1 |
280 |
jeremybenn |
/* Scanning of rtl byte level scanning for dataflow analysis.
|
2 |
|
|
Copyright (C) 2008 Free Software Foundation, Inc.
|
3 |
|
|
Contributed by Kenneth Zadeck (zadeck@naturalbridge.com).
|
4 |
|
|
|
5 |
|
|
This file is part of GCC.
|
6 |
|
|
|
7 |
|
|
GCC is free software; you can redistribute it and/or modify it under
|
8 |
|
|
the terms of the GNU General Public License as published by the Free
|
9 |
|
|
Software Foundation; either version 3, or (at your option) any later
|
10 |
|
|
version.
|
11 |
|
|
|
12 |
|
|
GCC is distributed in the hope that it will be useful, but WITHOUT ANY
|
13 |
|
|
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.
|
16 |
|
|
|
17 |
|
|
You should have received a copy of the GNU General Public License
|
18 |
|
|
along with GCC; see the file COPYING3. If not see
|
19 |
|
|
<http://www.gnu.org/licenses/>. */
|
20 |
|
|
|
21 |
|
|
#include "config.h"
|
22 |
|
|
#include "system.h"
|
23 |
|
|
#include "coretypes.h"
|
24 |
|
|
#include "tm.h"
|
25 |
|
|
#include "rtl.h"
|
26 |
|
|
#include "tm_p.h"
|
27 |
|
|
#include "df.h"
|
28 |
|
|
#include "output.h"
|
29 |
|
|
#include "dbgcnt.h"
|
30 |
|
|
|
31 |
|
|
/* The following suite of functions provides bytewise modeling of REFs
|
32 |
|
|
which are struct df_ref. START_BYTE and LAST_BYTE are returned.
|
33 |
|
|
These can be used as indexes into bitmaps. The indexes are
|
34 |
|
|
normalized so that 0 is the lowest numbered byte, of the inner
|
35 |
|
|
register according to the natural ordering of the machine.
|
36 |
|
|
|
37 |
|
|
This code is designed to be used in backwards scans (which is, of
|
38 |
|
|
course, the way all dataflow scanning should really be done). It
|
39 |
|
|
would require a lot of reworking of the api to make it work in a
|
40 |
|
|
forwards scanning world. */
|
41 |
|
|
|
42 |
|
|
|
43 |
|
|
/* Helper for df_compute_accessed_bytes. Ref is some sort of extract.
|
44 |
|
|
Return true if this effects the entire reg in REF. Return false if
|
45 |
|
|
otherwise and set START_BYTE and LAST_BYTE. See the description of
|
46 |
|
|
df_compute_accessed_bytes for a description of MM. */
|
47 |
|
|
|
48 |
|
|
static bool
|
49 |
|
|
df_compute_accessed_bytes_extract (df_ref ref,
|
50 |
|
|
enum df_mm mm ,
|
51 |
|
|
unsigned int *start_byte,
|
52 |
|
|
unsigned int *last_byte)
|
53 |
|
|
{
|
54 |
|
|
int start;
|
55 |
|
|
int last;
|
56 |
|
|
rtx reg = DF_REF_REG (ref);
|
57 |
|
|
enum machine_mode m1;
|
58 |
|
|
int m1_size;
|
59 |
|
|
enum machine_mode m2;
|
60 |
|
|
int m2_size;
|
61 |
|
|
|
62 |
|
|
/* (*_extract:M1 (reg:M2 X) WIDTH POS)
|
63 |
|
|
(*_extract:M1 (subreg:M1 (reg:M2 X N) WIDTH POS)
|
64 |
|
|
|
65 |
|
|
This is a bitfield extraction. The assignment clobbers/extracts
|
66 |
|
|
exactly the bits named by WIDTH and POS and does not affect the
|
67 |
|
|
other bits in register X. It is also technically possible that
|
68 |
|
|
the bits asked for are longer than units per word. */
|
69 |
|
|
|
70 |
|
|
int offset = DF_REF_EXTRACT_OFFSET (ref);
|
71 |
|
|
int width = DF_REF_EXTRACT_WIDTH (ref);
|
72 |
|
|
|
73 |
|
|
if (width == -1 || offset == -1)
|
74 |
|
|
return true;
|
75 |
|
|
|
76 |
|
|
m1 = DF_REF_EXTRACT_MODE (ref);
|
77 |
|
|
m1_size = GET_MODE_SIZE (m1);
|
78 |
|
|
|
79 |
|
|
gcc_assert (m1_size <= UNITS_PER_WORD);
|
80 |
|
|
|
81 |
|
|
/* There is nothing to do if this is a pure big or small endian
|
82 |
|
|
machine, but if the machine is a pastiche, we have to convert the
|
83 |
|
|
bit offsets into byte offsets. This is only possible because we
|
84 |
|
|
do not care about individual bits because this conversion may
|
85 |
|
|
make the bits non-contiguous. */
|
86 |
|
|
if (BYTES_BIG_ENDIAN != BITS_BIG_ENDIAN)
|
87 |
|
|
offset = GET_MODE_BITSIZE (m1_size) - (offset + width);
|
88 |
|
|
|
89 |
|
|
/* The offset is now in the same order as the subreg_byte. */
|
90 |
|
|
if (GET_CODE (reg) == SUBREG)
|
91 |
|
|
{
|
92 |
|
|
m2 = GET_MODE (SUBREG_REG (reg));
|
93 |
|
|
m2_size = GET_MODE_SIZE (m2);
|
94 |
|
|
if (m1_size > m2_size)
|
95 |
|
|
/* If it is paradoxical, subreg_byte will be zero. */
|
96 |
|
|
offset -= subreg_lowpart_offset (m2, m1) * BITS_PER_UNIT;
|
97 |
|
|
else
|
98 |
|
|
offset += SUBREG_BYTE (reg) * BITS_PER_UNIT;
|
99 |
|
|
}
|
100 |
|
|
else
|
101 |
|
|
{
|
102 |
|
|
m2 = GET_MODE (reg);
|
103 |
|
|
m2_size = GET_MODE_SIZE (m2);
|
104 |
|
|
}
|
105 |
|
|
|
106 |
|
|
if (mm == DF_MM_MUST)
|
107 |
|
|
{
|
108 |
|
|
/* For defs (generally), count the byte only if the whole byte
|
109 |
|
|
is touched. */
|
110 |
|
|
start = (offset + BITS_PER_UNIT - 1) / BITS_PER_UNIT;
|
111 |
|
|
last = (width + offset) / BITS_PER_UNIT;
|
112 |
|
|
|
113 |
|
|
/* In the case where there is nothing, start may be one larger
|
114 |
|
|
than last, we canonize this to return zeros. This keeps
|
115 |
|
|
computations of length from being negative. */
|
116 |
|
|
if (start >= last)
|
117 |
|
|
{
|
118 |
|
|
start = 0;
|
119 |
|
|
last = 0;
|
120 |
|
|
}
|
121 |
|
|
}
|
122 |
|
|
else
|
123 |
|
|
{
|
124 |
|
|
/* For uses (generally), count the byte if any part of the byte
|
125 |
|
|
is touched. */
|
126 |
|
|
start = offset / BITS_PER_UNIT;
|
127 |
|
|
last = (width + offset + BITS_PER_UNIT - 1) / BITS_PER_UNIT;
|
128 |
|
|
}
|
129 |
|
|
|
130 |
|
|
/* Paradoxical truncation. */
|
131 |
|
|
if (start < 0)
|
132 |
|
|
start = 0;
|
133 |
|
|
if (last > m2_size)
|
134 |
|
|
last = m2_size;
|
135 |
|
|
|
136 |
|
|
if (dump_file)
|
137 |
|
|
fprintf (dump_file, " cpb extract regno=%d start=%d last=%d\n",
|
138 |
|
|
DF_REF_REGNO (ref), start, last);
|
139 |
|
|
|
140 |
|
|
*start_byte = start;
|
141 |
|
|
*last_byte = last;
|
142 |
|
|
return false;
|
143 |
|
|
}
|
144 |
|
|
|
145 |
|
|
|
146 |
|
|
/* Helper for df_compute_accessed_bytes. Ref is a strict_low_part.
|
147 |
|
|
Return true if this effects the entire reg in REF. Return false if
|
148 |
|
|
otherwise and set START_BYTE and LAST_BYTE. */
|
149 |
|
|
|
150 |
|
|
static bool
|
151 |
|
|
df_compute_accessed_bytes_strict_low_part (df_ref ref,
|
152 |
|
|
unsigned int *start_byte,
|
153 |
|
|
unsigned int *last_byte)
|
154 |
|
|
{
|
155 |
|
|
int start;
|
156 |
|
|
int last;
|
157 |
|
|
rtx reg = DF_REF_REG (ref);
|
158 |
|
|
enum machine_mode m1;
|
159 |
|
|
int m1_size;
|
160 |
|
|
enum machine_mode m2;
|
161 |
|
|
int m2_size;
|
162 |
|
|
int offset;
|
163 |
|
|
|
164 |
|
|
/* In order to accommodate multiword subregs of a hardreg, df_scan
|
165 |
|
|
eats the subreg and it can only be found from the loc. */
|
166 |
|
|
if (REG_P (reg))
|
167 |
|
|
reg = *(DF_REF_LOC (ref));
|
168 |
|
|
|
169 |
|
|
m1 = GET_MODE (reg);
|
170 |
|
|
m1_size = GET_MODE_SIZE (m1);
|
171 |
|
|
m2 = GET_MODE (SUBREG_REG (reg));
|
172 |
|
|
m2_size = GET_MODE_SIZE (m2);
|
173 |
|
|
offset = SUBREG_BYTE (reg);
|
174 |
|
|
|
175 |
|
|
/* It does not seem to be meaningful to apply a strict_low_part of a
|
176 |
|
|
paradoxical register. */
|
177 |
|
|
gcc_assert (m1_size <= m2_size);
|
178 |
|
|
|
179 |
|
|
/* (set (strict_low_part (subreg:M1 (reg:M2 X) N)) ...)
|
180 |
|
|
|
181 |
|
|
This is a bitfield insertion. The assignment clobbers exactly the
|
182 |
|
|
bits named by the subreg--the M1 bits at position N. It is also
|
183 |
|
|
technically possible that the bits asked for are longer than units
|
184 |
|
|
per word. */
|
185 |
|
|
|
186 |
|
|
start = offset;
|
187 |
|
|
last = offset + m1_size;
|
188 |
|
|
|
189 |
|
|
if (dump_file)
|
190 |
|
|
fprintf (dump_file, " cpb strict low part regno=%d start=%d last=%d\n",
|
191 |
|
|
DF_REF_REGNO (ref), start, last);
|
192 |
|
|
|
193 |
|
|
*start_byte = start;
|
194 |
|
|
*last_byte = last;
|
195 |
|
|
return false;
|
196 |
|
|
}
|
197 |
|
|
|
198 |
|
|
/* Helper for df_compute_accessed_bytes. Ref is a naked subreg.
|
199 |
|
|
Return true if this effects the entire reg in REF. Return false if
|
200 |
|
|
otherwise and set START_BYTE and LAST_BYTE. */
|
201 |
|
|
|
202 |
|
|
static bool
|
203 |
|
|
df_compute_accessed_bytes_subreg (df_ref ref, unsigned int *start_byte,
|
204 |
|
|
unsigned int *last_byte)
|
205 |
|
|
|
206 |
|
|
{
|
207 |
|
|
/* (subreg:M1 (reg:M2 X) N) */
|
208 |
|
|
int start;
|
209 |
|
|
int last;
|
210 |
|
|
rtx reg = DF_REF_REG (ref);
|
211 |
|
|
|
212 |
|
|
enum machine_mode m1;
|
213 |
|
|
int m1_size;
|
214 |
|
|
enum machine_mode m2;
|
215 |
|
|
int m2_size;
|
216 |
|
|
|
217 |
|
|
/* In order to accommodate multiword subregs of a hardreg, df_scan
|
218 |
|
|
eats the subreg and it can only be found from the loc. */
|
219 |
|
|
if (REG_P (reg))
|
220 |
|
|
reg = *(DF_REF_LOC (ref));
|
221 |
|
|
|
222 |
|
|
m1 = GET_MODE (reg);
|
223 |
|
|
m1_size = GET_MODE_SIZE (m1);
|
224 |
|
|
m2 = GET_MODE (SUBREG_REG (reg));
|
225 |
|
|
m2_size = GET_MODE_SIZE (m2);
|
226 |
|
|
|
227 |
|
|
/* A simple paradoxical subreg just accesses the entire inner reg. */
|
228 |
|
|
if (m1_size >= m2_size)
|
229 |
|
|
return true;
|
230 |
|
|
|
231 |
|
|
/* Defs and uses are different in the amount of the reg that touch. */
|
232 |
|
|
if (DF_REF_REG_DEF_P (ref))
|
233 |
|
|
{
|
234 |
|
|
/* This is an lvalue. */
|
235 |
|
|
|
236 |
|
|
if (m2_size > UNITS_PER_WORD)
|
237 |
|
|
{
|
238 |
|
|
/* The assignment clobbers UNITS_PER_WORD segments of X.
|
239 |
|
|
Look at the bytes named by the subreg, and expand it to
|
240 |
|
|
cover a UNITS_PER_WORD part of register X. That part of
|
241 |
|
|
register X is clobbered, the rest is not.
|
242 |
|
|
|
243 |
|
|
E.g., (subreg:SI (reg:DI X) 0), where UNITS_PER_WORD is the
|
244 |
|
|
size of SImode, clobbers the first SImode part of X, and does
|
245 |
|
|
not affect the second SImode part.
|
246 |
|
|
|
247 |
|
|
E.g., (subreg:QI (reg:DI X) 0), where UNITS_PER_WORD is the
|
248 |
|
|
size of SImode, clobbers the first SImode part of X, and does
|
249 |
|
|
not affect the second SImode part. Here the QImode byte is
|
250 |
|
|
expanded to a UNITS_PER_WORD portion of the register for
|
251 |
|
|
purposes of determining what is clobbered.
|
252 |
|
|
|
253 |
|
|
If this is an rvalue, then it touches just the bytes that it
|
254 |
|
|
talks about. */
|
255 |
|
|
int offset = SUBREG_BYTE (reg);
|
256 |
|
|
|
257 |
|
|
start = offset & ~(UNITS_PER_WORD - 1);
|
258 |
|
|
last = (offset + m1_size + UNITS_PER_WORD - 1)
|
259 |
|
|
& ~(UNITS_PER_WORD - 1);
|
260 |
|
|
}
|
261 |
|
|
else
|
262 |
|
|
/* Whole register size M2 equal to or smaller than
|
263 |
|
|
UNITS_PER_WORD The assignment clobbers the entire register
|
264 |
|
|
X. */
|
265 |
|
|
return true;
|
266 |
|
|
}
|
267 |
|
|
else
|
268 |
|
|
{
|
269 |
|
|
/* This is an rvalue. It touches just the bytes they explicitly
|
270 |
|
|
mentioned. */
|
271 |
|
|
int offset = SUBREG_BYTE (reg);
|
272 |
|
|
start = offset;
|
273 |
|
|
last = start + m1_size;
|
274 |
|
|
}
|
275 |
|
|
|
276 |
|
|
if (dump_file)
|
277 |
|
|
fprintf (dump_file, " cpb subreg regno=%d start=%d last=%d\n",
|
278 |
|
|
DF_REF_REGNO (ref), start, last);
|
279 |
|
|
|
280 |
|
|
*start_byte = start;
|
281 |
|
|
*last_byte = last;
|
282 |
|
|
return false;
|
283 |
|
|
}
|
284 |
|
|
|
285 |
|
|
|
286 |
|
|
/* Compute the set of affected bytes by a store to a pseudo to REF.
|
287 |
|
|
MM is either DF_MM_MAY or DF_MM_MUST. This is only relevant for
|
288 |
|
|
the extracts which are not aligned to byte boundaries. The
|
289 |
|
|
DF_MM_MAY returns all of the bytes that any bit is set in and the
|
290 |
|
|
DF_MM_MUST returns only the bytes that are completely covered. In
|
291 |
|
|
general DF_MM_MAY is used for uses and DF_MM_MUST is used for defs,
|
292 |
|
|
but there are exceptions such as the inner loop of the byte level
|
293 |
|
|
dead code eliminator which needs DF_MM_MAY for the defs to see if
|
294 |
|
|
it any possible bit could be used.
|
295 |
|
|
|
296 |
|
|
If the store is to the whole register, just return TRUE, if it is
|
297 |
|
|
to part of the register, return FALSE and set START_BYTE and
|
298 |
|
|
LAST_BYTE properly. In the case where fabricated uses are passed
|
299 |
|
|
in, START_BYTE and LAST_BYTE are set to 0 and false is returned.
|
300 |
|
|
This means that this use can be ignored. */
|
301 |
|
|
|
302 |
|
|
bool
|
303 |
|
|
df_compute_accessed_bytes (df_ref ref, enum df_mm mm,
|
304 |
|
|
unsigned int *start_byte,
|
305 |
|
|
unsigned int *last_byte)
|
306 |
|
|
{
|
307 |
|
|
if (!dbg_cnt (df_byte_scan))
|
308 |
|
|
return true;
|
309 |
|
|
|
310 |
|
|
if (!DF_REF_REG_DEF_P (ref)
|
311 |
|
|
&& DF_REF_FLAGS_IS_SET (ref, DF_REF_READ_WRITE))
|
312 |
|
|
{
|
313 |
|
|
if (DF_REF_FLAGS_IS_SET (ref, DF_REF_PRE_POST_MODIFY))
|
314 |
|
|
/* Pre/post modify/inc/dec always read and write the entire
|
315 |
|
|
reg. */
|
316 |
|
|
return true;
|
317 |
|
|
else
|
318 |
|
|
{
|
319 |
|
|
/* DF_REF_READ_WRITE on a use (except for the
|
320 |
|
|
DF_REF_PRE_POST_MODIFY) means that this use is fabricated
|
321 |
|
|
from a def that is a partial set to a multiword reg.
|
322 |
|
|
Here, we only model those cases precisely so the only one
|
323 |
|
|
to consider is the use put on a auto inc and dec
|
324 |
|
|
insns. */
|
325 |
|
|
*start_byte = 0;
|
326 |
|
|
*last_byte = 0;
|
327 |
|
|
return false;
|
328 |
|
|
}
|
329 |
|
|
}
|
330 |
|
|
|
331 |
|
|
if (DF_REF_FLAGS_IS_SET (ref, DF_REF_SIGN_EXTRACT | DF_REF_ZERO_EXTRACT))
|
332 |
|
|
return df_compute_accessed_bytes_extract (ref, mm, start_byte, last_byte);
|
333 |
|
|
else if (DF_REF_FLAGS_IS_SET (ref, DF_REF_STRICT_LOW_PART))
|
334 |
|
|
return df_compute_accessed_bytes_strict_low_part (ref,
|
335 |
|
|
start_byte, last_byte);
|
336 |
|
|
else if (GET_CODE (DF_REF_REG (ref)) == SUBREG)
|
337 |
|
|
return df_compute_accessed_bytes_subreg (ref, start_byte, last_byte);
|
338 |
|
|
return true;
|
339 |
|
|
}
|
340 |
|
|
|