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)
45 /* Simplifications of operations with one constant operand and
46 simplifications to constants or single values. */
48 (for op (plus pointer_plus minus bit_ior bit_xor)
53 /* 0 +p index -> (type)index */
55 (pointer_plus integer_zerop @1)
56 (non_lvalue (convert @1)))
58 /* See if ARG1 is zero and X + ARG1 reduces to X.
59 Likewise if the operands are reversed. */
61 (plus:c @0 real_zerop@1)
62 (if (fold_real_zero_addition_p (type, @1, 0))
65 /* See if ARG1 is zero and X - ARG1 reduces to X. */
67 (minus @0 real_zerop@1)
68 (if (fold_real_zero_addition_p (type, @1, 1))
72 This is unsafe for certain floats even in non-IEEE formats.
73 In IEEE, it is unsafe because it does wrong for NaNs.
74 Also note that operand_equal_p is always false if an operand
78 (if (!FLOAT_TYPE_P (type) || !HONOR_NANS (type))
79 { build_zero_cst (type); }))
82 (mult @0 integer_zerop@1)
85 /* Maybe fold x * 0 to 0. The expressions aren't the same
86 when x is NaN, since x * 0 is also NaN. Nor are they the
87 same in modes with signed zeros, since multiplying a
88 negative value by 0 gives -0, not +0. */
90 (mult @0 real_zerop@1)
91 (if (!HONOR_NANS (type) && !HONOR_SIGNED_ZEROS (element_mode (type)))
94 /* In IEEE floating point, x*1 is not equivalent to x for snans.
95 Likewise for complex arithmetic with signed zeros. */
98 (if (!HONOR_SNANS (element_mode (type))
99 && (!HONOR_SIGNED_ZEROS (element_mode (type))
100 || !COMPLEX_FLOAT_TYPE_P (type)))
103 /* Transform x * -1.0 into -x. */
105 (mult @0 real_minus_onep)
106 (if (!HONOR_SNANS (element_mode (type))
107 && (!HONOR_SIGNED_ZEROS (element_mode (type))
108 || !COMPLEX_FLOAT_TYPE_P (type)))
111 /* Make sure to preserve divisions by zero. This is the reason why
112 we don't simplify x / x to 1 or 0 / x to 0. */
113 (for op (mult trunc_div ceil_div floor_div round_div exact_div)
119 (for div (trunc_div ceil_div floor_div round_div exact_div)
121 (div @0 integer_minus_onep@1)
122 (if (!TYPE_UNSIGNED (type))
125 /* For unsigned integral types, FLOOR_DIV_EXPR is the same as
126 TRUNC_DIV_EXPR. Rewrite into the latter in this case. */
129 (if ((INTEGRAL_TYPE_P (type) || VECTOR_INTEGER_TYPE_P (type))
130 && TYPE_UNSIGNED (type))
133 /* Combine two successive divisions. Note that combining ceil_div
134 and floor_div is trickier and combining round_div even more so. */
135 (for div (trunc_div exact_div)
137 (div (div @0 INTEGER_CST@1) INTEGER_CST@2)
140 wide_int mul = wi::mul (@1, @2, TYPE_SIGN (type), &overflow_p);
143 (div @0 { wide_int_to_tree (type, mul); }))
145 && (TYPE_UNSIGNED (type)
146 || mul != wi::min_value (TYPE_PRECISION (type), SIGNED)))
147 { build_zero_cst (type); }))))
149 /* Optimize A / A to 1.0 if we don't care about
150 NaNs or Infinities. */
153 (if (FLOAT_TYPE_P (type)
154 && ! HONOR_NANS (type)
155 && ! HONOR_INFINITIES (element_mode (type)))
156 { build_one_cst (type); }))
158 /* Optimize -A / A to -1.0 if we don't care about
159 NaNs or Infinities. */
161 (rdiv:c @0 (negate @0))
162 (if (FLOAT_TYPE_P (type)
163 && ! HONOR_NANS (type)
164 && ! HONOR_INFINITIES (element_mode (type)))
165 { build_minus_one_cst (type); }))
167 /* In IEEE floating point, x/1 is not equivalent to x for snans. */
170 (if (!HONOR_SNANS (element_mode (type)))
173 /* In IEEE floating point, x/-1 is not equivalent to -x for snans. */
175 (rdiv @0 real_minus_onep)
176 (if (!HONOR_SNANS (element_mode (type)))
179 /* If ARG1 is a constant, we can convert this to a multiply by the
180 reciprocal. This does not have the same rounding properties,
181 so only do this if -freciprocal-math. We can actually
182 always safely do it if ARG1 is a power of two, but it's hard to
183 tell if it is or not in a portable manner. */
184 (for cst (REAL_CST COMPLEX_CST VECTOR_CST)
188 (if (flag_reciprocal_math
191 { tree tem = const_binop (RDIV_EXPR, type, build_one_cst (type), @1); }
193 (mult @0 { tem; } ))))
194 (if (cst != COMPLEX_CST)
195 (with { tree inverse = exact_inverse (type, @1); }
197 (mult @0 { inverse; } )))))))
199 /* Same applies to modulo operations, but fold is inconsistent here
200 and simplifies 0 % x to 0, only preserving literal 0 % 0. */
201 (for mod (ceil_mod floor_mod round_mod trunc_mod)
202 /* 0 % X is always zero. */
204 (mod integer_zerop@0 @1)
205 /* But not for 0 % 0 so that we can get the proper warnings and errors. */
206 (if (!integer_zerop (@1))
208 /* X % 1 is always zero. */
210 (mod @0 integer_onep)
211 { build_zero_cst (type); })
212 /* X % -1 is zero. */
214 (mod @0 integer_minus_onep@1)
215 (if (!TYPE_UNSIGNED (type))
216 { build_zero_cst (type); }))
217 /* (X % Y) % Y is just X % Y. */
219 (mod (mod@2 @0 @1) @1)
222 /* X % -C is the same as X % C. */
224 (trunc_mod @0 INTEGER_CST@1)
225 (if (TYPE_SIGN (type) == SIGNED
226 && !TREE_OVERFLOW (@1)
228 && !TYPE_OVERFLOW_TRAPS (type)
229 /* Avoid this transformation if C is INT_MIN, i.e. C == -C. */
230 && !sign_bit_p (@1, @1))
231 (trunc_mod @0 (negate @1))))
233 /* X % -Y is the same as X % Y. */
235 (trunc_mod @0 (convert? (negate @1)))
236 (if (!TYPE_UNSIGNED (type)
237 && !TYPE_OVERFLOW_TRAPS (type)
238 && tree_nop_conversion_p (type, TREE_TYPE (@1)))
239 (trunc_mod @0 (convert @1))))
241 /* Optimize TRUNC_MOD_EXPR by a power of two into a BIT_AND_EXPR,
242 i.e. "X % C" into "X & (C - 1)", if X and C are positive.
243 Also optimize A % (C << N) where C is a power of 2,
244 to A & ((C << N) - 1). */
245 (match (power_of_two_cand @1)
247 (match (power_of_two_cand @1)
248 (lshift INTEGER_CST@1 @2))
249 (for mod (trunc_mod floor_mod)
251 (mod @0 (convert?@3 (power_of_two_cand@1 @2)))
252 (if ((TYPE_UNSIGNED (type)
253 || tree_expr_nonnegative_p (@0))
254 && tree_nop_conversion_p (type, TREE_TYPE (@3))
255 && integer_pow2p (@2) && tree_int_cst_sgn (@2) > 0)
256 (bit_and @0 (convert (minus @1 { build_int_cst (TREE_TYPE (@1), 1); }))))))
258 /* X % Y is smaller than Y. */
261 (cmp (trunc_mod @0 @1) @1)
262 (if (TYPE_UNSIGNED (TREE_TYPE (@0)))
263 { constant_boolean_node (cmp == LT_EXPR, type); })))
266 (cmp @1 (trunc_mod @0 @1))
267 (if (TYPE_UNSIGNED (TREE_TYPE (@0)))
268 { constant_boolean_node (cmp == GT_EXPR, type); })))
272 (bit_ior @0 integer_all_onesp@1)
277 (bit_and @0 integer_zerop@1)
283 { build_zero_cst (type); })
285 /* Canonicalize X ^ ~0 to ~X. */
287 (bit_xor @0 integer_all_onesp@1)
292 (bit_and @0 integer_all_onesp)
295 /* x & x -> x, x | x -> x */
296 (for bitop (bit_and bit_ior)
301 /* x + (x & 1) -> (x + 1) & ~1 */
303 (plus:c @0 (bit_and@2 @0 integer_onep@1))
304 (if (single_use (@2))
305 (bit_and (plus @0 @1) (bit_not @1))))
307 /* x & ~(x & y) -> x & ~y */
308 /* x | ~(x | y) -> x | ~y */
309 (for bitop (bit_and bit_ior)
311 (bitop:c @0 (bit_not (bitop:c@2 @0 @1)))
312 (if (single_use (@2))
313 (bitop @0 (bit_not @1)))))
315 /* (x | y) & ~x -> y & ~x */
316 /* (x & y) | ~x -> y | ~x */
317 (for bitop (bit_and bit_ior)
318 rbitop (bit_ior bit_and)
320 (bitop:c (rbitop:c @0 @1) (bit_not@2 @0))
323 /* (x & y) ^ (x | y) -> x ^ y */
325 (bit_xor:c (bit_and @0 @1) (bit_ior @0 @1))
328 /* (x ^ y) ^ (x | y) -> x & y */
330 (bit_xor:c (bit_xor @0 @1) (bit_ior @0 @1))
333 /* (x & y) + (x ^ y) -> x | y */
334 /* (x & y) | (x ^ y) -> x | y */
335 /* (x & y) ^ (x ^ y) -> x | y */
336 (for op (plus bit_ior bit_xor)
338 (op:c (bit_and @0 @1) (bit_xor @0 @1))
341 /* (x & y) + (x | y) -> x + y */
343 (plus:c (bit_and @0 @1) (bit_ior @0 @1))
346 /* (x + y) - (x | y) -> x & y */
348 (minus (plus @0 @1) (bit_ior @0 @1))
349 (if (!TYPE_OVERFLOW_SANITIZED (type) && !TYPE_OVERFLOW_TRAPS (type)
350 && !TYPE_SATURATING (type))
353 /* (x + y) - (x & y) -> x | y */
355 (minus (plus @0 @1) (bit_and @0 @1))
356 (if (!TYPE_OVERFLOW_SANITIZED (type) && !TYPE_OVERFLOW_TRAPS (type)
357 && !TYPE_SATURATING (type))
360 /* (x | y) - (x ^ y) -> x & y */
362 (minus (bit_ior @0 @1) (bit_xor @0 @1))
365 /* (x | y) - (x & y) -> x ^ y */
367 (minus (bit_ior @0 @1) (bit_and @0 @1))
374 (abs tree_expr_nonnegative_p@0)
378 /* Try to fold (type) X op CST -> (type) (X op ((type-x) CST))
380 For bitwise binary operations apply operand conversions to the
381 binary operation result instead of to the operands. This allows
382 to combine successive conversions and bitwise binary operations.
383 We combine the above two cases by using a conditional convert. */
384 (for bitop (bit_and bit_ior bit_xor)
386 (bitop (convert @0) (convert? @1))
387 (if (((TREE_CODE (@1) == INTEGER_CST
388 && INTEGRAL_TYPE_P (TREE_TYPE (@0))
389 && int_fits_type_p (@1, TREE_TYPE (@0)))
390 || types_match (@0, @1))
391 /* ??? This transform conflicts with fold-const.c doing
392 Convert (T)(x & c) into (T)x & (T)c, if c is an integer
393 constants (if x has signed type, the sign bit cannot be set
394 in c). This folds extension into the BIT_AND_EXPR.
395 Restrict it to GIMPLE to avoid endless recursions. */
396 && (bitop != BIT_AND_EXPR || GIMPLE)
397 && (/* That's a good idea if the conversion widens the operand, thus
398 after hoisting the conversion the operation will be narrower. */
399 TYPE_PRECISION (TREE_TYPE (@0)) < TYPE_PRECISION (type)
400 /* It's also a good idea if the conversion is to a non-integer
402 || GET_MODE_CLASS (TYPE_MODE (type)) != MODE_INT
403 /* Or if the precision of TO is not the same as the precision
405 || TYPE_PRECISION (type) != GET_MODE_PRECISION (TYPE_MODE (type))))
406 (convert (bitop @0 (convert @1))))))
408 /* Simplify (A & B) OP0 (C & B) to (A OP0 C) & B. */
409 (for bitop (bit_and bit_ior bit_xor)
411 (bitop (bit_and:c @0 @1) (bit_and @2 @1))
412 (bit_and (bitop @0 @2) @1)))
414 /* (x | CST1) & CST2 -> (x & CST2) | (CST1 & CST2) */
416 (bit_and (bit_ior @0 CONSTANT_CLASS_P@1) CONSTANT_CLASS_P@2)
417 (bit_ior (bit_and @0 @2) (bit_and @1 @2)))
419 /* Combine successive equal operations with constants. */
420 (for bitop (bit_and bit_ior bit_xor)
422 (bitop (bitop @0 CONSTANT_CLASS_P@1) CONSTANT_CLASS_P@2)
423 (bitop @0 (bitop @1 @2))))
425 /* Try simple folding for X op !X, and X op X with the help
426 of the truth_valued_p and logical_inverted_value predicates. */
427 (match truth_valued_p
429 (if (INTEGRAL_TYPE_P (type) && TYPE_PRECISION (type) == 1)))
430 (for op (tcc_comparison truth_and truth_andif truth_or truth_orif truth_xor)
431 (match truth_valued_p
433 (match truth_valued_p
436 (match (logical_inverted_value @0)
437 (bit_not truth_valued_p@0))
438 (match (logical_inverted_value @0)
439 (eq @0 integer_zerop))
440 (match (logical_inverted_value @0)
441 (ne truth_valued_p@0 integer_truep))
442 (match (logical_inverted_value @0)
443 (bit_xor truth_valued_p@0 integer_truep))
447 (bit_and:c @0 (logical_inverted_value @0))
448 { build_zero_cst (type); })
449 /* X | !X and X ^ !X -> 1, , if X is truth-valued. */
450 (for op (bit_ior bit_xor)
452 (op:c truth_valued_p@0 (logical_inverted_value @0))
453 { constant_boolean_node (true, type); }))
455 (for bitop (bit_and bit_ior)
456 rbitop (bit_ior bit_and)
457 /* (x | y) & x -> x */
458 /* (x & y) | x -> x */
460 (bitop:c (rbitop:c @0 @1) @0)
462 /* (~x | y) & x -> x & y */
463 /* (~x & y) | x -> x | y */
465 (bitop:c (rbitop:c (bit_not @0) @1) @0)
468 /* If arg1 and arg2 are booleans (or any single bit type)
469 then try to simplify:
476 But only do this if our result feeds into a comparison as
477 this transformation is not always a win, particularly on
478 targets with and-not instructions.
479 -> simplify_bitwise_binary_boolean */
481 (ne (bit_and:c (bit_not @0) @1) integer_zerop)
482 (if (INTEGRAL_TYPE_P (TREE_TYPE (@1))
483 && TYPE_PRECISION (TREE_TYPE (@1)) == 1)
486 (ne (bit_ior:c (bit_not @0) @1) integer_zerop)
487 (if (INTEGRAL_TYPE_P (TREE_TYPE (@1))
488 && TYPE_PRECISION (TREE_TYPE (@1)) == 1)
493 (bit_not (bit_not @0))
496 /* (x & ~m) | (y & m) -> ((x ^ y) & m) ^ x */
498 (bit_ior:c (bit_and:c@3 @0 (bit_not @2)) (bit_and:c@4 @1 @2))
499 (if (single_use (@3) && single_use (@4))
500 (bit_xor (bit_and (bit_xor @0 @1) @2) @0)))
503 /* Associate (p +p off1) +p off2 as (p +p (off1 + off2)). */
505 (pointer_plus (pointer_plus@2 @0 @1) @3)
506 (if (single_use (@2))
507 (pointer_plus @0 (plus @1 @3))))
513 tem4 = (unsigned long) tem3;
518 (pointer_plus @0 (convert?@2 (minus@3 (convert @1) (convert @0))))
519 /* Conditionally look through a sign-changing conversion. */
520 (if (TYPE_PRECISION (TREE_TYPE (@2)) == TYPE_PRECISION (TREE_TYPE (@3))
521 && ((GIMPLE && useless_type_conversion_p (type, TREE_TYPE (@1)))
522 || (GENERIC && type == TREE_TYPE (@1))))
526 tem = (sizetype) ptr;
530 and produce the simpler and easier to analyze with respect to alignment
531 ... = ptr & ~algn; */
533 (pointer_plus @0 (negate (bit_and (convert @0) INTEGER_CST@1)))
534 (with { tree algn = wide_int_to_tree (TREE_TYPE (@0), wi::bit_not (@1)); }
535 (bit_and @0 { algn; })))
538 /* We can't reassociate at all for saturating types. */
539 (if (!TYPE_SATURATING (type))
541 /* Contract negates. */
542 /* A + (-B) -> A - B */
544 (plus:c (convert1? @0) (convert2? (negate @1)))
545 /* Apply STRIP_NOPS on @0 and the negate. */
546 (if (tree_nop_conversion_p (type, TREE_TYPE (@0))
547 && tree_nop_conversion_p (type, TREE_TYPE (@1))
548 && !TYPE_OVERFLOW_SANITIZED (type))
549 (minus (convert @0) (convert @1))))
550 /* A - (-B) -> A + B */
552 (minus (convert1? @0) (convert2? (negate @1)))
553 (if (tree_nop_conversion_p (type, TREE_TYPE (@0))
554 && tree_nop_conversion_p (type, TREE_TYPE (@1))
555 && !TYPE_OVERFLOW_SANITIZED (type))
556 (plus (convert @0) (convert @1))))
559 (negate (convert? (negate @1)))
560 (if (tree_nop_conversion_p (type, TREE_TYPE (@1))
561 && !TYPE_OVERFLOW_SANITIZED (type))
564 /* We can't reassociate floating-point or fixed-point plus or minus
565 because of saturation to +-Inf. */
566 (if (!FLOAT_TYPE_P (type) && !FIXED_POINT_TYPE_P (type))
568 /* Match patterns that allow contracting a plus-minus pair
569 irrespective of overflow issues. */
570 /* (A +- B) - A -> +- B */
571 /* (A +- B) -+ B -> A */
572 /* A - (A +- B) -> -+ B */
573 /* A +- (B -+ A) -> +- B */
575 (minus (plus:c @0 @1) @0)
578 (minus (minus @0 @1) @0)
581 (plus:c (minus @0 @1) @1)
584 (minus @0 (plus:c @0 @1))
587 (minus @0 (minus @0 @1))
590 /* (A +- CST) +- CST -> A + CST */
591 (for outer_op (plus minus)
592 (for inner_op (plus minus)
594 (outer_op (inner_op @0 CONSTANT_CLASS_P@1) CONSTANT_CLASS_P@2)
595 /* If the constant operation overflows we cannot do the transform
596 as we would introduce undefined overflow, for example
597 with (a - 1) + INT_MIN. */
598 (with { tree cst = fold_binary (outer_op == inner_op
599 ? PLUS_EXPR : MINUS_EXPR, type, @1, @2); }
600 (if (cst && !TREE_OVERFLOW (cst))
601 (inner_op @0 { cst; } ))))))
603 /* (CST - A) +- CST -> CST - A */
604 (for outer_op (plus minus)
606 (outer_op (minus CONSTANT_CLASS_P@1 @0) CONSTANT_CLASS_P@2)
607 (with { tree cst = fold_binary (outer_op, type, @1, @2); }
608 (if (cst && !TREE_OVERFLOW (cst))
609 (minus { cst; } @0)))))
613 (plus:c (bit_not @0) @0)
614 (if (!TYPE_OVERFLOW_TRAPS (type))
615 { build_all_ones_cst (type); }))
619 (plus (convert? (bit_not @0)) integer_each_onep)
620 (if (tree_nop_conversion_p (type, TREE_TYPE (@0)))
621 (negate (convert @0))))
625 (minus (convert? (negate @0)) integer_each_onep)
626 (if (!TYPE_OVERFLOW_TRAPS (type)
627 && tree_nop_conversion_p (type, TREE_TYPE (@0)))
628 (bit_not (convert @0))))
632 (minus integer_all_onesp @0)
635 /* (T)(P + A) - (T)P -> (T) A */
636 (for add (plus pointer_plus)
638 (minus (convert (add @0 @1))
640 (if (element_precision (type) <= element_precision (TREE_TYPE (@1))
641 /* For integer types, if A has a smaller type
642 than T the result depends on the possible
644 E.g. T=size_t, A=(unsigned)429497295, P>0.
645 However, if an overflow in P + A would cause
646 undefined behavior, we can assume that there
648 || (INTEGRAL_TYPE_P (TREE_TYPE (@0))
649 && TYPE_OVERFLOW_UNDEFINED (TREE_TYPE (@0)))
650 /* For pointer types, if the conversion of A to the
651 final type requires a sign- or zero-extension,
652 then we have to punt - it is not defined which
654 || (POINTER_TYPE_P (TREE_TYPE (@0))
655 && TREE_CODE (@1) == INTEGER_CST
656 && tree_int_cst_sign_bit (@1) == 0))
660 /* Simplifications of MIN_EXPR and MAX_EXPR. */
662 (for minmax (min max)
668 (if (INTEGRAL_TYPE_P (type)
669 && TYPE_MIN_VALUE (type)
670 && operand_equal_p (@1, TYPE_MIN_VALUE (type), OEP_ONLY_CONST))
674 (if (INTEGRAL_TYPE_P (type)
675 && TYPE_MAX_VALUE (type)
676 && operand_equal_p (@1, TYPE_MAX_VALUE (type), OEP_ONLY_CONST))
680 /* Simplifications of shift and rotates. */
682 (for rotate (lrotate rrotate)
684 (rotate integer_all_onesp@0 @1)
687 /* Optimize -1 >> x for arithmetic right shifts. */
689 (rshift integer_all_onesp@0 @1)
690 (if (!TYPE_UNSIGNED (type)
691 && tree_expr_nonnegative_p (@1))
694 (for shiftrotate (lrotate rrotate lshift rshift)
696 (shiftrotate @0 integer_zerop)
699 (shiftrotate integer_zerop@0 @1)
701 /* Prefer vector1 << scalar to vector1 << vector2
702 if vector2 is uniform. */
703 (for vec (VECTOR_CST CONSTRUCTOR)
705 (shiftrotate @0 vec@1)
706 (with { tree tem = uniform_vector_p (@1); }
708 (shiftrotate @0 { tem; }))))))
710 /* Rewrite an LROTATE_EXPR by a constant into an
711 RROTATE_EXPR by a new constant. */
713 (lrotate @0 INTEGER_CST@1)
714 (rrotate @0 { fold_binary (MINUS_EXPR, TREE_TYPE (@1),
715 build_int_cst (TREE_TYPE (@1),
716 element_precision (type)), @1); }))
718 /* ((1 << A) & 1) != 0 -> A == 0
719 ((1 << A) & 1) == 0 -> A != 0 */
723 (cmp (bit_and (lshift integer_onep @0) integer_onep) integer_zerop)
724 (icmp @0 { build_zero_cst (TREE_TYPE (@0)); })))
726 /* (CST1 << A) == CST2 -> A == ctz (CST2) - ctz (CST1)
727 (CST1 << A) != CST2 -> A != ctz (CST2) - ctz (CST1)
731 (cmp (lshift INTEGER_CST@0 @1) INTEGER_CST@2)
732 (with { int cand = wi::ctz (@2) - wi::ctz (@0); }
734 || (!integer_zerop (@2)
735 && wi::ne_p (wi::lshift (@0, cand), @2)))
736 { constant_boolean_node (cmp == NE_EXPR, type); })
737 (if (!integer_zerop (@2)
738 && wi::eq_p (wi::lshift (@0, cand), @2))
739 (cmp @1 { build_int_cst (TREE_TYPE (@1), cand); })))))
741 /* Simplifications of conversions. */
743 /* Basic strip-useless-type-conversions / strip_nops. */
744 (for cvt (convert view_convert float fix_trunc)
747 (if ((GIMPLE && useless_type_conversion_p (type, TREE_TYPE (@0)))
748 || (GENERIC && type == TREE_TYPE (@0)))
751 /* Contract view-conversions. */
753 (view_convert (view_convert @0))
756 /* For integral conversions with the same precision or pointer
757 conversions use a NOP_EXPR instead. */
760 (if ((INTEGRAL_TYPE_P (type) || POINTER_TYPE_P (type))
761 && (INTEGRAL_TYPE_P (TREE_TYPE (@0)) || POINTER_TYPE_P (TREE_TYPE (@0)))
762 && TYPE_PRECISION (type) == TYPE_PRECISION (TREE_TYPE (@0)))
765 /* Strip inner integral conversions that do not change precision or size. */
767 (view_convert (convert@0 @1))
768 (if ((INTEGRAL_TYPE_P (TREE_TYPE (@0)) || POINTER_TYPE_P (TREE_TYPE (@0)))
769 && (INTEGRAL_TYPE_P (TREE_TYPE (@1)) || POINTER_TYPE_P (TREE_TYPE (@1)))
770 && (TYPE_PRECISION (TREE_TYPE (@0)) == TYPE_PRECISION (TREE_TYPE (@1)))
771 && (TYPE_SIZE (TREE_TYPE (@0)) == TYPE_SIZE (TREE_TYPE (@1))))
774 /* Re-association barriers around constants and other re-association
775 barriers can be removed. */
777 (paren CONSTANT_CLASS_P@0)
783 /* Handle cases of two conversions in a row. */
784 (for ocvt (convert float fix_trunc)
785 (for icvt (convert float)
790 tree inside_type = TREE_TYPE (@0);
791 tree inter_type = TREE_TYPE (@1);
792 int inside_int = INTEGRAL_TYPE_P (inside_type);
793 int inside_ptr = POINTER_TYPE_P (inside_type);
794 int inside_float = FLOAT_TYPE_P (inside_type);
795 int inside_vec = VECTOR_TYPE_P (inside_type);
796 unsigned int inside_prec = TYPE_PRECISION (inside_type);
797 int inside_unsignedp = TYPE_UNSIGNED (inside_type);
798 int inter_int = INTEGRAL_TYPE_P (inter_type);
799 int inter_ptr = POINTER_TYPE_P (inter_type);
800 int inter_float = FLOAT_TYPE_P (inter_type);
801 int inter_vec = VECTOR_TYPE_P (inter_type);
802 unsigned int inter_prec = TYPE_PRECISION (inter_type);
803 int inter_unsignedp = TYPE_UNSIGNED (inter_type);
804 int final_int = INTEGRAL_TYPE_P (type);
805 int final_ptr = POINTER_TYPE_P (type);
806 int final_float = FLOAT_TYPE_P (type);
807 int final_vec = VECTOR_TYPE_P (type);
808 unsigned int final_prec = TYPE_PRECISION (type);
809 int final_unsignedp = TYPE_UNSIGNED (type);
811 /* In addition to the cases of two conversions in a row
812 handled below, if we are converting something to its own
813 type via an object of identical or wider precision, neither
814 conversion is needed. */
815 (if (((GIMPLE && useless_type_conversion_p (type, inside_type))
817 && TYPE_MAIN_VARIANT (type) == TYPE_MAIN_VARIANT (inside_type)))
818 && (((inter_int || inter_ptr) && final_int)
819 || (inter_float && final_float))
820 && inter_prec >= final_prec)
823 /* Likewise, if the intermediate and initial types are either both
824 float or both integer, we don't need the middle conversion if the
825 former is wider than the latter and doesn't change the signedness
826 (for integers). Avoid this if the final type is a pointer since
827 then we sometimes need the middle conversion. Likewise if the
828 final type has a precision not equal to the size of its mode. */
829 (if (((inter_int && inside_int) || (inter_float && inside_float))
830 && (final_int || final_float)
831 && inter_prec >= inside_prec
832 && (inter_float || inter_unsignedp == inside_unsignedp)
833 && ! (final_prec != GET_MODE_PRECISION (TYPE_MODE (type))
834 && TYPE_MODE (type) == TYPE_MODE (inter_type)))
837 /* If we have a sign-extension of a zero-extended value, we can
838 replace that by a single zero-extension. Likewise if the
839 final conversion does not change precision we can drop the
840 intermediate conversion. */
841 (if (inside_int && inter_int && final_int
842 && ((inside_prec < inter_prec && inter_prec < final_prec
843 && inside_unsignedp && !inter_unsignedp)
844 || final_prec == inter_prec))
847 /* Two conversions in a row are not needed unless:
848 - some conversion is floating-point (overstrict for now), or
849 - some conversion is a vector (overstrict for now), or
850 - the intermediate type is narrower than both initial and
852 - the intermediate type and innermost type differ in signedness,
853 and the outermost type is wider than the intermediate, or
854 - the initial type is a pointer type and the precisions of the
855 intermediate and final types differ, or
856 - the final type is a pointer type and the precisions of the
857 initial and intermediate types differ. */
858 (if (! inside_float && ! inter_float && ! final_float
859 && ! inside_vec && ! inter_vec && ! final_vec
860 && (inter_prec >= inside_prec || inter_prec >= final_prec)
861 && ! (inside_int && inter_int
862 && inter_unsignedp != inside_unsignedp
863 && inter_prec < final_prec)
864 && ((inter_unsignedp && inter_prec > inside_prec)
865 == (final_unsignedp && final_prec > inter_prec))
866 && ! (inside_ptr && inter_prec != final_prec)
867 && ! (final_ptr && inside_prec != inter_prec)
868 && ! (final_prec != GET_MODE_PRECISION (TYPE_MODE (type))
869 && TYPE_MODE (type) == TYPE_MODE (inter_type)))
872 /* A truncation to an unsigned type (a zero-extension) should be
873 canonicalized as bitwise and of a mask. */
874 (if (final_int && inter_int && inside_int
875 && final_prec == inside_prec
876 && final_prec > inter_prec
878 (convert (bit_and @0 { wide_int_to_tree
880 wi::mask (inter_prec, false,
881 TYPE_PRECISION (inside_type))); })))
883 /* If we are converting an integer to a floating-point that can
884 represent it exactly and back to an integer, we can skip the
885 floating-point conversion. */
886 (if (GIMPLE /* PR66211 */
887 && inside_int && inter_float && final_int &&
888 (unsigned) significand_size (TYPE_MODE (inter_type))
889 >= inside_prec - !inside_unsignedp)
892 /* If we have a narrowing conversion to an integral type that is fed by a
893 BIT_AND_EXPR, we might be able to remove the BIT_AND_EXPR if it merely
894 masks off bits outside the final type (and nothing else). */
896 (convert (bit_and @0 INTEGER_CST@1))
897 (if (INTEGRAL_TYPE_P (type)
898 && INTEGRAL_TYPE_P (TREE_TYPE (@0))
899 && TYPE_PRECISION (type) <= TYPE_PRECISION (TREE_TYPE (@0))
900 && operand_equal_p (@1, build_low_bits_mask (TREE_TYPE (@1),
901 TYPE_PRECISION (type)), 0))
905 /* (X /[ex] A) * A -> X. */
907 (mult (convert? (exact_div @0 @1)) @1)
908 /* Look through a sign-changing conversion. */
911 /* Canonicalization of binary operations. */
913 /* Convert X + -C into X - C. */
916 (if (REAL_VALUE_NEGATIVE (TREE_REAL_CST (@1)))
917 (with { tree tem = fold_unary (NEGATE_EXPR, type, @1); }
918 (if (!TREE_OVERFLOW (tem) || !flag_trapping_math)
919 (minus @0 { tem; })))))
921 /* Convert x+x into x*2.0. */
924 (if (SCALAR_FLOAT_TYPE_P (type))
925 (mult @0 { build_real (type, dconst2); })))
928 (minus integer_zerop @1)
931 /* (ARG0 - ARG1) is the same as (-ARG1 + ARG0). So check whether
932 ARG0 is zero and X + ARG0 reduces to X, since that would mean
933 (-ARG1 + ARG0) reduces to -ARG1. */
935 (minus real_zerop@0 @1)
936 (if (fold_real_zero_addition_p (type, @0, 0))
939 /* Transform x * -1 into -x. */
941 (mult @0 integer_minus_onep)
944 /* COMPLEX_EXPR and REALPART/IMAGPART_EXPR cancellations. */
946 (complex (realpart @0) (imagpart @0))
949 (realpart (complex @0 @1))
952 (imagpart (complex @0 @1))
956 /* BSWAP simplifications, transforms checked by gcc.dg/builtin-bswap-8.c. */
957 (for bswap (BUILT_IN_BSWAP16 BUILT_IN_BSWAP32 BUILT_IN_BSWAP64)
962 (bswap (bit_not (bswap @0)))
964 (for bitop (bit_xor bit_ior bit_and)
966 (bswap (bitop:c (bswap @0) @1))
967 (bitop @0 (bswap @1)))))
970 /* Combine COND_EXPRs and VEC_COND_EXPRs. */
972 /* Simplify constant conditions.
973 Only optimize constant conditions when the selected branch
974 has the same type as the COND_EXPR. This avoids optimizing
975 away "c ? x : throw", where the throw has a void type.
976 Note that we cannot throw away the fold-const.c variant nor
977 this one as we depend on doing this transform before possibly
978 A ? B : B -> B triggers and the fold-const.c one can optimize
979 0 ? A : B to B even if A has side-effects. Something
980 genmatch cannot handle. */
982 (cond INTEGER_CST@0 @1 @2)
983 (if (integer_zerop (@0)
984 && (!VOID_TYPE_P (TREE_TYPE (@2))
985 || VOID_TYPE_P (type)))
987 (if (!integer_zerop (@0)
988 && (!VOID_TYPE_P (TREE_TYPE (@1))
989 || VOID_TYPE_P (type)))
992 (vec_cond VECTOR_CST@0 @1 @2)
993 (if (integer_all_onesp (@0))
995 (if (integer_zerop (@0))
998 (for cnd (cond vec_cond)
999 /* A ? B : (A ? X : C) -> A ? B : C. */
1001 (cnd @0 (cnd @0 @1 @2) @3)
1004 (cnd @0 @1 (cnd @0 @2 @3))
1007 /* A ? B : B -> B. */
1012 /* !A ? B : C -> A ? C : B. */
1014 (cnd (logical_inverted_value truth_valued_p@0) @1 @2)
1018 /* Simplifications of comparisons. */
1020 /* We can simplify a logical negation of a comparison to the
1021 inverted comparison. As we cannot compute an expression
1022 operator using invert_tree_comparison we have to simulate
1023 that with expression code iteration. */
1024 (for cmp (tcc_comparison)
1025 icmp (inverted_tcc_comparison)
1026 ncmp (inverted_tcc_comparison_with_nans)
1027 /* Ideally we'd like to combine the following two patterns
1028 and handle some more cases by using
1029 (logical_inverted_value (cmp @0 @1))
1030 here but for that genmatch would need to "inline" that.
1031 For now implement what forward_propagate_comparison did. */
1033 (bit_not (cmp @0 @1))
1034 (if (VECTOR_TYPE_P (type)
1035 || (INTEGRAL_TYPE_P (type) && TYPE_PRECISION (type) == 1))
1036 /* Comparison inversion may be impossible for trapping math,
1037 invert_tree_comparison will tell us. But we can't use
1038 a computed operator in the replacement tree thus we have
1039 to play the trick below. */
1040 (with { enum tree_code ic = invert_tree_comparison
1041 (cmp, HONOR_NANS (@0)); }
1047 (bit_xor (cmp @0 @1) integer_truep)
1048 (with { enum tree_code ic = invert_tree_comparison
1049 (cmp, HONOR_NANS (@0)); }
1055 /* Unordered tests if either argument is a NaN. */
1057 (bit_ior (unordered @0 @0) (unordered @1 @1))
1058 (if (types_match (@0, @1))
1061 (bit_and (ordered @0 @0) (ordered @1 @1))
1062 (if (types_match (@0, @1))
1065 (bit_ior:c (unordered @0 @0) (unordered:c@2 @0 @1))
1068 (bit_and:c (ordered @0 @0) (ordered:c@2 @0 @1))
1071 /* -A CMP -B -> B CMP A. */
1072 (for cmp (tcc_comparison)
1073 scmp (swapped_tcc_comparison)
1075 (cmp (negate @0) (negate @1))
1076 (if (FLOAT_TYPE_P (TREE_TYPE (@0))
1077 || (ANY_INTEGRAL_TYPE_P (TREE_TYPE (@0))
1078 && TYPE_OVERFLOW_UNDEFINED (TREE_TYPE (@0))))
1081 (cmp (negate @0) CONSTANT_CLASS_P@1)
1082 (if (FLOAT_TYPE_P (TREE_TYPE (@0))
1083 || (ANY_INTEGRAL_TYPE_P (TREE_TYPE (@0))
1084 && TYPE_OVERFLOW_UNDEFINED (TREE_TYPE (@0))))
1085 (with { tree tem = fold_unary (NEGATE_EXPR, TREE_TYPE (@0), @1); }
1086 (if (tem && !TREE_OVERFLOW (tem))
1087 (scmp @0 { tem; }))))))
1089 /* Simplification of math builtins. */
1091 (define_operator_list LOG BUILT_IN_LOGF BUILT_IN_LOG BUILT_IN_LOGL)
1092 (define_operator_list EXP BUILT_IN_EXPF BUILT_IN_EXP BUILT_IN_EXPL)
1093 (define_operator_list LOG2 BUILT_IN_LOG2F BUILT_IN_LOG2 BUILT_IN_LOG2L)
1094 (define_operator_list EXP2 BUILT_IN_EXP2F BUILT_IN_EXP2 BUILT_IN_EXP2L)
1095 (define_operator_list LOG10 BUILT_IN_LOG10F BUILT_IN_LOG10 BUILT_IN_LOG10L)
1096 (define_operator_list EXP10 BUILT_IN_EXP10F BUILT_IN_EXP10 BUILT_IN_EXP10L)
1097 (define_operator_list POW BUILT_IN_POWF BUILT_IN_POW BUILT_IN_POWL)
1098 (define_operator_list POW10 BUILT_IN_POW10F BUILT_IN_POW10 BUILT_IN_POW10L)
1099 (define_operator_list SQRT BUILT_IN_SQRTF BUILT_IN_SQRT BUILT_IN_SQRTL)
1100 (define_operator_list CBRT BUILT_IN_CBRTF BUILT_IN_CBRT BUILT_IN_CBRTL)
1103 /* fold_builtin_logarithm */
1104 (if (flag_unsafe_math_optimizations)
1105 /* Special case, optimize logN(expN(x)) = x. */
1106 (for logs (LOG LOG2 LOG10)
1107 exps (EXP EXP2 EXP10)
1111 /* Optimize logN(func()) for various exponential functions. We
1112 want to determine the value "x" and the power "exponent" in
1113 order to transform logN(x**exponent) into exponent*logN(x). */
1114 (for logs (LOG LOG LOG LOG
1116 LOG10 LOG10 LOG10 LOG10)
1117 exps (EXP EXP2 EXP10 POW10)
1124 CASE_FLT_FN (BUILT_IN_EXP):
1125 /* Prepare to do logN(exp(exponent) -> exponent*logN(e). */
1126 x = build_real (type, real_value_truncate (TYPE_MODE (type),
1129 CASE_FLT_FN (BUILT_IN_EXP2):
1130 /* Prepare to do logN(exp2(exponent) -> exponent*logN(2). */
1131 x = build_real (type, dconst2);
1133 CASE_FLT_FN (BUILT_IN_EXP10):
1134 CASE_FLT_FN (BUILT_IN_POW10):
1135 /* Prepare to do logN(exp10(exponent) -> exponent*logN(10). */
1137 REAL_VALUE_TYPE dconst10;
1138 real_from_integer (&dconst10, VOIDmode, 10, SIGNED);
1139 x = build_real (type, dconst10);
1144 (mult (logs { x; }) @0))))
1155 CASE_FLT_FN (BUILT_IN_SQRT):
1156 /* Prepare to do logN(sqrt(x) -> 0.5*logN(x). */
1157 x = build_real (type, dconsthalf);
1159 CASE_FLT_FN (BUILT_IN_CBRT):
1160 /* Prepare to do logN(cbrt(x) -> (1/3)*logN(x). */
1161 x = build_real (type, real_value_truncate (TYPE_MODE (type),
1166 (mult { x; } (logs @0)))))
1167 /* logN(pow(x,exponent) -> exponent*logN(x). */
1168 (for logs (LOG LOG2 LOG10)
1172 (mult @1 (logs @0)))))
1174 /* Narrowing of arithmetic and logical operations.
1176 These are conceptually similar to the transformations performed for
1177 the C/C++ front-ends by shorten_binary_op and shorten_compare. Long
1178 term we want to move all that code out of the front-ends into here. */
1180 /* If we have a narrowing conversion of an arithmetic operation where
1181 both operands are widening conversions from the same type as the outer
1182 narrowing conversion. Then convert the innermost operands to a suitable
1183 unsigned type (to avoid introducing undefined behaviour), perform the
1184 operation and convert the result to the desired type. */
1185 (for op (plus minus)
1187 (convert (op@4 (convert@2 @0) (convert@3 @1)))
1188 (if (INTEGRAL_TYPE_P (type)
1189 /* We check for type compatibility between @0 and @1 below,
1190 so there's no need to check that @1/@3 are integral types. */
1191 && INTEGRAL_TYPE_P (TREE_TYPE (@0))
1192 && INTEGRAL_TYPE_P (TREE_TYPE (@2))
1193 /* The precision of the type of each operand must match the
1194 precision of the mode of each operand, similarly for the
1196 && (TYPE_PRECISION (TREE_TYPE (@0))
1197 == GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (@0))))
1198 && (TYPE_PRECISION (TREE_TYPE (@1))
1199 == GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (@1))))
1200 && TYPE_PRECISION (type) == GET_MODE_PRECISION (TYPE_MODE (type))
1201 /* The inner conversion must be a widening conversion. */
1202 && TYPE_PRECISION (TREE_TYPE (@2)) > TYPE_PRECISION (TREE_TYPE (@0))
1203 && types_match (@0, @1)
1204 && types_match (@0, type)
1206 (if (TYPE_OVERFLOW_WRAPS (TREE_TYPE (@0)))
1207 (convert (op @0 @1)))
1208 (with { tree utype = unsigned_type_for (TREE_TYPE (@0)); }
1209 (convert (op (convert:utype @0) (convert:utype @1)))))))
1211 /* This is another case of narrowing, specifically when there's an outer
1212 BIT_AND_EXPR which masks off bits outside the type of the innermost
1213 operands. Like the previous case we have to convert the operands
1214 to unsigned types to avoid introducing undefined behaviour for the
1215 arithmetic operation. */
1216 (for op (minus plus)
1218 (bit_and (op@5 (convert@2 @0) (convert@3 @1)) INTEGER_CST@4)
1219 (if (INTEGRAL_TYPE_P (type)
1220 /* We check for type compatibility between @0 and @1 below,
1221 so there's no need to check that @1/@3 are integral types. */
1222 && INTEGRAL_TYPE_P (TREE_TYPE (@0))
1223 && INTEGRAL_TYPE_P (TREE_TYPE (@2))
1224 /* The precision of the type of each operand must match the
1225 precision of the mode of each operand, similarly for the
1227 && (TYPE_PRECISION (TREE_TYPE (@0))
1228 == GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (@0))))
1229 && (TYPE_PRECISION (TREE_TYPE (@1))
1230 == GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (@1))))
1231 && TYPE_PRECISION (type) == GET_MODE_PRECISION (TYPE_MODE (type))
1232 /* The inner conversion must be a widening conversion. */
1233 && TYPE_PRECISION (TREE_TYPE (@2)) > TYPE_PRECISION (TREE_TYPE (@0))
1234 && types_match (@0, @1)
1235 && (tree_int_cst_min_precision (@4, TYPE_SIGN (TREE_TYPE (@0)))
1236 <= TYPE_PRECISION (TREE_TYPE (@0)))
1237 && (TYPE_OVERFLOW_WRAPS (TREE_TYPE (@0))
1238 || tree_int_cst_sgn (@4) >= 0)
1240 (if (TYPE_OVERFLOW_WRAPS (TREE_TYPE (@0)))
1241 (with { tree ntype = TREE_TYPE (@0); }
1242 (convert (bit_and (op @0 @1) (convert:ntype @4)))))
1243 (with { tree utype = unsigned_type_for (TREE_TYPE (@0)); }
1244 (convert (bit_and (op (convert:utype @0) (convert:utype @1))
1245 (convert:utype @4)))))))