1 ;; ARM Cortex-A53 pipeline description
2 ;; Copyright (C) 2013-2014 Free Software Foundation, Inc.
4 ;; Contributed by ARM Ltd.
6 ;; This file is part of GCC.
8 ;; GCC is free software; you can redistribute it and/or modify it
9 ;; under the terms of the GNU General Public License as published by
10 ;; the Free Software Foundation; either version 3, or (at your option)
13 ;; GCC is distributed in the hope that it will be useful, but
14 ;; WITHOUT ANY WARRANTY; without even the implied warranty of
15 ;; MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
16 ;; General Public License for more details.
18 ;; You should have received a copy of the GNU General Public License
19 ;; along with GCC; see the file COPYING3. If not see
20 ;; <http://www.gnu.org/licenses/>.
22 (define_automaton "cortex_a53")
24 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
26 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
28 ;; There are two main integer execution pipelines, described as
29 ;; slot 0 and issue slot 1.
31 (define_cpu_unit "cortex_a53_slot0" "cortex_a53")
32 (define_cpu_unit "cortex_a53_slot1" "cortex_a53")
34 (define_reservation "cortex_a53_slot_any" "cortex_a53_slot0|cortex_a53_slot1")
35 (define_reservation "cortex_a53_single_issue" "cortex_a53_slot0+cortex_a53_slot1")
37 ;; The load/store pipeline. Load/store instructions can dual-issue from
38 ;; either pipeline, but two load/stores cannot simultaneously issue.
40 (define_cpu_unit "cortex_a53_ls" "cortex_a53")
42 ;; The store pipeline. Shared between both execution pipelines.
44 (define_cpu_unit "cortex_a53_store" "cortex_a53")
46 ;; The branch pipeline. Branches can dual-issue with other instructions
47 ;; (except when those instructions take multiple cycles to issue).
49 (define_cpu_unit "cortex_a53_branch" "cortex_a53")
51 ;; The integer divider.
53 (define_cpu_unit "cortex_a53_idiv" "cortex_a53")
55 ;; The floating-point add pipeline used to model the usage
56 ;; of the add pipeline by fmac instructions.
58 (define_cpu_unit "cortex_a53_fpadd_pipe" "cortex_a53")
60 ;; Floating-point div/sqrt (long latency, out-of-order completion).
62 (define_cpu_unit "cortex_a53_fp_div_sqrt" "cortex_a53")
64 ;; The Advanced SIMD pipelines.
66 (define_cpu_unit "cortex_a53_simd0" "cortex_a53")
67 (define_cpu_unit "cortex_a53_simd1" "cortex_a53")
69 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
71 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
73 (define_insn_reservation "cortex_a53_alu" 2
74 (and (eq_attr "tune" "cortexa53")
75 (eq_attr "type" "alu_imm,alus_imm,logic_imm,logics_imm,\
76 alu_reg,alus_reg,logic_reg,logics_reg,\
77 adc_imm,adcs_imm,adc_reg,adcs_reg,\
80 mov_imm,mov_reg,mvn_imm,mvn_reg,\
81 mrs,multiple,no_insn"))
82 "cortex_a53_slot_any")
84 (define_insn_reservation "cortex_a53_alu_shift" 2
85 (and (eq_attr "tune" "cortexa53")
86 (eq_attr "type" "alu_shift_imm,alus_shift_imm,\
87 logic_shift_imm,logics_shift_imm,\
88 alu_shift_reg,alus_shift_reg,\
89 logic_shift_reg,logics_shift_reg,\
90 extend,mov_shift,mov_shift_reg,\
91 mvn_shift,mvn_shift_reg"))
92 "cortex_a53_slot_any")
94 ;; Forwarding path for unshifted operands.
96 (define_bypass 1 "cortex_a53_alu,cortex_a53_alu_shift"
99 (define_bypass 1 "cortex_a53_alu,cortex_a53_alu_shift"
100 "cortex_a53_alu_shift"
101 "arm_no_early_alu_shift_dep")
103 ;; The multiplier pipeline can forward results so there's no need to specify
104 ;; bypasses. Multiplies can only single-issue currently.
106 (define_insn_reservation "cortex_a53_mul" 3
107 (and (eq_attr "tune" "cortexa53")
108 (ior (eq_attr "mul32" "yes")
109 (eq_attr "mul64" "yes")))
110 "cortex_a53_single_issue")
112 ;; A multiply with a single-register result or an MLA, followed by an
113 ;; MLA with an accumulator dependency, has its result forwarded so two
114 ;; such instructions can issue back-to-back.
116 (define_bypass 1 "cortex_a53_mul"
118 "arm_mac_accumulator_is_mul_result")
120 ;; Punt with a high enough latency for divides.
121 (define_insn_reservation "cortex_a53_udiv" 8
122 (and (eq_attr "tune" "cortexa53")
123 (eq_attr "type" "udiv"))
124 "(cortex_a53_slot0+cortex_a53_idiv),cortex_a53_idiv*7")
126 (define_insn_reservation "cortex_a53_sdiv" 9
127 (and (eq_attr "tune" "cortexa53")
128 (eq_attr "type" "sdiv"))
129 "(cortex_a53_slot0+cortex_a53_idiv),cortex_a53_idiv*8")
132 (define_bypass 2 "cortex_a53_mul,cortex_a53_udiv,cortex_a53_sdiv"
134 (define_bypass 2 "cortex_a53_mul,cortex_a53_udiv,cortex_a53_sdiv"
135 "cortex_a53_alu_shift"
136 "arm_no_early_alu_shift_dep")
138 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
139 ;; Load/store instructions.
140 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
142 ;; Address-generation happens in the issue stage.
144 (define_insn_reservation "cortex_a53_load1" 3
145 (and (eq_attr "tune" "cortexa53")
146 (eq_attr "type" "load_byte,load1,load_acq"))
147 "cortex_a53_slot_any+cortex_a53_ls")
149 (define_insn_reservation "cortex_a53_store1" 2
150 (and (eq_attr "tune" "cortexa53")
151 (eq_attr "type" "store1,store_rel"))
152 "cortex_a53_slot_any+cortex_a53_ls+cortex_a53_store")
154 (define_insn_reservation "cortex_a53_load2" 3
155 (and (eq_attr "tune" "cortexa53")
156 (eq_attr "type" "load2"))
157 "cortex_a53_single_issue+cortex_a53_ls")
159 (define_insn_reservation "cortex_a53_store2" 2
160 (and (eq_attr "tune" "cortexa53")
161 (eq_attr "type" "store2"))
162 "cortex_a53_single_issue+cortex_a53_ls+cortex_a53_store")
164 (define_insn_reservation "cortex_a53_load3plus" 4
165 (and (eq_attr "tune" "cortexa53")
166 (eq_attr "type" "load3,load4"))
167 "(cortex_a53_single_issue+cortex_a53_ls)*2")
169 (define_insn_reservation "cortex_a53_store3plus" 3
170 (and (eq_attr "tune" "cortexa53")
171 (eq_attr "type" "store3,store4"))
172 "(cortex_a53_single_issue+cortex_a53_ls+cortex_a53_store)*2")
174 ;; Load/store addresses are required early in Issue.
175 (define_bypass 3 "cortex_a53_load1,cortex_a53_load2,cortex_a53_load3plus,cortex_a53_alu,cortex_a53_alu_shift"
177 "arm_early_load_addr_dep")
178 (define_bypass 3 "cortex_a53_load1,cortex_a53_load2,cortex_a53_load3plus,cortex_a53_alu,cortex_a53_alu_shift"
180 "arm_early_store_addr_dep")
182 ;; Load data can forward in the ALU pipeline
183 (define_bypass 2 "cortex_a53_load1,cortex_a53_load2"
185 (define_bypass 2 "cortex_a53_load1,cortex_a53_load2"
186 "cortex_a53_alu_shift"
187 "arm_no_early_alu_shift_dep")
189 ;; ALU ops can forward to stores.
190 (define_bypass 0 "cortex_a53_alu,cortex_a53_alu_shift"
191 "cortex_a53_store1,cortex_a53_store2,cortex_a53_store3plus"
192 "arm_no_early_store_addr_dep")
194 (define_bypass 1 "cortex_a53_mul,cortex_a53_udiv,cortex_a53_sdiv,cortex_a53_load1,cortex_a53_load2,cortex_a53_load3plus"
195 "cortex_a53_store1,cortex_a53_store2,cortex_a53_store3plus"
196 "arm_no_early_store_addr_dep")
198 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
200 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
202 ;; Currently models all branches as dual-issuable from either execution
203 ;; slot, which isn't true for all cases. We still need to model indirect
206 (define_insn_reservation "cortex_a53_branch" 0
207 (and (eq_attr "tune" "cortexa53")
208 (eq_attr "type" "branch,call"))
209 "cortex_a53_slot_any+cortex_a53_branch")
211 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
212 ;; Floating-point arithmetic.
213 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
215 (define_insn_reservation "cortex_a53_fpalu" 4
216 (and (eq_attr "tune" "cortexa53")
217 (eq_attr "type" "ffariths, fadds, ffarithd, faddd, fmov, fmuls,\
218 f_cvt,f_cvtf2i,f_cvti2f,\
219 fcmps, fcmpd, fcsel"))
220 "cortex_a53_slot0+cortex_a53_fpadd_pipe")
222 (define_insn_reservation "cortex_a53_fconst" 2
223 (and (eq_attr "tune" "cortexa53")
224 (eq_attr "type" "fconsts,fconstd"))
225 "cortex_a53_slot0+cortex_a53_fpadd_pipe")
227 (define_insn_reservation "cortex_a53_fpmul" 4
228 (and (eq_attr "tune" "cortexa53")
229 (eq_attr "type" "fmuls,fmuld"))
232 ;; For single-precision multiply-accumulate, the add (accumulate) is issued after
233 ;; the multiply completes. Model that accordingly.
235 (define_insn_reservation "cortex_a53_fpmac" 8
236 (and (eq_attr "tune" "cortexa53")
237 (eq_attr "type" "fmacs,fmacd,ffmas,ffmad"))
238 "cortex_a53_slot0, nothing*3, cortex_a53_fpadd_pipe")
240 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
241 ;; Floating-point divide/square root instructions.
242 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
243 ;; fsqrt really takes one cycle less, but that is not modelled.
245 (define_insn_reservation "cortex_a53_fdivs" 14
246 (and (eq_attr "tune" "cortexa53")
247 (eq_attr "type" "fdivs, fsqrts"))
248 "cortex_a53_slot0, cortex_a53_fp_div_sqrt * 5")
250 (define_insn_reservation "cortex_a53_fdivd" 29
251 (and (eq_attr "tune" "cortexa53")
252 (eq_attr "type" "fdivd, fsqrtd"))
253 "cortex_a53_slot0, cortex_a53_fp_div_sqrt * 8")
255 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
256 ;; ARMv8-A Cryptographic extensions.
257 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
259 (define_insn_reservation "cortex_a53_crypto_aese" 2
260 (and (eq_attr "tune" "cortexa53")
261 (eq_attr "type" "crypto_aese"))
264 (define_insn_reservation "cortex_a53_crypto_aesmc" 2
265 (and (eq_attr "tune" "cortexa53")
266 (eq_attr "type" "crypto_aesmc"))
267 "cortex_a53_simd0 | cortex_a53_simd1")
269 (define_insn_reservation "cortex_a53_crypto_sha1_fast" 2
270 (and (eq_attr "tune" "cortexa53")
271 (eq_attr "type" "crypto_sha1_fast, crypto_sha256_fast"))
274 (define_insn_reservation "cortex_a53_crypto_sha1_xor" 3
275 (and (eq_attr "tune" "cortexa53")
276 (eq_attr "type" "crypto_sha1_xor"))
279 (define_insn_reservation "cortex_a53_crypto_sha_slow" 5
280 (and (eq_attr "tune" "cortexa53")
281 (eq_attr "type" "crypto_sha1_slow, crypto_sha256_slow"))
284 (define_bypass 0 "cortex_a53_crypto_aese"
285 "cortex_a53_crypto_aesmc"
286 "aarch_crypto_can_dual_issue")
288 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
289 ;; VFP to/from core transfers.
290 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
292 (define_insn_reservation "cortex_a53_r2f" 4
293 (and (eq_attr "tune" "cortexa53")
294 (eq_attr "type" "f_mcr,f_mcrr"))
297 (define_insn_reservation "cortex_a53_f2r" 2
298 (and (eq_attr "tune" "cortexa53")
299 (eq_attr "type" "f_mrc,f_mrrc"))
302 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
303 ;; VFP flag transfer.
304 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
306 (define_insn_reservation "cortex_a53_f_flags" 4
307 (and (eq_attr "tune" "cortexa53")
308 (eq_attr "type" "f_flag"))
311 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
313 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
315 (define_insn_reservation "cortex_a53_f_loads" 4
316 (and (eq_attr "tune" "cortexa53")
317 (eq_attr "type" "f_loads"))
320 (define_insn_reservation "cortex_a53_f_loadd" 5
321 (and (eq_attr "tune" "cortexa53")
322 (eq_attr "type" "f_loadd"))
325 (define_insn_reservation "cortex_a53_f_load_2reg" 5
326 (and (eq_attr "tune" "cortexa53")
327 (eq_attr "type" "neon_load2_2reg_q"))
328 "(cortex_a53_slot_any+cortex_a53_ls)*2")
330 (define_insn_reservation "cortex_a53_f_loadq" 5
331 (and (eq_attr "tune" "cortexa53")
332 (eq_attr "type" "neon_load1_1reg_q"))
333 "cortex_a53_slot_any+cortex_a53_ls")
335 (define_insn_reservation "cortex_a53_f_stores" 0
336 (and (eq_attr "tune" "cortexa53")
337 (eq_attr "type" "f_stores"))
340 (define_insn_reservation "cortex_a53_f_stored" 0
341 (and (eq_attr "tune" "cortexa53")
342 (eq_attr "type" "f_stored"))
345 ;; Load-to-use for floating-point values has a penalty of one cycle,
346 ;; i.e. a latency of two.
348 (define_bypass 2 "cortex_a53_f_loads"
349 "cortex_a53_fpalu, cortex_a53_fpmac, cortex_a53_fpmul,\
350 cortex_a53_fdivs, cortex_a53_fdivd,\
353 (define_bypass 2 "cortex_a53_f_loadd"
354 "cortex_a53_fpalu, cortex_a53_fpmac, cortex_a53_fpmul,\
355 cortex_a53_fdivs, cortex_a53_fdivd,\
358 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
359 ;; Crude Advanced SIMD approximation.
360 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
362 (define_insn_reservation "cortex_53_advsimd" 4
363 (and (eq_attr "tune" "cortexa53")
364 (eq_attr "is_neon_type" "yes"))