]>
Commit | Line | Data |
---|---|---|
e6a23add | 1 | /* Conditional Dead Call Elimination pass for the GNU compiler. |
d353bf18 | 2 | Copyright (C) 2008-2015 Free Software Foundation, Inc. |
e6a23add | 3 | Contributed by Xinliang David Li <davidxl@google.com> |
4 | ||
5 | This file is part of GCC. | |
48e1416a | 6 | |
e6a23add | 7 | GCC is free software; you can redistribute it and/or modify it |
8 | under the terms of the GNU General Public License as published by the | |
9 | Free Software Foundation; either version 3, or (at your option) any | |
10 | later version. | |
48e1416a | 11 | |
e6a23add | 12 | GCC is distributed in the hope that it will be useful, but WITHOUT |
13 | ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
14 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
15 | for more details. | |
48e1416a | 16 | |
e6a23add | 17 | You should have received a copy of the GNU General Public License |
18 | along with GCC; see the file COPYING3. If not see | |
19 | <http://www.gnu.org/licenses/>. */ | |
20 | ||
21 | #include "config.h" | |
22 | #include "system.h" | |
23 | #include "coretypes.h" | |
9ef16211 | 24 | #include "backend.h" |
7c29e30e | 25 | #include "hard-reg-set.h" |
9ef16211 | 26 | #include "tree.h" |
27 | #include "gimple.h" | |
7c29e30e | 28 | #include "cfghooks.h" |
29 | #include "tree-pass.h" | |
9ef16211 | 30 | #include "ssa.h" |
7c29e30e | 31 | #include "gimple-pretty-print.h" |
b20a8bb4 | 32 | #include "alias.h" |
b20a8bb4 | 33 | #include "fold-const.h" |
9ed99284 | 34 | #include "stor-layout.h" |
bc61cadb | 35 | #include "internal-fn.h" |
dcf1a1ec | 36 | #include "gimple-iterator.h" |
073c1fd5 | 37 | #include "tree-cfg.h" |
073c1fd5 | 38 | #include "tree-into-ssa.h" |
e6a23add | 39 | #include "flags.h" |
40 | \f | |
41 | ||
42 | /* Conditional dead call elimination | |
43 | ||
44 | Some builtin functions can set errno on error conditions, but they | |
45 | are otherwise pure. If the result of a call to such a function is | |
46 | not used, the compiler can still not eliminate the call without | |
47 | powerful interprocedural analysis to prove that the errno is not | |
48 | checked. However, if the conditions under which the error occurs | |
48e1416a | 49 | are known, the compiler can conditionally dead code eliminate the |
e6a23add | 50 | calls by shrink-wrapping the semi-dead calls into the error condition: |
51 | ||
52 | built_in_call (args) | |
53 | ==> | |
54 | if (error_cond (args)) | |
55 | built_in_call (args) | |
56 | ||
57 | An actual simple example is : | |
58 | log (x); // Mostly dead call | |
59 | ==> | |
1f5fe264 | 60 | if (x <= 0) |
e6a23add | 61 | log (x); |
62 | With this change, call to log (x) is effectively eliminated, as | |
63 | in majority of the cases, log won't be called with x out of | |
64 | range. The branch is totally predictable, so the branch cost | |
48e1416a | 65 | is low. |
e6a23add | 66 | |
67 | Note that library functions are not supposed to clear errno to zero without | |
68 | error. See IEEE Std 1003.1, section 2.3 Error Numbers, and section 7.5:3 of | |
69 | ISO/IEC 9899 (C99). | |
70 | ||
71 | The condition wrapping the builtin call is conservatively set to avoid too | |
72 | aggressive (wrong) shrink wrapping. The optimization is called conditional | |
73 | dead call elimination because the call is eliminated under the condition | |
74 | that the input arguments would not lead to domain or range error (for | |
75 | instance when x <= 0 for a log (x) call), however the chances that the error | |
76 | condition is hit is very low (those builtin calls which are conditionally | |
77 | dead are usually part of the C++ abstraction penalty exposed after | |
78 | inlining). */ | |
79 | ||
80 | ||
48e1416a | 81 | /* A structure for representing input domain of |
e6a23add | 82 | a function argument in integer. If the lower |
48e1416a | 83 | bound is -inf, has_lb is set to false. If the |
84 | upper bound is +inf, has_ub is false. | |
85 | is_lb_inclusive and is_ub_inclusive are flags | |
86 | to indicate if lb and ub value are inclusive | |
e6a23add | 87 | respectively. */ |
88 | ||
6dc50383 | 89 | struct inp_domain |
e6a23add | 90 | { |
91 | int lb; | |
92 | int ub; | |
93 | bool has_lb; | |
94 | bool has_ub; | |
95 | bool is_lb_inclusive; | |
96 | bool is_ub_inclusive; | |
6dc50383 | 97 | }; |
e6a23add | 98 | |
e6a23add | 99 | /* A helper function to construct and return an input |
48e1416a | 100 | domain object. LB is the lower bound, HAS_LB is |
e6a23add | 101 | a boolean flag indicating if the lower bound exists, |
102 | and LB_INCLUSIVE is a boolean flag indicating if the | |
103 | lower bound is inclusive or not. UB, HAS_UB, and | |
48e1416a | 104 | UB_INCLUSIVE have the same meaning, but for upper |
e6a23add | 105 | bound of the domain. */ |
106 | ||
107 | static inp_domain | |
108 | get_domain (int lb, bool has_lb, bool lb_inclusive, | |
109 | int ub, bool has_ub, bool ub_inclusive) | |
110 | { | |
111 | inp_domain domain; | |
112 | domain.lb = lb; | |
113 | domain.has_lb = has_lb; | |
114 | domain.is_lb_inclusive = lb_inclusive; | |
115 | domain.ub = ub; | |
116 | domain.has_ub = has_ub; | |
117 | domain.is_ub_inclusive = ub_inclusive; | |
118 | return domain; | |
119 | } | |
120 | ||
48e1416a | 121 | /* A helper function to check the target format for the |
e6a23add | 122 | argument type. In this implementation, only IEEE formats |
48e1416a | 123 | are supported. ARG is the call argument to be checked. |
e6a23add | 124 | Returns true if the format is supported. To support other |
125 | target formats, function get_no_error_domain needs to be | |
48e1416a | 126 | enhanced to have range bounds properly computed. Since |
127 | the check is cheap (very small number of candidates | |
e6a23add | 128 | to be checked), the result is not cached for each float type. */ |
129 | ||
130 | static bool | |
131 | check_target_format (tree arg) | |
132 | { | |
133 | tree type; | |
3754d046 | 134 | machine_mode mode; |
e6a23add | 135 | const struct real_format *rfmt; |
48e1416a | 136 | |
e6a23add | 137 | type = TREE_TYPE (arg); |
138 | mode = TYPE_MODE (type); | |
139 | rfmt = REAL_MODE_FORMAT (mode); | |
defc07a6 | 140 | if ((mode == SFmode |
b161ca01 | 141 | && (rfmt == &ieee_single_format || rfmt == &mips_single_format |
142 | || rfmt == &motorola_single_format)) | |
defc07a6 | 143 | || (mode == DFmode |
b161ca01 | 144 | && (rfmt == &ieee_double_format || rfmt == &mips_double_format |
145 | || rfmt == &motorola_double_format)) | |
e6a23add | 146 | /* For long double, we can not really check XFmode |
48e1416a | 147 | which is only defined on intel platforms. |
148 | Candidate pre-selection using builtin function | |
149 | code guarantees that we are checking formats | |
e6a23add | 150 | for long double modes: double, quad, and extended. */ |
48e1416a | 151 | || (mode != SFmode && mode != DFmode |
e6a23add | 152 | && (rfmt == &ieee_quad_format |
defc07a6 | 153 | || rfmt == &mips_quad_format |
b161ca01 | 154 | || rfmt == &ieee_extended_motorola_format |
48e1416a | 155 | || rfmt == &ieee_extended_intel_96_format |
156 | || rfmt == &ieee_extended_intel_128_format | |
e6a23add | 157 | || rfmt == &ieee_extended_intel_96_round_53_format))) |
158 | return true; | |
159 | ||
160 | return false; | |
161 | } | |
162 | ||
163 | \f | |
164 | /* A helper function to help select calls to pow that are suitable for | |
165 | conditional DCE transformation. It looks for pow calls that can be | |
166 | guided with simple conditions. Such calls either have constant base | |
48e1416a | 167 | values or base values converted from integers. Returns true if |
e6a23add | 168 | the pow call POW_CALL is a candidate. */ |
169 | ||
170 | /* The maximum integer bit size for base argument of a pow call | |
171 | that is suitable for shrink-wrapping transformation. */ | |
172 | #define MAX_BASE_INT_BIT_SIZE 32 | |
173 | ||
174 | static bool | |
1a91d914 | 175 | check_pow (gcall *pow_call) |
e6a23add | 176 | { |
177 | tree base, expn; | |
178 | enum tree_code bc, ec; | |
179 | ||
75a70cf9 | 180 | if (gimple_call_num_args (pow_call) != 2) |
e6a23add | 181 | return false; |
182 | ||
75a70cf9 | 183 | base = gimple_call_arg (pow_call, 0); |
184 | expn = gimple_call_arg (pow_call, 1); | |
e6a23add | 185 | |
186 | if (!check_target_format (expn)) | |
187 | return false; | |
188 | ||
189 | bc = TREE_CODE (base); | |
190 | ec = TREE_CODE (expn); | |
191 | ||
192 | /* Folding candidates are not interesting. | |
193 | Can actually assert that it is already folded. */ | |
194 | if (ec == REAL_CST && bc == REAL_CST) | |
195 | return false; | |
196 | ||
197 | if (bc == REAL_CST) | |
198 | { | |
199 | /* Only handle a fixed range of constant. */ | |
200 | REAL_VALUE_TYPE mv; | |
201 | REAL_VALUE_TYPE bcv = TREE_REAL_CST (base); | |
20cb53c9 | 202 | if (real_equal (&bcv, &dconst1)) |
e6a23add | 203 | return false; |
1b67971e | 204 | if (real_less (&bcv, &dconst1)) |
e6a23add | 205 | return false; |
e913b5cd | 206 | real_from_integer (&mv, TYPE_MODE (TREE_TYPE (base)), 256, UNSIGNED); |
1b67971e | 207 | if (real_less (&mv, &bcv)) |
e6a23add | 208 | return false; |
209 | return true; | |
210 | } | |
211 | else if (bc == SSA_NAME) | |
212 | { | |
7ecda5e8 | 213 | tree base_val0, type; |
42acab1c | 214 | gimple *base_def; |
e6a23add | 215 | int bit_sz; |
216 | ||
217 | /* Only handles cases where base value is converted | |
48e1416a | 218 | from integer values. */ |
e6a23add | 219 | base_def = SSA_NAME_DEF_STMT (base); |
75a70cf9 | 220 | if (gimple_code (base_def) != GIMPLE_ASSIGN) |
e6a23add | 221 | return false; |
222 | ||
75a70cf9 | 223 | if (gimple_assign_rhs_code (base_def) != FLOAT_EXPR) |
e6a23add | 224 | return false; |
75a70cf9 | 225 | base_val0 = gimple_assign_rhs1 (base_def); |
e6a23add | 226 | |
7ecda5e8 | 227 | type = TREE_TYPE (base_val0); |
e6a23add | 228 | if (TREE_CODE (type) != INTEGER_TYPE) |
229 | return false; | |
230 | bit_sz = TYPE_PRECISION (type); | |
231 | /* If the type of the base is too wide, | |
232 | the resulting shrink wrapping condition | |
233 | will be too conservative. */ | |
234 | if (bit_sz > MAX_BASE_INT_BIT_SIZE) | |
235 | return false; | |
236 | ||
237 | return true; | |
238 | } | |
239 | else | |
240 | return false; | |
241 | } | |
242 | ||
243 | /* A helper function to help select candidate function calls that are | |
244 | suitable for conditional DCE. Candidate functions must have single | |
245 | valid input domain in this implementation except for pow (see check_pow). | |
246 | Returns true if the function call is a candidate. */ | |
247 | ||
248 | static bool | |
1a91d914 | 249 | check_builtin_call (gcall *bcall) |
e6a23add | 250 | { |
251 | tree arg; | |
252 | ||
75a70cf9 | 253 | arg = gimple_call_arg (bcall, 0); |
e6a23add | 254 | return check_target_format (arg); |
255 | } | |
256 | ||
257 | /* A helper function to determine if a builtin function call is a | |
258 | candidate for conditional DCE. Returns true if the builtin call | |
259 | is a candidate. */ | |
260 | ||
261 | static bool | |
1a91d914 | 262 | is_call_dce_candidate (gcall *call) |
e6a23add | 263 | { |
264 | tree fn; | |
265 | enum built_in_function fnc; | |
266 | ||
75a70cf9 | 267 | /* Only potentially dead calls are considered. */ |
268 | if (gimple_call_lhs (call)) | |
e6a23add | 269 | return false; |
270 | ||
75a70cf9 | 271 | fn = gimple_call_fndecl (call); |
272 | if (!fn | |
48e1416a | 273 | || !DECL_BUILT_IN (fn) |
e6a23add | 274 | || (DECL_BUILT_IN_CLASS (fn) != BUILT_IN_NORMAL)) |
275 | return false; | |
276 | ||
277 | fnc = DECL_FUNCTION_CODE (fn); | |
278 | switch (fnc) | |
279 | { | |
280 | /* Trig functions. */ | |
281 | CASE_FLT_FN (BUILT_IN_ACOS): | |
282 | CASE_FLT_FN (BUILT_IN_ASIN): | |
283 | /* Hyperbolic functions. */ | |
284 | CASE_FLT_FN (BUILT_IN_ACOSH): | |
285 | CASE_FLT_FN (BUILT_IN_ATANH): | |
286 | CASE_FLT_FN (BUILT_IN_COSH): | |
287 | CASE_FLT_FN (BUILT_IN_SINH): | |
288 | /* Log functions. */ | |
289 | CASE_FLT_FN (BUILT_IN_LOG): | |
290 | CASE_FLT_FN (BUILT_IN_LOG2): | |
291 | CASE_FLT_FN (BUILT_IN_LOG10): | |
292 | CASE_FLT_FN (BUILT_IN_LOG1P): | |
293 | /* Exp functions. */ | |
294 | CASE_FLT_FN (BUILT_IN_EXP): | |
295 | CASE_FLT_FN (BUILT_IN_EXP2): | |
296 | CASE_FLT_FN (BUILT_IN_EXP10): | |
297 | CASE_FLT_FN (BUILT_IN_EXPM1): | |
298 | CASE_FLT_FN (BUILT_IN_POW10): | |
299 | /* Sqrt. */ | |
300 | CASE_FLT_FN (BUILT_IN_SQRT): | |
301 | return check_builtin_call (call); | |
302 | /* Special one: two argument pow. */ | |
303 | case BUILT_IN_POW: | |
304 | return check_pow (call); | |
305 | default: | |
306 | break; | |
307 | } | |
308 | ||
309 | return false; | |
310 | } | |
311 | ||
312 | \f | |
313 | /* A helper function to generate gimple statements for | |
314 | one bound comparison. ARG is the call argument to | |
315 | be compared with the bound, LBUB is the bound value | |
316 | in integer, TCODE is the tree_code of the comparison, | |
317 | TEMP_NAME1/TEMP_NAME2 are names of the temporaries, | |
318 | CONDS is a vector holding the produced GIMPLE statements, | |
319 | and NCONDS points to the variable holding the number | |
48e1416a | 320 | of logical comparisons. CONDS is either empty or |
e6a23add | 321 | a list ended with a null tree. */ |
322 | ||
323 | static void | |
48e1416a | 324 | gen_one_condition (tree arg, int lbub, |
e6a23add | 325 | enum tree_code tcode, |
326 | const char *temp_name1, | |
75a70cf9 | 327 | const char *temp_name2, |
42acab1c | 328 | vec<gimple *> conds, |
e6a23add | 329 | unsigned *nconds) |
330 | { | |
331 | tree lbub_real_cst, lbub_cst, float_type; | |
332 | tree temp, tempn, tempc, tempcn; | |
1a91d914 | 333 | gassign *stmt1; |
334 | gassign *stmt2; | |
335 | gcond *stmt3; | |
e6a23add | 336 | |
337 | float_type = TREE_TYPE (arg); | |
338 | lbub_cst = build_int_cst (integer_type_node, lbub); | |
339 | lbub_real_cst = build_real_from_int_cst (float_type, lbub_cst); | |
340 | ||
341 | temp = create_tmp_var (float_type, temp_name1); | |
75a70cf9 | 342 | stmt1 = gimple_build_assign (temp, arg); |
e6a23add | 343 | tempn = make_ssa_name (temp, stmt1); |
75a70cf9 | 344 | gimple_assign_set_lhs (stmt1, tempn); |
e6a23add | 345 | |
346 | tempc = create_tmp_var (boolean_type_node, temp_name2); | |
75a70cf9 | 347 | stmt2 = gimple_build_assign (tempc, |
348 | fold_build2 (tcode, | |
349 | boolean_type_node, | |
350 | tempn, lbub_real_cst)); | |
e6a23add | 351 | tempcn = make_ssa_name (tempc, stmt2); |
75a70cf9 | 352 | gimple_assign_set_lhs (stmt2, tempcn); |
353 | ||
354 | stmt3 = gimple_build_cond_from_tree (tempcn, NULL_TREE, NULL_TREE); | |
f1f41a6c | 355 | conds.quick_push (stmt1); |
356 | conds.quick_push (stmt2); | |
357 | conds.quick_push (stmt3); | |
e6a23add | 358 | (*nconds)++; |
359 | } | |
360 | ||
361 | /* A helper function to generate GIMPLE statements for | |
362 | out of input domain check. ARG is the call argument | |
363 | to be runtime checked, DOMAIN holds the valid domain | |
364 | for the given function, CONDS points to the vector | |
48e1416a | 365 | holding the result GIMPLE statements. *NCONDS is |
366 | the number of logical comparisons. This function | |
e6a23add | 367 | produces no more than two logical comparisons, one |
368 | for lower bound check, one for upper bound check. */ | |
369 | ||
370 | static void | |
371 | gen_conditions_for_domain (tree arg, inp_domain domain, | |
42acab1c | 372 | vec<gimple *> conds, |
e6a23add | 373 | unsigned *nconds) |
374 | { | |
375 | if (domain.has_lb) | |
376 | gen_one_condition (arg, domain.lb, | |
377 | (domain.is_lb_inclusive | |
378 | ? LT_EXPR : LE_EXPR), | |
379 | "DCE_COND_LB", "DCE_COND_LB_TEST", | |
380 | conds, nconds); | |
381 | ||
382 | if (domain.has_ub) | |
383 | { | |
384 | /* Now push a separator. */ | |
385 | if (domain.has_lb) | |
f1f41a6c | 386 | conds.quick_push (NULL); |
e6a23add | 387 | |
388 | gen_one_condition (arg, domain.ub, | |
389 | (domain.is_ub_inclusive | |
390 | ? GT_EXPR : GE_EXPR), | |
391 | "DCE_COND_UB", "DCE_COND_UB_TEST", | |
392 | conds, nconds); | |
393 | } | |
394 | } | |
395 | ||
396 | ||
397 | /* A helper function to generate condition | |
398 | code for the y argument in call pow (some_const, y). | |
48e1416a | 399 | See candidate selection in check_pow. Since the |
e6a23add | 400 | candidates' base values have a limited range, |
401 | the guarded code generated for y are simple: | |
402 | if (y > max_y) | |
403 | pow (const, y); | |
404 | Note max_y can be computed separately for each | |
405 | const base, but in this implementation, we | |
406 | choose to compute it using the max base | |
407 | in the allowed range for the purpose of | |
408 | simplicity. BASE is the constant base value, | |
409 | EXPN is the expression for the exponent argument, | |
410 | *CONDS is the vector to hold resulting statements, | |
411 | and *NCONDS is the number of logical conditions. */ | |
412 | ||
413 | static void | |
414 | gen_conditions_for_pow_cst_base (tree base, tree expn, | |
42acab1c | 415 | vec<gimple *> conds, |
e6a23add | 416 | unsigned *nconds) |
417 | { | |
48e1416a | 418 | inp_domain exp_domain; |
419 | /* Validate the range of the base constant to make | |
e6a23add | 420 | sure it is consistent with check_pow. */ |
421 | REAL_VALUE_TYPE mv; | |
422 | REAL_VALUE_TYPE bcv = TREE_REAL_CST (base); | |
20cb53c9 | 423 | gcc_assert (!real_equal (&bcv, &dconst1) |
1b67971e | 424 | && !real_less (&bcv, &dconst1)); |
e913b5cd | 425 | real_from_integer (&mv, TYPE_MODE (TREE_TYPE (base)), 256, UNSIGNED); |
1b67971e | 426 | gcc_assert (!real_less (&mv, &bcv)); |
e6a23add | 427 | |
428 | exp_domain = get_domain (0, false, false, | |
429 | 127, true, false); | |
430 | ||
431 | gen_conditions_for_domain (expn, exp_domain, | |
432 | conds, nconds); | |
433 | } | |
434 | ||
435 | /* Generate error condition code for pow calls with | |
436 | non constant base values. The candidates selected | |
437 | have their base argument value converted from | |
438 | integer (see check_pow) value (1, 2, 4 bytes), and | |
439 | the max exp value is computed based on the size | |
440 | of the integer type (i.e. max possible base value). | |
441 | The resulting input domain for exp argument is thus | |
48e1416a | 442 | conservative (smaller than the max value allowed by |
443 | the runtime value of the base). BASE is the integer | |
444 | base value, EXPN is the expression for the exponent | |
445 | argument, *CONDS is the vector to hold resulting | |
446 | statements, and *NCONDS is the number of logical | |
e6a23add | 447 | conditions. */ |
448 | ||
449 | static void | |
450 | gen_conditions_for_pow_int_base (tree base, tree expn, | |
42acab1c | 451 | vec<gimple *> conds, |
e6a23add | 452 | unsigned *nconds) |
453 | { | |
42acab1c | 454 | gimple *base_def; |
f018d957 | 455 | tree base_val0; |
7ecda5e8 | 456 | tree int_type; |
e6a23add | 457 | tree temp, tempn; |
75a70cf9 | 458 | tree cst0; |
42acab1c | 459 | gimple *stmt1, *stmt2; |
e6a23add | 460 | int bit_sz, max_exp; |
461 | inp_domain exp_domain; | |
462 | ||
463 | base_def = SSA_NAME_DEF_STMT (base); | |
75a70cf9 | 464 | base_val0 = gimple_assign_rhs1 (base_def); |
7ecda5e8 | 465 | int_type = TREE_TYPE (base_val0); |
e6a23add | 466 | bit_sz = TYPE_PRECISION (int_type); |
48e1416a | 467 | gcc_assert (bit_sz > 0 |
e6a23add | 468 | && bit_sz <= MAX_BASE_INT_BIT_SIZE); |
469 | ||
470 | /* Determine the max exp argument value according to | |
471 | the size of the base integer. The max exp value | |
472 | is conservatively estimated assuming IEEE754 double | |
473 | precision format. */ | |
474 | if (bit_sz == 8) | |
475 | max_exp = 128; | |
476 | else if (bit_sz == 16) | |
477 | max_exp = 64; | |
478 | else | |
f018d957 | 479 | { |
480 | gcc_assert (bit_sz == MAX_BASE_INT_BIT_SIZE); | |
481 | max_exp = 32; | |
482 | } | |
e6a23add | 483 | |
484 | /* For pow ((double)x, y), generate the following conditions: | |
485 | cond 1: | |
486 | temp1 = x; | |
487 | if (temp1 <= 0) | |
488 | ||
489 | cond 2: | |
490 | temp2 = y; | |
491 | if (temp2 > max_exp_real_cst) */ | |
492 | ||
493 | /* Generate condition in reverse order -- first | |
494 | the condition for the exp argument. */ | |
495 | ||
496 | exp_domain = get_domain (0, false, false, | |
497 | max_exp, true, true); | |
498 | ||
499 | gen_conditions_for_domain (expn, exp_domain, | |
500 | conds, nconds); | |
501 | ||
502 | /* Now generate condition for the base argument. | |
503 | Note it does not use the helper function | |
504 | gen_conditions_for_domain because the base | |
505 | type is integer. */ | |
506 | ||
507 | /* Push a separator. */ | |
f1f41a6c | 508 | conds.quick_push (NULL); |
e6a23add | 509 | |
510 | temp = create_tmp_var (int_type, "DCE_COND1"); | |
511 | cst0 = build_int_cst (int_type, 0); | |
75a70cf9 | 512 | stmt1 = gimple_build_assign (temp, base_val0); |
e6a23add | 513 | tempn = make_ssa_name (temp, stmt1); |
75a70cf9 | 514 | gimple_assign_set_lhs (stmt1, tempn); |
515 | stmt2 = gimple_build_cond (LE_EXPR, tempn, cst0, NULL_TREE, NULL_TREE); | |
e6a23add | 516 | |
f1f41a6c | 517 | conds.quick_push (stmt1); |
518 | conds.quick_push (stmt2); | |
e6a23add | 519 | (*nconds)++; |
520 | } | |
521 | ||
522 | /* Method to generate conditional statements for guarding conditionally | |
523 | dead calls to pow. One or more statements can be generated for | |
524 | each logical condition. Statement groups of different conditions | |
f1f41a6c | 525 | are separated by a NULL tree and they are stored in the vec |
e6a23add | 526 | conds. The number of logical conditions are stored in *nconds. |
527 | ||
528 | See C99 standard, 7.12.7.4:2, for description of pow (x, y). | |
529 | The precise condition for domain errors are complex. In this | |
530 | implementation, a simplified (but conservative) valid domain | |
531 | for x and y are used: x is positive to avoid dom errors, while | |
532 | y is smaller than a upper bound (depending on x) to avoid range | |
533 | errors. Runtime code is generated to check x (if not constant) | |
534 | and y against the valid domain. If it is out, jump to the call, | |
535 | otherwise the call is bypassed. POW_CALL is the call statement, | |
536 | *CONDS is a vector holding the resulting condition statements, | |
537 | and *NCONDS is the number of logical conditions. */ | |
538 | ||
539 | static void | |
42acab1c | 540 | gen_conditions_for_pow (gcall *pow_call, vec<gimple *> conds, |
e6a23add | 541 | unsigned *nconds) |
542 | { | |
543 | tree base, expn; | |
f018d957 | 544 | enum tree_code bc; |
e6a23add | 545 | |
1b4345f7 | 546 | gcc_checking_assert (check_pow (pow_call)); |
e6a23add | 547 | |
548 | *nconds = 0; | |
549 | ||
75a70cf9 | 550 | base = gimple_call_arg (pow_call, 0); |
551 | expn = gimple_call_arg (pow_call, 1); | |
e6a23add | 552 | |
553 | bc = TREE_CODE (base); | |
e6a23add | 554 | |
555 | if (bc == REAL_CST) | |
f018d957 | 556 | gen_conditions_for_pow_cst_base (base, expn, conds, nconds); |
e6a23add | 557 | else if (bc == SSA_NAME) |
f018d957 | 558 | gen_conditions_for_pow_int_base (base, expn, conds, nconds); |
e6a23add | 559 | else |
560 | gcc_unreachable (); | |
561 | } | |
562 | ||
563 | /* A helper routine to help computing the valid input domain | |
564 | for a builtin function. See C99 7.12.7 for details. In this | |
565 | implementation, we only handle single region domain. The | |
566 | resulting region can be conservative (smaller) than the actual | |
567 | one and rounded to integers. Some of the bounds are documented | |
568 | in the standard, while other limit constants are computed | |
48e1416a | 569 | assuming IEEE floating point format (for SF and DF modes). |
570 | Since IEEE only sets minimum requirements for long double format, | |
571 | different long double formats exist under different implementations | |
572 | (e.g, 64 bit double precision (DF), 80 bit double-extended | |
573 | precision (XF), and 128 bit quad precision (QF) ). For simplicity, | |
574 | in this implementation, the computed bounds for long double assume | |
575 | 64 bit format (DF), and are therefore conservative. Another | |
e6a23add | 576 | assumption is that single precision float type is always SF mode, |
48e1416a | 577 | and double type is DF mode. This function is quite |
e6a23add | 578 | implementation specific, so it may not be suitable to be part of |
579 | builtins.c. This needs to be revisited later to see if it can | |
580 | be leveraged in x87 assembly expansion. */ | |
581 | ||
582 | static inp_domain | |
583 | get_no_error_domain (enum built_in_function fnc) | |
584 | { | |
585 | switch (fnc) | |
586 | { | |
587 | /* Trig functions: return [-1, +1] */ | |
588 | CASE_FLT_FN (BUILT_IN_ACOS): | |
589 | CASE_FLT_FN (BUILT_IN_ASIN): | |
590 | return get_domain (-1, true, true, | |
591 | 1, true, true); | |
592 | /* Hyperbolic functions. */ | |
593 | CASE_FLT_FN (BUILT_IN_ACOSH): | |
594 | /* acosh: [1, +inf) */ | |
595 | return get_domain (1, true, true, | |
596 | 1, false, false); | |
597 | CASE_FLT_FN (BUILT_IN_ATANH): | |
598 | /* atanh: (-1, +1) */ | |
599 | return get_domain (-1, true, false, | |
600 | 1, true, false); | |
601 | case BUILT_IN_COSHF: | |
602 | case BUILT_IN_SINHF: | |
603 | /* coshf: (-89, +89) */ | |
604 | return get_domain (-89, true, false, | |
605 | 89, true, false); | |
606 | case BUILT_IN_COSH: | |
607 | case BUILT_IN_SINH: | |
608 | case BUILT_IN_COSHL: | |
609 | case BUILT_IN_SINHL: | |
610 | /* cosh: (-710, +710) */ | |
611 | return get_domain (-710, true, false, | |
612 | 710, true, false); | |
613 | /* Log functions: (0, +inf) */ | |
614 | CASE_FLT_FN (BUILT_IN_LOG): | |
615 | CASE_FLT_FN (BUILT_IN_LOG2): | |
616 | CASE_FLT_FN (BUILT_IN_LOG10): | |
617 | return get_domain (0, true, false, | |
618 | 0, false, false); | |
619 | CASE_FLT_FN (BUILT_IN_LOG1P): | |
620 | return get_domain (-1, true, false, | |
621 | 0, false, false); | |
622 | /* Exp functions. */ | |
623 | case BUILT_IN_EXPF: | |
624 | case BUILT_IN_EXPM1F: | |
625 | /* expf: (-inf, 88) */ | |
626 | return get_domain (-1, false, false, | |
627 | 88, true, false); | |
628 | case BUILT_IN_EXP: | |
629 | case BUILT_IN_EXPM1: | |
630 | case BUILT_IN_EXPL: | |
631 | case BUILT_IN_EXPM1L: | |
632 | /* exp: (-inf, 709) */ | |
633 | return get_domain (-1, false, false, | |
634 | 709, true, false); | |
635 | case BUILT_IN_EXP2F: | |
636 | /* exp2f: (-inf, 128) */ | |
637 | return get_domain (-1, false, false, | |
638 | 128, true, false); | |
639 | case BUILT_IN_EXP2: | |
640 | case BUILT_IN_EXP2L: | |
641 | /* exp2: (-inf, 1024) */ | |
642 | return get_domain (-1, false, false, | |
643 | 1024, true, false); | |
644 | case BUILT_IN_EXP10F: | |
645 | case BUILT_IN_POW10F: | |
646 | /* exp10f: (-inf, 38) */ | |
647 | return get_domain (-1, false, false, | |
648 | 38, true, false); | |
649 | case BUILT_IN_EXP10: | |
650 | case BUILT_IN_POW10: | |
651 | case BUILT_IN_EXP10L: | |
652 | case BUILT_IN_POW10L: | |
653 | /* exp10: (-inf, 308) */ | |
654 | return get_domain (-1, false, false, | |
655 | 308, true, false); | |
656 | /* sqrt: [0, +inf) */ | |
657 | CASE_FLT_FN (BUILT_IN_SQRT): | |
658 | return get_domain (0, true, true, | |
659 | 0, false, false); | |
660 | default: | |
48e1416a | 661 | gcc_unreachable (); |
e6a23add | 662 | } |
663 | ||
48e1416a | 664 | gcc_unreachable (); |
e6a23add | 665 | } |
666 | ||
667 | /* The function to generate shrink wrap conditions for a partially | |
668 | dead builtin call whose return value is not used anywhere, | |
669 | but has to be kept live due to potential error condition. | |
48e1416a | 670 | BI_CALL is the builtin call, CONDS is the vector of statements |
671 | for condition code, NCODES is the pointer to the number of | |
e6a23add | 672 | logical conditions. Statements belonging to different logical |
673 | condition are separated by NULL tree in the vector. */ | |
674 | ||
675 | static void | |
42acab1c | 676 | gen_shrink_wrap_conditions (gcall *bi_call, vec<gimple *> conds, |
e6a23add | 677 | unsigned int *nconds) |
678 | { | |
1a91d914 | 679 | gcall *call; |
75a70cf9 | 680 | tree fn; |
e6a23add | 681 | enum built_in_function fnc; |
682 | ||
f1f41a6c | 683 | gcc_assert (nconds && conds.exists ()); |
684 | gcc_assert (conds.length () == 0); | |
75a70cf9 | 685 | gcc_assert (is_gimple_call (bi_call)); |
e6a23add | 686 | |
687 | call = bi_call; | |
75a70cf9 | 688 | fn = gimple_call_fndecl (call); |
e6a23add | 689 | gcc_assert (fn && DECL_BUILT_IN (fn)); |
690 | fnc = DECL_FUNCTION_CODE (fn); | |
691 | *nconds = 0; | |
692 | ||
693 | if (fnc == BUILT_IN_POW) | |
694 | gen_conditions_for_pow (call, conds, nconds); | |
695 | else | |
696 | { | |
697 | tree arg; | |
698 | inp_domain domain = get_no_error_domain (fnc); | |
699 | *nconds = 0; | |
75a70cf9 | 700 | arg = gimple_call_arg (bi_call, 0); |
e6a23add | 701 | gen_conditions_for_domain (arg, domain, conds, nconds); |
702 | } | |
703 | ||
704 | return; | |
705 | } | |
706 | ||
707 | ||
708 | /* Probability of the branch (to the call) is taken. */ | |
709 | #define ERR_PROB 0.01 | |
710 | ||
48e1416a | 711 | /* The function to shrink wrap a partially dead builtin call |
712 | whose return value is not used anywhere, but has to be kept | |
e6a23add | 713 | live due to potential error condition. Returns true if the |
714 | transformation actually happens. */ | |
715 | ||
48e1416a | 716 | static bool |
1a91d914 | 717 | shrink_wrap_one_built_in_call (gcall *bi_call) |
e6a23add | 718 | { |
75a70cf9 | 719 | gimple_stmt_iterator bi_call_bsi; |
e6a23add | 720 | basic_block bi_call_bb, join_tgt_bb, guard_bb, guard_bb0; |
721 | edge join_tgt_in_edge_from_call, join_tgt_in_edge_fall_thru; | |
722 | edge bi_call_in_edge0, guard_bb_in_edge; | |
e6a23add | 723 | unsigned tn_cond_stmts, nconds; |
724 | unsigned ci; | |
42acab1c | 725 | gimple *cond_expr = NULL; |
726 | gimple *cond_expr_start; | |
e6a23add | 727 | tree bi_call_label_decl; |
42acab1c | 728 | gimple *bi_call_label; |
e6a23add | 729 | |
42acab1c | 730 | auto_vec<gimple *, 12> conds; |
e6a23add | 731 | gen_shrink_wrap_conditions (bi_call, conds, &nconds); |
732 | ||
733 | /* This can happen if the condition generator decides | |
734 | it is not beneficial to do the transformation. Just | |
48e1416a | 735 | return false and do not do any transformation for |
e6a23add | 736 | the call. */ |
737 | if (nconds == 0) | |
e85cf4e5 | 738 | return false; |
e6a23add | 739 | |
75a70cf9 | 740 | bi_call_bb = gimple_bb (bi_call); |
e6a23add | 741 | |
36ee0f30 | 742 | /* Now find the join target bb -- split bi_call_bb if needed. */ |
743 | if (stmt_ends_bb_p (bi_call)) | |
744 | { | |
745 | /* If the call must be the last in the bb, don't split the block, | |
746 | it could e.g. have EH edges. */ | |
747 | join_tgt_in_edge_from_call = find_fallthru_edge (bi_call_bb->succs); | |
748 | if (join_tgt_in_edge_from_call == NULL) | |
e85cf4e5 | 749 | return false; |
36ee0f30 | 750 | } |
751 | else | |
752 | join_tgt_in_edge_from_call = split_block (bi_call_bb, bi_call); | |
e6a23add | 753 | |
75a70cf9 | 754 | bi_call_bsi = gsi_for_stmt (bi_call); |
e6a23add | 755 | |
756 | join_tgt_bb = join_tgt_in_edge_from_call->dest; | |
757 | ||
758 | /* Now it is time to insert the first conditional expression | |
759 | into bi_call_bb and split this bb so that bi_call is | |
760 | shrink-wrapped. */ | |
f1f41a6c | 761 | tn_cond_stmts = conds.length (); |
e6a23add | 762 | cond_expr = NULL; |
f1f41a6c | 763 | cond_expr_start = conds[0]; |
e6a23add | 764 | for (ci = 0; ci < tn_cond_stmts; ci++) |
765 | { | |
42acab1c | 766 | gimple *c = conds[ci]; |
e6a23add | 767 | gcc_assert (c || ci != 0); |
768 | if (!c) | |
769 | break; | |
75a70cf9 | 770 | gsi_insert_before (&bi_call_bsi, c, GSI_SAME_STMT); |
e6a23add | 771 | cond_expr = c; |
772 | } | |
773 | nconds--; | |
774 | ci++; | |
75a70cf9 | 775 | gcc_assert (cond_expr && gimple_code (cond_expr) == GIMPLE_COND); |
e6a23add | 776 | |
777 | /* Now the label. */ | |
e60a6f7b | 778 | bi_call_label_decl = create_artificial_label (gimple_location (bi_call)); |
75a70cf9 | 779 | bi_call_label = gimple_build_label (bi_call_label_decl); |
780 | gsi_insert_before (&bi_call_bsi, bi_call_label, GSI_SAME_STMT); | |
e6a23add | 781 | |
782 | bi_call_in_edge0 = split_block (bi_call_bb, cond_expr); | |
783 | bi_call_in_edge0->flags &= ~EDGE_FALLTHRU; | |
784 | bi_call_in_edge0->flags |= EDGE_TRUE_VALUE; | |
785 | guard_bb0 = bi_call_bb; | |
786 | bi_call_bb = bi_call_in_edge0->dest; | |
48e1416a | 787 | join_tgt_in_edge_fall_thru = make_edge (guard_bb0, join_tgt_bb, |
e6a23add | 788 | EDGE_FALSE_VALUE); |
789 | ||
790 | bi_call_in_edge0->probability = REG_BR_PROB_BASE * ERR_PROB; | |
9231bca5 | 791 | bi_call_in_edge0->count = |
792 | apply_probability (guard_bb0->count, | |
793 | bi_call_in_edge0->probability); | |
e6a23add | 794 | join_tgt_in_edge_fall_thru->probability = |
9231bca5 | 795 | inverse_probability (bi_call_in_edge0->probability); |
796 | join_tgt_in_edge_fall_thru->count = | |
797 | guard_bb0->count - bi_call_in_edge0->count; | |
e6a23add | 798 | |
799 | /* Code generation for the rest of the conditions */ | |
800 | guard_bb = guard_bb0; | |
801 | while (nconds > 0) | |
802 | { | |
803 | unsigned ci0; | |
804 | edge bi_call_in_edge; | |
75a70cf9 | 805 | gimple_stmt_iterator guard_bsi = gsi_for_stmt (cond_expr_start); |
e6a23add | 806 | ci0 = ci; |
f1f41a6c | 807 | cond_expr_start = conds[ci0]; |
e6a23add | 808 | for (; ci < tn_cond_stmts; ci++) |
809 | { | |
42acab1c | 810 | gimple *c = conds[ci]; |
e6a23add | 811 | gcc_assert (c || ci != ci0); |
812 | if (!c) | |
813 | break; | |
75a70cf9 | 814 | gsi_insert_before (&guard_bsi, c, GSI_SAME_STMT); |
e6a23add | 815 | cond_expr = c; |
816 | } | |
817 | nconds--; | |
818 | ci++; | |
75a70cf9 | 819 | gcc_assert (cond_expr && gimple_code (cond_expr) == GIMPLE_COND); |
e6a23add | 820 | guard_bb_in_edge = split_block (guard_bb, cond_expr); |
821 | guard_bb_in_edge->flags &= ~EDGE_FALLTHRU; | |
822 | guard_bb_in_edge->flags |= EDGE_FALSE_VALUE; | |
823 | ||
824 | bi_call_in_edge = make_edge (guard_bb, bi_call_bb, EDGE_TRUE_VALUE); | |
825 | ||
826 | bi_call_in_edge->probability = REG_BR_PROB_BASE * ERR_PROB; | |
9231bca5 | 827 | bi_call_in_edge->count = |
828 | apply_probability (guard_bb->count, | |
829 | bi_call_in_edge->probability); | |
e6a23add | 830 | guard_bb_in_edge->probability = |
9231bca5 | 831 | inverse_probability (bi_call_in_edge->probability); |
832 | guard_bb_in_edge->count = guard_bb->count - bi_call_in_edge->count; | |
e6a23add | 833 | } |
834 | ||
e6a23add | 835 | if (dump_file && (dump_flags & TDF_DETAILS)) |
836 | { | |
837 | location_t loc; | |
75a70cf9 | 838 | loc = gimple_location (bi_call); |
e6a23add | 839 | fprintf (dump_file, |
840 | "%s:%d: note: function call is shrink-wrapped" | |
841 | " into error conditions.\n", | |
842 | LOCATION_FILE (loc), LOCATION_LINE (loc)); | |
843 | } | |
844 | ||
845 | return true; | |
846 | } | |
847 | ||
848 | /* The top level function for conditional dead code shrink | |
849 | wrapping transformation. */ | |
850 | ||
851 | static bool | |
1a91d914 | 852 | shrink_wrap_conditional_dead_built_in_calls (vec<gcall *> calls) |
e6a23add | 853 | { |
854 | bool changed = false; | |
855 | unsigned i = 0; | |
856 | ||
f1f41a6c | 857 | unsigned n = calls.length (); |
48e1416a | 858 | if (n == 0) |
e6a23add | 859 | return false; |
860 | ||
861 | for (; i < n ; i++) | |
862 | { | |
1a91d914 | 863 | gcall *bi_call = calls[i]; |
e6a23add | 864 | changed |= shrink_wrap_one_built_in_call (bi_call); |
865 | } | |
866 | ||
867 | return changed; | |
868 | } | |
869 | ||
65b0537f | 870 | namespace { |
871 | ||
872 | const pass_data pass_data_call_cdce = | |
873 | { | |
874 | GIMPLE_PASS, /* type */ | |
875 | "cdce", /* name */ | |
876 | OPTGROUP_NONE, /* optinfo_flags */ | |
65b0537f | 877 | TV_TREE_CALL_CDCE, /* tv_id */ |
878 | ( PROP_cfg | PROP_ssa ), /* properties_required */ | |
879 | 0, /* properties_provided */ | |
880 | 0, /* properties_destroyed */ | |
881 | 0, /* todo_flags_start */ | |
8b88439e | 882 | 0, /* todo_flags_finish */ |
65b0537f | 883 | }; |
e6a23add | 884 | |
65b0537f | 885 | class pass_call_cdce : public gimple_opt_pass |
886 | { | |
887 | public: | |
888 | pass_call_cdce (gcc::context *ctxt) | |
889 | : gimple_opt_pass (pass_data_call_cdce, ctxt) | |
890 | {} | |
891 | ||
892 | /* opt_pass methods: */ | |
893 | virtual bool gate (function *fun) | |
894 | { | |
895 | /* The limit constants used in the implementation | |
896 | assume IEEE floating point format. Other formats | |
897 | can be supported in the future if needed. */ | |
898 | return flag_tree_builtin_call_dce != 0 | |
899 | && optimize_function_for_speed_p (fun); | |
900 | } | |
e6a23add | 901 | |
65b0537f | 902 | virtual unsigned int execute (function *); |
903 | ||
904 | }; // class pass_call_cdce | |
905 | ||
906 | unsigned int | |
907 | pass_call_cdce::execute (function *fun) | |
e6a23add | 908 | { |
909 | basic_block bb; | |
75a70cf9 | 910 | gimple_stmt_iterator i; |
e6a23add | 911 | bool something_changed = false; |
1a91d914 | 912 | auto_vec<gcall *> cond_dead_built_in_calls; |
65b0537f | 913 | FOR_EACH_BB_FN (bb, fun) |
e6a23add | 914 | { |
915 | /* Collect dead call candidates. */ | |
75a70cf9 | 916 | for (i = gsi_start_bb (bb); !gsi_end_p (i); gsi_next (&i)) |
e6a23add | 917 | { |
1a91d914 | 918 | gcall *stmt = dyn_cast <gcall *> (gsi_stmt (i)); |
919 | if (stmt && is_call_dce_candidate (stmt)) | |
e6a23add | 920 | { |
921 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
922 | { | |
923 | fprintf (dump_file, "Found conditional dead call: "); | |
75a70cf9 | 924 | print_gimple_stmt (dump_file, stmt, 0, TDF_SLIM); |
e6a23add | 925 | fprintf (dump_file, "\n"); |
926 | } | |
f1f41a6c | 927 | if (!cond_dead_built_in_calls.exists ()) |
928 | cond_dead_built_in_calls.create (64); | |
929 | cond_dead_built_in_calls.safe_push (stmt); | |
e6a23add | 930 | } |
931 | } | |
932 | } | |
933 | ||
f1f41a6c | 934 | if (!cond_dead_built_in_calls.exists ()) |
86ed2c77 | 935 | return 0; |
936 | ||
937 | something_changed | |
938 | = shrink_wrap_conditional_dead_built_in_calls (cond_dead_built_in_calls); | |
e6a23add | 939 | |
e6a23add | 940 | if (something_changed) |
941 | { | |
942 | free_dominance_info (CDI_DOMINATORS); | |
943 | free_dominance_info (CDI_POST_DOMINATORS); | |
dd277d48 | 944 | /* As we introduced new control-flow we need to insert PHI-nodes |
945 | for the call-clobbers of the remaining call. */ | |
65b0537f | 946 | mark_virtual_operands_for_renaming (fun); |
560965e9 | 947 | return TODO_update_ssa; |
e6a23add | 948 | } |
560965e9 | 949 | |
950 | return 0; | |
e6a23add | 951 | } |
952 | ||
cbe8bda8 | 953 | } // anon namespace |
954 | ||
955 | gimple_opt_pass * | |
956 | make_pass_call_cdce (gcc::context *ctxt) | |
957 | { | |
958 | return new pass_call_cdce (ctxt); | |
959 | } |