decl.c (value_annotation_hasher::handle_cache_entry): Delete.
[gcc.git] / gcc / match.pd
1 /* Match-and-simplify patterns for shared GENERIC and GIMPLE folding.
2 This file is consumed by genmatch which produces gimple-match.c
3 and generic-match.c from it.
4
5 Copyright (C) 2014-2015 Free Software Foundation, Inc.
6 Contributed by Richard Biener <rguenther@suse.de>
7 and Prathamesh Kulkarni <bilbotheelffriend@gmail.com>
8
9 This file is part of GCC.
10
11 GCC is free software; you can redistribute it and/or modify it under
12 the terms of the GNU General Public License as published by the Free
13 Software Foundation; either version 3, or (at your option) any later
14 version.
15
16 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
17 WARRANTY; without even the implied warranty of MERCHANTABILITY or
18 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
19 for more details.
20
21 You should have received a copy of the GNU General Public License
22 along with GCC; see the file COPYING3. If not see
23 <http://www.gnu.org/licenses/>. */
24
25
26 /* Generic tree predicates we inherit. */
27 (define_predicates
28 integer_onep integer_zerop integer_all_onesp integer_minus_onep
29 integer_each_onep integer_truep
30 real_zerop real_onep real_minus_onep
31 CONSTANT_CLASS_P
32 tree_expr_nonnegative_p)
33
34 /* Operator lists. */
35 (define_operator_list tcc_comparison
36 lt le eq ne ge gt unordered ordered unlt unle ungt unge uneq ltgt)
37 (define_operator_list inverted_tcc_comparison
38 ge gt ne eq lt le ordered unordered ge gt le lt ltgt uneq)
39 (define_operator_list inverted_tcc_comparison_with_nans
40 unge ungt ne eq unlt unle ordered unordered ge gt le lt ltgt uneq)
41 (define_operator_list swapped_tcc_comparison
42 gt ge eq ne le lt unordered ordered ungt unge unlt unle uneq ltgt)
43
44
45 /* Simplifications of operations with one constant operand and
46 simplifications to constants or single values. */
47
48 (for op (plus pointer_plus minus bit_ior bit_xor)
49 (simplify
50 (op @0 integer_zerop)
51 (non_lvalue @0)))
52
53 /* 0 +p index -> (type)index */
54 (simplify
55 (pointer_plus integer_zerop @1)
56 (non_lvalue (convert @1)))
57
58 /* See if ARG1 is zero and X + ARG1 reduces to X.
59 Likewise if the operands are reversed. */
60 (simplify
61 (plus:c @0 real_zerop@1)
62 (if (fold_real_zero_addition_p (type, @1, 0))
63 (non_lvalue @0)))
64
65 /* See if ARG1 is zero and X - ARG1 reduces to X. */
66 (simplify
67 (minus @0 real_zerop@1)
68 (if (fold_real_zero_addition_p (type, @1, 1))
69 (non_lvalue @0)))
70
71 /* Simplify x - x.
72 This is unsafe for certain floats even in non-IEEE formats.
73 In IEEE, it is unsafe because it does wrong for NaNs.
74 Also note that operand_equal_p is always false if an operand
75 is volatile. */
76 (simplify
77 (minus @0 @0)
78 (if (!FLOAT_TYPE_P (type) || !HONOR_NANS (type))
79 { build_zero_cst (type); }))
80
81 (simplify
82 (mult @0 integer_zerop@1)
83 @1)
84
85 /* Maybe fold x * 0 to 0. The expressions aren't the same
86 when x is NaN, since x * 0 is also NaN. Nor are they the
87 same in modes with signed zeros, since multiplying a
88 negative value by 0 gives -0, not +0. */
89 (simplify
90 (mult @0 real_zerop@1)
91 (if (!HONOR_NANS (type) && !HONOR_SIGNED_ZEROS (element_mode (type)))
92 @1))
93
94 /* In IEEE floating point, x*1 is not equivalent to x for snans.
95 Likewise for complex arithmetic with signed zeros. */
96 (simplify
97 (mult @0 real_onep)
98 (if (!HONOR_SNANS (element_mode (type))
99 && (!HONOR_SIGNED_ZEROS (element_mode (type))
100 || !COMPLEX_FLOAT_TYPE_P (type)))
101 (non_lvalue @0)))
102
103 /* Transform x * -1.0 into -x. */
104 (simplify
105 (mult @0 real_minus_onep)
106 (if (!HONOR_SNANS (element_mode (type))
107 && (!HONOR_SIGNED_ZEROS (element_mode (type))
108 || !COMPLEX_FLOAT_TYPE_P (type)))
109 (negate @0)))
110
111 /* Make sure to preserve divisions by zero. This is the reason why
112 we don't simplify x / x to 1 or 0 / x to 0. */
113 (for op (mult trunc_div ceil_div floor_div round_div exact_div)
114 (simplify
115 (op @0 integer_onep)
116 (non_lvalue @0)))
117
118 /* X / -1 is -X. */
119 (for div (trunc_div ceil_div floor_div round_div exact_div)
120 (simplify
121 (div @0 integer_minus_onep@1)
122 (if (!TYPE_UNSIGNED (type))
123 (negate @0))))
124
125 /* For unsigned integral types, FLOOR_DIV_EXPR is the same as
126 TRUNC_DIV_EXPR. Rewrite into the latter in this case. */
127 (simplify
128 (floor_div @0 @1)
129 (if ((INTEGRAL_TYPE_P (type) || VECTOR_INTEGER_TYPE_P (type))
130 && TYPE_UNSIGNED (type))
131 (trunc_div @0 @1)))
132
133 /* Combine two successive divisions. Note that combining ceil_div
134 and floor_div is trickier and combining round_div even more so. */
135 (for div (trunc_div exact_div)
136 (simplify
137 (div (div @0 INTEGER_CST@1) INTEGER_CST@2)
138 (with {
139 bool overflow_p;
140 wide_int mul = wi::mul (@1, @2, TYPE_SIGN (type), &overflow_p);
141 }
142 (if (!overflow_p)
143 (div @0 { wide_int_to_tree (type, mul); }))
144 (if (overflow_p
145 && (TYPE_UNSIGNED (type)
146 || mul != wi::min_value (TYPE_PRECISION (type), SIGNED)))
147 { build_zero_cst (type); }))))
148
149 /* Optimize A / A to 1.0 if we don't care about
150 NaNs or Infinities. */
151 (simplify
152 (rdiv @0 @0)
153 (if (FLOAT_TYPE_P (type)
154 && ! HONOR_NANS (type)
155 && ! HONOR_INFINITIES (element_mode (type)))
156 { build_one_cst (type); }))
157
158 /* Optimize -A / A to -1.0 if we don't care about
159 NaNs or Infinities. */
160 (simplify
161 (rdiv:c @0 (negate @0))
162 (if (FLOAT_TYPE_P (type)
163 && ! HONOR_NANS (type)
164 && ! HONOR_INFINITIES (element_mode (type)))
165 { build_minus_one_cst (type); }))
166
167 /* In IEEE floating point, x/1 is not equivalent to x for snans. */
168 (simplify
169 (rdiv @0 real_onep)
170 (if (!HONOR_SNANS (element_mode (type)))
171 (non_lvalue @0)))
172
173 /* In IEEE floating point, x/-1 is not equivalent to -x for snans. */
174 (simplify
175 (rdiv @0 real_minus_onep)
176 (if (!HONOR_SNANS (element_mode (type)))
177 (negate @0)))
178
179 /* If ARG1 is a constant, we can convert this to a multiply by the
180 reciprocal. This does not have the same rounding properties,
181 so only do this if -freciprocal-math. We can actually
182 always safely do it if ARG1 is a power of two, but it's hard to
183 tell if it is or not in a portable manner. */
184 (for cst (REAL_CST COMPLEX_CST VECTOR_CST)
185 (simplify
186 (rdiv @0 cst@1)
187 (if (optimize)
188 (if (flag_reciprocal_math
189 && !real_zerop (@1))
190 (with
191 { tree tem = const_binop (RDIV_EXPR, type, build_one_cst (type), @1); }
192 (if (tem)
193 (mult @0 { tem; } ))))
194 (if (cst != COMPLEX_CST)
195 (with { tree inverse = exact_inverse (type, @1); }
196 (if (inverse)
197 (mult @0 { inverse; } )))))))
198
199 /* Same applies to modulo operations, but fold is inconsistent here
200 and simplifies 0 % x to 0, only preserving literal 0 % 0. */
201 (for mod (ceil_mod floor_mod round_mod trunc_mod)
202 /* 0 % X is always zero. */
203 (simplify
204 (mod integer_zerop@0 @1)
205 /* But not for 0 % 0 so that we can get the proper warnings and errors. */
206 (if (!integer_zerop (@1))
207 @0))
208 /* X % 1 is always zero. */
209 (simplify
210 (mod @0 integer_onep)
211 { build_zero_cst (type); })
212 /* X % -1 is zero. */
213 (simplify
214 (mod @0 integer_minus_onep@1)
215 (if (!TYPE_UNSIGNED (type))
216 { build_zero_cst (type); }))
217 /* (X % Y) % Y is just X % Y. */
218 (simplify
219 (mod (mod@2 @0 @1) @1)
220 @2))
221
222 /* X % -C is the same as X % C. */
223 (simplify
224 (trunc_mod @0 INTEGER_CST@1)
225 (if (TYPE_SIGN (type) == SIGNED
226 && !TREE_OVERFLOW (@1)
227 && wi::neg_p (@1)
228 && !TYPE_OVERFLOW_TRAPS (type)
229 /* Avoid this transformation if C is INT_MIN, i.e. C == -C. */
230 && !sign_bit_p (@1, @1))
231 (trunc_mod @0 (negate @1))))
232
233 /* X % -Y is the same as X % Y. */
234 (simplify
235 (trunc_mod @0 (convert? (negate @1)))
236 (if (!TYPE_UNSIGNED (type)
237 && !TYPE_OVERFLOW_TRAPS (type)
238 && tree_nop_conversion_p (type, TREE_TYPE (@1)))
239 (trunc_mod @0 (convert @1))))
240
241 /* Optimize TRUNC_MOD_EXPR by a power of two into a BIT_AND_EXPR,
242 i.e. "X % C" into "X & (C - 1)", if X and C are positive.
243 Also optimize A % (C << N) where C is a power of 2,
244 to A & ((C << N) - 1). */
245 (match (power_of_two_cand @1)
246 INTEGER_CST@1)
247 (match (power_of_two_cand @1)
248 (lshift INTEGER_CST@1 @2))
249 (for mod (trunc_mod floor_mod)
250 (simplify
251 (mod @0 (convert?@3 (power_of_two_cand@1 @2)))
252 (if ((TYPE_UNSIGNED (type)
253 || tree_expr_nonnegative_p (@0))
254 && tree_nop_conversion_p (type, TREE_TYPE (@3))
255 && integer_pow2p (@2) && tree_int_cst_sgn (@2) > 0)
256 (bit_and @0 (convert (minus @1 { build_int_cst (TREE_TYPE (@1), 1); }))))))
257
258 /* X % Y is smaller than Y. */
259 (for cmp (lt ge)
260 (simplify
261 (cmp (trunc_mod @0 @1) @1)
262 (if (TYPE_UNSIGNED (TREE_TYPE (@0)))
263 { constant_boolean_node (cmp == LT_EXPR, type); })))
264 (for cmp (gt le)
265 (simplify
266 (cmp @1 (trunc_mod @0 @1))
267 (if (TYPE_UNSIGNED (TREE_TYPE (@0)))
268 { constant_boolean_node (cmp == GT_EXPR, type); })))
269
270 /* x | ~0 -> ~0 */
271 (simplify
272 (bit_ior @0 integer_all_onesp@1)
273 @1)
274
275 /* x & 0 -> 0 */
276 (simplify
277 (bit_and @0 integer_zerop@1)
278 @1)
279
280 /* x ^ x -> 0 */
281 (simplify
282 (bit_xor @0 @0)
283 { build_zero_cst (type); })
284
285 /* Canonicalize X ^ ~0 to ~X. */
286 (simplify
287 (bit_xor @0 integer_all_onesp@1)
288 (bit_not @0))
289
290 /* x & ~0 -> x */
291 (simplify
292 (bit_and @0 integer_all_onesp)
293 (non_lvalue @0))
294
295 /* x & x -> x, x | x -> x */
296 (for bitop (bit_and bit_ior)
297 (simplify
298 (bitop @0 @0)
299 (non_lvalue @0)))
300
301 /* x + (x & 1) -> (x + 1) & ~1 */
302 (simplify
303 (plus:c @0 (bit_and@2 @0 integer_onep@1))
304 (if (single_use (@2))
305 (bit_and (plus @0 @1) (bit_not @1))))
306
307 /* x & ~(x & y) -> x & ~y */
308 /* x | ~(x | y) -> x | ~y */
309 (for bitop (bit_and bit_ior)
310 (simplify
311 (bitop:c @0 (bit_not (bitop:c@2 @0 @1)))
312 (if (single_use (@2))
313 (bitop @0 (bit_not @1)))))
314
315 /* (x | y) & ~x -> y & ~x */
316 /* (x & y) | ~x -> y | ~x */
317 (for bitop (bit_and bit_ior)
318 rbitop (bit_ior bit_and)
319 (simplify
320 (bitop:c (rbitop:c @0 @1) (bit_not@2 @0))
321 (bitop @1 @2)))
322
323 /* (x & y) ^ (x | y) -> x ^ y */
324 (simplify
325 (bit_xor:c (bit_and @0 @1) (bit_ior @0 @1))
326 (bit_xor @0 @1))
327
328 /* (x ^ y) ^ (x | y) -> x & y */
329 (simplify
330 (bit_xor:c (bit_xor @0 @1) (bit_ior @0 @1))
331 (bit_and @0 @1))
332
333 /* (x & y) + (x ^ y) -> x | y */
334 /* (x & y) | (x ^ y) -> x | y */
335 /* (x & y) ^ (x ^ y) -> x | y */
336 (for op (plus bit_ior bit_xor)
337 (simplify
338 (op:c (bit_and @0 @1) (bit_xor @0 @1))
339 (bit_ior @0 @1)))
340
341 /* (x & y) + (x | y) -> x + y */
342 (simplify
343 (plus:c (bit_and @0 @1) (bit_ior @0 @1))
344 (plus @0 @1))
345
346 /* (x + y) - (x | y) -> x & y */
347 (simplify
348 (minus (plus @0 @1) (bit_ior @0 @1))
349 (if (!TYPE_OVERFLOW_SANITIZED (type) && !TYPE_OVERFLOW_TRAPS (type)
350 && !TYPE_SATURATING (type))
351 (bit_and @0 @1)))
352
353 /* (x + y) - (x & y) -> x | y */
354 (simplify
355 (minus (plus @0 @1) (bit_and @0 @1))
356 (if (!TYPE_OVERFLOW_SANITIZED (type) && !TYPE_OVERFLOW_TRAPS (type)
357 && !TYPE_SATURATING (type))
358 (bit_ior @0 @1)))
359
360 /* (x | y) - (x ^ y) -> x & y */
361 (simplify
362 (minus (bit_ior @0 @1) (bit_xor @0 @1))
363 (bit_and @0 @1))
364
365 /* (x | y) - (x & y) -> x ^ y */
366 (simplify
367 (minus (bit_ior @0 @1) (bit_and @0 @1))
368 (bit_xor @0 @1))
369
370 (simplify
371 (abs (negate @0))
372 (abs @0))
373 (simplify
374 (abs tree_expr_nonnegative_p@0)
375 @0)
376
377
378 /* Try to fold (type) X op CST -> (type) (X op ((type-x) CST))
379 when profitable.
380 For bitwise binary operations apply operand conversions to the
381 binary operation result instead of to the operands. This allows
382 to combine successive conversions and bitwise binary operations.
383 We combine the above two cases by using a conditional convert. */
384 (for bitop (bit_and bit_ior bit_xor)
385 (simplify
386 (bitop (convert @0) (convert? @1))
387 (if (((TREE_CODE (@1) == INTEGER_CST
388 && INTEGRAL_TYPE_P (TREE_TYPE (@0))
389 && int_fits_type_p (@1, TREE_TYPE (@0)))
390 || types_match (@0, @1))
391 /* ??? This transform conflicts with fold-const.c doing
392 Convert (T)(x & c) into (T)x & (T)c, if c is an integer
393 constants (if x has signed type, the sign bit cannot be set
394 in c). This folds extension into the BIT_AND_EXPR.
395 Restrict it to GIMPLE to avoid endless recursions. */
396 && (bitop != BIT_AND_EXPR || GIMPLE)
397 && (/* That's a good idea if the conversion widens the operand, thus
398 after hoisting the conversion the operation will be narrower. */
399 TYPE_PRECISION (TREE_TYPE (@0)) < TYPE_PRECISION (type)
400 /* It's also a good idea if the conversion is to a non-integer
401 mode. */
402 || GET_MODE_CLASS (TYPE_MODE (type)) != MODE_INT
403 /* Or if the precision of TO is not the same as the precision
404 of its mode. */
405 || TYPE_PRECISION (type) != GET_MODE_PRECISION (TYPE_MODE (type))))
406 (convert (bitop @0 (convert @1))))))
407
408 /* Simplify (A & B) OP0 (C & B) to (A OP0 C) & B. */
409 (for bitop (bit_and bit_ior bit_xor)
410 (simplify
411 (bitop (bit_and:c @0 @1) (bit_and @2 @1))
412 (bit_and (bitop @0 @2) @1)))
413
414 /* (x | CST1) & CST2 -> (x & CST2) | (CST1 & CST2) */
415 (simplify
416 (bit_and (bit_ior @0 CONSTANT_CLASS_P@1) CONSTANT_CLASS_P@2)
417 (bit_ior (bit_and @0 @2) (bit_and @1 @2)))
418
419 /* Combine successive equal operations with constants. */
420 (for bitop (bit_and bit_ior bit_xor)
421 (simplify
422 (bitop (bitop @0 CONSTANT_CLASS_P@1) CONSTANT_CLASS_P@2)
423 (bitop @0 (bitop @1 @2))))
424
425 /* Try simple folding for X op !X, and X op X with the help
426 of the truth_valued_p and logical_inverted_value predicates. */
427 (match truth_valued_p
428 @0
429 (if (INTEGRAL_TYPE_P (type) && TYPE_PRECISION (type) == 1)))
430 (for op (tcc_comparison truth_and truth_andif truth_or truth_orif truth_xor)
431 (match truth_valued_p
432 (op @0 @1)))
433 (match truth_valued_p
434 (truth_not @0))
435
436 (match (logical_inverted_value @0)
437 (bit_not truth_valued_p@0))
438 (match (logical_inverted_value @0)
439 (eq @0 integer_zerop))
440 (match (logical_inverted_value @0)
441 (ne truth_valued_p@0 integer_truep))
442 (match (logical_inverted_value @0)
443 (bit_xor truth_valued_p@0 integer_truep))
444
445 /* X & !X -> 0. */
446 (simplify
447 (bit_and:c @0 (logical_inverted_value @0))
448 { build_zero_cst (type); })
449 /* X | !X and X ^ !X -> 1, , if X is truth-valued. */
450 (for op (bit_ior bit_xor)
451 (simplify
452 (op:c truth_valued_p@0 (logical_inverted_value @0))
453 { constant_boolean_node (true, type); }))
454
455 (for bitop (bit_and bit_ior)
456 rbitop (bit_ior bit_and)
457 /* (x | y) & x -> x */
458 /* (x & y) | x -> x */
459 (simplify
460 (bitop:c (rbitop:c @0 @1) @0)
461 @0)
462 /* (~x | y) & x -> x & y */
463 /* (~x & y) | x -> x | y */
464 (simplify
465 (bitop:c (rbitop:c (bit_not @0) @1) @0)
466 (bitop @0 @1)))
467
468 /* If arg1 and arg2 are booleans (or any single bit type)
469 then try to simplify:
470
471 (~X & Y) -> X < Y
472 (X & ~Y) -> Y < X
473 (~X | Y) -> X <= Y
474 (X | ~Y) -> Y <= X
475
476 But only do this if our result feeds into a comparison as
477 this transformation is not always a win, particularly on
478 targets with and-not instructions.
479 -> simplify_bitwise_binary_boolean */
480 (simplify
481 (ne (bit_and:c (bit_not @0) @1) integer_zerop)
482 (if (INTEGRAL_TYPE_P (TREE_TYPE (@1))
483 && TYPE_PRECISION (TREE_TYPE (@1)) == 1)
484 (lt @0 @1)))
485 (simplify
486 (ne (bit_ior:c (bit_not @0) @1) integer_zerop)
487 (if (INTEGRAL_TYPE_P (TREE_TYPE (@1))
488 && TYPE_PRECISION (TREE_TYPE (@1)) == 1)
489 (le @0 @1)))
490
491 /* ~~x -> x */
492 (simplify
493 (bit_not (bit_not @0))
494 @0)
495
496 /* (x & ~m) | (y & m) -> ((x ^ y) & m) ^ x */
497 (simplify
498 (bit_ior:c (bit_and:c@3 @0 (bit_not @2)) (bit_and:c@4 @1 @2))
499 (if (single_use (@3) && single_use (@4))
500 (bit_xor (bit_and (bit_xor @0 @1) @2) @0)))
501
502
503 /* Associate (p +p off1) +p off2 as (p +p (off1 + off2)). */
504 (simplify
505 (pointer_plus (pointer_plus@2 @0 @1) @3)
506 (if (single_use (@2))
507 (pointer_plus @0 (plus @1 @3))))
508
509 /* Pattern match
510 tem1 = (long) ptr1;
511 tem2 = (long) ptr2;
512 tem3 = tem2 - tem1;
513 tem4 = (unsigned long) tem3;
514 tem5 = ptr1 + tem4;
515 and produce
516 tem5 = ptr2; */
517 (simplify
518 (pointer_plus @0 (convert?@2 (minus@3 (convert @1) (convert @0))))
519 /* Conditionally look through a sign-changing conversion. */
520 (if (TYPE_PRECISION (TREE_TYPE (@2)) == TYPE_PRECISION (TREE_TYPE (@3))
521 && ((GIMPLE && useless_type_conversion_p (type, TREE_TYPE (@1)))
522 || (GENERIC && type == TREE_TYPE (@1))))
523 @1))
524
525 /* Pattern match
526 tem = (sizetype) ptr;
527 tem = tem & algn;
528 tem = -tem;
529 ... = ptr p+ tem;
530 and produce the simpler and easier to analyze with respect to alignment
531 ... = ptr & ~algn; */
532 (simplify
533 (pointer_plus @0 (negate (bit_and (convert @0) INTEGER_CST@1)))
534 (with { tree algn = wide_int_to_tree (TREE_TYPE (@0), wi::bit_not (@1)); }
535 (bit_and @0 { algn; })))
536
537
538 /* We can't reassociate at all for saturating types. */
539 (if (!TYPE_SATURATING (type))
540
541 /* Contract negates. */
542 /* A + (-B) -> A - B */
543 (simplify
544 (plus:c (convert1? @0) (convert2? (negate @1)))
545 /* Apply STRIP_NOPS on @0 and the negate. */
546 (if (tree_nop_conversion_p (type, TREE_TYPE (@0))
547 && tree_nop_conversion_p (type, TREE_TYPE (@1))
548 && !TYPE_OVERFLOW_SANITIZED (type))
549 (minus (convert @0) (convert @1))))
550 /* A - (-B) -> A + B */
551 (simplify
552 (minus (convert1? @0) (convert2? (negate @1)))
553 (if (tree_nop_conversion_p (type, TREE_TYPE (@0))
554 && tree_nop_conversion_p (type, TREE_TYPE (@1))
555 && !TYPE_OVERFLOW_SANITIZED (type))
556 (plus (convert @0) (convert @1))))
557 /* -(-A) -> A */
558 (simplify
559 (negate (convert? (negate @1)))
560 (if (tree_nop_conversion_p (type, TREE_TYPE (@1))
561 && !TYPE_OVERFLOW_SANITIZED (type))
562 (convert @1)))
563
564 /* We can't reassociate floating-point or fixed-point plus or minus
565 because of saturation to +-Inf. */
566 (if (!FLOAT_TYPE_P (type) && !FIXED_POINT_TYPE_P (type))
567
568 /* Match patterns that allow contracting a plus-minus pair
569 irrespective of overflow issues. */
570 /* (A +- B) - A -> +- B */
571 /* (A +- B) -+ B -> A */
572 /* A - (A +- B) -> -+ B */
573 /* A +- (B -+ A) -> +- B */
574 (simplify
575 (minus (plus:c @0 @1) @0)
576 @1)
577 (simplify
578 (minus (minus @0 @1) @0)
579 (negate @1))
580 (simplify
581 (plus:c (minus @0 @1) @1)
582 @0)
583 (simplify
584 (minus @0 (plus:c @0 @1))
585 (negate @1))
586 (simplify
587 (minus @0 (minus @0 @1))
588 @1)
589
590 /* (A +- CST) +- CST -> A + CST */
591 (for outer_op (plus minus)
592 (for inner_op (plus minus)
593 (simplify
594 (outer_op (inner_op @0 CONSTANT_CLASS_P@1) CONSTANT_CLASS_P@2)
595 /* If the constant operation overflows we cannot do the transform
596 as we would introduce undefined overflow, for example
597 with (a - 1) + INT_MIN. */
598 (with { tree cst = fold_binary (outer_op == inner_op
599 ? PLUS_EXPR : MINUS_EXPR, type, @1, @2); }
600 (if (cst && !TREE_OVERFLOW (cst))
601 (inner_op @0 { cst; } ))))))
602
603 /* (CST - A) +- CST -> CST - A */
604 (for outer_op (plus minus)
605 (simplify
606 (outer_op (minus CONSTANT_CLASS_P@1 @0) CONSTANT_CLASS_P@2)
607 (with { tree cst = fold_binary (outer_op, type, @1, @2); }
608 (if (cst && !TREE_OVERFLOW (cst))
609 (minus { cst; } @0)))))
610
611 /* ~A + A -> -1 */
612 (simplify
613 (plus:c (bit_not @0) @0)
614 (if (!TYPE_OVERFLOW_TRAPS (type))
615 { build_all_ones_cst (type); }))
616
617 /* ~A + 1 -> -A */
618 (simplify
619 (plus (convert? (bit_not @0)) integer_each_onep)
620 (if (tree_nop_conversion_p (type, TREE_TYPE (@0)))
621 (negate (convert @0))))
622
623 /* -A - 1 -> ~A */
624 (simplify
625 (minus (convert? (negate @0)) integer_each_onep)
626 (if (!TYPE_OVERFLOW_TRAPS (type)
627 && tree_nop_conversion_p (type, TREE_TYPE (@0)))
628 (bit_not (convert @0))))
629
630 /* -1 - A -> ~A */
631 (simplify
632 (minus integer_all_onesp @0)
633 (bit_not @0))
634
635 /* (T)(P + A) - (T)P -> (T) A */
636 (for add (plus pointer_plus)
637 (simplify
638 (minus (convert (add @0 @1))
639 (convert @0))
640 (if (element_precision (type) <= element_precision (TREE_TYPE (@1))
641 /* For integer types, if A has a smaller type
642 than T the result depends on the possible
643 overflow in P + A.
644 E.g. T=size_t, A=(unsigned)429497295, P>0.
645 However, if an overflow in P + A would cause
646 undefined behavior, we can assume that there
647 is no overflow. */
648 || (INTEGRAL_TYPE_P (TREE_TYPE (@0))
649 && TYPE_OVERFLOW_UNDEFINED (TREE_TYPE (@0)))
650 /* For pointer types, if the conversion of A to the
651 final type requires a sign- or zero-extension,
652 then we have to punt - it is not defined which
653 one is correct. */
654 || (POINTER_TYPE_P (TREE_TYPE (@0))
655 && TREE_CODE (@1) == INTEGER_CST
656 && tree_int_cst_sign_bit (@1) == 0))
657 (convert @1))))))
658
659
660 /* Simplifications of MIN_EXPR and MAX_EXPR. */
661
662 (for minmax (min max)
663 (simplify
664 (minmax @0 @0)
665 @0))
666 (simplify
667 (min @0 @1)
668 (if (INTEGRAL_TYPE_P (type)
669 && TYPE_MIN_VALUE (type)
670 && operand_equal_p (@1, TYPE_MIN_VALUE (type), OEP_ONLY_CONST))
671 @1))
672 (simplify
673 (max @0 @1)
674 (if (INTEGRAL_TYPE_P (type)
675 && TYPE_MAX_VALUE (type)
676 && operand_equal_p (@1, TYPE_MAX_VALUE (type), OEP_ONLY_CONST))
677 @1))
678
679
680 /* Simplifications of shift and rotates. */
681
682 (for rotate (lrotate rrotate)
683 (simplify
684 (rotate integer_all_onesp@0 @1)
685 @0))
686
687 /* Optimize -1 >> x for arithmetic right shifts. */
688 (simplify
689 (rshift integer_all_onesp@0 @1)
690 (if (!TYPE_UNSIGNED (type)
691 && tree_expr_nonnegative_p (@1))
692 @0))
693
694 (for shiftrotate (lrotate rrotate lshift rshift)
695 (simplify
696 (shiftrotate @0 integer_zerop)
697 (non_lvalue @0))
698 (simplify
699 (shiftrotate integer_zerop@0 @1)
700 @0)
701 /* Prefer vector1 << scalar to vector1 << vector2
702 if vector2 is uniform. */
703 (for vec (VECTOR_CST CONSTRUCTOR)
704 (simplify
705 (shiftrotate @0 vec@1)
706 (with { tree tem = uniform_vector_p (@1); }
707 (if (tem)
708 (shiftrotate @0 { tem; }))))))
709
710 /* Rewrite an LROTATE_EXPR by a constant into an
711 RROTATE_EXPR by a new constant. */
712 (simplify
713 (lrotate @0 INTEGER_CST@1)
714 (rrotate @0 { fold_binary (MINUS_EXPR, TREE_TYPE (@1),
715 build_int_cst (TREE_TYPE (@1),
716 element_precision (type)), @1); }))
717
718 /* ((1 << A) & 1) != 0 -> A == 0
719 ((1 << A) & 1) == 0 -> A != 0 */
720 (for cmp (ne eq)
721 icmp (eq ne)
722 (simplify
723 (cmp (bit_and (lshift integer_onep @0) integer_onep) integer_zerop)
724 (icmp @0 { build_zero_cst (TREE_TYPE (@0)); })))
725
726 /* (CST1 << A) == CST2 -> A == ctz (CST2) - ctz (CST1)
727 (CST1 << A) != CST2 -> A != ctz (CST2) - ctz (CST1)
728 if CST2 != 0. */
729 (for cmp (ne eq)
730 (simplify
731 (cmp (lshift INTEGER_CST@0 @1) INTEGER_CST@2)
732 (with { int cand = wi::ctz (@2) - wi::ctz (@0); }
733 (if (cand < 0
734 || (!integer_zerop (@2)
735 && wi::ne_p (wi::lshift (@0, cand), @2)))
736 { constant_boolean_node (cmp == NE_EXPR, type); })
737 (if (!integer_zerop (@2)
738 && wi::eq_p (wi::lshift (@0, cand), @2))
739 (cmp @1 { build_int_cst (TREE_TYPE (@1), cand); })))))
740
741 /* Fold (X << C1) & C2 into (X << C1) & (C2 | ((1 << C1) - 1))
742 (X >> C1) & C2 into (X >> C1) & (C2 | ~((type) -1 >> C1))
743 if the new mask might be further optimized. */
744 (for shift (lshift rshift)
745 (simplify
746 (bit_and (convert?@4 (shift@5 (convert1?@3 @0) INTEGER_CST@1)) INTEGER_CST@2)
747 (if (tree_nop_conversion_p (TREE_TYPE (@4), TREE_TYPE (@5))
748 && TYPE_PRECISION (type) <= HOST_BITS_PER_WIDE_INT
749 && tree_fits_uhwi_p (@1)
750 && tree_to_uhwi (@1) > 0
751 && tree_to_uhwi (@1) < TYPE_PRECISION (type))
752 (with
753 {
754 unsigned int shiftc = tree_to_uhwi (@1);
755 unsigned HOST_WIDE_INT mask = TREE_INT_CST_LOW (@2);
756 unsigned HOST_WIDE_INT newmask, zerobits = 0;
757 tree shift_type = TREE_TYPE (@3);
758 unsigned int prec;
759
760 if (shift == LSHIFT_EXPR)
761 zerobits = ((((unsigned HOST_WIDE_INT) 1) << shiftc) - 1);
762 else if (shift == RSHIFT_EXPR
763 && (TYPE_PRECISION (shift_type)
764 == GET_MODE_PRECISION (TYPE_MODE (shift_type))))
765 {
766 prec = TYPE_PRECISION (TREE_TYPE (@3));
767 tree arg00 = @0;
768 /* See if more bits can be proven as zero because of
769 zero extension. */
770 if (@3 != @0
771 && TYPE_UNSIGNED (TREE_TYPE (@0)))
772 {
773 tree inner_type = TREE_TYPE (@0);
774 if ((TYPE_PRECISION (inner_type)
775 == GET_MODE_PRECISION (TYPE_MODE (inner_type)))
776 && TYPE_PRECISION (inner_type) < prec)
777 {
778 prec = TYPE_PRECISION (inner_type);
779 /* See if we can shorten the right shift. */
780 if (shiftc < prec)
781 shift_type = inner_type;
782 /* Otherwise X >> C1 is all zeros, so we'll optimize
783 it into (X, 0) later on by making sure zerobits
784 is all ones. */
785 }
786 }
787 zerobits = ~(unsigned HOST_WIDE_INT) 0;
788 if (shiftc < prec)
789 {
790 zerobits >>= HOST_BITS_PER_WIDE_INT - shiftc;
791 zerobits <<= prec - shiftc;
792 }
793 /* For arithmetic shift if sign bit could be set, zerobits
794 can contain actually sign bits, so no transformation is
795 possible, unless MASK masks them all away. In that
796 case the shift needs to be converted into logical shift. */
797 if (!TYPE_UNSIGNED (TREE_TYPE (@3))
798 && prec == TYPE_PRECISION (TREE_TYPE (@3)))
799 {
800 if ((mask & zerobits) == 0)
801 shift_type = unsigned_type_for (TREE_TYPE (@3));
802 else
803 zerobits = 0;
804 }
805 }
806 }
807 /* ((X << 16) & 0xff00) is (X, 0). */
808 (if ((mask & zerobits) == mask)
809 { build_int_cst (type, 0); })
810 (with { newmask = mask | zerobits; }
811 (if (newmask != mask && (newmask & (newmask + 1)) == 0)
812 (with
813 {
814 /* Only do the transformation if NEWMASK is some integer
815 mode's mask. */
816 for (prec = BITS_PER_UNIT;
817 prec < HOST_BITS_PER_WIDE_INT; prec <<= 1)
818 if (newmask == (((unsigned HOST_WIDE_INT) 1) << prec) - 1)
819 break;
820 }
821 (if (prec < HOST_BITS_PER_WIDE_INT
822 || newmask == ~(unsigned HOST_WIDE_INT) 0)
823 (with
824 { tree newmaskt = build_int_cst_type (TREE_TYPE (@2), newmask); }
825 (if (!tree_int_cst_equal (newmaskt, @2))
826 (if (shift_type != TREE_TYPE (@3)
827 && single_use (@4) && single_use (@5))
828 (bit_and (convert (shift:shift_type (convert @3) @1)) { newmaskt; }))
829 (if (shift_type == TREE_TYPE (@3))
830 (bit_and @4 { newmaskt; }))))))))))))
831
832 /* Simplifications of conversions. */
833
834 /* Basic strip-useless-type-conversions / strip_nops. */
835 (for cvt (convert view_convert float fix_trunc)
836 (simplify
837 (cvt @0)
838 (if ((GIMPLE && useless_type_conversion_p (type, TREE_TYPE (@0)))
839 || (GENERIC && type == TREE_TYPE (@0)))
840 @0)))
841
842 /* Contract view-conversions. */
843 (simplify
844 (view_convert (view_convert @0))
845 (view_convert @0))
846
847 /* For integral conversions with the same precision or pointer
848 conversions use a NOP_EXPR instead. */
849 (simplify
850 (view_convert @0)
851 (if ((INTEGRAL_TYPE_P (type) || POINTER_TYPE_P (type))
852 && (INTEGRAL_TYPE_P (TREE_TYPE (@0)) || POINTER_TYPE_P (TREE_TYPE (@0)))
853 && TYPE_PRECISION (type) == TYPE_PRECISION (TREE_TYPE (@0)))
854 (convert @0)))
855
856 /* Strip inner integral conversions that do not change precision or size. */
857 (simplify
858 (view_convert (convert@0 @1))
859 (if ((INTEGRAL_TYPE_P (TREE_TYPE (@0)) || POINTER_TYPE_P (TREE_TYPE (@0)))
860 && (INTEGRAL_TYPE_P (TREE_TYPE (@1)) || POINTER_TYPE_P (TREE_TYPE (@1)))
861 && (TYPE_PRECISION (TREE_TYPE (@0)) == TYPE_PRECISION (TREE_TYPE (@1)))
862 && (TYPE_SIZE (TREE_TYPE (@0)) == TYPE_SIZE (TREE_TYPE (@1))))
863 (view_convert @1)))
864
865 /* Re-association barriers around constants and other re-association
866 barriers can be removed. */
867 (simplify
868 (paren CONSTANT_CLASS_P@0)
869 @0)
870 (simplify
871 (paren (paren@1 @0))
872 @1)
873
874 /* Handle cases of two conversions in a row. */
875 (for ocvt (convert float fix_trunc)
876 (for icvt (convert float)
877 (simplify
878 (ocvt (icvt@1 @0))
879 (with
880 {
881 tree inside_type = TREE_TYPE (@0);
882 tree inter_type = TREE_TYPE (@1);
883 int inside_int = INTEGRAL_TYPE_P (inside_type);
884 int inside_ptr = POINTER_TYPE_P (inside_type);
885 int inside_float = FLOAT_TYPE_P (inside_type);
886 int inside_vec = VECTOR_TYPE_P (inside_type);
887 unsigned int inside_prec = TYPE_PRECISION (inside_type);
888 int inside_unsignedp = TYPE_UNSIGNED (inside_type);
889 int inter_int = INTEGRAL_TYPE_P (inter_type);
890 int inter_ptr = POINTER_TYPE_P (inter_type);
891 int inter_float = FLOAT_TYPE_P (inter_type);
892 int inter_vec = VECTOR_TYPE_P (inter_type);
893 unsigned int inter_prec = TYPE_PRECISION (inter_type);
894 int inter_unsignedp = TYPE_UNSIGNED (inter_type);
895 int final_int = INTEGRAL_TYPE_P (type);
896 int final_ptr = POINTER_TYPE_P (type);
897 int final_float = FLOAT_TYPE_P (type);
898 int final_vec = VECTOR_TYPE_P (type);
899 unsigned int final_prec = TYPE_PRECISION (type);
900 int final_unsignedp = TYPE_UNSIGNED (type);
901 }
902 /* In addition to the cases of two conversions in a row
903 handled below, if we are converting something to its own
904 type via an object of identical or wider precision, neither
905 conversion is needed. */
906 (if (((GIMPLE && useless_type_conversion_p (type, inside_type))
907 || (GENERIC
908 && TYPE_MAIN_VARIANT (type) == TYPE_MAIN_VARIANT (inside_type)))
909 && (((inter_int || inter_ptr) && final_int)
910 || (inter_float && final_float))
911 && inter_prec >= final_prec)
912 (ocvt @0))
913
914 /* Likewise, if the intermediate and initial types are either both
915 float or both integer, we don't need the middle conversion if the
916 former is wider than the latter and doesn't change the signedness
917 (for integers). Avoid this if the final type is a pointer since
918 then we sometimes need the middle conversion. Likewise if the
919 final type has a precision not equal to the size of its mode. */
920 (if (((inter_int && inside_int) || (inter_float && inside_float))
921 && (final_int || final_float)
922 && inter_prec >= inside_prec
923 && (inter_float || inter_unsignedp == inside_unsignedp)
924 && ! (final_prec != GET_MODE_PRECISION (TYPE_MODE (type))
925 && TYPE_MODE (type) == TYPE_MODE (inter_type)))
926 (ocvt @0))
927
928 /* If we have a sign-extension of a zero-extended value, we can
929 replace that by a single zero-extension. Likewise if the
930 final conversion does not change precision we can drop the
931 intermediate conversion. */
932 (if (inside_int && inter_int && final_int
933 && ((inside_prec < inter_prec && inter_prec < final_prec
934 && inside_unsignedp && !inter_unsignedp)
935 || final_prec == inter_prec))
936 (ocvt @0))
937
938 /* Two conversions in a row are not needed unless:
939 - some conversion is floating-point (overstrict for now), or
940 - some conversion is a vector (overstrict for now), or
941 - the intermediate type is narrower than both initial and
942 final, or
943 - the intermediate type and innermost type differ in signedness,
944 and the outermost type is wider than the intermediate, or
945 - the initial type is a pointer type and the precisions of the
946 intermediate and final types differ, or
947 - the final type is a pointer type and the precisions of the
948 initial and intermediate types differ. */
949 (if (! inside_float && ! inter_float && ! final_float
950 && ! inside_vec && ! inter_vec && ! final_vec
951 && (inter_prec >= inside_prec || inter_prec >= final_prec)
952 && ! (inside_int && inter_int
953 && inter_unsignedp != inside_unsignedp
954 && inter_prec < final_prec)
955 && ((inter_unsignedp && inter_prec > inside_prec)
956 == (final_unsignedp && final_prec > inter_prec))
957 && ! (inside_ptr && inter_prec != final_prec)
958 && ! (final_ptr && inside_prec != inter_prec)
959 && ! (final_prec != GET_MODE_PRECISION (TYPE_MODE (type))
960 && TYPE_MODE (type) == TYPE_MODE (inter_type)))
961 (ocvt @0))
962
963 /* A truncation to an unsigned type (a zero-extension) should be
964 canonicalized as bitwise and of a mask. */
965 (if (final_int && inter_int && inside_int
966 && final_prec == inside_prec
967 && final_prec > inter_prec
968 && inter_unsignedp)
969 (convert (bit_and @0 { wide_int_to_tree
970 (inside_type,
971 wi::mask (inter_prec, false,
972 TYPE_PRECISION (inside_type))); })))
973
974 /* If we are converting an integer to a floating-point that can
975 represent it exactly and back to an integer, we can skip the
976 floating-point conversion. */
977 (if (GIMPLE /* PR66211 */
978 && inside_int && inter_float && final_int &&
979 (unsigned) significand_size (TYPE_MODE (inter_type))
980 >= inside_prec - !inside_unsignedp)
981 (convert @0))))))
982
983 /* If we have a narrowing conversion to an integral type that is fed by a
984 BIT_AND_EXPR, we might be able to remove the BIT_AND_EXPR if it merely
985 masks off bits outside the final type (and nothing else). */
986 (simplify
987 (convert (bit_and @0 INTEGER_CST@1))
988 (if (INTEGRAL_TYPE_P (type)
989 && INTEGRAL_TYPE_P (TREE_TYPE (@0))
990 && TYPE_PRECISION (type) <= TYPE_PRECISION (TREE_TYPE (@0))
991 && operand_equal_p (@1, build_low_bits_mask (TREE_TYPE (@1),
992 TYPE_PRECISION (type)), 0))
993 (convert @0)))
994
995
996 /* (X /[ex] A) * A -> X. */
997 (simplify
998 (mult (convert? (exact_div @0 @1)) @1)
999 /* Look through a sign-changing conversion. */
1000 (convert @0))
1001
1002 /* Canonicalization of binary operations. */
1003
1004 /* Convert X + -C into X - C. */
1005 (simplify
1006 (plus @0 REAL_CST@1)
1007 (if (REAL_VALUE_NEGATIVE (TREE_REAL_CST (@1)))
1008 (with { tree tem = fold_unary (NEGATE_EXPR, type, @1); }
1009 (if (!TREE_OVERFLOW (tem) || !flag_trapping_math)
1010 (minus @0 { tem; })))))
1011
1012 /* Convert x+x into x*2.0. */
1013 (simplify
1014 (plus @0 @0)
1015 (if (SCALAR_FLOAT_TYPE_P (type))
1016 (mult @0 { build_real (type, dconst2); })))
1017
1018 (simplify
1019 (minus integer_zerop @1)
1020 (negate @1))
1021
1022 /* (ARG0 - ARG1) is the same as (-ARG1 + ARG0). So check whether
1023 ARG0 is zero and X + ARG0 reduces to X, since that would mean
1024 (-ARG1 + ARG0) reduces to -ARG1. */
1025 (simplify
1026 (minus real_zerop@0 @1)
1027 (if (fold_real_zero_addition_p (type, @0, 0))
1028 (negate @1)))
1029
1030 /* Transform x * -1 into -x. */
1031 (simplify
1032 (mult @0 integer_minus_onep)
1033 (negate @0))
1034
1035 /* COMPLEX_EXPR and REALPART/IMAGPART_EXPR cancellations. */
1036 (simplify
1037 (complex (realpart @0) (imagpart @0))
1038 @0)
1039 (simplify
1040 (realpart (complex @0 @1))
1041 @0)
1042 (simplify
1043 (imagpart (complex @0 @1))
1044 @1)
1045
1046
1047 /* BSWAP simplifications, transforms checked by gcc.dg/builtin-bswap-8.c. */
1048 (for bswap (BUILT_IN_BSWAP16 BUILT_IN_BSWAP32 BUILT_IN_BSWAP64)
1049 (simplify
1050 (bswap (bswap @0))
1051 @0)
1052 (simplify
1053 (bswap (bit_not (bswap @0)))
1054 (bit_not @0))
1055 (for bitop (bit_xor bit_ior bit_and)
1056 (simplify
1057 (bswap (bitop:c (bswap @0) @1))
1058 (bitop @0 (bswap @1)))))
1059
1060
1061 /* Combine COND_EXPRs and VEC_COND_EXPRs. */
1062
1063 /* Simplify constant conditions.
1064 Only optimize constant conditions when the selected branch
1065 has the same type as the COND_EXPR. This avoids optimizing
1066 away "c ? x : throw", where the throw has a void type.
1067 Note that we cannot throw away the fold-const.c variant nor
1068 this one as we depend on doing this transform before possibly
1069 A ? B : B -> B triggers and the fold-const.c one can optimize
1070 0 ? A : B to B even if A has side-effects. Something
1071 genmatch cannot handle. */
1072 (simplify
1073 (cond INTEGER_CST@0 @1 @2)
1074 (if (integer_zerop (@0)
1075 && (!VOID_TYPE_P (TREE_TYPE (@2))
1076 || VOID_TYPE_P (type)))
1077 @2)
1078 (if (!integer_zerop (@0)
1079 && (!VOID_TYPE_P (TREE_TYPE (@1))
1080 || VOID_TYPE_P (type)))
1081 @1))
1082 (simplify
1083 (vec_cond VECTOR_CST@0 @1 @2)
1084 (if (integer_all_onesp (@0))
1085 @1)
1086 (if (integer_zerop (@0))
1087 @2))
1088
1089 (for cnd (cond vec_cond)
1090 /* A ? B : (A ? X : C) -> A ? B : C. */
1091 (simplify
1092 (cnd @0 (cnd @0 @1 @2) @3)
1093 (cnd @0 @1 @3))
1094 (simplify
1095 (cnd @0 @1 (cnd @0 @2 @3))
1096 (cnd @0 @1 @3))
1097
1098 /* A ? B : B -> B. */
1099 (simplify
1100 (cnd @0 @1 @1)
1101 @1)
1102
1103 /* !A ? B : C -> A ? C : B. */
1104 (simplify
1105 (cnd (logical_inverted_value truth_valued_p@0) @1 @2)
1106 (cnd @0 @2 @1)))
1107
1108 /* A + (B vcmp C ? 1 : 0) -> A - (B vcmp C), since vector comparisons
1109 return all-1 or all-0 results. */
1110 /* ??? We could instead convert all instances of the vec_cond to negate,
1111 but that isn't necessarily a win on its own. */
1112 (simplify
1113 (plus:c @3 (view_convert? (vec_cond @0 integer_each_onep@1 integer_zerop@2)))
1114 (if (VECTOR_TYPE_P (type)
1115 && TYPE_VECTOR_SUBPARTS (type) == TYPE_VECTOR_SUBPARTS (TREE_TYPE (@0))
1116 && (TYPE_MODE (TREE_TYPE (type))
1117 == TYPE_MODE (TREE_TYPE (TREE_TYPE (@0)))))
1118 (minus @3 (view_convert @0))))
1119
1120 /* ... likewise A - (B vcmp C ? 1 : 0) -> A + (B vcmp C). */
1121 (simplify
1122 (minus @3 (view_convert? (vec_cond @0 integer_each_onep@1 integer_zerop@2)))
1123 (if (VECTOR_TYPE_P (type)
1124 && TYPE_VECTOR_SUBPARTS (type) == TYPE_VECTOR_SUBPARTS (TREE_TYPE (@0))
1125 && (TYPE_MODE (TREE_TYPE (type))
1126 == TYPE_MODE (TREE_TYPE (TREE_TYPE (@0)))))
1127 (plus @3 (view_convert @0))))
1128
1129 /* Simplifications of comparisons. */
1130
1131 /* We can simplify a logical negation of a comparison to the
1132 inverted comparison. As we cannot compute an expression
1133 operator using invert_tree_comparison we have to simulate
1134 that with expression code iteration. */
1135 (for cmp (tcc_comparison)
1136 icmp (inverted_tcc_comparison)
1137 ncmp (inverted_tcc_comparison_with_nans)
1138 /* Ideally we'd like to combine the following two patterns
1139 and handle some more cases by using
1140 (logical_inverted_value (cmp @0 @1))
1141 here but for that genmatch would need to "inline" that.
1142 For now implement what forward_propagate_comparison did. */
1143 (simplify
1144 (bit_not (cmp @0 @1))
1145 (if (VECTOR_TYPE_P (type)
1146 || (INTEGRAL_TYPE_P (type) && TYPE_PRECISION (type) == 1))
1147 /* Comparison inversion may be impossible for trapping math,
1148 invert_tree_comparison will tell us. But we can't use
1149 a computed operator in the replacement tree thus we have
1150 to play the trick below. */
1151 (with { enum tree_code ic = invert_tree_comparison
1152 (cmp, HONOR_NANS (@0)); }
1153 (if (ic == icmp)
1154 (icmp @0 @1))
1155 (if (ic == ncmp)
1156 (ncmp @0 @1)))))
1157 (simplify
1158 (bit_xor (cmp @0 @1) integer_truep)
1159 (with { enum tree_code ic = invert_tree_comparison
1160 (cmp, HONOR_NANS (@0)); }
1161 (if (ic == icmp)
1162 (icmp @0 @1))
1163 (if (ic == ncmp)
1164 (ncmp @0 @1)))))
1165
1166 /* Unordered tests if either argument is a NaN. */
1167 (simplify
1168 (bit_ior (unordered @0 @0) (unordered @1 @1))
1169 (if (types_match (@0, @1))
1170 (unordered @0 @1)))
1171 (simplify
1172 (bit_and (ordered @0 @0) (ordered @1 @1))
1173 (if (types_match (@0, @1))
1174 (ordered @0 @1)))
1175 (simplify
1176 (bit_ior:c (unordered @0 @0) (unordered:c@2 @0 @1))
1177 @2)
1178 (simplify
1179 (bit_and:c (ordered @0 @0) (ordered:c@2 @0 @1))
1180 @2)
1181
1182 /* -A CMP -B -> B CMP A. */
1183 (for cmp (tcc_comparison)
1184 scmp (swapped_tcc_comparison)
1185 (simplify
1186 (cmp (negate @0) (negate @1))
1187 (if (FLOAT_TYPE_P (TREE_TYPE (@0))
1188 || (ANY_INTEGRAL_TYPE_P (TREE_TYPE (@0))
1189 && TYPE_OVERFLOW_UNDEFINED (TREE_TYPE (@0))))
1190 (scmp @0 @1)))
1191 (simplify
1192 (cmp (negate @0) CONSTANT_CLASS_P@1)
1193 (if (FLOAT_TYPE_P (TREE_TYPE (@0))
1194 || (ANY_INTEGRAL_TYPE_P (TREE_TYPE (@0))
1195 && TYPE_OVERFLOW_UNDEFINED (TREE_TYPE (@0))))
1196 (with { tree tem = fold_unary (NEGATE_EXPR, TREE_TYPE (@0), @1); }
1197 (if (tem && !TREE_OVERFLOW (tem))
1198 (scmp @0 { tem; }))))))
1199
1200 /* Simplification of math builtins. */
1201
1202 (define_operator_list LOG BUILT_IN_LOGF BUILT_IN_LOG BUILT_IN_LOGL)
1203 (define_operator_list EXP BUILT_IN_EXPF BUILT_IN_EXP BUILT_IN_EXPL)
1204 (define_operator_list LOG2 BUILT_IN_LOG2F BUILT_IN_LOG2 BUILT_IN_LOG2L)
1205 (define_operator_list EXP2 BUILT_IN_EXP2F BUILT_IN_EXP2 BUILT_IN_EXP2L)
1206 (define_operator_list LOG10 BUILT_IN_LOG10F BUILT_IN_LOG10 BUILT_IN_LOG10L)
1207 (define_operator_list EXP10 BUILT_IN_EXP10F BUILT_IN_EXP10 BUILT_IN_EXP10L)
1208 (define_operator_list POW BUILT_IN_POWF BUILT_IN_POW BUILT_IN_POWL)
1209 (define_operator_list POW10 BUILT_IN_POW10F BUILT_IN_POW10 BUILT_IN_POW10L)
1210 (define_operator_list SQRT BUILT_IN_SQRTF BUILT_IN_SQRT BUILT_IN_SQRTL)
1211 (define_operator_list CBRT BUILT_IN_CBRTF BUILT_IN_CBRT BUILT_IN_CBRTL)
1212
1213
1214 /* fold_builtin_logarithm */
1215 (if (flag_unsafe_math_optimizations)
1216 /* Special case, optimize logN(expN(x)) = x. */
1217 (for logs (LOG LOG2 LOG10)
1218 exps (EXP EXP2 EXP10)
1219 (simplify
1220 (logs (exps @0))
1221 @0))
1222 /* Optimize logN(func()) for various exponential functions. We
1223 want to determine the value "x" and the power "exponent" in
1224 order to transform logN(x**exponent) into exponent*logN(x). */
1225 (for logs (LOG LOG LOG LOG
1226 LOG2 LOG2 LOG2 LOG2
1227 LOG10 LOG10 LOG10 LOG10)
1228 exps (EXP EXP2 EXP10 POW10)
1229 (simplify
1230 (logs (exps @0))
1231 (with {
1232 tree x;
1233 switch (exps)
1234 {
1235 CASE_FLT_FN (BUILT_IN_EXP):
1236 /* Prepare to do logN(exp(exponent) -> exponent*logN(e). */
1237 x = build_real (type, real_value_truncate (TYPE_MODE (type),
1238 dconst_e ()));
1239 break;
1240 CASE_FLT_FN (BUILT_IN_EXP2):
1241 /* Prepare to do logN(exp2(exponent) -> exponent*logN(2). */
1242 x = build_real (type, dconst2);
1243 break;
1244 CASE_FLT_FN (BUILT_IN_EXP10):
1245 CASE_FLT_FN (BUILT_IN_POW10):
1246 /* Prepare to do logN(exp10(exponent) -> exponent*logN(10). */
1247 {
1248 REAL_VALUE_TYPE dconst10;
1249 real_from_integer (&dconst10, VOIDmode, 10, SIGNED);
1250 x = build_real (type, dconst10);
1251 }
1252 break;
1253 }
1254 }
1255 (mult (logs { x; }) @0))))
1256 (for logs (LOG LOG
1257 LOG2 LOG2
1258 LOG10 LOG10)
1259 exps (SQRT CBRT)
1260 (simplify
1261 (logs (exps @0))
1262 (with {
1263 tree x;
1264 switch (exps)
1265 {
1266 CASE_FLT_FN (BUILT_IN_SQRT):
1267 /* Prepare to do logN(sqrt(x) -> 0.5*logN(x). */
1268 x = build_real (type, dconsthalf);
1269 break;
1270 CASE_FLT_FN (BUILT_IN_CBRT):
1271 /* Prepare to do logN(cbrt(x) -> (1/3)*logN(x). */
1272 x = build_real (type, real_value_truncate (TYPE_MODE (type),
1273 dconst_third ()));
1274 break;
1275 }
1276 }
1277 (mult { x; } (logs @0)))))
1278 /* logN(pow(x,exponent) -> exponent*logN(x). */
1279 (for logs (LOG LOG2 LOG10)
1280 pows (POW)
1281 (simplify
1282 (logs (pows @0 @1))
1283 (mult @1 (logs @0)))))
1284
1285 /* Narrowing of arithmetic and logical operations.
1286
1287 These are conceptually similar to the transformations performed for
1288 the C/C++ front-ends by shorten_binary_op and shorten_compare. Long
1289 term we want to move all that code out of the front-ends into here. */
1290
1291 /* If we have a narrowing conversion of an arithmetic operation where
1292 both operands are widening conversions from the same type as the outer
1293 narrowing conversion. Then convert the innermost operands to a suitable
1294 unsigned type (to avoid introducing undefined behaviour), perform the
1295 operation and convert the result to the desired type. */
1296 (for op (plus minus)
1297 (simplify
1298 (convert (op@4 (convert@2 @0) (convert@3 @1)))
1299 (if (INTEGRAL_TYPE_P (type)
1300 /* We check for type compatibility between @0 and @1 below,
1301 so there's no need to check that @1/@3 are integral types. */
1302 && INTEGRAL_TYPE_P (TREE_TYPE (@0))
1303 && INTEGRAL_TYPE_P (TREE_TYPE (@2))
1304 /* The precision of the type of each operand must match the
1305 precision of the mode of each operand, similarly for the
1306 result. */
1307 && (TYPE_PRECISION (TREE_TYPE (@0))
1308 == GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (@0))))
1309 && (TYPE_PRECISION (TREE_TYPE (@1))
1310 == GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (@1))))
1311 && TYPE_PRECISION (type) == GET_MODE_PRECISION (TYPE_MODE (type))
1312 /* The inner conversion must be a widening conversion. */
1313 && TYPE_PRECISION (TREE_TYPE (@2)) > TYPE_PRECISION (TREE_TYPE (@0))
1314 && types_match (@0, @1)
1315 && types_match (@0, type)
1316 && single_use (@4))
1317 (if (TYPE_OVERFLOW_WRAPS (TREE_TYPE (@0)))
1318 (convert (op @0 @1)))
1319 (with { tree utype = unsigned_type_for (TREE_TYPE (@0)); }
1320 (convert (op (convert:utype @0) (convert:utype @1)))))))
1321
1322 /* This is another case of narrowing, specifically when there's an outer
1323 BIT_AND_EXPR which masks off bits outside the type of the innermost
1324 operands. Like the previous case we have to convert the operands
1325 to unsigned types to avoid introducing undefined behaviour for the
1326 arithmetic operation. */
1327 (for op (minus plus)
1328 (simplify
1329 (bit_and (op@5 (convert@2 @0) (convert@3 @1)) INTEGER_CST@4)
1330 (if (INTEGRAL_TYPE_P (type)
1331 /* We check for type compatibility between @0 and @1 below,
1332 so there's no need to check that @1/@3 are integral types. */
1333 && INTEGRAL_TYPE_P (TREE_TYPE (@0))
1334 && INTEGRAL_TYPE_P (TREE_TYPE (@2))
1335 /* The precision of the type of each operand must match the
1336 precision of the mode of each operand, similarly for the
1337 result. */
1338 && (TYPE_PRECISION (TREE_TYPE (@0))
1339 == GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (@0))))
1340 && (TYPE_PRECISION (TREE_TYPE (@1))
1341 == GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (@1))))
1342 && TYPE_PRECISION (type) == GET_MODE_PRECISION (TYPE_MODE (type))
1343 /* The inner conversion must be a widening conversion. */
1344 && TYPE_PRECISION (TREE_TYPE (@2)) > TYPE_PRECISION (TREE_TYPE (@0))
1345 && types_match (@0, @1)
1346 && (tree_int_cst_min_precision (@4, TYPE_SIGN (TREE_TYPE (@0)))
1347 <= TYPE_PRECISION (TREE_TYPE (@0)))
1348 && (TYPE_OVERFLOW_WRAPS (TREE_TYPE (@0))
1349 || tree_int_cst_sgn (@4) >= 0)
1350 && single_use (@5))
1351 (if (TYPE_OVERFLOW_WRAPS (TREE_TYPE (@0)))
1352 (with { tree ntype = TREE_TYPE (@0); }
1353 (convert (bit_and (op @0 @1) (convert:ntype @4)))))
1354 (with { tree utype = unsigned_type_for (TREE_TYPE (@0)); }
1355 (convert (bit_and (op (convert:utype @0) (convert:utype @1))
1356 (convert:utype @4)))))))
1357