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.
5 Copyright (C) 2014-2015 Free Software Foundation, Inc.
6 Contributed by Richard Biener <rguenther@suse.de>
7 and Prathamesh Kulkarni <bilbotheelffriend@gmail.com>
9 This file is part of GCC.
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
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
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/>. */
26 /* Generic tree predicates we inherit. */
28 integer_onep integer_zerop integer_all_onesp integer_minus_onep
29 integer_each_onep integer_truep
30 real_zerop real_onep real_minus_onep
32 tree_expr_nonnegative_p)
35 (define_operator_list tcc_comparison
36 lt le eq ne ge gt unordered ordered unlt unle ungt unge uneq ltgt)
37 (define_operator_list inverted_tcc_comparison
38 ge gt ne eq lt le ordered unordered ge gt le lt ltgt uneq)
39 (define_operator_list inverted_tcc_comparison_with_nans
40 unge ungt ne eq unlt unle ordered unordered ge gt le lt ltgt uneq)
41 (define_operator_list swapped_tcc_comparison
42 gt ge eq ne le lt unordered ordered ungt unge unlt unle uneq ltgt)
43 (define_operator_list simple_comparison lt le eq ne ge gt)
44 (define_operator_list swapped_simple_comparison gt ge eq ne le lt)
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)
58 /* Simplifications of operations with one constant operand and
59 simplifications to constants or single values. */
61 (for op (plus pointer_plus minus bit_ior bit_xor)
66 /* 0 +p index -> (type)index */
68 (pointer_plus integer_zerop @1)
69 (non_lvalue (convert @1)))
71 /* See if ARG1 is zero and X + ARG1 reduces to X.
72 Likewise if the operands are reversed. */
74 (plus:c @0 real_zerop@1)
75 (if (fold_real_zero_addition_p (type, @1, 0))
78 /* See if ARG1 is zero and X - ARG1 reduces to X. */
80 (minus @0 real_zerop@1)
81 (if (fold_real_zero_addition_p (type, @1, 1))
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
91 (if (!FLOAT_TYPE_P (type) || !HONOR_NANS (type))
92 { build_zero_cst (type); }))
95 (mult @0 integer_zerop@1)
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. */
103 (mult @0 real_zerop@1)
104 (if (!HONOR_NANS (type) && !HONOR_SIGNED_ZEROS (element_mode (type)))
107 /* In IEEE floating point, x*1 is not equivalent to x for snans.
108 Likewise for complex arithmetic with signed zeros. */
111 (if (!HONOR_SNANS (element_mode (type))
112 && (!HONOR_SIGNED_ZEROS (element_mode (type))
113 || !COMPLEX_FLOAT_TYPE_P (type)))
116 /* Transform x * -1.0 into -x. */
118 (mult @0 real_minus_onep)
119 (if (!HONOR_SNANS (element_mode (type))
120 && (!HONOR_SIGNED_ZEROS (element_mode (type))
121 || !COMPLEX_FLOAT_TYPE_P (type)))
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)
132 (for div (trunc_div ceil_div floor_div round_div exact_div)
134 (div @0 integer_minus_onep@1)
135 (if (!TYPE_UNSIGNED (type))
138 /* For unsigned integral types, FLOOR_DIV_EXPR is the same as
139 TRUNC_DIV_EXPR. Rewrite into the latter in this case. */
142 (if ((INTEGRAL_TYPE_P (type) || VECTOR_INTEGER_TYPE_P (type))
143 && TYPE_UNSIGNED (type))
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)
150 (div (div @0 INTEGER_CST@1) INTEGER_CST@2)
153 wide_int mul = wi::mul (@1, @2, TYPE_SIGN (type), &overflow_p);
156 (div @0 { wide_int_to_tree (type, mul); }))
158 && (TYPE_UNSIGNED (type)
159 || mul != wi::min_value (TYPE_PRECISION (type), SIGNED)))
160 { build_zero_cst (type); }))))
162 /* Optimize A / A to 1.0 if we don't care about
163 NaNs or Infinities. */
166 (if (FLOAT_TYPE_P (type)
167 && ! HONOR_NANS (type)
168 && ! HONOR_INFINITIES (element_mode (type)))
169 { build_one_cst (type); }))
171 /* Optimize -A / A to -1.0 if we don't care about
172 NaNs or Infinities. */
174 (rdiv:c @0 (negate @0))
175 (if (FLOAT_TYPE_P (type)
176 && ! HONOR_NANS (type)
177 && ! HONOR_INFINITIES (element_mode (type)))
178 { build_minus_one_cst (type); }))
180 /* In IEEE floating point, x/1 is not equivalent to x for snans. */
183 (if (!HONOR_SNANS (element_mode (type)))
186 /* In IEEE floating point, x/-1 is not equivalent to -x for snans. */
188 (rdiv @0 real_minus_onep)
189 (if (!HONOR_SNANS (element_mode (type)))
192 /* If ARG1 is a constant, we can convert this to a multiply by the
193 reciprocal. This does not have the same rounding properties,
194 so only do this if -freciprocal-math. We can actually
195 always safely do it if ARG1 is a power of two, but it's hard to
196 tell if it is or not in a portable manner. */
197 (for cst (REAL_CST COMPLEX_CST VECTOR_CST)
201 (if (flag_reciprocal_math
204 { tree tem = const_binop (RDIV_EXPR, type, build_one_cst (type), @1); }
206 (mult @0 { tem; } ))))
207 (if (cst != COMPLEX_CST)
208 (with { tree inverse = exact_inverse (type, @1); }
210 (mult @0 { inverse; } )))))))
212 /* Same applies to modulo operations, but fold is inconsistent here
213 and simplifies 0 % x to 0, only preserving literal 0 % 0. */
214 (for mod (ceil_mod floor_mod round_mod trunc_mod)
215 /* 0 % X is always zero. */
217 (mod integer_zerop@0 @1)
218 /* But not for 0 % 0 so that we can get the proper warnings and errors. */
219 (if (!integer_zerop (@1))
221 /* X % 1 is always zero. */
223 (mod @0 integer_onep)
224 { build_zero_cst (type); })
225 /* X % -1 is zero. */
227 (mod @0 integer_minus_onep@1)
228 (if (!TYPE_UNSIGNED (type))
229 { build_zero_cst (type); }))
230 /* (X % Y) % Y is just X % Y. */
232 (mod (mod@2 @0 @1) @1)
235 /* X % -C is the same as X % C. */
237 (trunc_mod @0 INTEGER_CST@1)
238 (if (TYPE_SIGN (type) == SIGNED
239 && !TREE_OVERFLOW (@1)
241 && !TYPE_OVERFLOW_TRAPS (type)
242 /* Avoid this transformation if C is INT_MIN, i.e. C == -C. */
243 && !sign_bit_p (@1, @1))
244 (trunc_mod @0 (negate @1))))
246 /* X % -Y is the same as X % Y. */
248 (trunc_mod @0 (convert? (negate @1)))
249 (if (!TYPE_UNSIGNED (type)
250 && !TYPE_OVERFLOW_TRAPS (type)
251 && tree_nop_conversion_p (type, TREE_TYPE (@1)))
252 (trunc_mod @0 (convert @1))))
254 /* X - (X / Y) * Y is the same as X % Y. */
256 (minus (convert1? @0) (convert2? (mult (trunc_div @0 @1) @1)))
257 (if (INTEGRAL_TYPE_P (type) || VECTOR_INTEGER_TYPE_P (type))
258 (trunc_mod (convert @0) (convert @1))))
260 /* Optimize TRUNC_MOD_EXPR by a power of two into a BIT_AND_EXPR,
261 i.e. "X % C" into "X & (C - 1)", if X and C are positive.
262 Also optimize A % (C << N) where C is a power of 2,
263 to A & ((C << N) - 1). */
264 (match (power_of_two_cand @1)
266 (match (power_of_two_cand @1)
267 (lshift INTEGER_CST@1 @2))
268 (for mod (trunc_mod floor_mod)
270 (mod @0 (convert?@3 (power_of_two_cand@1 @2)))
271 (if ((TYPE_UNSIGNED (type)
272 || tree_expr_nonnegative_p (@0))
273 && tree_nop_conversion_p (type, TREE_TYPE (@3))
274 && integer_pow2p (@2) && tree_int_cst_sgn (@2) > 0)
275 (bit_and @0 (convert (minus @1 { build_int_cst (TREE_TYPE (@1), 1); }))))))
277 /* X % Y is smaller than Y. */
280 (cmp (trunc_mod @0 @1) @1)
281 (if (TYPE_UNSIGNED (TREE_TYPE (@0)))
282 { constant_boolean_node (cmp == LT_EXPR, type); })))
285 (cmp @1 (trunc_mod @0 @1))
286 (if (TYPE_UNSIGNED (TREE_TYPE (@0)))
287 { constant_boolean_node (cmp == GT_EXPR, type); })))
291 (bit_ior @0 integer_all_onesp@1)
296 (bit_and @0 integer_zerop@1)
301 (bit_ior:c (convert? @0) (convert? (bit_not @0)))
302 (convert { build_all_ones_cst (TREE_TYPE (@0)); }))
307 { build_zero_cst (type); })
309 /* Canonicalize X ^ ~0 to ~X. */
311 (bit_xor @0 integer_all_onesp@1)
316 (bit_xor:c (bit_not @0) @0)
317 { build_all_ones_cst (type); })
321 (bit_and @0 integer_all_onesp)
324 /* x & x -> x, x | x -> x */
325 (for bitop (bit_and bit_ior)
330 /* x + (x & 1) -> (x + 1) & ~1 */
332 (plus:c @0 (bit_and@2 @0 integer_onep@1))
333 (if (single_use (@2))
334 (bit_and (plus @0 @1) (bit_not @1))))
336 /* x & ~(x & y) -> x & ~y */
337 /* x | ~(x | y) -> x | ~y */
338 (for bitop (bit_and bit_ior)
340 (bitop:c @0 (bit_not (bitop:c@2 @0 @1)))
341 (if (single_use (@2))
342 (bitop @0 (bit_not @1)))))
344 /* (x | y) & ~x -> y & ~x */
345 /* (x & y) | ~x -> y | ~x */
346 (for bitop (bit_and bit_ior)
347 rbitop (bit_ior bit_and)
349 (bitop:c (rbitop:c @0 @1) (bit_not@2 @0))
352 /* (x & y) ^ (x | y) -> x ^ y */
354 (bit_xor:c (bit_and @0 @1) (bit_ior @0 @1))
357 /* (x ^ y) ^ (x | y) -> x & y */
359 (bit_xor:c (bit_xor @0 @1) (bit_ior @0 @1))
362 /* (x & y) + (x ^ y) -> x | y */
363 /* (x & y) | (x ^ y) -> x | y */
364 /* (x & y) ^ (x ^ y) -> x | y */
365 (for op (plus bit_ior bit_xor)
367 (op:c (bit_and @0 @1) (bit_xor @0 @1))
370 /* (x & y) + (x | y) -> x + y */
372 (plus:c (bit_and @0 @1) (bit_ior @0 @1))
375 /* (x + y) - (x | y) -> x & y */
377 (minus (plus @0 @1) (bit_ior @0 @1))
378 (if (!TYPE_OVERFLOW_SANITIZED (type) && !TYPE_OVERFLOW_TRAPS (type)
379 && !TYPE_SATURATING (type))
382 /* (x + y) - (x & y) -> x | y */
384 (minus (plus @0 @1) (bit_and @0 @1))
385 (if (!TYPE_OVERFLOW_SANITIZED (type) && !TYPE_OVERFLOW_TRAPS (type)
386 && !TYPE_SATURATING (type))
389 /* (x | y) - (x ^ y) -> x & y */
391 (minus (bit_ior @0 @1) (bit_xor @0 @1))
394 /* (x | y) - (x & y) -> x ^ y */
396 (minus (bit_ior @0 @1) (bit_and @0 @1))
399 /* (x | y) & ~(x & y) -> x ^ y */
401 (bit_and:c (bit_ior @0 @1) (bit_not (bit_and @0 @1)))
404 /* (x | y) & (~x ^ y) -> x & y */
406 (bit_and:c (bit_ior:c @0 @1) (bit_xor:c @1 (bit_not @0)))
409 /* ~x & ~y -> ~(x | y)
410 ~x | ~y -> ~(x & y) */
411 (for op (bit_and bit_ior)
412 rop (bit_ior bit_and)
414 (op (convert1? (bit_not @0)) (convert2? (bit_not @1)))
415 (if (tree_nop_conversion_p (type, TREE_TYPE (@0))
416 && tree_nop_conversion_p (type, TREE_TYPE (@1)))
417 (bit_not (rop (convert @0) (convert @1))))))
419 /* If we are XORing two BIT_AND_EXPR's, both of which are and'ing
420 with a constant, and the two constants have no bits in common,
421 we should treat this as a BIT_IOR_EXPR since this may produce more
424 (bit_xor (convert1? (bit_and@4 @0 INTEGER_CST@1))
425 (convert2? (bit_and@5 @2 INTEGER_CST@3)))
426 (if (tree_nop_conversion_p (type, TREE_TYPE (@0))
427 && tree_nop_conversion_p (type, TREE_TYPE (@2))
428 && wi::bit_and (@1, @3) == 0)
429 (bit_ior (convert @4) (convert @5))))
431 /* (X | Y) ^ X -> Y & ~ X*/
433 (bit_xor:c (convert? (bit_ior:c @0 @1)) (convert? @0))
434 (if (tree_nop_conversion_p (type, TREE_TYPE (@0)))
435 (convert (bit_and @1 (bit_not @0)))))
437 /* Convert ~X ^ ~Y to X ^ Y. */
439 (bit_xor (convert1? (bit_not @0)) (convert2? (bit_not @1)))
440 (if (tree_nop_conversion_p (type, TREE_TYPE (@0))
441 && tree_nop_conversion_p (type, TREE_TYPE (@1)))
442 (bit_xor (convert @0) (convert @1))))
444 /* Convert ~X ^ C to X ^ ~C. */
446 (bit_xor (convert? (bit_not @0)) INTEGER_CST@1)
447 (bit_xor (convert @0) (bit_not @1)))
449 /* Fold (X & Y) ^ Y as ~X & Y. */
451 (bit_xor:c (bit_and:c @0 @1) @1)
452 (bit_and (bit_not @0) @1))
462 (abs tree_expr_nonnegative_p@0)
465 /* A - B -> A + (-B) if B is easily negatable. This just covers
466 very few cases of "easily negatable", effectively inlining negate_expr_p. */
468 (minus @0 INTEGER_CST@1)
469 (if ((INTEGRAL_TYPE_P (type)
470 && TYPE_OVERFLOW_WRAPS (type))
471 || (!TYPE_OVERFLOW_SANITIZED (type)
472 && may_negate_without_overflow_p (@1)))
473 (plus @0 (negate @1))))
475 (minus @0 REAL_CST@1)
476 (if (REAL_VALUE_NEGATIVE (TREE_REAL_CST (@1)))
477 (plus @0 (negate @1))))
479 (minus @0 VECTOR_CST@1)
480 (if (FLOAT_TYPE_P (type) || TYPE_OVERFLOW_WRAPS (type))
481 (plus @0 (negate @1))))
484 /* Try to fold (type) X op CST -> (type) (X op ((type-x) CST))
486 For bitwise binary operations apply operand conversions to the
487 binary operation result instead of to the operands. This allows
488 to combine successive conversions and bitwise binary operations.
489 We combine the above two cases by using a conditional convert. */
490 (for bitop (bit_and bit_ior bit_xor)
492 (bitop (convert @0) (convert? @1))
493 (if (((TREE_CODE (@1) == INTEGER_CST
494 && INTEGRAL_TYPE_P (TREE_TYPE (@0))
495 && int_fits_type_p (@1, TREE_TYPE (@0)))
496 || types_match (@0, @1))
497 /* ??? This transform conflicts with fold-const.c doing
498 Convert (T)(x & c) into (T)x & (T)c, if c is an integer
499 constants (if x has signed type, the sign bit cannot be set
500 in c). This folds extension into the BIT_AND_EXPR.
501 Restrict it to GIMPLE to avoid endless recursions. */
502 && (bitop != BIT_AND_EXPR || GIMPLE)
503 && (/* That's a good idea if the conversion widens the operand, thus
504 after hoisting the conversion the operation will be narrower. */
505 TYPE_PRECISION (TREE_TYPE (@0)) < TYPE_PRECISION (type)
506 /* It's also a good idea if the conversion is to a non-integer
508 || GET_MODE_CLASS (TYPE_MODE (type)) != MODE_INT
509 /* Or if the precision of TO is not the same as the precision
511 || TYPE_PRECISION (type) != GET_MODE_PRECISION (TYPE_MODE (type))))
512 (convert (bitop @0 (convert @1))))))
514 (for bitop (bit_and bit_ior)
515 rbitop (bit_ior bit_and)
516 /* (x | y) & x -> x */
517 /* (x & y) | x -> x */
519 (bitop:c (rbitop:c @0 @1) @0)
521 /* (~x | y) & x -> x & y */
522 /* (~x & y) | x -> x | y */
524 (bitop:c (rbitop:c (bit_not @0) @1) @0)
527 /* Simplify (A & B) OP0 (C & B) to (A OP0 C) & B. */
528 (for bitop (bit_and bit_ior bit_xor)
530 (bitop (bit_and:c @0 @1) (bit_and @2 @1))
531 (bit_and (bitop @0 @2) @1)))
533 /* (x | CST1) & CST2 -> (x & CST2) | (CST1 & CST2) */
535 (bit_and (bit_ior @0 CONSTANT_CLASS_P@1) CONSTANT_CLASS_P@2)
536 (bit_ior (bit_and @0 @2) (bit_and @1 @2)))
538 /* Combine successive equal operations with constants. */
539 (for bitop (bit_and bit_ior bit_xor)
541 (bitop (bitop @0 CONSTANT_CLASS_P@1) CONSTANT_CLASS_P@2)
542 (bitop @0 (bitop @1 @2))))
544 /* Try simple folding for X op !X, and X op X with the help
545 of the truth_valued_p and logical_inverted_value predicates. */
546 (match truth_valued_p
548 (if (INTEGRAL_TYPE_P (type) && TYPE_PRECISION (type) == 1)))
549 (for op (tcc_comparison truth_and truth_andif truth_or truth_orif truth_xor)
550 (match truth_valued_p
552 (match truth_valued_p
555 (match (logical_inverted_value @0)
556 (bit_not truth_valued_p@0))
557 (match (logical_inverted_value @0)
558 (eq @0 integer_zerop))
559 (match (logical_inverted_value @0)
560 (ne truth_valued_p@0 integer_truep))
561 (match (logical_inverted_value @0)
562 (bit_xor truth_valued_p@0 integer_truep))
566 (bit_and:c @0 (logical_inverted_value @0))
567 { build_zero_cst (type); })
568 /* X | !X and X ^ !X -> 1, , if X is truth-valued. */
569 (for op (bit_ior bit_xor)
571 (op:c truth_valued_p@0 (logical_inverted_value @0))
572 { constant_boolean_node (true, type); }))
574 /* If arg1 and arg2 are booleans (or any single bit type)
575 then try to simplify:
582 But only do this if our result feeds into a comparison as
583 this transformation is not always a win, particularly on
584 targets with and-not instructions.
585 -> simplify_bitwise_binary_boolean */
587 (ne (bit_and:c (bit_not @0) @1) integer_zerop)
588 (if (INTEGRAL_TYPE_P (TREE_TYPE (@1))
589 && TYPE_PRECISION (TREE_TYPE (@1)) == 1)
592 (ne (bit_ior:c (bit_not @0) @1) integer_zerop)
593 (if (INTEGRAL_TYPE_P (TREE_TYPE (@1))
594 && TYPE_PRECISION (TREE_TYPE (@1)) == 1)
599 (bit_not (bit_not @0))
602 /* Convert ~ (-A) to A - 1. */
604 (bit_not (convert? (negate @0)))
605 (if (tree_nop_conversion_p (type, TREE_TYPE (@0)))
606 (convert (minus @0 { build_one_cst (TREE_TYPE (@0)); }))))
608 /* Convert ~ (A - 1) or ~ (A + -1) to -A. */
610 (bit_not (convert? (minus @0 integer_onep)))
611 (if (tree_nop_conversion_p (type, TREE_TYPE (@0)))
612 (convert (negate @0))))
614 (bit_not (convert? (plus @0 integer_all_onesp)))
615 (if (tree_nop_conversion_p (type, TREE_TYPE (@0)))
616 (convert (negate @0))))
618 /* Part of convert ~(X ^ Y) to ~X ^ Y or X ^ ~Y if ~X or ~Y simplify. */
620 (bit_not (convert? (bit_xor @0 INTEGER_CST@1)))
621 (if (tree_nop_conversion_p (type, TREE_TYPE (@0)))
622 (convert (bit_xor @0 (bit_not @1)))))
624 (bit_not (convert? (bit_xor:c (bit_not @0) @1)))
625 (if (tree_nop_conversion_p (type, TREE_TYPE (@0)))
626 (convert (bit_xor @0 @1))))
628 /* (x & ~m) | (y & m) -> ((x ^ y) & m) ^ x */
630 (bit_ior:c (bit_and:c@3 @0 (bit_not @2)) (bit_and:c@4 @1 @2))
631 (if (single_use (@3) && single_use (@4))
632 (bit_xor (bit_and (bit_xor @0 @1) @2) @0)))
635 /* Associate (p +p off1) +p off2 as (p +p (off1 + off2)). */
637 (pointer_plus (pointer_plus@2 @0 @1) @3)
639 || (TREE_CODE (@1) == INTEGER_CST && TREE_CODE (@3) == INTEGER_CST))
640 (pointer_plus @0 (plus @1 @3))))
646 tem4 = (unsigned long) tem3;
651 (pointer_plus @0 (convert?@2 (minus@3 (convert @1) (convert @0))))
652 /* Conditionally look through a sign-changing conversion. */
653 (if (TYPE_PRECISION (TREE_TYPE (@2)) == TYPE_PRECISION (TREE_TYPE (@3))
654 && ((GIMPLE && useless_type_conversion_p (type, TREE_TYPE (@1)))
655 || (GENERIC && type == TREE_TYPE (@1))))
659 tem = (sizetype) ptr;
663 and produce the simpler and easier to analyze with respect to alignment
664 ... = ptr & ~algn; */
666 (pointer_plus @0 (negate (bit_and (convert @0) INTEGER_CST@1)))
667 (with { tree algn = wide_int_to_tree (TREE_TYPE (@0), wi::bit_not (@1)); }
668 (bit_and @0 { algn; })))
670 /* Try folding difference of addresses. */
672 (minus (convert ADDR_EXPR@0) (convert @1))
673 (if (tree_nop_conversion_p (type, TREE_TYPE (@0)))
674 (with { HOST_WIDE_INT diff; }
675 (if (ptr_difference_const (@0, @1, &diff))
676 { build_int_cst_type (type, diff); }))))
678 (minus (convert @0) (convert ADDR_EXPR@1))
679 (if (tree_nop_conversion_p (type, TREE_TYPE (@0)))
680 (with { HOST_WIDE_INT diff; }
681 (if (ptr_difference_const (@0, @1, &diff))
682 { build_int_cst_type (type, diff); }))))
684 /* If arg0 is derived from the address of an object or function, we may
685 be able to fold this expression using the object or function's
688 (bit_and (convert? @0) INTEGER_CST@1)
689 (if (POINTER_TYPE_P (TREE_TYPE (@0))
690 && tree_nop_conversion_p (type, TREE_TYPE (@0)))
694 unsigned HOST_WIDE_INT bitpos;
695 get_pointer_alignment_1 (@0, &align, &bitpos);
697 (if (wi::ltu_p (@1, align / BITS_PER_UNIT))
698 { wide_int_to_tree (type, wi::bit_and (@1, bitpos / BITS_PER_UNIT)); }))))
701 /* We can't reassociate at all for saturating types. */
702 (if (!TYPE_SATURATING (type))
704 /* Contract negates. */
705 /* A + (-B) -> A - B */
707 (plus:c (convert1? @0) (convert2? (negate @1)))
708 /* Apply STRIP_NOPS on @0 and the negate. */
709 (if (tree_nop_conversion_p (type, TREE_TYPE (@0))
710 && tree_nop_conversion_p (type, TREE_TYPE (@1))
711 && !TYPE_OVERFLOW_SANITIZED (type))
712 (minus (convert @0) (convert @1))))
713 /* A - (-B) -> A + B */
715 (minus (convert1? @0) (convert2? (negate @1)))
716 (if (tree_nop_conversion_p (type, TREE_TYPE (@0))
717 && tree_nop_conversion_p (type, TREE_TYPE (@1))
718 && !TYPE_OVERFLOW_SANITIZED (type))
719 (plus (convert @0) (convert @1))))
722 (negate (convert? (negate @1)))
723 (if (tree_nop_conversion_p (type, TREE_TYPE (@1))
724 && !TYPE_OVERFLOW_SANITIZED (type))
727 /* We can't reassociate floating-point unless -fassociative-math
728 or fixed-point plus or minus because of saturation to +-Inf. */
729 (if ((!FLOAT_TYPE_P (type) || flag_associative_math)
730 && !FIXED_POINT_TYPE_P (type))
732 /* Match patterns that allow contracting a plus-minus pair
733 irrespective of overflow issues. */
734 /* (A +- B) - A -> +- B */
735 /* (A +- B) -+ B -> A */
736 /* A - (A +- B) -> -+ B */
737 /* A +- (B -+ A) -> +- B */
739 (minus (plus:c @0 @1) @0)
742 (minus (minus @0 @1) @0)
745 (plus:c (minus @0 @1) @1)
748 (minus @0 (plus:c @0 @1))
751 (minus @0 (minus @0 @1))
754 /* (A +- CST) +- CST -> A + CST */
755 (for outer_op (plus minus)
756 (for inner_op (plus minus)
758 (outer_op (inner_op @0 CONSTANT_CLASS_P@1) CONSTANT_CLASS_P@2)
759 /* If the constant operation overflows we cannot do the transform
760 as we would introduce undefined overflow, for example
761 with (a - 1) + INT_MIN. */
762 (with { tree cst = fold_binary (outer_op == inner_op
763 ? PLUS_EXPR : MINUS_EXPR, type, @1, @2); }
764 (if (cst && !TREE_OVERFLOW (cst))
765 (inner_op @0 { cst; } ))))))
767 /* (CST - A) +- CST -> CST - A */
768 (for outer_op (plus minus)
770 (outer_op (minus CONSTANT_CLASS_P@1 @0) CONSTANT_CLASS_P@2)
771 (with { tree cst = fold_binary (outer_op, type, @1, @2); }
772 (if (cst && !TREE_OVERFLOW (cst))
773 (minus { cst; } @0)))))
777 (plus:c (bit_not @0) @0)
778 (if (!TYPE_OVERFLOW_TRAPS (type))
779 { build_all_ones_cst (type); }))
783 (plus (convert? (bit_not @0)) integer_each_onep)
784 (if (tree_nop_conversion_p (type, TREE_TYPE (@0)))
785 (negate (convert @0))))
789 (minus (convert? (negate @0)) integer_each_onep)
790 (if (!TYPE_OVERFLOW_TRAPS (type)
791 && tree_nop_conversion_p (type, TREE_TYPE (@0)))
792 (bit_not (convert @0))))
796 (minus integer_all_onesp @0)
799 /* (T)(P + A) - (T)P -> (T) A */
800 (for add (plus pointer_plus)
802 (minus (convert (add @0 @1))
804 (if (element_precision (type) <= element_precision (TREE_TYPE (@1))
805 /* For integer types, if A has a smaller type
806 than T the result depends on the possible
808 E.g. T=size_t, A=(unsigned)429497295, P>0.
809 However, if an overflow in P + A would cause
810 undefined behavior, we can assume that there
812 || (INTEGRAL_TYPE_P (TREE_TYPE (@0))
813 && TYPE_OVERFLOW_UNDEFINED (TREE_TYPE (@0)))
814 /* For pointer types, if the conversion of A to the
815 final type requires a sign- or zero-extension,
816 then we have to punt - it is not defined which
818 || (POINTER_TYPE_P (TREE_TYPE (@0))
819 && TREE_CODE (@1) == INTEGER_CST
820 && tree_int_cst_sign_bit (@1) == 0))
824 /* Simplifications of MIN_EXPR and MAX_EXPR. */
826 (for minmax (min max)
832 (if (INTEGRAL_TYPE_P (type)
833 && TYPE_MIN_VALUE (type)
834 && operand_equal_p (@1, TYPE_MIN_VALUE (type), OEP_ONLY_CONST))
838 (if (INTEGRAL_TYPE_P (type)
839 && TYPE_MAX_VALUE (type)
840 && operand_equal_p (@1, TYPE_MAX_VALUE (type), OEP_ONLY_CONST))
844 /* Simplifications of shift and rotates. */
846 (for rotate (lrotate rrotate)
848 (rotate integer_all_onesp@0 @1)
851 /* Optimize -1 >> x for arithmetic right shifts. */
853 (rshift integer_all_onesp@0 @1)
854 (if (!TYPE_UNSIGNED (type)
855 && tree_expr_nonnegative_p (@1))
858 (for shiftrotate (lrotate rrotate lshift rshift)
860 (shiftrotate @0 integer_zerop)
863 (shiftrotate integer_zerop@0 @1)
865 /* Prefer vector1 << scalar to vector1 << vector2
866 if vector2 is uniform. */
867 (for vec (VECTOR_CST CONSTRUCTOR)
869 (shiftrotate @0 vec@1)
870 (with { tree tem = uniform_vector_p (@1); }
872 (shiftrotate @0 { tem; }))))))
874 /* Rewrite an LROTATE_EXPR by a constant into an
875 RROTATE_EXPR by a new constant. */
877 (lrotate @0 INTEGER_CST@1)
878 (rrotate @0 { fold_binary (MINUS_EXPR, TREE_TYPE (@1),
879 build_int_cst (TREE_TYPE (@1),
880 element_precision (type)), @1); }))
882 /* ((1 << A) & 1) != 0 -> A == 0
883 ((1 << A) & 1) == 0 -> A != 0 */
887 (cmp (bit_and (lshift integer_onep @0) integer_onep) integer_zerop)
888 (icmp @0 { build_zero_cst (TREE_TYPE (@0)); })))
890 /* (CST1 << A) == CST2 -> A == ctz (CST2) - ctz (CST1)
891 (CST1 << A) != CST2 -> A != ctz (CST2) - ctz (CST1)
895 (cmp (lshift INTEGER_CST@0 @1) INTEGER_CST@2)
896 (with { int cand = wi::ctz (@2) - wi::ctz (@0); }
898 || (!integer_zerop (@2)
899 && wi::ne_p (wi::lshift (@0, cand), @2)))
900 { constant_boolean_node (cmp == NE_EXPR, type); })
901 (if (!integer_zerop (@2)
902 && wi::eq_p (wi::lshift (@0, cand), @2))
903 (cmp @1 { build_int_cst (TREE_TYPE (@1), cand); })))))
905 /* Fold (X << C1) & C2 into (X << C1) & (C2 | ((1 << C1) - 1))
906 (X >> C1) & C2 into (X >> C1) & (C2 | ~((type) -1 >> C1))
907 if the new mask might be further optimized. */
908 (for shift (lshift rshift)
910 (bit_and (convert?@4 (shift@5 (convert1?@3 @0) INTEGER_CST@1)) INTEGER_CST@2)
911 (if (tree_nop_conversion_p (TREE_TYPE (@4), TREE_TYPE (@5))
912 && TYPE_PRECISION (type) <= HOST_BITS_PER_WIDE_INT
913 && tree_fits_uhwi_p (@1)
914 && tree_to_uhwi (@1) > 0
915 && tree_to_uhwi (@1) < TYPE_PRECISION (type))
918 unsigned int shiftc = tree_to_uhwi (@1);
919 unsigned HOST_WIDE_INT mask = TREE_INT_CST_LOW (@2);
920 unsigned HOST_WIDE_INT newmask, zerobits = 0;
921 tree shift_type = TREE_TYPE (@3);
924 if (shift == LSHIFT_EXPR)
925 zerobits = ((((unsigned HOST_WIDE_INT) 1) << shiftc) - 1);
926 else if (shift == RSHIFT_EXPR
927 && (TYPE_PRECISION (shift_type)
928 == GET_MODE_PRECISION (TYPE_MODE (shift_type))))
930 prec = TYPE_PRECISION (TREE_TYPE (@3));
932 /* See if more bits can be proven as zero because of
935 && TYPE_UNSIGNED (TREE_TYPE (@0)))
937 tree inner_type = TREE_TYPE (@0);
938 if ((TYPE_PRECISION (inner_type)
939 == GET_MODE_PRECISION (TYPE_MODE (inner_type)))
940 && TYPE_PRECISION (inner_type) < prec)
942 prec = TYPE_PRECISION (inner_type);
943 /* See if we can shorten the right shift. */
945 shift_type = inner_type;
946 /* Otherwise X >> C1 is all zeros, so we'll optimize
947 it into (X, 0) later on by making sure zerobits
951 zerobits = ~(unsigned HOST_WIDE_INT) 0;
954 zerobits >>= HOST_BITS_PER_WIDE_INT - shiftc;
955 zerobits <<= prec - shiftc;
957 /* For arithmetic shift if sign bit could be set, zerobits
958 can contain actually sign bits, so no transformation is
959 possible, unless MASK masks them all away. In that
960 case the shift needs to be converted into logical shift. */
961 if (!TYPE_UNSIGNED (TREE_TYPE (@3))
962 && prec == TYPE_PRECISION (TREE_TYPE (@3)))
964 if ((mask & zerobits) == 0)
965 shift_type = unsigned_type_for (TREE_TYPE (@3));
971 /* ((X << 16) & 0xff00) is (X, 0). */
972 (if ((mask & zerobits) == mask)
973 { build_int_cst (type, 0); })
974 (with { newmask = mask | zerobits; }
975 (if (newmask != mask && (newmask & (newmask + 1)) == 0)
978 /* Only do the transformation if NEWMASK is some integer
980 for (prec = BITS_PER_UNIT;
981 prec < HOST_BITS_PER_WIDE_INT; prec <<= 1)
982 if (newmask == (((unsigned HOST_WIDE_INT) 1) << prec) - 1)
985 (if (prec < HOST_BITS_PER_WIDE_INT
986 || newmask == ~(unsigned HOST_WIDE_INT) 0)
988 { tree newmaskt = build_int_cst_type (TREE_TYPE (@2), newmask); }
989 (if (!tree_int_cst_equal (newmaskt, @2))
990 (if (shift_type != TREE_TYPE (@3)
991 && single_use (@4) && single_use (@5))
992 (bit_and (convert (shift:shift_type (convert @3) @1)) { newmaskt; }))
993 (if (shift_type == TREE_TYPE (@3))
994 (bit_and @4 { newmaskt; }))))))))))))
996 /* Simplifications of conversions. */
998 /* Basic strip-useless-type-conversions / strip_nops. */
999 (for cvt (convert view_convert float fix_trunc)
1002 (if ((GIMPLE && useless_type_conversion_p (type, TREE_TYPE (@0)))
1003 || (GENERIC && type == TREE_TYPE (@0)))
1006 /* Contract view-conversions. */
1008 (view_convert (view_convert @0))
1011 /* For integral conversions with the same precision or pointer
1012 conversions use a NOP_EXPR instead. */
1015 (if ((INTEGRAL_TYPE_P (type) || POINTER_TYPE_P (type))
1016 && (INTEGRAL_TYPE_P (TREE_TYPE (@0)) || POINTER_TYPE_P (TREE_TYPE (@0)))
1017 && TYPE_PRECISION (type) == TYPE_PRECISION (TREE_TYPE (@0)))
1020 /* Strip inner integral conversions that do not change precision or size. */
1022 (view_convert (convert@0 @1))
1023 (if ((INTEGRAL_TYPE_P (TREE_TYPE (@0)) || POINTER_TYPE_P (TREE_TYPE (@0)))
1024 && (INTEGRAL_TYPE_P (TREE_TYPE (@1)) || POINTER_TYPE_P (TREE_TYPE (@1)))
1025 && (TYPE_PRECISION (TREE_TYPE (@0)) == TYPE_PRECISION (TREE_TYPE (@1)))
1026 && (TYPE_SIZE (TREE_TYPE (@0)) == TYPE_SIZE (TREE_TYPE (@1))))
1029 /* Re-association barriers around constants and other re-association
1030 barriers can be removed. */
1032 (paren CONSTANT_CLASS_P@0)
1035 (paren (paren@1 @0))
1038 /* Handle cases of two conversions in a row. */
1039 (for ocvt (convert float fix_trunc)
1040 (for icvt (convert float)
1045 tree inside_type = TREE_TYPE (@0);
1046 tree inter_type = TREE_TYPE (@1);
1047 int inside_int = INTEGRAL_TYPE_P (inside_type);
1048 int inside_ptr = POINTER_TYPE_P (inside_type);
1049 int inside_float = FLOAT_TYPE_P (inside_type);
1050 int inside_vec = VECTOR_TYPE_P (inside_type);
1051 unsigned int inside_prec = TYPE_PRECISION (inside_type);
1052 int inside_unsignedp = TYPE_UNSIGNED (inside_type);
1053 int inter_int = INTEGRAL_TYPE_P (inter_type);
1054 int inter_ptr = POINTER_TYPE_P (inter_type);
1055 int inter_float = FLOAT_TYPE_P (inter_type);
1056 int inter_vec = VECTOR_TYPE_P (inter_type);
1057 unsigned int inter_prec = TYPE_PRECISION (inter_type);
1058 int inter_unsignedp = TYPE_UNSIGNED (inter_type);
1059 int final_int = INTEGRAL_TYPE_P (type);
1060 int final_ptr = POINTER_TYPE_P (type);
1061 int final_float = FLOAT_TYPE_P (type);
1062 int final_vec = VECTOR_TYPE_P (type);
1063 unsigned int final_prec = TYPE_PRECISION (type);
1064 int final_unsignedp = TYPE_UNSIGNED (type);
1066 /* In addition to the cases of two conversions in a row
1067 handled below, if we are converting something to its own
1068 type via an object of identical or wider precision, neither
1069 conversion is needed. */
1070 (if (((GIMPLE && useless_type_conversion_p (type, inside_type))
1072 && TYPE_MAIN_VARIANT (type) == TYPE_MAIN_VARIANT (inside_type)))
1073 && (((inter_int || inter_ptr) && final_int)
1074 || (inter_float && final_float))
1075 && inter_prec >= final_prec)
1078 /* Likewise, if the intermediate and initial types are either both
1079 float or both integer, we don't need the middle conversion if the
1080 former is wider than the latter and doesn't change the signedness
1081 (for integers). Avoid this if the final type is a pointer since
1082 then we sometimes need the middle conversion. Likewise if the
1083 final type has a precision not equal to the size of its mode. */
1084 (if (((inter_int && inside_int) || (inter_float && inside_float))
1085 && (final_int || final_float)
1086 && inter_prec >= inside_prec
1087 && (inter_float || inter_unsignedp == inside_unsignedp)
1088 && ! (final_prec != GET_MODE_PRECISION (TYPE_MODE (type))
1089 && TYPE_MODE (type) == TYPE_MODE (inter_type)))
1092 /* If we have a sign-extension of a zero-extended value, we can
1093 replace that by a single zero-extension. Likewise if the
1094 final conversion does not change precision we can drop the
1095 intermediate conversion. */
1096 (if (inside_int && inter_int && final_int
1097 && ((inside_prec < inter_prec && inter_prec < final_prec
1098 && inside_unsignedp && !inter_unsignedp)
1099 || final_prec == inter_prec))
1102 /* Two conversions in a row are not needed unless:
1103 - some conversion is floating-point (overstrict for now), or
1104 - some conversion is a vector (overstrict for now), or
1105 - the intermediate type is narrower than both initial and
1107 - the intermediate type and innermost type differ in signedness,
1108 and the outermost type is wider than the intermediate, or
1109 - the initial type is a pointer type and the precisions of the
1110 intermediate and final types differ, or
1111 - the final type is a pointer type and the precisions of the
1112 initial and intermediate types differ. */
1113 (if (! inside_float && ! inter_float && ! final_float
1114 && ! inside_vec && ! inter_vec && ! final_vec
1115 && (inter_prec >= inside_prec || inter_prec >= final_prec)
1116 && ! (inside_int && inter_int
1117 && inter_unsignedp != inside_unsignedp
1118 && inter_prec < final_prec)
1119 && ((inter_unsignedp && inter_prec > inside_prec)
1120 == (final_unsignedp && final_prec > inter_prec))
1121 && ! (inside_ptr && inter_prec != final_prec)
1122 && ! (final_ptr && inside_prec != inter_prec)
1123 && ! (final_prec != GET_MODE_PRECISION (TYPE_MODE (type))
1124 && TYPE_MODE (type) == TYPE_MODE (inter_type)))
1127 /* A truncation to an unsigned type (a zero-extension) should be
1128 canonicalized as bitwise and of a mask. */
1129 (if (final_int && inter_int && inside_int
1130 && final_prec == inside_prec
1131 && final_prec > inter_prec
1133 (convert (bit_and @0 { wide_int_to_tree
1135 wi::mask (inter_prec, false,
1136 TYPE_PRECISION (inside_type))); })))
1138 /* If we are converting an integer to a floating-point that can
1139 represent it exactly and back to an integer, we can skip the
1140 floating-point conversion. */
1141 (if (GIMPLE /* PR66211 */
1142 && inside_int && inter_float && final_int &&
1143 (unsigned) significand_size (TYPE_MODE (inter_type))
1144 >= inside_prec - !inside_unsignedp)
1147 /* If we have a narrowing conversion to an integral type that is fed by a
1148 BIT_AND_EXPR, we might be able to remove the BIT_AND_EXPR if it merely
1149 masks off bits outside the final type (and nothing else). */
1151 (convert (bit_and @0 INTEGER_CST@1))
1152 (if (INTEGRAL_TYPE_P (type)
1153 && INTEGRAL_TYPE_P (TREE_TYPE (@0))
1154 && TYPE_PRECISION (type) <= TYPE_PRECISION (TREE_TYPE (@0))
1155 && operand_equal_p (@1, build_low_bits_mask (TREE_TYPE (@1),
1156 TYPE_PRECISION (type)), 0))
1160 /* (X /[ex] A) * A -> X. */
1162 (mult (convert? (exact_div @0 @1)) @1)
1163 /* Look through a sign-changing conversion. */
1166 /* Canonicalization of binary operations. */
1168 /* Convert X + -C into X - C. */
1170 (plus @0 REAL_CST@1)
1171 (if (REAL_VALUE_NEGATIVE (TREE_REAL_CST (@1)))
1172 (with { tree tem = fold_unary (NEGATE_EXPR, type, @1); }
1173 (if (!TREE_OVERFLOW (tem) || !flag_trapping_math)
1174 (minus @0 { tem; })))))
1176 /* Convert x+x into x*2.0. */
1179 (if (SCALAR_FLOAT_TYPE_P (type))
1180 (mult @0 { build_real (type, dconst2); })))
1183 (minus integer_zerop @1)
1186 /* (ARG0 - ARG1) is the same as (-ARG1 + ARG0). So check whether
1187 ARG0 is zero and X + ARG0 reduces to X, since that would mean
1188 (-ARG1 + ARG0) reduces to -ARG1. */
1190 (minus real_zerop@0 @1)
1191 (if (fold_real_zero_addition_p (type, @0, 0))
1194 /* Transform x * -1 into -x. */
1196 (mult @0 integer_minus_onep)
1199 /* COMPLEX_EXPR and REALPART/IMAGPART_EXPR cancellations. */
1201 (complex (realpart @0) (imagpart @0))
1204 (realpart (complex @0 @1))
1207 (imagpart (complex @0 @1))
1211 /* BSWAP simplifications, transforms checked by gcc.dg/builtin-bswap-8.c. */
1212 (for bswap (BUILT_IN_BSWAP16 BUILT_IN_BSWAP32 BUILT_IN_BSWAP64)
1217 (bswap (bit_not (bswap @0)))
1219 (for bitop (bit_xor bit_ior bit_and)
1221 (bswap (bitop:c (bswap @0) @1))
1222 (bitop @0 (bswap @1)))))
1225 /* Combine COND_EXPRs and VEC_COND_EXPRs. */
1227 /* Simplify constant conditions.
1228 Only optimize constant conditions when the selected branch
1229 has the same type as the COND_EXPR. This avoids optimizing
1230 away "c ? x : throw", where the throw has a void type.
1231 Note that we cannot throw away the fold-const.c variant nor
1232 this one as we depend on doing this transform before possibly
1233 A ? B : B -> B triggers and the fold-const.c one can optimize
1234 0 ? A : B to B even if A has side-effects. Something
1235 genmatch cannot handle. */
1237 (cond INTEGER_CST@0 @1 @2)
1238 (if (integer_zerop (@0)
1239 && (!VOID_TYPE_P (TREE_TYPE (@2))
1240 || VOID_TYPE_P (type)))
1242 (if (!integer_zerop (@0)
1243 && (!VOID_TYPE_P (TREE_TYPE (@1))
1244 || VOID_TYPE_P (type)))
1247 (vec_cond VECTOR_CST@0 @1 @2)
1248 (if (integer_all_onesp (@0))
1250 (if (integer_zerop (@0))
1253 (for cnd (cond vec_cond)
1254 /* A ? B : (A ? X : C) -> A ? B : C. */
1256 (cnd @0 (cnd @0 @1 @2) @3)
1259 (cnd @0 @1 (cnd @0 @2 @3))
1262 /* A ? B : B -> B. */
1267 /* !A ? B : C -> A ? C : B. */
1269 (cnd (logical_inverted_value truth_valued_p@0) @1 @2)
1272 /* A + (B vcmp C ? 1 : 0) -> A - (B vcmp C), since vector comparisons
1273 return all-1 or all-0 results. */
1274 /* ??? We could instead convert all instances of the vec_cond to negate,
1275 but that isn't necessarily a win on its own. */
1277 (plus:c @3 (view_convert? (vec_cond @0 integer_each_onep@1 integer_zerop@2)))
1278 (if (VECTOR_TYPE_P (type)
1279 && TYPE_VECTOR_SUBPARTS (type) == TYPE_VECTOR_SUBPARTS (TREE_TYPE (@0))
1280 && (TYPE_MODE (TREE_TYPE (type))
1281 == TYPE_MODE (TREE_TYPE (TREE_TYPE (@0)))))
1282 (minus @3 (view_convert @0))))
1284 /* ... likewise A - (B vcmp C ? 1 : 0) -> A + (B vcmp C). */
1286 (minus @3 (view_convert? (vec_cond @0 integer_each_onep@1 integer_zerop@2)))
1287 (if (VECTOR_TYPE_P (type)
1288 && TYPE_VECTOR_SUBPARTS (type) == TYPE_VECTOR_SUBPARTS (TREE_TYPE (@0))
1289 && (TYPE_MODE (TREE_TYPE (type))
1290 == TYPE_MODE (TREE_TYPE (TREE_TYPE (@0)))))
1291 (plus @3 (view_convert @0))))
1294 /* Simplifications of comparisons. */
1296 /* We can simplify a logical negation of a comparison to the
1297 inverted comparison. As we cannot compute an expression
1298 operator using invert_tree_comparison we have to simulate
1299 that with expression code iteration. */
1300 (for cmp (tcc_comparison)
1301 icmp (inverted_tcc_comparison)
1302 ncmp (inverted_tcc_comparison_with_nans)
1303 /* Ideally we'd like to combine the following two patterns
1304 and handle some more cases by using
1305 (logical_inverted_value (cmp @0 @1))
1306 here but for that genmatch would need to "inline" that.
1307 For now implement what forward_propagate_comparison did. */
1309 (bit_not (cmp @0 @1))
1310 (if (VECTOR_TYPE_P (type)
1311 || (INTEGRAL_TYPE_P (type) && TYPE_PRECISION (type) == 1))
1312 /* Comparison inversion may be impossible for trapping math,
1313 invert_tree_comparison will tell us. But we can't use
1314 a computed operator in the replacement tree thus we have
1315 to play the trick below. */
1316 (with { enum tree_code ic = invert_tree_comparison
1317 (cmp, HONOR_NANS (@0)); }
1323 (bit_xor (cmp @0 @1) integer_truep)
1324 (with { enum tree_code ic = invert_tree_comparison
1325 (cmp, HONOR_NANS (@0)); }
1331 /* Transform comparisons of the form X - Y CMP 0 to X CMP Y.
1332 ??? The transformation is valid for the other operators if overflow
1333 is undefined for the type, but performing it here badly interacts
1334 with the transformation in fold_cond_expr_with_comparison which
1335 attempts to synthetize ABS_EXPR. */
1338 (cmp (minus @0 @1) integer_zerop)
1341 /* Transform comparisons of the form X * C1 CMP 0 to X CMP 0 in the
1342 signed arithmetic case. That form is created by the compiler
1343 often enough for folding it to be of value. One example is in
1344 computing loop trip counts after Operator Strength Reduction. */
1345 (for cmp (simple_comparison)
1346 scmp (swapped_simple_comparison)
1348 (cmp (mult @0 INTEGER_CST@1) integer_zerop@2)
1349 /* Handle unfolded multiplication by zero. */
1350 (if (integer_zerop (@1))
1352 (if (ANY_INTEGRAL_TYPE_P (TREE_TYPE (@0))
1353 && TYPE_OVERFLOW_UNDEFINED (TREE_TYPE (@0)))
1354 /* If @1 is negative we swap the sense of the comparison. */
1355 (if (tree_int_cst_sgn (@1) < 0)
1359 /* Simplify comparison of something with itself. For IEEE
1360 floating-point, we can only do some of these simplifications. */
1363 (if (! FLOAT_TYPE_P (TREE_TYPE (@0))
1364 || ! HONOR_NANS (TYPE_MODE (TREE_TYPE (@0))))
1365 { constant_boolean_node (true, type); }))
1374 || ! FLOAT_TYPE_P (TREE_TYPE (@0))
1375 || ! HONOR_NANS (TYPE_MODE (TREE_TYPE (@0))))
1376 { constant_boolean_node (false, type); })))
1378 /* Fold ~X op ~Y as Y op X. */
1379 (for cmp (simple_comparison)
1381 (cmp (bit_not @0) (bit_not @1))
1384 /* Fold ~X op C as X op' ~C, where op' is the swapped comparison. */
1385 (for cmp (simple_comparison)
1386 scmp (swapped_simple_comparison)
1388 (cmp (bit_not @0) CONSTANT_CLASS_P@1)
1389 (if (TREE_CODE (@1) == INTEGER_CST || TREE_CODE (@1) == VECTOR_CST)
1390 (scmp @0 (bit_not @1)))))
1392 (for cmp (simple_comparison)
1393 /* Fold (double)float1 CMP (double)float2 into float1 CMP float2. */
1395 (cmp (convert@2 @0) (convert? @1))
1396 (if (FLOAT_TYPE_P (TREE_TYPE (@0))
1397 && (DECIMAL_FLOAT_TYPE_P (TREE_TYPE (@2))
1398 == DECIMAL_FLOAT_TYPE_P (TREE_TYPE (@0)))
1399 && (DECIMAL_FLOAT_TYPE_P (TREE_TYPE (@2))
1400 == DECIMAL_FLOAT_TYPE_P (TREE_TYPE (@1))))
1403 tree type1 = TREE_TYPE (@1);
1404 if (TREE_CODE (@1) == REAL_CST && !DECIMAL_FLOAT_TYPE_P (type1))
1406 REAL_VALUE_TYPE orig = TREE_REAL_CST (@1);
1407 if (TYPE_PRECISION (type1) > TYPE_PRECISION (float_type_node)
1408 && exact_real_truncate (TYPE_MODE (float_type_node), &orig))
1409 type1 = float_type_node;
1410 if (TYPE_PRECISION (type1) > TYPE_PRECISION (double_type_node)
1411 && exact_real_truncate (TYPE_MODE (double_type_node), &orig))
1412 type1 = double_type_node;
1415 = (TYPE_PRECISION (TREE_TYPE (@0)) > TYPE_PRECISION (type1)
1416 ? TREE_TYPE (@0) : type1);
1418 (if (TYPE_PRECISION (TREE_TYPE (@2)) > TYPE_PRECISION (newtype))
1419 (cmp (convert:newtype @0) (convert:newtype @1))))))
1423 /* IEEE doesn't distinguish +0 and -0 in comparisons. */
1424 /* a CMP (-0) -> a CMP 0 */
1425 (if (REAL_VALUE_MINUS_ZERO (TREE_REAL_CST (@1)))
1426 (cmp @0 { build_real (TREE_TYPE (@1), dconst0); }))
1427 /* x != NaN is always true, other ops are always false. */
1428 (if (REAL_VALUE_ISNAN (TREE_REAL_CST (@1))
1429 && ! HONOR_SNANS (@1))
1430 { constant_boolean_node (cmp == NE_EXPR, type); })
1431 /* Fold comparisons against infinity. */
1432 (if (REAL_VALUE_ISINF (TREE_REAL_CST (@1))
1433 && MODE_HAS_INFINITIES (TYPE_MODE (TREE_TYPE (@1))))
1436 REAL_VALUE_TYPE max;
1437 enum tree_code code = cmp;
1438 bool neg = REAL_VALUE_NEGATIVE (TREE_REAL_CST (@1));
1440 code = swap_tree_comparison (code);
1442 /* x > +Inf is always false, if with ignore sNANs. */
1443 (if (code == GT_EXPR
1444 && ! HONOR_SNANS (@0))
1445 { constant_boolean_node (false, type); })
1446 (if (code == LE_EXPR)
1447 /* x <= +Inf is always true, if we don't case about NaNs. */
1448 (if (! HONOR_NANS (@0))
1449 { constant_boolean_node (true, type); })
1450 /* x <= +Inf is the same as x == x, i.e. isfinite(x). */
1452 /* x == +Inf and x >= +Inf are always equal to x > DBL_MAX. */
1453 (if (code == EQ_EXPR || code == GE_EXPR)
1454 (with { real_maxval (&max, neg, TYPE_MODE (TREE_TYPE (@0))); }
1456 (lt @0 { build_real (TREE_TYPE (@0), max); }))
1457 (gt @0 { build_real (TREE_TYPE (@0), max); })))
1458 /* x < +Inf is always equal to x <= DBL_MAX. */
1459 (if (code == LT_EXPR)
1460 (with { real_maxval (&max, neg, TYPE_MODE (TREE_TYPE (@0))); }
1462 (ge @0 { build_real (TREE_TYPE (@0), max); }))
1463 (le @0 { build_real (TREE_TYPE (@0), max); })))
1464 /* x != +Inf is always equal to !(x > DBL_MAX). */
1465 (if (code == NE_EXPR)
1466 (with { real_maxval (&max, neg, TYPE_MODE (TREE_TYPE (@0))); }
1467 (if (! HONOR_NANS (@0))
1469 (ge @0 { build_real (TREE_TYPE (@0), max); }))
1470 (le @0 { build_real (TREE_TYPE (@0), max); }))
1472 (bit_xor (lt @0 { build_real (TREE_TYPE (@0), max); })
1473 { build_one_cst (type); }))
1474 (bit_xor (gt @0 { build_real (TREE_TYPE (@0), max); })
1475 { build_one_cst (type); }))))))
1477 /* If this is a comparison of a real constant with a PLUS_EXPR
1478 or a MINUS_EXPR of a real constant, we can convert it into a
1479 comparison with a revised real constant as long as no overflow
1480 occurs when unsafe_math_optimizations are enabled. */
1481 (if (flag_unsafe_math_optimizations)
1482 (for op (plus minus)
1484 (cmp (op @0 REAL_CST@1) REAL_CST@2)
1487 tree tem = const_binop (op == PLUS_EXPR ? MINUS_EXPR : PLUS_EXPR,
1488 TREE_TYPE (@1), @2, @1);
1490 (if (!TREE_OVERFLOW (tem))
1491 (cmp @0 { tem; }))))))
1493 /* Likewise, we can simplify a comparison of a real constant with
1494 a MINUS_EXPR whose first operand is also a real constant, i.e.
1495 (c1 - x) < c2 becomes x > c1-c2. Reordering is allowed on
1496 floating-point types only if -fassociative-math is set. */
1497 (if (flag_associative_math)
1499 (cmp (minus REAL_CST@0 @1) @2)
1500 (with { tree tem = const_binop (MINUS_EXPR, TREE_TYPE (@1), @0, @2); }
1501 (if (!TREE_OVERFLOW (tem))
1502 (cmp { tem; } @1)))))
1504 /* Fold comparisons against built-in math functions. */
1505 (if (flag_unsafe_math_optimizations
1506 && ! flag_errno_math)
1509 (cmp (sq @0) REAL_CST@1)
1510 (if (REAL_VALUE_NEGATIVE (TREE_REAL_CST (@1)))
1511 /* sqrt(x) < y is always false, if y is negative. */
1512 (if (cmp == EQ_EXPR || cmp == LT_EXPR || cmp == LE_EXPR)
1513 { constant_boolean_node (false, type); })
1514 /* sqrt(x) > y is always true, if y is negative and we
1515 don't care about NaNs, i.e. negative values of x. */
1516 (if (cmp == NE_EXPR || !HONOR_NANS (@0))
1517 { constant_boolean_node (true, type); })
1518 /* sqrt(x) > y is the same as x >= 0, if y is negative. */
1519 (ge @0 { build_real (TREE_TYPE (@0), dconst0); }))
1520 (if (cmp == GT_EXPR || cmp == GE_EXPR)
1524 REAL_ARITHMETIC (c2, MULT_EXPR, TREE_REAL_CST (@1), TREE_REAL_CST (@1));
1525 real_convert (&c2, TYPE_MODE (TREE_TYPE (@0)), &c2);
1527 (if (REAL_VALUE_ISINF (c2))
1528 /* sqrt(x) > y is x == +Inf, when y is very large. */
1529 (if (HONOR_INFINITIES (@0))
1530 (eq @0 { build_real (TREE_TYPE (@0), c2); }))
1531 { constant_boolean_node (false, type); })
1532 /* sqrt(x) > c is the same as x > c*c. */
1533 (cmp @0 { build_real (TREE_TYPE (@0), c2); })))
1534 (if (cmp == LT_EXPR || cmp == LE_EXPR)
1538 REAL_ARITHMETIC (c2, MULT_EXPR, TREE_REAL_CST (@1), TREE_REAL_CST (@1));
1539 real_convert (&c2, TYPE_MODE (TREE_TYPE (@0)), &c2);
1541 (if (REAL_VALUE_ISINF (c2))
1542 /* sqrt(x) < y is always true, when y is a very large
1543 value and we don't care about NaNs or Infinities. */
1544 (if (! HONOR_NANS (@0) && ! HONOR_INFINITIES (@0))
1545 { constant_boolean_node (true, type); })
1546 /* sqrt(x) < y is x != +Inf when y is very large and we
1547 don't care about NaNs. */
1548 (if (! HONOR_NANS (@0))
1549 (ne @0 { build_real (TREE_TYPE (@0), c2); }))
1550 /* sqrt(x) < y is x >= 0 when y is very large and we
1551 don't care about Infinities. */
1552 (if (! HONOR_INFINITIES (@0))
1553 (ge @0 { build_real (TREE_TYPE (@0), dconst0); }))
1554 /* sqrt(x) < y is x >= 0 && x != +Inf, when y is large. */
1557 (ge @0 { build_real (TREE_TYPE (@0), dconst0); })
1558 (ne @0 { build_real (TREE_TYPE (@0), c2); }))))
1559 /* sqrt(x) < c is the same as x < c*c, if we ignore NaNs. */
1560 (if (! REAL_VALUE_ISINF (c2)
1561 && ! HONOR_NANS (@0))
1562 (cmp @0 { build_real (TREE_TYPE (@0), c2); }))
1563 /* sqrt(x) < c is the same as x >= 0 && x < c*c. */
1564 (if (! REAL_VALUE_ISINF (c2)
1567 (ge @0 { build_real (TREE_TYPE (@0), dconst0); })
1568 (cmp @0 { build_real (TREE_TYPE (@0), c2); })))))))))
1570 /* Unordered tests if either argument is a NaN. */
1572 (bit_ior (unordered @0 @0) (unordered @1 @1))
1573 (if (types_match (@0, @1))
1576 (bit_and (ordered @0 @0) (ordered @1 @1))
1577 (if (types_match (@0, @1))
1580 (bit_ior:c (unordered @0 @0) (unordered:c@2 @0 @1))
1583 (bit_and:c (ordered @0 @0) (ordered:c@2 @0 @1))
1586 /* -A CMP -B -> B CMP A. */
1587 (for cmp (tcc_comparison)
1588 scmp (swapped_tcc_comparison)
1590 (cmp (negate @0) (negate @1))
1591 (if (FLOAT_TYPE_P (TREE_TYPE (@0))
1592 || (ANY_INTEGRAL_TYPE_P (TREE_TYPE (@0))
1593 && TYPE_OVERFLOW_UNDEFINED (TREE_TYPE (@0))))
1596 (cmp (negate @0) CONSTANT_CLASS_P@1)
1597 (if (FLOAT_TYPE_P (TREE_TYPE (@0))
1598 || (ANY_INTEGRAL_TYPE_P (TREE_TYPE (@0))
1599 && TYPE_OVERFLOW_UNDEFINED (TREE_TYPE (@0))))
1600 (with { tree tem = fold_unary (NEGATE_EXPR, TREE_TYPE (@0), @1); }
1601 (if (tem && !TREE_OVERFLOW (tem))
1602 (scmp @0 { tem; }))))))
1605 /* Equality compare simplifications from fold_binary */
1608 /* If we have (A | C) == D where C & ~D != 0, convert this into 0.
1609 Similarly for NE_EXPR. */
1611 (cmp (convert?@3 (bit_ior @0 INTEGER_CST@1)) INTEGER_CST@2)
1612 (if (tree_nop_conversion_p (TREE_TYPE (@3), TREE_TYPE (@0))
1613 && wi::bit_and_not (@1, @2) != 0)
1614 { constant_boolean_node (cmp == NE_EXPR, type); }))
1616 /* (X ^ Y) == 0 becomes X == Y, and (X ^ Y) != 0 becomes X != Y. */
1618 (cmp (bit_xor @0 @1) integer_zerop)
1621 /* (X ^ Y) == Y becomes X == 0.
1622 Likewise (X ^ Y) == X becomes Y == 0. */
1624 (cmp:c (bit_xor:c @0 @1) @0)
1625 (cmp @1 { build_zero_cst (TREE_TYPE (@1)); }))
1627 /* (X ^ C1) op C2 can be rewritten as X op (C1 ^ C2). */
1629 (cmp (convert?@3 (bit_xor @0 INTEGER_CST@1)) INTEGER_CST@2)
1630 (if (tree_nop_conversion_p (TREE_TYPE (@3), TREE_TYPE (@0)))
1631 (cmp @0 (bit_xor @1 (convert @2))))))
1633 /* Simplification of math builtins. */
1635 /* fold_builtin_logarithm */
1636 (if (flag_unsafe_math_optimizations)
1637 /* Special case, optimize logN(expN(x)) = x. */
1638 (for logs (LOG LOG2 LOG10)
1639 exps (EXP EXP2 EXP10)
1643 /* Optimize logN(func()) for various exponential functions. We
1644 want to determine the value "x" and the power "exponent" in
1645 order to transform logN(x**exponent) into exponent*logN(x). */
1646 (for logs (LOG LOG LOG LOG
1648 LOG10 LOG10 LOG10 LOG10)
1649 exps (EXP EXP2 EXP10 POW10)
1656 CASE_FLT_FN (BUILT_IN_EXP):
1657 /* Prepare to do logN(exp(exponent) -> exponent*logN(e). */
1658 x = build_real (type, real_value_truncate (TYPE_MODE (type),
1661 CASE_FLT_FN (BUILT_IN_EXP2):
1662 /* Prepare to do logN(exp2(exponent) -> exponent*logN(2). */
1663 x = build_real (type, dconst2);
1665 CASE_FLT_FN (BUILT_IN_EXP10):
1666 CASE_FLT_FN (BUILT_IN_POW10):
1667 /* Prepare to do logN(exp10(exponent) -> exponent*logN(10). */
1669 REAL_VALUE_TYPE dconst10;
1670 real_from_integer (&dconst10, VOIDmode, 10, SIGNED);
1671 x = build_real (type, dconst10);
1676 (mult (logs { x; }) @0))))
1687 CASE_FLT_FN (BUILT_IN_SQRT):
1688 /* Prepare to do logN(sqrt(x) -> 0.5*logN(x). */
1689 x = build_real (type, dconsthalf);
1691 CASE_FLT_FN (BUILT_IN_CBRT):
1692 /* Prepare to do logN(cbrt(x) -> (1/3)*logN(x). */
1693 x = build_real (type, real_value_truncate (TYPE_MODE (type),
1698 (mult { x; } (logs @0)))))
1699 /* logN(pow(x,exponent) -> exponent*logN(x). */
1700 (for logs (LOG LOG2 LOG10)
1704 (mult @1 (logs @0)))))
1706 /* Narrowing of arithmetic and logical operations.
1708 These are conceptually similar to the transformations performed for
1709 the C/C++ front-ends by shorten_binary_op and shorten_compare. Long
1710 term we want to move all that code out of the front-ends into here. */
1712 /* If we have a narrowing conversion of an arithmetic operation where
1713 both operands are widening conversions from the same type as the outer
1714 narrowing conversion. Then convert the innermost operands to a suitable
1715 unsigned type (to avoid introducing undefined behaviour), perform the
1716 operation and convert the result to the desired type. */
1717 (for op (plus minus)
1719 (convert (op@4 (convert@2 @0) (convert@3 @1)))
1720 (if (INTEGRAL_TYPE_P (type)
1721 /* We check for type compatibility between @0 and @1 below,
1722 so there's no need to check that @1/@3 are integral types. */
1723 && INTEGRAL_TYPE_P (TREE_TYPE (@0))
1724 && INTEGRAL_TYPE_P (TREE_TYPE (@2))
1725 /* The precision of the type of each operand must match the
1726 precision of the mode of each operand, similarly for the
1728 && (TYPE_PRECISION (TREE_TYPE (@0))
1729 == GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (@0))))
1730 && (TYPE_PRECISION (TREE_TYPE (@1))
1731 == GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (@1))))
1732 && TYPE_PRECISION (type) == GET_MODE_PRECISION (TYPE_MODE (type))
1733 /* The inner conversion must be a widening conversion. */
1734 && TYPE_PRECISION (TREE_TYPE (@2)) > TYPE_PRECISION (TREE_TYPE (@0))
1735 && types_match (@0, @1)
1736 && types_match (@0, type)
1738 (if (TYPE_OVERFLOW_WRAPS (TREE_TYPE (@0)))
1739 (convert (op @0 @1)))
1740 (with { tree utype = unsigned_type_for (TREE_TYPE (@0)); }
1741 (convert (op (convert:utype @0) (convert:utype @1)))))))
1743 /* This is another case of narrowing, specifically when there's an outer
1744 BIT_AND_EXPR which masks off bits outside the type of the innermost
1745 operands. Like the previous case we have to convert the operands
1746 to unsigned types to avoid introducing undefined behaviour for the
1747 arithmetic operation. */
1748 (for op (minus plus)
1750 (bit_and (op@5 (convert@2 @0) (convert@3 @1)) INTEGER_CST@4)
1751 (if (INTEGRAL_TYPE_P (type)
1752 /* We check for type compatibility between @0 and @1 below,
1753 so there's no need to check that @1/@3 are integral types. */
1754 && INTEGRAL_TYPE_P (TREE_TYPE (@0))
1755 && INTEGRAL_TYPE_P (TREE_TYPE (@2))
1756 /* The precision of the type of each operand must match the
1757 precision of the mode of each operand, similarly for the
1759 && (TYPE_PRECISION (TREE_TYPE (@0))
1760 == GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (@0))))
1761 && (TYPE_PRECISION (TREE_TYPE (@1))
1762 == GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (@1))))
1763 && TYPE_PRECISION (type) == GET_MODE_PRECISION (TYPE_MODE (type))
1764 /* The inner conversion must be a widening conversion. */
1765 && TYPE_PRECISION (TREE_TYPE (@2)) > TYPE_PRECISION (TREE_TYPE (@0))
1766 && types_match (@0, @1)
1767 && (tree_int_cst_min_precision (@4, TYPE_SIGN (TREE_TYPE (@0)))
1768 <= TYPE_PRECISION (TREE_TYPE (@0)))
1769 && (TYPE_OVERFLOW_WRAPS (TREE_TYPE (@0))
1770 || tree_int_cst_sgn (@4) >= 0)
1772 (if (TYPE_OVERFLOW_WRAPS (TREE_TYPE (@0)))
1773 (with { tree ntype = TREE_TYPE (@0); }
1774 (convert (bit_and (op @0 @1) (convert:ntype @4)))))
1775 (with { tree utype = unsigned_type_for (TREE_TYPE (@0)); }
1776 (convert (bit_and (op (convert:utype @0) (convert:utype @1))
1777 (convert:utype @4)))))))