]>
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) | |
41 | ||
e0ee10ed RB |
42 | |
43 | /* Simplifications of operations with one constant operand and | |
36a60e48 | 44 | simplifications to constants or single values. */ |
e0ee10ed RB |
45 | |
46 | (for op (plus pointer_plus minus bit_ior bit_xor) | |
47 | (simplify | |
48 | (op @0 integer_zerop) | |
49 | (non_lvalue @0))) | |
50 | ||
a499aac5 RB |
51 | /* 0 +p index -> (type)index */ |
52 | (simplify | |
53 | (pointer_plus integer_zerop @1) | |
54 | (non_lvalue (convert @1))) | |
55 | ||
a7f24614 RB |
56 | /* See if ARG1 is zero and X + ARG1 reduces to X. |
57 | Likewise if the operands are reversed. */ | |
58 | (simplify | |
59 | (plus:c @0 real_zerop@1) | |
60 | (if (fold_real_zero_addition_p (type, @1, 0)) | |
61 | (non_lvalue @0))) | |
62 | ||
63 | /* See if ARG1 is zero and X - ARG1 reduces to X. */ | |
64 | (simplify | |
65 | (minus @0 real_zerop@1) | |
66 | (if (fold_real_zero_addition_p (type, @1, 1)) | |
67 | (non_lvalue @0))) | |
68 | ||
e0ee10ed RB |
69 | /* Simplify x - x. |
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 | |
73 | is volatile. */ | |
74 | (simplify | |
a7f24614 | 75 | (minus @0 @0) |
1b457aa4 | 76 | (if (!FLOAT_TYPE_P (type) || !HONOR_NANS (type)) |
a7f24614 | 77 | { build_zero_cst (type); })) |
e0ee10ed RB |
78 | |
79 | (simplify | |
a7f24614 RB |
80 | (mult @0 integer_zerop@1) |
81 | @1) | |
82 | ||
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. */ | |
87 | (simplify | |
88 | (mult @0 real_zerop@1) | |
1b457aa4 | 89 | (if (!HONOR_NANS (type) && !HONOR_SIGNED_ZEROS (element_mode (type))) |
a7f24614 RB |
90 | @1)) |
91 | ||
92 | /* In IEEE floating point, x*1 is not equivalent to x for snans. | |
93 | Likewise for complex arithmetic with signed zeros. */ | |
94 | (simplify | |
95 | (mult @0 real_onep) | |
09240451 MG |
96 | (if (!HONOR_SNANS (element_mode (type)) |
97 | && (!HONOR_SIGNED_ZEROS (element_mode (type)) | |
a7f24614 RB |
98 | || !COMPLEX_FLOAT_TYPE_P (type))) |
99 | (non_lvalue @0))) | |
100 | ||
101 | /* Transform x * -1.0 into -x. */ | |
102 | (simplify | |
103 | (mult @0 real_minus_onep) | |
09240451 MG |
104 | (if (!HONOR_SNANS (element_mode (type)) |
105 | && (!HONOR_SIGNED_ZEROS (element_mode (type)) | |
a7f24614 RB |
106 | || !COMPLEX_FLOAT_TYPE_P (type))) |
107 | (negate @0))) | |
e0ee10ed RB |
108 | |
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) | |
112 | (simplify | |
113 | (op @0 integer_onep) | |
114 | (non_lvalue @0))) | |
115 | ||
a7f24614 RB |
116 | /* X / -1 is -X. */ |
117 | (for div (trunc_div ceil_div floor_div round_div exact_div) | |
118 | (simplify | |
09240451 MG |
119 | (div @0 integer_minus_onep@1) |
120 | (if (!TYPE_UNSIGNED (type)) | |
a7f24614 RB |
121 | (negate @0)))) |
122 | ||
123 | /* For unsigned integral types, FLOOR_DIV_EXPR is the same as | |
124 | TRUNC_DIV_EXPR. Rewrite into the latter in this case. */ | |
125 | (simplify | |
126 | (floor_div @0 @1) | |
09240451 MG |
127 | (if ((INTEGRAL_TYPE_P (type) || VECTOR_INTEGER_TYPE_P (type)) |
128 | && TYPE_UNSIGNED (type)) | |
a7f24614 RB |
129 | (trunc_div @0 @1))) |
130 | ||
28093105 RB |
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) | |
c306cfaf RB |
134 | (simplify |
135 | (div (div @0 INTEGER_CST@1) INTEGER_CST@2) | |
136 | (with { | |
137 | bool overflow_p; | |
138 | wide_int mul = wi::mul (@1, @2, TYPE_SIGN (type), &overflow_p); | |
139 | } | |
140 | (if (!overflow_p) | |
141 | (div @0 { wide_int_to_tree (type, mul); })) | |
ac19a303 RB |
142 | (if (overflow_p |
143 | && (TYPE_UNSIGNED (type) | |
144 | || mul != wi::min_value (TYPE_PRECISION (type), SIGNED))) | |
c306cfaf RB |
145 | { build_zero_cst (type); })))) |
146 | ||
a7f24614 | 147 | /* Optimize A / A to 1.0 if we don't care about |
09240451 | 148 | NaNs or Infinities. */ |
a7f24614 RB |
149 | (simplify |
150 | (rdiv @0 @0) | |
09240451 | 151 | (if (FLOAT_TYPE_P (type) |
1b457aa4 | 152 | && ! HONOR_NANS (type) |
09240451 MG |
153 | && ! HONOR_INFINITIES (element_mode (type))) |
154 | { build_one_cst (type); })) | |
155 | ||
156 | /* Optimize -A / A to -1.0 if we don't care about | |
157 | NaNs or Infinities. */ | |
158 | (simplify | |
159 | (rdiv:c @0 (negate @0)) | |
160 | (if (FLOAT_TYPE_P (type) | |
1b457aa4 | 161 | && ! HONOR_NANS (type) |
09240451 MG |
162 | && ! HONOR_INFINITIES (element_mode (type))) |
163 | { build_minus_one_cst (type); })) | |
a7f24614 RB |
164 | |
165 | /* In IEEE floating point, x/1 is not equivalent to x for snans. */ | |
166 | (simplify | |
167 | (rdiv @0 real_onep) | |
09240451 | 168 | (if (!HONOR_SNANS (element_mode (type))) |
a7f24614 RB |
169 | (non_lvalue @0))) |
170 | ||
171 | /* In IEEE floating point, x/-1 is not equivalent to -x for snans. */ | |
172 | (simplify | |
173 | (rdiv @0 real_minus_onep) | |
09240451 | 174 | (if (!HONOR_SNANS (element_mode (type))) |
a7f24614 RB |
175 | (negate @0))) |
176 | ||
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) | |
183 | (simplify | |
184 | (rdiv @0 cst@1) | |
185 | (if (optimize) | |
53bc4b3a RB |
186 | (if (flag_reciprocal_math |
187 | && !real_zerop (@1)) | |
a7f24614 | 188 | (with |
249700b5 | 189 | { tree tem = const_binop (RDIV_EXPR, type, build_one_cst (type), @1); } |
a7f24614 RB |
190 | (if (tem) |
191 | (mult @0 { tem; } )))) | |
192 | (if (cst != COMPLEX_CST) | |
193 | (with { tree inverse = exact_inverse (type, @1); } | |
194 | (if (inverse) | |
195 | (mult @0 { inverse; } ))))))) | |
196 | ||
e0ee10ed RB |
197 | /* Same applies to modulo operations, but fold is inconsistent here |
198 | and simplifies 0 % x to 0, only preserving literal 0 % 0. */ | |
a7f24614 | 199 | (for mod (ceil_mod floor_mod round_mod trunc_mod) |
e0ee10ed RB |
200 | /* 0 % X is always zero. */ |
201 | (simplify | |
a7f24614 | 202 | (mod integer_zerop@0 @1) |
e0ee10ed RB |
203 | /* But not for 0 % 0 so that we can get the proper warnings and errors. */ |
204 | (if (!integer_zerop (@1)) | |
205 | @0)) | |
206 | /* X % 1 is always zero. */ | |
207 | (simplify | |
a7f24614 RB |
208 | (mod @0 integer_onep) |
209 | { build_zero_cst (type); }) | |
210 | /* X % -1 is zero. */ | |
211 | (simplify | |
09240451 MG |
212 | (mod @0 integer_minus_onep@1) |
213 | (if (!TYPE_UNSIGNED (type)) | |
bc4315fb MG |
214 | { build_zero_cst (type); })) |
215 | /* (X % Y) % Y is just X % Y. */ | |
216 | (simplify | |
217 | (mod (mod@2 @0 @1) @1) | |
218 | @2)) | |
a7f24614 RB |
219 | |
220 | /* X % -C is the same as X % C. */ | |
221 | (simplify | |
222 | (trunc_mod @0 INTEGER_CST@1) | |
223 | (if (TYPE_SIGN (type) == SIGNED | |
224 | && !TREE_OVERFLOW (@1) | |
225 | && wi::neg_p (@1) | |
226 | && !TYPE_OVERFLOW_TRAPS (type) | |
227 | /* Avoid this transformation if C is INT_MIN, i.e. C == -C. */ | |
228 | && !sign_bit_p (@1, @1)) | |
229 | (trunc_mod @0 (negate @1)))) | |
e0ee10ed | 230 | |
bc4315fb MG |
231 | /* X % Y is smaller than Y. */ |
232 | (for cmp (lt ge) | |
233 | (simplify | |
234 | (cmp (trunc_mod @0 @1) @1) | |
235 | (if (TYPE_UNSIGNED (TREE_TYPE (@0))) | |
236 | { constant_boolean_node (cmp == LT_EXPR, type); }))) | |
237 | (for cmp (gt le) | |
238 | (simplify | |
239 | (cmp @1 (trunc_mod @0 @1)) | |
240 | (if (TYPE_UNSIGNED (TREE_TYPE (@0))) | |
241 | { constant_boolean_node (cmp == GT_EXPR, type); }))) | |
242 | ||
e0ee10ed RB |
243 | /* x | ~0 -> ~0 */ |
244 | (simplify | |
245 | (bit_ior @0 integer_all_onesp@1) | |
246 | @1) | |
247 | ||
248 | /* x & 0 -> 0 */ | |
249 | (simplify | |
250 | (bit_and @0 integer_zerop@1) | |
251 | @1) | |
252 | ||
253 | /* x ^ x -> 0 */ | |
254 | (simplify | |
255 | (bit_xor @0 @0) | |
256 | { build_zero_cst (type); }) | |
257 | ||
36a60e48 RB |
258 | /* Canonicalize X ^ ~0 to ~X. */ |
259 | (simplify | |
260 | (bit_xor @0 integer_all_onesp@1) | |
261 | (bit_not @0)) | |
262 | ||
263 | /* x & ~0 -> x */ | |
264 | (simplify | |
265 | (bit_and @0 integer_all_onesp) | |
266 | (non_lvalue @0)) | |
267 | ||
268 | /* x & x -> x, x | x -> x */ | |
269 | (for bitop (bit_and bit_ior) | |
270 | (simplify | |
271 | (bitop @0 @0) | |
272 | (non_lvalue @0))) | |
273 | ||
0f770b01 RV |
274 | /* x + (x & 1) -> (x + 1) & ~1 */ |
275 | (simplify | |
276 | (plus:c @0 (bit_and@2 @0 integer_onep@1)) | |
277 | (if (TREE_CODE (@2) != SSA_NAME || has_single_use (@2)) | |
278 | (bit_and (plus @0 @1) (bit_not @1)))) | |
279 | ||
280 | /* x & ~(x & y) -> x & ~y */ | |
281 | /* x | ~(x | y) -> x | ~y */ | |
282 | (for bitop (bit_and bit_ior) | |
283 | (simplify | |
284 | (bitop:c @0 (bit_not (bitop:c@2 @0 @1))) | |
285 | (if (TREE_CODE (@2) != SSA_NAME || has_single_use (@2)) | |
286 | (bitop @0 (bit_not @1))))) | |
287 | ||
f3582e54 RB |
288 | (simplify |
289 | (abs (negate @0)) | |
290 | (abs @0)) | |
291 | (simplify | |
292 | (abs tree_expr_nonnegative_p@0) | |
293 | @0) | |
294 | ||
d4573ffe | 295 | |
5609420f RB |
296 | /* Try to fold (type) X op CST -> (type) (X op ((type-x) CST)) |
297 | when profitable. | |
298 | For bitwise binary operations apply operand conversions to the | |
299 | binary operation result instead of to the operands. This allows | |
300 | to combine successive conversions and bitwise binary operations. | |
301 | We combine the above two cases by using a conditional convert. */ | |
302 | (for bitop (bit_and bit_ior bit_xor) | |
303 | (simplify | |
304 | (bitop (convert @0) (convert? @1)) | |
305 | (if (((TREE_CODE (@1) == INTEGER_CST | |
306 | && INTEGRAL_TYPE_P (TREE_TYPE (@0)) | |
ad6f996c | 307 | && int_fits_type_p (@1, TREE_TYPE (@0))) |
aea417d7 | 308 | || types_match (@0, @1)) |
ad6f996c RB |
309 | /* ??? This transform conflicts with fold-const.c doing |
310 | Convert (T)(x & c) into (T)x & (T)c, if c is an integer | |
311 | constants (if x has signed type, the sign bit cannot be set | |
312 | in c). This folds extension into the BIT_AND_EXPR. | |
313 | Restrict it to GIMPLE to avoid endless recursions. */ | |
314 | && (bitop != BIT_AND_EXPR || GIMPLE) | |
5609420f RB |
315 | && (/* That's a good idea if the conversion widens the operand, thus |
316 | after hoisting the conversion the operation will be narrower. */ | |
317 | TYPE_PRECISION (TREE_TYPE (@0)) < TYPE_PRECISION (type) | |
318 | /* It's also a good idea if the conversion is to a non-integer | |
319 | mode. */ | |
320 | || GET_MODE_CLASS (TYPE_MODE (type)) != MODE_INT | |
321 | /* Or if the precision of TO is not the same as the precision | |
322 | of its mode. */ | |
323 | || TYPE_PRECISION (type) != GET_MODE_PRECISION (TYPE_MODE (type)))) | |
324 | (convert (bitop @0 (convert @1)))))) | |
325 | ||
326 | /* Simplify (A & B) OP0 (C & B) to (A OP0 C) & B. */ | |
327 | (for bitop (bit_and bit_ior bit_xor) | |
328 | (simplify | |
329 | (bitop (bit_and:c @0 @1) (bit_and @2 @1)) | |
330 | (bit_and (bitop @0 @2) @1))) | |
331 | ||
332 | /* (x | CST1) & CST2 -> (x & CST2) | (CST1 & CST2) */ | |
333 | (simplify | |
334 | (bit_and (bit_ior @0 CONSTANT_CLASS_P@1) CONSTANT_CLASS_P@2) | |
335 | (bit_ior (bit_and @0 @2) (bit_and @1 @2))) | |
336 | ||
337 | /* Combine successive equal operations with constants. */ | |
338 | (for bitop (bit_and bit_ior bit_xor) | |
339 | (simplify | |
340 | (bitop (bitop @0 CONSTANT_CLASS_P@1) CONSTANT_CLASS_P@2) | |
341 | (bitop @0 (bitop @1 @2)))) | |
342 | ||
343 | /* Try simple folding for X op !X, and X op X with the help | |
344 | of the truth_valued_p and logical_inverted_value predicates. */ | |
345 | (match truth_valued_p | |
346 | @0 | |
347 | (if (INTEGRAL_TYPE_P (type) && TYPE_PRECISION (type) == 1))) | |
f84e7fd6 | 348 | (for op (tcc_comparison truth_and truth_andif truth_or truth_orif truth_xor) |
5609420f RB |
349 | (match truth_valued_p |
350 | (op @0 @1))) | |
351 | (match truth_valued_p | |
352 | (truth_not @0)) | |
353 | ||
354 | (match (logical_inverted_value @0) | |
355 | (bit_not truth_valued_p@0)) | |
356 | (match (logical_inverted_value @0) | |
09240451 | 357 | (eq @0 integer_zerop)) |
5609420f | 358 | (match (logical_inverted_value @0) |
09240451 | 359 | (ne truth_valued_p@0 integer_truep)) |
5609420f | 360 | (match (logical_inverted_value @0) |
09240451 | 361 | (bit_xor truth_valued_p@0 integer_truep)) |
5609420f RB |
362 | |
363 | /* X & !X -> 0. */ | |
364 | (simplify | |
365 | (bit_and:c @0 (logical_inverted_value @0)) | |
366 | { build_zero_cst (type); }) | |
367 | /* X | !X and X ^ !X -> 1, , if X is truth-valued. */ | |
368 | (for op (bit_ior bit_xor) | |
369 | (simplify | |
370 | (op:c truth_valued_p@0 (logical_inverted_value @0)) | |
f84e7fd6 | 371 | { constant_boolean_node (true, type); })) |
5609420f RB |
372 | |
373 | (for bitop (bit_and bit_ior) | |
374 | rbitop (bit_ior bit_and) | |
375 | /* (x | y) & x -> x */ | |
376 | /* (x & y) | x -> x */ | |
377 | (simplify | |
378 | (bitop:c (rbitop:c @0 @1) @0) | |
379 | @0) | |
380 | /* (~x | y) & x -> x & y */ | |
381 | /* (~x & y) | x -> x | y */ | |
382 | (simplify | |
383 | (bitop:c (rbitop:c (bit_not @0) @1) @0) | |
384 | (bitop @0 @1))) | |
385 | ||
386 | /* If arg1 and arg2 are booleans (or any single bit type) | |
387 | then try to simplify: | |
388 | ||
389 | (~X & Y) -> X < Y | |
390 | (X & ~Y) -> Y < X | |
391 | (~X | Y) -> X <= Y | |
392 | (X | ~Y) -> Y <= X | |
393 | ||
394 | But only do this if our result feeds into a comparison as | |
395 | this transformation is not always a win, particularly on | |
396 | targets with and-not instructions. | |
397 | -> simplify_bitwise_binary_boolean */ | |
398 | (simplify | |
399 | (ne (bit_and:c (bit_not @0) @1) integer_zerop) | |
400 | (if (INTEGRAL_TYPE_P (TREE_TYPE (@1)) | |
401 | && TYPE_PRECISION (TREE_TYPE (@1)) == 1) | |
402 | (lt @0 @1))) | |
403 | (simplify | |
404 | (ne (bit_ior:c (bit_not @0) @1) integer_zerop) | |
405 | (if (INTEGRAL_TYPE_P (TREE_TYPE (@1)) | |
406 | && TYPE_PRECISION (TREE_TYPE (@1)) == 1) | |
407 | (le @0 @1))) | |
408 | ||
5609420f RB |
409 | /* ~~x -> x */ |
410 | (simplify | |
411 | (bit_not (bit_not @0)) | |
412 | @0) | |
413 | ||
f52baa7b MP |
414 | /* (x & ~m) | (y & m) -> ((x ^ y) & m) ^ x */ |
415 | (simplify | |
416 | (bit_ior:c (bit_and:c@3 @0 (bit_not @2)) (bit_and:c@4 @1 @2)) | |
417 | (if ((TREE_CODE (@3) != SSA_NAME || has_single_use (@3)) | |
418 | && (TREE_CODE (@4) != SSA_NAME || has_single_use (@4))) | |
419 | (bit_xor (bit_and (bit_xor @0 @1) @2) @0))) | |
420 | ||
5609420f | 421 | |
a499aac5 RB |
422 | /* Associate (p +p off1) +p off2 as (p +p (off1 + off2)). */ |
423 | (simplify | |
e6121733 RB |
424 | (pointer_plus (pointer_plus@2 @0 @1) @3) |
425 | (if (TREE_CODE (@2) != SSA_NAME || has_single_use (@2)) | |
426 | (pointer_plus @0 (plus @1 @3)))) | |
a499aac5 RB |
427 | |
428 | /* Pattern match | |
429 | tem1 = (long) ptr1; | |
430 | tem2 = (long) ptr2; | |
431 | tem3 = tem2 - tem1; | |
432 | tem4 = (unsigned long) tem3; | |
433 | tem5 = ptr1 + tem4; | |
434 | and produce | |
435 | tem5 = ptr2; */ | |
436 | (simplify | |
437 | (pointer_plus @0 (convert?@2 (minus@3 (convert @1) (convert @0)))) | |
438 | /* Conditionally look through a sign-changing conversion. */ | |
439 | (if (TYPE_PRECISION (TREE_TYPE (@2)) == TYPE_PRECISION (TREE_TYPE (@3)) | |
440 | && ((GIMPLE && useless_type_conversion_p (type, TREE_TYPE (@1))) | |
441 | || (GENERIC && type == TREE_TYPE (@1)))) | |
442 | @1)) | |
443 | ||
444 | /* Pattern match | |
445 | tem = (sizetype) ptr; | |
446 | tem = tem & algn; | |
447 | tem = -tem; | |
448 | ... = ptr p+ tem; | |
449 | and produce the simpler and easier to analyze with respect to alignment | |
450 | ... = ptr & ~algn; */ | |
451 | (simplify | |
452 | (pointer_plus @0 (negate (bit_and (convert @0) INTEGER_CST@1))) | |
453 | (with { tree algn = wide_int_to_tree (TREE_TYPE (@0), wi::bit_not (@1)); } | |
454 | (bit_and @0 { algn; }))) | |
455 | ||
456 | ||
cc7b5acf RB |
457 | /* We can't reassociate at all for saturating types. */ |
458 | (if (!TYPE_SATURATING (type)) | |
459 | ||
460 | /* Contract negates. */ | |
461 | /* A + (-B) -> A - B */ | |
462 | (simplify | |
463 | (plus:c (convert1? @0) (convert2? (negate @1))) | |
464 | /* Apply STRIP_NOPS on @0 and the negate. */ | |
465 | (if (tree_nop_conversion_p (type, TREE_TYPE (@0)) | |
466 | && tree_nop_conversion_p (type, TREE_TYPE (@1)) | |
6a4f0678 | 467 | && !TYPE_OVERFLOW_SANITIZED (type)) |
cc7b5acf RB |
468 | (minus (convert @0) (convert @1)))) |
469 | /* A - (-B) -> A + B */ | |
470 | (simplify | |
471 | (minus (convert1? @0) (convert2? (negate @1))) | |
472 | (if (tree_nop_conversion_p (type, TREE_TYPE (@0)) | |
2f68e8bc | 473 | && tree_nop_conversion_p (type, TREE_TYPE (@1)) |
6a4f0678 | 474 | && !TYPE_OVERFLOW_SANITIZED (type)) |
cc7b5acf RB |
475 | (plus (convert @0) (convert @1)))) |
476 | /* -(-A) -> A */ | |
477 | (simplify | |
478 | (negate (convert? (negate @1))) | |
479 | (if (tree_nop_conversion_p (type, TREE_TYPE (@1)) | |
6a4f0678 | 480 | && !TYPE_OVERFLOW_SANITIZED (type)) |
a0f12cf8 | 481 | (convert @1))) |
cc7b5acf RB |
482 | |
483 | /* We can't reassociate floating-point or fixed-point plus or minus | |
484 | because of saturation to +-Inf. */ | |
485 | (if (!FLOAT_TYPE_P (type) && !FIXED_POINT_TYPE_P (type)) | |
486 | ||
487 | /* Match patterns that allow contracting a plus-minus pair | |
488 | irrespective of overflow issues. */ | |
489 | /* (A +- B) - A -> +- B */ | |
490 | /* (A +- B) -+ B -> A */ | |
491 | /* A - (A +- B) -> -+ B */ | |
492 | /* A +- (B -+ A) -> +- B */ | |
493 | (simplify | |
494 | (minus (plus:c @0 @1) @0) | |
495 | @1) | |
496 | (simplify | |
497 | (minus (minus @0 @1) @0) | |
498 | (negate @1)) | |
499 | (simplify | |
500 | (plus:c (minus @0 @1) @1) | |
501 | @0) | |
502 | (simplify | |
503 | (minus @0 (plus:c @0 @1)) | |
504 | (negate @1)) | |
505 | (simplify | |
506 | (minus @0 (minus @0 @1)) | |
507 | @1) | |
508 | ||
509 | /* (A +- CST) +- CST -> A + CST */ | |
510 | (for outer_op (plus minus) | |
511 | (for inner_op (plus minus) | |
512 | (simplify | |
513 | (outer_op (inner_op @0 CONSTANT_CLASS_P@1) CONSTANT_CLASS_P@2) | |
514 | /* If the constant operation overflows we cannot do the transform | |
515 | as we would introduce undefined overflow, for example | |
516 | with (a - 1) + INT_MIN. */ | |
517 | (with { tree cst = fold_binary (outer_op == inner_op | |
518 | ? PLUS_EXPR : MINUS_EXPR, type, @1, @2); } | |
519 | (if (cst && !TREE_OVERFLOW (cst)) | |
520 | (inner_op @0 { cst; } )))))) | |
521 | ||
522 | /* (CST - A) +- CST -> CST - A */ | |
523 | (for outer_op (plus minus) | |
524 | (simplify | |
525 | (outer_op (minus CONSTANT_CLASS_P@1 @0) CONSTANT_CLASS_P@2) | |
526 | (with { tree cst = fold_binary (outer_op, type, @1, @2); } | |
527 | (if (cst && !TREE_OVERFLOW (cst)) | |
528 | (minus { cst; } @0))))) | |
529 | ||
530 | /* ~A + A -> -1 */ | |
531 | (simplify | |
532 | (plus:c (bit_not @0) @0) | |
533 | (if (!TYPE_OVERFLOW_TRAPS (type)) | |
534 | { build_all_ones_cst (type); })) | |
535 | ||
536 | /* ~A + 1 -> -A */ | |
537 | (simplify | |
e19740ae RB |
538 | (plus (convert? (bit_not @0)) integer_each_onep) |
539 | (if (tree_nop_conversion_p (type, TREE_TYPE (@0))) | |
540 | (negate (convert @0)))) | |
541 | ||
542 | /* -A - 1 -> ~A */ | |
543 | (simplify | |
544 | (minus (convert? (negate @0)) integer_each_onep) | |
545 | (if (!TYPE_OVERFLOW_TRAPS (type) | |
546 | && tree_nop_conversion_p (type, TREE_TYPE (@0))) | |
547 | (bit_not (convert @0)))) | |
548 | ||
549 | /* -1 - A -> ~A */ | |
550 | (simplify | |
551 | (minus integer_all_onesp @0) | |
bc4315fb | 552 | (bit_not @0)) |
cc7b5acf RB |
553 | |
554 | /* (T)(P + A) - (T)P -> (T) A */ | |
555 | (for add (plus pointer_plus) | |
556 | (simplify | |
557 | (minus (convert (add @0 @1)) | |
558 | (convert @0)) | |
09240451 | 559 | (if (element_precision (type) <= element_precision (TREE_TYPE (@1)) |
cc7b5acf RB |
560 | /* For integer types, if A has a smaller type |
561 | than T the result depends on the possible | |
562 | overflow in P + A. | |
563 | E.g. T=size_t, A=(unsigned)429497295, P>0. | |
564 | However, if an overflow in P + A would cause | |
565 | undefined behavior, we can assume that there | |
566 | is no overflow. */ | |
567 | || (INTEGRAL_TYPE_P (TREE_TYPE (@0)) | |
568 | && TYPE_OVERFLOW_UNDEFINED (TREE_TYPE (@0))) | |
569 | /* For pointer types, if the conversion of A to the | |
570 | final type requires a sign- or zero-extension, | |
571 | then we have to punt - it is not defined which | |
572 | one is correct. */ | |
573 | || (POINTER_TYPE_P (TREE_TYPE (@0)) | |
574 | && TREE_CODE (@1) == INTEGER_CST | |
575 | && tree_int_cst_sign_bit (@1) == 0)) | |
576 | (convert @1)))))) | |
577 | ||
578 | ||
a7f24614 RB |
579 | /* Simplifications of MIN_EXPR and MAX_EXPR. */ |
580 | ||
581 | (for minmax (min max) | |
582 | (simplify | |
583 | (minmax @0 @0) | |
584 | @0)) | |
585 | (simplify | |
586 | (min @0 @1) | |
587 | (if (INTEGRAL_TYPE_P (type) | |
588 | && TYPE_MIN_VALUE (type) | |
589 | && operand_equal_p (@1, TYPE_MIN_VALUE (type), OEP_ONLY_CONST)) | |
590 | @1)) | |
591 | (simplify | |
592 | (max @0 @1) | |
593 | (if (INTEGRAL_TYPE_P (type) | |
594 | && TYPE_MAX_VALUE (type) | |
595 | && operand_equal_p (@1, TYPE_MAX_VALUE (type), OEP_ONLY_CONST)) | |
596 | @1)) | |
597 | ||
598 | ||
599 | /* Simplifications of shift and rotates. */ | |
600 | ||
601 | (for rotate (lrotate rrotate) | |
602 | (simplify | |
603 | (rotate integer_all_onesp@0 @1) | |
604 | @0)) | |
605 | ||
606 | /* Optimize -1 >> x for arithmetic right shifts. */ | |
607 | (simplify | |
608 | (rshift integer_all_onesp@0 @1) | |
609 | (if (!TYPE_UNSIGNED (type) | |
610 | && tree_expr_nonnegative_p (@1)) | |
611 | @0)) | |
612 | ||
613 | (for shiftrotate (lrotate rrotate lshift rshift) | |
614 | (simplify | |
615 | (shiftrotate @0 integer_zerop) | |
616 | (non_lvalue @0)) | |
617 | (simplify | |
618 | (shiftrotate integer_zerop@0 @1) | |
619 | @0) | |
620 | /* Prefer vector1 << scalar to vector1 << vector2 | |
621 | if vector2 is uniform. */ | |
622 | (for vec (VECTOR_CST CONSTRUCTOR) | |
623 | (simplify | |
624 | (shiftrotate @0 vec@1) | |
625 | (with { tree tem = uniform_vector_p (@1); } | |
626 | (if (tem) | |
627 | (shiftrotate @0 { tem; })))))) | |
628 | ||
629 | /* Rewrite an LROTATE_EXPR by a constant into an | |
630 | RROTATE_EXPR by a new constant. */ | |
631 | (simplify | |
632 | (lrotate @0 INTEGER_CST@1) | |
633 | (rrotate @0 { fold_binary (MINUS_EXPR, TREE_TYPE (@1), | |
634 | build_int_cst (TREE_TYPE (@1), | |
635 | element_precision (type)), @1); })) | |
636 | ||
01ada710 MP |
637 | /* ((1 << A) & 1) != 0 -> A == 0 |
638 | ((1 << A) & 1) == 0 -> A != 0 */ | |
639 | (for cmp (ne eq) | |
640 | icmp (eq ne) | |
641 | (simplify | |
642 | (cmp (bit_and (lshift integer_onep @0) integer_onep) integer_zerop) | |
643 | (icmp @0 { build_zero_cst (TREE_TYPE (@0)); }))) | |
cc7b5acf | 644 | |
d4573ffe RB |
645 | /* Simplifications of conversions. */ |
646 | ||
647 | /* Basic strip-useless-type-conversions / strip_nops. */ | |
f3582e54 | 648 | (for cvt (convert view_convert float fix_trunc) |
d4573ffe RB |
649 | (simplify |
650 | (cvt @0) | |
651 | (if ((GIMPLE && useless_type_conversion_p (type, TREE_TYPE (@0))) | |
652 | || (GENERIC && type == TREE_TYPE (@0))) | |
653 | @0))) | |
654 | ||
655 | /* Contract view-conversions. */ | |
656 | (simplify | |
657 | (view_convert (view_convert @0)) | |
658 | (view_convert @0)) | |
659 | ||
660 | /* For integral conversions with the same precision or pointer | |
661 | conversions use a NOP_EXPR instead. */ | |
662 | (simplify | |
663 | (view_convert @0) | |
664 | (if ((INTEGRAL_TYPE_P (type) || POINTER_TYPE_P (type)) | |
665 | && (INTEGRAL_TYPE_P (TREE_TYPE (@0)) || POINTER_TYPE_P (TREE_TYPE (@0))) | |
666 | && TYPE_PRECISION (type) == TYPE_PRECISION (TREE_TYPE (@0))) | |
667 | (convert @0))) | |
668 | ||
669 | /* Strip inner integral conversions that do not change precision or size. */ | |
670 | (simplify | |
671 | (view_convert (convert@0 @1)) | |
672 | (if ((INTEGRAL_TYPE_P (TREE_TYPE (@0)) || POINTER_TYPE_P (TREE_TYPE (@0))) | |
673 | && (INTEGRAL_TYPE_P (TREE_TYPE (@1)) || POINTER_TYPE_P (TREE_TYPE (@1))) | |
674 | && (TYPE_PRECISION (TREE_TYPE (@0)) == TYPE_PRECISION (TREE_TYPE (@1))) | |
675 | && (TYPE_SIZE (TREE_TYPE (@0)) == TYPE_SIZE (TREE_TYPE (@1)))) | |
676 | (view_convert @1))) | |
677 | ||
678 | /* Re-association barriers around constants and other re-association | |
679 | barriers can be removed. */ | |
680 | (simplify | |
681 | (paren CONSTANT_CLASS_P@0) | |
682 | @0) | |
683 | (simplify | |
684 | (paren (paren@1 @0)) | |
685 | @1) | |
1e51d0a2 RB |
686 | |
687 | /* Handle cases of two conversions in a row. */ | |
688 | (for ocvt (convert float fix_trunc) | |
689 | (for icvt (convert float) | |
690 | (simplify | |
691 | (ocvt (icvt@1 @0)) | |
692 | (with | |
693 | { | |
694 | tree inside_type = TREE_TYPE (@0); | |
695 | tree inter_type = TREE_TYPE (@1); | |
696 | int inside_int = INTEGRAL_TYPE_P (inside_type); | |
697 | int inside_ptr = POINTER_TYPE_P (inside_type); | |
698 | int inside_float = FLOAT_TYPE_P (inside_type); | |
09240451 | 699 | int inside_vec = VECTOR_TYPE_P (inside_type); |
1e51d0a2 RB |
700 | unsigned int inside_prec = TYPE_PRECISION (inside_type); |
701 | int inside_unsignedp = TYPE_UNSIGNED (inside_type); | |
702 | int inter_int = INTEGRAL_TYPE_P (inter_type); | |
703 | int inter_ptr = POINTER_TYPE_P (inter_type); | |
704 | int inter_float = FLOAT_TYPE_P (inter_type); | |
09240451 | 705 | int inter_vec = VECTOR_TYPE_P (inter_type); |
1e51d0a2 RB |
706 | unsigned int inter_prec = TYPE_PRECISION (inter_type); |
707 | int inter_unsignedp = TYPE_UNSIGNED (inter_type); | |
708 | int final_int = INTEGRAL_TYPE_P (type); | |
709 | int final_ptr = POINTER_TYPE_P (type); | |
710 | int final_float = FLOAT_TYPE_P (type); | |
09240451 | 711 | int final_vec = VECTOR_TYPE_P (type); |
1e51d0a2 RB |
712 | unsigned int final_prec = TYPE_PRECISION (type); |
713 | int final_unsignedp = TYPE_UNSIGNED (type); | |
714 | } | |
715 | /* In addition to the cases of two conversions in a row | |
716 | handled below, if we are converting something to its own | |
717 | type via an object of identical or wider precision, neither | |
718 | conversion is needed. */ | |
719 | (if (((GIMPLE && useless_type_conversion_p (type, inside_type)) | |
720 | || (GENERIC | |
721 | && TYPE_MAIN_VARIANT (type) == TYPE_MAIN_VARIANT (inside_type))) | |
722 | && (((inter_int || inter_ptr) && final_int) | |
723 | || (inter_float && final_float)) | |
724 | && inter_prec >= final_prec) | |
725 | (ocvt @0)) | |
726 | ||
727 | /* Likewise, if the intermediate and initial types are either both | |
728 | float or both integer, we don't need the middle conversion if the | |
729 | former is wider than the latter and doesn't change the signedness | |
730 | (for integers). Avoid this if the final type is a pointer since | |
731 | then we sometimes need the middle conversion. Likewise if the | |
732 | final type has a precision not equal to the size of its mode. */ | |
d51a6714 JJ |
733 | (if (((inter_int && inside_int) || (inter_float && inside_float)) |
734 | && (final_int || final_float) | |
1e51d0a2 | 735 | && inter_prec >= inside_prec |
d51a6714 JJ |
736 | && (inter_float || inter_unsignedp == inside_unsignedp) |
737 | && ! (final_prec != GET_MODE_PRECISION (TYPE_MODE (type)) | |
738 | && TYPE_MODE (type) == TYPE_MODE (inter_type))) | |
1e51d0a2 RB |
739 | (ocvt @0)) |
740 | ||
741 | /* If we have a sign-extension of a zero-extended value, we can | |
742 | replace that by a single zero-extension. Likewise if the | |
743 | final conversion does not change precision we can drop the | |
744 | intermediate conversion. */ | |
745 | (if (inside_int && inter_int && final_int | |
746 | && ((inside_prec < inter_prec && inter_prec < final_prec | |
747 | && inside_unsignedp && !inter_unsignedp) | |
748 | || final_prec == inter_prec)) | |
749 | (ocvt @0)) | |
750 | ||
751 | /* Two conversions in a row are not needed unless: | |
752 | - some conversion is floating-point (overstrict for now), or | |
753 | - some conversion is a vector (overstrict for now), or | |
754 | - the intermediate type is narrower than both initial and | |
755 | final, or | |
756 | - the intermediate type and innermost type differ in signedness, | |
757 | and the outermost type is wider than the intermediate, or | |
758 | - the initial type is a pointer type and the precisions of the | |
759 | intermediate and final types differ, or | |
760 | - the final type is a pointer type and the precisions of the | |
761 | initial and intermediate types differ. */ | |
762 | (if (! inside_float && ! inter_float && ! final_float | |
763 | && ! inside_vec && ! inter_vec && ! final_vec | |
764 | && (inter_prec >= inside_prec || inter_prec >= final_prec) | |
765 | && ! (inside_int && inter_int | |
766 | && inter_unsignedp != inside_unsignedp | |
767 | && inter_prec < final_prec) | |
768 | && ((inter_unsignedp && inter_prec > inside_prec) | |
769 | == (final_unsignedp && final_prec > inter_prec)) | |
770 | && ! (inside_ptr && inter_prec != final_prec) | |
771 | && ! (final_ptr && inside_prec != inter_prec) | |
772 | && ! (final_prec != GET_MODE_PRECISION (TYPE_MODE (type)) | |
773 | && TYPE_MODE (type) == TYPE_MODE (inter_type))) | |
1f00c1b9 RB |
774 | (ocvt @0)) |
775 | ||
776 | /* A truncation to an unsigned type (a zero-extension) should be | |
777 | canonicalized as bitwise and of a mask. */ | |
778 | (if (final_int && inter_int && inside_int | |
779 | && final_prec == inside_prec | |
780 | && final_prec > inter_prec | |
781 | && inter_unsignedp) | |
782 | (convert (bit_and @0 { wide_int_to_tree | |
783 | (inside_type, | |
784 | wi::mask (inter_prec, false, | |
785 | TYPE_PRECISION (inside_type))); }))) | |
786 | ||
787 | /* If we are converting an integer to a floating-point that can | |
788 | represent it exactly and back to an integer, we can skip the | |
789 | floating-point conversion. */ | |
5ba3ae6d RB |
790 | (if (GIMPLE /* PR66211 */ |
791 | && inside_int && inter_float && final_int && | |
1f00c1b9 RB |
792 | (unsigned) significand_size (TYPE_MODE (inter_type)) |
793 | >= inside_prec - !inside_unsignedp) | |
794 | (convert @0)))))) | |
ea2042ba RB |
795 | |
796 | /* If we have a narrowing conversion to an integral type that is fed by a | |
797 | BIT_AND_EXPR, we might be able to remove the BIT_AND_EXPR if it merely | |
798 | masks off bits outside the final type (and nothing else). */ | |
799 | (simplify | |
800 | (convert (bit_and @0 INTEGER_CST@1)) | |
801 | (if (INTEGRAL_TYPE_P (type) | |
802 | && INTEGRAL_TYPE_P (TREE_TYPE (@0)) | |
803 | && TYPE_PRECISION (type) <= TYPE_PRECISION (TREE_TYPE (@0)) | |
804 | && operand_equal_p (@1, build_low_bits_mask (TREE_TYPE (@1), | |
805 | TYPE_PRECISION (type)), 0)) | |
806 | (convert @0))) | |
a25454ea RB |
807 | |
808 | ||
809 | /* (X /[ex] A) * A -> X. */ | |
810 | (simplify | |
811 | (mult (convert? (exact_div @0 @1)) @1) | |
812 | /* Look through a sign-changing conversion. */ | |
813 | (if (TYPE_PRECISION (TREE_TYPE (@0)) == TYPE_PRECISION (type)) | |
814 | (convert @0))) | |
eaeba53a | 815 | |
a7f24614 RB |
816 | /* Canonicalization of binary operations. */ |
817 | ||
818 | /* Convert X + -C into X - C. */ | |
819 | (simplify | |
820 | (plus @0 REAL_CST@1) | |
821 | (if (REAL_VALUE_NEGATIVE (TREE_REAL_CST (@1))) | |
822 | (with { tree tem = fold_unary (NEGATE_EXPR, type, @1); } | |
823 | (if (!TREE_OVERFLOW (tem) || !flag_trapping_math) | |
824 | (minus @0 { tem; }))))) | |
825 | ||
826 | /* Convert x+x into x*2.0. */ | |
827 | (simplify | |
828 | (plus @0 @0) | |
829 | (if (SCALAR_FLOAT_TYPE_P (type)) | |
830 | (mult @0 { build_real (type, dconst2); }))) | |
831 | ||
832 | (simplify | |
833 | (minus integer_zerop @1) | |
834 | (negate @1)) | |
835 | ||
836 | /* (ARG0 - ARG1) is the same as (-ARG1 + ARG0). So check whether | |
837 | ARG0 is zero and X + ARG0 reduces to X, since that would mean | |
838 | (-ARG1 + ARG0) reduces to -ARG1. */ | |
839 | (simplify | |
840 | (minus real_zerop@0 @1) | |
841 | (if (fold_real_zero_addition_p (type, @0, 0)) | |
842 | (negate @1))) | |
843 | ||
844 | /* Transform x * -1 into -x. */ | |
845 | (simplify | |
846 | (mult @0 integer_minus_onep) | |
847 | (negate @0)) | |
eaeba53a RB |
848 | |
849 | /* COMPLEX_EXPR and REALPART/IMAGPART_EXPR cancellations. */ | |
850 | (simplify | |
851 | (complex (realpart @0) (imagpart @0)) | |
852 | @0) | |
853 | (simplify | |
854 | (realpart (complex @0 @1)) | |
855 | @0) | |
856 | (simplify | |
857 | (imagpart (complex @0 @1)) | |
858 | @1) | |
83633539 RB |
859 | |
860 | ||
861 | /* BSWAP simplifications, transforms checked by gcc.dg/builtin-bswap-8.c. */ | |
862 | (for bswap (BUILT_IN_BSWAP16 BUILT_IN_BSWAP32 BUILT_IN_BSWAP64) | |
863 | (simplify | |
864 | (bswap (bswap @0)) | |
865 | @0) | |
866 | (simplify | |
867 | (bswap (bit_not (bswap @0))) | |
868 | (bit_not @0)) | |
869 | (for bitop (bit_xor bit_ior bit_and) | |
870 | (simplify | |
871 | (bswap (bitop:c (bswap @0) @1)) | |
872 | (bitop @0 (bswap @1))))) | |
96994de0 RB |
873 | |
874 | ||
875 | /* Combine COND_EXPRs and VEC_COND_EXPRs. */ | |
876 | ||
877 | /* Simplify constant conditions. | |
878 | Only optimize constant conditions when the selected branch | |
879 | has the same type as the COND_EXPR. This avoids optimizing | |
880 | away "c ? x : throw", where the throw has a void type. | |
881 | Note that we cannot throw away the fold-const.c variant nor | |
882 | this one as we depend on doing this transform before possibly | |
883 | A ? B : B -> B triggers and the fold-const.c one can optimize | |
884 | 0 ? A : B to B even if A has side-effects. Something | |
885 | genmatch cannot handle. */ | |
886 | (simplify | |
887 | (cond INTEGER_CST@0 @1 @2) | |
888 | (if (integer_zerop (@0) | |
889 | && (!VOID_TYPE_P (TREE_TYPE (@2)) | |
890 | || VOID_TYPE_P (type))) | |
891 | @2) | |
892 | (if (!integer_zerop (@0) | |
893 | && (!VOID_TYPE_P (TREE_TYPE (@1)) | |
894 | || VOID_TYPE_P (type))) | |
895 | @1)) | |
896 | (simplify | |
897 | (vec_cond VECTOR_CST@0 @1 @2) | |
898 | (if (integer_all_onesp (@0)) | |
899 | @1) | |
900 | (if (integer_zerop (@0)) | |
901 | @2)) | |
902 | ||
903 | (for cnd (cond vec_cond) | |
904 | /* A ? B : (A ? X : C) -> A ? B : C. */ | |
905 | (simplify | |
906 | (cnd @0 (cnd @0 @1 @2) @3) | |
907 | (cnd @0 @1 @3)) | |
908 | (simplify | |
909 | (cnd @0 @1 (cnd @0 @2 @3)) | |
910 | (cnd @0 @1 @3)) | |
911 | ||
912 | /* A ? B : B -> B. */ | |
913 | (simplify | |
914 | (cnd @0 @1 @1) | |
09240451 | 915 | @1) |
96994de0 | 916 | |
09240451 MG |
917 | /* !A ? B : C -> A ? C : B. */ |
918 | (simplify | |
919 | (cnd (logical_inverted_value truth_valued_p@0) @1 @2) | |
920 | (cnd @0 @2 @1))) | |
f84e7fd6 RB |
921 | |
922 | ||
923 | /* Simplifications of comparisons. */ | |
924 | ||
925 | /* We can simplify a logical negation of a comparison to the | |
926 | inverted comparison. As we cannot compute an expression | |
927 | operator using invert_tree_comparison we have to simulate | |
928 | that with expression code iteration. */ | |
929 | (for cmp (tcc_comparison) | |
930 | icmp (inverted_tcc_comparison) | |
931 | ncmp (inverted_tcc_comparison_with_nans) | |
932 | /* Ideally we'd like to combine the following two patterns | |
933 | and handle some more cases by using | |
934 | (logical_inverted_value (cmp @0 @1)) | |
935 | here but for that genmatch would need to "inline" that. | |
936 | For now implement what forward_propagate_comparison did. */ | |
937 | (simplify | |
938 | (bit_not (cmp @0 @1)) | |
939 | (if (VECTOR_TYPE_P (type) | |
940 | || (INTEGRAL_TYPE_P (type) && TYPE_PRECISION (type) == 1)) | |
941 | /* Comparison inversion may be impossible for trapping math, | |
942 | invert_tree_comparison will tell us. But we can't use | |
943 | a computed operator in the replacement tree thus we have | |
944 | to play the trick below. */ | |
945 | (with { enum tree_code ic = invert_tree_comparison | |
1b457aa4 | 946 | (cmp, HONOR_NANS (@0)); } |
f84e7fd6 RB |
947 | (if (ic == icmp) |
948 | (icmp @0 @1)) | |
949 | (if (ic == ncmp) | |
950 | (ncmp @0 @1))))) | |
951 | (simplify | |
09240451 MG |
952 | (bit_xor (cmp @0 @1) integer_truep) |
953 | (with { enum tree_code ic = invert_tree_comparison | |
1b457aa4 | 954 | (cmp, HONOR_NANS (@0)); } |
09240451 MG |
955 | (if (ic == icmp) |
956 | (icmp @0 @1)) | |
957 | (if (ic == ncmp) | |
958 | (ncmp @0 @1))))) | |
e18c1d66 | 959 | |
cfdc4f33 MG |
960 | /* Unordered tests if either argument is a NaN. */ |
961 | (simplify | |
962 | (bit_ior (unordered @0 @0) (unordered @1 @1)) | |
aea417d7 | 963 | (if (types_match (@0, @1)) |
cfdc4f33 MG |
964 | (unordered @0 @1))) |
965 | (simplify | |
966 | (bit_ior:c (unordered @0 @0) (unordered:c@2 @0 @1)) | |
967 | @2) | |
e18c1d66 RB |
968 | |
969 | /* Simplification of math builtins. */ | |
970 | ||
971 | (define_operator_list LOG BUILT_IN_LOGF BUILT_IN_LOG BUILT_IN_LOGL) | |
972 | (define_operator_list EXP BUILT_IN_EXPF BUILT_IN_EXP BUILT_IN_EXPL) | |
973 | (define_operator_list LOG2 BUILT_IN_LOG2F BUILT_IN_LOG2 BUILT_IN_LOG2L) | |
974 | (define_operator_list EXP2 BUILT_IN_EXP2F BUILT_IN_EXP2 BUILT_IN_EXP2L) | |
975 | (define_operator_list LOG10 BUILT_IN_LOG10F BUILT_IN_LOG10 BUILT_IN_LOG10L) | |
976 | (define_operator_list EXP10 BUILT_IN_EXP10F BUILT_IN_EXP10 BUILT_IN_EXP10L) | |
977 | (define_operator_list POW BUILT_IN_POWF BUILT_IN_POW BUILT_IN_POWL) | |
978 | (define_operator_list POW10 BUILT_IN_POW10F BUILT_IN_POW10 BUILT_IN_POW10L) | |
979 | (define_operator_list SQRT BUILT_IN_SQRTF BUILT_IN_SQRT BUILT_IN_SQRTL) | |
980 | (define_operator_list CBRT BUILT_IN_CBRTF BUILT_IN_CBRT BUILT_IN_CBRTL) | |
981 | ||
982 | ||
983 | /* fold_builtin_logarithm */ | |
984 | (if (flag_unsafe_math_optimizations) | |
985 | /* Special case, optimize logN(expN(x)) = x. */ | |
986 | (for logs (LOG LOG2 LOG10) | |
987 | exps (EXP EXP2 EXP10) | |
988 | (simplify | |
989 | (logs (exps @0)) | |
990 | @0)) | |
991 | /* Optimize logN(func()) for various exponential functions. We | |
992 | want to determine the value "x" and the power "exponent" in | |
993 | order to transform logN(x**exponent) into exponent*logN(x). */ | |
994 | (for logs (LOG LOG LOG LOG | |
995 | LOG2 LOG2 LOG2 LOG2 | |
996 | LOG10 LOG10 LOG10 LOG10) | |
997 | exps (EXP EXP2 EXP10 POW10) | |
998 | (simplify | |
999 | (logs (exps @0)) | |
1000 | (with { | |
1001 | tree x; | |
1002 | switch (exps) | |
1003 | { | |
1004 | CASE_FLT_FN (BUILT_IN_EXP): | |
1005 | /* Prepare to do logN(exp(exponent) -> exponent*logN(e). */ | |
1006 | x = build_real (type, real_value_truncate (TYPE_MODE (type), | |
1007 | dconst_e ())); | |
1008 | break; | |
1009 | CASE_FLT_FN (BUILT_IN_EXP2): | |
1010 | /* Prepare to do logN(exp2(exponent) -> exponent*logN(2). */ | |
1011 | x = build_real (type, dconst2); | |
1012 | break; | |
1013 | CASE_FLT_FN (BUILT_IN_EXP10): | |
1014 | CASE_FLT_FN (BUILT_IN_POW10): | |
1015 | /* Prepare to do logN(exp10(exponent) -> exponent*logN(10). */ | |
1016 | { | |
1017 | REAL_VALUE_TYPE dconst10; | |
1018 | real_from_integer (&dconst10, VOIDmode, 10, SIGNED); | |
1019 | x = build_real (type, dconst10); | |
1020 | } | |
1021 | break; | |
1022 | } | |
1023 | } | |
1024 | (mult (logs { x; }) @0)))) | |
1025 | (for logs (LOG LOG | |
1026 | LOG2 LOG2 | |
1027 | LOG10 LOG10) | |
1028 | exps (SQRT CBRT) | |
1029 | (simplify | |
1030 | (logs (exps @0)) | |
1031 | (with { | |
1032 | tree x; | |
1033 | switch (exps) | |
1034 | { | |
1035 | CASE_FLT_FN (BUILT_IN_SQRT): | |
1036 | /* Prepare to do logN(sqrt(x) -> 0.5*logN(x). */ | |
1037 | x = build_real (type, dconsthalf); | |
1038 | break; | |
1039 | CASE_FLT_FN (BUILT_IN_CBRT): | |
1040 | /* Prepare to do logN(cbrt(x) -> (1/3)*logN(x). */ | |
1041 | x = build_real (type, real_value_truncate (TYPE_MODE (type), | |
1042 | dconst_third ())); | |
1043 | break; | |
1044 | } | |
1045 | } | |
1046 | (mult { x; } (logs @0))))) | |
1047 | /* logN(pow(x,exponent) -> exponent*logN(x). */ | |
1048 | (for logs (LOG LOG2 LOG10) | |
1049 | pows (POW) | |
1050 | (simplify | |
1051 | (logs (pows @0 @1)) | |
1052 | (mult @1 (logs @0))))) | |
1053 | ||
be144838 JL |
1054 | /* Narrowing of arithmetic and logical operations. |
1055 | ||
1056 | These are conceptually similar to the transformations performed for | |
1057 | the C/C++ front-ends by shorten_binary_op and shorten_compare. Long | |
1058 | term we want to move all that code out of the front-ends into here. */ | |
1059 | ||
1060 | /* If we have a narrowing conversion of an arithmetic operation where | |
1061 | both operands are widening conversions from the same type as the outer | |
1062 | narrowing conversion. Then convert the innermost operands to a suitable | |
1063 | unsigned type (to avoid introducing undefined behaviour), perform the | |
1064 | operation and convert the result to the desired type. */ | |
1065 | (for op (plus minus) | |
1066 | (simplify | |
48451e8f | 1067 | (convert (op@4 (convert@2 @0) (convert@3 @1))) |
be144838 JL |
1068 | (if (INTEGRAL_TYPE_P (type) |
1069 | /* We check for type compatibility between @0 and @1 below, | |
1070 | so there's no need to check that @1/@3 are integral types. */ | |
1071 | && INTEGRAL_TYPE_P (TREE_TYPE (@0)) | |
1072 | && INTEGRAL_TYPE_P (TREE_TYPE (@2)) | |
1073 | /* The precision of the type of each operand must match the | |
1074 | precision of the mode of each operand, similarly for the | |
1075 | result. */ | |
1076 | && (TYPE_PRECISION (TREE_TYPE (@0)) | |
1077 | == GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (@0)))) | |
1078 | && (TYPE_PRECISION (TREE_TYPE (@1)) | |
1079 | == GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (@1)))) | |
1080 | && TYPE_PRECISION (type) == GET_MODE_PRECISION (TYPE_MODE (type)) | |
1081 | /* The inner conversion must be a widening conversion. */ | |
1082 | && TYPE_PRECISION (TREE_TYPE (@2)) > TYPE_PRECISION (TREE_TYPE (@0)) | |
aea417d7 MG |
1083 | && types_match (@0, @1) |
1084 | && types_match (@0, type) | |
48451e8f | 1085 | && single_use (@4)) |
be144838 JL |
1086 | (if (TYPE_OVERFLOW_WRAPS (TREE_TYPE (@0))) |
1087 | (convert (op @0 @1))) | |
1088 | (with { tree utype = unsigned_type_for (TREE_TYPE (@0)); } | |
1089 | (convert (op (convert:utype @0) (convert:utype @1))))))) | |
48451e8f JL |
1090 | |
1091 | /* This is another case of narrowing, specifically when there's an outer | |
1092 | BIT_AND_EXPR which masks off bits outside the type of the innermost | |
1093 | operands. Like the previous case we have to convert the operands | |
1094 | to unsigned types to avoid introducing undefined behaviour for the | |
1095 | arithmetic operation. */ | |
1096 | (for op (minus plus) | |
1097 | (simplify | |
1098 | (bit_and (op@5 (convert@2 @0) (convert@3 @1)) INTEGER_CST@4) | |
1099 | (if (INTEGRAL_TYPE_P (type) | |
1100 | /* We check for type compatibility between @0 and @1 below, | |
1101 | so there's no need to check that @1/@3 are integral types. */ | |
1102 | && INTEGRAL_TYPE_P (TREE_TYPE (@0)) | |
1103 | && INTEGRAL_TYPE_P (TREE_TYPE (@2)) | |
1104 | /* The precision of the type of each operand must match the | |
1105 | precision of the mode of each operand, similarly for the | |
1106 | result. */ | |
1107 | && (TYPE_PRECISION (TREE_TYPE (@0)) | |
1108 | == GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (@0)))) | |
1109 | && (TYPE_PRECISION (TREE_TYPE (@1)) | |
1110 | == GET_MODE_PRECISION (TYPE_MODE (TREE_TYPE (@1)))) | |
1111 | && TYPE_PRECISION (type) == GET_MODE_PRECISION (TYPE_MODE (type)) | |
1112 | /* The inner conversion must be a widening conversion. */ | |
1113 | && TYPE_PRECISION (TREE_TYPE (@2)) > TYPE_PRECISION (TREE_TYPE (@0)) | |
aea417d7 | 1114 | && types_match (@0, @1) |
a60c51fe | 1115 | && (tree_int_cst_min_precision (@4, TYPE_SIGN (TREE_TYPE (@0))) |
48451e8f | 1116 | <= TYPE_PRECISION (TREE_TYPE (@0))) |
a60c51fe JJ |
1117 | && (TYPE_OVERFLOW_WRAPS (TREE_TYPE (@0)) |
1118 | || tree_int_cst_sgn (@4) >= 0) | |
48451e8f JL |
1119 | && single_use (@5)) |
1120 | (if (TYPE_OVERFLOW_WRAPS (TREE_TYPE (@0))) | |
1121 | (with { tree ntype = TREE_TYPE (@0); } | |
1122 | (convert (bit_and (op @0 @1) (convert:ntype @4))))) | |
1123 | (with { tree utype = unsigned_type_for (TREE_TYPE (@0)); } | |
1124 | (convert (bit_and (op (convert:utype @0) (convert:utype @1)) | |
1125 | (convert:utype @4))))))) | |
1126 |