Support TI mode and soft float on PA64
[official-gcc.git] / gcc / tree-call-cdce.c
blobd9b9b4c6e843a022ab4956fe49d4158486372cd2
1 /* Conditional Dead Call Elimination pass for the GNU compiler.
2 Copyright (C) 2008-2021 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 "backend.h"
25 #include "tree.h"
26 #include "gimple.h"
27 #include "cfghooks.h"
28 #include "tree-pass.h"
29 #include "ssa.h"
30 #include "gimple-pretty-print.h"
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"
41 /* This pass serves two closely-related purposes:
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.
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.
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:
60 a. by protecting the original call with the same argument checks as (1)
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).
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.
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.
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.
80 An example of (1) is:
82 log (x); // Mostly dead call
83 ==>
84 if (__builtin_islessequal (x, 0))
85 log (x);
87 With this change, call to log (x) is effectively eliminated, as
88 in the majority of the cases, log won't be called with x out of
89 range. The branch is totally predictable, so the branch cost
90 is low.
92 An example of (2) is:
94 y = sqrt (x);
95 ==>
96 if (__builtin_isless (x, 0))
97 y = sqrt (x);
98 else
99 y = IFN_SQRT (x);
100 In the vast majority of cases we should then never need to call sqrt.
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).
106 The condition wrapping the builtin call is conservatively set to avoid too
107 aggressive (wrong) shrink wrapping. */
110 /* A structure for representing input domain of
111 a function argument in integer. If the lower
112 bound is -inf, has_lb is set to false. If the
113 upper bound is +inf, has_ub is false.
114 is_lb_inclusive and is_ub_inclusive are flags
115 to indicate if lb and ub value are inclusive
116 respectively. */
118 struct inp_domain
120 int lb;
121 int ub;
122 bool has_lb;
123 bool has_ub;
124 bool is_lb_inclusive;
125 bool is_ub_inclusive;
128 /* A helper function to construct and return an input
129 domain object. LB is the lower bound, HAS_LB is
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
133 UB_INCLUSIVE have the same meaning, but for upper
134 bound of the domain. */
136 static inp_domain
137 get_domain (int lb, bool has_lb, bool lb_inclusive,
138 int ub, bool has_ub, bool ub_inclusive)
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;
150 /* A helper function to check the target format for the
151 argument type. In this implementation, only IEEE formats
152 are supported. ARG is the call argument to be checked.
153 Returns true if the format is supported. To support other
154 target formats, function get_no_error_domain needs to be
155 enhanced to have range bounds properly computed. Since
156 the check is cheap (very small number of candidates
157 to be checked), the result is not cached for each float type. */
159 static bool
160 check_target_format (tree arg)
162 tree type;
163 machine_mode mode;
164 const struct real_format *rfmt;
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))
175 /* For long double, we cannot really check XFmode
176 which is only defined on intel platforms.
177 Candidate pre-selection using builtin function
178 code guarantees that we are checking formats
179 for long double modes: double, quad, and extended. */
180 || (mode != SFmode && mode != DFmode
181 && (rfmt == &ieee_quad_format
182 || rfmt == &mips_quad_format
183 || rfmt == &ieee_extended_motorola_format
184 || rfmt == &ieee_extended_intel_96_format
185 || rfmt == &ieee_extended_intel_128_format
186 || rfmt == &ieee_extended_intel_96_round_53_format)))
187 return true;
189 return false;
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
196 values or base values converted from integers. Returns true if
197 the pow call POW_CALL is a candidate. */
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
203 static bool
204 check_pow (gcall *pow_call)
206 tree base, expn;
207 enum tree_code bc, ec;
209 if (gimple_call_num_args (pow_call) != 2)
210 return false;
212 base = gimple_call_arg (pow_call, 0);
213 expn = gimple_call_arg (pow_call, 1);
215 if (!check_target_format (expn))
216 return false;
218 bc = TREE_CODE (base);
219 ec = TREE_CODE (expn);
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;
226 if (bc == REAL_CST)
228 /* Only handle a fixed range of constant. */
229 REAL_VALUE_TYPE mv;
230 REAL_VALUE_TYPE bcv = TREE_REAL_CST (base);
231 if (real_equal (&bcv, &dconst1))
232 return false;
233 if (real_less (&bcv, &dconst1))
234 return false;
235 real_from_integer (&mv, TYPE_MODE (TREE_TYPE (base)), 256, UNSIGNED);
236 if (real_less (&mv, &bcv))
237 return false;
238 return true;
240 else if (bc == SSA_NAME)
242 tree base_val0, type;
243 gimple *base_def;
244 int bit_sz;
246 /* Only handles cases where base value is converted
247 from integer values. */
248 base_def = SSA_NAME_DEF_STMT (base);
249 if (gimple_code (base_def) != GIMPLE_ASSIGN)
250 return false;
252 if (gimple_assign_rhs_code (base_def) != FLOAT_EXPR)
253 return false;
254 base_val0 = gimple_assign_rhs1 (base_def);
256 type = TREE_TYPE (base_val0);
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;
266 return true;
268 else
269 return false;
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. */
277 static bool
278 check_builtin_call (gcall *bcall)
280 tree arg;
282 arg = gimple_call_arg (bcall, 0);
283 return check_target_format (arg);
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. */
291 static bool
292 can_test_argument_range (gcall *call)
294 switch (DECL_FUNCTION_CODE (gimple_call_fndecl (call)))
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):
317 CASE_FLT_FN_FLOATN_NX (BUILT_IN_SQRT):
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;
326 return false;
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. */
334 static bool
335 edom_only_function (gcall *call)
337 switch (DECL_FUNCTION_CODE (gimple_call_fndecl (call)))
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):
346 CASE_FLT_FN_FLOATN_NX (BUILT_IN_SQRT):
347 CASE_FLT_FN (BUILT_IN_FMOD):
348 CASE_FLT_FN (BUILT_IN_REMAINDER):
349 return true;
351 default:
352 return false;
356 /* Return true if it is structurally possible to guard CALL. */
358 static bool
359 can_guard_call_p (gimple *call)
361 return (!stmt_ends_bb_p (call)
362 || find_fallthru_edge (gimple_bb (call)->succs));
365 /* For a comparison code return the comparison code we should use if we don't
366 HONOR_NANS. */
368 static enum tree_code
369 comparison_code_if_no_nans (tree_code code)
371 switch (code)
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;
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;
394 default:
395 gcc_unreachable ();
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. */
407 static void
408 gen_one_condition (tree arg, int lbub,
409 enum tree_code tcode,
410 const char *temp_name1,
411 const char *temp_name2,
412 vec<gimple *> conds,
413 unsigned *nconds)
415 if (!HONOR_NANS (arg))
416 tcode = comparison_code_if_no_nans (tcode);
418 tree lbub_real_cst, lbub_cst, float_type;
419 tree temp, tempn, tempc, tempcn;
420 gassign *stmt1;
421 gassign *stmt2;
422 gcond *stmt3;
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);
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);
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);
441 stmt3 = gimple_build_cond_from_tree (tempcn, NULL_TREE, NULL_TREE);
442 conds.quick_push (stmt1);
443 conds.quick_push (stmt2);
444 conds.quick_push (stmt3);
445 (*nconds)++;
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
452 holding the result GIMPLE statements. *NCONDS is
453 the number of logical comparisons. This function
454 produces no more than two logical comparisons, one
455 for lower bound check, one for upper bound check. */
457 static void
458 gen_conditions_for_domain (tree arg, inp_domain domain,
459 vec<gimple *> conds,
460 unsigned *nconds)
462 if (domain.has_lb)
463 gen_one_condition (arg, domain.lb,
464 (domain.is_lb_inclusive
465 ? UNGE_EXPR : UNGT_EXPR),
466 "DCE_COND_LB", "DCE_COND_LB_TEST",
467 conds, nconds);
469 if (domain.has_ub)
471 /* Now push a separator. */
472 if (domain.has_lb)
473 conds.quick_push (NULL);
475 gen_one_condition (arg, domain.ub,
476 (domain.is_ub_inclusive
477 ? UNLE_EXPR : UNLT_EXPR),
478 "DCE_COND_UB", "DCE_COND_UB_TEST",
479 conds, nconds);
484 /* A helper function to generate condition
485 code for the y argument in call pow (some_const, y).
486 See candidate selection in check_pow. Since the
487 candidates' base values have a limited range,
488 the guarded code generated for y are simple:
489 if (__builtin_isgreater (y, max_y))
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. */
500 static void
501 gen_conditions_for_pow_cst_base (tree base, tree expn,
502 vec<gimple *> conds,
503 unsigned *nconds)
505 inp_domain exp_domain;
506 /* Validate the range of the base constant to make
507 sure it is consistent with check_pow. */
508 REAL_VALUE_TYPE mv;
509 REAL_VALUE_TYPE bcv = TREE_REAL_CST (base);
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));
515 exp_domain = get_domain (0, false, false,
516 127, true, false);
518 gen_conditions_for_domain (expn, exp_domain,
519 conds, nconds);
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
529 conservative (smaller than the max value allowed by
530 the runtime value of the base). BASE is the integer
531 base value, EXPN is the expression for the exponent
532 argument, *CONDS is the vector to hold resulting
533 statements, and *NCONDS is the number of logical
534 conditions. */
536 static void
537 gen_conditions_for_pow_int_base (tree base, tree expn,
538 vec<gimple *> conds,
539 unsigned *nconds)
541 gimple *base_def;
542 tree base_val0;
543 tree int_type;
544 tree temp, tempn;
545 tree cst0;
546 gimple *stmt1, *stmt2;
547 int bit_sz, max_exp;
548 inp_domain exp_domain;
550 base_def = SSA_NAME_DEF_STMT (base);
551 base_val0 = gimple_assign_rhs1 (base_def);
552 int_type = TREE_TYPE (base_val0);
553 bit_sz = TYPE_PRECISION (int_type);
554 gcc_assert (bit_sz > 0
555 && bit_sz <= MAX_BASE_INT_BIT_SIZE);
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
567 gcc_assert (bit_sz == MAX_BASE_INT_BIT_SIZE);
568 max_exp = 32;
571 /* For pow ((double)x, y), generate the following conditions:
572 cond 1:
573 temp1 = x;
574 if (__builtin_islessequal (temp1, 0))
576 cond 2:
577 temp2 = y;
578 if (__builtin_isgreater (temp2, max_exp_real_cst)) */
580 /* Generate condition in reverse order -- first
581 the condition for the exp argument. */
583 exp_domain = get_domain (0, false, false,
584 max_exp, true, true);
586 gen_conditions_for_domain (expn, exp_domain,
587 conds, nconds);
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. */
594 /* Push a separator. */
595 conds.quick_push (NULL);
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);
602 stmt2 = gimple_build_cond (GT_EXPR, tempn, cst0, NULL_TREE, NULL_TREE);
604 conds.quick_push (stmt1);
605 conds.quick_push (stmt2);
606 (*nconds)++;
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
612 are separated by a NULL tree and they are stored in the vec
613 conds. The number of logical conditions are stored in *nconds.
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. */
626 static void
627 gen_conditions_for_pow (gcall *pow_call, vec<gimple *> conds,
628 unsigned *nconds)
630 tree base, expn;
631 enum tree_code bc;
633 gcc_checking_assert (check_pow (pow_call));
635 *nconds = 0;
637 base = gimple_call_arg (pow_call, 0);
638 expn = gimple_call_arg (pow_call, 1);
640 bc = TREE_CODE (base);
642 if (bc == REAL_CST)
643 gen_conditions_for_pow_cst_base (base, expn, conds, nconds);
644 else if (bc == SSA_NAME)
645 gen_conditions_for_pow_int_base (base, expn, conds, nconds);
646 else
647 gcc_unreachable ();
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
656 assuming IEEE floating point format (for SF and DF modes).
657 Since IEEE only sets minimum requirements for long double format,
658 different long double formats exist under different implementations
659 (e.g, 64 bit double precision (DF), 80 bit double-extended
660 precision (XF), and 128 bit quad precision (QF) ). For simplicity,
661 in this implementation, the computed bounds for long double assume
662 64 bit format (DF), and are therefore conservative. Another
663 assumption is that single precision float type is always SF mode,
664 and double type is DF mode. This function is quite
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. */
669 static inp_domain
670 get_no_error_domain (enum built_in_function fnc)
672 switch (fnc)
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):
745 CASE_FLT_FN_FLOATN_NX (BUILT_IN_SQRT):
746 return get_domain (0, true, true,
747 0, false, false);
748 default:
749 gcc_unreachable ();
752 gcc_unreachable ();
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.
758 BI_CALL is the builtin call, CONDS is the vector of statements
759 for condition code, NCODES is the pointer to the number of
760 logical conditions. Statements belonging to different logical
761 condition are separated by NULL tree in the vector. */
763 static void
764 gen_shrink_wrap_conditions (gcall *bi_call, const vec<gimple *> &conds,
765 unsigned int *nconds)
767 gcall *call;
768 tree fn;
769 enum built_in_function fnc;
771 gcc_assert (nconds && conds.exists ());
772 gcc_assert (conds.length () == 0);
773 gcc_assert (is_gimple_call (bi_call));
775 call = bi_call;
776 fn = gimple_call_fndecl (call);
777 gcc_assert (fn && fndecl_built_in_p (fn));
778 fnc = DECL_FUNCTION_CODE (fn);
779 *nconds = 0;
781 if (fnc == BUILT_IN_POW)
782 gen_conditions_for_pow (call, conds, nconds);
783 else
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);
792 return;
795 /* Shrink-wrap BI_CALL so that it is only called when one of the NCONDS
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. */
799 static void
800 shrink_wrap_one_built_in_call_with_conds (gcall *bi_call,
801 const vec <gimple *> &conds,
802 unsigned int nconds,
803 gcall *bi_newcall = NULL)
805 gimple_stmt_iterator bi_call_bsi;
806 basic_block bi_call_bb, bi_newcall_bb, join_tgt_bb, guard_bb;
807 edge join_tgt_in_edge_from_call, join_tgt_in_edge_fall_thru;
808 edge bi_call_in_edge0, guard_bb_in_edge;
809 unsigned tn_cond_stmts;
810 unsigned ci;
811 gimple *cond_expr = NULL;
812 gimple *cond_expr_start;
814 /* The cfg we want to create looks like this:
815 [guard n-1] <- guard_bb (old block)
817 | [guard n-2] }
818 | / \ }
819 | / ... } new blocks
820 | / [guard 0] }
821 | / / | }
822 [call] | <- bi_call_bb }
823 \ [newcall] <-bi_newcall_bb}
825 [join] <- join_tgt_bb (old iff call must end bb)
826 possible EH edges (only if [join] is old)
828 When [join] is new, the immediate dominators for these blocks are:
830 1. [guard n-1]: unchanged
831 2. [call]: [guard n-1]
832 3. [newcall]: [guard 0]
833 4. [guard m]: [guard m+1] for 0 <= m <= n-2
834 5. [join]: [guard n-1]
836 We punt for the more complex case of [join] being old and
837 simply free the dominance info. We also punt on postdominators,
838 which aren't expected to be available at this point anyway. */
839 bi_call_bb = gimple_bb (bi_call);
841 /* Now find the join target bb -- split bi_call_bb if needed. */
842 if (stmt_ends_bb_p (bi_call))
844 /* We checked that there was a fallthrough edge in
845 can_guard_call_p. */
846 join_tgt_in_edge_from_call = find_fallthru_edge (bi_call_bb->succs);
847 gcc_assert (join_tgt_in_edge_from_call);
848 /* We don't want to handle PHIs. */
849 if (EDGE_COUNT (join_tgt_in_edge_from_call->dest->preds) > 1)
850 join_tgt_bb = split_edge (join_tgt_in_edge_from_call);
851 else
853 join_tgt_bb = join_tgt_in_edge_from_call->dest;
854 /* We may have degenerate PHIs in the destination. Propagate
855 those out. */
856 for (gphi_iterator i = gsi_start_phis (join_tgt_bb); !gsi_end_p (i);)
858 gphi *phi = i.phi ();
859 replace_uses_by (gimple_phi_result (phi),
860 gimple_phi_arg_def (phi, 0));
861 remove_phi_node (&i, true);
865 else
867 join_tgt_in_edge_from_call = split_block (bi_call_bb, bi_call);
868 join_tgt_bb = join_tgt_in_edge_from_call->dest;
871 bi_call_bsi = gsi_for_stmt (bi_call);
873 /* Now it is time to insert the first conditional expression
874 into bi_call_bb and split this bb so that bi_call is
875 shrink-wrapped. */
876 tn_cond_stmts = conds.length ();
877 cond_expr = NULL;
878 cond_expr_start = conds[0];
879 for (ci = 0; ci < tn_cond_stmts; ci++)
881 gimple *c = conds[ci];
882 gcc_assert (c || ci != 0);
883 if (!c)
884 break;
885 gsi_insert_before (&bi_call_bsi, c, GSI_SAME_STMT);
886 cond_expr = c;
888 ci++;
889 gcc_assert (cond_expr && gimple_code (cond_expr) == GIMPLE_COND);
891 typedef std::pair<edge, edge> edge_pair;
892 auto_vec<edge_pair, 8> edges;
894 bi_call_in_edge0 = split_block (bi_call_bb, cond_expr);
895 bi_call_in_edge0->flags &= ~EDGE_FALLTHRU;
896 bi_call_in_edge0->flags |= EDGE_FALSE_VALUE;
897 guard_bb = bi_call_bb;
898 bi_call_bb = bi_call_in_edge0->dest;
899 join_tgt_in_edge_fall_thru = make_edge (guard_bb, join_tgt_bb,
900 EDGE_TRUE_VALUE);
902 edges.reserve (nconds);
903 edges.quick_push (edge_pair (bi_call_in_edge0, join_tgt_in_edge_fall_thru));
905 /* Code generation for the rest of the conditions */
906 for (unsigned int i = 1; i < nconds; ++i)
908 unsigned ci0;
909 edge bi_call_in_edge;
910 gimple_stmt_iterator guard_bsi = gsi_for_stmt (cond_expr_start);
911 ci0 = ci;
912 cond_expr_start = conds[ci0];
913 for (; ci < tn_cond_stmts; ci++)
915 gimple *c = conds[ci];
916 gcc_assert (c || ci != ci0);
917 if (!c)
918 break;
919 gsi_insert_before (&guard_bsi, c, GSI_SAME_STMT);
920 cond_expr = c;
922 ci++;
923 gcc_assert (cond_expr && gimple_code (cond_expr) == GIMPLE_COND);
924 guard_bb_in_edge = split_block (guard_bb, cond_expr);
925 guard_bb_in_edge->flags &= ~EDGE_FALLTHRU;
926 guard_bb_in_edge->flags |= EDGE_TRUE_VALUE;
928 bi_call_in_edge = make_edge (guard_bb, bi_call_bb, EDGE_FALSE_VALUE);
929 edges.quick_push (edge_pair (bi_call_in_edge, guard_bb_in_edge));
932 /* Move BI_NEWCALL to new basic block when it is non-null. */
933 if (bi_newcall)
935 /* Get bi_newcall_bb by split join_tgt_in_edge_fall_thru edge,
936 and move BI_NEWCALL to bi_newcall_bb. */
937 bi_newcall_bb = split_edge (join_tgt_in_edge_fall_thru);
938 gimple_stmt_iterator to_gsi = gsi_start_bb (bi_newcall_bb);
939 gimple_stmt_iterator from_gsi = gsi_for_stmt (bi_newcall);
940 gsi_move_before (&from_gsi, &to_gsi);
941 join_tgt_in_edge_fall_thru = EDGE_SUCC (bi_newcall_bb, 0);
942 join_tgt_bb = join_tgt_in_edge_fall_thru->dest;
944 tree bi_newcall_lhs = gimple_call_lhs (bi_newcall);
945 tree bi_call_lhs = gimple_call_lhs (bi_call);
946 if (!bi_call_lhs)
948 bi_call_lhs = copy_ssa_name (bi_newcall_lhs);
949 gimple_call_set_lhs (bi_call, bi_call_lhs);
950 SSA_NAME_DEF_STMT (bi_call_lhs) = bi_call;
953 /* Create phi node for lhs of BI_CALL and BI_NEWCALL. */
954 gphi *new_phi = create_phi_node (copy_ssa_name (bi_newcall_lhs),
955 join_tgt_bb);
956 SSA_NAME_OCCURS_IN_ABNORMAL_PHI (PHI_RESULT (new_phi))
957 = SSA_NAME_OCCURS_IN_ABNORMAL_PHI (bi_newcall_lhs);
958 add_phi_arg (new_phi, bi_call_lhs, join_tgt_in_edge_from_call,
959 gimple_location (bi_call));
960 add_phi_arg (new_phi, bi_newcall_lhs, join_tgt_in_edge_fall_thru,
961 gimple_location (bi_newcall));
963 /* Replace all use of original return value with result of phi node. */
964 use_operand_p use_p;
965 gimple *use_stmt;
966 imm_use_iterator iterator;
967 FOR_EACH_IMM_USE_STMT (use_stmt, iterator, bi_newcall_lhs)
968 if (use_stmt != new_phi)
969 FOR_EACH_IMM_USE_ON_STMT (use_p, iterator)
970 SET_USE (use_p, PHI_RESULT (new_phi));
973 /* Now update the probability and profile information, processing the
974 guards in order of execution.
976 There are two approaches we could take here. On the one hand we
977 could assign a probability of X to the call block and distribute
978 that probability among its incoming edges. On the other hand we
979 could assign a probability of X to each individual call edge.
981 The choice only affects calls that have more than one condition.
982 In those cases, the second approach would give the call block
983 a greater probability than the first. However, the difference
984 is only small, and our chosen X is a pure guess anyway.
986 Here we take the second approach because it's slightly simpler
987 and because it's easy to see that it doesn't lose profile counts. */
988 bi_call_bb->count = profile_count::zero ();
989 while (!edges.is_empty ())
991 edge_pair e = edges.pop ();
992 edge call_edge = e.first;
993 edge nocall_edge = e.second;
994 basic_block src_bb = call_edge->src;
995 gcc_assert (src_bb == nocall_edge->src);
997 call_edge->probability = profile_probability::very_unlikely ();
998 nocall_edge->probability = profile_probability::always ()
999 - call_edge->probability;
1001 bi_call_bb->count += call_edge->count ();
1003 if (nocall_edge->dest != join_tgt_bb)
1004 nocall_edge->dest->count = src_bb->count - bi_call_bb->count;
1007 if (dom_info_available_p (CDI_DOMINATORS))
1009 /* The split_blocks leave [guard 0] as the immediate dominator
1010 of [call] and [call] as the immediate dominator of [join].
1011 Fix them up. */
1012 set_immediate_dominator (CDI_DOMINATORS, bi_call_bb, guard_bb);
1013 set_immediate_dominator (CDI_DOMINATORS, join_tgt_bb, guard_bb);
1016 if (dump_file && (dump_flags & TDF_DETAILS))
1018 location_t loc;
1019 loc = gimple_location (bi_call);
1020 fprintf (dump_file,
1021 "%s:%d: note: function call is shrink-wrapped"
1022 " into error conditions.\n",
1023 LOCATION_FILE (loc), LOCATION_LINE (loc));
1027 /* Shrink-wrap BI_CALL so that it is only called when it might set errno
1028 (but is always called if it would set errno). */
1030 static void
1031 shrink_wrap_one_built_in_call (gcall *bi_call)
1033 unsigned nconds = 0;
1034 auto_vec<gimple *, 12> conds;
1035 gen_shrink_wrap_conditions (bi_call, conds, &nconds);
1036 gcc_assert (nconds != 0);
1037 shrink_wrap_one_built_in_call_with_conds (bi_call, conds, nconds);
1040 /* Return true if built-in function call CALL could be implemented using
1041 a combination of an internal function to compute the result and a
1042 separate call to set errno. */
1044 static bool
1045 can_use_internal_fn (gcall *call)
1047 /* Only replace calls that set errno. */
1048 if (!gimple_vdef (call))
1049 return false;
1051 /* See whether there is an internal function for this built-in. */
1052 if (replacement_internal_fn (call) == IFN_LAST)
1053 return false;
1055 /* See whether we can catch all cases where errno would be set,
1056 while still avoiding the call in most cases. */
1057 if (!can_test_argument_range (call)
1058 && !edom_only_function (call))
1059 return false;
1061 return true;
1064 /* Implement built-in function call CALL using an internal function. */
1066 static void
1067 use_internal_fn (gcall *call)
1069 /* We'll be inserting another call with the same arguments after the
1070 lhs has been set, so prevent any possible coalescing failure from
1071 having both values live at once. See PR 71020. */
1072 replace_abnormal_ssa_names (call);
1074 unsigned nconds = 0;
1075 auto_vec<gimple *, 12> conds;
1076 bool is_arg_conds = false;
1077 if (can_test_argument_range (call))
1079 gen_shrink_wrap_conditions (call, conds, &nconds);
1080 is_arg_conds = true;
1081 gcc_assert (nconds != 0);
1083 else
1084 gcc_assert (edom_only_function (call));
1086 internal_fn ifn = replacement_internal_fn (call);
1087 gcc_assert (ifn != IFN_LAST);
1089 /* Construct the new call, with the same arguments as the original one. */
1090 auto_vec <tree, 16> args;
1091 unsigned int nargs = gimple_call_num_args (call);
1092 for (unsigned int i = 0; i < nargs; ++i)
1093 args.safe_push (gimple_call_arg (call, i));
1094 gcall *new_call = gimple_build_call_internal_vec (ifn, args);
1095 gimple_set_location (new_call, gimple_location (call));
1096 gimple_call_set_nothrow (new_call, gimple_call_nothrow_p (call));
1098 /* Transfer the LHS to the new call. */
1099 tree lhs = gimple_call_lhs (call);
1100 gimple_call_set_lhs (new_call, lhs);
1101 gimple_call_set_lhs (call, NULL_TREE);
1102 SSA_NAME_DEF_STMT (lhs) = new_call;
1104 /* Insert the new call. */
1105 gimple_stmt_iterator gsi = gsi_for_stmt (call);
1106 gsi_insert_before (&gsi, new_call, GSI_SAME_STMT);
1108 if (nconds == 0)
1110 /* Skip the call if LHS == LHS. If we reach here, EDOM is the only
1111 valid errno value and it is used iff the result is NaN. */
1112 conds.quick_push (gimple_build_cond (EQ_EXPR, lhs, lhs,
1113 NULL_TREE, NULL_TREE));
1114 nconds++;
1116 /* Try replacing the original call with a direct assignment to
1117 errno, via an internal function. */
1118 if (set_edom_supported_p () && !stmt_ends_bb_p (call))
1120 gimple_stmt_iterator gsi = gsi_for_stmt (call);
1121 gcall *new_call = gimple_build_call_internal (IFN_SET_EDOM, 0);
1122 gimple_move_vops (new_call, call);
1123 gimple_set_location (new_call, gimple_location (call));
1124 gsi_replace (&gsi, new_call, false);
1125 call = new_call;
1128 shrink_wrap_one_built_in_call_with_conds (call, conds, nconds,
1129 is_arg_conds ? new_call : NULL);
1132 /* The top level function for conditional dead code shrink
1133 wrapping transformation. */
1135 static void
1136 shrink_wrap_conditional_dead_built_in_calls (const vec<gcall *> &calls)
1138 unsigned i = 0;
1140 unsigned n = calls.length ();
1141 for (; i < n ; i++)
1143 gcall *bi_call = calls[i];
1144 if (gimple_call_lhs (bi_call))
1145 use_internal_fn (bi_call);
1146 else
1147 shrink_wrap_one_built_in_call (bi_call);
1151 namespace {
1153 const pass_data pass_data_call_cdce =
1155 GIMPLE_PASS, /* type */
1156 "cdce", /* name */
1157 OPTGROUP_NONE, /* optinfo_flags */
1158 TV_TREE_CALL_CDCE, /* tv_id */
1159 ( PROP_cfg | PROP_ssa ), /* properties_required */
1160 0, /* properties_provided */
1161 0, /* properties_destroyed */
1162 0, /* todo_flags_start */
1163 0, /* todo_flags_finish */
1166 class pass_call_cdce : public gimple_opt_pass
1168 public:
1169 pass_call_cdce (gcc::context *ctxt)
1170 : gimple_opt_pass (pass_data_call_cdce, ctxt)
1173 /* opt_pass methods: */
1174 virtual bool gate (function *)
1176 /* The limit constants used in the implementation
1177 assume IEEE floating point format. Other formats
1178 can be supported in the future if needed. */
1179 return flag_tree_builtin_call_dce != 0;
1182 virtual unsigned int execute (function *);
1184 }; // class pass_call_cdce
1186 unsigned int
1187 pass_call_cdce::execute (function *fun)
1189 basic_block bb;
1190 gimple_stmt_iterator i;
1191 auto_vec<gcall *> cond_dead_built_in_calls;
1192 FOR_EACH_BB_FN (bb, fun)
1194 /* Skip blocks that are being optimized for size, since our
1195 transformation always increases code size. */
1196 if (optimize_bb_for_size_p (bb))
1197 continue;
1199 /* Collect dead call candidates. */
1200 for (i = gsi_start_bb (bb); !gsi_end_p (i); gsi_next (&i))
1202 gcall *stmt = dyn_cast <gcall *> (gsi_stmt (i));
1203 if (stmt
1204 && gimple_call_builtin_p (stmt, BUILT_IN_NORMAL)
1205 && (gimple_call_lhs (stmt)
1206 ? can_use_internal_fn (stmt)
1207 : can_test_argument_range (stmt))
1208 && can_guard_call_p (stmt))
1210 if (dump_file && (dump_flags & TDF_DETAILS))
1212 fprintf (dump_file, "Found conditional dead call: ");
1213 print_gimple_stmt (dump_file, stmt, 0, TDF_SLIM);
1214 fprintf (dump_file, "\n");
1216 if (!cond_dead_built_in_calls.exists ())
1217 cond_dead_built_in_calls.create (64);
1218 cond_dead_built_in_calls.safe_push (stmt);
1223 if (!cond_dead_built_in_calls.exists ())
1224 return 0;
1226 shrink_wrap_conditional_dead_built_in_calls (cond_dead_built_in_calls);
1227 free_dominance_info (CDI_POST_DOMINATORS);
1228 /* As we introduced new control-flow we need to insert PHI-nodes
1229 for the call-clobbers of the remaining call. */
1230 mark_virtual_operands_for_renaming (fun);
1231 return TODO_update_ssa;
1234 } // anon namespace
1236 gimple_opt_pass *
1237 make_pass_call_cdce (gcc::context *ctxt)
1239 return new pass_call_cdce (ctxt);