Mercurial > hg > CbC > CbC_gcc
annotate gcc/tree-call-cdce.c @ 158:494b0b89df80 default tip
...
author | Shinji KONO <kono@ie.u-ryukyu.ac.jp> |
---|---|
date | Mon, 25 May 2020 18:13:55 +0900 |
parents | 1830386684a0 |
children |
rev | line source |
---|---|
0 | 1 /* Conditional Dead Call Elimination pass for the GNU compiler. |
145 | 2 Copyright (C) 2008-2020 Free Software Foundation, Inc. |
0 | 3 Contributed by Xinliang David Li <davidxl@google.com> |
4 | |
5 This file is part of GCC. | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
6 |
0 | 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. | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
11 |
0 | 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. | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
16 |
0 | 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" | |
111 | 24 #include "backend.h" |
0 | 25 #include "tree.h" |
111 | 26 #include "gimple.h" |
27 #include "cfghooks.h" | |
28 #include "tree-pass.h" | |
29 #include "ssa.h" | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
30 #include "gimple-pretty-print.h" |
111 | 31 #include "fold-const.h" |
32 #include "stor-layout.h" | |
33 #include "gimple-iterator.h" | |
34 #include "tree-cfg.h" | |
35 #include "tree-into-ssa.h" | |
36 #include "builtins.h" | |
37 #include "internal-fn.h" | |
38 #include "tree-dfa.h" | |
0 | 39 |
40 | |
111 | 41 /* This pass serves two closely-related purposes: |
42 | |
43 1. It conditionally executes calls that set errno if (a) the result of | |
44 the call is unused and (b) a simple range check on the arguments can | |
45 detect most cases where errno does not need to be set. | |
46 | |
47 This is the "conditional dead-code elimination" that gave the pass | |
48 its original name, since the call is dead for most argument values. | |
49 The calls for which it helps are usually part of the C++ abstraction | |
50 penalty exposed after inlining. | |
0 | 51 |
111 | 52 2. It looks for calls to built-in functions that set errno and whose |
53 result is used. It checks whether there is an associated internal | |
54 function that doesn't set errno and whether the target supports | |
55 that internal function. If so, the pass uses the internal function | |
56 to compute the result of the built-in function but still arranges | |
57 for errno to be set when necessary. There are two ways of setting | |
58 errno: | |
59 | |
60 a. by protecting the original call with the same argument checks as (1) | |
61 | |
62 b. by protecting the original call with a check that the result | |
63 of the internal function is not equal to itself (i.e. is NaN). | |
0 | 64 |
111 | 65 (b) requires that NaNs are the only erroneous results. It is not |
66 appropriate for functions like log, which returns ERANGE for zero | |
67 arguments. (b) is also likely to perform worse than (a) because it | |
68 requires the result to be calculated first. The pass therefore uses | |
69 (a) when it can and uses (b) as a fallback. | |
70 | |
71 For (b) the pass can replace the original call with a call to | |
72 IFN_SET_EDOM, if the target supports direct assignments to errno. | |
0 | 73 |
111 | 74 In both cases, arguments that require errno to be set should occur |
75 rarely in practice. Checks of the errno result should also be rare, | |
76 but the compiler would need powerful interprocedural analysis to | |
77 prove that errno is not checked. It's much easier to add argument | |
78 checks or result checks instead. | |
79 | |
80 An example of (1) is: | |
81 | |
82 log (x); // Mostly dead call | |
0 | 83 ==> |
111 | 84 if (__builtin_islessequal (x, 0)) |
85 log (x); | |
86 | |
0 | 87 With this change, call to log (x) is effectively eliminated, as |
111 | 88 in the majority of the cases, log won't be called with x out of |
0 | 89 range. The branch is totally predictable, so the branch cost |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
90 is low. |
0 | 91 |
111 | 92 An example of (2) is: |
93 | |
94 y = sqrt (x); | |
95 ==> | |
96 if (__builtin_isless (x, 0)) | |
145 | 97 y = sqrt (x); |
98 else | |
99 y = IFN_SQRT (x); | |
111 | 100 In the vast majority of cases we should then never need to call sqrt. |
101 | |
0 | 102 Note that library functions are not supposed to clear errno to zero without |
103 error. See IEEE Std 1003.1, section 2.3 Error Numbers, and section 7.5:3 of | |
104 ISO/IEC 9899 (C99). | |
105 | |
106 The condition wrapping the builtin call is conservatively set to avoid too | |
111 | 107 aggressive (wrong) shrink wrapping. */ |
0 | 108 |
109 | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
110 /* A structure for representing input domain of |
0 | 111 a function argument in integer. If the lower |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
112 bound is -inf, has_lb is set to false. If the |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
113 upper bound is +inf, has_ub is false. |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
114 is_lb_inclusive and is_ub_inclusive are flags |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
115 to indicate if lb and ub value are inclusive |
0 | 116 respectively. */ |
117 | |
111 | 118 struct inp_domain |
0 | 119 { |
120 int lb; | |
121 int ub; | |
122 bool has_lb; | |
123 bool has_ub; | |
124 bool is_lb_inclusive; | |
125 bool is_ub_inclusive; | |
111 | 126 }; |
0 | 127 |
128 /* A helper function to construct and return an input | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
129 domain object. LB is the lower bound, HAS_LB is |
0 | 130 a boolean flag indicating if the lower bound exists, |
131 and LB_INCLUSIVE is a boolean flag indicating if the | |
132 lower bound is inclusive or not. UB, HAS_UB, and | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
133 UB_INCLUSIVE have the same meaning, but for upper |
0 | 134 bound of the domain. */ |
135 | |
136 static inp_domain | |
137 get_domain (int lb, bool has_lb, bool lb_inclusive, | |
138 int ub, bool has_ub, bool ub_inclusive) | |
139 { | |
140 inp_domain domain; | |
141 domain.lb = lb; | |
142 domain.has_lb = has_lb; | |
143 domain.is_lb_inclusive = lb_inclusive; | |
144 domain.ub = ub; | |
145 domain.has_ub = has_ub; | |
146 domain.is_ub_inclusive = ub_inclusive; | |
147 return domain; | |
148 } | |
149 | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
150 /* A helper function to check the target format for the |
0 | 151 argument type. In this implementation, only IEEE formats |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
152 are supported. ARG is the call argument to be checked. |
0 | 153 Returns true if the format is supported. To support other |
154 target formats, function get_no_error_domain needs to be | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
155 enhanced to have range bounds properly computed. Since |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
156 the check is cheap (very small number of candidates |
0 | 157 to be checked), the result is not cached for each float type. */ |
158 | |
159 static bool | |
160 check_target_format (tree arg) | |
161 { | |
162 tree type; | |
111 | 163 machine_mode mode; |
0 | 164 const struct real_format *rfmt; |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
165 |
0 | 166 type = TREE_TYPE (arg); |
167 mode = TYPE_MODE (type); | |
168 rfmt = REAL_MODE_FORMAT (mode); | |
169 if ((mode == SFmode | |
170 && (rfmt == &ieee_single_format || rfmt == &mips_single_format | |
171 || rfmt == &motorola_single_format)) | |
172 || (mode == DFmode | |
173 && (rfmt == &ieee_double_format || rfmt == &mips_double_format | |
174 || rfmt == &motorola_double_format)) | |
145 | 175 /* For long double, we cannot really check XFmode |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
176 which is only defined on intel platforms. |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
177 Candidate pre-selection using builtin function |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
178 code guarantees that we are checking formats |
0 | 179 for long double modes: double, quad, and extended. */ |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
180 || (mode != SFmode && mode != DFmode |
0 | 181 && (rfmt == &ieee_quad_format |
182 || rfmt == &mips_quad_format | |
183 || rfmt == &ieee_extended_motorola_format | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
184 || rfmt == &ieee_extended_intel_96_format |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
185 || rfmt == &ieee_extended_intel_128_format |
0 | 186 || rfmt == &ieee_extended_intel_96_round_53_format))) |
187 return true; | |
188 | |
189 return false; | |
190 } | |
191 | |
192 | |
193 /* A helper function to help select calls to pow that are suitable for | |
194 conditional DCE transformation. It looks for pow calls that can be | |
195 guided with simple conditions. Such calls either have constant base | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
196 values or base values converted from integers. Returns true if |
0 | 197 the pow call POW_CALL is a candidate. */ |
198 | |
199 /* The maximum integer bit size for base argument of a pow call | |
200 that is suitable for shrink-wrapping transformation. */ | |
201 #define MAX_BASE_INT_BIT_SIZE 32 | |
202 | |
203 static bool | |
111 | 204 check_pow (gcall *pow_call) |
0 | 205 { |
206 tree base, expn; | |
207 enum tree_code bc, ec; | |
208 | |
209 if (gimple_call_num_args (pow_call) != 2) | |
210 return false; | |
211 | |
212 base = gimple_call_arg (pow_call, 0); | |
213 expn = gimple_call_arg (pow_call, 1); | |
214 | |
215 if (!check_target_format (expn)) | |
216 return false; | |
217 | |
218 bc = TREE_CODE (base); | |
219 ec = TREE_CODE (expn); | |
220 | |
221 /* Folding candidates are not interesting. | |
222 Can actually assert that it is already folded. */ | |
223 if (ec == REAL_CST && bc == REAL_CST) | |
224 return false; | |
225 | |
226 if (bc == REAL_CST) | |
227 { | |
228 /* Only handle a fixed range of constant. */ | |
229 REAL_VALUE_TYPE mv; | |
230 REAL_VALUE_TYPE bcv = TREE_REAL_CST (base); | |
111 | 231 if (real_equal (&bcv, &dconst1)) |
0 | 232 return false; |
111 | 233 if (real_less (&bcv, &dconst1)) |
0 | 234 return false; |
111 | 235 real_from_integer (&mv, TYPE_MODE (TREE_TYPE (base)), 256, UNSIGNED); |
236 if (real_less (&mv, &bcv)) | |
0 | 237 return false; |
238 return true; | |
239 } | |
240 else if (bc == SSA_NAME) | |
241 { | |
111 | 242 tree base_val0, type; |
243 gimple *base_def; | |
0 | 244 int bit_sz; |
245 | |
246 /* Only handles cases where base value is converted | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
247 from integer values. */ |
0 | 248 base_def = SSA_NAME_DEF_STMT (base); |
249 if (gimple_code (base_def) != GIMPLE_ASSIGN) | |
250 return false; | |
251 | |
252 if (gimple_assign_rhs_code (base_def) != FLOAT_EXPR) | |
253 return false; | |
254 base_val0 = gimple_assign_rhs1 (base_def); | |
255 | |
111 | 256 type = TREE_TYPE (base_val0); |
0 | 257 if (TREE_CODE (type) != INTEGER_TYPE) |
258 return false; | |
259 bit_sz = TYPE_PRECISION (type); | |
260 /* If the type of the base is too wide, | |
261 the resulting shrink wrapping condition | |
262 will be too conservative. */ | |
263 if (bit_sz > MAX_BASE_INT_BIT_SIZE) | |
264 return false; | |
265 | |
266 return true; | |
267 } | |
268 else | |
269 return false; | |
270 } | |
271 | |
272 /* A helper function to help select candidate function calls that are | |
273 suitable for conditional DCE. Candidate functions must have single | |
274 valid input domain in this implementation except for pow (see check_pow). | |
275 Returns true if the function call is a candidate. */ | |
276 | |
277 static bool | |
111 | 278 check_builtin_call (gcall *bcall) |
0 | 279 { |
280 tree arg; | |
281 | |
282 arg = gimple_call_arg (bcall, 0); | |
283 return check_target_format (arg); | |
284 } | |
285 | |
111 | 286 /* Return true if built-in function call CALL calls a math function |
287 and if we know how to test the range of its arguments to detect _most_ | |
288 situations in which errno is not set. The test must err on the side | |
289 of treating non-erroneous values as potentially erroneous. */ | |
0 | 290 |
291 static bool | |
111 | 292 can_test_argument_range (gcall *call) |
0 | 293 { |
111 | 294 switch (DECL_FUNCTION_CODE (gimple_call_fndecl (call))) |
0 | 295 { |
296 /* Trig functions. */ | |
297 CASE_FLT_FN (BUILT_IN_ACOS): | |
298 CASE_FLT_FN (BUILT_IN_ASIN): | |
299 /* Hyperbolic functions. */ | |
300 CASE_FLT_FN (BUILT_IN_ACOSH): | |
301 CASE_FLT_FN (BUILT_IN_ATANH): | |
302 CASE_FLT_FN (BUILT_IN_COSH): | |
303 CASE_FLT_FN (BUILT_IN_SINH): | |
304 /* Log functions. */ | |
305 CASE_FLT_FN (BUILT_IN_LOG): | |
306 CASE_FLT_FN (BUILT_IN_LOG2): | |
307 CASE_FLT_FN (BUILT_IN_LOG10): | |
308 CASE_FLT_FN (BUILT_IN_LOG1P): | |
309 /* Exp functions. */ | |
310 CASE_FLT_FN (BUILT_IN_EXP): | |
311 CASE_FLT_FN (BUILT_IN_EXP2): | |
312 CASE_FLT_FN (BUILT_IN_EXP10): | |
313 CASE_FLT_FN (BUILT_IN_EXPM1): | |
314 CASE_FLT_FN (BUILT_IN_POW10): | |
315 /* Sqrt. */ | |
316 CASE_FLT_FN (BUILT_IN_SQRT): | |
131 | 317 CASE_FLT_FN_FLOATN_NX (BUILT_IN_SQRT): |
0 | 318 return check_builtin_call (call); |
319 /* Special one: two argument pow. */ | |
320 case BUILT_IN_POW: | |
321 return check_pow (call); | |
322 default: | |
323 break; | |
324 } | |
325 | |
326 return false; | |
327 } | |
328 | |
111 | 329 /* Return true if CALL can produce a domain error (EDOM) but can never |
330 produce a pole, range overflow or range underflow error (all ERANGE). | |
331 This means that we can tell whether a function would have set errno | |
332 by testing whether the result is a NaN. */ | |
333 | |
334 static bool | |
335 edom_only_function (gcall *call) | |
336 { | |
337 switch (DECL_FUNCTION_CODE (gimple_call_fndecl (call))) | |
338 { | |
339 CASE_FLT_FN (BUILT_IN_ACOS): | |
340 CASE_FLT_FN (BUILT_IN_ASIN): | |
341 CASE_FLT_FN (BUILT_IN_ATAN): | |
342 CASE_FLT_FN (BUILT_IN_COS): | |
343 CASE_FLT_FN (BUILT_IN_SIGNIFICAND): | |
344 CASE_FLT_FN (BUILT_IN_SIN): | |
345 CASE_FLT_FN (BUILT_IN_SQRT): | |
131 | 346 CASE_FLT_FN_FLOATN_NX (BUILT_IN_SQRT): |
111 | 347 CASE_FLT_FN (BUILT_IN_FMOD): |
348 CASE_FLT_FN (BUILT_IN_REMAINDER): | |
349 return true; | |
350 | |
351 default: | |
352 return false; | |
353 } | |
354 } | |
355 | |
356 /* Return true if it is structurally possible to guard CALL. */ | |
357 | |
358 static bool | |
359 can_guard_call_p (gimple *call) | |
360 { | |
361 return (!stmt_ends_bb_p (call) | |
362 || find_fallthru_edge (gimple_bb (call)->succs)); | |
363 } | |
0 | 364 |
145 | 365 /* For a comparison code return the comparison code we should use if we don't |
366 HONOR_NANS. */ | |
367 | |
368 static enum tree_code | |
369 comparison_code_if_no_nans (tree_code code) | |
370 { | |
371 switch (code) | |
372 { | |
373 case UNLT_EXPR: | |
374 return LT_EXPR; | |
375 case UNGT_EXPR: | |
376 return GT_EXPR; | |
377 case UNLE_EXPR: | |
378 return LE_EXPR; | |
379 case UNGE_EXPR: | |
380 return GE_EXPR; | |
381 case UNEQ_EXPR: | |
382 return EQ_EXPR; | |
383 case LTGT_EXPR: | |
384 return NE_EXPR; | |
385 | |
386 case LT_EXPR: | |
387 case GT_EXPR: | |
388 case LE_EXPR: | |
389 case GE_EXPR: | |
390 case EQ_EXPR: | |
391 case NE_EXPR: | |
392 return code; | |
393 | |
394 default: | |
395 gcc_unreachable (); | |
396 } | |
397 } | |
398 | |
111 | 399 /* A helper function to generate gimple statements for one bound |
400 comparison, so that the built-in function is called whenever | |
401 TCODE <ARG, LBUB> is *false*. TEMP_NAME1/TEMP_NAME2 are names | |
402 of the temporaries, CONDS is a vector holding the produced GIMPLE | |
403 statements, and NCONDS points to the variable holding the number of | |
404 logical comparisons. CONDS is either empty or a list ended with a | |
405 null tree. */ | |
0 | 406 |
407 static void | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
408 gen_one_condition (tree arg, int lbub, |
0 | 409 enum tree_code tcode, |
410 const char *temp_name1, | |
411 const char *temp_name2, | |
111 | 412 vec<gimple *> conds, |
0 | 413 unsigned *nconds) |
414 { | |
145 | 415 if (!HONOR_NANS (arg)) |
416 tcode = comparison_code_if_no_nans (tcode); | |
417 | |
0 | 418 tree lbub_real_cst, lbub_cst, float_type; |
419 tree temp, tempn, tempc, tempcn; | |
111 | 420 gassign *stmt1; |
421 gassign *stmt2; | |
422 gcond *stmt3; | |
0 | 423 |
424 float_type = TREE_TYPE (arg); | |
425 lbub_cst = build_int_cst (integer_type_node, lbub); | |
426 lbub_real_cst = build_real_from_int_cst (float_type, lbub_cst); | |
427 | |
428 temp = create_tmp_var (float_type, temp_name1); | |
429 stmt1 = gimple_build_assign (temp, arg); | |
430 tempn = make_ssa_name (temp, stmt1); | |
431 gimple_assign_set_lhs (stmt1, tempn); | |
432 | |
433 tempc = create_tmp_var (boolean_type_node, temp_name2); | |
434 stmt2 = gimple_build_assign (tempc, | |
435 fold_build2 (tcode, | |
436 boolean_type_node, | |
437 tempn, lbub_real_cst)); | |
438 tempcn = make_ssa_name (tempc, stmt2); | |
439 gimple_assign_set_lhs (stmt2, tempcn); | |
440 | |
441 stmt3 = gimple_build_cond_from_tree (tempcn, NULL_TREE, NULL_TREE); | |
111 | 442 conds.quick_push (stmt1); |
443 conds.quick_push (stmt2); | |
444 conds.quick_push (stmt3); | |
0 | 445 (*nconds)++; |
446 } | |
447 | |
448 /* A helper function to generate GIMPLE statements for | |
449 out of input domain check. ARG is the call argument | |
450 to be runtime checked, DOMAIN holds the valid domain | |
451 for the given function, CONDS points to the vector | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
452 holding the result GIMPLE statements. *NCONDS is |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
453 the number of logical comparisons. This function |
0 | 454 produces no more than two logical comparisons, one |
455 for lower bound check, one for upper bound check. */ | |
456 | |
457 static void | |
458 gen_conditions_for_domain (tree arg, inp_domain domain, | |
111 | 459 vec<gimple *> conds, |
0 | 460 unsigned *nconds) |
461 { | |
462 if (domain.has_lb) | |
463 gen_one_condition (arg, domain.lb, | |
464 (domain.is_lb_inclusive | |
111 | 465 ? UNGE_EXPR : UNGT_EXPR), |
0 | 466 "DCE_COND_LB", "DCE_COND_LB_TEST", |
467 conds, nconds); | |
468 | |
469 if (domain.has_ub) | |
470 { | |
471 /* Now push a separator. */ | |
472 if (domain.has_lb) | |
111 | 473 conds.quick_push (NULL); |
0 | 474 |
475 gen_one_condition (arg, domain.ub, | |
476 (domain.is_ub_inclusive | |
111 | 477 ? UNLE_EXPR : UNLT_EXPR), |
0 | 478 "DCE_COND_UB", "DCE_COND_UB_TEST", |
479 conds, nconds); | |
480 } | |
481 } | |
482 | |
483 | |
484 /* A helper function to generate condition | |
485 code for the y argument in call pow (some_const, y). | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
486 See candidate selection in check_pow. Since the |
0 | 487 candidates' base values have a limited range, |
488 the guarded code generated for y are simple: | |
111 | 489 if (__builtin_isgreater (y, max_y)) |
0 | 490 pow (const, y); |
491 Note max_y can be computed separately for each | |
492 const base, but in this implementation, we | |
493 choose to compute it using the max base | |
494 in the allowed range for the purpose of | |
495 simplicity. BASE is the constant base value, | |
496 EXPN is the expression for the exponent argument, | |
497 *CONDS is the vector to hold resulting statements, | |
498 and *NCONDS is the number of logical conditions. */ | |
499 | |
500 static void | |
501 gen_conditions_for_pow_cst_base (tree base, tree expn, | |
111 | 502 vec<gimple *> conds, |
0 | 503 unsigned *nconds) |
504 { | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
505 inp_domain exp_domain; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
506 /* Validate the range of the base constant to make |
0 | 507 sure it is consistent with check_pow. */ |
508 REAL_VALUE_TYPE mv; | |
509 REAL_VALUE_TYPE bcv = TREE_REAL_CST (base); | |
111 | 510 gcc_assert (!real_equal (&bcv, &dconst1) |
511 && !real_less (&bcv, &dconst1)); | |
512 real_from_integer (&mv, TYPE_MODE (TREE_TYPE (base)), 256, UNSIGNED); | |
513 gcc_assert (!real_less (&mv, &bcv)); | |
0 | 514 |
515 exp_domain = get_domain (0, false, false, | |
516 127, true, false); | |
517 | |
518 gen_conditions_for_domain (expn, exp_domain, | |
519 conds, nconds); | |
520 } | |
521 | |
522 /* Generate error condition code for pow calls with | |
523 non constant base values. The candidates selected | |
524 have their base argument value converted from | |
525 integer (see check_pow) value (1, 2, 4 bytes), and | |
526 the max exp value is computed based on the size | |
527 of the integer type (i.e. max possible base value). | |
528 The resulting input domain for exp argument is thus | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
529 conservative (smaller than the max value allowed by |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
530 the runtime value of the base). BASE is the integer |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
531 base value, EXPN is the expression for the exponent |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
532 argument, *CONDS is the vector to hold resulting |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
533 statements, and *NCONDS is the number of logical |
0 | 534 conditions. */ |
535 | |
536 static void | |
537 gen_conditions_for_pow_int_base (tree base, tree expn, | |
111 | 538 vec<gimple *> conds, |
0 | 539 unsigned *nconds) |
540 { | |
111 | 541 gimple *base_def; |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
542 tree base_val0; |
111 | 543 tree int_type; |
0 | 544 tree temp, tempn; |
545 tree cst0; | |
111 | 546 gimple *stmt1, *stmt2; |
0 | 547 int bit_sz, max_exp; |
548 inp_domain exp_domain; | |
549 | |
550 base_def = SSA_NAME_DEF_STMT (base); | |
551 base_val0 = gimple_assign_rhs1 (base_def); | |
111 | 552 int_type = TREE_TYPE (base_val0); |
0 | 553 bit_sz = TYPE_PRECISION (int_type); |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
554 gcc_assert (bit_sz > 0 |
0 | 555 && bit_sz <= MAX_BASE_INT_BIT_SIZE); |
556 | |
557 /* Determine the max exp argument value according to | |
558 the size of the base integer. The max exp value | |
559 is conservatively estimated assuming IEEE754 double | |
560 precision format. */ | |
561 if (bit_sz == 8) | |
562 max_exp = 128; | |
563 else if (bit_sz == 16) | |
564 max_exp = 64; | |
565 else | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
566 { |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
567 gcc_assert (bit_sz == MAX_BASE_INT_BIT_SIZE); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
568 max_exp = 32; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
569 } |
0 | 570 |
571 /* For pow ((double)x, y), generate the following conditions: | |
572 cond 1: | |
573 temp1 = x; | |
111 | 574 if (__builtin_islessequal (temp1, 0)) |
0 | 575 |
576 cond 2: | |
577 temp2 = y; | |
111 | 578 if (__builtin_isgreater (temp2, max_exp_real_cst)) */ |
0 | 579 |
580 /* Generate condition in reverse order -- first | |
581 the condition for the exp argument. */ | |
582 | |
583 exp_domain = get_domain (0, false, false, | |
584 max_exp, true, true); | |
585 | |
586 gen_conditions_for_domain (expn, exp_domain, | |
587 conds, nconds); | |
588 | |
589 /* Now generate condition for the base argument. | |
590 Note it does not use the helper function | |
591 gen_conditions_for_domain because the base | |
592 type is integer. */ | |
593 | |
594 /* Push a separator. */ | |
111 | 595 conds.quick_push (NULL); |
0 | 596 |
597 temp = create_tmp_var (int_type, "DCE_COND1"); | |
598 cst0 = build_int_cst (int_type, 0); | |
599 stmt1 = gimple_build_assign (temp, base_val0); | |
600 tempn = make_ssa_name (temp, stmt1); | |
601 gimple_assign_set_lhs (stmt1, tempn); | |
111 | 602 stmt2 = gimple_build_cond (GT_EXPR, tempn, cst0, NULL_TREE, NULL_TREE); |
0 | 603 |
111 | 604 conds.quick_push (stmt1); |
605 conds.quick_push (stmt2); | |
0 | 606 (*nconds)++; |
607 } | |
608 | |
609 /* Method to generate conditional statements for guarding conditionally | |
610 dead calls to pow. One or more statements can be generated for | |
611 each logical condition. Statement groups of different conditions | |
111 | 612 are separated by a NULL tree and they are stored in the vec |
0 | 613 conds. The number of logical conditions are stored in *nconds. |
614 | |
615 See C99 standard, 7.12.7.4:2, for description of pow (x, y). | |
616 The precise condition for domain errors are complex. In this | |
617 implementation, a simplified (but conservative) valid domain | |
618 for x and y are used: x is positive to avoid dom errors, while | |
619 y is smaller than a upper bound (depending on x) to avoid range | |
620 errors. Runtime code is generated to check x (if not constant) | |
621 and y against the valid domain. If it is out, jump to the call, | |
622 otherwise the call is bypassed. POW_CALL is the call statement, | |
623 *CONDS is a vector holding the resulting condition statements, | |
624 and *NCONDS is the number of logical conditions. */ | |
625 | |
626 static void | |
111 | 627 gen_conditions_for_pow (gcall *pow_call, vec<gimple *> conds, |
0 | 628 unsigned *nconds) |
629 { | |
630 tree base, expn; | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
631 enum tree_code bc; |
0 | 632 |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
633 gcc_checking_assert (check_pow (pow_call)); |
0 | 634 |
635 *nconds = 0; | |
636 | |
637 base = gimple_call_arg (pow_call, 0); | |
638 expn = gimple_call_arg (pow_call, 1); | |
639 | |
640 bc = TREE_CODE (base); | |
641 | |
642 if (bc == REAL_CST) | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
643 gen_conditions_for_pow_cst_base (base, expn, conds, nconds); |
0 | 644 else if (bc == SSA_NAME) |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
645 gen_conditions_for_pow_int_base (base, expn, conds, nconds); |
0 | 646 else |
647 gcc_unreachable (); | |
648 } | |
649 | |
650 /* A helper routine to help computing the valid input domain | |
651 for a builtin function. See C99 7.12.7 for details. In this | |
652 implementation, we only handle single region domain. The | |
653 resulting region can be conservative (smaller) than the actual | |
654 one and rounded to integers. Some of the bounds are documented | |
655 in the standard, while other limit constants are computed | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
656 assuming IEEE floating point format (for SF and DF modes). |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
657 Since IEEE only sets minimum requirements for long double format, |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
658 different long double formats exist under different implementations |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
659 (e.g, 64 bit double precision (DF), 80 bit double-extended |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
660 precision (XF), and 128 bit quad precision (QF) ). For simplicity, |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
661 in this implementation, the computed bounds for long double assume |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
662 64 bit format (DF), and are therefore conservative. Another |
0 | 663 assumption is that single precision float type is always SF mode, |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
664 and double type is DF mode. This function is quite |
0 | 665 implementation specific, so it may not be suitable to be part of |
666 builtins.c. This needs to be revisited later to see if it can | |
667 be leveraged in x87 assembly expansion. */ | |
668 | |
669 static inp_domain | |
670 get_no_error_domain (enum built_in_function fnc) | |
671 { | |
672 switch (fnc) | |
673 { | |
674 /* Trig functions: return [-1, +1] */ | |
675 CASE_FLT_FN (BUILT_IN_ACOS): | |
676 CASE_FLT_FN (BUILT_IN_ASIN): | |
677 return get_domain (-1, true, true, | |
678 1, true, true); | |
679 /* Hyperbolic functions. */ | |
680 CASE_FLT_FN (BUILT_IN_ACOSH): | |
681 /* acosh: [1, +inf) */ | |
682 return get_domain (1, true, true, | |
683 1, false, false); | |
684 CASE_FLT_FN (BUILT_IN_ATANH): | |
685 /* atanh: (-1, +1) */ | |
686 return get_domain (-1, true, false, | |
687 1, true, false); | |
688 case BUILT_IN_COSHF: | |
689 case BUILT_IN_SINHF: | |
690 /* coshf: (-89, +89) */ | |
691 return get_domain (-89, true, false, | |
692 89, true, false); | |
693 case BUILT_IN_COSH: | |
694 case BUILT_IN_SINH: | |
695 case BUILT_IN_COSHL: | |
696 case BUILT_IN_SINHL: | |
697 /* cosh: (-710, +710) */ | |
698 return get_domain (-710, true, false, | |
699 710, true, false); | |
700 /* Log functions: (0, +inf) */ | |
701 CASE_FLT_FN (BUILT_IN_LOG): | |
702 CASE_FLT_FN (BUILT_IN_LOG2): | |
703 CASE_FLT_FN (BUILT_IN_LOG10): | |
704 return get_domain (0, true, false, | |
705 0, false, false); | |
706 CASE_FLT_FN (BUILT_IN_LOG1P): | |
707 return get_domain (-1, true, false, | |
708 0, false, false); | |
709 /* Exp functions. */ | |
710 case BUILT_IN_EXPF: | |
711 case BUILT_IN_EXPM1F: | |
712 /* expf: (-inf, 88) */ | |
713 return get_domain (-1, false, false, | |
714 88, true, false); | |
715 case BUILT_IN_EXP: | |
716 case BUILT_IN_EXPM1: | |
717 case BUILT_IN_EXPL: | |
718 case BUILT_IN_EXPM1L: | |
719 /* exp: (-inf, 709) */ | |
720 return get_domain (-1, false, false, | |
721 709, true, false); | |
722 case BUILT_IN_EXP2F: | |
723 /* exp2f: (-inf, 128) */ | |
724 return get_domain (-1, false, false, | |
725 128, true, false); | |
726 case BUILT_IN_EXP2: | |
727 case BUILT_IN_EXP2L: | |
728 /* exp2: (-inf, 1024) */ | |
729 return get_domain (-1, false, false, | |
730 1024, true, false); | |
731 case BUILT_IN_EXP10F: | |
732 case BUILT_IN_POW10F: | |
733 /* exp10f: (-inf, 38) */ | |
734 return get_domain (-1, false, false, | |
735 38, true, false); | |
736 case BUILT_IN_EXP10: | |
737 case BUILT_IN_POW10: | |
738 case BUILT_IN_EXP10L: | |
739 case BUILT_IN_POW10L: | |
740 /* exp10: (-inf, 308) */ | |
741 return get_domain (-1, false, false, | |
742 308, true, false); | |
743 /* sqrt: [0, +inf) */ | |
744 CASE_FLT_FN (BUILT_IN_SQRT): | |
131 | 745 CASE_FLT_FN_FLOATN_NX (BUILT_IN_SQRT): |
0 | 746 return get_domain (0, true, true, |
747 0, false, false); | |
748 default: | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
749 gcc_unreachable (); |
0 | 750 } |
751 | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
752 gcc_unreachable (); |
0 | 753 } |
754 | |
755 /* The function to generate shrink wrap conditions for a partially | |
756 dead builtin call whose return value is not used anywhere, | |
757 but has to be kept live due to potential error condition. | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
758 BI_CALL is the builtin call, CONDS is the vector of statements |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
759 for condition code, NCODES is the pointer to the number of |
0 | 760 logical conditions. Statements belonging to different logical |
761 condition are separated by NULL tree in the vector. */ | |
762 | |
763 static void | |
111 | 764 gen_shrink_wrap_conditions (gcall *bi_call, vec<gimple *> conds, |
0 | 765 unsigned int *nconds) |
766 { | |
111 | 767 gcall *call; |
0 | 768 tree fn; |
769 enum built_in_function fnc; | |
770 | |
111 | 771 gcc_assert (nconds && conds.exists ()); |
772 gcc_assert (conds.length () == 0); | |
0 | 773 gcc_assert (is_gimple_call (bi_call)); |
774 | |
775 call = bi_call; | |
776 fn = gimple_call_fndecl (call); | |
131 | 777 gcc_assert (fn && fndecl_built_in_p (fn)); |
0 | 778 fnc = DECL_FUNCTION_CODE (fn); |
779 *nconds = 0; | |
780 | |
781 if (fnc == BUILT_IN_POW) | |
782 gen_conditions_for_pow (call, conds, nconds); | |
783 else | |
784 { | |
785 tree arg; | |
786 inp_domain domain = get_no_error_domain (fnc); | |
787 *nconds = 0; | |
788 arg = gimple_call_arg (bi_call, 0); | |
789 gen_conditions_for_domain (arg, domain, conds, nconds); | |
790 } | |
791 | |
792 return; | |
793 } | |
794 | |
111 | 795 /* Shrink-wrap BI_CALL so that it is only called when one of the NCONDS |
145 | 796 conditions in CONDS is false. Also move BI_NEWCALL to a new basic block |
797 when it is non-null, it is called while all of the CONDS are true. */ | |
0 | 798 |
111 | 799 static void |
800 shrink_wrap_one_built_in_call_with_conds (gcall *bi_call, vec <gimple *> conds, | |
145 | 801 unsigned int nconds, |
802 gcall *bi_newcall = NULL) | |
0 | 803 { |
804 gimple_stmt_iterator bi_call_bsi; | |
145 | 805 basic_block bi_call_bb, bi_newcall_bb, join_tgt_bb, guard_bb; |
0 | 806 edge join_tgt_in_edge_from_call, join_tgt_in_edge_fall_thru; |
807 edge bi_call_in_edge0, guard_bb_in_edge; | |
111 | 808 unsigned tn_cond_stmts; |
0 | 809 unsigned ci; |
111 | 810 gimple *cond_expr = NULL; |
811 gimple *cond_expr_start; | |
812 | |
813 /* The cfg we want to create looks like this: | |
145 | 814 [guard n-1] <- guard_bb (old block) |
815 | \ | |
816 | [guard n-2] } | |
817 | / \ } | |
818 | / ... } new blocks | |
819 | / [guard 0] } | |
820 | / / | } | |
821 [call] | <- bi_call_bb } | |
822 \ [newcall] <-bi_newcall_bb} | |
823 \ | | |
824 [join] <- join_tgt_bb (old iff call must end bb) | |
111 | 825 possible EH edges (only if [join] is old) |
0 | 826 |
111 | 827 When [join] is new, the immediate dominators for these blocks are: |
0 | 828 |
111 | 829 1. [guard n-1]: unchanged |
830 2. [call]: [guard n-1] | |
145 | 831 3. [newcall]: [guard 0] |
832 4. [guard m]: [guard m+1] for 0 <= m <= n-2 | |
833 5. [join]: [guard n-1] | |
111 | 834 |
835 We punt for the more complex case case of [join] being old and | |
836 simply free the dominance info. We also punt on postdominators, | |
837 which aren't expected to be available at this point anyway. */ | |
0 | 838 bi_call_bb = gimple_bb (bi_call); |
839 | |
111 | 840 /* Now find the join target bb -- split bi_call_bb if needed. */ |
841 if (stmt_ends_bb_p (bi_call)) | |
842 { | |
843 /* We checked that there was a fallthrough edge in | |
844 can_guard_call_p. */ | |
845 join_tgt_in_edge_from_call = find_fallthru_edge (bi_call_bb->succs); | |
846 gcc_assert (join_tgt_in_edge_from_call); | |
847 /* We don't want to handle PHIs. */ | |
848 if (EDGE_COUNT (join_tgt_in_edge_from_call->dest->preds) > 1) | |
849 join_tgt_bb = split_edge (join_tgt_in_edge_from_call); | |
850 else | |
851 { | |
852 join_tgt_bb = join_tgt_in_edge_from_call->dest; | |
853 /* We may have degenerate PHIs in the destination. Propagate | |
854 those out. */ | |
855 for (gphi_iterator i = gsi_start_phis (join_tgt_bb); !gsi_end_p (i);) | |
856 { | |
857 gphi *phi = i.phi (); | |
858 replace_uses_by (gimple_phi_result (phi), | |
859 gimple_phi_arg_def (phi, 0)); | |
860 remove_phi_node (&i, true); | |
861 } | |
862 } | |
863 } | |
864 else | |
865 { | |
866 join_tgt_in_edge_from_call = split_block (bi_call_bb, bi_call); | |
867 join_tgt_bb = join_tgt_in_edge_from_call->dest; | |
868 } | |
0 | 869 |
870 bi_call_bsi = gsi_for_stmt (bi_call); | |
871 | |
872 /* Now it is time to insert the first conditional expression | |
873 into bi_call_bb and split this bb so that bi_call is | |
874 shrink-wrapped. */ | |
111 | 875 tn_cond_stmts = conds.length (); |
0 | 876 cond_expr = NULL; |
111 | 877 cond_expr_start = conds[0]; |
0 | 878 for (ci = 0; ci < tn_cond_stmts; ci++) |
879 { | |
111 | 880 gimple *c = conds[ci]; |
0 | 881 gcc_assert (c || ci != 0); |
882 if (!c) | |
883 break; | |
884 gsi_insert_before (&bi_call_bsi, c, GSI_SAME_STMT); | |
885 cond_expr = c; | |
886 } | |
887 ci++; | |
888 gcc_assert (cond_expr && gimple_code (cond_expr) == GIMPLE_COND); | |
889 | |
111 | 890 typedef std::pair<edge, edge> edge_pair; |
891 auto_vec<edge_pair, 8> edges; | |
0 | 892 |
893 bi_call_in_edge0 = split_block (bi_call_bb, cond_expr); | |
894 bi_call_in_edge0->flags &= ~EDGE_FALLTHRU; | |
111 | 895 bi_call_in_edge0->flags |= EDGE_FALSE_VALUE; |
896 guard_bb = bi_call_bb; | |
0 | 897 bi_call_bb = bi_call_in_edge0->dest; |
111 | 898 join_tgt_in_edge_fall_thru = make_edge (guard_bb, join_tgt_bb, |
899 EDGE_TRUE_VALUE); | |
0 | 900 |
111 | 901 edges.reserve (nconds); |
902 edges.quick_push (edge_pair (bi_call_in_edge0, join_tgt_in_edge_fall_thru)); | |
0 | 903 |
904 /* Code generation for the rest of the conditions */ | |
111 | 905 for (unsigned int i = 1; i < nconds; ++i) |
0 | 906 { |
907 unsigned ci0; | |
908 edge bi_call_in_edge; | |
909 gimple_stmt_iterator guard_bsi = gsi_for_stmt (cond_expr_start); | |
910 ci0 = ci; | |
111 | 911 cond_expr_start = conds[ci0]; |
0 | 912 for (; ci < tn_cond_stmts; ci++) |
913 { | |
111 | 914 gimple *c = conds[ci]; |
0 | 915 gcc_assert (c || ci != ci0); |
916 if (!c) | |
917 break; | |
918 gsi_insert_before (&guard_bsi, c, GSI_SAME_STMT); | |
919 cond_expr = c; | |
920 } | |
921 ci++; | |
922 gcc_assert (cond_expr && gimple_code (cond_expr) == GIMPLE_COND); | |
923 guard_bb_in_edge = split_block (guard_bb, cond_expr); | |
924 guard_bb_in_edge->flags &= ~EDGE_FALLTHRU; | |
111 | 925 guard_bb_in_edge->flags |= EDGE_TRUE_VALUE; |
0 | 926 |
111 | 927 bi_call_in_edge = make_edge (guard_bb, bi_call_bb, EDGE_FALSE_VALUE); |
928 edges.quick_push (edge_pair (bi_call_in_edge, guard_bb_in_edge)); | |
0 | 929 } |
930 | |
145 | 931 /* Move BI_NEWCALL to new basic block when it is non-null. */ |
932 if (bi_newcall) | |
933 { | |
934 /* Get bi_newcall_bb by split join_tgt_in_edge_fall_thru edge, | |
935 and move BI_NEWCALL to bi_newcall_bb. */ | |
936 bi_newcall_bb = split_edge (join_tgt_in_edge_fall_thru); | |
937 gimple_stmt_iterator to_gsi = gsi_start_bb (bi_newcall_bb); | |
938 gimple_stmt_iterator from_gsi = gsi_for_stmt (bi_newcall); | |
939 gsi_move_before (&from_gsi, &to_gsi); | |
940 join_tgt_in_edge_fall_thru = EDGE_SUCC (bi_newcall_bb, 0); | |
941 join_tgt_bb = join_tgt_in_edge_fall_thru->dest; | |
942 | |
943 tree bi_newcall_lhs = gimple_call_lhs (bi_newcall); | |
944 tree bi_call_lhs = gimple_call_lhs (bi_call); | |
945 if (!bi_call_lhs) | |
946 { | |
947 bi_call_lhs = copy_ssa_name (bi_newcall_lhs); | |
948 gimple_call_set_lhs (bi_call, bi_call_lhs); | |
949 SSA_NAME_DEF_STMT (bi_call_lhs) = bi_call; | |
950 } | |
951 | |
952 /* Create phi node for lhs of BI_CALL and BI_NEWCALL. */ | |
953 gphi *new_phi = create_phi_node (copy_ssa_name (bi_newcall_lhs), | |
954 join_tgt_bb); | |
955 SSA_NAME_OCCURS_IN_ABNORMAL_PHI (PHI_RESULT (new_phi)) | |
956 = SSA_NAME_OCCURS_IN_ABNORMAL_PHI (bi_newcall_lhs); | |
957 add_phi_arg (new_phi, bi_call_lhs, join_tgt_in_edge_from_call, | |
958 gimple_location (bi_call)); | |
959 add_phi_arg (new_phi, bi_newcall_lhs, join_tgt_in_edge_fall_thru, | |
960 gimple_location (bi_newcall)); | |
961 | |
962 /* Replace all use of original return value with result of phi node. */ | |
963 use_operand_p use_p; | |
964 gimple *use_stmt; | |
965 imm_use_iterator iterator; | |
966 FOR_EACH_IMM_USE_STMT (use_stmt, iterator, bi_newcall_lhs) | |
967 if (use_stmt != new_phi) | |
968 FOR_EACH_IMM_USE_ON_STMT (use_p, iterator) | |
969 SET_USE (use_p, PHI_RESULT (new_phi)); | |
970 } | |
971 | |
111 | 972 /* Now update the probability and profile information, processing the |
973 guards in order of execution. | |
974 | |
975 There are two approaches we could take here. On the one hand we | |
976 could assign a probability of X to the call block and distribute | |
977 that probability among its incoming edges. On the other hand we | |
978 could assign a probability of X to each individual call edge. | |
979 | |
980 The choice only affects calls that have more than one condition. | |
981 In those cases, the second approach would give the call block | |
982 a greater probability than the first. However, the difference | |
983 is only small, and our chosen X is a pure guess anyway. | |
984 | |
985 Here we take the second approach because it's slightly simpler | |
986 and because it's easy to see that it doesn't lose profile counts. */ | |
987 bi_call_bb->count = profile_count::zero (); | |
988 while (!edges.is_empty ()) | |
989 { | |
990 edge_pair e = edges.pop (); | |
991 edge call_edge = e.first; | |
992 edge nocall_edge = e.second; | |
993 basic_block src_bb = call_edge->src; | |
994 gcc_assert (src_bb == nocall_edge->src); | |
995 | |
996 call_edge->probability = profile_probability::very_unlikely (); | |
997 nocall_edge->probability = profile_probability::always () | |
998 - call_edge->probability; | |
999 | |
1000 bi_call_bb->count += call_edge->count (); | |
1001 | |
1002 if (nocall_edge->dest != join_tgt_bb) | |
131 | 1003 nocall_edge->dest->count = src_bb->count - bi_call_bb->count; |
111 | 1004 } |
1005 | |
1006 if (dom_info_available_p (CDI_DOMINATORS)) | |
1007 { | |
1008 /* The split_blocks leave [guard 0] as the immediate dominator | |
1009 of [call] and [call] as the immediate dominator of [join]. | |
1010 Fix them up. */ | |
1011 set_immediate_dominator (CDI_DOMINATORS, bi_call_bb, guard_bb); | |
1012 set_immediate_dominator (CDI_DOMINATORS, join_tgt_bb, guard_bb); | |
1013 } | |
1014 | |
0 | 1015 if (dump_file && (dump_flags & TDF_DETAILS)) |
1016 { | |
1017 location_t loc; | |
1018 loc = gimple_location (bi_call); | |
1019 fprintf (dump_file, | |
1020 "%s:%d: note: function call is shrink-wrapped" | |
1021 " into error conditions.\n", | |
1022 LOCATION_FILE (loc), LOCATION_LINE (loc)); | |
1023 } | |
111 | 1024 } |
1025 | |
1026 /* Shrink-wrap BI_CALL so that it is only called when it might set errno | |
1027 (but is always called if it would set errno). */ | |
1028 | |
1029 static void | |
1030 shrink_wrap_one_built_in_call (gcall *bi_call) | |
1031 { | |
1032 unsigned nconds = 0; | |
1033 auto_vec<gimple *, 12> conds; | |
1034 gen_shrink_wrap_conditions (bi_call, conds, &nconds); | |
1035 gcc_assert (nconds != 0); | |
1036 shrink_wrap_one_built_in_call_with_conds (bi_call, conds, nconds); | |
1037 } | |
1038 | |
1039 /* Return true if built-in function call CALL could be implemented using | |
1040 a combination of an internal function to compute the result and a | |
1041 separate call to set errno. */ | |
1042 | |
1043 static bool | |
1044 can_use_internal_fn (gcall *call) | |
1045 { | |
1046 /* Only replace calls that set errno. */ | |
1047 if (!gimple_vdef (call)) | |
1048 return false; | |
1049 | |
1050 /* See whether there is an internal function for this built-in. */ | |
1051 if (replacement_internal_fn (call) == IFN_LAST) | |
1052 return false; | |
1053 | |
1054 /* See whether we can catch all cases where errno would be set, | |
1055 while still avoiding the call in most cases. */ | |
1056 if (!can_test_argument_range (call) | |
1057 && !edom_only_function (call)) | |
1058 return false; | |
0 | 1059 |
1060 return true; | |
1061 } | |
1062 | |
111 | 1063 /* Implement built-in function call CALL using an internal function. */ |
1064 | |
1065 static void | |
1066 use_internal_fn (gcall *call) | |
1067 { | |
1068 /* We'll be inserting another call with the same arguments after the | |
1069 lhs has been set, so prevent any possible coalescing failure from | |
1070 having both values live at once. See PR 71020. */ | |
1071 replace_abnormal_ssa_names (call); | |
1072 | |
1073 unsigned nconds = 0; | |
1074 auto_vec<gimple *, 12> conds; | |
145 | 1075 bool is_arg_conds = false; |
111 | 1076 if (can_test_argument_range (call)) |
1077 { | |
1078 gen_shrink_wrap_conditions (call, conds, &nconds); | |
145 | 1079 is_arg_conds = true; |
111 | 1080 gcc_assert (nconds != 0); |
1081 } | |
1082 else | |
1083 gcc_assert (edom_only_function (call)); | |
1084 | |
1085 internal_fn ifn = replacement_internal_fn (call); | |
1086 gcc_assert (ifn != IFN_LAST); | |
1087 | |
1088 /* Construct the new call, with the same arguments as the original one. */ | |
1089 auto_vec <tree, 16> args; | |
1090 unsigned int nargs = gimple_call_num_args (call); | |
1091 for (unsigned int i = 0; i < nargs; ++i) | |
1092 args.safe_push (gimple_call_arg (call, i)); | |
1093 gcall *new_call = gimple_build_call_internal_vec (ifn, args); | |
1094 gimple_set_location (new_call, gimple_location (call)); | |
1095 gimple_call_set_nothrow (new_call, gimple_call_nothrow_p (call)); | |
1096 | |
1097 /* Transfer the LHS to the new call. */ | |
1098 tree lhs = gimple_call_lhs (call); | |
1099 gimple_call_set_lhs (new_call, lhs); | |
1100 gimple_call_set_lhs (call, NULL_TREE); | |
1101 SSA_NAME_DEF_STMT (lhs) = new_call; | |
1102 | |
1103 /* Insert the new call. */ | |
1104 gimple_stmt_iterator gsi = gsi_for_stmt (call); | |
1105 gsi_insert_before (&gsi, new_call, GSI_SAME_STMT); | |
1106 | |
1107 if (nconds == 0) | |
1108 { | |
1109 /* Skip the call if LHS == LHS. If we reach here, EDOM is the only | |
1110 valid errno value and it is used iff the result is NaN. */ | |
1111 conds.quick_push (gimple_build_cond (EQ_EXPR, lhs, lhs, | |
1112 NULL_TREE, NULL_TREE)); | |
1113 nconds++; | |
1114 | |
1115 /* Try replacing the original call with a direct assignment to | |
1116 errno, via an internal function. */ | |
1117 if (set_edom_supported_p () && !stmt_ends_bb_p (call)) | |
1118 { | |
1119 gimple_stmt_iterator gsi = gsi_for_stmt (call); | |
1120 gcall *new_call = gimple_build_call_internal (IFN_SET_EDOM, 0); | |
145 | 1121 gimple_move_vops (new_call, call); |
111 | 1122 gimple_set_location (new_call, gimple_location (call)); |
1123 gsi_replace (&gsi, new_call, false); | |
1124 call = new_call; | |
1125 } | |
1126 } | |
145 | 1127 shrink_wrap_one_built_in_call_with_conds (call, conds, nconds, |
1128 is_arg_conds ? new_call : NULL); | |
111 | 1129 } |
1130 | |
0 | 1131 /* The top level function for conditional dead code shrink |
1132 wrapping transformation. */ | |
1133 | |
111 | 1134 static void |
1135 shrink_wrap_conditional_dead_built_in_calls (vec<gcall *> calls) | |
0 | 1136 { |
1137 unsigned i = 0; | |
1138 | |
111 | 1139 unsigned n = calls.length (); |
0 | 1140 for (; i < n ; i++) |
1141 { | |
111 | 1142 gcall *bi_call = calls[i]; |
1143 if (gimple_call_lhs (bi_call)) | |
1144 use_internal_fn (bi_call); | |
1145 else | |
1146 shrink_wrap_one_built_in_call (bi_call); | |
1147 } | |
1148 } | |
1149 | |
1150 namespace { | |
1151 | |
1152 const pass_data pass_data_call_cdce = | |
1153 { | |
1154 GIMPLE_PASS, /* type */ | |
1155 "cdce", /* name */ | |
1156 OPTGROUP_NONE, /* optinfo_flags */ | |
1157 TV_TREE_CALL_CDCE, /* tv_id */ | |
1158 ( PROP_cfg | PROP_ssa ), /* properties_required */ | |
1159 0, /* properties_provided */ | |
1160 0, /* properties_destroyed */ | |
1161 0, /* todo_flags_start */ | |
1162 0, /* todo_flags_finish */ | |
1163 }; | |
1164 | |
1165 class pass_call_cdce : public gimple_opt_pass | |
1166 { | |
1167 public: | |
1168 pass_call_cdce (gcc::context *ctxt) | |
1169 : gimple_opt_pass (pass_data_call_cdce, ctxt) | |
1170 {} | |
1171 | |
1172 /* opt_pass methods: */ | |
1173 virtual bool gate (function *) | |
1174 { | |
1175 /* The limit constants used in the implementation | |
1176 assume IEEE floating point format. Other formats | |
1177 can be supported in the future if needed. */ | |
1178 return flag_tree_builtin_call_dce != 0; | |
0 | 1179 } |
1180 | |
111 | 1181 virtual unsigned int execute (function *); |
0 | 1182 |
111 | 1183 }; // class pass_call_cdce |
0 | 1184 |
111 | 1185 unsigned int |
1186 pass_call_cdce::execute (function *fun) | |
0 | 1187 { |
1188 basic_block bb; | |
1189 gimple_stmt_iterator i; | |
111 | 1190 auto_vec<gcall *> cond_dead_built_in_calls; |
1191 FOR_EACH_BB_FN (bb, fun) | |
0 | 1192 { |
111 | 1193 /* Skip blocks that are being optimized for size, since our |
1194 transformation always increases code size. */ | |
1195 if (optimize_bb_for_size_p (bb)) | |
1196 continue; | |
1197 | |
0 | 1198 /* Collect dead call candidates. */ |
1199 for (i = gsi_start_bb (bb); !gsi_end_p (i); gsi_next (&i)) | |
1200 { | |
111 | 1201 gcall *stmt = dyn_cast <gcall *> (gsi_stmt (i)); |
1202 if (stmt | |
1203 && gimple_call_builtin_p (stmt, BUILT_IN_NORMAL) | |
1204 && (gimple_call_lhs (stmt) | |
1205 ? can_use_internal_fn (stmt) | |
1206 : can_test_argument_range (stmt)) | |
1207 && can_guard_call_p (stmt)) | |
0 | 1208 { |
1209 if (dump_file && (dump_flags & TDF_DETAILS)) | |
1210 { | |
1211 fprintf (dump_file, "Found conditional dead call: "); | |
1212 print_gimple_stmt (dump_file, stmt, 0, TDF_SLIM); | |
1213 fprintf (dump_file, "\n"); | |
1214 } | |
111 | 1215 if (!cond_dead_built_in_calls.exists ()) |
1216 cond_dead_built_in_calls.create (64); | |
1217 cond_dead_built_in_calls.safe_push (stmt); | |
0 | 1218 } |
1219 } | |
1220 } | |
1221 | |
111 | 1222 if (!cond_dead_built_in_calls.exists ()) |
0 | 1223 return 0; |
1224 | |
111 | 1225 shrink_wrap_conditional_dead_built_in_calls (cond_dead_built_in_calls); |
1226 free_dominance_info (CDI_POST_DOMINATORS); | |
1227 /* As we introduced new control-flow we need to insert PHI-nodes | |
1228 for the call-clobbers of the remaining call. */ | |
1229 mark_virtual_operands_for_renaming (fun); | |
1230 return TODO_update_ssa; | |
0 | 1231 } |
1232 | |
111 | 1233 } // anon namespace |
0 | 1234 |
111 | 1235 gimple_opt_pass * |
1236 make_pass_call_cdce (gcc::context *ctxt) | |
0 | 1237 { |
111 | 1238 return new pass_call_cdce (ctxt); |
1239 } |