summaryrefslogtreecommitdiff
path: root/gcc/config/arm/arm1136jfs.md
blob: 6ada7fac8ea4e9b1ce4aa2a1eadaa47382ab9045 (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
;; ARM 1136J[F]-S Pipeline Description
;; Copyright (C) 2003, 2007, 2012 Free Software Foundation, Inc.
;; Written by CodeSourcery, LLC.
;;
;; This file is part of GCC.
;;
;; GCC is free software; you can redistribute it and/or modify it
;; under the terms of the GNU General Public License as published by
;; the Free Software Foundation; either version 3, or (at your option)
;; any later version.
;;
;; GCC is distributed in the hope that it will be useful, but
;; WITHOUT ANY WARRANTY; without even the implied warranty of
;; MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
;; General Public License for more details.
;;
;; You should have received a copy of the GNU General Public License
;; along with GCC; see the file COPYING3.  If not see
;; <http://www.gnu.org/licenses/>.  */

;; These descriptions are based on the information contained in the
;; ARM1136JF-S Technical Reference Manual, Copyright (c) 2003 ARM
;; Limited.
;;

;; This automaton provides a pipeline description for the ARM
;; 1136J-S and 1136JF-S cores.
;;
;; The model given here assumes that the condition for all conditional
;; instructions is "true", i.e., that all of the instructions are
;; actually executed.

(define_automaton "arm1136jfs")

;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
;; Pipelines
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

;; There are three distinct pipelines (page 1-26 and following):
;;
;; - A 4-stage decode pipeline, shared by all three.  It has fetch (1),
;;   fetch (2), decode, and issue stages.  Since this is always involved,
;;   we do not model it in the scheduler.
;;
;; - A 4-stage ALU pipeline.  It has shifter, ALU (main integer operations),
;;   and saturation stages.  The fourth stage is writeback; see below.
;;
;; - A 4-stage multiply-accumulate pipeline.  It has three stages, called
;;   MAC1 through MAC3, and a fourth writeback stage.
;;
;;   The 4th-stage writeback is shared between the ALU and MAC pipelines,
;;   which operate in lockstep.  Results from either pipeline will be
;;   moved into the writeback stage.  Because the two pipelines operate
;;   in lockstep, we schedule them as a single "execute" pipeline.
;;
;; - A 4-stage LSU pipeline.  It has address generation, data cache (1),
;;   data cache (2), and writeback stages.  (Note that this pipeline,
;;   including the writeback stage, is independent from the ALU & LSU pipes.)  

(define_cpu_unit "e_1,e_2,e_3,e_wb" "arm1136jfs")     ; ALU and MAC
; e_1 = Sh/Mac1, e_2 = ALU/Mac2, e_3 = SAT/Mac3
(define_cpu_unit "l_a,l_dc1,l_dc2,l_wb" "arm1136jfs") ; Load/Store

;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
;; ALU Instructions
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

;; ALU instructions require eight cycles to execute, and use the ALU
;; pipeline in each of the eight stages.  The results are available
;; after the alu stage has finished.
;;
;; If the destination register is the PC, the pipelines are stalled
;; for several cycles.  That case is not modelled here.

;; ALU operations with no shifted operand
(define_insn_reservation "11_alu_op" 2
 (and (eq_attr "tune" "arm1136js,arm1136jfs")
      (eq_attr "type" "alu_reg,simple_alu_imm"))
 "e_1,e_2,e_3,e_wb")

;; ALU operations with a shift-by-constant operand
(define_insn_reservation "11_alu_shift_op" 2
 (and (eq_attr "tune" "arm1136js,arm1136jfs")
      (eq_attr "type" "simple_alu_shift,alu_shift"))
 "e_1,e_2,e_3,e_wb")

;; ALU operations with a shift-by-register operand
;; These really stall in the decoder, in order to read
;; the shift value in a second cycle. Pretend we take two cycles in
;; the shift stage.
(define_insn_reservation "11_alu_shift_reg_op" 3
 (and (eq_attr "tune" "arm1136js,arm1136jfs")
      (eq_attr "type" "alu_shift_reg"))
 "e_1*2,e_2,e_3,e_wb")

;; alu_ops can start sooner, if there is no shifter dependency
(define_bypass 1 "11_alu_op,11_alu_shift_op"
	       "11_alu_op")
(define_bypass 1 "11_alu_op,11_alu_shift_op"
	       "11_alu_shift_op"
	       "arm_no_early_alu_shift_value_dep")
(define_bypass 1 "11_alu_op,11_alu_shift_op"
	       "11_alu_shift_reg_op"
	       "arm_no_early_alu_shift_dep")
(define_bypass 2 "11_alu_shift_reg_op"
	       "11_alu_op")
(define_bypass 2 "11_alu_shift_reg_op"
	       "11_alu_shift_op"
	       "arm_no_early_alu_shift_value_dep")
(define_bypass 2 "11_alu_shift_reg_op"
	       "11_alu_shift_reg_op"
	       "arm_no_early_alu_shift_dep")

(define_bypass 1 "11_alu_op,11_alu_shift_op"
	       "11_mult1,11_mult2,11_mult3,11_mult4,11_mult5,11_mult6,11_mult7"
	       "arm_no_early_mul_dep")
(define_bypass 2 "11_alu_shift_reg_op"
	       "11_mult1,11_mult2,11_mult3,11_mult4,11_mult5,11_mult6,11_mult7"
	       "arm_no_early_mul_dep")

;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
;; Multiplication Instructions
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

;; Multiplication instructions loop in the first two execute stages until
;; the instruction has been passed through the multiplier array enough
;; times.

;; Multiply and multiply-accumulate results are available after four stages.
(define_insn_reservation "11_mult1" 4
 (and (eq_attr "tune" "arm1136js,arm1136jfs")
      (eq_attr "insn" "mul,mla"))
 "e_1*2,e_2,e_3,e_wb")

;; The *S variants set the condition flags, which requires three more cycles.
(define_insn_reservation "11_mult2" 4
 (and (eq_attr "tune" "arm1136js,arm1136jfs")
      (eq_attr "insn" "muls,mlas"))
 "e_1*2,e_2,e_3,e_wb")

(define_bypass 3 "11_mult1,11_mult2"
	       "11_mult1,11_mult2,11_mult3,11_mult4,11_mult5,11_mult6,11_mult7"
	       "arm_no_early_mul_dep")
(define_bypass 3 "11_mult1,11_mult2"
	       "11_alu_op")
(define_bypass 3 "11_mult1,11_mult2"
	       "11_alu_shift_op"
	       "arm_no_early_alu_shift_value_dep")
(define_bypass 3 "11_mult1,11_mult2"
	       "11_alu_shift_reg_op"
	       "arm_no_early_alu_shift_dep")
(define_bypass 3 "11_mult1,11_mult2"
	       "11_store1"
	       "arm_no_early_store_addr_dep")

;; Signed and unsigned multiply long results are available across two cycles;
;; the less significant word is available one cycle before the more significant
;; word.  Here we conservatively wait until both are available, which is
;; after three iterations and the memory cycle.  The same is also true of
;; the two multiply-accumulate instructions.
(define_insn_reservation "11_mult3" 5
 (and (eq_attr "tune" "arm1136js,arm1136jfs")
      (eq_attr "insn" "smull,umull,smlal,umlal"))
 "e_1*3,e_2,e_3,e_wb*2")

;; The *S variants set the condition flags, which requires three more cycles.
(define_insn_reservation "11_mult4" 5
 (and (eq_attr "tune" "arm1136js,arm1136jfs")
      (eq_attr "insn" "smulls,umulls,smlals,umlals"))
 "e_1*3,e_2,e_3,e_wb*2")

(define_bypass 4 "11_mult3,11_mult4"
	       "11_mult1,11_mult2,11_mult3,11_mult4,11_mult5,11_mult6,11_mult7"
	       "arm_no_early_mul_dep")
(define_bypass 4 "11_mult3,11_mult4"
	       "11_alu_op")
(define_bypass 4 "11_mult3,11_mult4"
	       "11_alu_shift_op"
	       "arm_no_early_alu_shift_value_dep")
(define_bypass 4 "11_mult3,11_mult4"
	       "11_alu_shift_reg_op"
	       "arm_no_early_alu_shift_dep")
(define_bypass 4 "11_mult3,11_mult4"
	       "11_store1"
	       "arm_no_early_store_addr_dep")

;; Various 16x16->32 multiplies and multiply-accumulates, using combinations
;; of high and low halves of the argument registers.  They take a single
;; pass through the pipeline and make the result available after three
;; cycles.
(define_insn_reservation "11_mult5" 3
 (and (eq_attr "tune" "arm1136js,arm1136jfs")
      (eq_attr "insn" "smulxy,smlaxy,smulwy,smlawy,smuad,smuadx,smlad,smladx,smusd,smusdx,smlsd,smlsdx"))
 "e_1,e_2,e_3,e_wb")

(define_bypass 2 "11_mult5"
	       "11_mult1,11_mult2,11_mult3,11_mult4,11_mult5,11_mult6,11_mult7"
	       "arm_no_early_mul_dep")
(define_bypass 2 "11_mult5"
	       "11_alu_op")
(define_bypass 2 "11_mult5"
	       "11_alu_shift_op"
	       "arm_no_early_alu_shift_value_dep")
(define_bypass 2 "11_mult5"
	       "11_alu_shift_reg_op"
	       "arm_no_early_alu_shift_dep")
(define_bypass 2 "11_mult5"
	       "11_store1"
	       "arm_no_early_store_addr_dep")

;; The same idea, then the 32-bit result is added to a 64-bit quantity.
(define_insn_reservation "11_mult6" 4
 (and (eq_attr "tune" "arm1136js,arm1136jfs")
      (eq_attr "insn" "smlalxy"))
 "e_1*2,e_2,e_3,e_wb*2")

;; Signed 32x32 multiply, then the most significant 32 bits are extracted
;; and are available after the memory stage.
(define_insn_reservation "11_mult7" 4
 (and (eq_attr "tune" "arm1136js,arm1136jfs")
      (eq_attr "insn" "smmul,smmulr"))
 "e_1*2,e_2,e_3,e_wb")

(define_bypass 3 "11_mult6,11_mult7"
	       "11_mult1,11_mult2,11_mult3,11_mult4,11_mult5,11_mult6,11_mult7"
	       "arm_no_early_mul_dep")
(define_bypass 3 "11_mult6,11_mult7"
	       "11_alu_op")
(define_bypass 3 "11_mult6,11_mult7"
	       "11_alu_shift_op"
	       "arm_no_early_alu_shift_value_dep")
(define_bypass 3 "11_mult6,11_mult7"
	       "11_alu_shift_reg_op"
	       "arm_no_early_alu_shift_dep")
(define_bypass 3 "11_mult6,11_mult7"
	       "11_store1"
	       "arm_no_early_store_addr_dep")

;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
;; Branch Instructions
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

;; These vary greatly depending on their arguments and the results of
;; stat prediction.  Cycle count ranges from zero (unconditional branch,
;; folded dynamic prediction) to seven (incorrect predictions, etc).  We
;; assume an optimal case for now, because the cost of a cache miss
;; overwhelms the cost of everything else anyhow.

(define_insn_reservation "11_branches" 0
 (and (eq_attr "tune" "arm1136js,arm1136jfs")
      (eq_attr "type" "branch"))
 "nothing")

;; Call latencies are not predictable.  A semi-arbitrary very large
;; number is used as "positive infinity" so that everything should be
;; finished by the time of return.
(define_insn_reservation "11_call" 32
 (and (eq_attr "tune" "arm1136js,arm1136jfs")
      (eq_attr "type" "call"))
 "nothing")

;; Branches are predicted. A correctly predicted branch will be no
;; cost, but we're conservative here, and use the timings a
;; late-register would give us.
(define_bypass 1 "11_alu_op,11_alu_shift_op"
	       "11_branches")
(define_bypass 2 "11_alu_shift_reg_op"
	       "11_branches")
(define_bypass 2 "11_load1,11_load2"
	       "11_branches")
(define_bypass 3 "11_load34"
	       "11_branches")

;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
;; Load/Store Instructions
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

;; The models for load/store instructions do not accurately describe
;; the difference between operations with a base register writeback.
;; These models assume that all memory references hit in dcache.  Also,
;; if the PC is one of the registers involved, there are additional stalls
;; not modelled here.  Addressing modes are also not modelled.

(define_insn_reservation "11_load1" 3
 (and (eq_attr "tune" "arm1136js,arm1136jfs")
      (eq_attr "type" "load1"))
 "l_a+e_1,l_dc1,l_dc2,l_wb")

;; Load byte results are not available until the writeback stage, where
;; the correct byte is extracted.

(define_insn_reservation "11_loadb" 4
 (and (eq_attr "tune" "arm1136js,arm1136jfs")
      (eq_attr "type" "load_byte"))
 "l_a+e_1,l_dc1,l_dc2,l_wb")

(define_insn_reservation "11_store1" 0
 (and (eq_attr "tune" "arm1136js,arm1136jfs")
      (eq_attr "type" "store1"))
 "l_a+e_1,l_dc1,l_dc2,l_wb")

;; Load/store double words into adjacent registers.  The timing and
;; latencies are different depending on whether the address is 64-bit
;; aligned.  This model assumes that it is.
(define_insn_reservation "11_load2" 3
 (and (eq_attr "tune" "arm1136js,arm1136jfs")
      (eq_attr "type" "load2"))
 "l_a+e_1,l_dc1,l_dc2,l_wb")

(define_insn_reservation "11_store2" 0
 (and (eq_attr "tune" "arm1136js,arm1136jfs")
      (eq_attr "type" "store2"))
 "l_a+e_1,l_dc1,l_dc2,l_wb")

;; Load/store multiple registers.  Two registers are stored per cycle.
;; Actual timing depends on how many registers are affected, so we
;; optimistically schedule a low latency.
(define_insn_reservation "11_load34" 4
 (and (eq_attr "tune" "arm1136js,arm1136jfs")
      (eq_attr "type" "load3,load4"))
 "l_a+e_1,l_dc1*2,l_dc2,l_wb")

(define_insn_reservation "11_store34" 0
 (and (eq_attr "tune" "arm1136js,arm1136jfs")
      (eq_attr "type" "store3,store4"))
 "l_a+e_1,l_dc1*2,l_dc2,l_wb")

;; A store can start immediately after an alu op, if that alu op does
;; not provide part of the address to access.
(define_bypass 1 "11_alu_op,11_alu_shift_op"
	       "11_store1"
	       "arm_no_early_store_addr_dep")
(define_bypass 2 "11_alu_shift_reg_op"
	       "11_store1"
	       "arm_no_early_store_addr_dep")

;; An alu op can start sooner after a load, if that alu op does not
;; have an early register dependency on the load
(define_bypass 2 "11_load1"
	       "11_alu_op")
(define_bypass 2 "11_load1"
	       "11_alu_shift_op"
	       "arm_no_early_alu_shift_value_dep")
(define_bypass 2 "11_load1"
	       "11_alu_shift_reg_op"
	       "arm_no_early_alu_shift_dep")

(define_bypass 3 "11_loadb"
	       "11_alu_op")
(define_bypass 3 "11_loadb"
	       "11_alu_shift_op"
	       "arm_no_early_alu_shift_value_dep")
(define_bypass 3 "11_loadb"
	       "11_alu_shift_reg_op"
	       "arm_no_early_alu_shift_dep")

;; A mul op can start sooner after a load, if that mul op does not
;; have an early multiply dependency
(define_bypass 2 "11_load1"
	       "11_mult1,11_mult2,11_mult3,11_mult4,11_mult5,11_mult6,11_mult7"
	       "arm_no_early_mul_dep")
(define_bypass 3 "11_load34"
	       "11_mult1,11_mult2,11_mult3,11_mult4,11_mult5,11_mult6,11_mult7"
	       "arm_no_early_mul_dep")
(define_bypass 3 "11_loadb"
	       "11_mult1,11_mult2,11_mult3,11_mult4,11_mult5,11_mult6,11_mult7"
	       "arm_no_early_mul_dep")

;; A store can start sooner after a load, if that load does not
;; produce part of the address to access
(define_bypass 2 "11_load1"
	       "11_store1"
	       "arm_no_early_store_addr_dep")
(define_bypass 3 "11_loadb"
	       "11_store1"
	       "arm_no_early_store_addr_dep")