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 (bit_ior @0 @1) (bit_xor @0 @1))
351 /* (x | y) - (x & y) -> x ^ y */
353 (minus (bit_ior @0 @1) (bit_and @0 @1))
360 (abs tree_expr_nonnegative_p@0)
364 /* Try to fold (type) X op CST -> (type) (X op ((type-x) CST))
366 For bitwise binary operations apply operand conversions to the
367 binary operation result instead of to the operands. This allows
368 to combine successive conversions and bitwise binary operations.
369 We combine the above two cases by using a conditional convert. */
370 (for bitop (bit_and bit_ior bit_xor)
372 (bitop (convert @0) (convert? @1))
373 (if (((TREE_CODE (@1) == INTEGER_CST
374 && INTEGRAL_TYPE_P (TREE_TYPE (@0))
375 && int_fits_type_p (@1, TREE_TYPE (@0)))
376 || types_match (@0, @1))
377 /* ??? This transform conflicts with fold-const.c doing
378 Convert (T)(x & c) into (T)x & (T)c, if c is an integer
379 constants (if x has signed type, the sign bit cannot be set
380 in c). This folds extension into the BIT_AND_EXPR.
381 Restrict it to GIMPLE to avoid endless recursions. */
382 && (bitop != BIT_AND_EXPR || GIMPLE)
383 && (/* That's a good idea if the conversion widens the operand, thus
384 after hoisting the conversion the operation will be narrower. */
385 TYPE_PRECISION (TREE_TYPE (@0)) < TYPE_PRECISION (type)
386 /* It's also a good idea if the conversion is to a non-integer
388 || GET_MODE_CLASS (TYPE_MODE (type)) != MODE_INT
389 /* Or if the precision of TO is not the same as the precision
391 || TYPE_PRECISION (type) != GET_MODE_PRECISION (TYPE_MODE (type))))
392 (convert (bitop @0 (convert @1))))))
394 /* Simplify (A & B) OP0 (C & B) to (A OP0 C) & B. */
395 (for bitop (bit_and bit_ior bit_xor)
397 (bitop (bit_and:c @0 @1) (bit_and @2 @1))
398 (bit_and (bitop @0 @2) @1)))
400 /* (x | CST1) & CST2 -> (x & CST2) | (CST1 & CST2) */
402 (bit_and (bit_ior @0 CONSTANT_CLASS_P@1) CONSTANT_CLASS_P@2)
403 (bit_ior (bit_and @0 @2) (bit_and @1 @2)))
405 /* Combine successive equal operations with constants. */
406 (for bitop (bit_and bit_ior bit_xor)
408 (bitop (bitop @0 CONSTANT_CLASS_P@1) CONSTANT_CLASS_P@2)
409 (bitop @0 (bitop @1 @2))))
411 /* Try simple folding for X op !X, and X op X with the help
412 of the truth_valued_p and logical_inverted_value predicates. */
413 (match truth_valued_p
415 (if (INTEGRAL_TYPE_P (type) && TYPE_PRECISION (type) == 1)))
416 (for op (tcc_comparison truth_and truth_andif truth_or truth_orif truth_xor)
417 (match truth_valued_p
419 (match truth_valued_p
422 (match (logical_inverted_value @0)
423 (bit_not truth_valued_p@0))
424 (match (logical_inverted_value @0)
425 (eq @0 integer_zerop))
426 (match (logical_inverted_value @0)
427 (ne truth_valued_p@0 integer_truep))
428 (match (logical_inverted_value @0)
429 (bit_xor truth_valued_p@0 integer_truep))
433 (bit_and:c @0 (logical_inverted_value @0))
434 { build_zero_cst (type); })
435 /* X | !X and X ^ !X -> 1, , if X is truth-valued. */
436 (for op (bit_ior bit_xor)
438 (op:c truth_valued_p@0 (logical_inverted_value @0))
439 { constant_boolean_node (true, type); }))
441 (for bitop (bit_and bit_ior)
442 rbitop (bit_ior bit_and)
443 /* (x | y) & x -> x */
444 /* (x & y) | x -> x */
446 (bitop:c (rbitop:c @0 @1) @0)
448 /* (~x | y) & x -> x & y */
449 /* (~x & y) | x -> x | y */
451 (bitop:c (rbitop:c (bit_not @0) @1) @0)
454 /* If arg1 and arg2 are booleans (or any single bit type)
455 then try to simplify:
462 But only do this if our result feeds into a comparison as
463 this transformation is not always a win, particularly on
464 targets with and-not instructions.
465 -> simplify_bitwise_binary_boolean */
467 (ne (bit_and:c (bit_not @0) @1) integer_zerop)
468 (if (INTEGRAL_TYPE_P (TREE_TYPE (@1))
469 && TYPE_PRECISION (TREE_TYPE (@1)) == 1)
472 (ne (bit_ior:c (bit_not @0) @1) integer_zerop)
473 (if (INTEGRAL_TYPE_P (TREE_TYPE (@1))
474 && TYPE_PRECISION (TREE_TYPE (@1)) == 1)
479 (bit_not (bit_not @0))
482 /* (x & ~m) | (y & m) -> ((x ^ y) & m) ^ x */
484 (bit_ior:c (bit_and:c@3 @0 (bit_not @2)) (bit_and:c@4 @1 @2))
485 (if (single_use (@3) && single_use (@4))
486 (bit_xor (bit_and (bit_xor @0 @1) @2) @0)))
489 /* Associate (p +p off1) +p off2 as (p +p (off1 + off2)). */
491 (pointer_plus (pointer_plus@2 @0 @1) @3)
492 (if (single_use (@2))
493 (pointer_plus @0 (plus @1 @3))))
499 tem4 = (unsigned long) tem3;
504 (pointer_plus @0 (convert?@2 (minus@3 (convert @1) (convert @0))))
505 /* Conditionally look through a sign-changing conversion. */
506 (if (TYPE_PRECISION (TREE_TYPE (@2)) == TYPE_PRECISION (TREE_TYPE (@3))
507 && ((GIMPLE && useless_type_conversion_p (type, TREE_TYPE (@1)))
508 || (GENERIC && type == TREE_TYPE (@1))))
512 tem = (sizetype) ptr;
516 and produce the simpler and easier to analyze with respect to alignment
517 ... = ptr & ~algn; */
519 (pointer_plus @0 (negate (bit_and (convert @0) INTEGER_CST@1)))
520 (with { tree algn = wide_int_to_tree (TREE_TYPE (@0), wi::bit_not (@1)); }
521 (bit_and @0 { algn; })))
524 /* We can't reassociate at all for saturating types. */
525 (if (!TYPE_SATURATING (type))
527 /* Contract negates. */
528 /* A + (-B) -> A - B */
530 (plus:c (convert1? @0) (convert2? (negate @1)))
531 /* Apply STRIP_NOPS on @0 and the negate. */
532 (if (tree_nop_conversion_p (type, TREE_TYPE (@0))
533 && tree_nop_conversion_p (type, TREE_TYPE (@1))
534 && !TYPE_OVERFLOW_SANITIZED (type))
535 (minus (convert @0) (convert @1))))
536 /* A - (-B) -> A + B */
538 (minus (convert1? @0) (convert2? (negate @1)))
539 (if (tree_nop_conversion_p (type, TREE_TYPE (@0))
540 && tree_nop_conversion_p (type, TREE_TYPE (@1))
541 && !TYPE_OVERFLOW_SANITIZED (type))
542 (plus (convert @0) (convert @1))))
545 (negate (convert? (negate @1)))
546 (if (tree_nop_conversion_p (type, TREE_TYPE (@1))
547 && !TYPE_OVERFLOW_SANITIZED (type))
550 /* We can't reassociate floating-point or fixed-point plus or minus
551 because of saturation to +-Inf. */
552 (if (!FLOAT_TYPE_P (type) && !FIXED_POINT_TYPE_P (type))
554 /* Match patterns that allow contracting a plus-minus pair
555 irrespective of overflow issues. */
556 /* (A +- B) - A -> +- B */
557 /* (A +- B) -+ B -> A */
558 /* A - (A +- B) -> -+ B */
559 /* A +- (B -+ A) -> +- B */
561 (minus (plus:c @0 @1) @0)
564 (minus (minus @0 @1) @0)
567 (plus:c (minus @0 @1) @1)
570 (minus @0 (plus:c @0 @1))
573 (minus @0 (minus @0 @1))
576 /* (A +- CST) +- CST -> A + CST */
577 (for outer_op (plus minus)
578 (for inner_op (plus minus)
580 (outer_op (inner_op @0 CONSTANT_CLASS_P@1) CONSTANT_CLASS_P@2)
581 /* If the constant operation overflows we cannot do the transform
582 as we would introduce undefined overflow, for example
583 with (a - 1) + INT_MIN. */
584 (with { tree cst = fold_binary (outer_op == inner_op
585 ? PLUS_EXPR : MINUS_EXPR, type, @1, @2); }
586 (if (cst && !TREE_OVERFLOW (cst))
587 (inner_op @0 { cst; } ))))))
589 /* (CST - A) +- CST -> CST - A */
590 (for outer_op (plus minus)
592 (outer_op (minus CONSTANT_CLASS_P@1 @0) CONSTANT_CLASS_P@2)
593 (with { tree cst = fold_binary (outer_op, type, @1, @2); }
594 (if (cst && !TREE_OVERFLOW (cst))
595 (minus { cst; } @0)))))
599 (plus:c (bit_not @0) @0)
600 (if (!TYPE_OVERFLOW_TRAPS (type))
601 { build_all_ones_cst (type); }))
605 (plus (convert? (bit_not @0)) integer_each_onep)
606 (if (tree_nop_conversion_p (type, TREE_TYPE (@0)))
607 (negate (convert @0))))
611 (minus (convert? (negate @0)) integer_each_onep)
612 (if (!TYPE_OVERFLOW_TRAPS (type)
613 && tree_nop_conversion_p (type, TREE_TYPE (@0)))
614 (bit_not (convert @0))))
618 (minus integer_all_onesp @0)
621 /* (T)(P + A) - (T)P -> (T) A */
622 (for add (plus pointer_plus)
624 (minus (convert (add @0 @1))
626 (if (element_precision (type) <= element_precision (TREE_TYPE (@1))
627 /* For integer types, if A has a smaller type
628 than T the result depends on the possible
630 E.g. T=size_t, A=(unsigned)429497295, P>0.
631 However, if an overflow in P + A would cause
632 undefined behavior, we can assume that there
634 || (INTEGRAL_TYPE_P (TREE_TYPE (@0))
635 && TYPE_OVERFLOW_UNDEFINED (TREE_TYPE (@0)))
636 /* For pointer types, if the conversion of A to the
637 final type requires a sign- or zero-extension,
638 then we have to punt - it is not defined which
640 || (POINTER_TYPE_P (TREE_TYPE (@0))
641 && TREE_CODE (@1) == INTEGER_CST
642 && tree_int_cst_sign_bit (@1) == 0))
646 /* Simplifications of MIN_EXPR and MAX_EXPR. */
648 (for minmax (min max)
654 (if (INTEGRAL_TYPE_P (type)
655 && TYPE_MIN_VALUE (type)
656 && operand_equal_p (@1, TYPE_MIN_VALUE (type), OEP_ONLY_CONST))
660 (if (INTEGRAL_TYPE_P (type)
661 && TYPE_MAX_VALUE (type)
662 && operand_equal_p (@1, TYPE_MAX_VALUE (type), OEP_ONLY_CONST))
666 /* Simplifications of shift and rotates. */
668 (for rotate (lrotate rrotate)
670 (rotate integer_all_onesp@0 @1)
673 /* Optimize -1 >> x for arithmetic right shifts. */
675 (rshift integer_all_onesp@0 @1)
676 (if (!TYPE_UNSIGNED (type)
677 && tree_expr_nonnegative_p (@1))
680 (for shiftrotate (lrotate rrotate lshift rshift)
682 (shiftrotate @0 integer_zerop)
685 (shiftrotate integer_zerop@0 @1)
687 /* Prefer vector1 << scalar to vector1 << vector2
688 if vector2 is uniform. */
689 (for vec (VECTOR_CST CONSTRUCTOR)
691 (shiftrotate @0 vec@1)
692 (with { tree tem = uniform_vector_p (@1); }
694 (shiftrotate @0 { tem; }))))))
696 /* Rewrite an LROTATE_EXPR by a constant into an
697 RROTATE_EXPR by a new constant. */
699 (lrotate @0 INTEGER_CST@1)
700 (rrotate @0 { fold_binary (MINUS_EXPR, TREE_TYPE (@1),
701 build_int_cst (TREE_TYPE (@1),
702 element_precision (type)), @1); }))
704 /* ((1 << A) & 1) != 0 -> A == 0
705 ((1 << A) & 1) == 0 -> A != 0 */
709 (cmp (bit_and (lshift integer_onep @0) integer_onep) integer_zerop)
710 (icmp @0 { build_zero_cst (TREE_TYPE (@0)); })))
712 /* (CST1 << A) == CST2 -> A == ctz (CST2) - ctz (CST1)
713 (CST1 << A) != CST2 -> A != ctz (CST2) - ctz (CST1)
717 (cmp (lshift INTEGER_CST@0 @1) INTEGER_CST@2)
718 (with { int cand = wi::ctz (@2) - wi::ctz (@0); }
720 || (!integer_zerop (@2)
721 && wi::ne_p (wi::lshift (@0, cand), @2)))
722 { constant_boolean_node (cmp == NE_EXPR, type); })
723 (if (!integer_zerop (@2)
724 && wi::eq_p (wi::lshift (@0, cand), @2))
725 (cmp @1 { build_int_cst (TREE_TYPE (@1), cand); })))))
727 /* Simplifications of conversions. */
729 /* Basic strip-useless-type-conversions / strip_nops. */
730 (for cvt (convert view_convert float fix_trunc)
733 (if ((GIMPLE && useless_type_conversion_p (type, TREE_TYPE (@0)))
734 || (GENERIC && type == TREE_TYPE (@0)))
737 /* Contract view-conversions. */
739 (view_convert (view_convert @0))
742 /* For integral conversions with the same precision or pointer
743 conversions use a NOP_EXPR instead. */
746 (if ((INTEGRAL_TYPE_P (type) || POINTER_TYPE_P (type))
747 && (INTEGRAL_TYPE_P (TREE_TYPE (@0)) || POINTER_TYPE_P (TREE_TYPE (@0)))
748 && TYPE_PRECISION (type) == TYPE_PRECISION (TREE_TYPE (@0)))
751 /* Strip inner integral conversions that do not change precision or size. */
753 (view_convert (convert@0 @1))
754 (if ((INTEGRAL_TYPE_P (TREE_TYPE (@0)) || POINTER_TYPE_P (TREE_TYPE (@0)))
755 && (INTEGRAL_TYPE_P (TREE_TYPE (@1)) || POINTER_TYPE_P (TREE_TYPE (@1)))
756 && (TYPE_PRECISION (TREE_TYPE (@0)) == TYPE_PRECISION (TREE_TYPE (@1)))
757 && (TYPE_SIZE (TREE_TYPE (@0)) == TYPE_SIZE (TREE_TYPE (@1))))
760 /* Re-association barriers around constants and other re-association
761 barriers can be removed. */
763 (paren CONSTANT_CLASS_P@0)
769 /* Handle cases of two conversions in a row. */
770 (for ocvt (convert float fix_trunc)
771 (for icvt (convert float)
776 tree inside_type = TREE_TYPE (@0);
777 tree inter_type = TREE_TYPE (@1);
778 int inside_int = INTEGRAL_TYPE_P (inside_type);
779 int inside_ptr = POINTER_TYPE_P (inside_type);
780 int inside_float = FLOAT_TYPE_P (inside_type);
781 int inside_vec = VECTOR_TYPE_P (inside_type);
782 unsigned int inside_prec = TYPE_PRECISION (inside_type);
783 int inside_unsignedp = TYPE_UNSIGNED (inside_type);
784 int inter_int = INTEGRAL_TYPE_P (inter_type);
785 int inter_ptr = POINTER_TYPE_P (inter_type);
786 int inter_float = FLOAT_TYPE_P (inter_type);
787 int inter_vec = VECTOR_TYPE_P (inter_type);
788 unsigned int inter_prec = TYPE_PRECISION (inter_type);
789 int inter_unsignedp = TYPE_UNSIGNED (inter_type);
790 int final_int = INTEGRAL_TYPE_P (type);
791 int final_ptr = POINTER_TYPE_P (type);
792 int final_float = FLOAT_TYPE_P (type);
793 int final_vec = VECTOR_TYPE_P (type);
794 unsigned int final_prec = TYPE_PRECISION (type);
795 int final_unsignedp = TYPE_UNSIGNED (type);
797 /* In addition to the cases of two conversions in a row
798 handled below, if we are converting something to its own
799 type via an object of identical or wider precision, neither
800 conversion is needed. */
801 (if (((GIMPLE && useless_type_conversion_p (type, inside_type))
803 && TYPE_MAIN_VARIANT (type) == TYPE_MAIN_VARIANT (inside_type)))
804 && (((inter_int || inter_ptr) && final_int)
805 || (inter_float && final_float))
806 && inter_prec >= final_prec)
809 /* Likewise, if the intermediate and initial types are either both
810 float or both integer, we don't need the middle conversion if the
811 former is wider than the latter and doesn't change the signedness
812 (for integers). Avoid this if the final type is a pointer since
813 then we sometimes need the middle conversion. Likewise if the
814 final type has a precision not equal to the size of its mode. */
815 (if (((inter_int && inside_int) || (inter_float && inside_float))
816 && (final_int || final_float)
817 && inter_prec >= inside_prec
818 && (inter_float || inter_unsignedp == inside_unsignedp)
819 && ! (final_prec != GET_MODE_PRECISION (TYPE_MODE (type))
820 && TYPE_MODE (type) == TYPE_MODE (inter_type)))
823 /* If we have a sign-extension of a zero-extended value, we can
824 replace that by a single zero-extension. Likewise if the
825 final conversion does not change precision we can drop the
826 intermediate conversion. */
827 (if (inside_int && inter_int && final_int
828 && ((inside_prec < inter_prec && inter_prec < final_prec
829 && inside_unsignedp && !inter_unsignedp)
830 || final_prec == inter_prec))
833 /* Two conversions in a row are not needed unless:
834 - some conversion is floating-point (overstrict for now), or
835 - some conversion is a vector (overstrict for now), or
836 - the intermediate type is narrower than both initial and
838 - the intermediate type and innermost type differ in signedness,
839 and the outermost type is wider than the intermediate, or
840 - the initial type is a pointer type and the precisions of the
841 intermediate and final types differ, or
842 - the final type is a pointer type and the precisions of the
843 initial and intermediate types differ. */
844 (if (! inside_float && ! inter_float && ! final_float
845 && ! inside_vec && ! inter_vec && ! final_vec
846 && (inter_prec >= inside_prec || inter_prec >= final_prec)
847 && ! (inside_int && inter_int
848 && inter_unsignedp != inside_unsignedp
849 && inter_prec < final_prec)
850 && ((inter_unsignedp && inter_prec > inside_prec)
851 == (final_unsignedp && final_prec > inter_prec))
852 && ! (inside_ptr && inter_prec != final_prec)
853 && ! (final_ptr && inside_prec != inter_prec)
854 && ! (final_prec != GET_MODE_PRECISION (TYPE_MODE (type))
855 && TYPE_MODE (type) == TYPE_MODE (inter_type)))
858 /* A truncation to an unsigned type (a zero-extension) should be
859 canonicalized as bitwise and of a mask. */
860 (if (final_int && inter_int && inside_int
861 && final_prec == inside_prec
862 && final_prec > inter_prec
864 (convert (bit_and @0 { wide_int_to_tree
866 wi::mask (inter_prec, false,
867 TYPE_PRECISION (inside_type))); })))
869 /* If we are converting an integer to a floating-point that can
870 represent it exactly and back to an integer, we can skip the
871 floating-point conversion. */
872 (if (GIMPLE /* PR66211 */
873 && inside_int && inter_float && final_int &&
874 (unsigned) significand_size (TYPE_MODE (inter_type))
875 >= inside_prec - !inside_unsignedp)
878 /* If we have a narrowing conversion to an integral type that is fed by a
879 BIT_AND_EXPR, we might be able to remove the BIT_AND_EXPR if it merely
880 masks off bits outside the final type (and nothing else). */
882 (convert (bit_and @0 INTEGER_CST@1))
883 (if (INTEGRAL_TYPE_P (type)
884 && INTEGRAL_TYPE_P (TREE_TYPE (@0))
885 && TYPE_PRECISION (type) <= TYPE_PRECISION (TREE_TYPE (@0))
886 && operand_equal_p (@1, build_low_bits_mask (TREE_TYPE (@1),
887 TYPE_PRECISION (type)), 0))
891 /* (X /[ex] A) * A -> X. */
893 (mult (convert? (exact_div @0 @1)) @1)
894 /* Look through a sign-changing conversion. */
897 /* Canonicalization of binary operations. */
899 /* Convert X + -C into X - C. */
902 (if (REAL_VALUE_NEGATIVE (TREE_REAL_CST (@1)))
903 (with { tree tem = fold_unary (NEGATE_EXPR, type, @1); }
904 (if (!TREE_OVERFLOW (tem) || !flag_trapping_math)
905 (minus @0 { tem; })))))
907 /* Convert x+x into x*2.0. */
910 (if (SCALAR_FLOAT_TYPE_P (type))
911 (mult @0 { build_real (type, dconst2); })))
914 (minus integer_zerop @1)
917 /* (ARG0 - ARG1) is the same as (-ARG1 + ARG0). So check whether
918 ARG0 is zero and X + ARG0 reduces to X, since that would mean
919 (-ARG1 + ARG0) reduces to -ARG1. */
921 (minus real_zerop@0 @1)
922 (if (fold_real_zero_addition_p (type, @0, 0))
925 /* Transform x * -1 into -x. */
927 (mult @0 integer_minus_onep)
930 /* COMPLEX_EXPR and REALPART/IMAGPART_EXPR cancellations. */
932 (complex (realpart @0) (imagpart @0))
935 (realpart (complex @0 @1))
938 (imagpart (complex @0 @1))
942 /* BSWAP simplifications, transforms checked by gcc.dg/builtin-bswap-8.c. */
943 (for bswap (BUILT_IN_BSWAP16 BUILT_IN_BSWAP32 BUILT_IN_BSWAP64)
948 (bswap (bit_not (bswap @0)))
950 (for bitop (bit_xor bit_ior bit_and)
952 (bswap (bitop:c (bswap @0) @1))
953 (bitop @0 (bswap @1)))))
956 /* Combine COND_EXPRs and VEC_COND_EXPRs. */
958 /* Simplify constant conditions.
959 Only optimize constant conditions when the selected branch
960 has the same type as the COND_EXPR. This avoids optimizing
961 away "c ? x : throw", where the throw has a void type.
962 Note that we cannot throw away the fold-const.c variant nor
963 this one as we depend on doing this transform before possibly
964 A ? B : B -> B triggers and the fold-const.c one can optimize
965 0 ? A : B to B even if A has side-effects. Something
966 genmatch cannot handle. */
968 (cond INTEGER_CST@0 @1 @2)
969 (if (integer_zerop (@0)
970 && (!VOID_TYPE_P (TREE_TYPE (@2))
971 || VOID_TYPE_P (type)))
973 (if (!integer_zerop (@0)
974 && (!VOID_TYPE_P (TREE_TYPE (@1))
975 || VOID_TYPE_P (type)))
978 (vec_cond VECTOR_CST@0 @1 @2)
979 (if (integer_all_onesp (@0))
981 (if (integer_zerop (@0))
984 (for cnd (cond vec_cond)
985 /* A ? B : (A ? X : C) -> A ? B : C. */
987 (cnd @0 (cnd @0 @1 @2) @3)
990 (cnd @0 @1 (cnd @0 @2 @3))
993 /* A ? B : B -> B. */
998 /* !A ? B : C -> A ? C : B. */
1000 (cnd (logical_inverted_value truth_valued_p@0) @1 @2)
1004 /* Simplifications of comparisons. */
1006 /* We can simplify a logical negation of a comparison to the
1007 inverted comparison. As we cannot compute an expression
1008 operator using invert_tree_comparison we have to simulate
1009 that with expression code iteration. */
1010 (for cmp (tcc_comparison)
1011 icmp (inverted_tcc_comparison)
1012 ncmp (inverted_tcc_comparison_with_nans)
1013 /* Ideally we'd like to combine the following two patterns
1014 and handle some more cases by using
1015 (logical_inverted_value (cmp @0 @1))
1016 here but for that genmatch would need to "inline" that.
1017 For now implement what forward_propagate_comparison did. */
1019 (bit_not (cmp @0 @1))
1020 (if (VECTOR_TYPE_P (type)
1021 || (INTEGRAL_TYPE_P (type) && TYPE_PRECISION (type) == 1))
1022 /* Comparison inversion may be impossible for trapping math,
1023 invert_tree_comparison will tell us. But we can't use
1024 a computed operator in the replacement tree thus we have
1025 to play the trick below. */
1026 (with { enum tree_code ic = invert_tree_comparison
1027 (cmp, HONOR_NANS (@0)); }
1033 (bit_xor (cmp @0 @1) integer_truep)
1034 (with { enum tree_code ic = invert_tree_comparison
1035 (cmp, HONOR_NANS (@0)); }
1041 /* Unordered tests if either argument is a NaN. */
1043 (bit_ior (unordered @0 @0) (unordered @1 @1))
1044 (if (types_match (@0, @1))
1047 (bit_and (ordered @0 @0) (ordered @1 @1))
1048 (if (types_match (@0, @1))
1051 (bit_ior:c (unordered @0 @0) (unordered:c@2 @0 @1))
1054 (bit_and:c (ordered @0 @0) (ordered:c@2 @0 @1))
1057 /* -A CMP -B -> B CMP A. */
1058 (for cmp (tcc_comparison)
1059 scmp (swapped_tcc_comparison)
1061 (cmp (negate @0) (negate @1))
1062 (if (FLOAT_TYPE_P (TREE_TYPE (@0))
1063 || (ANY_INTEGRAL_TYPE_P (TREE_TYPE (@0))
1064 && TYPE_OVERFLOW_UNDEFINED (TREE_TYPE (@0))))
1067 (cmp (negate @0) CONSTANT_CLASS_P@1)
1068 (if (FLOAT_TYPE_P (TREE_TYPE (@0))
1069 || (ANY_INTEGRAL_TYPE_P (TREE_TYPE (@0))
1070 && TYPE_OVERFLOW_UNDEFINED (TREE_TYPE (@0))))
1071 (with { tree tem = fold_unary (NEGATE_EXPR, TREE_TYPE (@0), @1); }
1072 (if (tem && !TREE_OVERFLOW (tem))
1073 (scmp @0 { tem; }))))))
1075 /* Simplification of math builtins. */
1077 (define_operator_list LOG BUILT_IN_LOGF BUILT_IN_LOG BUILT_IN_LOGL)
1078 (define_operator_list EXP BUILT_IN_EXPF BUILT_IN_EXP BUILT_IN_EXPL)
1079 (define_operator_list LOG2 BUILT_IN_LOG2F BUILT_IN_LOG2 BUILT_IN_LOG2L)
1080 (define_operator_list EXP2 BUILT_IN_EXP2F BUILT_IN_EXP2 BUILT_IN_EXP2L)
1081 (define_operator_list LOG10 BUILT_IN_LOG10F BUILT_IN_LOG10 BUILT_IN_LOG10L)
1082 (define_operator_list EXP10 BUILT_IN_EXP10F BUILT_IN_EXP10 BUILT_IN_EXP10L)
1083 (define_operator_list POW BUILT_IN_POWF BUILT_IN_POW BUILT_IN_POWL)
1084 (define_operator_list POW10 BUILT_IN_POW10F BUILT_IN_POW10 BUILT_IN_POW10L)
1085 (define_operator_list SQRT BUILT_IN_SQRTF BUILT_IN_SQRT BUILT_IN_SQRTL)
1086 (define_operator_list CBRT BUILT_IN_CBRTF BUILT_IN_CBRT BUILT_IN_CBRTL)
1089 /* fold_builtin_logarithm */
1090 (if (flag_unsafe_math_optimizations)
1091 /* Special case, optimize logN(expN(x)) = x. */
1092 (for logs (LOG LOG2 LOG10)
1093 exps (EXP EXP2 EXP10)
1097 /* Optimize logN(func()) for various exponential functions. We
1098 want to determine the value "x" and the power "exponent" in
1099 order to transform logN(x**exponent) into exponent*logN(x). */
1100 (for logs (LOG LOG LOG LOG
1102 LOG10 LOG10 LOG10 LOG10)
1103 exps (EXP EXP2 EXP10 POW10)
1110 CASE_FLT_FN (BUILT_IN_EXP):
1111 /* Prepare to do logN(exp(exponent) -> exponent*logN(e). */
1112 x = build_real (type, real_value_truncate (TYPE_MODE (type),
1115 CASE_FLT_FN (BUILT_IN_EXP2):
1116 /* Prepare to do logN(exp2(exponent) -> exponent*logN(2). */
1117 x = build_real (type, dconst2);
1119 CASE_FLT_FN (BUILT_IN_EXP10):
1120 CASE_FLT_FN (BUILT_IN_POW10):
1121 /* Prepare to do logN(exp10(exponent) -> exponent*logN(10). */
1123 REAL_VALUE_TYPE dconst10;
1124 real_from_integer (&dconst10, VOIDmode, 10, SIGNED);
1125 x = build_real (type, dconst10);
1130 (mult (logs { x; }) @0))))
1141 CASE_FLT_FN (BUILT_IN_SQRT):
1142 /* Prepare to do logN(sqrt(x) -> 0.5*logN(x). */
1143 x = build_real (type, dconsthalf);
1145 CASE_FLT_FN (BUILT_IN_CBRT):
1146 /* Prepare to do logN(cbrt(x) -> (1/3)*logN(x). */
1147 x = build_real (type, real_value_truncate (TYPE_MODE (type),
1152 (mult { x; } (logs @0)))))
1153 /* logN(pow(x,exponent) -> exponent*logN(x). */
1154 (for logs (LOG LOG2 LOG10)
1158 (mult @1 (logs @0)))))
1160 /* Narrowing of arithmetic and logical operations.
1162 These are conceptually similar to the transformations performed for
1163 the C/C++ front-ends by shorten_binary_op and shorten_compare. Long
1164 term we want to move all that code out of the front-ends into here. */
1166 /* If we have a narrowing conversion of an arithmetic operation where
1167 both operands are widening conversions from the same type as the outer
1168 narrowing conversion. Then convert the innermost operands to a suitable
1169 unsigned type (to avoid introducing undefined behaviour), perform the
1170 operation and convert the result to the desired type. */
1171 (for op (plus minus)
1173 (convert (op@4 (convert@2 @0) (convert@3 @1)))
1174 (if (INTEGRAL_TYPE_P (type)
1175 /* We check for type compatibility between @0 and @1 below,
1176 so there's no need to check that @1/@3 are integral types. */
1177 && INTEGRAL_TYPE_P (TREE_TYPE (@0))
1178 && INTEGRAL_TYPE_P (TREE_TYPE (@2))
1179 /* The precision of the type of each operand must match the
1180 precision of the mode of each operand, similarly for the
1182 && (TYPE_PRECISION (TREE_TYPE (@0))
1183 == GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (@0))))
1184 && (TYPE_PRECISION (TREE_TYPE (@1))
1185 == GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (@1))))
1186 && TYPE_PRECISION (type) == GET_MODE_PRECISION (TYPE_MODE (type))
1187 /* The inner conversion must be a widening conversion. */
1188 && TYPE_PRECISION (TREE_TYPE (@2)) > TYPE_PRECISION (TREE_TYPE (@0))
1189 && types_match (@0, @1)
1190 && types_match (@0, type)
1192 (if (TYPE_OVERFLOW_WRAPS (TREE_TYPE (@0)))
1193 (convert (op @0 @1)))
1194 (with { tree utype = unsigned_type_for (TREE_TYPE (@0)); }
1195 (convert (op (convert:utype @0) (convert:utype @1)))))))
1197 /* This is another case of narrowing, specifically when there's an outer
1198 BIT_AND_EXPR which masks off bits outside the type of the innermost
1199 operands. Like the previous case we have to convert the operands
1200 to unsigned types to avoid introducing undefined behaviour for the
1201 arithmetic operation. */
1202 (for op (minus plus)
1204 (bit_and (op@5 (convert@2 @0) (convert@3 @1)) INTEGER_CST@4)
1205 (if (INTEGRAL_TYPE_P (type)
1206 /* We check for type compatibility between @0 and @1 below,
1207 so there's no need to check that @1/@3 are integral types. */
1208 && INTEGRAL_TYPE_P (TREE_TYPE (@0))
1209 && INTEGRAL_TYPE_P (TREE_TYPE (@2))
1210 /* The precision of the type of each operand must match the
1211 precision of the mode of each operand, similarly for the
1213 && (TYPE_PRECISION (TREE_TYPE (@0))
1214 == GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (@0))))
1215 && (TYPE_PRECISION (TREE_TYPE (@1))
1216 == GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (@1))))
1217 && TYPE_PRECISION (type) == GET_MODE_PRECISION (TYPE_MODE (type))
1218 /* The inner conversion must be a widening conversion. */
1219 && TYPE_PRECISION (TREE_TYPE (@2)) > TYPE_PRECISION (TREE_TYPE (@0))
1220 && types_match (@0, @1)
1221 && (tree_int_cst_min_precision (@4, TYPE_SIGN (TREE_TYPE (@0)))
1222 <= TYPE_PRECISION (TREE_TYPE (@0)))
1223 && (TYPE_OVERFLOW_WRAPS (TREE_TYPE (@0))
1224 || tree_int_cst_sgn (@4) >= 0)
1226 (if (TYPE_OVERFLOW_WRAPS (TREE_TYPE (@0)))
1227 (with { tree ntype = TREE_TYPE (@0); }
1228 (convert (bit_and (op @0 @1) (convert:ntype @4)))))
1229 (with { tree utype = unsigned_type_for (TREE_TYPE (@0)); }
1230 (convert (bit_and (op (convert:utype @0) (convert:utype @1))
1231 (convert:utype @4)))))))