]> git.ipfire.org Git - thirdparty/gcc.git/blame - gcc/match.pd
re PR tree-optimization/66948 (Performance regression in bit manipulation code)
[thirdparty/gcc.git] / gcc / match.pd
CommitLineData
3d2cf79f
RB
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
5624e564 5 Copyright (C) 2014-2015 Free Software Foundation, Inc.
3d2cf79f
RB
6 Contributed by Richard Biener <rguenther@suse.de>
7 and Prathamesh Kulkarni <bilbotheelffriend@gmail.com>
8
9This file is part of GCC.
10
11GCC is free software; you can redistribute it and/or modify it under
12the terms of the GNU General Public License as published by the Free
13Software Foundation; either version 3, or (at your option) any later
14version.
15
16GCC is distributed in the hope that it will be useful, but WITHOUT ANY
17WARRANTY; without even the implied warranty of MERCHANTABILITY or
18FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
19for more details.
20
21You should have received a copy of the GNU General Public License
22along 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
cc7b5acf 28 integer_onep integer_zerop integer_all_onesp integer_minus_onep
09240451 29 integer_each_onep integer_truep
cc7b5acf 30 real_zerop real_onep real_minus_onep
f3582e54
RB
31 CONSTANT_CLASS_P
32 tree_expr_nonnegative_p)
e0ee10ed 33
f84e7fd6
RB
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)
534bd33b
MG
41(define_operator_list swapped_tcc_comparison
42 gt ge eq ne le lt unordered ordered ungt unge unlt unle uneq ltgt)
07cdc2b8
RB
43(define_operator_list simple_comparison lt le eq ne ge gt)
44(define_operator_list swapped_simple_comparison gt ge eq ne le lt)
45
46(define_operator_list LOG BUILT_IN_LOGF BUILT_IN_LOG BUILT_IN_LOGL)
47(define_operator_list EXP BUILT_IN_EXPF BUILT_IN_EXP BUILT_IN_EXPL)
48(define_operator_list LOG2 BUILT_IN_LOG2F BUILT_IN_LOG2 BUILT_IN_LOG2L)
49(define_operator_list EXP2 BUILT_IN_EXP2F BUILT_IN_EXP2 BUILT_IN_EXP2L)
50(define_operator_list LOG10 BUILT_IN_LOG10F BUILT_IN_LOG10 BUILT_IN_LOG10L)
51(define_operator_list EXP10 BUILT_IN_EXP10F BUILT_IN_EXP10 BUILT_IN_EXP10L)
52(define_operator_list POW BUILT_IN_POWF BUILT_IN_POW BUILT_IN_POWL)
53(define_operator_list POW10 BUILT_IN_POW10F BUILT_IN_POW10 BUILT_IN_POW10L)
54(define_operator_list SQRT BUILT_IN_SQRTF BUILT_IN_SQRT BUILT_IN_SQRTL)
55(define_operator_list CBRT BUILT_IN_CBRTF BUILT_IN_CBRT BUILT_IN_CBRTL)
f84e7fd6 56
e0ee10ed
RB
57
58/* Simplifications of operations with one constant operand and
36a60e48 59 simplifications to constants or single values. */
e0ee10ed
RB
60
61(for op (plus pointer_plus minus bit_ior bit_xor)
62 (simplify
63 (op @0 integer_zerop)
64 (non_lvalue @0)))
65
a499aac5
RB
66/* 0 +p index -> (type)index */
67(simplify
68 (pointer_plus integer_zerop @1)
69 (non_lvalue (convert @1)))
70
a7f24614
RB
71/* See if ARG1 is zero and X + ARG1 reduces to X.
72 Likewise if the operands are reversed. */
73(simplify
74 (plus:c @0 real_zerop@1)
75 (if (fold_real_zero_addition_p (type, @1, 0))
76 (non_lvalue @0)))
77
78/* See if ARG1 is zero and X - ARG1 reduces to X. */
79(simplify
80 (minus @0 real_zerop@1)
81 (if (fold_real_zero_addition_p (type, @1, 1))
82 (non_lvalue @0)))
83
e0ee10ed
RB
84/* Simplify x - x.
85 This is unsafe for certain floats even in non-IEEE formats.
86 In IEEE, it is unsafe because it does wrong for NaNs.
87 Also note that operand_equal_p is always false if an operand
88 is volatile. */
89(simplify
a7f24614 90 (minus @0 @0)
1b457aa4 91 (if (!FLOAT_TYPE_P (type) || !HONOR_NANS (type))
a7f24614 92 { build_zero_cst (type); }))
e0ee10ed
RB
93
94(simplify
a7f24614
RB
95 (mult @0 integer_zerop@1)
96 @1)
97
98/* Maybe fold x * 0 to 0. The expressions aren't the same
99 when x is NaN, since x * 0 is also NaN. Nor are they the
100 same in modes with signed zeros, since multiplying a
101 negative value by 0 gives -0, not +0. */
102(simplify
103 (mult @0 real_zerop@1)
8b5ee871 104 (if (!HONOR_NANS (type) && !HONOR_SIGNED_ZEROS (type))
a7f24614
RB
105 @1))
106
107/* In IEEE floating point, x*1 is not equivalent to x for snans.
108 Likewise for complex arithmetic with signed zeros. */
109(simplify
110 (mult @0 real_onep)
8b5ee871
MG
111 (if (!HONOR_SNANS (type)
112 && (!HONOR_SIGNED_ZEROS (type)
a7f24614
RB
113 || !COMPLEX_FLOAT_TYPE_P (type)))
114 (non_lvalue @0)))
115
116/* Transform x * -1.0 into -x. */
117(simplify
118 (mult @0 real_minus_onep)
8b5ee871
MG
119 (if (!HONOR_SNANS (type)
120 && (!HONOR_SIGNED_ZEROS (type)
a7f24614
RB
121 || !COMPLEX_FLOAT_TYPE_P (type)))
122 (negate @0)))
e0ee10ed
RB
123
124/* Make sure to preserve divisions by zero. This is the reason why
125 we don't simplify x / x to 1 or 0 / x to 0. */
126(for op (mult trunc_div ceil_div floor_div round_div exact_div)
127 (simplify
128 (op @0 integer_onep)
129 (non_lvalue @0)))
130
a7f24614
RB
131/* X / -1 is -X. */
132(for div (trunc_div ceil_div floor_div round_div exact_div)
133 (simplify
09240451
MG
134 (div @0 integer_minus_onep@1)
135 (if (!TYPE_UNSIGNED (type))
a7f24614
RB
136 (negate @0))))
137
138/* For unsigned integral types, FLOOR_DIV_EXPR is the same as
139 TRUNC_DIV_EXPR. Rewrite into the latter in this case. */
140(simplify
141 (floor_div @0 @1)
09240451
MG
142 (if ((INTEGRAL_TYPE_P (type) || VECTOR_INTEGER_TYPE_P (type))
143 && TYPE_UNSIGNED (type))
a7f24614
RB
144 (trunc_div @0 @1)))
145
28093105
RB
146/* Combine two successive divisions. Note that combining ceil_div
147 and floor_div is trickier and combining round_div even more so. */
148(for div (trunc_div exact_div)
c306cfaf
RB
149 (simplify
150 (div (div @0 INTEGER_CST@1) INTEGER_CST@2)
151 (with {
152 bool overflow_p;
153 wide_int mul = wi::mul (@1, @2, TYPE_SIGN (type), &overflow_p);
154 }
155 (if (!overflow_p)
8fdc6c67
RB
156 (div @0 { wide_int_to_tree (type, mul); })
157 (if (TYPE_UNSIGNED (type)
158 || mul != wi::min_value (TYPE_PRECISION (type), SIGNED))
159 { build_zero_cst (type); })))))
c306cfaf 160
a7f24614 161/* Optimize A / A to 1.0 if we don't care about
09240451 162 NaNs or Infinities. */
a7f24614
RB
163(simplify
164 (rdiv @0 @0)
09240451 165 (if (FLOAT_TYPE_P (type)
1b457aa4 166 && ! HONOR_NANS (type)
8b5ee871 167 && ! HONOR_INFINITIES (type))
09240451
MG
168 { build_one_cst (type); }))
169
170/* Optimize -A / A to -1.0 if we don't care about
171 NaNs or Infinities. */
172(simplify
173 (rdiv:c @0 (negate @0))
174 (if (FLOAT_TYPE_P (type)
1b457aa4 175 && ! HONOR_NANS (type)
8b5ee871 176 && ! HONOR_INFINITIES (type))
09240451 177 { build_minus_one_cst (type); }))
a7f24614
RB
178
179/* In IEEE floating point, x/1 is not equivalent to x for snans. */
180(simplify
181 (rdiv @0 real_onep)
8b5ee871 182 (if (!HONOR_SNANS (type))
a7f24614
RB
183 (non_lvalue @0)))
184
185/* In IEEE floating point, x/-1 is not equivalent to -x for snans. */
186(simplify
187 (rdiv @0 real_minus_onep)
8b5ee871 188 (if (!HONOR_SNANS (type))
a7f24614
RB
189 (negate @0)))
190
191/* If ARG1 is a constant, we can convert this to a multiply by the
192 reciprocal. This does not have the same rounding properties,
193 so only do this if -freciprocal-math. We can actually
194 always safely do it if ARG1 is a power of two, but it's hard to
195 tell if it is or not in a portable manner. */
196(for cst (REAL_CST COMPLEX_CST VECTOR_CST)
197 (simplify
198 (rdiv @0 cst@1)
199 (if (optimize)
53bc4b3a
RB
200 (if (flag_reciprocal_math
201 && !real_zerop (@1))
a7f24614 202 (with
249700b5 203 { tree tem = const_binop (RDIV_EXPR, type, build_one_cst (type), @1); }
a7f24614 204 (if (tem)
8fdc6c67
RB
205 (mult @0 { tem; } )))
206 (if (cst != COMPLEX_CST)
207 (with { tree inverse = exact_inverse (type, @1); }
208 (if (inverse)
209 (mult @0 { inverse; } ))))))))
a7f24614 210
e0ee10ed
RB
211/* Same applies to modulo operations, but fold is inconsistent here
212 and simplifies 0 % x to 0, only preserving literal 0 % 0. */
a7f24614 213(for mod (ceil_mod floor_mod round_mod trunc_mod)
e0ee10ed
RB
214 /* 0 % X is always zero. */
215 (simplify
a7f24614 216 (mod integer_zerop@0 @1)
e0ee10ed
RB
217 /* But not for 0 % 0 so that we can get the proper warnings and errors. */
218 (if (!integer_zerop (@1))
219 @0))
220 /* X % 1 is always zero. */
221 (simplify
a7f24614
RB
222 (mod @0 integer_onep)
223 { build_zero_cst (type); })
224 /* X % -1 is zero. */
225 (simplify
09240451
MG
226 (mod @0 integer_minus_onep@1)
227 (if (!TYPE_UNSIGNED (type))
bc4315fb
MG
228 { build_zero_cst (type); }))
229 /* (X % Y) % Y is just X % Y. */
230 (simplify
231 (mod (mod@2 @0 @1) @1)
98e30e51
RB
232 @2)
233 /* From extract_muldiv_1: (X * C1) % C2 is zero if C1 is a multiple of C2. */
234 (simplify
235 (mod (mult @0 INTEGER_CST@1) INTEGER_CST@2)
236 (if (ANY_INTEGRAL_TYPE_P (type)
237 && TYPE_OVERFLOW_UNDEFINED (type)
238 && wi::multiple_of_p (@1, @2, TYPE_SIGN (type)))
239 { build_zero_cst (type); })))
a7f24614
RB
240
241/* X % -C is the same as X % C. */
242(simplify
243 (trunc_mod @0 INTEGER_CST@1)
244 (if (TYPE_SIGN (type) == SIGNED
245 && !TREE_OVERFLOW (@1)
246 && wi::neg_p (@1)
247 && !TYPE_OVERFLOW_TRAPS (type)
248 /* Avoid this transformation if C is INT_MIN, i.e. C == -C. */
249 && !sign_bit_p (@1, @1))
250 (trunc_mod @0 (negate @1))))
e0ee10ed 251
8f0c696a
RB
252/* X % -Y is the same as X % Y. */
253(simplify
254 (trunc_mod @0 (convert? (negate @1)))
255 (if (!TYPE_UNSIGNED (type)
256 && !TYPE_OVERFLOW_TRAPS (type)
257 && tree_nop_conversion_p (type, TREE_TYPE (@1)))
258 (trunc_mod @0 (convert @1))))
259
f461569a
MP
260/* X - (X / Y) * Y is the same as X % Y. */
261(simplify
d3bc1d1b 262 (minus (convert1? @0) (convert2? (mult (trunc_div @0 @1) @1)))
f461569a 263 (if (INTEGRAL_TYPE_P (type) || VECTOR_INTEGER_TYPE_P (type))
d3bc1d1b 264 (trunc_mod (convert @0) (convert @1))))
f461569a 265
8f0c696a
RB
266/* Optimize TRUNC_MOD_EXPR by a power of two into a BIT_AND_EXPR,
267 i.e. "X % C" into "X & (C - 1)", if X and C are positive.
268 Also optimize A % (C << N) where C is a power of 2,
269 to A & ((C << N) - 1). */
270(match (power_of_two_cand @1)
271 INTEGER_CST@1)
272(match (power_of_two_cand @1)
273 (lshift INTEGER_CST@1 @2))
274(for mod (trunc_mod floor_mod)
275 (simplify
4ab1e111 276 (mod @0 (convert?@3 (power_of_two_cand@1 @2)))
8f0c696a
RB
277 (if ((TYPE_UNSIGNED (type)
278 || tree_expr_nonnegative_p (@0))
4ab1e111 279 && tree_nop_conversion_p (type, TREE_TYPE (@3))
8f0c696a 280 && integer_pow2p (@2) && tree_int_cst_sgn (@2) > 0)
4ab1e111 281 (bit_and @0 (convert (minus @1 { build_int_cst (TREE_TYPE (@1), 1); }))))))
8f0c696a 282
bc4315fb
MG
283/* X % Y is smaller than Y. */
284(for cmp (lt ge)
285 (simplify
286 (cmp (trunc_mod @0 @1) @1)
287 (if (TYPE_UNSIGNED (TREE_TYPE (@0)))
288 { constant_boolean_node (cmp == LT_EXPR, type); })))
289(for cmp (gt le)
290 (simplify
291 (cmp @1 (trunc_mod @0 @1))
292 (if (TYPE_UNSIGNED (TREE_TYPE (@0)))
293 { constant_boolean_node (cmp == GT_EXPR, type); })))
294
e0ee10ed
RB
295/* x | ~0 -> ~0 */
296(simplify
297 (bit_ior @0 integer_all_onesp@1)
298 @1)
299
300/* x & 0 -> 0 */
301(simplify
302 (bit_and @0 integer_zerop@1)
303 @1)
304
a4398a30 305/* ~x | x -> -1 */
8b5ee871
MG
306/* ~x ^ x -> -1 */
307/* ~x + x -> -1 */
308(for op (bit_ior bit_xor plus)
309 (simplify
310 (op:c (convert? @0) (convert? (bit_not @0)))
311 (convert { build_all_ones_cst (TREE_TYPE (@0)); })))
a4398a30 312
e0ee10ed
RB
313/* x ^ x -> 0 */
314(simplify
315 (bit_xor @0 @0)
316 { build_zero_cst (type); })
317
36a60e48
RB
318/* Canonicalize X ^ ~0 to ~X. */
319(simplify
320 (bit_xor @0 integer_all_onesp@1)
321 (bit_not @0))
322
323/* x & ~0 -> x */
324(simplify
325 (bit_and @0 integer_all_onesp)
326 (non_lvalue @0))
327
328/* x & x -> x, x | x -> x */
329(for bitop (bit_and bit_ior)
330 (simplify
331 (bitop @0 @0)
332 (non_lvalue @0)))
333
0f770b01
RV
334/* x + (x & 1) -> (x + 1) & ~1 */
335(simplify
44fc0a51
RB
336 (plus:c @0 (bit_and:s @0 integer_onep@1))
337 (bit_and (plus @0 @1) (bit_not @1)))
0f770b01
RV
338
339/* x & ~(x & y) -> x & ~y */
340/* x | ~(x | y) -> x | ~y */
341(for bitop (bit_and bit_ior)
af563d4b 342 (simplify
44fc0a51
RB
343 (bitop:c @0 (bit_not (bitop:cs @0 @1)))
344 (bitop @0 (bit_not @1))))
af563d4b
MG
345
346/* (x | y) & ~x -> y & ~x */
347/* (x & y) | ~x -> y | ~x */
348(for bitop (bit_and bit_ior)
349 rbitop (bit_ior bit_and)
350 (simplify
351 (bitop:c (rbitop:c @0 @1) (bit_not@2 @0))
352 (bitop @1 @2)))
0f770b01 353
f13c4673
MP
354/* (x & y) ^ (x | y) -> x ^ y */
355(simplify
2d6f2dce
MP
356 (bit_xor:c (bit_and @0 @1) (bit_ior @0 @1))
357 (bit_xor @0 @1))
f13c4673 358
9ea65ca6
MP
359/* (x ^ y) ^ (x | y) -> x & y */
360(simplify
361 (bit_xor:c (bit_xor @0 @1) (bit_ior @0 @1))
362 (bit_and @0 @1))
363
364/* (x & y) + (x ^ y) -> x | y */
365/* (x & y) | (x ^ y) -> x | y */
366/* (x & y) ^ (x ^ y) -> x | y */
367(for op (plus bit_ior bit_xor)
368 (simplify
369 (op:c (bit_and @0 @1) (bit_xor @0 @1))
370 (bit_ior @0 @1)))
371
372/* (x & y) + (x | y) -> x + y */
373(simplify
374 (plus:c (bit_and @0 @1) (bit_ior @0 @1))
375 (plus @0 @1))
376
9737efaf
MP
377/* (x + y) - (x | y) -> x & y */
378(simplify
379 (minus (plus @0 @1) (bit_ior @0 @1))
380 (if (!TYPE_OVERFLOW_SANITIZED (type) && !TYPE_OVERFLOW_TRAPS (type)
381 && !TYPE_SATURATING (type))
382 (bit_and @0 @1)))
383
384/* (x + y) - (x & y) -> x | y */
385(simplify
386 (minus (plus @0 @1) (bit_and @0 @1))
387 (if (!TYPE_OVERFLOW_SANITIZED (type) && !TYPE_OVERFLOW_TRAPS (type)
388 && !TYPE_SATURATING (type))
389 (bit_ior @0 @1)))
390
9ea65ca6
MP
391/* (x | y) - (x ^ y) -> x & y */
392(simplify
393 (minus (bit_ior @0 @1) (bit_xor @0 @1))
394 (bit_and @0 @1))
395
396/* (x | y) - (x & y) -> x ^ y */
397(simplify
398 (minus (bit_ior @0 @1) (bit_and @0 @1))
399 (bit_xor @0 @1))
400
66cc6273
MP
401/* (x | y) & ~(x & y) -> x ^ y */
402(simplify
403 (bit_and:c (bit_ior @0 @1) (bit_not (bit_and @0 @1)))
404 (bit_xor @0 @1))
405
406/* (x | y) & (~x ^ y) -> x & y */
407(simplify
408 (bit_and:c (bit_ior:c @0 @1) (bit_xor:c @1 (bit_not @0)))
409 (bit_and @0 @1))
410
5b00d921
RB
411/* ~x & ~y -> ~(x | y)
412 ~x | ~y -> ~(x & y) */
413(for op (bit_and bit_ior)
414 rop (bit_ior bit_and)
415 (simplify
416 (op (convert1? (bit_not @0)) (convert2? (bit_not @1)))
417 (if (tree_nop_conversion_p (type, TREE_TYPE (@0))
418 && tree_nop_conversion_p (type, TREE_TYPE (@1)))
419 (bit_not (rop (convert @0) (convert @1))))))
420
14ea9f92 421/* If we are XORing or adding two BIT_AND_EXPR's, both of which are and'ing
5b00d921
RB
422 with a constant, and the two constants have no bits in common,
423 we should treat this as a BIT_IOR_EXPR since this may produce more
424 simplifications. */
14ea9f92
RB
425(for op (bit_xor plus)
426 (simplify
427 (op (convert1? (bit_and@4 @0 INTEGER_CST@1))
428 (convert2? (bit_and@5 @2 INTEGER_CST@3)))
429 (if (tree_nop_conversion_p (type, TREE_TYPE (@0))
430 && tree_nop_conversion_p (type, TREE_TYPE (@2))
431 && wi::bit_and (@1, @3) == 0)
432 (bit_ior (convert @4) (convert @5)))))
5b00d921
RB
433
434/* (X | Y) ^ X -> Y & ~ X*/
435(simplify
436 (bit_xor:c (convert? (bit_ior:c @0 @1)) (convert? @0))
437 (if (tree_nop_conversion_p (type, TREE_TYPE (@0)))
438 (convert (bit_and @1 (bit_not @0)))))
439
440/* Convert ~X ^ ~Y to X ^ Y. */
441(simplify
442 (bit_xor (convert1? (bit_not @0)) (convert2? (bit_not @1)))
443 (if (tree_nop_conversion_p (type, TREE_TYPE (@0))
444 && tree_nop_conversion_p (type, TREE_TYPE (@1)))
445 (bit_xor (convert @0) (convert @1))))
446
447/* Convert ~X ^ C to X ^ ~C. */
448(simplify
449 (bit_xor (convert? (bit_not @0)) INTEGER_CST@1)
c8ba6498
EB
450 (if (tree_nop_conversion_p (type, TREE_TYPE (@0)))
451 (bit_xor (convert @0) (bit_not @1))))
5b00d921 452
97e77391
RB
453/* Fold (X & Y) ^ Y as ~X & Y. */
454(simplify
455 (bit_xor:c (bit_and:c @0 @1) @1)
456 (bit_and (bit_not @0) @1))
457
14ea9f92
RB
458/* Given a bit-wise operation CODE applied to ARG0 and ARG1, see if both
459 operands are another bit-wise operation with a common input. If so,
460 distribute the bit operations to save an operation and possibly two if
461 constants are involved. For example, convert
462 (A | B) & (A | C) into A | (B & C)
463 Further simplification will occur if B and C are constants. */
464(for op (bit_and bit_ior)
465 rop (bit_ior bit_and)
466 (simplify
467 (op (convert? (rop:c @0 @1)) (convert? (rop @0 @2)))
468 (if (tree_nop_conversion_p (type, TREE_TYPE (@0)))
469 (rop (convert @0) (op (convert @1) (convert @2))))))
470
5b00d921 471
b14a9c57
RB
472(simplify
473 (abs (abs@1 @0))
474 @1)
f3582e54
RB
475(simplify
476 (abs (negate @0))
477 (abs @0))
478(simplify
479 (abs tree_expr_nonnegative_p@0)
480 @0)
481
55cf3946
RB
482/* A few cases of fold-const.c negate_expr_p predicate. */
483(match negate_expr_p
484 INTEGER_CST
b14a9c57
RB
485 (if ((INTEGRAL_TYPE_P (type)
486 && TYPE_OVERFLOW_WRAPS (type))
487 || (!TYPE_OVERFLOW_SANITIZED (type)
55cf3946
RB
488 && may_negate_without_overflow_p (t)))))
489(match negate_expr_p
490 FIXED_CST)
491(match negate_expr_p
492 (negate @0)
493 (if (!TYPE_OVERFLOW_SANITIZED (type))))
494(match negate_expr_p
495 REAL_CST
496 (if (REAL_VALUE_NEGATIVE (TREE_REAL_CST (t)))))
497/* VECTOR_CST handling of non-wrapping types would recurse in unsupported
498 ways. */
499(match negate_expr_p
500 VECTOR_CST
501 (if (FLOAT_TYPE_P (TREE_TYPE (type)) || TYPE_OVERFLOW_WRAPS (type))))
502
503/* -(A + B) -> (-B) - A. */
b14a9c57 504(simplify
55cf3946
RB
505 (negate (plus:c @0 negate_expr_p@1))
506 (if (!HONOR_SIGN_DEPENDENT_ROUNDING (element_mode (type))
507 && !HONOR_SIGNED_ZEROS (element_mode (type)))
508 (minus (negate @1) @0)))
509
510/* A - B -> A + (-B) if B is easily negatable. */
b14a9c57 511(simplify
55cf3946 512 (minus @0 negate_expr_p@1)
e4e96a4f
KT
513 (if (!FIXED_POINT_TYPE_P (type))
514 (plus @0 (negate @1))))
d4573ffe 515
5609420f
RB
516/* Try to fold (type) X op CST -> (type) (X op ((type-x) CST))
517 when profitable.
518 For bitwise binary operations apply operand conversions to the
519 binary operation result instead of to the operands. This allows
520 to combine successive conversions and bitwise binary operations.
521 We combine the above two cases by using a conditional convert. */
522(for bitop (bit_and bit_ior bit_xor)
523 (simplify
524 (bitop (convert @0) (convert? @1))
525 (if (((TREE_CODE (@1) == INTEGER_CST
526 && INTEGRAL_TYPE_P (TREE_TYPE (@0))
ad6f996c 527 && int_fits_type_p (@1, TREE_TYPE (@0)))
aea417d7 528 || types_match (@0, @1))
ad6f996c
RB
529 /* ??? This transform conflicts with fold-const.c doing
530 Convert (T)(x & c) into (T)x & (T)c, if c is an integer
531 constants (if x has signed type, the sign bit cannot be set
532 in c). This folds extension into the BIT_AND_EXPR.
533 Restrict it to GIMPLE to avoid endless recursions. */
534 && (bitop != BIT_AND_EXPR || GIMPLE)
5609420f
RB
535 && (/* That's a good idea if the conversion widens the operand, thus
536 after hoisting the conversion the operation will be narrower. */
537 TYPE_PRECISION (TREE_TYPE (@0)) < TYPE_PRECISION (type)
538 /* It's also a good idea if the conversion is to a non-integer
539 mode. */
540 || GET_MODE_CLASS (TYPE_MODE (type)) != MODE_INT
541 /* Or if the precision of TO is not the same as the precision
542 of its mode. */
543 || TYPE_PRECISION (type) != GET_MODE_PRECISION (TYPE_MODE (type))))
544 (convert (bitop @0 (convert @1))))))
545
b14a9c57
RB
546(for bitop (bit_and bit_ior)
547 rbitop (bit_ior bit_and)
548 /* (x | y) & x -> x */
549 /* (x & y) | x -> x */
550 (simplify
551 (bitop:c (rbitop:c @0 @1) @0)
552 @0)
553 /* (~x | y) & x -> x & y */
554 /* (~x & y) | x -> x | y */
555 (simplify
556 (bitop:c (rbitop:c (bit_not @0) @1) @0)
557 (bitop @0 @1)))
558
5609420f
RB
559/* Simplify (A & B) OP0 (C & B) to (A OP0 C) & B. */
560(for bitop (bit_and bit_ior bit_xor)
561 (simplify
562 (bitop (bit_and:c @0 @1) (bit_and @2 @1))
563 (bit_and (bitop @0 @2) @1)))
564
565/* (x | CST1) & CST2 -> (x & CST2) | (CST1 & CST2) */
566(simplify
567 (bit_and (bit_ior @0 CONSTANT_CLASS_P@1) CONSTANT_CLASS_P@2)
568 (bit_ior (bit_and @0 @2) (bit_and @1 @2)))
569
570/* Combine successive equal operations with constants. */
571(for bitop (bit_and bit_ior bit_xor)
572 (simplify
573 (bitop (bitop @0 CONSTANT_CLASS_P@1) CONSTANT_CLASS_P@2)
574 (bitop @0 (bitop @1 @2))))
575
576/* Try simple folding for X op !X, and X op X with the help
577 of the truth_valued_p and logical_inverted_value predicates. */
578(match truth_valued_p
579 @0
580 (if (INTEGRAL_TYPE_P (type) && TYPE_PRECISION (type) == 1)))
f84e7fd6 581(for op (tcc_comparison truth_and truth_andif truth_or truth_orif truth_xor)
5609420f
RB
582 (match truth_valued_p
583 (op @0 @1)))
584(match truth_valued_p
585 (truth_not @0))
586
587(match (logical_inverted_value @0)
588 (bit_not truth_valued_p@0))
589(match (logical_inverted_value @0)
09240451 590 (eq @0 integer_zerop))
5609420f 591(match (logical_inverted_value @0)
09240451 592 (ne truth_valued_p@0 integer_truep))
5609420f 593(match (logical_inverted_value @0)
09240451 594 (bit_xor truth_valued_p@0 integer_truep))
5609420f
RB
595
596/* X & !X -> 0. */
597(simplify
598 (bit_and:c @0 (logical_inverted_value @0))
599 { build_zero_cst (type); })
600/* X | !X and X ^ !X -> 1, , if X is truth-valued. */
601(for op (bit_ior bit_xor)
602 (simplify
603 (op:c truth_valued_p@0 (logical_inverted_value @0))
f84e7fd6 604 { constant_boolean_node (true, type); }))
5609420f 605
5609420f
RB
606/* If arg1 and arg2 are booleans (or any single bit type)
607 then try to simplify:
608
609 (~X & Y) -> X < Y
610 (X & ~Y) -> Y < X
611 (~X | Y) -> X <= Y
612 (X | ~Y) -> Y <= X
613
614 But only do this if our result feeds into a comparison as
615 this transformation is not always a win, particularly on
616 targets with and-not instructions.
617 -> simplify_bitwise_binary_boolean */
618(simplify
619 (ne (bit_and:c (bit_not @0) @1) integer_zerop)
620 (if (INTEGRAL_TYPE_P (TREE_TYPE (@1))
621 && TYPE_PRECISION (TREE_TYPE (@1)) == 1)
622 (lt @0 @1)))
623(simplify
624 (ne (bit_ior:c (bit_not @0) @1) integer_zerop)
625 (if (INTEGRAL_TYPE_P (TREE_TYPE (@1))
626 && TYPE_PRECISION (TREE_TYPE (@1)) == 1)
627 (le @0 @1)))
628
5609420f
RB
629/* ~~x -> x */
630(simplify
631 (bit_not (bit_not @0))
632 @0)
633
b14a9c57
RB
634/* Convert ~ (-A) to A - 1. */
635(simplify
636 (bit_not (convert? (negate @0)))
637 (if (tree_nop_conversion_p (type, TREE_TYPE (@0)))
8b5ee871 638 (convert (minus @0 { build_each_one_cst (TREE_TYPE (@0)); }))))
b14a9c57
RB
639
640/* Convert ~ (A - 1) or ~ (A + -1) to -A. */
641(simplify
8b5ee871 642 (bit_not (convert? (minus @0 integer_each_onep)))
b14a9c57
RB
643 (if (tree_nop_conversion_p (type, TREE_TYPE (@0)))
644 (convert (negate @0))))
645(simplify
646 (bit_not (convert? (plus @0 integer_all_onesp)))
647 (if (tree_nop_conversion_p (type, TREE_TYPE (@0)))
648 (convert (negate @0))))
649
650/* Part of convert ~(X ^ Y) to ~X ^ Y or X ^ ~Y if ~X or ~Y simplify. */
651(simplify
652 (bit_not (convert? (bit_xor @0 INTEGER_CST@1)))
653 (if (tree_nop_conversion_p (type, TREE_TYPE (@0)))
654 (convert (bit_xor @0 (bit_not @1)))))
655(simplify
656 (bit_not (convert? (bit_xor:c (bit_not @0) @1)))
657 (if (tree_nop_conversion_p (type, TREE_TYPE (@0)))
658 (convert (bit_xor @0 @1))))
659
f52baa7b
MP
660/* (x & ~m) | (y & m) -> ((x ^ y) & m) ^ x */
661(simplify
44fc0a51
RB
662 (bit_ior:c (bit_and:cs @0 (bit_not @2)) (bit_and:cs @1 @2))
663 (bit_xor (bit_and (bit_xor @0 @1) @2) @0))
f52baa7b 664
f7b7b0aa
MP
665/* Fold A - (A & B) into ~B & A. */
666(simplify
667 (minus (convert? @0) (convert?:s (bit_and:cs @0 @1)))
668 (if (tree_nop_conversion_p (type, TREE_TYPE (@0))
669 && tree_nop_conversion_p (type, TREE_TYPE (@1)))
670 (convert (bit_and (bit_not @1) @0))))
5609420f 671
a499aac5
RB
672/* Associate (p +p off1) +p off2 as (p +p (off1 + off2)). */
673(simplify
44fc0a51
RB
674 (pointer_plus (pointer_plus:s @0 @1) @3)
675 (pointer_plus @0 (plus @1 @3)))
a499aac5
RB
676
677/* Pattern match
678 tem1 = (long) ptr1;
679 tem2 = (long) ptr2;
680 tem3 = tem2 - tem1;
681 tem4 = (unsigned long) tem3;
682 tem5 = ptr1 + tem4;
683 and produce
684 tem5 = ptr2; */
685(simplify
686 (pointer_plus @0 (convert?@2 (minus@3 (convert @1) (convert @0))))
687 /* Conditionally look through a sign-changing conversion. */
688 (if (TYPE_PRECISION (TREE_TYPE (@2)) == TYPE_PRECISION (TREE_TYPE (@3))
689 && ((GIMPLE && useless_type_conversion_p (type, TREE_TYPE (@1)))
690 || (GENERIC && type == TREE_TYPE (@1))))
691 @1))
692
693/* Pattern match
694 tem = (sizetype) ptr;
695 tem = tem & algn;
696 tem = -tem;
697 ... = ptr p+ tem;
698 and produce the simpler and easier to analyze with respect to alignment
699 ... = ptr & ~algn; */
700(simplify
701 (pointer_plus @0 (negate (bit_and (convert @0) INTEGER_CST@1)))
702 (with { tree algn = wide_int_to_tree (TREE_TYPE (@0), wi::bit_not (@1)); }
703 (bit_and @0 { algn; })))
704
99e943a2
RB
705/* Try folding difference of addresses. */
706(simplify
707 (minus (convert ADDR_EXPR@0) (convert @1))
708 (if (tree_nop_conversion_p (type, TREE_TYPE (@0)))
709 (with { HOST_WIDE_INT diff; }
710 (if (ptr_difference_const (@0, @1, &diff))
711 { build_int_cst_type (type, diff); }))))
712(simplify
713 (minus (convert @0) (convert ADDR_EXPR@1))
714 (if (tree_nop_conversion_p (type, TREE_TYPE (@0)))
715 (with { HOST_WIDE_INT diff; }
716 (if (ptr_difference_const (@0, @1, &diff))
717 { build_int_cst_type (type, diff); }))))
718
bab73f11
RB
719/* If arg0 is derived from the address of an object or function, we may
720 be able to fold this expression using the object or function's
721 alignment. */
722(simplify
723 (bit_and (convert? @0) INTEGER_CST@1)
724 (if (POINTER_TYPE_P (TREE_TYPE (@0))
725 && tree_nop_conversion_p (type, TREE_TYPE (@0)))
726 (with
727 {
728 unsigned int align;
729 unsigned HOST_WIDE_INT bitpos;
730 get_pointer_alignment_1 (@0, &align, &bitpos);
731 }
732 (if (wi::ltu_p (@1, align / BITS_PER_UNIT))
733 { wide_int_to_tree (type, wi::bit_and (@1, bitpos / BITS_PER_UNIT)); }))))
99e943a2 734
a499aac5 735
cc7b5acf
RB
736/* We can't reassociate at all for saturating types. */
737(if (!TYPE_SATURATING (type))
738
739 /* Contract negates. */
740 /* A + (-B) -> A - B */
741 (simplify
742 (plus:c (convert1? @0) (convert2? (negate @1)))
743 /* Apply STRIP_NOPS on @0 and the negate. */
744 (if (tree_nop_conversion_p (type, TREE_TYPE (@0))
745 && tree_nop_conversion_p (type, TREE_TYPE (@1))
6a4f0678 746 && !TYPE_OVERFLOW_SANITIZED (type))
cc7b5acf
RB
747 (minus (convert @0) (convert @1))))
748 /* A - (-B) -> A + B */
749 (simplify
750 (minus (convert1? @0) (convert2? (negate @1)))
751 (if (tree_nop_conversion_p (type, TREE_TYPE (@0))
2f68e8bc 752 && tree_nop_conversion_p (type, TREE_TYPE (@1))
6a4f0678 753 && !TYPE_OVERFLOW_SANITIZED (type))
cc7b5acf
RB
754 (plus (convert @0) (convert @1))))
755 /* -(-A) -> A */
756 (simplify
757 (negate (convert? (negate @1)))
758 (if (tree_nop_conversion_p (type, TREE_TYPE (@1))
6a4f0678 759 && !TYPE_OVERFLOW_SANITIZED (type))
a0f12cf8 760 (convert @1)))
cc7b5acf 761
7318e44f
RB
762 /* We can't reassociate floating-point unless -fassociative-math
763 or fixed-point plus or minus because of saturation to +-Inf. */
764 (if ((!FLOAT_TYPE_P (type) || flag_associative_math)
765 && !FIXED_POINT_TYPE_P (type))
cc7b5acf
RB
766
767 /* Match patterns that allow contracting a plus-minus pair
768 irrespective of overflow issues. */
769 /* (A +- B) - A -> +- B */
770 /* (A +- B) -+ B -> A */
771 /* A - (A +- B) -> -+ B */
772 /* A +- (B -+ A) -> +- B */
773 (simplify
774 (minus (plus:c @0 @1) @0)
775 @1)
776 (simplify
777 (minus (minus @0 @1) @0)
778 (negate @1))
779 (simplify
780 (plus:c (minus @0 @1) @1)
781 @0)
782 (simplify
783 (minus @0 (plus:c @0 @1))
784 (negate @1))
785 (simplify
786 (minus @0 (minus @0 @1))
787 @1)
788
789 /* (A +- CST) +- CST -> A + CST */
790 (for outer_op (plus minus)
791 (for inner_op (plus minus)
792 (simplify
793 (outer_op (inner_op @0 CONSTANT_CLASS_P@1) CONSTANT_CLASS_P@2)
794 /* If the constant operation overflows we cannot do the transform
795 as we would introduce undefined overflow, for example
796 with (a - 1) + INT_MIN. */
797 (with { tree cst = fold_binary (outer_op == inner_op
798 ? PLUS_EXPR : MINUS_EXPR, type, @1, @2); }
799 (if (cst && !TREE_OVERFLOW (cst))
800 (inner_op @0 { cst; } ))))))
801
802 /* (CST - A) +- CST -> CST - A */
803 (for outer_op (plus minus)
804 (simplify
805 (outer_op (minus CONSTANT_CLASS_P@1 @0) CONSTANT_CLASS_P@2)
806 (with { tree cst = fold_binary (outer_op, type, @1, @2); }
807 (if (cst && !TREE_OVERFLOW (cst))
808 (minus { cst; } @0)))))
809
810 /* ~A + A -> -1 */
811 (simplify
812 (plus:c (bit_not @0) @0)
813 (if (!TYPE_OVERFLOW_TRAPS (type))
814 { build_all_ones_cst (type); }))
815
816 /* ~A + 1 -> -A */
817 (simplify
e19740ae
RB
818 (plus (convert? (bit_not @0)) integer_each_onep)
819 (if (tree_nop_conversion_p (type, TREE_TYPE (@0)))
820 (negate (convert @0))))
821
822 /* -A - 1 -> ~A */
823 (simplify
824 (minus (convert? (negate @0)) integer_each_onep)
825 (if (!TYPE_OVERFLOW_TRAPS (type)
826 && tree_nop_conversion_p (type, TREE_TYPE (@0)))
827 (bit_not (convert @0))))
828
829 /* -1 - A -> ~A */
830 (simplify
831 (minus integer_all_onesp @0)
bc4315fb 832 (bit_not @0))
cc7b5acf
RB
833
834 /* (T)(P + A) - (T)P -> (T) A */
835 (for add (plus pointer_plus)
836 (simplify
837 (minus (convert (add @0 @1))
838 (convert @0))
09240451 839 (if (element_precision (type) <= element_precision (TREE_TYPE (@1))
cc7b5acf
RB
840 /* For integer types, if A has a smaller type
841 than T the result depends on the possible
842 overflow in P + A.
843 E.g. T=size_t, A=(unsigned)429497295, P>0.
844 However, if an overflow in P + A would cause
845 undefined behavior, we can assume that there
846 is no overflow. */
847 || (INTEGRAL_TYPE_P (TREE_TYPE (@0))
848 && TYPE_OVERFLOW_UNDEFINED (TREE_TYPE (@0)))
849 /* For pointer types, if the conversion of A to the
850 final type requires a sign- or zero-extension,
851 then we have to punt - it is not defined which
852 one is correct. */
853 || (POINTER_TYPE_P (TREE_TYPE (@0))
854 && TREE_CODE (@1) == INTEGER_CST
855 && tree_int_cst_sign_bit (@1) == 0))
856 (convert @1))))))
857
858
a7f24614
RB
859/* Simplifications of MIN_EXPR and MAX_EXPR. */
860
861(for minmax (min max)
862 (simplify
863 (minmax @0 @0)
864 @0))
865(simplify
866 (min @0 @1)
867 (if (INTEGRAL_TYPE_P (type)
868 && TYPE_MIN_VALUE (type)
869 && operand_equal_p (@1, TYPE_MIN_VALUE (type), OEP_ONLY_CONST))
870 @1))
871(simplify
872 (max @0 @1)
873 (if (INTEGRAL_TYPE_P (type)
874 && TYPE_MAX_VALUE (type)
875 && operand_equal_p (@1, TYPE_MAX_VALUE (type), OEP_ONLY_CONST))
876 @1))
877
878
879/* Simplifications of shift and rotates. */
880
881(for rotate (lrotate rrotate)
882 (simplify
883 (rotate integer_all_onesp@0 @1)
884 @0))
885
886/* Optimize -1 >> x for arithmetic right shifts. */
887(simplify
888 (rshift integer_all_onesp@0 @1)
889 (if (!TYPE_UNSIGNED (type)
890 && tree_expr_nonnegative_p (@1))
891 @0))
892
893(for shiftrotate (lrotate rrotate lshift rshift)
894 (simplify
895 (shiftrotate @0 integer_zerop)
896 (non_lvalue @0))
897 (simplify
898 (shiftrotate integer_zerop@0 @1)
899 @0)
900 /* Prefer vector1 << scalar to vector1 << vector2
901 if vector2 is uniform. */
902 (for vec (VECTOR_CST CONSTRUCTOR)
903 (simplify
904 (shiftrotate @0 vec@1)
905 (with { tree tem = uniform_vector_p (@1); }
906 (if (tem)
907 (shiftrotate @0 { tem; }))))))
908
909/* Rewrite an LROTATE_EXPR by a constant into an
910 RROTATE_EXPR by a new constant. */
911(simplify
912 (lrotate @0 INTEGER_CST@1)
913 (rrotate @0 { fold_binary (MINUS_EXPR, TREE_TYPE (@1),
914 build_int_cst (TREE_TYPE (@1),
915 element_precision (type)), @1); }))
916
14ea9f92
RB
917/* Turn (a OP c1) OP c2 into a OP (c1+c2). */
918(for op (lrotate rrotate rshift lshift)
919 (simplify
920 (op (op @0 INTEGER_CST@1) INTEGER_CST@2)
921 (with { unsigned int prec = element_precision (type); }
922 (if (wi::ge_p (@1, 0, TYPE_SIGN (TREE_TYPE (@1)))
923 && wi::lt_p (@1, prec, TYPE_SIGN (TREE_TYPE (@1)))
924 && wi::ge_p (@2, 0, TYPE_SIGN (TREE_TYPE (@2)))
925 && wi::lt_p (@2, prec, TYPE_SIGN (TREE_TYPE (@2))))
926 (with { unsigned int low = wi::add (@1, @2).to_uhwi (); }
927 /* Deal with a OP (c1 + c2) being undefined but (a OP c1) OP c2
928 being well defined. */
929 (if (low >= prec)
930 (if (op == LROTATE_EXPR || op == RROTATE_EXPR)
8fdc6c67
RB
931 (op @0 { build_int_cst (TREE_TYPE (@1), low % prec); })
932 (if (TYPE_UNSIGNED (type) || code == LSHIFT_EXPR)
933 { build_zero_cst (type); }
934 (op @0 { build_int_cst (TREE_TYPE (@1), prec - 1); })))
935 (op @0 { build_int_cst (TREE_TYPE (@1), low); })))))))
14ea9f92
RB
936
937
01ada710
MP
938/* ((1 << A) & 1) != 0 -> A == 0
939 ((1 << A) & 1) == 0 -> A != 0 */
940(for cmp (ne eq)
941 icmp (eq ne)
942 (simplify
943 (cmp (bit_and (lshift integer_onep @0) integer_onep) integer_zerop)
944 (icmp @0 { build_zero_cst (TREE_TYPE (@0)); })))
cc7b5acf 945
f2e609c3
MP
946/* (CST1 << A) == CST2 -> A == ctz (CST2) - ctz (CST1)
947 (CST1 << A) != CST2 -> A != ctz (CST2) - ctz (CST1)
948 if CST2 != 0. */
949(for cmp (ne eq)
950 (simplify
951 (cmp (lshift INTEGER_CST@0 @1) INTEGER_CST@2)
952 (with { int cand = wi::ctz (@2) - wi::ctz (@0); }
953 (if (cand < 0
954 || (!integer_zerop (@2)
955 && wi::ne_p (wi::lshift (@0, cand), @2)))
8fdc6c67
RB
956 { constant_boolean_node (cmp == NE_EXPR, type); }
957 (if (!integer_zerop (@2)
958 && wi::eq_p (wi::lshift (@0, cand), @2))
959 (cmp @1 { build_int_cst (TREE_TYPE (@1), cand); }))))))
f2e609c3 960
1ffbaa3f
RB
961/* Fold (X << C1) & C2 into (X << C1) & (C2 | ((1 << C1) - 1))
962 (X >> C1) & C2 into (X >> C1) & (C2 | ~((type) -1 >> C1))
963 if the new mask might be further optimized. */
964(for shift (lshift rshift)
965 (simplify
44fc0a51
RB
966 (bit_and (convert?:s@4 (shift:s@5 (convert1?@3 @0) INTEGER_CST@1))
967 INTEGER_CST@2)
1ffbaa3f
RB
968 (if (tree_nop_conversion_p (TREE_TYPE (@4), TREE_TYPE (@5))
969 && TYPE_PRECISION (type) <= HOST_BITS_PER_WIDE_INT
970 && tree_fits_uhwi_p (@1)
971 && tree_to_uhwi (@1) > 0
972 && tree_to_uhwi (@1) < TYPE_PRECISION (type))
973 (with
974 {
975 unsigned int shiftc = tree_to_uhwi (@1);
976 unsigned HOST_WIDE_INT mask = TREE_INT_CST_LOW (@2);
977 unsigned HOST_WIDE_INT newmask, zerobits = 0;
978 tree shift_type = TREE_TYPE (@3);
979 unsigned int prec;
980
981 if (shift == LSHIFT_EXPR)
982 zerobits = ((((unsigned HOST_WIDE_INT) 1) << shiftc) - 1);
983 else if (shift == RSHIFT_EXPR
984 && (TYPE_PRECISION (shift_type)
985 == GET_MODE_PRECISION (TYPE_MODE (shift_type))))
986 {
987 prec = TYPE_PRECISION (TREE_TYPE (@3));
988 tree arg00 = @0;
989 /* See if more bits can be proven as zero because of
990 zero extension. */
991 if (@3 != @0
992 && TYPE_UNSIGNED (TREE_TYPE (@0)))
993 {
994 tree inner_type = TREE_TYPE (@0);
995 if ((TYPE_PRECISION (inner_type)
996 == GET_MODE_PRECISION (TYPE_MODE (inner_type)))
997 && TYPE_PRECISION (inner_type) < prec)
998 {
999 prec = TYPE_PRECISION (inner_type);
1000 /* See if we can shorten the right shift. */
1001 if (shiftc < prec)
1002 shift_type = inner_type;
1003 /* Otherwise X >> C1 is all zeros, so we'll optimize
1004 it into (X, 0) later on by making sure zerobits
1005 is all ones. */
1006 }
1007 }
1008 zerobits = ~(unsigned HOST_WIDE_INT) 0;
1009 if (shiftc < prec)
1010 {
1011 zerobits >>= HOST_BITS_PER_WIDE_INT - shiftc;
1012 zerobits <<= prec - shiftc;
1013 }
1014 /* For arithmetic shift if sign bit could be set, zerobits
1015 can contain actually sign bits, so no transformation is
1016 possible, unless MASK masks them all away. In that
1017 case the shift needs to be converted into logical shift. */
1018 if (!TYPE_UNSIGNED (TREE_TYPE (@3))
1019 && prec == TYPE_PRECISION (TREE_TYPE (@3)))
1020 {
1021 if ((mask & zerobits) == 0)
1022 shift_type = unsigned_type_for (TREE_TYPE (@3));
1023 else
1024 zerobits = 0;
1025 }
1026 }
1027 }
1028 /* ((X << 16) & 0xff00) is (X, 0). */
1029 (if ((mask & zerobits) == mask)
8fdc6c67
RB
1030 { build_int_cst (type, 0); }
1031 (with { newmask = mask | zerobits; }
1032 (if (newmask != mask && (newmask & (newmask + 1)) == 0)
1033 (with
1034 {
1035 /* Only do the transformation if NEWMASK is some integer
1036 mode's mask. */
1037 for (prec = BITS_PER_UNIT;
1038 prec < HOST_BITS_PER_WIDE_INT; prec <<= 1)
1039 if (newmask == (((unsigned HOST_WIDE_INT) 1) << prec) - 1)
1040 break;
1041 }
1042 (if (prec < HOST_BITS_PER_WIDE_INT
1043 || newmask == ~(unsigned HOST_WIDE_INT) 0)
1044 (with
1045 { tree newmaskt = build_int_cst_type (TREE_TYPE (@2), newmask); }
1046 (if (!tree_int_cst_equal (newmaskt, @2))
1047 (if (shift_type != TREE_TYPE (@3))
1048 (bit_and (convert (shift:shift_type (convert @3) @1)) { newmaskt; })
1049 (bit_and @4 { newmaskt; })))))))))))))
1ffbaa3f 1050
98e30e51
RB
1051/* Fold (X & C2) << C1 into (X << C1) & (C2 << C1)
1052 (X & C2) >> C1 into (X >> C1) & (C2 >> C1). */
1053(for shift (lshift rshift)
1054 (simplify
2d799646 1055 (shift (convert?:s (bit_and:s @0 INTEGER_CST@2)) INTEGER_CST@1)
98e30e51
RB
1056 (if (tree_nop_conversion_p (type, TREE_TYPE (@0)))
1057 (with { tree mask = int_const_binop (shift, fold_convert (type, @2), @1); }
1058 (bit_and (shift (convert @0) @1) { mask; })))))
1059
1060
d4573ffe
RB
1061/* Simplifications of conversions. */
1062
1063/* Basic strip-useless-type-conversions / strip_nops. */
f3582e54 1064(for cvt (convert view_convert float fix_trunc)
d4573ffe
RB
1065 (simplify
1066 (cvt @0)
1067 (if ((GIMPLE && useless_type_conversion_p (type, TREE_TYPE (@0)))
1068 || (GENERIC && type == TREE_TYPE (@0)))
1069 @0)))
1070
1071/* Contract view-conversions. */
1072(simplify
1073 (view_convert (view_convert @0))
1074 (view_convert @0))
1075
1076/* For integral conversions with the same precision or pointer
1077 conversions use a NOP_EXPR instead. */
1078(simplify
1079 (view_convert @0)
1080 (if ((INTEGRAL_TYPE_P (type) || POINTER_TYPE_P (type))
1081 && (INTEGRAL_TYPE_P (TREE_TYPE (@0)) || POINTER_TYPE_P (TREE_TYPE (@0)))
1082 && TYPE_PRECISION (type) == TYPE_PRECISION (TREE_TYPE (@0)))
1083 (convert @0)))
1084
1085/* Strip inner integral conversions that do not change precision or size. */
1086(simplify
1087 (view_convert (convert@0 @1))
1088 (if ((INTEGRAL_TYPE_P (TREE_TYPE (@0)) || POINTER_TYPE_P (TREE_TYPE (@0)))
1089 && (INTEGRAL_TYPE_P (TREE_TYPE (@1)) || POINTER_TYPE_P (TREE_TYPE (@1)))
1090 && (TYPE_PRECISION (TREE_TYPE (@0)) == TYPE_PRECISION (TREE_TYPE (@1)))
1091 && (TYPE_SIZE (TREE_TYPE (@0)) == TYPE_SIZE (TREE_TYPE (@1))))
1092 (view_convert @1)))
1093
1094/* Re-association barriers around constants and other re-association
1095 barriers can be removed. */
1096(simplify
1097 (paren CONSTANT_CLASS_P@0)
1098 @0)
1099(simplify
1100 (paren (paren@1 @0))
1101 @1)
1e51d0a2
RB
1102
1103/* Handle cases of two conversions in a row. */
1104(for ocvt (convert float fix_trunc)
1105 (for icvt (convert float)
1106 (simplify
1107 (ocvt (icvt@1 @0))
1108 (with
1109 {
1110 tree inside_type = TREE_TYPE (@0);
1111 tree inter_type = TREE_TYPE (@1);
1112 int inside_int = INTEGRAL_TYPE_P (inside_type);
1113 int inside_ptr = POINTER_TYPE_P (inside_type);
1114 int inside_float = FLOAT_TYPE_P (inside_type);
09240451 1115 int inside_vec = VECTOR_TYPE_P (inside_type);
1e51d0a2
RB
1116 unsigned int inside_prec = TYPE_PRECISION (inside_type);
1117 int inside_unsignedp = TYPE_UNSIGNED (inside_type);
1118 int inter_int = INTEGRAL_TYPE_P (inter_type);
1119 int inter_ptr = POINTER_TYPE_P (inter_type);
1120 int inter_float = FLOAT_TYPE_P (inter_type);
09240451 1121 int inter_vec = VECTOR_TYPE_P (inter_type);
1e51d0a2
RB
1122 unsigned int inter_prec = TYPE_PRECISION (inter_type);
1123 int inter_unsignedp = TYPE_UNSIGNED (inter_type);
1124 int final_int = INTEGRAL_TYPE_P (type);
1125 int final_ptr = POINTER_TYPE_P (type);
1126 int final_float = FLOAT_TYPE_P (type);
09240451 1127 int final_vec = VECTOR_TYPE_P (type);
1e51d0a2
RB
1128 unsigned int final_prec = TYPE_PRECISION (type);
1129 int final_unsignedp = TYPE_UNSIGNED (type);
1130 }
64d3a1f0
RB
1131 (switch
1132 /* In addition to the cases of two conversions in a row
1133 handled below, if we are converting something to its own
1134 type via an object of identical or wider precision, neither
1135 conversion is needed. */
1136 (if (((GIMPLE && useless_type_conversion_p (type, inside_type))
1137 || (GENERIC
1138 && TYPE_MAIN_VARIANT (type) == TYPE_MAIN_VARIANT (inside_type)))
1139 && (((inter_int || inter_ptr) && final_int)
1140 || (inter_float && final_float))
1141 && inter_prec >= final_prec)
1142 (ocvt @0))
1143
1144 /* Likewise, if the intermediate and initial types are either both
1145 float or both integer, we don't need the middle conversion if the
1146 former is wider than the latter and doesn't change the signedness
1147 (for integers). Avoid this if the final type is a pointer since
1148 then we sometimes need the middle conversion. Likewise if the
1149 final type has a precision not equal to the size of its mode. */
1150 (if (((inter_int && inside_int) || (inter_float && inside_float))
1151 && (final_int || final_float)
1152 && inter_prec >= inside_prec
1153 && (inter_float || inter_unsignedp == inside_unsignedp)
1154 && ! (final_prec != GET_MODE_PRECISION (TYPE_MODE (type))
1155 && TYPE_MODE (type) == TYPE_MODE (inter_type)))
1156 (ocvt @0))
1157
1158 /* If we have a sign-extension of a zero-extended value, we can
1159 replace that by a single zero-extension. Likewise if the
1160 final conversion does not change precision we can drop the
1161 intermediate conversion. */
1162 (if (inside_int && inter_int && final_int
1163 && ((inside_prec < inter_prec && inter_prec < final_prec
1164 && inside_unsignedp && !inter_unsignedp)
1165 || final_prec == inter_prec))
1166 (ocvt @0))
1167
1168 /* Two conversions in a row are not needed unless:
1e51d0a2
RB
1169 - some conversion is floating-point (overstrict for now), or
1170 - some conversion is a vector (overstrict for now), or
1171 - the intermediate type is narrower than both initial and
1172 final, or
1173 - the intermediate type and innermost type differ in signedness,
1174 and the outermost type is wider than the intermediate, or
1175 - the initial type is a pointer type and the precisions of the
1176 intermediate and final types differ, or
1177 - the final type is a pointer type and the precisions of the
1178 initial and intermediate types differ. */
64d3a1f0
RB
1179 (if (! inside_float && ! inter_float && ! final_float
1180 && ! inside_vec && ! inter_vec && ! final_vec
1181 && (inter_prec >= inside_prec || inter_prec >= final_prec)
1182 && ! (inside_int && inter_int
1183 && inter_unsignedp != inside_unsignedp
1184 && inter_prec < final_prec)
1185 && ((inter_unsignedp && inter_prec > inside_prec)
1186 == (final_unsignedp && final_prec > inter_prec))
1187 && ! (inside_ptr && inter_prec != final_prec)
1188 && ! (final_ptr && inside_prec != inter_prec)
1189 && ! (final_prec != GET_MODE_PRECISION (TYPE_MODE (type))
1190 && TYPE_MODE (type) == TYPE_MODE (inter_type)))
1191 (ocvt @0))
1192
1193 /* A truncation to an unsigned type (a zero-extension) should be
1194 canonicalized as bitwise and of a mask. */
1195 (if (final_int && inter_int && inside_int
1196 && final_prec == inside_prec
1197 && final_prec > inter_prec
1198 && inter_unsignedp)
1199 (convert (bit_and @0 { wide_int_to_tree
1200 (inside_type,
1201 wi::mask (inter_prec, false,
1202 TYPE_PRECISION (inside_type))); })))
1203
1204 /* If we are converting an integer to a floating-point that can
1205 represent it exactly and back to an integer, we can skip the
1206 floating-point conversion. */
1207 (if (GIMPLE /* PR66211 */
1208 && inside_int && inter_float && final_int &&
1209 (unsigned) significand_size (TYPE_MODE (inter_type))
1210 >= inside_prec - !inside_unsignedp)
1211 (convert @0)))))))
ea2042ba
RB
1212
1213/* If we have a narrowing conversion to an integral type that is fed by a
1214 BIT_AND_EXPR, we might be able to remove the BIT_AND_EXPR if it merely
1215 masks off bits outside the final type (and nothing else). */
1216(simplify
1217 (convert (bit_and @0 INTEGER_CST@1))
1218 (if (INTEGRAL_TYPE_P (type)
1219 && INTEGRAL_TYPE_P (TREE_TYPE (@0))
1220 && TYPE_PRECISION (type) <= TYPE_PRECISION (TREE_TYPE (@0))
1221 && operand_equal_p (@1, build_low_bits_mask (TREE_TYPE (@1),
1222 TYPE_PRECISION (type)), 0))
1223 (convert @0)))
a25454ea
RB
1224
1225
1226/* (X /[ex] A) * A -> X. */
1227(simplify
1228 (mult (convert? (exact_div @0 @1)) @1)
1229 /* Look through a sign-changing conversion. */
257b01ba 1230 (convert @0))
eaeba53a 1231
a7f24614
RB
1232/* Canonicalization of binary operations. */
1233
1234/* Convert X + -C into X - C. */
1235(simplify
1236 (plus @0 REAL_CST@1)
1237 (if (REAL_VALUE_NEGATIVE (TREE_REAL_CST (@1)))
1238 (with { tree tem = fold_unary (NEGATE_EXPR, type, @1); }
1239 (if (!TREE_OVERFLOW (tem) || !flag_trapping_math)
1240 (minus @0 { tem; })))))
1241
1242/* Convert x+x into x*2.0. */
1243(simplify
1244 (plus @0 @0)
1245 (if (SCALAR_FLOAT_TYPE_P (type))
1246 (mult @0 { build_real (type, dconst2); })))
1247
1248(simplify
1249 (minus integer_zerop @1)
1250 (negate @1))
1251
1252/* (ARG0 - ARG1) is the same as (-ARG1 + ARG0). So check whether
1253 ARG0 is zero and X + ARG0 reduces to X, since that would mean
1254 (-ARG1 + ARG0) reduces to -ARG1. */
1255(simplify
1256 (minus real_zerop@0 @1)
1257 (if (fold_real_zero_addition_p (type, @0, 0))
1258 (negate @1)))
1259
1260/* Transform x * -1 into -x. */
1261(simplify
1262 (mult @0 integer_minus_onep)
1263 (negate @0))
eaeba53a
RB
1264
1265/* COMPLEX_EXPR and REALPART/IMAGPART_EXPR cancellations. */
1266(simplify
1267 (complex (realpart @0) (imagpart @0))
1268 @0)
1269(simplify
1270 (realpart (complex @0 @1))
1271 @0)
1272(simplify
1273 (imagpart (complex @0 @1))
1274 @1)
83633539
RB
1275
1276
1277/* BSWAP simplifications, transforms checked by gcc.dg/builtin-bswap-8.c. */
1278(for bswap (BUILT_IN_BSWAP16 BUILT_IN_BSWAP32 BUILT_IN_BSWAP64)
1279 (simplify
1280 (bswap (bswap @0))
1281 @0)
1282 (simplify
1283 (bswap (bit_not (bswap @0)))
1284 (bit_not @0))
1285 (for bitop (bit_xor bit_ior bit_and)
1286 (simplify
1287 (bswap (bitop:c (bswap @0) @1))
1288 (bitop @0 (bswap @1)))))
96994de0
RB
1289
1290
1291/* Combine COND_EXPRs and VEC_COND_EXPRs. */
1292
1293/* Simplify constant conditions.
1294 Only optimize constant conditions when the selected branch
1295 has the same type as the COND_EXPR. This avoids optimizing
1296 away "c ? x : throw", where the throw has a void type.
1297 Note that we cannot throw away the fold-const.c variant nor
1298 this one as we depend on doing this transform before possibly
1299 A ? B : B -> B triggers and the fold-const.c one can optimize
1300 0 ? A : B to B even if A has side-effects. Something
1301 genmatch cannot handle. */
1302(simplify
1303 (cond INTEGER_CST@0 @1 @2)
8fdc6c67
RB
1304 (if (integer_zerop (@0))
1305 (if (!VOID_TYPE_P (TREE_TYPE (@2)) || VOID_TYPE_P (type))
1306 @2)
1307 (if (!VOID_TYPE_P (TREE_TYPE (@1)) || VOID_TYPE_P (type))
1308 @1)))
96994de0
RB
1309(simplify
1310 (vec_cond VECTOR_CST@0 @1 @2)
1311 (if (integer_all_onesp (@0))
8fdc6c67
RB
1312 @1
1313 (if (integer_zerop (@0))
1314 @2)))
96994de0
RB
1315
1316(for cnd (cond vec_cond)
1317 /* A ? B : (A ? X : C) -> A ? B : C. */
1318 (simplify
1319 (cnd @0 (cnd @0 @1 @2) @3)
1320 (cnd @0 @1 @3))
1321 (simplify
1322 (cnd @0 @1 (cnd @0 @2 @3))
1323 (cnd @0 @1 @3))
1324
1325 /* A ? B : B -> B. */
1326 (simplify
1327 (cnd @0 @1 @1)
09240451 1328 @1)
96994de0 1329
09240451
MG
1330 /* !A ? B : C -> A ? C : B. */
1331 (simplify
1332 (cnd (logical_inverted_value truth_valued_p@0) @1 @2)
1333 (cnd @0 @2 @1)))
f84e7fd6 1334
f43d102e
RS
1335/* A + (B vcmp C ? 1 : 0) -> A - (B vcmp C), since vector comparisons
1336 return all-1 or all-0 results. */
1337/* ??? We could instead convert all instances of the vec_cond to negate,
1338 but that isn't necessarily a win on its own. */
1339(simplify
1340 (plus:c @3 (view_convert? (vec_cond @0 integer_each_onep@1 integer_zerop@2)))
1341 (if (VECTOR_TYPE_P (type)
1342 && TYPE_VECTOR_SUBPARTS (type) == TYPE_VECTOR_SUBPARTS (TREE_TYPE (@0))
1343 && (TYPE_MODE (TREE_TYPE (type))
1344 == TYPE_MODE (TREE_TYPE (TREE_TYPE (@0)))))
1345 (minus @3 (view_convert @0))))
1346
1347/* ... likewise A - (B vcmp C ? 1 : 0) -> A + (B vcmp C). */
1348(simplify
1349 (minus @3 (view_convert? (vec_cond @0 integer_each_onep@1 integer_zerop@2)))
1350 (if (VECTOR_TYPE_P (type)
1351 && TYPE_VECTOR_SUBPARTS (type) == TYPE_VECTOR_SUBPARTS (TREE_TYPE (@0))
1352 && (TYPE_MODE (TREE_TYPE (type))
1353 == TYPE_MODE (TREE_TYPE (TREE_TYPE (@0)))))
1354 (plus @3 (view_convert @0))))
f84e7fd6 1355
2ee05f1e 1356
f84e7fd6
RB
1357/* Simplifications of comparisons. */
1358
1359/* We can simplify a logical negation of a comparison to the
1360 inverted comparison. As we cannot compute an expression
1361 operator using invert_tree_comparison we have to simulate
1362 that with expression code iteration. */
1363(for cmp (tcc_comparison)
1364 icmp (inverted_tcc_comparison)
1365 ncmp (inverted_tcc_comparison_with_nans)
1366 /* Ideally we'd like to combine the following two patterns
1367 and handle some more cases by using
1368 (logical_inverted_value (cmp @0 @1))
1369 here but for that genmatch would need to "inline" that.
1370 For now implement what forward_propagate_comparison did. */
1371 (simplify
1372 (bit_not (cmp @0 @1))
1373 (if (VECTOR_TYPE_P (type)
1374 || (INTEGRAL_TYPE_P (type) && TYPE_PRECISION (type) == 1))
1375 /* Comparison inversion may be impossible for trapping math,
1376 invert_tree_comparison will tell us. But we can't use
1377 a computed operator in the replacement tree thus we have
1378 to play the trick below. */
1379 (with { enum tree_code ic = invert_tree_comparison
1b457aa4 1380 (cmp, HONOR_NANS (@0)); }
f84e7fd6 1381 (if (ic == icmp)
8fdc6c67
RB
1382 (icmp @0 @1)
1383 (if (ic == ncmp)
1384 (ncmp @0 @1))))))
f84e7fd6 1385 (simplify
09240451
MG
1386 (bit_xor (cmp @0 @1) integer_truep)
1387 (with { enum tree_code ic = invert_tree_comparison
1b457aa4 1388 (cmp, HONOR_NANS (@0)); }
09240451 1389 (if (ic == icmp)
8fdc6c67
RB
1390 (icmp @0 @1)
1391 (if (ic == ncmp)
1392 (ncmp @0 @1))))))
e18c1d66 1393
2ee05f1e
RB
1394/* Transform comparisons of the form X - Y CMP 0 to X CMP Y.
1395 ??? The transformation is valid for the other operators if overflow
1396 is undefined for the type, but performing it here badly interacts
1397 with the transformation in fold_cond_expr_with_comparison which
1398 attempts to synthetize ABS_EXPR. */
1399(for cmp (eq ne)
1400 (simplify
d9ba1961
RB
1401 (cmp (minus@2 @0 @1) integer_zerop)
1402 (if (single_use (@2))
1403 (cmp @0 @1))))
2ee05f1e
RB
1404
1405/* Transform comparisons of the form X * C1 CMP 0 to X CMP 0 in the
1406 signed arithmetic case. That form is created by the compiler
1407 often enough for folding it to be of value. One example is in
1408 computing loop trip counts after Operator Strength Reduction. */
07cdc2b8
RB
1409(for cmp (simple_comparison)
1410 scmp (swapped_simple_comparison)
2ee05f1e
RB
1411 (simplify
1412 (cmp (mult @0 INTEGER_CST@1) integer_zerop@2)
1413 /* Handle unfolded multiplication by zero. */
1414 (if (integer_zerop (@1))
8fdc6c67
RB
1415 (cmp @1 @2)
1416 (if (ANY_INTEGRAL_TYPE_P (TREE_TYPE (@0))
1417 && TYPE_OVERFLOW_UNDEFINED (TREE_TYPE (@0)))
1418 /* If @1 is negative we swap the sense of the comparison. */
1419 (if (tree_int_cst_sgn (@1) < 0)
1420 (scmp @0 @2)
1421 (cmp @0 @2))))))
2ee05f1e
RB
1422
1423/* Simplify comparison of something with itself. For IEEE
1424 floating-point, we can only do some of these simplifications. */
1425(simplify
1426 (eq @0 @0)
1427 (if (! FLOAT_TYPE_P (TREE_TYPE (@0))
1428 || ! HONOR_NANS (TYPE_MODE (TREE_TYPE (@0))))
1429 { constant_boolean_node (true, type); }))
1430(for cmp (ge le)
1431 (simplify
1432 (cmp @0 @0)
1433 (eq @0 @0)))
1434(for cmp (ne gt lt)
1435 (simplify
1436 (cmp @0 @0)
1437 (if (cmp != NE_EXPR
1438 || ! FLOAT_TYPE_P (TREE_TYPE (@0))
1439 || ! HONOR_NANS (TYPE_MODE (TREE_TYPE (@0))))
1440 { constant_boolean_node (false, type); })))
1441
1442/* Fold ~X op ~Y as Y op X. */
07cdc2b8 1443(for cmp (simple_comparison)
2ee05f1e
RB
1444 (simplify
1445 (cmp (bit_not @0) (bit_not @1))
1446 (cmp @1 @0)))
1447
1448/* Fold ~X op C as X op' ~C, where op' is the swapped comparison. */
07cdc2b8
RB
1449(for cmp (simple_comparison)
1450 scmp (swapped_simple_comparison)
2ee05f1e
RB
1451 (simplify
1452 (cmp (bit_not @0) CONSTANT_CLASS_P@1)
1453 (if (TREE_CODE (@1) == INTEGER_CST || TREE_CODE (@1) == VECTOR_CST)
1454 (scmp @0 (bit_not @1)))))
1455
07cdc2b8
RB
1456(for cmp (simple_comparison)
1457 /* Fold (double)float1 CMP (double)float2 into float1 CMP float2. */
1458 (simplify
1459 (cmp (convert@2 @0) (convert? @1))
1460 (if (FLOAT_TYPE_P (TREE_TYPE (@0))
1461 && (DECIMAL_FLOAT_TYPE_P (TREE_TYPE (@2))
1462 == DECIMAL_FLOAT_TYPE_P (TREE_TYPE (@0)))
1463 && (DECIMAL_FLOAT_TYPE_P (TREE_TYPE (@2))
1464 == DECIMAL_FLOAT_TYPE_P (TREE_TYPE (@1))))
1465 (with
1466 {
1467 tree type1 = TREE_TYPE (@1);
1468 if (TREE_CODE (@1) == REAL_CST && !DECIMAL_FLOAT_TYPE_P (type1))
1469 {
1470 REAL_VALUE_TYPE orig = TREE_REAL_CST (@1);
1471 if (TYPE_PRECISION (type1) > TYPE_PRECISION (float_type_node)
1472 && exact_real_truncate (TYPE_MODE (float_type_node), &orig))
1473 type1 = float_type_node;
1474 if (TYPE_PRECISION (type1) > TYPE_PRECISION (double_type_node)
1475 && exact_real_truncate (TYPE_MODE (double_type_node), &orig))
1476 type1 = double_type_node;
1477 }
1478 tree newtype
1479 = (TYPE_PRECISION (TREE_TYPE (@0)) > TYPE_PRECISION (type1)
1480 ? TREE_TYPE (@0) : type1);
1481 }
1482 (if (TYPE_PRECISION (TREE_TYPE (@2)) > TYPE_PRECISION (newtype))
1483 (cmp (convert:newtype @0) (convert:newtype @1))))))
1484
1485 (simplify
1486 (cmp @0 REAL_CST@1)
1487 /* IEEE doesn't distinguish +0 and -0 in comparisons. */
64d3a1f0
RB
1488 (switch
1489 /* a CMP (-0) -> a CMP 0 */
1490 (if (REAL_VALUE_MINUS_ZERO (TREE_REAL_CST (@1)))
1491 (cmp @0 { build_real (TREE_TYPE (@1), dconst0); }))
1492 /* x != NaN is always true, other ops are always false. */
1493 (if (REAL_VALUE_ISNAN (TREE_REAL_CST (@1))
1494 && ! HONOR_SNANS (@1))
1495 { constant_boolean_node (cmp == NE_EXPR, type); })
1496 /* Fold comparisons against infinity. */
1497 (if (REAL_VALUE_ISINF (TREE_REAL_CST (@1))
1498 && MODE_HAS_INFINITIES (TYPE_MODE (TREE_TYPE (@1))))
1499 (with
1500 {
1501 REAL_VALUE_TYPE max;
1502 enum tree_code code = cmp;
1503 bool neg = REAL_VALUE_NEGATIVE (TREE_REAL_CST (@1));
1504 if (neg)
1505 code = swap_tree_comparison (code);
1506 }
1507 (switch
1508 /* x > +Inf is always false, if with ignore sNANs. */
1509 (if (code == GT_EXPR
1510 && ! HONOR_SNANS (@0))
1511 { constant_boolean_node (false, type); })
1512 (if (code == LE_EXPR)
1513 /* x <= +Inf is always true, if we don't case about NaNs. */
1514 (if (! HONOR_NANS (@0))
1515 { constant_boolean_node (true, type); }
1516 /* x <= +Inf is the same as x == x, i.e. isfinite(x). */
1517 (eq @0 @0)))
1518 /* x == +Inf and x >= +Inf are always equal to x > DBL_MAX. */
1519 (if (code == EQ_EXPR || code == GE_EXPR)
1520 (with { real_maxval (&max, neg, TYPE_MODE (TREE_TYPE (@0))); }
1521 (if (neg)
1522 (lt @0 { build_real (TREE_TYPE (@0), max); })
1523 (gt @0 { build_real (TREE_TYPE (@0), max); }))))
1524 /* x < +Inf is always equal to x <= DBL_MAX. */
1525 (if (code == LT_EXPR)
1526 (with { real_maxval (&max, neg, TYPE_MODE (TREE_TYPE (@0))); }
1527 (if (neg)
1528 (ge @0 { build_real (TREE_TYPE (@0), max); })
1529 (le @0 { build_real (TREE_TYPE (@0), max); }))))
1530 /* x != +Inf is always equal to !(x > DBL_MAX). */
1531 (if (code == NE_EXPR)
1532 (with { real_maxval (&max, neg, TYPE_MODE (TREE_TYPE (@0))); }
1533 (if (! HONOR_NANS (@0))
1534 (if (neg)
1535 (ge @0 { build_real (TREE_TYPE (@0), max); })
1536 (le @0 { build_real (TREE_TYPE (@0), max); }))
1537 (if (neg)
1538 (bit_xor (lt @0 { build_real (TREE_TYPE (@0), max); })
1539 { build_one_cst (type); })
1540 (bit_xor (gt @0 { build_real (TREE_TYPE (@0), max); })
1541 { build_one_cst (type); }))))))))))
07cdc2b8
RB
1542
1543 /* If this is a comparison of a real constant with a PLUS_EXPR
1544 or a MINUS_EXPR of a real constant, we can convert it into a
1545 comparison with a revised real constant as long as no overflow
1546 occurs when unsafe_math_optimizations are enabled. */
1547 (if (flag_unsafe_math_optimizations)
1548 (for op (plus minus)
1549 (simplify
1550 (cmp (op @0 REAL_CST@1) REAL_CST@2)
1551 (with
1552 {
1553 tree tem = const_binop (op == PLUS_EXPR ? MINUS_EXPR : PLUS_EXPR,
1554 TREE_TYPE (@1), @2, @1);
1555 }
1556 (if (!TREE_OVERFLOW (tem))
1557 (cmp @0 { tem; }))))))
1558
1559 /* Likewise, we can simplify a comparison of a real constant with
1560 a MINUS_EXPR whose first operand is also a real constant, i.e.
1561 (c1 - x) < c2 becomes x > c1-c2. Reordering is allowed on
1562 floating-point types only if -fassociative-math is set. */
1563 (if (flag_associative_math)
1564 (simplify
0409237b 1565 (cmp (minus REAL_CST@0 @1) REAL_CST@2)
07cdc2b8
RB
1566 (with { tree tem = const_binop (MINUS_EXPR, TREE_TYPE (@1), @0, @2); }
1567 (if (!TREE_OVERFLOW (tem))
1568 (cmp { tem; } @1)))))
1569
1570 /* Fold comparisons against built-in math functions. */
1571 (if (flag_unsafe_math_optimizations
1572 && ! flag_errno_math)
1573 (for sq (SQRT)
1574 (simplify
1575 (cmp (sq @0) REAL_CST@1)
64d3a1f0
RB
1576 (switch
1577 (if (REAL_VALUE_NEGATIVE (TREE_REAL_CST (@1)))
1578 (switch
1579 /* sqrt(x) < y is always false, if y is negative. */
1580 (if (cmp == EQ_EXPR || cmp == LT_EXPR || cmp == LE_EXPR)
8fdc6c67 1581 { constant_boolean_node (false, type); })
64d3a1f0
RB
1582 /* sqrt(x) > y is always true, if y is negative and we
1583 don't care about NaNs, i.e. negative values of x. */
1584 (if (cmp == NE_EXPR || !HONOR_NANS (@0))
1585 { constant_boolean_node (true, type); })
1586 /* sqrt(x) > y is the same as x >= 0, if y is negative. */
1587 (ge @0 { build_real (TREE_TYPE (@0), dconst0); })))
1588 (if (cmp == GT_EXPR || cmp == GE_EXPR)
1589 (with
1590 {
1591 REAL_VALUE_TYPE c2;
1592 REAL_ARITHMETIC (c2, MULT_EXPR,
1593 TREE_REAL_CST (@1), TREE_REAL_CST (@1));
1594 real_convert (&c2, TYPE_MODE (TREE_TYPE (@0)), &c2);
1595 }
1596 (if (REAL_VALUE_ISINF (c2))
1597 /* sqrt(x) > y is x == +Inf, when y is very large. */
1598 (if (HONOR_INFINITIES (@0))
1599 (eq @0 { build_real (TREE_TYPE (@0), c2); })
1600 { constant_boolean_node (false, type); })
1601 /* sqrt(x) > c is the same as x > c*c. */
1602 (cmp @0 { build_real (TREE_TYPE (@0), c2); }))))
1603 (if (cmp == LT_EXPR || cmp == LE_EXPR)
1604 (with
1605 {
1606 REAL_VALUE_TYPE c2;
1607 REAL_ARITHMETIC (c2, MULT_EXPR,
1608 TREE_REAL_CST (@1), TREE_REAL_CST (@1));
1609 real_convert (&c2, TYPE_MODE (TREE_TYPE (@0)), &c2);
1610 }
1611 (if (REAL_VALUE_ISINF (c2))
1612 (switch
1613 /* sqrt(x) < y is always true, when y is a very large
1614 value and we don't care about NaNs or Infinities. */
1615 (if (! HONOR_NANS (@0) && ! HONOR_INFINITIES (@0))
1616 { constant_boolean_node (true, type); })
1617 /* sqrt(x) < y is x != +Inf when y is very large and we
1618 don't care about NaNs. */
1619 (if (! HONOR_NANS (@0))
1620 (ne @0 { build_real (TREE_TYPE (@0), c2); }))
1621 /* sqrt(x) < y is x >= 0 when y is very large and we
1622 don't care about Infinities. */
1623 (if (! HONOR_INFINITIES (@0))
1624 (ge @0 { build_real (TREE_TYPE (@0), dconst0); }))
1625 /* sqrt(x) < y is x >= 0 && x != +Inf, when y is large. */
1626 (if (GENERIC)
1627 (truth_andif
1628 (ge @0 { build_real (TREE_TYPE (@0), dconst0); })
1629 (ne @0 { build_real (TREE_TYPE (@0), c2); }))))
1630 /* sqrt(x) < c is the same as x < c*c, if we ignore NaNs. */
1631 (if (! HONOR_NANS (@0))
1632 (cmp @0 { build_real (TREE_TYPE (@0), c2); })
1633 /* sqrt(x) < c is the same as x >= 0 && x < c*c. */
1634 (if (GENERIC)
1635 (truth_andif
1636 (ge @0 { build_real (TREE_TYPE (@0), dconst0); })
1637 (cmp @0 { build_real (TREE_TYPE (@0), c2); }))))))))))))
2ee05f1e 1638
cfdc4f33
MG
1639/* Unordered tests if either argument is a NaN. */
1640(simplify
1641 (bit_ior (unordered @0 @0) (unordered @1 @1))
aea417d7 1642 (if (types_match (@0, @1))
cfdc4f33 1643 (unordered @0 @1)))
257b01ba
MG
1644(simplify
1645 (bit_and (ordered @0 @0) (ordered @1 @1))
1646 (if (types_match (@0, @1))
1647 (ordered @0 @1)))
cfdc4f33
MG
1648(simplify
1649 (bit_ior:c (unordered @0 @0) (unordered:c@2 @0 @1))
1650 @2)
257b01ba
MG
1651(simplify
1652 (bit_and:c (ordered @0 @0) (ordered:c@2 @0 @1))
1653 @2)
e18c1d66 1654
534bd33b
MG
1655/* -A CMP -B -> B CMP A. */
1656(for cmp (tcc_comparison)
1657 scmp (swapped_tcc_comparison)
1658 (simplify
1659 (cmp (negate @0) (negate @1))
1660 (if (FLOAT_TYPE_P (TREE_TYPE (@0))
1661 || (ANY_INTEGRAL_TYPE_P (TREE_TYPE (@0))
1662 && TYPE_OVERFLOW_UNDEFINED (TREE_TYPE (@0))))
1663 (scmp @0 @1)))
1664 (simplify
1665 (cmp (negate @0) CONSTANT_CLASS_P@1)
1666 (if (FLOAT_TYPE_P (TREE_TYPE (@0))
1667 || (ANY_INTEGRAL_TYPE_P (TREE_TYPE (@0))
1668 && TYPE_OVERFLOW_UNDEFINED (TREE_TYPE (@0))))
1669 (with { tree tem = fold_unary (NEGATE_EXPR, TREE_TYPE (@0), @1); }
1670 (if (tem && !TREE_OVERFLOW (tem))
1671 (scmp @0 { tem; }))))))
1672
79d4f7c6
RB
1673/* From fold_sign_changed_comparison and fold_widened_comparison. */
1674(for cmp (simple_comparison)
1675 (simplify
1676 (cmp (convert@0 @00) (convert?@1 @10))
1677 (if (TREE_CODE (TREE_TYPE (@0)) == INTEGER_TYPE
1678 /* Disable this optimization if we're casting a function pointer
1679 type on targets that require function pointer canonicalization. */
1680 && !(targetm.have_canonicalize_funcptr_for_compare ()
1681 && TREE_CODE (TREE_TYPE (@00)) == POINTER_TYPE
1682 && TREE_CODE (TREE_TYPE (TREE_TYPE (@00))) == FUNCTION_TYPE))
1683 (if (TYPE_PRECISION (TREE_TYPE (@00)) == TYPE_PRECISION (TREE_TYPE (@0))
1684 && (TREE_CODE (@10) == INTEGER_CST
1685 || (@1 != @10 && types_match (TREE_TYPE (@10), TREE_TYPE (@00))))
1686 && (TYPE_UNSIGNED (TREE_TYPE (@00)) == TYPE_UNSIGNED (TREE_TYPE (@0))
1687 || cmp == NE_EXPR
1688 || cmp == EQ_EXPR)
1689 && (POINTER_TYPE_P (TREE_TYPE (@00)) == POINTER_TYPE_P (TREE_TYPE (@0))))
1690 /* ??? The special-casing of INTEGER_CST conversion was in the original
1691 code and here to avoid a spurious overflow flag on the resulting
1692 constant which fold_convert produces. */
1693 (if (TREE_CODE (@1) == INTEGER_CST)
1694 (cmp @00 { force_fit_type (TREE_TYPE (@00), wi::to_widest (@1), 0,
1695 TREE_OVERFLOW (@1)); })
1696 (cmp @00 (convert @1)))
1697
1698 (if (TYPE_PRECISION (TREE_TYPE (@0)) > TYPE_PRECISION (TREE_TYPE (@00)))
1699 /* If possible, express the comparison in the shorter mode. */
1700 (if ((cmp == EQ_EXPR || cmp == NE_EXPR
1701 || TYPE_UNSIGNED (TREE_TYPE (@0)) == TYPE_UNSIGNED (TREE_TYPE (@00)))
1702 && (types_match (TREE_TYPE (@10), TREE_TYPE (@00))
1703 || ((TYPE_PRECISION (TREE_TYPE (@00))
1704 >= TYPE_PRECISION (TREE_TYPE (@10)))
1705 && (TYPE_UNSIGNED (TREE_TYPE (@00))
1706 == TYPE_UNSIGNED (TREE_TYPE (@10))))
1707 || (TREE_CODE (@10) == INTEGER_CST
1708 && (TREE_CODE (TREE_TYPE (@00)) == INTEGER_TYPE
1709 || TREE_CODE (TREE_TYPE (@00)) == BOOLEAN_TYPE)
1710 && int_fits_type_p (@10, TREE_TYPE (@00)))))
1711 (cmp @00 (convert @10))
1712 (if (TREE_CODE (@10) == INTEGER_CST
1713 && TREE_CODE (TREE_TYPE (@00)) == INTEGER_TYPE
1714 && !int_fits_type_p (@10, TREE_TYPE (@00)))
1715 (with
1716 {
1717 tree min = lower_bound_in_type (TREE_TYPE (@10), TREE_TYPE (@00));
1718 tree max = upper_bound_in_type (TREE_TYPE (@10), TREE_TYPE (@00));
1719 bool above = integer_nonzerop (const_binop (LT_EXPR, type, max, @10));
1720 bool below = integer_nonzerop (const_binop (LT_EXPR, type, @10, min));
1721 }
1722 (if (above || below)
1723 (if (cmp == EQ_EXPR || cmp == NE_EXPR)
1724 { constant_boolean_node (cmp == EQ_EXPR ? false : true, type); }
1725 (if (cmp == LT_EXPR || cmp == LE_EXPR)
1726 { constant_boolean_node (above ? true : false, type); }
1727 (if (cmp == GT_EXPR || cmp == GE_EXPR)
1728 { constant_boolean_node (above ? false : true, type); }))))))))))))
66e1cacf
RB
1729
1730/* Equality compare simplifications from fold_binary */
1731(for cmp (eq ne)
1732
1733 /* If we have (A | C) == D where C & ~D != 0, convert this into 0.
1734 Similarly for NE_EXPR. */
1735 (simplify
1736 (cmp (convert?@3 (bit_ior @0 INTEGER_CST@1)) INTEGER_CST@2)
1737 (if (tree_nop_conversion_p (TREE_TYPE (@3), TREE_TYPE (@0))
1738 && wi::bit_and_not (@1, @2) != 0)
1739 { constant_boolean_node (cmp == NE_EXPR, type); }))
1740
1741 /* (X ^ Y) == 0 becomes X == Y, and (X ^ Y) != 0 becomes X != Y. */
1742 (simplify
1743 (cmp (bit_xor @0 @1) integer_zerop)
1744 (cmp @0 @1))
1745
1746 /* (X ^ Y) == Y becomes X == 0.
1747 Likewise (X ^ Y) == X becomes Y == 0. */
1748 (simplify
99e943a2 1749 (cmp:c (bit_xor:c @0 @1) @0)
66e1cacf
RB
1750 (cmp @1 { build_zero_cst (TREE_TYPE (@1)); }))
1751
1752 /* (X ^ C1) op C2 can be rewritten as X op (C1 ^ C2). */
1753 (simplify
1754 (cmp (convert?@3 (bit_xor @0 INTEGER_CST@1)) INTEGER_CST@2)
1755 (if (tree_nop_conversion_p (TREE_TYPE (@3), TREE_TYPE (@0)))
1756 (cmp @0 (bit_xor @1 (convert @2))))))
1757
55cf3946
RB
1758/* bool_var != 0 becomes bool_var. */
1759(simplify
1760 (ne @0 integer_zerop@1)
1761 (if (TREE_CODE (TREE_TYPE (@0)) == BOOLEAN_TYPE
1762 && types_match (type, TREE_TYPE (@0)))
1763 (non_lvalue @0)))
1764/* bool_var == 1 becomes bool_var. */
1765(simplify
1766 (eq @0 integer_onep@1)
1767 (if (TREE_CODE (TREE_TYPE (@0)) == BOOLEAN_TYPE
1768 && types_match (type, TREE_TYPE (@0)))
1769 (non_lvalue @0)))
1770
1771
e18c1d66
RB
1772/* Simplification of math builtins. */
1773
e18c1d66
RB
1774/* fold_builtin_logarithm */
1775(if (flag_unsafe_math_optimizations)
1776 /* Special case, optimize logN(expN(x)) = x. */
1777 (for logs (LOG LOG2 LOG10)
1778 exps (EXP EXP2 EXP10)
1779 (simplify
1780 (logs (exps @0))
1781 @0))
1782 /* Optimize logN(func()) for various exponential functions. We
1783 want to determine the value "x" and the power "exponent" in
1784 order to transform logN(x**exponent) into exponent*logN(x). */
1785 (for logs (LOG LOG LOG LOG
1786 LOG2 LOG2 LOG2 LOG2
1787 LOG10 LOG10 LOG10 LOG10)
1788 exps (EXP EXP2 EXP10 POW10)
1789 (simplify
1790 (logs (exps @0))
1791 (with {
1792 tree x;
1793 switch (exps)
1794 {
1795 CASE_FLT_FN (BUILT_IN_EXP):
1796 /* Prepare to do logN(exp(exponent) -> exponent*logN(e). */
1797 x = build_real (type, real_value_truncate (TYPE_MODE (type),
1798 dconst_e ()));
1799 break;
1800 CASE_FLT_FN (BUILT_IN_EXP2):
1801 /* Prepare to do logN(exp2(exponent) -> exponent*logN(2). */
1802 x = build_real (type, dconst2);
1803 break;
1804 CASE_FLT_FN (BUILT_IN_EXP10):
1805 CASE_FLT_FN (BUILT_IN_POW10):
1806 /* Prepare to do logN(exp10(exponent) -> exponent*logN(10). */
1807 {
1808 REAL_VALUE_TYPE dconst10;
1809 real_from_integer (&dconst10, VOIDmode, 10, SIGNED);
1810 x = build_real (type, dconst10);
1811 }
1812 break;
1813 }
1814 }
1815 (mult (logs { x; }) @0))))
1816 (for logs (LOG LOG
1817 LOG2 LOG2
1818 LOG10 LOG10)
1819 exps (SQRT CBRT)
1820 (simplify
1821 (logs (exps @0))
1822 (with {
1823 tree x;
1824 switch (exps)
1825 {
1826 CASE_FLT_FN (BUILT_IN_SQRT):
1827 /* Prepare to do logN(sqrt(x) -> 0.5*logN(x). */
1828 x = build_real (type, dconsthalf);
1829 break;
1830 CASE_FLT_FN (BUILT_IN_CBRT):
1831 /* Prepare to do logN(cbrt(x) -> (1/3)*logN(x). */
1832 x = build_real (type, real_value_truncate (TYPE_MODE (type),
1833 dconst_third ()));
1834 break;
1835 }
1836 }
1837 (mult { x; } (logs @0)))))
1838 /* logN(pow(x,exponent) -> exponent*logN(x). */
1839 (for logs (LOG LOG2 LOG10)
1840 pows (POW)
1841 (simplify
1842 (logs (pows @0 @1))
1843 (mult @1 (logs @0)))))
1844
be144838
JL
1845/* Narrowing of arithmetic and logical operations.
1846
1847 These are conceptually similar to the transformations performed for
1848 the C/C++ front-ends by shorten_binary_op and shorten_compare. Long
1849 term we want to move all that code out of the front-ends into here. */
1850
1851/* If we have a narrowing conversion of an arithmetic operation where
1852 both operands are widening conversions from the same type as the outer
1853 narrowing conversion. Then convert the innermost operands to a suitable
1854 unsigned type (to avoid introducing undefined behaviour), perform the
1855 operation and convert the result to the desired type. */
1856(for op (plus minus)
1857 (simplify
44fc0a51 1858 (convert (op:s (convert@2 @0) (convert@3 @1)))
be144838
JL
1859 (if (INTEGRAL_TYPE_P (type)
1860 /* We check for type compatibility between @0 and @1 below,
1861 so there's no need to check that @1/@3 are integral types. */
1862 && INTEGRAL_TYPE_P (TREE_TYPE (@0))
1863 && INTEGRAL_TYPE_P (TREE_TYPE (@2))
1864 /* The precision of the type of each operand must match the
1865 precision of the mode of each operand, similarly for the
1866 result. */
1867 && (TYPE_PRECISION (TREE_TYPE (@0))
1868 == GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (@0))))
1869 && (TYPE_PRECISION (TREE_TYPE (@1))
1870 == GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (@1))))
1871 && TYPE_PRECISION (type) == GET_MODE_PRECISION (TYPE_MODE (type))
1872 /* The inner conversion must be a widening conversion. */
1873 && TYPE_PRECISION (TREE_TYPE (@2)) > TYPE_PRECISION (TREE_TYPE (@0))
aea417d7 1874 && types_match (@0, @1)
44fc0a51 1875 && types_match (@0, type))
be144838 1876 (if (TYPE_OVERFLOW_WRAPS (TREE_TYPE (@0)))
8fdc6c67
RB
1877 (convert (op @0 @1))
1878 (with { tree utype = unsigned_type_for (TREE_TYPE (@0)); }
1879 (convert (op (convert:utype @0) (convert:utype @1))))))))
48451e8f
JL
1880
1881/* This is another case of narrowing, specifically when there's an outer
1882 BIT_AND_EXPR which masks off bits outside the type of the innermost
1883 operands. Like the previous case we have to convert the operands
1884 to unsigned types to avoid introducing undefined behaviour for the
1885 arithmetic operation. */
1886(for op (minus plus)
8fdc6c67
RB
1887 (simplify
1888 (bit_and (op:s (convert@2 @0) (convert@3 @1)) INTEGER_CST@4)
1889 (if (INTEGRAL_TYPE_P (type)
1890 /* We check for type compatibility between @0 and @1 below,
1891 so there's no need to check that @1/@3 are integral types. */
1892 && INTEGRAL_TYPE_P (TREE_TYPE (@0))
1893 && INTEGRAL_TYPE_P (TREE_TYPE (@2))
1894 /* The precision of the type of each operand must match the
1895 precision of the mode of each operand, similarly for the
1896 result. */
1897 && (TYPE_PRECISION (TREE_TYPE (@0))
1898 == GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (@0))))
1899 && (TYPE_PRECISION (TREE_TYPE (@1))
1900 == GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (@1))))
1901 && TYPE_PRECISION (type) == GET_MODE_PRECISION (TYPE_MODE (type))
1902 /* The inner conversion must be a widening conversion. */
1903 && TYPE_PRECISION (TREE_TYPE (@2)) > TYPE_PRECISION (TREE_TYPE (@0))
1904 && types_match (@0, @1)
1905 && (tree_int_cst_min_precision (@4, TYPE_SIGN (TREE_TYPE (@0)))
1906 <= TYPE_PRECISION (TREE_TYPE (@0)))
1907 && (TYPE_OVERFLOW_WRAPS (TREE_TYPE (@0))
1908 || tree_int_cst_sgn (@4) >= 0))
1909 (if (TYPE_OVERFLOW_WRAPS (TREE_TYPE (@0)))
1910 (with { tree ntype = TREE_TYPE (@0); }
1911 (convert (bit_and (op @0 @1) (convert:ntype @4))))
1912 (with { tree utype = unsigned_type_for (TREE_TYPE (@0)); }
1913 (convert (bit_and (op (convert:utype @0) (convert:utype @1))
1914 (convert:utype @4))))))))