]>
Commit | Line | Data |
---|---|---|
b9d73ea6 | 1 | /* Functions to determine/estimate number of iterations of a loop. |
fbd26352 | 2 | Copyright (C) 2004-2019 Free Software Foundation, Inc. |
48e1416a | 3 | |
b9d73ea6 | 4 | This file is part of GCC. |
48e1416a | 5 | |
b9d73ea6 | 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 | |
8c4c00c1 | 8 | Free Software Foundation; either version 3, or (at your option) any |
b9d73ea6 | 9 | later version. |
48e1416a | 10 | |
b9d73ea6 | 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. | |
48e1416a | 15 | |
b9d73ea6 | 16 | You should have received a copy of the GNU General Public License |
8c4c00c1 | 17 | along with GCC; see the file COPYING3. If not see |
18 | <http://www.gnu.org/licenses/>. */ | |
b9d73ea6 | 19 | |
20 | #include "config.h" | |
21 | #include "system.h" | |
22 | #include "coretypes.h" | |
9ef16211 | 23 | #include "backend.h" |
7c29e30e | 24 | #include "rtl.h" |
b9d73ea6 | 25 | #include "tree.h" |
9ef16211 | 26 | #include "gimple.h" |
7c29e30e | 27 | #include "tree-pass.h" |
9ef16211 | 28 | #include "ssa.h" |
7c29e30e | 29 | #include "gimple-pretty-print.h" |
30 | #include "diagnostic-core.h" | |
05032b10 | 31 | #include "stor-layout.h" |
b20a8bb4 | 32 | #include "fold-const.h" |
9ed99284 | 33 | #include "calls.h" |
7e3cc681 | 34 | #include "intl.h" |
a8783bee | 35 | #include "gimplify.h" |
dcf1a1ec | 36 | #include "gimple-iterator.h" |
073c1fd5 | 37 | #include "tree-cfg.h" |
05d9c18a | 38 | #include "tree-ssa-loop-ivopts.h" |
39 | #include "tree-ssa-loop-niter.h" | |
073c1fd5 | 40 | #include "tree-ssa-loop.h" |
b9d73ea6 | 41 | #include "cfgloop.h" |
b9d73ea6 | 42 | #include "tree-chrec.h" |
43 | #include "tree-scalar-evolution.h" | |
44 | #include "params.h" | |
d64ba765 | 45 | #include "tree-dfa.h" |
b9d73ea6 | 46 | |
f86b328b | 47 | |
a463eaea | 48 | /* The maximum number of dominator BBs we search for conditions |
49 | of loop header copies we use for simplifying a conditional | |
50 | expression. */ | |
51 | #define MAX_DOMINATORS_TO_WALK 8 | |
b9d73ea6 | 52 | |
53 | /* | |
54 | ||
55 | Analysis of number of iterations of an affine exit test. | |
56 | ||
57 | */ | |
58 | ||
a463eaea | 59 | /* Bounds on some value, BELOW <= X <= UP. */ |
60 | ||
6dc50383 | 61 | struct bounds |
a463eaea | 62 | { |
63 | mpz_t below, up; | |
6dc50383 | 64 | }; |
a463eaea | 65 | |
c71993fc | 66 | static bool number_of_iterations_popcount (loop_p loop, edge exit, |
67 | enum tree_code code, | |
68 | struct tree_niter_desc *niter); | |
69 | ||
a463eaea | 70 | |
71 | /* Splits expression EXPR to a variable part VAR and constant OFFSET. */ | |
72 | ||
73 | static void | |
74 | split_to_var_and_offset (tree expr, tree *var, mpz_t offset) | |
75 | { | |
76 | tree type = TREE_TYPE (expr); | |
77 | tree op0, op1; | |
a463eaea | 78 | bool negate = false; |
79 | ||
80 | *var = expr; | |
81 | mpz_set_ui (offset, 0); | |
82 | ||
83 | switch (TREE_CODE (expr)) | |
84 | { | |
85 | case MINUS_EXPR: | |
86 | negate = true; | |
87 | /* Fallthru. */ | |
88 | ||
89 | case PLUS_EXPR: | |
0de36bdb | 90 | case POINTER_PLUS_EXPR: |
a463eaea | 91 | op0 = TREE_OPERAND (expr, 0); |
92 | op1 = TREE_OPERAND (expr, 1); | |
93 | ||
94 | if (TREE_CODE (op1) != INTEGER_CST) | |
95 | break; | |
96 | ||
97 | *var = op0; | |
98 | /* Always sign extend the offset. */ | |
e3d0f65c | 99 | wi::to_mpz (wi::to_wide (op1), offset, SIGNED); |
b6440e39 | 100 | if (negate) |
101 | mpz_neg (offset, offset); | |
a463eaea | 102 | break; |
103 | ||
104 | case INTEGER_CST: | |
105 | *var = build_int_cst_type (type, 0); | |
e3d0f65c | 106 | wi::to_mpz (wi::to_wide (expr), offset, TYPE_SIGN (type)); |
a463eaea | 107 | break; |
108 | ||
109 | default: | |
110 | break; | |
111 | } | |
112 | } | |
113 | ||
1eb4f3d2 | 114 | /* From condition C0 CMP C1 derives information regarding the value range |
115 | of VAR, which is of TYPE. Results are stored in to BELOW and UP. */ | |
116 | ||
117 | static void | |
118 | refine_value_range_using_guard (tree type, tree var, | |
119 | tree c0, enum tree_code cmp, tree c1, | |
120 | mpz_t below, mpz_t up) | |
121 | { | |
122 | tree varc0, varc1, ctype; | |
123 | mpz_t offc0, offc1; | |
124 | mpz_t mint, maxt, minc1, maxc1; | |
125 | wide_int minv, maxv; | |
126 | bool no_wrap = nowrap_type_p (type); | |
127 | bool c0_ok, c1_ok; | |
128 | signop sgn = TYPE_SIGN (type); | |
129 | ||
130 | switch (cmp) | |
131 | { | |
132 | case LT_EXPR: | |
133 | case LE_EXPR: | |
134 | case GT_EXPR: | |
135 | case GE_EXPR: | |
136 | STRIP_SIGN_NOPS (c0); | |
137 | STRIP_SIGN_NOPS (c1); | |
138 | ctype = TREE_TYPE (c0); | |
139 | if (!useless_type_conversion_p (ctype, type)) | |
140 | return; | |
141 | ||
142 | break; | |
143 | ||
144 | case EQ_EXPR: | |
145 | /* We could derive quite precise information from EQ_EXPR, however, | |
146 | such a guard is unlikely to appear, so we do not bother with | |
147 | handling it. */ | |
148 | return; | |
149 | ||
150 | case NE_EXPR: | |
151 | /* NE_EXPR comparisons do not contain much of useful information, | |
152 | except for cases of comparing with bounds. */ | |
153 | if (TREE_CODE (c1) != INTEGER_CST | |
154 | || !INTEGRAL_TYPE_P (type)) | |
155 | return; | |
156 | ||
157 | /* Ensure that the condition speaks about an expression in the same | |
158 | type as X and Y. */ | |
159 | ctype = TREE_TYPE (c0); | |
160 | if (TYPE_PRECISION (ctype) != TYPE_PRECISION (type)) | |
161 | return; | |
162 | c0 = fold_convert (type, c0); | |
163 | c1 = fold_convert (type, c1); | |
164 | ||
165 | if (operand_equal_p (var, c0, 0)) | |
166 | { | |
167 | mpz_t valc1; | |
168 | ||
169 | /* Case of comparing VAR with its below/up bounds. */ | |
170 | mpz_init (valc1); | |
e3d0f65c | 171 | wi::to_mpz (wi::to_wide (c1), valc1, TYPE_SIGN (type)); |
1eb4f3d2 | 172 | if (mpz_cmp (valc1, below) == 0) |
173 | cmp = GT_EXPR; | |
174 | if (mpz_cmp (valc1, up) == 0) | |
175 | cmp = LT_EXPR; | |
176 | ||
177 | mpz_clear (valc1); | |
178 | } | |
179 | else | |
180 | { | |
181 | /* Case of comparing with the bounds of the type. */ | |
182 | wide_int min = wi::min_value (type); | |
183 | wide_int max = wi::max_value (type); | |
184 | ||
e3d0f65c | 185 | if (wi::to_wide (c1) == min) |
1eb4f3d2 | 186 | cmp = GT_EXPR; |
e3d0f65c | 187 | if (wi::to_wide (c1) == max) |
1eb4f3d2 | 188 | cmp = LT_EXPR; |
189 | } | |
190 | ||
191 | /* Quick return if no useful information. */ | |
192 | if (cmp == NE_EXPR) | |
193 | return; | |
194 | ||
195 | break; | |
196 | ||
197 | default: | |
198 | return; | |
199 | } | |
200 | ||
201 | mpz_init (offc0); | |
202 | mpz_init (offc1); | |
203 | split_to_var_and_offset (expand_simple_operations (c0), &varc0, offc0); | |
204 | split_to_var_and_offset (expand_simple_operations (c1), &varc1, offc1); | |
205 | ||
206 | /* We are only interested in comparisons of expressions based on VAR. */ | |
207 | if (operand_equal_p (var, varc1, 0)) | |
208 | { | |
209 | std::swap (varc0, varc1); | |
210 | mpz_swap (offc0, offc1); | |
211 | cmp = swap_tree_comparison (cmp); | |
212 | } | |
213 | else if (!operand_equal_p (var, varc0, 0)) | |
214 | { | |
215 | mpz_clear (offc0); | |
216 | mpz_clear (offc1); | |
217 | return; | |
218 | } | |
219 | ||
220 | mpz_init (mint); | |
221 | mpz_init (maxt); | |
222 | get_type_static_bounds (type, mint, maxt); | |
223 | mpz_init (minc1); | |
224 | mpz_init (maxc1); | |
225 | /* Setup range information for varc1. */ | |
226 | if (integer_zerop (varc1)) | |
227 | { | |
e3d0f65c | 228 | wi::to_mpz (0, minc1, TYPE_SIGN (type)); |
229 | wi::to_mpz (0, maxc1, TYPE_SIGN (type)); | |
1eb4f3d2 | 230 | } |
231 | else if (TREE_CODE (varc1) == SSA_NAME | |
232 | && INTEGRAL_TYPE_P (type) | |
233 | && get_range_info (varc1, &minv, &maxv) == VR_RANGE) | |
234 | { | |
235 | gcc_assert (wi::le_p (minv, maxv, sgn)); | |
236 | wi::to_mpz (minv, minc1, sgn); | |
237 | wi::to_mpz (maxv, maxc1, sgn); | |
238 | } | |
239 | else | |
240 | { | |
241 | mpz_set (minc1, mint); | |
242 | mpz_set (maxc1, maxt); | |
243 | } | |
244 | ||
245 | /* Compute valid range information for varc1 + offc1. Note nothing | |
246 | useful can be derived if it overflows or underflows. Overflow or | |
247 | underflow could happen when: | |
248 | ||
249 | offc1 > 0 && varc1 + offc1 > MAX_VAL (type) | |
250 | offc1 < 0 && varc1 + offc1 < MIN_VAL (type). */ | |
251 | mpz_add (minc1, minc1, offc1); | |
252 | mpz_add (maxc1, maxc1, offc1); | |
253 | c1_ok = (no_wrap | |
254 | || mpz_sgn (offc1) == 0 | |
255 | || (mpz_sgn (offc1) < 0 && mpz_cmp (minc1, mint) >= 0) | |
256 | || (mpz_sgn (offc1) > 0 && mpz_cmp (maxc1, maxt) <= 0)); | |
257 | if (!c1_ok) | |
258 | goto end; | |
259 | ||
260 | if (mpz_cmp (minc1, mint) < 0) | |
261 | mpz_set (minc1, mint); | |
262 | if (mpz_cmp (maxc1, maxt) > 0) | |
263 | mpz_set (maxc1, maxt); | |
264 | ||
265 | if (cmp == LT_EXPR) | |
266 | { | |
267 | cmp = LE_EXPR; | |
268 | mpz_sub_ui (maxc1, maxc1, 1); | |
269 | } | |
270 | if (cmp == GT_EXPR) | |
271 | { | |
272 | cmp = GE_EXPR; | |
273 | mpz_add_ui (minc1, minc1, 1); | |
274 | } | |
275 | ||
276 | /* Compute range information for varc0. If there is no overflow, | |
277 | the condition implied that | |
278 | ||
279 | (varc0) cmp (varc1 + offc1 - offc0) | |
280 | ||
281 | We can possibly improve the upper bound of varc0 if cmp is LE_EXPR, | |
282 | or the below bound if cmp is GE_EXPR. | |
283 | ||
284 | To prove there is no overflow/underflow, we need to check below | |
285 | four cases: | |
286 | 1) cmp == LE_EXPR && offc0 > 0 | |
287 | ||
288 | (varc0 + offc0) doesn't overflow | |
289 | && (varc1 + offc1 - offc0) doesn't underflow | |
290 | ||
291 | 2) cmp == LE_EXPR && offc0 < 0 | |
292 | ||
293 | (varc0 + offc0) doesn't underflow | |
294 | && (varc1 + offc1 - offc0) doesn't overfloe | |
295 | ||
296 | In this case, (varc0 + offc0) will never underflow if we can | |
297 | prove (varc1 + offc1 - offc0) doesn't overflow. | |
298 | ||
299 | 3) cmp == GE_EXPR && offc0 < 0 | |
300 | ||
301 | (varc0 + offc0) doesn't underflow | |
302 | && (varc1 + offc1 - offc0) doesn't overflow | |
303 | ||
304 | 4) cmp == GE_EXPR && offc0 > 0 | |
305 | ||
306 | (varc0 + offc0) doesn't overflow | |
307 | && (varc1 + offc1 - offc0) doesn't underflow | |
308 | ||
309 | In this case, (varc0 + offc0) will never overflow if we can | |
310 | prove (varc1 + offc1 - offc0) doesn't underflow. | |
311 | ||
312 | Note we only handle case 2 and 4 in below code. */ | |
313 | ||
314 | mpz_sub (minc1, minc1, offc0); | |
315 | mpz_sub (maxc1, maxc1, offc0); | |
316 | c0_ok = (no_wrap | |
317 | || mpz_sgn (offc0) == 0 | |
318 | || (cmp == LE_EXPR | |
319 | && mpz_sgn (offc0) < 0 && mpz_cmp (maxc1, maxt) <= 0) | |
320 | || (cmp == GE_EXPR | |
321 | && mpz_sgn (offc0) > 0 && mpz_cmp (minc1, mint) >= 0)); | |
322 | if (!c0_ok) | |
323 | goto end; | |
324 | ||
325 | if (cmp == LE_EXPR) | |
326 | { | |
327 | if (mpz_cmp (up, maxc1) > 0) | |
328 | mpz_set (up, maxc1); | |
329 | } | |
330 | else | |
331 | { | |
332 | if (mpz_cmp (below, minc1) < 0) | |
333 | mpz_set (below, minc1); | |
334 | } | |
335 | ||
336 | end: | |
337 | mpz_clear (mint); | |
338 | mpz_clear (maxt); | |
339 | mpz_clear (minc1); | |
340 | mpz_clear (maxc1); | |
341 | mpz_clear (offc0); | |
342 | mpz_clear (offc1); | |
343 | } | |
344 | ||
a463eaea | 345 | /* Stores estimate on the minimum/maximum value of the expression VAR + OFF |
346 | in TYPE to MIN and MAX. */ | |
347 | ||
348 | static void | |
2c7e5f10 | 349 | determine_value_range (struct loop *loop, tree type, tree var, mpz_t off, |
a463eaea | 350 | mpz_t min, mpz_t max) |
351 | { | |
1eb4f3d2 | 352 | int cnt = 0; |
353 | mpz_t minm, maxm; | |
354 | basic_block bb; | |
9c1be15e | 355 | wide_int minv, maxv; |
be44111e | 356 | enum value_range_kind rtype = VR_VARYING; |
2c7e5f10 | 357 | |
a463eaea | 358 | /* If the expression is a constant, we know its value exactly. */ |
359 | if (integer_zerop (var)) | |
360 | { | |
361 | mpz_set (min, off); | |
362 | mpz_set (max, off); | |
363 | return; | |
364 | } | |
365 | ||
2c7e5f10 | 366 | get_type_static_bounds (type, min, max); |
367 | ||
368 | /* See if we have some range info from VRP. */ | |
369 | if (TREE_CODE (var) == SSA_NAME && INTEGRAL_TYPE_P (type)) | |
370 | { | |
371 | edge e = loop_preheader_edge (loop); | |
9c1be15e | 372 | signop sgn = TYPE_SIGN (type); |
1a91d914 | 373 | gphi_iterator gsi; |
2c7e5f10 | 374 | |
375 | /* Either for VAR itself... */ | |
376 | rtype = get_range_info (var, &minv, &maxv); | |
377 | /* Or for PHI results in loop->header where VAR is used as | |
378 | PHI argument from the loop preheader edge. */ | |
379 | for (gsi = gsi_start_phis (loop->header); !gsi_end_p (gsi); gsi_next (&gsi)) | |
380 | { | |
1a91d914 | 381 | gphi *phi = gsi.phi (); |
9c1be15e | 382 | wide_int minc, maxc; |
2c7e5f10 | 383 | if (PHI_ARG_DEF_FROM_EDGE (phi, e) == var |
384 | && (get_range_info (gimple_phi_result (phi), &minc, &maxc) | |
385 | == VR_RANGE)) | |
386 | { | |
387 | if (rtype != VR_RANGE) | |
388 | { | |
389 | rtype = VR_RANGE; | |
390 | minv = minc; | |
391 | maxv = maxc; | |
392 | } | |
393 | else | |
394 | { | |
9c1be15e | 395 | minv = wi::max (minv, minc, sgn); |
396 | maxv = wi::min (maxv, maxc, sgn); | |
64243ab0 | 397 | /* If the PHI result range are inconsistent with |
398 | the VAR range, give up on looking at the PHI | |
399 | results. This can happen if VR_UNDEFINED is | |
400 | involved. */ | |
a2d59721 | 401 | if (wi::gt_p (minv, maxv, sgn)) |
64243ab0 | 402 | { |
403 | rtype = get_range_info (var, &minv, &maxv); | |
404 | break; | |
405 | } | |
2c7e5f10 | 406 | } |
407 | } | |
408 | } | |
1eb4f3d2 | 409 | mpz_init (minm); |
410 | mpz_init (maxm); | |
411 | if (rtype != VR_RANGE) | |
412 | { | |
413 | mpz_set (minm, min); | |
414 | mpz_set (maxm, max); | |
415 | } | |
416 | else | |
2c7e5f10 | 417 | { |
9c1be15e | 418 | gcc_assert (wi::le_p (minv, maxv, sgn)); |
9c1be15e | 419 | wi::to_mpz (minv, minm, sgn); |
420 | wi::to_mpz (maxv, maxm, sgn); | |
1eb4f3d2 | 421 | } |
422 | /* Now walk the dominators of the loop header and use the entry | |
423 | guards to refine the estimates. */ | |
424 | for (bb = loop->header; | |
425 | bb != ENTRY_BLOCK_PTR_FOR_FN (cfun) && cnt < MAX_DOMINATORS_TO_WALK; | |
426 | bb = get_immediate_dominator (CDI_DOMINATORS, bb)) | |
427 | { | |
428 | edge e; | |
429 | tree c0, c1; | |
42acab1c | 430 | gimple *cond; |
1eb4f3d2 | 431 | enum tree_code cmp; |
432 | ||
433 | if (!single_pred_p (bb)) | |
434 | continue; | |
435 | e = single_pred_edge (bb); | |
436 | ||
437 | if (!(e->flags & (EDGE_TRUE_VALUE | EDGE_FALSE_VALUE))) | |
438 | continue; | |
439 | ||
440 | cond = last_stmt (e->src); | |
441 | c0 = gimple_cond_lhs (cond); | |
442 | cmp = gimple_cond_code (cond); | |
443 | c1 = gimple_cond_rhs (cond); | |
444 | ||
445 | if (e->flags & EDGE_FALSE_VALUE) | |
446 | cmp = invert_tree_comparison (cmp, false); | |
447 | ||
448 | refine_value_range_using_guard (type, var, c0, cmp, c1, minm, maxm); | |
449 | ++cnt; | |
450 | } | |
451 | ||
452 | mpz_add (minm, minm, off); | |
453 | mpz_add (maxm, maxm, off); | |
454 | /* If the computation may not wrap or off is zero, then this | |
455 | is always fine. If off is negative and minv + off isn't | |
456 | smaller than type's minimum, or off is positive and | |
457 | maxv + off isn't bigger than type's maximum, use the more | |
458 | precise range too. */ | |
459 | if (nowrap_type_p (type) | |
460 | || mpz_sgn (off) == 0 | |
461 | || (mpz_sgn (off) < 0 && mpz_cmp (minm, min) >= 0) | |
462 | || (mpz_sgn (off) > 0 && mpz_cmp (maxm, max) <= 0)) | |
463 | { | |
464 | mpz_set (min, minm); | |
465 | mpz_set (max, maxm); | |
2c7e5f10 | 466 | mpz_clear (minm); |
467 | mpz_clear (maxm); | |
1eb4f3d2 | 468 | return; |
2c7e5f10 | 469 | } |
1eb4f3d2 | 470 | mpz_clear (minm); |
471 | mpz_clear (maxm); | |
2c7e5f10 | 472 | } |
473 | ||
a463eaea | 474 | /* If the computation may wrap, we know nothing about the value, except for |
475 | the range of the type. */ | |
a463eaea | 476 | if (!nowrap_type_p (type)) |
477 | return; | |
478 | ||
479 | /* Since the addition of OFF does not wrap, if OFF is positive, then we may | |
480 | add it to MIN, otherwise to MAX. */ | |
481 | if (mpz_sgn (off) < 0) | |
482 | mpz_add (max, max, off); | |
483 | else | |
484 | mpz_add (min, min, off); | |
485 | } | |
486 | ||
487 | /* Stores the bounds on the difference of the values of the expressions | |
488 | (var + X) and (var + Y), computed in TYPE, to BNDS. */ | |
489 | ||
490 | static void | |
491 | bound_difference_of_offsetted_base (tree type, mpz_t x, mpz_t y, | |
492 | bounds *bnds) | |
493 | { | |
494 | int rel = mpz_cmp (x, y); | |
495 | bool may_wrap = !nowrap_type_p (type); | |
496 | mpz_t m; | |
497 | ||
498 | /* If X == Y, then the expressions are always equal. | |
499 | If X > Y, there are the following possibilities: | |
500 | a) neither of var + X and var + Y overflow or underflow, or both of | |
501 | them do. Then their difference is X - Y. | |
502 | b) var + X overflows, and var + Y does not. Then the values of the | |
503 | expressions are var + X - M and var + Y, where M is the range of | |
504 | the type, and their difference is X - Y - M. | |
505 | c) var + Y underflows and var + X does not. Their difference again | |
506 | is M - X + Y. | |
507 | Therefore, if the arithmetics in type does not overflow, then the | |
508 | bounds are (X - Y, X - Y), otherwise they are (X - Y - M, X - Y) | |
509 | Similarly, if X < Y, the bounds are either (X - Y, X - Y) or | |
510 | (X - Y, X - Y + M). */ | |
511 | ||
512 | if (rel == 0) | |
513 | { | |
514 | mpz_set_ui (bnds->below, 0); | |
515 | mpz_set_ui (bnds->up, 0); | |
516 | return; | |
517 | } | |
518 | ||
519 | mpz_init (m); | |
796b6678 | 520 | wi::to_mpz (wi::minus_one (TYPE_PRECISION (type)), m, UNSIGNED); |
a463eaea | 521 | mpz_add_ui (m, m, 1); |
522 | mpz_sub (bnds->up, x, y); | |
523 | mpz_set (bnds->below, bnds->up); | |
524 | ||
525 | if (may_wrap) | |
526 | { | |
527 | if (rel > 0) | |
528 | mpz_sub (bnds->below, bnds->below, m); | |
529 | else | |
530 | mpz_add (bnds->up, bnds->up, m); | |
531 | } | |
532 | ||
533 | mpz_clear (m); | |
534 | } | |
535 | ||
536 | /* From condition C0 CMP C1 derives information regarding the | |
537 | difference of values of VARX + OFFX and VARY + OFFY, computed in TYPE, | |
538 | and stores it to BNDS. */ | |
539 | ||
540 | static void | |
541 | refine_bounds_using_guard (tree type, tree varx, mpz_t offx, | |
542 | tree vary, mpz_t offy, | |
543 | tree c0, enum tree_code cmp, tree c1, | |
544 | bounds *bnds) | |
545 | { | |
a4f59596 | 546 | tree varc0, varc1, ctype; |
a463eaea | 547 | mpz_t offc0, offc1, loffx, loffy, bnd; |
548 | bool lbound = false; | |
549 | bool no_wrap = nowrap_type_p (type); | |
550 | bool x_ok, y_ok; | |
551 | ||
552 | switch (cmp) | |
553 | { | |
554 | case LT_EXPR: | |
555 | case LE_EXPR: | |
556 | case GT_EXPR: | |
557 | case GE_EXPR: | |
8cd53b4f | 558 | STRIP_SIGN_NOPS (c0); |
559 | STRIP_SIGN_NOPS (c1); | |
560 | ctype = TREE_TYPE (c0); | |
548044d8 | 561 | if (!useless_type_conversion_p (ctype, type)) |
8cd53b4f | 562 | return; |
563 | ||
a463eaea | 564 | break; |
565 | ||
566 | case EQ_EXPR: | |
567 | /* We could derive quite precise information from EQ_EXPR, however, such | |
8cd53b4f | 568 | a guard is unlikely to appear, so we do not bother with handling |
569 | it. */ | |
a463eaea | 570 | return; |
571 | ||
572 | case NE_EXPR: | |
8cd53b4f | 573 | /* NE_EXPR comparisons do not contain much of useful information, except for |
574 | special case of comparing with the bounds of the type. */ | |
575 | if (TREE_CODE (c1) != INTEGER_CST | |
576 | || !INTEGRAL_TYPE_P (type)) | |
577 | return; | |
578 | ||
579 | /* Ensure that the condition speaks about an expression in the same type | |
580 | as X and Y. */ | |
581 | ctype = TREE_TYPE (c0); | |
582 | if (TYPE_PRECISION (ctype) != TYPE_PRECISION (type)) | |
583 | return; | |
584 | c0 = fold_convert (type, c0); | |
585 | c1 = fold_convert (type, c1); | |
586 | ||
587 | if (TYPE_MIN_VALUE (type) | |
588 | && operand_equal_p (c1, TYPE_MIN_VALUE (type), 0)) | |
589 | { | |
590 | cmp = GT_EXPR; | |
591 | break; | |
592 | } | |
593 | if (TYPE_MAX_VALUE (type) | |
594 | && operand_equal_p (c1, TYPE_MAX_VALUE (type), 0)) | |
595 | { | |
596 | cmp = LT_EXPR; | |
597 | break; | |
598 | } | |
599 | ||
a463eaea | 600 | return; |
601 | default: | |
602 | return; | |
48e1416a | 603 | } |
a463eaea | 604 | |
605 | mpz_init (offc0); | |
606 | mpz_init (offc1); | |
607 | split_to_var_and_offset (expand_simple_operations (c0), &varc0, offc0); | |
608 | split_to_var_and_offset (expand_simple_operations (c1), &varc1, offc1); | |
609 | ||
610 | /* We are only interested in comparisons of expressions based on VARX and | |
611 | VARY. TODO -- we might also be able to derive some bounds from | |
612 | expressions containing just one of the variables. */ | |
613 | ||
614 | if (operand_equal_p (varx, varc1, 0)) | |
615 | { | |
a4f59596 | 616 | std::swap (varc0, varc1); |
a463eaea | 617 | mpz_swap (offc0, offc1); |
618 | cmp = swap_tree_comparison (cmp); | |
619 | } | |
620 | ||
621 | if (!operand_equal_p (varx, varc0, 0) | |
622 | || !operand_equal_p (vary, varc1, 0)) | |
623 | goto end; | |
624 | ||
625 | mpz_init_set (loffx, offx); | |
626 | mpz_init_set (loffy, offy); | |
627 | ||
628 | if (cmp == GT_EXPR || cmp == GE_EXPR) | |
629 | { | |
a4f59596 | 630 | std::swap (varx, vary); |
a463eaea | 631 | mpz_swap (offc0, offc1); |
632 | mpz_swap (loffx, loffy); | |
633 | cmp = swap_tree_comparison (cmp); | |
634 | lbound = true; | |
635 | } | |
636 | ||
637 | /* If there is no overflow, the condition implies that | |
638 | ||
639 | (VARX + OFFX) cmp (VARY + OFFY) + (OFFX - OFFY + OFFC1 - OFFC0). | |
640 | ||
641 | The overflows and underflows may complicate things a bit; each | |
642 | overflow decreases the appropriate offset by M, and underflow | |
643 | increases it by M. The above inequality would not necessarily be | |
644 | true if | |
48e1416a | 645 | |
a463eaea | 646 | -- VARX + OFFX underflows and VARX + OFFC0 does not, or |
647 | VARX + OFFC0 overflows, but VARX + OFFX does not. | |
648 | This may only happen if OFFX < OFFC0. | |
649 | -- VARY + OFFY overflows and VARY + OFFC1 does not, or | |
650 | VARY + OFFC1 underflows and VARY + OFFY does not. | |
651 | This may only happen if OFFY > OFFC1. */ | |
652 | ||
653 | if (no_wrap) | |
654 | { | |
655 | x_ok = true; | |
656 | y_ok = true; | |
657 | } | |
658 | else | |
659 | { | |
660 | x_ok = (integer_zerop (varx) | |
661 | || mpz_cmp (loffx, offc0) >= 0); | |
662 | y_ok = (integer_zerop (vary) | |
663 | || mpz_cmp (loffy, offc1) <= 0); | |
664 | } | |
665 | ||
666 | if (x_ok && y_ok) | |
667 | { | |
668 | mpz_init (bnd); | |
669 | mpz_sub (bnd, loffx, loffy); | |
670 | mpz_add (bnd, bnd, offc1); | |
671 | mpz_sub (bnd, bnd, offc0); | |
672 | ||
673 | if (cmp == LT_EXPR) | |
674 | mpz_sub_ui (bnd, bnd, 1); | |
675 | ||
676 | if (lbound) | |
677 | { | |
678 | mpz_neg (bnd, bnd); | |
679 | if (mpz_cmp (bnds->below, bnd) < 0) | |
680 | mpz_set (bnds->below, bnd); | |
681 | } | |
682 | else | |
683 | { | |
684 | if (mpz_cmp (bnd, bnds->up) < 0) | |
685 | mpz_set (bnds->up, bnd); | |
686 | } | |
687 | mpz_clear (bnd); | |
688 | } | |
689 | ||
690 | mpz_clear (loffx); | |
691 | mpz_clear (loffy); | |
692 | end: | |
693 | mpz_clear (offc0); | |
694 | mpz_clear (offc1); | |
695 | } | |
696 | ||
697 | /* Stores the bounds on the value of the expression X - Y in LOOP to BNDS. | |
698 | The subtraction is considered to be performed in arbitrary precision, | |
699 | without overflows. | |
48e1416a | 700 | |
a463eaea | 701 | We do not attempt to be too clever regarding the value ranges of X and |
702 | Y; most of the time, they are just integers or ssa names offsetted by | |
703 | integer. However, we try to use the information contained in the | |
704 | comparisons before the loop (usually created by loop header copying). */ | |
705 | ||
706 | static void | |
707 | bound_difference (struct loop *loop, tree x, tree y, bounds *bnds) | |
708 | { | |
709 | tree type = TREE_TYPE (x); | |
710 | tree varx, vary; | |
711 | mpz_t offx, offy; | |
712 | mpz_t minx, maxx, miny, maxy; | |
713 | int cnt = 0; | |
714 | edge e; | |
715 | basic_block bb; | |
75a70cf9 | 716 | tree c0, c1; |
42acab1c | 717 | gimple *cond; |
a463eaea | 718 | enum tree_code cmp; |
719 | ||
8cd53b4f | 720 | /* Get rid of unnecessary casts, but preserve the value of |
721 | the expressions. */ | |
722 | STRIP_SIGN_NOPS (x); | |
723 | STRIP_SIGN_NOPS (y); | |
724 | ||
a463eaea | 725 | mpz_init (bnds->below); |
726 | mpz_init (bnds->up); | |
727 | mpz_init (offx); | |
728 | mpz_init (offy); | |
729 | split_to_var_and_offset (x, &varx, offx); | |
730 | split_to_var_and_offset (y, &vary, offy); | |
731 | ||
732 | if (!integer_zerop (varx) | |
733 | && operand_equal_p (varx, vary, 0)) | |
734 | { | |
735 | /* Special case VARX == VARY -- we just need to compare the | |
736 | offsets. The matters are a bit more complicated in the | |
737 | case addition of offsets may wrap. */ | |
738 | bound_difference_of_offsetted_base (type, offx, offy, bnds); | |
739 | } | |
740 | else | |
741 | { | |
742 | /* Otherwise, use the value ranges to determine the initial | |
743 | estimates on below and up. */ | |
744 | mpz_init (minx); | |
745 | mpz_init (maxx); | |
746 | mpz_init (miny); | |
747 | mpz_init (maxy); | |
2c7e5f10 | 748 | determine_value_range (loop, type, varx, offx, minx, maxx); |
749 | determine_value_range (loop, type, vary, offy, miny, maxy); | |
a463eaea | 750 | |
751 | mpz_sub (bnds->below, minx, maxy); | |
752 | mpz_sub (bnds->up, maxx, miny); | |
753 | mpz_clear (minx); | |
754 | mpz_clear (maxx); | |
755 | mpz_clear (miny); | |
756 | mpz_clear (maxy); | |
757 | } | |
758 | ||
759 | /* If both X and Y are constants, we cannot get any more precise. */ | |
760 | if (integer_zerop (varx) && integer_zerop (vary)) | |
761 | goto end; | |
762 | ||
763 | /* Now walk the dominators of the loop header and use the entry | |
764 | guards to refine the estimates. */ | |
765 | for (bb = loop->header; | |
34154e27 | 766 | bb != ENTRY_BLOCK_PTR_FOR_FN (cfun) && cnt < MAX_DOMINATORS_TO_WALK; |
a463eaea | 767 | bb = get_immediate_dominator (CDI_DOMINATORS, bb)) |
768 | { | |
769 | if (!single_pred_p (bb)) | |
770 | continue; | |
771 | e = single_pred_edge (bb); | |
772 | ||
773 | if (!(e->flags & (EDGE_TRUE_VALUE | EDGE_FALSE_VALUE))) | |
774 | continue; | |
775 | ||
75a70cf9 | 776 | cond = last_stmt (e->src); |
777 | c0 = gimple_cond_lhs (cond); | |
778 | cmp = gimple_cond_code (cond); | |
779 | c1 = gimple_cond_rhs (cond); | |
a463eaea | 780 | |
781 | if (e->flags & EDGE_FALSE_VALUE) | |
782 | cmp = invert_tree_comparison (cmp, false); | |
783 | ||
784 | refine_bounds_using_guard (type, varx, offx, vary, offy, | |
785 | c0, cmp, c1, bnds); | |
786 | ++cnt; | |
787 | } | |
788 | ||
789 | end: | |
790 | mpz_clear (offx); | |
791 | mpz_clear (offy); | |
792 | } | |
793 | ||
794 | /* Update the bounds in BNDS that restrict the value of X to the bounds | |
795 | that restrict the value of X + DELTA. X can be obtained as a | |
796 | difference of two values in TYPE. */ | |
797 | ||
798 | static void | |
10c3fe8d | 799 | bounds_add (bounds *bnds, const widest_int &delta, tree type) |
a463eaea | 800 | { |
801 | mpz_t mdelta, max; | |
802 | ||
803 | mpz_init (mdelta); | |
796b6678 | 804 | wi::to_mpz (delta, mdelta, SIGNED); |
a463eaea | 805 | |
806 | mpz_init (max); | |
796b6678 | 807 | wi::to_mpz (wi::minus_one (TYPE_PRECISION (type)), max, UNSIGNED); |
a463eaea | 808 | |
809 | mpz_add (bnds->up, bnds->up, mdelta); | |
810 | mpz_add (bnds->below, bnds->below, mdelta); | |
811 | ||
812 | if (mpz_cmp (bnds->up, max) > 0) | |
813 | mpz_set (bnds->up, max); | |
814 | ||
815 | mpz_neg (max, max); | |
816 | if (mpz_cmp (bnds->below, max) < 0) | |
817 | mpz_set (bnds->below, max); | |
818 | ||
819 | mpz_clear (mdelta); | |
820 | mpz_clear (max); | |
821 | } | |
822 | ||
823 | /* Update the bounds in BNDS that restrict the value of X to the bounds | |
824 | that restrict the value of -X. */ | |
825 | ||
826 | static void | |
827 | bounds_negate (bounds *bnds) | |
828 | { | |
829 | mpz_t tmp; | |
830 | ||
831 | mpz_init_set (tmp, bnds->up); | |
832 | mpz_neg (bnds->up, bnds->below); | |
833 | mpz_neg (bnds->below, tmp); | |
834 | mpz_clear (tmp); | |
835 | } | |
836 | ||
b9d73ea6 | 837 | /* Returns inverse of X modulo 2^s, where MASK = 2^s-1. */ |
838 | ||
839 | static tree | |
840 | inverse (tree x, tree mask) | |
841 | { | |
842 | tree type = TREE_TYPE (x); | |
aa8fd7bc | 843 | tree rslt; |
844 | unsigned ctr = tree_floor_log2 (mask); | |
845 | ||
846 | if (TYPE_PRECISION (type) <= HOST_BITS_PER_WIDE_INT) | |
847 | { | |
848 | unsigned HOST_WIDE_INT ix; | |
849 | unsigned HOST_WIDE_INT imask; | |
850 | unsigned HOST_WIDE_INT irslt = 1; | |
851 | ||
e1d65c9f | 852 | gcc_assert (cst_and_fits_in_hwi (x)); |
853 | gcc_assert (cst_and_fits_in_hwi (mask)); | |
aa8fd7bc | 854 | |
855 | ix = int_cst_value (x); | |
856 | imask = int_cst_value (mask); | |
857 | ||
858 | for (; ctr; ctr--) | |
859 | { | |
860 | irslt *= ix; | |
861 | ix *= ix; | |
862 | } | |
863 | irslt &= imask; | |
b9d73ea6 | 864 | |
aa8fd7bc | 865 | rslt = build_int_cst_type (type, irslt); |
866 | } | |
867 | else | |
b9d73ea6 | 868 | { |
05db596e | 869 | rslt = build_int_cst (type, 1); |
aa8fd7bc | 870 | for (; ctr; ctr--) |
871 | { | |
317e2a67 | 872 | rslt = int_const_binop (MULT_EXPR, rslt, x); |
873 | x = int_const_binop (MULT_EXPR, x, x); | |
aa8fd7bc | 874 | } |
317e2a67 | 875 | rslt = int_const_binop (BIT_AND_EXPR, rslt, mask); |
b9d73ea6 | 876 | } |
877 | ||
878 | return rslt; | |
879 | } | |
880 | ||
a463eaea | 881 | /* Derives the upper bound BND on the number of executions of loop with exit |
d16175cc | 882 | condition S * i <> C. If NO_OVERFLOW is true, then the control variable of |
883 | the loop does not overflow. EXIT_MUST_BE_TAKEN is true if we are guaranteed | |
884 | that the loop ends through this exit, i.e., the induction variable ever | |
885 | reaches the value of C. | |
886 | ||
887 | The value C is equal to final - base, where final and base are the final and | |
888 | initial value of the actual induction variable in the analysed loop. BNDS | |
889 | bounds the value of this difference when computed in signed type with | |
890 | unbounded range, while the computation of C is performed in an unsigned | |
891 | type with the range matching the range of the type of the induction variable. | |
892 | In particular, BNDS.up contains an upper bound on C in the following cases: | |
893 | -- if the iv must reach its final value without overflow, i.e., if | |
894 | NO_OVERFLOW && EXIT_MUST_BE_TAKEN is true, or | |
895 | -- if final >= base, which we know to hold when BNDS.below >= 0. */ | |
a463eaea | 896 | |
897 | static void | |
898 | number_of_iterations_ne_max (mpz_t bnd, bool no_overflow, tree c, tree s, | |
d16175cc | 899 | bounds *bnds, bool exit_must_be_taken) |
a463eaea | 900 | { |
5de9d3ed | 901 | widest_int max; |
a463eaea | 902 | mpz_t d; |
e4800344 | 903 | tree type = TREE_TYPE (c); |
d16175cc | 904 | bool bnds_u_valid = ((no_overflow && exit_must_be_taken) |
905 | || mpz_sgn (bnds->below) >= 0); | |
a463eaea | 906 | |
e4800344 | 907 | if (integer_onep (s) |
908 | || (TREE_CODE (c) == INTEGER_CST | |
909 | && TREE_CODE (s) == INTEGER_CST | |
e3d0f65c | 910 | && wi::mod_trunc (wi::to_wide (c), wi::to_wide (s), |
911 | TYPE_SIGN (type)) == 0) | |
e913b5cd | 912 | || (TYPE_OVERFLOW_UNDEFINED (type) |
e4800344 | 913 | && multiple_of_p (type, c, s))) |
d16175cc | 914 | { |
915 | /* If C is an exact multiple of S, then its value will be reached before | |
916 | the induction variable overflows (unless the loop is exited in some | |
917 | other way before). Note that the actual induction variable in the | |
918 | loop (which ranges from base to final instead of from 0 to C) may | |
919 | overflow, in which case BNDS.up will not be giving a correct upper | |
920 | bound on C; thus, BNDS_U_VALID had to be computed in advance. */ | |
921 | no_overflow = true; | |
922 | exit_must_be_taken = true; | |
923 | } | |
924 | ||
925 | /* If the induction variable can overflow, the number of iterations is at | |
926 | most the period of the control variable (or infinite, but in that case | |
927 | the whole # of iterations analysis will fail). */ | |
928 | if (!no_overflow) | |
a463eaea | 929 | { |
e3d0f65c | 930 | max = wi::mask <widest_int> (TYPE_PRECISION (type) |
931 | - wi::ctz (wi::to_wide (s)), false); | |
796b6678 | 932 | wi::to_mpz (max, bnd, UNSIGNED); |
a463eaea | 933 | return; |
934 | } | |
935 | ||
d16175cc | 936 | /* Now we know that the induction variable does not overflow, so the loop |
937 | iterates at most (range of type / S) times. */ | |
796b6678 | 938 | wi::to_mpz (wi::minus_one (TYPE_PRECISION (type)), bnd, UNSIGNED); |
d16175cc | 939 | |
940 | /* If the induction variable is guaranteed to reach the value of C before | |
941 | overflow, ... */ | |
942 | if (exit_must_be_taken) | |
943 | { | |
9d75589a | 944 | /* ... then we can strengthen this to C / S, and possibly we can use |
d16175cc | 945 | the upper bound on C given by BNDS. */ |
946 | if (TREE_CODE (c) == INTEGER_CST) | |
e3d0f65c | 947 | wi::to_mpz (wi::to_wide (c), bnd, UNSIGNED); |
d16175cc | 948 | else if (bnds_u_valid) |
949 | mpz_set (bnd, bnds->up); | |
950 | } | |
a463eaea | 951 | |
952 | mpz_init (d); | |
e3d0f65c | 953 | wi::to_mpz (wi::to_wide (s), d, UNSIGNED); |
a463eaea | 954 | mpz_fdiv_q (bnd, bnd, d); |
955 | mpz_clear (d); | |
956 | } | |
957 | ||
0f81d1b1 | 958 | /* Determines number of iterations of loop whose ending condition |
959 | is IV <> FINAL. TYPE is the type of the iv. The number of | |
fffbaaf2 | 960 | iterations is stored to NITER. EXIT_MUST_BE_TAKEN is true if |
0840f4fe | 961 | we know that the exit must be taken eventually, i.e., that the IV |
962 | ever reaches the value FINAL (we derived this earlier, and possibly set | |
a463eaea | 963 | NITER->assumptions to make sure this is the case). BNDS contains the |
964 | bounds on the difference FINAL - IV->base. */ | |
b9d73ea6 | 965 | |
0f81d1b1 | 966 | static bool |
a94f1e5b | 967 | number_of_iterations_ne (struct loop *loop, tree type, affine_iv *iv, |
968 | tree final, struct tree_niter_desc *niter, | |
969 | bool exit_must_be_taken, bounds *bnds) | |
b9d73ea6 | 970 | { |
0f81d1b1 | 971 | tree niter_type = unsigned_type_for (type); |
972 | tree s, c, d, bits, assumption, tmp, bound; | |
a463eaea | 973 | mpz_t max; |
b9d73ea6 | 974 | |
b30560de | 975 | niter->control = *iv; |
976 | niter->bound = final; | |
977 | niter->cmp = NE_EXPR; | |
978 | ||
a463eaea | 979 | /* Rearrange the terms so that we get inequality S * i <> C, with S |
980 | positive. Also cast everything to the unsigned type. If IV does | |
981 | not overflow, BNDS bounds the value of C. Also, this is the | |
982 | case if the computation |FINAL - IV->base| does not overflow, i.e., | |
983 | if BNDS->below in the result is nonnegative. */ | |
0f81d1b1 | 984 | if (tree_int_cst_sign_bit (iv->step)) |
b9d73ea6 | 985 | { |
0f81d1b1 | 986 | s = fold_convert (niter_type, |
987 | fold_build1 (NEGATE_EXPR, type, iv->step)); | |
988 | c = fold_build2 (MINUS_EXPR, niter_type, | |
989 | fold_convert (niter_type, iv->base), | |
990 | fold_convert (niter_type, final)); | |
a463eaea | 991 | bounds_negate (bnds); |
b9d73ea6 | 992 | } |
553b9523 | 993 | else |
b9d73ea6 | 994 | { |
0f81d1b1 | 995 | s = fold_convert (niter_type, iv->step); |
996 | c = fold_build2 (MINUS_EXPR, niter_type, | |
997 | fold_convert (niter_type, final), | |
998 | fold_convert (niter_type, iv->base)); | |
999 | } | |
b9d73ea6 | 1000 | |
a463eaea | 1001 | mpz_init (max); |
d16175cc | 1002 | number_of_iterations_ne_max (max, iv->no_overflow, c, s, bnds, |
1003 | exit_must_be_taken); | |
5de9d3ed | 1004 | niter->max = widest_int::from (wi::from_mpz (niter_type, max, false), |
1005 | TYPE_SIGN (niter_type)); | |
a463eaea | 1006 | mpz_clear (max); |
1007 | ||
ad8a330c | 1008 | /* Compute no-overflow information for the control iv. This can be |
b441db40 | 1009 | proven when below two conditions are satisfied: |
ad8a330c | 1010 | |
b441db40 | 1011 | 1) IV evaluates toward FINAL at beginning, i.e: |
ad8a330c | 1012 | base <= FINAL ; step > 0 |
1013 | base >= FINAL ; step < 0 | |
1014 | ||
b441db40 | 1015 | 2) |FINAL - base| is an exact multiple of step. |
1016 | ||
1017 | Unfortunately, it's hard to prove above conditions after pass loop-ch | |
1018 | because loop with exit condition (IV != FINAL) usually will be guarded | |
1019 | by initial-condition (IV.base - IV.step != FINAL). In this case, we | |
1020 | can alternatively try to prove below conditions: | |
1021 | ||
1022 | 1') IV evaluates toward FINAL at beginning, i.e: | |
1023 | new_base = base - step < FINAL ; step > 0 | |
1024 | && base - step doesn't underflow | |
1025 | new_base = base - step > FINAL ; step < 0 | |
1026 | && base - step doesn't overflow | |
ad8a330c | 1027 | |
b441db40 | 1028 | 2') |FINAL - new_base| is an exact multiple of step. |
ad8a330c | 1029 | |
b441db40 | 1030 | Please refer to PR34114 as an example of loop-ch's impact, also refer |
1031 | to PR72817 as an example why condition 2') is necessary. | |
ad8a330c | 1032 | |
b441db40 | 1033 | Note, for NE_EXPR, base equals to FINAL is a special case, in |
ad8a330c | 1034 | which the loop exits immediately, and the iv does not overflow. */ |
1035 | if (!niter->control.no_overflow | |
1036 | && (integer_onep (s) || multiple_of_p (type, c, s))) | |
a94f1e5b | 1037 | { |
b441db40 | 1038 | tree t, cond, new_c, relaxed_cond = boolean_false_node; |
ad8a330c | 1039 | |
1040 | if (tree_int_cst_sign_bit (iv->step)) | |
1041 | { | |
1042 | cond = fold_build2 (GE_EXPR, boolean_type_node, iv->base, final); | |
1043 | if (TREE_CODE (type) == INTEGER_TYPE) | |
1044 | { | |
1045 | /* Only when base - step doesn't overflow. */ | |
1046 | t = TYPE_MAX_VALUE (type); | |
1047 | t = fold_build2 (PLUS_EXPR, type, t, iv->step); | |
1048 | t = fold_build2 (GE_EXPR, boolean_type_node, t, iv->base); | |
1049 | if (integer_nonzerop (t)) | |
1050 | { | |
1051 | t = fold_build2 (MINUS_EXPR, type, iv->base, iv->step); | |
b441db40 | 1052 | new_c = fold_build2 (MINUS_EXPR, niter_type, |
1053 | fold_convert (niter_type, t), | |
1054 | fold_convert (niter_type, final)); | |
1055 | if (multiple_of_p (type, new_c, s)) | |
1056 | relaxed_cond = fold_build2 (GT_EXPR, boolean_type_node, | |
1057 | t, final); | |
ad8a330c | 1058 | } |
1059 | } | |
1060 | } | |
1061 | else | |
1062 | { | |
1063 | cond = fold_build2 (LE_EXPR, boolean_type_node, iv->base, final); | |
1064 | if (TREE_CODE (type) == INTEGER_TYPE) | |
1065 | { | |
1066 | /* Only when base - step doesn't underflow. */ | |
1067 | t = TYPE_MIN_VALUE (type); | |
1068 | t = fold_build2 (PLUS_EXPR, type, t, iv->step); | |
1069 | t = fold_build2 (LE_EXPR, boolean_type_node, t, iv->base); | |
1070 | if (integer_nonzerop (t)) | |
1071 | { | |
1072 | t = fold_build2 (MINUS_EXPR, type, iv->base, iv->step); | |
b441db40 | 1073 | new_c = fold_build2 (MINUS_EXPR, niter_type, |
1074 | fold_convert (niter_type, final), | |
1075 | fold_convert (niter_type, t)); | |
1076 | if (multiple_of_p (type, new_c, s)) | |
1077 | relaxed_cond = fold_build2 (LT_EXPR, boolean_type_node, | |
1078 | t, final); | |
ad8a330c | 1079 | } |
1080 | } | |
1081 | } | |
1082 | ||
1083 | t = simplify_using_initial_conditions (loop, cond); | |
1084 | if (!t || !integer_onep (t)) | |
1085 | t = simplify_using_initial_conditions (loop, relaxed_cond); | |
1086 | ||
1087 | if (t && integer_onep (t)) | |
1088 | niter->control.no_overflow = true; | |
a94f1e5b | 1089 | } |
1090 | ||
0f81d1b1 | 1091 | /* First the trivial cases -- when the step is 1. */ |
1092 | if (integer_onep (s)) | |
1093 | { | |
1094 | niter->niter = c; | |
1095 | return true; | |
b9d73ea6 | 1096 | } |
ad8a330c | 1097 | if (niter->control.no_overflow && multiple_of_p (type, c, s)) |
1098 | { | |
1099 | niter->niter = fold_build2 (FLOOR_DIV_EXPR, niter_type, c, s); | |
1100 | return true; | |
1101 | } | |
b9d73ea6 | 1102 | |
0f81d1b1 | 1103 | /* Let nsd (step, size of mode) = d. If d does not divide c, the loop |
1104 | is infinite. Otherwise, the number of iterations is | |
1105 | (inverse(s/d) * (c/d)) mod (size of mode/d). */ | |
1106 | bits = num_ending_zeros (s); | |
1107 | bound = build_low_bits_mask (niter_type, | |
1108 | (TYPE_PRECISION (niter_type) | |
e913b5cd | 1109 | - tree_to_uhwi (bits))); |
b9d73ea6 | 1110 | |
0f81d1b1 | 1111 | d = fold_binary_to_constant (LSHIFT_EXPR, niter_type, |
05db596e | 1112 | build_int_cst (niter_type, 1), bits); |
0f81d1b1 | 1113 | s = fold_binary_to_constant (RSHIFT_EXPR, niter_type, s, bits); |
b9d73ea6 | 1114 | |
fffbaaf2 | 1115 | if (!exit_must_be_taken) |
0f81d1b1 | 1116 | { |
fffbaaf2 | 1117 | /* If we cannot assume that the exit is taken eventually, record the |
0f81d1b1 | 1118 | assumptions for divisibility of c. */ |
1119 | assumption = fold_build2 (FLOOR_MOD_EXPR, niter_type, c, d); | |
1120 | assumption = fold_build2 (EQ_EXPR, boolean_type_node, | |
1121 | assumption, build_int_cst (niter_type, 0)); | |
cd743a11 | 1122 | if (!integer_nonzerop (assumption)) |
0f81d1b1 | 1123 | niter->assumptions = fold_build2 (TRUTH_AND_EXPR, boolean_type_node, |
1124 | niter->assumptions, assumption); | |
b9d73ea6 | 1125 | } |
48e1416a | 1126 | |
0f81d1b1 | 1127 | c = fold_build2 (EXACT_DIV_EXPR, niter_type, c, d); |
2ecfcddd | 1128 | if (integer_onep (s)) |
1129 | { | |
1130 | niter->niter = c; | |
1131 | } | |
1132 | else | |
1133 | { | |
1134 | tmp = fold_build2 (MULT_EXPR, niter_type, c, inverse (s, bound)); | |
1135 | niter->niter = fold_build2 (BIT_AND_EXPR, niter_type, tmp, bound); | |
1136 | } | |
0f81d1b1 | 1137 | return true; |
1138 | } | |
b9d73ea6 | 1139 | |
0f81d1b1 | 1140 | /* Checks whether we can determine the final value of the control variable |
1141 | of the loop with ending condition IV0 < IV1 (computed in TYPE). | |
1142 | DELTA is the difference IV1->base - IV0->base, STEP is the absolute value | |
1143 | of the step. The assumptions necessary to ensure that the computation | |
1144 | of the final value does not overflow are recorded in NITER. If we | |
1145 | find the final value, we adjust DELTA and return TRUE. Otherwise | |
a463eaea | 1146 | we return false. BNDS bounds the value of IV1->base - IV0->base, |
fffbaaf2 | 1147 | and will be updated by the same amount as DELTA. EXIT_MUST_BE_TAKEN is |
1148 | true if we know that the exit must be taken eventually. */ | |
0f81d1b1 | 1149 | |
1150 | static bool | |
1151 | number_of_iterations_lt_to_ne (tree type, affine_iv *iv0, affine_iv *iv1, | |
1152 | struct tree_niter_desc *niter, | |
a463eaea | 1153 | tree *delta, tree step, |
fffbaaf2 | 1154 | bool exit_must_be_taken, bounds *bnds) |
0f81d1b1 | 1155 | { |
1156 | tree niter_type = TREE_TYPE (step); | |
1157 | tree mod = fold_build2 (FLOOR_MOD_EXPR, niter_type, *delta, step); | |
1158 | tree tmod; | |
e38da69c | 1159 | mpz_t mmod; |
1160 | tree assumption = boolean_true_node, bound, noloop; | |
1161 | bool ret = false, fv_comp_no_overflow; | |
1162 | tree type1 = type; | |
1163 | if (POINTER_TYPE_P (type)) | |
1164 | type1 = sizetype; | |
0f81d1b1 | 1165 | |
1166 | if (TREE_CODE (mod) != INTEGER_CST) | |
1167 | return false; | |
cd743a11 | 1168 | if (integer_nonzerop (mod)) |
0f81d1b1 | 1169 | mod = fold_build2 (MINUS_EXPR, niter_type, step, mod); |
0de36bdb | 1170 | tmod = fold_convert (type1, mod); |
0f81d1b1 | 1171 | |
e38da69c | 1172 | mpz_init (mmod); |
e3d0f65c | 1173 | wi::to_mpz (wi::to_wide (mod), mmod, UNSIGNED); |
e38da69c | 1174 | mpz_neg (mmod, mmod); |
1175 | ||
fffbaaf2 | 1176 | /* If the induction variable does not overflow and the exit is taken, |
e38da69c | 1177 | then the computation of the final value does not overflow. This is |
1178 | also obviously the case if the new final value is equal to the | |
1179 | current one. Finally, we postulate this for pointer type variables, | |
1180 | as the code cannot rely on the object to that the pointer points being | |
1181 | placed at the end of the address space (and more pragmatically, | |
1182 | TYPE_{MIN,MAX}_VALUE is not defined for pointers). */ | |
1183 | if (integer_zerop (mod) || POINTER_TYPE_P (type)) | |
1184 | fv_comp_no_overflow = true; | |
1185 | else if (!exit_must_be_taken) | |
1186 | fv_comp_no_overflow = false; | |
1187 | else | |
1188 | fv_comp_no_overflow = | |
1189 | (iv0->no_overflow && integer_nonzerop (iv0->step)) | |
1190 | || (iv1->no_overflow && integer_nonzerop (iv1->step)); | |
1191 | ||
1192 | if (integer_nonzerop (iv0->step)) | |
b9d73ea6 | 1193 | { |
e38da69c | 1194 | /* The final value of the iv is iv1->base + MOD, assuming that this |
1195 | computation does not overflow, and that | |
1196 | iv0->base <= iv1->base + MOD. */ | |
1197 | if (!fv_comp_no_overflow) | |
0f81d1b1 | 1198 | { |
1095b81d | 1199 | bound = fold_build2 (MINUS_EXPR, type1, |
0de36bdb | 1200 | TYPE_MAX_VALUE (type1), tmod); |
0f81d1b1 | 1201 | assumption = fold_build2 (LE_EXPR, boolean_type_node, |
1202 | iv1->base, bound); | |
e38da69c | 1203 | if (integer_zerop (assumption)) |
1204 | goto end; | |
0f81d1b1 | 1205 | } |
e38da69c | 1206 | if (mpz_cmp (mmod, bnds->below) < 0) |
1207 | noloop = boolean_false_node; | |
1208 | else if (POINTER_TYPE_P (type)) | |
1209 | noloop = fold_build2 (GT_EXPR, boolean_type_node, | |
1210 | iv0->base, | |
1211 | fold_build_pointer_plus (iv1->base, tmod)); | |
a463eaea | 1212 | else |
e38da69c | 1213 | noloop = fold_build2 (GT_EXPR, boolean_type_node, |
1214 | iv0->base, | |
1215 | fold_build2 (PLUS_EXPR, type1, | |
1216 | iv1->base, tmod)); | |
1217 | } | |
1218 | else | |
1219 | { | |
1220 | /* The final value of the iv is iv0->base - MOD, assuming that this | |
1221 | computation does not overflow, and that | |
1222 | iv0->base - MOD <= iv1->base. */ | |
1223 | if (!fv_comp_no_overflow) | |
0f81d1b1 | 1224 | { |
0de36bdb | 1225 | bound = fold_build2 (PLUS_EXPR, type1, |
1226 | TYPE_MIN_VALUE (type1), tmod); | |
0f81d1b1 | 1227 | assumption = fold_build2 (GE_EXPR, boolean_type_node, |
1228 | iv0->base, bound); | |
e38da69c | 1229 | if (integer_zerop (assumption)) |
1230 | goto end; | |
0f81d1b1 | 1231 | } |
e38da69c | 1232 | if (mpz_cmp (mmod, bnds->below) < 0) |
1233 | noloop = boolean_false_node; | |
1234 | else if (POINTER_TYPE_P (type)) | |
1235 | noloop = fold_build2 (GT_EXPR, boolean_type_node, | |
1236 | fold_build_pointer_plus (iv0->base, | |
1237 | fold_build1 (NEGATE_EXPR, | |
1238 | type1, tmod)), | |
1239 | iv1->base); | |
1240 | else | |
1241 | noloop = fold_build2 (GT_EXPR, boolean_type_node, | |
1242 | fold_build2 (MINUS_EXPR, type1, | |
1243 | iv0->base, tmod), | |
1244 | iv1->base); | |
b9d73ea6 | 1245 | } |
1246 | ||
e38da69c | 1247 | if (!integer_nonzerop (assumption)) |
1248 | niter->assumptions = fold_build2 (TRUTH_AND_EXPR, boolean_type_node, | |
1249 | niter->assumptions, | |
1250 | assumption); | |
1251 | if (!integer_zerop (noloop)) | |
1252 | niter->may_be_zero = fold_build2 (TRUTH_OR_EXPR, boolean_type_node, | |
1253 | niter->may_be_zero, | |
1254 | noloop); | |
5de9d3ed | 1255 | bounds_add (bnds, wi::to_widest (mod), type); |
0f81d1b1 | 1256 | *delta = fold_build2 (PLUS_EXPR, niter_type, *delta, mod); |
e38da69c | 1257 | |
1258 | ret = true; | |
1259 | end: | |
1260 | mpz_clear (mmod); | |
1261 | return ret; | |
0f81d1b1 | 1262 | } |
1263 | ||
1264 | /* Add assertions to NITER that ensure that the control variable of the loop | |
1265 | with ending condition IV0 < IV1 does not overflow. Types of IV0 and IV1 | |
1266 | are TYPE. Returns false if we can prove that there is an overflow, true | |
1267 | otherwise. STEP is the absolute value of the step. */ | |
b9d73ea6 | 1268 | |
0f81d1b1 | 1269 | static bool |
1270 | assert_no_overflow_lt (tree type, affine_iv *iv0, affine_iv *iv1, | |
1271 | struct tree_niter_desc *niter, tree step) | |
1272 | { | |
1273 | tree bound, d, assumption, diff; | |
1274 | tree niter_type = TREE_TYPE (step); | |
1275 | ||
7a973feb | 1276 | if (integer_nonzerop (iv0->step)) |
b9d73ea6 | 1277 | { |
0f81d1b1 | 1278 | /* for (i = iv0->base; i < iv1->base; i += iv0->step) */ |
1279 | if (iv0->no_overflow) | |
1280 | return true; | |
1281 | ||
1282 | /* If iv0->base is a constant, we can determine the last value before | |
1283 | overflow precisely; otherwise we conservatively assume | |
1284 | MAX - STEP + 1. */ | |
1285 | ||
1286 | if (TREE_CODE (iv0->base) == INTEGER_CST) | |
b9d73ea6 | 1287 | { |
0f81d1b1 | 1288 | d = fold_build2 (MINUS_EXPR, niter_type, |
1289 | fold_convert (niter_type, TYPE_MAX_VALUE (type)), | |
1290 | fold_convert (niter_type, iv0->base)); | |
1291 | diff = fold_build2 (FLOOR_MOD_EXPR, niter_type, d, step); | |
b9d73ea6 | 1292 | } |
1293 | else | |
0f81d1b1 | 1294 | diff = fold_build2 (MINUS_EXPR, niter_type, step, |
05db596e | 1295 | build_int_cst (niter_type, 1)); |
0f81d1b1 | 1296 | bound = fold_build2 (MINUS_EXPR, type, |
1297 | TYPE_MAX_VALUE (type), fold_convert (type, diff)); | |
1298 | assumption = fold_build2 (LE_EXPR, boolean_type_node, | |
1299 | iv1->base, bound); | |
1300 | } | |
1301 | else | |
1302 | { | |
1303 | /* for (i = iv1->base; i > iv0->base; i += iv1->step) */ | |
1304 | if (iv1->no_overflow) | |
1305 | return true; | |
1306 | ||
1307 | if (TREE_CODE (iv1->base) == INTEGER_CST) | |
b9d73ea6 | 1308 | { |
0f81d1b1 | 1309 | d = fold_build2 (MINUS_EXPR, niter_type, |
1310 | fold_convert (niter_type, iv1->base), | |
1311 | fold_convert (niter_type, TYPE_MIN_VALUE (type))); | |
1312 | diff = fold_build2 (FLOOR_MOD_EXPR, niter_type, d, step); | |
b9d73ea6 | 1313 | } |
0f81d1b1 | 1314 | else |
1315 | diff = fold_build2 (MINUS_EXPR, niter_type, step, | |
05db596e | 1316 | build_int_cst (niter_type, 1)); |
0f81d1b1 | 1317 | bound = fold_build2 (PLUS_EXPR, type, |
1318 | TYPE_MIN_VALUE (type), fold_convert (type, diff)); | |
1319 | assumption = fold_build2 (GE_EXPR, boolean_type_node, | |
1320 | iv0->base, bound); | |
b9d73ea6 | 1321 | } |
1322 | ||
cd743a11 | 1323 | if (integer_zerop (assumption)) |
0f81d1b1 | 1324 | return false; |
cd743a11 | 1325 | if (!integer_nonzerop (assumption)) |
0f81d1b1 | 1326 | niter->assumptions = fold_build2 (TRUTH_AND_EXPR, boolean_type_node, |
1327 | niter->assumptions, assumption); | |
48e1416a | 1328 | |
0f81d1b1 | 1329 | iv0->no_overflow = true; |
1330 | iv1->no_overflow = true; | |
1331 | return true; | |
1332 | } | |
b9d73ea6 | 1333 | |
0f81d1b1 | 1334 | /* Add an assumption to NITER that a loop whose ending condition |
a463eaea | 1335 | is IV0 < IV1 rolls. TYPE is the type of the control iv. BNDS |
1336 | bounds the value of IV1->base - IV0->base. */ | |
0f81d1b1 | 1337 | |
1338 | static void | |
1339 | assert_loop_rolls_lt (tree type, affine_iv *iv0, affine_iv *iv1, | |
a463eaea | 1340 | struct tree_niter_desc *niter, bounds *bnds) |
0f81d1b1 | 1341 | { |
1342 | tree assumption = boolean_true_node, bound, diff; | |
0de36bdb | 1343 | tree mbz, mbzl, mbzr, type1; |
a463eaea | 1344 | bool rolls_p, no_overflow_p; |
5de9d3ed | 1345 | widest_int dstep; |
a463eaea | 1346 | mpz_t mstep, max; |
1347 | ||
1348 | /* We are going to compute the number of iterations as | |
1349 | (iv1->base - iv0->base + step - 1) / step, computed in the unsigned | |
48e1416a | 1350 | variant of TYPE. This formula only works if |
1351 | ||
a463eaea | 1352 | -step + 1 <= (iv1->base - iv0->base) <= MAX - step + 1 |
48e1416a | 1353 | |
a463eaea | 1354 | (where MAX is the maximum value of the unsigned variant of TYPE, and |
d3746d81 | 1355 | the computations in this formula are performed in full precision, |
1356 | i.e., without overflows). | |
a463eaea | 1357 | |
1358 | Usually, for loops with exit condition iv0->base + step * i < iv1->base, | |
d3746d81 | 1359 | we have a condition of the form iv0->base - step < iv1->base before the loop, |
a463eaea | 1360 | and for loops iv0->base < iv1->base - step * i the condition |
1361 | iv0->base < iv1->base + step, due to loop header copying, which enable us | |
1362 | to prove the lower bound. | |
48e1416a | 1363 | |
a463eaea | 1364 | The upper bound is more complicated. Unless the expressions for initial |
1365 | and final value themselves contain enough information, we usually cannot | |
1366 | derive it from the context. */ | |
1367 | ||
1368 | /* First check whether the answer does not follow from the bounds we gathered | |
1369 | before. */ | |
1370 | if (integer_nonzerop (iv0->step)) | |
5de9d3ed | 1371 | dstep = wi::to_widest (iv0->step); |
a463eaea | 1372 | else |
1373 | { | |
5de9d3ed | 1374 | dstep = wi::sext (wi::to_widest (iv1->step), TYPE_PRECISION (type)); |
cf8f0e63 | 1375 | dstep = -dstep; |
a463eaea | 1376 | } |
1377 | ||
1378 | mpz_init (mstep); | |
796b6678 | 1379 | wi::to_mpz (dstep, mstep, UNSIGNED); |
a463eaea | 1380 | mpz_neg (mstep, mstep); |
1381 | mpz_add_ui (mstep, mstep, 1); | |
1382 | ||
1383 | rolls_p = mpz_cmp (mstep, bnds->below) <= 0; | |
1384 | ||
1385 | mpz_init (max); | |
796b6678 | 1386 | wi::to_mpz (wi::minus_one (TYPE_PRECISION (type)), max, UNSIGNED); |
a463eaea | 1387 | mpz_add (max, max, mstep); |
1388 | no_overflow_p = (mpz_cmp (bnds->up, max) <= 0 | |
1389 | /* For pointers, only values lying inside a single object | |
1390 | can be compared or manipulated by pointer arithmetics. | |
1391 | Gcc in general does not allow or handle objects larger | |
1392 | than half of the address space, hence the upper bound | |
1393 | is satisfied for pointers. */ | |
1394 | || POINTER_TYPE_P (type)); | |
1395 | mpz_clear (mstep); | |
1396 | mpz_clear (max); | |
1397 | ||
1398 | if (rolls_p && no_overflow_p) | |
1399 | return; | |
48e1416a | 1400 | |
0de36bdb | 1401 | type1 = type; |
1402 | if (POINTER_TYPE_P (type)) | |
1403 | type1 = sizetype; | |
a463eaea | 1404 | |
1405 | /* Now the hard part; we must formulate the assumption(s) as expressions, and | |
1406 | we must be careful not to introduce overflow. */ | |
0f81d1b1 | 1407 | |
7a973feb | 1408 | if (integer_nonzerop (iv0->step)) |
b9d73ea6 | 1409 | { |
0de36bdb | 1410 | diff = fold_build2 (MINUS_EXPR, type1, |
1411 | iv0->step, build_int_cst (type1, 1)); | |
b9d73ea6 | 1412 | |
0f81d1b1 | 1413 | /* We need to know that iv0->base >= MIN + iv0->step - 1. Since |
1414 | 0 address never belongs to any object, we can assume this for | |
1415 | pointers. */ | |
1416 | if (!POINTER_TYPE_P (type)) | |
b9d73ea6 | 1417 | { |
0de36bdb | 1418 | bound = fold_build2 (PLUS_EXPR, type1, |
0f81d1b1 | 1419 | TYPE_MIN_VALUE (type), diff); |
1420 | assumption = fold_build2 (GE_EXPR, boolean_type_node, | |
1421 | iv0->base, bound); | |
b9d73ea6 | 1422 | } |
1423 | ||
0f81d1b1 | 1424 | /* And then we can compute iv0->base - diff, and compare it with |
48e1416a | 1425 | iv1->base. */ |
1426 | mbzl = fold_build2 (MINUS_EXPR, type1, | |
e88bb328 | 1427 | fold_convert (type1, iv0->base), diff); |
1428 | mbzr = fold_convert (type1, iv1->base); | |
b9d73ea6 | 1429 | } |
0f81d1b1 | 1430 | else |
b9d73ea6 | 1431 | { |
0de36bdb | 1432 | diff = fold_build2 (PLUS_EXPR, type1, |
1433 | iv1->step, build_int_cst (type1, 1)); | |
0f81d1b1 | 1434 | |
1435 | if (!POINTER_TYPE_P (type)) | |
b9d73ea6 | 1436 | { |
0de36bdb | 1437 | bound = fold_build2 (PLUS_EXPR, type1, |
0f81d1b1 | 1438 | TYPE_MAX_VALUE (type), diff); |
1439 | assumption = fold_build2 (LE_EXPR, boolean_type_node, | |
1440 | iv1->base, bound); | |
b9d73ea6 | 1441 | } |
1442 | ||
e88bb328 | 1443 | mbzl = fold_convert (type1, iv0->base); |
1444 | mbzr = fold_build2 (MINUS_EXPR, type1, | |
1445 | fold_convert (type1, iv1->base), diff); | |
0f81d1b1 | 1446 | } |
b9d73ea6 | 1447 | |
cd743a11 | 1448 | if (!integer_nonzerop (assumption)) |
0f81d1b1 | 1449 | niter->assumptions = fold_build2 (TRUTH_AND_EXPR, boolean_type_node, |
1450 | niter->assumptions, assumption); | |
a463eaea | 1451 | if (!rolls_p) |
1452 | { | |
1453 | mbz = fold_build2 (GT_EXPR, boolean_type_node, mbzl, mbzr); | |
1454 | niter->may_be_zero = fold_build2 (TRUTH_OR_EXPR, boolean_type_node, | |
1455 | niter->may_be_zero, mbz); | |
1456 | } | |
0f81d1b1 | 1457 | } |
b9d73ea6 | 1458 | |
0f81d1b1 | 1459 | /* Determines number of iterations of loop whose ending condition |
1460 | is IV0 < IV1. TYPE is the type of the iv. The number of | |
a463eaea | 1461 | iterations is stored to NITER. BNDS bounds the difference |
fffbaaf2 | 1462 | IV1->base - IV0->base. EXIT_MUST_BE_TAKEN is true if we know |
1463 | that the exit must be taken eventually. */ | |
0f81d1b1 | 1464 | |
1465 | static bool | |
a94f1e5b | 1466 | number_of_iterations_lt (struct loop *loop, tree type, affine_iv *iv0, |
1467 | affine_iv *iv1, struct tree_niter_desc *niter, | |
fffbaaf2 | 1468 | bool exit_must_be_taken, bounds *bnds) |
0f81d1b1 | 1469 | { |
1470 | tree niter_type = unsigned_type_for (type); | |
1471 | tree delta, step, s; | |
a463eaea | 1472 | mpz_t mstep, tmp; |
0f81d1b1 | 1473 | |
7a973feb | 1474 | if (integer_nonzerop (iv0->step)) |
b30560de | 1475 | { |
1476 | niter->control = *iv0; | |
1477 | niter->cmp = LT_EXPR; | |
1478 | niter->bound = iv1->base; | |
1479 | } | |
1480 | else | |
1481 | { | |
1482 | niter->control = *iv1; | |
1483 | niter->cmp = GT_EXPR; | |
1484 | niter->bound = iv0->base; | |
1485 | } | |
1486 | ||
0f81d1b1 | 1487 | delta = fold_build2 (MINUS_EXPR, niter_type, |
1488 | fold_convert (niter_type, iv1->base), | |
1489 | fold_convert (niter_type, iv0->base)); | |
1490 | ||
1491 | /* First handle the special case that the step is +-1. */ | |
7a973feb | 1492 | if ((integer_onep (iv0->step) && integer_zerop (iv1->step)) |
1493 | || (integer_all_onesp (iv1->step) && integer_zerop (iv0->step))) | |
0f81d1b1 | 1494 | { |
1495 | /* for (i = iv0->base; i < iv1->base; i++) | |
1496 | ||
1497 | or | |
bb445479 | 1498 | |
0f81d1b1 | 1499 | for (i = iv1->base; i > iv0->base; i--). |
48e1416a | 1500 | |
0f81d1b1 | 1501 | In both cases # of iterations is iv1->base - iv0->base, assuming that |
a463eaea | 1502 | iv1->base >= iv0->base. |
1503 | ||
1504 | First try to derive a lower bound on the value of | |
1505 | iv1->base - iv0->base, computed in full precision. If the difference | |
1506 | is nonnegative, we are done, otherwise we must record the | |
1507 | condition. */ | |
1508 | ||
1509 | if (mpz_sgn (bnds->below) < 0) | |
1510 | niter->may_be_zero = fold_build2 (LT_EXPR, boolean_type_node, | |
1511 | iv1->base, iv0->base); | |
0f81d1b1 | 1512 | niter->niter = delta; |
5de9d3ed | 1513 | niter->max = widest_int::from (wi::from_mpz (niter_type, bnds->up, false), |
1514 | TYPE_SIGN (niter_type)); | |
05032b10 | 1515 | niter->control.no_overflow = true; |
0f81d1b1 | 1516 | return true; |
b9d73ea6 | 1517 | } |
0f81d1b1 | 1518 | |
7a973feb | 1519 | if (integer_nonzerop (iv0->step)) |
0f81d1b1 | 1520 | step = fold_convert (niter_type, iv0->step); |
b9d73ea6 | 1521 | else |
0f81d1b1 | 1522 | step = fold_convert (niter_type, |
1523 | fold_build1 (NEGATE_EXPR, type, iv1->step)); | |
1524 | ||
1525 | /* If we can determine the final value of the control iv exactly, we can | |
1526 | transform the condition to != comparison. In particular, this will be | |
1527 | the case if DELTA is constant. */ | |
a463eaea | 1528 | if (number_of_iterations_lt_to_ne (type, iv0, iv1, niter, &delta, step, |
fffbaaf2 | 1529 | exit_must_be_taken, bnds)) |
b9d73ea6 | 1530 | { |
0f81d1b1 | 1531 | affine_iv zps; |
1532 | ||
05db596e | 1533 | zps.base = build_int_cst (niter_type, 0); |
0f81d1b1 | 1534 | zps.step = step; |
1535 | /* number_of_iterations_lt_to_ne will add assumptions that ensure that | |
1536 | zps does not overflow. */ | |
1537 | zps.no_overflow = true; | |
1538 | ||
a94f1e5b | 1539 | return number_of_iterations_ne (loop, type, &zps, |
1540 | delta, niter, true, bnds); | |
b9d73ea6 | 1541 | } |
1542 | ||
0f81d1b1 | 1543 | /* Make sure that the control iv does not overflow. */ |
1544 | if (!assert_no_overflow_lt (type, iv0, iv1, niter, step)) | |
1545 | return false; | |
b9d73ea6 | 1546 | |
0f81d1b1 | 1547 | /* We determine the number of iterations as (delta + step - 1) / step. For |
1548 | this to work, we must know that iv1->base >= iv0->base - step + 1, | |
1549 | otherwise the loop does not roll. */ | |
a463eaea | 1550 | assert_loop_rolls_lt (type, iv0, iv1, niter, bnds); |
0f81d1b1 | 1551 | |
1552 | s = fold_build2 (MINUS_EXPR, niter_type, | |
05db596e | 1553 | step, build_int_cst (niter_type, 1)); |
0f81d1b1 | 1554 | delta = fold_build2 (PLUS_EXPR, niter_type, delta, s); |
1555 | niter->niter = fold_build2 (FLOOR_DIV_EXPR, niter_type, delta, step); | |
a463eaea | 1556 | |
1557 | mpz_init (mstep); | |
1558 | mpz_init (tmp); | |
e3d0f65c | 1559 | wi::to_mpz (wi::to_wide (step), mstep, UNSIGNED); |
a463eaea | 1560 | mpz_add (tmp, bnds->up, mstep); |
1561 | mpz_sub_ui (tmp, tmp, 1); | |
1562 | mpz_fdiv_q (tmp, tmp, mstep); | |
5de9d3ed | 1563 | niter->max = widest_int::from (wi::from_mpz (niter_type, tmp, false), |
1564 | TYPE_SIGN (niter_type)); | |
a463eaea | 1565 | mpz_clear (mstep); |
1566 | mpz_clear (tmp); | |
1567 | ||
0f81d1b1 | 1568 | return true; |
b9d73ea6 | 1569 | } |
1570 | ||
0f81d1b1 | 1571 | /* Determines number of iterations of loop whose ending condition |
1572 | is IV0 <= IV1. TYPE is the type of the iv. The number of | |
fffbaaf2 | 1573 | iterations is stored to NITER. EXIT_MUST_BE_TAKEN is true if |
0840f4fe | 1574 | we know that this condition must eventually become false (we derived this |
0f81d1b1 | 1575 | earlier, and possibly set NITER->assumptions to make sure this |
a463eaea | 1576 | is the case). BNDS bounds the difference IV1->base - IV0->base. */ |
0f81d1b1 | 1577 | |
1578 | static bool | |
a94f1e5b | 1579 | number_of_iterations_le (struct loop *loop, tree type, affine_iv *iv0, |
1580 | affine_iv *iv1, struct tree_niter_desc *niter, | |
1581 | bool exit_must_be_taken, bounds *bnds) | |
0f81d1b1 | 1582 | { |
1583 | tree assumption; | |
0de36bdb | 1584 | tree type1 = type; |
1585 | if (POINTER_TYPE_P (type)) | |
1586 | type1 = sizetype; | |
0f81d1b1 | 1587 | |
1588 | /* Say that IV0 is the control variable. Then IV0 <= IV1 iff | |
1589 | IV0 < IV1 + 1, assuming that IV1 is not equal to the greatest | |
1590 | value of the type. This we must know anyway, since if it is | |
fffbaaf2 | 1591 | equal to this value, the loop rolls forever. We do not check |
48e1416a | 1592 | this condition for pointer type ivs, as the code cannot rely on |
fffbaaf2 | 1593 | the object to that the pointer points being placed at the end of |
1594 | the address space (and more pragmatically, TYPE_{MIN,MAX}_VALUE is | |
1595 | not defined for pointers). */ | |
0f81d1b1 | 1596 | |
fffbaaf2 | 1597 | if (!exit_must_be_taken && !POINTER_TYPE_P (type)) |
0f81d1b1 | 1598 | { |
7a973feb | 1599 | if (integer_nonzerop (iv0->step)) |
0f81d1b1 | 1600 | assumption = fold_build2 (NE_EXPR, boolean_type_node, |
1095b81d | 1601 | iv1->base, TYPE_MAX_VALUE (type)); |
0f81d1b1 | 1602 | else |
1603 | assumption = fold_build2 (NE_EXPR, boolean_type_node, | |
1095b81d | 1604 | iv0->base, TYPE_MIN_VALUE (type)); |
0f81d1b1 | 1605 | |
cd743a11 | 1606 | if (integer_zerop (assumption)) |
0f81d1b1 | 1607 | return false; |
cd743a11 | 1608 | if (!integer_nonzerop (assumption)) |
0f81d1b1 | 1609 | niter->assumptions = fold_build2 (TRUTH_AND_EXPR, boolean_type_node, |
1610 | niter->assumptions, assumption); | |
1611 | } | |
1612 | ||
7a973feb | 1613 | if (integer_nonzerop (iv0->step)) |
1095b81d | 1614 | { |
1615 | if (POINTER_TYPE_P (type)) | |
2cc66f2a | 1616 | iv1->base = fold_build_pointer_plus_hwi (iv1->base, 1); |
1095b81d | 1617 | else |
1618 | iv1->base = fold_build2 (PLUS_EXPR, type1, iv1->base, | |
1619 | build_int_cst (type1, 1)); | |
1620 | } | |
1621 | else if (POINTER_TYPE_P (type)) | |
2cc66f2a | 1622 | iv0->base = fold_build_pointer_plus_hwi (iv0->base, -1); |
0f81d1b1 | 1623 | else |
0de36bdb | 1624 | iv0->base = fold_build2 (MINUS_EXPR, type1, |
1625 | iv0->base, build_int_cst (type1, 1)); | |
a463eaea | 1626 | |
e913b5cd | 1627 | bounds_add (bnds, 1, type1); |
a463eaea | 1628 | |
a94f1e5b | 1629 | return number_of_iterations_lt (loop, type, iv0, iv1, niter, exit_must_be_taken, |
fffbaaf2 | 1630 | bnds); |
a463eaea | 1631 | } |
1632 | ||
1633 | /* Dumps description of affine induction variable IV to FILE. */ | |
1634 | ||
1635 | static void | |
1636 | dump_affine_iv (FILE *file, affine_iv *iv) | |
1637 | { | |
1638 | if (!integer_zerop (iv->step)) | |
1639 | fprintf (file, "["); | |
1640 | ||
1641 | print_generic_expr (dump_file, iv->base, TDF_SLIM); | |
1642 | ||
1643 | if (!integer_zerop (iv->step)) | |
1644 | { | |
1645 | fprintf (file, ", + , "); | |
1646 | print_generic_expr (dump_file, iv->step, TDF_SLIM); | |
1647 | fprintf (file, "]%s", iv->no_overflow ? "(no_overflow)" : ""); | |
1648 | } | |
0f81d1b1 | 1649 | } |
b9d73ea6 | 1650 | |
75ae0e7f | 1651 | /* Given exit condition IV0 CODE IV1 in TYPE, this function adjusts |
1652 | the condition for loop-until-wrap cases. For example: | |
1653 | (unsigned){8, -1}_loop < 10 => {0, 1} != 9 | |
1654 | 10 < (unsigned){0, max - 7}_loop => {0, 1} != 8 | |
1655 | Return true if condition is successfully adjusted. */ | |
1656 | ||
1657 | static bool | |
1658 | adjust_cond_for_loop_until_wrap (tree type, affine_iv *iv0, tree_code *code, | |
1659 | affine_iv *iv1) | |
1660 | { | |
1661 | /* Only support simple cases for the moment. */ | |
1662 | if (TREE_CODE (iv0->base) != INTEGER_CST | |
1663 | || TREE_CODE (iv1->base) != INTEGER_CST) | |
1664 | return false; | |
1665 | ||
1666 | tree niter_type = unsigned_type_for (type), high, low; | |
1667 | /* Case: i-- < 10. */ | |
1668 | if (integer_zerop (iv1->step)) | |
1669 | { | |
1670 | /* TODO: Should handle case in which abs(step) != 1. */ | |
1671 | if (!integer_minus_onep (iv0->step)) | |
1672 | return false; | |
1673 | /* Give up on infinite loop. */ | |
1674 | if (*code == LE_EXPR | |
1675 | && tree_int_cst_equal (iv1->base, TYPE_MAX_VALUE (type))) | |
1676 | return false; | |
1677 | high = fold_build2 (PLUS_EXPR, niter_type, | |
1678 | fold_convert (niter_type, iv0->base), | |
1679 | build_int_cst (niter_type, 1)); | |
1680 | low = fold_convert (niter_type, TYPE_MIN_VALUE (type)); | |
1681 | } | |
1682 | else if (integer_zerop (iv0->step)) | |
1683 | { | |
1684 | /* TODO: Should handle case in which abs(step) != 1. */ | |
1685 | if (!integer_onep (iv1->step)) | |
1686 | return false; | |
1687 | /* Give up on infinite loop. */ | |
1688 | if (*code == LE_EXPR | |
1689 | && tree_int_cst_equal (iv0->base, TYPE_MIN_VALUE (type))) | |
1690 | return false; | |
1691 | high = fold_convert (niter_type, TYPE_MAX_VALUE (type)); | |
1692 | low = fold_build2 (MINUS_EXPR, niter_type, | |
1693 | fold_convert (niter_type, iv1->base), | |
1694 | build_int_cst (niter_type, 1)); | |
1695 | } | |
1696 | else | |
1697 | gcc_unreachable (); | |
1698 | ||
1699 | iv0->base = low; | |
1700 | iv0->step = fold_convert (niter_type, integer_one_node); | |
1701 | iv1->base = high; | |
1702 | iv1->step = build_int_cst (niter_type, 0); | |
1703 | *code = NE_EXPR; | |
1704 | return true; | |
1705 | } | |
1706 | ||
0f81d1b1 | 1707 | /* Determine the number of iterations according to condition (for staying |
1708 | inside loop) which compares two induction variables using comparison | |
1709 | operator CODE. The induction variable on left side of the comparison | |
1710 | is IV0, the right-hand side is IV1. Both induction variables must have | |
1711 | type TYPE, which must be an integer or pointer type. The steps of the | |
1712 | ivs must be constants (or NULL_TREE, which is interpreted as constant zero). | |
0840f4fe | 1713 | |
a463eaea | 1714 | LOOP is the loop whose number of iterations we are determining. |
1715 | ||
0840f4fe | 1716 | ONLY_EXIT is true if we are sure this is the only way the loop could be |
1717 | exited (including possibly non-returning function calls, exceptions, etc.) | |
1718 | -- in this case we can use the information whether the control induction | |
1719 | variables can overflow or not in a more efficient way. | |
48e1416a | 1720 | |
64552007 | 1721 | if EVERY_ITERATION is true, we know the test is executed on every iteration. |
1722 | ||
0f81d1b1 | 1723 | The results (number of iterations and assumptions as described in |
f6d8a42a | 1724 | comments at struct tree_niter_desc in tree-ssa-loop.h) are stored to NITER. |
0f81d1b1 | 1725 | Returns false if it fails to determine number of iterations, true if it |
1726 | was determined (possibly with some assumptions). */ | |
ae288d12 | 1727 | |
1728 | static bool | |
a463eaea | 1729 | number_of_iterations_cond (struct loop *loop, |
1730 | tree type, affine_iv *iv0, enum tree_code code, | |
0840f4fe | 1731 | affine_iv *iv1, struct tree_niter_desc *niter, |
64552007 | 1732 | bool only_exit, bool every_iteration) |
b9d73ea6 | 1733 | { |
fffbaaf2 | 1734 | bool exit_must_be_taken = false, ret; |
a463eaea | 1735 | bounds bnds; |
0f81d1b1 | 1736 | |
64552007 | 1737 | /* If the test is not executed every iteration, wrapping may make the test |
1738 | to pass again. | |
1739 | TODO: the overflow case can be still used as unreliable estimate of upper | |
1740 | bound. But we have no API to pass it down to number of iterations code | |
1741 | and, at present, it will not use it anyway. */ | |
1742 | if (!every_iteration | |
1743 | && (!iv0->no_overflow || !iv1->no_overflow | |
1744 | || code == NE_EXPR || code == EQ_EXPR)) | |
1745 | return false; | |
1746 | ||
0f81d1b1 | 1747 | /* The meaning of these assumptions is this: |
1748 | if !assumptions | |
1749 | then the rest of information does not have to be valid | |
1750 | if may_be_zero then the loop does not roll, even if | |
1751 | niter != 0. */ | |
1752 | niter->assumptions = boolean_true_node; | |
1753 | niter->may_be_zero = boolean_false_node; | |
1754 | niter->niter = NULL_TREE; | |
e913b5cd | 1755 | niter->max = 0; |
b30560de | 1756 | niter->bound = NULL_TREE; |
1757 | niter->cmp = ERROR_MARK; | |
1758 | ||
0f81d1b1 | 1759 | /* Make < comparison from > ones, and for NE_EXPR comparisons, ensure that |
1760 | the control variable is on lhs. */ | |
1761 | if (code == GE_EXPR || code == GT_EXPR | |
7a973feb | 1762 | || (code == NE_EXPR && integer_zerop (iv0->step))) |
ae288d12 | 1763 | { |
a4f59596 | 1764 | std::swap (iv0, iv1); |
ae288d12 | 1765 | code = swap_tree_comparison (code); |
1766 | } | |
b9d73ea6 | 1767 | |
0f81d1b1 | 1768 | if (POINTER_TYPE_P (type)) |
b9d73ea6 | 1769 | { |
0f81d1b1 | 1770 | /* Comparison of pointers is undefined unless both iv0 and iv1 point |
1771 | to the same object. If they do, the control variable cannot wrap | |
1772 | (as wrap around the bounds of memory will never return a pointer | |
1773 | that would be guaranteed to point to the same object, even if we | |
fffbaaf2 | 1774 | avoid undefined behavior by casting to size_t and back). */ |
0f81d1b1 | 1775 | iv0->no_overflow = true; |
1776 | iv1->no_overflow = true; | |
1777 | } | |
b9d73ea6 | 1778 | |
fffbaaf2 | 1779 | /* If the control induction variable does not overflow and the only exit |
1780 | from the loop is the one that we analyze, we know it must be taken | |
1781 | eventually. */ | |
1782 | if (only_exit) | |
1783 | { | |
1784 | if (!integer_zerop (iv0->step) && iv0->no_overflow) | |
1785 | exit_must_be_taken = true; | |
1786 | else if (!integer_zerop (iv1->step) && iv1->no_overflow) | |
1787 | exit_must_be_taken = true; | |
1788 | } | |
b9d73ea6 | 1789 | |
b2d4d2bb | 1790 | /* We can handle cases which neither of the sides of the comparison is |
1791 | invariant: | |
1792 | ||
1793 | {iv0.base, iv0.step} cmp_code {iv1.base, iv1.step} | |
1794 | as if: | |
1795 | {iv0.base, iv0.step - iv1.step} cmp_code {iv1.base, 0} | |
1796 | ||
1797 | provided that either below condition is satisfied: | |
1798 | ||
1799 | a) the test is NE_EXPR; | |
0fad202d | 1800 | b) iv0.step - iv1.step is integer and iv0/iv1 don't overflow. |
b2d4d2bb | 1801 | |
1802 | This rarely occurs in practice, but it is simple enough to manage. */ | |
7a973feb | 1803 | if (!integer_zerop (iv0->step) && !integer_zerop (iv1->step)) |
0f81d1b1 | 1804 | { |
a50fd9b7 | 1805 | tree step_type = POINTER_TYPE_P (type) ? sizetype : type; |
b2d4d2bb | 1806 | tree step = fold_binary_to_constant (MINUS_EXPR, step_type, |
1807 | iv0->step, iv1->step); | |
1808 | ||
1809 | /* No need to check sign of the new step since below code takes care | |
1810 | of this well. */ | |
0fad202d | 1811 | if (code != NE_EXPR |
1812 | && (TREE_CODE (step) != INTEGER_CST | |
1813 | || !iv0->no_overflow || !iv1->no_overflow)) | |
0f81d1b1 | 1814 | return false; |
b9d73ea6 | 1815 | |
b2d4d2bb | 1816 | iv0->step = step; |
1817 | if (!POINTER_TYPE_P (type)) | |
1818 | iv0->no_overflow = false; | |
1819 | ||
a50fd9b7 | 1820 | iv1->step = build_int_cst (step_type, 0); |
0f81d1b1 | 1821 | iv1->no_overflow = true; |
1822 | } | |
ae288d12 | 1823 | |
0f81d1b1 | 1824 | /* If the result of the comparison is a constant, the loop is weird. More |
1825 | precise handling would be possible, but the situation is not common enough | |
1826 | to waste time on it. */ | |
7a973feb | 1827 | if (integer_zerop (iv0->step) && integer_zerop (iv1->step)) |
0f81d1b1 | 1828 | return false; |
ae288d12 | 1829 | |
334ec2d8 | 1830 | /* If the loop exits immediately, there is nothing to do. */ |
e4800344 | 1831 | tree tem = fold_binary (code, boolean_type_node, iv0->base, iv1->base); |
1832 | if (tem && integer_zerop (tem)) | |
0f81d1b1 | 1833 | { |
d92b5509 | 1834 | if (!every_iteration) |
1835 | return false; | |
05db596e | 1836 | niter->niter = build_int_cst (unsigned_type_for (type), 0); |
e913b5cd | 1837 | niter->max = 0; |
0f81d1b1 | 1838 | return true; |
1839 | } | |
48e1416a | 1840 | |
75ae0e7f | 1841 | /* Handle special case loops: while (i-- < 10) and while (10 < i++) by |
1842 | adjusting iv0, iv1 and code. */ | |
1843 | if (code != NE_EXPR | |
1844 | && (tree_int_cst_sign_bit (iv0->step) | |
1845 | || (!integer_zerop (iv1->step) | |
1846 | && !tree_int_cst_sign_bit (iv1->step))) | |
1847 | && !adjust_cond_for_loop_until_wrap (type, iv0, &code, iv1)) | |
1848 | return false; | |
1849 | ||
0f81d1b1 | 1850 | /* OK, now we know we have a senseful loop. Handle several cases, depending |
1851 | on what comparison operator is used. */ | |
a463eaea | 1852 | bound_difference (loop, iv1->base, iv0->base, &bnds); |
1853 | ||
1854 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1855 | { | |
1856 | fprintf (dump_file, | |
666137bc | 1857 | "Analyzing # of iterations of loop %d\n", loop->num); |
a463eaea | 1858 | |
1859 | fprintf (dump_file, " exit condition "); | |
1860 | dump_affine_iv (dump_file, iv0); | |
1861 | fprintf (dump_file, " %s ", | |
1862 | code == NE_EXPR ? "!=" | |
1863 | : code == LT_EXPR ? "<" | |
1864 | : "<="); | |
1865 | dump_affine_iv (dump_file, iv1); | |
1866 | fprintf (dump_file, "\n"); | |
1867 | ||
1868 | fprintf (dump_file, " bounds on difference of bases: "); | |
1869 | mpz_out_str (dump_file, 10, bnds.below); | |
1870 | fprintf (dump_file, " ... "); | |
1871 | mpz_out_str (dump_file, 10, bnds.up); | |
1872 | fprintf (dump_file, "\n"); | |
1873 | } | |
1874 | ||
0f81d1b1 | 1875 | switch (code) |
1876 | { | |
1877 | case NE_EXPR: | |
7a973feb | 1878 | gcc_assert (integer_zerop (iv1->step)); |
a94f1e5b | 1879 | ret = number_of_iterations_ne (loop, type, iv0, iv1->base, niter, |
fffbaaf2 | 1880 | exit_must_be_taken, &bnds); |
a463eaea | 1881 | break; |
1882 | ||
0f81d1b1 | 1883 | case LT_EXPR: |
a94f1e5b | 1884 | ret = number_of_iterations_lt (loop, type, iv0, iv1, niter, |
1885 | exit_must_be_taken, &bnds); | |
a463eaea | 1886 | break; |
1887 | ||
0f81d1b1 | 1888 | case LE_EXPR: |
a94f1e5b | 1889 | ret = number_of_iterations_le (loop, type, iv0, iv1, niter, |
1890 | exit_must_be_taken, &bnds); | |
a463eaea | 1891 | break; |
1892 | ||
ae288d12 | 1893 | default: |
1894 | gcc_unreachable (); | |
1895 | } | |
a463eaea | 1896 | |
1897 | mpz_clear (bnds.up); | |
1898 | mpz_clear (bnds.below); | |
1899 | ||
1900 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1901 | { | |
1902 | if (ret) | |
1903 | { | |
1904 | fprintf (dump_file, " result:\n"); | |
1905 | if (!integer_nonzerop (niter->assumptions)) | |
1906 | { | |
1907 | fprintf (dump_file, " under assumptions "); | |
1908 | print_generic_expr (dump_file, niter->assumptions, TDF_SLIM); | |
1909 | fprintf (dump_file, "\n"); | |
1910 | } | |
1911 | ||
1912 | if (!integer_zerop (niter->may_be_zero)) | |
1913 | { | |
1914 | fprintf (dump_file, " zero if "); | |
1915 | print_generic_expr (dump_file, niter->may_be_zero, TDF_SLIM); | |
1916 | fprintf (dump_file, "\n"); | |
1917 | } | |
1918 | ||
1919 | fprintf (dump_file, " # of iterations "); | |
1920 | print_generic_expr (dump_file, niter->niter, TDF_SLIM); | |
1921 | fprintf (dump_file, ", bounded by "); | |
e913b5cd | 1922 | print_decu (niter->max, dump_file); |
a463eaea | 1923 | fprintf (dump_file, "\n"); |
1924 | } | |
1925 | else | |
1926 | fprintf (dump_file, " failed\n\n"); | |
1927 | } | |
1928 | return ret; | |
b9d73ea6 | 1929 | } |
1930 | ||
b51523c4 | 1931 | /* Substitute NEW_TREE for OLD in EXPR and fold the result. |
1932 | If VALUEIZE is non-NULL then OLD and NEW_TREE are ignored and instead | |
1933 | all SSA names are replaced with the result of calling the VALUEIZE | |
1934 | function with the SSA name as argument. */ | |
9a73db25 | 1935 | |
b51523c4 | 1936 | tree |
1937 | simplify_replace_tree (tree expr, tree old, tree new_tree, | |
1938 | tree (*valueize) (tree)) | |
9a73db25 | 1939 | { |
1940 | unsigned i, n; | |
1941 | tree ret = NULL_TREE, e, se; | |
1942 | ||
1943 | if (!expr) | |
1944 | return NULL_TREE; | |
1945 | ||
4b9edb63 | 1946 | /* Do not bother to replace constants. */ |
b51523c4 | 1947 | if (CONSTANT_CLASS_P (expr)) |
4b9edb63 | 1948 | return expr; |
1949 | ||
b51523c4 | 1950 | if (valueize) |
1951 | { | |
1952 | if (TREE_CODE (expr) == SSA_NAME) | |
1953 | { | |
1954 | new_tree = valueize (expr); | |
1955 | if (new_tree != expr) | |
1956 | return new_tree; | |
1957 | } | |
1958 | } | |
1959 | else if (expr == old | |
1960 | || operand_equal_p (expr, old, 0)) | |
f0d6e81c | 1961 | return unshare_expr (new_tree); |
9a73db25 | 1962 | |
75a70cf9 | 1963 | if (!EXPR_P (expr)) |
9a73db25 | 1964 | return expr; |
1965 | ||
c2f47e15 | 1966 | n = TREE_OPERAND_LENGTH (expr); |
9a73db25 | 1967 | for (i = 0; i < n; i++) |
1968 | { | |
1969 | e = TREE_OPERAND (expr, i); | |
b51523c4 | 1970 | se = simplify_replace_tree (e, old, new_tree, valueize); |
9a73db25 | 1971 | if (e == se) |
1972 | continue; | |
1973 | ||
1974 | if (!ret) | |
1975 | ret = copy_node (expr); | |
1976 | ||
1977 | TREE_OPERAND (ret, i) = se; | |
1978 | } | |
1979 | ||
1980 | return (ret ? fold (ret) : expr); | |
1981 | } | |
1982 | ||
1962ceba | 1983 | /* Expand definitions of ssa names in EXPR as long as they are simple |
02acb8f3 | 1984 | enough, and return the new expression. If STOP is specified, stop |
1985 | expanding if EXPR equals to it. */ | |
1962ceba | 1986 | |
c16eb701 | 1987 | static tree |
1988 | expand_simple_operations (tree expr, tree stop, hash_map<tree, tree> &cache) | |
1962ceba | 1989 | { |
1990 | unsigned i, n; | |
75a70cf9 | 1991 | tree ret = NULL_TREE, e, ee, e1; |
73a99144 | 1992 | enum tree_code code; |
42acab1c | 1993 | gimple *stmt; |
73a99144 | 1994 | |
1995 | if (expr == NULL_TREE) | |
1996 | return expr; | |
1962ceba | 1997 | |
1998 | if (is_gimple_min_invariant (expr)) | |
1999 | return expr; | |
2000 | ||
73a99144 | 2001 | code = TREE_CODE (expr); |
1962ceba | 2002 | if (IS_EXPR_CODE_CLASS (TREE_CODE_CLASS (code))) |
2003 | { | |
c2f47e15 | 2004 | n = TREE_OPERAND_LENGTH (expr); |
1962ceba | 2005 | for (i = 0; i < n; i++) |
2006 | { | |
2007 | e = TREE_OPERAND (expr, i); | |
c16eb701 | 2008 | /* SCEV analysis feeds us with a proper expression |
2009 | graph matching the SSA graph. Avoid turning it | |
2010 | into a tree here, thus handle tree sharing | |
2011 | properly. | |
2012 | ??? The SSA walk below still turns the SSA graph | |
2013 | into a tree but until we find a testcase do not | |
2014 | introduce additional tree sharing here. */ | |
2015 | bool existed_p; | |
2016 | tree &cee = cache.get_or_insert (e, &existed_p); | |
2017 | if (existed_p) | |
2018 | ee = cee; | |
2019 | else | |
2020 | { | |
2021 | cee = e; | |
2022 | ee = expand_simple_operations (e, stop, cache); | |
2023 | if (ee != e) | |
2024 | *cache.get (e) = ee; | |
2025 | } | |
1962ceba | 2026 | if (e == ee) |
2027 | continue; | |
2028 | ||
2029 | if (!ret) | |
2030 | ret = copy_node (expr); | |
2031 | ||
2032 | TREE_OPERAND (ret, i) = ee; | |
2033 | } | |
2034 | ||
add6ee5e | 2035 | if (!ret) |
2036 | return expr; | |
2037 | ||
2038 | fold_defer_overflow_warnings (); | |
2039 | ret = fold (ret); | |
2040 | fold_undefer_and_ignore_overflow_warnings (); | |
2041 | return ret; | |
1962ceba | 2042 | } |
2043 | ||
02acb8f3 | 2044 | /* Stop if it's not ssa name or the one we don't want to expand. */ |
2045 | if (TREE_CODE (expr) != SSA_NAME || expr == stop) | |
1962ceba | 2046 | return expr; |
2047 | ||
2048 | stmt = SSA_NAME_DEF_STMT (expr); | |
75a70cf9 | 2049 | if (gimple_code (stmt) == GIMPLE_PHI) |
a463eaea | 2050 | { |
2051 | basic_block src, dest; | |
2052 | ||
75a70cf9 | 2053 | if (gimple_phi_num_args (stmt) != 1) |
a463eaea | 2054 | return expr; |
2055 | e = PHI_ARG_DEF (stmt, 0); | |
2056 | ||
2057 | /* Avoid propagating through loop exit phi nodes, which | |
2058 | could break loop-closed SSA form restrictions. */ | |
75a70cf9 | 2059 | dest = gimple_bb (stmt); |
a463eaea | 2060 | src = single_pred (dest); |
2061 | if (TREE_CODE (e) == SSA_NAME | |
2062 | && src->loop_father != dest->loop_father) | |
2063 | return expr; | |
2064 | ||
c16eb701 | 2065 | return expand_simple_operations (e, stop, cache); |
a463eaea | 2066 | } |
75a70cf9 | 2067 | if (gimple_code (stmt) != GIMPLE_ASSIGN) |
1962ceba | 2068 | return expr; |
2069 | ||
d51283e3 | 2070 | /* Avoid expanding to expressions that contain SSA names that need |
2071 | to take part in abnormal coalescing. */ | |
2072 | ssa_op_iter iter; | |
2073 | FOR_EACH_SSA_TREE_OPERAND (e, stmt, iter, SSA_OP_USE) | |
2074 | if (SSA_NAME_OCCURS_IN_ABNORMAL_PHI (e)) | |
2075 | return expr; | |
2076 | ||
75a70cf9 | 2077 | e = gimple_assign_rhs1 (stmt); |
2078 | code = gimple_assign_rhs_code (stmt); | |
2079 | if (get_gimple_rhs_class (code) == GIMPLE_SINGLE_RHS) | |
2080 | { | |
2081 | if (is_gimple_min_invariant (e)) | |
2082 | return e; | |
2083 | ||
2084 | if (code == SSA_NAME) | |
c16eb701 | 2085 | return expand_simple_operations (e, stop, cache); |
d64ba765 | 2086 | else if (code == ADDR_EXPR) |
2087 | { | |
773078cb | 2088 | poly_int64 offset; |
d64ba765 | 2089 | tree base = get_addr_base_and_unit_offset (TREE_OPERAND (e, 0), |
2090 | &offset); | |
2091 | if (base | |
2092 | && TREE_CODE (base) == MEM_REF) | |
2093 | { | |
c16eb701 | 2094 | ee = expand_simple_operations (TREE_OPERAND (base, 0), stop, |
2095 | cache); | |
d64ba765 | 2096 | return fold_build2 (POINTER_PLUS_EXPR, TREE_TYPE (expr), ee, |
2097 | wide_int_to_tree (sizetype, | |
2098 | mem_ref_offset (base) | |
2099 | + offset)); | |
2100 | } | |
2101 | } | |
75a70cf9 | 2102 | |
2103 | return expr; | |
2104 | } | |
2105 | ||
2106 | switch (code) | |
2107 | { | |
d9659041 | 2108 | CASE_CONVERT: |
75a70cf9 | 2109 | /* Casts are simple. */ |
c16eb701 | 2110 | ee = expand_simple_operations (e, stop, cache); |
75a70cf9 | 2111 | return fold_build1 (code, TREE_TYPE (expr), ee); |
2112 | ||
2113 | case PLUS_EXPR: | |
2114 | case MINUS_EXPR: | |
782a35e1 | 2115 | if (ANY_INTEGRAL_TYPE_P (TREE_TYPE (expr)) |
2116 | && TYPE_OVERFLOW_TRAPS (TREE_TYPE (expr))) | |
66d11416 | 2117 | return expr; |
2118 | /* Fallthru. */ | |
75a70cf9 | 2119 | case POINTER_PLUS_EXPR: |
1962ceba | 2120 | /* And increments and decrements by a constant are simple. */ |
75a70cf9 | 2121 | e1 = gimple_assign_rhs2 (stmt); |
2122 | if (!is_gimple_min_invariant (e1)) | |
2123 | return expr; | |
2124 | ||
c16eb701 | 2125 | ee = expand_simple_operations (e, stop, cache); |
75a70cf9 | 2126 | return fold_build2 (code, TREE_TYPE (expr), ee, e1); |
1962ceba | 2127 | |
75a70cf9 | 2128 | default: |
2129 | return expr; | |
2130 | } | |
1962ceba | 2131 | } |
2132 | ||
c16eb701 | 2133 | tree |
2134 | expand_simple_operations (tree expr, tree stop) | |
2135 | { | |
2136 | hash_map<tree, tree> cache; | |
2137 | return expand_simple_operations (expr, stop, cache); | |
2138 | } | |
2139 | ||
b9d73ea6 | 2140 | /* Tries to simplify EXPR using the condition COND. Returns the simplified |
1962ceba | 2141 | expression (or EXPR unchanged, if no simplification was possible). */ |
b9d73ea6 | 2142 | |
2143 | static tree | |
6723abf5 | 2144 | tree_simplify_using_condition_1 (tree cond, tree expr) |
b9d73ea6 | 2145 | { |
2146 | bool changed; | |
6723abf5 | 2147 | tree e, e0, e1, e2, notcond; |
b9d73ea6 | 2148 | enum tree_code code = TREE_CODE (expr); |
2149 | ||
2150 | if (code == INTEGER_CST) | |
2151 | return expr; | |
2152 | ||
2153 | if (code == TRUTH_OR_EXPR | |
2154 | || code == TRUTH_AND_EXPR | |
2155 | || code == COND_EXPR) | |
2156 | { | |
2157 | changed = false; | |
2158 | ||
6723abf5 | 2159 | e0 = tree_simplify_using_condition_1 (cond, TREE_OPERAND (expr, 0)); |
b9d73ea6 | 2160 | if (TREE_OPERAND (expr, 0) != e0) |
2161 | changed = true; | |
2162 | ||
6723abf5 | 2163 | e1 = tree_simplify_using_condition_1 (cond, TREE_OPERAND (expr, 1)); |
b9d73ea6 | 2164 | if (TREE_OPERAND (expr, 1) != e1) |
2165 | changed = true; | |
2166 | ||
2167 | if (code == COND_EXPR) | |
2168 | { | |
6723abf5 | 2169 | e2 = tree_simplify_using_condition_1 (cond, TREE_OPERAND (expr, 2)); |
b9d73ea6 | 2170 | if (TREE_OPERAND (expr, 2) != e2) |
2171 | changed = true; | |
2172 | } | |
2173 | else | |
2174 | e2 = NULL_TREE; | |
2175 | ||
2176 | if (changed) | |
2177 | { | |
2178 | if (code == COND_EXPR) | |
ae288d12 | 2179 | expr = fold_build3 (code, boolean_type_node, e0, e1, e2); |
b9d73ea6 | 2180 | else |
ae288d12 | 2181 | expr = fold_build2 (code, boolean_type_node, e0, e1); |
b9d73ea6 | 2182 | } |
2183 | ||
2184 | return expr; | |
2185 | } | |
2186 | ||
9a73db25 | 2187 | /* In case COND is equality, we may be able to simplify EXPR by copy/constant |
2188 | propagation, and vice versa. Fold does not handle this, since it is | |
2189 | considered too expensive. */ | |
2190 | if (TREE_CODE (cond) == EQ_EXPR) | |
2191 | { | |
2192 | e0 = TREE_OPERAND (cond, 0); | |
2193 | e1 = TREE_OPERAND (cond, 1); | |
2194 | ||
2195 | /* We know that e0 == e1. Check whether we cannot simplify expr | |
2196 | using this fact. */ | |
2197 | e = simplify_replace_tree (expr, e0, e1); | |
cd743a11 | 2198 | if (integer_zerop (e) || integer_nonzerop (e)) |
9a73db25 | 2199 | return e; |
2200 | ||
2201 | e = simplify_replace_tree (expr, e1, e0); | |
cd743a11 | 2202 | if (integer_zerop (e) || integer_nonzerop (e)) |
9a73db25 | 2203 | return e; |
2204 | } | |
2205 | if (TREE_CODE (expr) == EQ_EXPR) | |
2206 | { | |
2207 | e0 = TREE_OPERAND (expr, 0); | |
2208 | e1 = TREE_OPERAND (expr, 1); | |
2209 | ||
2210 | /* If e0 == e1 (EXPR) implies !COND, then EXPR cannot be true. */ | |
2211 | e = simplify_replace_tree (cond, e0, e1); | |
cd743a11 | 2212 | if (integer_zerop (e)) |
9a73db25 | 2213 | return e; |
2214 | e = simplify_replace_tree (cond, e1, e0); | |
cd743a11 | 2215 | if (integer_zerop (e)) |
9a73db25 | 2216 | return e; |
2217 | } | |
2218 | if (TREE_CODE (expr) == NE_EXPR) | |
2219 | { | |
2220 | e0 = TREE_OPERAND (expr, 0); | |
2221 | e1 = TREE_OPERAND (expr, 1); | |
2222 | ||
2223 | /* If e0 == e1 (!EXPR) implies !COND, then EXPR must be true. */ | |
2224 | e = simplify_replace_tree (cond, e0, e1); | |
cd743a11 | 2225 | if (integer_zerop (e)) |
9a73db25 | 2226 | return boolean_true_node; |
2227 | e = simplify_replace_tree (cond, e1, e0); | |
cd743a11 | 2228 | if (integer_zerop (e)) |
9a73db25 | 2229 | return boolean_true_node; |
2230 | } | |
2231 | ||
b9d73ea6 | 2232 | /* Check whether COND ==> EXPR. */ |
2233 | notcond = invert_truthvalue (cond); | |
6723abf5 | 2234 | e = fold_binary (TRUTH_OR_EXPR, boolean_type_node, notcond, expr); |
cd743a11 | 2235 | if (e && integer_nonzerop (e)) |
b9d73ea6 | 2236 | return e; |
2237 | ||
2238 | /* Check whether COND ==> not EXPR. */ | |
6723abf5 | 2239 | e = fold_binary (TRUTH_AND_EXPR, boolean_type_node, cond, expr); |
cd743a11 | 2240 | if (e && integer_zerop (e)) |
b9d73ea6 | 2241 | return e; |
2242 | ||
2243 | return expr; | |
2244 | } | |
2245 | ||
1962ceba | 2246 | /* Tries to simplify EXPR using the condition COND. Returns the simplified |
2247 | expression (or EXPR unchanged, if no simplification was possible). | |
2248 | Wrapper around tree_simplify_using_condition_1 that ensures that chains | |
2249 | of simple operations in definitions of ssa names in COND are expanded, | |
2250 | so that things like casts or incrementing the value of the bound before | |
2251 | the loop do not cause us to fail. */ | |
2252 | ||
2253 | static tree | |
6723abf5 | 2254 | tree_simplify_using_condition (tree cond, tree expr) |
1962ceba | 2255 | { |
6723abf5 | 2256 | cond = expand_simple_operations (cond); |
1962ceba | 2257 | |
6723abf5 | 2258 | return tree_simplify_using_condition_1 (cond, expr); |
1962ceba | 2259 | } |
86159992 | 2260 | |
b9d73ea6 | 2261 | /* Tries to simplify EXPR using the conditions on entry to LOOP. |
b9d73ea6 | 2262 | Returns the simplified expression (or EXPR unchanged, if no |
6c48ce84 | 2263 | simplification was possible). */ |
b9d73ea6 | 2264 | |
6c48ce84 | 2265 | tree |
6723abf5 | 2266 | simplify_using_initial_conditions (struct loop *loop, tree expr) |
b9d73ea6 | 2267 | { |
2268 | edge e; | |
2269 | basic_block bb; | |
42acab1c | 2270 | gimple *stmt; |
6723abf5 | 2271 | tree cond, expanded, backup; |
86159992 | 2272 | int cnt = 0; |
b9d73ea6 | 2273 | |
2274 | if (TREE_CODE (expr) == INTEGER_CST) | |
2275 | return expr; | |
2276 | ||
6723abf5 | 2277 | backup = expanded = expand_simple_operations (expr); |
2278 | ||
86159992 | 2279 | /* Limit walking the dominators to avoid quadraticness in |
2280 | the number of BBs times the number of loops in degenerate | |
2281 | cases. */ | |
b9d73ea6 | 2282 | for (bb = loop->header; |
34154e27 | 2283 | bb != ENTRY_BLOCK_PTR_FOR_FN (cfun) && cnt < MAX_DOMINATORS_TO_WALK; |
b9d73ea6 | 2284 | bb = get_immediate_dominator (CDI_DOMINATORS, bb)) |
2285 | { | |
ea091dfd | 2286 | if (!single_pred_p (bb)) |
b9d73ea6 | 2287 | continue; |
ea091dfd | 2288 | e = single_pred_edge (bb); |
b9d73ea6 | 2289 | |
2290 | if (!(e->flags & (EDGE_TRUE_VALUE | EDGE_FALSE_VALUE))) | |
2291 | continue; | |
2292 | ||
75a70cf9 | 2293 | stmt = last_stmt (e->src); |
2294 | cond = fold_build2 (gimple_cond_code (stmt), | |
2295 | boolean_type_node, | |
2296 | gimple_cond_lhs (stmt), | |
2297 | gimple_cond_rhs (stmt)); | |
b9d73ea6 | 2298 | if (e->flags & EDGE_FALSE_VALUE) |
2299 | cond = invert_truthvalue (cond); | |
6723abf5 | 2300 | expanded = tree_simplify_using_condition (cond, expanded); |
1ebfc4d6 | 2301 | /* Break if EXPR is simplified to const values. */ |
6723abf5 | 2302 | if (expanded |
2303 | && (integer_zerop (expanded) || integer_nonzerop (expanded))) | |
2304 | return expanded; | |
1ebfc4d6 | 2305 | |
86159992 | 2306 | ++cnt; |
b9d73ea6 | 2307 | } |
2308 | ||
6723abf5 | 2309 | /* Return the original expression if no simplification is done. */ |
2310 | return operand_equal_p (backup, expanded, 0) ? expr : expanded; | |
b9d73ea6 | 2311 | } |
2312 | ||
ae288d12 | 2313 | /* Tries to simplify EXPR using the evolutions of the loop invariants |
2314 | in the superloops of LOOP. Returns the simplified expression | |
2315 | (or EXPR unchanged, if no simplification was possible). */ | |
2316 | ||
2317 | static tree | |
2318 | simplify_using_outer_evolutions (struct loop *loop, tree expr) | |
2319 | { | |
2320 | enum tree_code code = TREE_CODE (expr); | |
2321 | bool changed; | |
2322 | tree e, e0, e1, e2; | |
2323 | ||
2324 | if (is_gimple_min_invariant (expr)) | |
2325 | return expr; | |
2326 | ||
2327 | if (code == TRUTH_OR_EXPR | |
2328 | || code == TRUTH_AND_EXPR | |
2329 | || code == COND_EXPR) | |
2330 | { | |
2331 | changed = false; | |
2332 | ||
2333 | e0 = simplify_using_outer_evolutions (loop, TREE_OPERAND (expr, 0)); | |
2334 | if (TREE_OPERAND (expr, 0) != e0) | |
2335 | changed = true; | |
2336 | ||
2337 | e1 = simplify_using_outer_evolutions (loop, TREE_OPERAND (expr, 1)); | |
2338 | if (TREE_OPERAND (expr, 1) != e1) | |
2339 | changed = true; | |
2340 | ||
2341 | if (code == COND_EXPR) | |
2342 | { | |
2343 | e2 = simplify_using_outer_evolutions (loop, TREE_OPERAND (expr, 2)); | |
2344 | if (TREE_OPERAND (expr, 2) != e2) | |
2345 | changed = true; | |
2346 | } | |
2347 | else | |
2348 | e2 = NULL_TREE; | |
2349 | ||
2350 | if (changed) | |
2351 | { | |
2352 | if (code == COND_EXPR) | |
2353 | expr = fold_build3 (code, boolean_type_node, e0, e1, e2); | |
2354 | else | |
2355 | expr = fold_build2 (code, boolean_type_node, e0, e1); | |
2356 | } | |
2357 | ||
2358 | return expr; | |
2359 | } | |
2360 | ||
2361 | e = instantiate_parameters (loop, expr); | |
2362 | if (is_gimple_min_invariant (e)) | |
2363 | return e; | |
2364 | ||
2365 | return expr; | |
2366 | } | |
2367 | ||
0840f4fe | 2368 | /* Returns true if EXIT is the only possible exit from LOOP. */ |
2369 | ||
14f25899 | 2370 | bool |
1f1872fd | 2371 | loop_only_exit_p (const struct loop *loop, const_edge exit) |
0840f4fe | 2372 | { |
2373 | basic_block *body; | |
75a70cf9 | 2374 | gimple_stmt_iterator bsi; |
0840f4fe | 2375 | unsigned i; |
0840f4fe | 2376 | |
d9e7e1a2 | 2377 | if (exit != single_exit (loop)) |
0840f4fe | 2378 | return false; |
2379 | ||
2380 | body = get_loop_body (loop); | |
2381 | for (i = 0; i < loop->num_nodes; i++) | |
2382 | { | |
75a70cf9 | 2383 | for (bsi = gsi_start_bb (body[i]); !gsi_end_p (bsi); gsi_next (&bsi)) |
531f0ff8 | 2384 | if (stmt_can_terminate_bb_p (gsi_stmt (bsi))) |
118c9b7a | 2385 | { |
2386 | free (body); | |
2387 | return true; | |
2388 | } | |
0840f4fe | 2389 | } |
2390 | ||
2391 | free (body); | |
2392 | return true; | |
2393 | } | |
2394 | ||
b9d73ea6 | 2395 | /* Stores description of number of iterations of LOOP derived from |
a4d34e5e | 2396 | EXIT (an exit edge of the LOOP) in NITER. Returns true if some useful |
2397 | information could be derived (and fields of NITER have meaning described | |
2398 | in comments at struct tree_niter_desc declaration), false otherwise. | |
fd943c70 | 2399 | When EVERY_ITERATION is true, only tests that are known to be executed |
a4d34e5e | 2400 | every iteration are considered (i.e. only test that alone bounds the loop). |
1b793819 | 2401 | If AT_STMT is not NULL, this function stores LOOP's condition statement in |
2402 | it when returning true. */ | |
b9d73ea6 | 2403 | |
2404 | bool | |
a4d34e5e | 2405 | number_of_iterations_exit_assumptions (struct loop *loop, edge exit, |
2406 | struct tree_niter_desc *niter, | |
1b793819 | 2407 | gcond **at_stmt, bool every_iteration) |
b9d73ea6 | 2408 | { |
42acab1c | 2409 | gimple *last; |
1a91d914 | 2410 | gcond *stmt; |
75a70cf9 | 2411 | tree type; |
553b9523 | 2412 | tree op0, op1; |
b9d73ea6 | 2413 | enum tree_code code; |
553b9523 | 2414 | affine_iv iv0, iv1; |
64552007 | 2415 | bool safe; |
b9d73ea6 | 2416 | |
85168485 | 2417 | /* Nothing to analyze if the loop is known to be infinite. */ |
2418 | if (loop_constraint_set_p (loop, LOOP_C_INFINITE)) | |
2419 | return false; | |
2420 | ||
64552007 | 2421 | safe = dominated_by_p (CDI_DOMINATORS, loop->latch, exit->src); |
2422 | ||
2423 | if (every_iteration && !safe) | |
b9d73ea6 | 2424 | return false; |
2425 | ||
2426 | niter->assumptions = boolean_false_node; | |
05032b10 | 2427 | niter->control.base = NULL_TREE; |
2428 | niter->control.step = NULL_TREE; | |
2429 | niter->control.no_overflow = false; | |
1a91d914 | 2430 | last = last_stmt (exit->src); |
2431 | if (!last) | |
2432 | return false; | |
2433 | stmt = dyn_cast <gcond *> (last); | |
2434 | if (!stmt) | |
b9d73ea6 | 2435 | return false; |
2436 | ||
2437 | /* We want the condition for staying inside loop. */ | |
75a70cf9 | 2438 | code = gimple_cond_code (stmt); |
b9d73ea6 | 2439 | if (exit->flags & EDGE_TRUE_VALUE) |
75a70cf9 | 2440 | code = invert_tree_comparison (code, false); |
b9d73ea6 | 2441 | |
b9d73ea6 | 2442 | switch (code) |
2443 | { | |
2444 | case GT_EXPR: | |
2445 | case GE_EXPR: | |
b9d73ea6 | 2446 | case LT_EXPR: |
2447 | case LE_EXPR: | |
64552007 | 2448 | case NE_EXPR: |
b9d73ea6 | 2449 | break; |
2450 | ||
2451 | default: | |
2452 | return false; | |
2453 | } | |
48e1416a | 2454 | |
75a70cf9 | 2455 | op0 = gimple_cond_lhs (stmt); |
2456 | op1 = gimple_cond_rhs (stmt); | |
b9d73ea6 | 2457 | type = TREE_TYPE (op0); |
2458 | ||
2459 | if (TREE_CODE (type) != INTEGER_TYPE | |
4fef3950 | 2460 | && !POINTER_TYPE_P (type)) |
b9d73ea6 | 2461 | return false; |
48e1416a | 2462 | |
a4d34e5e | 2463 | tree iv0_niters = NULL_TREE; |
2464 | if (!simple_iv_with_niters (loop, loop_containing_stmt (stmt), | |
03d52952 | 2465 | op0, &iv0, safe ? &iv0_niters : NULL, false)) |
c71993fc | 2466 | return number_of_iterations_popcount (loop, exit, code, niter); |
a4d34e5e | 2467 | tree iv1_niters = NULL_TREE; |
2468 | if (!simple_iv_with_niters (loop, loop_containing_stmt (stmt), | |
03d52952 | 2469 | op1, &iv1, safe ? &iv1_niters : NULL, false)) |
b9d73ea6 | 2470 | return false; |
a4d34e5e | 2471 | /* Give up on complicated case. */ |
2472 | if (iv0_niters && iv1_niters) | |
b9d73ea6 | 2473 | return false; |
2474 | ||
add6ee5e | 2475 | /* We don't want to see undefined signed overflow warnings while |
85694bac | 2476 | computing the number of iterations. */ |
add6ee5e | 2477 | fold_defer_overflow_warnings (); |
2478 | ||
0f81d1b1 | 2479 | iv0.base = expand_simple_operations (iv0.base); |
2480 | iv1.base = expand_simple_operations (iv1.base); | |
a463eaea | 2481 | if (!number_of_iterations_cond (loop, type, &iv0, code, &iv1, niter, |
64552007 | 2482 | loop_only_exit_p (loop, exit), safe)) |
add6ee5e | 2483 | { |
2484 | fold_undefer_and_ignore_overflow_warnings (); | |
2485 | return false; | |
2486 | } | |
ae288d12 | 2487 | |
a4d34e5e | 2488 | /* Incorporate additional assumption implied by control iv. */ |
2489 | tree iv_niters = iv0_niters ? iv0_niters : iv1_niters; | |
2490 | if (iv_niters) | |
2491 | { | |
2492 | tree assumption = fold_build2 (LE_EXPR, boolean_type_node, niter->niter, | |
2493 | fold_convert (TREE_TYPE (niter->niter), | |
2494 | iv_niters)); | |
2495 | ||
2496 | if (!integer_nonzerop (assumption)) | |
2497 | niter->assumptions = fold_build2 (TRUTH_AND_EXPR, boolean_type_node, | |
2498 | niter->assumptions, assumption); | |
2499 | ||
2500 | /* Refine upper bound if possible. */ | |
2501 | if (TREE_CODE (iv_niters) == INTEGER_CST | |
2502 | && niter->max > wi::to_widest (iv_niters)) | |
2503 | niter->max = wi::to_widest (iv_niters); | |
2504 | } | |
2505 | ||
85168485 | 2506 | /* There is no assumptions if the loop is known to be finite. */ |
2507 | if (!integer_zerop (niter->assumptions) | |
2508 | && loop_constraint_set_p (loop, LOOP_C_FINITE)) | |
2509 | niter->assumptions = boolean_true_node; | |
2510 | ||
ae288d12 | 2511 | if (optimize >= 3) |
2512 | { | |
2513 | niter->assumptions = simplify_using_outer_evolutions (loop, | |
2514 | niter->assumptions); | |
2515 | niter->may_be_zero = simplify_using_outer_evolutions (loop, | |
2516 | niter->may_be_zero); | |
2517 | niter->niter = simplify_using_outer_evolutions (loop, niter->niter); | |
2518 | } | |
b9d73ea6 | 2519 | |
b9d73ea6 | 2520 | niter->assumptions |
2521 | = simplify_using_initial_conditions (loop, | |
a463eaea | 2522 | niter->assumptions); |
b9d73ea6 | 2523 | niter->may_be_zero |
2524 | = simplify_using_initial_conditions (loop, | |
a463eaea | 2525 | niter->may_be_zero); |
7e3cc681 | 2526 | |
add6ee5e | 2527 | fold_undefer_and_ignore_overflow_warnings (); |
2528 | ||
72276d01 | 2529 | /* If NITER has simplified into a constant, update MAX. */ |
2530 | if (TREE_CODE (niter->niter) == INTEGER_CST) | |
7f7d8378 | 2531 | niter->max = wi::to_widest (niter->niter); |
72276d01 | 2532 | |
1b793819 | 2533 | if (at_stmt) |
2534 | *at_stmt = stmt; | |
2535 | ||
a4d34e5e | 2536 | return (!integer_zerop (niter->assumptions)); |
2537 | } | |
48e1416a | 2538 | |
c71993fc | 2539 | |
2540 | /* Utility function to check if OP is defined by a stmt | |
2541 | that is a val - 1. */ | |
2542 | ||
2543 | static bool | |
2544 | ssa_defined_by_minus_one_stmt_p (tree op, tree val) | |
2545 | { | |
2546 | gimple *stmt; | |
2547 | return (TREE_CODE (op) == SSA_NAME | |
2548 | && (stmt = SSA_NAME_DEF_STMT (op)) | |
2549 | && is_gimple_assign (stmt) | |
2550 | && (gimple_assign_rhs_code (stmt) == PLUS_EXPR) | |
2551 | && val == gimple_assign_rhs1 (stmt) | |
2552 | && integer_minus_onep (gimple_assign_rhs2 (stmt))); | |
2553 | } | |
2554 | ||
2555 | ||
2556 | /* See if LOOP is a popcout implementation, determine NITER for the loop | |
2557 | ||
2558 | We match: | |
2559 | <bb 2> | |
2560 | goto <bb 4> | |
2561 | ||
2562 | <bb 3> | |
2563 | _1 = b_11 + -1 | |
2564 | b_6 = _1 & b_11 | |
2565 | ||
2566 | <bb 4> | |
2567 | b_11 = PHI <b_5(D)(2), b_6(3)> | |
2568 | ||
2569 | exit block | |
2570 | if (b_11 != 0) | |
2571 | goto <bb 3> | |
2572 | else | |
2573 | goto <bb 5> | |
2574 | ||
2575 | OR we match copy-header version: | |
2576 | if (b_5 != 0) | |
2577 | goto <bb 3> | |
2578 | else | |
2579 | goto <bb 4> | |
2580 | ||
2581 | <bb 3> | |
2582 | b_11 = PHI <b_5(2), b_6(3)> | |
2583 | _1 = b_11 + -1 | |
2584 | b_6 = _1 & b_11 | |
2585 | ||
2586 | exit block | |
2587 | if (b_6 != 0) | |
2588 | goto <bb 3> | |
2589 | else | |
2590 | goto <bb 4> | |
2591 | ||
2592 | If popcount pattern, update NITER accordingly. | |
2593 | i.e., set NITER to __builtin_popcount (b) | |
2594 | return true if we did, false otherwise. | |
2595 | ||
2596 | */ | |
2597 | ||
2598 | static bool | |
2599 | number_of_iterations_popcount (loop_p loop, edge exit, | |
2600 | enum tree_code code, | |
2601 | struct tree_niter_desc *niter) | |
2602 | { | |
2603 | bool adjust = true; | |
2604 | tree iter; | |
2605 | HOST_WIDE_INT max; | |
2606 | adjust = true; | |
2607 | tree fn = NULL_TREE; | |
2608 | ||
2609 | /* Check loop terminating branch is like | |
2610 | if (b != 0). */ | |
2611 | gimple *stmt = last_stmt (exit->src); | |
2612 | if (!stmt | |
2613 | || gimple_code (stmt) != GIMPLE_COND | |
2614 | || code != NE_EXPR | |
2615 | || !integer_zerop (gimple_cond_rhs (stmt)) | |
2616 | || TREE_CODE (gimple_cond_lhs (stmt)) != SSA_NAME) | |
2617 | return false; | |
2618 | ||
2619 | gimple *and_stmt = SSA_NAME_DEF_STMT (gimple_cond_lhs (stmt)); | |
2620 | ||
2621 | /* Depending on copy-header is performed, feeding PHI stmts might be in | |
2622 | the loop header or loop latch, handle this. */ | |
2623 | if (gimple_code (and_stmt) == GIMPLE_PHI | |
2624 | && gimple_bb (and_stmt) == loop->header | |
2625 | && gimple_phi_num_args (and_stmt) == 2 | |
2626 | && (TREE_CODE (gimple_phi_arg_def (and_stmt, | |
2627 | loop_latch_edge (loop)->dest_idx)) | |
2628 | == SSA_NAME)) | |
2629 | { | |
2630 | /* SSA used in exit condition is defined by PHI stmt | |
2631 | b_11 = PHI <b_5(D)(2), b_6(3)> | |
2632 | from the PHI stmt, get the and_stmt | |
2633 | b_6 = _1 & b_11. */ | |
2634 | tree t = gimple_phi_arg_def (and_stmt, loop_latch_edge (loop)->dest_idx); | |
2635 | and_stmt = SSA_NAME_DEF_STMT (t); | |
2636 | adjust = false; | |
2637 | } | |
2638 | ||
2639 | /* Make sure it is indeed an and stmt (b_6 = _1 & b_11). */ | |
2640 | if (!is_gimple_assign (and_stmt) | |
2641 | || gimple_assign_rhs_code (and_stmt) != BIT_AND_EXPR) | |
2642 | return false; | |
2643 | ||
2644 | tree b_11 = gimple_assign_rhs1 (and_stmt); | |
2645 | tree _1 = gimple_assign_rhs2 (and_stmt); | |
2646 | ||
2647 | /* Check that _1 is defined by _b11 + -1 (_1 = b_11 + -1). | |
2648 | Also make sure that b_11 is the same in and_stmt and _1 defining stmt. | |
2649 | Also canonicalize if _1 and _b11 are revrsed. */ | |
2650 | if (ssa_defined_by_minus_one_stmt_p (b_11, _1)) | |
2651 | std::swap (b_11, _1); | |
2652 | else if (ssa_defined_by_minus_one_stmt_p (_1, b_11)) | |
2653 | ; | |
2654 | else | |
2655 | return false; | |
2656 | /* Check the recurrence: | |
2657 | ... = PHI <b_5(2), b_6(3)>. */ | |
2658 | gimple *phi = SSA_NAME_DEF_STMT (b_11); | |
2659 | if (gimple_code (phi) != GIMPLE_PHI | |
ad3d7ab9 | 2660 | || (gimple_bb (phi) != loop_latch_edge (loop)->dest) |
c71993fc | 2661 | || (gimple_assign_lhs (and_stmt) |
2662 | != gimple_phi_arg_def (phi, loop_latch_edge (loop)->dest_idx))) | |
2663 | return false; | |
2664 | ||
2665 | /* We found a match. Get the corresponding popcount builtin. */ | |
2666 | tree src = gimple_phi_arg_def (phi, loop_preheader_edge (loop)->dest_idx); | |
2667 | if (TYPE_PRECISION (TREE_TYPE (src)) == TYPE_PRECISION (integer_type_node)) | |
2668 | fn = builtin_decl_implicit (BUILT_IN_POPCOUNT); | |
2669 | else if (TYPE_PRECISION (TREE_TYPE (src)) == TYPE_PRECISION | |
2670 | (long_integer_type_node)) | |
2671 | fn = builtin_decl_implicit (BUILT_IN_POPCOUNTL); | |
2672 | else if (TYPE_PRECISION (TREE_TYPE (src)) == TYPE_PRECISION | |
2673 | (long_long_integer_type_node)) | |
2674 | fn = builtin_decl_implicit (BUILT_IN_POPCOUNTLL); | |
2675 | ||
2676 | /* ??? Support promoting char/short to int. */ | |
2677 | if (!fn) | |
2678 | return false; | |
2679 | ||
2680 | /* Update NITER params accordingly */ | |
c71993fc | 2681 | tree utype = unsigned_type_for (TREE_TYPE (src)); |
2682 | src = fold_convert (utype, src); | |
2683 | tree call = fold_convert (utype, build_call_expr (fn, 1, src)); | |
2684 | if (adjust) | |
2685 | iter = fold_build2 (MINUS_EXPR, utype, | |
2686 | call, | |
2687 | build_int_cst (utype, 1)); | |
2688 | else | |
2689 | iter = call; | |
2690 | ||
35f05322 | 2691 | if (TREE_CODE (call) == INTEGER_CST) |
2692 | max = tree_to_uhwi (call); | |
2693 | else | |
81bfce6a | 2694 | max = TYPE_PRECISION (TREE_TYPE (src)); |
2695 | if (adjust) | |
2696 | max = max - 1; | |
35f05322 | 2697 | |
c71993fc | 2698 | niter->niter = iter; |
2699 | niter->assumptions = boolean_true_node; | |
af73b08a | 2700 | |
c71993fc | 2701 | if (adjust) |
af73b08a | 2702 | { |
2703 | tree may_be_zero = fold_build2 (EQ_EXPR, boolean_type_node, src, | |
c71993fc | 2704 | build_zero_cst |
2705 | (TREE_TYPE (src))); | |
af73b08a | 2706 | niter->may_be_zero = |
2707 | simplify_using_initial_conditions (loop, may_be_zero); | |
2708 | } | |
c71993fc | 2709 | else |
2710 | niter->may_be_zero = boolean_false_node; | |
2711 | ||
2712 | niter->max = max; | |
2713 | niter->bound = NULL_TREE; | |
2714 | niter->cmp = ERROR_MARK; | |
2715 | return true; | |
2716 | } | |
2717 | ||
2718 | ||
80ac98e8 | 2719 | /* Like number_of_iterations_exit_assumptions, but return TRUE only if |
2720 | the niter information holds unconditionally. */ | |
7e3cc681 | 2721 | |
a4d34e5e | 2722 | bool |
2723 | number_of_iterations_exit (struct loop *loop, edge exit, | |
2724 | struct tree_niter_desc *niter, | |
1b793819 | 2725 | bool warn, bool every_iteration) |
a4d34e5e | 2726 | { |
1b793819 | 2727 | gcond *stmt; |
a4d34e5e | 2728 | if (!number_of_iterations_exit_assumptions (loop, exit, niter, |
1b793819 | 2729 | &stmt, every_iteration)) |
a4d34e5e | 2730 | return false; |
7e3cc681 | 2731 | |
1b793819 | 2732 | if (integer_nonzerop (niter->assumptions)) |
2733 | return true; | |
2734 | ||
91f42adc | 2735 | if (warn && dump_enabled_p ()) |
c309657f | 2736 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, stmt, |
e06b96aa | 2737 | "missed loop optimization: niters analysis ends up " |
2738 | "with assumptions.\n"); | |
1b793819 | 2739 | |
2740 | return false; | |
b9d73ea6 | 2741 | } |
2742 | ||
b091dc59 | 2743 | /* Try to determine the number of iterations of LOOP. If we succeed, |
2744 | expression giving number of iterations is returned and *EXIT is | |
2745 | set to the edge from that the information is obtained. Otherwise | |
2746 | chrec_dont_know is returned. */ | |
2747 | ||
2748 | tree | |
2749 | find_loop_niter (struct loop *loop, edge *exit) | |
2750 | { | |
749ea85f | 2751 | unsigned i; |
f1f41a6c | 2752 | vec<edge> exits = get_loop_exit_edges (loop); |
b091dc59 | 2753 | edge ex; |
2754 | tree niter = NULL_TREE, aniter; | |
2755 | struct tree_niter_desc desc; | |
2756 | ||
2757 | *exit = NULL; | |
f1f41a6c | 2758 | FOR_EACH_VEC_ELT (exits, i, ex) |
b091dc59 | 2759 | { |
7e3cc681 | 2760 | if (!number_of_iterations_exit (loop, ex, &desc, false)) |
b091dc59 | 2761 | continue; |
2762 | ||
cd743a11 | 2763 | if (integer_nonzerop (desc.may_be_zero)) |
b091dc59 | 2764 | { |
2765 | /* We exit in the first iteration through this exit. | |
2766 | We won't find anything better. */ | |
05db596e | 2767 | niter = build_int_cst (unsigned_type_node, 0); |
b091dc59 | 2768 | *exit = ex; |
2769 | break; | |
2770 | } | |
2771 | ||
cd743a11 | 2772 | if (!integer_zerop (desc.may_be_zero)) |
b091dc59 | 2773 | continue; |
2774 | ||
2775 | aniter = desc.niter; | |
2776 | ||
2777 | if (!niter) | |
2778 | { | |
2779 | /* Nothing recorded yet. */ | |
2780 | niter = aniter; | |
2781 | *exit = ex; | |
2782 | continue; | |
2783 | } | |
2784 | ||
2785 | /* Prefer constants, the lower the better. */ | |
2786 | if (TREE_CODE (aniter) != INTEGER_CST) | |
2787 | continue; | |
2788 | ||
2789 | if (TREE_CODE (niter) != INTEGER_CST) | |
2790 | { | |
2791 | niter = aniter; | |
2792 | *exit = ex; | |
2793 | continue; | |
2794 | } | |
2795 | ||
2796 | if (tree_int_cst_lt (aniter, niter)) | |
2797 | { | |
2798 | niter = aniter; | |
2799 | *exit = ex; | |
2800 | continue; | |
2801 | } | |
2802 | } | |
f1f41a6c | 2803 | exits.release (); |
b091dc59 | 2804 | |
2805 | return niter ? niter : chrec_dont_know; | |
2806 | } | |
2807 | ||
df9b545b | 2808 | /* Return true if loop is known to have bounded number of iterations. */ |
2809 | ||
2810 | bool | |
2811 | finite_loop_p (struct loop *loop) | |
2812 | { | |
5de9d3ed | 2813 | widest_int nit; |
67fa4078 | 2814 | int flags; |
df9b545b | 2815 | |
67fa4078 | 2816 | flags = flags_from_decl_or_type (current_function_decl); |
2817 | if ((flags & (ECF_CONST|ECF_PURE)) && !(flags & ECF_LOOPING_CONST_OR_PURE)) | |
df9b545b | 2818 | { |
2819 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2820 | fprintf (dump_file, "Found loop %i to be finite: it is within pure or const function.\n", | |
2821 | loop->num); | |
2822 | return true; | |
2823 | } | |
48e1416a | 2824 | |
13e18258 | 2825 | if (loop->any_upper_bound |
2826 | || max_loop_iterations (loop, &nit)) | |
df9b545b | 2827 | { |
13e18258 | 2828 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2829 | fprintf (dump_file, "Found loop %i to be finite: upper bound found.\n", | |
2830 | loop->num); | |
2831 | return true; | |
df9b545b | 2832 | } |
3984c5cd | 2833 | |
2834 | if (flag_finite_loops) | |
2835 | { | |
2836 | unsigned i; | |
2837 | vec<edge> exits = get_loop_exit_edges (loop); | |
2838 | edge ex; | |
2839 | ||
2840 | /* If the loop has a normal exit, we can assume it will terminate. */ | |
2841 | FOR_EACH_VEC_ELT (exits, i, ex) | |
2842 | if (!(ex->flags & (EDGE_EH | EDGE_ABNORMAL | EDGE_FAKE))) | |
2843 | { | |
2844 | exits.release (); | |
2845 | if (dump_file) | |
2846 | fprintf (dump_file, "Assume loop %i to be finite: it has an exit " | |
2847 | "and -ffinite-loops is on.\n", loop->num); | |
2848 | return true; | |
2849 | } | |
2850 | ||
2851 | exits.release (); | |
2852 | } | |
2853 | ||
13e18258 | 2854 | return false; |
df9b545b | 2855 | } |
2856 | ||
b9d73ea6 | 2857 | /* |
2858 | ||
2859 | Analysis of a number of iterations of a loop by a brute-force evaluation. | |
2860 | ||
2861 | */ | |
2862 | ||
2863 | /* Bound on the number of iterations we try to evaluate. */ | |
2864 | ||
2865 | #define MAX_ITERATIONS_TO_TRACK \ | |
2866 | ((unsigned) PARAM_VALUE (PARAM_MAX_ITERATIONS_TO_TRACK)) | |
2867 | ||
2868 | /* Returns the loop phi node of LOOP such that ssa name X is derived from its | |
2869 | result by a chain of operations such that all but exactly one of their | |
2870 | operands are constants. */ | |
2871 | ||
1a91d914 | 2872 | static gphi * |
b9d73ea6 | 2873 | chain_of_csts_start (struct loop *loop, tree x) |
2874 | { | |
42acab1c | 2875 | gimple *stmt = SSA_NAME_DEF_STMT (x); |
b66731e8 | 2876 | tree use; |
75a70cf9 | 2877 | basic_block bb = gimple_bb (stmt); |
2878 | enum tree_code code; | |
b9d73ea6 | 2879 | |
2880 | if (!bb | |
2881 | || !flow_bb_inside_loop_p (loop, bb)) | |
75a70cf9 | 2882 | return NULL; |
48e1416a | 2883 | |
75a70cf9 | 2884 | if (gimple_code (stmt) == GIMPLE_PHI) |
b9d73ea6 | 2885 | { |
2886 | if (bb == loop->header) | |
1a91d914 | 2887 | return as_a <gphi *> (stmt); |
b9d73ea6 | 2888 | |
75a70cf9 | 2889 | return NULL; |
b9d73ea6 | 2890 | } |
2891 | ||
ec1d1b75 | 2892 | if (gimple_code (stmt) != GIMPLE_ASSIGN |
2893 | || gimple_assign_rhs_class (stmt) == GIMPLE_TERNARY_RHS) | |
75a70cf9 | 2894 | return NULL; |
b9d73ea6 | 2895 | |
75a70cf9 | 2896 | code = gimple_assign_rhs_code (stmt); |
2897 | if (gimple_references_memory_p (stmt) | |
75a70cf9 | 2898 | || TREE_CODE_CLASS (code) == tcc_reference |
dd277d48 | 2899 | || (code == ADDR_EXPR |
2900 | && !is_gimple_min_invariant (gimple_assign_rhs1 (stmt)))) | |
75a70cf9 | 2901 | return NULL; |
b66731e8 | 2902 | |
2903 | use = SINGLE_SSA_TREE_OPERAND (stmt, SSA_OP_USE); | |
dd277d48 | 2904 | if (use == NULL_TREE) |
75a70cf9 | 2905 | return NULL; |
b9d73ea6 | 2906 | |
b66731e8 | 2907 | return chain_of_csts_start (loop, use); |
b9d73ea6 | 2908 | } |
2909 | ||
2910 | /* Determines whether the expression X is derived from a result of a phi node | |
2911 | in header of LOOP such that | |
2912 | ||
2913 | * the derivation of X consists only from operations with constants | |
2914 | * the initial value of the phi node is constant | |
2915 | * the value of the phi node in the next iteration can be derived from the | |
13cd9f01 | 2916 | value in the current iteration by a chain of operations with constants, |
2917 | or is also a constant | |
48e1416a | 2918 | |
75a70cf9 | 2919 | If such phi node exists, it is returned, otherwise NULL is returned. */ |
b9d73ea6 | 2920 | |
1a91d914 | 2921 | static gphi * |
b9d73ea6 | 2922 | get_base_for (struct loop *loop, tree x) |
2923 | { | |
1a91d914 | 2924 | gphi *phi; |
75a70cf9 | 2925 | tree init, next; |
b9d73ea6 | 2926 | |
2927 | if (is_gimple_min_invariant (x)) | |
75a70cf9 | 2928 | return NULL; |
b9d73ea6 | 2929 | |
2930 | phi = chain_of_csts_start (loop, x); | |
2931 | if (!phi) | |
75a70cf9 | 2932 | return NULL; |
b9d73ea6 | 2933 | |
2934 | init = PHI_ARG_DEF_FROM_EDGE (phi, loop_preheader_edge (loop)); | |
2935 | next = PHI_ARG_DEF_FROM_EDGE (phi, loop_latch_edge (loop)); | |
2936 | ||
b9d73ea6 | 2937 | if (!is_gimple_min_invariant (init)) |
75a70cf9 | 2938 | return NULL; |
b9d73ea6 | 2939 | |
13cd9f01 | 2940 | if (TREE_CODE (next) == SSA_NAME |
2941 | && chain_of_csts_start (loop, next) != phi) | |
75a70cf9 | 2942 | return NULL; |
b9d73ea6 | 2943 | |
2944 | return phi; | |
2945 | } | |
2946 | ||
48e1416a | 2947 | /* Given an expression X, then |
2948 | ||
f8540d49 | 2949 | * if X is NULL_TREE, we return the constant BASE. |
13cd9f01 | 2950 | * if X is a constant, we return the constant X. |
b9d73ea6 | 2951 | * otherwise X is a SSA name, whose value in the considered loop is derived |
2952 | by a chain of operations with constant from a result of a phi node in | |
2953 | the header of the loop. Then we return value of X when the value of the | |
2954 | result of this phi node is given by the constant BASE. */ | |
2955 | ||
2956 | static tree | |
2957 | get_val_for (tree x, tree base) | |
2958 | { | |
42acab1c | 2959 | gimple *stmt; |
b9d73ea6 | 2960 | |
ec1d1b75 | 2961 | gcc_checking_assert (is_gimple_min_invariant (base)); |
f8540d49 | 2962 | |
b9d73ea6 | 2963 | if (!x) |
2964 | return base; | |
13cd9f01 | 2965 | else if (is_gimple_min_invariant (x)) |
2966 | return x; | |
b9d73ea6 | 2967 | |
2968 | stmt = SSA_NAME_DEF_STMT (x); | |
75a70cf9 | 2969 | if (gimple_code (stmt) == GIMPLE_PHI) |
b9d73ea6 | 2970 | return base; |
2971 | ||
ec1d1b75 | 2972 | gcc_checking_assert (is_gimple_assign (stmt)); |
75a70cf9 | 2973 | |
2974 | /* STMT must be either an assignment of a single SSA name or an | |
2975 | expression involving an SSA name and a constant. Try to fold that | |
2976 | expression using the value for the SSA name. */ | |
b8747787 | 2977 | if (gimple_assign_ssa_name_copy_p (stmt)) |
2978 | return get_val_for (gimple_assign_rhs1 (stmt), base); | |
2979 | else if (gimple_assign_rhs_class (stmt) == GIMPLE_UNARY_RHS | |
2980 | && TREE_CODE (gimple_assign_rhs1 (stmt)) == SSA_NAME) | |
13cd9f01 | 2981 | return fold_build1 (gimple_assign_rhs_code (stmt), |
2982 | gimple_expr_type (stmt), | |
2983 | get_val_for (gimple_assign_rhs1 (stmt), base)); | |
b8747787 | 2984 | else if (gimple_assign_rhs_class (stmt) == GIMPLE_BINARY_RHS) |
75a70cf9 | 2985 | { |
b8747787 | 2986 | tree rhs1 = gimple_assign_rhs1 (stmt); |
2987 | tree rhs2 = gimple_assign_rhs2 (stmt); | |
2988 | if (TREE_CODE (rhs1) == SSA_NAME) | |
2989 | rhs1 = get_val_for (rhs1, base); | |
2990 | else if (TREE_CODE (rhs2) == SSA_NAME) | |
2991 | rhs2 = get_val_for (rhs2, base); | |
2992 | else | |
2993 | gcc_unreachable (); | |
2994 | return fold_build2 (gimple_assign_rhs_code (stmt), | |
2995 | gimple_expr_type (stmt), rhs1, rhs2); | |
b66731e8 | 2996 | } |
75a70cf9 | 2997 | else |
b8747787 | 2998 | gcc_unreachable (); |
b9d73ea6 | 2999 | } |
3000 | ||
75a70cf9 | 3001 | |
b9d73ea6 | 3002 | /* Tries to count the number of iterations of LOOP till it exits by EXIT |
3003 | by brute force -- i.e. by determining the value of the operands of the | |
3004 | condition at EXIT in first few iterations of the loop (assuming that | |
3005 | these values are constant) and determining the first one in that the | |
3006 | condition is not satisfied. Returns the constant giving the number | |
3007 | of the iterations of LOOP if successful, chrec_dont_know otherwise. */ | |
3008 | ||
3009 | tree | |
3010 | loop_niter_by_eval (struct loop *loop, edge exit) | |
3011 | { | |
75a70cf9 | 3012 | tree acnd; |
3013 | tree op[2], val[2], next[2], aval[2]; | |
1a91d914 | 3014 | gphi *phi; |
42acab1c | 3015 | gimple *cond; |
b9d73ea6 | 3016 | unsigned i, j; |
3017 | enum tree_code cmp; | |
3018 | ||
3019 | cond = last_stmt (exit->src); | |
75a70cf9 | 3020 | if (!cond || gimple_code (cond) != GIMPLE_COND) |
b9d73ea6 | 3021 | return chrec_dont_know; |
3022 | ||
75a70cf9 | 3023 | cmp = gimple_cond_code (cond); |
b9d73ea6 | 3024 | if (exit->flags & EDGE_TRUE_VALUE) |
75a70cf9 | 3025 | cmp = invert_tree_comparison (cmp, false); |
b9d73ea6 | 3026 | |
b9d73ea6 | 3027 | switch (cmp) |
3028 | { | |
3029 | case EQ_EXPR: | |
3030 | case NE_EXPR: | |
3031 | case GT_EXPR: | |
3032 | case GE_EXPR: | |
3033 | case LT_EXPR: | |
3034 | case LE_EXPR: | |
75a70cf9 | 3035 | op[0] = gimple_cond_lhs (cond); |
3036 | op[1] = gimple_cond_rhs (cond); | |
b9d73ea6 | 3037 | break; |
3038 | ||
3039 | default: | |
3040 | return chrec_dont_know; | |
3041 | } | |
3042 | ||
3043 | for (j = 0; j < 2; j++) | |
3044 | { | |
75a70cf9 | 3045 | if (is_gimple_min_invariant (op[j])) |
b9d73ea6 | 3046 | { |
75a70cf9 | 3047 | val[j] = op[j]; |
3048 | next[j] = NULL_TREE; | |
3049 | op[j] = NULL_TREE; | |
b9d73ea6 | 3050 | } |
3051 | else | |
3052 | { | |
75a70cf9 | 3053 | phi = get_base_for (loop, op[j]); |
3054 | if (!phi) | |
3055 | return chrec_dont_know; | |
3056 | val[j] = PHI_ARG_DEF_FROM_EDGE (phi, loop_preheader_edge (loop)); | |
3057 | next[j] = PHI_ARG_DEF_FROM_EDGE (phi, loop_latch_edge (loop)); | |
b9d73ea6 | 3058 | } |
3059 | } | |
3060 | ||
add6ee5e | 3061 | /* Don't issue signed overflow warnings. */ |
3062 | fold_defer_overflow_warnings (); | |
3063 | ||
b9d73ea6 | 3064 | for (i = 0; i < MAX_ITERATIONS_TO_TRACK; i++) |
3065 | { | |
3066 | for (j = 0; j < 2; j++) | |
3067 | aval[j] = get_val_for (op[j], val[j]); | |
3068 | ||
9dfb8b49 | 3069 | acnd = fold_binary (cmp, boolean_type_node, aval[0], aval[1]); |
cd743a11 | 3070 | if (acnd && integer_zerop (acnd)) |
b9d73ea6 | 3071 | { |
add6ee5e | 3072 | fold_undefer_and_ignore_overflow_warnings (); |
b9d73ea6 | 3073 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3074 | fprintf (dump_file, | |
3075 | "Proved that loop %d iterates %d times using brute force.\n", | |
3076 | loop->num, i); | |
7016c612 | 3077 | return build_int_cst (unsigned_type_node, i); |
b9d73ea6 | 3078 | } |
3079 | ||
3080 | for (j = 0; j < 2; j++) | |
f8540d49 | 3081 | { |
13cd9f01 | 3082 | aval[j] = val[j]; |
f8540d49 | 3083 | val[j] = get_val_for (next[j], val[j]); |
3084 | if (!is_gimple_min_invariant (val[j])) | |
add6ee5e | 3085 | { |
3086 | fold_undefer_and_ignore_overflow_warnings (); | |
3087 | return chrec_dont_know; | |
3088 | } | |
f8540d49 | 3089 | } |
13cd9f01 | 3090 | |
3091 | /* If the next iteration would use the same base values | |
3092 | as the current one, there is no point looping further, | |
3093 | all following iterations will be the same as this one. */ | |
3094 | if (val[0] == aval[0] && val[1] == aval[1]) | |
3095 | break; | |
b9d73ea6 | 3096 | } |
3097 | ||
add6ee5e | 3098 | fold_undefer_and_ignore_overflow_warnings (); |
3099 | ||
b9d73ea6 | 3100 | return chrec_dont_know; |
3101 | } | |
3102 | ||
3103 | /* Finds the exit of the LOOP by that the loop exits after a constant | |
3104 | number of iterations and stores the exit edge to *EXIT. The constant | |
3105 | giving the number of iterations of LOOP is returned. The number of | |
3106 | iterations is determined using loop_niter_by_eval (i.e. by brute force | |
3107 | evaluation). If we are unable to find the exit for that loop_niter_by_eval | |
3108 | determines the number of iterations, chrec_dont_know is returned. */ | |
3109 | ||
3110 | tree | |
3111 | find_loop_niter_by_eval (struct loop *loop, edge *exit) | |
3112 | { | |
749ea85f | 3113 | unsigned i; |
f1f41a6c | 3114 | vec<edge> exits = get_loop_exit_edges (loop); |
b9d73ea6 | 3115 | edge ex; |
3116 | tree niter = NULL_TREE, aniter; | |
3117 | ||
3118 | *exit = NULL; | |
a974c7bc | 3119 | |
3120 | /* Loops with multiple exits are expensive to handle and less important. */ | |
3121 | if (!flag_expensive_optimizations | |
f1f41a6c | 3122 | && exits.length () > 1) |
83b709f2 | 3123 | { |
f1f41a6c | 3124 | exits.release (); |
83b709f2 | 3125 | return chrec_dont_know; |
3126 | } | |
a974c7bc | 3127 | |
f1f41a6c | 3128 | FOR_EACH_VEC_ELT (exits, i, ex) |
b9d73ea6 | 3129 | { |
b9d73ea6 | 3130 | if (!just_once_each_iteration_p (loop, ex->src)) |
3131 | continue; | |
3132 | ||
3133 | aniter = loop_niter_by_eval (loop, ex); | |
b091dc59 | 3134 | if (chrec_contains_undetermined (aniter)) |
b9d73ea6 | 3135 | continue; |
3136 | ||
3137 | if (niter | |
b091dc59 | 3138 | && !tree_int_cst_lt (aniter, niter)) |
b9d73ea6 | 3139 | continue; |
3140 | ||
3141 | niter = aniter; | |
3142 | *exit = ex; | |
3143 | } | |
f1f41a6c | 3144 | exits.release (); |
b9d73ea6 | 3145 | |
3146 | return niter ? niter : chrec_dont_know; | |
3147 | } | |
3148 | ||
3149 | /* | |
3150 | ||
3151 | Analysis of upper bounds on number of iterations of a loop. | |
3152 | ||
3153 | */ | |
3154 | ||
5de9d3ed | 3155 | static widest_int derive_constant_upper_bound_ops (tree, tree, |
3156 | enum tree_code, tree); | |
75a70cf9 | 3157 | |
3158 | /* Returns a constant upper bound on the value of the right-hand side of | |
3159 | an assignment statement STMT. */ | |
3160 | ||
5de9d3ed | 3161 | static widest_int |
42acab1c | 3162 | derive_constant_upper_bound_assign (gimple *stmt) |
75a70cf9 | 3163 | { |
3164 | enum tree_code code = gimple_assign_rhs_code (stmt); | |
3165 | tree op0 = gimple_assign_rhs1 (stmt); | |
3166 | tree op1 = gimple_assign_rhs2 (stmt); | |
3167 | ||
3168 | return derive_constant_upper_bound_ops (TREE_TYPE (gimple_assign_lhs (stmt)), | |
3169 | op0, code, op1); | |
3170 | } | |
3171 | ||
d8a99c17 | 3172 | /* Returns a constant upper bound on the value of expression VAL. VAL |
3173 | is considered to be unsigned. If its type is signed, its value must | |
a463eaea | 3174 | be nonnegative. */ |
48e1416a | 3175 | |
5de9d3ed | 3176 | static widest_int |
75a70cf9 | 3177 | derive_constant_upper_bound (tree val) |
3178 | { | |
3179 | enum tree_code code; | |
f9c352a1 | 3180 | tree op0, op1, op2; |
75a70cf9 | 3181 | |
f9c352a1 | 3182 | extract_ops_from_tree (val, &code, &op0, &op1, &op2); |
75a70cf9 | 3183 | return derive_constant_upper_bound_ops (TREE_TYPE (val), op0, code, op1); |
3184 | } | |
3185 | ||
3186 | /* Returns a constant upper bound on the value of expression OP0 CODE OP1, | |
3187 | whose type is TYPE. The expression is considered to be unsigned. If | |
3188 | its type is signed, its value must be nonnegative. */ | |
48e1416a | 3189 | |
5de9d3ed | 3190 | static widest_int |
75a70cf9 | 3191 | derive_constant_upper_bound_ops (tree type, tree op0, |
3192 | enum tree_code code, tree op1) | |
d1baf6ca | 3193 | { |
75a70cf9 | 3194 | tree subtype, maxt; |
36d081ee | 3195 | widest_int bnd, max, cst; |
42acab1c | 3196 | gimple *stmt; |
d8a99c17 | 3197 | |
3198 | if (INTEGRAL_TYPE_P (type)) | |
3199 | maxt = TYPE_MAX_VALUE (type); | |
3200 | else | |
3201 | maxt = upper_bound_in_type (type, type); | |
3202 | ||
5de9d3ed | 3203 | max = wi::to_widest (maxt); |
d8a99c17 | 3204 | |
75a70cf9 | 3205 | switch (code) |
d8a99c17 | 3206 | { |
3207 | case INTEGER_CST: | |
5de9d3ed | 3208 | return wi::to_widest (op0); |
d8a99c17 | 3209 | |
72dd6141 | 3210 | CASE_CONVERT: |
d8a99c17 | 3211 | subtype = TREE_TYPE (op0); |
3212 | if (!TYPE_UNSIGNED (subtype) | |
3213 | /* If TYPE is also signed, the fact that VAL is nonnegative implies | |
3214 | that OP0 is nonnegative. */ | |
3215 | && TYPE_UNSIGNED (type) | |
a463eaea | 3216 | && !tree_expr_nonnegative_p (op0)) |
d8a99c17 | 3217 | { |
3218 | /* If we cannot prove that the casted expression is nonnegative, | |
3219 | we cannot establish more useful upper bound than the precision | |
3220 | of the type gives us. */ | |
3221 | return max; | |
3222 | } | |
d1baf6ca | 3223 | |
d8a99c17 | 3224 | /* We now know that op0 is an nonnegative value. Try deriving an upper |
3225 | bound for it. */ | |
a463eaea | 3226 | bnd = derive_constant_upper_bound (op0); |
d8a99c17 | 3227 | |
3228 | /* If the bound does not fit in TYPE, max. value of TYPE could be | |
3229 | attained. */ | |
796b6678 | 3230 | if (wi::ltu_p (max, bnd)) |
d8a99c17 | 3231 | return max; |
3232 | ||
3233 | return bnd; | |
3234 | ||
3235 | case PLUS_EXPR: | |
0de36bdb | 3236 | case POINTER_PLUS_EXPR: |
d8a99c17 | 3237 | case MINUS_EXPR: |
d8a99c17 | 3238 | if (TREE_CODE (op1) != INTEGER_CST |
a463eaea | 3239 | || !tree_expr_nonnegative_p (op0)) |
d8a99c17 | 3240 | return max; |
3241 | ||
30b9be29 | 3242 | /* Canonicalize to OP0 - CST. Consider CST to be signed, in order to |
3243 | choose the most logical way how to treat this constant regardless | |
3244 | of the signedness of the type. */ | |
5de9d3ed | 3245 | cst = wi::sext (wi::to_widest (op1), TYPE_PRECISION (type)); |
75a70cf9 | 3246 | if (code != MINUS_EXPR) |
cf8f0e63 | 3247 | cst = -cst; |
d8a99c17 | 3248 | |
a463eaea | 3249 | bnd = derive_constant_upper_bound (op0); |
d8a99c17 | 3250 | |
796b6678 | 3251 | if (wi::neg_p (cst)) |
d8a99c17 | 3252 | { |
cf8f0e63 | 3253 | cst = -cst; |
d8a99c17 | 3254 | /* Avoid CST == 0x80000... */ |
796b6678 | 3255 | if (wi::neg_p (cst)) |
7082509e | 3256 | return max; |
d8a99c17 | 3257 | |
30b9be29 | 3258 | /* OP0 + CST. We need to check that |
d8a99c17 | 3259 | BND <= MAX (type) - CST. */ |
3260 | ||
36d081ee | 3261 | widest_int mmax = max - cst; |
3262 | if (wi::leu_p (bnd, mmax)) | |
d8a99c17 | 3263 | return max; |
3264 | ||
cf8f0e63 | 3265 | return bnd + cst; |
d8a99c17 | 3266 | } |
3267 | else | |
3268 | { | |
30b9be29 | 3269 | /* OP0 - CST, where CST >= 0. |
3270 | ||
3271 | If TYPE is signed, we have already verified that OP0 >= 0, and we | |
3272 | know that the result is nonnegative. This implies that | |
3273 | VAL <= BND - CST. | |
3274 | ||
3275 | If TYPE is unsigned, we must additionally know that OP0 >= CST, | |
3276 | otherwise the operation underflows. | |
3277 | */ | |
3278 | ||
3279 | /* This should only happen if the type is unsigned; however, for | |
a463eaea | 3280 | buggy programs that use overflowing signed arithmetics even with |
30b9be29 | 3281 | -fno-wrapv, this condition may also be true for signed values. */ |
796b6678 | 3282 | if (wi::ltu_p (bnd, cst)) |
d8a99c17 | 3283 | return max; |
3284 | ||
a463eaea | 3285 | if (TYPE_UNSIGNED (type)) |
3286 | { | |
3287 | tree tem = fold_binary (GE_EXPR, boolean_type_node, op0, | |
e913b5cd | 3288 | wide_int_to_tree (type, cst)); |
a463eaea | 3289 | if (!tem || integer_nonzerop (tem)) |
3290 | return max; | |
3291 | } | |
30b9be29 | 3292 | |
cf8f0e63 | 3293 | bnd -= cst; |
d8a99c17 | 3294 | } |
3295 | ||
3296 | return bnd; | |
3297 | ||
3298 | case FLOOR_DIV_EXPR: | |
3299 | case EXACT_DIV_EXPR: | |
d8a99c17 | 3300 | if (TREE_CODE (op1) != INTEGER_CST |
3301 | || tree_int_cst_sign_bit (op1)) | |
3302 | return max; | |
3303 | ||
a463eaea | 3304 | bnd = derive_constant_upper_bound (op0); |
5de9d3ed | 3305 | return wi::udiv_floor (bnd, wi::to_widest (op1)); |
d8a99c17 | 3306 | |
faa56cf9 | 3307 | case BIT_AND_EXPR: |
faa56cf9 | 3308 | if (TREE_CODE (op1) != INTEGER_CST |
3309 | || tree_int_cst_sign_bit (op1)) | |
3310 | return max; | |
5de9d3ed | 3311 | return wi::to_widest (op1); |
faa56cf9 | 3312 | |
3313 | case SSA_NAME: | |
75a70cf9 | 3314 | stmt = SSA_NAME_DEF_STMT (op0); |
3315 | if (gimple_code (stmt) != GIMPLE_ASSIGN | |
3316 | || gimple_assign_lhs (stmt) != op0) | |
faa56cf9 | 3317 | return max; |
75a70cf9 | 3318 | return derive_constant_upper_bound_assign (stmt); |
faa56cf9 | 3319 | |
48e1416a | 3320 | default: |
d8a99c17 | 3321 | return max; |
3322 | } | |
d1baf6ca | 3323 | } |
3324 | ||
228bf2b8 | 3325 | /* Emit a -Waggressive-loop-optimizations warning if needed. */ |
3326 | ||
3327 | static void | |
3328 | do_warn_aggressive_loop_optimizations (struct loop *loop, | |
42acab1c | 3329 | widest_int i_bound, gimple *stmt) |
228bf2b8 | 3330 | { |
3331 | /* Don't warn if the loop doesn't have known constant bound. */ | |
3332 | if (!loop->nb_iterations | |
3333 | || TREE_CODE (loop->nb_iterations) != INTEGER_CST | |
3334 | || !warn_aggressive_loop_optimizations | |
3335 | /* To avoid warning multiple times for the same loop, | |
3336 | only start warning when we preserve loops. */ | |
3337 | || (cfun->curr_properties & PROP_loops) == 0 | |
3338 | /* Only warn once per loop. */ | |
3339 | || loop->warned_aggressive_loop_optimizations | |
3340 | /* Only warn if undefined behavior gives us lower estimate than the | |
3341 | known constant bound. */ | |
cd9b5516 | 3342 | || wi::cmpu (i_bound, wi::to_widest (loop->nb_iterations)) >= 0 |
228bf2b8 | 3343 | /* And undefined behavior happens unconditionally. */ |
3344 | || !dominated_by_p (CDI_DOMINATORS, loop->latch, gimple_bb (stmt))) | |
3345 | return; | |
3346 | ||
3347 | edge e = single_exit (loop); | |
3348 | if (e == NULL) | |
3349 | return; | |
3350 | ||
42acab1c | 3351 | gimple *estmt = last_stmt (e->src); |
6ca34fe2 | 3352 | char buf[WIDE_INT_PRINT_BUFFER_SIZE]; |
3353 | print_dec (i_bound, buf, TYPE_UNSIGNED (TREE_TYPE (loop->nb_iterations)) | |
3354 | ? UNSIGNED : SIGNED); | |
bc35ef65 | 3355 | auto_diagnostic_group d; |
531a0a15 | 3356 | if (warning_at (gimple_location (stmt), OPT_Waggressive_loop_optimizations, |
6ca34fe2 | 3357 | "iteration %s invokes undefined behavior", buf)) |
3358 | inform (gimple_location (estmt), "within this loop"); | |
228bf2b8 | 3359 | loop->warned_aggressive_loop_optimizations = true; |
3360 | } | |
3361 | ||
a463eaea | 3362 | /* Records that AT_STMT is executed at most BOUND + 1 times in LOOP. IS_EXIT |
faa56cf9 | 3363 | is true if the loop is exited immediately after STMT, and this exit |
3364 | is taken at last when the STMT is executed BOUND + 1 times. | |
f0b5f617 | 3365 | REALISTIC is true if BOUND is expected to be close to the real number |
4b4ab846 | 3366 | of iterations. UPPER is true if we are sure the loop iterates at most |
10c3fe8d | 3367 | BOUND times. I_BOUND is a widest_int upper estimate on BOUND. */ |
b9d73ea6 | 3368 | |
faa56cf9 | 3369 | static void |
10c3fe8d | 3370 | record_estimate (struct loop *loop, tree bound, const widest_int &i_bound, |
42acab1c | 3371 | gimple *at_stmt, bool is_exit, bool realistic, bool upper) |
b9d73ea6 | 3372 | { |
5de9d3ed | 3373 | widest_int delta; |
b9d73ea6 | 3374 | |
3375 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
3376 | { | |
faa56cf9 | 3377 | fprintf (dump_file, "Statement %s", is_exit ? "(exit)" : ""); |
75a70cf9 | 3378 | print_gimple_stmt (dump_file, at_stmt, 0, TDF_SLIM); |
4b4ab846 | 3379 | fprintf (dump_file, " is %sexecuted at most ", |
3380 | upper ? "" : "probably "); | |
b9d73ea6 | 3381 | print_generic_expr (dump_file, bound, TDF_SLIM); |
d1baf6ca | 3382 | fprintf (dump_file, " (bounded by "); |
e913b5cd | 3383 | print_decu (i_bound, dump_file); |
faa56cf9 | 3384 | fprintf (dump_file, ") + 1 times in loop %d.\n", loop->num); |
b9d73ea6 | 3385 | } |
3386 | ||
4b4ab846 | 3387 | /* If the I_BOUND is just an estimate of BOUND, it rarely is close to the |
3388 | real number of iterations. */ | |
3389 | if (TREE_CODE (bound) != INTEGER_CST) | |
3390 | realistic = false; | |
72276d01 | 3391 | else |
5de9d3ed | 3392 | gcc_checking_assert (i_bound == wi::to_widest (bound)); |
4b4ab846 | 3393 | |
3394 | /* If we have a guaranteed upper bound, record it in the appropriate | |
228bf2b8 | 3395 | list, unless this is an !is_exit bound (i.e. undefined behavior in |
3396 | at_stmt) in a loop with known constant number of iterations. */ | |
3397 | if (upper | |
3398 | && (is_exit | |
3399 | || loop->nb_iterations == NULL_TREE | |
3400 | || TREE_CODE (loop->nb_iterations) != INTEGER_CST)) | |
4b4ab846 | 3401 | { |
25a27413 | 3402 | struct nb_iter_bound *elt = ggc_alloc<nb_iter_bound> (); |
4b4ab846 | 3403 | |
3404 | elt->bound = i_bound; | |
3405 | elt->stmt = at_stmt; | |
3406 | elt->is_exit = is_exit; | |
3407 | elt->next = loop->bounds; | |
3408 | loop->bounds = elt; | |
3409 | } | |
3410 | ||
fd943c70 | 3411 | /* If statement is executed on every path to the loop latch, we can directly |
3412 | infer the upper bound on the # of iterations of the loop. */ | |
3413 | if (!dominated_by_p (CDI_DOMINATORS, loop->latch, gimple_bb (at_stmt))) | |
8e3ffe30 | 3414 | upper = false; |
fd943c70 | 3415 | |
4b4ab846 | 3416 | /* Update the number of iteration estimates according to the bound. |
e3606104 | 3417 | If at_stmt is an exit then the loop latch is executed at most BOUND times, |
3418 | otherwise it can be executed BOUND + 1 times. We will lower the estimate | |
3419 | later if such statement must be executed on last iteration */ | |
3420 | if (is_exit) | |
e913b5cd | 3421 | delta = 0; |
4b4ab846 | 3422 | else |
e913b5cd | 3423 | delta = 1; |
10c3fe8d | 3424 | widest_int new_i_bound = i_bound + delta; |
4b4ab846 | 3425 | |
efec32e0 | 3426 | /* If an overflow occurred, ignore the result. */ |
10c3fe8d | 3427 | if (wi::ltu_p (new_i_bound, delta)) |
4b4ab846 | 3428 | return; |
3429 | ||
228bf2b8 | 3430 | if (upper && !is_exit) |
10c3fe8d | 3431 | do_warn_aggressive_loop_optimizations (loop, new_i_bound, at_stmt); |
3432 | record_niter_bound (loop, new_i_bound, realistic, upper); | |
b9d73ea6 | 3433 | } |
3434 | ||
05032b10 | 3435 | /* Records the control iv analyzed in NITER for LOOP if the iv is valid |
3436 | and doesn't overflow. */ | |
3437 | ||
3438 | static void | |
3439 | record_control_iv (struct loop *loop, struct tree_niter_desc *niter) | |
3440 | { | |
3441 | struct control_iv *iv; | |
3442 | ||
3443 | if (!niter->control.base || !niter->control.step) | |
3444 | return; | |
3445 | ||
3446 | if (!integer_onep (niter->assumptions) || !niter->control.no_overflow) | |
3447 | return; | |
3448 | ||
3449 | iv = ggc_alloc<control_iv> (); | |
3450 | iv->base = niter->control.base; | |
3451 | iv->step = niter->control.step; | |
3452 | iv->next = loop->control_ivs; | |
3453 | loop->control_ivs = iv; | |
3454 | ||
3455 | return; | |
3456 | } | |
3457 | ||
cf1b1fd7 | 3458 | /* This function returns TRUE if below conditions are satisfied: |
3459 | 1) VAR is SSA variable. | |
3460 | 2) VAR is an IV:{base, step} in its defining loop. | |
3461 | 3) IV doesn't overflow. | |
3462 | 4) Both base and step are integer constants. | |
3463 | 5) Base is the MIN/MAX value depends on IS_MIN. | |
3464 | Store value of base to INIT correspondingly. */ | |
3465 | ||
3466 | static bool | |
3467 | get_cst_init_from_scev (tree var, wide_int *init, bool is_min) | |
3468 | { | |
3469 | if (TREE_CODE (var) != SSA_NAME) | |
3470 | return false; | |
3471 | ||
3472 | gimple *def_stmt = SSA_NAME_DEF_STMT (var); | |
3473 | struct loop *loop = loop_containing_stmt (def_stmt); | |
3474 | ||
3475 | if (loop == NULL) | |
3476 | return false; | |
3477 | ||
3478 | affine_iv iv; | |
3479 | if (!simple_iv (loop, loop, var, &iv, false)) | |
3480 | return false; | |
3481 | ||
3482 | if (!iv.no_overflow) | |
3483 | return false; | |
3484 | ||
3485 | if (TREE_CODE (iv.base) != INTEGER_CST || TREE_CODE (iv.step) != INTEGER_CST) | |
3486 | return false; | |
3487 | ||
3488 | if (is_min == tree_int_cst_sign_bit (iv.step)) | |
3489 | return false; | |
3490 | ||
e3d0f65c | 3491 | *init = wi::to_wide (iv.base); |
cf1b1fd7 | 3492 | return true; |
3493 | } | |
3494 | ||
faa56cf9 | 3495 | /* Record the estimate on number of iterations of LOOP based on the fact that |
3496 | the induction variable BASE + STEP * i evaluated in STMT does not wrap and | |
4b4ab846 | 3497 | its values belong to the range <LOW, HIGH>. REALISTIC is true if the |
3498 | estimated number of iterations is expected to be close to the real one. | |
3499 | UPPER is true if we are sure the induction variable does not wrap. */ | |
faa56cf9 | 3500 | |
3501 | static void | |
42acab1c | 3502 | record_nonwrapping_iv (struct loop *loop, tree base, tree step, gimple *stmt, |
4b4ab846 | 3503 | tree low, tree high, bool realistic, bool upper) |
faa56cf9 | 3504 | { |
3505 | tree niter_bound, extreme, delta; | |
3506 | tree type = TREE_TYPE (base), unsigned_type; | |
97178c4f | 3507 | tree orig_base = base; |
faa56cf9 | 3508 | |
cd743a11 | 3509 | if (TREE_CODE (step) != INTEGER_CST || integer_zerop (step)) |
faa56cf9 | 3510 | return; |
3511 | ||
3512 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
3513 | { | |
3514 | fprintf (dump_file, "Induction variable ("); | |
3515 | print_generic_expr (dump_file, TREE_TYPE (base), TDF_SLIM); | |
3516 | fprintf (dump_file, ") "); | |
3517 | print_generic_expr (dump_file, base, TDF_SLIM); | |
3518 | fprintf (dump_file, " + "); | |
3519 | print_generic_expr (dump_file, step, TDF_SLIM); | |
3520 | fprintf (dump_file, " * iteration does not wrap in statement "); | |
75a70cf9 | 3521 | print_gimple_stmt (dump_file, stmt, 0, TDF_SLIM); |
faa56cf9 | 3522 | fprintf (dump_file, " in loop %d.\n", loop->num); |
3523 | } | |
3524 | ||
3525 | unsigned_type = unsigned_type_for (type); | |
3526 | base = fold_convert (unsigned_type, base); | |
3527 | step = fold_convert (unsigned_type, step); | |
3528 | ||
3529 | if (tree_int_cst_sign_bit (step)) | |
3530 | { | |
97178c4f | 3531 | wide_int min, max; |
faa56cf9 | 3532 | extreme = fold_convert (unsigned_type, low); |
97178c4f | 3533 | if (TREE_CODE (orig_base) == SSA_NAME |
3534 | && TREE_CODE (high) == INTEGER_CST | |
3535 | && INTEGRAL_TYPE_P (TREE_TYPE (orig_base)) | |
cf1b1fd7 | 3536 | && (get_range_info (orig_base, &min, &max) == VR_RANGE |
3537 | || get_cst_init_from_scev (orig_base, &max, false)) | |
e3d0f65c | 3538 | && wi::gts_p (wi::to_wide (high), max)) |
97178c4f | 3539 | base = wide_int_to_tree (unsigned_type, max); |
36d081ee | 3540 | else if (TREE_CODE (base) != INTEGER_CST |
3541 | && dominated_by_p (CDI_DOMINATORS, | |
3542 | loop->latch, gimple_bb (stmt))) | |
faa56cf9 | 3543 | base = fold_convert (unsigned_type, high); |
3544 | delta = fold_build2 (MINUS_EXPR, unsigned_type, base, extreme); | |
3545 | step = fold_build1 (NEGATE_EXPR, unsigned_type, step); | |
3546 | } | |
3547 | else | |
3548 | { | |
97178c4f | 3549 | wide_int min, max; |
faa56cf9 | 3550 | extreme = fold_convert (unsigned_type, high); |
97178c4f | 3551 | if (TREE_CODE (orig_base) == SSA_NAME |
3552 | && TREE_CODE (low) == INTEGER_CST | |
3553 | && INTEGRAL_TYPE_P (TREE_TYPE (orig_base)) | |
cf1b1fd7 | 3554 | && (get_range_info (orig_base, &min, &max) == VR_RANGE |
3555 | || get_cst_init_from_scev (orig_base, &min, true)) | |
e3d0f65c | 3556 | && wi::gts_p (min, wi::to_wide (low))) |
97178c4f | 3557 | base = wide_int_to_tree (unsigned_type, min); |
36d081ee | 3558 | else if (TREE_CODE (base) != INTEGER_CST |
3559 | && dominated_by_p (CDI_DOMINATORS, | |
3560 | loop->latch, gimple_bb (stmt))) | |
faa56cf9 | 3561 | base = fold_convert (unsigned_type, low); |
3562 | delta = fold_build2 (MINUS_EXPR, unsigned_type, extreme, base); | |
3563 | } | |
3564 | ||
3565 | /* STMT is executed at most NITER_BOUND + 1 times, since otherwise the value | |
3566 | would get out of the range. */ | |
3567 | niter_bound = fold_build2 (FLOOR_DIV_EXPR, unsigned_type, delta, step); | |
ab2c1de8 | 3568 | widest_int max = derive_constant_upper_bound (niter_bound); |
4b4ab846 | 3569 | record_estimate (loop, niter_bound, max, stmt, false, realistic, upper); |
d500fef3 | 3570 | } |
3571 | ||
faa56cf9 | 3572 | /* Determine information about number of iterations a LOOP from the index |
d97e22fb | 3573 | IDX of a data reference accessed in STMT. RELIABLE is true if STMT is |
3574 | guaranteed to be executed in every iteration of LOOP. Callback for | |
3575 | for_each_index. */ | |
faa56cf9 | 3576 | |
3577 | struct ilb_data | |
3578 | { | |
3579 | struct loop *loop; | |
42acab1c | 3580 | gimple *stmt; |
faa56cf9 | 3581 | }; |
3582 | ||
3583 | static bool | |
3584 | idx_infer_loop_bounds (tree base, tree *idx, void *dta) | |
3585 | { | |
f0d6e81c | 3586 | struct ilb_data *data = (struct ilb_data *) dta; |
faa56cf9 | 3587 | tree ev, init, step; |
3588 | tree low, high, type, next; | |
fd943c70 | 3589 | bool sign, upper = true, at_end = false; |
faa56cf9 | 3590 | struct loop *loop = data->loop; |
3591 | ||
4b4ab846 | 3592 | if (TREE_CODE (base) != ARRAY_REF) |
faa56cf9 | 3593 | return true; |
3594 | ||
4b4ab846 | 3595 | /* For arrays at the end of the structure, we are not guaranteed that they |
3596 | do not really extend over their declared size. However, for arrays of | |
3597 | size greater than one, this is unlikely to be intended. */ | |
3598 | if (array_at_struct_end_p (base)) | |
d97e22fb | 3599 | { |
3600 | at_end = true; | |
3601 | upper = false; | |
3602 | } | |
4b4ab846 | 3603 | |
6ee802ad | 3604 | struct loop *dloop = loop_containing_stmt (data->stmt); |
3605 | if (!dloop) | |
3606 | return true; | |
3607 | ||
3608 | ev = analyze_scalar_evolution (dloop, *idx); | |
3609 | ev = instantiate_parameters (loop, ev); | |
faa56cf9 | 3610 | init = initial_condition (ev); |
3611 | step = evolution_part_in_loop_num (ev, loop->num); | |
3612 | ||
3613 | if (!init | |
3614 | || !step | |
3615 | || TREE_CODE (step) != INTEGER_CST | |
cd743a11 | 3616 | || integer_zerop (step) |
faa56cf9 | 3617 | || tree_contains_chrecs (init, NULL) |
3618 | || chrec_contains_symbols_defined_in_loop (init, loop->num)) | |
3619 | return true; | |
3620 | ||
3621 | low = array_ref_low_bound (base); | |
3622 | high = array_ref_up_bound (base); | |
48e1416a | 3623 | |
faa56cf9 | 3624 | /* The case of nonconstant bounds could be handled, but it would be |
3625 | complicated. */ | |
3626 | if (TREE_CODE (low) != INTEGER_CST | |
3627 | || !high | |
3628 | || TREE_CODE (high) != INTEGER_CST) | |
3629 | return true; | |
3630 | sign = tree_int_cst_sign_bit (step); | |
3631 | type = TREE_TYPE (step); | |
4b4ab846 | 3632 | |
3633 | /* The array of length 1 at the end of a structure most likely extends | |
3634 | beyond its bounds. */ | |
d97e22fb | 3635 | if (at_end |
4b4ab846 | 3636 | && operand_equal_p (low, high, 0)) |
3637 | return true; | |
3638 | ||
faa56cf9 | 3639 | /* In case the relevant bound of the array does not fit in type, or |
3640 | it does, but bound + step (in type) still belongs into the range of the | |
3641 | array, the index may wrap and still stay within the range of the array | |
3642 | (consider e.g. if the array is indexed by the full range of | |
3643 | unsigned char). | |
3644 | ||
3645 | To make things simpler, we require both bounds to fit into type, although | |
fa7637bd | 3646 | there are cases where this would not be strictly necessary. */ |
faa56cf9 | 3647 | if (!int_fits_type_p (high, type) |
3648 | || !int_fits_type_p (low, type)) | |
3649 | return true; | |
3650 | low = fold_convert (type, low); | |
3651 | high = fold_convert (type, high); | |
3652 | ||
3653 | if (sign) | |
3654 | next = fold_binary (PLUS_EXPR, type, low, step); | |
3655 | else | |
3656 | next = fold_binary (PLUS_EXPR, type, high, step); | |
48e1416a | 3657 | |
faa56cf9 | 3658 | if (tree_int_cst_compare (low, next) <= 0 |
3659 | && tree_int_cst_compare (next, high) <= 0) | |
3660 | return true; | |
3661 | ||
9a2ace10 | 3662 | /* If access is not executed on every iteration, we must ensure that overlow |
3663 | may not make the access valid later. */ | |
64552007 | 3664 | if (!dominated_by_p (CDI_DOMINATORS, loop->latch, gimple_bb (data->stmt)) |
90261646 | 3665 | && scev_probably_wraps_p (NULL_TREE, |
3666 | initial_condition_in_loop_num (ev, loop->num), | |
64552007 | 3667 | step, data->stmt, loop, true)) |
9a2ace10 | 3668 | upper = false; |
64552007 | 3669 | |
9a2ace10 | 3670 | record_nonwrapping_iv (loop, init, step, data->stmt, low, high, false, upper); |
faa56cf9 | 3671 | return true; |
3672 | } | |
3673 | ||
3674 | /* Determine information about number of iterations a LOOP from the bounds | |
d97e22fb | 3675 | of arrays in the data reference REF accessed in STMT. RELIABLE is true if |
3676 | STMT is guaranteed to be executed in every iteration of LOOP.*/ | |
faa56cf9 | 3677 | |
3678 | static void | |
42acab1c | 3679 | infer_loop_bounds_from_ref (struct loop *loop, gimple *stmt, tree ref) |
faa56cf9 | 3680 | { |
3681 | struct ilb_data data; | |
3682 | ||
3683 | data.loop = loop; | |
3684 | data.stmt = stmt; | |
3685 | for_each_index (&ref, idx_infer_loop_bounds, &data); | |
3686 | } | |
3687 | ||
3688 | /* Determine information about number of iterations of a LOOP from the way | |
d97e22fb | 3689 | arrays are used in STMT. RELIABLE is true if STMT is guaranteed to be |
3690 | executed in every iteration of LOOP. */ | |
faa56cf9 | 3691 | |
3692 | static void | |
42acab1c | 3693 | infer_loop_bounds_from_array (struct loop *loop, gimple *stmt) |
faa56cf9 | 3694 | { |
75a70cf9 | 3695 | if (is_gimple_assign (stmt)) |
faa56cf9 | 3696 | { |
75a70cf9 | 3697 | tree op0 = gimple_assign_lhs (stmt); |
3698 | tree op1 = gimple_assign_rhs1 (stmt); | |
faa56cf9 | 3699 | |
3700 | /* For each memory access, analyze its access function | |
3701 | and record a bound on the loop iteration domain. */ | |
3702 | if (REFERENCE_CLASS_P (op0)) | |
fd943c70 | 3703 | infer_loop_bounds_from_ref (loop, stmt, op0); |
faa56cf9 | 3704 | |
3705 | if (REFERENCE_CLASS_P (op1)) | |
fd943c70 | 3706 | infer_loop_bounds_from_ref (loop, stmt, op1); |
faa56cf9 | 3707 | } |
75a70cf9 | 3708 | else if (is_gimple_call (stmt)) |
faa56cf9 | 3709 | { |
75a70cf9 | 3710 | tree arg, lhs; |
3711 | unsigned i, n = gimple_call_num_args (stmt); | |
faa56cf9 | 3712 | |
75a70cf9 | 3713 | lhs = gimple_call_lhs (stmt); |
3714 | if (lhs && REFERENCE_CLASS_P (lhs)) | |
fd943c70 | 3715 | infer_loop_bounds_from_ref (loop, stmt, lhs); |
75a70cf9 | 3716 | |
3717 | for (i = 0; i < n; i++) | |
3718 | { | |
3719 | arg = gimple_call_arg (stmt, i); | |
3720 | if (REFERENCE_CLASS_P (arg)) | |
fd943c70 | 3721 | infer_loop_bounds_from_ref (loop, stmt, arg); |
75a70cf9 | 3722 | } |
faa56cf9 | 3723 | } |
3724 | } | |
3725 | ||
dba7464a | 3726 | /* Determine information about number of iterations of a LOOP from the fact |
3727 | that pointer arithmetics in STMT does not overflow. */ | |
3728 | ||
3729 | static void | |
42acab1c | 3730 | infer_loop_bounds_from_pointer_arith (struct loop *loop, gimple *stmt) |
dba7464a | 3731 | { |
3732 | tree def, base, step, scev, type, low, high; | |
3733 | tree var, ptr; | |
3734 | ||
3735 | if (!is_gimple_assign (stmt) | |
3736 | || gimple_assign_rhs_code (stmt) != POINTER_PLUS_EXPR) | |
3737 | return; | |
3738 | ||
3739 | def = gimple_assign_lhs (stmt); | |
3740 | if (TREE_CODE (def) != SSA_NAME) | |
3741 | return; | |
3742 | ||
3743 | type = TREE_TYPE (def); | |
3744 | if (!nowrap_type_p (type)) | |
3745 | return; | |
3746 | ||
3747 | ptr = gimple_assign_rhs1 (stmt); | |
3748 | if (!expr_invariant_in_loop_p (loop, ptr)) | |
3749 | return; | |
3750 | ||
3751 | var = gimple_assign_rhs2 (stmt); | |
3752 | if (TYPE_PRECISION (type) != TYPE_PRECISION (TREE_TYPE (var))) | |
3753 | return; | |
3754 | ||
8427cc63 | 3755 | struct loop *uloop = loop_containing_stmt (stmt); |
3756 | scev = instantiate_parameters (loop, analyze_scalar_evolution (uloop, def)); | |
dba7464a | 3757 | if (chrec_contains_undetermined (scev)) |
3758 | return; | |
3759 | ||
3760 | base = initial_condition_in_loop_num (scev, loop->num); | |
3761 | step = evolution_part_in_loop_num (scev, loop->num); | |
3762 | ||
3763 | if (!base || !step | |
3764 | || TREE_CODE (step) != INTEGER_CST | |
3765 | || tree_contains_chrecs (base, NULL) | |
3766 | || chrec_contains_symbols_defined_in_loop (base, loop->num)) | |
3767 | return; | |
3768 | ||
3769 | low = lower_bound_in_type (type, type); | |
3770 | high = upper_bound_in_type (type, type); | |
3771 | ||
8767ae9c | 3772 | /* In C, pointer arithmetic p + 1 cannot use a NULL pointer, and p - 1 cannot |
3773 | produce a NULL pointer. The contrary would mean NULL points to an object, | |
3774 | while NULL is supposed to compare unequal with the address of all objects. | |
3775 | Furthermore, p + 1 cannot produce a NULL pointer and p - 1 cannot use a | |
3776 | NULL pointer since that would mean wrapping, which we assume here not to | |
3777 | happen. So, we can exclude NULL from the valid range of pointer | |
3778 | arithmetic. */ | |
3779 | if (flag_delete_null_pointer_checks && int_cst_value (low) == 0) | |
3780 | low = build_int_cstu (TREE_TYPE (low), TYPE_ALIGN_UNIT (TREE_TYPE (type))); | |
3781 | ||
dba7464a | 3782 | record_nonwrapping_iv (loop, base, step, stmt, low, high, false, true); |
3783 | } | |
3784 | ||
faa56cf9 | 3785 | /* Determine information about number of iterations of a LOOP from the fact |
3786 | that signed arithmetics in STMT does not overflow. */ | |
3787 | ||
3788 | static void | |
42acab1c | 3789 | infer_loop_bounds_from_signedness (struct loop *loop, gimple *stmt) |
faa56cf9 | 3790 | { |
3791 | tree def, base, step, scev, type, low, high; | |
3792 | ||
75a70cf9 | 3793 | if (gimple_code (stmt) != GIMPLE_ASSIGN) |
faa56cf9 | 3794 | return; |
3795 | ||
75a70cf9 | 3796 | def = gimple_assign_lhs (stmt); |
faa56cf9 | 3797 | |
3798 | if (TREE_CODE (def) != SSA_NAME) | |
3799 | return; | |
3800 | ||
3801 | type = TREE_TYPE (def); | |
3802 | if (!INTEGRAL_TYPE_P (type) | |
981eb798 | 3803 | || !TYPE_OVERFLOW_UNDEFINED (type)) |
faa56cf9 | 3804 | return; |
3805 | ||
3806 | scev = instantiate_parameters (loop, analyze_scalar_evolution (loop, def)); | |
3807 | if (chrec_contains_undetermined (scev)) | |
3808 | return; | |
3809 | ||
3810 | base = initial_condition_in_loop_num (scev, loop->num); | |
3811 | step = evolution_part_in_loop_num (scev, loop->num); | |
3812 | ||
3813 | if (!base || !step | |
3814 | || TREE_CODE (step) != INTEGER_CST | |
3815 | || tree_contains_chrecs (base, NULL) | |
3816 | || chrec_contains_symbols_defined_in_loop (base, loop->num)) | |
3817 | return; | |
3818 | ||
3819 | low = lower_bound_in_type (type, type); | |
3820 | high = upper_bound_in_type (type, type); | |
0b840c38 | 3821 | wide_int minv, maxv; |
3822 | if (get_range_info (def, &minv, &maxv) == VR_RANGE) | |
3823 | { | |
3824 | low = wide_int_to_tree (type, minv); | |
3825 | high = wide_int_to_tree (type, maxv); | |
3826 | } | |
faa56cf9 | 3827 | |
4b4ab846 | 3828 | record_nonwrapping_iv (loop, base, step, stmt, low, high, false, true); |
faa56cf9 | 3829 | } |
3830 | ||
903dae48 | 3831 | /* The following analyzers are extracting informations on the bounds |
3832 | of LOOP from the following undefined behaviors: | |
3833 | ||
3834 | - data references should not access elements over the statically | |
3835 | allocated size, | |
3836 | ||
3837 | - signed variables should not overflow when flag_wrapv is not set. | |
3838 | */ | |
3839 | ||
3840 | static void | |
3841 | infer_loop_bounds_from_undefined (struct loop *loop) | |
3842 | { | |
3843 | unsigned i; | |
faa56cf9 | 3844 | basic_block *bbs; |
75a70cf9 | 3845 | gimple_stmt_iterator bsi; |
faa56cf9 | 3846 | basic_block bb; |
d97e22fb | 3847 | bool reliable; |
48e1416a | 3848 | |
903dae48 | 3849 | bbs = get_loop_body (loop); |
3850 | ||
3851 | for (i = 0; i < loop->num_nodes; i++) | |
3852 | { | |
3853 | bb = bbs[i]; | |
3854 | ||
faa56cf9 | 3855 | /* If BB is not executed in each iteration of the loop, we cannot |
d97e22fb | 3856 | use the operations in it to infer reliable upper bound on the |
fd943c70 | 3857 | # of iterations of the loop. However, we can use it as a guess. |
3858 | Reliable guesses come only from array bounds. */ | |
d97e22fb | 3859 | reliable = dominated_by_p (CDI_DOMINATORS, loop->latch, bb); |
faa56cf9 | 3860 | |
75a70cf9 | 3861 | for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi)) |
faa56cf9 | 3862 | { |
42acab1c | 3863 | gimple *stmt = gsi_stmt (bsi); |
903dae48 | 3864 | |
fd943c70 | 3865 | infer_loop_bounds_from_array (loop, stmt); |
d97e22fb | 3866 | |
3867 | if (reliable) | |
dba7464a | 3868 | { |
3869 | infer_loop_bounds_from_signedness (loop, stmt); | |
3870 | infer_loop_bounds_from_pointer_arith (loop, stmt); | |
3871 | } | |
faa56cf9 | 3872 | } |
3873 | ||
903dae48 | 3874 | } |
3875 | ||
3876 | free (bbs); | |
3877 | } | |
3878 | ||
e913b5cd | 3879 | /* Compare wide ints, callback for qsort. */ |
6689221f | 3880 | |
f86b328b | 3881 | static int |
e913b5cd | 3882 | wide_int_cmp (const void *p1, const void *p2) |
6689221f | 3883 | { |
5de9d3ed | 3884 | const widest_int *d1 = (const widest_int *) p1; |
3885 | const widest_int *d2 = (const widest_int *) p2; | |
796b6678 | 3886 | return wi::cmpu (*d1, *d2); |
6689221f | 3887 | } |
3888 | ||
3889 | /* Return index of BOUND in BOUNDS array sorted in increasing order. | |
3890 | Lookup by binary search. */ | |
3891 | ||
f86b328b | 3892 | static int |
5de9d3ed | 3893 | bound_index (vec<widest_int> bounds, const widest_int &bound) |
6689221f | 3894 | { |
f1f41a6c | 3895 | unsigned int end = bounds.length (); |
6689221f | 3896 | unsigned int begin = 0; |
3897 | ||
3898 | /* Find a matching index by means of a binary search. */ | |
3899 | while (begin != end) | |
3900 | { | |
3901 | unsigned int middle = (begin + end) / 2; | |
5de9d3ed | 3902 | widest_int index = bounds[middle]; |
6689221f | 3903 | |
3904 | if (index == bound) | |
3905 | return middle; | |
796b6678 | 3906 | else if (wi::ltu_p (index, bound)) |
6689221f | 3907 | begin = middle + 1; |
3908 | else | |
3909 | end = middle; | |
3910 | } | |
3911 | gcc_unreachable (); | |
3912 | } | |
3913 | ||
6689221f | 3914 | /* We recorded loop bounds only for statements dominating loop latch (and thus |
3915 | executed each loop iteration). If there are any bounds on statements not | |
3916 | dominating the loop latch we can improve the estimate by walking the loop | |
3917 | body and seeing if every path from loop header to loop latch contains | |
3918 | some bounded statement. */ | |
3919 | ||
3920 | static void | |
3921 | discover_iteration_bound_by_body_walk (struct loop *loop) | |
3922 | { | |
6689221f | 3923 | struct nb_iter_bound *elt; |
e9258aee | 3924 | auto_vec<widest_int> bounds; |
b71850a1 | 3925 | vec<vec<basic_block> > queues = vNULL; |
3926 | vec<basic_block> queue = vNULL; | |
6689221f | 3927 | ptrdiff_t queue_index; |
3928 | ptrdiff_t latch_index = 0; | |
6689221f | 3929 | |
3930 | /* Discover what bounds may interest us. */ | |
3931 | for (elt = loop->bounds; elt; elt = elt->next) | |
3932 | { | |
5de9d3ed | 3933 | widest_int bound = elt->bound; |
6689221f | 3934 | |
3935 | /* Exit terminates loop at given iteration, while non-exits produce undefined | |
3936 | effect on the next iteration. */ | |
3937 | if (!elt->is_exit) | |
cc2263b1 | 3938 | { |
e913b5cd | 3939 | bound += 1; |
cc2263b1 | 3940 | /* If an overflow occurred, ignore the result. */ |
796b6678 | 3941 | if (bound == 0) |
cc2263b1 | 3942 | continue; |
3943 | } | |
6689221f | 3944 | |
3945 | if (!loop->any_upper_bound | |
796b6678 | 3946 | || wi::ltu_p (bound, loop->nb_iterations_upper_bound)) |
f1f41a6c | 3947 | bounds.safe_push (bound); |
6689221f | 3948 | } |
3949 | ||
3950 | /* Exit early if there is nothing to do. */ | |
f1f41a6c | 3951 | if (!bounds.exists ()) |
6689221f | 3952 | return; |
3953 | ||
3954 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
3955 | fprintf (dump_file, " Trying to walk loop body to reduce the bound.\n"); | |
3956 | ||
3957 | /* Sort the bounds in decreasing order. */ | |
90a2d741 | 3958 | bounds.qsort (wide_int_cmp); |
6689221f | 3959 | |
3960 | /* For every basic block record the lowest bound that is guaranteed to | |
3961 | terminate the loop. */ | |
3962 | ||
5f8841a5 | 3963 | hash_map<basic_block, ptrdiff_t> bb_bounds; |
6689221f | 3964 | for (elt = loop->bounds; elt; elt = elt->next) |
3965 | { | |
5de9d3ed | 3966 | widest_int bound = elt->bound; |
6689221f | 3967 | if (!elt->is_exit) |
cc2263b1 | 3968 | { |
e913b5cd | 3969 | bound += 1; |
cc2263b1 | 3970 | /* If an overflow occurred, ignore the result. */ |
796b6678 | 3971 | if (bound == 0) |
cc2263b1 | 3972 | continue; |
3973 | } | |
6689221f | 3974 | |
3975 | if (!loop->any_upper_bound | |
796b6678 | 3976 | || wi::ltu_p (bound, loop->nb_iterations_upper_bound)) |
6689221f | 3977 | { |
3978 | ptrdiff_t index = bound_index (bounds, bound); | |
5f8841a5 | 3979 | ptrdiff_t *entry = bb_bounds.get (gimple_bb (elt->stmt)); |
6689221f | 3980 | if (!entry) |
5f8841a5 | 3981 | bb_bounds.put (gimple_bb (elt->stmt), index); |
6689221f | 3982 | else if ((ptrdiff_t)*entry > index) |
5f8841a5 | 3983 | *entry = index; |
6689221f | 3984 | } |
3985 | } | |
3986 | ||
5f8841a5 | 3987 | hash_map<basic_block, ptrdiff_t> block_priority; |
6689221f | 3988 | |
3989 | /* Perform shortest path discovery loop->header ... loop->latch. | |
3990 | ||
3991 | The "distance" is given by the smallest loop bound of basic block | |
3992 | present in the path and we look for path with largest smallest bound | |
3993 | on it. | |
3994 | ||
b71850a1 | 3995 | To avoid the need for fibonacci heap on double ints we simply compress |
6689221f | 3996 | double ints into indexes to BOUNDS array and then represent the queue |
3997 | as arrays of queues for every index. | |
f1f41a6c | 3998 | Index of BOUNDS.length() means that the execution of given BB has |
6689221f | 3999 | no bounds determined. |
4000 | ||
4001 | VISITED is a pointer map translating basic block into smallest index | |
4002 | it was inserted into the priority queue with. */ | |
4003 | latch_index = -1; | |
4004 | ||
4005 | /* Start walk in loop header with index set to infinite bound. */ | |
f1f41a6c | 4006 | queue_index = bounds.length (); |
4007 | queues.safe_grow_cleared (queue_index + 1); | |
4008 | queue.safe_push (loop->header); | |
4009 | queues[queue_index] = queue; | |
5f8841a5 | 4010 | block_priority.put (loop->header, queue_index); |
6689221f | 4011 | |
4012 | for (; queue_index >= 0; queue_index--) | |
4013 | { | |
4014 | if (latch_index < queue_index) | |
4015 | { | |
f1f41a6c | 4016 | while (queues[queue_index].length ()) |
6689221f | 4017 | { |
4018 | basic_block bb; | |
4019 | ptrdiff_t bound_index = queue_index; | |
6689221f | 4020 | edge e; |
4021 | edge_iterator ei; | |
4022 | ||
f1f41a6c | 4023 | queue = queues[queue_index]; |
4024 | bb = queue.pop (); | |
6689221f | 4025 | |
4026 | /* OK, we later inserted the BB with lower priority, skip it. */ | |
5f8841a5 | 4027 | if (*block_priority.get (bb) > queue_index) |
6689221f | 4028 | continue; |
4029 | ||
4030 | /* See if we can improve the bound. */ | |
5f8841a5 | 4031 | ptrdiff_t *entry = bb_bounds.get (bb); |
4032 | if (entry && *entry < bound_index) | |
4033 | bound_index = *entry; | |
6689221f | 4034 | |
4035 | /* Insert succesors into the queue, watch for latch edge | |
4036 | and record greatest index we saw. */ | |
4037 | FOR_EACH_EDGE (e, ei, bb->succs) | |
4038 | { | |
4039 | bool insert = false; | |
6689221f | 4040 | |
4041 | if (loop_exit_edge_p (loop, e)) | |
4042 | continue; | |
4043 | ||
4044 | if (e == loop_latch_edge (loop) | |
4045 | && latch_index < bound_index) | |
4046 | latch_index = bound_index; | |
5f8841a5 | 4047 | else if (!(entry = block_priority.get (e->dest))) |
6689221f | 4048 | { |
4049 | insert = true; | |
5f8841a5 | 4050 | block_priority.put (e->dest, bound_index); |
6689221f | 4051 | } |
5f8841a5 | 4052 | else if (*entry < bound_index) |
6689221f | 4053 | { |
4054 | insert = true; | |
5f8841a5 | 4055 | *entry = bound_index; |
6689221f | 4056 | } |
4057 | ||
4058 | if (insert) | |
b71850a1 | 4059 | queues[bound_index].safe_push (e->dest); |
6689221f | 4060 | } |
4061 | } | |
4062 | } | |
b71850a1 | 4063 | queues[queue_index].release (); |
6689221f | 4064 | } |
4065 | ||
4066 | gcc_assert (latch_index >= 0); | |
f1f41a6c | 4067 | if ((unsigned)latch_index < bounds.length ()) |
6689221f | 4068 | { |
4069 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
4070 | { | |
4071 | fprintf (dump_file, "Found better loop bound "); | |
e913b5cd | 4072 | print_decu (bounds[latch_index], dump_file); |
6689221f | 4073 | fprintf (dump_file, "\n"); |
4074 | } | |
f1f41a6c | 4075 | record_niter_bound (loop, bounds[latch_index], false, true); |
6689221f | 4076 | } |
4077 | ||
f1f41a6c | 4078 | queues.release (); |
6689221f | 4079 | } |
4080 | ||
e3606104 | 4081 | /* See if every path cross the loop goes through a statement that is known |
4082 | to not execute at the last iteration. In that case we can decrese iteration | |
4083 | count by 1. */ | |
4084 | ||
4085 | static void | |
4086 | maybe_lower_iteration_bound (struct loop *loop) | |
4087 | { | |
42acab1c | 4088 | hash_set<gimple *> *not_executed_last_iteration = NULL; |
e3606104 | 4089 | struct nb_iter_bound *elt; |
4090 | bool found_exit = false; | |
e9258aee | 4091 | auto_vec<basic_block> queue; |
e3606104 | 4092 | bitmap visited; |
4093 | ||
4094 | /* Collect all statements with interesting (i.e. lower than | |
4095 | nb_iterations_upper_bound) bound on them. | |
4096 | ||
4097 | TODO: Due to the way record_estimate choose estimates to store, the bounds | |
4098 | will be always nb_iterations_upper_bound-1. We can change this to record | |
4099 | also statements not dominating the loop latch and update the walk bellow | |
2fbe7a32 | 4100 | to the shortest path algorithm. */ |
e3606104 | 4101 | for (elt = loop->bounds; elt; elt = elt->next) |
4102 | { | |
4103 | if (!elt->is_exit | |
796b6678 | 4104 | && wi::ltu_p (elt->bound, loop->nb_iterations_upper_bound)) |
e3606104 | 4105 | { |
4106 | if (!not_executed_last_iteration) | |
42acab1c | 4107 | not_executed_last_iteration = new hash_set<gimple *>; |
431205b7 | 4108 | not_executed_last_iteration->add (elt->stmt); |
e3606104 | 4109 | } |
4110 | } | |
4111 | if (!not_executed_last_iteration) | |
4112 | return; | |
4113 | ||
4114 | /* Start DFS walk in the loop header and see if we can reach the | |
4115 | loop latch or any of the exits (including statements with side | |
4116 | effects that may terminate the loop otherwise) without visiting | |
4117 | any of the statements known to have undefined effect on the last | |
4118 | iteration. */ | |
f1f41a6c | 4119 | queue.safe_push (loop->header); |
e3606104 | 4120 | visited = BITMAP_ALLOC (NULL); |
4121 | bitmap_set_bit (visited, loop->header->index); | |
4122 | found_exit = false; | |
4123 | ||
4124 | do | |
4125 | { | |
f1f41a6c | 4126 | basic_block bb = queue.pop (); |
e3606104 | 4127 | gimple_stmt_iterator gsi; |
4128 | bool stmt_found = false; | |
4129 | ||
4130 | /* Loop for possible exits and statements bounding the execution. */ | |
4131 | for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi)) | |
4132 | { | |
42acab1c | 4133 | gimple *stmt = gsi_stmt (gsi); |
431205b7 | 4134 | if (not_executed_last_iteration->contains (stmt)) |
e3606104 | 4135 | { |
4136 | stmt_found = true; | |
4137 | break; | |
4138 | } | |
4139 | if (gimple_has_side_effects (stmt)) | |
4140 | { | |
4141 | found_exit = true; | |
4142 | break; | |
4143 | } | |
4144 | } | |
4145 | if (found_exit) | |
4146 | break; | |
4147 | ||
4148 | /* If no bounding statement is found, continue the walk. */ | |
4149 | if (!stmt_found) | |
4150 | { | |
4151 | edge e; | |
4152 | edge_iterator ei; | |
4153 | ||
4154 | FOR_EACH_EDGE (e, ei, bb->succs) | |
4155 | { | |
4156 | if (loop_exit_edge_p (loop, e) | |
4157 | || e == loop_latch_edge (loop)) | |
4158 | { | |
4159 | found_exit = true; | |
4160 | break; | |
4161 | } | |
4162 | if (bitmap_set_bit (visited, e->dest->index)) | |
f1f41a6c | 4163 | queue.safe_push (e->dest); |
e3606104 | 4164 | } |
4165 | } | |
4166 | } | |
f1f41a6c | 4167 | while (queue.length () && !found_exit); |
e3606104 | 4168 | |
4169 | /* If every path through the loop reach bounding statement before exit, | |
4170 | then we know the last iteration of the loop will have undefined effect | |
4171 | and we can decrease number of iterations. */ | |
4172 | ||
4173 | if (!found_exit) | |
4174 | { | |
4175 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
4176 | fprintf (dump_file, "Reducing loop iteration estimate by 1; " | |
4177 | "undefined statement must be executed at the last iteration.\n"); | |
e913b5cd | 4178 | record_niter_bound (loop, loop->nb_iterations_upper_bound - 1, |
e3606104 | 4179 | false, true); |
4180 | } | |
f18e93a8 | 4181 | |
e3606104 | 4182 | BITMAP_FREE (visited); |
431205b7 | 4183 | delete not_executed_last_iteration; |
e3606104 | 4184 | } |
4185 | ||
9b53b827 | 4186 | /* Records estimates on numbers of iterations of LOOP. If USE_UNDEFINED_P |
4187 | is true also use estimates derived from undefined behavior. */ | |
b9d73ea6 | 4188 | |
46480a95 | 4189 | void |
4190 | estimate_numbers_of_iterations (struct loop *loop) | |
b9d73ea6 | 4191 | { |
f1f41a6c | 4192 | vec<edge> exits; |
b9d73ea6 | 4193 | tree niter, type; |
749ea85f | 4194 | unsigned i; |
b9d73ea6 | 4195 | struct tree_niter_desc niter_desc; |
749ea85f | 4196 | edge ex; |
5de9d3ed | 4197 | widest_int bound; |
3681186e | 4198 | edge likely_exit; |
b9d73ea6 | 4199 | |
aff6b8e2 | 4200 | /* Give up if we already have tried to compute an estimation. */ |
faa56cf9 | 4201 | if (loop->estimate_state != EST_NOT_COMPUTED) |
aff6b8e2 | 4202 | return; |
b6556916 | 4203 | |
4b4ab846 | 4204 | loop->estimate_state = EST_AVAILABLE; |
f5de8018 | 4205 | |
4206 | /* If we have a measured profile, use it to estimate the number of | |
4207 | iterations. Normally this is recorded by branch_prob right after | |
4208 | reading the profile. In case we however found a new loop, record the | |
4209 | information here. | |
4210 | ||
4211 | Explicitly check for profile status so we do not report | |
4212 | wrong prediction hitrates for guessed loop iterations heuristics. | |
4213 | Do not recompute already recorded bounds - we ought to be better on | |
4214 | updating iteration bounds than updating profile in general and thus | |
4215 | recomputing iteration bounds later in the compilation process will just | |
4216 | introduce random roundoff errors. */ | |
4217 | if (!loop->any_estimate | |
205ce1aa | 4218 | && loop->header->count.reliable_p ()) |
f5de8018 | 4219 | { |
4220 | gcov_type nit = expected_loop_iterations_unbounded (loop); | |
4221 | bound = gcov_type_to_wide_int (nit); | |
4222 | record_niter_bound (loop, bound, true, false); | |
4223 | } | |
aff6b8e2 | 4224 | |
228bf2b8 | 4225 | /* Ensure that loop->nb_iterations is computed if possible. If it turns out |
4226 | to be constant, we avoid undefined behavior implied bounds and instead | |
4227 | diagnose those loops with -Waggressive-loop-optimizations. */ | |
4228 | number_of_latch_executions (loop); | |
4229 | ||
749ea85f | 4230 | exits = get_loop_exit_edges (loop); |
3681186e | 4231 | likely_exit = single_likely_exit (loop); |
f1f41a6c | 4232 | FOR_EACH_VEC_ELT (exits, i, ex) |
b9d73ea6 | 4233 | { |
fd943c70 | 4234 | if (!number_of_iterations_exit (loop, ex, &niter_desc, false, false)) |
b9d73ea6 | 4235 | continue; |
4236 | ||
4237 | niter = niter_desc.niter; | |
4238 | type = TREE_TYPE (niter); | |
faa56cf9 | 4239 | if (TREE_CODE (niter_desc.may_be_zero) != INTEGER_CST) |
065d134e | 4240 | niter = build3 (COND_EXPR, type, niter_desc.may_be_zero, |
05db596e | 4241 | build_int_cst (type, 0), |
065d134e | 4242 | niter); |
a463eaea | 4243 | record_estimate (loop, niter, niter_desc.max, |
749ea85f | 4244 | last_stmt (ex->src), |
3681186e | 4245 | true, ex == likely_exit, true); |
05032b10 | 4246 | record_control_iv (loop, &niter_desc); |
b9d73ea6 | 4247 | } |
f1f41a6c | 4248 | exits.release (); |
48e1416a | 4249 | |
56dfa44a | 4250 | if (flag_aggressive_loop_optimizations) |
4251 | infer_loop_bounds_from_undefined (loop); | |
4b4ab846 | 4252 | |
6689221f | 4253 | discover_iteration_bound_by_body_walk (loop); |
4254 | ||
e3606104 | 4255 | maybe_lower_iteration_bound (loop); |
4256 | ||
228bf2b8 | 4257 | /* If we know the exact number of iterations of this loop, try to |
4258 | not break code with undefined behavior by not recording smaller | |
4259 | maximum number of iterations. */ | |
4260 | if (loop->nb_iterations | |
4261 | && TREE_CODE (loop->nb_iterations) == INTEGER_CST) | |
4262 | { | |
4263 | loop->any_upper_bound = true; | |
5de9d3ed | 4264 | loop->nb_iterations_upper_bound = wi::to_widest (loop->nb_iterations); |
228bf2b8 | 4265 | } |
b9d73ea6 | 4266 | } |
4267 | ||
8fe79ba5 | 4268 | /* Sets NIT to the estimated number of executions of the latch of the |
4269 | LOOP. If CONSERVATIVE is true, we must be sure that NIT is at least as | |
4270 | large as the number of iterations. If we have no reliable estimate, | |
4271 | the function returns false, otherwise returns true. */ | |
4272 | ||
4273 | bool | |
5de9d3ed | 4274 | estimated_loop_iterations (struct loop *loop, widest_int *nit) |
8fe79ba5 | 4275 | { |
a9ef9877 | 4276 | /* When SCEV information is available, try to update loop iterations |
4277 | estimate. Otherwise just return whatever we recorded earlier. */ | |
4278 | if (scev_initialized_p ()) | |
46480a95 | 4279 | estimate_numbers_of_iterations (loop); |
a9ef9877 | 4280 | |
f86b328b | 4281 | return (get_estimated_loop_iterations (loop, nit)); |
8fe79ba5 | 4282 | } |
4283 | ||
4284 | /* Similar to estimated_loop_iterations, but returns the estimate only | |
4285 | if it fits to HOST_WIDE_INT. If this is not the case, or the estimate | |
4286 | on the number of iterations of LOOP could not be derived, returns -1. */ | |
4287 | ||
4288 | HOST_WIDE_INT | |
fee017b3 | 4289 | estimated_loop_iterations_int (struct loop *loop) |
4290 | { | |
5de9d3ed | 4291 | widest_int nit; |
fee017b3 | 4292 | HOST_WIDE_INT hwi_nit; |
4293 | ||
4294 | if (!estimated_loop_iterations (loop, &nit)) | |
4295 | return -1; | |
4296 | ||
796b6678 | 4297 | if (!wi::fits_shwi_p (nit)) |
fee017b3 | 4298 | return -1; |
cf8f0e63 | 4299 | hwi_nit = nit.to_shwi (); |
fee017b3 | 4300 | |
4301 | return hwi_nit < 0 ? -1 : hwi_nit; | |
4302 | } | |
4303 | ||
4e948f5a | 4304 | |
fee017b3 | 4305 | /* Sets NIT to an upper bound for the maximum number of executions of the |
4306 | latch of the LOOP. If we have no reliable estimate, the function returns | |
4307 | false, otherwise returns true. */ | |
4308 | ||
4309 | bool | |
5de9d3ed | 4310 | max_loop_iterations (struct loop *loop, widest_int *nit) |
fee017b3 | 4311 | { |
a9ef9877 | 4312 | /* When SCEV information is available, try to update loop iterations |
4313 | estimate. Otherwise just return whatever we recorded earlier. */ | |
4314 | if (scev_initialized_p ()) | |
46480a95 | 4315 | estimate_numbers_of_iterations (loop); |
8fe79ba5 | 4316 | |
f86b328b | 4317 | return get_max_loop_iterations (loop, nit); |
fee017b3 | 4318 | } |
4319 | ||
4320 | /* Similar to max_loop_iterations, but returns the estimate only | |
4321 | if it fits to HOST_WIDE_INT. If this is not the case, or the estimate | |
4322 | on the number of iterations of LOOP could not be derived, returns -1. */ | |
4323 | ||
4324 | HOST_WIDE_INT | |
4325 | max_loop_iterations_int (struct loop *loop) | |
8fe79ba5 | 4326 | { |
5de9d3ed | 4327 | widest_int nit; |
8fe79ba5 | 4328 | HOST_WIDE_INT hwi_nit; |
4329 | ||
fee017b3 | 4330 | if (!max_loop_iterations (loop, &nit)) |
8fe79ba5 | 4331 | return -1; |
4332 | ||
796b6678 | 4333 | if (!wi::fits_shwi_p (nit)) |
8fe79ba5 | 4334 | return -1; |
cf8f0e63 | 4335 | hwi_nit = nit.to_shwi (); |
8fe79ba5 | 4336 | |
4337 | return hwi_nit < 0 ? -1 : hwi_nit; | |
4338 | } | |
4339 | ||
8e3ffe30 | 4340 | /* Sets NIT to an likely upper bound for the maximum number of executions of the |
4341 | latch of the LOOP. If we have no reliable estimate, the function returns | |
4342 | false, otherwise returns true. */ | |
4343 | ||
4344 | bool | |
4345 | likely_max_loop_iterations (struct loop *loop, widest_int *nit) | |
4346 | { | |
4347 | /* When SCEV information is available, try to update loop iterations | |
4348 | estimate. Otherwise just return whatever we recorded earlier. */ | |
4349 | if (scev_initialized_p ()) | |
46480a95 | 4350 | estimate_numbers_of_iterations (loop); |
8e3ffe30 | 4351 | |
4352 | return get_likely_max_loop_iterations (loop, nit); | |
4353 | } | |
4354 | ||
4355 | /* Similar to max_loop_iterations, but returns the estimate only | |
4356 | if it fits to HOST_WIDE_INT. If this is not the case, or the estimate | |
4357 | on the number of iterations of LOOP could not be derived, returns -1. */ | |
4358 | ||
4359 | HOST_WIDE_INT | |
4360 | likely_max_loop_iterations_int (struct loop *loop) | |
4361 | { | |
4362 | widest_int nit; | |
4363 | HOST_WIDE_INT hwi_nit; | |
4364 | ||
4365 | if (!likely_max_loop_iterations (loop, &nit)) | |
4366 | return -1; | |
4367 | ||
4368 | if (!wi::fits_shwi_p (nit)) | |
4369 | return -1; | |
4370 | hwi_nit = nit.to_shwi (); | |
4371 | ||
4372 | return hwi_nit < 0 ? -1 : hwi_nit; | |
4373 | } | |
4374 | ||
fee017b3 | 4375 | /* Returns an estimate for the number of executions of statements |
4376 | in the LOOP. For statements before the loop exit, this exceeds | |
4377 | the number of execution of the latch by one. */ | |
4378 | ||
4379 | HOST_WIDE_INT | |
4380 | estimated_stmt_executions_int (struct loop *loop) | |
4381 | { | |
4382 | HOST_WIDE_INT nit = estimated_loop_iterations_int (loop); | |
4383 | HOST_WIDE_INT snit; | |
4384 | ||
4385 | if (nit == -1) | |
4386 | return -1; | |
4387 | ||
4388 | snit = (HOST_WIDE_INT) ((unsigned HOST_WIDE_INT) nit + 1); | |
4389 | ||
4390 | /* If the computation overflows, return -1. */ | |
4391 | return snit < 0 ? -1 : snit; | |
4392 | } | |
4393 | ||
8e3ffe30 | 4394 | /* Sets NIT to the maximum number of executions of the latch of the |
fee017b3 | 4395 | LOOP, plus one. If we have no reliable estimate, the function returns |
4396 | false, otherwise returns true. */ | |
4397 | ||
4398 | bool | |
5de9d3ed | 4399 | max_stmt_executions (struct loop *loop, widest_int *nit) |
fee017b3 | 4400 | { |
5de9d3ed | 4401 | widest_int nit_minus_one; |
fee017b3 | 4402 | |
4403 | if (!max_loop_iterations (loop, nit)) | |
4404 | return false; | |
4405 | ||
4406 | nit_minus_one = *nit; | |
4407 | ||
e913b5cd | 4408 | *nit += 1; |
fee017b3 | 4409 | |
796b6678 | 4410 | return wi::gtu_p (*nit, nit_minus_one); |
fee017b3 | 4411 | } |
4412 | ||
8e3ffe30 | 4413 | /* Sets NIT to the estimated maximum number of executions of the latch of the |
4414 | LOOP, plus one. If we have no likely estimate, the function returns | |
4415 | false, otherwise returns true. */ | |
4416 | ||
4417 | bool | |
4418 | likely_max_stmt_executions (struct loop *loop, widest_int *nit) | |
4419 | { | |
4420 | widest_int nit_minus_one; | |
4421 | ||
4422 | if (!likely_max_loop_iterations (loop, nit)) | |
4423 | return false; | |
4424 | ||
4425 | nit_minus_one = *nit; | |
4426 | ||
4427 | *nit += 1; | |
4428 | ||
4429 | return wi::gtu_p (*nit, nit_minus_one); | |
4430 | } | |
4431 | ||
8fe79ba5 | 4432 | /* Sets NIT to the estimated number of executions of the latch of the |
fee017b3 | 4433 | LOOP, plus one. If we have no reliable estimate, the function returns |
4434 | false, otherwise returns true. */ | |
8fe79ba5 | 4435 | |
4436 | bool | |
5de9d3ed | 4437 | estimated_stmt_executions (struct loop *loop, widest_int *nit) |
8fe79ba5 | 4438 | { |
5de9d3ed | 4439 | widest_int nit_minus_one; |
8fe79ba5 | 4440 | |
fee017b3 | 4441 | if (!estimated_loop_iterations (loop, nit)) |
8fe79ba5 | 4442 | return false; |
4443 | ||
4444 | nit_minus_one = *nit; | |
4445 | ||
e913b5cd | 4446 | *nit += 1; |
8fe79ba5 | 4447 | |
796b6678 | 4448 | return wi::gtu_p (*nit, nit_minus_one); |
8fe79ba5 | 4449 | } |
4450 | ||
7194de72 | 4451 | /* Records estimates on numbers of iterations of loops. */ |
b9d73ea6 | 4452 | |
4453 | void | |
46480a95 | 4454 | estimate_numbers_of_iterations (function *fn) |
b9d73ea6 | 4455 | { |
b9d73ea6 | 4456 | struct loop *loop; |
4457 | ||
add6ee5e | 4458 | /* We don't want to issue signed overflow warnings while getting |
4459 | loop iteration estimates. */ | |
4460 | fold_defer_overflow_warnings (); | |
4461 | ||
46480a95 | 4462 | FOR_EACH_LOOP_FN (fn, loop, 0) |
4463 | estimate_numbers_of_iterations (loop); | |
add6ee5e | 4464 | |
4465 | fold_undefer_and_ignore_overflow_warnings (); | |
b9d73ea6 | 4466 | } |
4467 | ||
b9d73ea6 | 4468 | /* Returns true if statement S1 dominates statement S2. */ |
4469 | ||
ad4a85ad | 4470 | bool |
42acab1c | 4471 | stmt_dominates_stmt_p (gimple *s1, gimple *s2) |
b9d73ea6 | 4472 | { |
75a70cf9 | 4473 | basic_block bb1 = gimple_bb (s1), bb2 = gimple_bb (s2); |
b9d73ea6 | 4474 | |
4475 | if (!bb1 | |
4476 | || s1 == s2) | |
4477 | return true; | |
4478 | ||
4479 | if (bb1 == bb2) | |
4480 | { | |
75a70cf9 | 4481 | gimple_stmt_iterator bsi; |
b9d73ea6 | 4482 | |
dddf5036 | 4483 | if (gimple_code (s2) == GIMPLE_PHI) |
4484 | return false; | |
4485 | ||
4486 | if (gimple_code (s1) == GIMPLE_PHI) | |
4487 | return true; | |
4488 | ||
75a70cf9 | 4489 | for (bsi = gsi_start_bb (bb1); gsi_stmt (bsi) != s2; gsi_next (&bsi)) |
4490 | if (gsi_stmt (bsi) == s1) | |
b9d73ea6 | 4491 | return true; |
4492 | ||
4493 | return false; | |
4494 | } | |
4495 | ||
4496 | return dominated_by_p (CDI_DOMINATORS, bb2, bb1); | |
4497 | } | |
4498 | ||
d1baf6ca | 4499 | /* Returns true when we can prove that the number of executions of |
faa56cf9 | 4500 | STMT in the loop is at most NITER, according to the bound on |
4501 | the number of executions of the statement NITER_BOUND->stmt recorded in | |
64552007 | 4502 | NITER_BOUND and fact that NITER_BOUND->stmt dominate STMT. |
4503 | ||
4504 | ??? This code can become quite a CPU hog - we can have many bounds, | |
4505 | and large basic block forcing stmt_dominates_stmt_p to be queried | |
4506 | many times on a large basic blocks, so the whole thing is O(n^2) | |
4507 | for scev_probably_wraps_p invocation (that can be done n times). | |
4508 | ||
4509 | It would make more sense (and give better answers) to remember BB | |
4510 | bounds computed by discover_iteration_bound_by_body_walk. */ | |
b9d73ea6 | 4511 | |
b3786ab3 | 4512 | static bool |
42acab1c | 4513 | n_of_executions_at_most (gimple *stmt, |
48e1416a | 4514 | struct nb_iter_bound *niter_bound, |
0523b3f8 | 4515 | tree niter) |
b9d73ea6 | 4516 | { |
5de9d3ed | 4517 | widest_int bound = niter_bound->bound; |
cd743a11 | 4518 | tree nit_type = TREE_TYPE (niter), e; |
9dfb8b49 | 4519 | enum tree_code cmp; |
b3786ab3 | 4520 | |
faa56cf9 | 4521 | gcc_assert (TYPE_UNSIGNED (nit_type)); |
4522 | ||
4523 | /* If the bound does not even fit into NIT_TYPE, it cannot tell us that | |
4524 | the number of iterations is small. */ | |
796b6678 | 4525 | if (!wi::fits_to_tree_p (bound, nit_type)) |
faa56cf9 | 4526 | return false; |
4527 | ||
4528 | /* We know that NITER_BOUND->stmt is executed at most NITER_BOUND->bound + 1 | |
4529 | times. This means that: | |
48e1416a | 4530 | |
64552007 | 4531 | -- if NITER_BOUND->is_exit is true, then everything after |
4532 | it at most NITER_BOUND->bound times. | |
faa56cf9 | 4533 | |
4534 | -- If NITER_BOUND->is_exit is false, then if we can prove that when STMT | |
4535 | is executed, then NITER_BOUND->stmt is executed as well in the same | |
64552007 | 4536 | iteration then STMT is executed at most NITER_BOUND->bound + 1 times. |
4537 | ||
4538 | If we can determine that NITER_BOUND->stmt is always executed | |
4539 | after STMT, then STMT is executed at most NITER_BOUND->bound + 2 times. | |
4540 | We conclude that if both statements belong to the same | |
4541 | basic block and STMT is before NITER_BOUND->stmt and there are no | |
4542 | statements with side effects in between. */ | |
faa56cf9 | 4543 | |
4544 | if (niter_bound->is_exit) | |
4545 | { | |
64552007 | 4546 | if (stmt == niter_bound->stmt |
4547 | || !stmt_dominates_stmt_p (niter_bound->stmt, stmt)) | |
4548 | return false; | |
4549 | cmp = GE_EXPR; | |
faa56cf9 | 4550 | } |
b3786ab3 | 4551 | else |
faa56cf9 | 4552 | { |
64552007 | 4553 | if (!stmt_dominates_stmt_p (niter_bound->stmt, stmt)) |
faa56cf9 | 4554 | { |
64552007 | 4555 | gimple_stmt_iterator bsi; |
4556 | if (gimple_bb (stmt) != gimple_bb (niter_bound->stmt) | |
4557 | || gimple_code (stmt) == GIMPLE_PHI | |
4558 | || gimple_code (niter_bound->stmt) == GIMPLE_PHI) | |
4559 | return false; | |
4560 | ||
4561 | /* By stmt_dominates_stmt_p we already know that STMT appears | |
4562 | before NITER_BOUND->STMT. Still need to test that the loop | |
f4d3c071 | 4563 | cannot be terinated by a side effect in between. */ |
64552007 | 4564 | for (bsi = gsi_for_stmt (stmt); gsi_stmt (bsi) != niter_bound->stmt; |
4565 | gsi_next (&bsi)) | |
4566 | if (gimple_has_side_effects (gsi_stmt (bsi))) | |
4567 | return false; | |
e913b5cd | 4568 | bound += 1; |
796b6678 | 4569 | if (bound == 0 |
4570 | || !wi::fits_to_tree_p (bound, nit_type)) | |
faa56cf9 | 4571 | return false; |
4572 | } | |
4573 | cmp = GT_EXPR; | |
4574 | } | |
b3786ab3 | 4575 | |
cd743a11 | 4576 | e = fold_binary (cmp, boolean_type_node, |
e913b5cd | 4577 | niter, wide_int_to_tree (nit_type, bound)); |
cd743a11 | 4578 | return e && integer_nonzerop (e); |
b3786ab3 | 4579 | } |
4580 | ||
57e3f39a | 4581 | /* Returns true if the arithmetics in TYPE can be assumed not to wrap. */ |
b9d73ea6 | 4582 | |
57e3f39a | 4583 | bool |
4584 | nowrap_type_p (tree type) | |
903dae48 | 4585 | { |
e44f1a39 | 4586 | if (ANY_INTEGRAL_TYPE_P (type) |
981eb798 | 4587 | && TYPE_OVERFLOW_UNDEFINED (type)) |
57e3f39a | 4588 | return true; |
903dae48 | 4589 | |
57e3f39a | 4590 | if (POINTER_TYPE_P (type)) |
4591 | return true; | |
903dae48 | 4592 | |
903dae48 | 4593 | return false; |
4594 | } | |
4595 | ||
05032b10 | 4596 | /* Return true if we can prove LOOP is exited before evolution of induction |
80ac98e8 | 4597 | variable {BASE, STEP} overflows with respect to its type bound. */ |
05032b10 | 4598 | |
4599 | static bool | |
4600 | loop_exits_before_overflow (tree base, tree step, | |
42acab1c | 4601 | gimple *at_stmt, struct loop *loop) |
05032b10 | 4602 | { |
4603 | widest_int niter; | |
4604 | struct control_iv *civ; | |
4605 | struct nb_iter_bound *bound; | |
4606 | tree e, delta, step_abs, unsigned_base; | |
4607 | tree type = TREE_TYPE (step); | |
4608 | tree unsigned_type, valid_niter; | |
4609 | ||
4610 | /* Don't issue signed overflow warnings. */ | |
4611 | fold_defer_overflow_warnings (); | |
4612 | ||
4613 | /* Compute the number of iterations before we reach the bound of the | |
4614 | type, and verify that the loop is exited before this occurs. */ | |
4615 | unsigned_type = unsigned_type_for (type); | |
4616 | unsigned_base = fold_convert (unsigned_type, base); | |
4617 | ||
4618 | if (tree_int_cst_sign_bit (step)) | |
4619 | { | |
4620 | tree extreme = fold_convert (unsigned_type, | |
4621 | lower_bound_in_type (type, type)); | |
4622 | delta = fold_build2 (MINUS_EXPR, unsigned_type, unsigned_base, extreme); | |
4623 | step_abs = fold_build1 (NEGATE_EXPR, unsigned_type, | |
4624 | fold_convert (unsigned_type, step)); | |
4625 | } | |
4626 | else | |
4627 | { | |
4628 | tree extreme = fold_convert (unsigned_type, | |
4629 | upper_bound_in_type (type, type)); | |
4630 | delta = fold_build2 (MINUS_EXPR, unsigned_type, extreme, unsigned_base); | |
4631 | step_abs = fold_convert (unsigned_type, step); | |
4632 | } | |
4633 | ||
4634 | valid_niter = fold_build2 (FLOOR_DIV_EXPR, unsigned_type, delta, step_abs); | |
4635 | ||
46480a95 | 4636 | estimate_numbers_of_iterations (loop); |
05032b10 | 4637 | |
4638 | if (max_loop_iterations (loop, &niter) | |
4639 | && wi::fits_to_tree_p (niter, TREE_TYPE (valid_niter)) | |
4640 | && (e = fold_binary (GT_EXPR, boolean_type_node, valid_niter, | |
4641 | wide_int_to_tree (TREE_TYPE (valid_niter), | |
4642 | niter))) != NULL | |
4643 | && integer_nonzerop (e)) | |
4644 | { | |
4645 | fold_undefer_and_ignore_overflow_warnings (); | |
4646 | return true; | |
4647 | } | |
4648 | if (at_stmt) | |
4649 | for (bound = loop->bounds; bound; bound = bound->next) | |
4650 | { | |
4651 | if (n_of_executions_at_most (at_stmt, bound, valid_niter)) | |
4652 | { | |
4653 | fold_undefer_and_ignore_overflow_warnings (); | |
4654 | return true; | |
4655 | } | |
4656 | } | |
4657 | fold_undefer_and_ignore_overflow_warnings (); | |
4658 | ||
4659 | /* Try to prove loop is exited before {base, step} overflows with the | |
4660 | help of analyzed loop control IV. This is done only for IVs with | |
4661 | constant step because otherwise we don't have the information. */ | |
4662 | if (TREE_CODE (step) == INTEGER_CST) | |
6c48ce84 | 4663 | { |
6c48ce84 | 4664 | for (civ = loop->control_ivs; civ; civ = civ->next) |
4665 | { | |
4666 | enum tree_code code; | |
a031a844 | 4667 | tree civ_type = TREE_TYPE (civ->step); |
05032b10 | 4668 | |
6c48ce84 | 4669 | /* Have to consider type difference because operand_equal_p ignores |
4670 | that for constants. */ | |
4671 | if (TYPE_UNSIGNED (type) != TYPE_UNSIGNED (civ_type) | |
4672 | || element_precision (type) != element_precision (civ_type)) | |
05032b10 | 4673 | continue; |
05032b10 | 4674 | |
6c48ce84 | 4675 | /* Only consider control IV with same step. */ |
4676 | if (!operand_equal_p (step, civ->step, 0)) | |
4677 | continue; | |
05032b10 | 4678 | |
6c48ce84 | 4679 | /* Done proving if this is a no-overflow control IV. */ |
a031a844 | 4680 | if (operand_equal_p (base, civ->base, 0)) |
4681 | return true; | |
4682 | ||
4683 | /* Control IV is recorded after expanding simple operations, | |
4684 | Here we expand base and compare it too. */ | |
4685 | tree expanded_base = expand_simple_operations (base); | |
4686 | if (operand_equal_p (expanded_base, civ->base, 0)) | |
6c48ce84 | 4687 | return true; |
05032b10 | 4688 | |
6c48ce84 | 4689 | /* If this is a before stepping control IV, in other words, we have |
05032b10 | 4690 | |
6c48ce84 | 4691 | {civ_base, step} = {base + step, step} |
7cf4130d | 4692 | |
6c48ce84 | 4693 | Because civ {base + step, step} doesn't overflow during loop |
4694 | iterations, {base, step} will not overflow if we can prove the | |
4695 | operation "base + step" does not overflow. Specifically, we try | |
4696 | to prove below conditions are satisfied: | |
7cf4130d | 4697 | |
6c48ce84 | 4698 | base <= UPPER_BOUND (type) - step ;;step > 0 |
4699 | base >= LOWER_BOUND (type) - step ;;step < 0 | |
7cf4130d | 4700 | |
6c48ce84 | 4701 | by proving the reverse conditions are false using loop's initial |
4702 | condition. */ | |
4703 | if (POINTER_TYPE_P (TREE_TYPE (base))) | |
4704 | code = POINTER_PLUS_EXPR; | |
4705 | else | |
4706 | code = PLUS_EXPR; | |
7cf4130d | 4707 | |
a031a844 | 4708 | tree stepped = fold_build2 (code, TREE_TYPE (base), base, step); |
4709 | tree expanded_stepped = fold_build2 (code, TREE_TYPE (base), | |
4710 | expanded_base, step); | |
4711 | if (operand_equal_p (stepped, civ->base, 0) | |
4712 | || operand_equal_p (expanded_stepped, civ->base, 0)) | |
6c48ce84 | 4713 | { |
a031a844 | 4714 | tree extreme; |
4715 | ||
6c48ce84 | 4716 | if (tree_int_cst_sign_bit (step)) |
4717 | { | |
4718 | code = LT_EXPR; | |
4719 | extreme = lower_bound_in_type (type, type); | |
4720 | } | |
4721 | else | |
4722 | { | |
4723 | code = GT_EXPR; | |
4724 | extreme = upper_bound_in_type (type, type); | |
4725 | } | |
4726 | extreme = fold_build2 (MINUS_EXPR, type, extreme, step); | |
4727 | e = fold_build2 (code, boolean_type_node, base, extreme); | |
6723abf5 | 4728 | e = simplify_using_initial_conditions (loop, e); |
6c48ce84 | 4729 | if (integer_zerop (e)) |
4730 | return true; | |
4731 | } | |
4732 | } | |
4733 | } | |
05032b10 | 4734 | |
4735 | return false; | |
4736 | } | |
4737 | ||
90261646 | 4738 | /* VAR is scev variable whose evolution part is constant STEP, this function |
4739 | proves that VAR can't overflow by using value range info. If VAR's value | |
4740 | range is [MIN, MAX], it can be proven by: | |
4741 | MAX + step doesn't overflow ; if step > 0 | |
4742 | or | |
4743 | MIN + step doesn't underflow ; if step < 0. | |
4744 | ||
4745 | We can only do this if var is computed in every loop iteration, i.e, var's | |
4746 | definition has to dominate loop latch. Consider below example: | |
4747 | ||
4748 | { | |
4749 | unsigned int i; | |
4750 | ||
4751 | <bb 3>: | |
4752 | ||
4753 | <bb 4>: | |
4754 | # RANGE [0, 4294967294] NONZERO 65535 | |
4755 | # i_21 = PHI <0(3), i_18(9)> | |
4756 | if (i_21 != 0) | |
4757 | goto <bb 6>; | |
4758 | else | |
4759 | goto <bb 8>; | |
4760 | ||
4761 | <bb 6>: | |
4762 | # RANGE [0, 65533] NONZERO 65535 | |
4763 | _6 = i_21 + 4294967295; | |
4764 | # RANGE [0, 65533] NONZERO 65535 | |
4765 | _7 = (long unsigned int) _6; | |
4766 | # RANGE [0, 524264] NONZERO 524280 | |
4767 | _8 = _7 * 8; | |
4768 | # PT = nonlocal escaped | |
4769 | _9 = a_14 + _8; | |
4770 | *_9 = 0; | |
4771 | ||
4772 | <bb 8>: | |
4773 | # RANGE [1, 65535] NONZERO 65535 | |
4774 | i_18 = i_21 + 1; | |
4775 | if (i_18 >= 65535) | |
4776 | goto <bb 10>; | |
4777 | else | |
4778 | goto <bb 9>; | |
4779 | ||
4780 | <bb 9>: | |
4781 | goto <bb 4>; | |
4782 | ||
4783 | <bb 10>: | |
4784 | return; | |
4785 | } | |
4786 | ||
4787 | VAR _6 doesn't overflow only with pre-condition (i_21 != 0), here we | |
4788 | can't use _6 to prove no-overlfow for _7. In fact, var _7 takes value | |
4789 | sequence (4294967295, 0, 1, ..., 65533) in loop life time, rather than | |
4790 | (4294967295, 4294967296, ...). */ | |
4791 | ||
4792 | static bool | |
4793 | scev_var_range_cant_overflow (tree var, tree step, struct loop *loop) | |
4794 | { | |
4795 | tree type; | |
4796 | wide_int minv, maxv, diff, step_wi; | |
be44111e | 4797 | enum value_range_kind rtype; |
90261646 | 4798 | |
4799 | if (TREE_CODE (step) != INTEGER_CST || !INTEGRAL_TYPE_P (TREE_TYPE (var))) | |
4800 | return false; | |
4801 | ||
4802 | /* Check if VAR evaluates in every loop iteration. It's not the case | |
4803 | if VAR is default definition or does not dominate loop's latch. */ | |
4804 | basic_block def_bb = gimple_bb (SSA_NAME_DEF_STMT (var)); | |
4805 | if (!def_bb || !dominated_by_p (CDI_DOMINATORS, loop->latch, def_bb)) | |
4806 | return false; | |
4807 | ||
4808 | rtype = get_range_info (var, &minv, &maxv); | |
4809 | if (rtype != VR_RANGE) | |
4810 | return false; | |
4811 | ||
4812 | /* VAR is a scev whose evolution part is STEP and value range info | |
4813 | is [MIN, MAX], we can prove its no-overflowness by conditions: | |
4814 | ||
4815 | type_MAX - MAX >= step ; if step > 0 | |
4816 | MIN - type_MIN >= |step| ; if step < 0. | |
4817 | ||
4818 | Or VAR must take value outside of value range, which is not true. */ | |
e3d0f65c | 4819 | step_wi = wi::to_wide (step); |
90261646 | 4820 | type = TREE_TYPE (var); |
4821 | if (tree_int_cst_sign_bit (step)) | |
4822 | { | |
e3d0f65c | 4823 | diff = minv - wi::to_wide (lower_bound_in_type (type, type)); |
90261646 | 4824 | step_wi = - step_wi; |
4825 | } | |
4826 | else | |
e3d0f65c | 4827 | diff = wi::to_wide (upper_bound_in_type (type, type)) - maxv; |
90261646 | 4828 | |
4829 | return (wi::geu_p (diff, step_wi)); | |
4830 | } | |
4831 | ||
b3786ab3 | 4832 | /* Return false only when the induction variable BASE + STEP * I is |
4833 | known to not overflow: i.e. when the number of iterations is small | |
4834 | enough with respect to the step and initial condition in order to | |
4835 | keep the evolution confined in TYPEs bounds. Return true when the | |
4836 | iv is known to overflow or when the property is not computable. | |
48e1416a | 4837 | |
57e3f39a | 4838 | USE_OVERFLOW_SEMANTICS is true if this function should assume that |
4839 | the rules for overflow of the given language apply (e.g., that signed | |
90261646 | 4840 | arithmetics in C does not overflow). |
4841 | ||
4842 | If VAR is a ssa variable, this function also returns false if VAR can | |
4843 | be proven not overflow with value range info. */ | |
b3786ab3 | 4844 | |
4845 | bool | |
90261646 | 4846 | scev_probably_wraps_p (tree var, tree base, tree step, |
42acab1c | 4847 | gimple *at_stmt, struct loop *loop, |
499c8b1e | 4848 | bool use_overflow_semantics) |
b3786ab3 | 4849 | { |
57e3f39a | 4850 | /* FIXME: We really need something like |
4851 | http://gcc.gnu.org/ml/gcc-patches/2005-06/msg02025.html. | |
4852 | ||
4853 | We used to test for the following situation that frequently appears | |
4854 | during address arithmetics: | |
48e1416a | 4855 | |
903dae48 | 4856 | D.1621_13 = (long unsigned intD.4) D.1620_12; |
4857 | D.1622_14 = D.1621_13 * 8; | |
4858 | D.1623_15 = (doubleD.29 *) D.1622_14; | |
903dae48 | 4859 | |
57e3f39a | 4860 | And derived that the sequence corresponding to D_14 |
4861 | can be proved to not wrap because it is used for computing a | |
4862 | memory access; however, this is not really the case -- for example, | |
4863 | if D_12 = (unsigned char) [254,+,1], then D_14 has values | |
4864 | 2032, 2040, 0, 8, ..., but the code is still legal. */ | |
b3786ab3 | 4865 | |
9887dd18 | 4866 | if (chrec_contains_undetermined (base) |
45dbd04b | 4867 | || chrec_contains_undetermined (step)) |
57e3f39a | 4868 | return true; |
903dae48 | 4869 | |
cd743a11 | 4870 | if (integer_zerop (step)) |
57e3f39a | 4871 | return false; |
ecf68307 | 4872 | |
57e3f39a | 4873 | /* If we can use the fact that signed and pointer arithmetics does not |
4874 | wrap, we are done. */ | |
b718fe63 | 4875 | if (use_overflow_semantics && nowrap_type_p (TREE_TYPE (base))) |
57e3f39a | 4876 | return false; |
ecf68307 | 4877 | |
45dbd04b | 4878 | /* To be able to use estimates on number of iterations of the loop, |
4879 | we must have an upper bound on the absolute value of the step. */ | |
4880 | if (TREE_CODE (step) != INTEGER_CST) | |
4881 | return true; | |
4882 | ||
90261646 | 4883 | /* Check if var can be proven not overflow with value range info. */ |
4884 | if (var && TREE_CODE (var) == SSA_NAME | |
4885 | && scev_var_range_cant_overflow (var, step, loop)) | |
4886 | return false; | |
4887 | ||
05032b10 | 4888 | if (loop_exits_before_overflow (base, step, at_stmt, loop)) |
4889 | return false; | |
b3786ab3 | 4890 | |
4891 | /* At this point we still don't have a proof that the iv does not | |
4892 | overflow: give up. */ | |
4893 | return true; | |
b9d73ea6 | 4894 | } |
4895 | ||
b9d73ea6 | 4896 | /* Frees the information on upper bounds on numbers of iterations of LOOP. */ |
4897 | ||
70106280 | 4898 | void |
46480a95 | 4899 | free_numbers_of_iterations_estimates (struct loop *loop) |
b9d73ea6 | 4900 | { |
05032b10 | 4901 | struct control_iv *civ; |
4902 | struct nb_iter_bound *bound; | |
70106280 | 4903 | |
4904 | loop->nb_iterations = NULL; | |
faa56cf9 | 4905 | loop->estimate_state = EST_NOT_COMPUTED; |
05032b10 | 4906 | for (bound = loop->bounds; bound;) |
b9d73ea6 | 4907 | { |
05032b10 | 4908 | struct nb_iter_bound *next = bound->next; |
ccae4f9f | 4909 | ggc_free (bound); |
05032b10 | 4910 | bound = next; |
b9d73ea6 | 4911 | } |
b9d73ea6 | 4912 | loop->bounds = NULL; |
05032b10 | 4913 | |
4914 | for (civ = loop->control_ivs; civ;) | |
4915 | { | |
4916 | struct control_iv *next = civ->next; | |
4917 | ggc_free (civ); | |
4918 | civ = next; | |
4919 | } | |
4920 | loop->control_ivs = NULL; | |
b9d73ea6 | 4921 | } |
4922 | ||
7194de72 | 4923 | /* Frees the information on upper bounds on numbers of iterations of loops. */ |
b9d73ea6 | 4924 | |
4925 | void | |
d4f078b5 | 4926 | free_numbers_of_iterations_estimates (function *fn) |
b9d73ea6 | 4927 | { |
b9d73ea6 | 4928 | struct loop *loop; |
4929 | ||
d4f078b5 | 4930 | FOR_EACH_LOOP_FN (fn, loop, 0) |
46480a95 | 4931 | free_numbers_of_iterations_estimates (loop); |
b9d73ea6 | 4932 | } |
f3830c7c | 4933 | |
4934 | /* Substitute value VAL for ssa name NAME inside expressions held | |
4935 | at LOOP. */ | |
4936 | ||
4937 | void | |
4938 | substitute_in_loop_info (struct loop *loop, tree name, tree val) | |
4939 | { | |
f3830c7c | 4940 | loop->nb_iterations = simplify_replace_tree (loop->nb_iterations, name, val); |
f3830c7c | 4941 | } |