1 |
282 |
jeremybenn |
;; ARM Cortex-R4 scheduling description.
|
2 |
|
|
;; Copyright (C) 2007, 2008 Free Software Foundation, Inc.
|
3 |
|
|
;; Contributed by CodeSourcery.
|
4 |
|
|
|
5 |
|
|
;; This file is part of GCC.
|
6 |
|
|
|
7 |
|
|
;; GCC is free software; you can redistribute it and/or modify it
|
8 |
|
|
;; under the terms of the GNU General Public License as published
|
9 |
|
|
;; by the Free Software Foundation; either version 3, or (at your
|
10 |
|
|
;; option) any later version.
|
11 |
|
|
|
12 |
|
|
;; GCC is distributed in the hope that it will be useful, but WITHOUT
|
13 |
|
|
;; ANY WARRANTY; without even the implied warranty of MERCHANTABILITY
|
14 |
|
|
;; or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public
|
15 |
|
|
;; 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 |
|
|
;; .
|
20 |
|
|
|
21 |
|
|
(define_automaton "cortex_r4")
|
22 |
|
|
|
23 |
|
|
;; We approximate the dual-issue constraints of this core using four
|
24 |
|
|
;; "issue units" and a reservation matrix as follows. The numbers indicate
|
25 |
|
|
;; the instruction groups' preferences in order. Multiple entries for
|
26 |
|
|
;; the same numbered preference indicate units that must be reserved
|
27 |
|
|
;; together.
|
28 |
|
|
;;
|
29 |
|
|
;; Issue unit: A B C ALU
|
30 |
|
|
;;
|
31 |
|
|
;; ALU w/o reg shift 1st 2nd 1st and 2nd
|
32 |
|
|
;; ALU w/ reg shift 1st 2nd 2nd 1st and 2nd
|
33 |
|
|
;; Moves 1st 2nd 2nd
|
34 |
|
|
;; Multiplication 1st 1st
|
35 |
|
|
;; Division 1st 1st
|
36 |
|
|
;; Load/store single 1st 1st
|
37 |
|
|
;; Other load/store 1st 1st
|
38 |
|
|
;; Branches 1st
|
39 |
|
|
|
40 |
|
|
(define_cpu_unit "cortex_r4_issue_a" "cortex_r4")
|
41 |
|
|
(define_cpu_unit "cortex_r4_issue_b" "cortex_r4")
|
42 |
|
|
(define_cpu_unit "cortex_r4_issue_c" "cortex_r4")
|
43 |
|
|
(define_cpu_unit "cortex_r4_issue_alu" "cortex_r4")
|
44 |
|
|
|
45 |
|
|
(define_reservation "cortex_r4_alu"
|
46 |
|
|
"(cortex_r4_issue_a+cortex_r4_issue_alu)|\
|
47 |
|
|
(cortex_r4_issue_b+cortex_r4_issue_alu)")
|
48 |
|
|
(define_reservation "cortex_r4_alu_shift_reg"
|
49 |
|
|
"(cortex_r4_issue_a+cortex_r4_issue_alu)|\
|
50 |
|
|
(cortex_r4_issue_b+cortex_r4_issue_c+\
|
51 |
|
|
cortex_r4_issue_alu)")
|
52 |
|
|
(define_reservation "cortex_r4_mov"
|
53 |
|
|
"cortex_r4_issue_a|(cortex_r4_issue_b+\
|
54 |
|
|
cortex_r4_issue_alu)")
|
55 |
|
|
(define_reservation "cortex_r4_mul" "cortex_r4_issue_a+cortex_r4_issue_alu")
|
56 |
|
|
(define_reservation "cortex_r4_mul_2"
|
57 |
|
|
"(cortex_r4_issue_a+cortex_r4_issue_alu)*2")
|
58 |
|
|
;; Division instructions execute out-of-order with respect to the
|
59 |
|
|
;; rest of the pipeline and only require reservations on their first and
|
60 |
|
|
;; final cycles.
|
61 |
|
|
(define_reservation "cortex_r4_div_9"
|
62 |
|
|
"cortex_r4_issue_a+cortex_r4_issue_alu,\
|
63 |
|
|
nothing*7,\
|
64 |
|
|
cortex_r4_issue_a+cortex_r4_issue_alu")
|
65 |
|
|
(define_reservation "cortex_r4_div_10"
|
66 |
|
|
"cortex_r4_issue_a+cortex_r4_issue_alu,\
|
67 |
|
|
nothing*8,\
|
68 |
|
|
cortex_r4_issue_a+cortex_r4_issue_alu")
|
69 |
|
|
(define_reservation "cortex_r4_load_store"
|
70 |
|
|
"cortex_r4_issue_a+cortex_r4_issue_c")
|
71 |
|
|
(define_reservation "cortex_r4_load_store_2"
|
72 |
|
|
"(cortex_r4_issue_a+cortex_r4_issue_b)*2")
|
73 |
|
|
(define_reservation "cortex_r4_branch" "cortex_r4_issue_b")
|
74 |
|
|
|
75 |
|
|
;; We assume that all instructions are unconditional.
|
76 |
|
|
|
77 |
|
|
;; Data processing instructions. Moves without shifts are kept separate
|
78 |
|
|
;; for the purposes of the dual-issue constraints above.
|
79 |
|
|
(define_insn_reservation "cortex_r4_alu" 2
|
80 |
|
|
(and (eq_attr "tune_cortexr4" "yes")
|
81 |
|
|
(and (eq_attr "type" "alu")
|
82 |
|
|
(not (eq_attr "insn" "mov"))))
|
83 |
|
|
"cortex_r4_alu")
|
84 |
|
|
|
85 |
|
|
(define_insn_reservation "cortex_r4_mov" 2
|
86 |
|
|
(and (eq_attr "tune_cortexr4" "yes")
|
87 |
|
|
(and (eq_attr "type" "alu")
|
88 |
|
|
(eq_attr "insn" "mov")))
|
89 |
|
|
"cortex_r4_mov")
|
90 |
|
|
|
91 |
|
|
(define_insn_reservation "cortex_r4_alu_shift" 2
|
92 |
|
|
(and (eq_attr "tune_cortexr4" "yes")
|
93 |
|
|
(eq_attr "type" "alu_shift"))
|
94 |
|
|
"cortex_r4_alu")
|
95 |
|
|
|
96 |
|
|
(define_insn_reservation "cortex_r4_alu_shift_reg" 2
|
97 |
|
|
(and (eq_attr "tune_cortexr4" "yes")
|
98 |
|
|
(eq_attr "type" "alu_shift_reg"))
|
99 |
|
|
"cortex_r4_alu_shift_reg")
|
100 |
|
|
|
101 |
|
|
;; An ALU instruction followed by an ALU instruction with no early dep.
|
102 |
|
|
(define_bypass 1 "cortex_r4_alu,cortex_r4_alu_shift,cortex_r4_alu_shift_reg,\
|
103 |
|
|
cortex_r4_mov"
|
104 |
|
|
"cortex_r4_alu")
|
105 |
|
|
(define_bypass 1 "cortex_r4_alu,cortex_r4_alu_shift,cortex_r4_alu_shift_reg,\
|
106 |
|
|
cortex_r4_mov"
|
107 |
|
|
"cortex_r4_alu_shift"
|
108 |
|
|
"arm_no_early_alu_shift_dep")
|
109 |
|
|
(define_bypass 1 "cortex_r4_alu,cortex_r4_alu_shift,cortex_r4_alu_shift_reg,\
|
110 |
|
|
cortex_r4_mov"
|
111 |
|
|
"cortex_r4_alu_shift_reg"
|
112 |
|
|
"arm_no_early_alu_shift_value_dep")
|
113 |
|
|
|
114 |
|
|
;; In terms of availabilities, a consumer mov could theoretically be
|
115 |
|
|
;; issued together with a producer ALU instruction, without stalls.
|
116 |
|
|
;; In practice this cannot happen because mov;add (in that order) is not
|
117 |
|
|
;; eligible for dual issue and furthermore dual issue is not permitted
|
118 |
|
|
;; when a dependency is involved. We therefore note it as latency one.
|
119 |
|
|
;; A mov followed by another of the same is also latency one.
|
120 |
|
|
(define_bypass 1 "cortex_r4_alu,cortex_r4_alu_shift,cortex_r4_alu_shift_reg,\
|
121 |
|
|
cortex_r4_mov"
|
122 |
|
|
"cortex_r4_mov")
|
123 |
|
|
|
124 |
|
|
;; qadd, qdadd, qsub and qdsub are not currently emitted, and neither are
|
125 |
|
|
;; media data processing instructions nor sad instructions.
|
126 |
|
|
|
127 |
|
|
;; Multiplication instructions.
|
128 |
|
|
|
129 |
|
|
(define_insn_reservation "cortex_r4_mul_4" 4
|
130 |
|
|
(and (eq_attr "tune_cortexr4" "yes")
|
131 |
|
|
(eq_attr "insn" "mul,smmul"))
|
132 |
|
|
"cortex_r4_mul_2")
|
133 |
|
|
|
134 |
|
|
(define_insn_reservation "cortex_r4_mul_3" 3
|
135 |
|
|
(and (eq_attr "tune_cortexr4" "yes")
|
136 |
|
|
(eq_attr "insn" "smulxy,smulwy,smuad,smusd"))
|
137 |
|
|
"cortex_r4_mul")
|
138 |
|
|
|
139 |
|
|
(define_insn_reservation "cortex_r4_mla_4" 4
|
140 |
|
|
(and (eq_attr "tune_cortexr4" "yes")
|
141 |
|
|
(eq_attr "insn" "mla,smmla"))
|
142 |
|
|
"cortex_r4_mul_2")
|
143 |
|
|
|
144 |
|
|
(define_insn_reservation "cortex_r4_mla_3" 3
|
145 |
|
|
(and (eq_attr "tune_cortexr4" "yes")
|
146 |
|
|
(eq_attr "insn" "smlaxy,smlawy,smlad,smlsd"))
|
147 |
|
|
"cortex_r4_mul")
|
148 |
|
|
|
149 |
|
|
(define_insn_reservation "cortex_r4_smlald" 3
|
150 |
|
|
(and (eq_attr "tune_cortexr4" "yes")
|
151 |
|
|
(eq_attr "insn" "smlald,smlsld"))
|
152 |
|
|
"cortex_r4_mul")
|
153 |
|
|
|
154 |
|
|
(define_insn_reservation "cortex_r4_mull" 4
|
155 |
|
|
(and (eq_attr "tune_cortexr4" "yes")
|
156 |
|
|
(eq_attr "insn" "smull,umull,umlal,umaal"))
|
157 |
|
|
"cortex_r4_mul_2")
|
158 |
|
|
|
159 |
|
|
;; A multiply or an MLA with a single-register result, followed by an
|
160 |
|
|
;; MLA with an accumulator dependency, has its result forwarded.
|
161 |
|
|
(define_bypass 2 "cortex_r4_mul_3,cortex_r4_mla_3"
|
162 |
|
|
"cortex_r4_mla_3,cortex_r4_mla_4"
|
163 |
|
|
"arm_mac_accumulator_is_mul_result")
|
164 |
|
|
|
165 |
|
|
(define_bypass 3 "cortex_r4_mul_4,cortex_r4_mla_4"
|
166 |
|
|
"cortex_r4_mla_3,cortex_r4_mla_4"
|
167 |
|
|
"arm_mac_accumulator_is_mul_result")
|
168 |
|
|
|
169 |
|
|
;; A multiply followed by an ALU instruction needing the multiply
|
170 |
|
|
;; result only at ALU has lower latency than one needing it at Shift.
|
171 |
|
|
(define_bypass 2 "cortex_r4_mul_3,cortex_r4_mla_3,cortex_r4_smlald"
|
172 |
|
|
"cortex_r4_alu")
|
173 |
|
|
(define_bypass 2 "cortex_r4_mul_3,cortex_r4_mla_3,cortex_r4_smlald"
|
174 |
|
|
"cortex_r4_alu_shift"
|
175 |
|
|
"arm_no_early_alu_shift_dep")
|
176 |
|
|
(define_bypass 2 "cortex_r4_mul_3,cortex_r4_mla_3,cortex_r4_smlald"
|
177 |
|
|
"cortex_r4_alu_shift_reg"
|
178 |
|
|
"arm_no_early_alu_shift_value_dep")
|
179 |
|
|
(define_bypass 3 "cortex_r4_mul_4,cortex_r4_mla_4,cortex_r4_mull"
|
180 |
|
|
"cortex_r4_alu")
|
181 |
|
|
(define_bypass 3 "cortex_r4_mul_4,cortex_r4_mla_4,cortex_r4_mull"
|
182 |
|
|
"cortex_r4_alu_shift"
|
183 |
|
|
"arm_no_early_alu_shift_dep")
|
184 |
|
|
(define_bypass 3 "cortex_r4_mul_4,cortex_r4_mla_4,cortex_r4_mull"
|
185 |
|
|
"cortex_r4_alu_shift_reg"
|
186 |
|
|
"arm_no_early_alu_shift_value_dep")
|
187 |
|
|
|
188 |
|
|
;; A multiply followed by a mov has one cycle lower latency again.
|
189 |
|
|
(define_bypass 1 "cortex_r4_mul_3,cortex_r4_mla_3,cortex_r4_smlald"
|
190 |
|
|
"cortex_r4_mov")
|
191 |
|
|
(define_bypass 2 "cortex_r4_mul_4,cortex_r4_mla_4,cortex_r4_mull"
|
192 |
|
|
"cortex_r4_mov")
|
193 |
|
|
|
194 |
|
|
;; We guess that division of A/B using sdiv or udiv, on average,
|
195 |
|
|
;; is performed with B having ten more leading zeros than A.
|
196 |
|
|
;; This gives a latency of nine for udiv and ten for sdiv.
|
197 |
|
|
(define_insn_reservation "cortex_r4_udiv" 9
|
198 |
|
|
(and (eq_attr "tune_cortexr4" "yes")
|
199 |
|
|
(eq_attr "insn" "udiv"))
|
200 |
|
|
"cortex_r4_div_9")
|
201 |
|
|
|
202 |
|
|
(define_insn_reservation "cortex_r4_sdiv" 10
|
203 |
|
|
(and (eq_attr "tune_cortexr4" "yes")
|
204 |
|
|
(eq_attr "insn" "sdiv"))
|
205 |
|
|
"cortex_r4_div_10")
|
206 |
|
|
|
207 |
|
|
;; Branches. We assume correct prediction.
|
208 |
|
|
|
209 |
|
|
(define_insn_reservation "cortex_r4_branch" 0
|
210 |
|
|
(and (eq_attr "tune_cortexr4" "yes")
|
211 |
|
|
(eq_attr "type" "branch"))
|
212 |
|
|
"cortex_r4_branch")
|
213 |
|
|
|
214 |
|
|
;; Call latencies are not predictable. A semi-arbitrary very large
|
215 |
|
|
;; number is used as "positive infinity" so that everything should be
|
216 |
|
|
;; finished by the time of return.
|
217 |
|
|
(define_insn_reservation "cortex_r4_call" 32
|
218 |
|
|
(and (eq_attr "tune_cortexr4" "yes")
|
219 |
|
|
(eq_attr "type" "call"))
|
220 |
|
|
"nothing")
|
221 |
|
|
|
222 |
|
|
;; Status register access instructions are not currently emitted.
|
223 |
|
|
|
224 |
|
|
;; Load instructions.
|
225 |
|
|
;; We do not model the "addr_md_3cycle" cases and assume that
|
226 |
|
|
;; accesses following are correctly aligned.
|
227 |
|
|
|
228 |
|
|
(define_insn_reservation "cortex_r4_load_1_2" 3
|
229 |
|
|
(and (eq_attr "tune_cortexr4" "yes")
|
230 |
|
|
(eq_attr "type" "load1,load2"))
|
231 |
|
|
"cortex_r4_load_store")
|
232 |
|
|
|
233 |
|
|
(define_insn_reservation "cortex_r4_load_3_4" 4
|
234 |
|
|
(and (eq_attr "tune_cortexr4" "yes")
|
235 |
|
|
(eq_attr "type" "load3,load4"))
|
236 |
|
|
"cortex_r4_load_store_2")
|
237 |
|
|
|
238 |
|
|
;; If a producing load is followed by an instruction consuming only
|
239 |
|
|
;; as a Normal Reg, there is one fewer cycle of latency.
|
240 |
|
|
|
241 |
|
|
(define_bypass 2 "cortex_r4_load_1_2"
|
242 |
|
|
"cortex_r4_alu")
|
243 |
|
|
(define_bypass 2 "cortex_r4_load_1_2"
|
244 |
|
|
"cortex_r4_alu_shift"
|
245 |
|
|
"arm_no_early_alu_shift_dep")
|
246 |
|
|
(define_bypass 2 "cortex_r4_load_1_2"
|
247 |
|
|
"cortex_r4_alu_shift_reg"
|
248 |
|
|
"arm_no_early_alu_shift_value_dep")
|
249 |
|
|
|
250 |
|
|
(define_bypass 3 "cortex_r4_load_3_4"
|
251 |
|
|
"cortex_r4_alu")
|
252 |
|
|
(define_bypass 3 "cortex_r4_load_3_4"
|
253 |
|
|
"cortex_r4_alu_shift"
|
254 |
|
|
"arm_no_early_alu_shift_dep")
|
255 |
|
|
(define_bypass 3 "cortex_r4_load_3_4"
|
256 |
|
|
"cortex_r4_alu_shift_reg"
|
257 |
|
|
"arm_no_early_alu_shift_value_dep")
|
258 |
|
|
|
259 |
|
|
;; If a producing load is followed by an instruction consuming only
|
260 |
|
|
;; as a Late Reg, there are two fewer cycles of latency. Such consumer
|
261 |
|
|
;; instructions are moves and stores.
|
262 |
|
|
|
263 |
|
|
(define_bypass 1 "cortex_r4_load_1_2"
|
264 |
|
|
"cortex_r4_mov,cortex_r4_store_1_2,cortex_r4_store_3_4")
|
265 |
|
|
(define_bypass 2 "cortex_r4_load_3_4"
|
266 |
|
|
"cortex_r4_mov,cortex_r4_store_1_2,cortex_r4_store_3_4")
|
267 |
|
|
|
268 |
|
|
;; If a producer's result is required as the base or offset of a load,
|
269 |
|
|
;; there is an extra cycle latency.
|
270 |
|
|
|
271 |
|
|
(define_bypass 3 "cortex_r4_alu,cortex_r4_mov,cortex_r4_alu_shift,\
|
272 |
|
|
cortex_r4_alu_shift_reg"
|
273 |
|
|
"cortex_r4_load_1_2,cortex_r4_load_3_4")
|
274 |
|
|
|
275 |
|
|
(define_bypass 4 "cortex_r4_mul_3,cortex_r4_mla_3,cortex_r4_smlald"
|
276 |
|
|
"cortex_r4_load_1_2,cortex_r4_load_3_4")
|
277 |
|
|
|
278 |
|
|
(define_bypass 5 "cortex_r4_mul_4,cortex_r4_mla_4,cortex_r4_mull"
|
279 |
|
|
"cortex_r4_load_1_2,cortex_r4_load_3_4")
|
280 |
|
|
|
281 |
|
|
;; Store instructions.
|
282 |
|
|
|
283 |
|
|
(define_insn_reservation "cortex_r4_store_1_2" 0
|
284 |
|
|
(and (eq_attr "tune_cortexr4" "yes")
|
285 |
|
|
(eq_attr "type" "store1,store2"))
|
286 |
|
|
"cortex_r4_load_store")
|
287 |
|
|
|
288 |
|
|
(define_insn_reservation "cortex_r4_store_3_4" 0
|
289 |
|
|
(and (eq_attr "tune_cortexr4" "yes")
|
290 |
|
|
(eq_attr "type" "store3,store4"))
|
291 |
|
|
"cortex_r4_load_store_2")
|
292 |
|
|
|