1 ;; ARM Cortex-A7 pipeline description
2 ;; Copyright (C) 2012-2013 Free Software Foundation, Inc.
4 ;; Contributed by ARM Ltd.
5 ;; Based on cortex-a5.md which was originally contributed by CodeSourcery.
7 ;; This file is part of GCC.
9 ;; GCC is free software; you can redistribute it and/or modify it
10 ;; under the terms of the GNU General Public License as published by
11 ;; the Free Software Foundation; either version 3, or (at your option)
14 ;; GCC is distributed in the hope that it will be useful, but
15 ;; WITHOUT ANY WARRANTY; without even the implied warranty of
16 ;; MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
17 ;; General Public License for more details.
19 ;; You should have received a copy of the GNU General Public License
20 ;; along with GCC; see the file COPYING3. If not see
21 ;; <http://www.gnu.org/licenses/>.
23 (define_automaton "cortex_a7")
25 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
27 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
29 ;; The Cortex-A7 pipeline integer and vfp pipeline.
30 ;; The decode is the same for all instructions, so do not model it.
31 ;; We only model the first execution stage because
32 ;; instructions always advance one stage per cycle in order.
33 ;; We model all of the LS, Branch, ALU, MAC and FPU pipelines together.
35 (define_cpu_unit "cortex_a7_ex1, cortex_a7_ex2" "cortex_a7")
37 (define_reservation "cortex_a7_both" "cortex_a7_ex1+cortex_a7_ex2")
39 (define_cpu_unit "cortex_a7_branch" "cortex_a7")
41 ;; Cortex-A7 is in order and can dual-issue under limited circumstances.
42 ;; ex2 can be reserved only after ex1 is reserved.
44 (final_presence_set "cortex_a7_ex2" "cortex_a7_ex1")
46 ;; Pseudo-unit for blocking the multiply pipeline when a double-precision
47 ;; multiply is in progress.
49 (define_cpu_unit "cortex_a7_fpmul_pipe" "cortex_a7")
51 ;; The floating-point add pipeline (ex1/f1 stage), used to model the usage
52 ;; of the add pipeline by fmac instructions, etc.
54 (define_cpu_unit "cortex_a7_fpadd_pipe" "cortex_a7")
56 ;; Floating-point div/sqrt (long latency, out-of-order completion).
58 (define_cpu_unit "cortex_a7_fp_div_sqrt" "cortex_a7")
60 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
62 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
64 ;; A direct branch can dual issue either as younger or older instruction,
65 ;; but branches cannot dual issue with branches.
66 ;; No latency as there is no result.
68 (define_insn_reservation "cortex_a7_branch" 0
69 (and (eq_attr "tune" "cortexa7")
70 (and (eq_attr "type" "branch")
71 (eq_attr "neon_type" "none")))
72 "(cortex_a7_ex2|cortex_a7_ex1)+cortex_a7_branch")
74 ;; Call cannot dual-issue as an older instruction. It can dual-issue
75 ;; as a younger instruction, or single-issue. Call cannot dual-issue
76 ;; with another branch instruction. The result is available the next
78 (define_insn_reservation "cortex_a7_call" 1
79 (and (eq_attr "tune" "cortexa7")
80 (and (eq_attr "type" "call")
81 (eq_attr "neon_type" "none")))
82 "(cortex_a7_ex2|cortex_a7_both)+cortex_a7_branch")
84 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
86 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
88 ;; ALU instruction with an immediate operand can dual-issue.
89 (define_insn_reservation "cortex_a7_alu_imm" 2
90 (and (eq_attr "tune" "cortexa7")
91 (and (ior (eq_attr "type" "arlo_imm,mov_imm,mvn_imm")
92 (ior (eq_attr "type" "extend")
93 (and (eq_attr "type" "mov_reg,mov_shift,mov_shift_reg")
94 (not (eq_attr "length" "8")))))
95 (eq_attr "neon_type" "none")))
96 "cortex_a7_ex2|cortex_a7_ex1")
98 ;; ALU instruction with register operands can dual-issue
99 ;; with a younger immediate-based instruction.
100 (define_insn_reservation "cortex_a7_alu_reg" 2
101 (and (eq_attr "tune" "cortexa7")
102 (and (eq_attr "type" "arlo_reg,shift,shift_reg,mov_reg,mvn_reg")
103 (eq_attr "neon_type" "none")))
106 (define_insn_reservation "cortex_a7_alu_shift" 2
107 (and (eq_attr "tune" "cortexa7")
108 (and (eq_attr "type" "arlo_shift,arlo_shift_reg,\
109 mov_shift,mov_shift_reg,\
110 mvn_shift,mvn_shift_reg")
111 (eq_attr "neon_type" "none")))
114 ;; Forwarding path for unshifted operands.
115 (define_bypass 1 "cortex_a7_alu_imm,cortex_a7_alu_reg,cortex_a7_alu_shift"
116 "cortex_a7_alu_imm,cortex_a7_alu_reg,cortex_a7_mul")
118 (define_bypass 1 "cortex_a7_alu_imm,cortex_a7_alu_reg,cortex_a7_alu_shift"
120 "arm_no_early_store_addr_dep")
122 (define_bypass 1 "cortex_a7_alu_imm,cortex_a7_alu_reg,cortex_a7_alu_shift"
123 "cortex_a7_alu_shift"
124 "arm_no_early_alu_shift_dep")
126 ;; The multiplier pipeline can forward results from wr stage only so
127 ;; there's no need to specify bypasses.
128 ;; Multiply instructions cannot dual-issue.
130 (define_insn_reservation "cortex_a7_mul" 2
131 (and (eq_attr "tune" "cortexa7")
132 (and (eq_attr "neon_type" "none")
133 (ior (eq_attr "mul32" "yes")
134 (eq_attr "mul64" "yes"))))
137 ;; Forward the result of a multiply operation to the accumulator
138 ;; of the following multiply and accumulate instruction.
139 (define_bypass 1 "cortex_a7_mul"
141 "arm_mac_accumulator_is_result")
143 ;; The latency depends on the operands, so we use an estimate here.
144 (define_insn_reservation "cortex_a7_idiv" 5
145 (and (eq_attr "tune" "cortexa7")
146 (eq_attr "type" "udiv,sdiv"))
149 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
150 ;; Load/store instructions.
151 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
153 ;; Address-generation happens in the issue stage.
154 ;; Double-word accesses can be issued in a single cycle,
155 ;; and occupy only one pipeline stage.
157 (define_insn_reservation "cortex_a7_load1" 2
158 (and (eq_attr "tune" "cortexa7")
159 (and (eq_attr "type" "load_byte,load1")
160 (eq_attr "neon_type" "none")))
163 (define_insn_reservation "cortex_a7_store1" 0
164 (and (eq_attr "tune" "cortexa7")
165 (and (eq_attr "type" "store1")
166 (eq_attr "neon_type" "none")))
169 (define_insn_reservation "cortex_a7_load2" 2
170 (and (eq_attr "tune" "cortexa7")
171 (and (eq_attr "type" "load2")
172 (eq_attr "neon_type" "none")))
175 (define_insn_reservation "cortex_a7_store2" 0
176 (and (eq_attr "tune" "cortexa7")
177 (and (eq_attr "type" "store2")
178 (eq_attr "neon_type" "none")))
181 (define_insn_reservation "cortex_a7_load3" 3
182 (and (eq_attr "tune" "cortexa7")
183 (and (eq_attr "type" "load3")
184 (eq_attr "neon_type" "none")))
185 "cortex_a7_both, cortex_a7_ex1")
187 (define_insn_reservation "cortex_a7_store3" 0
188 (and (eq_attr "tune" "cortexa7")
189 (and (eq_attr "type" "store4")
190 (eq_attr "neon_type" "none")))
191 "cortex_a7_both, cortex_a7_ex1")
193 (define_insn_reservation "cortex_a7_load4" 3
194 (and (eq_attr "tune" "cortexa7")
195 (and (eq_attr "type" "load4")
196 (eq_attr "neon_type" "none")))
197 "cortex_a7_both, cortex_a7_both")
199 (define_insn_reservation "cortex_a7_store4" 0
200 (and (eq_attr "tune" "cortexa7")
201 (and (eq_attr "type" "store3")
202 (eq_attr "neon_type" "none")))
203 "cortex_a7_both, cortex_a7_both")
205 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
206 ;; Floating-point arithmetic.
207 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
208 ;; Neon integer, neon floating point, and single-precision floating
209 ;; point instructions of the same type have the same timing
210 ;; characteristics, but neon instructions cannot dual-issue.
212 (define_insn_reservation "cortex_a7_fpalu" 4
213 (and (eq_attr "tune" "cortexa7")
214 (and (eq_attr "type" "ffariths, fadds, ffarithd, faddd, fcpys,\
215 f_cvt, fcmps, fcmpd")
216 (eq_attr "neon_type" "none")))
217 "cortex_a7_ex1+cortex_a7_fpadd_pipe")
219 ;; For fconsts and fconstd, 8-bit immediate data is passed directly from
220 ;; f1 to f3 (which I think reduces the latency by one cycle).
222 (define_insn_reservation "cortex_a7_fconst" 3
223 (and (eq_attr "tune" "cortexa7")
224 (and (eq_attr "type" "fconsts,fconstd")
225 (eq_attr "neon_type" "none")))
226 "cortex_a7_ex1+cortex_a7_fpadd_pipe")
228 ;; We should try not to attempt to issue a single-precision multiplication in
229 ;; the middle of a double-precision multiplication operation (the usage of
230 ;; cortex_a7_fpmul_pipe).
232 (define_insn_reservation "cortex_a7_fpmuls" 4
233 (and (eq_attr "tune" "cortexa7")
234 (and (eq_attr "type" "fmuls")
235 (eq_attr "neon_type" "none")))
236 "cortex_a7_ex1+cortex_a7_fpmul_pipe")
238 (define_insn_reservation "cortex_a7_neon_mul" 4
239 (and (eq_attr "tune" "cortexa7")
241 "neon_mul_ddd_8_16_qdd_16_8_long_32_16_long,\
242 neon_mul_qqq_8_16_32_ddd_32,\
243 neon_mul_qdd_64_32_long_qqd_16_ddd_32_scalar_64_32_long_scalar,\
244 neon_mul_ddd_16_scalar_32_16_long_scalar,\
245 neon_mul_qqd_32_scalar,\
248 "(cortex_a7_both+cortex_a7_fpmul_pipe)*2")
250 (define_insn_reservation "cortex_a7_fpmacs" 8
251 (and (eq_attr "tune" "cortexa7")
252 (and (eq_attr "type" "fmacs,ffmas")
253 (eq_attr "neon_type" "none")))
254 "cortex_a7_ex1+cortex_a7_fpmul_pipe")
256 (define_insn_reservation "cortex_a7_neon_mla" 8
257 (and (eq_attr "tune" "cortexa7")
259 "neon_mla_ddd_8_16_qdd_16_8_long_32_16_long,\
261 neon_mla_ddd_32_qqd_16_ddd_32_scalar_qdd_64_32_long_scalar_qdd_64_32_long,\
262 neon_mla_qqq_32_qqd_32_scalar,\
263 neon_mla_ddd_16_scalar_qdd_32_16_long_scalar,\
266 neon_fp_vmla_ddd_scalar,\
267 neon_fp_vmla_qqq_scalar"))
268 "cortex_a7_both+cortex_a7_fpmul_pipe")
270 (define_bypass 4 "cortex_a7_fpmacs,cortex_a7_neon_mla"
271 "cortex_a7_fpmacs,cortex_a7_neon_mla"
272 "arm_mac_accumulator_is_result")
274 ;; Non-multiply instructions can issue between two cycles of a
275 ;; double-precision multiply.
277 (define_insn_reservation "cortex_a7_fpmuld" 7
278 (and (eq_attr "tune" "cortexa7")
279 (and (eq_attr "type" "fmuld")
280 (eq_attr "neon_type" "none")))
281 "cortex_a7_ex1+cortex_a7_fpmul_pipe, cortex_a7_fpmul_pipe*3")
283 (define_insn_reservation "cortex_a7_fpmacd" 11
284 (and (eq_attr "tune" "cortexa7")
285 (and (eq_attr "type" "fmacd")
286 (eq_attr "neon_type" "none")))
287 "cortex_a7_ex1+cortex_a7_fpmul_pipe, cortex_a7_fpmul_pipe*3")
289 (define_insn_reservation "cortex_a7_fpfmad" 8
290 (and (eq_attr "tune" "cortexa7")
291 (and (eq_attr "type" "ffmad")
292 (eq_attr "neon_type" "none")))
293 "cortex_a7_ex1+cortex_a7_fpmul_pipe, cortex_a7_fpmul_pipe*4")
295 (define_bypass 7 "cortex_a7_fpmacd"
296 "cortex_a7_fpmacd,cortex_a7_fpfmad"
297 "arm_mac_accumulator_is_result")
299 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
300 ;; Floating-point divide/square root instructions.
301 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
303 (define_insn_reservation "cortex_a7_fdivs" 16
304 (and (eq_attr "tune" "cortexa7")
305 (and (eq_attr "type" "fdivs")
306 (eq_attr "neon_type" "none")))
307 "cortex_a7_ex1+cortex_a7_fp_div_sqrt, cortex_a7_fp_div_sqrt * 13")
309 (define_insn_reservation "cortex_a7_fdivd" 31
310 (and (eq_attr "tune" "cortexa7")
311 (and (eq_attr "type" "fdivd")
312 (eq_attr "neon_type" "none")))
313 "cortex_a7_ex1+cortex_a7_fp_div_sqrt, cortex_a7_fp_div_sqrt * 28")
315 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
316 ;; VFP to/from core transfers.
317 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
319 ;; Core-to-VFP transfers.
321 (define_insn_reservation "cortex_a7_r2f" 4
322 (and (eq_attr "tune" "cortexa7")
323 (and (eq_attr "type" "r_2_f")
324 (eq_attr "neon_type" "none")))
327 (define_insn_reservation "cortex_a7_f2r" 2
328 (and (eq_attr "tune" "cortexa7")
329 (and (eq_attr "type" "f_2_r")
330 (eq_attr "neon_type" "none")))
333 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
334 ;; VFP flag transfer.
335 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
337 ;; Fuxne: The flag forwarding from fmstat to the second instruction is
338 ;; not modeled at present.
340 (define_insn_reservation "cortex_a7_f_flags" 4
341 (and (eq_attr "tune" "cortexa7")
342 (and (eq_attr "type" "f_flag")
343 (eq_attr "neon_type" "none")))
346 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
348 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
350 (define_insn_reservation "cortex_a7_f_loads" 4
351 (and (eq_attr "tune" "cortexa7")
352 (and (eq_attr "type" "f_loads")
353 (eq_attr "neon_type" "none")))
356 (define_insn_reservation "cortex_a7_f_loadd" 4
357 (and (eq_attr "tune" "cortexa7")
358 (and (eq_attr "type" "f_loadd")
359 (eq_attr "neon_type" "none")))
362 (define_insn_reservation "cortex_a7_f_stores" 0
363 (and (eq_attr "tune" "cortexa7")
364 (and (eq_attr "type" "f_stores")
365 (eq_attr "neon_type" "none")))
368 (define_insn_reservation "cortex_a7_f_stored" 0
369 (and (eq_attr "tune" "cortexa7")
370 (and (eq_attr "type" "f_stored")
371 (eq_attr "neon_type" "none")))
374 ;; Load-to-use for floating-point values has a penalty of one cycle,
375 ;; i.e. a latency of two.
377 (define_bypass 2 "cortex_a7_f_loads, cortex_a7_f_loadd"
379 cortex_a7_fpmuls,cortex_a7_fpmacs,\
380 cortex_a7_fpmuld,cortex_a7_fpmacd, cortex_a7_fpfmad,\
381 cortex_a7_fdivs, cortex_a7_fdivd,\
384 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
386 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
388 ;; Simple modeling for all neon instructions not covered earlier.
390 (define_insn_reservation "cortex_a7_neon" 4
391 (and (eq_attr "tune" "cortexa7")
394 neon_mul_ddd_8_16_qdd_16_8_long_32_16_long,\
395 neon_mul_qqq_8_16_32_ddd_32,\
396 neon_mul_qdd_64_32_long_qqd_16_ddd_32_scalar_64_32_long_scalar,\
397 neon_mla_ddd_8_16_qdd_16_8_long_32_16_long,\
399 neon_mla_ddd_32_qqd_16_ddd_32_scalar_qdd_64_32_long_scalar_qdd_64_32_long,\
400 neon_mla_qqq_32_qqd_32_scalar,\
401 neon_mul_ddd_16_scalar_32_16_long_scalar,\
402 neon_mul_qqd_32_scalar,\
403 neon_mla_ddd_16_scalar_qdd_32_16_long_scalar,\
408 neon_fp_vmla_ddd_scalar,\
409 neon_fp_vmla_qqq_scalar"))