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)
43 /* Simplifications of operations with one constant operand and
44 simplifications to constants or single values. */
46 (for op (plus pointer_plus minus bit_ior bit_xor)
51 /* 0 +p index -> (type)index */
53 (pointer_plus integer_zerop @1)
54 (non_lvalue (convert @1)))
56 /* See if ARG1 is zero and X + ARG1 reduces to X.
57 Likewise if the operands are reversed. */
59 (plus:c @0 real_zerop@1)
60 (if (fold_real_zero_addition_p (type, @1, 0))
63 /* See if ARG1 is zero and X - ARG1 reduces to X. */
65 (minus @0 real_zerop@1)
66 (if (fold_real_zero_addition_p (type, @1, 1))
70 This is unsafe for certain floats even in non-IEEE formats.
71 In IEEE, it is unsafe because it does wrong for NaNs.
72 Also note that operand_equal_p is always false if an operand
76 (if (!FLOAT_TYPE_P (type) || !HONOR_NANS (type))
77 { build_zero_cst (type); }))
80 (mult @0 integer_zerop@1)
83 /* Maybe fold x * 0 to 0. The expressions aren't the same
84 when x is NaN, since x * 0 is also NaN. Nor are they the
85 same in modes with signed zeros, since multiplying a
86 negative value by 0 gives -0, not +0. */
88 (mult @0 real_zerop@1)
89 (if (!HONOR_NANS (type) && !HONOR_SIGNED_ZEROS (element_mode (type)))
92 /* In IEEE floating point, x*1 is not equivalent to x for snans.
93 Likewise for complex arithmetic with signed zeros. */
96 (if (!HONOR_SNANS (element_mode (type))
97 && (!HONOR_SIGNED_ZEROS (element_mode (type))
98 || !COMPLEX_FLOAT_TYPE_P (type)))
101 /* Transform x * -1.0 into -x. */
103 (mult @0 real_minus_onep)
104 (if (!HONOR_SNANS (element_mode (type))
105 && (!HONOR_SIGNED_ZEROS (element_mode (type))
106 || !COMPLEX_FLOAT_TYPE_P (type)))
109 /* Make sure to preserve divisions by zero. This is the reason why
110 we don't simplify x / x to 1 or 0 / x to 0. */
111 (for op (mult trunc_div ceil_div floor_div round_div exact_div)
117 (for div (trunc_div ceil_div floor_div round_div exact_div)
119 (div @0 integer_minus_onep@1)
120 (if (!TYPE_UNSIGNED (type))
123 /* For unsigned integral types, FLOOR_DIV_EXPR is the same as
124 TRUNC_DIV_EXPR. Rewrite into the latter in this case. */
127 (if ((INTEGRAL_TYPE_P (type) || VECTOR_INTEGER_TYPE_P (type))
128 && TYPE_UNSIGNED (type))
131 /* Combine two successive divisions. Note that combining ceil_div
132 and floor_div is trickier and combining round_div even more so. */
133 (for div (trunc_div exact_div)
135 (div (div @0 INTEGER_CST@1) INTEGER_CST@2)
138 wide_int mul = wi::mul (@1, @2, TYPE_SIGN (type), &overflow_p);
141 (div @0 { wide_int_to_tree (type, mul); }))
143 && (TYPE_UNSIGNED (type)
144 || mul != wi::min_value (TYPE_PRECISION (type), SIGNED)))
145 { build_zero_cst (type); }))))
147 /* Optimize A / A to 1.0 if we don't care about
148 NaNs or Infinities. */
151 (if (FLOAT_TYPE_P (type)
152 && ! HONOR_NANS (type)
153 && ! HONOR_INFINITIES (element_mode (type)))
154 { build_one_cst (type); }))
156 /* Optimize -A / A to -1.0 if we don't care about
157 NaNs or Infinities. */
159 (rdiv:c @0 (negate @0))
160 (if (FLOAT_TYPE_P (type)
161 && ! HONOR_NANS (type)
162 && ! HONOR_INFINITIES (element_mode (type)))
163 { build_minus_one_cst (type); }))
165 /* In IEEE floating point, x/1 is not equivalent to x for snans. */
168 (if (!HONOR_SNANS (element_mode (type)))
171 /* In IEEE floating point, x/-1 is not equivalent to -x for snans. */
173 (rdiv @0 real_minus_onep)
174 (if (!HONOR_SNANS (element_mode (type)))
177 /* If ARG1 is a constant, we can convert this to a multiply by the
178 reciprocal. This does not have the same rounding properties,
179 so only do this if -freciprocal-math. We can actually
180 always safely do it if ARG1 is a power of two, but it's hard to
181 tell if it is or not in a portable manner. */
182 (for cst (REAL_CST COMPLEX_CST VECTOR_CST)
186 (if (flag_reciprocal_math
189 { tree tem = const_binop (RDIV_EXPR, type, build_one_cst (type), @1); }
191 (mult @0 { tem; } ))))
192 (if (cst != COMPLEX_CST)
193 (with { tree inverse = exact_inverse (type, @1); }
195 (mult @0 { inverse; } )))))))
197 /* Same applies to modulo operations, but fold is inconsistent here
198 and simplifies 0 % x to 0, only preserving literal 0 % 0. */
199 (for mod (ceil_mod floor_mod round_mod trunc_mod)
200 /* 0 % X is always zero. */
202 (mod integer_zerop@0 @1)
203 /* But not for 0 % 0 so that we can get the proper warnings and errors. */
204 (if (!integer_zerop (@1))
206 /* X % 1 is always zero. */
208 (mod @0 integer_onep)
209 { build_zero_cst (type); })
210 /* X % -1 is zero. */
212 (mod @0 integer_minus_onep@1)
213 (if (!TYPE_UNSIGNED (type))
214 { build_zero_cst (type); }))
215 /* (X % Y) % Y is just X % Y. */
217 (mod (mod@2 @0 @1) @1)
220 /* X % -C is the same as X % C. */
222 (trunc_mod @0 INTEGER_CST@1)
223 (if (TYPE_SIGN (type) == SIGNED
224 && !TREE_OVERFLOW (@1)
226 && !TYPE_OVERFLOW_TRAPS (type)
227 /* Avoid this transformation if C is INT_MIN, i.e. C == -C. */
228 && !sign_bit_p (@1, @1))
229 (trunc_mod @0 (negate @1))))
231 /* X % Y is smaller than Y. */
234 (cmp (trunc_mod @0 @1) @1)
235 (if (TYPE_UNSIGNED (TREE_TYPE (@0)))
236 { constant_boolean_node (cmp == LT_EXPR, type); })))
239 (cmp @1 (trunc_mod @0 @1))
240 (if (TYPE_UNSIGNED (TREE_TYPE (@0)))
241 { constant_boolean_node (cmp == GT_EXPR, type); })))
245 (bit_ior @0 integer_all_onesp@1)
250 (bit_and @0 integer_zerop@1)
256 { build_zero_cst (type); })
258 /* Canonicalize X ^ ~0 to ~X. */
260 (bit_xor @0 integer_all_onesp@1)
265 (bit_and @0 integer_all_onesp)
268 /* x & x -> x, x | x -> x */
269 (for bitop (bit_and bit_ior)
274 /* x + (x & 1) -> (x + 1) & ~1 */
276 (plus:c @0 (bit_and@2 @0 integer_onep@1))
277 (if (TREE_CODE (@2) != SSA_NAME || has_single_use (@2))
278 (bit_and (plus @0 @1) (bit_not @1))))
280 /* x & ~(x & y) -> x & ~y */
281 /* x | ~(x | y) -> x | ~y */
282 (for bitop (bit_and bit_ior)
284 (bitop:c @0 (bit_not (bitop:c@2 @0 @1)))
285 (if (TREE_CODE (@2) != SSA_NAME || has_single_use (@2))
286 (bitop @0 (bit_not @1)))))
292 (abs tree_expr_nonnegative_p@0)
296 /* Try to fold (type) X op CST -> (type) (X op ((type-x) CST))
298 For bitwise binary operations apply operand conversions to the
299 binary operation result instead of to the operands. This allows
300 to combine successive conversions and bitwise binary operations.
301 We combine the above two cases by using a conditional convert. */
302 (for bitop (bit_and bit_ior bit_xor)
304 (bitop (convert @0) (convert? @1))
305 (if (((TREE_CODE (@1) == INTEGER_CST
306 && INTEGRAL_TYPE_P (TREE_TYPE (@0))
307 && int_fits_type_p (@1, TREE_TYPE (@0)))
308 || types_match (@0, @1))
309 /* ??? This transform conflicts with fold-const.c doing
310 Convert (T)(x & c) into (T)x & (T)c, if c is an integer
311 constants (if x has signed type, the sign bit cannot be set
312 in c). This folds extension into the BIT_AND_EXPR.
313 Restrict it to GIMPLE to avoid endless recursions. */
314 && (bitop != BIT_AND_EXPR || GIMPLE)
315 && (/* That's a good idea if the conversion widens the operand, thus
316 after hoisting the conversion the operation will be narrower. */
317 TYPE_PRECISION (TREE_TYPE (@0)) < TYPE_PRECISION (type)
318 /* It's also a good idea if the conversion is to a non-integer
320 || GET_MODE_CLASS (TYPE_MODE (type)) != MODE_INT
321 /* Or if the precision of TO is not the same as the precision
323 || TYPE_PRECISION (type) != GET_MODE_PRECISION (TYPE_MODE (type))))
324 (convert (bitop @0 (convert @1))))))
326 /* Simplify (A & B) OP0 (C & B) to (A OP0 C) & B. */
327 (for bitop (bit_and bit_ior bit_xor)
329 (bitop (bit_and:c @0 @1) (bit_and @2 @1))
330 (bit_and (bitop @0 @2) @1)))
332 /* (x | CST1) & CST2 -> (x & CST2) | (CST1 & CST2) */
334 (bit_and (bit_ior @0 CONSTANT_CLASS_P@1) CONSTANT_CLASS_P@2)
335 (bit_ior (bit_and @0 @2) (bit_and @1 @2)))
337 /* Combine successive equal operations with constants. */
338 (for bitop (bit_and bit_ior bit_xor)
340 (bitop (bitop @0 CONSTANT_CLASS_P@1) CONSTANT_CLASS_P@2)
341 (bitop @0 (bitop @1 @2))))
343 /* Try simple folding for X op !X, and X op X with the help
344 of the truth_valued_p and logical_inverted_value predicates. */
345 (match truth_valued_p
347 (if (INTEGRAL_TYPE_P (type) && TYPE_PRECISION (type) == 1)))
348 (for op (tcc_comparison truth_and truth_andif truth_or truth_orif truth_xor)
349 (match truth_valued_p
351 (match truth_valued_p
354 (match (logical_inverted_value @0)
355 (bit_not truth_valued_p@0))
356 (match (logical_inverted_value @0)
357 (eq @0 integer_zerop))
358 (match (logical_inverted_value @0)
359 (ne truth_valued_p@0 integer_truep))
360 (match (logical_inverted_value @0)
361 (bit_xor truth_valued_p@0 integer_truep))
365 (bit_and:c @0 (logical_inverted_value @0))
366 { build_zero_cst (type); })
367 /* X | !X and X ^ !X -> 1, , if X is truth-valued. */
368 (for op (bit_ior bit_xor)
370 (op:c truth_valued_p@0 (logical_inverted_value @0))
371 { constant_boolean_node (true, type); }))
373 (for bitop (bit_and bit_ior)
374 rbitop (bit_ior bit_and)
375 /* (x | y) & x -> x */
376 /* (x & y) | x -> x */
378 (bitop:c (rbitop:c @0 @1) @0)
380 /* (~x | y) & x -> x & y */
381 /* (~x & y) | x -> x | y */
383 (bitop:c (rbitop:c (bit_not @0) @1) @0)
386 /* If arg1 and arg2 are booleans (or any single bit type)
387 then try to simplify:
394 But only do this if our result feeds into a comparison as
395 this transformation is not always a win, particularly on
396 targets with and-not instructions.
397 -> simplify_bitwise_binary_boolean */
399 (ne (bit_and:c (bit_not @0) @1) integer_zerop)
400 (if (INTEGRAL_TYPE_P (TREE_TYPE (@1))
401 && TYPE_PRECISION (TREE_TYPE (@1)) == 1)
404 (ne (bit_ior:c (bit_not @0) @1) integer_zerop)
405 (if (INTEGRAL_TYPE_P (TREE_TYPE (@1))
406 && TYPE_PRECISION (TREE_TYPE (@1)) == 1)
411 (bit_not (bit_not @0))
414 /* (x & ~m) | (y & m) -> ((x ^ y) & m) ^ x */
416 (bit_ior:c (bit_and:c@3 @0 (bit_not @2)) (bit_and:c@4 @1 @2))
417 (if ((TREE_CODE (@3) != SSA_NAME || has_single_use (@3))
418 && (TREE_CODE (@4) != SSA_NAME || has_single_use (@4)))
419 (bit_xor (bit_and (bit_xor @0 @1) @2) @0)))
422 /* Associate (p +p off1) +p off2 as (p +p (off1 + off2)). */
424 (pointer_plus (pointer_plus@2 @0 @1) @3)
425 (if (TREE_CODE (@2) != SSA_NAME || has_single_use (@2))
426 (pointer_plus @0 (plus @1 @3))))
432 tem4 = (unsigned long) tem3;
437 (pointer_plus @0 (convert?@2 (minus@3 (convert @1) (convert @0))))
438 /* Conditionally look through a sign-changing conversion. */
439 (if (TYPE_PRECISION (TREE_TYPE (@2)) == TYPE_PRECISION (TREE_TYPE (@3))
440 && ((GIMPLE && useless_type_conversion_p (type, TREE_TYPE (@1)))
441 || (GENERIC && type == TREE_TYPE (@1))))
445 tem = (sizetype) ptr;
449 and produce the simpler and easier to analyze with respect to alignment
450 ... = ptr & ~algn; */
452 (pointer_plus @0 (negate (bit_and (convert @0) INTEGER_CST@1)))
453 (with { tree algn = wide_int_to_tree (TREE_TYPE (@0), wi::bit_not (@1)); }
454 (bit_and @0 { algn; })))
457 /* We can't reassociate at all for saturating types. */
458 (if (!TYPE_SATURATING (type))
460 /* Contract negates. */
461 /* A + (-B) -> A - B */
463 (plus:c (convert1? @0) (convert2? (negate @1)))
464 /* Apply STRIP_NOPS on @0 and the negate. */
465 (if (tree_nop_conversion_p (type, TREE_TYPE (@0))
466 && tree_nop_conversion_p (type, TREE_TYPE (@1))
467 && !TYPE_OVERFLOW_SANITIZED (type))
468 (minus (convert @0) (convert @1))))
469 /* A - (-B) -> A + B */
471 (minus (convert1? @0) (convert2? (negate @1)))
472 (if (tree_nop_conversion_p (type, TREE_TYPE (@0))
473 && tree_nop_conversion_p (type, TREE_TYPE (@1))
474 && !TYPE_OVERFLOW_SANITIZED (type))
475 (plus (convert @0) (convert @1))))
478 (negate (convert? (negate @1)))
479 (if (tree_nop_conversion_p (type, TREE_TYPE (@1))
480 && !TYPE_OVERFLOW_SANITIZED (type))
483 /* We can't reassociate floating-point or fixed-point plus or minus
484 because of saturation to +-Inf. */
485 (if (!FLOAT_TYPE_P (type) && !FIXED_POINT_TYPE_P (type))
487 /* Match patterns that allow contracting a plus-minus pair
488 irrespective of overflow issues. */
489 /* (A +- B) - A -> +- B */
490 /* (A +- B) -+ B -> A */
491 /* A - (A +- B) -> -+ B */
492 /* A +- (B -+ A) -> +- B */
494 (minus (plus:c @0 @1) @0)
497 (minus (minus @0 @1) @0)
500 (plus:c (minus @0 @1) @1)
503 (minus @0 (plus:c @0 @1))
506 (minus @0 (minus @0 @1))
509 /* (A +- CST) +- CST -> A + CST */
510 (for outer_op (plus minus)
511 (for inner_op (plus minus)
513 (outer_op (inner_op @0 CONSTANT_CLASS_P@1) CONSTANT_CLASS_P@2)
514 /* If the constant operation overflows we cannot do the transform
515 as we would introduce undefined overflow, for example
516 with (a - 1) + INT_MIN. */
517 (with { tree cst = fold_binary (outer_op == inner_op
518 ? PLUS_EXPR : MINUS_EXPR, type, @1, @2); }
519 (if (cst && !TREE_OVERFLOW (cst))
520 (inner_op @0 { cst; } ))))))
522 /* (CST - A) +- CST -> CST - A */
523 (for outer_op (plus minus)
525 (outer_op (minus CONSTANT_CLASS_P@1 @0) CONSTANT_CLASS_P@2)
526 (with { tree cst = fold_binary (outer_op, type, @1, @2); }
527 (if (cst && !TREE_OVERFLOW (cst))
528 (minus { cst; } @0)))))
532 (plus:c (bit_not @0) @0)
533 (if (!TYPE_OVERFLOW_TRAPS (type))
534 { build_all_ones_cst (type); }))
538 (plus (convert? (bit_not @0)) integer_each_onep)
539 (if (tree_nop_conversion_p (type, TREE_TYPE (@0)))
540 (negate (convert @0))))
544 (minus (convert? (negate @0)) integer_each_onep)
545 (if (!TYPE_OVERFLOW_TRAPS (type)
546 && tree_nop_conversion_p (type, TREE_TYPE (@0)))
547 (bit_not (convert @0))))
551 (minus integer_all_onesp @0)
554 /* (T)(P + A) - (T)P -> (T) A */
555 (for add (plus pointer_plus)
557 (minus (convert (add @0 @1))
559 (if (element_precision (type) <= element_precision (TREE_TYPE (@1))
560 /* For integer types, if A has a smaller type
561 than T the result depends on the possible
563 E.g. T=size_t, A=(unsigned)429497295, P>0.
564 However, if an overflow in P + A would cause
565 undefined behavior, we can assume that there
567 || (INTEGRAL_TYPE_P (TREE_TYPE (@0))
568 && TYPE_OVERFLOW_UNDEFINED (TREE_TYPE (@0)))
569 /* For pointer types, if the conversion of A to the
570 final type requires a sign- or zero-extension,
571 then we have to punt - it is not defined which
573 || (POINTER_TYPE_P (TREE_TYPE (@0))
574 && TREE_CODE (@1) == INTEGER_CST
575 && tree_int_cst_sign_bit (@1) == 0))
579 /* Simplifications of MIN_EXPR and MAX_EXPR. */
581 (for minmax (min max)
587 (if (INTEGRAL_TYPE_P (type)
588 && TYPE_MIN_VALUE (type)
589 && operand_equal_p (@1, TYPE_MIN_VALUE (type), OEP_ONLY_CONST))
593 (if (INTEGRAL_TYPE_P (type)
594 && TYPE_MAX_VALUE (type)
595 && operand_equal_p (@1, TYPE_MAX_VALUE (type), OEP_ONLY_CONST))
599 /* Simplifications of shift and rotates. */
601 (for rotate (lrotate rrotate)
603 (rotate integer_all_onesp@0 @1)
606 /* Optimize -1 >> x for arithmetic right shifts. */
608 (rshift integer_all_onesp@0 @1)
609 (if (!TYPE_UNSIGNED (type)
610 && tree_expr_nonnegative_p (@1))
613 (for shiftrotate (lrotate rrotate lshift rshift)
615 (shiftrotate @0 integer_zerop)
618 (shiftrotate integer_zerop@0 @1)
620 /* Prefer vector1 << scalar to vector1 << vector2
621 if vector2 is uniform. */
622 (for vec (VECTOR_CST CONSTRUCTOR)
624 (shiftrotate @0 vec@1)
625 (with { tree tem = uniform_vector_p (@1); }
627 (shiftrotate @0 { tem; }))))))
629 /* Rewrite an LROTATE_EXPR by a constant into an
630 RROTATE_EXPR by a new constant. */
632 (lrotate @0 INTEGER_CST@1)
633 (rrotate @0 { fold_binary (MINUS_EXPR, TREE_TYPE (@1),
634 build_int_cst (TREE_TYPE (@1),
635 element_precision (type)), @1); }))
637 /* ((1 << A) & 1) != 0 -> A == 0
638 ((1 << A) & 1) == 0 -> A != 0 */
642 (cmp (bit_and (lshift integer_onep @0) integer_onep) integer_zerop)
643 (icmp @0 { build_zero_cst (TREE_TYPE (@0)); })))
645 /* Simplifications of conversions. */
647 /* Basic strip-useless-type-conversions / strip_nops. */
648 (for cvt (convert view_convert float fix_trunc)
651 (if ((GIMPLE && useless_type_conversion_p (type, TREE_TYPE (@0)))
652 || (GENERIC && type == TREE_TYPE (@0)))
655 /* Contract view-conversions. */
657 (view_convert (view_convert @0))
660 /* For integral conversions with the same precision or pointer
661 conversions use a NOP_EXPR instead. */
664 (if ((INTEGRAL_TYPE_P (type) || POINTER_TYPE_P (type))
665 && (INTEGRAL_TYPE_P (TREE_TYPE (@0)) || POINTER_TYPE_P (TREE_TYPE (@0)))
666 && TYPE_PRECISION (type) == TYPE_PRECISION (TREE_TYPE (@0)))
669 /* Strip inner integral conversions that do not change precision or size. */
671 (view_convert (convert@0 @1))
672 (if ((INTEGRAL_TYPE_P (TREE_TYPE (@0)) || POINTER_TYPE_P (TREE_TYPE (@0)))
673 && (INTEGRAL_TYPE_P (TREE_TYPE (@1)) || POINTER_TYPE_P (TREE_TYPE (@1)))
674 && (TYPE_PRECISION (TREE_TYPE (@0)) == TYPE_PRECISION (TREE_TYPE (@1)))
675 && (TYPE_SIZE (TREE_TYPE (@0)) == TYPE_SIZE (TREE_TYPE (@1))))
678 /* Re-association barriers around constants and other re-association
679 barriers can be removed. */
681 (paren CONSTANT_CLASS_P@0)
687 /* Handle cases of two conversions in a row. */
688 (for ocvt (convert float fix_trunc)
689 (for icvt (convert float)
694 tree inside_type = TREE_TYPE (@0);
695 tree inter_type = TREE_TYPE (@1);
696 int inside_int = INTEGRAL_TYPE_P (inside_type);
697 int inside_ptr = POINTER_TYPE_P (inside_type);
698 int inside_float = FLOAT_TYPE_P (inside_type);
699 int inside_vec = VECTOR_TYPE_P (inside_type);
700 unsigned int inside_prec = TYPE_PRECISION (inside_type);
701 int inside_unsignedp = TYPE_UNSIGNED (inside_type);
702 int inter_int = INTEGRAL_TYPE_P (inter_type);
703 int inter_ptr = POINTER_TYPE_P (inter_type);
704 int inter_float = FLOAT_TYPE_P (inter_type);
705 int inter_vec = VECTOR_TYPE_P (inter_type);
706 unsigned int inter_prec = TYPE_PRECISION (inter_type);
707 int inter_unsignedp = TYPE_UNSIGNED (inter_type);
708 int final_int = INTEGRAL_TYPE_P (type);
709 int final_ptr = POINTER_TYPE_P (type);
710 int final_float = FLOAT_TYPE_P (type);
711 int final_vec = VECTOR_TYPE_P (type);
712 unsigned int final_prec = TYPE_PRECISION (type);
713 int final_unsignedp = TYPE_UNSIGNED (type);
715 /* In addition to the cases of two conversions in a row
716 handled below, if we are converting something to its own
717 type via an object of identical or wider precision, neither
718 conversion is needed. */
719 (if (((GIMPLE && useless_type_conversion_p (type, inside_type))
721 && TYPE_MAIN_VARIANT (type) == TYPE_MAIN_VARIANT (inside_type)))
722 && (((inter_int || inter_ptr) && final_int)
723 || (inter_float && final_float))
724 && inter_prec >= final_prec)
727 /* Likewise, if the intermediate and initial types are either both
728 float or both integer, we don't need the middle conversion if the
729 former is wider than the latter and doesn't change the signedness
730 (for integers). Avoid this if the final type is a pointer since
731 then we sometimes need the middle conversion. Likewise if the
732 final type has a precision not equal to the size of its mode. */
733 (if (((inter_int && inside_int)
734 || (inter_float && inside_float)
735 || (inter_vec && inside_vec))
736 && inter_prec >= inside_prec
737 && (inter_float || inter_vec
738 || inter_unsignedp == inside_unsignedp)
739 && ! (final_prec != GET_MODE_PRECISION (element_mode (type))
740 && element_mode (type) == element_mode (inter_type))
742 && (! final_vec || inter_prec == inside_prec))
745 /* If we have a sign-extension of a zero-extended value, we can
746 replace that by a single zero-extension. Likewise if the
747 final conversion does not change precision we can drop the
748 intermediate conversion. */
749 (if (inside_int && inter_int && final_int
750 && ((inside_prec < inter_prec && inter_prec < final_prec
751 && inside_unsignedp && !inter_unsignedp)
752 || final_prec == inter_prec))
755 /* Two conversions in a row are not needed unless:
756 - some conversion is floating-point (overstrict for now), or
757 - some conversion is a vector (overstrict for now), or
758 - the intermediate type is narrower than both initial and
760 - the intermediate type and innermost type differ in signedness,
761 and the outermost type is wider than the intermediate, or
762 - the initial type is a pointer type and the precisions of the
763 intermediate and final types differ, or
764 - the final type is a pointer type and the precisions of the
765 initial and intermediate types differ. */
766 (if (! inside_float && ! inter_float && ! final_float
767 && ! inside_vec && ! inter_vec && ! final_vec
768 && (inter_prec >= inside_prec || inter_prec >= final_prec)
769 && ! (inside_int && inter_int
770 && inter_unsignedp != inside_unsignedp
771 && inter_prec < final_prec)
772 && ((inter_unsignedp && inter_prec > inside_prec)
773 == (final_unsignedp && final_prec > inter_prec))
774 && ! (inside_ptr && inter_prec != final_prec)
775 && ! (final_ptr && inside_prec != inter_prec)
776 && ! (final_prec != GET_MODE_PRECISION (TYPE_MODE (type))
777 && TYPE_MODE (type) == TYPE_MODE (inter_type)))
780 /* A truncation to an unsigned type (a zero-extension) should be
781 canonicalized as bitwise and of a mask. */
782 (if (final_int && inter_int && inside_int
783 && final_prec == inside_prec
784 && final_prec > inter_prec
786 (convert (bit_and @0 { wide_int_to_tree
788 wi::mask (inter_prec, false,
789 TYPE_PRECISION (inside_type))); })))
791 /* If we are converting an integer to a floating-point that can
792 represent it exactly and back to an integer, we can skip the
793 floating-point conversion. */
794 (if (GIMPLE /* PR66211 */
795 && inside_int && inter_float && final_int &&
796 (unsigned) significand_size (TYPE_MODE (inter_type))
797 >= inside_prec - !inside_unsignedp)
800 /* If we have a narrowing conversion to an integral type that is fed by a
801 BIT_AND_EXPR, we might be able to remove the BIT_AND_EXPR if it merely
802 masks off bits outside the final type (and nothing else). */
804 (convert (bit_and @0 INTEGER_CST@1))
805 (if (INTEGRAL_TYPE_P (type)
806 && INTEGRAL_TYPE_P (TREE_TYPE (@0))
807 && TYPE_PRECISION (type) <= TYPE_PRECISION (TREE_TYPE (@0))
808 && operand_equal_p (@1, build_low_bits_mask (TREE_TYPE (@1),
809 TYPE_PRECISION (type)), 0))
813 /* (X /[ex] A) * A -> X. */
815 (mult (convert? (exact_div @0 @1)) @1)
816 /* Look through a sign-changing conversion. */
817 (if (TYPE_PRECISION (TREE_TYPE (@0)) == TYPE_PRECISION (type))
820 /* Canonicalization of binary operations. */
822 /* Convert X + -C into X - C. */
825 (if (REAL_VALUE_NEGATIVE (TREE_REAL_CST (@1)))
826 (with { tree tem = fold_unary (NEGATE_EXPR, type, @1); }
827 (if (!TREE_OVERFLOW (tem) || !flag_trapping_math)
828 (minus @0 { tem; })))))
830 /* Convert x+x into x*2.0. */
833 (if (SCALAR_FLOAT_TYPE_P (type))
834 (mult @0 { build_real (type, dconst2); })))
837 (minus integer_zerop @1)
840 /* (ARG0 - ARG1) is the same as (-ARG1 + ARG0). So check whether
841 ARG0 is zero and X + ARG0 reduces to X, since that would mean
842 (-ARG1 + ARG0) reduces to -ARG1. */
844 (minus real_zerop@0 @1)
845 (if (fold_real_zero_addition_p (type, @0, 0))
848 /* Transform x * -1 into -x. */
850 (mult @0 integer_minus_onep)
853 /* COMPLEX_EXPR and REALPART/IMAGPART_EXPR cancellations. */
855 (complex (realpart @0) (imagpart @0))
858 (realpart (complex @0 @1))
861 (imagpart (complex @0 @1))
865 /* BSWAP simplifications, transforms checked by gcc.dg/builtin-bswap-8.c. */
866 (for bswap (BUILT_IN_BSWAP16 BUILT_IN_BSWAP32 BUILT_IN_BSWAP64)
871 (bswap (bit_not (bswap @0)))
873 (for bitop (bit_xor bit_ior bit_and)
875 (bswap (bitop:c (bswap @0) @1))
876 (bitop @0 (bswap @1)))))
879 /* Combine COND_EXPRs and VEC_COND_EXPRs. */
881 /* Simplify constant conditions.
882 Only optimize constant conditions when the selected branch
883 has the same type as the COND_EXPR. This avoids optimizing
884 away "c ? x : throw", where the throw has a void type.
885 Note that we cannot throw away the fold-const.c variant nor
886 this one as we depend on doing this transform before possibly
887 A ? B : B -> B triggers and the fold-const.c one can optimize
888 0 ? A : B to B even if A has side-effects. Something
889 genmatch cannot handle. */
891 (cond INTEGER_CST@0 @1 @2)
892 (if (integer_zerop (@0)
893 && (!VOID_TYPE_P (TREE_TYPE (@2))
894 || VOID_TYPE_P (type)))
896 (if (!integer_zerop (@0)
897 && (!VOID_TYPE_P (TREE_TYPE (@1))
898 || VOID_TYPE_P (type)))
901 (vec_cond VECTOR_CST@0 @1 @2)
902 (if (integer_all_onesp (@0))
904 (if (integer_zerop (@0))
907 (for cnd (cond vec_cond)
908 /* A ? B : (A ? X : C) -> A ? B : C. */
910 (cnd @0 (cnd @0 @1 @2) @3)
913 (cnd @0 @1 (cnd @0 @2 @3))
916 /* A ? B : B -> B. */
921 /* !A ? B : C -> A ? C : B. */
923 (cnd (logical_inverted_value truth_valued_p@0) @1 @2)
927 /* Simplifications of comparisons. */
929 /* We can simplify a logical negation of a comparison to the
930 inverted comparison. As we cannot compute an expression
931 operator using invert_tree_comparison we have to simulate
932 that with expression code iteration. */
933 (for cmp (tcc_comparison)
934 icmp (inverted_tcc_comparison)
935 ncmp (inverted_tcc_comparison_with_nans)
936 /* Ideally we'd like to combine the following two patterns
937 and handle some more cases by using
938 (logical_inverted_value (cmp @0 @1))
939 here but for that genmatch would need to "inline" that.
940 For now implement what forward_propagate_comparison did. */
942 (bit_not (cmp @0 @1))
943 (if (VECTOR_TYPE_P (type)
944 || (INTEGRAL_TYPE_P (type) && TYPE_PRECISION (type) == 1))
945 /* Comparison inversion may be impossible for trapping math,
946 invert_tree_comparison will tell us. But we can't use
947 a computed operator in the replacement tree thus we have
948 to play the trick below. */
949 (with { enum tree_code ic = invert_tree_comparison
950 (cmp, HONOR_NANS (@0)); }
956 (bit_xor (cmp @0 @1) integer_truep)
957 (with { enum tree_code ic = invert_tree_comparison
958 (cmp, HONOR_NANS (@0)); }
964 /* Unordered tests if either argument is a NaN. */
966 (bit_ior (unordered @0 @0) (unordered @1 @1))
967 (if (types_match (@0, @1))
970 (bit_ior:c (unordered @0 @0) (unordered:c@2 @0 @1))
973 /* Simplification of math builtins. */
975 (define_operator_list LOG BUILT_IN_LOGF BUILT_IN_LOG BUILT_IN_LOGL)
976 (define_operator_list EXP BUILT_IN_EXPF BUILT_IN_EXP BUILT_IN_EXPL)
977 (define_operator_list LOG2 BUILT_IN_LOG2F BUILT_IN_LOG2 BUILT_IN_LOG2L)
978 (define_operator_list EXP2 BUILT_IN_EXP2F BUILT_IN_EXP2 BUILT_IN_EXP2L)
979 (define_operator_list LOG10 BUILT_IN_LOG10F BUILT_IN_LOG10 BUILT_IN_LOG10L)
980 (define_operator_list EXP10 BUILT_IN_EXP10F BUILT_IN_EXP10 BUILT_IN_EXP10L)
981 (define_operator_list POW BUILT_IN_POWF BUILT_IN_POW BUILT_IN_POWL)
982 (define_operator_list POW10 BUILT_IN_POW10F BUILT_IN_POW10 BUILT_IN_POW10L)
983 (define_operator_list SQRT BUILT_IN_SQRTF BUILT_IN_SQRT BUILT_IN_SQRTL)
984 (define_operator_list CBRT BUILT_IN_CBRTF BUILT_IN_CBRT BUILT_IN_CBRTL)
987 /* fold_builtin_logarithm */
988 (if (flag_unsafe_math_optimizations)
989 /* Special case, optimize logN(expN(x)) = x. */
990 (for logs (LOG LOG2 LOG10)
991 exps (EXP EXP2 EXP10)
995 /* Optimize logN(func()) for various exponential functions. We
996 want to determine the value "x" and the power "exponent" in
997 order to transform logN(x**exponent) into exponent*logN(x). */
998 (for logs (LOG LOG LOG LOG
1000 LOG10 LOG10 LOG10 LOG10)
1001 exps (EXP EXP2 EXP10 POW10)
1008 CASE_FLT_FN (BUILT_IN_EXP):
1009 /* Prepare to do logN(exp(exponent) -> exponent*logN(e). */
1010 x = build_real (type, real_value_truncate (TYPE_MODE (type),
1013 CASE_FLT_FN (BUILT_IN_EXP2):
1014 /* Prepare to do logN(exp2(exponent) -> exponent*logN(2). */
1015 x = build_real (type, dconst2);
1017 CASE_FLT_FN (BUILT_IN_EXP10):
1018 CASE_FLT_FN (BUILT_IN_POW10):
1019 /* Prepare to do logN(exp10(exponent) -> exponent*logN(10). */
1021 REAL_VALUE_TYPE dconst10;
1022 real_from_integer (&dconst10, VOIDmode, 10, SIGNED);
1023 x = build_real (type, dconst10);
1028 (mult (logs { x; }) @0))))
1039 CASE_FLT_FN (BUILT_IN_SQRT):
1040 /* Prepare to do logN(sqrt(x) -> 0.5*logN(x). */
1041 x = build_real (type, dconsthalf);
1043 CASE_FLT_FN (BUILT_IN_CBRT):
1044 /* Prepare to do logN(cbrt(x) -> (1/3)*logN(x). */
1045 x = build_real (type, real_value_truncate (TYPE_MODE (type),
1050 (mult { x; } (logs @0)))))
1051 /* logN(pow(x,exponent) -> exponent*logN(x). */
1052 (for logs (LOG LOG2 LOG10)
1056 (mult @1 (logs @0)))))
1058 /* Narrowing of arithmetic and logical operations.
1060 These are conceptually similar to the transformations performed for
1061 the C/C++ front-ends by shorten_binary_op and shorten_compare. Long
1062 term we want to move all that code out of the front-ends into here. */
1064 /* If we have a narrowing conversion of an arithmetic operation where
1065 both operands are widening conversions from the same type as the outer
1066 narrowing conversion. Then convert the innermost operands to a suitable
1067 unsigned type (to avoid introducing undefined behaviour), perform the
1068 operation and convert the result to the desired type. */
1069 (for op (plus minus)
1071 (convert (op@4 (convert@2 @0) (convert@3 @1)))
1072 (if (INTEGRAL_TYPE_P (type)
1073 /* We check for type compatibility between @0 and @1 below,
1074 so there's no need to check that @1/@3 are integral types. */
1075 && INTEGRAL_TYPE_P (TREE_TYPE (@0))
1076 && INTEGRAL_TYPE_P (TREE_TYPE (@2))
1077 /* The precision of the type of each operand must match the
1078 precision of the mode of each operand, similarly for the
1080 && (TYPE_PRECISION (TREE_TYPE (@0))
1081 == GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (@0))))
1082 && (TYPE_PRECISION (TREE_TYPE (@1))
1083 == GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (@1))))
1084 && TYPE_PRECISION (type) == GET_MODE_PRECISION (TYPE_MODE (type))
1085 /* The inner conversion must be a widening conversion. */
1086 && TYPE_PRECISION (TREE_TYPE (@2)) > TYPE_PRECISION (TREE_TYPE (@0))
1087 && types_match (@0, @1)
1088 && types_match (@0, type)
1090 (if (TYPE_OVERFLOW_WRAPS (TREE_TYPE (@0)))
1091 (convert (op @0 @1)))
1092 (with { tree utype = unsigned_type_for (TREE_TYPE (@0)); }
1093 (convert (op (convert:utype @0) (convert:utype @1)))))))
1095 /* This is another case of narrowing, specifically when there's an outer
1096 BIT_AND_EXPR which masks off bits outside the type of the innermost
1097 operands. Like the previous case we have to convert the operands
1098 to unsigned types to avoid introducing undefined behaviour for the
1099 arithmetic operation. */
1100 (for op (minus plus)
1102 (bit_and (op@5 (convert@2 @0) (convert@3 @1)) INTEGER_CST@4)
1103 (if (INTEGRAL_TYPE_P (type)
1104 /* We check for type compatibility between @0 and @1 below,
1105 so there's no need to check that @1/@3 are integral types. */
1106 && INTEGRAL_TYPE_P (TREE_TYPE (@0))
1107 && INTEGRAL_TYPE_P (TREE_TYPE (@2))
1108 /* The precision of the type of each operand must match the
1109 precision of the mode of each operand, similarly for the
1111 && (TYPE_PRECISION (TREE_TYPE (@0))
1112 == GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (@0))))
1113 && (TYPE_PRECISION (TREE_TYPE (@1))
1114 == GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (@1))))
1115 && TYPE_PRECISION (type) == GET_MODE_PRECISION (TYPE_MODE (type))
1116 /* The inner conversion must be a widening conversion. */
1117 && TYPE_PRECISION (TREE_TYPE (@2)) > TYPE_PRECISION (TREE_TYPE (@0))
1118 && types_match (@0, @1)
1119 && (tree_int_cst_min_precision (@4, TYPE_SIGN (TREE_TYPE (@0)))
1120 <= TYPE_PRECISION (TREE_TYPE (@0)))
1121 && (TYPE_OVERFLOW_WRAPS (TREE_TYPE (@0))
1122 || tree_int_cst_sgn (@4) >= 0)
1124 (if (TYPE_OVERFLOW_WRAPS (TREE_TYPE (@0)))
1125 (with { tree ntype = TREE_TYPE (@0); }
1126 (convert (bit_and (op @0 @1) (convert:ntype @4)))))
1127 (with { tree utype = unsigned_type_for (TREE_TYPE (@0)); }
1128 (convert (bit_and (op (convert:utype @0) (convert:utype @1))
1129 (convert:utype @4)))))))