpz_set (bnds->below, max); | |
813 | ||
814 | mpz_clear (mdelta); | |
815 | mpz_clear (max); | |
816 | } | |
817 | ||
818 | /* Update the bounds in BNDS that restrict the value of X to the bounds | |
819 | that restrict the value of -X. */ | |
820 | ||
821 | static void | |
822 | bounds_negate (bounds *bnds) | |
823 | { | |
824 | mpz_t tmp; | |
825 | ||
826 | mpz_init_set (tmp, bnds->up); | |
827 | mpz_neg (bnds->up, bnds->below); | |
828 | mpz_neg (bnds->below, tmp); | |
829 | mpz_clear (tmp); | |
830 | } | |
831 | ||
e9eb809d ZD |
832 | /* Returns inverse of X modulo 2^s, where MASK = 2^s-1. */ |
833 | ||
834 | static tree | |
835 | inverse (tree x, tree mask) | |
836 | { | |
837 | tree type = TREE_TYPE (x); | |
26630a99 ZD |
838 | tree rslt; |
839 | unsigned ctr = tree_floor_log2 (mask); | |
840 | ||
841 | if (TYPE_PRECISION (type) <= HOST_BITS_PER_WIDE_INT) | |
842 | { | |
843 | unsigned HOST_WIDE_INT ix; | |
844 | unsigned HOST_WIDE_INT imask; | |
845 | unsigned HOST_WIDE_INT irslt = 1; | |
846 | ||
847 | gcc_assert (cst_and_fits_in_hwi (x)); | |
848 | gcc_assert (cst_and_fits_in_hwi (mask)); | |
849 | ||
850 | ix = int_cst_value (x); | |
851 | imask = int_cst_value (mask); | |
852 | ||
853 | for (; ctr; ctr--) | |
854 | { | |
855 | irslt *= ix; | |
856 | ix *= ix; | |
857 | } | |
858 | irslt &= imask; | |
e9eb809d | 859 | |
26630a99 ZD |
860 | rslt = build_int_cst_type (type, irslt); |
861 | } | |
862 | else | |
e9eb809d | 863 | { |
ff5e9a94 | 864 | rslt = build_int_cst (type, 1); |
26630a99 ZD |
865 | for (; ctr; ctr--) |
866 | { | |
d35936ab RG |
867 | rslt = int_const_binop (MULT_EXPR, rslt, x); |
868 | x = int_const_binop (MULT_EXPR, x, x); | |
26630a99 | 869 | } |
d35936ab | 870 | rslt = int_const_binop (BIT_AND_EXPR, rslt, mask); |
e9eb809d ZD |
871 | } |
872 | ||
873 | return rslt; | |
874 | } | |
875 | ||
b3ce5b6e | 876 | /* Derives the upper bound BND on the number of executions of loop with exit |
1987baa3 ZD |
877 | condition S * i <> C. If NO_OVERFLOW is true, then the control variable of |
878 | the loop does not overflow. EXIT_MUST_BE_TAKEN is true if we are guaranteed | |
879 | that the loop ends through this exit, i.e., the induction variable ever | |
880 | reaches the value of C. | |
881 | ||
882 | The value C is equal to final - base, where final and base are the final and | |
883 | initial value of the actual induction variable in the analysed loop. BNDS | |
884 | bounds the value of this difference when computed in signed type with | |
885 | unbounded range, while the computation of C is performed in an unsigned | |
886 | type with the range matching the range of the type of the induction variable. | |
887 | In particular, BNDS.up contains an upper bound on C in the following cases: | |
888 | -- if the iv must reach its final value without overflow, i.e., if | |
889 | NO_OVERFLOW && EXIT_MUST_BE_TAKEN is true, or | |
890 | -- if final >= base, which we know to hold when BNDS.below >= 0. */ | |
b3ce5b6e ZD |
891 | |
892 | static void | |
893 | number_of_iterations_ne_max (mpz_t bnd, bool no_overflow, tree c, tree s, | |
1987baa3 | 894 | bounds *bnds, bool exit_must_be_taken) |
b3ce5b6e | 895 | { |
807e902e | 896 | widest_int max; |
b3ce5b6e | 897 | mpz_t d; |
5a892248 | 898 | tree type = TREE_TYPE (c); |
1987baa3 ZD |
899 | bool bnds_u_valid = ((no_overflow && exit_must_be_taken) |
900 | || mpz_sgn (bnds->below) >= 0); | |
b3ce5b6e | 901 | |
5a892248 RB |
902 | if (integer_onep (s) |
903 | || (TREE_CODE (c) == INTEGER_CST | |
904 | && TREE_CODE (s) == INTEGER_CST | |
807e902e KZ |
905 | && wi::mod_trunc (c, s, TYPE_SIGN (type)) == 0) |
906 | || (TYPE_OVERFLOW_UNDEFINED (type) | |
5a892248 | 907 | && multiple_of_p (type, c, s))) |
1987baa3 ZD |
908 | { |
909 | /* If C is an exact multiple of S, then its value will be reached before | |
910 | the induction variable overflows (unless the loop is exited in some | |
911 | other way before). Note that the actual induction variable in the | |
912 | loop (which ranges from base to final instead of from 0 to C) may | |
913 | overflow, in which case BNDS.up will not be giving a correct upper | |
914 | bound on C; thus, BNDS_U_VALID had to be computed in advance. */ | |
915 | no_overflow = true; | |
916 | exit_must_be_taken = true; | |
917 | } | |
918 | ||
919 | /* If the induction variable can overflow, the number of iterations is at | |
920 | most the period of the control variable (or infinite, but in that case | |
921 | the whole # of iterations analysis will fail). */ | |
922 | if (!no_overflow) | |
b3ce5b6e | 923 | { |
807e902e KZ |
924 | max = wi::mask <widest_int> (TYPE_PRECISION (type) - wi::ctz (s), false); |
925 | wi::to_mpz (max, bnd, UNSIGNED); | |
b3ce5b6e ZD |
926 | return; |
927 | } | |
928 | ||
1987baa3 ZD |
929 | /* Now we know that the induction variable does not overflow, so the loop |
930 | iterates at most (range of type / S) times. */ | |
807e902e | 931 | wi::to_mpz (wi::minus_one (TYPE_PRECISION (type)), bnd, UNSIGNED); |
1987baa3 ZD |
932 | |
933 | /* If the induction variable is guaranteed to reach the value of C before | |
934 | overflow, ... */ | |
935 | if (exit_must_be_taken) | |
936 | { | |
073a8998 | 937 | /* ... then we can strengthen this to C / S, and possibly we can use |
1987baa3 ZD |
938 | the upper bound on C given by BNDS. */ |
939 | if (TREE_CODE (c) == INTEGER_CST) | |
807e902e | 940 | wi::to_mpz (c, bnd, UNSIGNED); |
1987baa3 ZD |
941 | else if (bnds_u_valid) |
942 | mpz_set (bnd, bnds->up); | |
943 | } | |
b3ce5b6e ZD |
944 | |
945 | mpz_init (d); | |
807e902e | 946 | wi::to_mpz (s, d, UNSIGNED); |
b3ce5b6e ZD |
947 | mpz_fdiv_q (bnd, bnd, d); |
948 | mpz_clear (d); | |
949 | } | |
950 | ||
7f17528a ZD |
951 | /* Determines number of iterations of loop whose ending condition |
952 | is IV <> FINAL. TYPE is the type of the iv. The number of | |
e36dc339 | 953 | iterations is stored to NITER. EXIT_MUST_BE_TAKEN is true if |
f08ac361 ZD |
954 | we know that the exit must be taken eventually, i.e., that the IV |
955 | ever reaches the value FINAL (we derived this earlier, and possibly set | |
b3ce5b6e ZD |
956 | NITER->assumptions to make sure this is the case). BNDS contains the |
957 | bounds on the difference FINAL - IV->base. */ | |
e9eb809d | 958 | |
7f17528a | 959 | static bool |
cdf66caf BC |
960 | number_of_iterations_ne (struct loop *loop, tree type, affine_iv *iv, |
961 | tree final, struct tree_niter_desc *niter, | |
962 | bool exit_must_be_taken, bounds *bnds) | |
e9eb809d | 963 | { |
7f17528a ZD |
964 | tree niter_type = unsigned_type_for (type); |
965 | tree s, c, d, bits, assumption, tmp, bound; | |
b3ce5b6e | 966 | mpz_t max; |
e9eb809d | 967 | |
17684618 ZD |
968 | niter->control = *iv; |
969 | niter->bound = final; | |
970 | niter->cmp = NE_EXPR; | |
971 | ||
b3ce5b6e ZD |
972 | /* Rearrange the terms so that we get inequality S * i <> C, with S |
973 | positive. Also cast everything to the unsigned type. If IV does | |
974 | not overflow, BNDS bounds the value of C. Also, this is the | |
975 | case if the computation |FINAL - IV->base| does not overflow, i.e., | |
976 | if BNDS->below in the result is nonnegative. */ | |
7f17528a | 977 | if (tree_int_cst_sign_bit (iv->step)) |
e9eb809d | 978 | { |
7f17528a ZD |
979 | s = fold_convert (niter_type, |
980 | fold_build1 (NEGATE_EXPR, type, iv->step)); | |
981 | c = fold_build2 (MINUS_EXPR, niter_type, | |
982 | fold_convert (niter_type, iv->base), | |
983 | fold_convert (niter_type, final)); | |
b3ce5b6e | 984 | bounds_negate (bnds); |
e9eb809d | 985 | } |
a6f778b2 | 986 | else |
e9eb809d | 987 | { |
7f17528a ZD |
988 | s = fold_convert (niter_type, iv->step); |
989 | c = fold_build2 (MINUS_EXPR, niter_type, | |
990 | fold_convert (niter_type, final), | |
991 | fold_convert (niter_type, iv->base)); | |
992 | } | |
e9eb809d | 993 | |
b3ce5b6e | 994 | mpz_init (max); |
1987baa3 ZD |
995 | number_of_iterations_ne_max (max, iv->no_overflow, c, s, bnds, |
996 | exit_must_be_taken); | |
807e902e KZ |
997 | niter->max = widest_int::from (wi::from_mpz (niter_type, max, false), |
998 | TYPE_SIGN (niter_type)); | |
b3ce5b6e ZD |
999 | mpz_clear (max); |
1000 | ||
69b806f6 | 1001 | /* Compute no-overflow information for the control iv. This can be |
8f21990a | 1002 | proven when below two conditions are satisfied: |
69b806f6 | 1003 | |
8f21990a | 1004 | 1) IV evaluates toward FINAL at beginning, i.e: |
69b806f6 BC |
1005 | base <= FINAL ; step > 0 |
1006 | base >= FINAL ; step < 0 | |
1007 | ||
8f21990a BC |
1008 | 2) |FINAL - base| is an exact multiple of step. |
1009 | ||
1010 | Unfortunately, it's hard to prove above conditions after pass loop-ch | |
1011 | because loop with exit condition (IV != FINAL) usually will be guarded | |
1012 | by initial-condition (IV.base - IV.step != FINAL). In this case, we | |
1013 | can alternatively try to prove below conditions: | |
1014 | ||
1015 | 1') IV evaluates toward FINAL at beginning, i.e: | |
1016 | new_base = base - step < FINAL ; step > 0 | |
1017 | && base - step doesn't underflow | |
1018 | new_base = base - step > FINAL ; step < 0 | |
1019 | && base - step doesn't overflow | |
69b806f6 | 1020 | |
8f21990a | 1021 | 2') |FINAL - new_base| is an exact multiple of step. |
69b806f6 | 1022 | |
8f21990a BC |
1023 | Please refer to PR34114 as an example of loop-ch's impact, also refer |
1024 | to PR72817 as an example why condition 2') is necessary. | |
69b806f6 | 1025 | |
8f21990a | 1026 | Note, for NE_EXPR, base equals to FINAL is a special case, in |
69b806f6 BC |
1027 | which the loop exits immediately, and the iv does not overflow. */ |
1028 | if (!niter->control.no_overflow | |
1029 | && (integer_onep (s) || multiple_of_p (type, c, s))) | |
cdf66caf | 1030 | { |
8f21990a | 1031 | tree t, cond, new_c, relaxed_cond = boolean_false_node; |
69b806f6 BC |
1032 | |
1033 | if (tree_int_cst_sign_bit (iv->step)) | |
1034 | { | |
1035 | cond = fold_build2 (GE_EXPR, boolean_type_node, iv->base, final); | |
1036 | if (TREE_CODE (type) == INTEGER_TYPE) | |
1037 | { | |
1038 | /* Only when base - step doesn't overflow. */ | |
1039 | t = TYPE_MAX_VALUE (type); | |
1040 | t = fold_build2 (PLUS_EXPR, type, t, iv->step); | |
1041 | t = fold_build2 (GE_EXPR, boolean_type_node, t, iv->base); | |
1042 | if (integer_nonzerop (t)) | |
1043 | { | |
1044 | t = fold_build2 (MINUS_EXPR, type, iv->base, iv->step); | |
8f21990a BC |
1045 | new_c = fold_build2 (MINUS_EXPR, niter_type, |
1046 | fold_convert (niter_type, t), | |
1047 | fold_convert (niter_type, final)); | |
1048 | if (multiple_of_p (type, new_c, s)) | |
1049 | relaxed_cond = fold_build2 (GT_EXPR, boolean_type_node, | |
1050 | t, final); | |
69b806f6 BC |
1051 | } |
1052 | } | |
1053 | } | |
1054 | else | |
1055 | { | |
1056 | cond = fold_build2 (LE_EXPR, boolean_type_node, iv->base, final); | |
1057 | if (TREE_CODE (type) == INTEGER_TYPE) | |
1058 | { | |
1059 | /* Only when base - step doesn't underflow. */ | |
1060 | t = TYPE_MIN_VALUE (type); | |
1061 | t = fold_build2 (PLUS_EXPR, type, t, iv->step); | |
1062 | t = fold_build2 (LE_EXPR, boolean_type_node, t, iv->base); | |
1063 | if (integer_nonzerop (t)) | |
1064 | { | |
1065 | t = fold_build2 (MINUS_EXPR, type, iv->base, iv->step); | |
8f21990a BC |
1066 | new_c = fold_build2 (MINUS_EXPR, niter_type, |
1067 | fold_convert (niter_type, final), | |
1068 | fold_convert (niter_type, t)); | |
1069 | if (multiple_of_p (type, new_c, s)) | |
1070 | relaxed_cond = fold_build2 (LT_EXPR, boolean_type_node, | |
1071 | t, final); | |
69b806f6 BC |
1072 | } |
1073 | } | |
1074 | } | |
1075 | ||
1076 | t = simplify_using_initial_conditions (loop, cond); | |
1077 | if (!t || !integer_onep (t)) | |
1078 | t = simplify_using_initial_conditions (loop, relaxed_cond); | |
1079 | ||
1080 | if (t && integer_onep (t)) | |
1081 | niter->control.no_overflow = true; | |
cdf66caf BC |
1082 | } |
1083 | ||
7f17528a ZD |
1084 | /* First the trivial cases -- when the step is 1. */ |
1085 | if (integer_onep (s)) | |
1086 | { | |
1087 | niter->niter = c; | |
1088 | return true; | |
e9eb809d | 1089 | } |
69b806f6 BC |
1090 | if (niter->control.no_overflow && multiple_of_p (type, c, s)) |
1091 | { | |
1092 | niter->niter = fold_build2 (FLOOR_DIV_EXPR, niter_type, c, s); | |
1093 | return true; | |
1094 | } | |
e9eb809d | 1095 | |
7f17528a ZD |
1096 | /* Let nsd (step, size of mode) = d. If d does not divide c, the loop |
1097 | is infinite. Otherwise, the number of iterations is | |
1098 | (inverse(s/d) * (c/d)) mod (size of mode/d). */ | |
1099 | bits = num_ending_zeros (s); | |
1100 | bound = build_low_bits_mask (niter_type, | |
1101 | (TYPE_PRECISION (niter_type) | |
ae7e9ddd | 1102 | - tree_to_uhwi (bits))); |
e9eb809d | 1103 | |
7f17528a | 1104 | d = fold_binary_to_constant (LSHIFT_EXPR, niter_type, |
ff5e9a94 | 1105 | build_int_cst (niter_type, 1), bits); |
7f17528a | 1106 | s = fold_binary_to_constant (RSHIFT_EXPR, niter_type, s, bits); |
e9eb809d | 1107 | |
e36dc339 | 1108 | if (!exit_must_be_taken) |
7f17528a | 1109 | { |
e36dc339 | 1110 | /* If we cannot assume that the exit is taken eventually, record the |
7f17528a ZD |
1111 | assumptions for divisibility of c. */ |
1112 | assumption = fold_build2 (FLOOR_MOD_EXPR, niter_type, c, d); | |
1113 | assumption = fold_build2 (EQ_EXPR, boolean_type_node, | |
1114 | assumption, build_int_cst (niter_type, 0)); | |
6e682d7e | 1115 | if (!integer_nonzerop (assumption)) |
7f17528a ZD |
1116 | niter->assumptions = fold_build2 (TRUTH_AND_EXPR, boolean_type_node, |
1117 | niter->assumptions, assumption); | |
e9eb809d | 1118 | } |
b8698a0f | 1119 | |
7f17528a ZD |
1120 | c = fold_build2 (EXACT_DIV_EXPR, niter_type, c, d); |
1121 | tmp = fold_build2 (MULT_EXPR, niter_type, c, inverse (s, bound)); | |
1122 | niter->niter = fold_build2 (BIT_AND_EXPR, niter_type, tmp, bound); | |
1123 | return true; | |
1124 | } | |
e9eb809d | 1125 | |
7f17528a ZD |
1126 | /* Checks whether we can determine the final value of the control variable |
1127 | of the loop with ending condition IV0 < IV1 (computed in TYPE). | |
1128 | DELTA is the difference IV1->base - IV0->base, STEP is the absolute value | |
1129 | of the step. The assumptions necessary to ensure that the computation | |
1130 | of the final value does not overflow are recorded in NITER. If we | |
1131 | find the final value, we adjust DELTA and return TRUE. Otherwise | |
b3ce5b6e | 1132 | we return false. BNDS bounds the value of IV1->base - IV0->base, |
e36dc339 ZD |
1133 | and will be updated by the same amount as DELTA. EXIT_MUST_BE_TAKEN is |
1134 | true if we know that the exit must be taken eventually. */ | |
7f17528a ZD |
1135 | |
1136 | static bool | |
1137 | number_of_iterations_lt_to_ne (tree type, affine_iv *iv0, affine_iv *iv1, | |
1138 | struct tree_niter_desc *niter, | |
b3ce5b6e | 1139 | tree *delta, tree step, |
e36dc339 | 1140 | bool exit_must_be_taken, bounds *bnds) |
7f17528a ZD |
1141 | { |
1142 | tree niter_type = TREE_TYPE (step); | |
1143 | tree mod = fold_build2 (FLOOR_MOD_EXPR, niter_type, *delta, step); | |
1144 | tree tmod; | |
106d07f8 BC |
1145 | tree assumption = boolean_true_node, bound; |
1146 | tree type1 = (POINTER_TYPE_P (type)) ? sizetype : type; | |
7f17528a ZD |
1147 | |
1148 | if (TREE_CODE (mod) != INTEGER_CST) | |
1149 | return false; | |
6e682d7e | 1150 | if (integer_nonzerop (mod)) |
7f17528a | 1151 | mod = fold_build2 (MINUS_EXPR, niter_type, step, mod); |
5be014d5 | 1152 | tmod = fold_convert (type1, mod); |
7f17528a | 1153 | |
e36dc339 | 1154 | /* If the induction variable does not overflow and the exit is taken, |
106d07f8 BC |
1155 | then the computation of the final value does not overflow. There |
1156 | are three cases: | |
1157 | 1) The case if the new final value is equal to the current one. | |
1158 | 2) Induction varaible has pointer type, as the code cannot rely | |
1159 | on the object to that the pointer points being placed at the | |
1160 | end of the address space (and more pragmatically, | |
1161 | TYPE_{MIN,MAX}_VALUE is not defined for pointers). | |
1162 | 3) EXIT_MUST_BE_TAKEN is true, note it implies that the induction | |
1163 | variable does not overflow. */ | |
1164 | if (!integer_zerop (mod) && !POINTER_TYPE_P (type) && !exit_must_be_taken) | |
e9eb809d | 1165 | { |
106d07f8 | 1166 | if (integer_nonzerop (iv0->step)) |
7f17528a | 1167 | { |
106d07f8 BC |
1168 | /* The final value of the iv is iv1->base + MOD, assuming |
1169 | that this computation does not overflow, and that | |
1170 | iv0->base <= iv1->base + MOD. */ | |
97b4ba9f | 1171 | bound = fold_build2 (MINUS_EXPR, type1, |
5be014d5 | 1172 | TYPE_MAX_VALUE (type1), tmod); |
7f17528a ZD |
1173 | assumption = fold_build2 (LE_EXPR, boolean_type_node, |
1174 | iv1->base, bound); | |
7f17528a | 1175 | } |
b3ce5b6e | 1176 | else |
7f17528a | 1177 | { |
106d07f8 BC |
1178 | /* The final value of the iv is iv0->base - MOD, assuming |
1179 | that this computation does not overflow, and that | |
1180 | iv0->base - MOD <= iv1->base. */ | |
5be014d5 AP |
1181 | bound = fold_build2 (PLUS_EXPR, type1, |
1182 | TYPE_MIN_VALUE (type1), tmod); | |
7f17528a ZD |
1183 | assumption = fold_build2 (GE_EXPR, boolean_type_node, |
1184 | iv0->base, bound); | |
7f17528a | 1185 | } |
106d07f8 BC |
1186 | if (integer_zerop (assumption)) |
1187 | return false; | |
1188 | else if (!integer_nonzerop (assumption)) | |
1189 | niter->assumptions = fold_build2 (TRUTH_AND_EXPR, boolean_type_node, | |
1190 | niter->assumptions, assumption); | |
e9eb809d ZD |
1191 | } |
1192 | ||
106d07f8 BC |
1193 | /* Since we are transforming LT to NE and DELTA is constant, there |
1194 | is no need to compute may_be_zero because this loop must roll. */ | |
1195 | ||
807e902e | 1196 | bounds_add (bnds, wi::to_widest (mod), type); |
7f17528a | 1197 | *delta = fold_build2 (PLUS_EXPR, niter_type, *delta, mod); |
106d07f8 | 1198 | return true; |
7f17528a ZD |
1199 | } |
1200 | ||
1201 | /* Add assertions to NITER that ensure that the control variable of the loop | |
1202 | with ending condition IV0 < IV1 does not overflow. Types of IV0 and IV1 | |
1203 | are TYPE. Returns false if we can prove that there is an overflow, true | |
1204 | otherwise. STEP is the absolute value of the step. */ | |
e9eb809d | 1205 | |
7f17528a ZD |
1206 | static bool |
1207 | assert_no_overflow_lt (tree type, affine_iv *iv0, affine_iv *iv1, | |
1208 | struct tree_niter_desc *niter, tree step) | |
1209 | { | |
1210 | tree bound, d, assumption, diff; | |
1211 | tree niter_type = TREE_TYPE (step); | |
1212 | ||
6e42ce54 | 1213 | if (integer_nonzerop (iv0->step)) |
e9eb809d | 1214 | { |
7f17528a ZD |
1215 | /* for (i = iv0->base; i < iv1->base; i += iv0->step) */ |
1216 | if (iv0->no_overflow) | |
1217 | return true; | |
1218 | ||
1219 | /* If iv0->base is a constant, we can determine the last value before | |
1220 | overflow precisely; otherwise we conservatively assume | |
1221 | MAX - STEP + 1. */ | |
1222 | ||
1223 | if (TREE_CODE (iv0->base) == INTEGER_CST) | |
e9eb809d | 1224 | { |
7f17528a ZD |
1225 | d = fold_build2 (MINUS_EXPR, niter_type, |
1226 | fold_convert (niter_type, TYPE_MAX_VALUE (type)), | |
1227 | fold_convert (niter_type, iv0->base)); | |
1228 | diff = fold_build2 (FLOOR_MOD_EXPR, niter_type, d, step); | |
e9eb809d ZD |
1229 | } |
1230 | else | |
7f17528a | 1231 | diff = fold_build2 (MINUS_EXPR, niter_type, step, |
ff5e9a94 | 1232 | build_int_cst (niter_type, 1)); |
7f17528a ZD |
1233 | bound = fold_build2 (MINUS_EXPR, type, |
1234 | TYPE_MAX_VALUE (type), fold_convert (type, diff)); | |
1235 | assumption = fold_build2 (LE_EXPR, boolean_type_node, | |
1236 | iv1->base, bound); | |
1237 | } | |
1238 | else | |
1239 | { | |
1240 | /* for (i = iv1->base; i > iv0->base; i += iv1->step) */ | |
1241 | if (iv1->no_overflow) | |
1242 | return true; | |
1243 | ||
1244 | if (TREE_CODE (iv1->base) == INTEGER_CST) | |
e9eb809d | 1245 | { |
7f17528a ZD |
1246 | d = fold_build2 (MINUS_EXPR, niter_type, |
1247 | fold_convert (niter_type, iv1->base), | |
1248 | fold_convert (niter_type, TYPE_MIN_VALUE (type))); | |
1249 | diff = fold_build2 (FLOOR_MOD_EXPR, niter_type, d, step); | |
e9eb809d | 1250 | } |
7f17528a ZD |
1251 | else |
1252 | diff = fold_build2 (MINUS_EXPR, niter_type, step, | |
ff5e9a94 | 1253 | build_int_cst (niter_type, 1)); |
7f17528a ZD |
1254 | bound = fold_build2 (PLUS_EXPR, type, |
1255 | TYPE_MIN_VALUE (type), fold_convert (type, diff)); | |
1256 | assumption = fold_build2 (GE_EXPR, boolean_type_node, | |
1257 | iv0->base, bound); | |
e9eb809d ZD |
1258 | } |
1259 | ||
6e682d7e | 1260 | if (integer_zerop (assumption)) |
7f17528a | 1261 | return false; |
6e682d7e | 1262 | if (!integer_nonzerop (assumption)) |
7f17528a ZD |
1263 | niter->assumptions = fold_build2 (TRUTH_AND_EXPR, boolean_type_node, |
1264 | niter->assumptions, assumption); | |
b8698a0f | 1265 | |
7f17528a ZD |
1266 | iv0->no_overflow = true; |
1267 | iv1->no_overflow = true; | |
1268 | return true; | |
1269 | } | |
e9eb809d | 1270 | |
7f17528a | 1271 | /* Add an assumption to NITER that a loop whose ending condition |
b3ce5b6e ZD |
1272 | is IV0 < IV1 rolls. TYPE is the type of the control iv. BNDS |
1273 | bounds the value of IV1->base - IV0->base. */ | |
7f17528a ZD |
1274 | |
1275 | static void | |
1276 | assert_loop_rolls_lt (tree type, affine_iv *iv0, affine_iv *iv1, | |
b3ce5b6e | 1277 | struct tree_niter_desc *niter, bounds *bnds) |
7f17528a ZD |
1278 | { |
1279 | tree assumption = boolean_true_node, bound, diff; | |
5be014d5 | 1280 | tree mbz, mbzl, mbzr, type1; |
b3ce5b6e | 1281 | bool rolls_p, no_overflow_p; |
807e902e | 1282 | widest_int dstep; |
b3ce5b6e ZD |
1283 | mpz_t mstep, max; |
1284 | ||
1285 | /* We are going to compute the number of iterations as | |
1286 | (iv1->base - iv0->base + step - 1) / step, computed in the unsigned | |
b8698a0f L |
1287 | variant of TYPE. This formula only works if |
1288 | ||
b3ce5b6e | 1289 | -step + 1 <= (iv1->base - iv0->base) <= MAX - step + 1 |
b8698a0f | 1290 | |
b3ce5b6e | 1291 | (where MAX is the maximum value of the unsigned variant of TYPE, and |
072edf07 SP |
1292 | the computations in this formula are performed in full precision, |
1293 | i.e., without overflows). | |
b3ce5b6e ZD |
1294 | |
1295 | Usually, for loops with exit condition iv0->base + step * i < iv1->base, | |
072edf07 | 1296 | we have a condition of the form iv0->base - step < iv1->base before the loop, |
b3ce5b6e ZD |
1297 | and for loops iv0->base < iv1->base - step * i the condition |
1298 | iv0->base < iv1->base + step, due to loop header copying, which enable us | |
1299 | to prove the lower bound. | |
b8698a0f | 1300 | |
b3ce5b6e ZD |
1301 | The upper bound is more complicated. Unless the expressions for initial |
1302 | and final value themselves contain enough information, we usually cannot | |
1303 | derive it from the context. */ | |
1304 | ||
1305 | /* First check whether the answer does not follow from the bounds we gathered | |
1306 | before. */ | |
1307 | if (integer_nonzerop (iv0->step)) | |
807e902e | 1308 | dstep = wi::to_widest (iv0->step); |
b3ce5b6e ZD |
1309 | else |
1310 | { | |
807e902e | 1311 | dstep = wi::sext (wi::to_widest (iv1->step), TYPE_PRECISION (type)); |
27bcd47c | 1312 | dstep = -dstep; |
b3ce5b6e ZD |
1313 | } |
1314 | ||
1315 | mpz_init (mstep); | |
807e902e | 1316 | wi::to_mpz (dstep, mstep, UNSIGNED); |
b3ce5b6e ZD |
1317 | mpz_neg (mstep, mstep); |
1318 | mpz_add_ui (mstep, mstep, 1); | |
1319 | ||
1320 | rolls_p = mpz_cmp (mstep, bnds->below) <= 0; | |
1321 | ||
1322 | mpz_init (max); | |
807e902e | 1323 | wi::to_mpz (wi::minus_one (TYPE_PRECISION (type)), max, UNSIGNED); |
b3ce5b6e ZD |
1324 | mpz_add (max, max, mstep); |
1325 | no_overflow_p = (mpz_cmp (bnds->up, max) <= 0 | |
1326 | /* For pointers, only values lying inside a single object | |
1327 | can be compared or manipulated by pointer arithmetics. | |
1328 | Gcc in general does not allow or handle objects larger | |
1329 | than half of the address space, hence the upper bound | |
1330 | is satisfied for pointers. */ | |
1331 | || POINTER_TYPE_P (type)); | |
1332 | mpz_clear (mstep); | |
1333 | mpz_clear (max); | |
1334 | ||
1335 | if (rolls_p && no_overflow_p) | |
1336 | return; | |
b8698a0f | 1337 | |
5be014d5 AP |
1338 | type1 = type; |
1339 | if (POINTER_TYPE_P (type)) | |
1340 | type1 = sizetype; | |
b3ce5b6e ZD |
1341 | |
1342 | /* Now the hard part; we must formulate the assumption(s) as expressions, and | |
1343 | we must be careful not to introduce overflow. */ | |
7f17528a | 1344 | |
6e42ce54 | 1345 | if (integer_nonzerop (iv0->step)) |
e9eb809d | 1346 | { |
5be014d5 AP |
1347 | diff = fold_build2 (MINUS_EXPR, type1, |
1348 | iv0->step, build_int_cst (type1, 1)); | |
e9eb809d | 1349 | |
7f17528a ZD |
1350 | /* We need to know that iv0->base >= MIN + iv0->step - 1. Since |
1351 | 0 address never belongs to any object, we can assume this for | |
1352 | pointers. */ | |
1353 | if (!POINTER_TYPE_P (type)) | |
e9eb809d | 1354 | { |
5be014d5 | 1355 | bound = fold_build2 (PLUS_EXPR, type1, |
7f17528a ZD |
1356 | TYPE_MIN_VALUE (type), diff); |
1357 | assumption = fold_build2 (GE_EXPR, boolean_type_node, | |
1358 | iv0->base, bound); | |
e9eb809d ZD |
1359 | } |
1360 | ||
7f17528a | 1361 | /* And then we can compute iv0->base - diff, and compare it with |
b8698a0f L |
1362 | iv1->base. */ |
1363 | mbzl = fold_build2 (MINUS_EXPR, type1, | |
d24a32a1 ZD |
1364 | fold_convert (type1, iv0->base), diff); |
1365 | mbzr = fold_convert (type1, iv1->base); | |
e9eb809d | 1366 | } |
7f17528a | 1367 | else |
e9eb809d | 1368 | { |
5be014d5 AP |
1369 | diff = fold_build2 (PLUS_EXPR, type1, |
1370 | iv1->step, build_int_cst (type1, 1)); | |
7f17528a ZD |
1371 | |
1372 | if (!POINTER_TYPE_P (type)) | |
e9eb809d | 1373 | { |
5be014d5 | 1374 | bound = fold_build2 (PLUS_EXPR, type1, |
7f17528a ZD |
1375 | TYPE_MAX_VALUE (type), diff); |
1376 | assumption = fold_build2 (LE_EXPR, boolean_type_node, | |
1377 | iv1->base, bound); | |
e9eb809d ZD |
1378 | } |
1379 | ||
d24a32a1 ZD |
1380 | mbzl = fold_convert (type1, iv0->base); |
1381 | mbzr = fold_build2 (MINUS_EXPR, type1, | |
1382 | fold_convert (type1, iv1->base), diff); | |
7f17528a | 1383 | } |
e9eb809d | 1384 | |
6e682d7e | 1385 | if (!integer_nonzerop (assumption)) |
7f17528a ZD |
1386 | niter->assumptions = fold_build2 (TRUTH_AND_EXPR, boolean_type_node, |
1387 | niter->assumptions, assumption); | |
b3ce5b6e ZD |
1388 | if (!rolls_p) |
1389 | { | |
1390 | mbz = fold_build2 (GT_EXPR, boolean_type_node, mbzl, mbzr); | |
1391 | niter->may_be_zero = fold_build2 (TRUTH_OR_EXPR, boolean_type_node, | |
1392 | niter->may_be_zero, mbz); | |
1393 | } | |
7f17528a | 1394 | } |
e9eb809d | 1395 | |
7f17528a ZD |
1396 | /* Determines number of iterations of loop whose ending condition |
1397 | is IV0 < IV1. TYPE is the type of the iv. The number of | |
b3ce5b6e | 1398 | iterations is stored to NITER. BNDS bounds the difference |
e36dc339 ZD |
1399 | IV1->base - IV0->base. EXIT_MUST_BE_TAKEN is true if we know |
1400 | that the exit must be taken eventually. */ | |
7f17528a ZD |
1401 | |
1402 | static bool | |
cdf66caf BC |
1403 | number_of_iterations_lt (struct loop *loop, tree type, affine_iv *iv0, |
1404 | affine_iv *iv1, struct tree_niter_desc *niter, | |
e36dc339 | 1405 | bool exit_must_be_taken, bounds *bnds) |
7f17528a ZD |
1406 | { |
1407 | tree niter_type = unsigned_type_for (type); | |
1408 | tree delta, step, s; | |
b3ce5b6e | 1409 | mpz_t mstep, tmp; |
7f17528a | 1410 | |
6e42ce54 | 1411 | if (integer_nonzerop (iv0->step)) |
17684618 ZD |
1412 | { |
1413 | niter->control = *iv0; | |
1414 | niter->cmp = LT_EXPR; | |
1415 | niter->bound = iv1->base; | |
1416 | } | |
1417 | else | |
1418 | { | |
1419 | niter->control = *iv1; | |
1420 | niter->cmp = GT_EXPR; | |
1421 | niter->bound = iv0->base; | |
1422 | } | |
1423 | ||
7f17528a ZD |
1424 | delta = fold_build2 (MINUS_EXPR, niter_type, |
1425 | fold_convert (niter_type, iv1->base), | |
1426 | fold_convert (niter_type, iv0->base)); | |
1427 | ||
1428 | /* First handle the special case that the step is +-1. */ | |
6e42ce54 ZD |
1429 | if ((integer_onep (iv0->step) && integer_zerop (iv1->step)) |
1430 | || (integer_all_onesp (iv1->step) && integer_zerop (iv0->step))) | |
7f17528a ZD |
1431 | { |
1432 | /* for (i = iv0->base; i < iv1->base; i++) | |
1433 | ||
1434 | or | |
82b85a85 | 1435 | |
7f17528a | 1436 | for (i = iv1->base; i > iv0->base; i--). |
b8698a0f | 1437 | |
7f17528a | 1438 | In both cases # of iterations is iv1->base - iv0->base, assuming that |
b3ce5b6e ZD |
1439 | iv1->base >= iv0->base. |
1440 | ||
1441 | First try to derive a lower bound on the value of | |
1442 | iv1->base - iv0->base, computed in full precision. If the difference | |
1443 | is nonnegative, we are done, otherwise we must record the | |
1444 | condition. */ | |
1445 | ||
1446 | if (mpz_sgn (bnds->below) < 0) | |
1447 | niter->may_be_zero = fold_build2 (LT_EXPR, boolean_type_node, | |
1448 | iv1->base, iv0->base); | |
7f17528a | 1449 | niter->niter = delta; |
807e902e KZ |
1450 | niter->max = widest_int::from (wi::from_mpz (niter_type, bnds->up, false), |
1451 | TYPE_SIGN (niter_type)); | |
2f07b722 | 1452 | niter->control.no_overflow = true; |
7f17528a | 1453 | return true; |
e9eb809d | 1454 | } |
7f17528a | 1455 | |
6e42ce54 | 1456 | if (integer_nonzerop (iv0->step)) |
7f17528a | 1457 | step = fold_convert (niter_type, iv0->step); |
e9eb809d | 1458 | else |
7f17528a ZD |
1459 | step = fold_convert (niter_type, |
1460 | fold_build1 (NEGATE_EXPR, type, iv1->step)); | |
1461 | ||
1462 | /* If we can determine the final value of the control iv exactly, we can | |
1463 | transform the condition to != comparison. In particular, this will be | |
1464 | the case if DELTA is constant. */ | |
b3ce5b6e | 1465 | if (number_of_iterations_lt_to_ne (type, iv0, iv1, niter, &delta, step, |
e36dc339 | 1466 | exit_must_be_taken, bnds)) |
e9eb809d | 1467 | { |
7f17528a ZD |
1468 | affine_iv zps; |
1469 | ||
ff5e9a94 | 1470 | zps.base = build_int_cst (niter_type, 0); |
7f17528a ZD |
1471 | zps.step = step; |
1472 | /* number_of_iterations_lt_to_ne will add assumptions that ensure that | |
1473 | zps does not overflow. */ | |
1474 | zps.no_overflow = true; | |
1475 | ||
cdf66caf BC |
1476 | return number_of_iterations_ne (loop, type, &zps, |
1477 | delta, niter, true, bnds); | |
e9eb809d ZD |
1478 | } |
1479 | ||
7f17528a ZD |
1480 | /* Make sure that the control iv does not overflow. */ |
1481 | if (!assert_no_overflow_lt (type, iv0, iv1, niter, step)) | |
1482 | return false; | |
e9eb809d | 1483 | |
7f17528a ZD |
1484 | /* We determine the number of iterations as (delta + step - 1) / step. For |
1485 | this to work, we must know that iv1->base >= iv0->base - step + 1, | |
1486 | otherwise the loop does not roll. */ | |
b3ce5b6e | 1487 | assert_loop_rolls_lt (type, iv0, iv1, niter, bnds); |
7f17528a ZD |
1488 | |
1489 | s = fold_build2 (MINUS_EXPR, niter_type, | |
ff5e9a94 | 1490 | step, build_int_cst (niter_type, 1)); |
7f17528a ZD |
1491 | delta = fold_build2 (PLUS_EXPR, niter_type, delta, s); |
1492 | niter->niter = fold_build2 (FLOOR_DIV_EXPR, niter_type, delta, step); | |
b3ce5b6e ZD |
1493 | |
1494 | mpz_init (mstep); | |
1495 | mpz_init (tmp); | |
807e902e | 1496 | wi::to_mpz (step, mstep, UNSIGNED); |
b3ce5b6e ZD |
1497 | mpz_add (tmp, bnds->up, mstep); |
1498 | mpz_sub_ui (tmp, tmp, 1); | |
1499 | mpz_fdiv_q (tmp, tmp, mstep); | |
807e902e KZ |
1500 | niter->max = widest_int::from (wi::from_mpz (niter_type, tmp, false), |
1501 | TYPE_SIGN (niter_type)); | |
b3ce5b6e ZD |
1502 | mpz_clear (mstep); |
1503 | mpz_clear (tmp); | |
1504 | ||
7f17528a | 1505 | return true; |
e9eb809d ZD |
1506 | } |
1507 | ||
7f17528a ZD |
1508 | /* Determines number of iterations of loop whose ending condition |
1509 | is IV0 <= IV1. TYPE is the type of the iv. The number of | |
e36dc339 | 1510 | iterations is stored to NITER. EXIT_MUST_BE_TAKEN is true if |
f08ac361 | 1511 | we know that this condition must eventually become false (we derived this |
7f17528a | 1512 | earlier, and possibly set NITER->assumptions to make sure this |
b3ce5b6e | 1513 | is the case). BNDS bounds the difference IV1->base - IV0->base. */ |
7f17528a ZD |
1514 | |
1515 | static bool | |
cdf66caf BC |
1516 | number_of_iterations_le (struct loop *loop, tree type, affine_iv *iv0, |
1517 | affine_iv *iv1, struct tree_niter_desc *niter, | |
1518 | bool exit_must_be_taken, bounds *bnds) | |
7f17528a ZD |
1519 | { |
1520 | tree assumption; | |
5be014d5 AP |
1521 | tree type1 = type; |
1522 | if (POINTER_TYPE_P (type)) | |
1523 | type1 = sizetype; | |
7f17528a ZD |
1524 | |
1525 | /* Say that IV0 is the control variable. Then IV0 <= IV1 iff | |
1526 | IV0 < IV1 + 1, assuming that IV1 is not equal to the greatest | |
1527 | value of the type. This we must know anyway, since if it is | |
e36dc339 | 1528 | equal to this value, the loop rolls forever. We do not check |
b8698a0f | 1529 | this condition for pointer type ivs, as the code cannot rely on |
e36dc339 ZD |
1530 | the object to that the pointer points being placed at the end of |
1531 | the address space (and more pragmatically, TYPE_{MIN,MAX}_VALUE is | |
1532 | not defined for pointers). */ | |
7f17528a | 1533 | |
e36dc339 | 1534 | if (!exit_must_be_taken && !POINTER_TYPE_P (type)) |
7f17528a | 1535 | { |
6e42ce54 | 1536 | if (integer_nonzerop (iv0->step)) |
7f17528a | 1537 | assumption = fold_build2 (NE_EXPR, boolean_type_node, |
97b4ba9f | 1538 | iv1->base, TYPE_MAX_VALUE (type)); |
7f17528a ZD |
1539 | else |
1540 | assumption = fold_build2 (NE_EXPR, boolean_type_node, | |
97b4ba9f | 1541 | iv0->base, TYPE_MIN_VALUE (type)); |
7f17528a | 1542 | |
6e682d7e | 1543 | if (integer_zerop (assumption)) |
7f17528a | 1544 | return false; |
6e682d7e | 1545 | if (!integer_nonzerop (assumption)) |
7f17528a ZD |
1546 | niter->assumptions = fold_build2 (TRUTH_AND_EXPR, boolean_type_node, |
1547 | niter->assumptions, assumption); | |
1548 | } | |
1549 | ||
6e42ce54 | 1550 | if (integer_nonzerop (iv0->step)) |
97b4ba9f JJ |
1551 | { |
1552 | if (POINTER_TYPE_P (type)) | |
5d49b6a7 | 1553 | iv1->base = fold_build_pointer_plus_hwi (iv1->base, 1); |
97b4ba9f JJ |
1554 | else |
1555 | iv1->base = fold_build2 (PLUS_EXPR, type1, iv1->base, | |
1556 | build_int_cst (type1, 1)); | |
1557 | } | |
1558 | else if (POINTER_TYPE_P (type)) | |
5d49b6a7 | 1559 | iv0->base = fold_build_pointer_plus_hwi (iv0->base, -1); |
7f17528a | 1560 | else |
5be014d5 AP |
1561 | iv0->base = fold_build2 (MINUS_EXPR, type1, |
1562 | iv0->base, build_int_cst (type1, 1)); | |
b3ce5b6e | 1563 | |
807e902e | 1564 | bounds_add (bnds, 1, type1); |
b3ce5b6e | 1565 | |
cdf66caf | 1566 | return number_of_iterations_lt (loop, type, iv0, iv1, niter, exit_must_be_taken, |
e36dc339 | 1567 | bnds); |
b3ce5b6e ZD |
1568 | } |
1569 | ||
1570 | /* Dumps description of affine induction variable IV to FILE. */ | |
1571 | ||
1572 | static void | |
1573 | dump_affine_iv (FILE *file, affine_iv *iv) | |
1574 | { | |
1575 | if (!integer_zerop (iv->step)) | |
1576 | fprintf (file, "["); | |
1577 | ||
1578 | print_generic_expr (dump_file, iv->base, TDF_SLIM); | |
1579 | ||
1580 | if (!integer_zerop (iv->step)) | |
1581 | { | |
1582 | fprintf (file, ", + , "); | |
1583 | print_generic_expr (dump_file, iv->step, TDF_SLIM); | |
1584 | fprintf (file, "]%s", iv->no_overflow ? "(no_overflow)" : ""); | |
1585 | } | |
7f17528a | 1586 | } |
e9eb809d | 1587 | |
7f17528a ZD |
1588 | /* Determine the number of iterations according to condition (for staying |
1589 | inside loop) which compares two induction variables using comparison | |
1590 | operator CODE. The induction variable on left side of the comparison | |
1591 | is IV0, the right-hand side is IV1. Both induction variables must have | |
1592 | type TYPE, which must be an integer or pointer type. The steps of the | |
1593 | ivs must be constants (or NULL_TREE, which is interpreted as constant zero). | |
f08ac361 | 1594 | |
b3ce5b6e ZD |
1595 | LOOP is the loop whose number of iterations we are determining. |
1596 | ||
f08ac361 ZD |
1597 | ONLY_EXIT is true if we are sure this is the only way the loop could be |
1598 | exited (including possibly non-returning function calls, exceptions, etc.) | |
1599 | -- in this case we can use the information whether the control induction | |
1600 | variables can overflow or not in a more efficient way. | |
b8698a0f | 1601 | |
870ca331 JH |
1602 | if EVERY_ITERATION is true, we know the test is executed on every iteration. |
1603 | ||
7f17528a | 1604 | The results (number of iterations and assumptions as described in |
3fadf78a | 1605 | comments at struct tree_niter_desc in tree-ssa-loop.h) are stored to NITER. |
7f17528a ZD |
1606 | Returns false if it fails to determine number of iterations, true if it |
1607 | was determined (possibly with some assumptions). */ | |
c33e657d ZD |
1608 | |
1609 | static bool | |
b3ce5b6e ZD |
1610 | number_of_iterations_cond (struct loop *loop, |
1611 | tree type, affine_iv *iv0, enum tree_code code, | |
f08ac361 | 1612 | affine_iv *iv1, struct tree_niter_desc *niter, |
870ca331 | 1613 | bool only_exit, bool every_iteration) |
e9eb809d | 1614 | { |
e36dc339 | 1615 | bool exit_must_be_taken = false, ret; |
b3ce5b6e | 1616 | bounds bnds; |
7f17528a | 1617 | |
870ca331 JH |
1618 | /* If the test is not executed every iteration, wrapping may make the test |
1619 | to pass again. | |
1620 | TODO: the overflow case can be still used as unreliable estimate of upper | |
1621 | bound. But we have no API to pass it down to number of iterations code | |
1622 | and, at present, it will not use it anyway. */ | |
1623 | if (!every_iteration | |
1624 | && (!iv0->no_overflow || !iv1->no_overflow | |
1625 | || code == NE_EXPR || code == EQ_EXPR)) | |
1626 | return false; | |
1627 | ||
7f17528a ZD |
1628 | /* The meaning of these assumptions is this: |
1629 | if !assumptions | |
1630 | then the rest of information does not have to be valid | |
1631 | if may_be_zero then the loop does not roll, even if | |
1632 | niter != 0. */ | |
1633 | niter->assumptions = boolean_true_node; | |
1634 | niter->may_be_zero = boolean_false_node; | |
1635 | niter->niter = NULL_TREE; | |
807e902e | 1636 | niter->max = 0; |
17684618 ZD |
1637 | niter->bound = NULL_TREE; |
1638 | niter->cmp = ERROR_MARK; | |
1639 | ||
7f17528a ZD |
1640 | /* Make < comparison from > ones, and for NE_EXPR comparisons, ensure that |
1641 | the control variable is on lhs. */ | |
1642 | if (code == GE_EXPR || code == GT_EXPR | |
6e42ce54 | 1643 | || (code == NE_EXPR && integer_zerop (iv0->step))) |
c33e657d | 1644 | { |
6b4db501 | 1645 | std::swap (iv0, iv1); |
c33e657d ZD |
1646 | code = swap_tree_comparison (code); |
1647 | } | |
e9eb809d | 1648 | |
7f17528a | 1649 | if (POINTER_TYPE_P (type)) |
e9eb809d | 1650 | { |
7f17528a ZD |
1651 | /* Comparison of pointers is undefined unless both iv0 and iv1 point |
1652 | to the same object. If they do, the control variable cannot wrap | |
1653 | (as wrap around the bounds of memory will never return a pointer | |
1654 | that would be guaranteed to point to the same object, even if we | |
e36dc339 | 1655 | avoid undefined behavior by casting to size_t and back). */ |
7f17528a ZD |
1656 | iv0->no_overflow = true; |
1657 | iv1->no_overflow = true; | |
1658 | } | |
e9eb809d | 1659 | |
e36dc339 ZD |
1660 | /* If the control induction variable does not overflow and the only exit |
1661 | from the loop is the one that we analyze, we know it must be taken | |
1662 | eventually. */ | |
1663 | if (only_exit) | |
1664 | { | |
1665 | if (!integer_zerop (iv0->step) && iv0->no_overflow) | |
1666 | exit_must_be_taken = true; | |
1667 | else if (!integer_zerop (iv1->step) && iv1->no_overflow) | |
1668 | exit_must_be_taken = true; | |
1669 | } | |
e9eb809d | 1670 | |
7f17528a ZD |
1671 | /* We can handle the case when neither of the sides of the comparison is |
1672 | invariant, provided that the test is NE_EXPR. This rarely occurs in | |
1673 | practice, but it is simple enough to manage. */ | |
6e42ce54 | 1674 | if (!integer_zerop (iv0->step) && !integer_zerop (iv1->step)) |
7f17528a | 1675 | { |
5ece9847 | 1676 | tree step_type = POINTER_TYPE_P (type) ? sizetype : type; |
7f17528a ZD |
1677 | if (code != NE_EXPR) |
1678 | return false; | |
e9eb809d | 1679 | |
5ece9847 | 1680 | iv0->step = fold_binary_to_constant (MINUS_EXPR, step_type, |
7f17528a ZD |
1681 | iv0->step, iv1->step); |
1682 | iv0->no_overflow = false; | |
5ece9847 | 1683 | iv1->step = build_int_cst (step_type, 0); |
7f17528a ZD |
1684 | iv1->no_overflow = true; |
1685 | } | |
c33e657d | 1686 | |
7f17528a ZD |
1687 | /* If the result of the comparison is a constant, the loop is weird. More |
1688 | precise handling would be possible, but the situation is not common enough | |
1689 | to waste time on it. */ | |
6e42ce54 | 1690 | if (integer_zerop (iv0->step) && integer_zerop (iv1->step)) |
7f17528a | 1691 | return false; |
c33e657d | 1692 | |
7f17528a ZD |
1693 | /* Ignore loops of while (i-- < 10) type. */ |
1694 | if (code != NE_EXPR) | |
1695 | { | |
1696 | if (iv0->step && tree_int_cst_sign_bit (iv0->step)) | |
c33e657d | 1697 | return false; |
c33e657d | 1698 | |
6e42ce54 | 1699 | if (!integer_zerop (iv1->step) && !tree_int_cst_sign_bit (iv1->step)) |
c33e657d | 1700 | return false; |
7f17528a | 1701 | } |
e9eb809d | 1702 | |
c0220ea4 | 1703 | /* If the loop exits immediately, there is nothing to do. */ |
5a892248 RB |
1704 | tree tem = fold_binary (code, boolean_type_node, iv0->base, iv1->base); |
1705 | if (tem && integer_zerop (tem)) | |
7f17528a | 1706 | { |
ff5e9a94 | 1707 | niter->niter = build_int_cst (unsigned_type_for (type), 0); |
807e902e | 1708 | niter->max = 0; |
7f17528a ZD |
1709 | return true; |
1710 | } | |
b8698a0f | 1711 | |
7f17528a ZD |
1712 | /* OK, now we know we have a senseful loop. Handle several cases, depending |
1713 | on what comparison operator is used. */ | |
b3ce5b6e ZD |
1714 | bound_difference (loop, iv1->base, iv0->base, &bnds); |
1715 | ||
1716 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1717 | { | |
1718 | fprintf (dump_file, | |
4dad0aca | 1719 | "Analyzing # of iterations of loop %d\n", loop->num); |
b3ce5b6e ZD |
1720 | |
1721 | fprintf (dump_file, " exit condition "); | |
1722 | dump_affine_iv (dump_file, iv0); | |
1723 | fprintf (dump_file, " %s ", | |
1724 | code == NE_EXPR ? "!=" | |
1725 | : code == LT_EXPR ? "<" | |
1726 | : "<="); | |
1727 | dump_affine_iv (dump_file, iv1); | |
1728 | fprintf (dump_file, "\n"); | |
1729 | ||
1730 | fprintf (dump_file, " bounds on difference of bases: "); | |
1731 | mpz_out_str (dump_file, 10, bnds.below); | |
1732 | fprintf (dump_file, " ... "); | |
1733 | mpz_out_str (dump_file, 10, bnds.up); | |
1734 | fprintf (dump_file, "\n"); | |
1735 | } | |
1736 | ||
7f17528a ZD |
1737 | switch (code) |
1738 | { | |
1739 | case NE_EXPR: | |
6e42ce54 | 1740 | gcc_assert (integer_zerop (iv1->step)); |
cdf66caf | 1741 | ret = number_of_iterations_ne (loop, type, iv0, iv1->base, niter, |
e36dc339 | 1742 | exit_must_be_taken, &bnds); |
b3ce5b6e ZD |
1743 | break; |
1744 | ||
7f17528a | 1745 | case LT_EXPR: |
cdf66caf BC |
1746 | ret = number_of_iterations_lt (loop, type, iv0, iv1, niter, |
1747 | exit_must_be_taken, &bnds); | |
b3ce5b6e ZD |
1748 | break; |
1749 | ||
7f17528a | 1750 | case LE_EXPR: |
cdf66caf BC |
1751 | ret = number_of_iterations_le (loop, type, iv0, iv1, niter, |
1752 | exit_must_be_taken, &bnds); | |
b3ce5b6e ZD |
1753 | break; |
1754 | ||
c33e657d ZD |
1755 | default: |
1756 | gcc_unreachable (); | |
1757 | } | |
b3ce5b6e ZD |
1758 | |
1759 | mpz_clear (bnds.up); | |
1760 | mpz_clear (bnds.below); | |
1761 | ||
1762 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1763 | { | |
1764 | if (ret) | |
1765 | { | |
1766 | fprintf (dump_file, " result:\n"); | |
1767 | if (!integer_nonzerop (niter->assumptions)) | |
1768 | { | |
1769 | fprintf (dump_file, " under assumptions "); | |
1770 | print_generic_expr (dump_file, niter->assumptions, TDF_SLIM); | |
1771 | fprintf (dump_file, "\n"); | |
1772 | } | |
1773 | ||
1774 | if (!integer_zerop (niter->may_be_zero)) | |
1775 | { | |
1776 | fprintf (dump_file, " zero if "); | |
1777 | print_generic_expr (dump_file, niter->may_be_zero, TDF_SLIM); | |
1778 | fprintf (dump_file, "\n"); | |
1779 | } | |
1780 | ||
1781 | fprintf (dump_file, " # of iterations "); | |
1782 | print_generic_expr (dump_file, niter->niter, TDF_SLIM); | |
1783 | fprintf (dump_file, ", bounded by "); | |
807e902e | 1784 | print_decu (niter->max, dump_file); |
b3ce5b6e ZD |
1785 | fprintf (dump_file, "\n"); |
1786 | } | |
1787 | else | |
1788 | fprintf (dump_file, " failed\n\n"); | |
1789 | } | |
1790 | return ret; | |
e9eb809d ZD |
1791 | } |
1792 | ||
1d481ba8 ZD |
1793 | /* Substitute NEW for OLD in EXPR and fold the result. */ |
1794 | ||
1795 | static tree | |
c22940cd | 1796 | simplify_replace_tree (tree expr, tree old, tree new_tree) |
1d481ba8 ZD |
1797 | { |
1798 | unsigned i, n; | |
1799 | tree ret = NULL_TREE, e, se; | |
1800 | ||
1801 | if (!expr) | |
1802 | return NULL_TREE; | |
1803 | ||
76c85743 RG |
1804 | /* Do not bother to replace constants. */ |
1805 | if (CONSTANT_CLASS_P (old)) | |
1806 | return expr; | |
1807 | ||
1d481ba8 ZD |
1808 | if (expr == old |
1809 | || operand_equal_p (expr, old, 0)) | |
c22940cd | 1810 | return unshare_expr (new_tree); |
1d481ba8 | 1811 | |
726a989a | 1812 | if (!EXPR_P (expr)) |
1d481ba8 ZD |
1813 | return expr; |
1814 | ||
5039610b | 1815 | n = TREE_OPERAND_LENGTH (expr); |
1d481ba8 ZD |
1816 | for (i = 0; i < n; i++) |
1817 | { | |
1818 | e = TREE_OPERAND (expr, i); | |
c22940cd | 1819 | se = simplify_replace_tree (e, old, new_tree); |
1d481ba8 ZD |
1820 | if (e == se) |
1821 | continue; | |
1822 | ||
1823 | if (!ret) | |
1824 | ret = copy_node (expr); | |
1825 | ||
1826 | TREE_OPERAND (ret, i) = se; | |
1827 | } | |
1828 | ||
1829 | return (ret ? fold (ret) : expr); | |
1830 | } | |
1831 | ||
be1b5cba | 1832 | /* Expand definitions of ssa names in EXPR as long as they are simple |
fc06280e BC |
1833 | enough, and return the new expression. If STOP is specified, stop |
1834 | expanding if EXPR equals to it. */ | |
be1b5cba | 1835 | |
d7bf3bcf | 1836 | tree |
fc06280e | 1837 | expand_simple_operations (tree expr, tree stop) |
be1b5cba ZD |
1838 | { |
1839 | unsigned i, n; | |
726a989a | 1840 | tree ret = NULL_TREE, e, ee, e1; |
6fff2603 | 1841 | enum tree_code code; |
355fe088 | 1842 | gimple *stmt; |
6fff2603 JJ |
1843 | |
1844 | if (expr == NULL_TREE) | |
1845 | return expr; | |
be1b5cba ZD |
1846 | |
1847 | if (is_gimple_min_invariant (expr)) | |
1848 | return expr; | |
1849 | ||
6fff2603 | 1850 | code = TREE_CODE (expr); |
be1b5cba ZD |
1851 | if (IS_EXPR_CODE_CLASS (TREE_CODE_CLASS (code))) |
1852 | { | |
5039610b | 1853 | n = TREE_OPERAND_LENGTH (expr); |
be1b5cba ZD |
1854 | for (i = 0; i < n; i++) |
1855 | { | |
1856 | e = TREE_OPERAND (expr, i); | |
fc06280e | 1857 | ee = expand_simple_operations (e, stop); |
be1b5cba ZD |
1858 | if (e == ee) |
1859 | continue; | |
1860 | ||
1861 | if (!ret) | |
1862 | ret = copy_node (expr); | |
1863 | ||
1864 | TREE_OPERAND (ret, i) = ee; | |
1865 | } | |
1866 | ||
6ac01510 ILT |
1867 | if (!ret) |
1868 | return expr; | |
1869 | ||
1870 | fold_defer_overflow_warnings (); | |
1871 | ret = fold (ret); | |
1872 | fold_undefer_and_ignore_overflow_warnings (); | |
1873 | return ret; | |
be1b5cba ZD |
1874 | } |
1875 | ||
fc06280e BC |
1876 | /* Stop if it's not ssa name or the one we don't want to expand. */ |
1877 | if (TREE_CODE (expr) != SSA_NAME || expr == stop) | |
be1b5cba ZD |
1878 | return expr; |
1879 | ||
1880 | stmt = SSA_NAME_DEF_STMT (expr); | |
726a989a | 1881 | if (gimple_code (stmt) == GIMPLE_PHI) |
b3ce5b6e ZD |
1882 | { |
1883 | basic_block src, dest; | |
1884 | ||
726a989a | 1885 | if (gimple_phi_num_args (stmt) != 1) |
b3ce5b6e ZD |
1886 | return expr; |
1887 | e = PHI_ARG_DEF (stmt, 0); | |
1888 | ||
1889 | /* Avoid propagating through loop exit phi nodes, which | |
1890 | could break loop-closed SSA form restrictions. */ | |
726a989a | 1891 | dest = gimple_bb (stmt); |
b3ce5b6e ZD |
1892 | src = single_pred (dest); |
1893 | if (TREE_CODE (e) == SSA_NAME | |
1894 | && src->loop_father != dest->loop_father) | |
1895 | return expr; | |
1896 | ||
fc06280e | 1897 | return expand_simple_operations (e, stop); |
b3ce5b6e | 1898 | } |
726a989a | 1899 | if (gimple_code (stmt) != GIMPLE_ASSIGN) |
be1b5cba ZD |
1900 | return expr; |
1901 | ||
c3a9b91b RB |
1902 | /* Avoid expanding to expressions that contain SSA names that need |
1903 | to take part in abnormal coalescing. */ | |
1904 | ssa_op_iter iter; | |
1905 | FOR_EACH_SSA_TREE_OPERAND (e, stmt, iter, SSA_OP_USE) | |
1906 | if (SSA_NAME_OCCURS_IN_ABNORMAL_PHI (e)) | |
1907 | return expr; | |
1908 | ||
726a989a RB |
1909 | e = gimple_assign_rhs1 (stmt); |
1910 | code = gimple_assign_rhs_code (stmt); | |
1911 | if (get_gimple_rhs_class (code) == GIMPLE_SINGLE_RHS) | |
1912 | { | |
1913 | if (is_gimple_min_invariant (e)) | |
1914 | return e; | |
1915 | ||
1916 | if (code == SSA_NAME) | |
fc06280e | 1917 | return expand_simple_operations (e, stop); |
726a989a RB |
1918 | |
1919 | return expr; | |
1920 | } | |
1921 | ||
1922 | switch (code) | |
1923 | { | |
1a87cf0c | 1924 | CASE_CONVERT: |
726a989a | 1925 | /* Casts are simple. */ |
fc06280e | 1926 | ee = expand_simple_operations (e, stop); |
726a989a RB |
1927 | return fold_build1 (code, TREE_TYPE (expr), ee); |
1928 | ||
1929 | case PLUS_EXPR: | |
1930 | case MINUS_EXPR: | |
20bd649a MP |
1931 | if (ANY_INTEGRAL_TYPE_P (TREE_TYPE (expr)) |
1932 | && TYPE_OVERFLOW_TRAPS (TREE_TYPE (expr))) | |
8b228266 RB |
1933 | return expr; |
1934 | /* Fallthru. */ | |
726a989a | 1935 | case POINTER_PLUS_EXPR: |
be1b5cba | 1936 | /* And increments and decrements by a constant are simple. */ |
726a989a RB |
1937 | e1 = gimple_assign_rhs2 (stmt); |
1938 | if (!is_gimple_min_invariant (e1)) | |
1939 | return expr; | |
1940 | ||
fc06280e | 1941 | ee = expand_simple_operations (e, stop); |
726a989a | 1942 | return fold_build2 (code, TREE_TYPE (expr), ee, e1); |
be1b5cba | 1943 | |
726a989a RB |
1944 | default: |
1945 | return expr; | |
1946 | } | |
be1b5cba ZD |
1947 | } |
1948 | ||
e9eb809d | 1949 | /* Tries to simplify EXPR using the condition COND. Returns the simplified |
be1b5cba | 1950 | expression (or EXPR unchanged, if no simplification was possible). */ |
e9eb809d ZD |
1951 | |
1952 | static tree | |
8aa46dd2 | 1953 | tree_simplify_using_condition_1 (tree cond, tree expr) |
e9eb809d ZD |
1954 | { |
1955 | bool changed; | |
8aa46dd2 | 1956 | tree e, e0, e1, e2, notcond; |
e9eb809d ZD |
1957 | enum tree_code code = TREE_CODE (expr); |
1958 | ||
1959 | if (code == INTEGER_CST) | |
1960 | return expr; | |
1961 | ||
1962 | if (code == TRUTH_OR_EXPR | |
1963 | || code == TRUTH_AND_EXPR | |
1964 | || code == COND_EXPR) | |
1965 | { | |
1966 | changed = false; | |
1967 | ||
8aa46dd2 | 1968 | e0 = tree_simplify_using_condition_1 (cond, TREE_OPERAND (expr, 0)); |
e9eb809d ZD |
1969 | if (TREE_OPERAND (expr, 0) != e0) |
1970 | changed = true; | |
1971 | ||
8aa46dd2 | 1972 | e1 = tree_simplify_using_condition_1 (cond, TREE_OPERAND (expr, 1)); |
e9eb809d ZD |
1973 | if (TREE_OPERAND (expr, 1) != e1) |
1974 | changed = true; | |
1975 | ||
1976 | if (code == COND_EXPR) | |
1977 | { | |
8aa46dd2 | 1978 | e2 = tree_simplify_using_condition_1 (cond, TREE_OPERAND (expr, 2)); |
e9eb809d ZD |
1979 | if (TREE_OPERAND (expr, 2) != e2) |
1980 | changed = true; | |
1981 | } | |
1982 | else | |
1983 | e2 = NULL_TREE; | |
1984 | ||
1985 | if (changed) | |
1986 | { | |
1987 | if (code == COND_EXPR) | |
c33e657d | 1988 | expr = fold_build3 (code, boolean_type_node, e0, e1, e2); |
e9eb809d | 1989 | else |
c33e657d | 1990 | expr = fold_build2 (code, boolean_type_node, e0, e1); |
e9eb809d ZD |
1991 | } |
1992 | ||
1993 | return expr; | |
1994 | } | |
1995 | ||
1d481ba8 ZD |
1996 | /* In case COND is equality, we may be able to simplify EXPR by copy/constant |
1997 | propagation, and vice versa. Fold does not handle this, since it is | |
1998 | considered too expensive. */ | |
1999 | if (TREE_CODE (cond) == EQ_EXPR) | |
2000 | { | |
2001 | e0 = TREE_OPERAND (cond, 0); | |
2002 | e1 = TREE_OPERAND (cond, 1); | |
2003 | ||
2004 | /* We know that e0 == e1. Check whether we cannot simplify expr | |
2005 | using this fact. */ | |
2006 | e = simplify_replace_tree (expr, e0, e1); | |
6e682d7e | 2007 | if (integer_zerop (e) || integer_nonzerop (e)) |
1d481ba8 ZD |
2008 | return e; |
2009 | ||
2010 | e = simplify_replace_tree (expr, e1, e0); | |
6e682d7e | 2011 | if (integer_zerop (e) || integer_nonzerop (e)) |
1d481ba8 ZD |
2012 | return e; |
2013 | } | |
2014 | if (TREE_CODE (expr) == EQ_EXPR) | |
2015 | { | |
2016 | e0 = TREE_OPERAND (expr, 0); | |
2017 | e1 = TREE_OPERAND (expr, 1); | |
2018 | ||
2019 | /* If e0 == e1 (EXPR) implies !COND, then EXPR cannot be true. */ | |
2020 | e = simplify_replace_tree (cond, e0, e1); | |
6e682d7e | 2021 | if (integer_zerop (e)) |
1d481ba8 ZD |
2022 | return e; |
2023 | e = simplify_replace_tree (cond, e1, e0); | |
6e682d7e | 2024 | if (integer_zerop (e)) |
1d481ba8 ZD |
2025 | return e; |
2026 | } | |
2027 | if (TREE_CODE (expr) == NE_EXPR) | |
2028 | { | |
2029 | e0 = TREE_OPERAND (expr, 0); | |
2030 | e1 = TREE_OPERAND (expr, 1); | |
2031 | ||
2032 | /* If e0 == e1 (!EXPR) implies !COND, then EXPR must be true. */ | |
2033 | e = simplify_replace_tree (cond, e0, e1); | |
6e682d7e | 2034 | if (integer_zerop (e)) |
1d481ba8 ZD |
2035 | return boolean_true_node; |
2036 | e = simplify_replace_tree (cond, e1, e0); | |
6e682d7e | 2037 | if (integer_zerop (e)) |
1d481ba8 ZD |
2038 | return boolean_true_node; |
2039 | } | |
2040 | ||
e9eb809d ZD |
2041 | /* Check whether COND ==> EXPR. */ |
2042 | notcond = invert_truthvalue (cond); | |
8aa46dd2 | 2043 | e = fold_binary (TRUTH_OR_EXPR, boolean_type_node, notcond, expr); |
6e682d7e | 2044 | if (e && integer_nonzerop (e)) |
e9eb809d ZD |
2045 | return e; |
2046 | ||
2047 | /* Check whether COND ==> not EXPR. */ | |
8aa46dd2 | 2048 | e = fold_binary (TRUTH_AND_EXPR, boolean_type_node, cond, expr); |
6e682d7e | 2049 | if (e && integer_zerop (e)) |
e9eb809d ZD |
2050 | return e; |
2051 | ||
2052 | return expr; | |
2053 | } | |
2054 | ||
be1b5cba ZD |
2055 | /* Tries to simplify EXPR using the condition COND. Returns the simplified |
2056 | expression (or EXPR unchanged, if no simplification was possible). | |
2057 | Wrapper around tree_simplify_using_condition_1 that ensures that chains | |
2058 | of simple operations in definitions of ssa names in COND are expanded, | |
2059 | so that things like casts or incrementing the value of the bound before | |
2060 | the loop do not cause us to fail. */ | |
2061 | ||
2062 | static tree | |
8aa46dd2 | 2063 | tree_simplify_using_condition (tree cond, tree expr) |
be1b5cba | 2064 | { |
8aa46dd2 | 2065 | cond = expand_simple_operations (cond); |
be1b5cba | 2066 | |
8aa46dd2 | 2067 | return tree_simplify_using_condition_1 (cond, expr); |
be1b5cba | 2068 | } |
b16fb82d | 2069 | |
e9eb809d | 2070 | /* Tries to simplify EXPR using the conditions on entry to LOOP. |
e9eb809d | 2071 | Returns the simplified expression (or EXPR unchanged, if no |
f3c5f3a3 | 2072 | simplification was possible). */ |
e9eb809d | 2073 | |
f3c5f3a3 | 2074 | tree |
8aa46dd2 | 2075 | simplify_using_initial_conditions (struct loop *loop, tree expr) |
e9eb809d ZD |
2076 | { |
2077 | edge e; | |
2078 | basic_block bb; | |
355fe088 | 2079 | gimple *stmt; |
8aa46dd2 | 2080 | tree cond, expanded, backup; |
b16fb82d | 2081 | int cnt = 0; |
e9eb809d ZD |
2082 | |
2083 | if (TREE_CODE (expr) == INTEGER_CST) | |
2084 | return expr; | |
2085 | ||
8aa46dd2 BC |
2086 | backup = expanded = expand_simple_operations (expr); |
2087 | ||
b16fb82d RG |
2088 | /* Limit walking the dominators to avoid quadraticness in |
2089 | the number of BBs times the number of loops in degenerate | |
2090 | cases. */ | |
e9eb809d | 2091 | for (bb = loop->header; |
fefa31b5 | 2092 | bb != ENTRY_BLOCK_PTR_FOR_FN (cfun) && cnt < MAX_DOMINATORS_TO_WALK; |
e9eb809d ZD |
2093 | bb = get_immediate_dominator (CDI_DOMINATORS, bb)) |
2094 | { | |
c5cbcccf | 2095 | if (!single_pred_p (bb)) |
e9eb809d | 2096 | continue; |
c5cbcccf | 2097 | e = single_pred_edge (bb); |
e9eb809d ZD |
2098 | |
2099 | if (!(e->flags & (EDGE_TRUE_VALUE | EDGE_FALSE_VALUE))) | |
2100 | continue; | |
2101 | ||
726a989a RB |
2102 | stmt = last_stmt (e->src); |
2103 | cond = fold_build2 (gimple_cond_code (stmt), | |
2104 | boolean_type_node, | |
2105 | gimple_cond_lhs (stmt), | |
2106 | gimple_cond_rhs (stmt)); | |
e9eb809d ZD |
2107 | if (e->flags & EDGE_FALSE_VALUE) |
2108 | cond = invert_truthvalue (cond); | |
8aa46dd2 | 2109 | expanded = tree_simplify_using_condition (cond, expanded); |
eff1e5af | 2110 | /* Break if EXPR is simplified to const values. */ |
8aa46dd2 BC |
2111 | if (expanded |
2112 | && (integer_zerop (expanded) || integer_nonzerop (expanded))) | |
2113 | return expanded; | |
eff1e5af | 2114 | |
b16fb82d | 2115 | ++cnt; |
e9eb809d ZD |
2116 | } |
2117 | ||
8aa46dd2 BC |
2118 | /* Return the original expression if no simplification is done. */ |
2119 | return operand_equal_p (backup, expanded, 0) ? expr : expanded; | |
e9eb809d ZD |
2120 | } |
2121 | ||
c33e657d ZD |
2122 | /* Tries to simplify EXPR using the evolutions of the loop invariants |
2123 | in the superloops of LOOP. Returns the simplified expression | |
2124 | (or EXPR unchanged, if no simplification was possible). */ | |
2125 | ||
2126 | static tree | |
2127 | simplify_using_outer_evolutions (struct loop *loop, tree expr) | |
2128 | { | |
2129 | enum tree_code code = TREE_CODE (expr); | |
2130 | bool changed; | |
2131 | tree e, e0, e1, e2; | |
2132 | ||
2133 | if (is_gimple_min_invariant (expr)) | |
2134 | return expr; | |
2135 | ||
2136 | if (code == TRUTH_OR_EXPR | |
2137 | || code == TRUTH_AND_EXPR | |
2138 | || code == COND_EXPR) | |
2139 | { | |
2140 | changed = false; | |
2141 | ||
2142 | e0 = simplify_using_outer_evolutions (loop, TREE_OPERAND (expr, 0)); | |
2143 | if (TREE_OPERAND (expr, 0) != e0) | |
2144 | changed = true; | |
2145 | ||
2146 | e1 = simplify_using_outer_evolutions (loop, TREE_OPERAND (expr, 1)); | |
2147 | if (TREE_OPERAND (expr, 1) != e1) | |
2148 | changed = true; | |
2149 | ||
2150 | if (code == COND_EXPR) | |
2151 | { | |
2152 | e2 = simplify_using_outer_evolutions (loop, TREE_OPERAND (expr, 2)); | |
2153 | if (TREE_OPERAND (expr, 2) != e2) | |
2154 | changed = true; | |
2155 | } | |
2156 | else | |
2157 | e2 = NULL_TREE; | |
2158 | ||
2159 | if (changed) | |
2160 | { | |
2161 | if (code == COND_EXPR) | |
2162 | expr = fold_build3 (code, boolean_type_node, e0, e1, e2); | |
2163 | else | |
2164 | expr = fold_build2 (code, boolean_type_node, e0, e1); | |
2165 | } | |
2166 | ||
2167 | return expr; | |
2168 | } | |
2169 | ||
2170 | e = instantiate_parameters (loop, expr); | |
2171 | if (is_gimple_min_invariant (e)) | |
2172 | return e; | |
2173 | ||
2174 | return expr; | |
2175 | } | |
2176 | ||
f08ac361 ZD |
2177 | /* Returns true if EXIT is the only possible exit from LOOP. */ |
2178 | ||
52778e2a | 2179 | bool |
22ea9ec0 | 2180 | loop_only_exit_p (const struct loop *loop, const_edge exit) |
f08ac361 ZD |
2181 | { |
2182 | basic_block *body; | |
726a989a | 2183 | gimple_stmt_iterator bsi; |
f08ac361 | 2184 | unsigned i; |
f08ac361 | 2185 | |
ac8f6c69 | 2186 | if (exit != single_exit (loop)) |
f08ac361 ZD |
2187 | return false; |
2188 | ||
2189 | body = get_loop_body (loop); | |
2190 | for (i = 0; i < loop->num_nodes; i++) | |
2191 | { | |
726a989a | 2192 | for (bsi = gsi_start_bb (body[i]); !gsi_end_p (bsi); gsi_next (&bsi)) |
21bcd7be | 2193 | if (stmt_can_terminate_bb_p (gsi_stmt (bsi))) |
ccae0c85 ML |
2194 | { |
2195 | free (body); | |
2196 | return true; | |
2197 | } | |
f08ac361 ZD |
2198 | } |
2199 | ||
2200 | free (body); | |
2201 | return true; | |
2202 | } | |
2203 | ||
e9eb809d | 2204 | /* Stores description of number of iterations of LOOP derived from |
43aabfcf BC |
2205 | EXIT (an exit edge of the LOOP) in NITER. Returns true if some useful |
2206 | information could be derived (and fields of NITER have meaning described | |
2207 | in comments at struct tree_niter_desc declaration), false otherwise. | |
cd0f6278 | 2208 | When EVERY_ITERATION is true, only tests that are known to be executed |
43aabfcf | 2209 | every iteration are considered (i.e. only test that alone bounds the loop). |
faa1612a BC |
2210 | If AT_STMT is not NULL, this function stores LOOP's condition statement in |
2211 | it when returning true. */ | |
e9eb809d ZD |
2212 | |
2213 | bool | |
43aabfcf BC |
2214 | number_of_iterations_exit_assumptions (struct loop *loop, edge exit, |
2215 | struct tree_niter_desc *niter, | |
faa1612a | 2216 | gcond **at_stmt, bool every_iteration) |
e9eb809d | 2217 | { |
355fe088 | 2218 | gimple *last; |
538dd0b7 | 2219 | gcond *stmt; |
726a989a | 2220 | tree type; |
a6f778b2 | 2221 | tree op0, op1; |
e9eb809d | 2222 | enum tree_code code; |
a6f778b2 | 2223 | affine_iv iv0, iv1; |
870ca331 | 2224 | bool safe; |
e9eb809d | 2225 | |
18767ebc BC |
2226 | /* Nothing to analyze if the loop is known to be infinite. */ |
2227 | if (loop_constraint_set_p (loop, LOOP_C_INFINITE)) | |
2228 | return false; | |
2229 | ||
870ca331 JH |
2230 | safe = dominated_by_p (CDI_DOMINATORS, loop->latch, exit->src); |
2231 | ||
2232 | if (every_iteration && !safe) | |
e9eb809d ZD |
2233 | return false; |
2234 | ||
2235 | niter->assumptions = boolean_false_node; | |
2f07b722 BC |
2236 | niter->control.base = NULL_TREE; |
2237 | niter->control.step = NULL_TREE; | |
2238 | niter->control.no_overflow = false; | |
538dd0b7 DM |
2239 | last = last_stmt (exit->src); |
2240 | if (!last) | |
2241 | return false; | |
2242 | stmt = dyn_cast <gcond *> (last); | |
2243 | if (!stmt) | |
e9eb809d ZD |
2244 | return false; |
2245 | ||
2246 | /* We want the condition for staying inside loop. */ | |
726a989a | 2247 | code = gimple_cond_code (stmt); |
e9eb809d | 2248 | if (exit->flags & EDGE_TRUE_VALUE) |
726a989a | 2249 | code = invert_tree_comparison (code, false); |
e9eb809d | 2250 | |
e9eb809d ZD |
2251 | switch (code) |
2252 | { | |
2253 | case GT_EXPR: | |
2254 | case GE_EXPR: | |
e9eb809d ZD |
2255 | case LT_EXPR: |
2256 | case LE_EXPR: | |
870ca331 | 2257 | case NE_EXPR: |
e9eb809d ZD |
2258 | break; |
2259 | ||
2260 | default: | |
2261 | return false; | |
2262 | } | |
b8698a0f | 2263 | |
726a989a RB |
2264 | op0 = gimple_cond_lhs (stmt); |
2265 | op1 = gimple_cond_rhs (stmt); | |
e9eb809d ZD |
2266 | type = TREE_TYPE (op0); |
2267 | ||
2268 | if (TREE_CODE (type) != INTEGER_TYPE | |
b3393f1f | 2269 | && !POINTER_TYPE_P (type)) |
e9eb809d | 2270 | return false; |
b8698a0f | 2271 | |
43aabfcf BC |
2272 | tree iv0_niters = NULL_TREE; |
2273 | if (!simple_iv_with_niters (loop, loop_containing_stmt (stmt), | |
2274 | op0, &iv0, &iv0_niters, false)) | |
2275 | return false; | |
2276 | tree iv1_niters = NULL_TREE; | |
2277 | if (!simple_iv_with_niters (loop, loop_containing_stmt (stmt), | |
2278 | op1, &iv1, &iv1_niters, false)) | |
e9eb809d | 2279 | return false; |
43aabfcf BC |
2280 | /* Give up on complicated case. */ |
2281 | if (iv0_niters && iv1_niters) | |
e9eb809d ZD |
2282 | return false; |
2283 | ||
6ac01510 | 2284 | /* We don't want to see undefined signed overflow warnings while |
ea2c620c | 2285 | computing the number of iterations. */ |
6ac01510 ILT |
2286 | fold_defer_overflow_warnings (); |
2287 | ||
7f17528a ZD |
2288 | iv0.base = expand_simple_operations (iv0.base); |
2289 | iv1.base = expand_simple_operations (iv1.base); | |
b3ce5b6e | 2290 | if (!number_of_iterations_cond (loop, type, &iv0, code, &iv1, niter, |
870ca331 | 2291 | loop_only_exit_p (loop, exit), safe)) |
6ac01510 ILT |
2292 | { |
2293 | fold_undefer_and_ignore_overflow_warnings (); | |
2294 | return false; | |
2295 | } | |
c33e657d | 2296 | |
43aabfcf BC |
2297 | /* Incorporate additional assumption implied by control iv. */ |
2298 | tree iv_niters = iv0_niters ? iv0_niters : iv1_niters; | |
2299 | if (iv_niters) | |
2300 | { | |
2301 | tree assumption = fold_build2 (LE_EXPR, boolean_type_node, niter->niter, | |
2302 | fold_convert (TREE_TYPE (niter->niter), | |
2303 | iv_niters)); | |
2304 | ||
2305 | if (!integer_nonzerop (assumption)) | |
2306 | niter->assumptions = fold_build2 (TRUTH_AND_EXPR, boolean_type_node, | |
2307 | niter->assumptions, assumption); | |
2308 | ||
2309 | /* Refine upper bound if possible. */ | |
2310 | if (TREE_CODE (iv_niters) == INTEGER_CST | |
2311 | && niter->max > wi::to_widest (iv_niters)) | |
2312 | niter->max = wi::to_widest (iv_niters); | |
2313 | } | |
2314 | ||
18767ebc BC |
2315 | /* There is no assumptions if the loop is known to be finite. */ |
2316 | if (!integer_zerop (niter->assumptions) | |
2317 | && loop_constraint_set_p (loop, LOOP_C_FINITE)) | |
2318 | niter->assumptions = boolean_true_node; | |
2319 | ||
c33e657d ZD |
2320 | if (optimize >= 3) |
2321 | { | |
2322 | niter->assumptions = simplify_using_outer_evolutions (loop, | |
2323 | niter->assumptions); | |
2324 | niter->may_be_zero = simplify_using_outer_evolutions (loop, | |
2325 | niter->may_be_zero); | |
2326 | niter->niter = simplify_using_outer_evolutions (loop, niter->niter); | |
2327 | } | |
e9eb809d | 2328 | |
e9eb809d ZD |
2329 | niter->assumptions |
2330 | = simplify_using_initial_conditions (loop, | |
b3ce5b6e | 2331 | niter->assumptions); |
e9eb809d ZD |
2332 | niter->may_be_zero |
2333 | = simplify_using_initial_conditions (loop, | |
b3ce5b6e | 2334 | niter->may_be_zero); |
f9cc1a70 | 2335 | |
6ac01510 ILT |
2336 | fold_undefer_and_ignore_overflow_warnings (); |
2337 | ||
f2a1b469 JH |
2338 | /* If NITER has simplified into a constant, update MAX. */ |
2339 | if (TREE_CODE (niter->niter) == INTEGER_CST) | |
018b22f3 | 2340 | niter->max = wi::to_widest (niter->niter); |
f2a1b469 | 2341 | |
faa1612a BC |
2342 | if (at_stmt) |
2343 | *at_stmt = stmt; | |
2344 | ||
43aabfcf BC |
2345 | return (!integer_zerop (niter->assumptions)); |
2346 | } | |
b8698a0f | 2347 | |
70e1d145 AH |
2348 | /* Like number_of_iterations_exit_assumptions, but return TRUE only if |
2349 | the niter information holds unconditionally. */ | |
f9cc1a70 | 2350 | |
43aabfcf BC |
2351 | bool |
2352 | number_of_iterations_exit (struct loop *loop, edge exit, | |
2353 | struct tree_niter_desc *niter, | |
faa1612a | 2354 | bool warn, bool every_iteration) |
43aabfcf | 2355 | { |
faa1612a | 2356 | gcond *stmt; |
43aabfcf | 2357 | if (!number_of_iterations_exit_assumptions (loop, exit, niter, |
faa1612a | 2358 | &stmt, every_iteration)) |
43aabfcf | 2359 | return false; |
f9cc1a70 | 2360 | |
faa1612a BC |
2361 | if (integer_nonzerop (niter->assumptions)) |
2362 | return true; | |
2363 | ||
2364 | if (warn) | |
9d975cb6 JJ |
2365 | warning_at (gimple_location_safe (stmt), |
2366 | OPT_Wunsafe_loop_optimizations, | |
2367 | "missed loop optimization, the loop counter may overflow"); | |
faa1612a BC |
2368 | |
2369 | return false; | |
e9eb809d ZD |
2370 | } |
2371 | ||
ca4c3169 ZD |
2372 | /* Try to determine the number of iterations of LOOP. If we succeed, |
2373 | expression giving number of iterations is returned and *EXIT is | |
2374 | set to the edge from that the information is obtained. Otherwise | |
2375 | chrec_dont_know is returned. */ | |
2376 | ||
2377 | tree | |
2378 | find_loop_niter (struct loop *loop, edge *exit) | |
2379 | { | |
ca83d385 | 2380 | unsigned i; |
9771b263 | 2381 | vec<edge> exits = get_loop_exit_edges (loop); |
ca4c3169 ZD |
2382 | edge ex; |
2383 | tree niter = NULL_TREE, aniter; | |
2384 | struct tree_niter_desc desc; | |
2385 | ||
2386 | *exit = NULL; | |
9771b263 | 2387 | FOR_EACH_VEC_ELT (exits, i, ex) |
ca4c3169 | 2388 | { |
f9cc1a70 | 2389 | if (!number_of_iterations_exit (loop, ex, &desc, false)) |
ca4c3169 ZD |
2390 | continue; |
2391 | ||
6e682d7e | 2392 | if (integer_nonzerop (desc.may_be_zero)) |
ca4c3169 ZD |
2393 | { |
2394 | /* We exit in the first iteration through this exit. | |
2395 | We won't find anything better. */ | |
ff5e9a94 | 2396 | niter = build_int_cst (unsigned_type_node, 0); |
ca4c3169 ZD |
2397 | *exit = ex; |
2398 | break; | |
2399 | } | |
2400 | ||
6e682d7e | 2401 | if (!integer_zerop (desc.may_be_zero)) |
ca4c3169 ZD |
2402 | continue; |
2403 | ||
2404 | aniter = desc.niter; | |
2405 | ||
2406 | if (!niter) | |
2407 | { | |
2408 | /* Nothing recorded yet. */ | |
2409 | niter = aniter; | |
2410 | *exit = ex; | |
2411 | continue; | |
2412 | } | |
2413 | ||
2414 | /* Prefer constants, the lower the better. */ | |
2415 | if (TREE_CODE (aniter) != INTEGER_CST) | |
2416 | continue; | |
2417 | ||
2418 | if (TREE_CODE (niter) != INTEGER_CST) | |
2419 | { | |
2420 | niter = aniter; | |
2421 | *exit = ex; | |
2422 | continue; | |
2423 | } | |
2424 | ||
2425 | if (tree_int_cst_lt (aniter, niter)) | |
2426 | { | |
2427 | niter = aniter; | |
2428 | *exit = ex; | |
2429 | continue; | |
2430 | } | |
2431 | } | |
9771b263 | 2432 | exits.release (); |
ca4c3169 ZD |
2433 | |
2434 | return niter ? niter : chrec_dont_know; | |
2435 | } | |
2436 | ||
f87c9042 JH |
2437 | /* Return true if loop is known to have bounded number of iterations. */ |
2438 | ||
2439 | bool | |
2440 | finite_loop_p (struct loop *loop) | |
2441 | { | |
807e902e | 2442 | widest_int nit; |
9e3920e9 | 2443 | int flags; |
f87c9042 | 2444 | |
9e3920e9 JJ |
2445 | flags = flags_from_decl_or_type (current_function_decl); |
2446 | if ((flags & (ECF_CONST|ECF_PURE)) && !(flags & ECF_LOOPING_CONST_OR_PURE)) | |
f87c9042 JH |
2447 | { |
2448 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2449 | fprintf (dump_file, "Found loop %i to be finite: it is within pure or const function.\n", | |
2450 | loop->num); | |
2451 | return true; | |
2452 | } | |
b8698a0f | 2453 | |
1bc60b18 JH |
2454 | if (loop->any_upper_bound |
2455 | || max_loop_iterations (loop, &nit)) | |
f87c9042 | 2456 | { |
1bc60b18 JH |
2457 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2458 | fprintf (dump_file, "Found loop %i to be finite: upper bound found.\n", | |
2459 | loop->num); | |
2460 | return true; | |
f87c9042 | 2461 | } |
1bc60b18 | 2462 | return false; |
f87c9042 JH |
2463 | } |
2464 | ||
e9eb809d ZD |
2465 | /* |
2466 | ||
2467 | Analysis of a number of iterations of a loop by a brute-force evaluation. | |
2468 | ||
2469 | */ | |
2470 | ||
2471 | /* Bound on the number of iterations we try to evaluate. */ | |
2472 | ||
2473 | #define MAX_ITERATIONS_TO_TRACK \ | |
2474 | ((unsigned) PARAM_VALUE (PARAM_MAX_ITERATIONS_TO_TRACK)) | |
2475 | ||
2476 | /* Returns the loop phi node of LOOP such that ssa name X is derived from its | |
2477 | result by a chain of operations such that all but exactly one of their | |
2478 | operands are constants. */ | |
2479 | ||
538dd0b7 | 2480 | static gphi * |
e9eb809d ZD |
2481 | chain_of_csts_start (struct loop *loop, tree x) |
2482 | { | |
355fe088 | 2483 | gimple *stmt = SSA_NAME_DEF_STMT (x); |
f47c96aa | 2484 | tree use; |
726a989a RB |
2485 | basic_block bb = gimple_bb (stmt); |
2486 | enum tree_code code; | |
e9eb809d ZD |
2487 | |
2488 | if (!bb | |
2489 | || !flow_bb_inside_loop_p (loop, bb)) | |
726a989a | 2490 | return NULL; |
b8698a0f | 2491 | |
726a989a | 2492 | if (gimple_code (stmt) == GIMPLE_PHI) |
e9eb809d ZD |
2493 | { |
2494 | if (bb == loop->header) | |
538dd0b7 | 2495 | return as_a <gphi *> (stmt); |
e9eb809d | 2496 | |
726a989a | 2497 | return NULL; |
e9eb809d ZD |
2498 | } |
2499 | ||
100f09a5 RB |
2500 | if (gimple_code (stmt) != GIMPLE_ASSIGN |
2501 | || gimple_assign_rhs_class (stmt) == GIMPLE_TERNARY_RHS) | |
726a989a | 2502 | return NULL; |
e9eb809d | 2503 | |
726a989a RB |
2504 | code = gimple_assign_rhs_code (stmt); |
2505 | if (gimple_references_memory_p (stmt) | |
726a989a | 2506 | || TREE_CODE_CLASS (code) == tcc_reference |
5006671f RG |
2507 | || (code == ADDR_EXPR |
2508 | && !is_gimple_min_invariant (gimple_assign_rhs1 (stmt)))) | |
726a989a | 2509 | return NULL; |
f47c96aa AM |
2510 | |
2511 | use = SINGLE_SSA_TREE_OPERAND (stmt, SSA_OP_USE); | |
5006671f | 2512 | if (use == NULL_TREE) |
726a989a | 2513 | return NULL; |
e9eb809d | 2514 | |
f47c96aa | 2515 | return chain_of_csts_start (loop, use); |
e9eb809d ZD |
2516 | } |
2517 | ||
2518 | /* Determines whether the expression X is derived from a result of a phi node | |
2519 | in header of LOOP such that | |
2520 | ||
2521 | * the derivation of X consists only from operations with constants | |
2522 | * the initial value of the phi node is constant | |
2523 | * the value of the phi node in the next iteration can be derived from the | |
5558f089 JJ |
2524 | value in the current iteration by a chain of operations with constants, |
2525 | or is also a constant | |
b8698a0f | 2526 | |
726a989a | 2527 | If such phi node exists, it is returned, otherwise NULL is returned. */ |
e9eb809d | 2528 | |
538dd0b7 | 2529 | static gphi * |
e9eb809d ZD |
2530 | get_base_for (struct loop *loop, tree x) |
2531 | { | |
538dd0b7 | 2532 | gphi *phi; |
726a989a | 2533 | tree init, next; |
e9eb809d ZD |
2534 | |
2535 | if (is_gimple_min_invariant (x)) | |
726a989a | 2536 | return NULL; |
e9eb809d ZD |
2537 | |
2538 | phi = chain_of_csts_start (loop, x); | |
2539 | if (!phi) | |
726a989a | 2540 | return NULL; |
e9eb809d ZD |
2541 | |
2542 | init = PHI_ARG_DEF_FROM_EDGE (phi, loop_preheader_edge (loop)); | |
2543 | next = PHI_ARG_DEF_FROM_EDGE (phi, loop_latch_edge (loop)); | |
2544 | ||
e9eb809d | 2545 | if (!is_gimple_min_invariant (init)) |
726a989a | 2546 | return NULL; |
e9eb809d | 2547 | |
5558f089 JJ |
2548 | if (TREE_CODE (next) == SSA_NAME |
2549 | && chain_of_csts_start (loop, next) != phi) | |
726a989a | 2550 | return NULL; |
e9eb809d ZD |
2551 | |
2552 | return phi; | |
2553 | } | |
2554 | ||
b8698a0f L |
2555 | /* Given an expression X, then |
2556 | ||
ed52affe | 2557 | * if X is NULL_TREE, we return the constant BASE. |
5558f089 | 2558 | * if X is a constant, we return the constant X. |
e9eb809d ZD |
2559 | * otherwise X is a SSA name, whose value in the considered loop is derived |
2560 | by a chain of operations with constant from a result of a phi node in | |
2561 | the header of the loop. Then we return value of X when the value of the | |
2562 | result of this phi node is given by the constant BASE. */ | |
2563 | ||
2564 | static tree | |
2565 | get_val_for (tree x, tree base) | |
2566 | { | |
355fe088 | 2567 | gimple *stmt; |
e9eb809d | 2568 | |
100f09a5 | 2569 | gcc_checking_assert (is_gimple_min_invariant (base)); |
ed52affe | 2570 | |
e9eb809d ZD |
2571 | if (!x) |
2572 | return base; | |
5558f089 JJ |
2573 | else if (is_gimple_min_invariant (x)) |
2574 | return x; | |
e9eb809d ZD |
2575 | |
2576 | stmt = SSA_NAME_DEF_STMT (x); | |
726a989a | 2577 | if (gimple_code (stmt) == GIMPLE_PHI) |
e9eb809d ZD |
2578 | return base; |
2579 | ||
100f09a5 | 2580 | gcc_checking_assert (is_gimple_assign (stmt)); |
726a989a RB |
2581 | |
2582 | /* STMT must be either an assignment of a single SSA name or an | |
2583 | expression involving an SSA name and a constant. Try to fold that | |
2584 | expression using the value for the SSA name. */ | |
0f336c35 RG |
2585 | if (gimple_assign_ssa_name_copy_p (stmt)) |
2586 | return get_val_for (gimple_assign_rhs1 (stmt), base); | |
2587 | else if (gimple_assign_rhs_class (stmt) == GIMPLE_UNARY_RHS | |
2588 | && TREE_CODE (gimple_assign_rhs1 (stmt)) == SSA_NAME) | |
5558f089 JJ |
2589 | return fold_build1 (gimple_assign_rhs_code (stmt), |
2590 | gimple_expr_type (stmt), | |
2591 | get_val_for (gimple_assign_rhs1 (stmt), base)); | |
0f336c35 | 2592 | else if (gimple_assign_rhs_class (stmt) == GIMPLE_BINARY_RHS) |
726a989a | 2593 | { |
0f336c35 RG |
2594 | tree rhs1 = gimple_assign_rhs1 (stmt); |
2595 | tree rhs2 = gimple_assign_rhs2 (stmt); | |
2596 | if (TREE_CODE (rhs1) == SSA_NAME) | |
2597 | rhs1 = get_val_for (rhs1, base); | |
2598 | else if (TREE_CODE (rhs2) == SSA_NAME) | |
2599 | rhs2 = get_val_for (rhs2, base); | |
2600 | else | |
2601 | gcc_unreachable (); | |
2602 | return fold_build2 (gimple_assign_rhs_code (stmt), | |
2603 | gimple_expr_type (stmt), rhs1, rhs2); | |
f47c96aa | 2604 | } |
726a989a | 2605 | else |
0f336c35 | 2606 | gcc_unreachable (); |
e9eb809d ZD |
2607 | } |
2608 | ||
726a989a | 2609 | |
e9eb809d ZD |
2610 | /* Tries to count the number of iterations of LOOP till it exits by EXIT |
2611 | by brute force -- i.e. by determining the value of the operands of the | |
2612 | condition at EXIT in first few iterations of the loop (assuming that | |
2613 | these values are constant) and determining the first one in that the | |
2614 | condition is not satisfied. Returns the constant giving the number | |
2615 | of the iterations of LOOP if successful, chrec_dont_know otherwise. */ | |
2616 | ||
2617 | tree | |
2618 | loop_niter_by_eval (struct loop *loop, edge exit) | |
2619 | { | |
726a989a RB |
2620 | tree acnd; |
2621 | tree op[2], val[2], next[2], aval[2]; | |
538dd0b7 | 2622 | gphi *phi; |
355fe088 | 2623 | gimple *cond; |
e9eb809d ZD |
2624 | unsigned i, j; |
2625 | enum tree_code cmp; | |
2626 | ||
2627 | cond = last_stmt (exit->src); | |
726a989a | 2628 | if (!cond || gimple_code (cond) != GIMPLE_COND) |
e9eb809d ZD |
2629 | return chrec_dont_know; |
2630 | ||
726a989a | 2631 | cmp = gimple_cond_code (cond); |
e9eb809d | 2632 | if (exit->flags & EDGE_TRUE_VALUE) |
726a989a | 2633 | cmp = invert_tree_comparison (cmp, false); |
e9eb809d | 2634 | |
e9eb809d ZD |
2635 | switch (cmp) |
2636 | { | |
2637 | case EQ_EXPR: | |
2638 | case NE_EXPR: | |
2639 | case GT_EXPR: | |
2640 | case GE_EXPR: | |
2641 | case LT_EXPR: | |
2642 | case LE_EXPR: | |
726a989a RB |
2643 | op[0] = gimple_cond_lhs (cond); |
2644 | op[1] = gimple_cond_rhs (cond); | |
e9eb809d ZD |
2645 | break; |
2646 | ||
2647 | default: | |
2648 | return chrec_dont_know; | |
2649 | } | |
2650 | ||
2651 | for (j = 0; j < 2; j++) | |
2652 | { | |
726a989a | 2653 | if (is_gimple_min_invariant (op[j])) |
e9eb809d | 2654 | { |
726a989a RB |
2655 | val[j] = op[j]; |
2656 | next[j] = NULL_TREE; | |
2657 | op[j] = NULL_TREE; | |
e9eb809d ZD |
2658 | } |
2659 | else | |
2660 | { | |
726a989a RB |
2661 | phi = get_base_for (loop, op[j]); |
2662 | if (!phi) | |
2663 | return chrec_dont_know; | |
2664 | val[j] = PHI_ARG_DEF_FROM_EDGE (phi, loop_preheader_edge (loop)); | |
2665 | next[j] = PHI_ARG_DEF_FROM_EDGE (phi, loop_latch_edge (loop)); | |
e9eb809d ZD |
2666 | } |
2667 | } | |
2668 | ||
6ac01510 ILT |
2669 | /* Don't issue signed overflow warnings. */ |
2670 | fold_defer_overflow_warnings (); | |
2671 | ||
e9eb809d ZD |
2672 | for (i = 0; i < MAX_ITERATIONS_TO_TRACK; i++) |
2673 | { | |
2674 | for (j = 0; j < 2; j++) | |
2675 | aval[j] = get_val_for (op[j], val[j]); | |
2676 | ||
2f133f46 | 2677 | acnd = fold_binary (cmp, boolean_type_node, aval[0], aval[1]); |
6e682d7e | 2678 | if (acnd && integer_zerop (acnd)) |
e9eb809d | 2679 | { |
6ac01510 | 2680 | fold_undefer_and_ignore_overflow_warnings (); |
e9eb809d ZD |
2681 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2682 | fprintf (dump_file, | |
2683 | "Proved that loop %d iterates %d times using brute force.\n", | |
2684 | loop->num, i); | |
7d60be94 | 2685 | return build_int_cst (unsigned_type_node, i); |
e9eb809d ZD |
2686 | } |
2687 | ||
2688 | for (j = 0; j < 2; j++) | |
ed52affe | 2689 | { |
5558f089 | 2690 | aval[j] = val[j]; |
ed52affe RG |
2691 | val[j] = get_val_for (next[j], val[j]); |
2692 | if (!is_gimple_min_invariant (val[j])) | |
6ac01510 ILT |
2693 | { |
2694 | fold_undefer_and_ignore_overflow_warnings (); | |
2695 | return chrec_dont_know; | |
2696 | } | |
ed52affe | 2697 | } |
5558f089 JJ |
2698 | |
2699 | /* If the next iteration would use the same base values | |
2700 | as the current one, there is no point looping further, | |
2701 | all following iterations will be the same as this one. */ | |
2702 | if (val[0] == aval[0] && val[1] == aval[1]) | |
2703 | break; | |
e9eb809d ZD |
2704 | } |
2705 | ||
6ac01510 ILT |
2706 | fold_undefer_and_ignore_overflow_warnings (); |
2707 | ||
e9eb809d ZD |
2708 | return chrec_dont_know; |
2709 | } | |
2710 | ||
2711 | /* Finds the exit of the LOOP by that the loop exits after a constant | |
2712 | number of iterations and stores the exit edge to *EXIT. The constant | |
2713 | giving the number of iterations of LOOP is returned. The number of | |
2714 | iterations is determined using loop_niter_by_eval (i.e. by brute force | |
2715 | evaluation). If we are unable to find the exit for that loop_niter_by_eval | |
2716 | determines the number of iterations, chrec_dont_know is returned. */ | |
2717 | ||
2718 | tree | |
2719 | find_loop_niter_by_eval (struct loop *loop, edge *exit) | |
2720 | { | |
ca83d385 | 2721 | unsigned i; |
9771b263 | 2722 | vec<edge> exits = get_loop_exit_edges (loop); |
e9eb809d ZD |
2723 | edge ex; |
2724 | tree niter = NULL_TREE, aniter; | |
2725 | ||
2726 | *exit = NULL; | |
2cee1509 RG |
2727 | |
2728 | /* Loops with multiple exits are expensive to handle and less important. */ | |
2729 | if (!flag_expensive_optimizations | |
9771b263 | 2730 | && exits.length () > 1) |
f5843d08 | 2731 | { |
9771b263 | 2732 | exits.release (); |
f5843d08 RG |
2733 | return chrec_dont_know; |
2734 | } | |
2cee1509 | 2735 | |
9771b263 | 2736 | FOR_EACH_VEC_ELT (exits, i, ex) |
e9eb809d | 2737 | { |
e9eb809d ZD |
2738 | if (!just_once_each_iteration_p (loop, ex->src)) |
2739 | continue; | |
2740 | ||
2741 | aniter = loop_niter_by_eval (loop, ex); | |
ca4c3169 | 2742 | if (chrec_contains_undetermined (aniter)) |
e9eb809d ZD |
2743 | continue; |
2744 | ||
2745 | if (niter | |
ca4c3169 | 2746 | && !tree_int_cst_lt (aniter, niter)) |
e9eb809d ZD |
2747 | continue; |
2748 | ||
2749 | niter = aniter; | |
2750 | *exit = ex; | |
2751 | } | |
9771b263 | 2752 | exits.release (); |
e9eb809d ZD |
2753 | |
2754 | return niter ? niter : chrec_dont_know; | |
2755 | } | |
2756 | ||
2757 | /* | |
2758 | ||
2759 | Analysis of upper bounds on number of iterations of a loop. | |
2760 | ||
2761 | */ | |
2762 | ||
807e902e | 2763 | static widest_int derive_constant_upper_bound_ops (tree, tree, |
726a989a RB |
2764 | enum tree_code, tree); |
2765 | ||
2766 | /* Returns a constant upper bound on the value of the right-hand side of | |
2767 | an assignment statement STMT. */ | |
2768 | ||
807e902e | 2769 | static widest_int |
355fe088 | 2770 | derive_constant_upper_bound_assign (gimple *stmt) |
726a989a RB |
2771 | { |
2772 | enum tree_code code = gimple_assign_rhs_code (stmt); | |
2773 | tree op0 = gimple_assign_rhs1 (stmt); | |
2774 | tree op1 = gimple_assign_rhs2 (stmt); | |
2775 | ||
2776 | return derive_constant_upper_bound_ops (TREE_TYPE (gimple_assign_lhs (stmt)), | |
2777 | op0, code, op1); | |
2778 | } | |
2779 | ||
0ad1d5a1 ZD |
2780 | /* Returns a constant upper bound on the value of expression VAL. VAL |
2781 | is considered to be unsigned. If its type is signed, its value must | |
b3ce5b6e | 2782 | be nonnegative. */ |
b8698a0f | 2783 | |
807e902e | 2784 | static widest_int |
726a989a RB |
2785 | derive_constant_upper_bound (tree val) |
2786 | { | |
2787 | enum tree_code code; | |
d1e2bb2d | 2788 | tree op0, op1, op2; |
726a989a | 2789 | |
d1e2bb2d | 2790 | extract_ops_from_tree (val, &code, &op0, &op1, &op2); |
726a989a RB |
2791 | return derive_constant_upper_bound_ops (TREE_TYPE (val), op0, code, op1); |
2792 | } | |
2793 | ||
2794 | /* Returns a constant upper bound on the value of expression OP0 CODE OP1, | |
2795 | whose type is TYPE. The expression is considered to be unsigned. If | |
2796 | its type is signed, its value must be nonnegative. */ | |
b8698a0f | 2797 | |
807e902e | 2798 | static widest_int |
726a989a RB |
2799 | derive_constant_upper_bound_ops (tree type, tree op0, |
2800 | enum tree_code code, tree op1) | |
763f4527 | 2801 | { |
726a989a | 2802 | tree subtype, maxt; |
e53d562a | 2803 | widest_int bnd, max, cst; |
355fe088 | 2804 | gimple *stmt; |
0ad1d5a1 ZD |
2805 | |
2806 | if (INTEGRAL_TYPE_P (type)) | |
2807 | maxt = TYPE_MAX_VALUE (type); | |
2808 | else | |
2809 | maxt = upper_bound_in_type (type, type); | |
2810 | ||
807e902e | 2811 | max = wi::to_widest (maxt); |
0ad1d5a1 | 2812 | |
726a989a | 2813 | switch (code) |
0ad1d5a1 ZD |
2814 | { |
2815 | case INTEGER_CST: | |
807e902e | 2816 | return wi::to_widest (op0); |
0ad1d5a1 | 2817 | |
1043771b | 2818 | CASE_CONVERT: |
0ad1d5a1 ZD |
2819 | subtype = TREE_TYPE (op0); |
2820 | if (!TYPE_UNSIGNED (subtype) | |
2821 | /* If TYPE is also signed, the fact that VAL is nonnegative implies | |
2822 | that OP0 is nonnegative. */ | |
2823 | && TYPE_UNSIGNED (type) | |
b3ce5b6e | 2824 | && !tree_expr_nonnegative_p (op0)) |
0ad1d5a1 ZD |
2825 | { |
2826 | /* If we cannot prove that the casted expression is nonnegative, | |
2827 | we cannot establish more useful upper bound than the precision | |
2828 | of the type gives us. */ | |
2829 | return max; | |
2830 | } | |
763f4527 | 2831 | |
0ad1d5a1 ZD |
2832 | /* We now know that op0 is an nonnegative value. Try deriving an upper |
2833 | bound for it. */ | |
b3ce5b6e | 2834 | bnd = derive_constant_upper_bound (op0); |
0ad1d5a1 ZD |
2835 | |
2836 | /* If the bound does not fit in TYPE, max. value of TYPE could be | |
2837 | attained. */ | |
807e902e | 2838 | if (wi::ltu_p (max, bnd)) |
0ad1d5a1 ZD |
2839 | return max; |
2840 | ||
2841 | return bnd; | |
2842 | ||
2843 | case PLUS_EXPR: | |
5be014d5 | 2844 | case POINTER_PLUS_EXPR: |
0ad1d5a1 | 2845 | case MINUS_EXPR: |
0ad1d5a1 | 2846 | if (TREE_CODE (op1) != INTEGER_CST |
b3ce5b6e | 2847 | || !tree_expr_nonnegative_p (op0)) |
0ad1d5a1 ZD |
2848 | return max; |
2849 | ||
20fb52af ZD |
2850 | /* Canonicalize to OP0 - CST. Consider CST to be signed, in order to |
2851 | choose the most logical way how to treat this constant regardless | |
2852 | of the signedness of the type. */ | |
807e902e | 2853 | cst = wi::sext (wi::to_widest (op1), TYPE_PRECISION (type)); |
726a989a | 2854 | if (code != MINUS_EXPR) |
27bcd47c | 2855 | cst = -cst; |
0ad1d5a1 | 2856 | |
b3ce5b6e | 2857 | bnd = derive_constant_upper_bound (op0); |
0ad1d5a1 | 2858 | |
807e902e | 2859 | if (wi::neg_p (cst)) |
0ad1d5a1 | 2860 | { |
27bcd47c | 2861 | cst = -cst; |
0ad1d5a1 | 2862 | /* Avoid CST == 0x80000... */ |
807e902e | 2863 | if (wi::neg_p (cst)) |
6f3d1a5e | 2864 | return max; |
0ad1d5a1 | 2865 | |
20fb52af | 2866 | /* OP0 + CST. We need to check that |
0ad1d5a1 ZD |
2867 | BND <= MAX (type) - CST. */ |
2868 | ||
e53d562a RB |
2869 | widest_int mmax = max - cst; |
2870 | if (wi::leu_p (bnd, mmax)) | |
0ad1d5a1 ZD |
2871 | return max; |
2872 | ||
27bcd47c | 2873 | return bnd + cst; |
0ad1d5a1 ZD |
2874 | } |
2875 | else | |
2876 | { | |
20fb52af ZD |
2877 | /* OP0 - CST, where CST >= 0. |
2878 | ||
2879 | If TYPE is signed, we have already verified that OP0 >= 0, and we | |
2880 | know that the result is nonnegative. This implies that | |
2881 | VAL <= BND - CST. | |
2882 | ||
2883 | If TYPE is unsigned, we must additionally know that OP0 >= CST, | |
2884 | otherwise the operation underflows. | |
2885 | */ | |
2886 | ||
2887 | /* This should only happen if the type is unsigned; however, for | |
b3ce5b6e | 2888 | buggy programs that use overflowing signed arithmetics even with |
20fb52af | 2889 | -fno-wrapv, this condition may also be true for signed values. */ |
807e902e | 2890 | if (wi::ltu_p (bnd, cst)) |
0ad1d5a1 ZD |
2891 | return max; |
2892 | ||
b3ce5b6e ZD |
2893 | if (TYPE_UNSIGNED (type)) |
2894 | { | |
2895 | tree tem = fold_binary (GE_EXPR, boolean_type_node, op0, | |
807e902e | 2896 | wide_int_to_tree (type, cst)); |
b3ce5b6e ZD |
2897 | if (!tem || integer_nonzerop (tem)) |
2898 | return max; | |
2899 | } | |
20fb52af | 2900 | |
27bcd47c | 2901 | bnd -= cst; |
0ad1d5a1 ZD |
2902 | } |
2903 | ||
2904 | return bnd; | |
2905 | ||
2906 | case FLOOR_DIV_EXPR: | |
2907 | case EXACT_DIV_EXPR: | |
0ad1d5a1 ZD |
2908 | if (TREE_CODE (op1) != INTEGER_CST |
2909 | || tree_int_cst_sign_bit (op1)) | |
2910 | return max; | |
2911 | ||
b3ce5b6e | 2912 | bnd = derive_constant_upper_bound (op0); |
807e902e | 2913 | return wi::udiv_floor (bnd, wi::to_widest (op1)); |
0ad1d5a1 | 2914 | |
946e1bc7 | 2915 | case BIT_AND_EXPR: |
946e1bc7 ZD |
2916 | if (TREE_CODE (op1) != INTEGER_CST |
2917 | || tree_int_cst_sign_bit (op1)) | |
2918 | return max; | |
807e902e | 2919 | return wi::to_widest (op1); |
946e1bc7 ZD |
2920 | |
2921 | case SSA_NAME: | |
726a989a RB |
2922 | stmt = SSA_NAME_DEF_STMT (op0); |
2923 | if (gimple_code (stmt) != GIMPLE_ASSIGN | |
2924 | || gimple_assign_lhs (stmt) != op0) | |
946e1bc7 | 2925 | return max; |
726a989a | 2926 | return derive_constant_upper_bound_assign (stmt); |
946e1bc7 | 2927 | |
b8698a0f | 2928 | default: |
0ad1d5a1 ZD |
2929 | return max; |
2930 | } | |
763f4527 ZD |
2931 | } |
2932 | ||
fbd28bc3 JJ |
2933 | /* Emit a -Waggressive-loop-optimizations warning if needed. */ |
2934 | ||
2935 | static void | |
2936 | do_warn_aggressive_loop_optimizations (struct loop *loop, | |
355fe088 | 2937 | widest_int i_bound, gimple *stmt) |
fbd28bc3 JJ |
2938 | { |
2939 | /* Don't warn if the loop doesn't have known constant bound. */ | |
2940 | if (!loop->nb_iterations | |
2941 | || TREE_CODE (loop->nb_iterations) != INTEGER_CST | |
2942 | || !warn_aggressive_loop_optimizations | |
2943 | /* To avoid warning multiple times for the same loop, | |
2944 | only start warning when we preserve loops. */ | |
2945 | || (cfun->curr_properties & PROP_loops) == 0 | |
2946 | /* Only warn once per loop. */ | |
2947 | || loop->warned_aggressive_loop_optimizations | |
2948 | /* Only warn if undefined behavior gives us lower estimate than the | |
2949 | known constant bound. */ | |
807e902e | 2950 | || wi::cmpu (i_bound, wi::to_widest (loop->nb_iterations)) >= 0 |
fbd28bc3 JJ |
2951 | /* And undefined behavior happens unconditionally. */ |
2952 | || !dominated_by_p (CDI_DOMINATORS, loop->latch, gimple_bb (stmt))) | |
2953 | return; | |
2954 | ||
2955 | edge e = single_exit (loop); | |
2956 | if (e == NULL) | |
2957 | return; | |
2958 | ||
355fe088 | 2959 | gimple *estmt = last_stmt (e->src); |
973dabae MLI |
2960 | char buf[WIDE_INT_PRINT_BUFFER_SIZE]; |
2961 | print_dec (i_bound, buf, TYPE_UNSIGNED (TREE_TYPE (loop->nb_iterations)) | |
2962 | ? UNSIGNED : SIGNED); | |
44398cbe | 2963 | if (warning_at (gimple_location (stmt), OPT_Waggressive_loop_optimizations, |
973dabae MLI |
2964 | "iteration %s invokes undefined behavior", buf)) |
2965 | inform (gimple_location (estmt), "within this loop"); | |
fbd28bc3 JJ |
2966 | loop->warned_aggressive_loop_optimizations = true; |
2967 | } | |
2968 | ||
b3ce5b6e | 2969 | /* Records that AT_STMT is executed at most BOUND + 1 times in LOOP. IS_EXIT |
946e1bc7 ZD |
2970 | is true if the loop is exited immediately after STMT, and this exit |
2971 | is taken at last when the STMT is executed BOUND + 1 times. | |
fa10beec | 2972 | REALISTIC is true if BOUND is expected to be close to the real number |
9bdb685e | 2973 | of iterations. UPPER is true if we are sure the loop iterates at most |
807e902e | 2974 | BOUND times. I_BOUND is a widest_int upper estimate on BOUND. */ |
e9eb809d | 2975 | |
946e1bc7 | 2976 | static void |
807e902e | 2977 | record_estimate (struct loop *loop, tree bound, const widest_int &i_bound, |
355fe088 | 2978 | gimple *at_stmt, bool is_exit, bool realistic, bool upper) |
e9eb809d | 2979 | { |
807e902e | 2980 | widest_int delta; |
e9eb809d ZD |
2981 | |
2982 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2983 | { | |
946e1bc7 | 2984 | fprintf (dump_file, "Statement %s", is_exit ? "(exit)" : ""); |
726a989a | 2985 | print_gimple_stmt (dump_file, at_stmt, 0, TDF_SLIM); |
9bdb685e ZD |
2986 | fprintf (dump_file, " is %sexecuted at most ", |
2987 | upper ? "" : "probably "); | |
e9eb809d | 2988 | print_generic_expr (dump_file, bound, TDF_SLIM); |
763f4527 | 2989 | fprintf (dump_file, " (bounded by "); |
807e902e | 2990 | print_decu (i_bound, dump_file); |
946e1bc7 | 2991 | fprintf (dump_file, ") + 1 times in loop %d.\n", loop->num); |
e9eb809d ZD |
2992 | } |
2993 | ||
9bdb685e ZD |
2994 | /* If the I_BOUND is just an estimate of BOUND, it rarely is close to the |
2995 | real number of iterations. */ | |
2996 | if (TREE_CODE (bound) != INTEGER_CST) | |
2997 | realistic = false; | |
f2a1b469 | 2998 | else |
807e902e | 2999 | gcc_checking_assert (i_bound == wi::to_widest (bound)); |
9bdb685e ZD |
3000 | |
3001 | /* If we have a guaranteed upper bound, record it in the appropriate | |
fbd28bc3 JJ |
3002 | list, unless this is an !is_exit bound (i.e. undefined behavior in |
3003 | at_stmt) in a loop with known constant number of iterations. */ | |
3004 | if (upper | |
3005 | && (is_exit | |
3006 | || loop->nb_iterations == NULL_TREE | |
3007 | || TREE_CODE (loop->nb_iterations) != INTEGER_CST)) | |
9bdb685e | 3008 | { |
766090c2 | 3009 | struct nb_iter_bound *elt = ggc_alloc<nb_iter_bound> (); |
9bdb685e ZD |
3010 | |
3011 | elt->bound = i_bound; | |
3012 | elt->stmt = at_stmt; | |
3013 | elt->is_exit = is_exit; | |
3014 | elt->next = loop->bounds; | |
3015 | loop->bounds = elt; | |
3016 | } | |
3017 | ||
cd0f6278 JH |
3018 | /* If statement is executed on every path to the loop latch, we can directly |
3019 | infer the upper bound on the # of iterations of the loop. */ | |
3020 | if (!dominated_by_p (CDI_DOMINATORS, loop->latch, gimple_bb (at_stmt))) | |
105e29c5 | 3021 | upper = false; |
cd0f6278 | 3022 | |
9bdb685e | 3023 | /* Update the number of iteration estimates according to the bound. |
05322355 JH |
3024 | If at_stmt is an exit then the loop latch is executed at most BOUND times, |
3025 | otherwise it can be executed BOUND + 1 times. We will lower the estimate | |
3026 | later if such statement must be executed on last iteration */ | |
3027 | if (is_exit) | |
807e902e | 3028 | delta = 0; |
9bdb685e | 3029 | else |
807e902e KZ |
3030 | delta = 1; |
3031 | widest_int new_i_bound = i_bound + delta; | |
9bdb685e | 3032 | |
7fa7289d | 3033 | /* If an overflow occurred, ignore the result. */ |
807e902e | 3034 | if (wi::ltu_p (new_i_bound, delta)) |
9bdb685e ZD |
3035 | return; |
3036 | ||
fbd28bc3 | 3037 | if (upper && !is_exit) |
807e902e KZ |
3038 | do_warn_aggressive_loop_optimizations (loop, new_i_bound, at_stmt); |
3039 | record_niter_bound (loop, new_i_bound, realistic, upper); | |
e9eb809d ZD |
3040 | } |
3041 | ||
2f07b722 BC |
3042 | /* Records the control iv analyzed in NITER for LOOP if the iv is valid |
3043 | and doesn't overflow. */ | |
3044 | ||
3045 | static void | |
3046 | record_control_iv (struct loop *loop, struct tree_niter_desc *niter) | |
3047 | { | |
3048 | struct control_iv *iv; | |
3049 | ||
3050 | if (!niter->control.base || !niter->control.step) | |
3051 | return; | |
3052 | ||
3053 | if (!integer_onep (niter->assumptions) || !niter->control.no_overflow) | |
3054 | return; | |
3055 | ||
3056 | iv = ggc_alloc<control_iv> (); | |
3057 | iv->base = niter->control.base; | |
3058 | iv->step = niter->control.step; | |
3059 | iv->next = loop->control_ivs; | |
3060 | loop->control_ivs = iv; | |
3061 | ||
3062 | return; | |
3063 | } | |
3064 | ||
cf8d19de BC |
3065 | /* This function returns TRUE if below conditions are satisfied: |
3066 | 1) VAR is SSA variable. | |
3067 | 2) VAR is an IV:{base, step} in its defining loop. | |
3068 | 3) IV doesn't overflow. | |
3069 | 4) Both base and step are integer constants. | |
3070 | 5) Base is the MIN/MAX value depends on IS_MIN. | |
3071 | Store value of base to INIT correspondingly. */ | |
3072 | ||
3073 | static bool | |
3074 | get_cst_init_from_scev (tree var, wide_int *init, bool is_min) | |
3075 | { | |
3076 | if (TREE_CODE (var) != SSA_NAME) | |
3077 | return false; | |
3078 | ||
3079 | gimple *def_stmt = SSA_NAME_DEF_STMT (var); | |
3080 | struct loop *loop = loop_containing_stmt (def_stmt); | |
3081 | ||
3082 | if (loop == NULL) | |
3083 | return false; | |
3084 | ||
3085 | affine_iv iv; | |
3086 | if (!simple_iv (loop, loop, var, &iv, false)) | |
3087 | return false; | |
3088 | ||
3089 | if (!iv.no_overflow) | |
3090 | return false; | |
3091 | ||
3092 | if (TREE_CODE (iv.base) != INTEGER_CST || TREE_CODE (iv.step) != INTEGER_CST) | |
3093 | return false; | |
3094 | ||
3095 | if (is_min == tree_int_cst_sign_bit (iv.step)) | |
3096 | return false; | |
3097 | ||
3098 | *init = iv.base; | |
3099 | return true; | |
3100 | } | |
3101 | ||
946e1bc7 ZD |
3102 | /* Record the estimate on number of iterations of LOOP based on the fact that |
3103 | the induction variable BASE + STEP * i evaluated in STMT does not wrap and | |
9bdb685e ZD |
3104 | its values belong to the range <LOW, HIGH>. REALISTIC is true if the |
3105 | estimated number of iterations is expected to be close to the real one. | |
3106 | UPPER is true if we are sure the induction variable does not wrap. */ | |
946e1bc7 ZD |
3107 | |
3108 | static void | |
355fe088 | 3109 | record_nonwrapping_iv (struct loop *loop, tree base, tree step, gimple *stmt, |
9bdb685e | 3110 | tree low, tree high, bool realistic, bool upper) |
946e1bc7 ZD |
3111 | { |
3112 | tree niter_bound, extreme, delta; | |
3113 | tree type = TREE_TYPE (base), unsigned_type; | |
fa8e5051 | 3114 | tree orig_base = base; |
946e1bc7 | 3115 | |
6e682d7e | 3116 | if (TREE_CODE (step) != INTEGER_CST || integer_zerop (step)) |
946e1bc7 ZD |
3117 | return; |
3118 | ||
3119 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
3120 | { | |
3121 | fprintf (dump_file, "Induction variable ("); | |
3122 | print_generic_expr (dump_file, TREE_TYPE (base), TDF_SLIM); | |
3123 | fprintf (dump_file, ") "); | |
3124 | print_generic_expr (dump_file, base, TDF_SLIM); | |
3125 | fprintf (dump_file, " + "); | |
3126 | print_generic_expr (dump_file, step, TDF_SLIM); | |
3127 | fprintf (dump_file, " * iteration does not wrap in statement "); | |
726a989a | 3128 | print_gimple_stmt (dump_file, stmt, 0, TDF_SLIM); |
946e1bc7 ZD |
3129 | fprintf (dump_file, " in loop %d.\n", loop->num); |
3130 | } | |
3131 | ||
3132 | unsigned_type = unsigned_type_for (type); | |
3133 | base = fold_convert (unsigned_type, base); | |
3134 | step = fold_convert (unsigned_type, step); | |
3135 | ||
3136 | if (tree_int_cst_sign_bit (step)) | |
3137 | { | |
fa8e5051 | 3138 | wide_int min, max; |
946e1bc7 | 3139 | extreme = fold_convert (unsigned_type, low); |
fa8e5051 IE |
3140 | if (TREE_CODE (orig_base) == SSA_NAME |
3141 | && TREE_CODE (high) == INTEGER_CST | |
3142 | && INTEGRAL_TYPE_P (TREE_TYPE (orig_base)) | |
cf8d19de BC |
3143 | && (get_range_info (orig_base, &min, &max) == VR_RANGE |
3144 | || get_cst_init_from_scev (orig_base, &max, false)) | |
fa8e5051 IE |
3145 | && wi::gts_p (high, max)) |
3146 | base = wide_int_to_tree (unsigned_type, max); | |
e53d562a RB |
3147 | else if (TREE_CODE (base) != INTEGER_CST |
3148 | && dominated_by_p (CDI_DOMINATORS, | |
3149 | loop->latch, gimple_bb (stmt))) | |
946e1bc7 ZD |
3150 | base = fold_convert (unsigned_type, high); |
3151 | delta = fold_build2 (MINUS_EXPR, unsigned_type, base, extreme); | |
3152 | step = fold_build1 (NEGATE_EXPR, unsigned_type, step); | |
3153 | } | |
3154 | else | |
3155 | { | |
fa8e5051 | 3156 | wide_int min, max; |
946e1bc7 | 3157 | extreme = fold_convert (unsigned_type, high); |
fa8e5051 IE |
3158 | if (TREE_CODE (orig_base) == SSA_NAME |
3159 | && TREE_CODE (low) == INTEGER_CST | |
3160 | && INTEGRAL_TYPE_P (TREE_TYPE (orig_base)) | |
cf8d19de BC |
3161 | && (get_range_info (orig_base, &min, &max) == VR_RANGE |
3162 | || get_cst_init_from_scev (orig_base, &min, true)) | |
fa8e5051 IE |
3163 | && wi::gts_p (min, low)) |
3164 | base = wide_int_to_tree (unsigned_type, min); | |
e53d562a RB |
3165 | else if (TREE_CODE (base) != INTEGER_CST |
3166 | && dominated_by_p (CDI_DOMINATORS, | |
3167 | loop->latch, gimple_bb (stmt))) | |
946e1bc7 ZD |
3168 | base = fold_convert (unsigned_type, low); |
3169 | delta = fold_build2 (MINUS_EXPR, unsigned_type, extreme, base); | |
3170 | } | |
3171 | ||
3172 | /* STMT is executed at most NITER_BOUND + 1 times, since otherwise the value | |
3173 | would get out of the range. */ | |
3174 | niter_bound = fold_build2 (FLOOR_DIV_EXPR, unsigned_type, delta, step); | |
807e902e | 3175 | widest_int max = derive_constant_upper_bound (niter_bound); |
9bdb685e | 3176 | record_estimate (loop, niter_bound, max, stmt, false, realistic, upper); |
4839cb59 ZD |
3177 | } |
3178 | ||
946e1bc7 | 3179 | /* Determine information about number of iterations a LOOP from the index |
ac84e05e ZD |
3180 | IDX of a data reference accessed in STMT. RELIABLE is true if STMT is |
3181 | guaranteed to be executed in every iteration of LOOP. Callback for | |
3182 | for_each_index. */ | |
946e1bc7 ZD |
3183 | |
3184 | struct ilb_data | |
3185 | { | |
3186 | struct loop *loop; | |
355fe088 | 3187 | gimple *stmt; |
946e1bc7 ZD |
3188 | }; |
3189 | ||
3190 | static bool | |
3191 | idx_infer_loop_bounds (tree base, tree *idx, void *dta) | |
3192 | { | |
c22940cd | 3193 | struct ilb_data *data = (struct ilb_data *) dta; |
946e1bc7 ZD |
3194 | tree ev, init, step; |
3195 | tree low, high, type, next; | |
cd0f6278 | 3196 | bool sign, upper = true, at_end = false; |
946e1bc7 ZD |
3197 | struct loop *loop = data->loop; |
3198 | ||
9bdb685e | 3199 | if (TREE_CODE (base) != ARRAY_REF) |
946e1bc7 ZD |
3200 | return true; |
3201 | ||
9bdb685e ZD |
3202 | /* For arrays at the end of the structure, we are not guaranteed that they |
3203 | do not really extend over their declared size. However, for arrays of | |
3204 | size greater than one, this is unlikely to be intended. */ | |
3205 | if (array_at_struct_end_p (base)) | |
ac84e05e ZD |
3206 | { |
3207 | at_end = true; | |
3208 | upper = false; | |
3209 | } | |
9bdb685e | 3210 | |
8b679c9b RB |
3211 | struct loop *dloop = loop_containing_stmt (data->stmt); |
3212 | if (!dloop) | |
3213 | return true; | |
3214 | ||
3215 | ev = analyze_scalar_evolution (dloop, *idx); | |
3216 | ev = instantiate_parameters (loop, ev); | |
946e1bc7 ZD |
3217 | init = initial_condition (ev); |
3218 | step = evolution_part_in_loop_num (ev, loop->num); | |
3219 | ||
3220 | if (!init | |
3221 | || !step | |
3222 | || TREE_CODE (step) != INTEGER_CST | |
6e682d7e | 3223 | || integer_zerop (step) |
946e1bc7 ZD |
3224 | || tree_contains_chrecs (init, NULL) |
3225 | || chrec_contains_symbols_defined_in_loop (init, loop->num)) | |
3226 | return true; | |
3227 | ||
3228 | low = array_ref_low_bound (base); | |
3229 | high = array_ref_up_bound (base); | |
b8698a0f | 3230 | |
946e1bc7 ZD |
3231 | /* The case of nonconstant bounds could be handled, but it would be |
3232 | complicated. */ | |
3233 | if (TREE_CODE (low) != INTEGER_CST | |
3234 | || !high | |
3235 | || TREE_CODE (high) != INTEGER_CST) | |
3236 | return true; | |
3237 | sign = tree_int_cst_sign_bit (step); | |
3238 | type = TREE_TYPE (step); | |
9bdb685e ZD |
3239 | |
3240 | /* The array of length 1 at the end of a structure most likely extends | |
3241 | beyond its bounds. */ | |
ac84e05e | 3242 | if (at_end |
9bdb685e ZD |
3243 | && operand_equal_p (low, high, 0)) |
3244 | return true; | |
3245 | ||
946e1bc7 ZD |
3246 | /* In case the relevant bound of the array does not fit in type, or |
3247 | it does, but bound + step (in type) still belongs into the range of the | |
3248 | array, the index may wrap and still stay within the range of the array | |
3249 | (consider e.g. if the array is indexed by the full range of | |
3250 | unsigned char). | |
3251 | ||
3252 | To make things simpler, we require both bounds to fit into type, although | |
2f8e468b | 3253 | there are cases where this would not be strictly necessary. */ |
946e1bc7 ZD |
3254 | if (!int_fits_type_p (high, type) |
3255 | || !int_fits_type_p (low, type)) | |
3256 | return true; | |
3257 | low = fold_convert (type, low); | |
3258 | high = fold_convert (type, high); | |
3259 | ||
3260 | if (sign) | |
3261 | next = fold_binary (PLUS_EXPR, type, low, step); | |
3262 | else | |
3263 | next = fold_binary (PLUS_EXPR, type, high, step); | |
b8698a0f | 3264 | |
946e1bc7 ZD |
3265 | if (tree_int_cst_compare (low, next) <= 0 |
3266 | && tree_int_cst_compare (next, high) <= 0) | |
3267 | return true; | |
3268 | ||
77c9d5b4 JH |
3269 | /* If access is not executed on every iteration, we must ensure that overlow |
3270 | may not make the access valid later. */ | |
870ca331 | 3271 | if (!dominated_by_p (CDI_DOMINATORS, loop->latch, gimple_bb (data->stmt)) |
b24d9420 BC |
3272 | && scev_probably_wraps_p (NULL_TREE, |
3273 | initial_condition_in_loop_num (ev, loop->num), | |
870ca331 | 3274 | step, data->stmt, loop, true)) |
77c9d5b4 | 3275 | upper = false; |
870ca331 | 3276 | |
77c9d5b4 | 3277 | record_nonwrapping_iv (loop, init, step, data->stmt, low, high, false, upper); |
946e1bc7 ZD |
3278 | return true; |
3279 | } | |
3280 | ||
3281 | /* Determine information about number of iterations a LOOP from the bounds | |
ac84e05e ZD |
3282 | of arrays in the data reference REF accessed in STMT. RELIABLE is true if |
3283 | STMT is guaranteed to be executed in every iteration of LOOP.*/ | |
946e1bc7 ZD |
3284 | |
3285 | static void | |
355fe088 | 3286 | infer_loop_bounds_from_ref (struct loop *loop, gimple *stmt, tree ref) |
946e1bc7 ZD |
3287 | { |
3288 | struct ilb_data data; | |
3289 | ||
3290 | data.loop = loop; | |
3291 | data.stmt = stmt; | |
3292 | for_each_index (&ref, idx_infer_loop_bounds, &data); | |
3293 | } | |
3294 | ||
3295 | /* Determine information about number of iterations of a LOOP from the way | |
ac84e05e ZD |
3296 | arrays are used in STMT. RELIABLE is true if STMT is guaranteed to be |
3297 | executed in every iteration of LOOP. */ | |
946e1bc7 ZD |
3298 | |
3299 | static void | |
355fe088 | 3300 | infer_loop_bounds_from_array (struct loop *loop, gimple *stmt) |
946e1bc7 | 3301 | { |
726a989a | 3302 | if (is_gimple_assign (stmt)) |
946e1bc7 | 3303 | { |
726a989a RB |
3304 | tree op0 = gimple_assign_lhs (stmt); |
3305 | tree op1 = gimple_assign_rhs1 (stmt); | |
946e1bc7 ZD |
3306 | |
3307 | /* For each memory access, analyze its access function | |
3308 | and record a bound on the loop iteration domain. */ | |
3309 | if (REFERENCE_CLASS_P (op0)) | |
cd0f6278 | 3310 | infer_loop_bounds_from_ref (loop, stmt, op0); |
946e1bc7 ZD |
3311 | |
3312 | if (REFERENCE_CLASS_P (op1)) | |
cd0f6278 | 3313 | infer_loop_bounds_from_ref (loop, stmt, op1); |
946e1bc7 | 3314 | } |
726a989a | 3315 | else if (is_gimple_call (stmt)) |
946e1bc7 | 3316 | { |
726a989a RB |
3317 | tree arg, lhs; |
3318 | unsigned i, n = gimple_call_num_args (stmt); | |
946e1bc7 | 3319 | |
726a989a RB |
3320 | lhs = gimple_call_lhs (stmt); |
3321 | if (lhs && REFERENCE_CLASS_P (lhs)) | |
cd0f6278 | 3322 | infer_loop_bounds_from_ref (loop, stmt, lhs); |
726a989a RB |
3323 | |
3324 | for (i = 0; i < n; i++) | |
3325 | { | |
3326 | arg = gimple_call_arg (stmt, i); | |
3327 | if (REFERENCE_CLASS_P (arg)) | |
cd0f6278 | 3328 | infer_loop_bounds_from_ref (loop, stmt, arg); |
726a989a | 3329 | } |
946e1bc7 ZD |
3330 | } |
3331 | } | |
3332 | ||
bc69f7ff TV |
3333 | /* Determine information about number of iterations of a LOOP from the fact |
3334 | that pointer arithmetics in STMT does not overflow. */ | |
3335 | ||
3336 | static void | |
355fe088 | 3337 | infer_loop_bounds_from_pointer_arith (struct loop *loop, gimple *stmt) |
bc69f7ff TV |
3338 | { |
3339 | tree def, base, step, scev, type, low, high; | |
3340 | tree var, ptr; | |
3341 | ||
3342 | if (!is_gimple_assign (stmt) | |
3343 | || gimple_assign_rhs_code (stmt) != POINTER_PLUS_EXPR) | |
3344 | return; | |
3345 | ||
3346 | def = gimple_assign_lhs (stmt); | |
3347 | if (TREE_CODE (def) != SSA_NAME) | |
3348 | return; | |
3349 | ||
3350 | type = TREE_TYPE (def); | |
3351 | if (!nowrap_type_p (type)) | |
3352 | return; | |
3353 | ||
3354 | ptr = gimple_assign_rhs1 (stmt); | |
3355 | if (!expr_invariant_in_loop_p (loop, ptr)) | |
3356 | return; | |
3357 | ||
3358 | var = gimple_assign_rhs2 (stmt); | |
3359 | if (TYPE_PRECISION (type) != TYPE_PRECISION (TREE_TYPE (var))) | |
3360 | return; | |
3361 | ||
3362 | scev = instantiate_parameters (loop, analyze_scalar_evolution (loop, def)); | |
3363 | if (chrec_contains_undetermined (scev)) | |
3364 | return; | |
3365 | ||
3366 | base = initial_condition_in_loop_num (scev, loop->num); | |
3367 | step = evolution_part_in_loop_num (scev, loop->num); | |
3368 | ||
3369 | if (!base || !step | |
3370 | || TREE_CODE (step) != INTEGER_CST | |
3371 | || tree_contains_chrecs (base, NULL) | |
3372 | || chrec_contains_symbols_defined_in_loop (base, loop->num)) | |
3373 | return; | |
3374 | ||
3375 | low = lower_bound_in_type (type, type); | |
3376 | high = upper_bound_in_type (type, type); | |
3377 | ||
0703f020 TV |
3378 | /* In C, pointer arithmetic p + 1 cannot use a NULL pointer, and p - 1 cannot |
3379 | produce a NULL pointer. The contrary would mean NULL points to an object, | |
3380 | while NULL is supposed to compare unequal with the address of all objects. | |
3381 | Furthermore, p + 1 cannot produce a NULL pointer and p - 1 cannot use a | |
3382 | NULL pointer since that would mean wrapping, which we assume here not to | |
3383 | happen. So, we can exclude NULL from the valid range of pointer | |
3384 | arithmetic. */ | |
3385 | if (flag_delete_null_pointer_checks && int_cst_value (low) == 0) | |
3386 | low = build_int_cstu (TREE_TYPE (low), TYPE_ALIGN_UNIT (TREE_TYPE (type))); | |
3387 | ||
bc69f7ff TV |
3388 | record_nonwrapping_iv (loop, base, step, stmt, low, high, false, true); |
3389 | } | |
3390 | ||
946e1bc7 ZD |
3391 | /* Determine information about number of iterations of a LOOP from the fact |
3392 | that signed arithmetics in STMT does not overflow. */ | |
3393 | ||
3394 | static void | |
355fe088 | 3395 | infer_loop_bounds_from_signedness (struct loop *loop, gimple *stmt) |
946e1bc7 ZD |
3396 | { |
3397 | tree def, base, step, scev, type, low, high; | |
3398 | ||
726a989a | 3399 | if (gimple_code (stmt) != GIMPLE_ASSIGN) |
946e1bc7 ZD |
3400 | return; |
3401 | ||
726a989a | 3402 | def = gimple_assign_lhs (stmt); |
946e1bc7 ZD |
3403 | |
3404 | if (TREE_CODE (def) != SSA_NAME) | |
3405 | return; | |
3406 | ||
3407 | type = TREE_TYPE (def); | |
3408 | if (!INTEGRAL_TYPE_P (type) | |
eeef0e45 | 3409 | || !TYPE_OVERFLOW_UNDEFINED (type)) |
946e1bc7 ZD |
3410 | return; |
3411 | ||
3412 | scev = instantiate_parameters (loop, analyze_scalar_evolution (loop, def)); | |
3413 | if (chrec_contains_undetermined (scev)) | |
3414 | return; | |
3415 | ||
3416 | base = initial_condition_in_loop_num (scev, loop->num); | |
3417 | step = evolution_part_in_loop_num (scev, loop->num); | |
3418 | ||
3419 | if (!base || !step | |
3420 | || TREE_CODE (step) != INTEGER_CST | |
3421 | || tree_contains_chrecs (base, NULL) | |
3422 | || chrec_contains_symbols_defined_in_loop (base, loop->num)) | |
3423 | return; | |
3424 | ||
3425 | low = lower_bound_in_type (type, type); | |
3426 | high = upper_bound_in_type (type, type); | |
3427 | ||
9bdb685e | 3428 | record_nonwrapping_iv (loop, base, step, stmt, low, high, false, true); |
946e1bc7 ZD |
3429 | } |
3430 | ||
d7770457 SP |
3431 | /* The following analyzers are extracting informations on the bounds |
3432 | of LOOP from the following undefined behaviors: | |
3433 | ||
3434 | - data references should not access elements over the statically | |
3435 | allocated size, | |
3436 | ||
3437 | - signed variables should not overflow when flag_wrapv is not set. | |
3438 | */ | |
3439 | ||
3440 | static void | |
3441 | infer_loop_bounds_from_undefined (struct loop *loop) | |
3442 | { | |
3443 | unsigned i; | |
946e1bc7 | 3444 | basic_block *bbs; |
726a989a | 3445 | gimple_stmt_iterator bsi; |
946e1bc7 | 3446 | basic_block bb; |
ac84e05e | 3447 | bool reliable; |
b8698a0f | 3448 | |
d7770457 SP |
3449 | bbs = get_loop_body (loop); |
3450 | ||
3451 | for (i = 0; i < loop->num_nodes; i++) | |
3452 | { | |
3453 | bb = bbs[i]; | |
3454 | ||
946e1bc7 | 3455 | /* If BB is not executed in each iteration of the loop, we cannot |
ac84e05e | 3456 | use the operations in it to infer reliable upper bound on the |
cd0f6278 JH |
3457 | # of iterations of the loop. However, we can use it as a guess. |
3458 | Reliable guesses come only from array bounds. */ | |
ac84e05e | 3459 | reliable = dominated_by_p (CDI_DOMINATORS, loop->latch, bb); |
946e1bc7 | 3460 | |
726a989a | 3461 | for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi)) |
946e1bc7 | 3462 | { |
355fe088 | 3463 | gimple *stmt = gsi_stmt (bsi); |
d7770457 | 3464 | |
cd0f6278 | 3465 | infer_loop_bounds_from_array (loop, stmt); |
ac84e05e ZD |
3466 | |
3467 | if (reliable) | |
bc69f7ff TV |
3468 | { |
3469 | infer_loop_bounds_from_signedness (loop, stmt); | |
3470 | infer_loop_bounds_from_pointer_arith (loop, stmt); | |
3471 | } | |
946e1bc7 ZD |
3472 | } |
3473 | ||
d7770457 SP |
3474 | } |
3475 | ||
3476 | free (bbs); | |
3477 | } | |
3478 | ||
807e902e | 3479 | /* Compare wide ints, callback for qsort. */ |
73ddf95b | 3480 | |
71343877 | 3481 | static int |
807e902e | 3482 | wide_int_cmp (const void *p1, const void *p2) |
73ddf95b | 3483 | { |
807e902e KZ |
3484 | const widest_int *d1 = (const widest_int *) p1; |
3485 | const widest_int *d2 = (const widest_int *) p2; | |
3486 | return wi::cmpu (*d1, *d2); | |
73ddf95b JH |
3487 | } |
3488 | ||
3489 | /* Return index of BOUND in BOUNDS array sorted in increasing order. | |
3490 | Lookup by binary search. */ | |
3491 | ||
71343877 | 3492 | static int |
807e902e | 3493 | bound_index (vec<widest_int> bounds, const widest_int &bound) |
73ddf95b | 3494 | { |
9771b263 | 3495 | unsigned int end = bounds.length (); |
73ddf95b JH |
3496 | unsigned int begin = 0; |
3497 | ||
3498 | /* Find a matching index by means of a binary search. */ | |
3499 | while (begin != end) | |
3500 | { | |
3501 | unsigned int middle = (begin + end) / 2; | |
807e902e | 3502 | widest_int index = bounds[middle]; |
73ddf95b JH |
3503 | |
3504 | if (index == bound) | |
3505 | return middle; | |
807e902e | 3506 | else if (wi::ltu_p (index, bound)) |
73ddf95b JH |
3507 | begin = middle + 1; |
3508 | else | |
3509 | end = middle; | |
3510 | } | |
3511 | gcc_unreachable (); | |
3512 | } | |
3513 | ||
73ddf95b JH |
3514 | /* We recorded loop bounds only for statements dominating loop latch (and thus |
3515 | executed each loop iteration). If there are any bounds on statements not | |
3516 | dominating the loop latch we can improve the estimate by walking the loop | |
3517 | body and seeing if every path from loop header to loop latch contains | |
3518 | some bounded statement. */ | |
3519 | ||
3520 | static void | |
3521 | discover_iteration_bound_by_body_walk (struct loop *loop) | |
3522 | { | |
73ddf95b | 3523 | struct nb_iter_bound *elt; |
8c681247 | 3524 | auto_vec<widest_int> bounds; |
b4f9786b JJ |
3525 | vec<vec<basic_block> > queues = vNULL; |
3526 | vec<basic_block> queue = vNULL; | |
73ddf95b JH |
3527 | ptrdiff_t queue_index; |
3528 | ptrdiff_t latch_index = 0; | |
73ddf95b JH |
3529 | |
3530 | /* Discover what bounds may interest us. */ | |
3531 | for (elt = loop->bounds; elt; elt = elt->next) | |
3532 | { | |
807e902e | 3533 | widest_int bound = elt->bound; |
73ddf95b JH |
3534 | |
3535 | /* Exit terminates loop at given iteration, while non-exits produce undefined | |
3536 | effect on the next iteration. */ | |
3537 | if (!elt->is_exit) | |
4c052539 | 3538 | { |
807e902e | 3539 | bound += 1; |
4c052539 | 3540 | /* If an overflow occurred, ignore the result. */ |
807e902e | 3541 | if (bound == 0) |
4c052539 JJ |
3542 | continue; |
3543 | } | |
73ddf95b JH |
3544 | |
3545 | if (!loop->any_upper_bound | |
807e902e | 3546 | || wi::ltu_p (bound, loop->nb_iterations_upper_bound)) |
9771b263 | 3547 | bounds.safe_push (bound); |
73ddf95b JH |
3548 | } |
3549 | ||
3550 | /* Exit early if there is nothing to do. */ | |
9771b263 | 3551 | if (!bounds.exists ()) |
73ddf95b JH |
3552 | return; |
3553 | ||
3554 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
3555 | fprintf (dump_file, " Trying to walk loop body to reduce the bound.\n"); | |
3556 | ||
3557 | /* Sort the bounds in decreasing order. */ | |
75509ba2 | 3558 | bounds.qsort (wide_int_cmp); |
73ddf95b JH |
3559 | |
3560 | /* For every basic block record the lowest bound that is guaranteed to | |
3561 | terminate the loop. */ | |
3562 | ||
39c8aaa4 | 3563 | hash_map<basic_block, ptrdiff_t> bb_bounds; |
73ddf95b JH |
3564 | for (elt = loop->bounds; elt; elt = elt->next) |
3565 | { | |
807e902e | 3566 | widest_int bound = elt->bound; |
73ddf95b | 3567 | if (!elt->is_exit) |
4c052539 | 3568 | { |
807e902e | 3569 | bound += 1; |
4c052539 | 3570 | /* If an overflow occurred, ignore the result. */ |
807e902e | 3571 | if (bound == 0) |
4c052539 JJ |
3572 | continue; |
3573 | } | |
73ddf95b JH |
3574 | |
3575 | if (!loop->any_upper_bound | |
807e902e | 3576 | || wi::ltu_p (bound, loop->nb_iterations_upper_bound)) |
73ddf95b JH |
3577 | { |
3578 | ptrdiff_t index = bound_index (bounds, bound); | |
39c8aaa4 | 3579 | ptrdiff_t *entry = bb_bounds.get (gimple_bb (elt->stmt)); |
73ddf95b | 3580 | if (!entry) |
39c8aaa4 | 3581 | bb_bounds.put (gimple_bb (elt->stmt), index); |
73ddf95b | 3582 | else if ((ptrdiff_t)*entry > index) |
39c8aaa4 | 3583 | *entry = index; |
73ddf95b JH |
3584 | } |
3585 | } | |
3586 | ||
39c8aaa4 | 3587 | hash_map<basic_block, ptrdiff_t> block_priority; |
73ddf95b JH |
3588 | |
3589 | /* Perform shortest path discovery loop->header ... loop->latch. | |
3590 | ||
3591 | The "distance" is given by the smallest loop bound of basic block | |
3592 | present in the path and we look for path with largest smallest bound | |
3593 | on it. | |
3594 | ||
b4f9786b | 3595 | To avoid the need for fibonacci heap on double ints we simply compress |
73ddf95b JH |
3596 | double ints into indexes to BOUNDS array and then represent the queue |
3597 | as arrays of queues for every index. | |
9771b263 | 3598 | Index of BOUNDS.length() means that the execution of given BB has |
73ddf95b JH |
3599 | no bounds determined. |
3600 | ||
3601 | VISITED is a pointer map translating basic block into smallest index | |
3602 | it was inserted into the priority queue with. */ | |
3603 | latch_index = -1; | |
3604 | ||
3605 | /* Start walk in loop header with index set to infinite bound. */ | |
9771b263 DN |
3606 | queue_index = bounds.length (); |
3607 | queues.safe_grow_cleared (queue_index + 1); | |
3608 | queue.safe_push (loop->header); | |
3609 | queues[queue_index] = queue; | |
39c8aaa4 | 3610 | block_priority.put (loop->header, queue_index); |
73ddf95b JH |
3611 | |
3612 | for (; queue_index >= 0; queue_index--) | |
3613 | { | |
3614 | if (latch_index < queue_index) | |
3615 | { | |
9771b263 | 3616 | while (queues[queue_index].length ()) |
73ddf95b JH |
3617 | { |
3618 | basic_block bb; | |
3619 | ptrdiff_t bound_index = queue_index; | |
73ddf95b JH |
3620 | edge e; |
3621 | edge_iterator ei; | |
3622 | ||
9771b263 DN |
3623 | queue = queues[queue_index]; |
3624 | bb = queue.pop (); | |
73ddf95b JH |
3625 | |
3626 | /* OK, we later inserted the BB with lower priority, skip it. */ | |
39c8aaa4 | 3627 | if (*block_priority.get (bb) > queue_index) |
73ddf95b JH |
3628 | continue; |
3629 | ||
3630 | /* See if we can improve the bound. */ | |
39c8aaa4 TS |
3631 | ptrdiff_t *entry = bb_bounds.get (bb); |
3632 | if (entry && *entry < bound_index) | |
3633 | bound_index = *entry; | |
73ddf95b JH |
3634 | |
3635 | /* Insert succesors into the queue, watch for latch edge | |
3636 | and record greatest index we saw. */ | |
3637 | FOR_EACH_EDGE (e, ei, bb->succs) | |
3638 | { | |
3639 | bool insert = false; | |
73ddf95b JH |
3640 | |
3641 | if (loop_exit_edge_p (loop, e)) | |
3642 | continue; | |
3643 | ||
3644 | if (e == loop_latch_edge (loop) | |
3645 | && latch_index < bound_index) | |
3646 | latch_index = bound_index; | |
39c8aaa4 | 3647 | else if (!(entry = block_priority.get (e->dest))) |
73ddf95b JH |
3648 | { |
3649 | insert = true; | |
39c8aaa4 | 3650 | block_priority.put (e->dest, bound_index); |
73ddf95b | 3651 | } |
39c8aaa4 | 3652 | else if (*entry < bound_index) |
73ddf95b JH |
3653 | { |
3654 | insert = true; | |
39c8aaa4 | 3655 | *entry = bound_index; |
73ddf95b JH |
3656 | } |
3657 | ||
3658 | if (insert) | |
b4f9786b | 3659 | queues[bound_index].safe_push (e->dest); |
73ddf95b JH |
3660 | } |
3661 | } | |
3662 | } | |
b4f9786b | 3663 | queues[queue_index].release (); |
73ddf95b JH |
3664 | } |
3665 | ||
3666 | gcc_assert (latch_index >= 0); | |
9771b263 | 3667 | if ((unsigned)latch_index < bounds.length ()) |
73ddf95b JH |
3668 | { |
3669 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
3670 | { | |
3671 | fprintf (dump_file, "Found better loop bound "); | |
807e902e | 3672 | print_decu (bounds[latch_index], dump_file); |
73ddf95b JH |
3673 | fprintf (dump_file, "\n"); |
3674 | } | |
9771b263 | 3675 | record_niter_bound (loop, bounds[latch_index], false, true); |
73ddf95b JH |
3676 | } |
3677 | ||
9771b263 | 3678 | queues.release (); |
73ddf95b JH |
3679 | } |
3680 | ||
05322355 JH |
3681 | /* See if every path cross the loop goes through a statement that is known |
3682 | to not execute at the last iteration. In that case we can decrese iteration | |
3683 | count by 1. */ | |
3684 | ||
3685 | static void | |
3686 | maybe_lower_iteration_bound (struct loop *loop) | |
3687 | { | |
355fe088 | 3688 | hash_set<gimple *> *not_executed_last_iteration = NULL; |
05322355 JH |
3689 | struct nb_iter_bound *elt; |
3690 | bool found_exit = false; | |
8c681247 | 3691 | auto_vec<basic_block> queue; |
05322355 JH |
3692 | bitmap visited; |
3693 | ||
3694 | /* Collect all statements with interesting (i.e. lower than | |
3695 | nb_iterations_upper_bound) bound on them. | |
3696 | ||
3697 | TODO: Due to the way record_estimate choose estimates to store, the bounds | |
3698 | will be always nb_iterations_upper_bound-1. We can change this to record | |
3699 | also statements not dominating the loop latch and update the walk bellow | |
5764ee3c | 3700 | to the shortest path algorithm. */ |
05322355 JH |
3701 | for (elt = loop->bounds; elt; elt = elt->next) |
3702 | { | |
3703 | if (!elt->is_exit | |
807e902e | 3704 | && wi::ltu_p (elt->bound, loop->nb_iterations_upper_bound)) |
05322355 JH |
3705 | { |
3706 | if (!not_executed_last_iteration) | |
355fe088 | 3707 | not_executed_last_iteration = new hash_set<gimple *>; |
6e2830c3 | 3708 | not_executed_last_iteration->add (elt->stmt); |
05322355 JH |
3709 | } |
3710 | } | |
3711 | if (!not_executed_last_iteration) | |
3712 | return; | |
3713 | ||
3714 | /* Start DFS walk in the loop header and see if we can reach the | |
3715 | loop latch or any of the exits (including statements with side | |
3716 | effects that may terminate the loop otherwise) without visiting | |
3717 | any of the statements known to have undefined effect on the last | |
3718 | iteration. */ | |
9771b263 | 3719 | queue.safe_push (loop->header); |
05322355 JH |
3720 | visited = BITMAP_ALLOC (NULL); |
3721 | bitmap_set_bit (visited, loop->header->index); | |
3722 | found_exit = false; | |
3723 | ||
3724 | do | |
3725 | { | |
9771b263 | 3726 | basic_block bb = queue.pop (); |
05322355 JH |
3727 | gimple_stmt_iterator gsi; |
3728 | bool stmt_found = false; | |
3729 | ||
3730 | /* Loop for possible exits and statements bounding the execution. */ | |
3731 | for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi)) | |
3732 | { | |
355fe088 | 3733 | gimple *stmt = gsi_stmt (gsi); |
6e2830c3 | 3734 | if (not_executed_last_iteration->contains (stmt)) |
05322355 JH |
3735 | { |
3736 | stmt_found = true; | |
3737 | break; | |
3738 | } | |
3739 | if (gimple_has_side_effects (stmt)) | |
3740 | { | |
3741 | found_exit = true; | |
3742 | break; | |
3743 | } | |
3744 | } | |
3745 | if (found_exit) | |
3746 | break; | |
3747 | ||
3748 | /* If no bounding statement is found, continue the walk. */ | |
3749 | if (!stmt_found) | |
3750 | { | |
3751 | edge e; | |
3752 | edge_iterator ei; | |
3753 | ||
3754 | FOR_EACH_EDGE (e, ei, bb->succs) | |
3755 | { | |
3756 | if (loop_exit_edge_p (loop, e) | |
3757 | || e == loop_latch_edge (loop)) | |
3758 | { | |
3759 | found_exit = true; | |
3760 | break; | |
3761 | } | |
3762 | if (bitmap_set_bit (visited, e->dest->index)) | |
9771b263 | 3763 | queue.safe_push (e->dest); |
05322355 JH |
3764 | } |
3765 | } | |
3766 | } | |
9771b263 | 3767 | while (queue.length () && !found_exit); |
05322355 JH |
3768 | |
3769 | /* If every path through the loop reach bounding statement before exit, | |
3770 | then we know the last iteration of the loop will have undefined effect | |
3771 | and we can decrease number of iterations. */ | |
3772 | ||
3773 | if (!found_exit) | |
3774 | { | |
3775 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
3776 | fprintf (dump_file, "Reducing loop iteration estimate by 1; " | |
3777 | "undefined statement must be executed at the last iteration.\n"); | |
807e902e | 3778 | record_niter_bound (loop, loop->nb_iterations_upper_bound - 1, |
05322355 JH |
3779 | false, true); |
3780 | } | |
48067724 | 3781 | |
05322355 | 3782 | BITMAP_FREE (visited); |
6e2830c3 | 3783 | delete not_executed_last_iteration; |
05322355 JH |
3784 | } |
3785 | ||
e3488283 RG |
3786 | /* Records estimates on numbers of iterations of LOOP. If USE_UNDEFINED_P |
3787 | is true also use estimates derived from undefined behavior. */ | |
e9eb809d | 3788 | |
adb7eaa2 RB |
3789 | void |
3790 | estimate_numbers_of_iterations (struct loop *loop) | |
e9eb809d | 3791 | { |
9771b263 | 3792 | vec<edge> exits; |
e9eb809d | 3793 | tree niter, type; |
ca83d385 | 3794 | unsigned i; |
e9eb809d | 3795 | struct tree_niter_desc niter_desc; |
ca83d385 | 3796 | edge ex; |
807e902e | 3797 | widest_int bound; |
f9bf4777 | 3798 | edge likely_exit; |
e9eb809d | 3799 | |
79ebd55c | 3800 | /* Give up if we already have tried to compute an estimation. */ |
946e1bc7 | 3801 | if (loop->estimate_state != EST_NOT_COMPUTED) |
79ebd55c | 3802 | return; |
03fd03d5 | 3803 | |
9bdb685e | 3804 | loop->estimate_state = EST_AVAILABLE; |
aade5c72 JH |
3805 | |
3806 | /* If we have a measured profile, use it to estimate the number of | |
3807 | iterations. Normally this is recorded by branch_prob right after | |
3808 | reading the profile. In case we however found a new loop, record the | |
3809 | information here. | |
3810 | ||
3811 | Explicitly check for profile status so we do not report | |
3812 | wrong prediction hitrates for guessed loop iterations heuristics. | |
3813 | Do not recompute already recorded bounds - we ought to be better on | |
3814 | updating iteration bounds than updating profile in general and thus | |
3815 | recomputing iteration bounds later in the compilation process will just | |
3816 | introduce random roundoff errors. */ | |
3817 | if (!loop->any_estimate | |
3995f3a2 | 3818 | && loop->header->count > 0) |
aade5c72 JH |
3819 | { |
3820 | gcov_type nit = expected_loop_iterations_unbounded (loop); | |
3821 | bound = gcov_type_to_wide_int (nit); | |
3822 | record_niter_bound (loop, bound, true, false); | |
3823 | } | |
79ebd55c | 3824 | |
fbd28bc3 JJ |
3825 | /* Ensure that loop->nb_iterations is computed if possible. If it turns out |
3826 | to be constant, we avoid undefined behavior implied bounds and instead | |
3827 | diagnose those loops with -Waggressive-loop-optimizations. */ | |
3828 | number_of_latch_executions (loop); | |
3829 | ||
ca83d385 | 3830 | exits = get_loop_exit_edges (loop); |
f9bf4777 | 3831 | likely_exit = single_likely_exit (loop); |
9771b263 | 3832 | FOR_EACH_VEC_ELT (exits, i, ex) |
e9eb809d | 3833 | { |
cd0f6278 | 3834 | if (!number_of_iterations_exit (loop, ex, &niter_desc, false, false)) |
e9eb809d ZD |
3835 | continue; |
3836 | ||
3837 | niter = niter_desc.niter; | |
3838 | type = TREE_TYPE (niter); | |
946e1bc7 | 3839 | if (TREE_CODE (niter_desc.may_be_zero) != INTEGER_CST) |
e6845c23 | 3840 | niter = build3 (COND_EXPR, type, niter_desc.may_be_zero, |
ff5e9a94 | 3841 | build_int_cst (type, 0), |
e6845c23 | 3842 | niter); |
b3ce5b6e | 3843 | record_estimate (loop, niter, niter_desc.max, |
ca83d385 | 3844 | last_stmt (ex->src), |
f9bf4777 | 3845 | true, ex == likely_exit, true); |
2f07b722 | 3846 | record_control_iv (loop, &niter_desc); |
e9eb809d | 3847 | } |
9771b263 | 3848 | exits.release (); |
b8698a0f | 3849 | |
6e616110 RB |
3850 | if (flag_aggressive_loop_optimizations) |
3851 | infer_loop_bounds_from_undefined (loop); | |
9bdb685e | 3852 | |
73ddf95b JH |
3853 | discover_iteration_bound_by_body_walk (loop); |
3854 | ||
05322355 JH |
3855 | maybe_lower_iteration_bound (loop); |
3856 | ||
fbd28bc3 JJ |
3857 | /* If we know the exact number of iterations of this loop, try to |
3858 | not break code with undefined behavior by not recording smaller | |
3859 | maximum number of iterations. */ | |
3860 | if (loop->nb_iterations | |
3861 | && TREE_CODE (loop->nb_iterations) == INTEGER_CST) | |
3862 | { | |
3863 | loop->any_upper_bound = true; | |
807e902e | 3864 | loop->nb_iterations_upper_bound = wi::to_widest (loop->nb_iterations); |
fbd28bc3 | 3865 | } |
e9eb809d ZD |
3866 | } |
3867 | ||
b4a9343c ZD |
3868 | /* Sets NIT to the estimated number of executions of the latch of the |
3869 | LOOP. If CONSERVATIVE is true, we must be sure that NIT is at least as | |
3870 | large as the number of iterations. If we have no reliable estimate, | |
3871 | the function returns false, otherwise returns true. */ | |
3872 | ||
3873 | bool | |
807e902e | 3874 | estimated_loop_iterations (struct loop *loop, widest_int *nit) |
b4a9343c | 3875 | { |
e3a8f1fa JH |
3876 | /* When SCEV information is available, try to update loop iterations |
3877 | estimate. Otherwise just return whatever we recorded earlier. */ | |
3878 | if (scev_initialized_p ()) | |
adb7eaa2 | 3879 | estimate_numbers_of_iterations (loop); |
e3a8f1fa | 3880 | |
71343877 | 3881 | return (get_estimated_loop_iterations (loop, nit)); |
652c4c71 | 3882 | } |
b4a9343c | 3883 | |
1ef88893 AM |
3884 | /* Similar to estimated_loop_iterations, but returns the estimate only |
3885 | if it fits to HOST_WIDE_INT. If this is not the case, or the estimate | |
3886 | on the number of iterations of LOOP could not be derived, returns -1. */ | |
3887 | ||
3888 | HOST_WIDE_INT | |
3889 | estimated_loop_iterations_int (struct loop *loop) | |
3890 | { | |
807e902e | 3891 | widest_int nit; |
1ef88893 AM |
3892 | HOST_WIDE_INT hwi_nit; |
3893 | ||
3894 | if (!estimated_loop_iterations (loop, &nit)) | |
3895 | return -1; | |
3896 | ||
807e902e | 3897 | if (!wi::fits_shwi_p (nit)) |
1ef88893 AM |
3898 | return -1; |
3899 | hwi_nit = nit.to_shwi (); | |
3900 | ||
3901 | return hwi_nit < 0 ? -1 : hwi_nit; | |
3902 | } | |
3903 | ||
3904 | ||
652c4c71 RG |
3905 | /* Sets NIT to an upper bound for the maximum number of executions of the |
3906 | latch of the LOOP. If we have no reliable estimate, the function returns | |
3907 | false, otherwise returns true. */ | |
3908 | ||
3909 | bool | |
807e902e | 3910 | max_loop_iterations (struct loop *loop, widest_int *nit) |
652c4c71 | 3911 | { |
e3a8f1fa JH |
3912 | /* When SCEV information is available, try to update loop iterations |
3913 | estimate. Otherwise just return whatever we recorded earlier. */ | |
3914 | if (scev_initialized_p ()) | |
adb7eaa2 | 3915 | estimate_numbers_of_iterations (loop); |
b4a9343c | 3916 | |
71343877 | 3917 | return get_max_loop_iterations (loop, nit); |
652c4c71 RG |
3918 | } |
3919 | ||
3920 | /* Similar to max_loop_iterations, but returns the estimate only | |
3921 | if it fits to HOST_WIDE_INT. If this is not the case, or the estimate | |
3922 | on the number of iterations of LOOP could not be derived, returns -1. */ | |
3923 | ||
3924 | HOST_WIDE_INT | |
3925 | max_loop_iterations_int (struct loop *loop) | |
b4a9343c | 3926 | { |
807e902e | 3927 | widest_int nit; |
b4a9343c ZD |
3928 | HOST_WIDE_INT hwi_nit; |
3929 | ||
652c4c71 | 3930 | if (!max_loop_iterations (loop, &nit)) |
b4a9343c ZD |
3931 | return -1; |
3932 | ||
807e902e | 3933 | if (!wi::fits_shwi_p (nit)) |
b4a9343c | 3934 | return -1; |
27bcd47c | 3935 | hwi_nit = nit.to_shwi (); |
b4a9343c ZD |
3936 | |
3937 | return hwi_nit < 0 ? -1 : hwi_nit; | |
3938 | } | |
3939 | ||
105e29c5 JH |
3940 | /* Sets NIT to an likely upper bound for the maximum number of executions of the |
3941 | latch of the LOOP. If we have no reliable estimate, the function returns | |
3942 | false, otherwise returns true. */ | |
3943 | ||
3944 | bool | |
3945 | likely_max_loop_iterations (struct loop *loop, widest_int *nit) | |
3946 | { | |
3947 | /* When SCEV information is available, try to update loop iterations | |
3948 | estimate. Otherwise just return whatever we recorded earlier. */ | |
3949 | if (scev_initialized_p ()) | |
adb7eaa2 | 3950 | estimate_numbers_of_iterations (loop); |
105e29c5 JH |
3951 | |
3952 | return get_likely_max_loop_iterations (loop, nit); | |
3953 | } | |
3954 | ||
3955 | /* Similar to max_loop_iterations, but returns the estimate only | |
3956 | if it fits to HOST_WIDE_INT. If this is not the case, or the estimate | |
3957 | on the number of iterations of LOOP could not be derived, returns -1. */ | |
3958 | ||
3959 | HOST_WIDE_INT | |
3960 | likely_max_loop_iterations_int (struct loop *loop) | |
3961 | { | |
3962 | widest_int nit; | |
3963 | HOST_WIDE_INT hwi_nit; | |
3964 | ||
3965 | if (!likely_max_loop_iterations (loop, &nit)) | |
3966 | return -1; | |
3967 | ||
3968 | if (!wi::fits_shwi_p (nit)) | |
3969 | return -1; | |
3970 | hwi_nit = nit.to_shwi (); | |
3971 | ||
3972 | return hwi_nit < 0 ? -1 : hwi_nit; | |
3973 | } | |
3974 | ||
652c4c71 RG |
3975 | /* Returns an estimate for the number of executions of statements |
3976 | in the LOOP. For statements before the loop exit, this exceeds | |
3977 | the number of execution of the latch by one. */ | |
3978 | ||
3979 | HOST_WIDE_INT | |
3980 | estimated_stmt_executions_int (struct loop *loop) | |
3981 | { | |
3982 | HOST_WIDE_INT nit = estimated_loop_iterations_int (loop); | |
3983 | HOST_WIDE_INT snit; | |
3984 | ||
3985 | if (nit == -1) | |
3986 | return -1; | |
3987 | ||
3988 | snit = (HOST_WIDE_INT) ((unsigned HOST_WIDE_INT) nit + 1); | |
3989 | ||
3990 | /* If the computation overflows, return -1. */ | |
3991 | return snit < 0 ? -1 : snit; | |
3992 | } | |
3993 | ||
105e29c5 | 3994 | /* Sets NIT to the maximum number of executions of the latch of the |
652c4c71 RG |
3995 | LOOP, plus one. If we have no reliable estimate, the function returns |
3996 | false, otherwise returns true. */ | |
3997 | ||
3998 | bool | |
807e902e | 3999 | max_stmt_executions (struct loop *loop, widest_int *nit) |
652c4c71 | 4000 | { |
807e902e | 4001 | widest_int nit_minus_one; |
652c4c71 RG |
4002 | |
4003 | if (!max_loop_iterations (loop, nit)) | |
4004 | return false; | |
4005 | ||
4006 | nit_minus_one = *nit; | |
4007 | ||
807e902e | 4008 | *nit += 1; |
652c4c71 | 4009 | |
807e902e | 4010 | return wi::gtu_p (*nit, nit_minus_one); |
652c4c71 RG |
4011 | } |
4012 | ||
105e29c5 JH |
4013 | /* Sets NIT to the estimated maximum number of executions of the latch of the |
4014 | LOOP, plus one. If we have no likely estimate, the function returns | |
4015 | false, otherwise returns true. */ | |
4016 | ||
4017 | bool | |
4018 | likely_max_stmt_executions (struct loop *loop, widest_int *nit) | |
4019 | { | |
4020 | widest_int nit_minus_one; | |
4021 | ||
4022 | if (!likely_max_loop_iterations (loop, nit)) | |
4023 | return false; | |
4024 | ||
4025 | nit_minus_one = *nit; | |
4026 | ||
4027 | *nit += 1; | |
4028 | ||
4029 | return wi::gtu_p (*nit, nit_minus_one); | |
4030 | } | |
4031 | ||
b4a9343c | 4032 | /* Sets NIT to the estimated number of executions of the latch of the |
652c4c71 RG |
4033 | LOOP, plus one. If we have no reliable estimate, the function returns |
4034 | false, otherwise returns true. */ | |
b4a9343c ZD |
4035 | |
4036 | bool | |
807e902e | 4037 | estimated_stmt_executions (struct loop *loop, widest_int *nit) |
b4a9343c | 4038 | { |
807e902e | 4039 | widest_int nit_minus_one; |
b4a9343c | 4040 | |
652c4c71 | 4041 | if (!estimated_loop_iterations (loop, nit)) |
b4a9343c ZD |
4042 | return false; |
4043 | ||
4044 | nit_minus_one = *nit; | |
4045 | ||
807e902e | 4046 | *nit += 1; |
b4a9343c | 4047 | |
807e902e | 4048 | return wi::gtu_p (*nit, nit_minus_one); |
b4a9343c ZD |
4049 | } |
4050 | ||
d73be268 | 4051 | /* Records estimates on numbers of iterations of loops. */ |
e9eb809d ZD |
4052 | |
4053 | void | |
adb7eaa2 | 4054 | estimate_numbers_of_iterations (function *fn) |
e9eb809d | 4055 | { |
e9eb809d ZD |
4056 | struct loop *loop; |
4057 | ||
6ac01510 ILT |
4058 | /* We don't want to issue signed overflow warnings while getting |
4059 | loop iteration estimates. */ | |
4060 | fold_defer_overflow_warnings (); | |
4061 | ||
adb7eaa2 RB |
4062 | FOR_EACH_LOOP_FN (fn, loop, 0) |
4063 | estimate_numbers_of_iterations (loop); | |
6ac01510 ILT |
4064 | |
4065 | fold_undefer_and_ignore_overflow_warnings (); | |
e9eb809d ZD |
4066 | } |
4067 | ||
e9eb809d ZD |
4068 | /* Returns true if statement S1 dominates statement S2. */ |
4069 | ||
bbc8a8dc | 4070 | bool |
355fe088 | 4071 | stmt_dominates_stmt_p (gimple *s1, gimple *s2) |
e9eb809d | 4072 | { |
726a989a | 4073 | basic_block bb1 = gimple_bb (s1), bb2 = gimple_bb (s2); |
e9eb809d ZD |
4074 | |
4075 | if (!bb1 | |
4076 | || s1 == s2) | |
4077 | return true; | |
4078 | ||
4079 | if (bb1 == bb2) | |
4080 | { | |
726a989a | 4081 | gimple_stmt_iterator bsi; |
e9eb809d | 4082 | |
25c6036a RG |
4083 | if (gimple_code (s2) == GIMPLE_PHI) |
4084 | return false; | |
4085 | ||
4086 | if (gimple_code (s1) == GIMPLE_PHI) | |
4087 | return true; | |
4088 | ||
726a989a RB |
4089 | for (bsi = gsi_start_bb (bb1); gsi_stmt (bsi) != s2; gsi_next (&bsi)) |
4090 | if (gsi_stmt (bsi) == s1) | |
e9eb809d ZD |
4091 | return true; |
4092 | ||
4093 | return false; | |
4094 | } | |
4095 | ||
4096 | return dominated_by_p (CDI_DOMINATORS, bb2, bb1); | |
4097 | } | |
4098 | ||
763f4527 | 4099 | /* Returns true when we can prove that the number of executions of |
946e1bc7 ZD |
4100 | STMT in the loop is at most NITER, according to the bound on |
4101 | the number of executions of the statement NITER_BOUND->stmt recorded in | |
870ca331 JH |
4102 | NITER_BOUND and fact that NITER_BOUND->stmt dominate STMT. |
4103 | ||
4104 | ??? This code can become quite a CPU hog - we can have many bounds, | |
4105 | and large basic block forcing stmt_dominates_stmt_p to be queried | |
4106 | many times on a large basic blocks, so the whole thing is O(n^2) | |
4107 | for scev_probably_wraps_p invocation (that can be done n times). | |
4108 | ||
4109 | It would make more sense (and give better answers) to remember BB | |
4110 | bounds computed by discover_iteration_bound_by_body_walk. */ | |
e9eb809d | 4111 | |
1e8552eb | 4112 | static bool |
355fe088 | 4113 | n_of_executions_at_most (gimple *stmt, |
b8698a0f | 4114 | struct nb_iter_bound *niter_bound, |
7aa20a86 | 4115 | tree niter) |
e9eb809d | 4116 | { |
807e902e | 4117 | widest_int bound = niter_bound->bound; |
6e682d7e | 4118 | tree nit_type = TREE_TYPE (niter), e; |
2f133f46 | 4119 | enum tree_code cmp; |
1e8552eb | 4120 | |
946e1bc7 ZD |
4121 | gcc_assert (TYPE_UNSIGNED (nit_type)); |
4122 | ||
4123 | /* If the bound does not even fit into NIT_TYPE, it cannot tell us that | |
4124 | the number of iterations is small. */ | |
807e902e | 4125 | if (!wi::fits_to_tree_p (bound, nit_type)) |
946e1bc7 ZD |
4126 | return false; |
4127 | ||
4128 | /* We know that NITER_BOUND->stmt is executed at most NITER_BOUND->bound + 1 | |
4129 | times. This means that: | |
b8698a0f | 4130 | |
870ca331 JH |
4131 | -- if NITER_BOUND->is_exit is true, then everything after |
4132 | it at most NITER_BOUND->bound times. | |
946e1bc7 ZD |
4133 | |
4134 | -- If NITER_BOUND->is_exit is false, then if we can prove that when STMT | |
4135 | is executed, then NITER_BOUND->stmt is executed as well in the same | |
870ca331 JH |
4136 | iteration then STMT is executed at most NITER_BOUND->bound + 1 times. |
4137 | ||
4138 | If we can determine that NITER_BOUND->stmt is always executed | |
4139 | after STMT, then STMT is executed at most NITER_BOUND->bound + 2 times. | |
4140 | We conclude that if both statements belong to the same | |
4141 | basic block and STMT is before NITER_BOUND->stmt and there are no | |
4142 | statements with side effects in between. */ | |
946e1bc7 ZD |
4143 | |
4144 | if (niter_bound->is_exit) | |
4145 | { | |
870ca331 JH |
4146 | if (stmt == niter_bound->stmt |
4147 | || !stmt_dominates_stmt_p (niter_bound->stmt, stmt)) | |
4148 | return false; | |
4149 | cmp = GE_EXPR; | |
946e1bc7 | 4150 | } |
1e8552eb | 4151 | else |
946e1bc7 | 4152 | { |
870ca331 | 4153 | if (!stmt_dominates_stmt_p (niter_bound->stmt, stmt)) |
946e1bc7 | 4154 | { |
870ca331 JH |
4155 | gimple_stmt_iterator bsi; |
4156 | if (gimple_bb (stmt) != gimple_bb (niter_bound->stmt) | |
4157 | || gimple_code (stmt) == GIMPLE_PHI | |
4158 | || gimple_code (niter_bound->stmt) == GIMPLE_PHI) | |
4159 | return false; | |
4160 | ||
4161 | /* By stmt_dominates_stmt_p we already know that STMT appears | |
4162 | before NITER_BOUND->STMT. Still need to test that the loop | |
4163 | can not be terinated by a side effect in between. */ | |
4164 | for (bsi = gsi_for_stmt (stmt); gsi_stmt (bsi) != niter_bound->stmt; | |
4165 | gsi_next (&bsi)) | |
4166 | if (gimple_has_side_effects (gsi_stmt (bsi))) | |
4167 | return false; | |
807e902e KZ |
4168 | bound += 1; |
4169 | if (bound == 0 | |
4170 | || !wi::fits_to_tree_p (bound, nit_type)) | |
946e1bc7 ZD |
4171 | return false; |
4172 | } | |
4173 | cmp = GT_EXPR; | |
4174 | } | |
1e8552eb | 4175 | |
6e682d7e | 4176 | e = fold_binary (cmp, boolean_type_node, |
807e902e | 4177 | niter, wide_int_to_tree (nit_type, bound)); |
6e682d7e | 4178 | return e && integer_nonzerop (e); |
1e8552eb SP |
4179 | } |
4180 | ||
d7f5de76 | 4181 | /* Returns true if the arithmetics in TYPE can be assumed not to wrap. */ |
e9eb809d | 4182 | |
d7f5de76 ZD |
4183 | bool |
4184 | nowrap_type_p (tree type) | |
d7770457 | 4185 | { |
341c5337 | 4186 | if (ANY_INTEGRAL_TYPE_P (type) |
eeef0e45 | 4187 | && TYPE_OVERFLOW_UNDEFINED (type)) |
d7f5de76 | 4188 | return true; |
d7770457 | 4189 | |
d7f5de76 ZD |
4190 | if (POINTER_TYPE_P (type)) |
4191 | return true; | |
d7770457 | 4192 | |
d7770457 SP |
4193 | return false; |
4194 | } | |
4195 | ||
2f07b722 | 4196 | /* Return true if we can prove LOOP is exited before evolution of induction |
70e1d145 | 4197 | variable {BASE, STEP} overflows with respect to its type bound. */ |
2f07b722 BC |
4198 | |
4199 | static bool | |
4200 | loop_exits_before_overflow (tree base, tree step, | |
355fe088 | 4201 | gimple *at_stmt, struct loop *loop) |
2f07b722 BC |
4202 | { |
4203 | widest_int niter; | |
4204 | struct control_iv *civ; | |
4205 | struct nb_iter_bound *bound; | |
4206 | tree e, delta, step_abs, unsigned_base; | |
4207 | tree type = TREE_TYPE (step); | |
4208 | tree unsigned_type, valid_niter; | |
4209 | ||
4210 | /* Don't issue signed overflow warnings. */ | |
4211 | fold_defer_overflow_warnings (); | |
4212 | ||
4213 | /* Compute the number of iterations before we reach the bound of the | |
4214 | type, and verify that the loop is exited before this occurs. */ | |
4215 | unsigned_type = unsigned_type_for (type); | |
4216 | unsigned_base = fold_convert (unsigned_type, base); | |
4217 | ||
4218 | if (tree_int_cst_sign_bit (step)) | |
4219 | { | |
4220 | tree extreme = fold_convert (unsigned_type, | |
4221 | lower_bound_in_type (type, type)); | |
4222 | delta = fold_build2 (MINUS_EXPR, unsigned_type, unsigned_base, extreme); | |
4223 | step_abs = fold_build1 (NEGATE_EXPR, unsigned_type, | |
4224 | fold_convert (unsigned_type, step)); | |
4225 | } | |
4226 | else | |
4227 | { | |
4228 | tree extreme = fold_convert (unsigned_type, | |
4229 | upper_bound_in_type (type, type)); | |
4230 | delta = fold_build2 (MINUS_EXPR, unsigned_type, extreme, unsigned_base); | |
4231 | step_abs = fold_convert (unsigned_type, step); | |
4232 | } | |
4233 | ||
4234 | valid_niter = fold_build2 (FLOOR_DIV_EXPR, unsigned_type, delta, step_abs); | |
4235 | ||
adb7eaa2 | 4236 | estimate_numbers_of_iterations (loop); |
2f07b722 BC |
4237 | |
4238 | if (max_loop_iterations (loop, &niter) | |
4239 | && wi::fits_to_tree_p (niter, TREE_TYPE (valid_niter)) | |
4240 | && (e = fold_binary (GT_EXPR, boolean_type_node, valid_niter, | |
4241 | wide_int_to_tree (TREE_TYPE (valid_niter), | |
4242 | niter))) != NULL | |
4243 | && integer_nonzerop (e)) | |
4244 | { | |
4245 | fold_undefer_and_ignore_overflow_warnings (); | |
4246 | return true; | |
4247 | } | |
4248 | if (at_stmt) | |
4249 | for (bound = loop->bounds; bound; bound = bound->next) | |
4250 | { | |
4251 | if (n_of_executions_at_most (at_stmt, bound, valid_niter)) | |
4252 | { | |
4253 | fold_undefer_and_ignore_overflow_warnings (); | |
4254 | return true; | |
4255 | } | |
4256 | } | |
4257 | fold_undefer_and_ignore_overflow_warnings (); | |
4258 | ||
4259 | /* Try to prove loop is exited before {base, step} overflows with the | |
4260 | help of analyzed loop control IV. This is done only for IVs with | |
4261 | constant step because otherwise we don't have the information. */ | |
4262 | if (TREE_CODE (step) == INTEGER_CST) | |
f3c5f3a3 | 4263 | { |
f3c5f3a3 BC |
4264 | for (civ = loop->control_ivs; civ; civ = civ->next) |
4265 | { | |
4266 | enum tree_code code; | |
42970a17 | 4267 | tree civ_type = TREE_TYPE (civ->step); |
2f07b722 | 4268 | |
f3c5f3a3 BC |
4269 | /* Have to consider type difference because operand_equal_p ignores |
4270 | that for constants. */ | |
4271 | if (TYPE_UNSIGNED (type) != TYPE_UNSIGNED (civ_type) | |
4272 | || element_precision (type) != element_precision (civ_type)) | |
2f07b722 | 4273 | continue; |
2f07b722 | 4274 | |
f3c5f3a3 BC |
4275 | /* Only consider control IV with same step. */ |
4276 | if (!operand_equal_p (step, civ->step, 0)) | |
4277 | continue; | |
2f07b722 | 4278 | |
f3c5f3a3 | 4279 | /* Done proving if this is a no-overflow control IV. */ |
42970a17 BC |
4280 | if (operand_equal_p (base, civ->base, 0)) |
4281 | return true; | |
4282 | ||
4283 | /* Control IV is recorded after expanding simple operations, | |
4284 | Here we expand base and compare it too. */ | |
4285 | tree expanded_base = expand_simple_operations (base); | |
4286 | if (operand_equal_p (expanded_base, civ->base, 0)) | |
f3c5f3a3 | 4287 | return true; |
2f07b722 | 4288 | |
f3c5f3a3 | 4289 | /* If this is a before stepping control IV, in other words, we have |
2f07b722 | 4290 | |
f3c5f3a3 | 4291 | {civ_base, step} = {base + step, step} |
8710e302 | 4292 | |
f3c5f3a3 BC |
4293 | Because civ {base + step, step} doesn't overflow during loop |
4294 | iterations, {base, step} will not overflow if we can prove the | |
4295 | operation "base + step" does not overflow. Specifically, we try | |
4296 | to prove below conditions are satisfied: | |
8710e302 | 4297 | |
f3c5f3a3 BC |
4298 | base <= UPPER_BOUND (type) - step ;;step > 0 |
4299 | base >= LOWER_BOUND (type) - step ;;step < 0 | |
8710e302 | 4300 | |
f3c5f3a3 BC |
4301 | by proving the reverse conditions are false using loop's initial |
4302 | condition. */ | |
4303 | if (POINTER_TYPE_P (TREE_TYPE (base))) | |
4304 | code = POINTER_PLUS_EXPR; | |
4305 | else | |
4306 | code = PLUS_EXPR; | |
8710e302 | 4307 | |
42970a17 BC |
4308 | tree stepped = fold_build2 (code, TREE_TYPE (base), base, step); |
4309 | tree expanded_stepped = fold_build2 (code, TREE_TYPE (base), | |
4310 | expanded_base, step); | |
4311 | if (operand_equal_p (stepped, civ->base, 0) | |
4312 | || operand_equal_p (expanded_stepped, civ->base, 0)) | |
f3c5f3a3 | 4313 | { |
42970a17 BC |
4314 | tree extreme; |
4315 | ||
f3c5f3a3 BC |
4316 | if (tree_int_cst_sign_bit (step)) |
4317 | { | |
4318 | code = LT_EXPR; | |
4319 | extreme = lower_bound_in_type (type, type); | |
4320 | } | |
4321 | else | |
4322 | { | |
4323 | code = GT_EXPR; | |
4324 | extreme = upper_bound_in_type (type, type); | |
4325 | } | |
4326 | extreme = fold_build2 (MINUS_EXPR, type, extreme, step); | |
4327 | e = fold_build2 (code, boolean_type_node, base, extreme); | |
8aa46dd2 | 4328 | e = simplify_using_initial_conditions (loop, e); |
f3c5f3a3 BC |
4329 | if (integer_zerop (e)) |
4330 | return true; | |
4331 | } | |
4332 | } | |
4333 | } | |
2f07b722 BC |
4334 | |
4335 | return false; | |
4336 | } | |
4337 | ||
b24d9420 BC |
4338 | /* VAR is scev variable whose evolution part is constant STEP, this function |
4339 | proves that VAR can't overflow by using value range info. If VAR's value | |
4340 | range is [MIN, MAX], it can be proven by: | |
4341 | MAX + step doesn't overflow ; if step > 0 | |
4342 | or | |
4343 | MIN + step doesn't underflow ; if step < 0. | |
4344 | ||
4345 | We can only do this if var is computed in every loop iteration, i.e, var's | |
4346 | definition has to dominate loop latch. Consider below example: | |
4347 | ||
4348 | { | |
4349 | unsigned int i; | |
4350 | ||
4351 | <bb 3>: | |
4352 | ||
4353 | <bb 4>: | |
4354 | # RANGE [0, 4294967294] NONZERO 65535 | |
4355 | # i_21 = PHI <0(3), i_18(9)> | |
4356 | if (i_21 != 0) | |
4357 | goto <bb 6>; | |
4358 | else | |
4359 | goto <bb 8>; | |
4360 | ||
4361 | <bb 6>: | |
4362 | # RANGE [0, 65533] NONZERO 65535 | |
4363 | _6 = i_21 + 4294967295; | |
4364 | # RANGE [0, 65533] NONZERO 65535 | |
4365 | _7 = (long unsigned int) _6; | |
4366 | # RANGE [0, 524264] NONZERO 524280 | |
4367 | _8 = _7 * 8; | |
4368 | # PT = nonlocal escaped | |
4369 | _9 = a_14 + _8; | |
4370 | *_9 = 0; | |
4371 | ||
4372 | <bb 8>: | |
4373 | # RANGE [1, 65535] NONZERO 65535 | |
4374 | i_18 = i_21 + 1; | |
4375 | if (i_18 >= 65535) | |
4376 | goto <bb 10>; | |
4377 | else | |
4378 | goto <bb 9>; | |
4379 | ||
4380 | <bb 9>: | |
4381 | goto <bb 4>; | |
4382 | ||
4383 | <bb 10>: | |
4384 | return; | |
4385 | } | |
4386 | ||
4387 | VAR _6 doesn't overflow only with pre-condition (i_21 != 0), here we | |
4388 | can't use _6 to prove no-overlfow for _7. In fact, var _7 takes value | |
4389 | sequence (4294967295, 0, 1, ..., 65533) in loop life time, rather than | |
4390 | (4294967295, 4294967296, ...). */ | |
4391 | ||
4392 | static bool | |
4393 | scev_var_range_cant_overflow (tree var, tree step, struct loop *loop) | |
4394 | { | |
4395 | tree type; | |
4396 | wide_int minv, maxv, diff, step_wi; | |
4397 | enum value_range_type rtype; | |
4398 | ||
4399 | if (TREE_CODE (step) != INTEGER_CST || !INTEGRAL_TYPE_P (TREE_TYPE (var))) | |
4400 | return false; | |
4401 | ||
4402 | /* Check if VAR evaluates in every loop iteration. It's not the case | |
4403 | if VAR is default definition or does not dominate loop's latch. */ | |
4404 | basic_block def_bb = gimple_bb (SSA_NAME_DEF_STMT (var)); | |
4405 | if (!def_bb || !dominated_by_p (CDI_DOMINATORS, loop->latch, def_bb)) | |
4406 | return false; | |
4407 | ||
4408 | rtype = get_range_info (var, &minv, &maxv); | |
4409 | if (rtype != VR_RANGE) | |
4410 | return false; | |
4411 | ||
4412 | /* VAR is a scev whose evolution part is STEP and value range info | |
4413 | is [MIN, MAX], we can prove its no-overflowness by conditions: | |
4414 | ||
4415 | type_MAX - MAX >= step ; if step > 0 | |
4416 | MIN - type_MIN >= |step| ; if step < 0. | |
4417 | ||
4418 | Or VAR must take value outside of value range, which is not true. */ | |
4419 | step_wi = step; | |
4420 | type = TREE_TYPE (var); | |
4421 | if (tree_int_cst_sign_bit (step)) | |
4422 | { | |
4423 | diff = lower_bound_in_type (type, type); | |
4424 | diff = minv - diff; | |
4425 | step_wi = - step_wi; | |
4426 | } | |
4427 | else | |
4428 | { | |
4429 | diff = upper_bound_in_type (type, type); | |
4430 | diff = diff - maxv; | |
4431 | } | |
4432 | ||
4433 | return (wi::geu_p (diff, step_wi)); | |
4434 | } | |
4435 | ||
1e8552eb SP |
4436 | /* Return false only when the induction variable BASE + STEP * I is |
4437 | known to not overflow: i.e. when the number of iterations is small | |
4438 | enough with respect to the step and initial condition in order to | |
4439 | keep the evolution confined in TYPEs bounds. Return true when the | |
4440 | iv is known to overflow or when the property is not computable. | |
b8698a0f | 4441 | |
d7f5de76 ZD |
4442 | USE_OVERFLOW_SEMANTICS is true if this function should assume that |
4443 | the rules for overflow of the given language apply (e.g., that signed | |
b24d9420 BC |
4444 | arithmetics in C does not overflow). |
4445 | ||
4446 | If VAR is a ssa variable, this function also returns false if VAR can | |
4447 | be proven not overflow with value range info. */ | |
1e8552eb SP |
4448 | |
4449 | bool | |
b24d9420 | 4450 | scev_probably_wraps_p (tree var, tree base, tree step, |
355fe088 | 4451 | gimple *at_stmt, struct loop *loop, |
525dc87d | 4452 | bool use_overflow_semantics) |
1e8552eb | 4453 | { |
d7f5de76 ZD |
4454 | /* FIXME: We really need something like |
4455 | http://gcc.gnu.org/ml/gcc-patches/2005-06/msg02025.html. | |
4456 | ||
4457 | We used to test for the following situation that frequently appears | |
4458 | during address arithmetics: | |
b8698a0f | 4459 | |
d7770457 SP |
4460 | D.1621_13 = (long unsigned intD.4) D.1620_12; |
4461 | D.1622_14 = D.1621_13 * 8; | |
4462 | D.1623_15 = (doubleD.29 *) D.1622_14; | |
d7770457 | 4463 | |
d7f5de76 ZD |
4464 | And derived that the sequence corresponding to D_14 |
4465 | can be proved to not wrap because it is used for computing a | |
4466 | memory access; however, this is not really the case -- for example, | |
4467 | if D_12 = (unsigned char) [254,+,1], then D_14 has values | |
4468 | 2032, 2040, 0, 8, ..., but the code is still legal. */ | |
1e8552eb | 4469 | |
18aed06a | 4470 | if (chrec_contains_undetermined (base) |
24938ce9 | 4471 | || chrec_contains_undetermined (step)) |
d7f5de76 | 4472 | return true; |
d7770457 | 4473 | |
6e682d7e | 4474 | if (integer_zerop (step)) |
d7f5de76 | 4475 | return false; |
ab02cc4e | 4476 | |
d7f5de76 ZD |
4477 | /* If we can use the fact that signed and pointer arithmetics does not |
4478 | wrap, we are done. */ | |
dc5b3407 | 4479 | if (use_overflow_semantics && nowrap_type_p (TREE_TYPE (base))) |
d7f5de76 | 4480 | return false; |
ab02cc4e | 4481 | |
24938ce9 ZD |
4482 | /* To be able to use estimates on number of iterations of the loop, |
4483 | we must have an upper bound on the absolute value of the step. */ | |
4484 | if (TREE_CODE (step) != INTEGER_CST) | |
4485 | return true; | |
4486 | ||
b24d9420 BC |
4487 | /* Check if var can be proven not overflow with value range info. */ |
4488 | if (var && TREE_CODE (var) == SSA_NAME | |
4489 | && scev_var_range_cant_overflow (var, step, loop)) | |
4490 | return false; | |
4491 | ||
2f07b722 BC |
4492 | if (loop_exits_before_overflow (base, step, at_stmt, loop)) |
4493 | return false; | |
1e8552eb SP |
4494 | |
4495 | /* At this point we still don't have a proof that the iv does not | |
4496 | overflow: give up. */ | |
4497 | return true; | |
e9eb809d ZD |
4498 | } |
4499 | ||
e9eb809d ZD |
4500 | /* Frees the information on upper bounds on numbers of iterations of LOOP. */ |
4501 | ||
c9639aae | 4502 | void |
adb7eaa2 | 4503 | free_numbers_of_iterations_estimates (struct loop *loop) |
e9eb809d | 4504 | { |
2f07b722 BC |
4505 | struct control_iv *civ; |
4506 | struct nb_iter_bound *bound; | |
c9639aae ZD |
4507 | |
4508 | loop->nb_iterations = NULL; | |
946e1bc7 | 4509 | loop->estimate_state = EST_NOT_COMPUTED; |
2f07b722 | 4510 | for (bound = loop->bounds; bound;) |
e9eb809d | 4511 | { |
2f07b722 | 4512 | struct nb_iter_bound *next = bound->next; |
9e2f83a5 | 4513 | ggc_free (bound); |
2f07b722 | 4514 | bound = next; |
e9eb809d | 4515 | } |
e9eb809d | 4516 | loop->bounds = NULL; |
2f07b722 BC |
4517 | |
4518 | for (civ = loop->control_ivs; civ;) | |
4519 | { | |
4520 | struct control_iv *next = civ->next; | |
4521 | ggc_free (civ); | |
4522 | civ = next; | |
4523 | } | |
4524 | loop->control_ivs = NULL; | |
e9eb809d ZD |
4525 | } |
4526 | ||
d73be268 | 4527 | /* Frees the information on upper bounds on numbers of iterations of loops. */ |
e9eb809d ZD |
4528 | |
4529 | void | |
61183076 | 4530 | free_numbers_of_iterations_estimates (function *fn) |
e9eb809d | 4531 | { |
e9eb809d ZD |
4532 | struct loop *loop; |
4533 | ||
61183076 | 4534 | FOR_EACH_LOOP_FN (fn, loop, 0) |
adb7eaa2 | 4535 | free_numbers_of_iterations_estimates (loop); |
e9eb809d | 4536 | } |
d5ab5675 ZD |
4537 | |
4538 | /* Substitute value VAL for ssa name NAME inside expressions held | |
4539 | at LOOP. */ | |
4540 | ||
4541 | void | |
4542 | substitute_in_loop_info (struct loop *loop, tree name, tree val) | |
4543 | { | |
d5ab5675 | 4544 | loop->nb_iterations = simplify_replace_tree (loop->nb_iterations, name, val); |
d5ab5675 | 4545 | } |