gcc/config/rs6000/vsx.md

   1 ;; VSX patterns.
   2 ;; Copyright (C) 2009, 2010
   3 ;; Free Software Foundation, Inc.
   4 ;; Contributed by Michael Meissner <meissner@linux.vnet.ibm.com>
   5
   6 ;; This file is part of GCC.
   7
   8 ;; GCC is free software; you can redistribute it and/or modify it
   9 ;; under the terms of the GNU General Public License as published
  10 ;; by the Free Software Foundation; either version 3, or (at your
  11 ;; option) any later version.
  12
  13 ;; GCC is distributed in the hope that it will be useful, but WITHOUT
  14 ;; ANY WARRANTY; without even the implied warranty of MERCHANTABILITY
  15 ;; or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public
  16 ;; License for more details.
  17
  18 ;; You should have received a copy of the GNU General Public License
  19 ;; along with GCC; see the file COPYING3.  If not see
  20 ;; <http://www.gnu.org/licenses/>.
  21
  22 ;; Iterator for both scalar and vector floating point types supported by VSX
  23 (define_mode_iterator VSX_B [DF V4SF V2DF])
  24
  25 ;; Iterator for the 2 64-bit vector types
  26 (define_mode_iterator VSX_D [V2DF V2DI])
  27
  28 ;; Iterator for the 2 32-bit vector types
  29 (define_mode_iterator VSX_W [V4SF V4SI])
  30
  31 ;; Iterator for the DF types
  32 (define_mode_iterator VSX_DF [V2DF DF])
  33
  34 ;; Iterator for vector floating point types supported by VSX
  35 (define_mode_iterator VSX_F [V4SF V2DF])
  36
  37 ;; Iterator for logical types supported by VSX
  38 (define_mode_iterator VSX_L [V16QI V8HI V4SI V2DI V4SF V2DF TI])
  39
  40 ;; Iterator for memory move.  Handle TImode specially to allow
  41 ;; it to use gprs as well as vsx registers.
  42 (define_mode_iterator VSX_M [V16QI V8HI V4SI V2DI V4SF V2DF])
  43
  44 ;; Map into the appropriate load/store name based on the type
  45 (define_mode_attr VSm  [(V16QI "vw4")
  46                         (V8HI  "vw4")
  47                         (V4SI  "vw4")
  48                         (V4SF  "vw4")
  49                         (V2DF  "vd2")
  50                         (V2DI  "vd2")
  51                         (DF    "d")
  52                         (TI    "vw4")])
  53
  54 ;; Map into the appropriate suffix based on the type
  55 (define_mode_attr VSs   [(V16QI "sp")
  56                          (V8HI  "sp")
  57                          (V4SI  "sp")
  58                          (V4SF  "sp")
  59                          (V2DF  "dp")
  60                          (V2DI  "dp")
  61                          (DF    "dp")
  62                          (SF    "sp")
  63                          (TI    "sp")])
  64
  65 ;; Map the register class used
  66 (define_mode_attr VSr   [(V16QI "v")
  67                          (V8HI  "v")
  68                          (V4SI  "v")
  69                          (V4SF  "wf")
  70                          (V2DI  "wd")
  71                          (V2DF  "wd")
  72                          (DF    "ws")
  73                          (SF    "d")
  74                          (TI    "wd")])
  75
  76 ;; Map the register class used for float<->int conversions
  77 (define_mode_attr VSr2  [(V2DF  "wd")
  78                          (V4SF  "wf")
  79                          (DF    "ws")])
  80
  81 (define_mode_attr VSr3  [(V2DF  "wa")
  82                          (V4SF  "wa")
  83                          (DF    "ws")])
  84
  85 ;; Map the register class for sp<->dp float conversions, destination
  86 (define_mode_attr VSr4  [(SF    "ws")
  87                          (DF    "f")
  88                          (V2DF  "wd")
  89                          (V4SF  "v")])
  90
  91 ;; Map the register class for sp<->dp float conversions, destination
  92 (define_mode_attr VSr5  [(SF    "ws")
  93                          (DF    "f")
  94                          (V2DF  "v")
  95                          (V4SF  "wd")])
  96
  97 ;; Same size integer type for floating point data
  98 (define_mode_attr VSi [(V4SF  "v4si")
  99                        (V2DF  "v2di")
 100                        (DF    "di")])
 101
 102 (define_mode_attr VSI [(V4SF  "V4SI")
 103                        (V2DF  "V2DI")
 104                        (DF    "DI")])
 105
 106 ;; Word size for same size conversion
 107 (define_mode_attr VSc [(V4SF "w")
 108                        (V2DF "d")
 109                        (DF   "d")])
 110
 111 ;; Map into either s or v, depending on whether this is a scalar or vector
 112 ;; operation
 113 (define_mode_attr VSv   [(V16QI "v")
 114                          (V8HI  "v")
 115                          (V4SI  "v")
 116                          (V4SF  "v")
 117                          (V2DI  "v")
 118                          (V2DF  "v")
 119                          (TI    "v")
 120                          (DF    "s")])
 121
 122 ;; Appropriate type for add ops (and other simple FP ops)
 123 (define_mode_attr VStype_simple [(V2DF "vecfloat")
 124                                  (V4SF "vecfloat")
 125                                  (DF   "fp")])
 126
 127 (define_mode_attr VSfptype_simple [(V2DF "fp_addsub_d")
 128                                    (V4SF "fp_addsub_s")
 129                                    (DF   "fp_addsub_d")])
 130
 131 ;; Appropriate type for multiply ops
 132 (define_mode_attr VStype_mul    [(V2DF "vecfloat")
 133                                  (V4SF "vecfloat")
 134                                  (DF   "dmul")])
 135
 136 (define_mode_attr VSfptype_mul  [(V2DF "fp_mul_d")
 137                                  (V4SF "fp_mul_s")
 138                                  (DF   "fp_mul_d")])
 139
 140 ;; Appropriate type for divide ops.  For now, just lump the vector divide with
 141 ;; the scalar divides
 142 (define_mode_attr VStype_div    [(V2DF "ddiv")
 143                                  (V4SF "sdiv")
 144                                  (DF   "ddiv")])
 145
 146 (define_mode_attr VSfptype_div  [(V2DF "fp_div_d")
 147                                  (V4SF "fp_div_s")
 148                                  (DF   "fp_div_d")])
 149
 150 ;; Appropriate type for sqrt ops.  For now, just lump the vector sqrt with
 151 ;; the scalar sqrt
 152 (define_mode_attr VStype_sqrt   [(V2DF "dsqrt")
 153                                  (V4SF "sdiv")
 154                                  (DF   "ddiv")])
 155
 156 (define_mode_attr VSfptype_sqrt [(V2DF "fp_sqrt_d")
 157                                  (V4SF "fp_sqrt_s")
 158                                  (DF   "fp_sqrt_d")])
 159
 160 ;; Iterator and modes for sp<->dp conversions
 161 ;; Because scalar SF values are represented internally as double, use the
 162 ;; V4SF type to represent this than SF.
 163 (define_mode_iterator VSX_SPDP [DF V4SF V2DF])
 164
 165 (define_mode_attr VS_spdp_res [(DF      "V4SF")
 166                                (V4SF    "V2DF")
 167                                (V2DF    "V4SF")])
 168
 169 (define_mode_attr VS_spdp_insn [(DF     "xscvdpsp")
 170                                 (V4SF   "xvcvspdp")
 171                                 (V2DF   "xvcvdpsp")])
 172
 173 (define_mode_attr VS_spdp_type [(DF     "fp")
 174                                 (V4SF   "vecfloat")
 175                                 (V2DF   "vecfloat")])
 176
 177 ;; Map the scalar mode for a vector type
 178 (define_mode_attr VS_scalar [(V2DF      "DF")
 179                              (V2DI      "DI")
 180                              (V4SF      "SF")
 181                              (V4SI      "SI")
 182                              (V8HI      "HI")
 183                              (V16QI     "QI")])
 184
 185 ;; Constants for creating unspecs
 186 (define_constants
 187   [(UNSPEC_VSX_CONCAT           500)
 188    (UNSPEC_VSX_CVDPSXWS         501)
 189    (UNSPEC_VSX_CVDPUXWS         502)
 190    (UNSPEC_VSX_CVSPDP           503)
 191    (UNSPEC_VSX_CVSXWDP          504)
 192    (UNSPEC_VSX_CVUXWDP          505)
 193    (UNSPEC_VSX_CVSXDSP          506)
 194    (UNSPEC_VSX_CVUXDSP          507)
 195    (UNSPEC_VSX_CVSPSXDS         508)
 196    (UNSPEC_VSX_CVSPUXDS         509)
 197    ;; 510-514 deleted
 198    (UNSPEC_VSX_TDIV             515)
 199    (UNSPEC_VSX_TSQRT            516)
 200    (UNSPEC_VSX_XXPERMDI         517)
 201    (UNSPEC_VSX_SET              518)
 202    (UNSPEC_VSX_ROUND_I          519)
 203    (UNSPEC_VSX_ROUND_IC         520)
 204    (UNSPEC_VSX_SLDWI            521)])
 205
 206 ;; VSX moves
 207 (define_insn "*vsx_mov<mode>"
 208   [(set (match_operand:VSX_M 0 "nonimmediate_operand" "=Z,<VSr>,<VSr>,?Z,?wa,?wa,*o,*r,*r,<VSr>,?wa,v,wZ,v")
 209         (match_operand:VSX_M 1 "input_operand" "<VSr>,Z,<VSr>,wa,Z,wa,r,o,r,j,j,W,v,wZ"))]
 210   "VECTOR_MEM_VSX_P (<MODE>mode)
 211    && (register_operand (operands[0], <MODE>mode)
 212        || register_operand (operands[1], <MODE>mode))"
 213 {
 214   switch (which_alternative)
 215     {
 216     case 0:
 217     case 3:
 218       gcc_assert (MEM_P (operands[0])
 219                   && GET_CODE (XEXP (operands[0], 0)) != PRE_INC
 220                   && GET_CODE (XEXP (operands[0], 0)) != PRE_DEC
 221                   && GET_CODE (XEXP (operands[0], 0)) != PRE_MODIFY);
 222       return "stx<VSm>x %x1,%y0";
 223
 224     case 1:
 225     case 4:
 226       gcc_assert (MEM_P (operands[1])
 227                   && GET_CODE (XEXP (operands[1], 0)) != PRE_INC
 228                   && GET_CODE (XEXP (operands[1], 0)) != PRE_DEC
 229                   && GET_CODE (XEXP (operands[1], 0)) != PRE_MODIFY);
 230       return "lx<VSm>x %x0,%y1";
 231
 232     case 2:
 233     case 5:
 234       return "xxlor %x0,%x1,%x1";
 235
 236     case 6:
 237     case 7:
 238     case 8:
 239       return "#";
 240
 241     case 9:
 242     case 10:
 243       return "xxlxor %x0,%x0,%x0";
 244
 245     case 11:
 246       return output_vec_const_move (operands);
 247
 248     case 12:
 249       gcc_assert (MEM_P (operands[0])
 250                   && GET_CODE (XEXP (operands[0], 0)) != PRE_INC
 251                   && GET_CODE (XEXP (operands[0], 0)) != PRE_DEC
 252                   && GET_CODE (XEXP (operands[0], 0)) != PRE_MODIFY);
 253       return "stvx %1,%y0";
 254
 255     case 13:
 256       gcc_assert (MEM_P (operands[0])
 257                   && GET_CODE (XEXP (operands[0], 0)) != PRE_INC
 258                   && GET_CODE (XEXP (operands[0], 0)) != PRE_DEC
 259                   && GET_CODE (XEXP (operands[0], 0)) != PRE_MODIFY);
 260       return "lvx %0,%y1";
 261
 262     default:
 263       gcc_unreachable ();
 264     }
 265 }
 266   [(set_attr "type" "vecstore,vecload,vecsimple,vecstore,vecload,vecsimple,*,*,*,vecsimple,vecsimple,*,vecstore,vecload")])
 267
 268 ;; Unlike other VSX moves, allow the GPRs, since a normal use of TImode is for
 269 ;; unions.  However for plain data movement, slightly favor the vector loads
 270 (define_insn "*vsx_movti"
 271   [(set (match_operand:TI 0 "nonimmediate_operand" "=Z,wa,wa,?o,?r,?r,wa,v,v,wZ")
 272         (match_operand:TI 1 "input_operand" "wa,Z,wa,r,o,r,j,W,wZ,v"))]
 273   "VECTOR_MEM_VSX_P (TImode)
 274    && (register_operand (operands[0], TImode)
 275        || register_operand (operands[1], TImode))"
 276 {
 277   switch (which_alternative)
 278     {
 279     case 0:
 280       return "stxvd2x %x1,%y0";
 281
 282     case 1:
 283       return "lxvd2x %x0,%y1";
 284
 285     case 2:
 286       return "xxlor %x0,%x1,%x1";
 287
 288     case 3:
 289     case 4:
 290     case 5:
 291       return "#";
 292
 293     case 6:
 294       return "xxlxor %x0,%x0,%x0";
 295
 296     case 7:
 297       return output_vec_const_move (operands);
 298
 299     case 8:
 300       return "stvx %1,%y0";
 301
 302     case 9:
 303       return "lvx %0,%y1";
 304
 305     default:
 306       gcc_unreachable ();
 307     }
 308 }
 309   [(set_attr "type" "vecstore,vecload,vecsimple,*,*,*,vecsimple,*,vecstore,vecload")])
 310
 311 \f
 312 ;; VSX scalar and vector floating point arithmetic instructions
 313 (define_insn "*vsx_add<mode>3"
 314   [(set (match_operand:VSX_B 0 "vsx_register_operand" "=<VSr>,?wa")
 315         (plus:VSX_B (match_operand:VSX_B 1 "vsx_register_operand" "<VSr>,wa")
 316                     (match_operand:VSX_B 2 "vsx_register_operand" "<VSr>,wa")))]
 317   "VECTOR_UNIT_VSX_P (<MODE>mode)"
 318   "x<VSv>add<VSs> %x0,%x1,%x2"
 319   [(set_attr "type" "<VStype_simple>")
 320    (set_attr "fp_type" "<VSfptype_simple>")])
 321
 322 (define_insn "*vsx_sub<mode>3"
 323   [(set (match_operand:VSX_B 0 "vsx_register_operand" "=<VSr>,?wa")
 324         (minus:VSX_B (match_operand:VSX_B 1 "vsx_register_operand" "<VSr>,wa")
 325                      (match_operand:VSX_B 2 "vsx_register_operand" "<VSr>,wa")))]
 326   "VECTOR_UNIT_VSX_P (<MODE>mode)"
 327   "x<VSv>sub<VSs> %x0,%x1,%x2"
 328   [(set_attr "type" "<VStype_simple>")
 329    (set_attr "fp_type" "<VSfptype_simple>")])
 330
 331 (define_insn "*vsx_mul<mode>3"
 332   [(set (match_operand:VSX_B 0 "vsx_register_operand" "=<VSr>,?wa")
 333         (mult:VSX_B (match_operand:VSX_B 1 "vsx_register_operand" "<VSr>,wa")
 334                     (match_operand:VSX_B 2 "vsx_register_operand" "<VSr>,wa")))]
 335   "VECTOR_UNIT_VSX_P (<MODE>mode)"
 336   "x<VSv>mul<VSs> %x0,%x1,%x2"
 337   [(set_attr "type" "<VStype_mul>")
 338    (set_attr "fp_type" "<VSfptype_mul>")])
 339
 340 (define_insn "*vsx_div<mode>3"
 341   [(set (match_operand:VSX_B 0 "vsx_register_operand" "=<VSr>,?wa")
 342         (div:VSX_B (match_operand:VSX_B 1 "vsx_register_operand" "<VSr>,wa")
 343                    (match_operand:VSX_B 2 "vsx_register_operand" "<VSr>,wa")))]
 344   "VECTOR_UNIT_VSX_P (<MODE>mode)"
 345   "x<VSv>div<VSs> %x0,%x1,%x2"
 346   [(set_attr "type" "<VStype_div>")
 347    (set_attr "fp_type" "<VSfptype_div>")])
 348
 349 ;; *tdiv* instruction returning the FG flag
 350 (define_expand "vsx_tdiv<mode>3_fg"
 351   [(set (match_dup 3)
 352         (unspec:CCFP [(match_operand:VSX_B 1 "vsx_register_operand" "")
 353                       (match_operand:VSX_B 2 "vsx_register_operand" "")]
 354                      UNSPEC_VSX_TDIV))
 355    (set (match_operand:SI 0 "gpc_reg_operand" "")
 356         (gt:SI (match_dup 3)
 357                (const_int 0)))]
 358   "VECTOR_UNIT_VSX_P (<MODE>mode)"
 359 {
 360   operands[3] = gen_reg_rtx (CCFPmode);
 361 })
 362
 363 ;; *tdiv* instruction returning the FE flag
 364 (define_expand "vsx_tdiv<mode>3_fe"
 365   [(set (match_dup 3)
 366         (unspec:CCFP [(match_operand:VSX_B 1 "vsx_register_operand" "")
 367                       (match_operand:VSX_B 2 "vsx_register_operand" "")]
 368                      UNSPEC_VSX_TDIV))
 369    (set (match_operand:SI 0 "gpc_reg_operand" "")
 370         (eq:SI (match_dup 3)
 371                (const_int 0)))]
 372   "VECTOR_UNIT_VSX_P (<MODE>mode)"
 373 {
 374   operands[3] = gen_reg_rtx (CCFPmode);
 375 })
 376
 377 (define_insn "*vsx_tdiv<mode>3_internal"
 378   [(set (match_operand:CCFP 0 "cc_reg_operand" "=x,x")
 379         (unspec:CCFP [(match_operand:VSX_B 1 "vsx_register_operand" "<VSr>,wa")
 380                       (match_operand:VSX_B 2 "vsx_register_operand" "<VSr>,wa")]
 381                    UNSPEC_VSX_TDIV))]
 382   "VECTOR_UNIT_VSX_P (<MODE>mode)"
 383   "x<VSv>tdiv<VSs> %0,%x1,%x2"
 384   [(set_attr "type" "<VStype_simple>")
 385    (set_attr "fp_type" "<VSfptype_simple>")])
 386
 387 (define_insn "vsx_fre<mode>2"
 388   [(set (match_operand:VSX_B 0 "vsx_register_operand" "=<VSr>,?wa")
 389         (unspec:VSX_B [(match_operand:VSX_B 1 "vsx_register_operand" "<VSr>,wa")]
 390                       UNSPEC_FRES))]
 391   "VECTOR_UNIT_VSX_P (<MODE>mode)"
 392   "x<VSv>re<VSs> %x0,%x1"
 393   [(set_attr "type" "<VStype_simple>")
 394    (set_attr "fp_type" "<VSfptype_simple>")])
 395
 396 (define_insn "*vsx_neg<mode>2"
 397   [(set (match_operand:VSX_B 0 "vsx_register_operand" "=<VSr>,?wa")
 398         (neg:VSX_B (match_operand:VSX_B 1 "vsx_register_operand" "<VSr>,wa")))]
 399   "VECTOR_UNIT_VSX_P (<MODE>mode)"
 400   "x<VSv>neg<VSs> %x0,%x1"
 401   [(set_attr "type" "<VStype_simple>")
 402    (set_attr "fp_type" "<VSfptype_simple>")])
 403
 404 (define_insn "*vsx_abs<mode>2"
 405   [(set (match_operand:VSX_B 0 "vsx_register_operand" "=<VSr>,?wa")
 406         (abs:VSX_B (match_operand:VSX_B 1 "vsx_register_operand" "<VSr>,wa")))]
 407   "VECTOR_UNIT_VSX_P (<MODE>mode)"
 408   "x<VSv>abs<VSs> %x0,%x1"
 409   [(set_attr "type" "<VStype_simple>")
 410    (set_attr "fp_type" "<VSfptype_simple>")])
 411
 412 (define_insn "vsx_nabs<mode>2"
 413   [(set (match_operand:VSX_B 0 "vsx_register_operand" "=<VSr>,?wa")
 414         (neg:VSX_B
 415          (abs:VSX_B
 416           (match_operand:VSX_B 1 "vsx_register_operand" "<VSr>,wa"))))]
 417   "VECTOR_UNIT_VSX_P (<MODE>mode)"
 418   "x<VSv>nabs<VSs> %x0,%x1"
 419   [(set_attr "type" "<VStype_simple>")
 420    (set_attr "fp_type" "<VSfptype_simple>")])
 421
 422 (define_insn "vsx_smax<mode>3"
 423   [(set (match_operand:VSX_B 0 "vsx_register_operand" "=<VSr>,?wa")
 424         (smax:VSX_B (match_operand:VSX_B 1 "vsx_register_operand" "<VSr>,wa")
 425                     (match_operand:VSX_B 2 "vsx_register_operand" "<VSr>,wa")))]
 426   "VECTOR_UNIT_VSX_P (<MODE>mode)"
 427   "x<VSv>max<VSs> %x0,%x1,%x2"
 428   [(set_attr "type" "<VStype_simple>")
 429    (set_attr "fp_type" "<VSfptype_simple>")])
 430
 431 (define_insn "*vsx_smin<mode>3"
 432   [(set (match_operand:VSX_B 0 "vsx_register_operand" "=<VSr>,?wa")
 433         (smin:VSX_B (match_operand:VSX_B 1 "vsx_register_operand" "<VSr>,wa")
 434                     (match_operand:VSX_B 2 "vsx_register_operand" "<VSr>,wa")))]
 435   "VECTOR_UNIT_VSX_P (<MODE>mode)"
 436   "x<VSv>min<VSs> %x0,%x1,%x2"
 437   [(set_attr "type" "<VStype_simple>")
 438    (set_attr "fp_type" "<VSfptype_simple>")])
 439
 440 ;; Special VSX version of smin/smax for single precision floating point.  Since
 441 ;; both numbers are rounded to single precision, we can just use the DP version
 442 ;; of the instruction.
 443
 444 (define_insn "*vsx_smaxsf3"
 445   [(set (match_operand:SF 0 "vsx_register_operand" "=f")
 446         (smax:SF (match_operand:SF 1 "vsx_register_operand" "f")
 447                  (match_operand:SF 2 "vsx_register_operand" "f")))]
 448   "VECTOR_UNIT_VSX_P (DFmode)"
 449   "xsmaxdp %x0,%x1,%x2"
 450   [(set_attr "type" "fp")
 451    (set_attr "fp_type" "fp_addsub_d")])
 452
 453 (define_insn "*vsx_sminsf3"
 454   [(set (match_operand:SF 0 "vsx_register_operand" "=f")
 455         (smin:SF (match_operand:SF 1 "vsx_register_operand" "f")
 456                  (match_operand:SF 2 "vsx_register_operand" "f")))]
 457   "VECTOR_UNIT_VSX_P (DFmode)"
 458   "xsmindp %x0,%x1,%x2"
 459   [(set_attr "type" "fp")
 460    (set_attr "fp_type" "fp_addsub_d")])
 461
 462 (define_insn "*vsx_sqrt<mode>2"
 463   [(set (match_operand:VSX_B 0 "vsx_register_operand" "=<VSr>,?wa")
 464         (sqrt:VSX_B (match_operand:VSX_B 1 "vsx_register_operand" "<VSr>,wa")))]
 465   "VECTOR_UNIT_VSX_P (<MODE>mode)"
 466   "x<VSv>sqrt<VSs> %x0,%x1"
 467   [(set_attr "type" "<VStype_sqrt>")
 468    (set_attr "fp_type" "<VSfptype_sqrt>")])
 469
 470 (define_insn "*vsx_rsqrte<mode>2"
 471   [(set (match_operand:VSX_B 0 "vsx_register_operand" "=<VSr>,?wa")
 472         (unspec:VSX_B [(match_operand:VSX_B 1 "vsx_register_operand" "<VSr>,wa")]
 473                       UNSPEC_RSQRT))]
 474   "VECTOR_UNIT_VSX_P (<MODE>mode)"
 475   "x<VSv>rsqrte<VSs> %x0,%x1"
 476   [(set_attr "type" "<VStype_simple>")
 477    (set_attr "fp_type" "<VSfptype_simple>")])
 478
 479 ;; *tsqrt* returning the fg flag
 480 (define_expand "vsx_tsqrt<mode>2_fg"
 481   [(set (match_dup 3)
 482         (unspec:CCFP [(match_operand:VSX_B 1 "vsx_register_operand" "")]
 483                      UNSPEC_VSX_TSQRT))
 484    (set (match_operand:SI 0 "gpc_reg_operand" "")
 485         (gt:SI (match_dup 3)
 486                (const_int 0)))]
 487   "VECTOR_UNIT_VSX_P (<MODE>mode)"
 488 {
 489   operands[3] = gen_reg_rtx (CCFPmode);
 490 })
 491
 492 ;; *tsqrt* returning the fe flag
 493 (define_expand "vsx_tsqrt<mode>2_fe"
 494   [(set (match_dup 3)
 495         (unspec:CCFP [(match_operand:VSX_B 1 "vsx_register_operand" "")]
 496                      UNSPEC_VSX_TSQRT))
 497    (set (match_operand:SI 0 "gpc_reg_operand" "")
 498         (eq:SI (match_dup 3)
 499                (const_int 0)))]
 500   "VECTOR_UNIT_VSX_P (<MODE>mode)"
 501 {
 502   operands[3] = gen_reg_rtx (CCFPmode);
 503 })
 504
 505 (define_insn "*vsx_tsqrt<mode>2_internal"
 506   [(set (match_operand:CCFP 0 "cc_reg_operand" "=x,x")
 507         (unspec:CCFP [(match_operand:VSX_B 1 "vsx_register_operand" "<VSr>,wa")]
 508                      UNSPEC_VSX_TSQRT))]
 509   "VECTOR_UNIT_VSX_P (<MODE>mode)"
 510   "x<VSv>tsqrt<VSs> %0,%x1"
 511   [(set_attr "type" "<VStype_simple>")
 512    (set_attr "fp_type" "<VSfptype_simple>")])
 513
 514 ;; Fused vector multiply/add instructions
 515
 516 (define_insn "*vsx_fma<mode>4"
 517   [(set (match_operand:VSX_B 0 "vsx_register_operand" "=<VSr>,<VSr>,?wa,?wa")
 518         (fma:VSX_B
 519           (match_operand:VSX_B 1 "vsx_register_operand" "%<VSr>,<VSr>,wa,wa")
 520           (match_operand:VSX_B 2 "vsx_register_operand" "<VSr>,0,wa,0")
 521           (match_operand:VSX_B 3 "vsx_register_operand" "0,<VSr>,0,wa")))]
 522   "VECTOR_UNIT_VSX_P (<MODE>mode)"
 523   "@
 524    x<VSv>madda<VSs> %x0,%x1,%x2
 525    x<VSv>maddm<VSs> %x0,%x1,%x3
 526    x<VSv>madda<VSs> %x0,%x1,%x2
 527    x<VSv>maddm<VSs> %x0,%x1,%x3"
 528   [(set_attr "type" "<VStype_mul>")
 529    (set_attr "fp_type" "<VSfptype_mul>")])
 530
 531 (define_insn "*vsx_fms<mode>4"
 532   [(set (match_operand:VSX_B 0 "vsx_register_operand" "=<VSr>,<VSr>,?wa,?wa")
 533         (fma:VSX_B
 534           (match_operand:VSX_B 1 "vsx_register_operand" "%<VSr>,<VSr>,wa,wa")
 535           (match_operand:VSX_B 2 "vsx_register_operand" "<VSr>,0,wa,0")
 536           (neg:VSX_B
 537             (match_operand:VSX_B 3 "vsx_register_operand" "0,<VSr>,0,wa"))))]
 538   "VECTOR_UNIT_VSX_P (<MODE>mode)"
 539   "@
 540    x<VSv>msuba<VSs> %x0,%x1,%x2
 541    x<VSv>msubm<VSs> %x0,%x1,%x3
 542    x<VSv>msuba<VSs> %x0,%x1,%x2
 543    x<VSv>msubm<VSs> %x0,%x1,%x3"
 544   [(set_attr "type" "<VStype_mul>")
 545    (set_attr "fp_type" "<VSfptype_mul>")])
 546
 547 (define_insn "*vsx_nfma<mode>4"
 548   [(set (match_operand:VSX_B 0 "vsx_register_operand" "=<VSr>,<VSr>,?wa,?wa")
 549         (neg:VSX_B
 550          (fma:VSX_B
 551           (match_operand:VSX_B 1 "vsx_register_operand" "<VSr>,<VSr>,wa,wa")
 552           (match_operand:VSX_B 2 "vsx_register_operand" "<VSr>,0,wa,0")
 553           (match_operand:VSX_B 3 "vsx_register_operand" "0,<VSr>,0,wa"))))]
 554   "VECTOR_UNIT_VSX_P (<MODE>mode)"
 555   "@
 556    x<VSv>nmadda<VSs> %x0,%x1,%x2
 557    x<VSv>nmaddm<VSs> %x0,%x1,%x3
 558    x<VSv>nmadda<VSs> %x0,%x1,%x2
 559    x<VSv>nmaddm<VSs> %x0,%x1,%x3"
 560   [(set_attr "type" "<VStype_mul>")
 561    (set_attr "fp_type" "<VSfptype_mul>")])
 562
 563 (define_insn "*vsx_nfms<mode>4"
 564   [(set (match_operand:VSX_B 0 "vsx_register_operand" "=<VSr>,<VSr>,?wa,?wa")
 565         (neg:VSX_B
 566          (fma:VSX_B
 567            (match_operand:VSX_B 1 "vsx_register_operand" "%<VSr>,<VSr>,wa,wa")
 568            (match_operand:VSX_B 2 "vsx_register_operand" "<VSr>,0,wa,0")
 569            (neg:VSX_B
 570              (match_operand:VSX_B 3 "vsx_register_operand" "0,<VSr>,0,wa")))))]
 571   "VECTOR_UNIT_VSX_P (<MODE>mode)"
 572   "@
 573    x<VSv>nmsuba<VSs> %x0,%x1,%x2
 574    x<VSv>nmsubm<VSs> %x0,%x1,%x3
 575    x<VSv>nmsuba<VSs> %x0,%x1,%x2
 576    x<VSv>nmsubm<VSs> %x0,%x1,%x3"
 577   [(set_attr "type" "<VStype_mul>")
 578    (set_attr "fp_type" "<VSfptype_mul>")])
 579
 580 ;; Vector conditional expressions (no scalar version for these instructions)
 581 (define_insn "vsx_eq<mode>"
 582   [(set (match_operand:VSX_F 0 "vsx_register_operand" "=<VSr>,?wa")
 583         (eq:VSX_F (match_operand:VSX_F 1 "vsx_register_operand" "<VSr>,wa")
 584                   (match_operand:VSX_F 2 "vsx_register_operand" "<VSr>,wa")))]
 585   "VECTOR_UNIT_VSX_P (<MODE>mode)"
 586   "xvcmpeq<VSs> %x0,%x1,%x2"
 587   [(set_attr "type" "<VStype_simple>")
 588    (set_attr "fp_type" "<VSfptype_simple>")])
 589
 590 (define_insn "vsx_gt<mode>"
 591   [(set (match_operand:VSX_F 0 "vsx_register_operand" "=<VSr>,?wa")
 592         (gt:VSX_F (match_operand:VSX_F 1 "vsx_register_operand" "<VSr>,wa")
 593                   (match_operand:VSX_F 2 "vsx_register_operand" "<VSr>,wa")))]
 594   "VECTOR_UNIT_VSX_P (<MODE>mode)"
 595   "xvcmpgt<VSs> %x0,%x1,%x2"
 596   [(set_attr "type" "<VStype_simple>")
 597    (set_attr "fp_type" "<VSfptype_simple>")])
 598
 599 (define_insn "*vsx_ge<mode>"
 600   [(set (match_operand:VSX_F 0 "vsx_register_operand" "=<VSr>,?wa")
 601         (ge:VSX_F (match_operand:VSX_F 1 "vsx_register_operand" "<VSr>,wa")
 602                   (match_operand:VSX_F 2 "vsx_register_operand" "<VSr>,wa")))]
 603   "VECTOR_UNIT_VSX_P (<MODE>mode)"
 604   "xvcmpge<VSs> %x0,%x1,%x2"
 605   [(set_attr "type" "<VStype_simple>")
 606    (set_attr "fp_type" "<VSfptype_simple>")])
 607
 608 ;; Floating point scalar compare
 609 (define_insn "*vsx_cmpdf_internal1"
 610   [(set (match_operand:CCFP 0 "cc_reg_operand" "=y,?y")
 611         (compare:CCFP (match_operand:DF 1 "gpc_reg_operand" "ws,wa")
 612                       (match_operand:DF 2 "gpc_reg_operand" "ws,wa")))]
 613   "TARGET_HARD_FLOAT && TARGET_FPRS && TARGET_DOUBLE_FLOAT
 614    && VECTOR_UNIT_VSX_P (DFmode)"
 615   "xscmpudp %0,%x1,%x2"
 616   [(set_attr "type" "fpcompare")])
 617
 618 ;; Compare vectors producing a vector result and a predicate, setting CR6 to
 619 ;; indicate a combined status
 620 (define_insn "*vsx_eq_<mode>_p"
 621   [(set (reg:CC 74)
 622         (unspec:CC
 623          [(eq:CC (match_operand:VSX_F 1 "vsx_register_operand" "<VSr>,?wa")
 624                  (match_operand:VSX_F 2 "vsx_register_operand" "<VSr>,?wa"))]
 625          UNSPEC_PREDICATE))
 626    (set (match_operand:VSX_F 0 "vsx_register_operand" "=<VSr>,?wa")
 627         (eq:VSX_F (match_dup 1)
 628                   (match_dup 2)))]
 629   "VECTOR_UNIT_VSX_P (<MODE>mode)"
 630   "xvcmpeq<VSs>. %x0,%x1,%x2"
 631   [(set_attr "type" "veccmp")])
 632
 633 (define_insn "*vsx_gt_<mode>_p"
 634   [(set (reg:CC 74)
 635         (unspec:CC
 636          [(gt:CC (match_operand:VSX_F 1 "vsx_register_operand" "<VSr>,?wa")
 637                  (match_operand:VSX_F 2 "vsx_register_operand" "<VSr>,?wa"))]
 638          UNSPEC_PREDICATE))
 639    (set (match_operand:VSX_F 0 "vsx_register_operand" "=<VSr>,?wa")
 640         (gt:VSX_F (match_dup 1)
 641                   (match_dup 2)))]
 642   "VECTOR_UNIT_VSX_P (<MODE>mode)"
 643   "xvcmpgt<VSs>. %x0,%x1,%x2"
 644   [(set_attr "type" "veccmp")])
 645
 646 (define_insn "*vsx_ge_<mode>_p"
 647   [(set (reg:CC 74)
 648         (unspec:CC
 649          [(ge:CC (match_operand:VSX_F 1 "vsx_register_operand" "<VSr>,?wa")
 650                  (match_operand:VSX_F 2 "vsx_register_operand" "<VSr>,?wa"))]
 651          UNSPEC_PREDICATE))
 652    (set (match_operand:VSX_F 0 "vsx_register_operand" "=<VSr>,?wa")
 653         (ge:VSX_F (match_dup 1)
 654                   (match_dup 2)))]
 655   "VECTOR_UNIT_VSX_P (<MODE>mode)"
 656   "xvcmpge<VSs>. %x0,%x1,%x2"
 657   [(set_attr "type" "veccmp")])
 658
 659 ;; Vector select
 660 (define_insn "*vsx_xxsel<mode>"
 661   [(set (match_operand:VSX_L 0 "vsx_register_operand" "=<VSr>,?wa")
 662         (if_then_else:VSX_L
 663          (ne:CC (match_operand:VSX_L 1 "vsx_register_operand" "<VSr>,wa")
 664                 (const_int 0))
 665          (match_operand:VSX_L 2 "vsx_register_operand" "<VSr>,wa")
 666          (match_operand:VSX_L 3 "vsx_register_operand" "<VSr>,wa")))]
 667   "VECTOR_MEM_VSX_P (<MODE>mode)"
 668   "xxsel %x0,%x3,%x2,%x1"
 669   [(set_attr "type" "vecperm")])
 670
 671 (define_insn "*vsx_xxsel<mode>_uns"
 672   [(set (match_operand:VSX_L 0 "vsx_register_operand" "=<VSr>,?wa")
 673         (if_then_else:VSX_L
 674          (ne:CCUNS (match_operand:VSX_L 1 "vsx_register_operand" "<VSr>,wa")
 675                    (const_int 0))
 676          (match_operand:VSX_L 2 "vsx_register_operand" "<VSr>,wa")
 677          (match_operand:VSX_L 3 "vsx_register_operand" "<VSr>,wa")))]
 678   "VECTOR_MEM_VSX_P (<MODE>mode)"
 679   "xxsel %x0,%x3,%x2,%x1"
 680   [(set_attr "type" "vecperm")])
 681
 682 ;; Copy sign
 683 (define_insn "vsx_copysign<mode>3"
 684   [(set (match_operand:VSX_B 0 "vsx_register_operand" "=<VSr>,?wa")
 685         (unspec:VSX_B
 686          [(match_operand:VSX_B 1 "vsx_register_operand" "<VSr>,wa")
 687           (match_operand:VSX_B 2 "vsx_register_operand" "<VSr>,wa")]
 688          UNSPEC_COPYSIGN))]
 689   "VECTOR_UNIT_VSX_P (<MODE>mode)"
 690   "x<VSv>cpsgn<VSs> %x0,%x2,%x1"
 691   [(set_attr "type" "<VStype_simple>")
 692    (set_attr "fp_type" "<VSfptype_simple>")])
 693
 694 ;; For the conversions, limit the register class for the integer value to be
 695 ;; the fprs because we don't want to add the altivec registers to movdi/movsi.
 696 ;; For the unsigned tests, there isn't a generic double -> unsigned conversion
 697 ;; in rs6000.md so don't test VECTOR_UNIT_VSX_P, just test against VSX.
 698 (define_insn "vsx_float<VSi><mode>2"
 699   [(set (match_operand:VSX_B 0 "vsx_register_operand" "=<VSr>,?wa")
 700         (float:VSX_B (match_operand:<VSI> 1 "vsx_register_operand" "<VSr2>,<VSr3>")))]
 701   "VECTOR_UNIT_VSX_P (<MODE>mode)"
 702   "x<VSv>cvsx<VSc><VSs> %x0,%x1"
 703   [(set_attr "type" "<VStype_simple>")
 704    (set_attr "fp_type" "<VSfptype_simple>")])
 705
 706 (define_insn "vsx_floatuns<VSi><mode>2"
 707   [(set (match_operand:VSX_B 0 "vsx_register_operand" "=<VSr>,?wa")
 708         (unsigned_float:VSX_B (match_operand:<VSI> 1 "vsx_register_operand" "<VSr2>,<VSr3>")))]
 709   "VECTOR_UNIT_VSX_P (<MODE>mode)"
 710   "x<VSv>cvux<VSc><VSs> %x0,%x1"
 711   [(set_attr "type" "<VStype_simple>")
 712    (set_attr "fp_type" "<VSfptype_simple>")])
 713
 714 (define_insn "vsx_fix_trunc<mode><VSi>2"
 715   [(set (match_operand:<VSI> 0 "vsx_register_operand" "=<VSr2>,?<VSr3>")
 716         (fix:<VSI> (match_operand:VSX_B 1 "vsx_register_operand" "<VSr>,wa")))]
 717   "VECTOR_UNIT_VSX_P (<MODE>mode)"
 718   "x<VSv>cv<VSs>sx<VSc>s %x0,%x1"
 719   [(set_attr "type" "<VStype_simple>")
 720    (set_attr "fp_type" "<VSfptype_simple>")])
 721
 722 (define_insn "vsx_fixuns_trunc<mode><VSi>2"
 723   [(set (match_operand:<VSI> 0 "vsx_register_operand" "=<VSr2>,?<VSr3>")
 724         (unsigned_fix:<VSI> (match_operand:VSX_B 1 "vsx_register_operand" "<VSr>,wa")))]
 725   "VECTOR_UNIT_VSX_P (<MODE>mode)"
 726   "x<VSv>cv<VSs>ux<VSc>s %x0,%x1"
 727   [(set_attr "type" "<VStype_simple>")
 728    (set_attr "fp_type" "<VSfptype_simple>")])
 729
 730 ;; Math rounding functions
 731 (define_insn "vsx_x<VSv>r<VSs>i"
 732   [(set (match_operand:VSX_B 0 "vsx_register_operand" "=<VSr>,?wa")
 733         (unspec:VSX_B [(match_operand:VSX_B 1 "vsx_register_operand" "<VSr>,wa")]
 734                       UNSPEC_VSX_ROUND_I))]
 735   "VECTOR_UNIT_VSX_P (<MODE>mode)"
 736   "x<VSv>r<VSs>i %x0,%x1"
 737   [(set_attr "type" "<VStype_simple>")
 738    (set_attr "fp_type" "<VSfptype_simple>")])
 739
 740 (define_insn "vsx_x<VSv>r<VSs>ic"
 741   [(set (match_operand:VSX_B 0 "vsx_register_operand" "=<VSr>,?wa")
 742         (unspec:VSX_B [(match_operand:VSX_B 1 "vsx_register_operand" "<VSr>,wa")]
 743                       UNSPEC_VSX_ROUND_IC))]
 744   "VECTOR_UNIT_VSX_P (<MODE>mode)"
 745   "x<VSv>r<VSs>ic %x0,%x1"
 746   [(set_attr "type" "<VStype_simple>")
 747    (set_attr "fp_type" "<VSfptype_simple>")])
 748
 749 (define_insn "vsx_btrunc<mode>2"
 750   [(set (match_operand:VSX_B 0 "vsx_register_operand" "=<VSr>,?wa")
 751         (fix:VSX_B (match_operand:VSX_B 1 "vsx_register_operand" "<VSr>,wa")))]
 752   "VECTOR_UNIT_VSX_P (<MODE>mode)"
 753   "x<VSv>r<VSs>iz %x0,%x1"
 754   [(set_attr "type" "<VStype_simple>")
 755    (set_attr "fp_type" "<VSfptype_simple>")])
 756
 757 (define_insn "*vsx_b2trunc<mode>2"
 758   [(set (match_operand:VSX_B 0 "vsx_register_operand" "=<VSr>,?wa")
 759         (unspec:VSX_B [(match_operand:VSX_B 1 "vsx_register_operand" "<VSr>,wa")]
 760                       UNSPEC_FRIZ))]
 761   "VECTOR_UNIT_VSX_P (<MODE>mode)"
 762   "x<VSv>r<VSs>iz %x0,%x1"
 763   [(set_attr "type" "<VStype_simple>")
 764    (set_attr "fp_type" "<VSfptype_simple>")])
 765
 766 (define_insn "vsx_floor<mode>2"
 767   [(set (match_operand:VSX_B 0 "vsx_register_operand" "=<VSr>,?wa")
 768         (unspec:VSX_B [(match_operand:VSX_B 1 "vsx_register_operand" "<VSr>,wa")]
 769                       UNSPEC_FRIM))]
 770   "VECTOR_UNIT_VSX_P (<MODE>mode)"
 771   "x<VSv>r<VSs>im %x0,%x1"
 772   [(set_attr "type" "<VStype_simple>")
 773    (set_attr "fp_type" "<VSfptype_simple>")])
 774
 775 (define_insn "vsx_ceil<mode>2"
 776   [(set (match_operand:VSX_B 0 "vsx_register_operand" "=<VSr>,?wa")
 777         (unspec:VSX_B [(match_operand:VSX_B 1 "vsx_register_operand" "<VSr>,wa")]
 778                       UNSPEC_FRIP))]
 779   "VECTOR_UNIT_VSX_P (<MODE>mode)"
 780   "x<VSv>r<VSs>ip %x0,%x1"
 781   [(set_attr "type" "<VStype_simple>")
 782    (set_attr "fp_type" "<VSfptype_simple>")])
 783
 784 \f
 785 ;; VSX convert to/from double vector
 786
 787 ;; Convert between single and double precision
 788 ;; Don't use xscvspdp and xscvdpsp for scalar conversions, since the normal
 789 ;; scalar single precision instructions internally use the double format.
 790 ;; Prefer the altivec registers, since we likely will need to do a vperm
 791 (define_insn "vsx_<VS_spdp_insn>"
 792   [(set (match_operand:<VS_spdp_res> 0 "vsx_register_operand" "=<VSr4>,?wa")
 793         (unspec:<VS_spdp_res> [(match_operand:VSX_SPDP 1 "vsx_register_operand" "<VSr5>,wa")]
 794                               UNSPEC_VSX_CVSPDP))]
 795   "VECTOR_UNIT_VSX_P (<MODE>mode)"
 796   "<VS_spdp_insn> %x0,%x1"
 797   [(set_attr "type" "<VS_spdp_type>")])
 798
 799 ;; xscvspdp, represent the scalar SF type as V4SF
 800 (define_insn "vsx_xscvspdp"
 801   [(set (match_operand:DF 0 "vsx_register_operand" "=ws,?wa")
 802         (unspec:DF [(match_operand:V4SF 1 "vsx_register_operand" "wa,wa")]
 803                    UNSPEC_VSX_CVSPDP))]
 804   "VECTOR_UNIT_VSX_P (DFmode)"
 805   "xscvspdp %x0,%x1"
 806   [(set_attr "type" "fp")])
 807
 808 ;; xscvdpsp used for splat'ing a scalar to V4SF, knowing that the internal SF
 809 ;; format of scalars is actually DF.
 810 (define_insn "vsx_xscvdpsp_scalar"
 811   [(set (match_operand:V4SF 0 "vsx_register_operand" "=wa")
 812         (unspec:V4SF [(match_operand:SF 1 "vsx_register_operand" "f")]
 813                      UNSPEC_VSX_CVSPDP))]
 814   "VECTOR_UNIT_VSX_P (DFmode)"
 815   "xscvdpsp %x0,%x1"
 816   [(set_attr "type" "fp")])
 817
 818 ;; Convert from 64-bit to 32-bit types
 819 ;; Note, favor the Altivec registers since the usual use of these instructions
 820 ;; is in vector converts and we need to use the Altivec vperm instruction.
 821
 822 (define_insn "vsx_xvcvdpsxws"
 823   [(set (match_operand:V4SI 0 "vsx_register_operand" "=v,?wa")
 824         (unspec:V4SI [(match_operand:V2DF 1 "vsx_register_operand" "wd,wa")]
 825                      UNSPEC_VSX_CVDPSXWS))]
 826   "VECTOR_UNIT_VSX_P (V2DFmode)"
 827   "xvcvdpsxws %x0,%x1"
 828   [(set_attr "type" "vecfloat")])
 829
 830 (define_insn "vsx_xvcvdpuxws"
 831   [(set (match_operand:V4SI 0 "vsx_register_operand" "=v,?wa")
 832         (unspec:V4SI [(match_operand:V2DF 1 "vsx_register_operand" "wd,wa")]
 833                      UNSPEC_VSX_CVDPUXWS))]
 834   "VECTOR_UNIT_VSX_P (V2DFmode)"
 835   "xvcvdpuxws %x0,%x1"
 836   [(set_attr "type" "vecfloat")])
 837
 838 (define_insn "vsx_xvcvsxdsp"
 839   [(set (match_operand:V4SI 0 "vsx_register_operand" "=wd,?wa")
 840         (unspec:V4SI [(match_operand:V2DF 1 "vsx_register_operand" "wf,wa")]
 841                      UNSPEC_VSX_CVSXDSP))]
 842   "VECTOR_UNIT_VSX_P (V2DFmode)"
 843   "xvcvsxdsp %x0,%x1"
 844   [(set_attr "type" "vecfloat")])
 845
 846 (define_insn "vsx_xvcvuxdsp"
 847   [(set (match_operand:V4SI 0 "vsx_register_operand" "=wd,?wa")
 848         (unspec:V4SI [(match_operand:V2DF 1 "vsx_register_operand" "wf,wa")]
 849                      UNSPEC_VSX_CVUXDSP))]
 850   "VECTOR_UNIT_VSX_P (V2DFmode)"
 851   "xvcvuxwdp %x0,%x1"
 852   [(set_attr "type" "vecfloat")])
 853
 854 ;; Convert from 32-bit to 64-bit types
 855 (define_insn "vsx_xvcvsxwdp"
 856   [(set (match_operand:V2DF 0 "vsx_register_operand" "=wd,?wa")
 857         (unspec:V2DF [(match_operand:V4SI 1 "vsx_register_operand" "wf,wa")]
 858                      UNSPEC_VSX_CVSXWDP))]
 859   "VECTOR_UNIT_VSX_P (V2DFmode)"
 860   "xvcvsxwdp %x0,%x1"
 861   [(set_attr "type" "vecfloat")])
 862
 863 (define_insn "vsx_xvcvuxwdp"
 864   [(set (match_operand:V2DF 0 "vsx_register_operand" "=wd,?wa")
 865         (unspec:V2DF [(match_operand:V4SI 1 "vsx_register_operand" "wf,wa")]
 866                      UNSPEC_VSX_CVUXWDP))]
 867   "VECTOR_UNIT_VSX_P (V2DFmode)"
 868   "xvcvuxwdp %x0,%x1"
 869   [(set_attr "type" "vecfloat")])
 870
 871 (define_insn "vsx_xvcvspsxds"
 872   [(set (match_operand:V2DI 0 "vsx_register_operand" "=v,?wa")
 873         (unspec:V2DI [(match_operand:V4SF 1 "vsx_register_operand" "wd,wa")]
 874                      UNSPEC_VSX_CVSPSXDS))]
 875   "VECTOR_UNIT_VSX_P (V2DFmode)"
 876   "xvcvspsxds %x0,%x1"
 877   [(set_attr "type" "vecfloat")])
 878
 879 (define_insn "vsx_xvcvspuxds"
 880   [(set (match_operand:V2DI 0 "vsx_register_operand" "=v,?wa")
 881         (unspec:V2DI [(match_operand:V4SF 1 "vsx_register_operand" "wd,wa")]
 882                      UNSPEC_VSX_CVSPUXDS))]
 883   "VECTOR_UNIT_VSX_P (V2DFmode)"
 884   "xvcvspuxds %x0,%x1"
 885   [(set_attr "type" "vecfloat")])
 886
 887 ;; Only optimize (float (fix x)) -> frz if we are in fast-math mode, since
 888 ;; since the xsrdpiz instruction does not truncate the value if the floating
 889 ;; point value is < LONG_MIN or > LONG_MAX.
 890 (define_insn "*vsx_float_fix_<mode>2"
 891   [(set (match_operand:VSX_DF 0 "vsx_register_operand" "=<VSr>,?wa")
 892         (float:VSX_DF
 893          (fix:<VSI>
 894           (match_operand:VSX_DF 1 "vsx_register_operand" "<VSr>,?wa"))))]
 895   "TARGET_HARD_FLOAT && TARGET_FPRS && TARGET_DOUBLE_FLOAT
 896    && VECTOR_UNIT_VSX_P (<MODE>mode) && flag_unsafe_math_optimizations
 897    && !flag_trapping_math && TARGET_FRIZ"
 898   "x<VSv>r<VSs>iz %x0,%x1"
 899   [(set_attr "type" "<VStype_simple>")
 900    (set_attr "fp_type" "<VSfptype_simple>")])
 901
 902 \f
 903 ;; Logical and permute operations
 904 (define_insn "*vsx_and<mode>3"
 905   [(set (match_operand:VSX_L 0 "vsx_register_operand" "=<VSr>,?wa")
 906         (and:VSX_L
 907          (match_operand:VSX_L 1 "vsx_register_operand" "<VSr>,?wa")
 908          (match_operand:VSX_L 2 "vsx_register_operand" "<VSr>,?wa")))]
 909   "VECTOR_MEM_VSX_P (<MODE>mode)"
 910   "xxland %x0,%x1,%x2"
 911   [(set_attr "type" "vecsimple")])
 912
 913 (define_insn "*vsx_ior<mode>3"
 914   [(set (match_operand:VSX_L 0 "vsx_register_operand" "=<VSr>,?wa")
 915         (ior:VSX_L (match_operand:VSX_L 1 "vsx_register_operand" "<VSr>,?wa")
 916                    (match_operand:VSX_L 2 "vsx_register_operand" "<VSr>,?wa")))]
 917   "VECTOR_MEM_VSX_P (<MODE>mode)"
 918   "xxlor %x0,%x1,%x2"
 919   [(set_attr "type" "vecsimple")])
 920
 921 (define_insn "*vsx_xor<mode>3"
 922   [(set (match_operand:VSX_L 0 "vsx_register_operand" "=<VSr>,?wa")
 923         (xor:VSX_L
 924          (match_operand:VSX_L 1 "vsx_register_operand" "<VSr>,?wa")
 925          (match_operand:VSX_L 2 "vsx_register_operand" "<VSr>,?wa")))]
 926   "VECTOR_MEM_VSX_P (<MODE>mode)"
 927   "xxlxor %x0,%x1,%x2"
 928   [(set_attr "type" "vecsimple")])
 929
 930 (define_insn "*vsx_one_cmpl<mode>2"
 931   [(set (match_operand:VSX_L 0 "vsx_register_operand" "=<VSr>,?wa")
 932         (not:VSX_L
 933          (match_operand:VSX_L 1 "vsx_register_operand" "<VSr>,?wa")))]
 934   "VECTOR_MEM_VSX_P (<MODE>mode)"
 935   "xxlnor %x0,%x1,%x1"
 936   [(set_attr "type" "vecsimple")])
 937
 938 (define_insn "*vsx_nor<mode>3"
 939   [(set (match_operand:VSX_L 0 "vsx_register_operand" "=<VSr>,?wa")
 940         (not:VSX_L
 941          (ior:VSX_L
 942           (match_operand:VSX_L 1 "vsx_register_operand" "<VSr>,?wa")
 943           (match_operand:VSX_L 2 "vsx_register_operand" "<VSr>,?wa"))))]
 944   "VECTOR_MEM_VSX_P (<MODE>mode)"
 945   "xxlnor %x0,%x1,%x2"
 946   [(set_attr "type" "vecsimple")])
 947
 948 (define_insn "*vsx_andc<mode>3"
 949   [(set (match_operand:VSX_L 0 "vsx_register_operand" "=<VSr>,?wa")
 950         (and:VSX_L
 951          (not:VSX_L
 952           (match_operand:VSX_L 2 "vsx_register_operand" "<VSr>,?wa"))
 953          (match_operand:VSX_L 1 "vsx_register_operand" "<VSr>,?wa")))]
 954   "VECTOR_MEM_VSX_P (<MODE>mode)"
 955   "xxlandc %x0,%x1,%x2"
 956   [(set_attr "type" "vecsimple")])
 957
 958 \f
 959 ;; Permute operations
 960
 961 ;; Build a V2DF/V2DI vector from two scalars
 962 (define_insn "vsx_concat_<mode>"
 963   [(set (match_operand:VSX_D 0 "vsx_register_operand" "=wd,?wa")
 964         (unspec:VSX_D
 965          [(match_operand:<VS_scalar> 1 "vsx_register_operand" "ws,wa")
 966           (match_operand:<VS_scalar> 2 "vsx_register_operand" "ws,wa")]
 967          UNSPEC_VSX_CONCAT))]
 968   "VECTOR_MEM_VSX_P (<MODE>mode)"
 969   "xxpermdi %x0,%x1,%x2,0"
 970   [(set_attr "type" "vecperm")])
 971
 972 ;; Special purpose concat using xxpermdi to glue two single precision values
 973 ;; together, relying on the fact that internally scalar floats are represented
 974 ;; as doubles.  This is used to initialize a V4SF vector with 4 floats
 975 (define_insn "vsx_concat_v2sf"
 976   [(set (match_operand:V2DF 0 "vsx_register_operand" "=wd,?wa")
 977         (unspec:V2DF
 978          [(match_operand:SF 1 "vsx_register_operand" "f,f")
 979           (match_operand:SF 2 "vsx_register_operand" "f,f")]
 980          UNSPEC_VSX_CONCAT))]
 981   "VECTOR_MEM_VSX_P (V2DFmode)"
 982   "xxpermdi %x0,%x1,%x2,0"
 983   [(set_attr "type" "vecperm")])
 984
 985 ;; Set the element of a V2DI/VD2F mode
 986 (define_insn "vsx_set_<mode>"
 987   [(set (match_operand:VSX_D 0 "vsx_register_operand" "=wd,?wa")
 988         (unspec:VSX_D [(match_operand:VSX_D 1 "vsx_register_operand" "wd,wa")
 989                        (match_operand:<VS_scalar> 2 "vsx_register_operand" "ws,wa")
 990                        (match_operand:QI 3 "u5bit_cint_operand" "i,i")]
 991                       UNSPEC_VSX_SET))]
 992   "VECTOR_MEM_VSX_P (<MODE>mode)"
 993 {
 994   if (INTVAL (operands[3]) == 0)
 995     return \"xxpermdi %x0,%x1,%x2,1\";
 996   else if (INTVAL (operands[3]) == 1)
 997     return \"xxpermdi %x0,%x2,%x1,0\";
 998   else
 999     gcc_unreachable ();
1000 }
1001   [(set_attr "type" "vecperm")])
1002
1003 ;; Extract a DF/DI element from V2DF/V2DI
1004 (define_insn "vsx_extract_<mode>"
1005   [(set (match_operand:<VS_scalar> 0 "vsx_register_operand" "=ws,d,?wa")
1006         (vec_select:<VS_scalar> (match_operand:VSX_D 1 "vsx_register_operand" "wd,wd,wa")
1007                        (parallel
1008                         [(match_operand:QI 2 "u5bit_cint_operand" "i,i,i")])))]
1009   "VECTOR_MEM_VSX_P (<MODE>mode)"
1010 {
1011   gcc_assert (UINTVAL (operands[2]) <= 1);
1012   operands[3] = GEN_INT (INTVAL (operands[2]) << 1);
1013   return \"xxpermdi %x0,%x1,%x1,%3\";
1014 }
1015   [(set_attr "type" "vecperm")])
1016
1017 ;; Optimize extracting element 0 from memory
1018 (define_insn "*vsx_extract_<mode>_zero"
1019   [(set (match_operand:<VS_scalar> 0 "vsx_register_operand" "=ws,d,?wa")
1020         (vec_select:<VS_scalar>
1021          (match_operand:VSX_D 1 "indexed_or_indirect_operand" "Z,Z,Z")
1022          (parallel [(const_int 0)])))]
1023   "VECTOR_MEM_VSX_P (<MODE>mode) && WORDS_BIG_ENDIAN"
1024   "lxsd%U1x %x0,%y1"
1025   [(set_attr "type" "fpload")
1026    (set_attr "length" "4")])
1027
1028 ;; General double word oriented permute, allow the other vector types for
1029 ;; optimizing the permute instruction.
1030 (define_insn "vsx_xxpermdi_<mode>"
1031   [(set (match_operand:VSX_L 0 "vsx_register_operand" "=wd,?wa")
1032         (unspec:VSX_L [(match_operand:VSX_L 1 "vsx_register_operand" "wd,wa")
1033                        (match_operand:VSX_L 2 "vsx_register_operand" "wd,wa")
1034                        (match_operand:QI 3 "u5bit_cint_operand" "i,i")]
1035                       UNSPEC_VSX_XXPERMDI))]
1036   "VECTOR_MEM_VSX_P (<MODE>mode)"
1037   "xxpermdi %x0,%x1,%x2,%3"
1038   [(set_attr "type" "vecperm")])
1039
1040 ;; Varient of xxpermdi that is emitted by the vec_interleave functions
1041 (define_insn "*vsx_xxpermdi2_<mode>"
1042   [(set (match_operand:VSX_D 0 "vsx_register_operand" "=wd")
1043         (vec_concat:VSX_D
1044          (vec_select:<VS_scalar>
1045           (match_operand:VSX_D 1 "vsx_register_operand" "wd")
1046           (parallel
1047            [(match_operand:QI 2 "u5bit_cint_operand" "i")]))
1048          (vec_select:<VS_scalar>
1049           (match_operand:VSX_D 3 "vsx_register_operand" "wd")
1050           (parallel
1051            [(match_operand:QI 4 "u5bit_cint_operand" "i")]))))]
1052   "VECTOR_MEM_VSX_P (<MODE>mode)"
1053 {
1054   gcc_assert ((UINTVAL (operands[2]) <= 1) && (UINTVAL (operands[4]) <= 1));
1055   operands[5] = GEN_INT (((INTVAL (operands[2]) & 1) << 1)
1056                          | (INTVAL (operands[4]) & 1));
1057   return \"xxpermdi %x0,%x1,%x3,%5\";
1058 }
1059   [(set_attr "type" "vecperm")])
1060
1061 ;; V2DF/V2DI splat
1062 (define_insn "vsx_splat_<mode>"
1063   [(set (match_operand:VSX_D 0 "vsx_register_operand" "=wd,wd,wd,?wa,?wa,?wa")
1064         (vec_duplicate:VSX_D
1065          (match_operand:<VS_scalar> 1 "input_operand" "ws,f,Z,wa,wa,Z")))]
1066   "VECTOR_MEM_VSX_P (<MODE>mode)"
1067   "@
1068    xxpermdi %x0,%x1,%x1,0
1069    xxpermdi %x0,%x1,%x1,0
1070    lxvdsx %x0,%y1
1071    xxpermdi %x0,%x1,%x1,0
1072    xxpermdi %x0,%x1,%x1,0
1073    lxvdsx %x0,%y1"
1074   [(set_attr "type" "vecperm,vecperm,vecload,vecperm,vecperm,vecload")])
1075
1076 ;; V4SF/V4SI splat
1077 (define_insn "vsx_xxspltw_<mode>"
1078   [(set (match_operand:VSX_W 0 "vsx_register_operand" "=wf,?wa")
1079         (vec_duplicate:VSX_W
1080          (vec_select:<VS_scalar>
1081           (match_operand:VSX_W 1 "vsx_register_operand" "wf,wa")
1082           (parallel
1083            [(match_operand:QI 2 "u5bit_cint_operand" "i,i")]))))]
1084   "VECTOR_MEM_VSX_P (<MODE>mode)"
1085   "xxspltw %x0,%x1,%2"
1086   [(set_attr "type" "vecperm")])
1087
1088 ;; V4SF/V4SI interleave
1089 (define_insn "vsx_xxmrghw_<mode>"
1090   [(set (match_operand:VSX_W 0 "vsx_register_operand" "=wf,?wa")
1091         (vec_merge:VSX_W
1092          (vec_select:VSX_W
1093           (match_operand:VSX_W 1 "vsx_register_operand" "wf,wa")
1094           (parallel [(const_int 0)
1095                      (const_int 2)
1096                      (const_int 1)
1097                      (const_int 3)]))
1098          (vec_select:VSX_W
1099           (match_operand:VSX_W 2 "vsx_register_operand" "wf,wa")
1100           (parallel [(const_int 2)
1101                      (const_int 0)
1102                      (const_int 3)
1103                      (const_int 1)]))
1104          (const_int 5)))]
1105   "VECTOR_MEM_VSX_P (<MODE>mode)"
1106   "xxmrghw %x0,%x1,%x2"
1107   [(set_attr "type" "vecperm")])
1108
1109 (define_insn "vsx_xxmrglw_<mode>"
1110   [(set (match_operand:VSX_W 0 "vsx_register_operand" "=wf,?wa")
1111         (vec_merge:VSX_W
1112          (vec_select:VSX_W
1113           (match_operand:VSX_W 1 "vsx_register_operand" "wf,wa")
1114           (parallel [(const_int 2)
1115                      (const_int 0)
1116                      (const_int 3)
1117                      (const_int 1)]))
1118          (vec_select:VSX_W
1119           (match_operand:VSX_W 2 "vsx_register_operand" "wf,?wa")
1120           (parallel [(const_int 0)
1121                      (const_int 2)
1122                      (const_int 1)
1123                      (const_int 3)]))
1124          (const_int 5)))]
1125   "VECTOR_MEM_VSX_P (<MODE>mode)"
1126   "xxmrglw %x0,%x1,%x2"
1127   [(set_attr "type" "vecperm")])
1128
1129 ;; Shift left double by word immediate
1130 (define_insn "vsx_xxsldwi_<mode>"
1131   [(set (match_operand:VSX_L 0 "vsx_register_operand" "=wa")
1132         (unspec:VSX_L [(match_operand:VSX_L 1 "vsx_register_operand" "wa")
1133                        (match_operand:VSX_L 2 "vsx_register_operand" "wa")
1134                        (match_operand:QI 3 "u5bit_cint_operand" "i")]
1135                       UNSPEC_VSX_SLDWI))]
1136   "VECTOR_MEM_VSX_P (<MODE>mode)"
1137   "xxsldwi %x0,%x1,%x2,%3"
1138   [(set_attr "type" "vecperm")])