]>
Commit | Line | Data |
---|---|---|
c2699190 | 1 | /* Conditional Dead Call Elimination pass for the GNU compiler. |
23a5b65a | 2 | Copyright (C) 2008-2014 Free Software Foundation, Inc. |
c2699190 XDL |
3 | Contributed by Xinliang David Li <davidxl@google.com> |
4 | ||
5 | This file is part of GCC. | |
b8698a0f | 6 | |
c2699190 XDL |
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. | |
b8698a0f | 11 | |
c2699190 XDL |
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. | |
b8698a0f | 16 | |
c2699190 XDL |
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" | |
24 | #include "tm.h" | |
60393bbc AM |
25 | #include "predict.h" |
26 | #include "vec.h" | |
27 | #include "hashtab.h" | |
28 | #include "hash-set.h" | |
29 | #include "machmode.h" | |
30 | #include "hard-reg-set.h" | |
31 | #include "input.h" | |
32 | #include "function.h" | |
33 | #include "dominance.h" | |
34 | #include "cfg.h" | |
c2699190 | 35 | #include "basic-block.h" |
c2699190 | 36 | #include "tree.h" |
d8a2d370 | 37 | #include "stor-layout.h" |
cf835838 | 38 | #include "gimple-pretty-print.h" |
2fb9a547 AM |
39 | #include "tree-ssa-alias.h" |
40 | #include "internal-fn.h" | |
41 | #include "gimple-expr.h" | |
42 | #include "is-a.h" | |
726a989a | 43 | #include "gimple.h" |
5be5c238 | 44 | #include "gimple-iterator.h" |
442b4905 AM |
45 | #include "gimple-ssa.h" |
46 | #include "tree-cfg.h" | |
d8a2d370 | 47 | #include "stringpool.h" |
442b4905 AM |
48 | #include "tree-ssanames.h" |
49 | #include "tree-into-ssa.h" | |
c2699190 | 50 | #include "tree-pass.h" |
c2699190 XDL |
51 | #include "flags.h" |
52 | \f | |
53 | ||
54 | /* Conditional dead call elimination | |
55 | ||
56 | Some builtin functions can set errno on error conditions, but they | |
57 | are otherwise pure. If the result of a call to such a function is | |
58 | not used, the compiler can still not eliminate the call without | |
59 | powerful interprocedural analysis to prove that the errno is not | |
60 | checked. However, if the conditions under which the error occurs | |
b8698a0f | 61 | are known, the compiler can conditionally dead code eliminate the |
c2699190 XDL |
62 | calls by shrink-wrapping the semi-dead calls into the error condition: |
63 | ||
64 | built_in_call (args) | |
65 | ==> | |
66 | if (error_cond (args)) | |
67 | built_in_call (args) | |
68 | ||
69 | An actual simple example is : | |
70 | log (x); // Mostly dead call | |
71 | ==> | |
72 | if (x < 0) | |
73 | log (x); | |
74 | With this change, call to log (x) is effectively eliminated, as | |
75 | in majority of the cases, log won't be called with x out of | |
76 | range. The branch is totally predictable, so the branch cost | |
b8698a0f | 77 | is low. |
c2699190 XDL |
78 | |
79 | Note that library functions are not supposed to clear errno to zero without | |
80 | error. See IEEE Std 1003.1, section 2.3 Error Numbers, and section 7.5:3 of | |
81 | ISO/IEC 9899 (C99). | |
82 | ||
83 | The condition wrapping the builtin call is conservatively set to avoid too | |
84 | aggressive (wrong) shrink wrapping. The optimization is called conditional | |
85 | dead call elimination because the call is eliminated under the condition | |
86 | that the input arguments would not lead to domain or range error (for | |
87 | instance when x <= 0 for a log (x) call), however the chances that the error | |
88 | condition is hit is very low (those builtin calls which are conditionally | |
89 | dead are usually part of the C++ abstraction penalty exposed after | |
90 | inlining). */ | |
91 | ||
92 | ||
b8698a0f | 93 | /* A structure for representing input domain of |
c2699190 | 94 | a function argument in integer. If the lower |
b8698a0f L |
95 | bound is -inf, has_lb is set to false. If the |
96 | upper bound is +inf, has_ub is false. | |
97 | is_lb_inclusive and is_ub_inclusive are flags | |
98 | to indicate if lb and ub value are inclusive | |
c2699190 XDL |
99 | respectively. */ |
100 | ||
101 | typedef struct input_domain | |
102 | { | |
103 | int lb; | |
104 | int ub; | |
105 | bool has_lb; | |
106 | bool has_ub; | |
107 | bool is_lb_inclusive; | |
108 | bool is_ub_inclusive; | |
109 | } inp_domain; | |
110 | ||
c2699190 | 111 | /* A helper function to construct and return an input |
b8698a0f | 112 | domain object. LB is the lower bound, HAS_LB is |
c2699190 XDL |
113 | a boolean flag indicating if the lower bound exists, |
114 | and LB_INCLUSIVE is a boolean flag indicating if the | |
115 | lower bound is inclusive or not. UB, HAS_UB, and | |
b8698a0f | 116 | UB_INCLUSIVE have the same meaning, but for upper |
c2699190 XDL |
117 | bound of the domain. */ |
118 | ||
119 | static inp_domain | |
120 | get_domain (int lb, bool has_lb, bool lb_inclusive, | |
121 | int ub, bool has_ub, bool ub_inclusive) | |
122 | { | |
123 | inp_domain domain; | |
124 | domain.lb = lb; | |
125 | domain.has_lb = has_lb; | |
126 | domain.is_lb_inclusive = lb_inclusive; | |
127 | domain.ub = ub; | |
128 | domain.has_ub = has_ub; | |
129 | domain.is_ub_inclusive = ub_inclusive; | |
130 | return domain; | |
131 | } | |
132 | ||
b8698a0f | 133 | /* A helper function to check the target format for the |
c2699190 | 134 | argument type. In this implementation, only IEEE formats |
b8698a0f | 135 | are supported. ARG is the call argument to be checked. |
c2699190 XDL |
136 | Returns true if the format is supported. To support other |
137 | target formats, function get_no_error_domain needs to be | |
b8698a0f L |
138 | enhanced to have range bounds properly computed. Since |
139 | the check is cheap (very small number of candidates | |
c2699190 XDL |
140 | to be checked), the result is not cached for each float type. */ |
141 | ||
142 | static bool | |
143 | check_target_format (tree arg) | |
144 | { | |
145 | tree type; | |
ef4bddc2 | 146 | machine_mode mode; |
c2699190 | 147 | const struct real_format *rfmt; |
b8698a0f | 148 | |
c2699190 XDL |
149 | type = TREE_TYPE (arg); |
150 | mode = TYPE_MODE (type); | |
151 | rfmt = REAL_MODE_FORMAT (mode); | |
5d94a6d0 | 152 | if ((mode == SFmode |
ea8bce02 AS |
153 | && (rfmt == &ieee_single_format || rfmt == &mips_single_format |
154 | || rfmt == &motorola_single_format)) | |
5d94a6d0 | 155 | || (mode == DFmode |
ea8bce02 AS |
156 | && (rfmt == &ieee_double_format || rfmt == &mips_double_format |
157 | || rfmt == &motorola_double_format)) | |
c2699190 | 158 | /* For long double, we can not really check XFmode |
b8698a0f L |
159 | which is only defined on intel platforms. |
160 | Candidate pre-selection using builtin function | |
161 | code guarantees that we are checking formats | |
c2699190 | 162 | for long double modes: double, quad, and extended. */ |
b8698a0f | 163 | || (mode != SFmode && mode != DFmode |
c2699190 | 164 | && (rfmt == &ieee_quad_format |
5d94a6d0 | 165 | || rfmt == &mips_quad_format |
ea8bce02 | 166 | || rfmt == &ieee_extended_motorola_format |
b8698a0f L |
167 | || rfmt == &ieee_extended_intel_96_format |
168 | || rfmt == &ieee_extended_intel_128_format | |
c2699190 XDL |
169 | || rfmt == &ieee_extended_intel_96_round_53_format))) |
170 | return true; | |
171 | ||
172 | return false; | |
173 | } | |
174 | ||
175 | \f | |
176 | /* A helper function to help select calls to pow that are suitable for | |
177 | conditional DCE transformation. It looks for pow calls that can be | |
178 | guided with simple conditions. Such calls either have constant base | |
b8698a0f | 179 | values or base values converted from integers. Returns true if |
c2699190 XDL |
180 | the pow call POW_CALL is a candidate. */ |
181 | ||
182 | /* The maximum integer bit size for base argument of a pow call | |
183 | that is suitable for shrink-wrapping transformation. */ | |
184 | #define MAX_BASE_INT_BIT_SIZE 32 | |
185 | ||
186 | static bool | |
726a989a | 187 | check_pow (gimple pow_call) |
c2699190 XDL |
188 | { |
189 | tree base, expn; | |
190 | enum tree_code bc, ec; | |
191 | ||
726a989a | 192 | if (gimple_call_num_args (pow_call) != 2) |
c2699190 XDL |
193 | return false; |
194 | ||
726a989a RB |
195 | base = gimple_call_arg (pow_call, 0); |
196 | expn = gimple_call_arg (pow_call, 1); | |
c2699190 XDL |
197 | |
198 | if (!check_target_format (expn)) | |
199 | return false; | |
200 | ||
201 | bc = TREE_CODE (base); | |
202 | ec = TREE_CODE (expn); | |
203 | ||
204 | /* Folding candidates are not interesting. | |
205 | Can actually assert that it is already folded. */ | |
206 | if (ec == REAL_CST && bc == REAL_CST) | |
207 | return false; | |
208 | ||
209 | if (bc == REAL_CST) | |
210 | { | |
211 | /* Only handle a fixed range of constant. */ | |
212 | REAL_VALUE_TYPE mv; | |
213 | REAL_VALUE_TYPE bcv = TREE_REAL_CST (base); | |
214 | if (REAL_VALUES_EQUAL (bcv, dconst1)) | |
215 | return false; | |
216 | if (REAL_VALUES_LESS (bcv, dconst1)) | |
217 | return false; | |
807e902e | 218 | real_from_integer (&mv, TYPE_MODE (TREE_TYPE (base)), 256, UNSIGNED); |
c2699190 XDL |
219 | if (REAL_VALUES_LESS (mv, bcv)) |
220 | return false; | |
221 | return true; | |
222 | } | |
223 | else if (bc == SSA_NAME) | |
224 | { | |
6b4a85ad | 225 | tree base_val0, type; |
726a989a | 226 | gimple base_def; |
c2699190 XDL |
227 | int bit_sz; |
228 | ||
229 | /* Only handles cases where base value is converted | |
b8698a0f | 230 | from integer values. */ |
c2699190 | 231 | base_def = SSA_NAME_DEF_STMT (base); |
726a989a | 232 | if (gimple_code (base_def) != GIMPLE_ASSIGN) |
c2699190 XDL |
233 | return false; |
234 | ||
726a989a | 235 | if (gimple_assign_rhs_code (base_def) != FLOAT_EXPR) |
c2699190 | 236 | return false; |
726a989a | 237 | base_val0 = gimple_assign_rhs1 (base_def); |
c2699190 | 238 | |
6b4a85ad | 239 | type = TREE_TYPE (base_val0); |
c2699190 XDL |
240 | if (TREE_CODE (type) != INTEGER_TYPE) |
241 | return false; | |
242 | bit_sz = TYPE_PRECISION (type); | |
243 | /* If the type of the base is too wide, | |
244 | the resulting shrink wrapping condition | |
245 | will be too conservative. */ | |
246 | if (bit_sz > MAX_BASE_INT_BIT_SIZE) | |
247 | return false; | |
248 | ||
249 | return true; | |
250 | } | |
251 | else | |
252 | return false; | |
253 | } | |
254 | ||
255 | /* A helper function to help select candidate function calls that are | |
256 | suitable for conditional DCE. Candidate functions must have single | |
257 | valid input domain in this implementation except for pow (see check_pow). | |
258 | Returns true if the function call is a candidate. */ | |
259 | ||
260 | static bool | |
726a989a | 261 | check_builtin_call (gimple bcall) |
c2699190 XDL |
262 | { |
263 | tree arg; | |
264 | ||
726a989a | 265 | arg = gimple_call_arg (bcall, 0); |
c2699190 XDL |
266 | return check_target_format (arg); |
267 | } | |
268 | ||
269 | /* A helper function to determine if a builtin function call is a | |
270 | candidate for conditional DCE. Returns true if the builtin call | |
271 | is a candidate. */ | |
272 | ||
273 | static bool | |
726a989a | 274 | is_call_dce_candidate (gimple call) |
c2699190 XDL |
275 | { |
276 | tree fn; | |
277 | enum built_in_function fnc; | |
278 | ||
726a989a RB |
279 | /* Only potentially dead calls are considered. */ |
280 | if (gimple_call_lhs (call)) | |
c2699190 XDL |
281 | return false; |
282 | ||
726a989a RB |
283 | fn = gimple_call_fndecl (call); |
284 | if (!fn | |
b8698a0f | 285 | || !DECL_BUILT_IN (fn) |
c2699190 XDL |
286 | || (DECL_BUILT_IN_CLASS (fn) != BUILT_IN_NORMAL)) |
287 | return false; | |
288 | ||
289 | fnc = DECL_FUNCTION_CODE (fn); | |
290 | switch (fnc) | |
291 | { | |
292 | /* Trig functions. */ | |
293 | CASE_FLT_FN (BUILT_IN_ACOS): | |
294 | CASE_FLT_FN (BUILT_IN_ASIN): | |
295 | /* Hyperbolic functions. */ | |
296 | CASE_FLT_FN (BUILT_IN_ACOSH): | |
297 | CASE_FLT_FN (BUILT_IN_ATANH): | |
298 | CASE_FLT_FN (BUILT_IN_COSH): | |
299 | CASE_FLT_FN (BUILT_IN_SINH): | |
300 | /* Log functions. */ | |
301 | CASE_FLT_FN (BUILT_IN_LOG): | |
302 | CASE_FLT_FN (BUILT_IN_LOG2): | |
303 | CASE_FLT_FN (BUILT_IN_LOG10): | |
304 | CASE_FLT_FN (BUILT_IN_LOG1P): | |
305 | /* Exp functions. */ | |
306 | CASE_FLT_FN (BUILT_IN_EXP): | |
307 | CASE_FLT_FN (BUILT_IN_EXP2): | |
308 | CASE_FLT_FN (BUILT_IN_EXP10): | |
309 | CASE_FLT_FN (BUILT_IN_EXPM1): | |
310 | CASE_FLT_FN (BUILT_IN_POW10): | |
311 | /* Sqrt. */ | |
312 | CASE_FLT_FN (BUILT_IN_SQRT): | |
313 | return check_builtin_call (call); | |
314 | /* Special one: two argument pow. */ | |
315 | case BUILT_IN_POW: | |
316 | return check_pow (call); | |
317 | default: | |
318 | break; | |
319 | } | |
320 | ||
321 | return false; | |
322 | } | |
323 | ||
324 | \f | |
325 | /* A helper function to generate gimple statements for | |
326 | one bound comparison. ARG is the call argument to | |
327 | be compared with the bound, LBUB is the bound value | |
328 | in integer, TCODE is the tree_code of the comparison, | |
329 | TEMP_NAME1/TEMP_NAME2 are names of the temporaries, | |
330 | CONDS is a vector holding the produced GIMPLE statements, | |
331 | and NCONDS points to the variable holding the number | |
b8698a0f | 332 | of logical comparisons. CONDS is either empty or |
c2699190 XDL |
333 | a list ended with a null tree. */ |
334 | ||
335 | static void | |
b8698a0f | 336 | gen_one_condition (tree arg, int lbub, |
c2699190 XDL |
337 | enum tree_code tcode, |
338 | const char *temp_name1, | |
726a989a | 339 | const char *temp_name2, |
9771b263 | 340 | vec<gimple> conds, |
c2699190 XDL |
341 | unsigned *nconds) |
342 | { | |
343 | tree lbub_real_cst, lbub_cst, float_type; | |
344 | tree temp, tempn, tempc, tempcn; | |
726a989a | 345 | gimple stmt1, stmt2, stmt3; |
c2699190 XDL |
346 | |
347 | float_type = TREE_TYPE (arg); | |
348 | lbub_cst = build_int_cst (integer_type_node, lbub); | |
349 | lbub_real_cst = build_real_from_int_cst (float_type, lbub_cst); | |
350 | ||
351 | temp = create_tmp_var (float_type, temp_name1); | |
726a989a | 352 | stmt1 = gimple_build_assign (temp, arg); |
c2699190 | 353 | tempn = make_ssa_name (temp, stmt1); |
726a989a | 354 | gimple_assign_set_lhs (stmt1, tempn); |
c2699190 XDL |
355 | |
356 | tempc = create_tmp_var (boolean_type_node, temp_name2); | |
726a989a RB |
357 | stmt2 = gimple_build_assign (tempc, |
358 | fold_build2 (tcode, | |
359 | boolean_type_node, | |
360 | tempn, lbub_real_cst)); | |
c2699190 | 361 | tempcn = make_ssa_name (tempc, stmt2); |
726a989a RB |
362 | gimple_assign_set_lhs (stmt2, tempcn); |
363 | ||
364 | stmt3 = gimple_build_cond_from_tree (tempcn, NULL_TREE, NULL_TREE); | |
9771b263 DN |
365 | conds.quick_push (stmt1); |
366 | conds.quick_push (stmt2); | |
367 | conds.quick_push (stmt3); | |
c2699190 XDL |
368 | (*nconds)++; |
369 | } | |
370 | ||
371 | /* A helper function to generate GIMPLE statements for | |
372 | out of input domain check. ARG is the call argument | |
373 | to be runtime checked, DOMAIN holds the valid domain | |
374 | for the given function, CONDS points to the vector | |
b8698a0f L |
375 | holding the result GIMPLE statements. *NCONDS is |
376 | the number of logical comparisons. This function | |
c2699190 XDL |
377 | produces no more than two logical comparisons, one |
378 | for lower bound check, one for upper bound check. */ | |
379 | ||
380 | static void | |
381 | gen_conditions_for_domain (tree arg, inp_domain domain, | |
9771b263 | 382 | vec<gimple> conds, |
c2699190 XDL |
383 | unsigned *nconds) |
384 | { | |
385 | if (domain.has_lb) | |
386 | gen_one_condition (arg, domain.lb, | |
387 | (domain.is_lb_inclusive | |
388 | ? LT_EXPR : LE_EXPR), | |
389 | "DCE_COND_LB", "DCE_COND_LB_TEST", | |
390 | conds, nconds); | |
391 | ||
392 | if (domain.has_ub) | |
393 | { | |
394 | /* Now push a separator. */ | |
395 | if (domain.has_lb) | |
9771b263 | 396 | conds.quick_push (NULL); |
c2699190 XDL |
397 | |
398 | gen_one_condition (arg, domain.ub, | |
399 | (domain.is_ub_inclusive | |
400 | ? GT_EXPR : GE_EXPR), | |
401 | "DCE_COND_UB", "DCE_COND_UB_TEST", | |
402 | conds, nconds); | |
403 | } | |
404 | } | |
405 | ||
406 | ||
407 | /* A helper function to generate condition | |
408 | code for the y argument in call pow (some_const, y). | |
b8698a0f | 409 | See candidate selection in check_pow. Since the |
c2699190 XDL |
410 | candidates' base values have a limited range, |
411 | the guarded code generated for y are simple: | |
412 | if (y > max_y) | |
413 | pow (const, y); | |
414 | Note max_y can be computed separately for each | |
415 | const base, but in this implementation, we | |
416 | choose to compute it using the max base | |
417 | in the allowed range for the purpose of | |
418 | simplicity. BASE is the constant base value, | |
419 | EXPN is the expression for the exponent argument, | |
420 | *CONDS is the vector to hold resulting statements, | |
421 | and *NCONDS is the number of logical conditions. */ | |
422 | ||
423 | static void | |
424 | gen_conditions_for_pow_cst_base (tree base, tree expn, | |
9771b263 | 425 | vec<gimple> conds, |
c2699190 XDL |
426 | unsigned *nconds) |
427 | { | |
b8698a0f L |
428 | inp_domain exp_domain; |
429 | /* Validate the range of the base constant to make | |
c2699190 XDL |
430 | sure it is consistent with check_pow. */ |
431 | REAL_VALUE_TYPE mv; | |
432 | REAL_VALUE_TYPE bcv = TREE_REAL_CST (base); | |
433 | gcc_assert (!REAL_VALUES_EQUAL (bcv, dconst1) | |
434 | && !REAL_VALUES_LESS (bcv, dconst1)); | |
807e902e | 435 | real_from_integer (&mv, TYPE_MODE (TREE_TYPE (base)), 256, UNSIGNED); |
c2699190 XDL |
436 | gcc_assert (!REAL_VALUES_LESS (mv, bcv)); |
437 | ||
438 | exp_domain = get_domain (0, false, false, | |
439 | 127, true, false); | |
440 | ||
441 | gen_conditions_for_domain (expn, exp_domain, | |
442 | conds, nconds); | |
443 | } | |
444 | ||
445 | /* Generate error condition code for pow calls with | |
446 | non constant base values. The candidates selected | |
447 | have their base argument value converted from | |
448 | integer (see check_pow) value (1, 2, 4 bytes), and | |
449 | the max exp value is computed based on the size | |
450 | of the integer type (i.e. max possible base value). | |
451 | The resulting input domain for exp argument is thus | |
b8698a0f L |
452 | conservative (smaller than the max value allowed by |
453 | the runtime value of the base). BASE is the integer | |
454 | base value, EXPN is the expression for the exponent | |
455 | argument, *CONDS is the vector to hold resulting | |
456 | statements, and *NCONDS is the number of logical | |
c2699190 XDL |
457 | conditions. */ |
458 | ||
459 | static void | |
460 | gen_conditions_for_pow_int_base (tree base, tree expn, | |
9771b263 | 461 | vec<gimple> conds, |
c2699190 XDL |
462 | unsigned *nconds) |
463 | { | |
726a989a | 464 | gimple base_def; |
0f900dfa | 465 | tree base_val0; |
6b4a85ad | 466 | tree int_type; |
c2699190 | 467 | tree temp, tempn; |
726a989a RB |
468 | tree cst0; |
469 | gimple stmt1, stmt2; | |
c2699190 XDL |
470 | int bit_sz, max_exp; |
471 | inp_domain exp_domain; | |
472 | ||
473 | base_def = SSA_NAME_DEF_STMT (base); | |
726a989a | 474 | base_val0 = gimple_assign_rhs1 (base_def); |
6b4a85ad | 475 | int_type = TREE_TYPE (base_val0); |
c2699190 | 476 | bit_sz = TYPE_PRECISION (int_type); |
b8698a0f | 477 | gcc_assert (bit_sz > 0 |
c2699190 XDL |
478 | && bit_sz <= MAX_BASE_INT_BIT_SIZE); |
479 | ||
480 | /* Determine the max exp argument value according to | |
481 | the size of the base integer. The max exp value | |
482 | is conservatively estimated assuming IEEE754 double | |
483 | precision format. */ | |
484 | if (bit_sz == 8) | |
485 | max_exp = 128; | |
486 | else if (bit_sz == 16) | |
487 | max_exp = 64; | |
488 | else | |
0f900dfa JJ |
489 | { |
490 | gcc_assert (bit_sz == MAX_BASE_INT_BIT_SIZE); | |
491 | max_exp = 32; | |
492 | } | |
c2699190 XDL |
493 | |
494 | /* For pow ((double)x, y), generate the following conditions: | |
495 | cond 1: | |
496 | temp1 = x; | |
497 | if (temp1 <= 0) | |
498 | ||
499 | cond 2: | |
500 | temp2 = y; | |
501 | if (temp2 > max_exp_real_cst) */ | |
502 | ||
503 | /* Generate condition in reverse order -- first | |
504 | the condition for the exp argument. */ | |
505 | ||
506 | exp_domain = get_domain (0, false, false, | |
507 | max_exp, true, true); | |
508 | ||
509 | gen_conditions_for_domain (expn, exp_domain, | |
510 | conds, nconds); | |
511 | ||
512 | /* Now generate condition for the base argument. | |
513 | Note it does not use the helper function | |
514 | gen_conditions_for_domain because the base | |
515 | type is integer. */ | |
516 | ||
517 | /* Push a separator. */ | |
9771b263 | 518 | conds.quick_push (NULL); |
c2699190 XDL |
519 | |
520 | temp = create_tmp_var (int_type, "DCE_COND1"); | |
521 | cst0 = build_int_cst (int_type, 0); | |
726a989a | 522 | stmt1 = gimple_build_assign (temp, base_val0); |
c2699190 | 523 | tempn = make_ssa_name (temp, stmt1); |
726a989a RB |
524 | gimple_assign_set_lhs (stmt1, tempn); |
525 | stmt2 = gimple_build_cond (LE_EXPR, tempn, cst0, NULL_TREE, NULL_TREE); | |
c2699190 | 526 | |
9771b263 DN |
527 | conds.quick_push (stmt1); |
528 | conds.quick_push (stmt2); | |
c2699190 XDL |
529 | (*nconds)++; |
530 | } | |
531 | ||
532 | /* Method to generate conditional statements for guarding conditionally | |
533 | dead calls to pow. One or more statements can be generated for | |
534 | each logical condition. Statement groups of different conditions | |
9771b263 | 535 | are separated by a NULL tree and they are stored in the vec |
c2699190 XDL |
536 | conds. The number of logical conditions are stored in *nconds. |
537 | ||
538 | See C99 standard, 7.12.7.4:2, for description of pow (x, y). | |
539 | The precise condition for domain errors are complex. In this | |
540 | implementation, a simplified (but conservative) valid domain | |
541 | for x and y are used: x is positive to avoid dom errors, while | |
542 | y is smaller than a upper bound (depending on x) to avoid range | |
543 | errors. Runtime code is generated to check x (if not constant) | |
544 | and y against the valid domain. If it is out, jump to the call, | |
545 | otherwise the call is bypassed. POW_CALL is the call statement, | |
546 | *CONDS is a vector holding the resulting condition statements, | |
547 | and *NCONDS is the number of logical conditions. */ | |
548 | ||
549 | static void | |
9771b263 | 550 | gen_conditions_for_pow (gimple pow_call, vec<gimple> conds, |
c2699190 XDL |
551 | unsigned *nconds) |
552 | { | |
553 | tree base, expn; | |
0f900dfa | 554 | enum tree_code bc; |
c2699190 | 555 | |
77a74ed7 | 556 | gcc_checking_assert (check_pow (pow_call)); |
c2699190 XDL |
557 | |
558 | *nconds = 0; | |
559 | ||
726a989a RB |
560 | base = gimple_call_arg (pow_call, 0); |
561 | expn = gimple_call_arg (pow_call, 1); | |
c2699190 XDL |
562 | |
563 | bc = TREE_CODE (base); | |
c2699190 XDL |
564 | |
565 | if (bc == REAL_CST) | |
0f900dfa | 566 | gen_conditions_for_pow_cst_base (base, expn, conds, nconds); |
c2699190 | 567 | else if (bc == SSA_NAME) |
0f900dfa | 568 | gen_conditions_for_pow_int_base (base, expn, conds, nconds); |
c2699190 XDL |
569 | else |
570 | gcc_unreachable (); | |
571 | } | |
572 | ||
573 | /* A helper routine to help computing the valid input domain | |
574 | for a builtin function. See C99 7.12.7 for details. In this | |
575 | implementation, we only handle single region domain. The | |
576 | resulting region can be conservative (smaller) than the actual | |
577 | one and rounded to integers. Some of the bounds are documented | |
578 | in the standard, while other limit constants are computed | |
b8698a0f L |
579 | assuming IEEE floating point format (for SF and DF modes). |
580 | Since IEEE only sets minimum requirements for long double format, | |
581 | different long double formats exist under different implementations | |
582 | (e.g, 64 bit double precision (DF), 80 bit double-extended | |
583 | precision (XF), and 128 bit quad precision (QF) ). For simplicity, | |
584 | in this implementation, the computed bounds for long double assume | |
585 | 64 bit format (DF), and are therefore conservative. Another | |
c2699190 | 586 | assumption is that single precision float type is always SF mode, |
b8698a0f | 587 | and double type is DF mode. This function is quite |
c2699190 XDL |
588 | implementation specific, so it may not be suitable to be part of |
589 | builtins.c. This needs to be revisited later to see if it can | |
590 | be leveraged in x87 assembly expansion. */ | |
591 | ||
592 | static inp_domain | |
593 | get_no_error_domain (enum built_in_function fnc) | |
594 | { | |
595 | switch (fnc) | |
596 | { | |
597 | /* Trig functions: return [-1, +1] */ | |
598 | CASE_FLT_FN (BUILT_IN_ACOS): | |
599 | CASE_FLT_FN (BUILT_IN_ASIN): | |
600 | return get_domain (-1, true, true, | |
601 | 1, true, true); | |
602 | /* Hyperbolic functions. */ | |
603 | CASE_FLT_FN (BUILT_IN_ACOSH): | |
604 | /* acosh: [1, +inf) */ | |
605 | return get_domain (1, true, true, | |
606 | 1, false, false); | |
607 | CASE_FLT_FN (BUILT_IN_ATANH): | |
608 | /* atanh: (-1, +1) */ | |
609 | return get_domain (-1, true, false, | |
610 | 1, true, false); | |
611 | case BUILT_IN_COSHF: | |
612 | case BUILT_IN_SINHF: | |
613 | /* coshf: (-89, +89) */ | |
614 | return get_domain (-89, true, false, | |
615 | 89, true, false); | |
616 | case BUILT_IN_COSH: | |
617 | case BUILT_IN_SINH: | |
618 | case BUILT_IN_COSHL: | |
619 | case BUILT_IN_SINHL: | |
620 | /* cosh: (-710, +710) */ | |
621 | return get_domain (-710, true, false, | |
622 | 710, true, false); | |
623 | /* Log functions: (0, +inf) */ | |
624 | CASE_FLT_FN (BUILT_IN_LOG): | |
625 | CASE_FLT_FN (BUILT_IN_LOG2): | |
626 | CASE_FLT_FN (BUILT_IN_LOG10): | |
627 | return get_domain (0, true, false, | |
628 | 0, false, false); | |
629 | CASE_FLT_FN (BUILT_IN_LOG1P): | |
630 | return get_domain (-1, true, false, | |
631 | 0, false, false); | |
632 | /* Exp functions. */ | |
633 | case BUILT_IN_EXPF: | |
634 | case BUILT_IN_EXPM1F: | |
635 | /* expf: (-inf, 88) */ | |
636 | return get_domain (-1, false, false, | |
637 | 88, true, false); | |
638 | case BUILT_IN_EXP: | |
639 | case BUILT_IN_EXPM1: | |
640 | case BUILT_IN_EXPL: | |
641 | case BUILT_IN_EXPM1L: | |
642 | /* exp: (-inf, 709) */ | |
643 | return get_domain (-1, false, false, | |
644 | 709, true, false); | |
645 | case BUILT_IN_EXP2F: | |
646 | /* exp2f: (-inf, 128) */ | |
647 | return get_domain (-1, false, false, | |
648 | 128, true, false); | |
649 | case BUILT_IN_EXP2: | |
650 | case BUILT_IN_EXP2L: | |
651 | /* exp2: (-inf, 1024) */ | |
652 | return get_domain (-1, false, false, | |
653 | 1024, true, false); | |
654 | case BUILT_IN_EXP10F: | |
655 | case BUILT_IN_POW10F: | |
656 | /* exp10f: (-inf, 38) */ | |
657 | return get_domain (-1, false, false, | |
658 | 38, true, false); | |
659 | case BUILT_IN_EXP10: | |
660 | case BUILT_IN_POW10: | |
661 | case BUILT_IN_EXP10L: | |
662 | case BUILT_IN_POW10L: | |
663 | /* exp10: (-inf, 308) */ | |
664 | return get_domain (-1, false, false, | |
665 | 308, true, false); | |
666 | /* sqrt: [0, +inf) */ | |
667 | CASE_FLT_FN (BUILT_IN_SQRT): | |
668 | return get_domain (0, true, true, | |
669 | 0, false, false); | |
670 | default: | |
b8698a0f | 671 | gcc_unreachable (); |
c2699190 XDL |
672 | } |
673 | ||
b8698a0f | 674 | gcc_unreachable (); |
c2699190 XDL |
675 | } |
676 | ||
677 | /* The function to generate shrink wrap conditions for a partially | |
678 | dead builtin call whose return value is not used anywhere, | |
679 | but has to be kept live due to potential error condition. | |
b8698a0f L |
680 | BI_CALL is the builtin call, CONDS is the vector of statements |
681 | for condition code, NCODES is the pointer to the number of | |
c2699190 XDL |
682 | logical conditions. Statements belonging to different logical |
683 | condition are separated by NULL tree in the vector. */ | |
684 | ||
685 | static void | |
9771b263 | 686 | gen_shrink_wrap_conditions (gimple bi_call, vec<gimple> conds, |
c2699190 XDL |
687 | unsigned int *nconds) |
688 | { | |
726a989a RB |
689 | gimple call; |
690 | tree fn; | |
c2699190 XDL |
691 | enum built_in_function fnc; |
692 | ||
9771b263 DN |
693 | gcc_assert (nconds && conds.exists ()); |
694 | gcc_assert (conds.length () == 0); | |
726a989a | 695 | gcc_assert (is_gimple_call (bi_call)); |
c2699190 XDL |
696 | |
697 | call = bi_call; | |
726a989a | 698 | fn = gimple_call_fndecl (call); |
c2699190 XDL |
699 | gcc_assert (fn && DECL_BUILT_IN (fn)); |
700 | fnc = DECL_FUNCTION_CODE (fn); | |
701 | *nconds = 0; | |
702 | ||
703 | if (fnc == BUILT_IN_POW) | |
704 | gen_conditions_for_pow (call, conds, nconds); | |
705 | else | |
706 | { | |
707 | tree arg; | |
708 | inp_domain domain = get_no_error_domain (fnc); | |
709 | *nconds = 0; | |
726a989a | 710 | arg = gimple_call_arg (bi_call, 0); |
c2699190 XDL |
711 | gen_conditions_for_domain (arg, domain, conds, nconds); |
712 | } | |
713 | ||
714 | return; | |
715 | } | |
716 | ||
717 | ||
718 | /* Probability of the branch (to the call) is taken. */ | |
719 | #define ERR_PROB 0.01 | |
720 | ||
b8698a0f L |
721 | /* The function to shrink wrap a partially dead builtin call |
722 | whose return value is not used anywhere, but has to be kept | |
c2699190 XDL |
723 | live due to potential error condition. Returns true if the |
724 | transformation actually happens. */ | |
725 | ||
b8698a0f | 726 | static bool |
726a989a | 727 | shrink_wrap_one_built_in_call (gimple bi_call) |
c2699190 | 728 | { |
726a989a | 729 | gimple_stmt_iterator bi_call_bsi; |
c2699190 XDL |
730 | basic_block bi_call_bb, join_tgt_bb, guard_bb, guard_bb0; |
731 | edge join_tgt_in_edge_from_call, join_tgt_in_edge_fall_thru; | |
732 | edge bi_call_in_edge0, guard_bb_in_edge; | |
c2699190 XDL |
733 | unsigned tn_cond_stmts, nconds; |
734 | unsigned ci; | |
726a989a RB |
735 | gimple cond_expr = NULL; |
736 | gimple cond_expr_start; | |
c2699190 | 737 | tree bi_call_label_decl; |
726a989a | 738 | gimple bi_call_label; |
c2699190 | 739 | |
00f96dc9 | 740 | auto_vec<gimple, 12> conds; |
c2699190 XDL |
741 | gen_shrink_wrap_conditions (bi_call, conds, &nconds); |
742 | ||
743 | /* This can happen if the condition generator decides | |
744 | it is not beneficial to do the transformation. Just | |
b8698a0f | 745 | return false and do not do any transformation for |
c2699190 XDL |
746 | the call. */ |
747 | if (nconds == 0) | |
07687835 | 748 | return false; |
c2699190 | 749 | |
726a989a | 750 | bi_call_bb = gimple_bb (bi_call); |
c2699190 | 751 | |
7a460594 JJ |
752 | /* Now find the join target bb -- split bi_call_bb if needed. */ |
753 | if (stmt_ends_bb_p (bi_call)) | |
754 | { | |
755 | /* If the call must be the last in the bb, don't split the block, | |
756 | it could e.g. have EH edges. */ | |
757 | join_tgt_in_edge_from_call = find_fallthru_edge (bi_call_bb->succs); | |
758 | if (join_tgt_in_edge_from_call == NULL) | |
07687835 | 759 | return false; |
7a460594 JJ |
760 | } |
761 | else | |
762 | join_tgt_in_edge_from_call = split_block (bi_call_bb, bi_call); | |
c2699190 | 763 | |
726a989a | 764 | bi_call_bsi = gsi_for_stmt (bi_call); |
c2699190 XDL |
765 | |
766 | join_tgt_bb = join_tgt_in_edge_from_call->dest; | |
767 | ||
768 | /* Now it is time to insert the first conditional expression | |
769 | into bi_call_bb and split this bb so that bi_call is | |
770 | shrink-wrapped. */ | |
9771b263 | 771 | tn_cond_stmts = conds.length (); |
c2699190 | 772 | cond_expr = NULL; |
9771b263 | 773 | cond_expr_start = conds[0]; |
c2699190 XDL |
774 | for (ci = 0; ci < tn_cond_stmts; ci++) |
775 | { | |
9771b263 | 776 | gimple c = conds[ci]; |
c2699190 XDL |
777 | gcc_assert (c || ci != 0); |
778 | if (!c) | |
779 | break; | |
726a989a | 780 | gsi_insert_before (&bi_call_bsi, c, GSI_SAME_STMT); |
c2699190 XDL |
781 | cond_expr = c; |
782 | } | |
783 | nconds--; | |
784 | ci++; | |
726a989a | 785 | gcc_assert (cond_expr && gimple_code (cond_expr) == GIMPLE_COND); |
c2699190 XDL |
786 | |
787 | /* Now the label. */ | |
c2255bc4 | 788 | bi_call_label_decl = create_artificial_label (gimple_location (bi_call)); |
726a989a RB |
789 | bi_call_label = gimple_build_label (bi_call_label_decl); |
790 | gsi_insert_before (&bi_call_bsi, bi_call_label, GSI_SAME_STMT); | |
c2699190 XDL |
791 | |
792 | bi_call_in_edge0 = split_block (bi_call_bb, cond_expr); | |
793 | bi_call_in_edge0->flags &= ~EDGE_FALLTHRU; | |
794 | bi_call_in_edge0->flags |= EDGE_TRUE_VALUE; | |
795 | guard_bb0 = bi_call_bb; | |
796 | bi_call_bb = bi_call_in_edge0->dest; | |
b8698a0f | 797 | join_tgt_in_edge_fall_thru = make_edge (guard_bb0, join_tgt_bb, |
c2699190 XDL |
798 | EDGE_FALSE_VALUE); |
799 | ||
800 | bi_call_in_edge0->probability = REG_BR_PROB_BASE * ERR_PROB; | |
245ef160 JH |
801 | bi_call_in_edge0->count = |
802 | apply_probability (guard_bb0->count, | |
803 | bi_call_in_edge0->probability); | |
c2699190 | 804 | join_tgt_in_edge_fall_thru->probability = |
245ef160 JH |
805 | inverse_probability (bi_call_in_edge0->probability); |
806 | join_tgt_in_edge_fall_thru->count = | |
807 | guard_bb0->count - bi_call_in_edge0->count; | |
c2699190 XDL |
808 | |
809 | /* Code generation for the rest of the conditions */ | |
810 | guard_bb = guard_bb0; | |
811 | while (nconds > 0) | |
812 | { | |
813 | unsigned ci0; | |
814 | edge bi_call_in_edge; | |
726a989a | 815 | gimple_stmt_iterator guard_bsi = gsi_for_stmt (cond_expr_start); |
c2699190 | 816 | ci0 = ci; |
9771b263 | 817 | cond_expr_start = conds[ci0]; |
c2699190 XDL |
818 | for (; ci < tn_cond_stmts; ci++) |
819 | { | |
9771b263 | 820 | gimple c = conds[ci]; |
c2699190 XDL |
821 | gcc_assert (c || ci != ci0); |
822 | if (!c) | |
823 | break; | |
726a989a | 824 | gsi_insert_before (&guard_bsi, c, GSI_SAME_STMT); |
c2699190 XDL |
825 | cond_expr = c; |
826 | } | |
827 | nconds--; | |
828 | ci++; | |
726a989a | 829 | gcc_assert (cond_expr && gimple_code (cond_expr) == GIMPLE_COND); |
c2699190 XDL |
830 | guard_bb_in_edge = split_block (guard_bb, cond_expr); |
831 | guard_bb_in_edge->flags &= ~EDGE_FALLTHRU; | |
832 | guard_bb_in_edge->flags |= EDGE_FALSE_VALUE; | |
833 | ||
834 | bi_call_in_edge = make_edge (guard_bb, bi_call_bb, EDGE_TRUE_VALUE); | |
835 | ||
836 | bi_call_in_edge->probability = REG_BR_PROB_BASE * ERR_PROB; | |
245ef160 JH |
837 | bi_call_in_edge->count = |
838 | apply_probability (guard_bb->count, | |
839 | bi_call_in_edge->probability); | |
c2699190 | 840 | guard_bb_in_edge->probability = |
245ef160 JH |
841 | inverse_probability (bi_call_in_edge->probability); |
842 | guard_bb_in_edge->count = guard_bb->count - bi_call_in_edge->count; | |
c2699190 XDL |
843 | } |
844 | ||
c2699190 XDL |
845 | if (dump_file && (dump_flags & TDF_DETAILS)) |
846 | { | |
847 | location_t loc; | |
726a989a | 848 | loc = gimple_location (bi_call); |
c2699190 XDL |
849 | fprintf (dump_file, |
850 | "%s:%d: note: function call is shrink-wrapped" | |
851 | " into error conditions.\n", | |
852 | LOCATION_FILE (loc), LOCATION_LINE (loc)); | |
853 | } | |
854 | ||
855 | return true; | |
856 | } | |
857 | ||
858 | /* The top level function for conditional dead code shrink | |
859 | wrapping transformation. */ | |
860 | ||
861 | static bool | |
9771b263 | 862 | shrink_wrap_conditional_dead_built_in_calls (vec<gimple> calls) |
c2699190 XDL |
863 | { |
864 | bool changed = false; | |
865 | unsigned i = 0; | |
866 | ||
9771b263 | 867 | unsigned n = calls.length (); |
b8698a0f | 868 | if (n == 0) |
c2699190 XDL |
869 | return false; |
870 | ||
871 | for (; i < n ; i++) | |
872 | { | |
9771b263 | 873 | gimple bi_call = calls[i]; |
c2699190 XDL |
874 | changed |= shrink_wrap_one_built_in_call (bi_call); |
875 | } | |
876 | ||
877 | return changed; | |
878 | } | |
879 | ||
be55bfe6 TS |
880 | namespace { |
881 | ||
882 | const pass_data pass_data_call_cdce = | |
883 | { | |
884 | GIMPLE_PASS, /* type */ | |
885 | "cdce", /* name */ | |
886 | OPTGROUP_NONE, /* optinfo_flags */ | |
be55bfe6 TS |
887 | TV_TREE_CALL_CDCE, /* tv_id */ |
888 | ( PROP_cfg | PROP_ssa ), /* properties_required */ | |
889 | 0, /* properties_provided */ | |
890 | 0, /* properties_destroyed */ | |
891 | 0, /* todo_flags_start */ | |
3bea341f | 892 | 0, /* todo_flags_finish */ |
be55bfe6 TS |
893 | }; |
894 | ||
895 | class pass_call_cdce : public gimple_opt_pass | |
896 | { | |
897 | public: | |
898 | pass_call_cdce (gcc::context *ctxt) | |
899 | : gimple_opt_pass (pass_data_call_cdce, ctxt) | |
900 | {} | |
901 | ||
902 | /* opt_pass methods: */ | |
903 | virtual bool gate (function *fun) | |
904 | { | |
905 | /* The limit constants used in the implementation | |
906 | assume IEEE floating point format. Other formats | |
907 | can be supported in the future if needed. */ | |
908 | return flag_tree_builtin_call_dce != 0 | |
909 | && optimize_function_for_speed_p (fun); | |
910 | } | |
c2699190 | 911 | |
be55bfe6 TS |
912 | virtual unsigned int execute (function *); |
913 | ||
914 | }; // class pass_call_cdce | |
915 | ||
916 | unsigned int | |
917 | pass_call_cdce::execute (function *fun) | |
c2699190 XDL |
918 | { |
919 | basic_block bb; | |
726a989a | 920 | gimple_stmt_iterator i; |
c2699190 | 921 | bool something_changed = false; |
ef062b13 | 922 | auto_vec<gimple> cond_dead_built_in_calls; |
be55bfe6 | 923 | FOR_EACH_BB_FN (bb, fun) |
c2699190 XDL |
924 | { |
925 | /* Collect dead call candidates. */ | |
726a989a | 926 | for (i = gsi_start_bb (bb); !gsi_end_p (i); gsi_next (&i)) |
c2699190 | 927 | { |
726a989a RB |
928 | gimple stmt = gsi_stmt (i); |
929 | if (is_gimple_call (stmt) | |
c2699190 XDL |
930 | && is_call_dce_candidate (stmt)) |
931 | { | |
932 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
933 | { | |
934 | fprintf (dump_file, "Found conditional dead call: "); | |
726a989a | 935 | print_gimple_stmt (dump_file, stmt, 0, TDF_SLIM); |
c2699190 XDL |
936 | fprintf (dump_file, "\n"); |
937 | } | |
9771b263 DN |
938 | if (!cond_dead_built_in_calls.exists ()) |
939 | cond_dead_built_in_calls.create (64); | |
940 | cond_dead_built_in_calls.safe_push (stmt); | |
c2699190 XDL |
941 | } |
942 | } | |
943 | } | |
944 | ||
9771b263 | 945 | if (!cond_dead_built_in_calls.exists ()) |
6b672a29 JJ |
946 | return 0; |
947 | ||
948 | something_changed | |
949 | = shrink_wrap_conditional_dead_built_in_calls (cond_dead_built_in_calls); | |
c2699190 | 950 | |
c2699190 XDL |
951 | if (something_changed) |
952 | { | |
953 | free_dominance_info (CDI_DOMINATORS); | |
954 | free_dominance_info (CDI_POST_DOMINATORS); | |
5006671f RG |
955 | /* As we introduced new control-flow we need to insert PHI-nodes |
956 | for the call-clobbers of the remaining call. */ | |
be55bfe6 | 957 | mark_virtual_operands_for_renaming (fun); |
c0e50f72 | 958 | return TODO_update_ssa; |
c2699190 | 959 | } |
c0e50f72 RB |
960 | |
961 | return 0; | |
c2699190 XDL |
962 | } |
963 | ||
27a4cd48 DM |
964 | } // anon namespace |
965 | ||
966 | gimple_opt_pass * | |
967 | make_pass_call_cdce (gcc::context *ctxt) | |
968 | { | |
969 | return new pass_call_cdce (ctxt); | |
970 | } |