match.pd: Allow associating FLOAT_TYPE_P when flag_associative_math.
[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 || (TREE_CODE (@1) == INTEGER_CST && TREE_CODE (@3) == INTEGER_CST))
508 (pointer_plus @0 (plus @1 @3))))
509
510 /* Pattern match
511 tem1 = (long) ptr1;
512 tem2 = (long) ptr2;
513 tem3 = tem2 - tem1;
514 tem4 = (unsigned long) tem3;
515 tem5 = ptr1 + tem4;
516 and produce
517 tem5 = ptr2; */
518 (simplify
519 (pointer_plus @0 (convert?@2 (minus@3 (convert @1) (convert @0))))
520 /* Conditionally look through a sign-changing conversion. */
521 (if (TYPE_PRECISION (TREE_TYPE (@2)) == TYPE_PRECISION (TREE_TYPE (@3))
522 && ((GIMPLE && useless_type_conversion_p (type, TREE_TYPE (@1)))
523 || (GENERIC && type == TREE_TYPE (@1))))
524 @1))
525
526 /* Pattern match
527 tem = (sizetype) ptr;
528 tem = tem & algn;
529 tem = -tem;
530 ... = ptr p+ tem;
531 and produce the simpler and easier to analyze with respect to alignment
532 ... = ptr & ~algn; */
533 (simplify
534 (pointer_plus @0 (negate (bit_and (convert @0) INTEGER_CST@1)))
535 (with { tree algn = wide_int_to_tree (TREE_TYPE (@0), wi::bit_not (@1)); }
536 (bit_and @0 { algn; })))
537
538
539 /* We can't reassociate at all for saturating types. */
540 (if (!TYPE_SATURATING (type))
541
542 /* Contract negates. */
543 /* A + (-B) -> A - B */
544 (simplify
545 (plus:c (convert1? @0) (convert2? (negate @1)))
546 /* Apply STRIP_NOPS on @0 and the negate. */
547 (if (tree_nop_conversion_p (type, TREE_TYPE (@0))
548 && tree_nop_conversion_p (type, TREE_TYPE (@1))
549 && !TYPE_OVERFLOW_SANITIZED (type))
550 (minus (convert @0) (convert @1))))
551 /* A - (-B) -> A + B */
552 (simplify
553 (minus (convert1? @0) (convert2? (negate @1)))
554 (if (tree_nop_conversion_p (type, TREE_TYPE (@0))
555 && tree_nop_conversion_p (type, TREE_TYPE (@1))
556 && !TYPE_OVERFLOW_SANITIZED (type))
557 (plus (convert @0) (convert @1))))
558 /* -(-A) -> A */
559 (simplify
560 (negate (convert? (negate @1)))
561 (if (tree_nop_conversion_p (type, TREE_TYPE (@1))
562 && !TYPE_OVERFLOW_SANITIZED (type))
563 (convert @1)))
564
565 /* We can't reassociate floating-point unless -fassociative-math
566 or fixed-point plus or minus because of saturation to +-Inf. */
567 (if ((!FLOAT_TYPE_P (type) || flag_associative_math)
568 && !FIXED_POINT_TYPE_P (type))
569
570 /* Match patterns that allow contracting a plus-minus pair
571 irrespective of overflow issues. */
572 /* (A +- B) - A -> +- B */
573 /* (A +- B) -+ B -> A */
574 /* A - (A +- B) -> -+ B */
575 /* A +- (B -+ A) -> +- B */
576 (simplify
577 (minus (plus:c @0 @1) @0)
578 @1)
579 (simplify
580 (minus (minus @0 @1) @0)
581 (negate @1))
582 (simplify
583 (plus:c (minus @0 @1) @1)
584 @0)
585 (simplify
586 (minus @0 (plus:c @0 @1))
587 (negate @1))
588 (simplify
589 (minus @0 (minus @0 @1))
590 @1)
591
592 /* (A +- CST) +- CST -> A + CST */
593 (for outer_op (plus minus)
594 (for inner_op (plus minus)
595 (simplify
596 (outer_op (inner_op @0 CONSTANT_CLASS_P@1) CONSTANT_CLASS_P@2)
597 /* If the constant operation overflows we cannot do the transform
598 as we would introduce undefined overflow, for example
599 with (a - 1) + INT_MIN. */
600 (with { tree cst = fold_binary (outer_op == inner_op
601 ? PLUS_EXPR : MINUS_EXPR, type, @1, @2); }
602 (if (cst && !TREE_OVERFLOW (cst))
603 (inner_op @0 { cst; } ))))))
604
605 /* (CST - A) +- CST -> CST - A */
606 (for outer_op (plus minus)
607 (simplify
608 (outer_op (minus CONSTANT_CLASS_P@1 @0) CONSTANT_CLASS_P@2)
609 (with { tree cst = fold_binary (outer_op, type, @1, @2); }
610 (if (cst && !TREE_OVERFLOW (cst))
611 (minus { cst; } @0)))))
612
613 /* ~A + A -> -1 */
614 (simplify
615 (plus:c (bit_not @0) @0)
616 (if (!TYPE_OVERFLOW_TRAPS (type))
617 { build_all_ones_cst (type); }))
618
619 /* ~A + 1 -> -A */
620 (simplify
621 (plus (convert? (bit_not @0)) integer_each_onep)
622 (if (tree_nop_conversion_p (type, TREE_TYPE (@0)))
623 (negate (convert @0))))
624
625 /* -A - 1 -> ~A */
626 (simplify
627 (minus (convert? (negate @0)) integer_each_onep)
628 (if (!TYPE_OVERFLOW_TRAPS (type)
629 && tree_nop_conversion_p (type, TREE_TYPE (@0)))
630 (bit_not (convert @0))))
631
632 /* -1 - A -> ~A */
633 (simplify
634 (minus integer_all_onesp @0)
635 (bit_not @0))
636
637 /* (T)(P + A) - (T)P -> (T) A */
638 (for add (plus pointer_plus)
639 (simplify
640 (minus (convert (add @0 @1))
641 (convert @0))
642 (if (element_precision (type) <= element_precision (TREE_TYPE (@1))
643 /* For integer types, if A has a smaller type
644 than T the result depends on the possible
645 overflow in P + A.
646 E.g. T=size_t, A=(unsigned)429497295, P>0.
647 However, if an overflow in P + A would cause
648 undefined behavior, we can assume that there
649 is no overflow. */
650 || (INTEGRAL_TYPE_P (TREE_TYPE (@0))
651 && TYPE_OVERFLOW_UNDEFINED (TREE_TYPE (@0)))
652 /* For pointer types, if the conversion of A to the
653 final type requires a sign- or zero-extension,
654 then we have to punt - it is not defined which
655 one is correct. */
656 || (POINTER_TYPE_P (TREE_TYPE (@0))
657 && TREE_CODE (@1) == INTEGER_CST
658 && tree_int_cst_sign_bit (@1) == 0))
659 (convert @1))))))
660
661
662 /* Simplifications of MIN_EXPR and MAX_EXPR. */
663
664 (for minmax (min max)
665 (simplify
666 (minmax @0 @0)
667 @0))
668 (simplify
669 (min @0 @1)
670 (if (INTEGRAL_TYPE_P (type)
671 && TYPE_MIN_VALUE (type)
672 && operand_equal_p (@1, TYPE_MIN_VALUE (type), OEP_ONLY_CONST))
673 @1))
674 (simplify
675 (max @0 @1)
676 (if (INTEGRAL_TYPE_P (type)
677 && TYPE_MAX_VALUE (type)
678 && operand_equal_p (@1, TYPE_MAX_VALUE (type), OEP_ONLY_CONST))
679 @1))
680
681
682 /* Simplifications of shift and rotates. */
683
684 (for rotate (lrotate rrotate)
685 (simplify
686 (rotate integer_all_onesp@0 @1)
687 @0))
688
689 /* Optimize -1 >> x for arithmetic right shifts. */
690 (simplify
691 (rshift integer_all_onesp@0 @1)
692 (if (!TYPE_UNSIGNED (type)
693 && tree_expr_nonnegative_p (@1))
694 @0))
695
696 (for shiftrotate (lrotate rrotate lshift rshift)
697 (simplify
698 (shiftrotate @0 integer_zerop)
699 (non_lvalue @0))
700 (simplify
701 (shiftrotate integer_zerop@0 @1)
702 @0)
703 /* Prefer vector1 << scalar to vector1 << vector2
704 if vector2 is uniform. */
705 (for vec (VECTOR_CST CONSTRUCTOR)
706 (simplify
707 (shiftrotate @0 vec@1)
708 (with { tree tem = uniform_vector_p (@1); }
709 (if (tem)
710 (shiftrotate @0 { tem; }))))))
711
712 /* Rewrite an LROTATE_EXPR by a constant into an
713 RROTATE_EXPR by a new constant. */
714 (simplify
715 (lrotate @0 INTEGER_CST@1)
716 (rrotate @0 { fold_binary (MINUS_EXPR, TREE_TYPE (@1),
717 build_int_cst (TREE_TYPE (@1),
718 element_precision (type)), @1); }))
719
720 /* ((1 << A) & 1) != 0 -> A == 0
721 ((1 << A) & 1) == 0 -> A != 0 */
722 (for cmp (ne eq)
723 icmp (eq ne)
724 (simplify
725 (cmp (bit_and (lshift integer_onep @0) integer_onep) integer_zerop)
726 (icmp @0 { build_zero_cst (TREE_TYPE (@0)); })))
727
728 /* (CST1 << A) == CST2 -> A == ctz (CST2) - ctz (CST1)
729 (CST1 << A) != CST2 -> A != ctz (CST2) - ctz (CST1)
730 if CST2 != 0. */
731 (for cmp (ne eq)
732 (simplify
733 (cmp (lshift INTEGER_CST@0 @1) INTEGER_CST@2)
734 (with { int cand = wi::ctz (@2) - wi::ctz (@0); }
735 (if (cand < 0
736 || (!integer_zerop (@2)
737 && wi::ne_p (wi::lshift (@0, cand), @2)))
738 { constant_boolean_node (cmp == NE_EXPR, type); })
739 (if (!integer_zerop (@2)
740 && wi::eq_p (wi::lshift (@0, cand), @2))
741 (cmp @1 { build_int_cst (TREE_TYPE (@1), cand); })))))
742
743 /* Fold (X << C1) & C2 into (X << C1) & (C2 | ((1 << C1) - 1))
744 (X >> C1) & C2 into (X >> C1) & (C2 | ~((type) -1 >> C1))
745 if the new mask might be further optimized. */
746 (for shift (lshift rshift)
747 (simplify
748 (bit_and (convert?@4 (shift@5 (convert1?@3 @0) INTEGER_CST@1)) INTEGER_CST@2)
749 (if (tree_nop_conversion_p (TREE_TYPE (@4), TREE_TYPE (@5))
750 && TYPE_PRECISION (type) <= HOST_BITS_PER_WIDE_INT
751 && tree_fits_uhwi_p (@1)
752 && tree_to_uhwi (@1) > 0
753 && tree_to_uhwi (@1) < TYPE_PRECISION (type))
754 (with
755 {
756 unsigned int shiftc = tree_to_uhwi (@1);
757 unsigned HOST_WIDE_INT mask = TREE_INT_CST_LOW (@2);
758 unsigned HOST_WIDE_INT newmask, zerobits = 0;
759 tree shift_type = TREE_TYPE (@3);
760 unsigned int prec;
761
762 if (shift == LSHIFT_EXPR)
763 zerobits = ((((unsigned HOST_WIDE_INT) 1) << shiftc) - 1);
764 else if (shift == RSHIFT_EXPR
765 && (TYPE_PRECISION (shift_type)
766 == GET_MODE_PRECISION (TYPE_MODE (shift_type))))
767 {
768 prec = TYPE_PRECISION (TREE_TYPE (@3));
769 tree arg00 = @0;
770 /* See if more bits can be proven as zero because of
771 zero extension. */
772 if (@3 != @0
773 && TYPE_UNSIGNED (TREE_TYPE (@0)))
774 {
775 tree inner_type = TREE_TYPE (@0);
776 if ((TYPE_PRECISION (inner_type)
777 == GET_MODE_PRECISION (TYPE_MODE (inner_type)))
778 && TYPE_PRECISION (inner_type) < prec)
779 {
780 prec = TYPE_PRECISION (inner_type);
781 /* See if we can shorten the right shift. */
782 if (shiftc < prec)
783 shift_type = inner_type;
784 /* Otherwise X >> C1 is all zeros, so we'll optimize
785 it into (X, 0) later on by making sure zerobits
786 is all ones. */
787 }
788 }
789 zerobits = ~(unsigned HOST_WIDE_INT) 0;
790 if (shiftc < prec)
791 {
792 zerobits >>= HOST_BITS_PER_WIDE_INT - shiftc;
793 zerobits <<= prec - shiftc;
794 }
795 /* For arithmetic shift if sign bit could be set, zerobits
796 can contain actually sign bits, so no transformation is
797 possible, unless MASK masks them all away. In that
798 case the shift needs to be converted into logical shift. */
799 if (!TYPE_UNSIGNED (TREE_TYPE (@3))
800 && prec == TYPE_PRECISION (TREE_TYPE (@3)))
801 {
802 if ((mask & zerobits) == 0)
803 shift_type = unsigned_type_for (TREE_TYPE (@3));
804 else
805 zerobits = 0;
806 }
807 }
808 }
809 /* ((X << 16) & 0xff00) is (X, 0). */
810 (if ((mask & zerobits) == mask)
811 { build_int_cst (type, 0); })
812 (with { newmask = mask | zerobits; }
813 (if (newmask != mask && (newmask & (newmask + 1)) == 0)
814 (with
815 {
816 /* Only do the transformation if NEWMASK is some integer
817 mode's mask. */
818 for (prec = BITS_PER_UNIT;
819 prec < HOST_BITS_PER_WIDE_INT; prec <<= 1)
820 if (newmask == (((unsigned HOST_WIDE_INT) 1) << prec) - 1)
821 break;
822 }
823 (if (prec < HOST_BITS_PER_WIDE_INT
824 || newmask == ~(unsigned HOST_WIDE_INT) 0)
825 (with
826 { tree newmaskt = build_int_cst_type (TREE_TYPE (@2), newmask); }
827 (if (!tree_int_cst_equal (newmaskt, @2))
828 (if (shift_type != TREE_TYPE (@3)
829 && single_use (@4) && single_use (@5))
830 (bit_and (convert (shift:shift_type (convert @3) @1)) { newmaskt; }))
831 (if (shift_type == TREE_TYPE (@3))
832 (bit_and @4 { newmaskt; }))))))))))))
833
834 /* Simplifications of conversions. */
835
836 /* Basic strip-useless-type-conversions / strip_nops. */
837 (for cvt (convert view_convert float fix_trunc)
838 (simplify
839 (cvt @0)
840 (if ((GIMPLE && useless_type_conversion_p (type, TREE_TYPE (@0)))
841 || (GENERIC && type == TREE_TYPE (@0)))
842 @0)))
843
844 /* Contract view-conversions. */
845 (simplify
846 (view_convert (view_convert @0))
847 (view_convert @0))
848
849 /* For integral conversions with the same precision or pointer
850 conversions use a NOP_EXPR instead. */
851 (simplify
852 (view_convert @0)
853 (if ((INTEGRAL_TYPE_P (type) || POINTER_TYPE_P (type))
854 && (INTEGRAL_TYPE_P (TREE_TYPE (@0)) || POINTER_TYPE_P (TREE_TYPE (@0)))
855 && TYPE_PRECISION (type) == TYPE_PRECISION (TREE_TYPE (@0)))
856 (convert @0)))
857
858 /* Strip inner integral conversions that do not change precision or size. */
859 (simplify
860 (view_convert (convert@0 @1))
861 (if ((INTEGRAL_TYPE_P (TREE_TYPE (@0)) || POINTER_TYPE_P (TREE_TYPE (@0)))
862 && (INTEGRAL_TYPE_P (TREE_TYPE (@1)) || POINTER_TYPE_P (TREE_TYPE (@1)))
863 && (TYPE_PRECISION (TREE_TYPE (@0)) == TYPE_PRECISION (TREE_TYPE (@1)))
864 && (TYPE_SIZE (TREE_TYPE (@0)) == TYPE_SIZE (TREE_TYPE (@1))))
865 (view_convert @1)))
866
867 /* Re-association barriers around constants and other re-association
868 barriers can be removed. */
869 (simplify
870 (paren CONSTANT_CLASS_P@0)
871 @0)
872 (simplify
873 (paren (paren@1 @0))
874 @1)
875
876 /* Handle cases of two conversions in a row. */
877 (for ocvt (convert float fix_trunc)
878 (for icvt (convert float)
879 (simplify
880 (ocvt (icvt@1 @0))
881 (with
882 {
883 tree inside_type = TREE_TYPE (@0);
884 tree inter_type = TREE_TYPE (@1);
885 int inside_int = INTEGRAL_TYPE_P (inside_type);
886 int inside_ptr = POINTER_TYPE_P (inside_type);
887 int inside_float = FLOAT_TYPE_P (inside_type);
888 int inside_vec = VECTOR_TYPE_P (inside_type);
889 unsigned int inside_prec = TYPE_PRECISION (inside_type);
890 int inside_unsignedp = TYPE_UNSIGNED (inside_type);
891 int inter_int = INTEGRAL_TYPE_P (inter_type);
892 int inter_ptr = POINTER_TYPE_P (inter_type);
893 int inter_float = FLOAT_TYPE_P (inter_type);
894 int inter_vec = VECTOR_TYPE_P (inter_type);
895 unsigned int inter_prec = TYPE_PRECISION (inter_type);
896 int inter_unsignedp = TYPE_UNSIGNED (inter_type);
897 int final_int = INTEGRAL_TYPE_P (type);
898 int final_ptr = POINTER_TYPE_P (type);
899 int final_float = FLOAT_TYPE_P (type);
900 int final_vec = VECTOR_TYPE_P (type);
901 unsigned int final_prec = TYPE_PRECISION (type);
902 int final_unsignedp = TYPE_UNSIGNED (type);
903 }
904 /* In addition to the cases of two conversions in a row
905 handled below, if we are converting something to its own
906 type via an object of identical or wider precision, neither
907 conversion is needed. */
908 (if (((GIMPLE && useless_type_conversion_p (type, inside_type))
909 || (GENERIC
910 && TYPE_MAIN_VARIANT (type) == TYPE_MAIN_VARIANT (inside_type)))
911 && (((inter_int || inter_ptr) && final_int)
912 || (inter_float && final_float))
913 && inter_prec >= final_prec)
914 (ocvt @0))
915
916 /* Likewise, if the intermediate and initial types are either both
917 float or both integer, we don't need the middle conversion if the
918 former is wider than the latter and doesn't change the signedness
919 (for integers). Avoid this if the final type is a pointer since
920 then we sometimes need the middle conversion. Likewise if the
921 final type has a precision not equal to the size of its mode. */
922 (if (((inter_int && inside_int) || (inter_float && inside_float))
923 && (final_int || final_float)
924 && inter_prec >= inside_prec
925 && (inter_float || inter_unsignedp == inside_unsignedp)
926 && ! (final_prec != GET_MODE_PRECISION (TYPE_MODE (type))
927 && TYPE_MODE (type) == TYPE_MODE (inter_type)))
928 (ocvt @0))
929
930 /* If we have a sign-extension of a zero-extended value, we can
931 replace that by a single zero-extension. Likewise if the
932 final conversion does not change precision we can drop the
933 intermediate conversion. */
934 (if (inside_int && inter_int && final_int
935 && ((inside_prec < inter_prec && inter_prec < final_prec
936 && inside_unsignedp && !inter_unsignedp)
937 || final_prec == inter_prec))
938 (ocvt @0))
939
940 /* Two conversions in a row are not needed unless:
941 - some conversion is floating-point (overstrict for now), or
942 - some conversion is a vector (overstrict for now), or
943 - the intermediate type is narrower than both initial and
944 final, or
945 - the intermediate type and innermost type differ in signedness,
946 and the outermost type is wider than the intermediate, or
947 - the initial type is a pointer type and the precisions of the
948 intermediate and final types differ, or
949 - the final type is a pointer type and the precisions of the
950 initial and intermediate types differ. */
951 (if (! inside_float && ! inter_float && ! final_float
952 && ! inside_vec && ! inter_vec && ! final_vec
953 && (inter_prec >= inside_prec || inter_prec >= final_prec)
954 && ! (inside_int && inter_int
955 && inter_unsignedp != inside_unsignedp
956 && inter_prec < final_prec)
957 && ((inter_unsignedp && inter_prec > inside_prec)
958 == (final_unsignedp && final_prec > inter_prec))
959 && ! (inside_ptr && inter_prec != final_prec)
960 && ! (final_ptr && inside_prec != inter_prec)
961 && ! (final_prec != GET_MODE_PRECISION (TYPE_MODE (type))
962 && TYPE_MODE (type) == TYPE_MODE (inter_type)))
963 (ocvt @0))
964
965 /* A truncation to an unsigned type (a zero-extension) should be
966 canonicalized as bitwise and of a mask. */
967 (if (final_int && inter_int && inside_int
968 && final_prec == inside_prec
969 && final_prec > inter_prec
970 && inter_unsignedp)
971 (convert (bit_and @0 { wide_int_to_tree
972 (inside_type,
973 wi::mask (inter_prec, false,
974 TYPE_PRECISION (inside_type))); })))
975
976 /* If we are converting an integer to a floating-point that can
977 represent it exactly and back to an integer, we can skip the
978 floating-point conversion. */
979 (if (GIMPLE /* PR66211 */
980 && inside_int && inter_float && final_int &&
981 (unsigned) significand_size (TYPE_MODE (inter_type))
982 >= inside_prec - !inside_unsignedp)
983 (convert @0))))))
984
985 /* If we have a narrowing conversion to an integral type that is fed by a
986 BIT_AND_EXPR, we might be able to remove the BIT_AND_EXPR if it merely
987 masks off bits outside the final type (and nothing else). */
988 (simplify
989 (convert (bit_and @0 INTEGER_CST@1))
990 (if (INTEGRAL_TYPE_P (type)
991 && INTEGRAL_TYPE_P (TREE_TYPE (@0))
992 && TYPE_PRECISION (type) <= TYPE_PRECISION (TREE_TYPE (@0))
993 && operand_equal_p (@1, build_low_bits_mask (TREE_TYPE (@1),
994 TYPE_PRECISION (type)), 0))
995 (convert @0)))
996
997
998 /* (X /[ex] A) * A -> X. */
999 (simplify
1000 (mult (convert? (exact_div @0 @1)) @1)
1001 /* Look through a sign-changing conversion. */
1002 (convert @0))
1003
1004 /* Canonicalization of binary operations. */
1005
1006 /* Convert X + -C into X - C. */
1007 (simplify
1008 (plus @0 REAL_CST@1)
1009 (if (REAL_VALUE_NEGATIVE (TREE_REAL_CST (@1)))
1010 (with { tree tem = fold_unary (NEGATE_EXPR, type, @1); }
1011 (if (!TREE_OVERFLOW (tem) || !flag_trapping_math)
1012 (minus @0 { tem; })))))
1013
1014 /* Convert x+x into x*2.0. */
1015 (simplify
1016 (plus @0 @0)
1017 (if (SCALAR_FLOAT_TYPE_P (type))
1018 (mult @0 { build_real (type, dconst2); })))
1019
1020 (simplify
1021 (minus integer_zerop @1)
1022 (negate @1))
1023
1024 /* (ARG0 - ARG1) is the same as (-ARG1 + ARG0). So check whether
1025 ARG0 is zero and X + ARG0 reduces to X, since that would mean
1026 (-ARG1 + ARG0) reduces to -ARG1. */
1027 (simplify
1028 (minus real_zerop@0 @1)
1029 (if (fold_real_zero_addition_p (type, @0, 0))
1030 (negate @1)))
1031
1032 /* Transform x * -1 into -x. */
1033 (simplify
1034 (mult @0 integer_minus_onep)
1035 (negate @0))
1036
1037 /* COMPLEX_EXPR and REALPART/IMAGPART_EXPR cancellations. */
1038 (simplify
1039 (complex (realpart @0) (imagpart @0))
1040 @0)
1041 (simplify
1042 (realpart (complex @0 @1))
1043 @0)
1044 (simplify
1045 (imagpart (complex @0 @1))
1046 @1)
1047
1048
1049 /* BSWAP simplifications, transforms checked by gcc.dg/builtin-bswap-8.c. */
1050 (for bswap (BUILT_IN_BSWAP16 BUILT_IN_BSWAP32 BUILT_IN_BSWAP64)
1051 (simplify
1052 (bswap (bswap @0))
1053 @0)
1054 (simplify
1055 (bswap (bit_not (bswap @0)))
1056 (bit_not @0))
1057 (for bitop (bit_xor bit_ior bit_and)
1058 (simplify
1059 (bswap (bitop:c (bswap @0) @1))
1060 (bitop @0 (bswap @1)))))
1061
1062
1063 /* Combine COND_EXPRs and VEC_COND_EXPRs. */
1064
1065 /* Simplify constant conditions.
1066 Only optimize constant conditions when the selected branch
1067 has the same type as the COND_EXPR. This avoids optimizing
1068 away "c ? x : throw", where the throw has a void type.
1069 Note that we cannot throw away the fold-const.c variant nor
1070 this one as we depend on doing this transform before possibly
1071 A ? B : B -> B triggers and the fold-const.c one can optimize
1072 0 ? A : B to B even if A has side-effects. Something
1073 genmatch cannot handle. */
1074 (simplify
1075 (cond INTEGER_CST@0 @1 @2)
1076 (if (integer_zerop (@0)
1077 && (!VOID_TYPE_P (TREE_TYPE (@2))
1078 || VOID_TYPE_P (type)))
1079 @2)
1080 (if (!integer_zerop (@0)
1081 && (!VOID_TYPE_P (TREE_TYPE (@1))
1082 || VOID_TYPE_P (type)))
1083 @1))
1084 (simplify
1085 (vec_cond VECTOR_CST@0 @1 @2)
1086 (if (integer_all_onesp (@0))
1087 @1)
1088 (if (integer_zerop (@0))
1089 @2))
1090
1091 (for cnd (cond vec_cond)
1092 /* A ? B : (A ? X : C) -> A ? B : C. */
1093 (simplify
1094 (cnd @0 (cnd @0 @1 @2) @3)
1095 (cnd @0 @1 @3))
1096 (simplify
1097 (cnd @0 @1 (cnd @0 @2 @3))
1098 (cnd @0 @1 @3))
1099
1100 /* A ? B : B -> B. */
1101 (simplify
1102 (cnd @0 @1 @1)
1103 @1)
1104
1105 /* !A ? B : C -> A ? C : B. */
1106 (simplify
1107 (cnd (logical_inverted_value truth_valued_p@0) @1 @2)
1108 (cnd @0 @2 @1)))
1109
1110 /* A + (B vcmp C ? 1 : 0) -> A - (B vcmp C), since vector comparisons
1111 return all-1 or all-0 results. */
1112 /* ??? We could instead convert all instances of the vec_cond to negate,
1113 but that isn't necessarily a win on its own. */
1114 (simplify
1115 (plus:c @3 (view_convert? (vec_cond @0 integer_each_onep@1 integer_zerop@2)))
1116 (if (VECTOR_TYPE_P (type)
1117 && TYPE_VECTOR_SUBPARTS (type) == TYPE_VECTOR_SUBPARTS (TREE_TYPE (@0))
1118 && (TYPE_MODE (TREE_TYPE (type))
1119 == TYPE_MODE (TREE_TYPE (TREE_TYPE (@0)))))
1120 (minus @3 (view_convert @0))))
1121
1122 /* ... likewise A - (B vcmp C ? 1 : 0) -> A + (B vcmp C). */
1123 (simplify
1124 (minus @3 (view_convert? (vec_cond @0 integer_each_onep@1 integer_zerop@2)))
1125 (if (VECTOR_TYPE_P (type)
1126 && TYPE_VECTOR_SUBPARTS (type) == TYPE_VECTOR_SUBPARTS (TREE_TYPE (@0))
1127 && (TYPE_MODE (TREE_TYPE (type))
1128 == TYPE_MODE (TREE_TYPE (TREE_TYPE (@0)))))
1129 (plus @3 (view_convert @0))))
1130
1131 /* Simplifications of comparisons. */
1132
1133 /* We can simplify a logical negation of a comparison to the
1134 inverted comparison. As we cannot compute an expression
1135 operator using invert_tree_comparison we have to simulate
1136 that with expression code iteration. */
1137 (for cmp (tcc_comparison)
1138 icmp (inverted_tcc_comparison)
1139 ncmp (inverted_tcc_comparison_with_nans)
1140 /* Ideally we'd like to combine the following two patterns
1141 and handle some more cases by using
1142 (logical_inverted_value (cmp @0 @1))
1143 here but for that genmatch would need to "inline" that.
1144 For now implement what forward_propagate_comparison did. */
1145 (simplify
1146 (bit_not (cmp @0 @1))
1147 (if (VECTOR_TYPE_P (type)
1148 || (INTEGRAL_TYPE_P (type) && TYPE_PRECISION (type) == 1))
1149 /* Comparison inversion may be impossible for trapping math,
1150 invert_tree_comparison will tell us. But we can't use
1151 a computed operator in the replacement tree thus we have
1152 to play the trick below. */
1153 (with { enum tree_code ic = invert_tree_comparison
1154 (cmp, HONOR_NANS (@0)); }
1155 (if (ic == icmp)
1156 (icmp @0 @1))
1157 (if (ic == ncmp)
1158 (ncmp @0 @1)))))
1159 (simplify
1160 (bit_xor (cmp @0 @1) integer_truep)
1161 (with { enum tree_code ic = invert_tree_comparison
1162 (cmp, HONOR_NANS (@0)); }
1163 (if (ic == icmp)
1164 (icmp @0 @1))
1165 (if (ic == ncmp)
1166 (ncmp @0 @1)))))
1167
1168 /* Unordered tests if either argument is a NaN. */
1169 (simplify
1170 (bit_ior (unordered @0 @0) (unordered @1 @1))
1171 (if (types_match (@0, @1))
1172 (unordered @0 @1)))
1173 (simplify
1174 (bit_and (ordered @0 @0) (ordered @1 @1))
1175 (if (types_match (@0, @1))
1176 (ordered @0 @1)))
1177 (simplify
1178 (bit_ior:c (unordered @0 @0) (unordered:c@2 @0 @1))
1179 @2)
1180 (simplify
1181 (bit_and:c (ordered @0 @0) (ordered:c@2 @0 @1))
1182 @2)
1183
1184 /* -A CMP -B -> B CMP A. */
1185 (for cmp (tcc_comparison)
1186 scmp (swapped_tcc_comparison)
1187 (simplify
1188 (cmp (negate @0) (negate @1))
1189 (if (FLOAT_TYPE_P (TREE_TYPE (@0))
1190 || (ANY_INTEGRAL_TYPE_P (TREE_TYPE (@0))
1191 && TYPE_OVERFLOW_UNDEFINED (TREE_TYPE (@0))))
1192 (scmp @0 @1)))
1193 (simplify
1194 (cmp (negate @0) CONSTANT_CLASS_P@1)
1195 (if (FLOAT_TYPE_P (TREE_TYPE (@0))
1196 || (ANY_INTEGRAL_TYPE_P (TREE_TYPE (@0))
1197 && TYPE_OVERFLOW_UNDEFINED (TREE_TYPE (@0))))
1198 (with { tree tem = fold_unary (NEGATE_EXPR, TREE_TYPE (@0), @1); }
1199 (if (tem && !TREE_OVERFLOW (tem))
1200 (scmp @0 { tem; }))))))
1201
1202 /* Simplification of math builtins. */
1203
1204 (define_operator_list LOG BUILT_IN_LOGF BUILT_IN_LOG BUILT_IN_LOGL)
1205 (define_operator_list EXP BUILT_IN_EXPF BUILT_IN_EXP BUILT_IN_EXPL)
1206 (define_operator_list LOG2 BUILT_IN_LOG2F BUILT_IN_LOG2 BUILT_IN_LOG2L)
1207 (define_operator_list EXP2 BUILT_IN_EXP2F BUILT_IN_EXP2 BUILT_IN_EXP2L)
1208 (define_operator_list LOG10 BUILT_IN_LOG10F BUILT_IN_LOG10 BUILT_IN_LOG10L)
1209 (define_operator_list EXP10 BUILT_IN_EXP10F BUILT_IN_EXP10 BUILT_IN_EXP10L)
1210 (define_operator_list POW BUILT_IN_POWF BUILT_IN_POW BUILT_IN_POWL)
1211 (define_operator_list POW10 BUILT_IN_POW10F BUILT_IN_POW10 BUILT_IN_POW10L)
1212 (define_operator_list SQRT BUILT_IN_SQRTF BUILT_IN_SQRT BUILT_IN_SQRTL)
1213 (define_operator_list CBRT BUILT_IN_CBRTF BUILT_IN_CBRT BUILT_IN_CBRTL)
1214
1215
1216 /* fold_builtin_logarithm */
1217 (if (flag_unsafe_math_optimizations)
1218 /* Special case, optimize logN(expN(x)) = x. */
1219 (for logs (LOG LOG2 LOG10)
1220 exps (EXP EXP2 EXP10)
1221 (simplify
1222 (logs (exps @0))
1223 @0))
1224 /* Optimize logN(func()) for various exponential functions. We
1225 want to determine the value "x" and the power "exponent" in
1226 order to transform logN(x**exponent) into exponent*logN(x). */
1227 (for logs (LOG LOG LOG LOG
1228 LOG2 LOG2 LOG2 LOG2
1229 LOG10 LOG10 LOG10 LOG10)
1230 exps (EXP EXP2 EXP10 POW10)
1231 (simplify
1232 (logs (exps @0))
1233 (with {
1234 tree x;
1235 switch (exps)
1236 {
1237 CASE_FLT_FN (BUILT_IN_EXP):
1238 /* Prepare to do logN(exp(exponent) -> exponent*logN(e). */
1239 x = build_real (type, real_value_truncate (TYPE_MODE (type),
1240 dconst_e ()));
1241 break;
1242 CASE_FLT_FN (BUILT_IN_EXP2):
1243 /* Prepare to do logN(exp2(exponent) -> exponent*logN(2). */
1244 x = build_real (type, dconst2);
1245 break;
1246 CASE_FLT_FN (BUILT_IN_EXP10):
1247 CASE_FLT_FN (BUILT_IN_POW10):
1248 /* Prepare to do logN(exp10(exponent) -> exponent*logN(10). */
1249 {
1250 REAL_VALUE_TYPE dconst10;
1251 real_from_integer (&dconst10, VOIDmode, 10, SIGNED);
1252 x = build_real (type, dconst10);
1253 }
1254 break;
1255 }
1256 }
1257 (mult (logs { x; }) @0))))
1258 (for logs (LOG LOG
1259 LOG2 LOG2
1260 LOG10 LOG10)
1261 exps (SQRT CBRT)
1262 (simplify
1263 (logs (exps @0))
1264 (with {
1265 tree x;
1266 switch (exps)
1267 {
1268 CASE_FLT_FN (BUILT_IN_SQRT):
1269 /* Prepare to do logN(sqrt(x) -> 0.5*logN(x). */
1270 x = build_real (type, dconsthalf);
1271 break;
1272 CASE_FLT_FN (BUILT_IN_CBRT):
1273 /* Prepare to do logN(cbrt(x) -> (1/3)*logN(x). */
1274 x = build_real (type, real_value_truncate (TYPE_MODE (type),
1275 dconst_third ()));
1276 break;
1277 }
1278 }
1279 (mult { x; } (logs @0)))))
1280 /* logN(pow(x,exponent) -> exponent*logN(x). */
1281 (for logs (LOG LOG2 LOG10)
1282 pows (POW)
1283 (simplify
1284 (logs (pows @0 @1))
1285 (mult @1 (logs @0)))))
1286
1287 /* Narrowing of arithmetic and logical operations.
1288
1289 These are conceptually similar to the transformations performed for
1290 the C/C++ front-ends by shorten_binary_op and shorten_compare. Long
1291 term we want to move all that code out of the front-ends into here. */
1292
1293 /* If we have a narrowing conversion of an arithmetic operation where
1294 both operands are widening conversions from the same type as the outer
1295 narrowing conversion. Then convert the innermost operands to a suitable
1296 unsigned type (to avoid introducing undefined behaviour), perform the
1297 operation and convert the result to the desired type. */
1298 (for op (plus minus)
1299 (simplify
1300 (convert (op@4 (convert@2 @0) (convert@3 @1)))
1301 (if (INTEGRAL_TYPE_P (type)
1302 /* We check for type compatibility between @0 and @1 below,
1303 so there's no need to check that @1/@3 are integral types. */
1304 && INTEGRAL_TYPE_P (TREE_TYPE (@0))
1305 && INTEGRAL_TYPE_P (TREE_TYPE (@2))
1306 /* The precision of the type of each operand must match the
1307 precision of the mode of each operand, similarly for the
1308 result. */
1309 && (TYPE_PRECISION (TREE_TYPE (@0))
1310 == GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (@0))))
1311 && (TYPE_PRECISION (TREE_TYPE (@1))
1312 == GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (@1))))
1313 && TYPE_PRECISION (type) == GET_MODE_PRECISION (TYPE_MODE (type))
1314 /* The inner conversion must be a widening conversion. */
1315 && TYPE_PRECISION (TREE_TYPE (@2)) > TYPE_PRECISION (TREE_TYPE (@0))
1316 && types_match (@0, @1)
1317 && types_match (@0, type)
1318 && single_use (@4))
1319 (if (TYPE_OVERFLOW_WRAPS (TREE_TYPE (@0)))
1320 (convert (op @0 @1)))
1321 (with { tree utype = unsigned_type_for (TREE_TYPE (@0)); }
1322 (convert (op (convert:utype @0) (convert:utype @1)))))))
1323
1324 /* This is another case of narrowing, specifically when there's an outer
1325 BIT_AND_EXPR which masks off bits outside the type of the innermost
1326 operands. Like the previous case we have to convert the operands
1327 to unsigned types to avoid introducing undefined behaviour for the
1328 arithmetic operation. */
1329 (for op (minus plus)
1330 (simplify
1331 (bit_and (op@5 (convert@2 @0) (convert@3 @1)) INTEGER_CST@4)
1332 (if (INTEGRAL_TYPE_P (type)
1333 /* We check for type compatibility between @0 and @1 below,
1334 so there's no need to check that @1/@3 are integral types. */
1335 && INTEGRAL_TYPE_P (TREE_TYPE (@0))
1336 && INTEGRAL_TYPE_P (TREE_TYPE (@2))
1337 /* The precision of the type of each operand must match the
1338 precision of the mode of each operand, similarly for the
1339 result. */
1340 && (TYPE_PRECISION (TREE_TYPE (@0))
1341 == GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (@0))))
1342 && (TYPE_PRECISION (TREE_TYPE (@1))
1343 == GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (@1))))
1344 && TYPE_PRECISION (type) == GET_MODE_PRECISION (TYPE_MODE (type))
1345 /* The inner conversion must be a widening conversion. */
1346 && TYPE_PRECISION (TREE_TYPE (@2)) > TYPE_PRECISION (TREE_TYPE (@0))
1347 && types_match (@0, @1)
1348 && (tree_int_cst_min_precision (@4, TYPE_SIGN (TREE_TYPE (@0)))
1349 <= TYPE_PRECISION (TREE_TYPE (@0)))
1350 && (TYPE_OVERFLOW_WRAPS (TREE_TYPE (@0))
1351 || tree_int_cst_sgn (@4) >= 0)
1352 && single_use (@5))
1353 (if (TYPE_OVERFLOW_WRAPS (TREE_TYPE (@0)))
1354 (with { tree ntype = TREE_TYPE (@0); }
1355 (convert (bit_and (op @0 @1) (convert:ntype @4)))))
1356 (with { tree utype = unsigned_type_for (TREE_TYPE (@0)); }
1357 (convert (bit_and (op (convert:utype @0) (convert:utype @1))
1358 (convert:utype @4)))))))
1359