gnu/dist/gcc4/gcc/config/rs6000/power5.md

   1 ;; Scheduling description for IBM POWER5 processor.
   2 ;;   Copyright (C) 2003, 2004 Free Software Foundation, Inc.
   3 ;;
   4 ;; This file is part of GCC.
   5 ;;
   6 ;; GCC is free software; you can redistribute it and/or modify it
   7 ;; under the terms of the GNU General Public License as published
   8 ;; by the Free Software Foundation; either version 2, or (at your
   9 ;; option) any later version.
  10 ;;
  11 ;; GCC is distributed in the hope that it will be useful, but WITHOUT
  12 ;; ANY WARRANTY; without even the implied warranty of MERCHANTABILITY
  13 ;; or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public
  14 ;; License for more details.
  15 ;;
  16 ;; You should have received a copy of the GNU General Public License
  17 ;; along with GCC; see the file COPYING.  If not, write to the
  18 ;; Free Software Foundation, 51 Franklin Street, Fifth Floor, Boston,
  19 ;; MA 02110-1301, USA.
  20
  21 ;; Sources: IBM Red Book and White Paper on POWER5
  22
  23 ;; The POWER5 has 2 iu, 2 fpu, 2 lsu per engine (2 engines per chip).
  24 ;; Instructions that update more than one register get broken into two
  25 ;; (split) or more internal ops.  The chip can issue up to 5
  26 ;; internal ops per cycle.
  27
  28 (define_automaton "power5iu,power5fpu,power5misc")
  29
  30 (define_cpu_unit "iu1_power5,iu2_power5" "power5iu")
  31 (define_cpu_unit "lsu1_power5,lsu2_power5" "power5misc")
  32 (define_cpu_unit "fpu1_power5,fpu2_power5" "power5fpu")
  33 (define_cpu_unit "bpu_power5,cru_power5" "power5misc")
  34 (define_cpu_unit "du1_power5,du2_power5,du3_power5,du4_power5,du5_power5"
  35                  "power5misc")
  36
  37 (define_reservation "lsq_power5"
  38                     "(du1_power5,lsu1_power5)\
  39                     |(du2_power5,lsu2_power5)\
  40                     |(du3_power5,lsu2_power5)\
  41                     |(du4_power5,lsu1_power5)")
  42
  43 (define_reservation "iq_power5"
  44                     "(du1_power5,iu1_power5)\
  45                     |(du2_power5,iu2_power5)\
  46                     |(du3_power5,iu2_power5)\
  47                     |(du4_power5,iu1_power5)")
  48
  49 (define_reservation "fpq_power5"
  50                     "(du1_power5,fpu1_power5)\
  51                     |(du2_power5,fpu2_power5)\
  52                     |(du3_power5,fpu2_power5)\
  53                     |(du4_power5,fpu1_power5)")
  54
  55 ; Dispatch slots are allocated in order conforming to program order.
  56 (absence_set "du1_power5" "du2_power5,du3_power5,du4_power5,du5_power5")
  57 (absence_set "du2_power5" "du3_power5,du4_power5,du5_power5")
  58 (absence_set "du3_power5" "du4_power5,du5_power5")
  59 (absence_set "du4_power5" "du5_power5")
  60
  61
  62 ; Load/store
  63 (define_insn_reservation "power5-load" 4 ; 3
  64   (and (eq_attr "type" "load")
  65        (eq_attr "cpu" "power5"))
  66   "lsq_power5")
  67
  68 (define_insn_reservation "power5-load-ext" 5
  69   (and (eq_attr "type" "load_ext")
  70        (eq_attr "cpu" "power5"))
  71   "du1_power5+du2_power5,lsu1_power5,nothing,nothing,iu2_power5")
  72
  73 (define_insn_reservation "power5-load-ext-update" 5
  74   (and (eq_attr "type" "load_ext_u")
  75        (eq_attr "cpu" "power5"))
  76   "du1_power5+du2_power5+du3_power5+du4_power5,\
  77    lsu1_power5+iu2_power5,nothing,nothing,iu2_power5")
  78
  79 (define_insn_reservation "power5-load-ext-update-indexed" 5
  80   (and (eq_attr "type" "load_ext_ux")
  81        (eq_attr "cpu" "power5"))
  82   "du1_power5+du2_power5+du3_power5+du4_power5,\
  83    iu1_power5,lsu2_power5+iu1_power5,nothing,nothing,iu2_power5")
  84
  85 (define_insn_reservation "power5-load-update-indexed" 3
  86   (and (eq_attr "type" "load_ux")
  87        (eq_attr "cpu" "power5"))
  88   "du1_power5+du2_power5+du3_power5+du4_power5,\
  89    iu1_power5,lsu2_power5+iu2_power5")
  90
  91 (define_insn_reservation "power5-load-update" 4 ; 3
  92   (and (eq_attr "type" "load_u")
  93        (eq_attr "cpu" "power5"))
  94   "du1_power5+du2_power5,lsu1_power5+iu2_power5")
  95
  96 (define_insn_reservation "power5-fpload" 6 ; 5
  97   (and (eq_attr "type" "fpload")
  98        (eq_attr "cpu" "power5"))
  99   "lsq_power5")
 100
 101 (define_insn_reservation "power5-fpload-update" 6 ; 5
 102   (and (eq_attr "type" "fpload_u,fpload_ux")
 103        (eq_attr "cpu" "power5"))
 104   "du1_power5+du2_power5,lsu1_power5+iu2_power5")
 105
 106 (define_insn_reservation "power5-store" 12
 107   (and (eq_attr "type" "store")
 108        (eq_attr "cpu" "power5"))
 109   "(du1_power5,lsu1_power5,iu1_power5)\
 110   |(du2_power5,lsu2_power5,iu2_power5)\
 111   |(du3_power5,lsu2_power5,iu2_power5)\
 112   |(du4_power5,lsu1_power5,iu1_power5)")
 113
 114 (define_insn_reservation "power5-store-update" 12
 115   (and (eq_attr "type" "store_u")
 116        (eq_attr "cpu" "power5"))
 117   "du1_power5+du2_power5,lsu1_power5+iu2_power5,iu1_power5")
 118
 119 (define_insn_reservation "power5-store-update-indexed" 12
 120   (and (eq_attr "type" "store_ux")
 121        (eq_attr "cpu" "power5"))
 122    "du1_power5+du2_power5+du3_power5+du4_power5,\
 123     iu1_power5,lsu2_power5+iu2_power5,iu2_power5")
 124
 125 (define_insn_reservation "power5-fpstore" 12
 126   (and (eq_attr "type" "fpstore")
 127        (eq_attr "cpu" "power5"))
 128   "(du1_power5,lsu1_power5,fpu1_power5)\
 129   |(du2_power5,lsu2_power5,fpu2_power5)\
 130   |(du3_power5,lsu2_power5,fpu2_power5)\
 131   |(du4_power5,lsu1_power5,fpu1_power5)")
 132
 133 (define_insn_reservation "power5-fpstore-update" 12
 134   (and (eq_attr "type" "fpstore_u,fpstore_ux")
 135        (eq_attr "cpu" "power5"))
 136   "du1_power5+du2_power5,lsu1_power5+iu2_power5,fpu1_power5")
 137
 138 (define_insn_reservation "power5-llsc" 11
 139   (and (eq_attr "type" "load_l,store_c,sync")
 140        (eq_attr "cpu" "power5"))
 141   "du1_power5+du2_power5+du3_power5+du4_power5,\
 142   lsu1_power5")
 143
 144
 145 ; Integer latency is 2 cycles
 146 (define_insn_reservation "power5-integer" 2
 147   (and (eq_attr "type" "integer")
 148        (eq_attr "cpu" "power5"))
 149   "iq_power5")
 150
 151 (define_insn_reservation "power5-two" 2
 152   (and (eq_attr "type" "two")
 153        (eq_attr "cpu" "power5"))
 154   "(du1_power5+du2_power5,iu1_power5,nothing,iu2_power5)\
 155   |(du2_power5+du3_power5,iu2_power5,nothing,iu2_power5)\
 156   |(du3_power5+du4_power5,iu2_power5,nothing,iu1_power5)\
 157   |(du4_power5+du1_power5,iu1_power5,nothing,iu1_power5)")
 158
 159 (define_insn_reservation "power5-three" 2
 160   (and (eq_attr "type" "three")
 161        (eq_attr "cpu" "power5"))
 162   "(du1_power5+du2_power5+du3_power5,\
 163     iu1_power5,nothing,iu2_power5,nothing,iu2_power5)\
 164   |(du2_power5+du3_power5+du4_power5,\
 165     iu2_power5,nothing,iu2_power5,nothing,iu1_power5)\
 166   |(du3_power5+du4_power5+du1_power5,\
 167     iu2_power5,nothing,iu1_power5,nothing,iu1_power5)\
 168   |(du4_power5+du1_power5+du2_power5,\
 169     iu1_power5,nothing,iu2_power5,nothing,iu2_power5)")
 170
 171 (define_insn_reservation "power5-insert" 4
 172   (and (eq_attr "type" "insert_word")
 173        (eq_attr "cpu" "power5"))
 174   "du1_power5+du2_power5,iu1_power5,nothing,iu2_power5")
 175
 176 (define_insn_reservation "power5-cmp" 3
 177   (and (eq_attr "type" "cmp,fast_compare")
 178        (eq_attr "cpu" "power5"))
 179   "iq_power5")
 180
 181 (define_insn_reservation "power5-compare" 2
 182   (and (eq_attr "type" "compare,delayed_compare")
 183        (eq_attr "cpu" "power5"))
 184   "du1_power5+du2_power5,iu1_power5,iu2_power5")
 185
 186 (define_bypass 4 "power5-compare" "power5-branch,power5-crlogical,power5-delayedcr,power5-mfcr,power5-mfcrf")
 187
 188 (define_insn_reservation "power5-lmul-cmp" 7
 189   (and (eq_attr "type" "lmul_compare")
 190        (eq_attr "cpu" "power5"))
 191   "du1_power5+du2_power5,iu1_power5*6,iu2_power5")
 192
 193 (define_bypass 10 "power5-lmul-cmp" "power5-branch,power5-crlogical,power5-delayedcr,power5-mfcr,power5-mfcrf")
 194
 195 (define_insn_reservation "power5-imul-cmp" 5
 196   (and (eq_attr "type" "imul_compare")
 197        (eq_attr "cpu" "power5"))
 198   "du1_power5+du2_power5,iu1_power5*4,iu2_power5")
 199
 200 (define_bypass 8 "power5-imul-cmp" "power5-branch,power5-crlogical,power5-delayedcr,power5-mfcr,power5-mfcrf")
 201
 202 (define_insn_reservation "power5-lmul" 7
 203   (and (eq_attr "type" "lmul")
 204        (eq_attr "cpu" "power5"))
 205   "(du1_power5,iu1_power5*6)\
 206   |(du2_power5,iu2_power5*6)\
 207   |(du3_power5,iu2_power5*6)\
 208   |(du4_power5,iu1_power5*6)")
 209
 210 (define_insn_reservation "power5-imul" 5
 211   (and (eq_attr "type" "imul")
 212        (eq_attr "cpu" "power5"))
 213   "(du1_power5,iu1_power5*4)\
 214   |(du2_power5,iu2_power5*4)\
 215   |(du3_power5,iu2_power5*4)\
 216   |(du4_power5,iu1_power5*4)")
 217
 218 (define_insn_reservation "power5-imul3" 4
 219   (and (eq_attr "type" "imul2,imul3")
 220        (eq_attr "cpu" "power5"))
 221   "(du1_power5,iu1_power5*3)\
 222   |(du2_power5,iu2_power5*3)\
 223   |(du3_power5,iu2_power5*3)\
 224   |(du4_power5,iu1_power5*3)")
 225
 226
 227 ; SPR move only executes in first IU.
 228 ; Integer division only executes in second IU.
 229 (define_insn_reservation "power5-idiv" 36
 230   (and (eq_attr "type" "idiv")
 231        (eq_attr "cpu" "power5"))
 232   "du1_power5+du2_power5,iu2_power5*35")
 233
 234 (define_insn_reservation "power5-ldiv" 68
 235   (and (eq_attr "type" "ldiv")
 236        (eq_attr "cpu" "power5"))
 237   "du1_power5+du2_power5,iu2_power5*67")
 238
 239
 240 (define_insn_reservation "power5-mtjmpr" 3
 241   (and (eq_attr "type" "mtjmpr,mfjmpr")
 242        (eq_attr "cpu" "power5"))
 243   "du1_power5,bpu_power5")
 244
 245
 246 ; Branches take dispatch Slot 4.  The presence_sets prevent other insn from
 247 ; grabbing previous dispatch slots once this is assigned.
 248 (define_insn_reservation "power5-branch" 2
 249   (and (eq_attr "type" "jmpreg,branch")
 250        (eq_attr "cpu" "power5"))
 251   "(du5_power5\
 252    |du4_power5+du5_power5\
 253    |du3_power5+du4_power5+du5_power5\
 254    |du2_power5+du3_power5+du4_power5+du5_power5\
 255    |du1_power5+du2_power5+du3_power5+du4_power5+du5_power5),bpu_power5")
 256
 257
 258 ; Condition Register logical ops are split if non-destructive (RT != RB)
 259 (define_insn_reservation "power5-crlogical" 2
 260   (and (eq_attr "type" "cr_logical")
 261        (eq_attr "cpu" "power5"))
 262   "du1_power5,cru_power5")
 263
 264 (define_insn_reservation "power5-delayedcr" 4
 265   (and (eq_attr "type" "delayed_cr")
 266        (eq_attr "cpu" "power5"))
 267   "du1_power5+du2_power5,cru_power5,cru_power5")
 268
 269 ; 4 mfcrf (each 3 cyc, 1/cyc) + 3 fxu
 270 (define_insn_reservation "power5-mfcr" 6
 271   (and (eq_attr "type" "mfcr")
 272        (eq_attr "cpu" "power5"))
 273   "du1_power5+du2_power5+du3_power5+du4_power5,\
 274    du1_power5+du2_power5+du3_power5+du4_power5+cru_power5,\
 275    cru_power5,cru_power5,cru_power5")
 276
 277 ; mfcrf (1 field)
 278 (define_insn_reservation "power5-mfcrf" 3
 279   (and (eq_attr "type" "mfcrf")
 280        (eq_attr "cpu" "power5"))
 281   "du1_power5,cru_power5")
 282
 283 ; mtcrf (1 field)
 284 (define_insn_reservation "power5-mtcr" 4
 285   (and (eq_attr "type" "mtcr")
 286        (eq_attr "cpu" "power5"))
 287   "du1_power5,iu1_power5")
 288
 289 ; Basic FP latency is 6 cycles
 290 (define_insn_reservation "power5-fp" 6
 291   (and (eq_attr "type" "fp,dmul")
 292        (eq_attr "cpu" "power5"))
 293   "fpq_power5")
 294
 295 (define_insn_reservation "power5-fpcompare" 5
 296   (and (eq_attr "type" "fpcompare")
 297        (eq_attr "cpu" "power5"))
 298   "fpq_power5")
 299
 300 (define_insn_reservation "power5-sdiv" 33
 301   (and (eq_attr "type" "sdiv,ddiv")
 302        (eq_attr "cpu" "power5"))
 303   "(du1_power5,fpu1_power5*28)\
 304   |(du2_power5,fpu2_power5*28)\
 305   |(du3_power5,fpu2_power5*28)\
 306   |(du4_power5,fpu1_power5*28)")
 307
 308 (define_insn_reservation "power5-sqrt" 40
 309   (and (eq_attr "type" "ssqrt,dsqrt")
 310        (eq_attr "cpu" "power5"))
 311   "(du1_power5,fpu1_power5*35)\
 312   |(du2_power5,fpu2_power5*35)\
 313   |(du3_power5,fpu2_power5*35)\
 314   |(du4_power5,fpu2_power5*35)")
 315
 316 (define_insn_reservation "power5-isync" 2
 317   (and (eq_attr "type" "isync")
 318        (eq_attr "cpu" "power5"))
 319   "du1_power5+du2_power5+du3_power5+du4_power5,\
 320   lsu1_power5")
 321