]>
Commit | Line | Data |
---|---|---|
ac534736 | 1 | /* Tail call optimization on trees. |
ebb07520 | 2 | Copyright (C) 2003, 2004, 2005, 2006, 2007 Free Software Foundation, Inc. |
6de9cd9a DN |
3 | |
4 | This file is part of GCC. | |
5 | ||
6 | GCC is free software; you can redistribute it and/or modify | |
7 | it under the terms of the GNU General Public License as published by | |
9dcd6f09 | 8 | the Free Software Foundation; either version 3, or (at your option) |
6de9cd9a DN |
9 | any later version. |
10 | ||
11 | GCC is distributed in the hope that it will be useful, | |
12 | but WITHOUT ANY WARRANTY; without even the implied warranty of | |
13 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | |
14 | GNU General Public License for more details. | |
15 | ||
16 | You should have received a copy of the GNU General Public License | |
9dcd6f09 NC |
17 | along with GCC; see the file COPYING3. If not see |
18 | <http://www.gnu.org/licenses/>. */ | |
6de9cd9a DN |
19 | |
20 | #include "config.h" | |
21 | #include "system.h" | |
22 | #include "coretypes.h" | |
23 | #include "tm.h" | |
24 | #include "tree.h" | |
25 | #include "rtl.h" | |
26 | #include "tm_p.h" | |
27 | #include "hard-reg-set.h" | |
28 | #include "basic-block.h" | |
29 | #include "function.h" | |
30 | #include "tree-flow.h" | |
31 | #include "tree-dump.h" | |
32 | #include "diagnostic.h" | |
33 | #include "except.h" | |
34 | #include "tree-pass.h" | |
35 | #include "flags.h" | |
36 | #include "langhooks.h" | |
6fb5fa3c | 37 | #include "dbgcnt.h" |
6de9cd9a DN |
38 | |
39 | /* The file implements the tail recursion elimination. It is also used to | |
1ea7e6ad | 40 | analyze the tail calls in general, passing the results to the rtl level |
6de9cd9a DN |
41 | where they are used for sibcall optimization. |
42 | ||
43 | In addition to the standard tail recursion elimination, we handle the most | |
44 | trivial cases of making the call tail recursive by creating accumulators. | |
45 | For example the following function | |
46 | ||
47 | int sum (int n) | |
48 | { | |
49 | if (n > 0) | |
50 | return n + sum (n - 1); | |
51 | else | |
52 | return 0; | |
53 | } | |
54 | ||
55 | is transformed into | |
56 | ||
57 | int sum (int n) | |
58 | { | |
59 | int acc = 0; | |
60 | ||
61 | while (n > 0) | |
62 | acc += n--; | |
63 | ||
64 | return acc; | |
65 | } | |
66 | ||
67 | To do this, we maintain two accumulators (a_acc and m_acc) that indicate | |
68 | when we reach the return x statement, we should return a_acc + x * m_acc | |
69 | instead. They are initially initialized to 0 and 1, respectively, | |
70 | so the semantics of the function is obviously preserved. If we are | |
71 | guaranteed that the value of the accumulator never change, we | |
72 | omit the accumulator. | |
73 | ||
74 | There are three cases how the function may exit. The first one is | |
6ce2bcb7 | 75 | handled in adjust_return_value, the other two in adjust_accumulator_values |
6de9cd9a DN |
76 | (the second case is actually a special case of the third one and we |
77 | present it separately just for clarity): | |
78 | ||
79 | 1) Just return x, where x is not in any of the remaining special shapes. | |
80 | We rewrite this to a gimple equivalent of return m_acc * x + a_acc. | |
81 | ||
82 | 2) return f (...), where f is the current function, is rewritten in a | |
1ea7e6ad | 83 | classical tail-recursion elimination way, into assignment of arguments |
6de9cd9a DN |
84 | and jump to the start of the function. Values of the accumulators |
85 | are unchanged. | |
86 | ||
87 | 3) return a + m * f(...), where a and m do not depend on call to f. | |
88 | To preserve the semantics described before we want this to be rewritten | |
89 | in such a way that we finally return | |
90 | ||
91 | a_acc + (a + m * f(...)) * m_acc = (a_acc + a * m_acc) + (m * m_acc) * f(...). | |
92 | ||
93 | I.e. we increase a_acc by a * m_acc, multiply m_acc by m and | |
94 | eliminate the tail call to f. Special cases when the value is just | |
95 | added or just multiplied are obtained by setting a = 0 or m = 1. | |
96 | ||
97 | TODO -- it is possible to do similar tricks for other operations. */ | |
98 | ||
99 | /* A structure that describes the tailcall. */ | |
100 | ||
101 | struct tailcall | |
102 | { | |
6de9cd9a | 103 | /* The iterator pointing to the call statement. */ |
726a989a | 104 | gimple_stmt_iterator call_gsi; |
6de9cd9a DN |
105 | |
106 | /* True if it is a call to the current function. */ | |
107 | bool tail_recursion; | |
108 | ||
109 | /* The return value of the caller is mult * f + add, where f is the return | |
110 | value of the call. */ | |
111 | tree mult, add; | |
112 | ||
113 | /* Next tailcall in the chain. */ | |
114 | struct tailcall *next; | |
115 | }; | |
116 | ||
117 | /* The variables holding the value of multiplicative and additive | |
118 | accumulator. */ | |
119 | static tree m_acc, a_acc; | |
120 | ||
121 | static bool suitable_for_tail_opt_p (void); | |
122 | static bool optimize_tail_call (struct tailcall *, bool); | |
123 | static void eliminate_tail_call (struct tailcall *); | |
124 | static void find_tail_calls (basic_block, struct tailcall **); | |
125 | ||
126 | /* Returns false when the function is not suitable for tail call optimization | |
127 | from some reason (e.g. if it takes variable number of arguments). */ | |
128 | ||
129 | static bool | |
130 | suitable_for_tail_opt_p (void) | |
131 | { | |
a3648cfc DB |
132 | referenced_var_iterator rvi; |
133 | tree var; | |
6de9cd9a | 134 | |
e3b5732b | 135 | if (cfun->stdarg) |
6de9cd9a DN |
136 | return false; |
137 | ||
115340c7 | 138 | /* No local variable nor structure field should be call-used. We |
9044951e | 139 | ignore any kind of memory tag, as these are not real variables. */ |
a3648cfc DB |
140 | |
141 | FOR_EACH_REFERENCED_VAR (var, rvi) | |
6de9cd9a | 142 | { |
326eda4b | 143 | if (!is_global_var (var) |
5611cf0b | 144 | && !MTAG_P (var) |
115340c7 | 145 | && (gimple_aliases_computed_p (cfun)? is_call_used (var) |
7faade0f | 146 | : TREE_ADDRESSABLE (var))) |
6de9cd9a DN |
147 | return false; |
148 | } | |
149 | ||
150 | return true; | |
151 | } | |
152 | /* Returns false when the function is not suitable for tail call optimization | |
153 | from some reason (e.g. if it takes variable number of arguments). | |
154 | This test must pass in addition to suitable_for_tail_opt_p in order to make | |
155 | tail call discovery happen. */ | |
156 | ||
157 | static bool | |
158 | suitable_for_tail_call_opt_p (void) | |
159 | { | |
ead553a1 SB |
160 | tree param; |
161 | ||
6de9cd9a DN |
162 | /* alloca (until we have stack slot life analysis) inhibits |
163 | sibling call optimizations, but not tail recursion. */ | |
e3b5732b | 164 | if (cfun->calls_alloca) |
6de9cd9a DN |
165 | return false; |
166 | ||
167 | /* If we are using sjlj exceptions, we may need to add a call to | |
168 | _Unwind_SjLj_Unregister at exit of the function. Which means | |
169 | that we cannot do any sibcall transformations. */ | |
170 | if (USING_SJLJ_EXCEPTIONS && current_function_has_exception_handlers ()) | |
171 | return false; | |
172 | ||
173 | /* Any function that calls setjmp might have longjmp called from | |
174 | any called function. ??? We really should represent this | |
175 | properly in the CFG so that this needn't be special cased. */ | |
e3b5732b | 176 | if (cfun->calls_setjmp) |
6de9cd9a DN |
177 | return false; |
178 | ||
ead553a1 SB |
179 | /* ??? It is OK if the argument of a function is taken in some cases, |
180 | but not in all cases. See PR15387 and PR19616. Revisit for 4.1. */ | |
181 | for (param = DECL_ARGUMENTS (current_function_decl); | |
182 | param; | |
183 | param = TREE_CHAIN (param)) | |
184 | if (TREE_ADDRESSABLE (param)) | |
185 | return false; | |
186 | ||
6de9cd9a DN |
187 | return true; |
188 | } | |
189 | ||
190 | /* Checks whether the expression EXPR in stmt AT is independent of the | |
726a989a RB |
191 | statement pointed to by GSI (in a sense that we already know EXPR's value |
192 | at GSI). We use the fact that we are only called from the chain of | |
6de9cd9a | 193 | basic blocks that have only single successor. Returns the expression |
726a989a | 194 | containing the value of EXPR at GSI. */ |
6de9cd9a DN |
195 | |
196 | static tree | |
726a989a | 197 | independent_of_stmt_p (tree expr, gimple at, gimple_stmt_iterator gsi) |
6de9cd9a DN |
198 | { |
199 | basic_block bb, call_bb, at_bb; | |
200 | edge e; | |
628f6a4e | 201 | edge_iterator ei; |
6de9cd9a DN |
202 | |
203 | if (is_gimple_min_invariant (expr)) | |
204 | return expr; | |
205 | ||
206 | if (TREE_CODE (expr) != SSA_NAME) | |
207 | return NULL_TREE; | |
208 | ||
209 | /* Mark the blocks in the chain leading to the end. */ | |
726a989a RB |
210 | at_bb = gimple_bb (at); |
211 | call_bb = gimple_bb (gsi_stmt (gsi)); | |
c5cbcccf | 212 | for (bb = call_bb; bb != at_bb; bb = single_succ (bb)) |
6de9cd9a DN |
213 | bb->aux = &bb->aux; |
214 | bb->aux = &bb->aux; | |
215 | ||
216 | while (1) | |
217 | { | |
218 | at = SSA_NAME_DEF_STMT (expr); | |
726a989a | 219 | bb = gimple_bb (at); |
6de9cd9a | 220 | |
61ada8ae | 221 | /* The default definition or defined before the chain. */ |
6de9cd9a DN |
222 | if (!bb || !bb->aux) |
223 | break; | |
224 | ||
225 | if (bb == call_bb) | |
226 | { | |
726a989a RB |
227 | for (; !gsi_end_p (gsi); gsi_next (&gsi)) |
228 | if (gsi_stmt (gsi) == at) | |
6de9cd9a DN |
229 | break; |
230 | ||
726a989a | 231 | if (!gsi_end_p (gsi)) |
6de9cd9a DN |
232 | expr = NULL_TREE; |
233 | break; | |
234 | } | |
235 | ||
726a989a | 236 | if (gimple_code (at) != GIMPLE_PHI) |
6de9cd9a DN |
237 | { |
238 | expr = NULL_TREE; | |
239 | break; | |
240 | } | |
241 | ||
628f6a4e | 242 | FOR_EACH_EDGE (e, ei, bb->preds) |
6de9cd9a DN |
243 | if (e->src->aux) |
244 | break; | |
1e128c5f | 245 | gcc_assert (e); |
6de9cd9a | 246 | |
d00ad49b | 247 | expr = PHI_ARG_DEF_FROM_EDGE (at, e); |
066a0344 ZD |
248 | if (TREE_CODE (expr) != SSA_NAME) |
249 | { | |
250 | /* The value is a constant. */ | |
251 | break; | |
252 | } | |
6de9cd9a DN |
253 | } |
254 | ||
255 | /* Unmark the blocks. */ | |
c5cbcccf | 256 | for (bb = call_bb; bb != at_bb; bb = single_succ (bb)) |
6de9cd9a DN |
257 | bb->aux = NULL; |
258 | bb->aux = NULL; | |
259 | ||
260 | return expr; | |
261 | } | |
262 | ||
726a989a RB |
263 | /* Simulates the effect of an assignment STMT on the return value of the tail |
264 | recursive CALL passed in ASS_VAR. M and A are the multiplicative and the | |
265 | additive factor for the real return value. */ | |
6de9cd9a DN |
266 | |
267 | static bool | |
726a989a | 268 | process_assignment (gimple stmt, gimple_stmt_iterator call, tree *m, |
6de9cd9a DN |
269 | tree *a, tree *ass_var) |
270 | { | |
271 | tree op0, op1, non_ass_var; | |
726a989a RB |
272 | tree dest = gimple_assign_lhs (stmt); |
273 | enum tree_code code = gimple_assign_rhs_code (stmt); | |
274 | enum gimple_rhs_class rhs_class = get_gimple_rhs_class (code); | |
275 | tree src_var = gimple_assign_rhs1 (stmt); | |
276 | ||
b89e96ac SB |
277 | /* See if this is a simple copy operation of an SSA name to the function |
278 | result. In that case we may have a simple tail call. Ignore type | |
279 | conversions that can never produce extra code between the function | |
280 | call and the function return. */ | |
726a989a RB |
281 | if ((rhs_class == GIMPLE_SINGLE_RHS || gimple_assign_cast_p (stmt)) |
282 | && (TREE_CODE (src_var) == SSA_NAME)) | |
6de9cd9a | 283 | { |
726a989a RB |
284 | /* Reject a tailcall if the type conversion might need |
285 | additional code. */ | |
7352e27c | 286 | if (gimple_assign_cast_p (stmt) |
726a989a RB |
287 | && TYPE_MODE (TREE_TYPE (dest)) != TYPE_MODE (TREE_TYPE (src_var))) |
288 | return false; | |
289 | ||
b89e96ac | 290 | if (src_var != *ass_var) |
6de9cd9a DN |
291 | return false; |
292 | ||
293 | *ass_var = dest; | |
294 | return true; | |
295 | } | |
296 | ||
726a989a | 297 | if (rhs_class != GIMPLE_BINARY_RHS) |
6de9cd9a DN |
298 | return false; |
299 | ||
85d62520 RS |
300 | /* Accumulator optimizations will reverse the order of operations. |
301 | We can only do that for floating-point types if we're assuming | |
302 | that addition and multiplication are associative. */ | |
a1a82611 | 303 | if (!flag_associative_math) |
85d62520 RS |
304 | if (FLOAT_TYPE_P (TREE_TYPE (DECL_RESULT (current_function_decl)))) |
305 | return false; | |
306 | ||
6de9cd9a DN |
307 | /* We only handle the code like |
308 | ||
309 | x = call (); | |
310 | y = m * x; | |
311 | z = y + a; | |
312 | return z; | |
313 | ||
314 | TODO -- Extend it for cases where the linear transformation of the output | |
315 | is expressed in a more complicated way. */ | |
316 | ||
726a989a RB |
317 | op0 = gimple_assign_rhs1 (stmt); |
318 | op1 = gimple_assign_rhs2 (stmt); | |
6de9cd9a DN |
319 | |
320 | if (op0 == *ass_var | |
321 | && (non_ass_var = independent_of_stmt_p (op1, stmt, call))) | |
322 | ; | |
323 | else if (op1 == *ass_var | |
324 | && (non_ass_var = independent_of_stmt_p (op0, stmt, call))) | |
325 | ; | |
326 | else | |
327 | return false; | |
328 | ||
b89e96ac | 329 | switch (code) |
6de9cd9a DN |
330 | { |
331 | case PLUS_EXPR: | |
332 | /* There should be no previous addition. TODO -- it should be fairly | |
333 | straightforward to lift this restriction -- just allow storing | |
334 | more complicated expressions in *A, and gimplify it in | |
335 | adjust_accumulator_values. */ | |
336 | if (*a) | |
337 | return false; | |
338 | *a = non_ass_var; | |
339 | *ass_var = dest; | |
340 | return true; | |
341 | ||
342 | case MULT_EXPR: | |
343 | /* Similar remark applies here. Handling multiplication after addition | |
344 | is just slightly more complicated -- we need to multiply both *A and | |
345 | *M. */ | |
346 | if (*a || *m) | |
347 | return false; | |
348 | *m = non_ass_var; | |
349 | *ass_var = dest; | |
350 | return true; | |
351 | ||
726a989a RB |
352 | /* TODO -- Handle other codes (NEGATE_EXPR, MINUS_EXPR, |
353 | POINTER_PLUS_EXPR). */ | |
6de9cd9a DN |
354 | |
355 | default: | |
356 | return false; | |
357 | } | |
358 | } | |
359 | ||
360 | /* Propagate VAR through phis on edge E. */ | |
361 | ||
362 | static tree | |
363 | propagate_through_phis (tree var, edge e) | |
364 | { | |
365 | basic_block dest = e->dest; | |
726a989a RB |
366 | gimple_stmt_iterator gsi; |
367 | ||
368 | for (gsi = gsi_start_phis (dest); !gsi_end_p (gsi); gsi_next (&gsi)) | |
369 | { | |
370 | gimple phi = gsi_stmt (gsi); | |
371 | if (PHI_ARG_DEF_FROM_EDGE (phi, e) == var) | |
372 | return PHI_RESULT (phi); | |
373 | } | |
6de9cd9a DN |
374 | return var; |
375 | } | |
376 | ||
377 | /* Finds tailcalls falling into basic block BB. The list of found tailcalls is | |
378 | added to the start of RET. */ | |
379 | ||
380 | static void | |
381 | find_tail_calls (basic_block bb, struct tailcall **ret) | |
382 | { | |
726a989a RB |
383 | tree ass_var = NULL_TREE, ret_var, func, param; |
384 | gimple stmt, call = NULL; | |
385 | gimple_stmt_iterator gsi, agsi; | |
6de9cd9a DN |
386 | bool tail_recursion; |
387 | struct tailcall *nw; | |
388 | edge e; | |
389 | tree m, a; | |
390 | basic_block abb; | |
726a989a | 391 | size_t idx; |
6de9cd9a | 392 | |
c5cbcccf | 393 | if (!single_succ_p (bb)) |
6de9cd9a DN |
394 | return; |
395 | ||
726a989a | 396 | for (gsi = gsi_last_bb (bb); !gsi_end_p (gsi); gsi_prev (&gsi)) |
6de9cd9a | 397 | { |
726a989a | 398 | stmt = gsi_stmt (gsi); |
6de9cd9a DN |
399 | |
400 | /* Ignore labels. */ | |
726a989a | 401 | if (gimple_code (stmt) == GIMPLE_LABEL) |
6de9cd9a DN |
402 | continue; |
403 | ||
6de9cd9a | 404 | /* Check for a call. */ |
726a989a | 405 | if (is_gimple_call (stmt)) |
6de9cd9a | 406 | { |
6de9cd9a | 407 | call = stmt; |
726a989a RB |
408 | ass_var = gimple_call_lhs (stmt); |
409 | break; | |
6de9cd9a DN |
410 | } |
411 | ||
ba2e294d | 412 | /* If the statement has virtual or volatile operands, fail. */ |
f47c96aa | 413 | if (!ZERO_SSA_OPERANDS (stmt, (SSA_OP_VUSE | SSA_OP_VIRTUAL_DEFS)) |
726a989a RB |
414 | || gimple_has_volatile_ops (stmt) |
415 | || (!gimple_aliases_computed_p (cfun) | |
416 | && gimple_references_memory_p (stmt))) | |
6de9cd9a DN |
417 | return; |
418 | } | |
419 | ||
726a989a | 420 | if (gsi_end_p (gsi)) |
6de9cd9a | 421 | { |
628f6a4e | 422 | edge_iterator ei; |
6de9cd9a | 423 | /* Recurse to the predecessors. */ |
628f6a4e | 424 | FOR_EACH_EDGE (e, ei, bb->preds) |
6de9cd9a DN |
425 | find_tail_calls (e->src, ret); |
426 | ||
427 | return; | |
428 | } | |
429 | ||
bd42a56b MM |
430 | /* If the LHS of our call is not just a simple register, we can't |
431 | transform this into a tail or sibling call. This situation happens, | |
432 | in (e.g.) "*p = foo()" where foo returns a struct. In this case | |
433 | we won't have a temporary here, but we need to carry out the side | |
434 | effect anyway, so tailcall is impossible. | |
435 | ||
436 | ??? In some situations (when the struct is returned in memory via | |
437 | invisible argument) we could deal with this, e.g. by passing 'p' | |
438 | itself as that argument to foo, but it's too early to do this here, | |
439 | and expand_call() will not handle it anyway. If it ever can, then | |
440 | we need to revisit this here, to allow that situation. */ | |
441 | if (ass_var && !is_gimple_reg (ass_var)) | |
442 | return; | |
443 | ||
6de9cd9a DN |
444 | /* We found the call, check whether it is suitable. */ |
445 | tail_recursion = false; | |
726a989a | 446 | func = gimple_call_fndecl (call); |
6de9cd9a DN |
447 | if (func == current_function_decl) |
448 | { | |
5039610b | 449 | tree arg; |
726a989a RB |
450 | for (param = DECL_ARGUMENTS (func), idx = 0; |
451 | param && idx < gimple_call_num_args (call); | |
452 | param = TREE_CHAIN (param), idx ++) | |
6de9cd9a | 453 | { |
726a989a | 454 | arg = gimple_call_arg (call, idx); |
14de86fa ZD |
455 | if (param != arg) |
456 | { | |
457 | /* Make sure there are no problems with copying. The parameter | |
6de9cd9a DN |
458 | have a copyable type and the two arguments must have reasonably |
459 | equivalent types. The latter requirement could be relaxed if | |
460 | we emitted a suitable type conversion statement. */ | |
14de86fa | 461 | if (!is_gimple_reg_type (TREE_TYPE (param)) |
f4088621 | 462 | || !useless_type_conversion_p (TREE_TYPE (param), |
726a989a | 463 | TREE_TYPE (arg))) |
14de86fa ZD |
464 | break; |
465 | ||
466 | /* The parameter should be a real operand, so that phi node | |
467 | created for it at the start of the function has the meaning | |
468 | of copying the value. This test implies is_gimple_reg_type | |
469 | from the previous condition, however this one could be | |
470 | relaxed by being more careful with copying the new value | |
726a989a | 471 | of the parameter (emitting appropriate GIMPLE_ASSIGN and |
14de86fa ZD |
472 | updating the virtual operands). */ |
473 | if (!is_gimple_reg (param)) | |
474 | break; | |
475 | } | |
6de9cd9a | 476 | } |
726a989a | 477 | if (idx == gimple_call_num_args (call) && !param) |
6de9cd9a DN |
478 | tail_recursion = true; |
479 | } | |
480 | ||
481 | /* Now check the statements after the call. None of them has virtual | |
482 | operands, so they may only depend on the call through its return | |
483 | value. The return value should also be dependent on each of them, | |
484 | since we are running after dce. */ | |
485 | m = NULL_TREE; | |
486 | a = NULL_TREE; | |
487 | ||
488 | abb = bb; | |
726a989a | 489 | agsi = gsi; |
6de9cd9a DN |
490 | while (1) |
491 | { | |
726a989a | 492 | gsi_next (&agsi); |
6de9cd9a | 493 | |
726a989a | 494 | while (gsi_end_p (agsi)) |
6de9cd9a | 495 | { |
c5cbcccf ZD |
496 | ass_var = propagate_through_phis (ass_var, single_succ_edge (abb)); |
497 | abb = single_succ (abb); | |
726a989a | 498 | agsi = gsi_start_bb (abb); |
6de9cd9a DN |
499 | } |
500 | ||
726a989a | 501 | stmt = gsi_stmt (agsi); |
6de9cd9a | 502 | |
726a989a | 503 | if (gimple_code (stmt) == GIMPLE_LABEL) |
6de9cd9a DN |
504 | continue; |
505 | ||
726a989a | 506 | if (gimple_code (stmt) == GIMPLE_RETURN) |
6de9cd9a DN |
507 | break; |
508 | ||
726a989a | 509 | if (gimple_code (stmt) != GIMPLE_ASSIGN) |
6de9cd9a DN |
510 | return; |
511 | ||
726a989a RB |
512 | /* This is a gimple assign. */ |
513 | if (! process_assignment (stmt, gsi, &m, &a, &ass_var)) | |
6de9cd9a DN |
514 | return; |
515 | } | |
516 | ||
b89e96ac | 517 | /* See if this is a tail call we can handle. */ |
726a989a | 518 | ret_var = gimple_return_retval (stmt); |
6de9cd9a DN |
519 | |
520 | /* We may proceed if there either is no return value, or the return value | |
521 | is identical to the call's return. */ | |
522 | if (ret_var | |
523 | && (ret_var != ass_var)) | |
524 | return; | |
525 | ||
4b5b9613 RH |
526 | /* If this is not a tail recursive call, we cannot handle addends or |
527 | multiplicands. */ | |
528 | if (!tail_recursion && (m || a)) | |
529 | return; | |
530 | ||
e1111e8e | 531 | nw = XNEW (struct tailcall); |
6de9cd9a | 532 | |
726a989a | 533 | nw->call_gsi = gsi; |
6de9cd9a DN |
534 | |
535 | nw->tail_recursion = tail_recursion; | |
536 | ||
537 | nw->mult = m; | |
538 | nw->add = a; | |
539 | ||
540 | nw->next = *ret; | |
541 | *ret = nw; | |
542 | } | |
543 | ||
726a989a | 544 | /* Helper to insert PHI_ARGH to the phi of VAR in the destination of edge E. */ |
6de9cd9a DN |
545 | |
546 | static void | |
726a989a RB |
547 | add_successor_phi_arg (edge e, tree var, tree phi_arg) |
548 | { | |
549 | gimple_stmt_iterator gsi; | |
550 | ||
551 | for (gsi = gsi_start_phis (e->dest); !gsi_end_p (gsi); gsi_next (&gsi)) | |
552 | if (PHI_RESULT (gsi_stmt (gsi)) == var) | |
553 | break; | |
554 | ||
555 | gcc_assert (!gsi_end_p (gsi)); | |
556 | add_phi_arg (gsi_stmt (gsi), phi_arg, e); | |
557 | } | |
558 | ||
559 | /* Creates a GIMPLE statement which computes the operation specified by | |
560 | CODE, OP0 and OP1 to a new variable with name LABEL and inserts the | |
561 | statement in the position specified by GSI and UPDATE. Returns the | |
562 | tree node of the statement's result. */ | |
563 | ||
564 | static tree | |
565 | adjust_return_value_with_ops (enum tree_code code, const char *label, | |
566 | tree op0, tree op1, gimple_stmt_iterator gsi, | |
567 | enum gsi_iterator_update update) | |
6de9cd9a | 568 | { |
726a989a | 569 | |
6de9cd9a | 570 | tree ret_type = TREE_TYPE (DECL_RESULT (current_function_decl)); |
726a989a RB |
571 | tree tmp = create_tmp_var (ret_type, label); |
572 | gimple stmt = gimple_build_assign_with_ops (code, tmp, op0, op1); | |
573 | tree result; | |
574 | ||
575 | add_referenced_var (tmp); | |
576 | result = make_ssa_name (tmp, stmt); | |
577 | gimple_assign_set_lhs (stmt, result); | |
578 | update_stmt (stmt); | |
579 | gsi_insert_before (&gsi, stmt, update); | |
580 | return result; | |
581 | } | |
582 | ||
583 | /* Creates a new GIMPLE statement that adjusts the value of accumulator ACC by | |
584 | the computation specified by CODE and OP1 and insert the statement | |
585 | at the position specified by GSI as a new statement. Returns new SSA name | |
586 | of updated accumulator. */ | |
587 | ||
588 | static tree | |
589 | update_accumulator_with_ops (enum tree_code code, tree acc, tree op1, | |
590 | gimple_stmt_iterator gsi) | |
591 | { | |
592 | gimple stmt = gimple_build_assign_with_ops (code, SSA_NAME_VAR (acc), acc, | |
593 | op1); | |
594 | tree var = make_ssa_name (SSA_NAME_VAR (acc), stmt); | |
595 | gimple_assign_set_lhs (stmt, var); | |
596 | update_stmt (stmt); | |
597 | gsi_insert_after (&gsi, stmt, GSI_NEW_STMT); | |
598 | return var; | |
599 | } | |
600 | ||
601 | /* Adjust the accumulator values according to A and M after GSI, and update | |
602 | the phi nodes on edge BACK. */ | |
603 | ||
604 | static void | |
605 | adjust_accumulator_values (gimple_stmt_iterator gsi, tree m, tree a, edge back) | |
606 | { | |
607 | tree var, a_acc_arg = a_acc, m_acc_arg = m_acc; | |
6de9cd9a DN |
608 | |
609 | if (a) | |
610 | { | |
611 | if (m_acc) | |
612 | { | |
613 | if (integer_onep (a)) | |
614 | var = m_acc; | |
615 | else | |
726a989a RB |
616 | var = adjust_return_value_with_ops (MULT_EXPR, "acc_tmp", m_acc, |
617 | a, gsi, GSI_NEW_STMT); | |
6de9cd9a DN |
618 | } |
619 | else | |
620 | var = a; | |
621 | ||
726a989a | 622 | a_acc_arg = update_accumulator_with_ops (PLUS_EXPR, a_acc, var, gsi); |
6de9cd9a DN |
623 | } |
624 | ||
625 | if (m) | |
726a989a | 626 | m_acc_arg = update_accumulator_with_ops (MULT_EXPR, m_acc, m, gsi); |
6de9cd9a DN |
627 | |
628 | if (a_acc) | |
726a989a | 629 | add_successor_phi_arg (back, a_acc, a_acc_arg); |
6de9cd9a DN |
630 | |
631 | if (m_acc) | |
726a989a | 632 | add_successor_phi_arg (back, m_acc, m_acc_arg); |
6de9cd9a DN |
633 | } |
634 | ||
1ea7e6ad | 635 | /* Adjust value of the return at the end of BB according to M and A |
6de9cd9a DN |
636 | accumulators. */ |
637 | ||
638 | static void | |
639 | adjust_return_value (basic_block bb, tree m, tree a) | |
640 | { | |
726a989a RB |
641 | tree retval; |
642 | gimple ret_stmt = gimple_seq_last_stmt (bb_seq (bb)); | |
643 | gimple_stmt_iterator gsi = gsi_last_bb (bb); | |
6de9cd9a | 644 | |
726a989a | 645 | gcc_assert (gimple_code (ret_stmt) == GIMPLE_RETURN); |
6de9cd9a | 646 | |
726a989a RB |
647 | retval = gimple_return_retval (ret_stmt); |
648 | if (!retval || retval == error_mark_node) | |
6de9cd9a DN |
649 | return; |
650 | ||
6de9cd9a | 651 | if (m) |
726a989a RB |
652 | retval = adjust_return_value_with_ops (MULT_EXPR, "mul_tmp", m_acc, retval, |
653 | gsi, GSI_SAME_STMT); | |
6de9cd9a | 654 | if (a) |
726a989a RB |
655 | retval = adjust_return_value_with_ops (PLUS_EXPR, "acc_tmp", a_acc, retval, |
656 | gsi, GSI_SAME_STMT); | |
657 | gimple_return_set_retval (ret_stmt, retval); | |
f430bae8 | 658 | update_stmt (ret_stmt); |
6de9cd9a DN |
659 | } |
660 | ||
23504559 JH |
661 | /* Subtract COUNT and FREQUENCY from the basic block and it's |
662 | outgoing edge. */ | |
663 | static void | |
664 | decrease_profile (basic_block bb, gcov_type count, int frequency) | |
665 | { | |
666 | edge e; | |
667 | bb->count -= count; | |
668 | if (bb->count < 0) | |
669 | bb->count = 0; | |
670 | bb->frequency -= frequency; | |
671 | if (bb->frequency < 0) | |
672 | bb->frequency = 0; | |
673 | if (!single_succ_p (bb)) | |
674 | { | |
675 | gcc_assert (!EDGE_COUNT (bb->succs)); | |
676 | return; | |
677 | } | |
678 | e = single_succ_edge (bb); | |
679 | e->count -= count; | |
680 | if (e->count < 0) | |
681 | e->count = 0; | |
682 | } | |
683 | ||
016925bc ZD |
684 | /* Returns true if argument PARAM of the tail recursive call needs to be copied |
685 | when the call is eliminated. */ | |
686 | ||
687 | static bool | |
688 | arg_needs_copy_p (tree param) | |
689 | { | |
690 | tree def; | |
691 | ||
692 | if (!is_gimple_reg (param) || !var_ann (param)) | |
693 | return false; | |
694 | ||
695 | /* Parameters that are only defined but never used need not be copied. */ | |
5cd4ec7f | 696 | def = gimple_default_def (cfun, param); |
016925bc ZD |
697 | if (!def) |
698 | return false; | |
699 | ||
700 | return true; | |
701 | } | |
702 | ||
6de9cd9a DN |
703 | /* Eliminates tail call described by T. TMP_VARS is a list of |
704 | temporary variables used to copy the function arguments. */ | |
705 | ||
706 | static void | |
707 | eliminate_tail_call (struct tailcall *t) | |
708 | { | |
726a989a RB |
709 | tree param, rslt; |
710 | gimple stmt, call; | |
5039610b | 711 | tree arg; |
726a989a | 712 | size_t idx; |
6de9cd9a DN |
713 | basic_block bb, first; |
714 | edge e; | |
726a989a RB |
715 | gimple phi; |
716 | gimple_stmt_iterator gsi; | |
717 | gimple orig_stmt; | |
6de9cd9a | 718 | |
726a989a RB |
719 | stmt = orig_stmt = gsi_stmt (t->call_gsi); |
720 | bb = gsi_bb (t->call_gsi); | |
6de9cd9a DN |
721 | |
722 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
723 | { | |
724 | fprintf (dump_file, "Eliminated tail recursion in bb %d : ", | |
725 | bb->index); | |
726a989a | 726 | print_gimple_stmt (dump_file, stmt, 0, TDF_SLIM); |
6de9cd9a DN |
727 | fprintf (dump_file, "\n"); |
728 | } | |
729 | ||
726a989a | 730 | gcc_assert (is_gimple_call (stmt)); |
6de9cd9a | 731 | |
c5cbcccf | 732 | first = single_succ (ENTRY_BLOCK_PTR); |
6de9cd9a | 733 | |
726a989a | 734 | /* Remove the code after call_gsi that will become unreachable. The |
8d3d51b5 ZD |
735 | possibly unreachable code in other blocks is removed later in |
736 | cfg cleanup. */ | |
726a989a RB |
737 | gsi = t->call_gsi; |
738 | gsi_next (&gsi); | |
739 | while (!gsi_end_p (gsi)) | |
8d3d51b5 | 740 | { |
726a989a | 741 | gimple t = gsi_stmt (gsi); |
8d3d51b5 ZD |
742 | /* Do not remove the return statement, so that redirect_edge_and_branch |
743 | sees how the block ends. */ | |
726a989a | 744 | if (gimple_code (t) == GIMPLE_RETURN) |
8d3d51b5 ZD |
745 | break; |
746 | ||
726a989a | 747 | gsi_remove (&gsi, true); |
87527e4b | 748 | release_defs (t); |
8d3d51b5 ZD |
749 | } |
750 | ||
23504559 | 751 | /* Number of executions of function has reduced by the tailcall. */ |
726a989a | 752 | e = single_succ_edge (gsi_bb (t->call_gsi)); |
23504559 JH |
753 | decrease_profile (EXIT_BLOCK_PTR, e->count, EDGE_FREQUENCY (e)); |
754 | decrease_profile (ENTRY_BLOCK_PTR, e->count, EDGE_FREQUENCY (e)); | |
755 | if (e->dest != EXIT_BLOCK_PTR) | |
756 | decrease_profile (e->dest, e->count, EDGE_FREQUENCY (e)); | |
757 | ||
6de9cd9a | 758 | /* Replace the call by a jump to the start of function. */ |
726a989a RB |
759 | e = redirect_edge_and_branch (single_succ_edge (gsi_bb (t->call_gsi)), |
760 | first); | |
1e128c5f | 761 | gcc_assert (e); |
726a989a | 762 | PENDING_STMT (e) = NULL; |
6de9cd9a | 763 | |
016925bc ZD |
764 | /* Add phi node entries for arguments. The ordering of the phi nodes should |
765 | be the same as the ordering of the arguments. */ | |
6de9cd9a | 766 | for (param = DECL_ARGUMENTS (current_function_decl), |
726a989a | 767 | idx = 0, gsi = gsi_start_phis (first); |
6de9cd9a | 768 | param; |
726a989a | 769 | param = TREE_CHAIN (param), idx++) |
6de9cd9a | 770 | { |
016925bc | 771 | if (!arg_needs_copy_p (param)) |
6de9cd9a | 772 | continue; |
726a989a RB |
773 | |
774 | arg = gimple_call_arg (stmt, idx); | |
775 | phi = gsi_stmt (gsi); | |
016925bc | 776 | gcc_assert (param == SSA_NAME_VAR (PHI_RESULT (phi))); |
6de9cd9a | 777 | |
5039610b | 778 | add_phi_arg (phi, arg, e); |
726a989a | 779 | gsi_next (&gsi); |
6de9cd9a DN |
780 | } |
781 | ||
782 | /* Update the values of accumulators. */ | |
726a989a | 783 | adjust_accumulator_values (t->call_gsi, t->mult, t->add, e); |
6de9cd9a | 784 | |
726a989a RB |
785 | call = gsi_stmt (t->call_gsi); |
786 | rslt = gimple_call_lhs (call); | |
787 | if (rslt != NULL_TREE) | |
6de9cd9a | 788 | { |
6de9cd9a DN |
789 | /* Result of the call will no longer be defined. So adjust the |
790 | SSA_NAME_DEF_STMT accordingly. */ | |
726a989a | 791 | SSA_NAME_DEF_STMT (rslt) = gimple_build_nop (); |
6de9cd9a DN |
792 | } |
793 | ||
726a989a | 794 | gsi_remove (&t->call_gsi, true); |
87527e4b | 795 | release_defs (call); |
6de9cd9a DN |
796 | } |
797 | ||
016925bc ZD |
798 | /* Add phi nodes for the virtual operands defined in the function to the |
799 | header of the loop created by tail recursion elimination. | |
800 | ||
801 | Originally, we used to add phi nodes only for call clobbered variables, | |
802 | as the value of the non-call clobbered ones obviously cannot be used | |
803 | or changed within the recursive call. However, the local variables | |
804 | from multiple calls now share the same location, so the virtual ssa form | |
805 | requires us to say that the location dies on further iterations of the loop, | |
806 | which requires adding phi nodes. | |
807 | */ | |
808 | static void | |
809 | add_virtual_phis (void) | |
810 | { | |
811 | referenced_var_iterator rvi; | |
812 | tree var; | |
813 | ||
814 | /* The problematic part is that there is no way how to know what | |
815 | to put into phi nodes (there in fact does not have to be such | |
816 | ssa name available). A solution would be to have an artificial | |
817 | use/kill for all virtual operands in EXIT node. Unless we have | |
818 | this, we cannot do much better than to rebuild the ssa form for | |
819 | possibly affected virtual ssa names from scratch. */ | |
820 | ||
821 | FOR_EACH_REFERENCED_VAR (var, rvi) | |
822 | { | |
5cd4ec7f | 823 | if (!is_gimple_reg (var) && gimple_default_def (cfun, var) != NULL_TREE) |
016925bc ZD |
824 | mark_sym_for_renaming (var); |
825 | } | |
016925bc ZD |
826 | } |
827 | ||
6de9cd9a DN |
828 | /* Optimizes the tailcall described by T. If OPT_TAILCALLS is true, also |
829 | mark the tailcalls for the sibcall optimization. */ | |
830 | ||
831 | static bool | |
832 | optimize_tail_call (struct tailcall *t, bool opt_tailcalls) | |
833 | { | |
834 | if (t->tail_recursion) | |
835 | { | |
836 | eliminate_tail_call (t); | |
837 | return true; | |
838 | } | |
839 | ||
840 | if (opt_tailcalls) | |
841 | { | |
726a989a | 842 | gimple stmt = gsi_stmt (t->call_gsi); |
6de9cd9a | 843 | |
726a989a | 844 | gimple_call_set_tail (stmt, true); |
6de9cd9a DN |
845 | if (dump_file && (dump_flags & TDF_DETAILS)) |
846 | { | |
847 | fprintf (dump_file, "Found tail call "); | |
726a989a RB |
848 | print_gimple_stmt (dump_file, stmt, 0, dump_flags); |
849 | fprintf (dump_file, " in bb %i\n", (gsi_bb (t->call_gsi))->index); | |
6de9cd9a DN |
850 | } |
851 | } | |
852 | ||
853 | return false; | |
854 | } | |
855 | ||
726a989a RB |
856 | /* Creates a tail-call accumulator of the same type as the return type of the |
857 | current function. LABEL is the name used to creating the temporary | |
858 | variable for the accumulator. The accumulator will be inserted in the | |
859 | phis of a basic block BB with single predecessor with an initial value | |
860 | INIT converted to the current function return type. */ | |
861 | ||
862 | static tree | |
863 | create_tailcall_accumulator (const char *label, basic_block bb, tree init) | |
864 | { | |
865 | tree ret_type = TREE_TYPE (DECL_RESULT (current_function_decl)); | |
866 | tree tmp = create_tmp_var (ret_type, label); | |
867 | gimple phi; | |
868 | ||
869 | add_referenced_var (tmp); | |
870 | phi = create_phi_node (tmp, bb); | |
871 | /* RET_TYPE can be a float when -ffast-maths is enabled. */ | |
872 | add_phi_arg (phi, fold_convert (ret_type, init), single_pred_edge (bb)); | |
873 | return PHI_RESULT (phi); | |
874 | } | |
875 | ||
6de9cd9a DN |
876 | /* Optimizes tail calls in the function, turning the tail recursion |
877 | into iteration. */ | |
878 | ||
1994bfea | 879 | static unsigned int |
6de9cd9a DN |
880 | tree_optimize_tail_calls_1 (bool opt_tailcalls) |
881 | { | |
882 | edge e; | |
883 | bool phis_constructed = false; | |
884 | struct tailcall *tailcalls = NULL, *act, *next; | |
885 | bool changed = false; | |
c5cbcccf | 886 | basic_block first = single_succ (ENTRY_BLOCK_PTR); |
726a989a RB |
887 | tree param; |
888 | gimple stmt; | |
628f6a4e | 889 | edge_iterator ei; |
6de9cd9a DN |
890 | |
891 | if (!suitable_for_tail_opt_p ()) | |
1994bfea | 892 | return 0; |
6de9cd9a DN |
893 | if (opt_tailcalls) |
894 | opt_tailcalls = suitable_for_tail_call_opt_p (); | |
895 | ||
628f6a4e | 896 | FOR_EACH_EDGE (e, ei, EXIT_BLOCK_PTR->preds) |
6de9cd9a DN |
897 | { |
898 | /* Only traverse the normal exits, i.e. those that end with return | |
899 | statement. */ | |
900 | stmt = last_stmt (e->src); | |
901 | ||
902 | if (stmt | |
726a989a | 903 | && gimple_code (stmt) == GIMPLE_RETURN) |
6de9cd9a DN |
904 | find_tail_calls (e->src, &tailcalls); |
905 | } | |
906 | ||
907 | /* Construct the phi nodes and accumulators if necessary. */ | |
908 | a_acc = m_acc = NULL_TREE; | |
909 | for (act = tailcalls; act; act = act->next) | |
910 | { | |
911 | if (!act->tail_recursion) | |
912 | continue; | |
913 | ||
914 | if (!phis_constructed) | |
915 | { | |
916 | /* Ensure that there is only one predecessor of the block. */ | |
c5cbcccf ZD |
917 | if (!single_pred_p (first)) |
918 | first = split_edge (single_succ_edge (ENTRY_BLOCK_PTR)); | |
6de9cd9a DN |
919 | |
920 | /* Copy the args if needed. */ | |
921 | for (param = DECL_ARGUMENTS (current_function_decl); | |
922 | param; | |
923 | param = TREE_CHAIN (param)) | |
016925bc ZD |
924 | if (arg_needs_copy_p (param)) |
925 | { | |
5cd4ec7f | 926 | tree name = gimple_default_def (cfun, param); |
016925bc | 927 | tree new_name = make_ssa_name (param, SSA_NAME_DEF_STMT (name)); |
726a989a | 928 | gimple phi; |
016925bc ZD |
929 | |
930 | set_default_def (param, new_name); | |
931 | phi = create_phi_node (name, first); | |
932 | SSA_NAME_DEF_STMT (name) = phi; | |
933 | add_phi_arg (phi, new_name, single_pred_edge (first)); | |
934 | } | |
6de9cd9a DN |
935 | phis_constructed = true; |
936 | } | |
937 | ||
938 | if (act->add && !a_acc) | |
726a989a RB |
939 | a_acc = create_tailcall_accumulator ("add_acc", first, |
940 | integer_zero_node); | |
6de9cd9a DN |
941 | |
942 | if (act->mult && !m_acc) | |
726a989a RB |
943 | m_acc = create_tailcall_accumulator ("mult_acc", first, |
944 | integer_one_node); | |
016925bc ZD |
945 | } |
946 | ||
6de9cd9a DN |
947 | for (; tailcalls; tailcalls = next) |
948 | { | |
949 | next = tailcalls->next; | |
950 | changed |= optimize_tail_call (tailcalls, opt_tailcalls); | |
951 | free (tailcalls); | |
952 | } | |
953 | ||
954 | if (a_acc || m_acc) | |
955 | { | |
956 | /* Modify the remaining return statements. */ | |
628f6a4e | 957 | FOR_EACH_EDGE (e, ei, EXIT_BLOCK_PTR->preds) |
6de9cd9a DN |
958 | { |
959 | stmt = last_stmt (e->src); | |
960 | ||
961 | if (stmt | |
726a989a | 962 | && gimple_code (stmt) == GIMPLE_RETURN) |
6de9cd9a DN |
963 | adjust_return_value (e->src, m_acc, a_acc); |
964 | } | |
965 | } | |
966 | ||
967 | if (changed) | |
1994bfea | 968 | free_dominance_info (CDI_DOMINATORS); |
016925bc ZD |
969 | |
970 | if (phis_constructed) | |
971 | add_virtual_phis (); | |
1994bfea JH |
972 | if (changed) |
973 | return TODO_cleanup_cfg | TODO_update_ssa_only_virtuals; | |
974 | return 0; | |
6de9cd9a DN |
975 | } |
976 | ||
c2924966 | 977 | static unsigned int |
6de9cd9a DN |
978 | execute_tail_recursion (void) |
979 | { | |
1994bfea | 980 | return tree_optimize_tail_calls_1 (false); |
6de9cd9a DN |
981 | } |
982 | ||
983 | static bool | |
984 | gate_tail_calls (void) | |
985 | { | |
6fb5fa3c | 986 | return flag_optimize_sibling_calls != 0 && dbg_cnt (tail_call); |
6de9cd9a DN |
987 | } |
988 | ||
c2924966 | 989 | static unsigned int |
6de9cd9a DN |
990 | execute_tail_calls (void) |
991 | { | |
1994bfea | 992 | return tree_optimize_tail_calls_1 (true); |
6de9cd9a DN |
993 | } |
994 | ||
8ddbbcae | 995 | struct gimple_opt_pass pass_tail_recursion = |
6de9cd9a | 996 | { |
8ddbbcae JH |
997 | { |
998 | GIMPLE_PASS, | |
6de9cd9a | 999 | "tailr", /* name */ |
e8c3ff97 | 1000 | gate_tail_calls, /* gate */ |
6de9cd9a DN |
1001 | execute_tail_recursion, /* execute */ |
1002 | NULL, /* sub */ | |
1003 | NULL, /* next */ | |
1004 | 0, /* static_pass_number */ | |
1005 | 0, /* tv_id */ | |
7faade0f | 1006 | PROP_cfg | PROP_ssa, /* properties_required */ |
6de9cd9a DN |
1007 | 0, /* properties_provided */ |
1008 | 0, /* properties_destroyed */ | |
1009 | 0, /* todo_flags_start */ | |
8ddbbcae JH |
1010 | TODO_dump_func | TODO_verify_ssa /* todo_flags_finish */ |
1011 | } | |
6de9cd9a DN |
1012 | }; |
1013 | ||
8ddbbcae | 1014 | struct gimple_opt_pass pass_tail_calls = |
6de9cd9a | 1015 | { |
8ddbbcae JH |
1016 | { |
1017 | GIMPLE_PASS, | |
6de9cd9a DN |
1018 | "tailc", /* name */ |
1019 | gate_tail_calls, /* gate */ | |
1020 | execute_tail_calls, /* execute */ | |
1021 | NULL, /* sub */ | |
1022 | NULL, /* next */ | |
1023 | 0, /* static_pass_number */ | |
1024 | 0, /* tv_id */ | |
c1b763fa | 1025 | PROP_cfg | PROP_ssa | PROP_alias, /* properties_required */ |
6de9cd9a DN |
1026 | 0, /* properties_provided */ |
1027 | 0, /* properties_destroyed */ | |
1028 | 0, /* todo_flags_start */ | |
8ddbbcae JH |
1029 | TODO_dump_func | TODO_verify_ssa /* todo_flags_finish */ |
1030 | } | |
6de9cd9a | 1031 | }; |