2015-06-30 Sandra Loosemore <sandra@codesourcery.com>
[official-gcc.git] / gcc / tree-call-cdce.c
blob03a8b1e227731055fa52e9058f5c315725326c11
1 /* Conditional Dead Call Elimination pass for the GNU compiler.
2 Copyright (C) 2008-2015 Free Software Foundation, Inc.
3 Contributed by Xinliang David Li <davidxl@google.com>
5 This file is part of GCC.
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.
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.
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/>. */
21 #include "config.h"
22 #include "system.h"
23 #include "coretypes.h"
24 #include "tm.h"
25 #include "predict.h"
26 #include "hard-reg-set.h"
27 #include "function.h"
28 #include "dominance.h"
29 #include "cfg.h"
30 #include "basic-block.h"
31 #include "symtab.h"
32 #include "alias.h"
33 #include "tree.h"
34 #include "fold-const.h"
35 #include "stor-layout.h"
36 #include "gimple-pretty-print.h"
37 #include "tree-ssa-alias.h"
38 #include "internal-fn.h"
39 #include "gimple-expr.h"
40 #include "gimple.h"
41 #include "gimple-iterator.h"
42 #include "gimple-ssa.h"
43 #include "tree-cfg.h"
44 #include "stringpool.h"
45 #include "tree-ssanames.h"
46 #include "tree-into-ssa.h"
47 #include "tree-pass.h"
48 #include "flags.h"
51 /* Conditional dead call elimination
53 Some builtin functions can set errno on error conditions, but they
54 are otherwise pure. If the result of a call to such a function is
55 not used, the compiler can still not eliminate the call without
56 powerful interprocedural analysis to prove that the errno is not
57 checked. However, if the conditions under which the error occurs
58 are known, the compiler can conditionally dead code eliminate the
59 calls by shrink-wrapping the semi-dead calls into the error condition:
61 built_in_call (args)
62 ==>
63 if (error_cond (args))
64 built_in_call (args)
66 An actual simple example is :
67 log (x); // Mostly dead call
68 ==>
69 if (x <= 0)
70 log (x);
71 With this change, call to log (x) is effectively eliminated, as
72 in majority of the cases, log won't be called with x out of
73 range. The branch is totally predictable, so the branch cost
74 is low.
76 Note that library functions are not supposed to clear errno to zero without
77 error. See IEEE Std 1003.1, section 2.3 Error Numbers, and section 7.5:3 of
78 ISO/IEC 9899 (C99).
80 The condition wrapping the builtin call is conservatively set to avoid too
81 aggressive (wrong) shrink wrapping. The optimization is called conditional
82 dead call elimination because the call is eliminated under the condition
83 that the input arguments would not lead to domain or range error (for
84 instance when x <= 0 for a log (x) call), however the chances that the error
85 condition is hit is very low (those builtin calls which are conditionally
86 dead are usually part of the C++ abstraction penalty exposed after
87 inlining). */
90 /* A structure for representing input domain of
91 a function argument in integer. If the lower
92 bound is -inf, has_lb is set to false. If the
93 upper bound is +inf, has_ub is false.
94 is_lb_inclusive and is_ub_inclusive are flags
95 to indicate if lb and ub value are inclusive
96 respectively. */
98 typedef struct input_domain
100 int lb;
101 int ub;
102 bool has_lb;
103 bool has_ub;
104 bool is_lb_inclusive;
105 bool is_ub_inclusive;
106 } inp_domain;
108 /* A helper function to construct and return an input
109 domain object. LB is the lower bound, HAS_LB is
110 a boolean flag indicating if the lower bound exists,
111 and LB_INCLUSIVE is a boolean flag indicating if the
112 lower bound is inclusive or not. UB, HAS_UB, and
113 UB_INCLUSIVE have the same meaning, but for upper
114 bound of the domain. */
116 static inp_domain
117 get_domain (int lb, bool has_lb, bool lb_inclusive,
118 int ub, bool has_ub, bool ub_inclusive)
120 inp_domain domain;
121 domain.lb = lb;
122 domain.has_lb = has_lb;
123 domain.is_lb_inclusive = lb_inclusive;
124 domain.ub = ub;
125 domain.has_ub = has_ub;
126 domain.is_ub_inclusive = ub_inclusive;
127 return domain;
130 /* A helper function to check the target format for the
131 argument type. In this implementation, only IEEE formats
132 are supported. ARG is the call argument to be checked.
133 Returns true if the format is supported. To support other
134 target formats, function get_no_error_domain needs to be
135 enhanced to have range bounds properly computed. Since
136 the check is cheap (very small number of candidates
137 to be checked), the result is not cached for each float type. */
139 static bool
140 check_target_format (tree arg)
142 tree type;
143 machine_mode mode;
144 const struct real_format *rfmt;
146 type = TREE_TYPE (arg);
147 mode = TYPE_MODE (type);
148 rfmt = REAL_MODE_FORMAT (mode);
149 if ((mode == SFmode
150 && (rfmt == &ieee_single_format || rfmt == &mips_single_format
151 || rfmt == &motorola_single_format))
152 || (mode == DFmode
153 && (rfmt == &ieee_double_format || rfmt == &mips_double_format
154 || rfmt == &motorola_double_format))
155 /* For long double, we can not really check XFmode
156 which is only defined on intel platforms.
157 Candidate pre-selection using builtin function
158 code guarantees that we are checking formats
159 for long double modes: double, quad, and extended. */
160 || (mode != SFmode && mode != DFmode
161 && (rfmt == &ieee_quad_format
162 || rfmt == &mips_quad_format
163 || rfmt == &ieee_extended_motorola_format
164 || rfmt == &ieee_extended_intel_96_format
165 || rfmt == &ieee_extended_intel_128_format
166 || rfmt == &ieee_extended_intel_96_round_53_format)))
167 return true;
169 return false;
173 /* A helper function to help select calls to pow that are suitable for
174 conditional DCE transformation. It looks for pow calls that can be
175 guided with simple conditions. Such calls either have constant base
176 values or base values converted from integers. Returns true if
177 the pow call POW_CALL is a candidate. */
179 /* The maximum integer bit size for base argument of a pow call
180 that is suitable for shrink-wrapping transformation. */
181 #define MAX_BASE_INT_BIT_SIZE 32
183 static bool
184 check_pow (gcall *pow_call)
186 tree base, expn;
187 enum tree_code bc, ec;
189 if (gimple_call_num_args (pow_call) != 2)
190 return false;
192 base = gimple_call_arg (pow_call, 0);
193 expn = gimple_call_arg (pow_call, 1);
195 if (!check_target_format (expn))
196 return false;
198 bc = TREE_CODE (base);
199 ec = TREE_CODE (expn);
201 /* Folding candidates are not interesting.
202 Can actually assert that it is already folded. */
203 if (ec == REAL_CST && bc == REAL_CST)
204 return false;
206 if (bc == REAL_CST)
208 /* Only handle a fixed range of constant. */
209 REAL_VALUE_TYPE mv;
210 REAL_VALUE_TYPE bcv = TREE_REAL_CST (base);
211 if (REAL_VALUES_EQUAL (bcv, dconst1))
212 return false;
213 if (REAL_VALUES_LESS (bcv, dconst1))
214 return false;
215 real_from_integer (&mv, TYPE_MODE (TREE_TYPE (base)), 256, UNSIGNED);
216 if (REAL_VALUES_LESS (mv, bcv))
217 return false;
218 return true;
220 else if (bc == SSA_NAME)
222 tree base_val0, type;
223 gimple base_def;
224 int bit_sz;
226 /* Only handles cases where base value is converted
227 from integer values. */
228 base_def = SSA_NAME_DEF_STMT (base);
229 if (gimple_code (base_def) != GIMPLE_ASSIGN)
230 return false;
232 if (gimple_assign_rhs_code (base_def) != FLOAT_EXPR)
233 return false;
234 base_val0 = gimple_assign_rhs1 (base_def);
236 type = TREE_TYPE (base_val0);
237 if (TREE_CODE (type) != INTEGER_TYPE)
238 return false;
239 bit_sz = TYPE_PRECISION (type);
240 /* If the type of the base is too wide,
241 the resulting shrink wrapping condition
242 will be too conservative. */
243 if (bit_sz > MAX_BASE_INT_BIT_SIZE)
244 return false;
246 return true;
248 else
249 return false;
252 /* A helper function to help select candidate function calls that are
253 suitable for conditional DCE. Candidate functions must have single
254 valid input domain in this implementation except for pow (see check_pow).
255 Returns true if the function call is a candidate. */
257 static bool
258 check_builtin_call (gcall *bcall)
260 tree arg;
262 arg = gimple_call_arg (bcall, 0);
263 return check_target_format (arg);
266 /* A helper function to determine if a builtin function call is a
267 candidate for conditional DCE. Returns true if the builtin call
268 is a candidate. */
270 static bool
271 is_call_dce_candidate (gcall *call)
273 tree fn;
274 enum built_in_function fnc;
276 /* Only potentially dead calls are considered. */
277 if (gimple_call_lhs (call))
278 return false;
280 fn = gimple_call_fndecl (call);
281 if (!fn
282 || !DECL_BUILT_IN (fn)
283 || (DECL_BUILT_IN_CLASS (fn) != BUILT_IN_NORMAL))
284 return false;
286 fnc = DECL_FUNCTION_CODE (fn);
287 switch (fnc)
289 /* Trig functions. */
290 CASE_FLT_FN (BUILT_IN_ACOS):
291 CASE_FLT_FN (BUILT_IN_ASIN):
292 /* Hyperbolic functions. */
293 CASE_FLT_FN (BUILT_IN_ACOSH):
294 CASE_FLT_FN (BUILT_IN_ATANH):
295 CASE_FLT_FN (BUILT_IN_COSH):
296 CASE_FLT_FN (BUILT_IN_SINH):
297 /* Log functions. */
298 CASE_FLT_FN (BUILT_IN_LOG):
299 CASE_FLT_FN (BUILT_IN_LOG2):
300 CASE_FLT_FN (BUILT_IN_LOG10):
301 CASE_FLT_FN (BUILT_IN_LOG1P):
302 /* Exp functions. */
303 CASE_FLT_FN (BUILT_IN_EXP):
304 CASE_FLT_FN (BUILT_IN_EXP2):
305 CASE_FLT_FN (BUILT_IN_EXP10):
306 CASE_FLT_FN (BUILT_IN_EXPM1):
307 CASE_FLT_FN (BUILT_IN_POW10):
308 /* Sqrt. */
309 CASE_FLT_FN (BUILT_IN_SQRT):
310 return check_builtin_call (call);
311 /* Special one: two argument pow. */
312 case BUILT_IN_POW:
313 return check_pow (call);
314 default:
315 break;
318 return false;
322 /* A helper function to generate gimple statements for
323 one bound comparison. ARG is the call argument to
324 be compared with the bound, LBUB is the bound value
325 in integer, TCODE is the tree_code of the comparison,
326 TEMP_NAME1/TEMP_NAME2 are names of the temporaries,
327 CONDS is a vector holding the produced GIMPLE statements,
328 and NCONDS points to the variable holding the number
329 of logical comparisons. CONDS is either empty or
330 a list ended with a null tree. */
332 static void
333 gen_one_condition (tree arg, int lbub,
334 enum tree_code tcode,
335 const char *temp_name1,
336 const char *temp_name2,
337 vec<gimple> conds,
338 unsigned *nconds)
340 tree lbub_real_cst, lbub_cst, float_type;
341 tree temp, tempn, tempc, tempcn;
342 gassign *stmt1;
343 gassign *stmt2;
344 gcond *stmt3;
346 float_type = TREE_TYPE (arg);
347 lbub_cst = build_int_cst (integer_type_node, lbub);
348 lbub_real_cst = build_real_from_int_cst (float_type, lbub_cst);
350 temp = create_tmp_var (float_type, temp_name1);
351 stmt1 = gimple_build_assign (temp, arg);
352 tempn = make_ssa_name (temp, stmt1);
353 gimple_assign_set_lhs (stmt1, tempn);
355 tempc = create_tmp_var (boolean_type_node, temp_name2);
356 stmt2 = gimple_build_assign (tempc,
357 fold_build2 (tcode,
358 boolean_type_node,
359 tempn, lbub_real_cst));
360 tempcn = make_ssa_name (tempc, stmt2);
361 gimple_assign_set_lhs (stmt2, tempcn);
363 stmt3 = gimple_build_cond_from_tree (tempcn, NULL_TREE, NULL_TREE);
364 conds.quick_push (stmt1);
365 conds.quick_push (stmt2);
366 conds.quick_push (stmt3);
367 (*nconds)++;
370 /* A helper function to generate GIMPLE statements for
371 out of input domain check. ARG is the call argument
372 to be runtime checked, DOMAIN holds the valid domain
373 for the given function, CONDS points to the vector
374 holding the result GIMPLE statements. *NCONDS is
375 the number of logical comparisons. This function
376 produces no more than two logical comparisons, one
377 for lower bound check, one for upper bound check. */
379 static void
380 gen_conditions_for_domain (tree arg, inp_domain domain,
381 vec<gimple> conds,
382 unsigned *nconds)
384 if (domain.has_lb)
385 gen_one_condition (arg, domain.lb,
386 (domain.is_lb_inclusive
387 ? LT_EXPR : LE_EXPR),
388 "DCE_COND_LB", "DCE_COND_LB_TEST",
389 conds, nconds);
391 if (domain.has_ub)
393 /* Now push a separator. */
394 if (domain.has_lb)
395 conds.quick_push (NULL);
397 gen_one_condition (arg, domain.ub,
398 (domain.is_ub_inclusive
399 ? GT_EXPR : GE_EXPR),
400 "DCE_COND_UB", "DCE_COND_UB_TEST",
401 conds, nconds);
406 /* A helper function to generate condition
407 code for the y argument in call pow (some_const, y).
408 See candidate selection in check_pow. Since the
409 candidates' base values have a limited range,
410 the guarded code generated for y are simple:
411 if (y > max_y)
412 pow (const, y);
413 Note max_y can be computed separately for each
414 const base, but in this implementation, we
415 choose to compute it using the max base
416 in the allowed range for the purpose of
417 simplicity. BASE is the constant base value,
418 EXPN is the expression for the exponent argument,
419 *CONDS is the vector to hold resulting statements,
420 and *NCONDS is the number of logical conditions. */
422 static void
423 gen_conditions_for_pow_cst_base (tree base, tree expn,
424 vec<gimple> conds,
425 unsigned *nconds)
427 inp_domain exp_domain;
428 /* Validate the range of the base constant to make
429 sure it is consistent with check_pow. */
430 REAL_VALUE_TYPE mv;
431 REAL_VALUE_TYPE bcv = TREE_REAL_CST (base);
432 gcc_assert (!REAL_VALUES_EQUAL (bcv, dconst1)
433 && !REAL_VALUES_LESS (bcv, dconst1));
434 real_from_integer (&mv, TYPE_MODE (TREE_TYPE (base)), 256, UNSIGNED);
435 gcc_assert (!REAL_VALUES_LESS (mv, bcv));
437 exp_domain = get_domain (0, false, false,
438 127, true, false);
440 gen_conditions_for_domain (expn, exp_domain,
441 conds, nconds);
444 /* Generate error condition code for pow calls with
445 non constant base values. The candidates selected
446 have their base argument value converted from
447 integer (see check_pow) value (1, 2, 4 bytes), and
448 the max exp value is computed based on the size
449 of the integer type (i.e. max possible base value).
450 The resulting input domain for exp argument is thus
451 conservative (smaller than the max value allowed by
452 the runtime value of the base). BASE is the integer
453 base value, EXPN is the expression for the exponent
454 argument, *CONDS is the vector to hold resulting
455 statements, and *NCONDS is the number of logical
456 conditions. */
458 static void
459 gen_conditions_for_pow_int_base (tree base, tree expn,
460 vec<gimple> conds,
461 unsigned *nconds)
463 gimple base_def;
464 tree base_val0;
465 tree int_type;
466 tree temp, tempn;
467 tree cst0;
468 gimple stmt1, stmt2;
469 int bit_sz, max_exp;
470 inp_domain exp_domain;
472 base_def = SSA_NAME_DEF_STMT (base);
473 base_val0 = gimple_assign_rhs1 (base_def);
474 int_type = TREE_TYPE (base_val0);
475 bit_sz = TYPE_PRECISION (int_type);
476 gcc_assert (bit_sz > 0
477 && bit_sz <= MAX_BASE_INT_BIT_SIZE);
479 /* Determine the max exp argument value according to
480 the size of the base integer. The max exp value
481 is conservatively estimated assuming IEEE754 double
482 precision format. */
483 if (bit_sz == 8)
484 max_exp = 128;
485 else if (bit_sz == 16)
486 max_exp = 64;
487 else
489 gcc_assert (bit_sz == MAX_BASE_INT_BIT_SIZE);
490 max_exp = 32;
493 /* For pow ((double)x, y), generate the following conditions:
494 cond 1:
495 temp1 = x;
496 if (temp1 <= 0)
498 cond 2:
499 temp2 = y;
500 if (temp2 > max_exp_real_cst) */
502 /* Generate condition in reverse order -- first
503 the condition for the exp argument. */
505 exp_domain = get_domain (0, false, false,
506 max_exp, true, true);
508 gen_conditions_for_domain (expn, exp_domain,
509 conds, nconds);
511 /* Now generate condition for the base argument.
512 Note it does not use the helper function
513 gen_conditions_for_domain because the base
514 type is integer. */
516 /* Push a separator. */
517 conds.quick_push (NULL);
519 temp = create_tmp_var (int_type, "DCE_COND1");
520 cst0 = build_int_cst (int_type, 0);
521 stmt1 = gimple_build_assign (temp, base_val0);
522 tempn = make_ssa_name (temp, stmt1);
523 gimple_assign_set_lhs (stmt1, tempn);
524 stmt2 = gimple_build_cond (LE_EXPR, tempn, cst0, NULL_TREE, NULL_TREE);
526 conds.quick_push (stmt1);
527 conds.quick_push (stmt2);
528 (*nconds)++;
531 /* Method to generate conditional statements for guarding conditionally
532 dead calls to pow. One or more statements can be generated for
533 each logical condition. Statement groups of different conditions
534 are separated by a NULL tree and they are stored in the vec
535 conds. The number of logical conditions are stored in *nconds.
537 See C99 standard, 7.12.7.4:2, for description of pow (x, y).
538 The precise condition for domain errors are complex. In this
539 implementation, a simplified (but conservative) valid domain
540 for x and y are used: x is positive to avoid dom errors, while
541 y is smaller than a upper bound (depending on x) to avoid range
542 errors. Runtime code is generated to check x (if not constant)
543 and y against the valid domain. If it is out, jump to the call,
544 otherwise the call is bypassed. POW_CALL is the call statement,
545 *CONDS is a vector holding the resulting condition statements,
546 and *NCONDS is the number of logical conditions. */
548 static void
549 gen_conditions_for_pow (gcall *pow_call, vec<gimple> conds,
550 unsigned *nconds)
552 tree base, expn;
553 enum tree_code bc;
555 gcc_checking_assert (check_pow (pow_call));
557 *nconds = 0;
559 base = gimple_call_arg (pow_call, 0);
560 expn = gimple_call_arg (pow_call, 1);
562 bc = TREE_CODE (base);
564 if (bc == REAL_CST)
565 gen_conditions_for_pow_cst_base (base, expn, conds, nconds);
566 else if (bc == SSA_NAME)
567 gen_conditions_for_pow_int_base (base, expn, conds, nconds);
568 else
569 gcc_unreachable ();
572 /* A helper routine to help computing the valid input domain
573 for a builtin function. See C99 7.12.7 for details. In this
574 implementation, we only handle single region domain. The
575 resulting region can be conservative (smaller) than the actual
576 one and rounded to integers. Some of the bounds are documented
577 in the standard, while other limit constants are computed
578 assuming IEEE floating point format (for SF and DF modes).
579 Since IEEE only sets minimum requirements for long double format,
580 different long double formats exist under different implementations
581 (e.g, 64 bit double precision (DF), 80 bit double-extended
582 precision (XF), and 128 bit quad precision (QF) ). For simplicity,
583 in this implementation, the computed bounds for long double assume
584 64 bit format (DF), and are therefore conservative. Another
585 assumption is that single precision float type is always SF mode,
586 and double type is DF mode. This function is quite
587 implementation specific, so it may not be suitable to be part of
588 builtins.c. This needs to be revisited later to see if it can
589 be leveraged in x87 assembly expansion. */
591 static inp_domain
592 get_no_error_domain (enum built_in_function fnc)
594 switch (fnc)
596 /* Trig functions: return [-1, +1] */
597 CASE_FLT_FN (BUILT_IN_ACOS):
598 CASE_FLT_FN (BUILT_IN_ASIN):
599 return get_domain (-1, true, true,
600 1, true, true);
601 /* Hyperbolic functions. */
602 CASE_FLT_FN (BUILT_IN_ACOSH):
603 /* acosh: [1, +inf) */
604 return get_domain (1, true, true,
605 1, false, false);
606 CASE_FLT_FN (BUILT_IN_ATANH):
607 /* atanh: (-1, +1) */
608 return get_domain (-1, true, false,
609 1, true, false);
610 case BUILT_IN_COSHF:
611 case BUILT_IN_SINHF:
612 /* coshf: (-89, +89) */
613 return get_domain (-89, true, false,
614 89, true, false);
615 case BUILT_IN_COSH:
616 case BUILT_IN_SINH:
617 case BUILT_IN_COSHL:
618 case BUILT_IN_SINHL:
619 /* cosh: (-710, +710) */
620 return get_domain (-710, true, false,
621 710, true, false);
622 /* Log functions: (0, +inf) */
623 CASE_FLT_FN (BUILT_IN_LOG):
624 CASE_FLT_FN (BUILT_IN_LOG2):
625 CASE_FLT_FN (BUILT_IN_LOG10):
626 return get_domain (0, true, false,
627 0, false, false);
628 CASE_FLT_FN (BUILT_IN_LOG1P):
629 return get_domain (-1, true, false,
630 0, false, false);
631 /* Exp functions. */
632 case BUILT_IN_EXPF:
633 case BUILT_IN_EXPM1F:
634 /* expf: (-inf, 88) */
635 return get_domain (-1, false, false,
636 88, true, false);
637 case BUILT_IN_EXP:
638 case BUILT_IN_EXPM1:
639 case BUILT_IN_EXPL:
640 case BUILT_IN_EXPM1L:
641 /* exp: (-inf, 709) */
642 return get_domain (-1, false, false,
643 709, true, false);
644 case BUILT_IN_EXP2F:
645 /* exp2f: (-inf, 128) */
646 return get_domain (-1, false, false,
647 128, true, false);
648 case BUILT_IN_EXP2:
649 case BUILT_IN_EXP2L:
650 /* exp2: (-inf, 1024) */
651 return get_domain (-1, false, false,
652 1024, true, false);
653 case BUILT_IN_EXP10F:
654 case BUILT_IN_POW10F:
655 /* exp10f: (-inf, 38) */
656 return get_domain (-1, false, false,
657 38, true, false);
658 case BUILT_IN_EXP10:
659 case BUILT_IN_POW10:
660 case BUILT_IN_EXP10L:
661 case BUILT_IN_POW10L:
662 /* exp10: (-inf, 308) */
663 return get_domain (-1, false, false,
664 308, true, false);
665 /* sqrt: [0, +inf) */
666 CASE_FLT_FN (BUILT_IN_SQRT):
667 return get_domain (0, true, true,
668 0, false, false);
669 default:
670 gcc_unreachable ();
673 gcc_unreachable ();
676 /* The function to generate shrink wrap conditions for a partially
677 dead builtin call whose return value is not used anywhere,
678 but has to be kept live due to potential error condition.
679 BI_CALL is the builtin call, CONDS is the vector of statements
680 for condition code, NCODES is the pointer to the number of
681 logical conditions. Statements belonging to different logical
682 condition are separated by NULL tree in the vector. */
684 static void
685 gen_shrink_wrap_conditions (gcall *bi_call, vec<gimple> conds,
686 unsigned int *nconds)
688 gcall *call;
689 tree fn;
690 enum built_in_function fnc;
692 gcc_assert (nconds && conds.exists ());
693 gcc_assert (conds.length () == 0);
694 gcc_assert (is_gimple_call (bi_call));
696 call = bi_call;
697 fn = gimple_call_fndecl (call);
698 gcc_assert (fn && DECL_BUILT_IN (fn));
699 fnc = DECL_FUNCTION_CODE (fn);
700 *nconds = 0;
702 if (fnc == BUILT_IN_POW)
703 gen_conditions_for_pow (call, conds, nconds);
704 else
706 tree arg;
707 inp_domain domain = get_no_error_domain (fnc);
708 *nconds = 0;
709 arg = gimple_call_arg (bi_call, 0);
710 gen_conditions_for_domain (arg, domain, conds, nconds);
713 return;
717 /* Probability of the branch (to the call) is taken. */
718 #define ERR_PROB 0.01
720 /* The function to shrink wrap a partially dead builtin call
721 whose return value is not used anywhere, but has to be kept
722 live due to potential error condition. Returns true if the
723 transformation actually happens. */
725 static bool
726 shrink_wrap_one_built_in_call (gcall *bi_call)
728 gimple_stmt_iterator bi_call_bsi;
729 basic_block bi_call_bb, join_tgt_bb, guard_bb, guard_bb0;
730 edge join_tgt_in_edge_from_call, join_tgt_in_edge_fall_thru;
731 edge bi_call_in_edge0, guard_bb_in_edge;
732 unsigned tn_cond_stmts, nconds;
733 unsigned ci;
734 gimple cond_expr = NULL;
735 gimple cond_expr_start;
736 tree bi_call_label_decl;
737 gimple bi_call_label;
739 auto_vec<gimple, 12> conds;
740 gen_shrink_wrap_conditions (bi_call, conds, &nconds);
742 /* This can happen if the condition generator decides
743 it is not beneficial to do the transformation. Just
744 return false and do not do any transformation for
745 the call. */
746 if (nconds == 0)
747 return false;
749 bi_call_bb = gimple_bb (bi_call);
751 /* Now find the join target bb -- split bi_call_bb if needed. */
752 if (stmt_ends_bb_p (bi_call))
754 /* If the call must be the last in the bb, don't split the block,
755 it could e.g. have EH edges. */
756 join_tgt_in_edge_from_call = find_fallthru_edge (bi_call_bb->succs);
757 if (join_tgt_in_edge_from_call == NULL)
758 return false;
760 else
761 join_tgt_in_edge_from_call = split_block (bi_call_bb, bi_call);
763 bi_call_bsi = gsi_for_stmt (bi_call);
765 join_tgt_bb = join_tgt_in_edge_from_call->dest;
767 /* Now it is time to insert the first conditional expression
768 into bi_call_bb and split this bb so that bi_call is
769 shrink-wrapped. */
770 tn_cond_stmts = conds.length ();
771 cond_expr = NULL;
772 cond_expr_start = conds[0];
773 for (ci = 0; ci < tn_cond_stmts; ci++)
775 gimple c = conds[ci];
776 gcc_assert (c || ci != 0);
777 if (!c)
778 break;
779 gsi_insert_before (&bi_call_bsi, c, GSI_SAME_STMT);
780 cond_expr = c;
782 nconds--;
783 ci++;
784 gcc_assert (cond_expr && gimple_code (cond_expr) == GIMPLE_COND);
786 /* Now the label. */
787 bi_call_label_decl = create_artificial_label (gimple_location (bi_call));
788 bi_call_label = gimple_build_label (bi_call_label_decl);
789 gsi_insert_before (&bi_call_bsi, bi_call_label, GSI_SAME_STMT);
791 bi_call_in_edge0 = split_block (bi_call_bb, cond_expr);
792 bi_call_in_edge0->flags &= ~EDGE_FALLTHRU;
793 bi_call_in_edge0->flags |= EDGE_TRUE_VALUE;
794 guard_bb0 = bi_call_bb;
795 bi_call_bb = bi_call_in_edge0->dest;
796 join_tgt_in_edge_fall_thru = make_edge (guard_bb0, join_tgt_bb,
797 EDGE_FALSE_VALUE);
799 bi_call_in_edge0->probability = REG_BR_PROB_BASE * ERR_PROB;
800 bi_call_in_edge0->count =
801 apply_probability (guard_bb0->count,
802 bi_call_in_edge0->probability);
803 join_tgt_in_edge_fall_thru->probability =
804 inverse_probability (bi_call_in_edge0->probability);
805 join_tgt_in_edge_fall_thru->count =
806 guard_bb0->count - bi_call_in_edge0->count;
808 /* Code generation for the rest of the conditions */
809 guard_bb = guard_bb0;
810 while (nconds > 0)
812 unsigned ci0;
813 edge bi_call_in_edge;
814 gimple_stmt_iterator guard_bsi = gsi_for_stmt (cond_expr_start);
815 ci0 = ci;
816 cond_expr_start = conds[ci0];
817 for (; ci < tn_cond_stmts; ci++)
819 gimple c = conds[ci];
820 gcc_assert (c || ci != ci0);
821 if (!c)
822 break;
823 gsi_insert_before (&guard_bsi, c, GSI_SAME_STMT);
824 cond_expr = c;
826 nconds--;
827 ci++;
828 gcc_assert (cond_expr && gimple_code (cond_expr) == GIMPLE_COND);
829 guard_bb_in_edge = split_block (guard_bb, cond_expr);
830 guard_bb_in_edge->flags &= ~EDGE_FALLTHRU;
831 guard_bb_in_edge->flags |= EDGE_FALSE_VALUE;
833 bi_call_in_edge = make_edge (guard_bb, bi_call_bb, EDGE_TRUE_VALUE);
835 bi_call_in_edge->probability = REG_BR_PROB_BASE * ERR_PROB;
836 bi_call_in_edge->count =
837 apply_probability (guard_bb->count,
838 bi_call_in_edge->probability);
839 guard_bb_in_edge->probability =
840 inverse_probability (bi_call_in_edge->probability);
841 guard_bb_in_edge->count = guard_bb->count - bi_call_in_edge->count;
844 if (dump_file && (dump_flags & TDF_DETAILS))
846 location_t loc;
847 loc = gimple_location (bi_call);
848 fprintf (dump_file,
849 "%s:%d: note: function call is shrink-wrapped"
850 " into error conditions.\n",
851 LOCATION_FILE (loc), LOCATION_LINE (loc));
854 return true;
857 /* The top level function for conditional dead code shrink
858 wrapping transformation. */
860 static bool
861 shrink_wrap_conditional_dead_built_in_calls (vec<gcall *> calls)
863 bool changed = false;
864 unsigned i = 0;
866 unsigned n = calls.length ();
867 if (n == 0)
868 return false;
870 for (; i < n ; i++)
872 gcall *bi_call = calls[i];
873 changed |= shrink_wrap_one_built_in_call (bi_call);
876 return changed;
879 namespace {
881 const pass_data pass_data_call_cdce =
883 GIMPLE_PASS, /* type */
884 "cdce", /* name */
885 OPTGROUP_NONE, /* optinfo_flags */
886 TV_TREE_CALL_CDCE, /* tv_id */
887 ( PROP_cfg | PROP_ssa ), /* properties_required */
888 0, /* properties_provided */
889 0, /* properties_destroyed */
890 0, /* todo_flags_start */
891 0, /* todo_flags_finish */
894 class pass_call_cdce : public gimple_opt_pass
896 public:
897 pass_call_cdce (gcc::context *ctxt)
898 : gimple_opt_pass (pass_data_call_cdce, ctxt)
901 /* opt_pass methods: */
902 virtual bool gate (function *fun)
904 /* The limit constants used in the implementation
905 assume IEEE floating point format. Other formats
906 can be supported in the future if needed. */
907 return flag_tree_builtin_call_dce != 0
908 && optimize_function_for_speed_p (fun);
911 virtual unsigned int execute (function *);
913 }; // class pass_call_cdce
915 unsigned int
916 pass_call_cdce::execute (function *fun)
918 basic_block bb;
919 gimple_stmt_iterator i;
920 bool something_changed = false;
921 auto_vec<gcall *> cond_dead_built_in_calls;
922 FOR_EACH_BB_FN (bb, fun)
924 /* Collect dead call candidates. */
925 for (i = gsi_start_bb (bb); !gsi_end_p (i); gsi_next (&i))
927 gcall *stmt = dyn_cast <gcall *> (gsi_stmt (i));
928 if (stmt && is_call_dce_candidate (stmt))
930 if (dump_file && (dump_flags & TDF_DETAILS))
932 fprintf (dump_file, "Found conditional dead call: ");
933 print_gimple_stmt (dump_file, stmt, 0, TDF_SLIM);
934 fprintf (dump_file, "\n");
936 if (!cond_dead_built_in_calls.exists ())
937 cond_dead_built_in_calls.create (64);
938 cond_dead_built_in_calls.safe_push (stmt);
943 if (!cond_dead_built_in_calls.exists ())
944 return 0;
946 something_changed
947 = shrink_wrap_conditional_dead_built_in_calls (cond_dead_built_in_calls);
949 if (something_changed)
951 free_dominance_info (CDI_DOMINATORS);
952 free_dominance_info (CDI_POST_DOMINATORS);
953 /* As we introduced new control-flow we need to insert PHI-nodes
954 for the call-clobbers of the remaining call. */
955 mark_virtual_operands_for_renaming (fun);
956 return TODO_update_ssa;
959 return 0;
962 } // anon namespace
964 gimple_opt_pass *
965 make_pass_call_cdce (gcc::context *ctxt)
967 return new pass_call_cdce (ctxt);