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 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); })))
216 /* X % -C is the same as X % C. */
218 (trunc_mod @0 INTEGER_CST@1)
219 (if (TYPE_SIGN (type) == SIGNED
220 && !TREE_OVERFLOW (@1)
222 && !TYPE_OVERFLOW_TRAPS (type)
223 /* Avoid this transformation if C is INT_MIN, i.e. C == -C. */
224 && !sign_bit_p (@1, @1))
225 (trunc_mod @0 (negate @1))))
229 (bit_ior @0 integer_all_onesp@1)
234 (bit_and @0 integer_zerop@1)
240 { build_zero_cst (type); })
242 /* Canonicalize X ^ ~0 to ~X. */
244 (bit_xor @0 integer_all_onesp@1)
249 (bit_and @0 integer_all_onesp)
252 /* x & x -> x, x | x -> x */
253 (for bitop (bit_and bit_ior)
262 (abs tree_expr_nonnegative_p@0)
266 /* Try to fold (type) X op CST -> (type) (X op ((type-x) CST))
268 For bitwise binary operations apply operand conversions to the
269 binary operation result instead of to the operands. This allows
270 to combine successive conversions and bitwise binary operations.
271 We combine the above two cases by using a conditional convert. */
272 (for bitop (bit_and bit_ior bit_xor)
274 (bitop (convert @0) (convert? @1))
275 (if (((TREE_CODE (@1) == INTEGER_CST
276 && INTEGRAL_TYPE_P (TREE_TYPE (@0))
277 && int_fits_type_p (@1, TREE_TYPE (@0)))
278 || (GIMPLE && types_compatible_p (TREE_TYPE (@0), TREE_TYPE (@1)))
279 || (GENERIC && TREE_TYPE (@0) == TREE_TYPE (@1)))
280 /* ??? This transform conflicts with fold-const.c doing
281 Convert (T)(x & c) into (T)x & (T)c, if c is an integer
282 constants (if x has signed type, the sign bit cannot be set
283 in c). This folds extension into the BIT_AND_EXPR.
284 Restrict it to GIMPLE to avoid endless recursions. */
285 && (bitop != BIT_AND_EXPR || GIMPLE)
286 && (/* That's a good idea if the conversion widens the operand, thus
287 after hoisting the conversion the operation will be narrower. */
288 TYPE_PRECISION (TREE_TYPE (@0)) < TYPE_PRECISION (type)
289 /* It's also a good idea if the conversion is to a non-integer
291 || GET_MODE_CLASS (TYPE_MODE (type)) != MODE_INT
292 /* Or if the precision of TO is not the same as the precision
294 || TYPE_PRECISION (type) != GET_MODE_PRECISION (TYPE_MODE (type))))
295 (convert (bitop @0 (convert @1))))))
297 /* Simplify (A & B) OP0 (C & B) to (A OP0 C) & B. */
298 (for bitop (bit_and bit_ior bit_xor)
300 (bitop (bit_and:c @0 @1) (bit_and @2 @1))
301 (bit_and (bitop @0 @2) @1)))
303 /* (x | CST1) & CST2 -> (x & CST2) | (CST1 & CST2) */
305 (bit_and (bit_ior @0 CONSTANT_CLASS_P@1) CONSTANT_CLASS_P@2)
306 (bit_ior (bit_and @0 @2) (bit_and @1 @2)))
308 /* Combine successive equal operations with constants. */
309 (for bitop (bit_and bit_ior bit_xor)
311 (bitop (bitop @0 CONSTANT_CLASS_P@1) CONSTANT_CLASS_P@2)
312 (bitop @0 (bitop @1 @2))))
314 /* Try simple folding for X op !X, and X op X with the help
315 of the truth_valued_p and logical_inverted_value predicates. */
316 (match truth_valued_p
318 (if (INTEGRAL_TYPE_P (type) && TYPE_PRECISION (type) == 1)))
319 (for op (tcc_comparison truth_and truth_andif truth_or truth_orif truth_xor)
320 (match truth_valued_p
322 (match truth_valued_p
325 (match (logical_inverted_value @0)
326 (bit_not truth_valued_p@0))
327 (match (logical_inverted_value @0)
328 (eq @0 integer_zerop))
329 (match (logical_inverted_value @0)
330 (ne truth_valued_p@0 integer_truep))
331 (match (logical_inverted_value @0)
332 (bit_xor truth_valued_p@0 integer_truep))
336 (bit_and:c @0 (logical_inverted_value @0))
337 { build_zero_cst (type); })
338 /* X | !X and X ^ !X -> 1, , if X is truth-valued. */
339 (for op (bit_ior bit_xor)
341 (op:c truth_valued_p@0 (logical_inverted_value @0))
342 { constant_boolean_node (true, type); }))
344 (for bitop (bit_and bit_ior)
345 rbitop (bit_ior bit_and)
346 /* (x | y) & x -> x */
347 /* (x & y) | x -> x */
349 (bitop:c (rbitop:c @0 @1) @0)
351 /* (~x | y) & x -> x & y */
352 /* (~x & y) | x -> x | y */
354 (bitop:c (rbitop:c (bit_not @0) @1) @0)
357 /* If arg1 and arg2 are booleans (or any single bit type)
358 then try to simplify:
365 But only do this if our result feeds into a comparison as
366 this transformation is not always a win, particularly on
367 targets with and-not instructions.
368 -> simplify_bitwise_binary_boolean */
370 (ne (bit_and:c (bit_not @0) @1) integer_zerop)
371 (if (INTEGRAL_TYPE_P (TREE_TYPE (@1))
372 && TYPE_PRECISION (TREE_TYPE (@1)) == 1)
375 (ne (bit_ior:c (bit_not @0) @1) integer_zerop)
376 (if (INTEGRAL_TYPE_P (TREE_TYPE (@1))
377 && TYPE_PRECISION (TREE_TYPE (@1)) == 1)
382 (bit_not (bit_not @0))
385 /* (x & ~m) | (y & m) -> ((x ^ y) & m) ^ x */
387 (bit_ior:c (bit_and:c@3 @0 (bit_not @2)) (bit_and:c@4 @1 @2))
388 (if ((TREE_CODE (@3) != SSA_NAME || has_single_use (@3))
389 && (TREE_CODE (@4) != SSA_NAME || has_single_use (@4)))
390 (bit_xor (bit_and (bit_xor @0 @1) @2) @0)))
393 /* Associate (p +p off1) +p off2 as (p +p (off1 + off2)). */
395 (pointer_plus (pointer_plus@2 @0 @1) @3)
396 (if (TREE_CODE (@2) != SSA_NAME || has_single_use (@2))
397 (pointer_plus @0 (plus @1 @3))))
403 tem4 = (unsigned long) tem3;
408 (pointer_plus @0 (convert?@2 (minus@3 (convert @1) (convert @0))))
409 /* Conditionally look through a sign-changing conversion. */
410 (if (TYPE_PRECISION (TREE_TYPE (@2)) == TYPE_PRECISION (TREE_TYPE (@3))
411 && ((GIMPLE && useless_type_conversion_p (type, TREE_TYPE (@1)))
412 || (GENERIC && type == TREE_TYPE (@1))))
416 tem = (sizetype) ptr;
420 and produce the simpler and easier to analyze with respect to alignment
421 ... = ptr & ~algn; */
423 (pointer_plus @0 (negate (bit_and (convert @0) INTEGER_CST@1)))
424 (with { tree algn = wide_int_to_tree (TREE_TYPE (@0), wi::bit_not (@1)); }
425 (bit_and @0 { algn; })))
428 /* We can't reassociate at all for saturating types. */
429 (if (!TYPE_SATURATING (type))
431 /* Contract negates. */
432 /* A + (-B) -> A - B */
434 (plus:c (convert1? @0) (convert2? (negate @1)))
435 /* Apply STRIP_NOPS on @0 and the negate. */
436 (if (tree_nop_conversion_p (type, TREE_TYPE (@0))
437 && tree_nop_conversion_p (type, TREE_TYPE (@1))
438 && !TYPE_OVERFLOW_SANITIZED (type))
439 (minus (convert @0) (convert @1))))
440 /* A - (-B) -> A + B */
442 (minus (convert1? @0) (convert2? (negate @1)))
443 (if (tree_nop_conversion_p (type, TREE_TYPE (@0))
444 && tree_nop_conversion_p (type, TREE_TYPE (@1))
445 && !TYPE_OVERFLOW_SANITIZED (type))
446 (plus (convert @0) (convert @1))))
449 (negate (convert? (negate @1)))
450 (if (tree_nop_conversion_p (type, TREE_TYPE (@1))
451 && !TYPE_OVERFLOW_SANITIZED (type))
454 /* We can't reassociate floating-point or fixed-point plus or minus
455 because of saturation to +-Inf. */
456 (if (!FLOAT_TYPE_P (type) && !FIXED_POINT_TYPE_P (type))
458 /* Match patterns that allow contracting a plus-minus pair
459 irrespective of overflow issues. */
460 /* (A +- B) - A -> +- B */
461 /* (A +- B) -+ B -> A */
462 /* A - (A +- B) -> -+ B */
463 /* A +- (B -+ A) -> +- B */
465 (minus (plus:c @0 @1) @0)
468 (minus (minus @0 @1) @0)
471 (plus:c (minus @0 @1) @1)
474 (minus @0 (plus:c @0 @1))
477 (minus @0 (minus @0 @1))
480 /* (A +- CST) +- CST -> A + CST */
481 (for outer_op (plus minus)
482 (for inner_op (plus minus)
484 (outer_op (inner_op @0 CONSTANT_CLASS_P@1) CONSTANT_CLASS_P@2)
485 /* If the constant operation overflows we cannot do the transform
486 as we would introduce undefined overflow, for example
487 with (a - 1) + INT_MIN. */
488 (with { tree cst = fold_binary (outer_op == inner_op
489 ? PLUS_EXPR : MINUS_EXPR, type, @1, @2); }
490 (if (cst && !TREE_OVERFLOW (cst))
491 (inner_op @0 { cst; } ))))))
493 /* (CST - A) +- CST -> CST - A */
494 (for outer_op (plus minus)
496 (outer_op (minus CONSTANT_CLASS_P@1 @0) CONSTANT_CLASS_P@2)
497 (with { tree cst = fold_binary (outer_op, type, @1, @2); }
498 (if (cst && !TREE_OVERFLOW (cst))
499 (minus { cst; } @0)))))
503 (plus:c (bit_not @0) @0)
504 (if (!TYPE_OVERFLOW_TRAPS (type))
505 { build_all_ones_cst (type); }))
509 (plus (convert? (bit_not @0)) integer_each_onep)
510 (if (tree_nop_conversion_p (type, TREE_TYPE (@0)))
511 (negate (convert @0))))
515 (minus (convert? (negate @0)) integer_each_onep)
516 (if (!TYPE_OVERFLOW_TRAPS (type)
517 && tree_nop_conversion_p (type, TREE_TYPE (@0)))
518 (bit_not (convert @0))))
522 (minus integer_all_onesp @0)
523 (if (TREE_CODE (type) != COMPLEX_TYPE)
526 /* (T)(P + A) - (T)P -> (T) A */
527 (for add (plus pointer_plus)
529 (minus (convert (add @0 @1))
531 (if (element_precision (type) <= element_precision (TREE_TYPE (@1))
532 /* For integer types, if A has a smaller type
533 than T the result depends on the possible
535 E.g. T=size_t, A=(unsigned)429497295, P>0.
536 However, if an overflow in P + A would cause
537 undefined behavior, we can assume that there
539 || (INTEGRAL_TYPE_P (TREE_TYPE (@0))
540 && TYPE_OVERFLOW_UNDEFINED (TREE_TYPE (@0)))
541 /* For pointer types, if the conversion of A to the
542 final type requires a sign- or zero-extension,
543 then we have to punt - it is not defined which
545 || (POINTER_TYPE_P (TREE_TYPE (@0))
546 && TREE_CODE (@1) == INTEGER_CST
547 && tree_int_cst_sign_bit (@1) == 0))
551 /* Simplifications of MIN_EXPR and MAX_EXPR. */
553 (for minmax (min max)
559 (if (INTEGRAL_TYPE_P (type)
560 && TYPE_MIN_VALUE (type)
561 && operand_equal_p (@1, TYPE_MIN_VALUE (type), OEP_ONLY_CONST))
565 (if (INTEGRAL_TYPE_P (type)
566 && TYPE_MAX_VALUE (type)
567 && operand_equal_p (@1, TYPE_MAX_VALUE (type), OEP_ONLY_CONST))
571 /* Simplifications of shift and rotates. */
573 (for rotate (lrotate rrotate)
575 (rotate integer_all_onesp@0 @1)
578 /* Optimize -1 >> x for arithmetic right shifts. */
580 (rshift integer_all_onesp@0 @1)
581 (if (!TYPE_UNSIGNED (type)
582 && tree_expr_nonnegative_p (@1))
585 (for shiftrotate (lrotate rrotate lshift rshift)
587 (shiftrotate @0 integer_zerop)
590 (shiftrotate integer_zerop@0 @1)
592 /* Prefer vector1 << scalar to vector1 << vector2
593 if vector2 is uniform. */
594 (for vec (VECTOR_CST CONSTRUCTOR)
596 (shiftrotate @0 vec@1)
597 (with { tree tem = uniform_vector_p (@1); }
599 (shiftrotate @0 { tem; }))))))
601 /* Rewrite an LROTATE_EXPR by a constant into an
602 RROTATE_EXPR by a new constant. */
604 (lrotate @0 INTEGER_CST@1)
605 (rrotate @0 { fold_binary (MINUS_EXPR, TREE_TYPE (@1),
606 build_int_cst (TREE_TYPE (@1),
607 element_precision (type)), @1); }))
609 /* ((1 << A) & 1) != 0 -> A == 0
610 ((1 << A) & 1) == 0 -> A != 0 */
614 (cmp (bit_and (lshift integer_onep @0) integer_onep) integer_zerop)
615 (icmp @0 { build_zero_cst (TREE_TYPE (@0)); })))
617 /* Simplifications of conversions. */
619 /* Basic strip-useless-type-conversions / strip_nops. */
620 (for cvt (convert view_convert float fix_trunc)
623 (if ((GIMPLE && useless_type_conversion_p (type, TREE_TYPE (@0)))
624 || (GENERIC && type == TREE_TYPE (@0)))
627 /* Contract view-conversions. */
629 (view_convert (view_convert @0))
632 /* For integral conversions with the same precision or pointer
633 conversions use a NOP_EXPR instead. */
636 (if ((INTEGRAL_TYPE_P (type) || POINTER_TYPE_P (type))
637 && (INTEGRAL_TYPE_P (TREE_TYPE (@0)) || POINTER_TYPE_P (TREE_TYPE (@0)))
638 && TYPE_PRECISION (type) == TYPE_PRECISION (TREE_TYPE (@0)))
641 /* Strip inner integral conversions that do not change precision or size. */
643 (view_convert (convert@0 @1))
644 (if ((INTEGRAL_TYPE_P (TREE_TYPE (@0)) || POINTER_TYPE_P (TREE_TYPE (@0)))
645 && (INTEGRAL_TYPE_P (TREE_TYPE (@1)) || POINTER_TYPE_P (TREE_TYPE (@1)))
646 && (TYPE_PRECISION (TREE_TYPE (@0)) == TYPE_PRECISION (TREE_TYPE (@1)))
647 && (TYPE_SIZE (TREE_TYPE (@0)) == TYPE_SIZE (TREE_TYPE (@1))))
650 /* Re-association barriers around constants and other re-association
651 barriers can be removed. */
653 (paren CONSTANT_CLASS_P@0)
659 /* Handle cases of two conversions in a row. */
660 (for ocvt (convert float fix_trunc)
661 (for icvt (convert float)
666 tree inside_type = TREE_TYPE (@0);
667 tree inter_type = TREE_TYPE (@1);
668 int inside_int = INTEGRAL_TYPE_P (inside_type);
669 int inside_ptr = POINTER_TYPE_P (inside_type);
670 int inside_float = FLOAT_TYPE_P (inside_type);
671 int inside_vec = VECTOR_TYPE_P (inside_type);
672 unsigned int inside_prec = TYPE_PRECISION (inside_type);
673 int inside_unsignedp = TYPE_UNSIGNED (inside_type);
674 int inter_int = INTEGRAL_TYPE_P (inter_type);
675 int inter_ptr = POINTER_TYPE_P (inter_type);
676 int inter_float = FLOAT_TYPE_P (inter_type);
677 int inter_vec = VECTOR_TYPE_P (inter_type);
678 unsigned int inter_prec = TYPE_PRECISION (inter_type);
679 int inter_unsignedp = TYPE_UNSIGNED (inter_type);
680 int final_int = INTEGRAL_TYPE_P (type);
681 int final_ptr = POINTER_TYPE_P (type);
682 int final_float = FLOAT_TYPE_P (type);
683 int final_vec = VECTOR_TYPE_P (type);
684 unsigned int final_prec = TYPE_PRECISION (type);
685 int final_unsignedp = TYPE_UNSIGNED (type);
687 /* In addition to the cases of two conversions in a row
688 handled below, if we are converting something to its own
689 type via an object of identical or wider precision, neither
690 conversion is needed. */
691 (if (((GIMPLE && useless_type_conversion_p (type, inside_type))
693 && TYPE_MAIN_VARIANT (type) == TYPE_MAIN_VARIANT (inside_type)))
694 && (((inter_int || inter_ptr) && final_int)
695 || (inter_float && final_float))
696 && inter_prec >= final_prec)
699 /* Likewise, if the intermediate and initial types are either both
700 float or both integer, we don't need the middle conversion if the
701 former is wider than the latter and doesn't change the signedness
702 (for integers). Avoid this if the final type is a pointer since
703 then we sometimes need the middle conversion. Likewise if the
704 final type has a precision not equal to the size of its mode. */
705 (if (((inter_int && inside_int)
706 || (inter_float && inside_float)
707 || (inter_vec && inside_vec))
708 && inter_prec >= inside_prec
709 && (inter_float || inter_vec
710 || inter_unsignedp == inside_unsignedp)
711 && ! (final_prec != GET_MODE_PRECISION (element_mode (type))
712 && element_mode (type) == element_mode (inter_type))
714 && (! final_vec || inter_prec == inside_prec))
717 /* If we have a sign-extension of a zero-extended value, we can
718 replace that by a single zero-extension. Likewise if the
719 final conversion does not change precision we can drop the
720 intermediate conversion. */
721 (if (inside_int && inter_int && final_int
722 && ((inside_prec < inter_prec && inter_prec < final_prec
723 && inside_unsignedp && !inter_unsignedp)
724 || final_prec == inter_prec))
727 /* Two conversions in a row are not needed unless:
728 - some conversion is floating-point (overstrict for now), or
729 - some conversion is a vector (overstrict for now), or
730 - the intermediate type is narrower than both initial and
732 - the intermediate type and innermost type differ in signedness,
733 and the outermost type is wider than the intermediate, or
734 - the initial type is a pointer type and the precisions of the
735 intermediate and final types differ, or
736 - the final type is a pointer type and the precisions of the
737 initial and intermediate types differ. */
738 (if (! inside_float && ! inter_float && ! final_float
739 && ! inside_vec && ! inter_vec && ! final_vec
740 && (inter_prec >= inside_prec || inter_prec >= final_prec)
741 && ! (inside_int && inter_int
742 && inter_unsignedp != inside_unsignedp
743 && inter_prec < final_prec)
744 && ((inter_unsignedp && inter_prec > inside_prec)
745 == (final_unsignedp && final_prec > inter_prec))
746 && ! (inside_ptr && inter_prec != final_prec)
747 && ! (final_ptr && inside_prec != inter_prec)
748 && ! (final_prec != GET_MODE_PRECISION (TYPE_MODE (type))
749 && TYPE_MODE (type) == TYPE_MODE (inter_type)))
752 /* A truncation to an unsigned type (a zero-extension) should be
753 canonicalized as bitwise and of a mask. */
754 (if (final_int && inter_int && inside_int
755 && final_prec == inside_prec
756 && final_prec > inter_prec
758 (convert (bit_and @0 { wide_int_to_tree
760 wi::mask (inter_prec, false,
761 TYPE_PRECISION (inside_type))); })))
763 /* If we are converting an integer to a floating-point that can
764 represent it exactly and back to an integer, we can skip the
765 floating-point conversion. */
766 (if (inside_int && inter_float && final_int &&
767 (unsigned) significand_size (TYPE_MODE (inter_type))
768 >= inside_prec - !inside_unsignedp)
771 /* If we have a narrowing conversion to an integral type that is fed by a
772 BIT_AND_EXPR, we might be able to remove the BIT_AND_EXPR if it merely
773 masks off bits outside the final type (and nothing else). */
775 (convert (bit_and @0 INTEGER_CST@1))
776 (if (INTEGRAL_TYPE_P (type)
777 && INTEGRAL_TYPE_P (TREE_TYPE (@0))
778 && TYPE_PRECISION (type) <= TYPE_PRECISION (TREE_TYPE (@0))
779 && operand_equal_p (@1, build_low_bits_mask (TREE_TYPE (@1),
780 TYPE_PRECISION (type)), 0))
784 /* (X /[ex] A) * A -> X. */
786 (mult (convert? (exact_div @0 @1)) @1)
787 /* Look through a sign-changing conversion. */
788 (if (TYPE_PRECISION (TREE_TYPE (@0)) == TYPE_PRECISION (type))
791 /* Canonicalization of binary operations. */
793 /* Convert X + -C into X - C. */
796 (if (REAL_VALUE_NEGATIVE (TREE_REAL_CST (@1)))
797 (with { tree tem = fold_unary (NEGATE_EXPR, type, @1); }
798 (if (!TREE_OVERFLOW (tem) || !flag_trapping_math)
799 (minus @0 { tem; })))))
801 /* Convert x+x into x*2.0. */
804 (if (SCALAR_FLOAT_TYPE_P (type))
805 (mult @0 { build_real (type, dconst2); })))
808 (minus integer_zerop @1)
811 /* (ARG0 - ARG1) is the same as (-ARG1 + ARG0). So check whether
812 ARG0 is zero and X + ARG0 reduces to X, since that would mean
813 (-ARG1 + ARG0) reduces to -ARG1. */
815 (minus real_zerop@0 @1)
816 (if (fold_real_zero_addition_p (type, @0, 0))
819 /* Transform x * -1 into -x. */
821 (mult @0 integer_minus_onep)
824 /* COMPLEX_EXPR and REALPART/IMAGPART_EXPR cancellations. */
826 (complex (realpart @0) (imagpart @0))
829 (realpart (complex @0 @1))
832 (imagpart (complex @0 @1))
836 /* BSWAP simplifications, transforms checked by gcc.dg/builtin-bswap-8.c. */
837 (for bswap (BUILT_IN_BSWAP16 BUILT_IN_BSWAP32 BUILT_IN_BSWAP64)
842 (bswap (bit_not (bswap @0)))
844 (for bitop (bit_xor bit_ior bit_and)
846 (bswap (bitop:c (bswap @0) @1))
847 (bitop @0 (bswap @1)))))
850 /* Combine COND_EXPRs and VEC_COND_EXPRs. */
852 /* Simplify constant conditions.
853 Only optimize constant conditions when the selected branch
854 has the same type as the COND_EXPR. This avoids optimizing
855 away "c ? x : throw", where the throw has a void type.
856 Note that we cannot throw away the fold-const.c variant nor
857 this one as we depend on doing this transform before possibly
858 A ? B : B -> B triggers and the fold-const.c one can optimize
859 0 ? A : B to B even if A has side-effects. Something
860 genmatch cannot handle. */
862 (cond INTEGER_CST@0 @1 @2)
863 (if (integer_zerop (@0)
864 && (!VOID_TYPE_P (TREE_TYPE (@2))
865 || VOID_TYPE_P (type)))
867 (if (!integer_zerop (@0)
868 && (!VOID_TYPE_P (TREE_TYPE (@1))
869 || VOID_TYPE_P (type)))
872 (vec_cond VECTOR_CST@0 @1 @2)
873 (if (integer_all_onesp (@0))
875 (if (integer_zerop (@0))
878 (for cnd (cond vec_cond)
879 /* A ? B : (A ? X : C) -> A ? B : C. */
881 (cnd @0 (cnd @0 @1 @2) @3)
884 (cnd @0 @1 (cnd @0 @2 @3))
887 /* A ? B : B -> B. */
892 /* !A ? B : C -> A ? C : B. */
894 (cnd (logical_inverted_value truth_valued_p@0) @1 @2)
898 /* Simplifications of comparisons. */
900 /* We can simplify a logical negation of a comparison to the
901 inverted comparison. As we cannot compute an expression
902 operator using invert_tree_comparison we have to simulate
903 that with expression code iteration. */
904 (for cmp (tcc_comparison)
905 icmp (inverted_tcc_comparison)
906 ncmp (inverted_tcc_comparison_with_nans)
907 /* Ideally we'd like to combine the following two patterns
908 and handle some more cases by using
909 (logical_inverted_value (cmp @0 @1))
910 here but for that genmatch would need to "inline" that.
911 For now implement what forward_propagate_comparison did. */
913 (bit_not (cmp @0 @1))
914 (if (VECTOR_TYPE_P (type)
915 || (INTEGRAL_TYPE_P (type) && TYPE_PRECISION (type) == 1))
916 /* Comparison inversion may be impossible for trapping math,
917 invert_tree_comparison will tell us. But we can't use
918 a computed operator in the replacement tree thus we have
919 to play the trick below. */
920 (with { enum tree_code ic = invert_tree_comparison
921 (cmp, HONOR_NANS (@0)); }
927 (bit_xor (cmp @0 @1) integer_truep)
928 (with { enum tree_code ic = invert_tree_comparison
929 (cmp, HONOR_NANS (@0)); }
936 /* Simplification of math builtins. */
938 (define_operator_list LOG BUILT_IN_LOGF BUILT_IN_LOG BUILT_IN_LOGL)
939 (define_operator_list EXP BUILT_IN_EXPF BUILT_IN_EXP BUILT_IN_EXPL)
940 (define_operator_list LOG2 BUILT_IN_LOG2F BUILT_IN_LOG2 BUILT_IN_LOG2L)
941 (define_operator_list EXP2 BUILT_IN_EXP2F BUILT_IN_EXP2 BUILT_IN_EXP2L)
942 (define_operator_list LOG10 BUILT_IN_LOG10F BUILT_IN_LOG10 BUILT_IN_LOG10L)
943 (define_operator_list EXP10 BUILT_IN_EXP10F BUILT_IN_EXP10 BUILT_IN_EXP10L)
944 (define_operator_list POW BUILT_IN_POWF BUILT_IN_POW BUILT_IN_POWL)
945 (define_operator_list POW10 BUILT_IN_POW10F BUILT_IN_POW10 BUILT_IN_POW10L)
946 (define_operator_list SQRT BUILT_IN_SQRTF BUILT_IN_SQRT BUILT_IN_SQRTL)
947 (define_operator_list CBRT BUILT_IN_CBRTF BUILT_IN_CBRT BUILT_IN_CBRTL)
950 /* fold_builtin_logarithm */
951 (if (flag_unsafe_math_optimizations)
952 /* Special case, optimize logN(expN(x)) = x. */
953 (for logs (LOG LOG2 LOG10)
954 exps (EXP EXP2 EXP10)
958 /* Optimize logN(func()) for various exponential functions. We
959 want to determine the value "x" and the power "exponent" in
960 order to transform logN(x**exponent) into exponent*logN(x). */
961 (for logs (LOG LOG LOG LOG
963 LOG10 LOG10 LOG10 LOG10)
964 exps (EXP EXP2 EXP10 POW10)
971 CASE_FLT_FN (BUILT_IN_EXP):
972 /* Prepare to do logN(exp(exponent) -> exponent*logN(e). */
973 x = build_real (type, real_value_truncate (TYPE_MODE (type),
976 CASE_FLT_FN (BUILT_IN_EXP2):
977 /* Prepare to do logN(exp2(exponent) -> exponent*logN(2). */
978 x = build_real (type, dconst2);
980 CASE_FLT_FN (BUILT_IN_EXP10):
981 CASE_FLT_FN (BUILT_IN_POW10):
982 /* Prepare to do logN(exp10(exponent) -> exponent*logN(10). */
984 REAL_VALUE_TYPE dconst10;
985 real_from_integer (&dconst10, VOIDmode, 10, SIGNED);
986 x = build_real (type, dconst10);
991 (mult (logs { x; }) @0))))
1002 CASE_FLT_FN (BUILT_IN_SQRT):
1003 /* Prepare to do logN(sqrt(x) -> 0.5*logN(x). */
1004 x = build_real (type, dconsthalf);
1006 CASE_FLT_FN (BUILT_IN_CBRT):
1007 /* Prepare to do logN(cbrt(x) -> (1/3)*logN(x). */
1008 x = build_real (type, real_value_truncate (TYPE_MODE (type),
1013 (mult { x; } (logs @0)))))
1014 /* logN(pow(x,exponent) -> exponent*logN(x). */
1015 (for logs (LOG LOG2 LOG10)
1019 (mult @1 (logs @0)))))