gcc/config/rs6000/power4.md

   1 ;; Scheduling description for IBM Power4 and PowerPC 970 processors.
   2 ;;   Copyright (C) 2003-2013 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 3, 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 COPYING3.  If not see
  18 ;; <http://www.gnu.org/licenses/>.
  19
  20 ;; Sources: IBM Red Book and White Paper on POWER4
  21
  22 ;; The POWER4 has 2 iu, 2 fpu, 2 lsu per engine (2 engines per chip).
  23 ;; Instructions that update more than one register get broken into two
  24 ;; (split) or more internal ops.  The chip can issue up to 5
  25 ;; internal ops per cycle.
  26
  27 (define_automaton "power4iu,power4fpu,power4vec,power4misc")
  28
  29 (define_cpu_unit "iu1_power4,iu2_power4" "power4iu")
  30 (define_cpu_unit "lsu1_power4,lsu2_power4" "power4misc")
  31 (define_cpu_unit "fpu1_power4,fpu2_power4" "power4fpu")
  32 (define_cpu_unit "bpu_power4,cru_power4" "power4misc")
  33 (define_cpu_unit "vec_power4,vecperm_power4" "power4vec")
  34 (define_cpu_unit "du1_power4,du2_power4,du3_power4,du4_power4,du5_power4"
  35                  "power4misc")
  36
  37 (define_reservation "lsq_power4"
  38                     "(du1_power4,lsu1_power4)\
  39                     |(du2_power4,lsu2_power4)\
  40                     |(du3_power4,lsu2_power4)\
  41                     |(du4_power4,lsu1_power4)")
  42
  43 (define_reservation "lsuq_power4"
  44                     "((du1_power4+du2_power4,lsu1_power4)\
  45                       |(du2_power4+du3_power4,lsu2_power4)\
  46                       |(du3_power4+du4_power4,lsu2_power4))\
  47                      +(nothing,iu2_power4|nothing,iu1_power4)")
  48
  49 (define_reservation "iq_power4"
  50                     "(du1_power4|du2_power4|du3_power4|du4_power4),\
  51                      (iu1_power4|iu2_power4)")
  52
  53 (define_reservation "fpq_power4"
  54                     "(du1_power4|du2_power4|du3_power4|du4_power4),\
  55                      (fpu1_power4|fpu2_power4)")
  56
  57 (define_reservation "vq_power4"
  58                     "(du1_power4,vec_power4)\
  59                     |(du2_power4,vec_power4)\
  60                     |(du3_power4,vec_power4)\
  61                     |(du4_power4,vec_power4)")
  62
  63 (define_reservation "vpq_power4"
  64                     "(du1_power4,vecperm_power4)\
  65                     |(du2_power4,vecperm_power4)\
  66                     |(du3_power4,vecperm_power4)\
  67                     |(du4_power4,vecperm_power4)")
  68
  69
  70 ; Dispatch slots are allocated in order conforming to program order.
  71 (absence_set "du1_power4" "du2_power4,du3_power4,du4_power4,du5_power4")
  72 (absence_set "du2_power4" "du3_power4,du4_power4,du5_power4")
  73 (absence_set "du3_power4" "du4_power4,du5_power4")
  74 (absence_set "du4_power4" "du5_power4")
  75
  76
  77 ; Load/store
  78 (define_insn_reservation "power4-load" 4 ; 3
  79   (and (eq_attr "type" "load")
  80        (eq_attr "cpu" "power4"))
  81   "lsq_power4")
  82
  83 (define_insn_reservation "power4-load-ext" 5
  84   (and (eq_attr "type" "load_ext")
  85        (eq_attr "cpu" "power4"))
  86   "(du1_power4+du2_power4,lsu1_power4\
  87     |du2_power4+du3_power4,lsu2_power4\
  88     |du3_power4+du4_power4,lsu2_power4),\
  89    nothing,nothing,\
  90    (iu2_power4|iu1_power4)")
  91
  92 (define_insn_reservation "power4-load-ext-update" 5
  93   (and (eq_attr "type" "load_ext_u")
  94        (eq_attr "cpu" "power4"))
  95   "du1_power4+du2_power4+du3_power4+du4_power4,\
  96    lsu1_power4+iu2_power4,nothing,nothing,iu2_power4")
  97
  98 (define_insn_reservation "power4-load-ext-update-indexed" 5
  99   (and (eq_attr "type" "load_ext_ux")
 100        (eq_attr "cpu" "power4"))
 101   "du1_power4+du2_power4+du3_power4+du4_power4,\
 102    iu1_power4,lsu2_power4+iu1_power4,nothing,nothing,iu2_power4")
 103
 104 (define_insn_reservation "power4-load-update-indexed" 3
 105   (and (eq_attr "type" "load_ux")
 106        (eq_attr "cpu" "power4"))
 107   "du1_power4+du2_power4+du3_power4+du4_power4,\
 108    iu1_power4,lsu2_power4+iu2_power4")
 109
 110 (define_insn_reservation "power4-load-update" 4 ; 3
 111   (and (eq_attr "type" "load_u")
 112        (eq_attr "cpu" "power4"))
 113   "lsuq_power4")
 114
 115 (define_insn_reservation "power4-fpload" 6 ; 5
 116   (and (eq_attr "type" "fpload")
 117        (eq_attr "cpu" "power4"))
 118   "lsq_power4")
 119
 120 (define_insn_reservation "power4-fpload-update" 6 ; 5
 121   (and (eq_attr "type" "fpload_u,fpload_ux")
 122        (eq_attr "cpu" "power4"))
 123   "lsuq_power4")
 124
 125 (define_insn_reservation "power4-vecload" 6 ; 5
 126   (and (eq_attr "type" "vecload")
 127        (eq_attr "cpu" "power4"))
 128   "lsq_power4")
 129
 130 (define_insn_reservation "power4-store" 12
 131   (and (eq_attr "type" "store")
 132        (eq_attr "cpu" "power4"))
 133   "((du1_power4,lsu1_power4)\
 134     |(du2_power4,lsu2_power4)\
 135     |(du3_power4,lsu2_power4)\
 136     |(du4_power4,lsu1_power4)),\
 137    (iu1_power4|iu2_power4)")
 138
 139 (define_insn_reservation "power4-store-update" 12
 140   (and (eq_attr "type" "store_u")
 141        (eq_attr "cpu" "power4"))
 142   "((du1_power4+du2_power4,lsu1_power4)\
 143     |(du2_power4+du3_power4,lsu2_power4)\
 144     |(du3_power4+du4_power4,lsu2_power4))+\
 145    ((nothing,iu1_power4,iu2_power4)\
 146     |(nothing,iu2_power4,iu2_power4)\
 147     |(nothing,iu2_power4,iu1_power4))")
 148
 149 (define_insn_reservation "power4-store-update-indexed" 12
 150   (and (eq_attr "type" "store_ux")
 151        (eq_attr "cpu" "power4"))
 152    "du1_power4+du2_power4+du3_power4+du4_power4,\
 153     iu1_power4,lsu2_power4+iu2_power4,iu2_power4")
 154
 155 (define_insn_reservation "power4-fpstore" 12
 156   (and (eq_attr "type" "fpstore")
 157        (eq_attr "cpu" "power4"))
 158   "((du1_power4,lsu1_power4)\
 159     |(du2_power4,lsu2_power4)\
 160     |(du3_power4,lsu2_power4)\
 161     |(du4_power4,lsu1_power4)),\
 162    (fpu1_power4|fpu2_power4)")
 163
 164 (define_insn_reservation "power4-fpstore-update" 12
 165   (and (eq_attr "type" "fpstore_u,fpstore_ux")
 166        (eq_attr "cpu" "power4"))
 167   "((du1_power4+du2_power4,lsu1_power4)\
 168     |(du2_power4+du3_power4,lsu2_power4)\
 169     |(du3_power4+du4_power4,lsu2_power4))\
 170    +(nothing,(iu1_power4|iu2_power4),(fpu1_power4|fpu2_power4))")
 171
 172 (define_insn_reservation "power4-vecstore" 12
 173   (and (eq_attr "type" "vecstore")
 174        (eq_attr "cpu" "power4"))
 175   "(du1_power4,lsu1_power4,vec_power4)\
 176   |(du2_power4,lsu2_power4,vec_power4)\
 177   |(du3_power4,lsu2_power4,vec_power4)\
 178   |(du4_power4,lsu1_power4,vec_power4)")
 179
 180 (define_insn_reservation "power4-llsc" 11
 181   (and (eq_attr "type" "load_l,store_c,sync")
 182        (eq_attr "cpu" "power4"))
 183   "du1_power4+du2_power4+du3_power4+du4_power4,lsu1_power4")
 184
 185
 186 ; Integer latency is 2 cycles
 187 (define_insn_reservation "power4-integer" 2
 188   (and (eq_attr "type" "integer,insert_dword,shift,trap,\
 189                         var_shift_rotate,cntlz,exts,isel")
 190        (eq_attr "cpu" "power4"))
 191   "iq_power4")
 192
 193 (define_insn_reservation "power4-two" 2
 194   (and (eq_attr "type" "two")
 195        (eq_attr "cpu" "power4"))
 196   "((du1_power4+du2_power4)\
 197     |(du2_power4+du3_power4)\
 198     |(du3_power4+du4_power4)\
 199     |(du4_power4+du1_power4)),\
 200     ((iu1_power4,nothing,iu2_power4)\
 201      |(iu2_power4,nothing,iu2_power4)\
 202      |(iu2_power4,nothing,iu1_power4)\
 203      |(iu1_power4,nothing,iu1_power4))")
 204
 205 (define_insn_reservation "power4-three" 2
 206   (and (eq_attr "type" "three")
 207        (eq_attr "cpu" "power4"))
 208   "(du1_power4+du2_power4+du3_power4|du2_power4+du3_power4+du4_power4\
 209     |du3_power4+du4_power4+du1_power4|du4_power4+du1_power4+du2_power4),\
 210    ((iu1_power4,nothing,iu2_power4,nothing,iu2_power4)\
 211     |(iu2_power4,nothing,iu2_power4,nothing,iu1_power4)\
 212     |(iu2_power4,nothing,iu1_power4,nothing,iu1_power4)\
 213     |(iu1_power4,nothing,iu1_power4,nothing,iu2_power4))")
 214
 215 (define_insn_reservation "power4-insert" 4
 216   (and (eq_attr "type" "insert_word")
 217        (eq_attr "cpu" "power4"))
 218   "(du1_power4+du2_power4|du2_power4+du3_power4|du3_power4+du4_power4),\
 219    ((iu1_power4,nothing,iu2_power4)\
 220     |(iu2_power4,nothing,iu2_power4)\
 221     |(iu2_power4,nothing,iu1_power4))")
 222
 223 (define_insn_reservation "power4-cmp" 3
 224   (and (eq_attr "type" "cmp,fast_compare")
 225        (eq_attr "cpu" "power4"))
 226   "iq_power4")
 227
 228 (define_insn_reservation "power4-compare" 2
 229   (and (eq_attr "type" "compare,delayed_compare,var_delayed_compare")
 230        (eq_attr "cpu" "power4"))
 231   "(du1_power4+du2_power4|du2_power4+du3_power4|du3_power4+du4_power4),\
 232    ((iu1_power4,iu2_power4)\
 233     |(iu2_power4,iu2_power4)\
 234     |(iu2_power4,iu1_power4))")
 235
 236 (define_bypass 4 "power4-compare" "power4-branch,power4-crlogical,power4-delayedcr,power4-mfcr,power4-mfcrf")
 237
 238 (define_insn_reservation "power4-lmul-cmp" 7
 239   (and (eq_attr "type" "lmul_compare")
 240        (eq_attr "cpu" "power4"))
 241   "(du1_power4+du2_power4|du2_power4+du3_power4|du3_power4+du4_power4),\
 242    ((iu1_power4*6,iu2_power4)\
 243     |(iu2_power4*6,iu2_power4)\
 244     |(iu2_power4*6,iu1_power4))")
 245
 246 (define_bypass 10 "power4-lmul-cmp" "power4-branch,power4-crlogical,power4-delayedcr,power4-mfcr,power4-mfcrf")
 247
 248 (define_insn_reservation "power4-imul-cmp" 5
 249   (and (eq_attr "type" "imul_compare")
 250        (eq_attr "cpu" "power4"))
 251   "(du1_power4+du2_power4|du2_power4+du3_power4|du3_power4+du4_power4),\
 252    ((iu1_power4*4,iu2_power4)\
 253     |(iu2_power4*4,iu2_power4)\
 254     |(iu2_power4*4,iu1_power4))")
 255
 256 (define_bypass 8 "power4-imul-cmp" "power4-branch,power4-crlogical,power4-delayedcr,power4-mfcr,power4-mfcrf")
 257
 258 (define_insn_reservation "power4-lmul" 7
 259   (and (eq_attr "type" "lmul")
 260        (eq_attr "cpu" "power4"))
 261   "(du1_power4|du2_power4|du3_power4|du4_power4),\
 262    (iu1_power4*6|iu2_power4*6)")
 263
 264 (define_insn_reservation "power4-imul" 5
 265   (and (eq_attr "type" "imul")
 266        (eq_attr "cpu" "power4"))
 267   "(du1_power4|du2_power4|du3_power4|du4_power4),\
 268    (iu1_power4*4|iu2_power4*4)")
 269
 270 (define_insn_reservation "power4-imul3" 4
 271   (and (eq_attr "type" "imul2,imul3")
 272        (eq_attr "cpu" "power4"))
 273   "(du1_power4|du2_power4|du3_power4|du4_power4),\
 274    (iu1_power4*3|iu2_power4*3)")
 275
 276
 277 ; SPR move only executes in first IU.
 278 ; Integer division only executes in second IU.
 279 (define_insn_reservation "power4-idiv" 36
 280   (and (eq_attr "type" "idiv")
 281        (eq_attr "cpu" "power4"))
 282   "du1_power4+du2_power4,iu2_power4*35")
 283
 284 (define_insn_reservation "power4-ldiv" 68
 285   (and (eq_attr "type" "ldiv")
 286        (eq_attr "cpu" "power4"))
 287   "du1_power4+du2_power4,iu2_power4*67")
 288
 289
 290 (define_insn_reservation "power4-mtjmpr" 3
 291   (and (eq_attr "type" "mtjmpr,mfjmpr")
 292        (eq_attr "cpu" "power4"))
 293   "du1_power4,bpu_power4")
 294
 295
 296 ; Branches take dispatch Slot 4.  The presence_sets prevent other insn from
 297 ; grabbing previous dispatch slots once this is assigned.
 298 (define_insn_reservation "power4-branch" 2
 299   (and (eq_attr "type" "jmpreg,branch")
 300        (eq_attr "cpu" "power4"))
 301   "(du5_power4\
 302    |du4_power4+du5_power4\
 303    |du3_power4+du4_power4+du5_power4\
 304    |du2_power4+du3_power4+du4_power4+du5_power4\
 305    |du1_power4+du2_power4+du3_power4+du4_power4+du5_power4),bpu_power4")
 306
 307
 308 ; Condition Register logical ops are split if non-destructive (RT != RB)
 309 (define_insn_reservation "power4-crlogical" 2
 310   (and (eq_attr "type" "cr_logical")
 311        (eq_attr "cpu" "power4"))
 312   "du1_power4,cru_power4")
 313
 314 (define_insn_reservation "power4-delayedcr" 4
 315   (and (eq_attr "type" "delayed_cr")
 316        (eq_attr "cpu" "power4"))
 317   "du1_power4+du2_power4,cru_power4,cru_power4")
 318
 319 ; 4 mfcrf (each 3 cyc, 1/cyc) + 3 fxu
 320 (define_insn_reservation "power4-mfcr" 6
 321   (and (eq_attr "type" "mfcr")
 322        (eq_attr "cpu" "power4"))
 323   "du1_power4+du2_power4+du3_power4+du4_power4,\
 324    du1_power4+du2_power4+du3_power4+du4_power4+cru_power4,\
 325    cru_power4,cru_power4,cru_power4")
 326
 327 ; mfcrf (1 field)
 328 (define_insn_reservation "power4-mfcrf" 3
 329   (and (eq_attr "type" "mfcrf")
 330        (eq_attr "cpu" "power4"))
 331   "du1_power4,cru_power4")
 332
 333 ; mtcrf (1 field)
 334 (define_insn_reservation "power4-mtcr" 4
 335   (and (eq_attr "type" "mtcr")
 336        (eq_attr "cpu" "power4"))
 337   "du1_power4,iu1_power4")
 338
 339 ; Basic FP latency is 6 cycles
 340 (define_insn_reservation "power4-fp" 6
 341   (and (eq_attr "type" "fp,dmul")
 342        (eq_attr "cpu" "power4"))
 343   "fpq_power4")
 344
 345 (define_insn_reservation "power4-fpcompare" 5
 346   (and (eq_attr "type" "fpcompare")
 347        (eq_attr "cpu" "power4"))
 348   "fpq_power4")
 349
 350 (define_insn_reservation "power4-sdiv" 33
 351   (and (eq_attr "type" "sdiv,ddiv")
 352        (eq_attr "cpu" "power4"))
 353   "(du1_power4|du2_power4|du3_power4|du4_power4),\
 354    (fpu1_power4*28|fpu2_power4*28)")
 355
 356 (define_insn_reservation "power4-sqrt" 40
 357   (and (eq_attr "type" "ssqrt,dsqrt")
 358        (eq_attr "cpu" "power4"))
 359   "(du1_power4|du2_power4|du3_power4|du4_power4),\
 360    (fpu1_power4*35|fpu2_power4*35)")
 361
 362 (define_insn_reservation "power4-isync" 2
 363   (and (eq_attr "type" "isync")
 364        (eq_attr "cpu" "power4"))
 365   "du1_power4+du2_power4+du3_power4+du4_power4,lsu1_power4")
 366
 367
 368 ; VMX
 369 (define_insn_reservation "power4-vecsimple" 2
 370   (and (eq_attr "type" "vecsimple")
 371        (eq_attr "cpu" "power4"))
 372   "vq_power4")
 373
 374 (define_insn_reservation "power4-veccomplex" 5
 375   (and (eq_attr "type" "veccomplex")
 376        (eq_attr "cpu" "power4"))
 377   "vq_power4")
 378
 379 ; vecfp compare
 380 (define_insn_reservation "power4-veccmp" 8
 381   (and (eq_attr "type" "veccmp")
 382        (eq_attr "cpu" "power4"))
 383   "vq_power4")
 384
 385 (define_insn_reservation "power4-vecfloat" 8
 386   (and (eq_attr "type" "vecfloat")
 387        (eq_attr "cpu" "power4"))
 388   "vq_power4")
 389
 390 (define_insn_reservation "power4-vecperm" 2
 391   (and (eq_attr "type" "vecperm")
 392        (eq_attr "cpu" "power4"))
 393   "vpq_power4")
 394
 395 (define_bypass 4 "power4-vecload" "power4-vecperm")
 396
 397 (define_bypass 3 "power4-vecsimple" "power4-vecperm")
 398 (define_bypass 6 "power4-veccomplex" "power4-vecperm")
 399 (define_bypass 3 "power4-vecperm"
 400                  "power4-vecsimple,power4-veccomplex,power4-vecfloat")
 401 (define_bypass 9 "power4-vecfloat" "power4-vecperm")
 402
 403 (define_bypass 5 "power4-vecsimple,power4-veccomplex"
 404                  "power4-branch,power4-crlogical,power4-delayedcr,power4-mfcr,power4-mfcrf")
 405
 406 (define_bypass 4 "power4-vecsimple,power4-vecperm" "power4-vecstore")
 407 (define_bypass 7 "power4-veccomplex" "power4-vecstore")
 408 (define_bypass 10 "power4-vecfloat" "power4-vecstore")