]>
Commit | Line | Data |
---|---|---|
3d2cf79f RB |
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. | |
4 | ||
5624e564 | 5 | Copyright (C) 2014-2015 Free Software Foundation, Inc. |
3d2cf79f RB |
6 | Contributed by Richard Biener <rguenther@suse.de> |
7 | and Prathamesh Kulkarni <bilbotheelffriend@gmail.com> | |
8 | ||
9 | This file is part of GCC. | |
10 | ||
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 | |
14 | version. | |
15 | ||
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 | |
19 | for more details. | |
20 | ||
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/>. */ | |
24 | ||
25 | ||
26 | /* Generic tree predicates we inherit. */ | |
27 | (define_predicates | |
cc7b5acf | 28 | integer_onep integer_zerop integer_all_onesp integer_minus_onep |
09240451 | 29 | integer_each_onep integer_truep |
cc7b5acf | 30 | real_zerop real_onep real_minus_onep |
f3582e54 RB |
31 | CONSTANT_CLASS_P |
32 | tree_expr_nonnegative_p) | |
e0ee10ed | 33 | |
f84e7fd6 RB |
34 | /* Operator lists. */ |
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) | |
534bd33b MG |
41 | (define_operator_list swapped_tcc_comparison |
42 | gt ge eq ne le lt unordered ordered ungt unge unlt unle uneq ltgt) | |
f84e7fd6 | 43 | |
e0ee10ed RB |
44 | |
45 | /* Simplifications of operations with one constant operand and | |
36a60e48 | 46 | simplifications to constants or single values. */ |
e0ee10ed RB |
47 | |
48 | (for op (plus pointer_plus minus bit_ior bit_xor) | |
49 | (simplify | |
50 | (op @0 integer_zerop) | |
51 | (non_lvalue @0))) | |
52 | ||
a499aac5 RB |
53 | /* 0 +p index -> (type)index */ |
54 | (simplify | |
55 | (pointer_plus integer_zerop @1) | |
56 | (non_lvalue (convert @1))) | |
57 | ||
a7f24614 RB |
58 | /* See if ARG1 is zero and X + ARG1 reduces to X. |
59 | Likewise if the operands are reversed. */ | |
60 | (simplify | |
61 | (plus:c @0 real_zerop@1) | |
62 | (if (fold_real_zero_addition_p (type, @1, 0)) | |
63 | (non_lvalue @0))) | |
64 | ||
65 | /* See if ARG1 is zero and X - ARG1 reduces to X. */ | |
66 | (simplify | |
67 | (minus @0 real_zerop@1) | |
68 | (if (fold_real_zero_addition_p (type, @1, 1)) | |
69 | (non_lvalue @0))) | |
70 | ||
e0ee10ed RB |
71 | /* Simplify x - x. |
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 | |
75 | is volatile. */ | |
76 | (simplify | |
a7f24614 | 77 | (minus @0 @0) |
1b457aa4 | 78 | (if (!FLOAT_TYPE_P (type) || !HONOR_NANS (type)) |
a7f24614 | 79 | { build_zero_cst (type); })) |
e0ee10ed RB |
80 | |
81 | (simplify | |
a7f24614 RB |
82 | (mult @0 integer_zerop@1) |
83 | @1) | |
84 | ||
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. */ | |
89 | (simplify | |
90 | (mult @0 real_zerop@1) | |
1b457aa4 | 91 | (if (!HONOR_NANS (type) && !HONOR_SIGNED_ZEROS (element_mode (type))) |
a7f24614 RB |
92 | @1)) |
93 | ||
94 | /* In IEEE floating point, x*1 is not equivalent to x for snans. | |
95 | Likewise for complex arithmetic with signed zeros. */ | |
96 | (simplify | |
97 | (mult @0 real_onep) | |
09240451 MG |
98 | (if (!HONOR_SNANS (element_mode (type)) |
99 | && (!HONOR_SIGNED_ZEROS (element_mode (type)) | |
a7f24614 RB |
100 | || !COMPLEX_FLOAT_TYPE_P (type))) |
101 | (non_lvalue @0))) | |
102 | ||
103 | /* Transform x * -1.0 into -x. */ | |
104 | (simplify | |
105 | (mult @0 real_minus_onep) | |
09240451 MG |
106 | (if (!HONOR_SNANS (element_mode (type)) |
107 | && (!HONOR_SIGNED_ZEROS (element_mode (type)) | |
a7f24614 RB |
108 | || !COMPLEX_FLOAT_TYPE_P (type))) |
109 | (negate @0))) | |
e0ee10ed RB |
110 | |
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) | |
114 | (simplify | |
115 | (op @0 integer_onep) | |
116 | (non_lvalue @0))) | |
117 | ||
a7f24614 RB |
118 | /* X / -1 is -X. */ |
119 | (for div (trunc_div ceil_div floor_div round_div exact_div) | |
120 | (simplify | |
09240451 MG |
121 | (div @0 integer_minus_onep@1) |
122 | (if (!TYPE_UNSIGNED (type)) | |
a7f24614 RB |
123 | (negate @0)))) |
124 | ||
125 | /* For unsigned integral types, FLOOR_DIV_EXPR is the same as | |
126 | TRUNC_DIV_EXPR. Rewrite into the latter in this case. */ | |
127 | (simplify | |
128 | (floor_div @0 @1) | |
09240451 MG |
129 | (if ((INTEGRAL_TYPE_P (type) || VECTOR_INTEGER_TYPE_P (type)) |
130 | && TYPE_UNSIGNED (type)) | |
a7f24614 RB |
131 | (trunc_div @0 @1))) |
132 | ||
28093105 RB |
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) | |
c306cfaf RB |
136 | (simplify |
137 | (div (div @0 INTEGER_CST@1) INTEGER_CST@2) | |
138 | (with { | |
139 | bool overflow_p; | |
140 | wide_int mul = wi::mul (@1, @2, TYPE_SIGN (type), &overflow_p); | |
141 | } | |
142 | (if (!overflow_p) | |
143 | (div @0 { wide_int_to_tree (type, mul); })) | |
ac19a303 RB |
144 | (if (overflow_p |
145 | && (TYPE_UNSIGNED (type) | |
146 | || mul != wi::min_value (TYPE_PRECISION (type), SIGNED))) | |
c306cfaf RB |
147 | { build_zero_cst (type); })))) |
148 | ||
a7f24614 | 149 | /* Optimize A / A to 1.0 if we don't care about |
09240451 | 150 | NaNs or Infinities. */ |
a7f24614 RB |
151 | (simplify |
152 | (rdiv @0 @0) | |
09240451 | 153 | (if (FLOAT_TYPE_P (type) |
1b457aa4 | 154 | && ! HONOR_NANS (type) |
09240451 MG |
155 | && ! HONOR_INFINITIES (element_mode (type))) |
156 | { build_one_cst (type); })) | |
157 | ||
158 | /* Optimize -A / A to -1.0 if we don't care about | |
159 | NaNs or Infinities. */ | |
160 | (simplify | |
161 | (rdiv:c @0 (negate @0)) | |
162 | (if (FLOAT_TYPE_P (type) | |
1b457aa4 | 163 | && ! HONOR_NANS (type) |
09240451 MG |
164 | && ! HONOR_INFINITIES (element_mode (type))) |
165 | { build_minus_one_cst (type); })) | |
a7f24614 RB |
166 | |
167 | /* In IEEE floating point, x/1 is not equivalent to x for snans. */ | |
168 | (simplify | |
169 | (rdiv @0 real_onep) | |
09240451 | 170 | (if (!HONOR_SNANS (element_mode (type))) |
a7f24614 RB |
171 | (non_lvalue @0))) |
172 | ||
173 | /* In IEEE floating point, x/-1 is not equivalent to -x for snans. */ | |
174 | (simplify | |
175 | (rdiv @0 real_minus_onep) | |
09240451 | 176 | (if (!HONOR_SNANS (element_mode (type))) |
a7f24614 RB |
177 | (negate @0))) |
178 | ||
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) | |
185 | (simplify | |
186 | (rdiv @0 cst@1) | |
187 | (if (optimize) | |
53bc4b3a RB |
188 | (if (flag_reciprocal_math |
189 | && !real_zerop (@1)) | |
a7f24614 | 190 | (with |
249700b5 | 191 | { tree tem = const_binop (RDIV_EXPR, type, build_one_cst (type), @1); } |
a7f24614 RB |
192 | (if (tem) |
193 | (mult @0 { tem; } )))) | |
194 | (if (cst != COMPLEX_CST) | |
195 | (with { tree inverse = exact_inverse (type, @1); } | |
196 | (if (inverse) | |
197 | (mult @0 { inverse; } ))))))) | |
198 | ||
e0ee10ed RB |
199 | /* Same applies to modulo operations, but fold is inconsistent here |
200 | and simplifies 0 % x to 0, only preserving literal 0 % 0. */ | |
a7f24614 | 201 | (for mod (ceil_mod floor_mod round_mod trunc_mod) |
e0ee10ed RB |
202 | /* 0 % X is always zero. */ |
203 | (simplify | |
a7f24614 | 204 | (mod integer_zerop@0 @1) |
e0ee10ed RB |
205 | /* But not for 0 % 0 so that we can get the proper warnings and errors. */ |
206 | (if (!integer_zerop (@1)) | |
207 | @0)) | |
208 | /* X % 1 is always zero. */ | |
209 | (simplify | |
a7f24614 RB |
210 | (mod @0 integer_onep) |
211 | { build_zero_cst (type); }) | |
212 | /* X % -1 is zero. */ | |
213 | (simplify | |
09240451 MG |
214 | (mod @0 integer_minus_onep@1) |
215 | (if (!TYPE_UNSIGNED (type)) | |
bc4315fb MG |
216 | { build_zero_cst (type); })) |
217 | /* (X % Y) % Y is just X % Y. */ | |
218 | (simplify | |
219 | (mod (mod@2 @0 @1) @1) | |
220 | @2)) | |
a7f24614 RB |
221 | |
222 | /* X % -C is the same as X % C. */ | |
223 | (simplify | |
224 | (trunc_mod @0 INTEGER_CST@1) | |
225 | (if (TYPE_SIGN (type) == SIGNED | |
226 | && !TREE_OVERFLOW (@1) | |
227 | && wi::neg_p (@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)))) | |
e0ee10ed | 232 | |
8f0c696a RB |
233 | /* X % -Y is the same as X % Y. */ |
234 | (simplify | |
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)))) | |
240 | ||
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) | |
246 | INTEGER_CST@1) | |
247 | (match (power_of_two_cand @1) | |
248 | (lshift INTEGER_CST@1 @2)) | |
249 | (for mod (trunc_mod floor_mod) | |
250 | (simplify | |
4ab1e111 | 251 | (mod @0 (convert?@3 (power_of_two_cand@1 @2))) |
8f0c696a RB |
252 | (if ((TYPE_UNSIGNED (type) |
253 | || tree_expr_nonnegative_p (@0)) | |
4ab1e111 | 254 | && tree_nop_conversion_p (type, TREE_TYPE (@3)) |
8f0c696a | 255 | && integer_pow2p (@2) && tree_int_cst_sgn (@2) > 0) |
4ab1e111 | 256 | (bit_and @0 (convert (minus @1 { build_int_cst (TREE_TYPE (@1), 1); })))))) |
8f0c696a | 257 | |
bc4315fb MG |
258 | /* X % Y is smaller than Y. */ |
259 | (for cmp (lt ge) | |
260 | (simplify | |
261 | (cmp (trunc_mod @0 @1) @1) | |
262 | (if (TYPE_UNSIGNED (TREE_TYPE (@0))) | |
263 | { constant_boolean_node (cmp == LT_EXPR, type); }))) | |
264 | (for cmp (gt le) | |
265 | (simplify | |
266 | (cmp @1 (trunc_mod @0 @1)) | |
267 | (if (TYPE_UNSIGNED (TREE_TYPE (@0))) | |
268 | { constant_boolean_node (cmp == GT_EXPR, type); }))) | |
269 | ||
e0ee10ed RB |
270 | /* x | ~0 -> ~0 */ |
271 | (simplify | |
272 | (bit_ior @0 integer_all_onesp@1) | |
273 | @1) | |
274 | ||
275 | /* x & 0 -> 0 */ | |
276 | (simplify | |
277 | (bit_and @0 integer_zerop@1) | |
278 | @1) | |
279 | ||
280 | /* x ^ x -> 0 */ | |
281 | (simplify | |
282 | (bit_xor @0 @0) | |
283 | { build_zero_cst (type); }) | |
284 | ||
36a60e48 RB |
285 | /* Canonicalize X ^ ~0 to ~X. */ |
286 | (simplify | |
287 | (bit_xor @0 integer_all_onesp@1) | |
288 | (bit_not @0)) | |
289 | ||
290 | /* x & ~0 -> x */ | |
291 | (simplify | |
292 | (bit_and @0 integer_all_onesp) | |
293 | (non_lvalue @0)) | |
294 | ||
295 | /* x & x -> x, x | x -> x */ | |
296 | (for bitop (bit_and bit_ior) | |
297 | (simplify | |
298 | (bitop @0 @0) | |
299 | (non_lvalue @0))) | |
300 | ||
0f770b01 RV |
301 | /* x + (x & 1) -> (x + 1) & ~1 */ |
302 | (simplify | |
303 | (plus:c @0 (bit_and@2 @0 integer_onep@1)) | |
6e28e516 | 304 | (if (single_use (@2)) |
0f770b01 RV |
305 | (bit_and (plus @0 @1) (bit_not @1)))) |
306 | ||
307 | /* x & ~(x & y) -> x & ~y */ | |
308 | /* x | ~(x | y) -> x | ~y */ | |
309 | (for bitop (bit_and bit_ior) | |
af563d4b MG |
310 | (simplify |
311 | (bitop:c @0 (bit_not (bitop:c@2 @0 @1))) | |
6e28e516 | 312 | (if (single_use (@2)) |
af563d4b MG |
313 | (bitop @0 (bit_not @1))))) |
314 | ||
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) | |
319 | (simplify | |
320 | (bitop:c (rbitop:c @0 @1) (bit_not@2 @0)) | |
321 | (bitop @1 @2))) | |
0f770b01 | 322 | |
f13c4673 MP |
323 | /* (x & y) ^ (x | y) -> x ^ y */ |
324 | (simplify | |
2d6f2dce MP |
325 | (bit_xor:c (bit_and @0 @1) (bit_ior @0 @1)) |
326 | (bit_xor @0 @1)) | |
f13c4673 | 327 | |
9ea65ca6 MP |
328 | /* (x ^ y) ^ (x | y) -> x & y */ |
329 | (simplify | |
330 | (bit_xor:c (bit_xor @0 @1) (bit_ior @0 @1)) | |
331 | (bit_and @0 @1)) | |
332 | ||
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) | |
337 | (simplify | |
338 | (op:c (bit_and @0 @1) (bit_xor @0 @1)) | |
339 | (bit_ior @0 @1))) | |
340 | ||
341 | /* (x & y) + (x | y) -> x + y */ | |
342 | (simplify | |
343 | (plus:c (bit_and @0 @1) (bit_ior @0 @1)) | |
344 | (plus @0 @1)) | |
345 | ||
9737efaf MP |
346 | /* (x + y) - (x | y) -> x & y */ |
347 | (simplify | |
348 | (minus (plus @0 @1) (bit_ior @0 @1)) | |
349 | (if (!TYPE_OVERFLOW_SANITIZED (type) && !TYPE_OVERFLOW_TRAPS (type) | |
350 | && !TYPE_SATURATING (type)) | |
351 | (bit_and @0 @1))) | |
352 | ||
353 | /* (x + y) - (x & y) -> x | y */ | |
354 | (simplify | |
355 | (minus (plus @0 @1) (bit_and @0 @1)) | |
356 | (if (!TYPE_OVERFLOW_SANITIZED (type) && !TYPE_OVERFLOW_TRAPS (type) | |
357 | && !TYPE_SATURATING (type)) | |
358 | (bit_ior @0 @1))) | |
359 | ||
9ea65ca6 MP |
360 | /* (x | y) - (x ^ y) -> x & y */ |
361 | (simplify | |
362 | (minus (bit_ior @0 @1) (bit_xor @0 @1)) | |
363 | (bit_and @0 @1)) | |
364 | ||
365 | /* (x | y) - (x & y) -> x ^ y */ | |
366 | (simplify | |
367 | (minus (bit_ior @0 @1) (bit_and @0 @1)) | |
368 | (bit_xor @0 @1)) | |
369 | ||
66cc6273 MP |
370 | /* (x | y) & ~(x & y) -> x ^ y */ |
371 | (simplify | |
372 | (bit_and:c (bit_ior @0 @1) (bit_not (bit_and @0 @1))) | |
373 | (bit_xor @0 @1)) | |
374 | ||
375 | /* (x | y) & (~x ^ y) -> x & y */ | |
376 | (simplify | |
377 | (bit_and:c (bit_ior:c @0 @1) (bit_xor:c @1 (bit_not @0))) | |
378 | (bit_and @0 @1)) | |
379 | ||
f3582e54 RB |
380 | (simplify |
381 | (abs (negate @0)) | |
382 | (abs @0)) | |
383 | (simplify | |
384 | (abs tree_expr_nonnegative_p@0) | |
385 | @0) | |
386 | ||
d4573ffe | 387 | |
5609420f RB |
388 | /* Try to fold (type) X op CST -> (type) (X op ((type-x) CST)) |
389 | when profitable. | |
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) | |
395 | (simplify | |
396 | (bitop (convert @0) (convert? @1)) | |
397 | (if (((TREE_CODE (@1) == INTEGER_CST | |
398 | && INTEGRAL_TYPE_P (TREE_TYPE (@0)) | |
ad6f996c | 399 | && int_fits_type_p (@1, TREE_TYPE (@0))) |
aea417d7 | 400 | || types_match (@0, @1)) |
ad6f996c RB |
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) | |
5609420f RB |
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 | |
411 | mode. */ | |
412 | || GET_MODE_CLASS (TYPE_MODE (type)) != MODE_INT | |
413 | /* Or if the precision of TO is not the same as the precision | |
414 | of its mode. */ | |
415 | || TYPE_PRECISION (type) != GET_MODE_PRECISION (TYPE_MODE (type)))) | |
416 | (convert (bitop @0 (convert @1)))))) | |
417 | ||
418 | /* Simplify (A & B) OP0 (C & B) to (A OP0 C) & B. */ | |
419 | (for bitop (bit_and bit_ior bit_xor) | |
420 | (simplify | |
421 | (bitop (bit_and:c @0 @1) (bit_and @2 @1)) | |
422 | (bit_and (bitop @0 @2) @1))) | |
423 | ||
424 | /* (x | CST1) & CST2 -> (x & CST2) | (CST1 & CST2) */ | |
425 | (simplify | |
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))) | |
428 | ||
429 | /* Combine successive equal operations with constants. */ | |
430 | (for bitop (bit_and bit_ior bit_xor) | |
431 | (simplify | |
432 | (bitop (bitop @0 CONSTANT_CLASS_P@1) CONSTANT_CLASS_P@2) | |
433 | (bitop @0 (bitop @1 @2)))) | |
434 | ||
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 | |
438 | @0 | |
439 | (if (INTEGRAL_TYPE_P (type) && TYPE_PRECISION (type) == 1))) | |
f84e7fd6 | 440 | (for op (tcc_comparison truth_and truth_andif truth_or truth_orif truth_xor) |
5609420f RB |
441 | (match truth_valued_p |
442 | (op @0 @1))) | |
443 | (match truth_valued_p | |
444 | (truth_not @0)) | |
445 | ||
446 | (match (logical_inverted_value @0) | |
447 | (bit_not truth_valued_p@0)) | |
448 | (match (logical_inverted_value @0) | |
09240451 | 449 | (eq @0 integer_zerop)) |
5609420f | 450 | (match (logical_inverted_value @0) |
09240451 | 451 | (ne truth_valued_p@0 integer_truep)) |
5609420f | 452 | (match (logical_inverted_value @0) |
09240451 | 453 | (bit_xor truth_valued_p@0 integer_truep)) |
5609420f RB |
454 | |
455 | /* X & !X -> 0. */ | |
456 | (simplify | |
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) | |
461 | (simplify | |
462 | (op:c truth_valued_p@0 (logical_inverted_value @0)) | |
f84e7fd6 | 463 | { constant_boolean_node (true, type); })) |
5609420f RB |
464 | |
465 | (for bitop (bit_and bit_ior) | |
466 | rbitop (bit_ior bit_and) | |
467 | /* (x | y) & x -> x */ | |
468 | /* (x & y) | x -> x */ | |
469 | (simplify | |
470 | (bitop:c (rbitop:c @0 @1) @0) | |
471 | @0) | |
472 | /* (~x | y) & x -> x & y */ | |
473 | /* (~x & y) | x -> x | y */ | |
474 | (simplify | |
475 | (bitop:c (rbitop:c (bit_not @0) @1) @0) | |
476 | (bitop @0 @1))) | |
477 | ||
478 | /* If arg1 and arg2 are booleans (or any single bit type) | |
479 | then try to simplify: | |
480 | ||
481 | (~X & Y) -> X < Y | |
482 | (X & ~Y) -> Y < X | |
483 | (~X | Y) -> X <= Y | |
484 | (X | ~Y) -> Y <= X | |
485 | ||
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 */ | |
490 | (simplify | |
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) | |
494 | (lt @0 @1))) | |
495 | (simplify | |
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) | |
499 | (le @0 @1))) | |
500 | ||
5609420f RB |
501 | /* ~~x -> x */ |
502 | (simplify | |
503 | (bit_not (bit_not @0)) | |
504 | @0) | |
505 | ||
f52baa7b MP |
506 | /* (x & ~m) | (y & m) -> ((x ^ y) & m) ^ x */ |
507 | (simplify | |
508 | (bit_ior:c (bit_and:c@3 @0 (bit_not @2)) (bit_and:c@4 @1 @2)) | |
6e28e516 | 509 | (if (single_use (@3) && single_use (@4)) |
f52baa7b MP |
510 | (bit_xor (bit_and (bit_xor @0 @1) @2) @0))) |
511 | ||
5609420f | 512 | |
a499aac5 RB |
513 | /* Associate (p +p off1) +p off2 as (p +p (off1 + off2)). */ |
514 | (simplify | |
e6121733 | 515 | (pointer_plus (pointer_plus@2 @0 @1) @3) |
7318e44f RB |
516 | (if (single_use (@2) |
517 | || (TREE_CODE (@1) == INTEGER_CST && TREE_CODE (@3) == INTEGER_CST)) | |
e6121733 | 518 | (pointer_plus @0 (plus @1 @3)))) |
a499aac5 RB |
519 | |
520 | /* Pattern match | |
521 | tem1 = (long) ptr1; | |
522 | tem2 = (long) ptr2; | |
523 | tem3 = tem2 - tem1; | |
524 | tem4 = (unsigned long) tem3; | |
525 | tem5 = ptr1 + tem4; | |
526 | and produce | |
527 | tem5 = ptr2; */ | |
528 | (simplify | |
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)))) | |
534 | @1)) | |
535 | ||
536 | /* Pattern match | |
537 | tem = (sizetype) ptr; | |
538 | tem = tem & algn; | |
539 | tem = -tem; | |
540 | ... = ptr p+ tem; | |
541 | and produce the simpler and easier to analyze with respect to alignment | |
542 | ... = ptr & ~algn; */ | |
543 | (simplify | |
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; }))) | |
547 | ||
548 | ||
cc7b5acf RB |
549 | /* We can't reassociate at all for saturating types. */ |
550 | (if (!TYPE_SATURATING (type)) | |
551 | ||
552 | /* Contract negates. */ | |
553 | /* A + (-B) -> A - B */ | |
554 | (simplify | |
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)) | |
6a4f0678 | 559 | && !TYPE_OVERFLOW_SANITIZED (type)) |
cc7b5acf RB |
560 | (minus (convert @0) (convert @1)))) |
561 | /* A - (-B) -> A + B */ | |
562 | (simplify | |
563 | (minus (convert1? @0) (convert2? (negate @1))) | |
564 | (if (tree_nop_conversion_p (type, TREE_TYPE (@0)) | |
2f68e8bc | 565 | && tree_nop_conversion_p (type, TREE_TYPE (@1)) |
6a4f0678 | 566 | && !TYPE_OVERFLOW_SANITIZED (type)) |
cc7b5acf RB |
567 | (plus (convert @0) (convert @1)))) |
568 | /* -(-A) -> A */ | |
569 | (simplify | |
570 | (negate (convert? (negate @1))) | |
571 | (if (tree_nop_conversion_p (type, TREE_TYPE (@1)) | |
6a4f0678 | 572 | && !TYPE_OVERFLOW_SANITIZED (type)) |
a0f12cf8 | 573 | (convert @1))) |
cc7b5acf | 574 | |
7318e44f RB |
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)) | |
cc7b5acf RB |
579 | |
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 */ | |
586 | (simplify | |
587 | (minus (plus:c @0 @1) @0) | |
588 | @1) | |
589 | (simplify | |
590 | (minus (minus @0 @1) @0) | |
591 | (negate @1)) | |
592 | (simplify | |
593 | (plus:c (minus @0 @1) @1) | |
594 | @0) | |
595 | (simplify | |
596 | (minus @0 (plus:c @0 @1)) | |
597 | (negate @1)) | |
598 | (simplify | |
599 | (minus @0 (minus @0 @1)) | |
600 | @1) | |
601 | ||
602 | /* (A +- CST) +- CST -> A + CST */ | |
603 | (for outer_op (plus minus) | |
604 | (for inner_op (plus minus) | |
605 | (simplify | |
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; } )))))) | |
614 | ||
615 | /* (CST - A) +- CST -> CST - A */ | |
616 | (for outer_op (plus minus) | |
617 | (simplify | |
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))))) | |
622 | ||
623 | /* ~A + A -> -1 */ | |
624 | (simplify | |
625 | (plus:c (bit_not @0) @0) | |
626 | (if (!TYPE_OVERFLOW_TRAPS (type)) | |
627 | { build_all_ones_cst (type); })) | |
628 | ||
629 | /* ~A + 1 -> -A */ | |
630 | (simplify | |
e19740ae RB |
631 | (plus (convert? (bit_not @0)) integer_each_onep) |
632 | (if (tree_nop_conversion_p (type, TREE_TYPE (@0))) | |
633 | (negate (convert @0)))) | |
634 | ||
635 | /* -A - 1 -> ~A */ | |
636 | (simplify | |
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)))) | |
641 | ||
642 | /* -1 - A -> ~A */ | |
643 | (simplify | |
644 | (minus integer_all_onesp @0) | |
bc4315fb | 645 | (bit_not @0)) |
cc7b5acf RB |
646 | |
647 | /* (T)(P + A) - (T)P -> (T) A */ | |
648 | (for add (plus pointer_plus) | |
649 | (simplify | |
650 | (minus (convert (add @0 @1)) | |
651 | (convert @0)) | |
09240451 | 652 | (if (element_precision (type) <= element_precision (TREE_TYPE (@1)) |
cc7b5acf RB |
653 | /* For integer types, if A has a smaller type |
654 | than T the result depends on the possible | |
655 | overflow in P + A. | |
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 | |
659 | is no overflow. */ | |
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 | |
665 | one is correct. */ | |
666 | || (POINTER_TYPE_P (TREE_TYPE (@0)) | |
667 | && TREE_CODE (@1) == INTEGER_CST | |
668 | && tree_int_cst_sign_bit (@1) == 0)) | |
669 | (convert @1)))))) | |
670 | ||
671 | ||
a7f24614 RB |
672 | /* Simplifications of MIN_EXPR and MAX_EXPR. */ |
673 | ||
674 | (for minmax (min max) | |
675 | (simplify | |
676 | (minmax @0 @0) | |
677 | @0)) | |
678 | (simplify | |
679 | (min @0 @1) | |
680 | (if (INTEGRAL_TYPE_P (type) | |
681 | && TYPE_MIN_VALUE (type) | |
682 | && operand_equal_p (@1, TYPE_MIN_VALUE (type), OEP_ONLY_CONST)) | |
683 | @1)) | |
684 | (simplify | |
685 | (max @0 @1) | |
686 | (if (INTEGRAL_TYPE_P (type) | |
687 | && TYPE_MAX_VALUE (type) | |
688 | && operand_equal_p (@1, TYPE_MAX_VALUE (type), OEP_ONLY_CONST)) | |
689 | @1)) | |
690 | ||
691 | ||
692 | /* Simplifications of shift and rotates. */ | |
693 | ||
694 | (for rotate (lrotate rrotate) | |
695 | (simplify | |
696 | (rotate integer_all_onesp@0 @1) | |
697 | @0)) | |
698 | ||
699 | /* Optimize -1 >> x for arithmetic right shifts. */ | |
700 | (simplify | |
701 | (rshift integer_all_onesp@0 @1) | |
702 | (if (!TYPE_UNSIGNED (type) | |
703 | && tree_expr_nonnegative_p (@1)) | |
704 | @0)) | |
705 | ||
706 | (for shiftrotate (lrotate rrotate lshift rshift) | |
707 | (simplify | |
708 | (shiftrotate @0 integer_zerop) | |
709 | (non_lvalue @0)) | |
710 | (simplify | |
711 | (shiftrotate integer_zerop@0 @1) | |
712 | @0) | |
713 | /* Prefer vector1 << scalar to vector1 << vector2 | |
714 | if vector2 is uniform. */ | |
715 | (for vec (VECTOR_CST CONSTRUCTOR) | |
716 | (simplify | |
717 | (shiftrotate @0 vec@1) | |
718 | (with { tree tem = uniform_vector_p (@1); } | |
719 | (if (tem) | |
720 | (shiftrotate @0 { tem; })))))) | |
721 | ||
722 | /* Rewrite an LROTATE_EXPR by a constant into an | |
723 | RROTATE_EXPR by a new constant. */ | |
724 | (simplify | |
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); })) | |
729 | ||
01ada710 MP |
730 | /* ((1 << A) & 1) != 0 -> A == 0 |
731 | ((1 << A) & 1) == 0 -> A != 0 */ | |
732 | (for cmp (ne eq) | |
733 | icmp (eq ne) | |
734 | (simplify | |
735 | (cmp (bit_and (lshift integer_onep @0) integer_onep) integer_zerop) | |
736 | (icmp @0 { build_zero_cst (TREE_TYPE (@0)); }))) | |
cc7b5acf | 737 | |
f2e609c3 MP |
738 | /* (CST1 << A) == CST2 -> A == ctz (CST2) - ctz (CST1) |
739 | (CST1 << A) != CST2 -> A != ctz (CST2) - ctz (CST1) | |
740 | if CST2 != 0. */ | |
741 | (for cmp (ne eq) | |
742 | (simplify | |
743 | (cmp (lshift INTEGER_CST@0 @1) INTEGER_CST@2) | |
744 | (with { int cand = wi::ctz (@2) - wi::ctz (@0); } | |
745 | (if (cand < 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); }))))) | |
752 | ||
1ffbaa3f RB |
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) | |
757 | (simplify | |
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)) | |
764 | (with | |
765 | { | |
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); | |
770 | unsigned int prec; | |
771 | ||
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)))) | |
777 | { | |
778 | prec = TYPE_PRECISION (TREE_TYPE (@3)); | |
779 | tree arg00 = @0; | |
780 | /* See if more bits can be proven as zero because of | |
781 | zero extension. */ | |
782 | if (@3 != @0 | |
783 | && TYPE_UNSIGNED (TREE_TYPE (@0))) | |
784 | { | |
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) | |
789 | { | |
790 | prec = TYPE_PRECISION (inner_type); | |
791 | /* See if we can shorten the right shift. */ | |
792 | if (shiftc < prec) | |
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 | |
796 | is all ones. */ | |
797 | } | |
798 | } | |
799 | zerobits = ~(unsigned HOST_WIDE_INT) 0; | |
800 | if (shiftc < prec) | |
801 | { | |
802 | zerobits >>= HOST_BITS_PER_WIDE_INT - shiftc; | |
803 | zerobits <<= prec - shiftc; | |
804 | } | |
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))) | |
811 | { | |
812 | if ((mask & zerobits) == 0) | |
813 | shift_type = unsigned_type_for (TREE_TYPE (@3)); | |
814 | else | |
815 | zerobits = 0; | |
816 | } | |
817 | } | |
818 | } | |
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) | |
824 | (with | |
825 | { | |
826 | /* Only do the transformation if NEWMASK is some integer | |
827 | mode's mask. */ | |
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) | |
831 | break; | |
832 | } | |
833 | (if (prec < HOST_BITS_PER_WIDE_INT | |
834 | || newmask == ~(unsigned HOST_WIDE_INT) 0) | |
835 | (with | |
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; })))))))))))) | |
843 | ||
d4573ffe RB |
844 | /* Simplifications of conversions. */ |
845 | ||
846 | /* Basic strip-useless-type-conversions / strip_nops. */ | |
f3582e54 | 847 | (for cvt (convert view_convert float fix_trunc) |
d4573ffe RB |
848 | (simplify |
849 | (cvt @0) | |
850 | (if ((GIMPLE && useless_type_conversion_p (type, TREE_TYPE (@0))) | |
851 | || (GENERIC && type == TREE_TYPE (@0))) | |
852 | @0))) | |
853 | ||
854 | /* Contract view-conversions. */ | |
855 | (simplify | |
856 | (view_convert (view_convert @0)) | |
857 | (view_convert @0)) | |
858 | ||
859 | /* For integral conversions with the same precision or pointer | |
860 | conversions use a NOP_EXPR instead. */ | |
861 | (simplify | |
862 | (view_convert @0) | |
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))) | |
866 | (convert @0))) | |
867 | ||
868 | /* Strip inner integral conversions that do not change precision or size. */ | |
869 | (simplify | |
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)))) | |
875 | (view_convert @1))) | |
876 | ||
877 | /* Re-association barriers around constants and other re-association | |
878 | barriers can be removed. */ | |
879 | (simplify | |
880 | (paren CONSTANT_CLASS_P@0) | |
881 | @0) | |
882 | (simplify | |
883 | (paren (paren@1 @0)) | |
884 | @1) | |
1e51d0a2 RB |
885 | |
886 | /* Handle cases of two conversions in a row. */ | |
887 | (for ocvt (convert float fix_trunc) | |
888 | (for icvt (convert float) | |
889 | (simplify | |
890 | (ocvt (icvt@1 @0)) | |
891 | (with | |
892 | { | |
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); | |
09240451 | 898 | int inside_vec = VECTOR_TYPE_P (inside_type); |
1e51d0a2 RB |
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); | |
09240451 | 904 | int inter_vec = VECTOR_TYPE_P (inter_type); |
1e51d0a2 RB |
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); | |
09240451 | 910 | int final_vec = VECTOR_TYPE_P (type); |
1e51d0a2 RB |
911 | unsigned int final_prec = TYPE_PRECISION (type); |
912 | int final_unsignedp = TYPE_UNSIGNED (type); | |
913 | } | |
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)) | |
919 | || (GENERIC | |
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) | |
924 | (ocvt @0)) | |
925 | ||
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. */ | |
d51a6714 JJ |
932 | (if (((inter_int && inside_int) || (inter_float && inside_float)) |
933 | && (final_int || final_float) | |
1e51d0a2 | 934 | && inter_prec >= inside_prec |
d51a6714 JJ |
935 | && (inter_float || inter_unsignedp == inside_unsignedp) |
936 | && ! (final_prec != GET_MODE_PRECISION (TYPE_MODE (type)) | |
937 | && TYPE_MODE (type) == TYPE_MODE (inter_type))) | |
1e51d0a2 RB |
938 | (ocvt @0)) |
939 | ||
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)) | |
948 | (ocvt @0)) | |
949 | ||
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 | |
954 | final, or | |
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))) | |
1f00c1b9 RB |
973 | (ocvt @0)) |
974 | ||
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 | |
980 | && inter_unsignedp) | |
981 | (convert (bit_and @0 { wide_int_to_tree | |
982 | (inside_type, | |
983 | wi::mask (inter_prec, false, | |
984 | TYPE_PRECISION (inside_type))); }))) | |
985 | ||
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. */ | |
5ba3ae6d RB |
989 | (if (GIMPLE /* PR66211 */ |
990 | && inside_int && inter_float && final_int && | |
1f00c1b9 RB |
991 | (unsigned) significand_size (TYPE_MODE (inter_type)) |
992 | >= inside_prec - !inside_unsignedp) | |
993 | (convert @0)))))) | |
ea2042ba RB |
994 | |
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). */ | |
998 | (simplify | |
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)) | |
1005 | (convert @0))) | |
a25454ea RB |
1006 | |
1007 | ||
1008 | /* (X /[ex] A) * A -> X. */ | |
1009 | (simplify | |
1010 | (mult (convert? (exact_div @0 @1)) @1) | |
1011 | /* Look through a sign-changing conversion. */ | |
257b01ba | 1012 | (convert @0)) |
eaeba53a | 1013 | |
a7f24614 RB |
1014 | /* Canonicalization of binary operations. */ |
1015 | ||
1016 | /* Convert X + -C into X - C. */ | |
1017 | (simplify | |
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; }))))) | |
1023 | ||
1024 | /* Convert x+x into x*2.0. */ | |
1025 | (simplify | |
1026 | (plus @0 @0) | |
1027 | (if (SCALAR_FLOAT_TYPE_P (type)) | |
1028 | (mult @0 { build_real (type, dconst2); }))) | |
1029 | ||
1030 | (simplify | |
1031 | (minus integer_zerop @1) | |
1032 | (negate @1)) | |
1033 | ||
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. */ | |
1037 | (simplify | |
1038 | (minus real_zerop@0 @1) | |
1039 | (if (fold_real_zero_addition_p (type, @0, 0)) | |
1040 | (negate @1))) | |
1041 | ||
1042 | /* Transform x * -1 into -x. */ | |
1043 | (simplify | |
1044 | (mult @0 integer_minus_onep) | |
1045 | (negate @0)) | |
eaeba53a RB |
1046 | |
1047 | /* COMPLEX_EXPR and REALPART/IMAGPART_EXPR cancellations. */ | |
1048 | (simplify | |
1049 | (complex (realpart @0) (imagpart @0)) | |
1050 | @0) | |
1051 | (simplify | |
1052 | (realpart (complex @0 @1)) | |
1053 | @0) | |
1054 | (simplify | |
1055 | (imagpart (complex @0 @1)) | |
1056 | @1) | |
83633539 RB |
1057 | |
1058 | ||
1059 | /* BSWAP simplifications, transforms checked by gcc.dg/builtin-bswap-8.c. */ | |
1060 | (for bswap (BUILT_IN_BSWAP16 BUILT_IN_BSWAP32 BUILT_IN_BSWAP64) | |
1061 | (simplify | |
1062 | (bswap (bswap @0)) | |
1063 | @0) | |
1064 | (simplify | |
1065 | (bswap (bit_not (bswap @0))) | |
1066 | (bit_not @0)) | |
1067 | (for bitop (bit_xor bit_ior bit_and) | |
1068 | (simplify | |
1069 | (bswap (bitop:c (bswap @0) @1)) | |
1070 | (bitop @0 (bswap @1))))) | |
96994de0 RB |
1071 | |
1072 | ||
1073 | /* Combine COND_EXPRs and VEC_COND_EXPRs. */ | |
1074 | ||
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. */ | |
1084 | (simplify | |
1085 | (cond INTEGER_CST@0 @1 @2) | |
1086 | (if (integer_zerop (@0) | |
1087 | && (!VOID_TYPE_P (TREE_TYPE (@2)) | |
1088 | || VOID_TYPE_P (type))) | |
1089 | @2) | |
1090 | (if (!integer_zerop (@0) | |
1091 | && (!VOID_TYPE_P (TREE_TYPE (@1)) | |
1092 | || VOID_TYPE_P (type))) | |
1093 | @1)) | |
1094 | (simplify | |
1095 | (vec_cond VECTOR_CST@0 @1 @2) | |
1096 | (if (integer_all_onesp (@0)) | |
1097 | @1) | |
1098 | (if (integer_zerop (@0)) | |
1099 | @2)) | |
1100 | ||
1101 | (for cnd (cond vec_cond) | |
1102 | /* A ? B : (A ? X : C) -> A ? B : C. */ | |
1103 | (simplify | |
1104 | (cnd @0 (cnd @0 @1 @2) @3) | |
1105 | (cnd @0 @1 @3)) | |
1106 | (simplify | |
1107 | (cnd @0 @1 (cnd @0 @2 @3)) | |
1108 | (cnd @0 @1 @3)) | |
1109 | ||
1110 | /* A ? B : B -> B. */ | |
1111 | (simplify | |
1112 | (cnd @0 @1 @1) | |
09240451 | 1113 | @1) |
96994de0 | 1114 | |
09240451 MG |
1115 | /* !A ? B : C -> A ? C : B. */ |
1116 | (simplify | |
1117 | (cnd (logical_inverted_value truth_valued_p@0) @1 @2) | |
1118 | (cnd @0 @2 @1))) | |
f84e7fd6 | 1119 | |
f43d102e RS |
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. */ | |
1124 | (simplify | |
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)))) | |
1131 | ||
1132 | /* ... likewise A - (B vcmp C ? 1 : 0) -> A + (B vcmp C). */ | |
1133 | (simplify | |
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)))) | |
f84e7fd6 RB |
1140 | |
1141 | /* Simplifications of comparisons. */ | |
1142 | ||
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. */ | |
1155 | (simplify | |
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 | |
1b457aa4 | 1164 | (cmp, HONOR_NANS (@0)); } |
f84e7fd6 RB |
1165 | (if (ic == icmp) |
1166 | (icmp @0 @1)) | |
1167 | (if (ic == ncmp) | |
1168 | (ncmp @0 @1))))) | |
1169 | (simplify | |
09240451 MG |
1170 | (bit_xor (cmp @0 @1) integer_truep) |
1171 | (with { enum tree_code ic = invert_tree_comparison | |
1b457aa4 | 1172 | (cmp, HONOR_NANS (@0)); } |
09240451 MG |
1173 | (if (ic == icmp) |
1174 | (icmp @0 @1)) | |
1175 | (if (ic == ncmp) | |
1176 | (ncmp @0 @1))))) | |
e18c1d66 | 1177 | |
cfdc4f33 MG |
1178 | /* Unordered tests if either argument is a NaN. */ |
1179 | (simplify | |
1180 | (bit_ior (unordered @0 @0) (unordered @1 @1)) | |
aea417d7 | 1181 | (if (types_match (@0, @1)) |
cfdc4f33 | 1182 | (unordered @0 @1))) |
257b01ba MG |
1183 | (simplify |
1184 | (bit_and (ordered @0 @0) (ordered @1 @1)) | |
1185 | (if (types_match (@0, @1)) | |
1186 | (ordered @0 @1))) | |
cfdc4f33 MG |
1187 | (simplify |
1188 | (bit_ior:c (unordered @0 @0) (unordered:c@2 @0 @1)) | |
1189 | @2) | |
257b01ba MG |
1190 | (simplify |
1191 | (bit_and:c (ordered @0 @0) (ordered:c@2 @0 @1)) | |
1192 | @2) | |
e18c1d66 | 1193 | |
534bd33b MG |
1194 | /* -A CMP -B -> B CMP A. */ |
1195 | (for cmp (tcc_comparison) | |
1196 | scmp (swapped_tcc_comparison) | |
1197 | (simplify | |
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)))) | |
1202 | (scmp @0 @1))) | |
1203 | (simplify | |
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; })))))) | |
1211 | ||
e18c1d66 RB |
1212 | /* Simplification of math builtins. */ |
1213 | ||
1214 | (define_operator_list LOG BUILT_IN_LOGF BUILT_IN_LOG BUILT_IN_LOGL) | |
1215 | (define_operator_list EXP BUILT_IN_EXPF BUILT_IN_EXP BUILT_IN_EXPL) | |
1216 | (define_operator_list LOG2 BUILT_IN_LOG2F BUILT_IN_LOG2 BUILT_IN_LOG2L) | |
1217 | (define_operator_list EXP2 BUILT_IN_EXP2F BUILT_IN_EXP2 BUILT_IN_EXP2L) | |
1218 | (define_operator_list LOG10 BUILT_IN_LOG10F BUILT_IN_LOG10 BUILT_IN_LOG10L) | |
1219 | (define_operator_list EXP10 BUILT_IN_EXP10F BUILT_IN_EXP10 BUILT_IN_EXP10L) | |
1220 | (define_operator_list POW BUILT_IN_POWF BUILT_IN_POW BUILT_IN_POWL) | |
1221 | (define_operator_list POW10 BUILT_IN_POW10F BUILT_IN_POW10 BUILT_IN_POW10L) | |
1222 | (define_operator_list SQRT BUILT_IN_SQRTF BUILT_IN_SQRT BUILT_IN_SQRTL) | |
1223 | (define_operator_list CBRT BUILT_IN_CBRTF BUILT_IN_CBRT BUILT_IN_CBRTL) | |
1224 | ||
1225 | ||
1226 | /* fold_builtin_logarithm */ | |
1227 | (if (flag_unsafe_math_optimizations) | |
1228 | /* Special case, optimize logN(expN(x)) = x. */ | |
1229 | (for logs (LOG LOG2 LOG10) | |
1230 | exps (EXP EXP2 EXP10) | |
1231 | (simplify | |
1232 | (logs (exps @0)) | |
1233 | @0)) | |
1234 | /* Optimize logN(func()) for various exponential functions. We | |
1235 | want to determine the value "x" and the power "exponent" in | |
1236 | order to transform logN(x**exponent) into exponent*logN(x). */ | |
1237 | (for logs (LOG LOG LOG LOG | |
1238 | LOG2 LOG2 LOG2 LOG2 | |
1239 | LOG10 LOG10 LOG10 LOG10) | |
1240 | exps (EXP EXP2 EXP10 POW10) | |
1241 | (simplify | |
1242 | (logs (exps @0)) | |
1243 | (with { | |
1244 | tree x; | |
1245 | switch (exps) | |
1246 | { | |
1247 | CASE_FLT_FN (BUILT_IN_EXP): | |
1248 | /* Prepare to do logN(exp(exponent) -> exponent*logN(e). */ | |
1249 | x = build_real (type, real_value_truncate (TYPE_MODE (type), | |
1250 | dconst_e ())); | |
1251 | break; | |
1252 | CASE_FLT_FN (BUILT_IN_EXP2): | |
1253 | /* Prepare to do logN(exp2(exponent) -> exponent*logN(2). */ | |
1254 | x = build_real (type, dconst2); | |
1255 | break; | |
1256 | CASE_FLT_FN (BUILT_IN_EXP10): | |
1257 | CASE_FLT_FN (BUILT_IN_POW10): | |
1258 | /* Prepare to do logN(exp10(exponent) -> exponent*logN(10). */ | |
1259 | { | |
1260 | REAL_VALUE_TYPE dconst10; | |
1261 | real_from_integer (&dconst10, VOIDmode, 10, SIGNED); | |
1262 | x = build_real (type, dconst10); | |
1263 | } | |
1264 | break; | |
1265 | } | |
1266 | } | |
1267 | (mult (logs { x; }) @0)))) | |
1268 | (for logs (LOG LOG | |
1269 | LOG2 LOG2 | |
1270 | LOG10 LOG10) | |
1271 | exps (SQRT CBRT) | |
1272 | (simplify | |
1273 | (logs (exps @0)) | |
1274 | (with { | |
1275 | tree x; | |
1276 | switch (exps) | |
1277 | { | |
1278 | CASE_FLT_FN (BUILT_IN_SQRT): | |
1279 | /* Prepare to do logN(sqrt(x) -> 0.5*logN(x). */ | |
1280 | x = build_real (type, dconsthalf); | |
1281 | break; | |
1282 | CASE_FLT_FN (BUILT_IN_CBRT): | |
1283 | /* Prepare to do logN(cbrt(x) -> (1/3)*logN(x). */ | |
1284 | x = build_real (type, real_value_truncate (TYPE_MODE (type), | |
1285 | dconst_third ())); | |
1286 | break; | |
1287 | } | |
1288 | } | |
1289 | (mult { x; } (logs @0))))) | |
1290 | /* logN(pow(x,exponent) -> exponent*logN(x). */ | |
1291 | (for logs (LOG LOG2 LOG10) | |
1292 | pows (POW) | |
1293 | (simplify | |
1294 | (logs (pows @0 @1)) | |
1295 | (mult @1 (logs @0))))) | |
1296 | ||
be144838 JL |
1297 | /* Narrowing of arithmetic and logical operations. |
1298 | ||
1299 | These are conceptually similar to the transformations performed for | |
1300 | the C/C++ front-ends by shorten_binary_op and shorten_compare. Long | |
1301 | term we want to move all that code out of the front-ends into here. */ | |
1302 | ||
1303 | /* If we have a narrowing conversion of an arithmetic operation where | |
1304 | both operands are widening conversions from the same type as the outer | |
1305 | narrowing conversion. Then convert the innermost operands to a suitable | |
1306 | unsigned type (to avoid introducing undefined behaviour), perform the | |
1307 | operation and convert the result to the desired type. */ | |
1308 | (for op (plus minus) | |
1309 | (simplify | |
48451e8f | 1310 | (convert (op@4 (convert@2 @0) (convert@3 @1))) |
be144838 JL |
1311 | (if (INTEGRAL_TYPE_P (type) |
1312 | /* We check for type compatibility between @0 and @1 below, | |
1313 | so there's no need to check that @1/@3 are integral types. */ | |
1314 | && INTEGRAL_TYPE_P (TREE_TYPE (@0)) | |
1315 | && INTEGRAL_TYPE_P (TREE_TYPE (@2)) | |
1316 | /* The precision of the type of each operand must match the | |
1317 | precision of the mode of each operand, similarly for the | |
1318 | result. */ | |
1319 | && (TYPE_PRECISION (TREE_TYPE (@0)) | |
1320 | == GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (@0)))) | |
1321 | && (TYPE_PRECISION (TREE_TYPE (@1)) | |
1322 | == GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (@1)))) | |
1323 | && TYPE_PRECISION (type) == GET_MODE_PRECISION (TYPE_MODE (type)) | |
1324 | /* The inner conversion must be a widening conversion. */ | |
1325 | && TYPE_PRECISION (TREE_TYPE (@2)) > TYPE_PRECISION (TREE_TYPE (@0)) | |
aea417d7 MG |
1326 | && types_match (@0, @1) |
1327 | && types_match (@0, type) | |
48451e8f | 1328 | && single_use (@4)) |
be144838 JL |
1329 | (if (TYPE_OVERFLOW_WRAPS (TREE_TYPE (@0))) |
1330 | (convert (op @0 @1))) | |
1331 | (with { tree utype = unsigned_type_for (TREE_TYPE (@0)); } | |
1332 | (convert (op (convert:utype @0) (convert:utype @1))))))) | |
48451e8f JL |
1333 | |
1334 | /* This is another case of narrowing, specifically when there's an outer | |
1335 | BIT_AND_EXPR which masks off bits outside the type of the innermost | |
1336 | operands. Like the previous case we have to convert the operands | |
1337 | to unsigned types to avoid introducing undefined behaviour for the | |
1338 | arithmetic operation. */ | |
1339 | (for op (minus plus) | |
1340 | (simplify | |
1341 | (bit_and (op@5 (convert@2 @0) (convert@3 @1)) INTEGER_CST@4) | |
1342 | (if (INTEGRAL_TYPE_P (type) | |
1343 | /* We check for type compatibility between @0 and @1 below, | |
1344 | so there's no need to check that @1/@3 are integral types. */ | |
1345 | && INTEGRAL_TYPE_P (TREE_TYPE (@0)) | |
1346 | && INTEGRAL_TYPE_P (TREE_TYPE (@2)) | |
1347 | /* The precision of the type of each operand must match the | |
1348 | precision of the mode of each operand, similarly for the | |
1349 | result. */ | |
1350 | && (TYPE_PRECISION (TREE_TYPE (@0)) | |
1351 | == GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (@0)))) | |
1352 | && (TYPE_PRECISION (TREE_TYPE (@1)) | |
1353 | == GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (@1)))) | |
1354 | && TYPE_PRECISION (type) == GET_MODE_PRECISION (TYPE_MODE (type)) | |
1355 | /* The inner conversion must be a widening conversion. */ | |
1356 | && TYPE_PRECISION (TREE_TYPE (@2)) > TYPE_PRECISION (TREE_TYPE (@0)) | |
aea417d7 | 1357 | && types_match (@0, @1) |
a60c51fe | 1358 | && (tree_int_cst_min_precision (@4, TYPE_SIGN (TREE_TYPE (@0))) |
48451e8f | 1359 | <= TYPE_PRECISION (TREE_TYPE (@0))) |
a60c51fe JJ |
1360 | && (TYPE_OVERFLOW_WRAPS (TREE_TYPE (@0)) |
1361 | || tree_int_cst_sgn (@4) >= 0) | |
48451e8f JL |
1362 | && single_use (@5)) |
1363 | (if (TYPE_OVERFLOW_WRAPS (TREE_TYPE (@0))) | |
1364 | (with { tree ntype = TREE_TYPE (@0); } | |
1365 | (convert (bit_and (op @0 @1) (convert:ntype @4))))) | |
1366 | (with { tree utype = unsigned_type_for (TREE_TYPE (@0)); } | |
1367 | (convert (bit_and (op (convert:utype @0) (convert:utype @1)) | |
1368 | (convert:utype @4))))))) | |
1369 |