]>
Commit | Line | Data |
---|---|---|
21ed4444 HPN |
1 | ;; GCC machine description for CRIS atomic memory sequences. |
2 | ;; Copyright (C) 2012 | |
3 | ;; Free Software Foundation, Inc. | |
4 | ;; | |
5 | ;; This file is part of GCC. | |
6 | ;; | |
7 | ;; GCC is free software; you can redistribute it and/or modify | |
8 | ;; it under the terms of the GNU General Public License as published by | |
9 | ;; the Free Software Foundation; either version 3, or (at your option) | |
10 | ;; any later version. | |
11 | ;; | |
12 | ;; GCC is distributed in the hope that it will be useful, | |
13 | ;; but WITHOUT ANY WARRANTY; without even the implied warranty of | |
14 | ;; MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | |
15 | ;; GNU General Public License 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 | ;; The CRIS atomic support yields code in three flavors, depending on | |
22 | ;; the CPU for which code is generated: | |
23 | ;; | |
24 | ;; - Plain old CRIS v0 (..v8) | |
25 | ;; - CRIS v10 (as used in ETRAX 100 LX) | |
26 | ;; - CRIS v32 (as used in ETRAX FS) | |
27 | ;; | |
28 | ;; The last two alternatives are similar, of LL/SC type. They may | |
29 | ;; fail for other reasons; an exception, a cache miss or a bus request | |
30 | ;; from other parts of the system. The difference between them is | |
31 | ;; just in what condition-codes are used to track LL and success or | |
32 | ;; failure for the store. See the chapter on integral read-write | |
33 | ;; operations, chapter 1.13 in "ETRAX 100LX Programmers Manual", | |
34 | ;; <http://www.axis.com/files/tech_notes/etrax_100lx_prog_man-050519.pdf> | |
35 | ;; and chapter 2.1 in "ETRAX FS Designer's reference", | |
36 | ;; <http://www.axis.com/files/manuals/etrax_fs_des_ref-070821.pdf>. | |
37 | ;; Note that the datum being stored has to be contained fully within a | |
38 | ;; cache-line to be integral. A failure to store the data integrally | |
39 | ;; will be flagged, but the store may still have happened in part, | |
40 | ;; which translates most usefully into the data having to be | |
41 | ;; "naturally aligned" to work. Natural alignment is verified in the | |
42 | ;; generated code and will by default cause for unaligned pointers a | |
43 | ;; "break 8" to be executed or optionally a call to abort(). Beware | |
44 | ;; that options -m16bit and -m8bit may cause data to be unaligned | |
45 | ;; where it was otherwise aligned. Data has a better chance of being | |
46 | ;; aligned if it is declared with e.g. __attribute__ ((__align__ (4))). | |
47 | ;; | |
48 | ;; The "plain old v0..v8 flavor" just assumes there's a single CPU in | |
49 | ;; the system, that no other parts of the system have access to memory | |
50 | ;; used for atomic accesses and since there's no user mode without | |
51 | ;; access to interrupt flags (another assumption), it just turns off | |
52 | ;; interrupts while doing the access. Here, alignment is neither | |
53 | ;; required nor asserted. | |
54 | ||
55 | (define_c_enum "" | |
56 | [ | |
57 | CRIS_UNSPEC_ATOMIC_OP | |
58 | CRIS_UNSPEC_ATOMIC_SWAP_MEM | |
59 | CRIS_UNSPEC_ATOMIC_SWAP_BOOL | |
60 | ]) | |
61 | ||
62 | (define_constants [(CRIS_CCR_INTERRUPT_BIT 5)]) | |
63 | ||
64 | ;; We use "mult" as a placeholder for "nand" (which does not have a | |
65 | ;; separate binary rtx operation) so we can use an iterator in the | |
66 | ;; define_expand and define_insn and avoid having a separate | |
67 | ;; mostly-identical copy. You will see the "mult" operator in rtl | |
68 | ;; dumps, but it shouldn't matter as its use has one of its operands | |
69 | ;; inside an unspec_volatile. | |
70 | ||
71 | (define_code_iterator atomic_op [plus minus ior and xor mult]) | |
72 | ||
73 | (define_code_attr atomic_op_name | |
74 | [(plus "add") (minus "sub") (and "and") (ior "or") (xor "xor") (mult "nand")]) | |
75 | ||
76 | ;; Pairs of these are used to insert the "not" after the "and" for nand. | |
77 | (define_code_attr atomic_op_mnem_pre ;; Upper-case only to sinplify testing. | |
78 | [(plus "Add.d") (minus "Sub.d") (and "And.d") (ior "Or.d") (xor "Xor") | |
79 | (mult "aNd.d")]) | |
80 | (define_code_attr atomic_op_mnem_post_op3 | |
81 | [(plus "") (minus "") (and "") (ior "") (xor "") (mult "not %3\;")]) | |
82 | ||
83 | (define_expand "atomic_fetch_<atomic_op_name><mode>" | |
84 | [(match_operand:BWD 0 "register_operand") | |
85 | (match_operand:BWD 1 "memory_operand") | |
86 | (match_operand:BWD 2 "register_operand") | |
87 | (match_operand 3) | |
88 | (atomic_op:BWD (match_dup 0) (match_dup 1))] | |
89 | "" | |
90 | { | |
2f352e3d HPN |
91 | enum memmodel mmodel = (enum memmodel) INTVAL (operands[3]); |
92 | ||
21ed4444 HPN |
93 | if (<MODE>mode != QImode && TARGET_TRAP_UNALIGNED_ATOMIC) |
94 | cris_emit_trap_for_misalignment (operands[1]); | |
95 | ||
fdb6272c HPN |
96 | if (need_atomic_barrier_p (mmodel, true)) |
97 | expand_mem_thread_fence (mmodel); | |
98 | ||
21ed4444 HPN |
99 | emit_insn (gen_cris_atomic_fetch_<atomic_op_name><mode>_1 (operands[0], |
100 | operands[1], | |
101 | operands[2])); | |
fdb6272c HPN |
102 | if (need_atomic_barrier_p (mmodel, false)) |
103 | expand_mem_thread_fence (mmodel); | |
104 | ||
21ed4444 HPN |
105 | DONE; |
106 | }) | |
107 | ||
108 | (define_insn "cris_atomic_fetch_<atomic_op_name><mode>_1" | |
109 | [(set (match_operand:BWD 1 "memory_operand" "+Q") | |
110 | (atomic_op:BWD | |
111 | (unspec_volatile:BWD [(match_dup 1)] CRIS_UNSPEC_ATOMIC_OP) | |
112 | ;; FIXME: relax this for plus, minus, and, ior. | |
113 | (match_operand:BWD 2 "register_operand" "r"))) | |
114 | (set (match_operand:BWD 0 "register_operand" "=&r") | |
115 | (match_dup 1)) | |
116 | (clobber (match_scratch:SI 3 "=&r"))] | |
117 | "" | |
118 | { | |
119 | /* Can't be too sure; better ICE if this happens. */ | |
120 | gcc_assert (!reg_overlap_mentioned_p (operands[2], operands[1])); | |
121 | ||
122 | if (TARGET_V32) | |
123 | return | |
124 | "clearf p\n" | |
125 | ".Lsync.%=:\;" | |
126 | "move<m> %1,%0\;" | |
127 | "move.d %0,%3\;" | |
128 | "<atomic_op_mnem_pre> %2,%3\;<atomic_op_mnem_post_op3>" | |
129 | "ax\;" | |
130 | "move<m> %3,%1\;" | |
131 | "bcs .Lsync.%=\;" | |
132 | "clearf p"; | |
133 | else if (cris_cpu_version == 10) | |
134 | return | |
135 | "clearf\n" | |
136 | ".Lsync.%=:\;" | |
137 | "move<m> %1,%0\;" | |
138 | "move.d %0,%3\;" | |
139 | "<atomic_op_mnem_pre> %2,%3\;<atomic_op_mnem_post_op3>" | |
140 | "ax\;" | |
141 | "move<m> %3,%1\;" | |
142 | "bwf .Lsync.%=\;" | |
143 | "clearf"; | |
144 | else | |
145 | { | |
146 | /* This one is for CRIS versions without load-locked-store-conditional | |
147 | machinery; assume single-core-non-shared-memory without user | |
148 | mode/supervisor mode distinction, and just disable interrupts | |
149 | while performing the operation. | |
150 | Rather than making this pattern more complex by freeing another | |
151 | register or stack position to save condition codes (the value | |
152 | of the interrupt-enabled bit), we check whether interrupts were | |
153 | enabled before we disabled them and branch to a version | |
154 | with/without afterwards re-enabling them. */ | |
155 | rtx ops[5]; | |
156 | ||
157 | /* We have no available macro to stringify CRIS_CCR_INTERRUPT_BIT. */ | |
158 | memcpy (ops, operands, sizeof(ops)); | |
159 | ops[4] = GEN_INT (CRIS_CCR_INTERRUPT_BIT); | |
160 | ||
161 | output_asm_insn ("move $ccr,%3\;" | |
162 | "di\;" | |
163 | "move<m> %1,%0\;" | |
164 | "btstq %4,%3", | |
165 | ops); | |
166 | return | |
167 | "bmi .Lsync.irqon.%=\;" | |
168 | "move.d %0,%3\;" | |
169 | ||
170 | "<atomic_op_mnem_pre> %2,%3\;<atomic_op_mnem_post_op3>" | |
171 | "ba .Lsync.irqoff.%=\;" | |
172 | "move<m> %3,%1\n" | |
173 | ||
174 | ".Lsync.irqon.%=:\;" | |
175 | "<atomic_op_mnem_pre> %2,%3\;<atomic_op_mnem_post_op3>" | |
176 | "move<m> %3,%1\;" | |
177 | "ei\n" | |
178 | ".Lsync.irqoff.%=:"; | |
179 | } | |
180 | }) | |
181 | ||
182 | ;; This pattern is more-or-less assumed to always exist if any of the | |
183 | ;; other atomic patterns exist (see e.g. comment at the | |
184 | ;; can_compare_and_swap_p call in omp-low.c, 4.8 era). We'd slightly | |
185 | ;; prefer atomic_exchange<mode> over this, but having both would be | |
186 | ;; redundant. | |
187 | (define_expand "atomic_compare_and_swap<mode>" | |
188 | [(match_operand:SI 0 "register_operand") | |
189 | (match_operand:BWD 1 "register_operand") | |
190 | (match_operand:BWD 2 "memory_operand") | |
191 | (match_operand:BWD 3 "general_operand") | |
192 | (match_operand:BWD 4 "register_operand") | |
193 | (match_operand 5) | |
194 | (match_operand 6) | |
195 | (match_operand 7)] | |
196 | "" | |
197 | { | |
2f352e3d HPN |
198 | enum memmodel mmodel = (enum memmodel) INTVAL (operands[6]); |
199 | ||
21ed4444 HPN |
200 | if (<MODE>mode != QImode && TARGET_TRAP_UNALIGNED_ATOMIC) |
201 | cris_emit_trap_for_misalignment (operands[2]); | |
202 | ||
fdb6272c HPN |
203 | if (need_atomic_barrier_p (mmodel, true)) |
204 | expand_mem_thread_fence (mmodel); | |
205 | ||
21ed4444 HPN |
206 | emit_insn (gen_cris_atomic_compare_and_swap<mode>_1 (operands[0], |
207 | operands[1], | |
208 | operands[2], | |
209 | operands[3], | |
210 | operands[4])); | |
fdb6272c HPN |
211 | if (need_atomic_barrier_p (mmodel, false)) |
212 | expand_mem_thread_fence (mmodel); | |
213 | ||
21ed4444 HPN |
214 | DONE; |
215 | }) | |
216 | ||
217 | (define_insn "cris_atomic_compare_and_swap<mode>_1" | |
218 | [(set (match_operand:SI 0 "register_operand" "=&r") | |
219 | (unspec_volatile:SI | |
220 | [(match_operand:BWD 2 "memory_operand" "+Q") | |
221 | (match_operand:BWD 3 "general_operand" "g")] | |
222 | CRIS_UNSPEC_ATOMIC_SWAP_BOOL)) | |
223 | (set (match_operand:BWD 1 "register_operand" "=&r") (match_dup 2)) | |
224 | (set (match_dup 2) | |
225 | (unspec_volatile:BWD | |
226 | [(match_dup 2) | |
227 | (match_dup 3) | |
228 | (match_operand:BWD 4 "register_operand" "r")] | |
229 | CRIS_UNSPEC_ATOMIC_SWAP_MEM))] | |
230 | "" | |
231 | { | |
232 | if (TARGET_V32) | |
233 | return | |
234 | "clearf p\n" | |
235 | ".Lsync.repeat.%=:\;" | |
236 | "move<m> %2,%1\;" | |
237 | "cmp<m> %3,%1\;" | |
238 | "bne .Lsync.after.%=\;" | |
239 | "seq %0\;" | |
240 | ||
241 | "ax\;" | |
242 | "move<m> %4,%2\;" | |
243 | "bcs .Lsync.repeat.%=\;" | |
244 | "clearf p\n" | |
245 | ".Lsync.after.%=:"; | |
246 | else if (cris_cpu_version == 10) | |
247 | return | |
248 | "clearf\n" | |
249 | ".Lsync.repeat.%=:\;" | |
250 | "move<m> %2,%1\;" | |
251 | "cmp<m> %3,%1\;" | |
252 | "bne .Lsync.after.%=\;" | |
253 | "seq %0\;" | |
254 | ||
255 | "ax\;" | |
256 | "move<m> %4,%2\;" | |
257 | "bwf .Lsync.repeat.%=\;" | |
258 | "clearf\n" | |
259 | ".Lsync.after.%=:"; | |
260 | else | |
261 | { | |
262 | /* This one is for CRIS versions without load-locked-store-conditional | |
263 | machinery; assume single-core-non-shared-memory without user | |
264 | mode/supervisor mode distinction, and just disable interrupts | |
265 | while performing the operation. | |
266 | Rather than making this pattern more complex by freeing another | |
267 | register or stack position to save condition codes (the value | |
268 | of the interrupt-enabled bit), we check whether interrupts were | |
269 | enabled before we disabled them and branch to a version | |
270 | with/without afterwards re-enabling them. */ | |
271 | rtx ops[4]; | |
272 | ||
273 | /* We have no available macro to stringify CRIS_CCR_INTERRUPT_BIT. */ | |
274 | memcpy (ops, operands, sizeof(ops)); | |
275 | ops[3] = GEN_INT (CRIS_CCR_INTERRUPT_BIT); | |
276 | ||
277 | output_asm_insn ("move $ccr,%0\;" | |
278 | "di\;" | |
279 | "move<m> %2,%1\;" | |
280 | "btstq %3,%0", | |
281 | ops); | |
282 | return | |
283 | "bmi .Lsync.irqon.%=\;" | |
284 | "nop\;" | |
285 | ||
286 | "cmp<m> %3,%1\;" | |
287 | "bne .Lsync.after.%=\;" | |
288 | "seq %0\;" | |
289 | "ba .Lsync.after.%=\;" | |
290 | "move<m> %4,%2\n" | |
291 | ||
292 | ".Lsync.irqon.%=:\;" | |
293 | "cmp<m> %3,%1\;" | |
294 | "bne .Lsync.after.%=\;" | |
295 | "seq %0\;" | |
296 | "move<m> %4,%2\;" | |
297 | "ei\n" | |
298 | ".Lsync.after.%=:"; | |
299 | } | |
300 | }) |