1 |
280 |
jeremybenn |
/* Language independent return value optimizations
|
2 |
|
|
Copyright (C) 2004, 2005, 2007, 2008, 2009 Free Software Foundation, Inc.
|
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
|
8 |
|
|
the Free Software Foundation; either version 3, or (at your option)
|
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
|
17 |
|
|
along with GCC; see the file COPYING3. If not see
|
18 |
|
|
<http://www.gnu.org/licenses/>. */
|
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 "function.h"
|
27 |
|
|
#include "basic-block.h"
|
28 |
|
|
#include "expr.h"
|
29 |
|
|
#include "diagnostic.h"
|
30 |
|
|
#include "tree-flow.h"
|
31 |
|
|
#include "timevar.h"
|
32 |
|
|
#include "tree-dump.h"
|
33 |
|
|
#include "tree-pass.h"
|
34 |
|
|
#include "langhooks.h"
|
35 |
|
|
|
36 |
|
|
/* This file implements return value optimizations for functions which
|
37 |
|
|
return aggregate types.
|
38 |
|
|
|
39 |
|
|
Basically this pass searches the function for return statements which
|
40 |
|
|
return a local aggregate. When converted to RTL such statements will
|
41 |
|
|
generate a copy from the local aggregate to final return value destination
|
42 |
|
|
mandated by the target's ABI.
|
43 |
|
|
|
44 |
|
|
That copy can often be avoided by directly constructing the return value
|
45 |
|
|
into the final destination mandated by the target's ABI.
|
46 |
|
|
|
47 |
|
|
This is basically a generic equivalent to the C++ front-end's
|
48 |
|
|
Named Return Value optimization. */
|
49 |
|
|
|
50 |
|
|
struct nrv_data
|
51 |
|
|
{
|
52 |
|
|
/* This is the temporary (a VAR_DECL) which appears in all of
|
53 |
|
|
this function's RETURN_EXPR statements. */
|
54 |
|
|
tree var;
|
55 |
|
|
|
56 |
|
|
/* This is the function's RESULT_DECL. We will replace all occurrences
|
57 |
|
|
of VAR with RESULT_DECL when we apply this optimization. */
|
58 |
|
|
tree result;
|
59 |
|
|
int modified;
|
60 |
|
|
};
|
61 |
|
|
|
62 |
|
|
static tree finalize_nrv_r (tree *, int *, void *);
|
63 |
|
|
|
64 |
|
|
/* Callback for the tree walker.
|
65 |
|
|
|
66 |
|
|
If TP refers to a RETURN_EXPR, then set the expression being returned
|
67 |
|
|
to nrv_data->result.
|
68 |
|
|
|
69 |
|
|
If TP refers to nrv_data->var, then replace nrv_data->var with
|
70 |
|
|
nrv_data->result.
|
71 |
|
|
|
72 |
|
|
If we reach a node where we know all the subtrees are uninteresting,
|
73 |
|
|
then set *WALK_SUBTREES to zero. */
|
74 |
|
|
|
75 |
|
|
static tree
|
76 |
|
|
finalize_nrv_r (tree *tp, int *walk_subtrees, void *data)
|
77 |
|
|
{
|
78 |
|
|
struct walk_stmt_info *wi = (struct walk_stmt_info *) data;
|
79 |
|
|
struct nrv_data *dp = (struct nrv_data *) wi->info;
|
80 |
|
|
|
81 |
|
|
/* No need to walk into types. */
|
82 |
|
|
if (TYPE_P (*tp))
|
83 |
|
|
*walk_subtrees = 0;
|
84 |
|
|
|
85 |
|
|
/* Otherwise replace all occurrences of VAR with RESULT. */
|
86 |
|
|
else if (*tp == dp->var)
|
87 |
|
|
{
|
88 |
|
|
*tp = dp->result;
|
89 |
|
|
dp->modified = 1;
|
90 |
|
|
}
|
91 |
|
|
|
92 |
|
|
/* Keep iterating. */
|
93 |
|
|
return NULL_TREE;
|
94 |
|
|
}
|
95 |
|
|
|
96 |
|
|
/* Main entry point for return value optimizations.
|
97 |
|
|
|
98 |
|
|
If this function always returns the same local variable, and that
|
99 |
|
|
local variable is an aggregate type, then replace the variable with
|
100 |
|
|
the function's DECL_RESULT.
|
101 |
|
|
|
102 |
|
|
This is the equivalent of the C++ named return value optimization
|
103 |
|
|
applied to optimized trees in a language independent form. If we
|
104 |
|
|
ever encounter languages which prevent this kind of optimization,
|
105 |
|
|
then we could either have the languages register the optimization or
|
106 |
|
|
we could change the gating function to check the current language. */
|
107 |
|
|
|
108 |
|
|
static unsigned int
|
109 |
|
|
tree_nrv (void)
|
110 |
|
|
{
|
111 |
|
|
tree result = DECL_RESULT (current_function_decl);
|
112 |
|
|
tree result_type = TREE_TYPE (result);
|
113 |
|
|
tree found = NULL;
|
114 |
|
|
basic_block bb;
|
115 |
|
|
gimple_stmt_iterator gsi;
|
116 |
|
|
struct nrv_data data;
|
117 |
|
|
|
118 |
|
|
/* If this function does not return an aggregate type in memory, then
|
119 |
|
|
there is nothing to do. */
|
120 |
|
|
if (!aggregate_value_p (result, current_function_decl))
|
121 |
|
|
return 0;
|
122 |
|
|
|
123 |
|
|
/* If a GIMPLE type is returned in memory, finalize_nrv_r might create
|
124 |
|
|
non-GIMPLE. */
|
125 |
|
|
if (is_gimple_reg_type (result_type))
|
126 |
|
|
return 0;
|
127 |
|
|
|
128 |
|
|
/* If the front end already did something like this, don't do it here. */
|
129 |
|
|
if (DECL_NAME (result))
|
130 |
|
|
return 0;
|
131 |
|
|
|
132 |
|
|
/* If the result has its address taken then it might be modified
|
133 |
|
|
by means not detected in the following loop. Bail out in this
|
134 |
|
|
case. */
|
135 |
|
|
if (TREE_ADDRESSABLE (result))
|
136 |
|
|
return 0;
|
137 |
|
|
|
138 |
|
|
/* Look through each block for assignments to the RESULT_DECL. */
|
139 |
|
|
FOR_EACH_BB (bb)
|
140 |
|
|
{
|
141 |
|
|
for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
|
142 |
|
|
{
|
143 |
|
|
gimple stmt = gsi_stmt (gsi);
|
144 |
|
|
tree ret_val;
|
145 |
|
|
|
146 |
|
|
if (gimple_code (stmt) == GIMPLE_RETURN)
|
147 |
|
|
{
|
148 |
|
|
/* In a function with an aggregate return value, the
|
149 |
|
|
gimplifier has changed all non-empty RETURN_EXPRs to
|
150 |
|
|
return the RESULT_DECL. */
|
151 |
|
|
ret_val = gimple_return_retval (stmt);
|
152 |
|
|
if (ret_val)
|
153 |
|
|
gcc_assert (ret_val == result);
|
154 |
|
|
}
|
155 |
|
|
else if (gimple_has_lhs (stmt)
|
156 |
|
|
&& gimple_get_lhs (stmt) == result)
|
157 |
|
|
{
|
158 |
|
|
tree rhs;
|
159 |
|
|
|
160 |
|
|
if (!gimple_assign_copy_p (stmt))
|
161 |
|
|
return 0;
|
162 |
|
|
|
163 |
|
|
rhs = gimple_assign_rhs1 (stmt);
|
164 |
|
|
|
165 |
|
|
/* Now verify that this return statement uses the same value
|
166 |
|
|
as any previously encountered return statement. */
|
167 |
|
|
if (found != NULL)
|
168 |
|
|
{
|
169 |
|
|
/* If we found a return statement using a different variable
|
170 |
|
|
than previous return statements, then we can not perform
|
171 |
|
|
NRV optimizations. */
|
172 |
|
|
if (found != rhs)
|
173 |
|
|
return 0;
|
174 |
|
|
}
|
175 |
|
|
else
|
176 |
|
|
found = rhs;
|
177 |
|
|
|
178 |
|
|
/* The returned value must be a local automatic variable of the
|
179 |
|
|
same type and alignment as the function's result. */
|
180 |
|
|
if (TREE_CODE (found) != VAR_DECL
|
181 |
|
|
|| TREE_THIS_VOLATILE (found)
|
182 |
|
|
|| DECL_CONTEXT (found) != current_function_decl
|
183 |
|
|
|| TREE_STATIC (found)
|
184 |
|
|
|| TREE_ADDRESSABLE (found)
|
185 |
|
|
|| DECL_ALIGN (found) > DECL_ALIGN (result)
|
186 |
|
|
|| !useless_type_conversion_p (result_type,
|
187 |
|
|
TREE_TYPE (found)))
|
188 |
|
|
return 0;
|
189 |
|
|
}
|
190 |
|
|
else if (gimple_has_lhs (stmt))
|
191 |
|
|
{
|
192 |
|
|
tree addr = get_base_address (gimple_get_lhs (stmt));
|
193 |
|
|
/* If there's any MODIFY of component of RESULT,
|
194 |
|
|
then bail out. */
|
195 |
|
|
if (addr && addr == result)
|
196 |
|
|
return 0;
|
197 |
|
|
}
|
198 |
|
|
}
|
199 |
|
|
}
|
200 |
|
|
|
201 |
|
|
if (!found)
|
202 |
|
|
return 0;
|
203 |
|
|
|
204 |
|
|
/* If dumping details, then note once and only the NRV replacement. */
|
205 |
|
|
if (dump_file && (dump_flags & TDF_DETAILS))
|
206 |
|
|
{
|
207 |
|
|
fprintf (dump_file, "NRV Replaced: ");
|
208 |
|
|
print_generic_expr (dump_file, found, dump_flags);
|
209 |
|
|
fprintf (dump_file, " with: ");
|
210 |
|
|
print_generic_expr (dump_file, result, dump_flags);
|
211 |
|
|
fprintf (dump_file, "\n");
|
212 |
|
|
}
|
213 |
|
|
|
214 |
|
|
/* At this point we know that all the return statements return the
|
215 |
|
|
same local which has suitable attributes for NRV. Copy debugging
|
216 |
|
|
information from FOUND to RESULT if it will be useful. But don't set
|
217 |
|
|
DECL_ABSTRACT_ORIGIN to point at another function. */
|
218 |
|
|
if (!DECL_IGNORED_P (found)
|
219 |
|
|
&& !(DECL_ABSTRACT_ORIGIN (found)
|
220 |
|
|
&& DECL_CONTEXT (DECL_ABSTRACT_ORIGIN (found)) != current_function_decl))
|
221 |
|
|
{
|
222 |
|
|
DECL_NAME (result) = DECL_NAME (found);
|
223 |
|
|
DECL_SOURCE_LOCATION (result) = DECL_SOURCE_LOCATION (found);
|
224 |
|
|
DECL_ABSTRACT_ORIGIN (result) = DECL_ABSTRACT_ORIGIN (found);
|
225 |
|
|
}
|
226 |
|
|
|
227 |
|
|
TREE_ADDRESSABLE (result) |= TREE_ADDRESSABLE (found);
|
228 |
|
|
|
229 |
|
|
/* Now walk through the function changing all references to VAR to be
|
230 |
|
|
RESULT. */
|
231 |
|
|
data.var = found;
|
232 |
|
|
data.result = result;
|
233 |
|
|
FOR_EACH_BB (bb)
|
234 |
|
|
{
|
235 |
|
|
for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); )
|
236 |
|
|
{
|
237 |
|
|
gimple stmt = gsi_stmt (gsi);
|
238 |
|
|
/* If this is a copy from VAR to RESULT, remove it. */
|
239 |
|
|
if (gimple_assign_copy_p (stmt)
|
240 |
|
|
&& gimple_assign_lhs (stmt) == result
|
241 |
|
|
&& gimple_assign_rhs1 (stmt) == found)
|
242 |
|
|
{
|
243 |
|
|
unlink_stmt_vdef (stmt);
|
244 |
|
|
gsi_remove (&gsi, true);
|
245 |
|
|
}
|
246 |
|
|
else
|
247 |
|
|
{
|
248 |
|
|
struct walk_stmt_info wi;
|
249 |
|
|
memset (&wi, 0, sizeof (wi));
|
250 |
|
|
wi.info = &data;
|
251 |
|
|
data.modified = 0;
|
252 |
|
|
walk_gimple_op (stmt, finalize_nrv_r, &wi);
|
253 |
|
|
if (data.modified)
|
254 |
|
|
update_stmt (stmt);
|
255 |
|
|
gsi_next (&gsi);
|
256 |
|
|
}
|
257 |
|
|
}
|
258 |
|
|
}
|
259 |
|
|
|
260 |
|
|
/* FOUND is no longer used. Ensure it gets removed. */
|
261 |
|
|
var_ann (found)->used = 0;
|
262 |
|
|
return 0;
|
263 |
|
|
}
|
264 |
|
|
|
265 |
|
|
static bool
|
266 |
|
|
gate_pass_return_slot (void)
|
267 |
|
|
{
|
268 |
|
|
return optimize > 0;
|
269 |
|
|
}
|
270 |
|
|
|
271 |
|
|
struct gimple_opt_pass pass_nrv =
|
272 |
|
|
{
|
273 |
|
|
{
|
274 |
|
|
GIMPLE_PASS,
|
275 |
|
|
"nrv", /* name */
|
276 |
|
|
gate_pass_return_slot, /* gate */
|
277 |
|
|
tree_nrv, /* execute */
|
278 |
|
|
NULL, /* sub */
|
279 |
|
|
NULL, /* next */
|
280 |
|
|
0, /* static_pass_number */
|
281 |
|
|
TV_TREE_NRV, /* tv_id */
|
282 |
|
|
PROP_ssa | PROP_cfg, /* properties_required */
|
283 |
|
|
0, /* properties_provided */
|
284 |
|
|
0, /* properties_destroyed */
|
285 |
|
|
0, /* todo_flags_start */
|
286 |
|
|
TODO_dump_func | TODO_ggc_collect /* todo_flags_finish */
|
287 |
|
|
}
|
288 |
|
|
};
|
289 |
|
|
|
290 |
|
|
/* Determine (pessimistically) whether DEST is available for NRV
|
291 |
|
|
optimization, where DEST is expected to be the LHS of a modify
|
292 |
|
|
expression where the RHS is a function returning an aggregate.
|
293 |
|
|
|
294 |
|
|
We search for a base VAR_DECL and look to see if it is call clobbered.
|
295 |
|
|
Note that we could do better, for example, by
|
296 |
|
|
attempting to doing points-to analysis on INDIRECT_REFs. */
|
297 |
|
|
|
298 |
|
|
static bool
|
299 |
|
|
dest_safe_for_nrv_p (tree dest)
|
300 |
|
|
{
|
301 |
|
|
while (handled_component_p (dest))
|
302 |
|
|
dest = TREE_OPERAND (dest, 0);
|
303 |
|
|
|
304 |
|
|
if (! SSA_VAR_P (dest))
|
305 |
|
|
return false;
|
306 |
|
|
|
307 |
|
|
if (TREE_CODE (dest) == SSA_NAME)
|
308 |
|
|
dest = SSA_NAME_VAR (dest);
|
309 |
|
|
|
310 |
|
|
if (is_call_used (dest))
|
311 |
|
|
return false;
|
312 |
|
|
|
313 |
|
|
return true;
|
314 |
|
|
}
|
315 |
|
|
|
316 |
|
|
/* Walk through the function looking for GIMPLE_ASSIGNs with calls that
|
317 |
|
|
return in memory on the RHS. For each of these, determine whether it is
|
318 |
|
|
safe to pass the address of the LHS as the return slot, and mark the
|
319 |
|
|
call appropriately if so.
|
320 |
|
|
|
321 |
|
|
The NRV shares the return slot with a local variable in the callee; this
|
322 |
|
|
optimization shares the return slot with the target of the call within
|
323 |
|
|
the caller. If the NRV is performed (which we can't know in general),
|
324 |
|
|
this optimization is safe if the address of the target has not
|
325 |
|
|
escaped prior to the call. If it has, modifications to the local
|
326 |
|
|
variable will produce visible changes elsewhere, as in PR c++/19317. */
|
327 |
|
|
|
328 |
|
|
static unsigned int
|
329 |
|
|
execute_return_slot_opt (void)
|
330 |
|
|
{
|
331 |
|
|
basic_block bb;
|
332 |
|
|
|
333 |
|
|
FOR_EACH_BB (bb)
|
334 |
|
|
{
|
335 |
|
|
gimple_stmt_iterator gsi;
|
336 |
|
|
for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
|
337 |
|
|
{
|
338 |
|
|
gimple stmt = gsi_stmt (gsi);
|
339 |
|
|
bool slot_opt_p;
|
340 |
|
|
|
341 |
|
|
if (is_gimple_call (stmt)
|
342 |
|
|
&& gimple_call_lhs (stmt)
|
343 |
|
|
&& !gimple_call_return_slot_opt_p (stmt)
|
344 |
|
|
&& aggregate_value_p (TREE_TYPE (gimple_call_lhs (stmt)),
|
345 |
|
|
gimple_call_fndecl (stmt))
|
346 |
|
|
)
|
347 |
|
|
{
|
348 |
|
|
/* Check if the location being assigned to is
|
349 |
|
|
call-clobbered. */
|
350 |
|
|
slot_opt_p = dest_safe_for_nrv_p (gimple_call_lhs (stmt));
|
351 |
|
|
gimple_call_set_return_slot_opt (stmt, slot_opt_p);
|
352 |
|
|
}
|
353 |
|
|
}
|
354 |
|
|
}
|
355 |
|
|
return 0;
|
356 |
|
|
}
|
357 |
|
|
|
358 |
|
|
struct gimple_opt_pass pass_return_slot =
|
359 |
|
|
{
|
360 |
|
|
{
|
361 |
|
|
GIMPLE_PASS,
|
362 |
|
|
"retslot", /* name */
|
363 |
|
|
NULL, /* gate */
|
364 |
|
|
execute_return_slot_opt, /* execute */
|
365 |
|
|
NULL, /* sub */
|
366 |
|
|
NULL, /* next */
|
367 |
|
|
0, /* static_pass_number */
|
368 |
|
|
TV_NONE, /* tv_id */
|
369 |
|
|
PROP_ssa, /* properties_required */
|
370 |
|
|
0, /* properties_provided */
|
371 |
|
|
0, /* properties_destroyed */
|
372 |
|
|
0, /* todo_flags_start */
|
373 |
|
|
|
374 |
|
|
}
|
375 |
|
|
};
|