| 1 |
12 |
jlechner |
;; AMD K6/K6-2 Scheduling
|
| 2 |
|
|
;; Copyright (C) 2002, 2004
|
| 3 |
|
|
;; Free Software Foundation, Inc.
|
| 4 |
|
|
;;
|
| 5 |
|
|
;; This file is part of GCC.
|
| 6 |
|
|
;;
|
| 7 |
|
|
;; GCC is free software; you can redistribute it and/or modify
|
| 8 |
|
|
;; it under the terms of the GNU General Public License as published by
|
| 9 |
|
|
;; the Free Software Foundation; either version 2, or (at your option)
|
| 10 |
|
|
;; any later version.
|
| 11 |
|
|
;;
|
| 12 |
|
|
;; GCC is distributed in the hope that it will be useful,
|
| 13 |
|
|
;; but WITHOUT ANY WARRANTY; without even the implied warranty of
|
| 14 |
|
|
;; MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
| 15 |
|
|
;; GNU General Public 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 COPYING. If not, write to
|
| 19 |
|
|
;; the Free Software Foundation, 51 Franklin Street, Fifth Floor,
|
| 20 |
|
|
;; Boston, MA 02110-1301, USA.
|
| 21 |
|
|
;;
|
| 22 |
|
|
;; The K6 architecture is quite similar to PPro. Important difference is
|
| 23 |
|
|
;; that there are only two decoders and they seems to be much slower than
|
| 24 |
|
|
;; any of the execution units. So we have to pay much more attention to
|
| 25 |
|
|
;; proper scheduling for the decoders.
|
| 26 |
|
|
;; FIXME: We don't do that right now. A good start would be to sort the
|
| 27 |
|
|
;; instructions based on length.
|
| 28 |
|
|
;;
|
| 29 |
|
|
;; This description is based on data from the following documents:
|
| 30 |
|
|
;;
|
| 31 |
|
|
;; "AMD-K6 Processor Data Sheet (Preliminary information)"
|
| 32 |
|
|
;; Advanced Micro Devices, Inc., 1998.
|
| 33 |
|
|
;;
|
| 34 |
|
|
;; "AMD-K6 Processor Code Optimization Application Note"
|
| 35 |
|
|
;; Advanced Micro Devices, Inc., 2000.
|
| 36 |
|
|
;;
|
| 37 |
|
|
;; CPU execution units of the K6:
|
| 38 |
|
|
;;
|
| 39 |
|
|
;; store describes the Store unit. This unit is not modelled
|
| 40 |
|
|
;; completely and it is only used to model lea operation.
|
| 41 |
|
|
;; Otherwise it lies outside of any critical path.
|
| 42 |
|
|
;; load describes the Load unit
|
| 43 |
|
|
;; alux describes the Integer X unit
|
| 44 |
|
|
;; mm describes the Multimedia unit, which shares a pipe
|
| 45 |
|
|
;; with the Integer X unit. This unit is used for MMX,
|
| 46 |
|
|
;; which is not implemented for K6.
|
| 47 |
|
|
;; aluy describes the Integer Y unit
|
| 48 |
|
|
;; fpu describes the FPU unit
|
| 49 |
|
|
;; branch describes the Branch unit
|
| 50 |
|
|
;;
|
| 51 |
|
|
;; The fp unit is not pipelined, and it can only do one operation per two
|
| 52 |
|
|
;; cycles, including fxcg.
|
| 53 |
|
|
;;
|
| 54 |
|
|
;; Generally this is a very poor description, but at least no worse than
|
| 55 |
|
|
;; the old description, and a lot easier to extend to something more
|
| 56 |
|
|
;; reasonable if anyone still cares enough about this architecture in 2004.
|
| 57 |
|
|
;;
|
| 58 |
|
|
;; ??? fxch isn't handled; not an issue until sched3 after reg-stack is real.
|
| 59 |
|
|
|
| 60 |
|
|
(define_automaton "k6_decoder,k6_load_unit,k6_store_unit,k6_integer_units,k6_fpu_unit,k6_branch_unit")
|
| 61 |
|
|
|
| 62 |
|
|
;; The K6 instruction decoding begins before the on-chip instruction cache is
|
| 63 |
|
|
;; filled. Depending on the length of the instruction, two simple instructions
|
| 64 |
|
|
;; can be decoded in two parallel short decoders, or one complex instruction can
|
| 65 |
|
|
;; be decoded in either the long or the vector decoder. For all practical
|
| 66 |
|
|
;; purposes, the long and vector decoder can be modelled as one decoder.
|
| 67 |
|
|
(define_cpu_unit "k6_decode_short0" "k6_decoder")
|
| 68 |
|
|
(define_cpu_unit "k6_decode_short1" "k6_decoder")
|
| 69 |
|
|
(define_cpu_unit "k6_decode_long" "k6_decoder")
|
| 70 |
|
|
(exclusion_set "k6_decode_long" "k6_decode_short0,k6_decode_short1")
|
| 71 |
|
|
(define_reservation "k6_decode_short" "k6_decode_short0|k6_decode_short1")
|
| 72 |
|
|
(define_reservation "k6_decode_vector" "k6_decode_long")
|
| 73 |
|
|
|
| 74 |
|
|
(define_cpu_unit "k6_store" "k6_store_unit")
|
| 75 |
|
|
(define_cpu_unit "k6_load" "k6_load_unit")
|
| 76 |
|
|
(define_cpu_unit "k6_alux,k6_aluy" "k6_integer_units")
|
| 77 |
|
|
(define_cpu_unit "k6_fpu" "k6_fpu_unit")
|
| 78 |
|
|
(define_cpu_unit "k6_branch" "k6_branch_unit")
|
| 79 |
|
|
|
| 80 |
|
|
;; Shift instructions and certain arithmetic are issued only on Integer X.
|
| 81 |
|
|
(define_insn_reservation "k6_alux_only" 1
|
| 82 |
|
|
(and (eq_attr "cpu" "k6")
|
| 83 |
|
|
(and (eq_attr "type" "ishift,ishift1,rotate,rotate1,alu1,negnot,cld")
|
| 84 |
|
|
(eq_attr "memory" "none")))
|
| 85 |
|
|
"k6_decode_short,k6_alux")
|
| 86 |
|
|
|
| 87 |
|
|
(define_insn_reservation "k6_alux_only_load" 3
|
| 88 |
|
|
(and (eq_attr "cpu" "k6")
|
| 89 |
|
|
(and (eq_attr "type" "ishift,ishift1,rotate,rotate1,alu1,negnot,cld")
|
| 90 |
|
|
(eq_attr "memory" "load")))
|
| 91 |
|
|
"k6_decode_short,k6_load,k6_alux")
|
| 92 |
|
|
|
| 93 |
|
|
(define_insn_reservation "k6_alux_only_store" 3
|
| 94 |
|
|
(and (eq_attr "cpu" "k6")
|
| 95 |
|
|
(and (eq_attr "type" "ishift,ishift1,rotate,rotate1,alu1,negnot,cld")
|
| 96 |
|
|
(eq_attr "memory" "store,both,unknown")))
|
| 97 |
|
|
"k6_decode_long,k6_load,k6_alux,k6_store")
|
| 98 |
|
|
|
| 99 |
|
|
;; Integer divide and multiply can only be issued on Integer X, too.
|
| 100 |
|
|
(define_insn_reservation "k6_alu_imul" 2
|
| 101 |
|
|
(and (eq_attr "cpu" "k6")
|
| 102 |
|
|
(eq_attr "type" "imul"))
|
| 103 |
|
|
"k6_decode_vector,k6_alux*3")
|
| 104 |
|
|
|
| 105 |
|
|
(define_insn_reservation "k6_alu_imul_load" 4
|
| 106 |
|
|
(and (eq_attr "cpu" "k6")
|
| 107 |
|
|
(and (eq_attr "type" "imul")
|
| 108 |
|
|
(eq_attr "memory" "load")))
|
| 109 |
|
|
"k6_decode_vector,k6_load,k6_alux*3")
|
| 110 |
|
|
|
| 111 |
|
|
(define_insn_reservation "k6_alu_imul_store" 4
|
| 112 |
|
|
(and (eq_attr "cpu" "k6")
|
| 113 |
|
|
(and (eq_attr "type" "imul")
|
| 114 |
|
|
(eq_attr "memory" "store,both,unknown")))
|
| 115 |
|
|
"k6_decode_vector,k6_load,k6_alux*3,k6_store")
|
| 116 |
|
|
|
| 117 |
|
|
;; ??? Guessed latencies based on the old pipeline description.
|
| 118 |
|
|
(define_insn_reservation "k6_alu_idiv" 17
|
| 119 |
|
|
(and (eq_attr "cpu" "k6")
|
| 120 |
|
|
(and (eq_attr "type" "idiv")
|
| 121 |
|
|
(eq_attr "memory" "none")))
|
| 122 |
|
|
"k6_decode_vector,k6_alux*17")
|
| 123 |
|
|
|
| 124 |
|
|
(define_insn_reservation "k6_alu_idiv_mem" 19
|
| 125 |
|
|
(and (eq_attr "cpu" "k6")
|
| 126 |
|
|
(and (eq_attr "type" "idiv")
|
| 127 |
|
|
(eq_attr "memory" "!none")))
|
| 128 |
|
|
"k6_decode_vector,k6_load,k6_alux*17")
|
| 129 |
|
|
|
| 130 |
|
|
;; Basic word and doubleword ALU ops can be issued on both Integer units.
|
| 131 |
|
|
(define_insn_reservation "k6_alu" 1
|
| 132 |
|
|
(and (eq_attr "cpu" "k6")
|
| 133 |
|
|
(and (eq_attr "type" "alu,alu1,negnot,icmp,test,imovx,incdec,setcc")
|
| 134 |
|
|
(eq_attr "memory" "none")))
|
| 135 |
|
|
"k6_decode_short,k6_alux|k6_aluy")
|
| 136 |
|
|
|
| 137 |
|
|
(define_insn_reservation "k6_alu_load" 3
|
| 138 |
|
|
(and (eq_attr "cpu" "k6")
|
| 139 |
|
|
(and (eq_attr "type" "alu,alu1,negnot,icmp,test,imovx,incdec,setcc")
|
| 140 |
|
|
(eq_attr "memory" "load")))
|
| 141 |
|
|
"k6_decode_short,k6_load,k6_alux|k6_aluy")
|
| 142 |
|
|
|
| 143 |
|
|
(define_insn_reservation "k6_alu_store" 3
|
| 144 |
|
|
(and (eq_attr "cpu" "k6")
|
| 145 |
|
|
(and (eq_attr "type" "alu,alu1,negnot,icmp,test,imovx,incdec,setcc")
|
| 146 |
|
|
(eq_attr "memory" "store,both,unknown")))
|
| 147 |
|
|
"k6_decode_long,k6_load,k6_alux|k6_aluy,k6_store")
|
| 148 |
|
|
|
| 149 |
|
|
;; A "load immediate" operation does not require execution at all,
|
| 150 |
|
|
;; it is available immediately after decoding. Special-case this.
|
| 151 |
|
|
(define_insn_reservation "k6_alu_imov" 1
|
| 152 |
|
|
(and (eq_attr "cpu" "k6")
|
| 153 |
|
|
(and (eq_attr "type" "imov")
|
| 154 |
|
|
(and (eq_attr "memory" "none")
|
| 155 |
|
|
(match_operand 1 "nonimmediate_operand"))))
|
| 156 |
|
|
"k6_decode_short,k6_alux|k6_aluy")
|
| 157 |
|
|
|
| 158 |
|
|
(define_insn_reservation "k6_alu_imov_imm" 0
|
| 159 |
|
|
(and (eq_attr "cpu" "k6")
|
| 160 |
|
|
(and (eq_attr "type" "imov")
|
| 161 |
|
|
(and (eq_attr "memory" "none")
|
| 162 |
|
|
(match_operand 1 "immediate_operand"))))
|
| 163 |
|
|
"k6_decode_short")
|
| 164 |
|
|
|
| 165 |
|
|
(define_insn_reservation "k6_alu_imov_load" 2
|
| 166 |
|
|
(and (eq_attr "cpu" "k6")
|
| 167 |
|
|
(and (eq_attr "type" "imov")
|
| 168 |
|
|
(eq_attr "memory" "load")))
|
| 169 |
|
|
"k6_decode_short,k6_load")
|
| 170 |
|
|
|
| 171 |
|
|
(define_insn_reservation "k6_alu_imov_store" 1
|
| 172 |
|
|
(and (eq_attr "cpu" "k6")
|
| 173 |
|
|
(and (eq_attr "type" "imov")
|
| 174 |
|
|
(eq_attr "memory" "store")))
|
| 175 |
|
|
"k6_decode_short,k6_store")
|
| 176 |
|
|
|
| 177 |
|
|
(define_insn_reservation "k6_alu_imov_both" 2
|
| 178 |
|
|
(and (eq_attr "cpu" "k6")
|
| 179 |
|
|
(and (eq_attr "type" "imov")
|
| 180 |
|
|
(eq_attr "memory" "both,unknown")))
|
| 181 |
|
|
"k6_decode_long,k6_load,k6_alux|k6_aluy")
|
| 182 |
|
|
|
| 183 |
|
|
;; The branch unit.
|
| 184 |
|
|
(define_insn_reservation "k6_branch_call" 1
|
| 185 |
|
|
(and (eq_attr "cpu" "k6")
|
| 186 |
|
|
(eq_attr "type" "call,callv"))
|
| 187 |
|
|
"k6_decode_vector,k6_branch")
|
| 188 |
|
|
|
| 189 |
|
|
(define_insn_reservation "k6_branch_branch" 1
|
| 190 |
|
|
(and (eq_attr "cpu" "k6")
|
| 191 |
|
|
(eq_attr "type" "ibr"))
|
| 192 |
|
|
"k6_decode_short,k6_branch")
|
| 193 |
|
|
|
| 194 |
|
|
;; The load and units have two pipeline stages. The load latency is
|
| 195 |
|
|
;; two cycles.
|
| 196 |
|
|
(define_insn_reservation "k6_load_pop" 3
|
| 197 |
|
|
(and (eq_attr "cpu" "k6")
|
| 198 |
|
|
(ior (eq_attr "type" "pop")
|
| 199 |
|
|
(eq_attr "memory" "load,both")))
|
| 200 |
|
|
"k6_decode_short,k6_load")
|
| 201 |
|
|
|
| 202 |
|
|
(define_insn_reservation "k6_load_leave" 5
|
| 203 |
|
|
(and (eq_attr "cpu" "k6")
|
| 204 |
|
|
(eq_attr "type" "leave"))
|
| 205 |
|
|
"k6_decode_long,k6_load,(k6_alux|k6_aluy)*2")
|
| 206 |
|
|
|
| 207 |
|
|
;; ??? From the old pipeline description. Egad!
|
| 208 |
|
|
;; ??? Apparently we take care of this reservation in adjust_cost.
|
| 209 |
|
|
(define_insn_reservation "k6_load_str" 10
|
| 210 |
|
|
(and (eq_attr "cpu" "k6")
|
| 211 |
|
|
(and (eq_attr "type" "str")
|
| 212 |
|
|
(eq_attr "memory" "load,both")))
|
| 213 |
|
|
"k6_decode_vector,k6_load*10")
|
| 214 |
|
|
|
| 215 |
|
|
;; The store unit handles lea and push. It is otherwise unmodelled.
|
| 216 |
|
|
(define_insn_reservation "k6_store_lea" 2
|
| 217 |
|
|
(and (eq_attr "cpu" "k6")
|
| 218 |
|
|
(eq_attr "type" "lea"))
|
| 219 |
|
|
"k6_decode_short,k6_store,k6_alux|k6_aluy")
|
| 220 |
|
|
|
| 221 |
|
|
(define_insn_reservation "k6_store_push" 2
|
| 222 |
|
|
(and (eq_attr "cpu" "k6")
|
| 223 |
|
|
(ior (eq_attr "type" "push")
|
| 224 |
|
|
(eq_attr "memory" "store,both")))
|
| 225 |
|
|
"k6_decode_short,k6_store")
|
| 226 |
|
|
|
| 227 |
|
|
(define_insn_reservation "k6_store_str" 10
|
| 228 |
|
|
(and (eq_attr "cpu" "k6")
|
| 229 |
|
|
(eq_attr "type" "str"))
|
| 230 |
|
|
"k6_store*10")
|
| 231 |
|
|
|
| 232 |
|
|
;; Most FPU instructions have latency 2 and throughput 2.
|
| 233 |
|
|
(define_insn_reservation "k6_fpu" 2
|
| 234 |
|
|
(and (eq_attr "cpu" "k6")
|
| 235 |
|
|
(and (eq_attr "type" "fop,fmov,fcmp,fistp")
|
| 236 |
|
|
(eq_attr "memory" "none")))
|
| 237 |
|
|
"k6_decode_vector,k6_fpu*2")
|
| 238 |
|
|
|
| 239 |
|
|
(define_insn_reservation "k6_fpu_load" 6
|
| 240 |
|
|
(and (eq_attr "cpu" "k6")
|
| 241 |
|
|
(and (eq_attr "type" "fop,fmov,fcmp,fistp")
|
| 242 |
|
|
(eq_attr "memory" "load,both")))
|
| 243 |
|
|
"k6_decode_short,k6_load,k6_fpu*2")
|
| 244 |
|
|
|
| 245 |
|
|
(define_insn_reservation "k6_fpu_store" 6
|
| 246 |
|
|
(and (eq_attr "cpu" "k6")
|
| 247 |
|
|
(and (eq_attr "type" "fop,fmov,fcmp,fistp")
|
| 248 |
|
|
(eq_attr "memory" "store")))
|
| 249 |
|
|
"k6_decode_short,k6_store,k6_fpu*2")
|
| 250 |
|
|
|
| 251 |
|
|
(define_insn_reservation "k6_fpu_fmul" 2
|
| 252 |
|
|
(and (eq_attr "cpu" "k6")
|
| 253 |
|
|
(and (eq_attr "type" "fmul")
|
| 254 |
|
|
(eq_attr "memory" "none")))
|
| 255 |
|
|
"k6_decode_short,k6_fpu*2")
|
| 256 |
|
|
|
| 257 |
|
|
(define_insn_reservation "k6_fpu_fmul_load" 2
|
| 258 |
|
|
(and (eq_attr "cpu" "k6")
|
| 259 |
|
|
(and (eq_attr "type" "fmul")
|
| 260 |
|
|
(eq_attr "memory" "load,both")))
|
| 261 |
|
|
"k6_decode_short,k6_load,k6_fpu*2")
|
| 262 |
|
|
|
| 263 |
|
|
;; ??? Guessed latencies from the old pipeline description.
|
| 264 |
|
|
(define_insn_reservation "k6_fpu_expensive" 56
|
| 265 |
|
|
(and (eq_attr "cpu" "k6")
|
| 266 |
|
|
(eq_attr "type" "fdiv,fpspc"))
|
| 267 |
|
|
"k6_decode_short,k6_fpu*56")
|
| 268 |
|
|
|
