Mercurial > hg > CbC > CbC_gcc
comparison gcc/config/arm/cortex-a9.md @ 111:04ced10e8804
gcc 7
author | kono |
---|---|
date | Fri, 27 Oct 2017 22:46:09 +0900 |
parents | f6334be47118 |
children | 84e7813d76e9 |
comparison
equal
deleted
inserted
replaced
68:561a7518be6b | 111:04ced10e8804 |
---|---|
1 ;; ARM Cortex-A9 pipeline description | 1 ;; ARM Cortex-A9 pipeline description |
2 ;; Copyright (C) 2008, 2009, 2010 Free Software Foundation, Inc. | 2 ;; Copyright (C) 2008-2017 Free Software Foundation, Inc. |
3 ;; Originally written by CodeSourcery for VFP. | 3 ;; Originally written by CodeSourcery for VFP. |
4 ;; | 4 ;; |
5 ;; Rewritten by Ramana Radhakrishnan <ramana.radhakrishnan@arm.com> | 5 ;; Rewritten by Ramana Radhakrishnan <ramana.radhakrishnan@arm.com> |
6 ;; Integer Pipeline description contributed by ARM Ltd. | 6 ;; Integer Pipeline description contributed by ARM Ltd. |
7 ;; VFP Pipeline description rewritten and contributed by ARM Ltd. | 7 ;; VFP Pipeline description rewritten and contributed by ARM Ltd. |
66 "cortex_a9_multcycle1, cortex_a9_mac_m2, cortex_a9_p0_wb") | 66 "cortex_a9_multcycle1, cortex_a9_mac_m2, cortex_a9_p0_wb") |
67 (define_reservation "cortex_a9_mult" | 67 (define_reservation "cortex_a9_mult" |
68 "cortex_a9_mac_m1*2, cortex_a9_mac_m2, cortex_a9_p0_wb") | 68 "cortex_a9_mac_m1*2, cortex_a9_mac_m2, cortex_a9_p0_wb") |
69 (define_reservation "cortex_a9_mac" | 69 (define_reservation "cortex_a9_mac" |
70 "cortex_a9_multcycle1*2 ,cortex_a9_mac_m2, cortex_a9_p0_wb") | 70 "cortex_a9_multcycle1*2 ,cortex_a9_mac_m2, cortex_a9_p0_wb") |
71 | 71 (define_reservation "cortex_a9_mult_long" |
72 "cortex_a9_mac_m1*3, cortex_a9_mac_m2, cortex_a9_p0_wb") | |
72 | 73 |
73 ;; Issue at the same time along the load store pipeline and | 74 ;; Issue at the same time along the load store pipeline and |
74 ;; the VFP / Neon pipeline is not possible. | 75 ;; the VFP / Neon pipeline is not possible. |
75 (exclusion_set "cortex_a9_ls" "ca9_issue_vfp_neon") | 76 (exclusion_set "cortex_a9_ls" "ca9_issue_vfp_neon") |
76 | 77 |
77 ;; Default data processing instruction without any shift | 78 ;; Default data processing instruction without any shift |
78 ;; The only exception to this is the mov instruction | 79 ;; The only exception to this is the mov instruction |
79 ;; which can go down E2 without any problem. | 80 ;; which can go down E2 without any problem. |
80 (define_insn_reservation "cortex_a9_dp" 2 | 81 (define_insn_reservation "cortex_a9_dp" 2 |
81 (and (eq_attr "tune" "cortexa9") | 82 (and (eq_attr "tune" "cortexa9") |
82 (ior (and (eq_attr "type" "alu") | 83 (eq_attr "type" "alu_imm,alus_imm,logic_imm,logics_imm,\ |
83 (eq_attr "neon_type" "none")) | 84 alu_sreg,alus_sreg,logic_reg,logics_reg,\ |
84 (and (and (eq_attr "type" "alu_shift_reg, alu_shift") | 85 adc_imm,adcs_imm,adc_reg,adcs_reg,\ |
85 (eq_attr "insn" "mov")) | 86 adr,bfm,clz,rbit,rev,alu_dsp_reg,\ |
86 (eq_attr "neon_type" "none")))) | 87 shift_imm,shift_reg,\ |
88 mov_imm,mov_reg,mvn_imm,mvn_reg,\ | |
89 mov_shift_reg,mov_shift,\ | |
90 mrs,multiple,no_insn")) | |
87 "cortex_a9_p0_default|cortex_a9_p1_default") | 91 "cortex_a9_p0_default|cortex_a9_p1_default") |
88 | 92 |
89 ;; An instruction using the shifter will go down E1. | 93 ;; An instruction using the shifter will go down E1. |
90 (define_insn_reservation "cortex_a9_dp_shift" 3 | 94 (define_insn_reservation "cortex_a9_dp_shift" 3 |
91 (and (eq_attr "tune" "cortexa9") | 95 (and (eq_attr "tune" "cortexa9") |
92 (and (eq_attr "type" "alu_shift_reg, alu_shift") | 96 (eq_attr "type" "alu_shift_imm,alus_shift_imm,\ |
93 (not (eq_attr "insn" "mov")))) | 97 logic_shift_imm,logics_shift_imm,\ |
98 alu_shift_reg,alus_shift_reg,\ | |
99 logic_shift_reg,logics_shift_reg,\ | |
100 extend,mvn_shift,mvn_shift_reg")) | |
94 "cortex_a9_p0_shift | cortex_a9_p1_shift") | 101 "cortex_a9_p0_shift | cortex_a9_p1_shift") |
95 | 102 |
96 ;; Loads have a latency of 4 cycles. | 103 ;; Loads have a latency of 4 cycles. |
97 ;; We don't model autoincrement instructions. These | 104 ;; We don't model autoincrement instructions. These |
98 ;; instructions use the load store pipeline and 1 of | 105 ;; instructions use the load store pipeline and 1 of |
99 ;; the E2 units to write back the result of the increment. | 106 ;; the E2 units to write back the result of the increment. |
100 | 107 |
101 (define_insn_reservation "cortex_a9_load1_2" 4 | 108 (define_insn_reservation "cortex_a9_load1_2" 4 |
102 (and (eq_attr "tune" "cortexa9") | 109 (and (eq_attr "tune" "cortexa9") |
103 (eq_attr "type" "load1, load2, load_byte, f_loads, f_loadd")) | 110 (eq_attr "type" "load_4, load_8, load_byte, f_loads, f_loadd")) |
104 "cortex_a9_ls") | 111 "cortex_a9_ls") |
105 | 112 |
106 ;; Loads multiples and store multiples can't be issued for 2 cycles in a | 113 ;; Loads multiples and store multiples can't be issued for 2 cycles in a |
107 ;; row. The description below assumes that addresses are 64 bit aligned. | 114 ;; row. The description below assumes that addresses are 64 bit aligned. |
108 ;; If not, there is an extra cycle latency which is not modelled. | 115 ;; If not, there is an extra cycle latency which is not modelled. |
109 | 116 |
110 (define_insn_reservation "cortex_a9_load3_4" 5 | 117 (define_insn_reservation "cortex_a9_load3_4" 5 |
111 (and (eq_attr "tune" "cortexa9") | 118 (and (eq_attr "tune" "cortexa9") |
112 (eq_attr "type" "load3, load4")) | 119 (eq_attr "type" "load_12, load_16")) |
113 "cortex_a9_ls, cortex_a9_ls") | 120 "cortex_a9_ls, cortex_a9_ls") |
114 | 121 |
115 (define_insn_reservation "cortex_a9_store1_2" 0 | 122 (define_insn_reservation "cortex_a9_store1_2" 0 |
116 (and (eq_attr "tune" "cortexa9") | 123 (and (eq_attr "tune" "cortexa9") |
117 (eq_attr "type" "store1, store2, f_stores, f_stored")) | 124 (eq_attr "type" "store_4, store_8, f_stores, f_stored")) |
118 "cortex_a9_ls") | 125 "cortex_a9_ls") |
119 | 126 |
120 ;; Almost all our store multiples use an auto-increment | 127 ;; Almost all our store multiples use an auto-increment |
121 ;; form. Don't issue back to back load and store multiples | 128 ;; form. Don't issue back to back load and store multiples |
122 ;; because the load store unit will stall. | 129 ;; because the load store unit will stall. |
123 | 130 |
124 (define_insn_reservation "cortex_a9_store3_4" 0 | 131 (define_insn_reservation "cortex_a9_store3_4" 0 |
125 (and (eq_attr "tune" "cortexa9") | 132 (and (eq_attr "tune" "cortexa9") |
126 (eq_attr "type" "store3, store4")) | 133 (eq_attr "type" "store_12, store_16")) |
127 "cortex_a9_ls+(cortex_a9_p0_default | cortex_a9_p1_default), cortex_a9_ls") | 134 "cortex_a9_ls+(cortex_a9_p0_default | cortex_a9_p1_default), cortex_a9_ls") |
128 | 135 |
129 ;; We get 16*16 multiply / mac results in 3 cycles. | 136 ;; We get 16*16 multiply / mac results in 3 cycles. |
130 (define_insn_reservation "cortex_a9_mult16" 3 | 137 (define_insn_reservation "cortex_a9_mult16" 3 |
131 (and (eq_attr "tune" "cortexa9") | 138 (and (eq_attr "tune" "cortexa9") |
132 (eq_attr "insn" "smulxy")) | 139 (eq_attr "type" "smulxy")) |
133 "cortex_a9_mult16") | 140 "cortex_a9_mult16") |
134 | 141 |
135 ;; The 16*16 mac is slightly different that it | 142 ;; The 16*16 mac is slightly different that it |
136 ;; reserves M1 and M2 in the same cycle. | 143 ;; reserves M1 and M2 in the same cycle. |
137 (define_insn_reservation "cortex_a9_mac16" 3 | 144 (define_insn_reservation "cortex_a9_mac16" 3 |
138 (and (eq_attr "tune" "cortexa9") | 145 (and (eq_attr "tune" "cortexa9") |
139 (eq_attr "insn" "smlaxy")) | 146 (eq_attr "type" "smlaxy")) |
140 "cortex_a9_mac16") | 147 "cortex_a9_mac16") |
141 | 148 |
142 | |
143 (define_insn_reservation "cortex_a9_multiply" 4 | 149 (define_insn_reservation "cortex_a9_multiply" 4 |
144 (and (eq_attr "tune" "cortexa9") | 150 (and (eq_attr "tune" "cortexa9") |
145 (eq_attr "insn" "mul")) | 151 (eq_attr "type" "mul,smmul,smmulr")) |
146 "cortex_a9_mult") | 152 "cortex_a9_mult") |
147 | 153 |
148 (define_insn_reservation "cortex_a9_mac" 4 | 154 (define_insn_reservation "cortex_a9_mac" 4 |
149 (and (eq_attr "tune" "cortexa9") | 155 (and (eq_attr "tune" "cortexa9") |
150 (eq_attr "insn" "mla")) | 156 (eq_attr "type" "mla,smmla")) |
151 "cortex_a9_mac") | 157 "cortex_a9_mac") |
158 | |
159 (define_insn_reservation "cortex_a9_multiply_long" 5 | |
160 (and (eq_attr "tune" "cortexa9") | |
161 (eq_attr "type" "smull,umull,smulls,umulls,smlal,smlals,umlal,umlals")) | |
162 "cortex_a9_mult_long") | |
152 | 163 |
153 ;; An instruction with a result in E2 can be forwarded | 164 ;; An instruction with a result in E2 can be forwarded |
154 ;; to E2 or E1 or M1 or the load store unit in the next cycle. | 165 ;; to E2 or E1 or M1 or the load store unit in the next cycle. |
155 | 166 |
156 (define_bypass 1 "cortex_a9_dp" | 167 (define_bypass 1 "cortex_a9_dp" |
157 "cortex_a9_dp_shift, cortex_a9_multiply, | 168 "cortex_a9_dp_shift, cortex_a9_multiply, |
158 cortex_a9_load1_2, cortex_a9_dp, cortex_a9_store1_2, | 169 cortex_a9_load1_2, cortex_a9_dp, cortex_a9_store1_2, |
159 cortex_a9_mult16, cortex_a9_mac16, cortex_a9_mac, cortex_a9_store3_4, cortex_a9_load3_4") | 170 cortex_a9_mult16, cortex_a9_mac16, cortex_a9_mac, cortex_a9_store3_4, cortex_a9_load3_4, |
171 cortex_a9_multiply_long") | |
160 | 172 |
161 (define_bypass 2 "cortex_a9_dp_shift" | 173 (define_bypass 2 "cortex_a9_dp_shift" |
162 "cortex_a9_dp_shift, cortex_a9_multiply, | 174 "cortex_a9_dp_shift, cortex_a9_multiply, |
163 cortex_a9_load1_2, cortex_a9_dp, cortex_a9_store1_2, | 175 cortex_a9_load1_2, cortex_a9_dp, cortex_a9_store1_2, |
164 cortex_a9_mult16, cortex_a9_mac16, cortex_a9_mac, cortex_a9_store3_4, cortex_a9_load3_4") | 176 cortex_a9_mult16, cortex_a9_mac16, cortex_a9_mac, cortex_a9_store3_4, cortex_a9_load3_4, |
177 cortex_a9_multiply_long") | |
165 | 178 |
166 ;; An instruction in the load store pipeline can provide | 179 ;; An instruction in the load store pipeline can provide |
167 ;; read access to a DP instruction in the P0 default pipeline | 180 ;; read access to a DP instruction in the P0 default pipeline |
168 ;; before the writeback stage. | 181 ;; before the writeback stage. |
169 | 182 |
192 | 205 |
193 | 206 |
194 ;; Pipelining for VFP instructions. | 207 ;; Pipelining for VFP instructions. |
195 ;; Issue happens either along load store unit or the VFP / Neon unit. | 208 ;; Issue happens either along load store unit or the VFP / Neon unit. |
196 ;; Pipeline Instruction Classification. | 209 ;; Pipeline Instruction Classification. |
197 ;; FPS - fcpys, ffariths, ffarithd,r_2_f,f_2_r | 210 ;; FPS - fmov, ffariths, ffarithd,f_mcr,f_mcrr,f_mrc,f_mrrc |
198 ;; FP_ADD - fadds, faddd, fcmps (1) | 211 ;; FP_ADD - fadds, faddd, fcmps (1) |
199 ;; FPMUL - fmul{s,d}, fmac{s,d} | 212 ;; FPMUL - fmul{s,d}, fmac{s,d}, ffma{s,d} |
200 ;; FPDIV - fdiv{s,d} | 213 ;; FPDIV - fdiv{s,d} |
201 (define_cpu_unit "ca9fps" "cortex_a9") | 214 (define_cpu_unit "ca9fps" "cortex_a9") |
202 (define_cpu_unit "ca9fp_add1, ca9fp_add2, ca9fp_add3, ca9fp_add4" "cortex_a9") | 215 (define_cpu_unit "ca9fp_add1, ca9fp_add2, ca9fp_add3, ca9fp_add4" "cortex_a9") |
203 (define_cpu_unit "ca9fp_mul1, ca9fp_mul2 , ca9fp_mul3, ca9fp_mul4" "cortex_a9") | 216 (define_cpu_unit "ca9fp_mul1, ca9fp_mul2 , ca9fp_mul3, ca9fp_mul4" "cortex_a9") |
204 (define_cpu_unit "ca9fp_ds1" "cortex_a9") | 217 (define_cpu_unit "ca9fp_ds1" "cortex_a9") |
205 | 218 |
206 | 219 |
207 ;; fmrs, fmrrd, fmstat and fmrx - The data is available after 1 cycle. | 220 ;; fmrs, fmrrd, fmstat and fmrx - The data is available after 1 cycle. |
208 (define_insn_reservation "cortex_a9_fps" 2 | 221 (define_insn_reservation "cortex_a9_fps" 2 |
209 (and (eq_attr "tune" "cortexa9") | 222 (and (eq_attr "tune" "cortexa9") |
210 (eq_attr "type" "fcpys, fconsts, fconstd, ffariths, ffarithd, r_2_f, f_2_r, f_flag")) | 223 (eq_attr "type" "fmov, fconsts, fconstd, ffariths, ffarithd,\ |
224 f_mcr, f_mcrr, f_mrc, f_mrrc, f_flag")) | |
211 "ca9_issue_vfp_neon + ca9fps") | 225 "ca9_issue_vfp_neon + ca9fps") |
212 | 226 |
213 (define_bypass 1 | 227 (define_bypass 1 |
214 "cortex_a9_fps" | 228 "cortex_a9_fps" |
215 "cortex_a9_fadd, cortex_a9_fps, cortex_a9_fcmp, cortex_a9_dp, cortex_a9_dp_shift, cortex_a9_multiply") | 229 "cortex_a9_fadd, cortex_a9_fps, cortex_a9_fcmp, cortex_a9_dp, cortex_a9_dp_shift, cortex_a9_multiply, cortex_a9_multiply_long") |
216 | 230 |
217 ;; Scheduling on the FP_ADD pipeline. | 231 ;; Scheduling on the FP_ADD pipeline. |
218 (define_reservation "ca9fp_add" "ca9_issue_vfp_neon + ca9fp_add1, ca9fp_add2, ca9fp_add3, ca9fp_add4") | 232 (define_reservation "ca9fp_add" "ca9_issue_vfp_neon + ca9fp_add1, ca9fp_add2, ca9fp_add3, ca9fp_add4") |
219 | 233 |
220 (define_insn_reservation "cortex_a9_fadd" 4 | 234 (define_insn_reservation "cortex_a9_fadd" 4 |
221 (and (eq_attr "tune" "cortexa9") | 235 (and (eq_attr "tune" "cortexa9") |
222 (eq_attr "type" "fadds, faddd, f_cvt")) | 236 (eq_attr "type" "fadds, faddd, f_cvt, f_cvtf2i, f_cvti2f")) |
223 "ca9fp_add") | 237 "ca9fp_add") |
224 | 238 |
225 (define_insn_reservation "cortex_a9_fcmp" 1 | 239 (define_insn_reservation "cortex_a9_fcmp" 1 |
226 (and (eq_attr "tune" "cortexa9") | 240 (and (eq_attr "tune" "cortexa9") |
227 (eq_attr "type" "fcmps, fcmpd")) | 241 (eq_attr "type" "fcmps, fcmpd")) |
244 (eq_attr "type" "fmuld")) | 258 (eq_attr "type" "fmuld")) |
245 "ca9fmuld") | 259 "ca9fmuld") |
246 | 260 |
247 (define_insn_reservation "cortex_a9_fmacs" 8 | 261 (define_insn_reservation "cortex_a9_fmacs" 8 |
248 (and (eq_attr "tune" "cortexa9") | 262 (and (eq_attr "tune" "cortexa9") |
249 (eq_attr "type" "fmacs")) | 263 (eq_attr "type" "fmacs,ffmas")) |
250 "ca9fmuls, ca9fp_add") | 264 "ca9fmuls, ca9fp_add") |
251 | 265 |
252 (define_insn_reservation "cortex_a9_fmacd" 9 | 266 (define_insn_reservation "cortex_a9_fmacd" 9 |
253 (and (eq_attr "tune" "cortexa9") | 267 (and (eq_attr "tune" "cortexa9") |
254 (eq_attr "type" "fmacd")) | 268 (eq_attr "type" "fmacd,ffmad")) |
255 "ca9fmuld, ca9fp_add") | 269 "ca9fmuld, ca9fp_add") |
256 | 270 |
257 ;; Division pipeline description. | 271 ;; Division pipeline description. |
258 (define_insn_reservation "cortex_a9_fdivs" 15 | 272 (define_insn_reservation "cortex_a9_fdivs" 15 |
259 (and (eq_attr "tune" "cortexa9") | 273 (and (eq_attr "tune" "cortexa9") |
260 (eq_attr "type" "fdivs")) | 274 (eq_attr "type" "fdivs, fsqrts")) |
261 "ca9fp_ds1 + ca9_issue_vfp_neon, nothing*14") | 275 "ca9fp_ds1 + ca9_issue_vfp_neon, nothing*14") |
262 | 276 |
263 (define_insn_reservation "cortex_a9_fdivd" 25 | 277 (define_insn_reservation "cortex_a9_fdivd" 25 |
264 (and (eq_attr "tune" "cortexa9") | 278 (and (eq_attr "tune" "cortexa9") |
265 (eq_attr "type" "fdivd")) | 279 (eq_attr "type" "fdivd, fsqrtd")) |
266 "ca9fp_ds1 + ca9_issue_vfp_neon, nothing*24") | 280 "ca9fp_ds1 + ca9_issue_vfp_neon, nothing*24") |
267 | 281 |
268 ;; Include Neon pipeline description | 282 ;; Include Neon pipeline description |
269 (include "cortex-a9-neon.md") | 283 (include "cortex-a9-neon.md") |