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))
370 /* (x | y) & ~(x & y) -> x ^ y */
372 (bit_and:c (bit_ior @0 @1) (bit_not (bit_and @0 @1)))
375 /* (x | y) & (~x ^ y) -> x & y */
377 (bit_and:c (bit_ior:c @0 @1) (bit_xor:c @1 (bit_not @0)))
384 (abs tree_expr_nonnegative_p@0)
388 /* Try to fold (type) X op CST -> (type) (X op ((type-x) CST))
390 For bitwise binary operations apply operand conversions to the
391 binary operation result instead of to the operands. This allows
392 to combine successive conversions and bitwise binary operations.
393 We combine the above two cases by using a conditional convert. */
394 (for bitop (bit_and bit_ior bit_xor)
396 (bitop (convert @0) (convert? @1))
397 (if (((TREE_CODE (@1) == INTEGER_CST
398 && INTEGRAL_TYPE_P (TREE_TYPE (@0))
399 && int_fits_type_p (@1, TREE_TYPE (@0)))
400 || types_match (@0, @1))
401 /* ??? This transform conflicts with fold-const.c doing
402 Convert (T)(x & c) into (T)x & (T)c, if c is an integer
403 constants (if x has signed type, the sign bit cannot be set
404 in c). This folds extension into the BIT_AND_EXPR.
405 Restrict it to GIMPLE to avoid endless recursions. */
406 && (bitop != BIT_AND_EXPR || GIMPLE)
407 && (/* That's a good idea if the conversion widens the operand, thus
408 after hoisting the conversion the operation will be narrower. */
409 TYPE_PRECISION (TREE_TYPE (@0)) < TYPE_PRECISION (type)
410 /* It's also a good idea if the conversion is to a non-integer
412 || GET_MODE_CLASS (TYPE_MODE (type)) != MODE_INT
413 /* Or if the precision of TO is not the same as the precision
415 || TYPE_PRECISION (type) != GET_MODE_PRECISION (TYPE_MODE (type))))
416 (convert (bitop @0 (convert @1))))))
418 /* Simplify (A & B) OP0 (C & B) to (A OP0 C) & B. */
419 (for bitop (bit_and bit_ior bit_xor)
421 (bitop (bit_and:c @0 @1) (bit_and @2 @1))
422 (bit_and (bitop @0 @2) @1)))
424 /* (x | CST1) & CST2 -> (x & CST2) | (CST1 & CST2) */
426 (bit_and (bit_ior @0 CONSTANT_CLASS_P@1) CONSTANT_CLASS_P@2)
427 (bit_ior (bit_and @0 @2) (bit_and @1 @2)))
429 /* Combine successive equal operations with constants. */
430 (for bitop (bit_and bit_ior bit_xor)
432 (bitop (bitop @0 CONSTANT_CLASS_P@1) CONSTANT_CLASS_P@2)
433 (bitop @0 (bitop @1 @2))))
435 /* Try simple folding for X op !X, and X op X with the help
436 of the truth_valued_p and logical_inverted_value predicates. */
437 (match truth_valued_p
439 (if (INTEGRAL_TYPE_P (type) && TYPE_PRECISION (type) == 1)))
440 (for op (tcc_comparison truth_and truth_andif truth_or truth_orif truth_xor)
441 (match truth_valued_p
443 (match truth_valued_p
446 (match (logical_inverted_value @0)
447 (bit_not truth_valued_p@0))
448 (match (logical_inverted_value @0)
449 (eq @0 integer_zerop))
450 (match (logical_inverted_value @0)
451 (ne truth_valued_p@0 integer_truep))
452 (match (logical_inverted_value @0)
453 (bit_xor truth_valued_p@0 integer_truep))
457 (bit_and:c @0 (logical_inverted_value @0))
458 { build_zero_cst (type); })
459 /* X | !X and X ^ !X -> 1, , if X is truth-valued. */
460 (for op (bit_ior bit_xor)
462 (op:c truth_valued_p@0 (logical_inverted_value @0))
463 { constant_boolean_node (true, type); }))
465 (for bitop (bit_and bit_ior)
466 rbitop (bit_ior bit_and)
467 /* (x | y) & x -> x */
468 /* (x & y) | x -> x */
470 (bitop:c (rbitop:c @0 @1) @0)
472 /* (~x | y) & x -> x & y */
473 /* (~x & y) | x -> x | y */
475 (bitop:c (rbitop:c (bit_not @0) @1) @0)
478 /* If arg1 and arg2 are booleans (or any single bit type)
479 then try to simplify:
486 But only do this if our result feeds into a comparison as
487 this transformation is not always a win, particularly on
488 targets with and-not instructions.
489 -> simplify_bitwise_binary_boolean */
491 (ne (bit_and:c (bit_not @0) @1) integer_zerop)
492 (if (INTEGRAL_TYPE_P (TREE_TYPE (@1))
493 && TYPE_PRECISION (TREE_TYPE (@1)) == 1)
496 (ne (bit_ior:c (bit_not @0) @1) integer_zerop)
497 (if (INTEGRAL_TYPE_P (TREE_TYPE (@1))
498 && TYPE_PRECISION (TREE_TYPE (@1)) == 1)
503 (bit_not (bit_not @0))
506 /* (x & ~m) | (y & m) -> ((x ^ y) & m) ^ x */
508 (bit_ior:c (bit_and:c@3 @0 (bit_not @2)) (bit_and:c@4 @1 @2))
509 (if (single_use (@3) && single_use (@4))
510 (bit_xor (bit_and (bit_xor @0 @1) @2) @0)))
513 /* Associate (p +p off1) +p off2 as (p +p (off1 + off2)). */
515 (pointer_plus (pointer_plus@2 @0 @1) @3)
517 || (TREE_CODE (@1) == INTEGER_CST && TREE_CODE (@3) == INTEGER_CST))
518 (pointer_plus @0 (plus @1 @3))))
524 tem4 = (unsigned long) tem3;
529 (pointer_plus @0 (convert?@2 (minus@3 (convert @1) (convert @0))))
530 /* Conditionally look through a sign-changing conversion. */
531 (if (TYPE_PRECISION (TREE_TYPE (@2)) == TYPE_PRECISION (TREE_TYPE (@3))
532 && ((GIMPLE && useless_type_conversion_p (type, TREE_TYPE (@1)))
533 || (GENERIC && type == TREE_TYPE (@1))))
537 tem = (sizetype) ptr;
541 and produce the simpler and easier to analyze with respect to alignment
542 ... = ptr & ~algn; */
544 (pointer_plus @0 (negate (bit_and (convert @0) INTEGER_CST@1)))
545 (with { tree algn = wide_int_to_tree (TREE_TYPE (@0), wi::bit_not (@1)); }
546 (bit_and @0 { algn; })))
549 /* We can't reassociate at all for saturating types. */
550 (if (!TYPE_SATURATING (type))
552 /* Contract negates. */
553 /* A + (-B) -> A - B */
555 (plus:c (convert1? @0) (convert2? (negate @1)))
556 /* Apply STRIP_NOPS on @0 and the negate. */
557 (if (tree_nop_conversion_p (type, TREE_TYPE (@0))
558 && tree_nop_conversion_p (type, TREE_TYPE (@1))
559 && !TYPE_OVERFLOW_SANITIZED (type))
560 (minus (convert @0) (convert @1))))
561 /* A - (-B) -> A + B */
563 (minus (convert1? @0) (convert2? (negate @1)))
564 (if (tree_nop_conversion_p (type, TREE_TYPE (@0))
565 && tree_nop_conversion_p (type, TREE_TYPE (@1))
566 && !TYPE_OVERFLOW_SANITIZED (type))
567 (plus (convert @0) (convert @1))))
570 (negate (convert? (negate @1)))
571 (if (tree_nop_conversion_p (type, TREE_TYPE (@1))
572 && !TYPE_OVERFLOW_SANITIZED (type))
575 /* We can't reassociate floating-point unless -fassociative-math
576 or fixed-point plus or minus because of saturation to +-Inf. */
577 (if ((!FLOAT_TYPE_P (type) || flag_associative_math)
578 && !FIXED_POINT_TYPE_P (type))
580 /* Match patterns that allow contracting a plus-minus pair
581 irrespective of overflow issues. */
582 /* (A +- B) - A -> +- B */
583 /* (A +- B) -+ B -> A */
584 /* A - (A +- B) -> -+ B */
585 /* A +- (B -+ A) -> +- B */
587 (minus (plus:c @0 @1) @0)
590 (minus (minus @0 @1) @0)
593 (plus:c (minus @0 @1) @1)
596 (minus @0 (plus:c @0 @1))
599 (minus @0 (minus @0 @1))
602 /* (A +- CST) +- CST -> A + CST */
603 (for outer_op (plus minus)
604 (for inner_op (plus minus)
606 (outer_op (inner_op @0 CONSTANT_CLASS_P@1) CONSTANT_CLASS_P@2)
607 /* If the constant operation overflows we cannot do the transform
608 as we would introduce undefined overflow, for example
609 with (a - 1) + INT_MIN. */
610 (with { tree cst = fold_binary (outer_op == inner_op
611 ? PLUS_EXPR : MINUS_EXPR, type, @1, @2); }
612 (if (cst && !TREE_OVERFLOW (cst))
613 (inner_op @0 { cst; } ))))))
615 /* (CST - A) +- CST -> CST - A */
616 (for outer_op (plus minus)
618 (outer_op (minus CONSTANT_CLASS_P@1 @0) CONSTANT_CLASS_P@2)
619 (with { tree cst = fold_binary (outer_op, type, @1, @2); }
620 (if (cst && !TREE_OVERFLOW (cst))
621 (minus { cst; } @0)))))
625 (plus:c (bit_not @0) @0)
626 (if (!TYPE_OVERFLOW_TRAPS (type))
627 { build_all_ones_cst (type); }))
631 (plus (convert? (bit_not @0)) integer_each_onep)
632 (if (tree_nop_conversion_p (type, TREE_TYPE (@0)))
633 (negate (convert @0))))
637 (minus (convert? (negate @0)) integer_each_onep)
638 (if (!TYPE_OVERFLOW_TRAPS (type)
639 && tree_nop_conversion_p (type, TREE_TYPE (@0)))
640 (bit_not (convert @0))))
644 (minus integer_all_onesp @0)
647 /* (T)(P + A) - (T)P -> (T) A */
648 (for add (plus pointer_plus)
650 (minus (convert (add @0 @1))
652 (if (element_precision (type) <= element_precision (TREE_TYPE (@1))
653 /* For integer types, if A has a smaller type
654 than T the result depends on the possible
656 E.g. T=size_t, A=(unsigned)429497295, P>0.
657 However, if an overflow in P + A would cause
658 undefined behavior, we can assume that there
660 || (INTEGRAL_TYPE_P (TREE_TYPE (@0))
661 && TYPE_OVERFLOW_UNDEFINED (TREE_TYPE (@0)))
662 /* For pointer types, if the conversion of A to the
663 final type requires a sign- or zero-extension,
664 then we have to punt - it is not defined which
666 || (POINTER_TYPE_P (TREE_TYPE (@0))
667 && TREE_CODE (@1) == INTEGER_CST
668 && tree_int_cst_sign_bit (@1) == 0))
672 /* Simplifications of MIN_EXPR and MAX_EXPR. */
674 (for minmax (min max)
680 (if (INTEGRAL_TYPE_P (type)
681 && TYPE_MIN_VALUE (type)
682 && operand_equal_p (@1, TYPE_MIN_VALUE (type), OEP_ONLY_CONST))
686 (if (INTEGRAL_TYPE_P (type)
687 && TYPE_MAX_VALUE (type)
688 && operand_equal_p (@1, TYPE_MAX_VALUE (type), OEP_ONLY_CONST))
692 /* Simplifications of shift and rotates. */
694 (for rotate (lrotate rrotate)
696 (rotate integer_all_onesp@0 @1)
699 /* Optimize -1 >> x for arithmetic right shifts. */
701 (rshift integer_all_onesp@0 @1)
702 (if (!TYPE_UNSIGNED (type)
703 && tree_expr_nonnegative_p (@1))
706 (for shiftrotate (lrotate rrotate lshift rshift)
708 (shiftrotate @0 integer_zerop)
711 (shiftrotate integer_zerop@0 @1)
713 /* Prefer vector1 << scalar to vector1 << vector2
714 if vector2 is uniform. */
715 (for vec (VECTOR_CST CONSTRUCTOR)
717 (shiftrotate @0 vec@1)
718 (with { tree tem = uniform_vector_p (@1); }
720 (shiftrotate @0 { tem; }))))))
722 /* Rewrite an LROTATE_EXPR by a constant into an
723 RROTATE_EXPR by a new constant. */
725 (lrotate @0 INTEGER_CST@1)
726 (rrotate @0 { fold_binary (MINUS_EXPR, TREE_TYPE (@1),
727 build_int_cst (TREE_TYPE (@1),
728 element_precision (type)), @1); }))
730 /* ((1 << A) & 1) != 0 -> A == 0
731 ((1 << A) & 1) == 0 -> A != 0 */
735 (cmp (bit_and (lshift integer_onep @0) integer_onep) integer_zerop)
736 (icmp @0 { build_zero_cst (TREE_TYPE (@0)); })))
738 /* (CST1 << A) == CST2 -> A == ctz (CST2) - ctz (CST1)
739 (CST1 << A) != CST2 -> A != ctz (CST2) - ctz (CST1)
743 (cmp (lshift INTEGER_CST@0 @1) INTEGER_CST@2)
744 (with { int cand = wi::ctz (@2) - wi::ctz (@0); }
746 || (!integer_zerop (@2)
747 && wi::ne_p (wi::lshift (@0, cand), @2)))
748 { constant_boolean_node (cmp == NE_EXPR, type); })
749 (if (!integer_zerop (@2)
750 && wi::eq_p (wi::lshift (@0, cand), @2))
751 (cmp @1 { build_int_cst (TREE_TYPE (@1), cand); })))))
753 /* Fold (X << C1) & C2 into (X << C1) & (C2 | ((1 << C1) - 1))
754 (X >> C1) & C2 into (X >> C1) & (C2 | ~((type) -1 >> C1))
755 if the new mask might be further optimized. */
756 (for shift (lshift rshift)
758 (bit_and (convert?@4 (shift@5 (convert1?@3 @0) INTEGER_CST@1)) INTEGER_CST@2)
759 (if (tree_nop_conversion_p (TREE_TYPE (@4), TREE_TYPE (@5))
760 && TYPE_PRECISION (type) <= HOST_BITS_PER_WIDE_INT
761 && tree_fits_uhwi_p (@1)
762 && tree_to_uhwi (@1) > 0
763 && tree_to_uhwi (@1) < TYPE_PRECISION (type))
766 unsigned int shiftc = tree_to_uhwi (@1);
767 unsigned HOST_WIDE_INT mask = TREE_INT_CST_LOW (@2);
768 unsigned HOST_WIDE_INT newmask, zerobits = 0;
769 tree shift_type = TREE_TYPE (@3);
772 if (shift == LSHIFT_EXPR)
773 zerobits = ((((unsigned HOST_WIDE_INT) 1) << shiftc) - 1);
774 else if (shift == RSHIFT_EXPR
775 && (TYPE_PRECISION (shift_type)
776 == GET_MODE_PRECISION (TYPE_MODE (shift_type))))
778 prec = TYPE_PRECISION (TREE_TYPE (@3));
780 /* See if more bits can be proven as zero because of
783 && TYPE_UNSIGNED (TREE_TYPE (@0)))
785 tree inner_type = TREE_TYPE (@0);
786 if ((TYPE_PRECISION (inner_type)
787 == GET_MODE_PRECISION (TYPE_MODE (inner_type)))
788 && TYPE_PRECISION (inner_type) < prec)
790 prec = TYPE_PRECISION (inner_type);
791 /* See if we can shorten the right shift. */
793 shift_type = inner_type;
794 /* Otherwise X >> C1 is all zeros, so we'll optimize
795 it into (X, 0) later on by making sure zerobits
799 zerobits = ~(unsigned HOST_WIDE_INT) 0;
802 zerobits >>= HOST_BITS_PER_WIDE_INT - shiftc;
803 zerobits <<= prec - shiftc;
805 /* For arithmetic shift if sign bit could be set, zerobits
806 can contain actually sign bits, so no transformation is
807 possible, unless MASK masks them all away. In that
808 case the shift needs to be converted into logical shift. */
809 if (!TYPE_UNSIGNED (TREE_TYPE (@3))
810 && prec == TYPE_PRECISION (TREE_TYPE (@3)))
812 if ((mask & zerobits) == 0)
813 shift_type = unsigned_type_for (TREE_TYPE (@3));
819 /* ((X << 16) & 0xff00) is (X, 0). */
820 (if ((mask & zerobits) == mask)
821 { build_int_cst (type, 0); })
822 (with { newmask = mask | zerobits; }
823 (if (newmask != mask && (newmask & (newmask + 1)) == 0)
826 /* Only do the transformation if NEWMASK is some integer
828 for (prec = BITS_PER_UNIT;
829 prec < HOST_BITS_PER_WIDE_INT; prec <<= 1)
830 if (newmask == (((unsigned HOST_WIDE_INT) 1) << prec) - 1)
833 (if (prec < HOST_BITS_PER_WIDE_INT
834 || newmask == ~(unsigned HOST_WIDE_INT) 0)
836 { tree newmaskt = build_int_cst_type (TREE_TYPE (@2), newmask); }
837 (if (!tree_int_cst_equal (newmaskt, @2))
838 (if (shift_type != TREE_TYPE (@3)
839 && single_use (@4) && single_use (@5))
840 (bit_and (convert (shift:shift_type (convert @3) @1)) { newmaskt; }))
841 (if (shift_type == TREE_TYPE (@3))
842 (bit_and @4 { newmaskt; }))))))))))))
844 /* Simplifications of conversions. */
846 /* Basic strip-useless-type-conversions / strip_nops. */
847 (for cvt (convert view_convert float fix_trunc)
850 (if ((GIMPLE && useless_type_conversion_p (type, TREE_TYPE (@0)))
851 || (GENERIC && type == TREE_TYPE (@0)))
854 /* Contract view-conversions. */
856 (view_convert (view_convert @0))
859 /* For integral conversions with the same precision or pointer
860 conversions use a NOP_EXPR instead. */
863 (if ((INTEGRAL_TYPE_P (type) || POINTER_TYPE_P (type))
864 && (INTEGRAL_TYPE_P (TREE_TYPE (@0)) || POINTER_TYPE_P (TREE_TYPE (@0)))
865 && TYPE_PRECISION (type) == TYPE_PRECISION (TREE_TYPE (@0)))
868 /* Strip inner integral conversions that do not change precision or size. */
870 (view_convert (convert@0 @1))
871 (if ((INTEGRAL_TYPE_P (TREE_TYPE (@0)) || POINTER_TYPE_P (TREE_TYPE (@0)))
872 && (INTEGRAL_TYPE_P (TREE_TYPE (@1)) || POINTER_TYPE_P (TREE_TYPE (@1)))
873 && (TYPE_PRECISION (TREE_TYPE (@0)) == TYPE_PRECISION (TREE_TYPE (@1)))
874 && (TYPE_SIZE (TREE_TYPE (@0)) == TYPE_SIZE (TREE_TYPE (@1))))
877 /* Re-association barriers around constants and other re-association
878 barriers can be removed. */
880 (paren CONSTANT_CLASS_P@0)
886 /* Handle cases of two conversions in a row. */
887 (for ocvt (convert float fix_trunc)
888 (for icvt (convert float)
893 tree inside_type = TREE_TYPE (@0);
894 tree inter_type = TREE_TYPE (@1);
895 int inside_int = INTEGRAL_TYPE_P (inside_type);
896 int inside_ptr = POINTER_TYPE_P (inside_type);
897 int inside_float = FLOAT_TYPE_P (inside_type);
898 int inside_vec = VECTOR_TYPE_P (inside_type);
899 unsigned int inside_prec = TYPE_PRECISION (inside_type);
900 int inside_unsignedp = TYPE_UNSIGNED (inside_type);
901 int inter_int = INTEGRAL_TYPE_P (inter_type);
902 int inter_ptr = POINTER_TYPE_P (inter_type);
903 int inter_float = FLOAT_TYPE_P (inter_type);
904 int inter_vec = VECTOR_TYPE_P (inter_type);
905 unsigned int inter_prec = TYPE_PRECISION (inter_type);
906 int inter_unsignedp = TYPE_UNSIGNED (inter_type);
907 int final_int = INTEGRAL_TYPE_P (type);
908 int final_ptr = POINTER_TYPE_P (type);
909 int final_float = FLOAT_TYPE_P (type);
910 int final_vec = VECTOR_TYPE_P (type);
911 unsigned int final_prec = TYPE_PRECISION (type);
912 int final_unsignedp = TYPE_UNSIGNED (type);
914 /* In addition to the cases of two conversions in a row
915 handled below, if we are converting something to its own
916 type via an object of identical or wider precision, neither
917 conversion is needed. */
918 (if (((GIMPLE && useless_type_conversion_p (type, inside_type))
920 && TYPE_MAIN_VARIANT (type) == TYPE_MAIN_VARIANT (inside_type)))
921 && (((inter_int || inter_ptr) && final_int)
922 || (inter_float && final_float))
923 && inter_prec >= final_prec)
926 /* Likewise, if the intermediate and initial types are either both
927 float or both integer, we don't need the middle conversion if the
928 former is wider than the latter and doesn't change the signedness
929 (for integers). Avoid this if the final type is a pointer since
930 then we sometimes need the middle conversion. Likewise if the
931 final type has a precision not equal to the size of its mode. */
932 (if (((inter_int && inside_int) || (inter_float && inside_float))
933 && (final_int || final_float)
934 && inter_prec >= inside_prec
935 && (inter_float || inter_unsignedp == inside_unsignedp)
936 && ! (final_prec != GET_MODE_PRECISION (TYPE_MODE (type))
937 && TYPE_MODE (type) == TYPE_MODE (inter_type)))
940 /* If we have a sign-extension of a zero-extended value, we can
941 replace that by a single zero-extension. Likewise if the
942 final conversion does not change precision we can drop the
943 intermediate conversion. */
944 (if (inside_int && inter_int && final_int
945 && ((inside_prec < inter_prec && inter_prec < final_prec
946 && inside_unsignedp && !inter_unsignedp)
947 || final_prec == inter_prec))
950 /* Two conversions in a row are not needed unless:
951 - some conversion is floating-point (overstrict for now), or
952 - some conversion is a vector (overstrict for now), or
953 - the intermediate type is narrower than both initial and
955 - the intermediate type and innermost type differ in signedness,
956 and the outermost type is wider than the intermediate, or
957 - the initial type is a pointer type and the precisions of the
958 intermediate and final types differ, or
959 - the final type is a pointer type and the precisions of the
960 initial and intermediate types differ. */
961 (if (! inside_float && ! inter_float && ! final_float
962 && ! inside_vec && ! inter_vec && ! final_vec
963 && (inter_prec >= inside_prec || inter_prec >= final_prec)
964 && ! (inside_int && inter_int
965 && inter_unsignedp != inside_unsignedp
966 && inter_prec < final_prec)
967 && ((inter_unsignedp && inter_prec > inside_prec)
968 == (final_unsignedp && final_prec > inter_prec))
969 && ! (inside_ptr && inter_prec != final_prec)
970 && ! (final_ptr && inside_prec != inter_prec)
971 && ! (final_prec != GET_MODE_PRECISION (TYPE_MODE (type))
972 && TYPE_MODE (type) == TYPE_MODE (inter_type)))
975 /* A truncation to an unsigned type (a zero-extension) should be
976 canonicalized as bitwise and of a mask. */
977 (if (final_int && inter_int && inside_int
978 && final_prec == inside_prec
979 && final_prec > inter_prec
981 (convert (bit_and @0 { wide_int_to_tree
983 wi::mask (inter_prec, false,
984 TYPE_PRECISION (inside_type))); })))
986 /* If we are converting an integer to a floating-point that can
987 represent it exactly and back to an integer, we can skip the
988 floating-point conversion. */
989 (if (GIMPLE /* PR66211 */
990 && inside_int && inter_float && final_int &&
991 (unsigned) significand_size (TYPE_MODE (inter_type))
992 >= inside_prec - !inside_unsignedp)
995 /* If we have a narrowing conversion to an integral type that is fed by a
996 BIT_AND_EXPR, we might be able to remove the BIT_AND_EXPR if it merely
997 masks off bits outside the final type (and nothing else). */
999 (convert (bit_and @0 INTEGER_CST@1))
1000 (if (INTEGRAL_TYPE_P (type)
1001 && INTEGRAL_TYPE_P (TREE_TYPE (@0))
1002 && TYPE_PRECISION (type) <= TYPE_PRECISION (TREE_TYPE (@0))
1003 && operand_equal_p (@1, build_low_bits_mask (TREE_TYPE (@1),
1004 TYPE_PRECISION (type)), 0))
1008 /* (X /[ex] A) * A -> X. */
1010 (mult (convert? (exact_div @0 @1)) @1)
1011 /* Look through a sign-changing conversion. */
1014 /* Canonicalization of binary operations. */
1016 /* Convert X + -C into X - C. */
1018 (plus @0 REAL_CST@1)
1019 (if (REAL_VALUE_NEGATIVE (TREE_REAL_CST (@1)))
1020 (with { tree tem = fold_unary (NEGATE_EXPR, type, @1); }
1021 (if (!TREE_OVERFLOW (tem) || !flag_trapping_math)
1022 (minus @0 { tem; })))))
1024 /* Convert x+x into x*2.0. */
1027 (if (SCALAR_FLOAT_TYPE_P (type))
1028 (mult @0 { build_real (type, dconst2); })))
1031 (minus integer_zerop @1)
1034 /* (ARG0 - ARG1) is the same as (-ARG1 + ARG0). So check whether
1035 ARG0 is zero and X + ARG0 reduces to X, since that would mean
1036 (-ARG1 + ARG0) reduces to -ARG1. */
1038 (minus real_zerop@0 @1)
1039 (if (fold_real_zero_addition_p (type, @0, 0))
1042 /* Transform x * -1 into -x. */
1044 (mult @0 integer_minus_onep)
1047 /* COMPLEX_EXPR and REALPART/IMAGPART_EXPR cancellations. */
1049 (complex (realpart @0) (imagpart @0))
1052 (realpart (complex @0 @1))
1055 (imagpart (complex @0 @1))
1059 /* BSWAP simplifications, transforms checked by gcc.dg/builtin-bswap-8.c. */
1060 (for bswap (BUILT_IN_BSWAP16 BUILT_IN_BSWAP32 BUILT_IN_BSWAP64)
1065 (bswap (bit_not (bswap @0)))
1067 (for bitop (bit_xor bit_ior bit_and)
1069 (bswap (bitop:c (bswap @0) @1))
1070 (bitop @0 (bswap @1)))))
1073 /* Combine COND_EXPRs and VEC_COND_EXPRs. */
1075 /* Simplify constant conditions.
1076 Only optimize constant conditions when the selected branch
1077 has the same type as the COND_EXPR. This avoids optimizing
1078 away "c ? x : throw", where the throw has a void type.
1079 Note that we cannot throw away the fold-const.c variant nor
1080 this one as we depend on doing this transform before possibly
1081 A ? B : B -> B triggers and the fold-const.c one can optimize
1082 0 ? A : B to B even if A has side-effects. Something
1083 genmatch cannot handle. */
1085 (cond INTEGER_CST@0 @1 @2)
1086 (if (integer_zerop (@0)
1087 && (!VOID_TYPE_P (TREE_TYPE (@2))
1088 || VOID_TYPE_P (type)))
1090 (if (!integer_zerop (@0)
1091 && (!VOID_TYPE_P (TREE_TYPE (@1))
1092 || VOID_TYPE_P (type)))
1095 (vec_cond VECTOR_CST@0 @1 @2)
1096 (if (integer_all_onesp (@0))
1098 (if (integer_zerop (@0))
1101 (for cnd (cond vec_cond)
1102 /* A ? B : (A ? X : C) -> A ? B : C. */
1104 (cnd @0 (cnd @0 @1 @2) @3)
1107 (cnd @0 @1 (cnd @0 @2 @3))
1110 /* A ? B : B -> B. */
1115 /* !A ? B : C -> A ? C : B. */
1117 (cnd (logical_inverted_value truth_valued_p@0) @1 @2)
1120 /* A + (B vcmp C ? 1 : 0) -> A - (B vcmp C), since vector comparisons
1121 return all-1 or all-0 results. */
1122 /* ??? We could instead convert all instances of the vec_cond to negate,
1123 but that isn't necessarily a win on its own. */
1125 (plus:c @3 (view_convert? (vec_cond @0 integer_each_onep@1 integer_zerop@2)))
1126 (if (VECTOR_TYPE_P (type)
1127 && TYPE_VECTOR_SUBPARTS (type) == TYPE_VECTOR_SUBPARTS (TREE_TYPE (@0))
1128 && (TYPE_MODE (TREE_TYPE (type))
1129 == TYPE_MODE (TREE_TYPE (TREE_TYPE (@0)))))
1130 (minus @3 (view_convert @0))))
1132 /* ... likewise A - (B vcmp C ? 1 : 0) -> A + (B vcmp C). */
1134 (minus @3 (view_convert? (vec_cond @0 integer_each_onep@1 integer_zerop@2)))
1135 (if (VECTOR_TYPE_P (type)
1136 && TYPE_VECTOR_SUBPARTS (type) == TYPE_VECTOR_SUBPARTS (TREE_TYPE (@0))
1137 && (TYPE_MODE (TREE_TYPE (type))
1138 == TYPE_MODE (TREE_TYPE (TREE_TYPE (@0)))))
1139 (plus @3 (view_convert @0))))
1141 /* Simplifications of comparisons. */
1143 /* We can simplify a logical negation of a comparison to the
1144 inverted comparison. As we cannot compute an expression
1145 operator using invert_tree_comparison we have to simulate
1146 that with expression code iteration. */
1147 (for cmp (tcc_comparison)
1148 icmp (inverted_tcc_comparison)
1149 ncmp (inverted_tcc_comparison_with_nans)
1150 /* Ideally we'd like to combine the following two patterns
1151 and handle some more cases by using
1152 (logical_inverted_value (cmp @0 @1))
1153 here but for that genmatch would need to "inline" that.
1154 For now implement what forward_propagate_comparison did. */
1156 (bit_not (cmp @0 @1))
1157 (if (VECTOR_TYPE_P (type)
1158 || (INTEGRAL_TYPE_P (type) && TYPE_PRECISION (type) == 1))
1159 /* Comparison inversion may be impossible for trapping math,
1160 invert_tree_comparison will tell us. But we can't use
1161 a computed operator in the replacement tree thus we have
1162 to play the trick below. */
1163 (with { enum tree_code ic = invert_tree_comparison
1164 (cmp, HONOR_NANS (@0)); }
1170 (bit_xor (cmp @0 @1) integer_truep)
1171 (with { enum tree_code ic = invert_tree_comparison
1172 (cmp, HONOR_NANS (@0)); }
1178 /* Unordered tests if either argument is a NaN. */
1180 (bit_ior (unordered @0 @0) (unordered @1 @1))
1181 (if (types_match (@0, @1))
1184 (bit_and (ordered @0 @0) (ordered @1 @1))
1185 (if (types_match (@0, @1))
1188 (bit_ior:c (unordered @0 @0) (unordered:c@2 @0 @1))
1191 (bit_and:c (ordered @0 @0) (ordered:c@2 @0 @1))
1194 /* -A CMP -B -> B CMP A. */
1195 (for cmp (tcc_comparison)
1196 scmp (swapped_tcc_comparison)
1198 (cmp (negate @0) (negate @1))
1199 (if (FLOAT_TYPE_P (TREE_TYPE (@0))
1200 || (ANY_INTEGRAL_TYPE_P (TREE_TYPE (@0))
1201 && TYPE_OVERFLOW_UNDEFINED (TREE_TYPE (@0))))
1204 (cmp (negate @0) CONSTANT_CLASS_P@1)
1205 (if (FLOAT_TYPE_P (TREE_TYPE (@0))
1206 || (ANY_INTEGRAL_TYPE_P (TREE_TYPE (@0))
1207 && TYPE_OVERFLOW_UNDEFINED (TREE_TYPE (@0))))
1208 (with { tree tem = fold_unary (NEGATE_EXPR, TREE_TYPE (@0), @1); }
1209 (if (tem && !TREE_OVERFLOW (tem))
1210 (scmp @0 { tem; }))))))
1213 /* Equality compare simplifications from fold_binary */
1216 /* If we have (A | C) == D where C & ~D != 0, convert this into 0.
1217 Similarly for NE_EXPR. */
1219 (cmp (convert?@3 (bit_ior @0 INTEGER_CST@1)) INTEGER_CST@2)
1220 (if (tree_nop_conversion_p (TREE_TYPE (@3), TREE_TYPE (@0))
1221 && wi::bit_and_not (@1, @2) != 0)
1222 { constant_boolean_node (cmp == NE_EXPR, type); }))
1224 /* (X ^ Y) == 0 becomes X == Y, and (X ^ Y) != 0 becomes X != Y. */
1226 (cmp (bit_xor @0 @1) integer_zerop)
1229 /* (X ^ Y) == Y becomes X == 0.
1230 Likewise (X ^ Y) == X becomes Y == 0. */
1232 (cmp (bit_xor:c @0 @1) @0)
1233 (cmp @1 { build_zero_cst (TREE_TYPE (@1)); }))
1235 /* (X ^ C1) op C2 can be rewritten as X op (C1 ^ C2). */
1237 (cmp (convert?@3 (bit_xor @0 INTEGER_CST@1)) INTEGER_CST@2)
1238 (if (tree_nop_conversion_p (TREE_TYPE (@3), TREE_TYPE (@0)))
1239 (cmp @0 (bit_xor @1 (convert @2))))))
1241 /* Simplification of math builtins. */
1243 (define_operator_list LOG BUILT_IN_LOGF BUILT_IN_LOG BUILT_IN_LOGL)
1244 (define_operator_list EXP BUILT_IN_EXPF BUILT_IN_EXP BUILT_IN_EXPL)
1245 (define_operator_list LOG2 BUILT_IN_LOG2F BUILT_IN_LOG2 BUILT_IN_LOG2L)
1246 (define_operator_list EXP2 BUILT_IN_EXP2F BUILT_IN_EXP2 BUILT_IN_EXP2L)
1247 (define_operator_list LOG10 BUILT_IN_LOG10F BUILT_IN_LOG10 BUILT_IN_LOG10L)
1248 (define_operator_list EXP10 BUILT_IN_EXP10F BUILT_IN_EXP10 BUILT_IN_EXP10L)
1249 (define_operator_list POW BUILT_IN_POWF BUILT_IN_POW BUILT_IN_POWL)
1250 (define_operator_list POW10 BUILT_IN_POW10F BUILT_IN_POW10 BUILT_IN_POW10L)
1251 (define_operator_list SQRT BUILT_IN_SQRTF BUILT_IN_SQRT BUILT_IN_SQRTL)
1252 (define_operator_list CBRT BUILT_IN_CBRTF BUILT_IN_CBRT BUILT_IN_CBRTL)
1255 /* fold_builtin_logarithm */
1256 (if (flag_unsafe_math_optimizations)
1257 /* Special case, optimize logN(expN(x)) = x. */
1258 (for logs (LOG LOG2 LOG10)
1259 exps (EXP EXP2 EXP10)
1263 /* Optimize logN(func()) for various exponential functions. We
1264 want to determine the value "x" and the power "exponent" in
1265 order to transform logN(x**exponent) into exponent*logN(x). */
1266 (for logs (LOG LOG LOG LOG
1268 LOG10 LOG10 LOG10 LOG10)
1269 exps (EXP EXP2 EXP10 POW10)
1276 CASE_FLT_FN (BUILT_IN_EXP):
1277 /* Prepare to do logN(exp(exponent) -> exponent*logN(e). */
1278 x = build_real (type, real_value_truncate (TYPE_MODE (type),
1281 CASE_FLT_FN (BUILT_IN_EXP2):
1282 /* Prepare to do logN(exp2(exponent) -> exponent*logN(2). */
1283 x = build_real (type, dconst2);
1285 CASE_FLT_FN (BUILT_IN_EXP10):
1286 CASE_FLT_FN (BUILT_IN_POW10):
1287 /* Prepare to do logN(exp10(exponent) -> exponent*logN(10). */
1289 REAL_VALUE_TYPE dconst10;
1290 real_from_integer (&dconst10, VOIDmode, 10, SIGNED);
1291 x = build_real (type, dconst10);
1296 (mult (logs { x; }) @0))))
1307 CASE_FLT_FN (BUILT_IN_SQRT):
1308 /* Prepare to do logN(sqrt(x) -> 0.5*logN(x). */
1309 x = build_real (type, dconsthalf);
1311 CASE_FLT_FN (BUILT_IN_CBRT):
1312 /* Prepare to do logN(cbrt(x) -> (1/3)*logN(x). */
1313 x = build_real (type, real_value_truncate (TYPE_MODE (type),
1318 (mult { x; } (logs @0)))))
1319 /* logN(pow(x,exponent) -> exponent*logN(x). */
1320 (for logs (LOG LOG2 LOG10)
1324 (mult @1 (logs @0)))))
1326 /* Narrowing of arithmetic and logical operations.
1328 These are conceptually similar to the transformations performed for
1329 the C/C++ front-ends by shorten_binary_op and shorten_compare. Long
1330 term we want to move all that code out of the front-ends into here. */
1332 /* If we have a narrowing conversion of an arithmetic operation where
1333 both operands are widening conversions from the same type as the outer
1334 narrowing conversion. Then convert the innermost operands to a suitable
1335 unsigned type (to avoid introducing undefined behaviour), perform the
1336 operation and convert the result to the desired type. */
1337 (for op (plus minus)
1339 (convert (op@4 (convert@2 @0) (convert@3 @1)))
1340 (if (INTEGRAL_TYPE_P (type)
1341 /* We check for type compatibility between @0 and @1 below,
1342 so there's no need to check that @1/@3 are integral types. */
1343 && INTEGRAL_TYPE_P (TREE_TYPE (@0))
1344 && INTEGRAL_TYPE_P (TREE_TYPE (@2))
1345 /* The precision of the type of each operand must match the
1346 precision of the mode of each operand, similarly for the
1348 && (TYPE_PRECISION (TREE_TYPE (@0))
1349 == GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (@0))))
1350 && (TYPE_PRECISION (TREE_TYPE (@1))
1351 == GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (@1))))
1352 && TYPE_PRECISION (type) == GET_MODE_PRECISION (TYPE_MODE (type))
1353 /* The inner conversion must be a widening conversion. */
1354 && TYPE_PRECISION (TREE_TYPE (@2)) > TYPE_PRECISION (TREE_TYPE (@0))
1355 && types_match (@0, @1)
1356 && types_match (@0, type)
1358 (if (TYPE_OVERFLOW_WRAPS (TREE_TYPE (@0)))
1359 (convert (op @0 @1)))
1360 (with { tree utype = unsigned_type_for (TREE_TYPE (@0)); }
1361 (convert (op (convert:utype @0) (convert:utype @1)))))))
1363 /* This is another case of narrowing, specifically when there's an outer
1364 BIT_AND_EXPR which masks off bits outside the type of the innermost
1365 operands. Like the previous case we have to convert the operands
1366 to unsigned types to avoid introducing undefined behaviour for the
1367 arithmetic operation. */
1368 (for op (minus plus)
1370 (bit_and (op@5 (convert@2 @0) (convert@3 @1)) INTEGER_CST@4)
1371 (if (INTEGRAL_TYPE_P (type)
1372 /* We check for type compatibility between @0 and @1 below,
1373 so there's no need to check that @1/@3 are integral types. */
1374 && INTEGRAL_TYPE_P (TREE_TYPE (@0))
1375 && INTEGRAL_TYPE_P (TREE_TYPE (@2))
1376 /* The precision of the type of each operand must match the
1377 precision of the mode of each operand, similarly for the
1379 && (TYPE_PRECISION (TREE_TYPE (@0))
1380 == GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (@0))))
1381 && (TYPE_PRECISION (TREE_TYPE (@1))
1382 == GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (@1))))
1383 && TYPE_PRECISION (type) == GET_MODE_PRECISION (TYPE_MODE (type))
1384 /* The inner conversion must be a widening conversion. */
1385 && TYPE_PRECISION (TREE_TYPE (@2)) > TYPE_PRECISION (TREE_TYPE (@0))
1386 && types_match (@0, @1)
1387 && (tree_int_cst_min_precision (@4, TYPE_SIGN (TREE_TYPE (@0)))
1388 <= TYPE_PRECISION (TREE_TYPE (@0)))
1389 && (TYPE_OVERFLOW_WRAPS (TREE_TYPE (@0))
1390 || tree_int_cst_sgn (@4) >= 0)
1392 (if (TYPE_OVERFLOW_WRAPS (TREE_TYPE (@0)))
1393 (with { tree ntype = TREE_TYPE (@0); }
1394 (convert (bit_and (op @0 @1) (convert:ntype @4)))))
1395 (with { tree utype = unsigned_type_for (TREE_TYPE (@0)); }
1396 (convert (bit_and (op (convert:utype @0) (convert:utype @1))
1397 (convert:utype @4)))))))