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 (element_mode (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 (element_mode (type))
90 && !HONOR_SIGNED_ZEROS (element_mode (type)))
93 /* In IEEE floating point, x*1 is not equivalent to x for snans.
94 Likewise for complex arithmetic with signed zeros. */
97 (if (!HONOR_SNANS (element_mode (type))
98 && (!HONOR_SIGNED_ZEROS (element_mode (type))
99 || !COMPLEX_FLOAT_TYPE_P (type)))
102 /* Transform x * -1.0 into -x. */
104 (mult @0 real_minus_onep)
105 (if (!HONOR_SNANS (element_mode (type))
106 && (!HONOR_SIGNED_ZEROS (element_mode (type))
107 || !COMPLEX_FLOAT_TYPE_P (type)))
110 /* Make sure to preserve divisions by zero. This is the reason why
111 we don't simplify x / x to 1 or 0 / x to 0. */
112 (for op (mult trunc_div ceil_div floor_div round_div exact_div)
118 (for div (trunc_div ceil_div floor_div round_div exact_div)
120 (div @0 integer_minus_onep@1)
121 (if (!TYPE_UNSIGNED (type))
124 /* For unsigned integral types, FLOOR_DIV_EXPR is the same as
125 TRUNC_DIV_EXPR. Rewrite into the latter in this case. */
128 (if ((INTEGRAL_TYPE_P (type) || VECTOR_INTEGER_TYPE_P (type))
129 && TYPE_UNSIGNED (type))
132 /* Combine two successive divisions. */
133 (for div (trunc_div ceil_div floor_div round_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 { build_zero_cst (type); }))))
145 /* Optimize A / A to 1.0 if we don't care about
146 NaNs or Infinities. */
149 (if (FLOAT_TYPE_P (type)
150 && ! HONOR_NANS (element_mode (type))
151 && ! HONOR_INFINITIES (element_mode (type)))
152 { build_one_cst (type); }))
154 /* Optimize -A / A to -1.0 if we don't care about
155 NaNs or Infinities. */
157 (rdiv:c @0 (negate @0))
158 (if (FLOAT_TYPE_P (type)
159 && ! HONOR_NANS (element_mode (type))
160 && ! HONOR_INFINITIES (element_mode (type)))
161 { build_minus_one_cst (type); }))
163 /* In IEEE floating point, x/1 is not equivalent to x for snans. */
166 (if (!HONOR_SNANS (element_mode (type)))
169 /* In IEEE floating point, x/-1 is not equivalent to -x for snans. */
171 (rdiv @0 real_minus_onep)
172 (if (!HONOR_SNANS (element_mode (type)))
175 /* If ARG1 is a constant, we can convert this to a multiply by the
176 reciprocal. This does not have the same rounding properties,
177 so only do this if -freciprocal-math. We can actually
178 always safely do it if ARG1 is a power of two, but it's hard to
179 tell if it is or not in a portable manner. */
180 (for cst (REAL_CST COMPLEX_CST VECTOR_CST)
184 (if (flag_reciprocal_math
187 { tree tem = fold_binary (RDIV_EXPR, type, build_one_cst (type), @1); }
189 (mult @0 { tem; } ))))
190 (if (cst != COMPLEX_CST)
191 (with { tree inverse = exact_inverse (type, @1); }
193 (mult @0 { inverse; } )))))))
195 /* Same applies to modulo operations, but fold is inconsistent here
196 and simplifies 0 % x to 0, only preserving literal 0 % 0. */
197 (for mod (ceil_mod floor_mod round_mod trunc_mod)
198 /* 0 % X is always zero. */
200 (mod integer_zerop@0 @1)
201 /* But not for 0 % 0 so that we can get the proper warnings and errors. */
202 (if (!integer_zerop (@1))
204 /* X % 1 is always zero. */
206 (mod @0 integer_onep)
207 { build_zero_cst (type); })
208 /* X % -1 is zero. */
210 (mod @0 integer_minus_onep@1)
211 (if (!TYPE_UNSIGNED (type))
212 { build_zero_cst (type); })))
214 /* X % -C is the same as X % C. */
216 (trunc_mod @0 INTEGER_CST@1)
217 (if (TYPE_SIGN (type) == SIGNED
218 && !TREE_OVERFLOW (@1)
220 && !TYPE_OVERFLOW_TRAPS (type)
221 /* Avoid this transformation if C is INT_MIN, i.e. C == -C. */
222 && !sign_bit_p (@1, @1))
223 (trunc_mod @0 (negate @1))))
227 (bit_ior @0 integer_all_onesp@1)
232 (bit_and @0 integer_zerop@1)
238 { build_zero_cst (type); })
240 /* Canonicalize X ^ ~0 to ~X. */
242 (bit_xor @0 integer_all_onesp@1)
247 (bit_and @0 integer_all_onesp)
250 /* x & x -> x, x | x -> x */
251 (for bitop (bit_and bit_ior)
260 (abs tree_expr_nonnegative_p@0)
264 /* Try to fold (type) X op CST -> (type) (X op ((type-x) CST))
266 For bitwise binary operations apply operand conversions to the
267 binary operation result instead of to the operands. This allows
268 to combine successive conversions and bitwise binary operations.
269 We combine the above two cases by using a conditional convert. */
270 (for bitop (bit_and bit_ior bit_xor)
272 (bitop (convert @0) (convert? @1))
273 (if (((TREE_CODE (@1) == INTEGER_CST
274 && INTEGRAL_TYPE_P (TREE_TYPE (@0))
275 && int_fits_type_p (@1, TREE_TYPE (@0)))
276 || (GIMPLE && types_compatible_p (TREE_TYPE (@0), TREE_TYPE (@1)))
277 || (GENERIC && TREE_TYPE (@0) == TREE_TYPE (@1)))
278 /* ??? This transform conflicts with fold-const.c doing
279 Convert (T)(x & c) into (T)x & (T)c, if c is an integer
280 constants (if x has signed type, the sign bit cannot be set
281 in c). This folds extension into the BIT_AND_EXPR.
282 Restrict it to GIMPLE to avoid endless recursions. */
283 && (bitop != BIT_AND_EXPR || GIMPLE)
284 && (/* That's a good idea if the conversion widens the operand, thus
285 after hoisting the conversion the operation will be narrower. */
286 TYPE_PRECISION (TREE_TYPE (@0)) < TYPE_PRECISION (type)
287 /* It's also a good idea if the conversion is to a non-integer
289 || GET_MODE_CLASS (TYPE_MODE (type)) != MODE_INT
290 /* Or if the precision of TO is not the same as the precision
292 || TYPE_PRECISION (type) != GET_MODE_PRECISION (TYPE_MODE (type))))
293 (convert (bitop @0 (convert @1))))))
295 /* Simplify (A & B) OP0 (C & B) to (A OP0 C) & B. */
296 (for bitop (bit_and bit_ior bit_xor)
298 (bitop (bit_and:c @0 @1) (bit_and @2 @1))
299 (bit_and (bitop @0 @2) @1)))
301 /* (x | CST1) & CST2 -> (x & CST2) | (CST1 & CST2) */
303 (bit_and (bit_ior @0 CONSTANT_CLASS_P@1) CONSTANT_CLASS_P@2)
304 (bit_ior (bit_and @0 @2) (bit_and @1 @2)))
306 /* Combine successive equal operations with constants. */
307 (for bitop (bit_and bit_ior bit_xor)
309 (bitop (bitop @0 CONSTANT_CLASS_P@1) CONSTANT_CLASS_P@2)
310 (bitop @0 (bitop @1 @2))))
312 /* Try simple folding for X op !X, and X op X with the help
313 of the truth_valued_p and logical_inverted_value predicates. */
314 (match truth_valued_p
316 (if (INTEGRAL_TYPE_P (type) && TYPE_PRECISION (type) == 1)))
317 (for op (tcc_comparison truth_and truth_andif truth_or truth_orif truth_xor)
318 (match truth_valued_p
320 (match truth_valued_p
323 (match (logical_inverted_value @0)
324 (bit_not truth_valued_p@0))
325 (match (logical_inverted_value @0)
326 (eq @0 integer_zerop))
327 (match (logical_inverted_value @0)
328 (ne truth_valued_p@0 integer_truep))
329 (match (logical_inverted_value @0)
330 (bit_xor truth_valued_p@0 integer_truep))
334 (bit_and:c @0 (logical_inverted_value @0))
335 { build_zero_cst (type); })
336 /* X | !X and X ^ !X -> 1, , if X is truth-valued. */
337 (for op (bit_ior bit_xor)
339 (op:c truth_valued_p@0 (logical_inverted_value @0))
340 { constant_boolean_node (true, type); }))
342 (for bitop (bit_and bit_ior)
343 rbitop (bit_ior bit_and)
344 /* (x | y) & x -> x */
345 /* (x & y) | x -> x */
347 (bitop:c (rbitop:c @0 @1) @0)
349 /* (~x | y) & x -> x & y */
350 /* (~x & y) | x -> x | y */
352 (bitop:c (rbitop:c (bit_not @0) @1) @0)
355 /* If arg1 and arg2 are booleans (or any single bit type)
356 then try to simplify:
363 But only do this if our result feeds into a comparison as
364 this transformation is not always a win, particularly on
365 targets with and-not instructions.
366 -> simplify_bitwise_binary_boolean */
368 (ne (bit_and:c (bit_not @0) @1) integer_zerop)
369 (if (INTEGRAL_TYPE_P (TREE_TYPE (@1))
370 && TYPE_PRECISION (TREE_TYPE (@1)) == 1)
373 (ne (bit_ior:c (bit_not @0) @1) integer_zerop)
374 (if (INTEGRAL_TYPE_P (TREE_TYPE (@1))
375 && TYPE_PRECISION (TREE_TYPE (@1)) == 1)
380 (bit_not (bit_not @0))
384 /* Associate (p +p off1) +p off2 as (p +p (off1 + off2)). */
386 (pointer_plus (pointer_plus@2 @0 @1) @3)
387 (if (TREE_CODE (@2) != SSA_NAME || has_single_use (@2))
388 (pointer_plus @0 (plus @1 @3))))
394 tem4 = (unsigned long) tem3;
399 (pointer_plus @0 (convert?@2 (minus@3 (convert @1) (convert @0))))
400 /* Conditionally look through a sign-changing conversion. */
401 (if (TYPE_PRECISION (TREE_TYPE (@2)) == TYPE_PRECISION (TREE_TYPE (@3))
402 && ((GIMPLE && useless_type_conversion_p (type, TREE_TYPE (@1)))
403 || (GENERIC && type == TREE_TYPE (@1))))
407 tem = (sizetype) ptr;
411 and produce the simpler and easier to analyze with respect to alignment
412 ... = ptr & ~algn; */
414 (pointer_plus @0 (negate (bit_and (convert @0) INTEGER_CST@1)))
415 (with { tree algn = wide_int_to_tree (TREE_TYPE (@0), wi::bit_not (@1)); }
416 (bit_and @0 { algn; })))
419 /* We can't reassociate at all for saturating types. */
420 (if (!TYPE_SATURATING (type))
422 /* Contract negates. */
423 /* A + (-B) -> A - B */
425 (plus:c (convert1? @0) (convert2? (negate @1)))
426 /* Apply STRIP_NOPS on @0 and the negate. */
427 (if (tree_nop_conversion_p (type, TREE_TYPE (@0))
428 && tree_nop_conversion_p (type, TREE_TYPE (@1))
429 && !TYPE_OVERFLOW_SANITIZED (type))
430 (minus (convert @0) (convert @1))))
431 /* A - (-B) -> A + B */
433 (minus (convert1? @0) (convert2? (negate @1)))
434 (if (tree_nop_conversion_p (type, TREE_TYPE (@0))
435 && tree_nop_conversion_p (type, TREE_TYPE (@1))
436 && !TYPE_OVERFLOW_SANITIZED (type))
437 (plus (convert @0) (convert @1))))
440 (negate (convert? (negate @1)))
441 (if (tree_nop_conversion_p (type, TREE_TYPE (@1))
442 && !TYPE_OVERFLOW_SANITIZED (type))
445 /* We can't reassociate floating-point or fixed-point plus or minus
446 because of saturation to +-Inf. */
447 (if (!FLOAT_TYPE_P (type) && !FIXED_POINT_TYPE_P (type))
449 /* Match patterns that allow contracting a plus-minus pair
450 irrespective of overflow issues. */
451 /* (A +- B) - A -> +- B */
452 /* (A +- B) -+ B -> A */
453 /* A - (A +- B) -> -+ B */
454 /* A +- (B -+ A) -> +- B */
456 (minus (plus:c @0 @1) @0)
459 (minus (minus @0 @1) @0)
462 (plus:c (minus @0 @1) @1)
465 (minus @0 (plus:c @0 @1))
468 (minus @0 (minus @0 @1))
471 /* (A +- CST) +- CST -> A + CST */
472 (for outer_op (plus minus)
473 (for inner_op (plus minus)
475 (outer_op (inner_op @0 CONSTANT_CLASS_P@1) CONSTANT_CLASS_P@2)
476 /* If the constant operation overflows we cannot do the transform
477 as we would introduce undefined overflow, for example
478 with (a - 1) + INT_MIN. */
479 (with { tree cst = fold_binary (outer_op == inner_op
480 ? PLUS_EXPR : MINUS_EXPR, type, @1, @2); }
481 (if (cst && !TREE_OVERFLOW (cst))
482 (inner_op @0 { cst; } ))))))
484 /* (CST - A) +- CST -> CST - A */
485 (for outer_op (plus minus)
487 (outer_op (minus CONSTANT_CLASS_P@1 @0) CONSTANT_CLASS_P@2)
488 (with { tree cst = fold_binary (outer_op, type, @1, @2); }
489 (if (cst && !TREE_OVERFLOW (cst))
490 (minus { cst; } @0)))))
494 (plus:c (bit_not @0) @0)
495 (if (!TYPE_OVERFLOW_TRAPS (type))
496 { build_all_ones_cst (type); }))
500 (plus (convert? (bit_not @0)) integer_each_onep)
501 (if (tree_nop_conversion_p (type, TREE_TYPE (@0)))
502 (negate (convert @0))))
506 (minus (convert? (negate @0)) integer_each_onep)
507 (if (!TYPE_OVERFLOW_TRAPS (type)
508 && tree_nop_conversion_p (type, TREE_TYPE (@0)))
509 (bit_not (convert @0))))
513 (minus integer_all_onesp @0)
514 (if (TREE_CODE (type) != COMPLEX_TYPE)
517 /* (T)(P + A) - (T)P -> (T) A */
518 (for add (plus pointer_plus)
520 (minus (convert (add @0 @1))
522 (if (element_precision (type) <= element_precision (TREE_TYPE (@1))
523 /* For integer types, if A has a smaller type
524 than T the result depends on the possible
526 E.g. T=size_t, A=(unsigned)429497295, P>0.
527 However, if an overflow in P + A would cause
528 undefined behavior, we can assume that there
530 || (INTEGRAL_TYPE_P (TREE_TYPE (@0))
531 && TYPE_OVERFLOW_UNDEFINED (TREE_TYPE (@0)))
532 /* For pointer types, if the conversion of A to the
533 final type requires a sign- or zero-extension,
534 then we have to punt - it is not defined which
536 || (POINTER_TYPE_P (TREE_TYPE (@0))
537 && TREE_CODE (@1) == INTEGER_CST
538 && tree_int_cst_sign_bit (@1) == 0))
542 /* Simplifications of MIN_EXPR and MAX_EXPR. */
544 (for minmax (min max)
550 (if (INTEGRAL_TYPE_P (type)
551 && TYPE_MIN_VALUE (type)
552 && operand_equal_p (@1, TYPE_MIN_VALUE (type), OEP_ONLY_CONST))
556 (if (INTEGRAL_TYPE_P (type)
557 && TYPE_MAX_VALUE (type)
558 && operand_equal_p (@1, TYPE_MAX_VALUE (type), OEP_ONLY_CONST))
562 /* Simplifications of shift and rotates. */
564 (for rotate (lrotate rrotate)
566 (rotate integer_all_onesp@0 @1)
569 /* Optimize -1 >> x for arithmetic right shifts. */
571 (rshift integer_all_onesp@0 @1)
572 (if (!TYPE_UNSIGNED (type)
573 && tree_expr_nonnegative_p (@1))
576 (for shiftrotate (lrotate rrotate lshift rshift)
578 (shiftrotate @0 integer_zerop)
581 (shiftrotate integer_zerop@0 @1)
583 /* Prefer vector1 << scalar to vector1 << vector2
584 if vector2 is uniform. */
585 (for vec (VECTOR_CST CONSTRUCTOR)
587 (shiftrotate @0 vec@1)
588 (with { tree tem = uniform_vector_p (@1); }
590 (shiftrotate @0 { tem; }))))))
592 /* Rewrite an LROTATE_EXPR by a constant into an
593 RROTATE_EXPR by a new constant. */
595 (lrotate @0 INTEGER_CST@1)
596 (rrotate @0 { fold_binary (MINUS_EXPR, TREE_TYPE (@1),
597 build_int_cst (TREE_TYPE (@1),
598 element_precision (type)), @1); }))
601 /* Simplifications of conversions. */
603 /* Basic strip-useless-type-conversions / strip_nops. */
604 (for cvt (convert view_convert float fix_trunc)
607 (if ((GIMPLE && useless_type_conversion_p (type, TREE_TYPE (@0)))
608 || (GENERIC && type == TREE_TYPE (@0)))
611 /* Contract view-conversions. */
613 (view_convert (view_convert @0))
616 /* For integral conversions with the same precision or pointer
617 conversions use a NOP_EXPR instead. */
620 (if ((INTEGRAL_TYPE_P (type) || POINTER_TYPE_P (type))
621 && (INTEGRAL_TYPE_P (TREE_TYPE (@0)) || POINTER_TYPE_P (TREE_TYPE (@0)))
622 && TYPE_PRECISION (type) == TYPE_PRECISION (TREE_TYPE (@0)))
625 /* Strip inner integral conversions that do not change precision or size. */
627 (view_convert (convert@0 @1))
628 (if ((INTEGRAL_TYPE_P (TREE_TYPE (@0)) || POINTER_TYPE_P (TREE_TYPE (@0)))
629 && (INTEGRAL_TYPE_P (TREE_TYPE (@1)) || POINTER_TYPE_P (TREE_TYPE (@1)))
630 && (TYPE_PRECISION (TREE_TYPE (@0)) == TYPE_PRECISION (TREE_TYPE (@1)))
631 && (TYPE_SIZE (TREE_TYPE (@0)) == TYPE_SIZE (TREE_TYPE (@1))))
634 /* Re-association barriers around constants and other re-association
635 barriers can be removed. */
637 (paren CONSTANT_CLASS_P@0)
643 /* Handle cases of two conversions in a row. */
644 (for ocvt (convert float fix_trunc)
645 (for icvt (convert float)
650 tree inside_type = TREE_TYPE (@0);
651 tree inter_type = TREE_TYPE (@1);
652 int inside_int = INTEGRAL_TYPE_P (inside_type);
653 int inside_ptr = POINTER_TYPE_P (inside_type);
654 int inside_float = FLOAT_TYPE_P (inside_type);
655 int inside_vec = VECTOR_TYPE_P (inside_type);
656 unsigned int inside_prec = TYPE_PRECISION (inside_type);
657 int inside_unsignedp = TYPE_UNSIGNED (inside_type);
658 int inter_int = INTEGRAL_TYPE_P (inter_type);
659 int inter_ptr = POINTER_TYPE_P (inter_type);
660 int inter_float = FLOAT_TYPE_P (inter_type);
661 int inter_vec = VECTOR_TYPE_P (inter_type);
662 unsigned int inter_prec = TYPE_PRECISION (inter_type);
663 int inter_unsignedp = TYPE_UNSIGNED (inter_type);
664 int final_int = INTEGRAL_TYPE_P (type);
665 int final_ptr = POINTER_TYPE_P (type);
666 int final_float = FLOAT_TYPE_P (type);
667 int final_vec = VECTOR_TYPE_P (type);
668 unsigned int final_prec = TYPE_PRECISION (type);
669 int final_unsignedp = TYPE_UNSIGNED (type);
671 /* In addition to the cases of two conversions in a row
672 handled below, if we are converting something to its own
673 type via an object of identical or wider precision, neither
674 conversion is needed. */
675 (if (((GIMPLE && useless_type_conversion_p (type, inside_type))
677 && TYPE_MAIN_VARIANT (type) == TYPE_MAIN_VARIANT (inside_type)))
678 && (((inter_int || inter_ptr) && final_int)
679 || (inter_float && final_float))
680 && inter_prec >= final_prec)
683 /* Likewise, if the intermediate and initial types are either both
684 float or both integer, we don't need the middle conversion if the
685 former is wider than the latter and doesn't change the signedness
686 (for integers). Avoid this if the final type is a pointer since
687 then we sometimes need the middle conversion. Likewise if the
688 final type has a precision not equal to the size of its mode. */
689 (if (((inter_int && inside_int)
690 || (inter_float && inside_float)
691 || (inter_vec && inside_vec))
692 && inter_prec >= inside_prec
693 && (inter_float || inter_vec
694 || inter_unsignedp == inside_unsignedp)
695 && ! (final_prec != GET_MODE_PRECISION (element_mode (type))
696 && element_mode (type) == element_mode (inter_type))
698 && (! final_vec || inter_prec == inside_prec))
701 /* If we have a sign-extension of a zero-extended value, we can
702 replace that by a single zero-extension. Likewise if the
703 final conversion does not change precision we can drop the
704 intermediate conversion. */
705 (if (inside_int && inter_int && final_int
706 && ((inside_prec < inter_prec && inter_prec < final_prec
707 && inside_unsignedp && !inter_unsignedp)
708 || final_prec == inter_prec))
711 /* Two conversions in a row are not needed unless:
712 - some conversion is floating-point (overstrict for now), or
713 - some conversion is a vector (overstrict for now), or
714 - the intermediate type is narrower than both initial and
716 - the intermediate type and innermost type differ in signedness,
717 and the outermost type is wider than the intermediate, or
718 - the initial type is a pointer type and the precisions of the
719 intermediate and final types differ, or
720 - the final type is a pointer type and the precisions of the
721 initial and intermediate types differ. */
722 (if (! inside_float && ! inter_float && ! final_float
723 && ! inside_vec && ! inter_vec && ! final_vec
724 && (inter_prec >= inside_prec || inter_prec >= final_prec)
725 && ! (inside_int && inter_int
726 && inter_unsignedp != inside_unsignedp
727 && inter_prec < final_prec)
728 && ((inter_unsignedp && inter_prec > inside_prec)
729 == (final_unsignedp && final_prec > inter_prec))
730 && ! (inside_ptr && inter_prec != final_prec)
731 && ! (final_ptr && inside_prec != inter_prec)
732 && ! (final_prec != GET_MODE_PRECISION (TYPE_MODE (type))
733 && TYPE_MODE (type) == TYPE_MODE (inter_type)))
736 /* A truncation to an unsigned type (a zero-extension) should be
737 canonicalized as bitwise and of a mask. */
738 (if (final_int && inter_int && inside_int
739 && final_prec == inside_prec
740 && final_prec > inter_prec
742 (convert (bit_and @0 { wide_int_to_tree
744 wi::mask (inter_prec, false,
745 TYPE_PRECISION (inside_type))); })))
747 /* If we are converting an integer to a floating-point that can
748 represent it exactly and back to an integer, we can skip the
749 floating-point conversion. */
750 (if (inside_int && inter_float && final_int &&
751 (unsigned) significand_size (TYPE_MODE (inter_type))
752 >= inside_prec - !inside_unsignedp)
755 /* If we have a narrowing conversion to an integral type that is fed by a
756 BIT_AND_EXPR, we might be able to remove the BIT_AND_EXPR if it merely
757 masks off bits outside the final type (and nothing else). */
759 (convert (bit_and @0 INTEGER_CST@1))
760 (if (INTEGRAL_TYPE_P (type)
761 && INTEGRAL_TYPE_P (TREE_TYPE (@0))
762 && TYPE_PRECISION (type) <= TYPE_PRECISION (TREE_TYPE (@0))
763 && operand_equal_p (@1, build_low_bits_mask (TREE_TYPE (@1),
764 TYPE_PRECISION (type)), 0))
768 /* (X /[ex] A) * A -> X. */
770 (mult (convert? (exact_div @0 @1)) @1)
771 /* Look through a sign-changing conversion. */
772 (if (TYPE_PRECISION (TREE_TYPE (@0)) == TYPE_PRECISION (type))
775 /* Canonicalization of binary operations. */
777 /* Convert X + -C into X - C. */
780 (if (REAL_VALUE_NEGATIVE (TREE_REAL_CST (@1)))
781 (with { tree tem = fold_unary (NEGATE_EXPR, type, @1); }
782 (if (!TREE_OVERFLOW (tem) || !flag_trapping_math)
783 (minus @0 { tem; })))))
785 /* Convert x+x into x*2.0. */
788 (if (SCALAR_FLOAT_TYPE_P (type))
789 (mult @0 { build_real (type, dconst2); })))
792 (minus integer_zerop @1)
795 /* (ARG0 - ARG1) is the same as (-ARG1 + ARG0). So check whether
796 ARG0 is zero and X + ARG0 reduces to X, since that would mean
797 (-ARG1 + ARG0) reduces to -ARG1. */
799 (minus real_zerop@0 @1)
800 (if (fold_real_zero_addition_p (type, @0, 0))
803 /* Transform x * -1 into -x. */
805 (mult @0 integer_minus_onep)
808 /* COMPLEX_EXPR and REALPART/IMAGPART_EXPR cancellations. */
810 (complex (realpart @0) (imagpart @0))
813 (realpart (complex @0 @1))
816 (imagpart (complex @0 @1))
820 /* BSWAP simplifications, transforms checked by gcc.dg/builtin-bswap-8.c. */
821 (for bswap (BUILT_IN_BSWAP16 BUILT_IN_BSWAP32 BUILT_IN_BSWAP64)
826 (bswap (bit_not (bswap @0)))
828 (for bitop (bit_xor bit_ior bit_and)
830 (bswap (bitop:c (bswap @0) @1))
831 (bitop @0 (bswap @1)))))
834 /* Combine COND_EXPRs and VEC_COND_EXPRs. */
836 /* Simplify constant conditions.
837 Only optimize constant conditions when the selected branch
838 has the same type as the COND_EXPR. This avoids optimizing
839 away "c ? x : throw", where the throw has a void type.
840 Note that we cannot throw away the fold-const.c variant nor
841 this one as we depend on doing this transform before possibly
842 A ? B : B -> B triggers and the fold-const.c one can optimize
843 0 ? A : B to B even if A has side-effects. Something
844 genmatch cannot handle. */
846 (cond INTEGER_CST@0 @1 @2)
847 (if (integer_zerop (@0)
848 && (!VOID_TYPE_P (TREE_TYPE (@2))
849 || VOID_TYPE_P (type)))
851 (if (!integer_zerop (@0)
852 && (!VOID_TYPE_P (TREE_TYPE (@1))
853 || VOID_TYPE_P (type)))
856 (vec_cond VECTOR_CST@0 @1 @2)
857 (if (integer_all_onesp (@0))
859 (if (integer_zerop (@0))
862 (for cnd (cond vec_cond)
863 /* A ? B : (A ? X : C) -> A ? B : C. */
865 (cnd @0 (cnd @0 @1 @2) @3)
868 (cnd @0 @1 (cnd @0 @2 @3))
871 /* A ? B : B -> B. */
876 /* !A ? B : C -> A ? C : B. */
878 (cnd (logical_inverted_value truth_valued_p@0) @1 @2)
882 /* Simplifications of comparisons. */
884 /* We can simplify a logical negation of a comparison to the
885 inverted comparison. As we cannot compute an expression
886 operator using invert_tree_comparison we have to simulate
887 that with expression code iteration. */
888 (for cmp (tcc_comparison)
889 icmp (inverted_tcc_comparison)
890 ncmp (inverted_tcc_comparison_with_nans)
891 /* Ideally we'd like to combine the following two patterns
892 and handle some more cases by using
893 (logical_inverted_value (cmp @0 @1))
894 here but for that genmatch would need to "inline" that.
895 For now implement what forward_propagate_comparison did. */
897 (bit_not (cmp @0 @1))
898 (if (VECTOR_TYPE_P (type)
899 || (INTEGRAL_TYPE_P (type) && TYPE_PRECISION (type) == 1))
900 /* Comparison inversion may be impossible for trapping math,
901 invert_tree_comparison will tell us. But we can't use
902 a computed operator in the replacement tree thus we have
903 to play the trick below. */
904 (with { enum tree_code ic = invert_tree_comparison
905 (cmp, HONOR_NANS (element_mode (@0))); }
911 (bit_xor (cmp @0 @1) integer_truep)
912 (with { enum tree_code ic = invert_tree_comparison
913 (cmp, HONOR_NANS (element_mode (@0))); }