PR67945: Fix oscillation between pow representations
[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 integer_nonzerop
30 real_zerop real_onep real_minus_onep
31 zerop
32 CONSTANT_CLASS_P
33 tree_expr_nonnegative_p
34 integer_pow2p
35 HONOR_NANS)
36
37 /* Operator lists. */
38 (define_operator_list tcc_comparison
39 lt le eq ne ge gt unordered ordered unlt unle ungt unge uneq ltgt)
40 (define_operator_list inverted_tcc_comparison
41 ge gt ne eq lt le ordered unordered ge gt le lt ltgt uneq)
42 (define_operator_list inverted_tcc_comparison_with_nans
43 unge ungt ne eq unlt unle ordered unordered ge gt le lt ltgt uneq)
44 (define_operator_list swapped_tcc_comparison
45 gt ge eq ne le lt unordered ordered ungt unge unlt unle uneq ltgt)
46 (define_operator_list simple_comparison lt le eq ne ge gt)
47 (define_operator_list swapped_simple_comparison gt ge eq ne le lt)
48
49 (define_operator_list LOG BUILT_IN_LOGF BUILT_IN_LOG BUILT_IN_LOGL)
50 (define_operator_list EXP BUILT_IN_EXPF BUILT_IN_EXP BUILT_IN_EXPL)
51 (define_operator_list LOG2 BUILT_IN_LOG2F BUILT_IN_LOG2 BUILT_IN_LOG2L)
52 (define_operator_list EXP2 BUILT_IN_EXP2F BUILT_IN_EXP2 BUILT_IN_EXP2L)
53 (define_operator_list LOG10 BUILT_IN_LOG10F BUILT_IN_LOG10 BUILT_IN_LOG10L)
54 (define_operator_list EXP10 BUILT_IN_EXP10F BUILT_IN_EXP10 BUILT_IN_EXP10L)
55 (define_operator_list POW BUILT_IN_POWF BUILT_IN_POW BUILT_IN_POWL)
56 (define_operator_list POW10 BUILT_IN_POW10F BUILT_IN_POW10 BUILT_IN_POW10L)
57 (define_operator_list SQRT BUILT_IN_SQRTF BUILT_IN_SQRT BUILT_IN_SQRTL)
58 (define_operator_list CBRT BUILT_IN_CBRTF BUILT_IN_CBRT BUILT_IN_CBRTL)
59 (define_operator_list SIN BUILT_IN_SINF BUILT_IN_SIN BUILT_IN_SINL)
60 (define_operator_list COS BUILT_IN_COSF BUILT_IN_COS BUILT_IN_COSL)
61 (define_operator_list TAN BUILT_IN_TANF BUILT_IN_TAN BUILT_IN_TANL)
62 (define_operator_list COSH BUILT_IN_COSHF BUILT_IN_COSH BUILT_IN_COSHL)
63 (define_operator_list CEXPI BUILT_IN_CEXPIF BUILT_IN_CEXPI BUILT_IN_CEXPIL)
64
65 /* Simplifications of operations with one constant operand and
66 simplifications to constants or single values. */
67
68 (for op (plus pointer_plus minus bit_ior bit_xor)
69 (simplify
70 (op @0 integer_zerop)
71 (non_lvalue @0)))
72
73 /* 0 +p index -> (type)index */
74 (simplify
75 (pointer_plus integer_zerop @1)
76 (non_lvalue (convert @1)))
77
78 /* See if ARG1 is zero and X + ARG1 reduces to X.
79 Likewise if the operands are reversed. */
80 (simplify
81 (plus:c @0 real_zerop@1)
82 (if (fold_real_zero_addition_p (type, @1, 0))
83 (non_lvalue @0)))
84
85 /* See if ARG1 is zero and X - ARG1 reduces to X. */
86 (simplify
87 (minus @0 real_zerop@1)
88 (if (fold_real_zero_addition_p (type, @1, 1))
89 (non_lvalue @0)))
90
91 /* Simplify x - x.
92 This is unsafe for certain floats even in non-IEEE formats.
93 In IEEE, it is unsafe because it does wrong for NaNs.
94 Also note that operand_equal_p is always false if an operand
95 is volatile. */
96 (simplify
97 (minus @0 @0)
98 (if (!FLOAT_TYPE_P (type) || !HONOR_NANS (type))
99 { build_zero_cst (type); }))
100
101 (simplify
102 (mult @0 integer_zerop@1)
103 @1)
104
105 /* Maybe fold x * 0 to 0. The expressions aren't the same
106 when x is NaN, since x * 0 is also NaN. Nor are they the
107 same in modes with signed zeros, since multiplying a
108 negative value by 0 gives -0, not +0. */
109 (simplify
110 (mult @0 real_zerop@1)
111 (if (!HONOR_NANS (type) && !HONOR_SIGNED_ZEROS (type))
112 @1))
113
114 /* In IEEE floating point, x*1 is not equivalent to x for snans.
115 Likewise for complex arithmetic with signed zeros. */
116 (simplify
117 (mult @0 real_onep)
118 (if (!HONOR_SNANS (type)
119 && (!HONOR_SIGNED_ZEROS (type)
120 || !COMPLEX_FLOAT_TYPE_P (type)))
121 (non_lvalue @0)))
122
123 /* Transform x * -1.0 into -x. */
124 (simplify
125 (mult @0 real_minus_onep)
126 (if (!HONOR_SNANS (type)
127 && (!HONOR_SIGNED_ZEROS (type)
128 || !COMPLEX_FLOAT_TYPE_P (type)))
129 (negate @0)))
130
131 /* Make sure to preserve divisions by zero. This is the reason why
132 we don't simplify x / x to 1 or 0 / x to 0. */
133 (for op (mult trunc_div ceil_div floor_div round_div exact_div)
134 (simplify
135 (op @0 integer_onep)
136 (non_lvalue @0)))
137
138 /* X / -1 is -X. */
139 (for div (trunc_div ceil_div floor_div round_div exact_div)
140 (simplify
141 (div @0 integer_minus_onep@1)
142 (if (!TYPE_UNSIGNED (type))
143 (negate @0))))
144
145 /* For unsigned integral types, FLOOR_DIV_EXPR is the same as
146 TRUNC_DIV_EXPR. Rewrite into the latter in this case. */
147 (simplify
148 (floor_div @0 @1)
149 (if ((INTEGRAL_TYPE_P (type) || VECTOR_INTEGER_TYPE_P (type))
150 && TYPE_UNSIGNED (type))
151 (trunc_div @0 @1)))
152
153 /* Combine two successive divisions. Note that combining ceil_div
154 and floor_div is trickier and combining round_div even more so. */
155 (for div (trunc_div exact_div)
156 (simplify
157 (div (div @0 INTEGER_CST@1) INTEGER_CST@2)
158 (with {
159 bool overflow_p;
160 wide_int mul = wi::mul (@1, @2, TYPE_SIGN (type), &overflow_p);
161 }
162 (if (!overflow_p)
163 (div @0 { wide_int_to_tree (type, mul); })
164 (if (TYPE_UNSIGNED (type)
165 || mul != wi::min_value (TYPE_PRECISION (type), SIGNED))
166 { build_zero_cst (type); })))))
167
168 /* Optimize A / A to 1.0 if we don't care about
169 NaNs or Infinities. */
170 (simplify
171 (rdiv @0 @0)
172 (if (FLOAT_TYPE_P (type)
173 && ! HONOR_NANS (type)
174 && ! HONOR_INFINITIES (type))
175 { build_one_cst (type); }))
176
177 /* Optimize -A / A to -1.0 if we don't care about
178 NaNs or Infinities. */
179 (simplify
180 (rdiv:c @0 (negate @0))
181 (if (FLOAT_TYPE_P (type)
182 && ! HONOR_NANS (type)
183 && ! HONOR_INFINITIES (type))
184 { build_minus_one_cst (type); }))
185
186 /* In IEEE floating point, x/1 is not equivalent to x for snans. */
187 (simplify
188 (rdiv @0 real_onep)
189 (if (!HONOR_SNANS (type))
190 (non_lvalue @0)))
191
192 /* In IEEE floating point, x/-1 is not equivalent to -x for snans. */
193 (simplify
194 (rdiv @0 real_minus_onep)
195 (if (!HONOR_SNANS (type))
196 (negate @0)))
197
198 /* If ARG1 is a constant, we can convert this to a multiply by the
199 reciprocal. This does not have the same rounding properties,
200 so only do this if -freciprocal-math. We can actually
201 always safely do it if ARG1 is a power of two, but it's hard to
202 tell if it is or not in a portable manner. */
203 (for cst (REAL_CST COMPLEX_CST VECTOR_CST)
204 (simplify
205 (rdiv @0 cst@1)
206 (if (optimize)
207 (if (flag_reciprocal_math
208 && !real_zerop (@1))
209 (with
210 { tree tem = const_binop (RDIV_EXPR, type, build_one_cst (type), @1); }
211 (if (tem)
212 (mult @0 { tem; } )))
213 (if (cst != COMPLEX_CST)
214 (with { tree inverse = exact_inverse (type, @1); }
215 (if (inverse)
216 (mult @0 { inverse; } ))))))))
217
218 /* Same applies to modulo operations, but fold is inconsistent here
219 and simplifies 0 % x to 0, only preserving literal 0 % 0. */
220 (for mod (ceil_mod floor_mod round_mod trunc_mod)
221 /* 0 % X is always zero. */
222 (simplify
223 (mod integer_zerop@0 @1)
224 /* But not for 0 % 0 so that we can get the proper warnings and errors. */
225 (if (!integer_zerop (@1))
226 @0))
227 /* X % 1 is always zero. */
228 (simplify
229 (mod @0 integer_onep)
230 { build_zero_cst (type); })
231 /* X % -1 is zero. */
232 (simplify
233 (mod @0 integer_minus_onep@1)
234 (if (!TYPE_UNSIGNED (type))
235 { build_zero_cst (type); }))
236 /* (X % Y) % Y is just X % Y. */
237 (simplify
238 (mod (mod@2 @0 @1) @1)
239 @2)
240 /* From extract_muldiv_1: (X * C1) % C2 is zero if C1 is a multiple of C2. */
241 (simplify
242 (mod (mult @0 INTEGER_CST@1) INTEGER_CST@2)
243 (if (ANY_INTEGRAL_TYPE_P (type)
244 && TYPE_OVERFLOW_UNDEFINED (type)
245 && wi::multiple_of_p (@1, @2, TYPE_SIGN (type)))
246 { build_zero_cst (type); })))
247
248 /* X % -C is the same as X % C. */
249 (simplify
250 (trunc_mod @0 INTEGER_CST@1)
251 (if (TYPE_SIGN (type) == SIGNED
252 && !TREE_OVERFLOW (@1)
253 && wi::neg_p (@1)
254 && !TYPE_OVERFLOW_TRAPS (type)
255 /* Avoid this transformation if C is INT_MIN, i.e. C == -C. */
256 && !sign_bit_p (@1, @1))
257 (trunc_mod @0 (negate @1))))
258
259 /* X % -Y is the same as X % Y. */
260 (simplify
261 (trunc_mod @0 (convert? (negate @1)))
262 (if (!TYPE_UNSIGNED (type)
263 && !TYPE_OVERFLOW_TRAPS (type)
264 && tree_nop_conversion_p (type, TREE_TYPE (@1)))
265 (trunc_mod @0 (convert @1))))
266
267 /* X - (X / Y) * Y is the same as X % Y. */
268 (simplify
269 (minus (convert1? @0) (convert2? (mult (trunc_div @0 @1) @1)))
270 (if ((INTEGRAL_TYPE_P (type) || VECTOR_INTEGER_TYPE_P (type))
271 && TYPE_UNSIGNED (TREE_TYPE (@0)) == TYPE_UNSIGNED (type))
272 (trunc_mod (convert @0) (convert @1))))
273
274 /* Optimize TRUNC_MOD_EXPR by a power of two into a BIT_AND_EXPR,
275 i.e. "X % C" into "X & (C - 1)", if X and C are positive.
276 Also optimize A % (C << N) where C is a power of 2,
277 to A & ((C << N) - 1). */
278 (match (power_of_two_cand @1)
279 INTEGER_CST@1)
280 (match (power_of_two_cand @1)
281 (lshift INTEGER_CST@1 @2))
282 (for mod (trunc_mod floor_mod)
283 (simplify
284 (mod @0 (convert?@3 (power_of_two_cand@1 @2)))
285 (if ((TYPE_UNSIGNED (type)
286 || tree_expr_nonnegative_p (@0))
287 && tree_nop_conversion_p (type, TREE_TYPE (@3))
288 && integer_pow2p (@2) && tree_int_cst_sgn (@2) > 0)
289 (bit_and @0 (convert (minus @1 { build_int_cst (TREE_TYPE (@1), 1); }))))))
290
291 /* Simplify (unsigned t * 2)/2 -> unsigned t & 0x7FFFFFFF. */
292 (simplify
293 (trunc_div (mult @0 integer_pow2p@1) @1)
294 (if (TYPE_UNSIGNED (TREE_TYPE (@0)))
295 (bit_and @0 { wide_int_to_tree
296 (type, wi::mask (TYPE_PRECISION (type) - wi::exact_log2 (@1),
297 false, TYPE_PRECISION (type))); })))
298
299 /* Simplify (unsigned t / 2) * 2 -> unsigned t & ~1. */
300 (simplify
301 (mult (trunc_div @0 integer_pow2p@1) @1)
302 (if (TYPE_UNSIGNED (TREE_TYPE (@0)))
303 (bit_and @0 (negate @1))))
304
305 /* Simplify (t * 2) / 2) -> t. */
306 (for div (trunc_div ceil_div floor_div round_div exact_div)
307 (simplify
308 (div (mult @0 @1) @1)
309 (if (ANY_INTEGRAL_TYPE_P (type)
310 && TYPE_OVERFLOW_UNDEFINED (type))
311 @0)))
312
313 (for op (negate abs)
314 /* Simplify cos(-x) and cos(|x|) -> cos(x). Similarly for cosh. */
315 (for coss (COS COSH)
316 (simplify
317 (coss (op @0))
318 (coss @0)))
319 /* Simplify pow(-x, y) and pow(|x|,y) -> pow(x,y) if y is an even integer. */
320 (for pows (POW)
321 (simplify
322 (pows (op @0) REAL_CST@1)
323 (with { HOST_WIDE_INT n; }
324 (if (real_isinteger (&TREE_REAL_CST (@1), &n) && (n & 1) == 0)
325 (pows @0 @1))))))
326
327 /* X % Y is smaller than Y. */
328 (for cmp (lt ge)
329 (simplify
330 (cmp (trunc_mod @0 @1) @1)
331 (if (TYPE_UNSIGNED (TREE_TYPE (@0)))
332 { constant_boolean_node (cmp == LT_EXPR, type); })))
333 (for cmp (gt le)
334 (simplify
335 (cmp @1 (trunc_mod @0 @1))
336 (if (TYPE_UNSIGNED (TREE_TYPE (@0)))
337 { constant_boolean_node (cmp == GT_EXPR, type); })))
338
339 /* x | ~0 -> ~0 */
340 (simplify
341 (bit_ior @0 integer_all_onesp@1)
342 @1)
343
344 /* x & 0 -> 0 */
345 (simplify
346 (bit_and @0 integer_zerop@1)
347 @1)
348
349 /* ~x | x -> -1 */
350 /* ~x ^ x -> -1 */
351 /* ~x + x -> -1 */
352 (for op (bit_ior bit_xor plus)
353 (simplify
354 (op:c (convert? @0) (convert? (bit_not @0)))
355 (convert { build_all_ones_cst (TREE_TYPE (@0)); })))
356
357 /* x ^ x -> 0 */
358 (simplify
359 (bit_xor @0 @0)
360 { build_zero_cst (type); })
361
362 /* Canonicalize X ^ ~0 to ~X. */
363 (simplify
364 (bit_xor @0 integer_all_onesp@1)
365 (bit_not @0))
366
367 /* x & ~0 -> x */
368 (simplify
369 (bit_and @0 integer_all_onesp)
370 (non_lvalue @0))
371
372 /* x & x -> x, x | x -> x */
373 (for bitop (bit_and bit_ior)
374 (simplify
375 (bitop @0 @0)
376 (non_lvalue @0)))
377
378 /* x + (x & 1) -> (x + 1) & ~1 */
379 (simplify
380 (plus:c @0 (bit_and:s @0 integer_onep@1))
381 (bit_and (plus @0 @1) (bit_not @1)))
382
383 /* x & ~(x & y) -> x & ~y */
384 /* x | ~(x | y) -> x | ~y */
385 (for bitop (bit_and bit_ior)
386 (simplify
387 (bitop:c @0 (bit_not (bitop:cs @0 @1)))
388 (bitop @0 (bit_not @1))))
389
390 /* (x | y) & ~x -> y & ~x */
391 /* (x & y) | ~x -> y | ~x */
392 (for bitop (bit_and bit_ior)
393 rbitop (bit_ior bit_and)
394 (simplify
395 (bitop:c (rbitop:c @0 @1) (bit_not@2 @0))
396 (bitop @1 @2)))
397
398 /* (x & y) ^ (x | y) -> x ^ y */
399 (simplify
400 (bit_xor:c (bit_and @0 @1) (bit_ior @0 @1))
401 (bit_xor @0 @1))
402
403 /* (x ^ y) ^ (x | y) -> x & y */
404 (simplify
405 (bit_xor:c (bit_xor @0 @1) (bit_ior @0 @1))
406 (bit_and @0 @1))
407
408 /* (x & y) + (x ^ y) -> x | y */
409 /* (x & y) | (x ^ y) -> x | y */
410 /* (x & y) ^ (x ^ y) -> x | y */
411 (for op (plus bit_ior bit_xor)
412 (simplify
413 (op:c (bit_and @0 @1) (bit_xor @0 @1))
414 (bit_ior @0 @1)))
415
416 /* (x & y) + (x | y) -> x + y */
417 (simplify
418 (plus:c (bit_and @0 @1) (bit_ior @0 @1))
419 (plus @0 @1))
420
421 /* (x + y) - (x | y) -> x & y */
422 (simplify
423 (minus (plus @0 @1) (bit_ior @0 @1))
424 (if (!TYPE_OVERFLOW_SANITIZED (type) && !TYPE_OVERFLOW_TRAPS (type)
425 && !TYPE_SATURATING (type))
426 (bit_and @0 @1)))
427
428 /* (x + y) - (x & y) -> x | y */
429 (simplify
430 (minus (plus @0 @1) (bit_and @0 @1))
431 (if (!TYPE_OVERFLOW_SANITIZED (type) && !TYPE_OVERFLOW_TRAPS (type)
432 && !TYPE_SATURATING (type))
433 (bit_ior @0 @1)))
434
435 /* (x | y) - (x ^ y) -> x & y */
436 (simplify
437 (minus (bit_ior @0 @1) (bit_xor @0 @1))
438 (bit_and @0 @1))
439
440 /* (x | y) - (x & y) -> x ^ y */
441 (simplify
442 (minus (bit_ior @0 @1) (bit_and @0 @1))
443 (bit_xor @0 @1))
444
445 /* (x | y) & ~(x & y) -> x ^ y */
446 (simplify
447 (bit_and:c (bit_ior @0 @1) (bit_not (bit_and @0 @1)))
448 (bit_xor @0 @1))
449
450 /* (x | y) & (~x ^ y) -> x & y */
451 (simplify
452 (bit_and:c (bit_ior:c @0 @1) (bit_xor:c @1 (bit_not @0)))
453 (bit_and @0 @1))
454
455 /* ~x & ~y -> ~(x | y)
456 ~x | ~y -> ~(x & y) */
457 (for op (bit_and bit_ior)
458 rop (bit_ior bit_and)
459 (simplify
460 (op (convert1? (bit_not @0)) (convert2? (bit_not @1)))
461 (if (tree_nop_conversion_p (type, TREE_TYPE (@0))
462 && tree_nop_conversion_p (type, TREE_TYPE (@1)))
463 (bit_not (rop (convert @0) (convert @1))))))
464
465 /* If we are XORing or adding two BIT_AND_EXPR's, both of which are and'ing
466 with a constant, and the two constants have no bits in common,
467 we should treat this as a BIT_IOR_EXPR since this may produce more
468 simplifications. */
469 (for op (bit_xor plus)
470 (simplify
471 (op (convert1? (bit_and@4 @0 INTEGER_CST@1))
472 (convert2? (bit_and@5 @2 INTEGER_CST@3)))
473 (if (tree_nop_conversion_p (type, TREE_TYPE (@0))
474 && tree_nop_conversion_p (type, TREE_TYPE (@2))
475 && wi::bit_and (@1, @3) == 0)
476 (bit_ior (convert @4) (convert @5)))))
477
478 /* (X | Y) ^ X -> Y & ~ X*/
479 (simplify
480 (bit_xor:c (convert? (bit_ior:c @0 @1)) (convert? @0))
481 (if (tree_nop_conversion_p (type, TREE_TYPE (@0)))
482 (convert (bit_and @1 (bit_not @0)))))
483
484 /* Convert ~X ^ ~Y to X ^ Y. */
485 (simplify
486 (bit_xor (convert1? (bit_not @0)) (convert2? (bit_not @1)))
487 (if (tree_nop_conversion_p (type, TREE_TYPE (@0))
488 && tree_nop_conversion_p (type, TREE_TYPE (@1)))
489 (bit_xor (convert @0) (convert @1))))
490
491 /* Convert ~X ^ C to X ^ ~C. */
492 (simplify
493 (bit_xor (convert? (bit_not @0)) INTEGER_CST@1)
494 (if (tree_nop_conversion_p (type, TREE_TYPE (@0)))
495 (bit_xor (convert @0) (bit_not @1))))
496
497 /* Fold (X & Y) ^ Y as ~X & Y. */
498 (simplify
499 (bit_xor:c (bit_and:c @0 @1) @1)
500 (bit_and (bit_not @0) @1))
501
502 /* Given a bit-wise operation CODE applied to ARG0 and ARG1, see if both
503 operands are another bit-wise operation with a common input. If so,
504 distribute the bit operations to save an operation and possibly two if
505 constants are involved. For example, convert
506 (A | B) & (A | C) into A | (B & C)
507 Further simplification will occur if B and C are constants. */
508 (for op (bit_and bit_ior)
509 rop (bit_ior bit_and)
510 (simplify
511 (op (convert? (rop:c @0 @1)) (convert? (rop @0 @2)))
512 (if (tree_nop_conversion_p (type, TREE_TYPE (@0)))
513 (rop (convert @0) (op (convert @1) (convert @2))))))
514
515
516 (simplify
517 (abs (abs@1 @0))
518 @1)
519 (simplify
520 (abs (negate @0))
521 (abs @0))
522 (simplify
523 (abs tree_expr_nonnegative_p@0)
524 @0)
525
526 /* A few cases of fold-const.c negate_expr_p predicate. */
527 (match negate_expr_p
528 INTEGER_CST
529 (if ((INTEGRAL_TYPE_P (type)
530 && TYPE_OVERFLOW_WRAPS (type))
531 || (!TYPE_OVERFLOW_SANITIZED (type)
532 && may_negate_without_overflow_p (t)))))
533 (match negate_expr_p
534 FIXED_CST)
535 (match negate_expr_p
536 (negate @0)
537 (if (!TYPE_OVERFLOW_SANITIZED (type))))
538 (match negate_expr_p
539 REAL_CST
540 (if (REAL_VALUE_NEGATIVE (TREE_REAL_CST (t)))))
541 /* VECTOR_CST handling of non-wrapping types would recurse in unsupported
542 ways. */
543 (match negate_expr_p
544 VECTOR_CST
545 (if (FLOAT_TYPE_P (TREE_TYPE (type)) || TYPE_OVERFLOW_WRAPS (type))))
546
547 /* -(A + B) -> (-B) - A. */
548 (simplify
549 (negate (plus:c @0 negate_expr_p@1))
550 (if (!HONOR_SIGN_DEPENDENT_ROUNDING (element_mode (type))
551 && !HONOR_SIGNED_ZEROS (element_mode (type)))
552 (minus (negate @1) @0)))
553
554 /* A - B -> A + (-B) if B is easily negatable. */
555 (simplify
556 (minus @0 negate_expr_p@1)
557 (if (!FIXED_POINT_TYPE_P (type))
558 (plus @0 (negate @1))))
559
560 /* Try to fold (type) X op CST -> (type) (X op ((type-x) CST))
561 when profitable.
562 For bitwise binary operations apply operand conversions to the
563 binary operation result instead of to the operands. This allows
564 to combine successive conversions and bitwise binary operations.
565 We combine the above two cases by using a conditional convert. */
566 (for bitop (bit_and bit_ior bit_xor)
567 (simplify
568 (bitop (convert @0) (convert? @1))
569 (if (((TREE_CODE (@1) == INTEGER_CST
570 && INTEGRAL_TYPE_P (TREE_TYPE (@0))
571 && int_fits_type_p (@1, TREE_TYPE (@0)))
572 || types_match (@0, @1))
573 /* ??? This transform conflicts with fold-const.c doing
574 Convert (T)(x & c) into (T)x & (T)c, if c is an integer
575 constants (if x has signed type, the sign bit cannot be set
576 in c). This folds extension into the BIT_AND_EXPR.
577 Restrict it to GIMPLE to avoid endless recursions. */
578 && (bitop != BIT_AND_EXPR || GIMPLE)
579 && (/* That's a good idea if the conversion widens the operand, thus
580 after hoisting the conversion the operation will be narrower. */
581 TYPE_PRECISION (TREE_TYPE (@0)) < TYPE_PRECISION (type)
582 /* It's also a good idea if the conversion is to a non-integer
583 mode. */
584 || GET_MODE_CLASS (TYPE_MODE (type)) != MODE_INT
585 /* Or if the precision of TO is not the same as the precision
586 of its mode. */
587 || TYPE_PRECISION (type) != GET_MODE_PRECISION (TYPE_MODE (type))))
588 (convert (bitop @0 (convert @1))))))
589
590 (for bitop (bit_and bit_ior)
591 rbitop (bit_ior bit_and)
592 /* (x | y) & x -> x */
593 /* (x & y) | x -> x */
594 (simplify
595 (bitop:c (rbitop:c @0 @1) @0)
596 @0)
597 /* (~x | y) & x -> x & y */
598 /* (~x & y) | x -> x | y */
599 (simplify
600 (bitop:c (rbitop:c (bit_not @0) @1) @0)
601 (bitop @0 @1)))
602
603 /* Simplify (A & B) OP0 (C & B) to (A OP0 C) & B. */
604 (for bitop (bit_and bit_ior bit_xor)
605 (simplify
606 (bitop (bit_and:c @0 @1) (bit_and @2 @1))
607 (bit_and (bitop @0 @2) @1)))
608
609 /* (x | CST1) & CST2 -> (x & CST2) | (CST1 & CST2) */
610 (simplify
611 (bit_and (bit_ior @0 CONSTANT_CLASS_P@1) CONSTANT_CLASS_P@2)
612 (bit_ior (bit_and @0 @2) (bit_and @1 @2)))
613
614 /* Combine successive equal operations with constants. */
615 (for bitop (bit_and bit_ior bit_xor)
616 (simplify
617 (bitop (bitop @0 CONSTANT_CLASS_P@1) CONSTANT_CLASS_P@2)
618 (bitop @0 (bitop @1 @2))))
619
620 /* Try simple folding for X op !X, and X op X with the help
621 of the truth_valued_p and logical_inverted_value predicates. */
622 (match truth_valued_p
623 @0
624 (if (INTEGRAL_TYPE_P (type) && TYPE_PRECISION (type) == 1)))
625 (for op (tcc_comparison truth_and truth_andif truth_or truth_orif truth_xor)
626 (match truth_valued_p
627 (op @0 @1)))
628 (match truth_valued_p
629 (truth_not @0))
630
631 (match (logical_inverted_value @0)
632 (bit_not truth_valued_p@0))
633 (match (logical_inverted_value @0)
634 (eq @0 integer_zerop))
635 (match (logical_inverted_value @0)
636 (ne truth_valued_p@0 integer_truep))
637 (match (logical_inverted_value @0)
638 (bit_xor truth_valued_p@0 integer_truep))
639
640 /* X & !X -> 0. */
641 (simplify
642 (bit_and:c @0 (logical_inverted_value @0))
643 { build_zero_cst (type); })
644 /* X | !X and X ^ !X -> 1, , if X is truth-valued. */
645 (for op (bit_ior bit_xor)
646 (simplify
647 (op:c truth_valued_p@0 (logical_inverted_value @0))
648 { constant_boolean_node (true, type); }))
649 /* X ==/!= !X is false/true. */
650 (for op (eq ne)
651 (simplify
652 (op:c truth_valued_p@0 (logical_inverted_value @0))
653 { constant_boolean_node (op == NE_EXPR ? true : false, type); }))
654
655 /* If arg1 and arg2 are booleans (or any single bit type)
656 then try to simplify:
657
658 (~X & Y) -> X < Y
659 (X & ~Y) -> Y < X
660 (~X | Y) -> X <= Y
661 (X | ~Y) -> Y <= X
662
663 But only do this if our result feeds into a comparison as
664 this transformation is not always a win, particularly on
665 targets with and-not instructions.
666 -> simplify_bitwise_binary_boolean */
667 (simplify
668 (ne (bit_and:c (bit_not @0) @1) integer_zerop)
669 (if (INTEGRAL_TYPE_P (TREE_TYPE (@1))
670 && TYPE_PRECISION (TREE_TYPE (@1)) == 1)
671 (lt @0 @1)))
672 (simplify
673 (ne (bit_ior:c (bit_not @0) @1) integer_zerop)
674 (if (INTEGRAL_TYPE_P (TREE_TYPE (@1))
675 && TYPE_PRECISION (TREE_TYPE (@1)) == 1)
676 (le @0 @1)))
677
678 /* ~~x -> x */
679 (simplify
680 (bit_not (bit_not @0))
681 @0)
682
683 /* Convert ~ (-A) to A - 1. */
684 (simplify
685 (bit_not (convert? (negate @0)))
686 (if (tree_nop_conversion_p (type, TREE_TYPE (@0)))
687 (convert (minus @0 { build_each_one_cst (TREE_TYPE (@0)); }))))
688
689 /* Convert ~ (A - 1) or ~ (A + -1) to -A. */
690 (simplify
691 (bit_not (convert? (minus @0 integer_each_onep)))
692 (if (tree_nop_conversion_p (type, TREE_TYPE (@0)))
693 (convert (negate @0))))
694 (simplify
695 (bit_not (convert? (plus @0 integer_all_onesp)))
696 (if (tree_nop_conversion_p (type, TREE_TYPE (@0)))
697 (convert (negate @0))))
698
699 /* Part of convert ~(X ^ Y) to ~X ^ Y or X ^ ~Y if ~X or ~Y simplify. */
700 (simplify
701 (bit_not (convert? (bit_xor @0 INTEGER_CST@1)))
702 (if (tree_nop_conversion_p (type, TREE_TYPE (@0)))
703 (convert (bit_xor @0 (bit_not @1)))))
704 (simplify
705 (bit_not (convert? (bit_xor:c (bit_not @0) @1)))
706 (if (tree_nop_conversion_p (type, TREE_TYPE (@0)))
707 (convert (bit_xor @0 @1))))
708
709 /* (x & ~m) | (y & m) -> ((x ^ y) & m) ^ x */
710 (simplify
711 (bit_ior:c (bit_and:cs @0 (bit_not @2)) (bit_and:cs @1 @2))
712 (bit_xor (bit_and (bit_xor @0 @1) @2) @0))
713
714 /* Fold A - (A & B) into ~B & A. */
715 (simplify
716 (minus (convert? @0) (convert?:s (bit_and:cs @0 @1)))
717 (if (tree_nop_conversion_p (type, TREE_TYPE (@0))
718 && tree_nop_conversion_p (type, TREE_TYPE (@1)))
719 (convert (bit_and (bit_not @1) @0))))
720
721
722
723 /* ((X inner_op C0) outer_op C1)
724 With X being a tree where value_range has reasoned certain bits to always be
725 zero throughout its computed value range,
726 inner_op = {|,^}, outer_op = {|,^} and inner_op != outer_op
727 where zero_mask has 1's for all bits that are sure to be 0 in
728 and 0's otherwise.
729 if (inner_op == '^') C0 &= ~C1;
730 if ((C0 & ~zero_mask) == 0) then emit (X outer_op (C0 outer_op C1)
731 if ((C1 & ~zero_mask) == 0) then emit (X inner_op (C0 outer_op C1)
732 */
733 (for inner_op (bit_ior bit_xor)
734 outer_op (bit_xor bit_ior)
735 (simplify
736 (outer_op
737 (inner_op:s @2 INTEGER_CST@0) INTEGER_CST@1)
738 (with
739 {
740 bool fail = false;
741 wide_int zero_mask_not;
742 wide_int C0;
743 wide_int cst_emit;
744
745 if (TREE_CODE (@2) == SSA_NAME)
746 zero_mask_not = get_nonzero_bits (@2);
747 else
748 fail = true;
749
750 if (inner_op == BIT_XOR_EXPR)
751 {
752 C0 = wi::bit_and_not (@0, @1);
753 cst_emit = wi::bit_or (C0, @1);
754 }
755 else
756 {
757 C0 = @0;
758 cst_emit = wi::bit_xor (@0, @1);
759 }
760 }
761 (if (!fail && wi::bit_and (C0, zero_mask_not) == 0)
762 (outer_op @2 { wide_int_to_tree (type, cst_emit); })
763 (if (!fail && wi::bit_and (@1, zero_mask_not) == 0)
764 (inner_op @2 { wide_int_to_tree (type, cst_emit); }))))))
765
766 /* Associate (p +p off1) +p off2 as (p +p (off1 + off2)). */
767 (simplify
768 (pointer_plus (pointer_plus:s @0 @1) @3)
769 (pointer_plus @0 (plus @1 @3)))
770
771 /* Pattern match
772 tem1 = (long) ptr1;
773 tem2 = (long) ptr2;
774 tem3 = tem2 - tem1;
775 tem4 = (unsigned long) tem3;
776 tem5 = ptr1 + tem4;
777 and produce
778 tem5 = ptr2; */
779 (simplify
780 (pointer_plus @0 (convert?@2 (minus@3 (convert @1) (convert @0))))
781 /* Conditionally look through a sign-changing conversion. */
782 (if (TYPE_PRECISION (TREE_TYPE (@2)) == TYPE_PRECISION (TREE_TYPE (@3))
783 && ((GIMPLE && useless_type_conversion_p (type, TREE_TYPE (@1)))
784 || (GENERIC && type == TREE_TYPE (@1))))
785 @1))
786
787 /* Pattern match
788 tem = (sizetype) ptr;
789 tem = tem & algn;
790 tem = -tem;
791 ... = ptr p+ tem;
792 and produce the simpler and easier to analyze with respect to alignment
793 ... = ptr & ~algn; */
794 (simplify
795 (pointer_plus @0 (negate (bit_and (convert @0) INTEGER_CST@1)))
796 (with { tree algn = wide_int_to_tree (TREE_TYPE (@0), wi::bit_not (@1)); }
797 (bit_and @0 { algn; })))
798
799 /* Try folding difference of addresses. */
800 (simplify
801 (minus (convert ADDR_EXPR@0) (convert @1))
802 (if (tree_nop_conversion_p (type, TREE_TYPE (@0)))
803 (with { HOST_WIDE_INT diff; }
804 (if (ptr_difference_const (@0, @1, &diff))
805 { build_int_cst_type (type, diff); }))))
806 (simplify
807 (minus (convert @0) (convert ADDR_EXPR@1))
808 (if (tree_nop_conversion_p (type, TREE_TYPE (@0)))
809 (with { HOST_WIDE_INT diff; }
810 (if (ptr_difference_const (@0, @1, &diff))
811 { build_int_cst_type (type, diff); }))))
812
813 /* If arg0 is derived from the address of an object or function, we may
814 be able to fold this expression using the object or function's
815 alignment. */
816 (simplify
817 (bit_and (convert? @0) INTEGER_CST@1)
818 (if (POINTER_TYPE_P (TREE_TYPE (@0))
819 && tree_nop_conversion_p (type, TREE_TYPE (@0)))
820 (with
821 {
822 unsigned int align;
823 unsigned HOST_WIDE_INT bitpos;
824 get_pointer_alignment_1 (@0, &align, &bitpos);
825 }
826 (if (wi::ltu_p (@1, align / BITS_PER_UNIT))
827 { wide_int_to_tree (type, wi::bit_and (@1, bitpos / BITS_PER_UNIT)); }))))
828
829
830 /* We can't reassociate at all for saturating types. */
831 (if (!TYPE_SATURATING (type))
832
833 /* Contract negates. */
834 /* A + (-B) -> A - B */
835 (simplify
836 (plus:c (convert1? @0) (convert2? (negate @1)))
837 /* Apply STRIP_NOPS on @0 and the negate. */
838 (if (tree_nop_conversion_p (type, TREE_TYPE (@0))
839 && tree_nop_conversion_p (type, TREE_TYPE (@1))
840 && !TYPE_OVERFLOW_SANITIZED (type))
841 (minus (convert @0) (convert @1))))
842 /* A - (-B) -> A + B */
843 (simplify
844 (minus (convert1? @0) (convert2? (negate @1)))
845 (if (tree_nop_conversion_p (type, TREE_TYPE (@0))
846 && tree_nop_conversion_p (type, TREE_TYPE (@1))
847 && !TYPE_OVERFLOW_SANITIZED (type))
848 (plus (convert @0) (convert @1))))
849 /* -(-A) -> A */
850 (simplify
851 (negate (convert? (negate @1)))
852 (if (tree_nop_conversion_p (type, TREE_TYPE (@1))
853 && !TYPE_OVERFLOW_SANITIZED (type))
854 (convert @1)))
855
856 /* We can't reassociate floating-point unless -fassociative-math
857 or fixed-point plus or minus because of saturation to +-Inf. */
858 (if ((!FLOAT_TYPE_P (type) || flag_associative_math)
859 && !FIXED_POINT_TYPE_P (type))
860
861 /* Match patterns that allow contracting a plus-minus pair
862 irrespective of overflow issues. */
863 /* (A +- B) - A -> +- B */
864 /* (A +- B) -+ B -> A */
865 /* A - (A +- B) -> -+ B */
866 /* A +- (B -+ A) -> +- B */
867 (simplify
868 (minus (plus:c @0 @1) @0)
869 @1)
870 (simplify
871 (minus (minus @0 @1) @0)
872 (negate @1))
873 (simplify
874 (plus:c (minus @0 @1) @1)
875 @0)
876 (simplify
877 (minus @0 (plus:c @0 @1))
878 (negate @1))
879 (simplify
880 (minus @0 (minus @0 @1))
881 @1)
882
883 /* (A +- CST) +- CST -> A + CST */
884 (for outer_op (plus minus)
885 (for inner_op (plus minus)
886 (simplify
887 (outer_op (inner_op @0 CONSTANT_CLASS_P@1) CONSTANT_CLASS_P@2)
888 /* If the constant operation overflows we cannot do the transform
889 as we would introduce undefined overflow, for example
890 with (a - 1) + INT_MIN. */
891 (with { tree cst = fold_binary (outer_op == inner_op
892 ? PLUS_EXPR : MINUS_EXPR, type, @1, @2); }
893 (if (cst && !TREE_OVERFLOW (cst))
894 (inner_op @0 { cst; } ))))))
895
896 /* (CST - A) +- CST -> CST - A */
897 (for outer_op (plus minus)
898 (simplify
899 (outer_op (minus CONSTANT_CLASS_P@1 @0) CONSTANT_CLASS_P@2)
900 (with { tree cst = fold_binary (outer_op, type, @1, @2); }
901 (if (cst && !TREE_OVERFLOW (cst))
902 (minus { cst; } @0)))))
903
904 /* ~A + A -> -1 */
905 (simplify
906 (plus:c (bit_not @0) @0)
907 (if (!TYPE_OVERFLOW_TRAPS (type))
908 { build_all_ones_cst (type); }))
909
910 /* ~A + 1 -> -A */
911 (simplify
912 (plus (convert? (bit_not @0)) integer_each_onep)
913 (if (tree_nop_conversion_p (type, TREE_TYPE (@0)))
914 (negate (convert @0))))
915
916 /* -A - 1 -> ~A */
917 (simplify
918 (minus (convert? (negate @0)) integer_each_onep)
919 (if (!TYPE_OVERFLOW_TRAPS (type)
920 && tree_nop_conversion_p (type, TREE_TYPE (@0)))
921 (bit_not (convert @0))))
922
923 /* -1 - A -> ~A */
924 (simplify
925 (minus integer_all_onesp @0)
926 (bit_not @0))
927
928 /* (T)(P + A) - (T)P -> (T) A */
929 (for add (plus pointer_plus)
930 (simplify
931 (minus (convert (add @0 @1))
932 (convert @0))
933 (if (element_precision (type) <= element_precision (TREE_TYPE (@1))
934 /* For integer types, if A has a smaller type
935 than T the result depends on the possible
936 overflow in P + A.
937 E.g. T=size_t, A=(unsigned)429497295, P>0.
938 However, if an overflow in P + A would cause
939 undefined behavior, we can assume that there
940 is no overflow. */
941 || (INTEGRAL_TYPE_P (TREE_TYPE (@0))
942 && TYPE_OVERFLOW_UNDEFINED (TREE_TYPE (@0)))
943 /* For pointer types, if the conversion of A to the
944 final type requires a sign- or zero-extension,
945 then we have to punt - it is not defined which
946 one is correct. */
947 || (POINTER_TYPE_P (TREE_TYPE (@0))
948 && TREE_CODE (@1) == INTEGER_CST
949 && tree_int_cst_sign_bit (@1) == 0))
950 (convert @1))))))
951
952
953 /* Simplifications of MIN_EXPR and MAX_EXPR. */
954
955 (for minmax (min max)
956 (simplify
957 (minmax @0 @0)
958 @0))
959 (simplify
960 (min @0 @1)
961 (if (INTEGRAL_TYPE_P (type)
962 && TYPE_MIN_VALUE (type)
963 && operand_equal_p (@1, TYPE_MIN_VALUE (type), OEP_ONLY_CONST))
964 @1))
965 (simplify
966 (max @0 @1)
967 (if (INTEGRAL_TYPE_P (type)
968 && TYPE_MAX_VALUE (type)
969 && operand_equal_p (@1, TYPE_MAX_VALUE (type), OEP_ONLY_CONST))
970 @1))
971
972
973 /* Simplifications of shift and rotates. */
974
975 (for rotate (lrotate rrotate)
976 (simplify
977 (rotate integer_all_onesp@0 @1)
978 @0))
979
980 /* Optimize -1 >> x for arithmetic right shifts. */
981 (simplify
982 (rshift integer_all_onesp@0 @1)
983 (if (!TYPE_UNSIGNED (type)
984 && tree_expr_nonnegative_p (@1))
985 @0))
986
987 /* Optimize (x >> c) << c into x & (-1<<c). */
988 (simplify
989 (lshift (rshift @0 INTEGER_CST@1) @1)
990 (if (wi::ltu_p (@1, element_precision (type)))
991 (bit_and @0 (lshift { build_minus_one_cst (type); } @1))))
992
993 /* Optimize (x << c) >> c into x & ((unsigned)-1 >> c) for unsigned
994 types. */
995 (simplify
996 (rshift (lshift @0 INTEGER_CST@1) @1)
997 (if (TYPE_UNSIGNED (type)
998 && (wi::ltu_p (@1, element_precision (type))))
999 (bit_and @0 (rshift { build_minus_one_cst (type); } @1))))
1000
1001 (for shiftrotate (lrotate rrotate lshift rshift)
1002 (simplify
1003 (shiftrotate @0 integer_zerop)
1004 (non_lvalue @0))
1005 (simplify
1006 (shiftrotate integer_zerop@0 @1)
1007 @0)
1008 /* Prefer vector1 << scalar to vector1 << vector2
1009 if vector2 is uniform. */
1010 (for vec (VECTOR_CST CONSTRUCTOR)
1011 (simplify
1012 (shiftrotate @0 vec@1)
1013 (with { tree tem = uniform_vector_p (@1); }
1014 (if (tem)
1015 (shiftrotate @0 { tem; }))))))
1016
1017 /* Rewrite an LROTATE_EXPR by a constant into an
1018 RROTATE_EXPR by a new constant. */
1019 (simplify
1020 (lrotate @0 INTEGER_CST@1)
1021 (rrotate @0 { fold_binary (MINUS_EXPR, TREE_TYPE (@1),
1022 build_int_cst (TREE_TYPE (@1),
1023 element_precision (type)), @1); }))
1024
1025 /* Turn (a OP c1) OP c2 into a OP (c1+c2). */
1026 (for op (lrotate rrotate rshift lshift)
1027 (simplify
1028 (op (op @0 INTEGER_CST@1) INTEGER_CST@2)
1029 (with { unsigned int prec = element_precision (type); }
1030 (if (wi::ge_p (@1, 0, TYPE_SIGN (TREE_TYPE (@1)))
1031 && wi::lt_p (@1, prec, TYPE_SIGN (TREE_TYPE (@1)))
1032 && wi::ge_p (@2, 0, TYPE_SIGN (TREE_TYPE (@2)))
1033 && wi::lt_p (@2, prec, TYPE_SIGN (TREE_TYPE (@2))))
1034 (with { unsigned int low = wi::add (@1, @2).to_uhwi (); }
1035 /* Deal with a OP (c1 + c2) being undefined but (a OP c1) OP c2
1036 being well defined. */
1037 (if (low >= prec)
1038 (if (op == LROTATE_EXPR || op == RROTATE_EXPR)
1039 (op @0 { build_int_cst (TREE_TYPE (@1), low % prec); })
1040 (if (TYPE_UNSIGNED (type) || op == LSHIFT_EXPR)
1041 { build_zero_cst (type); }
1042 (op @0 { build_int_cst (TREE_TYPE (@1), prec - 1); })))
1043 (op @0 { build_int_cst (TREE_TYPE (@1), low); })))))))
1044
1045
1046 /* ((1 << A) & 1) != 0 -> A == 0
1047 ((1 << A) & 1) == 0 -> A != 0 */
1048 (for cmp (ne eq)
1049 icmp (eq ne)
1050 (simplify
1051 (cmp (bit_and (lshift integer_onep @0) integer_onep) integer_zerop)
1052 (icmp @0 { build_zero_cst (TREE_TYPE (@0)); })))
1053
1054 /* (CST1 << A) == CST2 -> A == ctz (CST2) - ctz (CST1)
1055 (CST1 << A) != CST2 -> A != ctz (CST2) - ctz (CST1)
1056 if CST2 != 0. */
1057 (for cmp (ne eq)
1058 (simplify
1059 (cmp (lshift INTEGER_CST@0 @1) INTEGER_CST@2)
1060 (with { int cand = wi::ctz (@2) - wi::ctz (@0); }
1061 (if (cand < 0
1062 || (!integer_zerop (@2)
1063 && wi::ne_p (wi::lshift (@0, cand), @2)))
1064 { constant_boolean_node (cmp == NE_EXPR, type); }
1065 (if (!integer_zerop (@2)
1066 && wi::eq_p (wi::lshift (@0, cand), @2))
1067 (cmp @1 { build_int_cst (TREE_TYPE (@1), cand); }))))))
1068
1069 /* Fold (X << C1) & C2 into (X << C1) & (C2 | ((1 << C1) - 1))
1070 (X >> C1) & C2 into (X >> C1) & (C2 | ~((type) -1 >> C1))
1071 if the new mask might be further optimized. */
1072 (for shift (lshift rshift)
1073 (simplify
1074 (bit_and (convert?:s@4 (shift:s@5 (convert1?@3 @0) INTEGER_CST@1))
1075 INTEGER_CST@2)
1076 (if (tree_nop_conversion_p (TREE_TYPE (@4), TREE_TYPE (@5))
1077 && TYPE_PRECISION (type) <= HOST_BITS_PER_WIDE_INT
1078 && tree_fits_uhwi_p (@1)
1079 && tree_to_uhwi (@1) > 0
1080 && tree_to_uhwi (@1) < TYPE_PRECISION (type))
1081 (with
1082 {
1083 unsigned int shiftc = tree_to_uhwi (@1);
1084 unsigned HOST_WIDE_INT mask = TREE_INT_CST_LOW (@2);
1085 unsigned HOST_WIDE_INT newmask, zerobits = 0;
1086 tree shift_type = TREE_TYPE (@3);
1087 unsigned int prec;
1088
1089 if (shift == LSHIFT_EXPR)
1090 zerobits = ((((unsigned HOST_WIDE_INT) 1) << shiftc) - 1);
1091 else if (shift == RSHIFT_EXPR
1092 && (TYPE_PRECISION (shift_type)
1093 == GET_MODE_PRECISION (TYPE_MODE (shift_type))))
1094 {
1095 prec = TYPE_PRECISION (TREE_TYPE (@3));
1096 tree arg00 = @0;
1097 /* See if more bits can be proven as zero because of
1098 zero extension. */
1099 if (@3 != @0
1100 && TYPE_UNSIGNED (TREE_TYPE (@0)))
1101 {
1102 tree inner_type = TREE_TYPE (@0);
1103 if ((TYPE_PRECISION (inner_type)
1104 == GET_MODE_PRECISION (TYPE_MODE (inner_type)))
1105 && TYPE_PRECISION (inner_type) < prec)
1106 {
1107 prec = TYPE_PRECISION (inner_type);
1108 /* See if we can shorten the right shift. */
1109 if (shiftc < prec)
1110 shift_type = inner_type;
1111 /* Otherwise X >> C1 is all zeros, so we'll optimize
1112 it into (X, 0) later on by making sure zerobits
1113 is all ones. */
1114 }
1115 }
1116 zerobits = ~(unsigned HOST_WIDE_INT) 0;
1117 if (shiftc < prec)
1118 {
1119 zerobits >>= HOST_BITS_PER_WIDE_INT - shiftc;
1120 zerobits <<= prec - shiftc;
1121 }
1122 /* For arithmetic shift if sign bit could be set, zerobits
1123 can contain actually sign bits, so no transformation is
1124 possible, unless MASK masks them all away. In that
1125 case the shift needs to be converted into logical shift. */
1126 if (!TYPE_UNSIGNED (TREE_TYPE (@3))
1127 && prec == TYPE_PRECISION (TREE_TYPE (@3)))
1128 {
1129 if ((mask & zerobits) == 0)
1130 shift_type = unsigned_type_for (TREE_TYPE (@3));
1131 else
1132 zerobits = 0;
1133 }
1134 }
1135 }
1136 /* ((X << 16) & 0xff00) is (X, 0). */
1137 (if ((mask & zerobits) == mask)
1138 { build_int_cst (type, 0); }
1139 (with { newmask = mask | zerobits; }
1140 (if (newmask != mask && (newmask & (newmask + 1)) == 0)
1141 (with
1142 {
1143 /* Only do the transformation if NEWMASK is some integer
1144 mode's mask. */
1145 for (prec = BITS_PER_UNIT;
1146 prec < HOST_BITS_PER_WIDE_INT; prec <<= 1)
1147 if (newmask == (((unsigned HOST_WIDE_INT) 1) << prec) - 1)
1148 break;
1149 }
1150 (if (prec < HOST_BITS_PER_WIDE_INT
1151 || newmask == ~(unsigned HOST_WIDE_INT) 0)
1152 (with
1153 { tree newmaskt = build_int_cst_type (TREE_TYPE (@2), newmask); }
1154 (if (!tree_int_cst_equal (newmaskt, @2))
1155 (if (shift_type != TREE_TYPE (@3))
1156 (bit_and (convert (shift:shift_type (convert @3) @1)) { newmaskt; })
1157 (bit_and @4 { newmaskt; })))))))))))))
1158
1159 /* Fold (X {&,^,|} C2) << C1 into (X << C1) {&,^,|} (C2 << C1)
1160 (X {&,^,|} C2) >> C1 into (X >> C1) & (C2 >> C1). */
1161 (for shift (lshift rshift)
1162 (for bit_op (bit_and bit_xor bit_ior)
1163 (simplify
1164 (shift (convert?:s (bit_op:s @0 INTEGER_CST@2)) INTEGER_CST@1)
1165 (if (tree_nop_conversion_p (type, TREE_TYPE (@0)))
1166 (with { tree mask = int_const_binop (shift, fold_convert (type, @2), @1); }
1167 (bit_op (shift (convert @0) @1) { mask; }))))))
1168
1169
1170 /* Simplifications of conversions. */
1171
1172 /* Basic strip-useless-type-conversions / strip_nops. */
1173 (for cvt (convert view_convert float fix_trunc)
1174 (simplify
1175 (cvt @0)
1176 (if ((GIMPLE && useless_type_conversion_p (type, TREE_TYPE (@0)))
1177 || (GENERIC && type == TREE_TYPE (@0)))
1178 @0)))
1179
1180 /* Contract view-conversions. */
1181 (simplify
1182 (view_convert (view_convert @0))
1183 (view_convert @0))
1184
1185 /* For integral conversions with the same precision or pointer
1186 conversions use a NOP_EXPR instead. */
1187 (simplify
1188 (view_convert @0)
1189 (if ((INTEGRAL_TYPE_P (type) || POINTER_TYPE_P (type))
1190 && (INTEGRAL_TYPE_P (TREE_TYPE (@0)) || POINTER_TYPE_P (TREE_TYPE (@0)))
1191 && TYPE_PRECISION (type) == TYPE_PRECISION (TREE_TYPE (@0)))
1192 (convert @0)))
1193
1194 /* Strip inner integral conversions that do not change precision or size. */
1195 (simplify
1196 (view_convert (convert@0 @1))
1197 (if ((INTEGRAL_TYPE_P (TREE_TYPE (@0)) || POINTER_TYPE_P (TREE_TYPE (@0)))
1198 && (INTEGRAL_TYPE_P (TREE_TYPE (@1)) || POINTER_TYPE_P (TREE_TYPE (@1)))
1199 && (TYPE_PRECISION (TREE_TYPE (@0)) == TYPE_PRECISION (TREE_TYPE (@1)))
1200 && (TYPE_SIZE (TREE_TYPE (@0)) == TYPE_SIZE (TREE_TYPE (@1))))
1201 (view_convert @1)))
1202
1203 /* Re-association barriers around constants and other re-association
1204 barriers can be removed. */
1205 (simplify
1206 (paren CONSTANT_CLASS_P@0)
1207 @0)
1208 (simplify
1209 (paren (paren@1 @0))
1210 @1)
1211
1212 /* Handle cases of two conversions in a row. */
1213 (for ocvt (convert float fix_trunc)
1214 (for icvt (convert float)
1215 (simplify
1216 (ocvt (icvt@1 @0))
1217 (with
1218 {
1219 tree inside_type = TREE_TYPE (@0);
1220 tree inter_type = TREE_TYPE (@1);
1221 int inside_int = INTEGRAL_TYPE_P (inside_type);
1222 int inside_ptr = POINTER_TYPE_P (inside_type);
1223 int inside_float = FLOAT_TYPE_P (inside_type);
1224 int inside_vec = VECTOR_TYPE_P (inside_type);
1225 unsigned int inside_prec = TYPE_PRECISION (inside_type);
1226 int inside_unsignedp = TYPE_UNSIGNED (inside_type);
1227 int inter_int = INTEGRAL_TYPE_P (inter_type);
1228 int inter_ptr = POINTER_TYPE_P (inter_type);
1229 int inter_float = FLOAT_TYPE_P (inter_type);
1230 int inter_vec = VECTOR_TYPE_P (inter_type);
1231 unsigned int inter_prec = TYPE_PRECISION (inter_type);
1232 int inter_unsignedp = TYPE_UNSIGNED (inter_type);
1233 int final_int = INTEGRAL_TYPE_P (type);
1234 int final_ptr = POINTER_TYPE_P (type);
1235 int final_float = FLOAT_TYPE_P (type);
1236 int final_vec = VECTOR_TYPE_P (type);
1237 unsigned int final_prec = TYPE_PRECISION (type);
1238 int final_unsignedp = TYPE_UNSIGNED (type);
1239 }
1240 (switch
1241 /* In addition to the cases of two conversions in a row
1242 handled below, if we are converting something to its own
1243 type via an object of identical or wider precision, neither
1244 conversion is needed. */
1245 (if (((GIMPLE && useless_type_conversion_p (type, inside_type))
1246 || (GENERIC
1247 && TYPE_MAIN_VARIANT (type) == TYPE_MAIN_VARIANT (inside_type)))
1248 && (((inter_int || inter_ptr) && final_int)
1249 || (inter_float && final_float))
1250 && inter_prec >= final_prec)
1251 (ocvt @0))
1252
1253 /* Likewise, if the intermediate and initial types are either both
1254 float or both integer, we don't need the middle conversion if the
1255 former is wider than the latter and doesn't change the signedness
1256 (for integers). Avoid this if the final type is a pointer since
1257 then we sometimes need the middle conversion. Likewise if the
1258 final type has a precision not equal to the size of its mode. */
1259 (if (((inter_int && inside_int) || (inter_float && inside_float))
1260 && (final_int || final_float)
1261 && inter_prec >= inside_prec
1262 && (inter_float || inter_unsignedp == inside_unsignedp)
1263 && ! (final_prec != GET_MODE_PRECISION (TYPE_MODE (type))
1264 && TYPE_MODE (type) == TYPE_MODE (inter_type)))
1265 (ocvt @0))
1266
1267 /* If we have a sign-extension of a zero-extended value, we can
1268 replace that by a single zero-extension. Likewise if the
1269 final conversion does not change precision we can drop the
1270 intermediate conversion. */
1271 (if (inside_int && inter_int && final_int
1272 && ((inside_prec < inter_prec && inter_prec < final_prec
1273 && inside_unsignedp && !inter_unsignedp)
1274 || final_prec == inter_prec))
1275 (ocvt @0))
1276
1277 /* Two conversions in a row are not needed unless:
1278 - some conversion is floating-point (overstrict for now), or
1279 - some conversion is a vector (overstrict for now), or
1280 - the intermediate type is narrower than both initial and
1281 final, or
1282 - the intermediate type and innermost type differ in signedness,
1283 and the outermost type is wider than the intermediate, or
1284 - the initial type is a pointer type and the precisions of the
1285 intermediate and final types differ, or
1286 - the final type is a pointer type and the precisions of the
1287 initial and intermediate types differ. */
1288 (if (! inside_float && ! inter_float && ! final_float
1289 && ! inside_vec && ! inter_vec && ! final_vec
1290 && (inter_prec >= inside_prec || inter_prec >= final_prec)
1291 && ! (inside_int && inter_int
1292 && inter_unsignedp != inside_unsignedp
1293 && inter_prec < final_prec)
1294 && ((inter_unsignedp && inter_prec > inside_prec)
1295 == (final_unsignedp && final_prec > inter_prec))
1296 && ! (inside_ptr && inter_prec != final_prec)
1297 && ! (final_ptr && inside_prec != inter_prec)
1298 && ! (final_prec != GET_MODE_PRECISION (TYPE_MODE (type))
1299 && TYPE_MODE (type) == TYPE_MODE (inter_type)))
1300 (ocvt @0))
1301
1302 /* A truncation to an unsigned type (a zero-extension) should be
1303 canonicalized as bitwise and of a mask. */
1304 (if (final_int && inter_int && inside_int
1305 && final_prec == inside_prec
1306 && final_prec > inter_prec
1307 && inter_unsignedp)
1308 (convert (bit_and @0 { wide_int_to_tree
1309 (inside_type,
1310 wi::mask (inter_prec, false,
1311 TYPE_PRECISION (inside_type))); })))
1312
1313 /* If we are converting an integer to a floating-point that can
1314 represent it exactly and back to an integer, we can skip the
1315 floating-point conversion. */
1316 (if (GIMPLE /* PR66211 */
1317 && inside_int && inter_float && final_int &&
1318 (unsigned) significand_size (TYPE_MODE (inter_type))
1319 >= inside_prec - !inside_unsignedp)
1320 (convert @0)))))))
1321
1322 /* If we have a narrowing conversion to an integral type that is fed by a
1323 BIT_AND_EXPR, we might be able to remove the BIT_AND_EXPR if it merely
1324 masks off bits outside the final type (and nothing else). */
1325 (simplify
1326 (convert (bit_and @0 INTEGER_CST@1))
1327 (if (INTEGRAL_TYPE_P (type)
1328 && INTEGRAL_TYPE_P (TREE_TYPE (@0))
1329 && TYPE_PRECISION (type) <= TYPE_PRECISION (TREE_TYPE (@0))
1330 && operand_equal_p (@1, build_low_bits_mask (TREE_TYPE (@1),
1331 TYPE_PRECISION (type)), 0))
1332 (convert @0)))
1333
1334
1335 /* (X /[ex] A) * A -> X. */
1336 (simplify
1337 (mult (convert? (exact_div @0 @1)) @1)
1338 /* Look through a sign-changing conversion. */
1339 (convert @0))
1340
1341 /* Canonicalization of binary operations. */
1342
1343 /* Convert X + -C into X - C. */
1344 (simplify
1345 (plus @0 REAL_CST@1)
1346 (if (REAL_VALUE_NEGATIVE (TREE_REAL_CST (@1)))
1347 (with { tree tem = fold_unary (NEGATE_EXPR, type, @1); }
1348 (if (!TREE_OVERFLOW (tem) || !flag_trapping_math)
1349 (minus @0 { tem; })))))
1350
1351 /* Convert x+x into x*2.0. */
1352 (simplify
1353 (plus @0 @0)
1354 (if (SCALAR_FLOAT_TYPE_P (type))
1355 (mult @0 { build_real (type, dconst2); })))
1356
1357 (simplify
1358 (minus integer_zerop @1)
1359 (negate @1))
1360
1361 /* (ARG0 - ARG1) is the same as (-ARG1 + ARG0). So check whether
1362 ARG0 is zero and X + ARG0 reduces to X, since that would mean
1363 (-ARG1 + ARG0) reduces to -ARG1. */
1364 (simplify
1365 (minus real_zerop@0 @1)
1366 (if (fold_real_zero_addition_p (type, @0, 0))
1367 (negate @1)))
1368
1369 /* Transform x * -1 into -x. */
1370 (simplify
1371 (mult @0 integer_minus_onep)
1372 (negate @0))
1373
1374 /* COMPLEX_EXPR and REALPART/IMAGPART_EXPR cancellations. */
1375 (simplify
1376 (complex (realpart @0) (imagpart @0))
1377 @0)
1378 (simplify
1379 (realpart (complex @0 @1))
1380 @0)
1381 (simplify
1382 (imagpart (complex @0 @1))
1383 @1)
1384
1385 /* Sometimes we only care about half of a complex expression. */
1386 (simplify
1387 (realpart (convert?:s (conj:s @0)))
1388 (convert (realpart @0)))
1389 (simplify
1390 (imagpart (convert?:s (conj:s @0)))
1391 (convert (negate (imagpart @0))))
1392 (for part (realpart imagpart)
1393 (for op (plus minus)
1394 (simplify
1395 (part (convert?:s@2 (op:s @0 @1)))
1396 (convert (op (part @0) (part @1))))))
1397 (simplify
1398 (realpart (convert?:s (CEXPI:s @0)))
1399 (convert (COS @0)))
1400 (simplify
1401 (imagpart (convert?:s (CEXPI:s @0)))
1402 (convert (SIN @0)))
1403
1404 /* conj(conj(x)) -> x */
1405 (simplify
1406 (conj (convert? (conj @0)))
1407 (if (tree_nop_conversion_p (TREE_TYPE (@0), type))
1408 (convert @0)))
1409
1410 /* conj({x,y}) -> {x,-y} */
1411 (simplify
1412 (conj (convert?:s (complex:s @0 @1)))
1413 (with { tree itype = TREE_TYPE (type); }
1414 (complex (convert:itype @0) (negate (convert:itype @1)))))
1415
1416 /* BSWAP simplifications, transforms checked by gcc.dg/builtin-bswap-8.c. */
1417 (for bswap (BUILT_IN_BSWAP16 BUILT_IN_BSWAP32 BUILT_IN_BSWAP64)
1418 (simplify
1419 (bswap (bswap @0))
1420 @0)
1421 (simplify
1422 (bswap (bit_not (bswap @0)))
1423 (bit_not @0))
1424 (for bitop (bit_xor bit_ior bit_and)
1425 (simplify
1426 (bswap (bitop:c (bswap @0) @1))
1427 (bitop @0 (bswap @1)))))
1428
1429
1430 /* Combine COND_EXPRs and VEC_COND_EXPRs. */
1431
1432 /* Simplify constant conditions.
1433 Only optimize constant conditions when the selected branch
1434 has the same type as the COND_EXPR. This avoids optimizing
1435 away "c ? x : throw", where the throw has a void type.
1436 Note that we cannot throw away the fold-const.c variant nor
1437 this one as we depend on doing this transform before possibly
1438 A ? B : B -> B triggers and the fold-const.c one can optimize
1439 0 ? A : B to B even if A has side-effects. Something
1440 genmatch cannot handle. */
1441 (simplify
1442 (cond INTEGER_CST@0 @1 @2)
1443 (if (integer_zerop (@0))
1444 (if (!VOID_TYPE_P (TREE_TYPE (@2)) || VOID_TYPE_P (type))
1445 @2)
1446 (if (!VOID_TYPE_P (TREE_TYPE (@1)) || VOID_TYPE_P (type))
1447 @1)))
1448 (simplify
1449 (vec_cond VECTOR_CST@0 @1 @2)
1450 (if (integer_all_onesp (@0))
1451 @1
1452 (if (integer_zerop (@0))
1453 @2)))
1454
1455 (for cnd (cond vec_cond)
1456 /* A ? B : (A ? X : C) -> A ? B : C. */
1457 (simplify
1458 (cnd @0 (cnd @0 @1 @2) @3)
1459 (cnd @0 @1 @3))
1460 (simplify
1461 (cnd @0 @1 (cnd @0 @2 @3))
1462 (cnd @0 @1 @3))
1463
1464 /* A ? B : B -> B. */
1465 (simplify
1466 (cnd @0 @1 @1)
1467 @1)
1468
1469 /* !A ? B : C -> A ? C : B. */
1470 (simplify
1471 (cnd (logical_inverted_value truth_valued_p@0) @1 @2)
1472 (cnd @0 @2 @1)))
1473
1474 /* A + (B vcmp C ? 1 : 0) -> A - (B vcmp C), since vector comparisons
1475 return all-1 or all-0 results. */
1476 /* ??? We could instead convert all instances of the vec_cond to negate,
1477 but that isn't necessarily a win on its own. */
1478 (simplify
1479 (plus:c @3 (view_convert? (vec_cond @0 integer_each_onep@1 integer_zerop@2)))
1480 (if (VECTOR_TYPE_P (type)
1481 && TYPE_VECTOR_SUBPARTS (type) == TYPE_VECTOR_SUBPARTS (TREE_TYPE (@0))
1482 && (TYPE_MODE (TREE_TYPE (type))
1483 == TYPE_MODE (TREE_TYPE (TREE_TYPE (@0)))))
1484 (minus @3 (view_convert @0))))
1485
1486 /* ... likewise A - (B vcmp C ? 1 : 0) -> A + (B vcmp C). */
1487 (simplify
1488 (minus @3 (view_convert? (vec_cond @0 integer_each_onep@1 integer_zerop@2)))
1489 (if (VECTOR_TYPE_P (type)
1490 && TYPE_VECTOR_SUBPARTS (type) == TYPE_VECTOR_SUBPARTS (TREE_TYPE (@0))
1491 && (TYPE_MODE (TREE_TYPE (type))
1492 == TYPE_MODE (TREE_TYPE (TREE_TYPE (@0)))))
1493 (plus @3 (view_convert @0))))
1494
1495
1496 /* Simplifications of comparisons. */
1497
1498 /* See if we can reduce the magnitude of a constant involved in a
1499 comparison by changing the comparison code. This is a canonicalization
1500 formerly done by maybe_canonicalize_comparison_1. */
1501 (for cmp (le gt)
1502 acmp (lt ge)
1503 (simplify
1504 (cmp @0 INTEGER_CST@1)
1505 (if (tree_int_cst_sgn (@1) == -1)
1506 (acmp @0 { wide_int_to_tree (TREE_TYPE (@1), wi::add (@1, 1)); }))))
1507 (for cmp (ge lt)
1508 acmp (gt le)
1509 (simplify
1510 (cmp @0 INTEGER_CST@1)
1511 (if (tree_int_cst_sgn (@1) == 1)
1512 (acmp @0 { wide_int_to_tree (TREE_TYPE (@1), wi::sub (@1, 1)); }))))
1513
1514
1515 /* We can simplify a logical negation of a comparison to the
1516 inverted comparison. As we cannot compute an expression
1517 operator using invert_tree_comparison we have to simulate
1518 that with expression code iteration. */
1519 (for cmp (tcc_comparison)
1520 icmp (inverted_tcc_comparison)
1521 ncmp (inverted_tcc_comparison_with_nans)
1522 /* Ideally we'd like to combine the following two patterns
1523 and handle some more cases by using
1524 (logical_inverted_value (cmp @0 @1))
1525 here but for that genmatch would need to "inline" that.
1526 For now implement what forward_propagate_comparison did. */
1527 (simplify
1528 (bit_not (cmp @0 @1))
1529 (if (VECTOR_TYPE_P (type)
1530 || (INTEGRAL_TYPE_P (type) && TYPE_PRECISION (type) == 1))
1531 /* Comparison inversion may be impossible for trapping math,
1532 invert_tree_comparison will tell us. But we can't use
1533 a computed operator in the replacement tree thus we have
1534 to play the trick below. */
1535 (with { enum tree_code ic = invert_tree_comparison
1536 (cmp, HONOR_NANS (@0)); }
1537 (if (ic == icmp)
1538 (icmp @0 @1)
1539 (if (ic == ncmp)
1540 (ncmp @0 @1))))))
1541 (simplify
1542 (bit_xor (cmp @0 @1) integer_truep)
1543 (with { enum tree_code ic = invert_tree_comparison
1544 (cmp, HONOR_NANS (@0)); }
1545 (if (ic == icmp)
1546 (icmp @0 @1)
1547 (if (ic == ncmp)
1548 (ncmp @0 @1))))))
1549
1550 /* Transform comparisons of the form X - Y CMP 0 to X CMP Y.
1551 ??? The transformation is valid for the other operators if overflow
1552 is undefined for the type, but performing it here badly interacts
1553 with the transformation in fold_cond_expr_with_comparison which
1554 attempts to synthetize ABS_EXPR. */
1555 (for cmp (eq ne)
1556 (simplify
1557 (cmp (minus@2 @0 @1) integer_zerop)
1558 (if (single_use (@2))
1559 (cmp @0 @1))))
1560
1561 /* Transform comparisons of the form X * C1 CMP 0 to X CMP 0 in the
1562 signed arithmetic case. That form is created by the compiler
1563 often enough for folding it to be of value. One example is in
1564 computing loop trip counts after Operator Strength Reduction. */
1565 (for cmp (simple_comparison)
1566 scmp (swapped_simple_comparison)
1567 (simplify
1568 (cmp (mult @0 INTEGER_CST@1) integer_zerop@2)
1569 /* Handle unfolded multiplication by zero. */
1570 (if (integer_zerop (@1))
1571 (cmp @1 @2)
1572 (if (ANY_INTEGRAL_TYPE_P (TREE_TYPE (@0))
1573 && TYPE_OVERFLOW_UNDEFINED (TREE_TYPE (@0)))
1574 /* If @1 is negative we swap the sense of the comparison. */
1575 (if (tree_int_cst_sgn (@1) < 0)
1576 (scmp @0 @2)
1577 (cmp @0 @2))))))
1578
1579 /* Simplify comparison of something with itself. For IEEE
1580 floating-point, we can only do some of these simplifications. */
1581 (simplify
1582 (eq @0 @0)
1583 (if (! FLOAT_TYPE_P (TREE_TYPE (@0))
1584 || ! HONOR_NANS (TYPE_MODE (TREE_TYPE (@0))))
1585 { constant_boolean_node (true, type); }))
1586 (for cmp (ge le)
1587 (simplify
1588 (cmp @0 @0)
1589 (eq @0 @0)))
1590 (for cmp (ne gt lt)
1591 (simplify
1592 (cmp @0 @0)
1593 (if (cmp != NE_EXPR
1594 || ! FLOAT_TYPE_P (TREE_TYPE (@0))
1595 || ! HONOR_NANS (TYPE_MODE (TREE_TYPE (@0))))
1596 { constant_boolean_node (false, type); })))
1597 (for cmp (unle unge uneq)
1598 (simplify
1599 (cmp @0 @0)
1600 { constant_boolean_node (true, type); }))
1601 (simplify
1602 (ltgt @0 @0)
1603 (if (!flag_trapping_math)
1604 { constant_boolean_node (false, type); }))
1605
1606 /* Fold ~X op ~Y as Y op X. */
1607 (for cmp (simple_comparison)
1608 (simplify
1609 (cmp (bit_not @0) (bit_not @1))
1610 (cmp @1 @0)))
1611
1612 /* Fold ~X op C as X op' ~C, where op' is the swapped comparison. */
1613 (for cmp (simple_comparison)
1614 scmp (swapped_simple_comparison)
1615 (simplify
1616 (cmp (bit_not @0) CONSTANT_CLASS_P@1)
1617 (if (TREE_CODE (@1) == INTEGER_CST || TREE_CODE (@1) == VECTOR_CST)
1618 (scmp @0 (bit_not @1)))))
1619
1620 (for cmp (simple_comparison)
1621 /* Fold (double)float1 CMP (double)float2 into float1 CMP float2. */
1622 (simplify
1623 (cmp (convert@2 @0) (convert? @1))
1624 (if (FLOAT_TYPE_P (TREE_TYPE (@0))
1625 && (DECIMAL_FLOAT_TYPE_P (TREE_TYPE (@2))
1626 == DECIMAL_FLOAT_TYPE_P (TREE_TYPE (@0)))
1627 && (DECIMAL_FLOAT_TYPE_P (TREE_TYPE (@2))
1628 == DECIMAL_FLOAT_TYPE_P (TREE_TYPE (@1))))
1629 (with
1630 {
1631 tree type1 = TREE_TYPE (@1);
1632 if (TREE_CODE (@1) == REAL_CST && !DECIMAL_FLOAT_TYPE_P (type1))
1633 {
1634 REAL_VALUE_TYPE orig = TREE_REAL_CST (@1);
1635 if (TYPE_PRECISION (type1) > TYPE_PRECISION (float_type_node)
1636 && exact_real_truncate (TYPE_MODE (float_type_node), &orig))
1637 type1 = float_type_node;
1638 if (TYPE_PRECISION (type1) > TYPE_PRECISION (double_type_node)
1639 && exact_real_truncate (TYPE_MODE (double_type_node), &orig))
1640 type1 = double_type_node;
1641 }
1642 tree newtype
1643 = (TYPE_PRECISION (TREE_TYPE (@0)) > TYPE_PRECISION (type1)
1644 ? TREE_TYPE (@0) : type1);
1645 }
1646 (if (TYPE_PRECISION (TREE_TYPE (@2)) > TYPE_PRECISION (newtype))
1647 (cmp (convert:newtype @0) (convert:newtype @1))))))
1648
1649 (simplify
1650 (cmp @0 REAL_CST@1)
1651 /* IEEE doesn't distinguish +0 and -0 in comparisons. */
1652 (switch
1653 /* a CMP (-0) -> a CMP 0 */
1654 (if (REAL_VALUE_MINUS_ZERO (TREE_REAL_CST (@1)))
1655 (cmp @0 { build_real (TREE_TYPE (@1), dconst0); }))
1656 /* x != NaN is always true, other ops are always false. */
1657 (if (REAL_VALUE_ISNAN (TREE_REAL_CST (@1))
1658 && ! HONOR_SNANS (@1))
1659 { constant_boolean_node (cmp == NE_EXPR, type); })
1660 /* Fold comparisons against infinity. */
1661 (if (REAL_VALUE_ISINF (TREE_REAL_CST (@1))
1662 && MODE_HAS_INFINITIES (TYPE_MODE (TREE_TYPE (@1))))
1663 (with
1664 {
1665 REAL_VALUE_TYPE max;
1666 enum tree_code code = cmp;
1667 bool neg = REAL_VALUE_NEGATIVE (TREE_REAL_CST (@1));
1668 if (neg)
1669 code = swap_tree_comparison (code);
1670 }
1671 (switch
1672 /* x > +Inf is always false, if with ignore sNANs. */
1673 (if (code == GT_EXPR
1674 && ! HONOR_SNANS (@0))
1675 { constant_boolean_node (false, type); })
1676 (if (code == LE_EXPR)
1677 /* x <= +Inf is always true, if we don't case about NaNs. */
1678 (if (! HONOR_NANS (@0))
1679 { constant_boolean_node (true, type); }
1680 /* x <= +Inf is the same as x == x, i.e. !isnan(x). */
1681 (eq @0 @0)))
1682 /* x == +Inf and x >= +Inf are always equal to x > DBL_MAX. */
1683 (if (code == EQ_EXPR || code == GE_EXPR)
1684 (with { real_maxval (&max, neg, TYPE_MODE (TREE_TYPE (@0))); }
1685 (if (neg)
1686 (lt @0 { build_real (TREE_TYPE (@0), max); })
1687 (gt @0 { build_real (TREE_TYPE (@0), max); }))))
1688 /* x < +Inf is always equal to x <= DBL_MAX. */
1689 (if (code == LT_EXPR)
1690 (with { real_maxval (&max, neg, TYPE_MODE (TREE_TYPE (@0))); }
1691 (if (neg)
1692 (ge @0 { build_real (TREE_TYPE (@0), max); })
1693 (le @0 { build_real (TREE_TYPE (@0), max); }))))
1694 /* x != +Inf is always equal to !(x > DBL_MAX). */
1695 (if (code == NE_EXPR)
1696 (with { real_maxval (&max, neg, TYPE_MODE (TREE_TYPE (@0))); }
1697 (if (! HONOR_NANS (@0))
1698 (if (neg)
1699 (ge @0 { build_real (TREE_TYPE (@0), max); })
1700 (le @0 { build_real (TREE_TYPE (@0), max); }))
1701 (if (neg)
1702 (bit_xor (lt @0 { build_real (TREE_TYPE (@0), max); })
1703 { build_one_cst (type); })
1704 (bit_xor (gt @0 { build_real (TREE_TYPE (@0), max); })
1705 { build_one_cst (type); }))))))))))
1706
1707 /* If this is a comparison of a real constant with a PLUS_EXPR
1708 or a MINUS_EXPR of a real constant, we can convert it into a
1709 comparison with a revised real constant as long as no overflow
1710 occurs when unsafe_math_optimizations are enabled. */
1711 (if (flag_unsafe_math_optimizations)
1712 (for op (plus minus)
1713 (simplify
1714 (cmp (op @0 REAL_CST@1) REAL_CST@2)
1715 (with
1716 {
1717 tree tem = const_binop (op == PLUS_EXPR ? MINUS_EXPR : PLUS_EXPR,
1718 TREE_TYPE (@1), @2, @1);
1719 }
1720 (if (tem && !TREE_OVERFLOW (tem))
1721 (cmp @0 { tem; }))))))
1722
1723 /* Likewise, we can simplify a comparison of a real constant with
1724 a MINUS_EXPR whose first operand is also a real constant, i.e.
1725 (c1 - x) < c2 becomes x > c1-c2. Reordering is allowed on
1726 floating-point types only if -fassociative-math is set. */
1727 (if (flag_associative_math)
1728 (simplify
1729 (cmp (minus REAL_CST@0 @1) REAL_CST@2)
1730 (with { tree tem = const_binop (MINUS_EXPR, TREE_TYPE (@1), @0, @2); }
1731 (if (tem && !TREE_OVERFLOW (tem))
1732 (cmp { tem; } @1)))))
1733
1734 /* Fold comparisons against built-in math functions. */
1735 (if (flag_unsafe_math_optimizations
1736 && ! flag_errno_math)
1737 (for sq (SQRT)
1738 (simplify
1739 (cmp (sq @0) REAL_CST@1)
1740 (switch
1741 (if (REAL_VALUE_NEGATIVE (TREE_REAL_CST (@1)))
1742 (switch
1743 /* sqrt(x) < y is always false, if y is negative. */
1744 (if (cmp == EQ_EXPR || cmp == LT_EXPR || cmp == LE_EXPR)
1745 { constant_boolean_node (false, type); })
1746 /* sqrt(x) > y is always true, if y is negative and we
1747 don't care about NaNs, i.e. negative values of x. */
1748 (if (cmp == NE_EXPR || !HONOR_NANS (@0))
1749 { constant_boolean_node (true, type); })
1750 /* sqrt(x) > y is the same as x >= 0, if y is negative. */
1751 (ge @0 { build_real (TREE_TYPE (@0), dconst0); })))
1752 (if (cmp == GT_EXPR || cmp == GE_EXPR)
1753 (with
1754 {
1755 REAL_VALUE_TYPE c2;
1756 real_arithmetic (&c2, MULT_EXPR,
1757 &TREE_REAL_CST (@1), &TREE_REAL_CST (@1));
1758 real_convert (&c2, TYPE_MODE (TREE_TYPE (@0)), &c2);
1759 }
1760 (if (REAL_VALUE_ISINF (c2))
1761 /* sqrt(x) > y is x == +Inf, when y is very large. */
1762 (if (HONOR_INFINITIES (@0))
1763 (eq @0 { build_real (TREE_TYPE (@0), c2); })
1764 { constant_boolean_node (false, type); })
1765 /* sqrt(x) > c is the same as x > c*c. */
1766 (cmp @0 { build_real (TREE_TYPE (@0), c2); }))))
1767 (if (cmp == LT_EXPR || cmp == LE_EXPR)
1768 (with
1769 {
1770 REAL_VALUE_TYPE c2;
1771 real_arithmetic (&c2, MULT_EXPR,
1772 &TREE_REAL_CST (@1), &TREE_REAL_CST (@1));
1773 real_convert (&c2, TYPE_MODE (TREE_TYPE (@0)), &c2);
1774 }
1775 (if (REAL_VALUE_ISINF (c2))
1776 (switch
1777 /* sqrt(x) < y is always true, when y is a very large
1778 value and we don't care about NaNs or Infinities. */
1779 (if (! HONOR_NANS (@0) && ! HONOR_INFINITIES (@0))
1780 { constant_boolean_node (true, type); })
1781 /* sqrt(x) < y is x != +Inf when y is very large and we
1782 don't care about NaNs. */
1783 (if (! HONOR_NANS (@0))
1784 (ne @0 { build_real (TREE_TYPE (@0), c2); }))
1785 /* sqrt(x) < y is x >= 0 when y is very large and we
1786 don't care about Infinities. */
1787 (if (! HONOR_INFINITIES (@0))
1788 (ge @0 { build_real (TREE_TYPE (@0), dconst0); }))
1789 /* sqrt(x) < y is x >= 0 && x != +Inf, when y is large. */
1790 (if (GENERIC)
1791 (truth_andif
1792 (ge @0 { build_real (TREE_TYPE (@0), dconst0); })
1793 (ne @0 { build_real (TREE_TYPE (@0), c2); }))))
1794 /* sqrt(x) < c is the same as x < c*c, if we ignore NaNs. */
1795 (if (! HONOR_NANS (@0))
1796 (cmp @0 { build_real (TREE_TYPE (@0), c2); })
1797 /* sqrt(x) < c is the same as x >= 0 && x < c*c. */
1798 (if (GENERIC)
1799 (truth_andif
1800 (ge @0 { build_real (TREE_TYPE (@0), dconst0); })
1801 (cmp @0 { build_real (TREE_TYPE (@0), c2); }))))))))))))
1802
1803 /* Unordered tests if either argument is a NaN. */
1804 (simplify
1805 (bit_ior (unordered @0 @0) (unordered @1 @1))
1806 (if (types_match (@0, @1))
1807 (unordered @0 @1)))
1808 (simplify
1809 (bit_and (ordered @0 @0) (ordered @1 @1))
1810 (if (types_match (@0, @1))
1811 (ordered @0 @1)))
1812 (simplify
1813 (bit_ior:c (unordered @0 @0) (unordered:c@2 @0 @1))
1814 @2)
1815 (simplify
1816 (bit_and:c (ordered @0 @0) (ordered:c@2 @0 @1))
1817 @2)
1818
1819 /* -A CMP -B -> B CMP A. */
1820 (for cmp (tcc_comparison)
1821 scmp (swapped_tcc_comparison)
1822 (simplify
1823 (cmp (negate @0) (negate @1))
1824 (if (FLOAT_TYPE_P (TREE_TYPE (@0))
1825 || (ANY_INTEGRAL_TYPE_P (TREE_TYPE (@0))
1826 && TYPE_OVERFLOW_UNDEFINED (TREE_TYPE (@0))))
1827 (scmp @0 @1)))
1828 (simplify
1829 (cmp (negate @0) CONSTANT_CLASS_P@1)
1830 (if (FLOAT_TYPE_P (TREE_TYPE (@0))
1831 || (ANY_INTEGRAL_TYPE_P (TREE_TYPE (@0))
1832 && TYPE_OVERFLOW_UNDEFINED (TREE_TYPE (@0))))
1833 (with { tree tem = fold_unary (NEGATE_EXPR, TREE_TYPE (@0), @1); }
1834 (if (tem && !TREE_OVERFLOW (tem))
1835 (scmp @0 { tem; }))))))
1836
1837 /* Convert ABS_EXPR<x> == 0 or ABS_EXPR<x> != 0 to x == 0 or x != 0. */
1838 (for op (eq ne)
1839 (simplify
1840 (op (abs @0) zerop@1)
1841 (op @0 @1)))
1842
1843 /* From fold_sign_changed_comparison and fold_widened_comparison. */
1844 (for cmp (simple_comparison)
1845 (simplify
1846 (cmp (convert@0 @00) (convert?@1 @10))
1847 (if (TREE_CODE (TREE_TYPE (@0)) == INTEGER_TYPE
1848 /* Disable this optimization if we're casting a function pointer
1849 type on targets that require function pointer canonicalization. */
1850 && !(targetm.have_canonicalize_funcptr_for_compare ()
1851 && TREE_CODE (TREE_TYPE (@00)) == POINTER_TYPE
1852 && TREE_CODE (TREE_TYPE (TREE_TYPE (@00))) == FUNCTION_TYPE)
1853 && single_use (@0))
1854 (if (TYPE_PRECISION (TREE_TYPE (@00)) == TYPE_PRECISION (TREE_TYPE (@0))
1855 && (TREE_CODE (@10) == INTEGER_CST
1856 || (@1 != @10 && types_match (TREE_TYPE (@10), TREE_TYPE (@00))))
1857 && (TYPE_UNSIGNED (TREE_TYPE (@00)) == TYPE_UNSIGNED (TREE_TYPE (@0))
1858 || cmp == NE_EXPR
1859 || cmp == EQ_EXPR)
1860 && (POINTER_TYPE_P (TREE_TYPE (@00)) == POINTER_TYPE_P (TREE_TYPE (@0))))
1861 /* ??? The special-casing of INTEGER_CST conversion was in the original
1862 code and here to avoid a spurious overflow flag on the resulting
1863 constant which fold_convert produces. */
1864 (if (TREE_CODE (@1) == INTEGER_CST)
1865 (cmp @00 { force_fit_type (TREE_TYPE (@00), wi::to_widest (@1), 0,
1866 TREE_OVERFLOW (@1)); })
1867 (cmp @00 (convert @1)))
1868
1869 (if (TYPE_PRECISION (TREE_TYPE (@0)) > TYPE_PRECISION (TREE_TYPE (@00)))
1870 /* If possible, express the comparison in the shorter mode. */
1871 (if ((cmp == EQ_EXPR || cmp == NE_EXPR
1872 || TYPE_UNSIGNED (TREE_TYPE (@0)) == TYPE_UNSIGNED (TREE_TYPE (@00)))
1873 && (types_match (TREE_TYPE (@10), TREE_TYPE (@00))
1874 || ((TYPE_PRECISION (TREE_TYPE (@00))
1875 >= TYPE_PRECISION (TREE_TYPE (@10)))
1876 && (TYPE_UNSIGNED (TREE_TYPE (@00))
1877 == TYPE_UNSIGNED (TREE_TYPE (@10))))
1878 || (TREE_CODE (@10) == INTEGER_CST
1879 && (TREE_CODE (TREE_TYPE (@00)) == INTEGER_TYPE
1880 || TREE_CODE (TREE_TYPE (@00)) == BOOLEAN_TYPE)
1881 && int_fits_type_p (@10, TREE_TYPE (@00)))))
1882 (cmp @00 (convert @10))
1883 (if (TREE_CODE (@10) == INTEGER_CST
1884 && TREE_CODE (TREE_TYPE (@00)) == INTEGER_TYPE
1885 && !int_fits_type_p (@10, TREE_TYPE (@00)))
1886 (with
1887 {
1888 tree min = lower_bound_in_type (TREE_TYPE (@10), TREE_TYPE (@00));
1889 tree max = upper_bound_in_type (TREE_TYPE (@10), TREE_TYPE (@00));
1890 bool above = integer_nonzerop (const_binop (LT_EXPR, type, max, @10));
1891 bool below = integer_nonzerop (const_binop (LT_EXPR, type, @10, min));
1892 }
1893 (if (above || below)
1894 (if (cmp == EQ_EXPR || cmp == NE_EXPR)
1895 { constant_boolean_node (cmp == EQ_EXPR ? false : true, type); }
1896 (if (cmp == LT_EXPR || cmp == LE_EXPR)
1897 { constant_boolean_node (above ? true : false, type); }
1898 (if (cmp == GT_EXPR || cmp == GE_EXPR)
1899 { constant_boolean_node (above ? false : true, type); }))))))))))))
1900
1901 (for cmp (eq ne)
1902 /* A local variable can never be pointed to by
1903 the default SSA name of an incoming parameter.
1904 SSA names are canonicalized to 2nd place. */
1905 (simplify
1906 (cmp addr@0 SSA_NAME@1)
1907 (if (SSA_NAME_IS_DEFAULT_DEF (@1)
1908 && TREE_CODE (SSA_NAME_VAR (@1)) == PARM_DECL)
1909 (with { tree base = get_base_address (TREE_OPERAND (@0, 0)); }
1910 (if (TREE_CODE (base) == VAR_DECL
1911 && auto_var_in_fn_p (base, current_function_decl))
1912 (if (cmp == NE_EXPR)
1913 { constant_boolean_node (true, type); }
1914 { constant_boolean_node (false, type); }))))))
1915
1916 /* Equality compare simplifications from fold_binary */
1917 (for cmp (eq ne)
1918
1919 /* If we have (A | C) == D where C & ~D != 0, convert this into 0.
1920 Similarly for NE_EXPR. */
1921 (simplify
1922 (cmp (convert?@3 (bit_ior @0 INTEGER_CST@1)) INTEGER_CST@2)
1923 (if (tree_nop_conversion_p (TREE_TYPE (@3), TREE_TYPE (@0))
1924 && wi::bit_and_not (@1, @2) != 0)
1925 { constant_boolean_node (cmp == NE_EXPR, type); }))
1926
1927 /* (X ^ Y) == 0 becomes X == Y, and (X ^ Y) != 0 becomes X != Y. */
1928 (simplify
1929 (cmp (bit_xor @0 @1) integer_zerop)
1930 (cmp @0 @1))
1931
1932 /* (X ^ Y) == Y becomes X == 0.
1933 Likewise (X ^ Y) == X becomes Y == 0. */
1934 (simplify
1935 (cmp:c (bit_xor:c @0 @1) @0)
1936 (cmp @1 { build_zero_cst (TREE_TYPE (@1)); }))
1937
1938 /* (X ^ C1) op C2 can be rewritten as X op (C1 ^ C2). */
1939 (simplify
1940 (cmp (convert?@3 (bit_xor @0 INTEGER_CST@1)) INTEGER_CST@2)
1941 (if (tree_nop_conversion_p (TREE_TYPE (@3), TREE_TYPE (@0)))
1942 (cmp @0 (bit_xor @1 (convert @2)))))
1943
1944 (simplify
1945 (cmp (convert? addr@0) integer_zerop)
1946 (if (tree_single_nonzero_warnv_p (@0, NULL))
1947 { constant_boolean_node (cmp == NE_EXPR, type); })))
1948
1949 /* If we have (A & C) == C where C is a power of 2, convert this into
1950 (A & C) != 0. Similarly for NE_EXPR. */
1951 (for cmp (eq ne)
1952 icmp (ne eq)
1953 (simplify
1954 (cmp (bit_and@2 @0 integer_pow2p@1) @1)
1955 (icmp @2 { build_zero_cst (TREE_TYPE (@0)); })))
1956
1957 /* If we have (A & C) != 0 where C is the sign bit of A, convert
1958 this into A < 0. Similarly for (A & C) == 0 into A >= 0. */
1959 (for cmp (eq ne)
1960 ncmp (ge lt)
1961 (simplify
1962 (cmp (bit_and (convert?@2 @0) integer_pow2p@1) integer_zerop)
1963 (if (INTEGRAL_TYPE_P (TREE_TYPE (@0))
1964 && (TYPE_PRECISION (TREE_TYPE (@0))
1965 == GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (@0))))
1966 && element_precision (@2) >= element_precision (@0)
1967 && wi::only_sign_bit_p (@1, element_precision (@0)))
1968 (with { tree stype = signed_type_for (TREE_TYPE (@0)); }
1969 (ncmp (convert:stype @0) { build_zero_cst (stype); })))))
1970
1971 /* When the addresses are not directly of decls compare base and offset.
1972 This implements some remaining parts of fold_comparison address
1973 comparisons but still no complete part of it. Still it is good
1974 enough to make fold_stmt not regress when not dispatching to fold_binary. */
1975 (for cmp (simple_comparison)
1976 (simplify
1977 (cmp (convert1?@2 addr@0) (convert2? addr@1))
1978 (with
1979 {
1980 HOST_WIDE_INT off0, off1;
1981 tree base0 = get_addr_base_and_unit_offset (TREE_OPERAND (@0, 0), &off0);
1982 tree base1 = get_addr_base_and_unit_offset (TREE_OPERAND (@1, 0), &off1);
1983 if (base0 && TREE_CODE (base0) == MEM_REF)
1984 {
1985 off0 += mem_ref_offset (base0).to_short_addr ();
1986 base0 = TREE_OPERAND (base0, 0);
1987 }
1988 if (base1 && TREE_CODE (base1) == MEM_REF)
1989 {
1990 off1 += mem_ref_offset (base1).to_short_addr ();
1991 base1 = TREE_OPERAND (base1, 0);
1992 }
1993 }
1994 (if (base0 && base1)
1995 (with
1996 {
1997 int equal = 2;
1998 if (decl_in_symtab_p (base0)
1999 && decl_in_symtab_p (base1))
2000 equal = symtab_node::get_create (base0)
2001 ->equal_address_to (symtab_node::get_create (base1));
2002 else if ((DECL_P (base0)
2003 || TREE_CODE (base0) == SSA_NAME
2004 || TREE_CODE (base0) == STRING_CST)
2005 && (DECL_P (base1)
2006 || TREE_CODE (base1) == SSA_NAME
2007 || TREE_CODE (base1) == STRING_CST))
2008 equal = (base0 == base1);
2009 }
2010 (if (equal == 1
2011 && (cmp == EQ_EXPR || cmp == NE_EXPR
2012 /* If the offsets are equal we can ignore overflow. */
2013 || off0 == off1
2014 || POINTER_TYPE_OVERFLOW_UNDEFINED
2015 /* Or if we compare using pointers to decls or strings. */
2016 || (POINTER_TYPE_P (TREE_TYPE (@2))
2017 && (DECL_P (base0) || TREE_CODE (base0) == STRING_CST))))
2018 (switch
2019 (if (cmp == EQ_EXPR)
2020 { constant_boolean_node (off0 == off1, type); })
2021 (if (cmp == NE_EXPR)
2022 { constant_boolean_node (off0 != off1, type); })
2023 (if (cmp == LT_EXPR)
2024 { constant_boolean_node (off0 < off1, type); })
2025 (if (cmp == LE_EXPR)
2026 { constant_boolean_node (off0 <= off1, type); })
2027 (if (cmp == GE_EXPR)
2028 { constant_boolean_node (off0 >= off1, type); })
2029 (if (cmp == GT_EXPR)
2030 { constant_boolean_node (off0 > off1, type); }))
2031 (if (equal == 0
2032 && DECL_P (base0) && DECL_P (base1)
2033 /* If we compare this as integers require equal offset. */
2034 && (!INTEGRAL_TYPE_P (TREE_TYPE (@2))
2035 || off0 == off1))
2036 (switch
2037 (if (cmp == EQ_EXPR)
2038 { constant_boolean_node (false, type); })
2039 (if (cmp == NE_EXPR)
2040 { constant_boolean_node (true, type); })))))))))
2041
2042 /* Non-equality compare simplifications from fold_binary */
2043 (for cmp (lt gt le ge)
2044 /* Comparisons with the highest or lowest possible integer of
2045 the specified precision will have known values. */
2046 (simplify
2047 (cmp (convert?@2 @0) INTEGER_CST@1)
2048 (if ((INTEGRAL_TYPE_P (TREE_TYPE (@1)) || POINTER_TYPE_P (TREE_TYPE (@1)))
2049 && tree_nop_conversion_p (TREE_TYPE (@2), TREE_TYPE (@0)))
2050 (with
2051 {
2052 tree arg1_type = TREE_TYPE (@1);
2053 unsigned int prec = TYPE_PRECISION (arg1_type);
2054 wide_int max = wi::max_value (arg1_type);
2055 wide_int signed_max = wi::max_value (prec, SIGNED);
2056 wide_int min = wi::min_value (arg1_type);
2057 }
2058 (switch
2059 (if (wi::eq_p (@1, max))
2060 (switch
2061 (if (cmp == GT_EXPR)
2062 { constant_boolean_node (false, type); })
2063 (if (cmp == GE_EXPR)
2064 (eq @2 @1))
2065 (if (cmp == LE_EXPR)
2066 { constant_boolean_node (true, type); })
2067 (if (cmp == LT_EXPR)
2068 (ne @2 @1))))
2069 (if (wi::eq_p (@1, min))
2070 (switch
2071 (if (cmp == LT_EXPR)
2072 { constant_boolean_node (false, type); })
2073 (if (cmp == LE_EXPR)
2074 (eq @2 @1))
2075 (if (cmp == GE_EXPR)
2076 { constant_boolean_node (true, type); })
2077 (if (cmp == GT_EXPR)
2078 (ne @2 @1))))
2079 (if (wi::eq_p (@1, max - 1))
2080 (switch
2081 (if (cmp == GT_EXPR)
2082 (eq @2 { wide_int_to_tree (TREE_TYPE (@1), wi::add (@1, 1)); }))
2083 (if (cmp == LE_EXPR)
2084 (ne @2 { wide_int_to_tree (TREE_TYPE (@1), wi::add (@1, 1)); }))))
2085 (if (wi::eq_p (@1, min + 1))
2086 (switch
2087 (if (cmp == GE_EXPR)
2088 (ne @2 { wide_int_to_tree (TREE_TYPE (@1), wi::sub (@1, 1)); }))
2089 (if (cmp == LT_EXPR)
2090 (eq @2 { wide_int_to_tree (TREE_TYPE (@1), wi::sub (@1, 1)); }))))
2091 (if (wi::eq_p (@1, signed_max)
2092 && TYPE_UNSIGNED (arg1_type)
2093 /* We will flip the signedness of the comparison operator
2094 associated with the mode of @1, so the sign bit is
2095 specified by this mode. Check that @1 is the signed
2096 max associated with this sign bit. */
2097 && prec == GET_MODE_PRECISION (TYPE_MODE (arg1_type))
2098 /* signed_type does not work on pointer types. */
2099 && INTEGRAL_TYPE_P (arg1_type))
2100 /* The following case also applies to X < signed_max+1
2101 and X >= signed_max+1 because previous transformations. */
2102 (if (cmp == LE_EXPR || cmp == GT_EXPR)
2103 (with { tree st = signed_type_for (arg1_type); }
2104 (if (cmp == LE_EXPR)
2105 (ge (convert:st @0) { build_zero_cst (st); })
2106 (lt (convert:st @0) { build_zero_cst (st); }))))))))))
2107
2108 (for cmp (unordered ordered unlt unle ungt unge uneq ltgt)
2109 /* If the second operand is NaN, the result is constant. */
2110 (simplify
2111 (cmp @0 REAL_CST@1)
2112 (if (REAL_VALUE_ISNAN (TREE_REAL_CST (@1))
2113 && (cmp != LTGT_EXPR || ! flag_trapping_math))
2114 { constant_boolean_node (cmp == ORDERED_EXPR || cmp == LTGT_EXPR
2115 ? false : true, type); })))
2116
2117 /* bool_var != 0 becomes bool_var. */
2118 (simplify
2119 (ne @0 integer_zerop)
2120 (if (TREE_CODE (TREE_TYPE (@0)) == BOOLEAN_TYPE
2121 && types_match (type, TREE_TYPE (@0)))
2122 (non_lvalue @0)))
2123 /* bool_var == 1 becomes bool_var. */
2124 (simplify
2125 (eq @0 integer_onep)
2126 (if (TREE_CODE (TREE_TYPE (@0)) == BOOLEAN_TYPE
2127 && types_match (type, TREE_TYPE (@0)))
2128 (non_lvalue @0)))
2129 /* Do not handle
2130 bool_var == 0 becomes !bool_var or
2131 bool_var != 1 becomes !bool_var
2132 here because that only is good in assignment context as long
2133 as we require a tcc_comparison in GIMPLE_CONDs where we'd
2134 replace if (x == 0) with tem = ~x; if (tem != 0) which is
2135 clearly less optimal and which we'll transform again in forwprop. */
2136
2137
2138 /* Simplification of math builtins. These rules must all be optimizations
2139 as well as IL simplifications. If there is a possibility that the new
2140 form could be a pessimization, the rule should go in the canonicalization
2141 section that follows this one.
2142
2143 Rules can generally go in this section if they satisfy one of
2144 the following:
2145
2146 - the rule describes an identity
2147
2148 - the rule replaces calls with something as simple as addition or
2149 multiplication
2150
2151 - the rule contains unary calls only and simplifies the surrounding
2152 arithmetic. (The idea here is to exclude non-unary calls in which
2153 one operand is constant and in which the call is known to be cheap
2154 when the operand has that value.) */
2155
2156 (if (flag_unsafe_math_optimizations)
2157 /* Simplify sqrt(x) * sqrt(x) -> x. */
2158 (simplify
2159 (mult (SQRT@1 @0) @1)
2160 (if (!HONOR_SNANS (type))
2161 @0))
2162
2163 /* Simplify sqrt(x) * sqrt(y) -> sqrt(x*y). */
2164 (for root (SQRT CBRT)
2165 (simplify
2166 (mult (root:s @0) (root:s @1))
2167 (root (mult @0 @1))))
2168
2169 /* Simplify expN(x) * expN(y) -> expN(x+y). */
2170 (for exps (EXP EXP2 EXP10 POW10)
2171 (simplify
2172 (mult (exps:s @0) (exps:s @1))
2173 (exps (plus @0 @1))))
2174
2175 /* Simplify a/root(b/c) into a*root(c/b). */
2176 (for root (SQRT CBRT)
2177 (simplify
2178 (rdiv @0 (root:s (rdiv:s @1 @2)))
2179 (mult @0 (root (rdiv @2 @1)))))
2180
2181 /* Simplify x/expN(y) into x*expN(-y). */
2182 (for exps (EXP EXP2 EXP10 POW10)
2183 (simplify
2184 (rdiv @0 (exps:s @1))
2185 (mult @0 (exps (negate @1)))))
2186
2187 /* Special case, optimize logN(expN(x)) = x. */
2188 (for logs (LOG LOG2 LOG10 LOG10)
2189 exps (EXP EXP2 EXP10 POW10)
2190 (simplify
2191 (logs (exps @0))
2192 @0))
2193
2194 /* Optimize logN(func()) for various exponential functions. We
2195 want to determine the value "x" and the power "exponent" in
2196 order to transform logN(x**exponent) into exponent*logN(x). */
2197 (for logs (LOG LOG LOG LOG2 LOG2 LOG2 LOG10 LOG10)
2198 exps (EXP2 EXP10 POW10 EXP EXP10 POW10 EXP EXP2)
2199 (simplify
2200 (logs (exps @0))
2201 (with {
2202 tree x;
2203 switch (exps)
2204 {
2205 CASE_FLT_FN (BUILT_IN_EXP):
2206 /* Prepare to do logN(exp(exponent)) -> exponent*logN(e). */
2207 x = build_real_truncate (type, dconst_e ());
2208 break;
2209 CASE_FLT_FN (BUILT_IN_EXP2):
2210 /* Prepare to do logN(exp2(exponent)) -> exponent*logN(2). */
2211 x = build_real (type, dconst2);
2212 break;
2213 CASE_FLT_FN (BUILT_IN_EXP10):
2214 CASE_FLT_FN (BUILT_IN_POW10):
2215 /* Prepare to do logN(exp10(exponent)) -> exponent*logN(10). */
2216 {
2217 REAL_VALUE_TYPE dconst10;
2218 real_from_integer (&dconst10, VOIDmode, 10, SIGNED);
2219 x = build_real (type, dconst10);
2220 }
2221 break;
2222 default:
2223 gcc_unreachable ();
2224 }
2225 }
2226 (mult (logs { x; }) @0))))
2227
2228 (for logs (LOG LOG
2229 LOG2 LOG2
2230 LOG10 LOG10)
2231 exps (SQRT CBRT)
2232 (simplify
2233 (logs (exps @0))
2234 (with {
2235 tree x;
2236 switch (exps)
2237 {
2238 CASE_FLT_FN (BUILT_IN_SQRT):
2239 /* Prepare to do logN(sqrt(x)) -> 0.5*logN(x). */
2240 x = build_real (type, dconsthalf);
2241 break;
2242 CASE_FLT_FN (BUILT_IN_CBRT):
2243 /* Prepare to do logN(cbrt(x)) -> (1/3)*logN(x). */
2244 x = build_real_truncate (type, dconst_third ());
2245 break;
2246 default:
2247 gcc_unreachable ();
2248 }
2249 }
2250 (mult { x; } (logs @0)))))
2251
2252 /* logN(pow(x,exponent)) -> exponent*logN(x). */
2253 (for logs (LOG LOG2 LOG10)
2254 pows (POW)
2255 (simplify
2256 (logs (pows @0 @1))
2257 (mult @1 (logs @0))))
2258
2259 (for sqrts (SQRT)
2260 cbrts (CBRT)
2261 exps (EXP EXP2 EXP10 POW10)
2262 /* sqrt(expN(x)) -> expN(x*0.5). */
2263 (simplify
2264 (sqrts (exps @0))
2265 (exps (mult @0 { build_real (type, dconsthalf); })))
2266 /* cbrt(expN(x)) -> expN(x/3). */
2267 (simplify
2268 (cbrts (exps @0))
2269 (exps (mult @0 { build_real_truncate (type, dconst_third ()); })))))
2270
2271 /* Canonicalization of sequences of math builtins. These rules represent
2272 IL simplifications but are not necessarily optimizations.
2273
2274 The sincos pass is responsible for picking "optimal" implementations
2275 of math builtins, which may be more complicated and can sometimes go
2276 the other way, e.g. converting pow into a sequence of sqrts.
2277 We only want to do these canonicalizations before the pass has run. */
2278
2279 (if (flag_unsafe_math_optimizations && canonicalize_math_p ())
2280 /* Simplify tan(x) * cos(x) -> sin(x). */
2281 (simplify
2282 (mult:c (TAN:s @0) (COS:s @0))
2283 (SIN @0))
2284
2285 /* Simplify x * pow(x,c) -> pow(x,c+1). */
2286 (simplify
2287 (mult @0 (POW:s @0 REAL_CST@1))
2288 (if (!TREE_OVERFLOW (@1))
2289 (POW @0 (plus @1 { build_one_cst (type); }))))
2290
2291 /* Simplify sin(x) / cos(x) -> tan(x). */
2292 (simplify
2293 (rdiv (SIN:s @0) (COS:s @0))
2294 (TAN @0))
2295
2296 /* Simplify cos(x) / sin(x) -> 1 / tan(x). */
2297 (simplify
2298 (rdiv (COS:s @0) (SIN:s @0))
2299 (rdiv { build_one_cst (type); } (TAN @0)))
2300
2301 /* Simplify sin(x) / tan(x) -> cos(x). */
2302 (simplify
2303 (rdiv (SIN:s @0) (TAN:s @0))
2304 (if (! HONOR_NANS (@0)
2305 && ! HONOR_INFINITIES (@0))
2306 (cos @0)))
2307
2308 /* Simplify tan(x) / sin(x) -> 1.0 / cos(x). */
2309 (simplify
2310 (rdiv (TAN:s @0) (SIN:s @0))
2311 (if (! HONOR_NANS (@0)
2312 && ! HONOR_INFINITIES (@0))
2313 (rdiv { build_one_cst (type); } (COS @0))))
2314
2315 /* Simplify pow(x,y) * pow(x,z) -> pow(x,y+z). */
2316 (simplify
2317 (mult (POW:s @0 @1) (POW:s @0 @2))
2318 (POW @0 (plus @1 @2)))
2319
2320 /* Simplify pow(x,y) * pow(z,y) -> pow(x*z,y). */
2321 (simplify
2322 (mult (POW:s @0 @1) (POW:s @2 @1))
2323 (POW (mult @0 @2) @1))
2324
2325 /* Simplify pow(x,c) / x -> pow(x,c-1). */
2326 (simplify
2327 (rdiv (POW:s @0 REAL_CST@1) @0)
2328 (if (!TREE_OVERFLOW (@1))
2329 (POW @0 (minus @1 { build_one_cst (type); }))))
2330
2331 /* Simplify x / pow (y,z) -> x * pow(y,-z). */
2332 (simplify
2333 (rdiv @0 (POW:s @1 @2))
2334 (mult @0 (POW @1 (negate @2))))
2335
2336 (for sqrts (SQRT)
2337 cbrts (CBRT)
2338 pows (POW)
2339 /* sqrt(sqrt(x)) -> pow(x,1/4). */
2340 (simplify
2341 (sqrts (sqrts @0))
2342 (pows @0 { build_real (type, dconst_quarter ()); }))
2343 /* sqrt(cbrt(x)) -> pow(x,1/6). */
2344 (simplify
2345 (sqrts (cbrts @0))
2346 (pows @0 { build_real_truncate (type, dconst_sixth ()); }))
2347 /* cbrt(sqrt(x)) -> pow(x,1/6). */
2348 (simplify
2349 (cbrts (sqrts @0))
2350 (pows @0 { build_real_truncate (type, dconst_sixth ()); }))
2351 /* cbrt(cbrt(x)) -> pow(x,1/9), iff x is nonnegative. */
2352 (simplify
2353 (cbrts (cbrts tree_expr_nonnegative_p@0))
2354 (pows @0 { build_real_truncate (type, dconst_ninth ()); }))
2355 /* sqrt(pow(x,y)) -> pow(|x|,y*0.5). */
2356 (simplify
2357 (sqrts (pows @0 @1))
2358 (pows (abs @0) (mult @1 { build_real (type, dconsthalf); })))
2359 /* cbrt(pow(x,y)) -> pow(x,y/3), iff x is nonnegative. */
2360 (simplify
2361 (cbrts (pows tree_expr_nonnegative_p@0 @1))
2362 (pows @0 (mult @1 { build_real_truncate (type, dconst_third ()); })))))
2363
2364 /* Narrowing of arithmetic and logical operations.
2365
2366 These are conceptually similar to the transformations performed for
2367 the C/C++ front-ends by shorten_binary_op and shorten_compare. Long
2368 term we want to move all that code out of the front-ends into here. */
2369
2370 /* If we have a narrowing conversion of an arithmetic operation where
2371 both operands are widening conversions from the same type as the outer
2372 narrowing conversion. Then convert the innermost operands to a suitable
2373 unsigned type (to avoid introducing undefined behaviour), perform the
2374 operation and convert the result to the desired type. */
2375 (for op (plus minus)
2376 (simplify
2377 (convert (op:s (convert@2 @0) (convert@3 @1)))
2378 (if (INTEGRAL_TYPE_P (type)
2379 /* We check for type compatibility between @0 and @1 below,
2380 so there's no need to check that @1/@3 are integral types. */
2381 && INTEGRAL_TYPE_P (TREE_TYPE (@0))
2382 && INTEGRAL_TYPE_P (TREE_TYPE (@2))
2383 /* The precision of the type of each operand must match the
2384 precision of the mode of each operand, similarly for the
2385 result. */
2386 && (TYPE_PRECISION (TREE_TYPE (@0))
2387 == GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (@0))))
2388 && (TYPE_PRECISION (TREE_TYPE (@1))
2389 == GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (@1))))
2390 && TYPE_PRECISION (type) == GET_MODE_PRECISION (TYPE_MODE (type))
2391 /* The inner conversion must be a widening conversion. */
2392 && TYPE_PRECISION (TREE_TYPE (@2)) > TYPE_PRECISION (TREE_TYPE (@0))
2393 && types_match (@0, @1)
2394 && types_match (@0, type))
2395 (if (TYPE_OVERFLOW_WRAPS (TREE_TYPE (@0)))
2396 (convert (op @0 @1))
2397 (with { tree utype = unsigned_type_for (TREE_TYPE (@0)); }
2398 (convert (op (convert:utype @0) (convert:utype @1))))))))
2399
2400 /* This is another case of narrowing, specifically when there's an outer
2401 BIT_AND_EXPR which masks off bits outside the type of the innermost
2402 operands. Like the previous case we have to convert the operands
2403 to unsigned types to avoid introducing undefined behaviour for the
2404 arithmetic operation. */
2405 (for op (minus plus)
2406 (simplify
2407 (bit_and (op:s (convert@2 @0) (convert@3 @1)) INTEGER_CST@4)
2408 (if (INTEGRAL_TYPE_P (type)
2409 /* We check for type compatibility between @0 and @1 below,
2410 so there's no need to check that @1/@3 are integral types. */
2411 && INTEGRAL_TYPE_P (TREE_TYPE (@0))
2412 && INTEGRAL_TYPE_P (TREE_TYPE (@2))
2413 /* The precision of the type of each operand must match the
2414 precision of the mode of each operand, similarly for the
2415 result. */
2416 && (TYPE_PRECISION (TREE_TYPE (@0))
2417 == GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (@0))))
2418 && (TYPE_PRECISION (TREE_TYPE (@1))
2419 == GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (@1))))
2420 && TYPE_PRECISION (type) == GET_MODE_PRECISION (TYPE_MODE (type))
2421 /* The inner conversion must be a widening conversion. */
2422 && TYPE_PRECISION (TREE_TYPE (@2)) > TYPE_PRECISION (TREE_TYPE (@0))
2423 && types_match (@0, @1)
2424 && (tree_int_cst_min_precision (@4, TYPE_SIGN (TREE_TYPE (@0)))
2425 <= TYPE_PRECISION (TREE_TYPE (@0)))
2426 && (TYPE_OVERFLOW_WRAPS (TREE_TYPE (@0))
2427 || tree_int_cst_sgn (@4) >= 0))
2428 (if (TYPE_OVERFLOW_WRAPS (TREE_TYPE (@0)))
2429 (with { tree ntype = TREE_TYPE (@0); }
2430 (convert (bit_and (op @0 @1) (convert:ntype @4))))
2431 (with { tree utype = unsigned_type_for (TREE_TYPE (@0)); }
2432 (convert (bit_and (op (convert:utype @0) (convert:utype @1))
2433 (convert:utype @4))))))))