lib/Target/X86/X86InstrFragmentsSIMD.td

   1 //===-- X86InstrFragmentsSIMD.td - x86 SIMD ISA ------------*- tablegen -*-===//
   2 //
   3 //                     The LLVM Compiler Infrastructure
   4 //
   5 // This file is distributed under the University of Illinois Open Source
   6 // License. See LICENSE.TXT for details.
   7 //
   8 //===----------------------------------------------------------------------===//
   9 //
  10 // This file provides pattern fragments useful for SIMD instructions.
  11 //
  12 //===----------------------------------------------------------------------===//
  13
  14 //===----------------------------------------------------------------------===//
  15 // MMX specific DAG Nodes.
  16 //===----------------------------------------------------------------------===//
  17
  18 // Low word of MMX to GPR.
  19 def MMX_X86movd2w : SDNode<"X86ISD::MMX_MOVD2W", SDTypeProfile<1, 1,
  20                             [SDTCisVT<0, i32>, SDTCisVT<1, x86mmx>]>>;
  21 // GPR to low word of MMX.
  22 def MMX_X86movw2d : SDNode<"X86ISD::MMX_MOVW2D", SDTypeProfile<1, 1,
  23                             [SDTCisVT<0, x86mmx>, SDTCisVT<1, i32>]>>;
  24
  25 //===----------------------------------------------------------------------===//
  26 // MMX Pattern Fragments
  27 //===----------------------------------------------------------------------===//
  28
  29 def load_mmx : PatFrag<(ops node:$ptr), (x86mmx (load node:$ptr))>;
  30 def load_mvmmx : PatFrag<(ops node:$ptr),
  31                          (x86mmx (MMX_X86movw2d (load node:$ptr)))>;
  32 def bc_mmx  : PatFrag<(ops node:$in), (x86mmx  (bitconvert node:$in))>;
  33
  34 //===----------------------------------------------------------------------===//
  35 // SSE specific DAG Nodes.
  36 //===----------------------------------------------------------------------===//
  37
  38 def SDTX86VFCMP : SDTypeProfile<1, 3, [SDTCisInt<0>, SDTCisSameAs<1, 2>,
  39                                        SDTCisFP<1>, SDTCisVT<3, i8>,
  40                                        SDTCisVec<1>]>;
  41
  42 def X86fmin    : SDNode<"X86ISD::FMIN",      SDTFPBinOp>;
  43 def X86fmax    : SDNode<"X86ISD::FMAX",      SDTFPBinOp>;
  44
  45 // Commutative and Associative FMIN and FMAX.
  46 def X86fminc    : SDNode<"X86ISD::FMINC", SDTFPBinOp,
  47     [SDNPCommutative, SDNPAssociative]>;
  48 def X86fmaxc    : SDNode<"X86ISD::FMAXC", SDTFPBinOp,
  49     [SDNPCommutative, SDNPAssociative]>;
  50
  51 def X86fand    : SDNode<"X86ISD::FAND",      SDTFPBinOp,
  52                         [SDNPCommutative, SDNPAssociative]>;
  53 def X86for     : SDNode<"X86ISD::FOR",       SDTFPBinOp,
  54                         [SDNPCommutative, SDNPAssociative]>;
  55 def X86fxor    : SDNode<"X86ISD::FXOR",      SDTFPBinOp,
  56                         [SDNPCommutative, SDNPAssociative]>;
  57 def X86fandn   : SDNode<"X86ISD::FANDN",     SDTFPBinOp,
  58                         [SDNPCommutative, SDNPAssociative]>;
  59 def X86frsqrt  : SDNode<"X86ISD::FRSQRT",    SDTFPUnaryOp>;
  60 def X86frcp    : SDNode<"X86ISD::FRCP",      SDTFPUnaryOp>;
  61 def X86frsqrt14s: SDNode<"X86ISD::FRSQRT",  SDTFPBinOp>;
  62 def X86frcp14s : SDNode<"X86ISD::FRCP",    SDTFPBinOp>;
  63 def X86fgetsign: SDNode<"X86ISD::FGETSIGNx86",SDTFPToIntOp>;
  64 def X86fhadd   : SDNode<"X86ISD::FHADD",     SDTFPBinOp>;
  65 def X86fhsub   : SDNode<"X86ISD::FHSUB",     SDTFPBinOp>;
  66 def X86hadd    : SDNode<"X86ISD::HADD",      SDTIntBinOp>;
  67 def X86hsub    : SDNode<"X86ISD::HSUB",      SDTIntBinOp>;
  68 def X86comi    : SDNode<"X86ISD::COMI",      SDTX86CmpTest>;
  69 def X86ucomi   : SDNode<"X86ISD::UCOMI",     SDTX86CmpTest>;
  70 def X86cmps    : SDNode<"X86ISD::FSETCC",     SDTX86Cmps>;
  71 //def X86cmpsd   : SDNode<"X86ISD::FSETCCsd",    SDTX86Cmpsd>;
  72 def X86cvtdq2pd: SDNode<"X86ISD::CVTDQ2PD",
  73                  SDTypeProfile<1, 1, [SDTCisVT<0, v2f64>,
  74                                       SDTCisVT<1, v4i32>]>>;
  75 def X86cvtudq2pd: SDNode<"X86ISD::CVTUDQ2PD",
  76                  SDTypeProfile<1, 1, [SDTCisVT<0, v2f64>,
  77                                       SDTCisVT<1, v4i32>]>>;
  78 def X86pshufb  : SDNode<"X86ISD::PSHUFB",
  79                  SDTypeProfile<1, 2, [SDTCisVec<0>, SDTCisSameAs<0,1>,
  80                                       SDTCisSameAs<0,2>]>>;
  81 def X86psadbw  : SDNode<"X86ISD::PSADBW",
  82                  SDTypeProfile<1, 2, [SDTCisVec<0>, SDTCisSameAs<0,1>,
  83                                       SDTCisSameAs<0,2>]>>;
  84 def X86dbpsadbw : SDNode<"X86ISD::DBPSADBW",
  85                   SDTypeProfile<1, 3, [SDTCisVec<0>, SDTCisVec<1>,
  86                                        SDTCisSameAs<1,2>, SDTCisInt<3>]>>;
  87 def X86andnp   : SDNode<"X86ISD::ANDNP",
  88                  SDTypeProfile<1, 2, [SDTCisVec<0>, SDTCisSameAs<0,1>,
  89                                       SDTCisSameAs<0,2>]>>;
  90 def X86psign   : SDNode<"X86ISD::PSIGN",
  91                  SDTypeProfile<1, 2, [SDTCisVec<0>, SDTCisSameAs<0,1>,
  92                                       SDTCisSameAs<0,2>]>>;
  93 def X86pextrb  : SDNode<"X86ISD::PEXTRB",
  94                  SDTypeProfile<1, 2, [SDTCisVT<0, i32>, SDTCisPtrTy<2>]>>;
  95 def X86pextrw  : SDNode<"X86ISD::PEXTRW",
  96                  SDTypeProfile<1, 2, [SDTCisVT<0, i32>, SDTCisPtrTy<2>]>>;
  97 def X86pinsrb  : SDNode<"X86ISD::PINSRB",
  98                  SDTypeProfile<1, 3, [SDTCisVT<0, v16i8>, SDTCisSameAs<0,1>,
  99                                       SDTCisVT<2, i32>, SDTCisPtrTy<3>]>>;
 100 def X86pinsrw  : SDNode<"X86ISD::PINSRW",
 101                  SDTypeProfile<1, 3, [SDTCisVT<0, v8i16>, SDTCisSameAs<0,1>,
 102                                       SDTCisVT<2, i32>, SDTCisPtrTy<3>]>>;
 103 def X86insertps : SDNode<"X86ISD::INSERTPS",
 104                  SDTypeProfile<1, 3, [SDTCisVT<0, v4f32>, SDTCisSameAs<0,1>,
 105                                       SDTCisVT<2, v4f32>, SDTCisVT<3, i8>]>>;
 106 def X86vzmovl  : SDNode<"X86ISD::VZEXT_MOVL",
 107                  SDTypeProfile<1, 1, [SDTCisSameAs<0,1>]>>;
 108
 109 def X86vzload  : SDNode<"X86ISD::VZEXT_LOAD", SDTLoad,
 110                         [SDNPHasChain, SDNPMayLoad, SDNPMemOperand]>;
 111
 112 def X86vzext   : SDNode<"X86ISD::VZEXT",
 113                          SDTypeProfile<1, 1, [SDTCisVec<0>, SDTCisVec<1>,
 114                                               SDTCisInt<0>, SDTCisInt<1>,
 115                                               SDTCisOpSmallerThanOp<1, 0>]>>;
 116
 117 def X86vsext   : SDNode<"X86ISD::VSEXT",
 118                          SDTypeProfile<1, 1, [SDTCisVec<0>, SDTCisVec<1>,
 119                                               SDTCisInt<0>, SDTCisInt<1>,
 120                                               SDTCisOpSmallerThanOp<1, 0>]>>;
 121
 122 def SDTVtrunc    : SDTypeProfile<1, 1, [SDTCisVec<0>, SDTCisVec<1>,
 123                                        SDTCisInt<0>, SDTCisInt<1>,
 124                                        SDTCisOpSmallerThanOp<0, 1>]>;
 125
 126 def X86vtrunc    : SDNode<"X86ISD::VTRUNC",   SDTVtrunc>;
 127 def X86vtruncs   : SDNode<"X86ISD::VTRUNCS",  SDTVtrunc>;
 128 def X86vtruncus  : SDNode<"X86ISD::VTRUNCUS", SDTVtrunc>;
 129
 130 def X86trunc    : SDNode<"X86ISD::TRUNC",
 131                          SDTypeProfile<1, 1, [SDTCisInt<0>, SDTCisInt<1>,
 132                                               SDTCisOpSmallerThanOp<0, 1>]>>;
 133 def X86vfpext  : SDNode<"X86ISD::VFPEXT",
 134                         SDTypeProfile<1, 1, [SDTCisVec<0>, SDTCisVec<1>,
 135                                              SDTCisFP<0>, SDTCisFP<1>,
 136                                              SDTCisOpSmallerThanOp<1, 0>]>>;
 137 def X86vfpround: SDNode<"X86ISD::VFPROUND",
 138                         SDTypeProfile<1, 1, [SDTCisVec<0>, SDTCisVec<1>,
 139                                              SDTCisFP<0>, SDTCisFP<1>,
 140                                              SDTCisOpSmallerThanOp<0, 1>]>>;
 141
 142 def X86fround: SDNode<"X86ISD::VFPROUND",
 143                         SDTypeProfile<1, 2, [SDTCisFP<0>, SDTCisFP<1>,SDTCisFP<2>,
 144                                              SDTCVecEltisVT<0, f32>,
 145                                              SDTCVecEltisVT<1, f64>,
 146                                              SDTCVecEltisVT<2, f64>,
 147                                              SDTCisOpSmallerThanOp<0, 1>]>>;
 148 def X86froundRnd: SDNode<"X86ISD::VFPROUND",
 149                         SDTypeProfile<1, 3, [SDTCisFP<0>, SDTCisFP<1>,SDTCisFP<2>,
 150                                              SDTCVecEltisVT<0, f32>,
 151                                              SDTCVecEltisVT<1, f64>,
 152                                              SDTCVecEltisVT<2, f64>,
 153                                              SDTCisOpSmallerThanOp<0, 1>,
 154                                              SDTCisInt<3>]>>;
 155
 156 def X86fpext  : SDNode<"X86ISD::VFPEXT",
 157                         SDTypeProfile<1, 2, [SDTCisFP<0>, SDTCisFP<1>,SDTCisFP<2>,
 158                                              SDTCVecEltisVT<0, f64>,
 159                                              SDTCVecEltisVT<1, f32>,
 160                                              SDTCVecEltisVT<2, f32>,
 161                                              SDTCisOpSmallerThanOp<1, 0>]>>;
 162
 163 def X86fpextRnd  : SDNode<"X86ISD::VFPEXT",
 164                         SDTypeProfile<1, 3, [SDTCisFP<0>, SDTCisFP<1>,SDTCisFP<2>,
 165                                              SDTCVecEltisVT<0, f64>,
 166                                              SDTCVecEltisVT<1, f32>,
 167                                              SDTCVecEltisVT<2, f32>,
 168                                              SDTCisOpSmallerThanOp<1, 0>,
 169                                              SDTCisInt<3>]>>;
 170
 171 def X86vshldq  : SDNode<"X86ISD::VSHLDQ",    SDTIntShiftOp>;
 172 def X86vshrdq  : SDNode<"X86ISD::VSRLDQ",    SDTIntShiftOp>;
 173 def X86cmpp    : SDNode<"X86ISD::CMPP",      SDTX86VFCMP>;
 174 def X86pcmpeq  : SDNode<"X86ISD::PCMPEQ", SDTIntBinOp, [SDNPCommutative]>;
 175 def X86pcmpgt  : SDNode<"X86ISD::PCMPGT", SDTIntBinOp>;
 176
 177 def X86IntCmpMask : SDTypeProfile<1, 2,
 178     [SDTCisVec<0>, SDTCisSameAs<1, 2>, SDTCisInt<1>]>;
 179 def X86pcmpeqm  : SDNode<"X86ISD::PCMPEQM", X86IntCmpMask, [SDNPCommutative]>;
 180 def X86pcmpgtm  : SDNode<"X86ISD::PCMPGTM", X86IntCmpMask>;
 181
 182 def X86CmpMaskCC :
 183       SDTypeProfile<1, 3, [SDTCisVec<0>, SDTCVecEltisVT<0, i1>,
 184                        SDTCisVec<1>, SDTCisSameAs<2, 1>,
 185                        SDTCisSameNumEltsAs<0, 1>, SDTCisVT<3, i8>]>;
 186 def X86CmpMaskCCRound :
 187       SDTypeProfile<1, 4, [SDTCisVec<0>,SDTCVecEltisVT<0, i1>,
 188                        SDTCisVec<1>, SDTCisSameAs<2, 1>,
 189                        SDTCisSameNumEltsAs<0, 1>, SDTCisVT<3, i8>,
 190                        SDTCisInt<4>]>;
 191 def X86CmpMaskCCScalar :
 192       SDTypeProfile<1, 3, [SDTCisInt<0>, SDTCisSameAs<1, 2>, SDTCisVT<3, i8>]>;
 193
 194 def X86CmpMaskCCScalarRound :
 195       SDTypeProfile<1, 4, [SDTCisInt<0>, SDTCisSameAs<1, 2>, SDTCisVT<3, i8>,
 196                            SDTCisInt<4>]>;
 197
 198 def X86cmpm     : SDNode<"X86ISD::CMPM",     X86CmpMaskCC>;
 199 def X86cmpmRnd  : SDNode<"X86ISD::CMPM_RND", X86CmpMaskCCRound>;
 200 def X86cmpmu    : SDNode<"X86ISD::CMPMU",    X86CmpMaskCC>;
 201 def X86cmpms    : SDNode<"X86ISD::FSETCC",   X86CmpMaskCCScalar>;
 202 def X86cmpmsRnd : SDNode<"X86ISD::FSETCC",   X86CmpMaskCCScalarRound>;
 203
 204 def X86vshl    : SDNode<"X86ISD::VSHL",
 205                         SDTypeProfile<1, 2, [SDTCisVec<0>, SDTCisSameAs<0,1>,
 206                                       SDTCisVec<2>]>>;
 207 def X86vsrl    : SDNode<"X86ISD::VSRL",
 208                         SDTypeProfile<1, 2, [SDTCisVec<0>, SDTCisSameAs<0,1>,
 209                                       SDTCisVec<2>]>>;
 210 def X86vsra    : SDNode<"X86ISD::VSRA",
 211                         SDTypeProfile<1, 2, [SDTCisVec<0>, SDTCisSameAs<0,1>,
 212                                       SDTCisVec<2>]>>;
 213
 214 def X86vshli   : SDNode<"X86ISD::VSHLI", SDTIntShiftOp>;
 215 def X86vsrli   : SDNode<"X86ISD::VSRLI", SDTIntShiftOp>;
 216 def X86vsrai   : SDNode<"X86ISD::VSRAI", SDTIntShiftOp>;
 217
 218 def X86vpshl   : SDNode<"X86ISD::VPSHL",
 219                         SDTypeProfile<1, 2, [SDTCisVec<0>, SDTCisSameAs<0,1>,
 220                                       SDTCisVec<2>]>>;
 221 def X86vpsha   : SDNode<"X86ISD::VPSHA",
 222                         SDTypeProfile<1, 2, [SDTCisVec<0>, SDTCisSameAs<0,1>,
 223                                       SDTCisVec<2>]>>;
 224
 225 def SDTX86CmpPTest : SDTypeProfile<1, 2, [SDTCisVT<0, i32>,
 226                                           SDTCisVec<1>,
 227                                           SDTCisSameAs<2, 1>]>;
 228 def X86addus   : SDNode<"X86ISD::ADDUS", SDTIntBinOp>;
 229 def X86subus   : SDNode<"X86ISD::SUBUS", SDTIntBinOp>;
 230 def X86adds    : SDNode<"X86ISD::ADDS", SDTIntBinOp>;
 231 def X86subs    : SDNode<"X86ISD::SUBS", SDTIntBinOp>;
 232 def X86mulhrs  : SDNode<"X86ISD::MULHRS" , SDTIntBinOp>;
 233 def X86avg     : SDNode<"X86ISD::AVG" , SDTIntBinOp>;
 234 def X86ptest   : SDNode<"X86ISD::PTEST", SDTX86CmpPTest>;
 235 def X86testp   : SDNode<"X86ISD::TESTP", SDTX86CmpPTest>;
 236 def X86kortest : SDNode<"X86ISD::KORTEST", SDTX86CmpPTest>;
 237 def X86ktest   : SDNode<"X86ISD::KTEST", SDTX86CmpPTest>;
 238 def X86testm   : SDNode<"X86ISD::TESTM", SDTypeProfile<1, 2, [SDTCisVec<0>,
 239                                           SDTCisVec<1>, SDTCisSameAs<2, 1>,
 240                                           SDTCVecEltisVT<0, i1>,
 241                                           SDTCisSameNumEltsAs<0, 1>]>>;
 242 def X86testnm  : SDNode<"X86ISD::TESTNM", SDTypeProfile<1, 2, [SDTCisVec<0>,
 243                                           SDTCisVec<1>, SDTCisSameAs<2, 1>,
 244                                           SDTCVecEltisVT<0, i1>,
 245                                           SDTCisSameNumEltsAs<0, 1>]>>;
 246 def X86select  : SDNode<"X86ISD::SELECT"     , SDTSelect>;
 247
 248 def X86pmuludq : SDNode<"X86ISD::PMULUDQ",
 249                         SDTypeProfile<1, 2, [SDTCisVec<0>, SDTCisVec<1>,
 250                                       SDTCisSameAs<1,2>]>>;
 251 def X86pmuldq  : SDNode<"X86ISD::PMULDQ",
 252                          SDTypeProfile<1, 2, [SDTCisVec<0>, SDTCisVec<1>,
 253                                        SDTCisSameAs<1,2>]>>;
 254
 255 def X86extrqi : SDNode<"X86ISD::EXTRQI",
 256                   SDTypeProfile<1, 3, [SDTCisVT<0, v2i64>, SDTCisSameAs<0,1>,
 257                                        SDTCisVT<2, i8>, SDTCisVT<3, i8>]>>;
 258 def X86insertqi : SDNode<"X86ISD::INSERTQI",
 259                     SDTypeProfile<1, 4, [SDTCisVT<0, v2i64>, SDTCisSameAs<0,1>,
 260                                          SDTCisSameAs<1,2>, SDTCisVT<3, i8>,
 261                                          SDTCisVT<4, i8>]>>;
 262
 263 // Specific shuffle nodes - At some point ISD::VECTOR_SHUFFLE will always get
 264 // translated into one of the target nodes below during lowering.
 265 // Note: this is a work in progress...
 266 def SDTShuff1Op : SDTypeProfile<1, 1, [SDTCisVec<0>, SDTCisSameAs<0,1>]>;
 267 def SDTShuff2Op : SDTypeProfile<1, 2, [SDTCisVec<0>, SDTCisSameAs<0,1>,
 268                                 SDTCisSameAs<0,2>]>;
 269 def SDTShuff3Op : SDTypeProfile<1, 3, [SDTCisVec<0>, SDTCisSameAs<0,1>,
 270                                 SDTCisSameAs<0,2>, SDTCisSameAs<0,3>]>;
 271
 272 def SDTShuff2OpM : SDTypeProfile<1, 2, [SDTCisVec<0>, SDTCisSameAs<0,1>,
 273                                         SDTCisVec<2>]>;
 274 def SDTShuff2OpI : SDTypeProfile<1, 2, [SDTCisVec<0>,
 275                                  SDTCisSameAs<0,1>, SDTCisInt<2>]>;
 276 def SDTShuff3OpI : SDTypeProfile<1, 3, [SDTCisVec<0>, SDTCisSameAs<0,1>,
 277                                  SDTCisSameAs<0,2>, SDTCisInt<3>]>;
 278 def SDTFPBinOpImmRound: SDTypeProfile<1, 4, [SDTCisVec<0>, SDTCisSameAs<0,1>,
 279                              SDTCisSameAs<0,2>, SDTCisInt<3>, SDTCisInt<4>]>;
 280 def SDTFPUnaryOpImmRound: SDTypeProfile<1, 3, [SDTCisVec<0>, SDTCisSameAs<0,1>,
 281                               SDTCisInt<2>, SDTCisInt<3>]>;
 282
 283 def SDTVBroadcast  : SDTypeProfile<1, 1, [SDTCisVec<0>]>;
 284 def SDTVBroadcastm : SDTypeProfile<1, 2, [SDTCisVec<0>, SDTCisVec<1>]>;
 285
 286 def SDTBlend : SDTypeProfile<1, 3, [SDTCisVec<0>, SDTCisSameAs<0,1>,
 287                              SDTCisSameAs<1,2>, SDTCisVT<3, i8>]>;
 288
 289 def SDTFPBinOpRound : SDTypeProfile<1, 3, [      // fadd_round, fmul_round, etc.
 290   SDTCisSameAs<0, 1>, SDTCisSameAs<0, 2>, SDTCisFP<0>, SDTCisInt<3>]>;
 291
 292 def SDTFPUnaryOpRound : SDTypeProfile<1, 2, [      // fsqrt_round, fgetexp_round, etc.
 293   SDTCisSameAs<0, 1>, SDTCisFP<0>, SDTCisInt<2>]>;
 294
 295 def SDTFma : SDTypeProfile<1, 3, [SDTCisSameAs<0,1>,
 296                            SDTCisSameAs<1,2>, SDTCisSameAs<1,3>]>;
 297 def SDTFmaRound : SDTypeProfile<1, 4, [SDTCisSameAs<0,1>,
 298                            SDTCisSameAs<1,2>, SDTCisSameAs<1,3>, SDTCisInt<4>]>;
 299 def STDFp1SrcRm : SDTypeProfile<1, 2, [SDTCisSameAs<0,1>,
 300                            SDTCisVec<0>, SDTCisInt<2>]>;
 301 def STDFp2SrcRm : SDTypeProfile<1, 3, [SDTCisSameAs<0,1>,
 302                            SDTCisVec<0>, SDTCisInt<3>]>;
 303 def STDFp3SrcRm : SDTypeProfile<1, 4, [SDTCisSameAs<0,1>,
 304                            SDTCisVec<0>, SDTCisInt<3>, SDTCisInt<4>]>;
 305
 306 def X86PAlignr : SDNode<"X86ISD::PALIGNR", SDTShuff3OpI>;
 307 def X86VAlign  : SDNode<"X86ISD::VALIGN", SDTShuff3OpI>;
 308
 309 def X86Abs      : SDNode<"X86ISD::ABS", SDTIntUnaryOp>;
 310 def X86Conflict : SDNode<"X86ISD::CONFLICT", SDTIntUnaryOp>;
 311
 312 def X86PShufd  : SDNode<"X86ISD::PSHUFD", SDTShuff2OpI>;
 313 def X86PShufhw : SDNode<"X86ISD::PSHUFHW", SDTShuff2OpI>;
 314 def X86PShuflw : SDNode<"X86ISD::PSHUFLW", SDTShuff2OpI>;
 315
 316 def X86Shufp   : SDNode<"X86ISD::SHUFP", SDTShuff3OpI>;
 317 def X86Shuf128 : SDNode<"X86ISD::SHUF128", SDTShuff3OpI>;
 318
 319 def X86Movddup  : SDNode<"X86ISD::MOVDDUP", SDTShuff1Op>;
 320 def X86Movshdup : SDNode<"X86ISD::MOVSHDUP", SDTShuff1Op>;
 321 def X86Movsldup : SDNode<"X86ISD::MOVSLDUP", SDTShuff1Op>;
 322
 323 def X86Movsd : SDNode<"X86ISD::MOVSD", SDTShuff2Op>;
 324 def X86Movss : SDNode<"X86ISD::MOVSS", SDTShuff2Op>;
 325
 326 def X86Movlhps : SDNode<"X86ISD::MOVLHPS", SDTShuff2Op>;
 327 def X86Movlhpd : SDNode<"X86ISD::MOVLHPD", SDTShuff2Op>;
 328 def X86Movhlps : SDNode<"X86ISD::MOVHLPS", SDTShuff2Op>;
 329
 330 def X86Movlps : SDNode<"X86ISD::MOVLPS", SDTShuff2Op>;
 331 def X86Movlpd : SDNode<"X86ISD::MOVLPD", SDTShuff2Op>;
 332
 333 def SDTPack : SDTypeProfile<1, 2, [SDTCisVec<0>, SDTCisVec<1>, SDTCisSameAs<2, 1>]>;
 334 def X86Packss : SDNode<"X86ISD::PACKSS", SDTPack>;
 335 def X86Packus : SDNode<"X86ISD::PACKUS", SDTPack>;
 336
 337 def X86Unpckl : SDNode<"X86ISD::UNPCKL", SDTShuff2Op>;
 338 def X86Unpckh : SDNode<"X86ISD::UNPCKH", SDTShuff2Op>;
 339
 340 def X86vpmaddubsw  : SDNode<"X86ISD::VPMADDUBSW" , SDTPack>;
 341 def X86vpmaddwd    : SDNode<"X86ISD::VPMADDWD"   , SDTPack>;
 342
 343 def X86VPermilpv  : SDNode<"X86ISD::VPERMILPV", SDTShuff2OpM>;
 344 def X86VPermilpi  : SDNode<"X86ISD::VPERMILPI", SDTShuff2OpI>;
 345 def X86VPermv     : SDNode<"X86ISD::VPERMV",    SDTShuff2Op>;
 346 def X86VPermi     : SDNode<"X86ISD::VPERMI",    SDTShuff2OpI>;
 347 def X86VPermv3    : SDNode<"X86ISD::VPERMV3",   SDTShuff3Op>;
 348 def X86VPermiv3   : SDNode<"X86ISD::VPERMIV3",  SDTShuff3Op>;
 349
 350 def X86VPerm2x128 : SDNode<"X86ISD::VPERM2X128", SDTShuff3OpI>;
 351
 352 def X86VFixupimm   : SDNode<"X86ISD::VFIXUPIMM", SDTFPBinOpImmRound>;
 353 def X86VRange      : SDNode<"X86ISD::VRANGE",    SDTFPBinOpImmRound>;
 354 def X86VReduce     : SDNode<"X86ISD::VREDUCE",   SDTFPUnaryOpImmRound>;
 355 def X86VRndScale   : SDNode<"X86ISD::VRNDSCALE", SDTFPUnaryOpImmRound>;
 356 def X86VGetMant    : SDNode<"X86ISD::VGETMANT",  SDTFPUnaryOpImmRound>;
 357 def X86Vfpclass    : SDNode<"X86ISD::VFPCLASS",
 358                        SDTypeProfile<1, 2, [SDTCisVec<0>, SDTCVecEltisVT<0, i1>,
 359                                             SDTCisVec<1>, SDTCisInt<2>]>, []>;
 360
 361 def X86SubVBroadcast : SDNode<"X86ISD::SUBV_BROADCAST",
 362                     SDTypeProfile<1, 1, [SDTCisVec<0>, SDTCisVec<1>,
 363                                          SDTCisSubVecOfVec<1, 0>]>, []>;
 364 def X86VBroadcast : SDNode<"X86ISD::VBROADCAST", SDTVBroadcast>;
 365 def X86Vinsert   : SDNode<"X86ISD::VINSERT",  SDTypeProfile<1, 3,
 366                               [SDTCisSameAs<0, 1>, SDTCisPtrTy<3>]>, []>;
 367 def X86Vextract   : SDNode<"X86ISD::VEXTRACT",  SDTypeProfile<1, 2,
 368                               [SDTCisVec<1>, SDTCisPtrTy<2>]>, []>;
 369
 370 def X86Blendi    : SDNode<"X86ISD::BLENDI",   SDTBlend>;
 371
 372 def X86Addsub    : SDNode<"X86ISD::ADDSUB", SDTFPBinOp>;
 373
 374 def X86faddRnd   : SDNode<"X86ISD::FADD_RND",  SDTFPBinOpRound>;
 375 def X86fsubRnd   : SDNode<"X86ISD::FSUB_RND",  SDTFPBinOpRound>;
 376 def X86fmulRnd   : SDNode<"X86ISD::FMUL_RND",  SDTFPBinOpRound>;
 377 def X86fdivRnd   : SDNode<"X86ISD::FDIV_RND",  SDTFPBinOpRound>;
 378 def X86fmaxRnd   : SDNode<"X86ISD::FMAX_RND",       SDTFPBinOpRound>;
 379 def X86scalef    : SDNode<"X86ISD::SCALEF",         SDTFPBinOpRound>;
 380 def X86fminRnd   : SDNode<"X86ISD::FMIN_RND",       SDTFPBinOpRound>;
 381 def X86fsqrtRnd     : SDNode<"X86ISD::FSQRT_RND",   SDTFPUnaryOpRound>;
 382 def X86fsqrtRnds    : SDNode<"X86ISD::FSQRT_RND",   STDFp2SrcRm>;
 383 def X86fgetexpRnd   : SDNode<"X86ISD::FGETEXP_RND", SDTFPUnaryOpRound>;
 384 def X86fgetexpRnds  : SDNode<"X86ISD::FGETEXP_RND", STDFp2SrcRm>;
 385
 386 def X86Fmadd     : SDNode<"X86ISD::FMADD",     SDTFma>;
 387 def X86Fnmadd    : SDNode<"X86ISD::FNMADD",    SDTFma>;
 388 def X86Fmsub     : SDNode<"X86ISD::FMSUB",     SDTFma>;
 389 def X86Fnmsub    : SDNode<"X86ISD::FNMSUB",    SDTFma>;
 390 def X86Fmaddsub  : SDNode<"X86ISD::FMADDSUB",  SDTFma>;
 391 def X86Fmsubadd  : SDNode<"X86ISD::FMSUBADD",  SDTFma>;
 392
 393 def X86FmaddRnd     : SDNode<"X86ISD::FMADD_RND",     SDTFmaRound>;
 394 def X86FnmaddRnd    : SDNode<"X86ISD::FNMADD_RND",    SDTFmaRound>;
 395 def X86FmsubRnd     : SDNode<"X86ISD::FMSUB_RND",     SDTFmaRound>;
 396 def X86FnmsubRnd    : SDNode<"X86ISD::FNMSUB_RND",    SDTFmaRound>;
 397 def X86FmaddsubRnd  : SDNode<"X86ISD::FMADDSUB_RND",  SDTFmaRound>;
 398 def X86FmsubaddRnd  : SDNode<"X86ISD::FMSUBADD_RND",  SDTFmaRound>;
 399
 400 def X86rsqrt28   : SDNode<"X86ISD::RSQRT28",  STDFp1SrcRm>;
 401 def X86rcp28     : SDNode<"X86ISD::RCP28",    STDFp1SrcRm>;
 402 def X86exp2      : SDNode<"X86ISD::EXP2",     STDFp1SrcRm>;
 403
 404 def X86rsqrt28s  : SDNode<"X86ISD::RSQRT28",   STDFp2SrcRm>;
 405 def X86rcp28s    : SDNode<"X86ISD::RCP28",     STDFp2SrcRm>;
 406 def X86RndScales : SDNode<"X86ISD::VRNDSCALE", STDFp3SrcRm>;
 407 def X86Reduces   : SDNode<"X86ISD::VREDUCE",   STDFp3SrcRm>;
 408 def X86GetMants  : SDNode<"X86ISD::VGETMANT",  STDFp3SrcRm>;
 409
 410 def SDT_PCMPISTRI : SDTypeProfile<2, 3, [SDTCisVT<0, i32>, SDTCisVT<1, i32>,
 411                                          SDTCisVT<2, v16i8>, SDTCisVT<3, v16i8>,
 412                                          SDTCisVT<4, i8>]>;
 413 def SDT_PCMPESTRI : SDTypeProfile<2, 5, [SDTCisVT<0, i32>, SDTCisVT<1, i32>,
 414                                          SDTCisVT<2, v16i8>, SDTCisVT<3, i32>,
 415                                          SDTCisVT<4, v16i8>, SDTCisVT<5, i32>,
 416                                          SDTCisVT<6, i8>]>;
 417
 418 def X86pcmpistri : SDNode<"X86ISD::PCMPISTRI", SDT_PCMPISTRI>;
 419 def X86pcmpestri : SDNode<"X86ISD::PCMPESTRI", SDT_PCMPESTRI>;
 420
 421 def X86compress: SDNode<"X86ISD::COMPRESS", SDTypeProfile<1, 1,
 422                               [SDTCisSameAs<0, 1>, SDTCisVec<1>]>, []>;
 423 def X86expand  : SDNode<"X86ISD::EXPAND", SDTypeProfile<1, 1,
 424                               [SDTCisSameAs<0, 1>, SDTCisVec<1>]>, []>;
 425
 426 def SDTintToFPRound: SDTypeProfile<1, 3, [SDTCisVec<0>, SDTCisFP<0>,
 427                                SDTCisSameAs<0,1>, SDTCisInt<2>, SDTCisInt<3>]>;
 428
 429 def SDTDoubleToInt: SDTypeProfile<1, 1, [SDTCisVec<0>, SDTCisVec<1>,
 430                                          SDTCisInt<0>, SDTCVecEltisVT<1, f64>]>;
 431 def SDTFloatToInt: SDTypeProfile<1, 1, [SDTCisVec<0>, SDTCisVec<1>,
 432                                          SDTCisInt<0>, SDTCVecEltisVT<1, f32>]>;
 433
 434 def SDTDoubleToIntRnd: SDTypeProfile<1, 2, [SDTCisVec<0>, SDTCisVec<1>,
 435                                          SDTCisInt<0>, SDTCVecEltisVT<1, f64>]>;
 436 def SDTSDoubleToIntRnd: SDTypeProfile<1, 2, [SDTCisInt<0>,SDTCisFP<1>,
 437                                              SDTCVecEltisVT<1, f64>, SDTCisInt<2>]>;
 438 def SDTFloatToIntRnd: SDTypeProfile<1, 2, [SDTCisVec<0>, SDTCisVec<1>,
 439                                          SDTCisInt<0>, SDTCVecEltisVT<1, f32>]>;
 440 def SDTSFloatToIntRnd: SDTypeProfile<1, 2, [SDTCisInt<0>, SDTCisFP<1>,
 441                                             SDTCVecEltisVT<1, f32>, SDTCisInt<2>]>;
 442 def SDTVintToFPRound: SDTypeProfile<1, 2, [SDTCisVec<0>, SDTCisVec<1>,
 443                                            SDTCisFP<0>, SDTCVecEltisVT<1, i32>,
 444                                            SDTCisInt<2>]>;
 445 def SDTVlongToFPRound: SDTypeProfile<1, 2, [SDTCisVec<0>, SDTCisVec<1>,
 446                                            SDTCisFP<0>, SDTCVecEltisVT<1, i64>,
 447                                            SDTCisInt<2>]>;
 448
 449 def SDTVFPToIntRound: SDTypeProfile<1, 2, [SDTCisVec<0>, SDTCisVec<1>,
 450                                            SDTCisFP<1>, SDTCVecEltisVT<0, i32>,
 451                                            SDTCisInt<2>]>;
 452 def SDTVFPToLongRound: SDTypeProfile<1, 2, [SDTCisVec<0>, SDTCisVec<1>,
 453                                            SDTCisFP<1>, SDTCVecEltisVT<0, i64>,
 454                                            SDTCisInt<2>]>;
 455
 456 // Scalar
 457 def X86SintToFpRnd  : SDNode<"X86ISD::SINT_TO_FP_RND",  SDTintToFPRound>;
 458 def X86UintToFpRnd  : SDNode<"X86ISD::UINT_TO_FP_RND",  SDTintToFPRound>;
 459
 460 def X86cvttss2IntRnd      : SDNode<"X86ISD::FP_TO_SINT_RND",  SDTSFloatToIntRnd>;
 461 def X86cvttss2UIntRnd     : SDNode<"X86ISD::FP_TO_UINT_RND",  SDTSFloatToIntRnd>;
 462 def X86cvttsd2IntRnd      : SDNode<"X86ISD::FP_TO_SINT_RND",  SDTSDoubleToIntRnd>;
 463 def X86cvttsd2UIntRnd     : SDNode<"X86ISD::FP_TO_UINT_RND",  SDTSDoubleToIntRnd>;
 464 // Vector with rounding mode
 465
 466 // cvtt fp-to-int staff
 467 def X86VFpToSintRnd   : SDNode<"ISD::FP_TO_SINT",  SDTVFPToIntRound>;
 468 def X86VFpToUintRnd   : SDNode<"ISD::FP_TO_UINT",  SDTVFPToIntRound>;
 469 def X86VFpToSlongRnd  : SDNode<"ISD::FP_TO_SINT",  SDTVFPToLongRound>;
 470 def X86VFpToUlongRnd  : SDNode<"ISD::FP_TO_UINT",  SDTVFPToLongRound>;
 471
 472 def X86VSintToFpRnd   : SDNode<"ISD::SINT_TO_FP",  SDTVintToFPRound>;
 473 def X86VUintToFpRnd   : SDNode<"ISD::UINT_TO_FP",  SDTVintToFPRound>;
 474 def X86VSlongToFpRnd  : SDNode<"ISD::SINT_TO_FP",  SDTVlongToFPRound>;
 475 def X86VUlongToFpRnd  : SDNode<"ISD::UINT_TO_FP",  SDTVlongToFPRound>;
 476
 477 // cvt fp-to-int staff
 478 def X86cvtps2IntRnd      : SDNode<"X86ISD::FP_TO_SINT_RND",  SDTFloatToIntRnd>;
 479 def X86cvtps2UIntRnd     : SDNode<"X86ISD::FP_TO_UINT_RND",  SDTFloatToIntRnd>;
 480 def X86cvtpd2IntRnd      : SDNode<"X86ISD::FP_TO_SINT_RND",  SDTDoubleToIntRnd>;
 481 def X86cvtpd2UIntRnd     : SDNode<"X86ISD::FP_TO_UINT_RND",  SDTDoubleToIntRnd>;
 482
 483 // Vector without rounding mode
 484 def X86cvtps2Int      : SDNode<"X86ISD::FP_TO_SINT_RND",  SDTFloatToInt>;
 485 def X86cvtps2UInt     : SDNode<"X86ISD::FP_TO_UINT_RND",  SDTFloatToInt>;
 486 def X86cvtpd2Int      : SDNode<"X86ISD::FP_TO_SINT_RND",  SDTDoubleToInt>;
 487 def X86cvtpd2UInt     : SDNode<"X86ISD::FP_TO_UINT_RND",  SDTDoubleToInt>;
 488
 489 def X86vfpextRnd  : SDNode<"X86ISD::VFPEXT",
 490                         SDTypeProfile<1, 2, [SDTCisVec<0>, SDTCisVec<1>,
 491                                              SDTCisFP<0>, SDTCisFP<1>,
 492                                              SDTCisOpSmallerThanOp<1, 0>,
 493                                              SDTCisInt<2>]>>;
 494 def X86vfproundRnd: SDNode<"X86ISD::VFPROUND",
 495                         SDTypeProfile<1, 2, [SDTCisVec<0>, SDTCisVec<1>,
 496                                              SDTCisFP<0>, SDTCisFP<1>,
 497                                              SDTCVecEltisVT<0, f32>,
 498                                              SDTCVecEltisVT<1, f64>,
 499                                              SDTCisInt<2>]>>;
 500
 501 //===----------------------------------------------------------------------===//
 502 // SSE Complex Patterns
 503 //===----------------------------------------------------------------------===//
 504
 505 // These are 'extloads' from a scalar to the low element of a vector, zeroing
 506 // the top elements.  These are used for the SSE 'ss' and 'sd' instruction
 507 // forms.
 508 def sse_load_f32 : ComplexPattern<v4f32, 5, "SelectScalarSSELoad", [],
 509                                   [SDNPHasChain, SDNPMayLoad, SDNPMemOperand,
 510                                    SDNPWantRoot]>;
 511 def sse_load_f64 : ComplexPattern<v2f64, 5, "SelectScalarSSELoad", [],
 512                                   [SDNPHasChain, SDNPMayLoad, SDNPMemOperand,
 513                                    SDNPWantRoot]>;
 514
 515 def ssmem : Operand<v4f32> {
 516   let PrintMethod = "printf32mem";
 517   let MIOperandInfo = (ops ptr_rc, i8imm, ptr_rc_nosp, i32imm, i8imm);
 518   let ParserMatchClass = X86Mem32AsmOperand;
 519   let OperandType = "OPERAND_MEMORY";
 520 }
 521 def sdmem : Operand<v2f64> {
 522   let PrintMethod = "printf64mem";
 523   let MIOperandInfo = (ops ptr_rc, i8imm, ptr_rc_nosp, i32imm, i8imm);
 524   let ParserMatchClass = X86Mem64AsmOperand;
 525   let OperandType = "OPERAND_MEMORY";
 526 }
 527
 528 //===----------------------------------------------------------------------===//
 529 // SSE pattern fragments
 530 //===----------------------------------------------------------------------===//
 531
 532 // 128-bit load pattern fragments
 533 // NOTE: all 128-bit integer vector loads are promoted to v2i64
 534 def loadv4f32    : PatFrag<(ops node:$ptr), (v4f32 (load node:$ptr))>;
 535 def loadv2f64    : PatFrag<(ops node:$ptr), (v2f64 (load node:$ptr))>;
 536 def loadv2i64    : PatFrag<(ops node:$ptr), (v2i64 (load node:$ptr))>;
 537
 538 // 256-bit load pattern fragments
 539 // NOTE: all 256-bit integer vector loads are promoted to v4i64
 540 def loadv8f32    : PatFrag<(ops node:$ptr), (v8f32 (load node:$ptr))>;
 541 def loadv4f64    : PatFrag<(ops node:$ptr), (v4f64 (load node:$ptr))>;
 542 def loadv4i64    : PatFrag<(ops node:$ptr), (v4i64 (load node:$ptr))>;
 543
 544 // 512-bit load pattern fragments
 545 def loadv16f32   : PatFrag<(ops node:$ptr), (v16f32 (load node:$ptr))>;
 546 def loadv8f64    : PatFrag<(ops node:$ptr), (v8f64 (load node:$ptr))>;
 547 def loadv64i8    : PatFrag<(ops node:$ptr), (v64i8 (load node:$ptr))>;
 548 def loadv32i16   : PatFrag<(ops node:$ptr), (v32i16 (load node:$ptr))>;
 549 def loadv16i32   : PatFrag<(ops node:$ptr), (v16i32 (load node:$ptr))>;
 550 def loadv8i64    : PatFrag<(ops node:$ptr), (v8i64 (load node:$ptr))>;
 551
 552 // 128-/256-/512-bit extload pattern fragments
 553 def extloadv2f32 : PatFrag<(ops node:$ptr), (v2f64 (extloadvf32 node:$ptr))>;
 554 def extloadv4f32 : PatFrag<(ops node:$ptr), (v4f64 (extloadvf32 node:$ptr))>;
 555 def extloadv8f32 : PatFrag<(ops node:$ptr), (v8f64 (extloadvf32 node:$ptr))>;
 556
 557 // These are needed to match a scalar load that is used in a vector-only
 558 // math instruction such as the FP logical ops: andps, andnps, orps, xorps.
 559 // The memory operand is required to be a 128-bit load, so it must be converted
 560 // from a vector to a scalar.
 561 def loadf32_128 : PatFrag<(ops node:$ptr),
 562   (f32 (vector_extract (loadv4f32 node:$ptr), (iPTR 0)))>;
 563 def loadf64_128 : PatFrag<(ops node:$ptr),
 564   (f64 (vector_extract (loadv2f64 node:$ptr), (iPTR 0)))>;
 565
 566 // Like 'store', but always requires 128-bit vector alignment.
 567 def alignedstore : PatFrag<(ops node:$val, node:$ptr),
 568                            (store node:$val, node:$ptr), [{
 569   return cast<StoreSDNode>(N)->getAlignment() >= 16;
 570 }]>;
 571
 572 // Like 'store', but always requires 256-bit vector alignment.
 573 def alignedstore256 : PatFrag<(ops node:$val, node:$ptr),
 574                               (store node:$val, node:$ptr), [{
 575   return cast<StoreSDNode>(N)->getAlignment() >= 32;
 576 }]>;
 577
 578 // Like 'store', but always requires 512-bit vector alignment.
 579 def alignedstore512 : PatFrag<(ops node:$val, node:$ptr),
 580                               (store node:$val, node:$ptr), [{
 581   return cast<StoreSDNode>(N)->getAlignment() >= 64;
 582 }]>;
 583
 584 // Like 'load', but always requires 128-bit vector alignment.
 585 def alignedload : PatFrag<(ops node:$ptr), (load node:$ptr), [{
 586   return cast<LoadSDNode>(N)->getAlignment() >= 16;
 587 }]>;
 588
 589 // Like 'X86vzload', but always requires 128-bit vector alignment.
 590 def alignedX86vzload : PatFrag<(ops node:$ptr), (X86vzload node:$ptr), [{
 591   return cast<MemSDNode>(N)->getAlignment() >= 16;
 592 }]>;
 593
 594 // Like 'load', but always requires 256-bit vector alignment.
 595 def alignedload256 : PatFrag<(ops node:$ptr), (load node:$ptr), [{
 596   return cast<LoadSDNode>(N)->getAlignment() >= 32;
 597 }]>;
 598
 599 // Like 'load', but always requires 512-bit vector alignment.
 600 def alignedload512 : PatFrag<(ops node:$ptr), (load node:$ptr), [{
 601   return cast<LoadSDNode>(N)->getAlignment() >= 64;
 602 }]>;
 603
 604 def alignedloadfsf32 : PatFrag<(ops node:$ptr),
 605                                (f32 (alignedload node:$ptr))>;
 606 def alignedloadfsf64 : PatFrag<(ops node:$ptr),
 607                                (f64 (alignedload node:$ptr))>;
 608
 609 // 128-bit aligned load pattern fragments
 610 // NOTE: all 128-bit integer vector loads are promoted to v2i64
 611 def alignedloadv4f32 : PatFrag<(ops node:$ptr),
 612                                (v4f32 (alignedload node:$ptr))>;
 613 def alignedloadv2f64 : PatFrag<(ops node:$ptr),
 614                                (v2f64 (alignedload node:$ptr))>;
 615 def alignedloadv2i64 : PatFrag<(ops node:$ptr),
 616                                (v2i64 (alignedload node:$ptr))>;
 617
 618 // 256-bit aligned load pattern fragments
 619 // NOTE: all 256-bit integer vector loads are promoted to v4i64
 620 def alignedloadv8f32 : PatFrag<(ops node:$ptr),
 621                                (v8f32 (alignedload256 node:$ptr))>;
 622 def alignedloadv4f64 : PatFrag<(ops node:$ptr),
 623                                (v4f64 (alignedload256 node:$ptr))>;
 624 def alignedloadv4i64 : PatFrag<(ops node:$ptr),
 625                                (v4i64 (alignedload256 node:$ptr))>;
 626
 627 // 512-bit aligned load pattern fragments
 628 def alignedloadv16f32 : PatFrag<(ops node:$ptr),
 629                                 (v16f32 (alignedload512 node:$ptr))>;
 630 def alignedloadv16i32 : PatFrag<(ops node:$ptr),
 631                                 (v16i32 (alignedload512 node:$ptr))>;
 632 def alignedloadv8f64  : PatFrag<(ops node:$ptr),
 633                                 (v8f64  (alignedload512 node:$ptr))>;
 634 def alignedloadv8i64  : PatFrag<(ops node:$ptr),
 635                                 (v8i64  (alignedload512 node:$ptr))>;
 636
 637 // Like 'load', but uses special alignment checks suitable for use in
 638 // memory operands in most SSE instructions, which are required to
 639 // be naturally aligned on some targets but not on others.  If the subtarget
 640 // allows unaligned accesses, match any load, though this may require
 641 // setting a feature bit in the processor (on startup, for example).
 642 // Opteron 10h and later implement such a feature.
 643 def memop : PatFrag<(ops node:$ptr), (load node:$ptr), [{
 644   return    Subtarget->hasSSEUnalignedMem()
 645          || cast<LoadSDNode>(N)->getAlignment() >= 16;
 646 }]>;
 647
 648 def memopfsf32 : PatFrag<(ops node:$ptr), (f32   (memop node:$ptr))>;
 649 def memopfsf64 : PatFrag<(ops node:$ptr), (f64   (memop node:$ptr))>;
 650
 651 // 128-bit memop pattern fragments
 652 // NOTE: all 128-bit integer vector loads are promoted to v2i64
 653 def memopv4f32 : PatFrag<(ops node:$ptr), (v4f32 (memop node:$ptr))>;
 654 def memopv2f64 : PatFrag<(ops node:$ptr), (v2f64 (memop node:$ptr))>;
 655 def memopv2i64 : PatFrag<(ops node:$ptr), (v2i64 (memop node:$ptr))>;
 656
 657 // These are needed to match a scalar memop that is used in a vector-only
 658 // math instruction such as the FP logical ops: andps, andnps, orps, xorps.
 659 // The memory operand is required to be a 128-bit load, so it must be converted
 660 // from a vector to a scalar.
 661 def memopfsf32_128 : PatFrag<(ops node:$ptr),
 662   (f32 (vector_extract (memopv4f32 node:$ptr), (iPTR 0)))>;
 663 def memopfsf64_128 : PatFrag<(ops node:$ptr),
 664   (f64 (vector_extract (memopv2f64 node:$ptr), (iPTR 0)))>;
 665
 666
 667 // SSSE3 uses MMX registers for some instructions. They aren't aligned on a
 668 // 16-byte boundary.
 669 // FIXME: 8 byte alignment for mmx reads is not required
 670 def memop64 : PatFrag<(ops node:$ptr), (load node:$ptr), [{
 671   return cast<LoadSDNode>(N)->getAlignment() >= 8;
 672 }]>;
 673
 674 def memopmmx  : PatFrag<(ops node:$ptr), (x86mmx  (memop64 node:$ptr))>;
 675
 676 def mgatherv4i32 : PatFrag<(ops node:$src1, node:$src2, node:$src3),
 677   (masked_gather node:$src1, node:$src2, node:$src3) , [{
 678   if (MaskedGatherSDNode *Mgt = dyn_cast<MaskedGatherSDNode>(N))
 679     return (Mgt->getIndex().getValueType() == MVT::v4i32 ||
 680             Mgt->getBasePtr().getValueType() == MVT::v4i32);
 681   return false;
 682 }]>;
 683
 684 def mgatherv8i32 : PatFrag<(ops node:$src1, node:$src2, node:$src3),
 685   (masked_gather node:$src1, node:$src2, node:$src3) , [{
 686   if (MaskedGatherSDNode *Mgt = dyn_cast<MaskedGatherSDNode>(N))
 687     return (Mgt->getIndex().getValueType() == MVT::v8i32 ||
 688             Mgt->getBasePtr().getValueType() == MVT::v8i32);
 689   return false;
 690 }]>;
 691
 692 def mgatherv2i64 : PatFrag<(ops node:$src1, node:$src2, node:$src3),
 693   (masked_gather node:$src1, node:$src2, node:$src3) , [{
 694   if (MaskedGatherSDNode *Mgt = dyn_cast<MaskedGatherSDNode>(N))
 695     return (Mgt->getIndex().getValueType() == MVT::v2i64 ||
 696             Mgt->getBasePtr().getValueType() == MVT::v2i64);
 697   return false;
 698 }]>;
 699 def mgatherv4i64 : PatFrag<(ops node:$src1, node:$src2, node:$src3),
 700   (masked_gather node:$src1, node:$src2, node:$src3) , [{
 701   if (MaskedGatherSDNode *Mgt = dyn_cast<MaskedGatherSDNode>(N))
 702     return (Mgt->getIndex().getValueType() == MVT::v4i64 ||
 703             Mgt->getBasePtr().getValueType() == MVT::v4i64);
 704   return false;
 705 }]>;
 706 def mgatherv8i64 : PatFrag<(ops node:$src1, node:$src2, node:$src3),
 707   (masked_gather node:$src1, node:$src2, node:$src3) , [{
 708   if (MaskedGatherSDNode *Mgt = dyn_cast<MaskedGatherSDNode>(N))
 709     return (Mgt->getIndex().getValueType() == MVT::v8i64 ||
 710             Mgt->getBasePtr().getValueType() == MVT::v8i64);
 711   return false;
 712 }]>;
 713 def mgatherv16i32 : PatFrag<(ops node:$src1, node:$src2, node:$src3),
 714   (masked_gather node:$src1, node:$src2, node:$src3) , [{
 715   if (MaskedGatherSDNode *Mgt = dyn_cast<MaskedGatherSDNode>(N))
 716     return (Mgt->getIndex().getValueType() == MVT::v16i32 ||
 717             Mgt->getBasePtr().getValueType() == MVT::v16i32);
 718   return false;
 719 }]>;
 720
 721 def mscatterv2i64 : PatFrag<(ops node:$src1, node:$src2, node:$src3),
 722   (masked_scatter node:$src1, node:$src2, node:$src3) , [{
 723   if (MaskedScatterSDNode *Sc = dyn_cast<MaskedScatterSDNode>(N))
 724     return (Sc->getIndex().getValueType() == MVT::v2i64 ||
 725             Sc->getBasePtr().getValueType() == MVT::v2i64);
 726   return false;
 727 }]>;
 728
 729 def mscatterv4i32 : PatFrag<(ops node:$src1, node:$src2, node:$src3),
 730   (masked_scatter node:$src1, node:$src2, node:$src3) , [{
 731   if (MaskedScatterSDNode *Sc = dyn_cast<MaskedScatterSDNode>(N))
 732     return (Sc->getIndex().getValueType() == MVT::v4i32 ||
 733             Sc->getBasePtr().getValueType() == MVT::v4i32);
 734   return false;
 735 }]>;
 736
 737 def mscatterv4i64 : PatFrag<(ops node:$src1, node:$src2, node:$src3),
 738   (masked_scatter node:$src1, node:$src2, node:$src3) , [{
 739   if (MaskedScatterSDNode *Sc = dyn_cast<MaskedScatterSDNode>(N))
 740     return (Sc->getIndex().getValueType() == MVT::v4i64 ||
 741             Sc->getBasePtr().getValueType() == MVT::v4i64);
 742   return false;
 743 }]>;
 744
 745 def mscatterv8i32 : PatFrag<(ops node:$src1, node:$src2, node:$src3),
 746   (masked_scatter node:$src1, node:$src2, node:$src3) , [{
 747   if (MaskedScatterSDNode *Sc = dyn_cast<MaskedScatterSDNode>(N))
 748     return (Sc->getIndex().getValueType() == MVT::v8i32 ||
 749             Sc->getBasePtr().getValueType() == MVT::v8i32);
 750   return false;
 751 }]>;
 752
 753 def mscatterv8i64 : PatFrag<(ops node:$src1, node:$src2, node:$src3),
 754   (masked_scatter node:$src1, node:$src2, node:$src3) , [{
 755   if (MaskedScatterSDNode *Sc = dyn_cast<MaskedScatterSDNode>(N))
 756     return (Sc->getIndex().getValueType() == MVT::v8i64 ||
 757             Sc->getBasePtr().getValueType() == MVT::v8i64);
 758   return false;
 759 }]>;
 760 def mscatterv16i32 : PatFrag<(ops node:$src1, node:$src2, node:$src3),
 761   (masked_scatter node:$src1, node:$src2, node:$src3) , [{
 762   if (MaskedScatterSDNode *Sc = dyn_cast<MaskedScatterSDNode>(N))
 763     return (Sc->getIndex().getValueType() == MVT::v16i32 ||
 764             Sc->getBasePtr().getValueType() == MVT::v16i32);
 765   return false;
 766 }]>;
 767
 768 // 128-bit bitconvert pattern fragments
 769 def bc_v4f32 : PatFrag<(ops node:$in), (v4f32 (bitconvert node:$in))>;
 770 def bc_v2f64 : PatFrag<(ops node:$in), (v2f64 (bitconvert node:$in))>;
 771 def bc_v16i8 : PatFrag<(ops node:$in), (v16i8 (bitconvert node:$in))>;
 772 def bc_v8i16 : PatFrag<(ops node:$in), (v8i16 (bitconvert node:$in))>;
 773 def bc_v4i32 : PatFrag<(ops node:$in), (v4i32 (bitconvert node:$in))>;
 774 def bc_v2i64 : PatFrag<(ops node:$in), (v2i64 (bitconvert node:$in))>;
 775
 776 // 256-bit bitconvert pattern fragments
 777 def bc_v32i8 : PatFrag<(ops node:$in), (v32i8 (bitconvert node:$in))>;
 778 def bc_v16i16 : PatFrag<(ops node:$in), (v16i16 (bitconvert node:$in))>;
 779 def bc_v8i32 : PatFrag<(ops node:$in), (v8i32 (bitconvert node:$in))>;
 780 def bc_v4i64 : PatFrag<(ops node:$in), (v4i64 (bitconvert node:$in))>;
 781 def bc_v8f32 : PatFrag<(ops node:$in), (v8f32 (bitconvert node:$in))>;
 782
 783 // 512-bit bitconvert pattern fragments
 784 def bc_v16i32 : PatFrag<(ops node:$in), (v16i32 (bitconvert node:$in))>;
 785 def bc_v8i64 : PatFrag<(ops node:$in), (v8i64 (bitconvert node:$in))>;
 786 def bc_v8f64 : PatFrag<(ops node:$in), (v8f64 (bitconvert node:$in))>;
 787 def bc_v16f32 : PatFrag<(ops node:$in), (v16f32 (bitconvert node:$in))>;
 788
 789 def vzmovl_v2i64 : PatFrag<(ops node:$src),
 790                            (bitconvert (v2i64 (X86vzmovl
 791                              (v2i64 (scalar_to_vector (loadi64 node:$src))))))>;
 792 def vzmovl_v4i32 : PatFrag<(ops node:$src),
 793                            (bitconvert (v4i32 (X86vzmovl
 794                              (v4i32 (scalar_to_vector (loadi32 node:$src))))))>;
 795
 796 def vzload_v2i64 : PatFrag<(ops node:$src),
 797                            (bitconvert (v2i64 (X86vzload node:$src)))>;
 798
 799
 800 def fp32imm0 : PatLeaf<(f32 fpimm), [{
 801   return N->isExactlyValue(+0.0);
 802 }]>;
 803
 804 def I8Imm : SDNodeXForm<imm, [{
 805   // Transformation function: get the low 8 bits.
 806   return getI8Imm((uint8_t)N->getZExtValue(), SDLoc(N));
 807 }]>;
 808
 809 def FROUND_NO_EXC : ImmLeaf<i32, [{ return Imm == 8; }]>;
 810 def FROUND_CURRENT : ImmLeaf<i32, [{
 811   return Imm == X86::STATIC_ROUNDING::CUR_DIRECTION;
 812 }]>;
 813
 814 // BYTE_imm - Transform bit immediates into byte immediates.
 815 def BYTE_imm  : SDNodeXForm<imm, [{
 816   // Transformation function: imm >> 3
 817   return getI32Imm(N->getZExtValue() >> 3, SDLoc(N));
 818 }]>;
 819
 820 // EXTRACT_get_vextract128_imm xform function: convert extract_subvector index
 821 // to VEXTRACTF128/VEXTRACTI128 imm.
 822 def EXTRACT_get_vextract128_imm : SDNodeXForm<extract_subvector, [{
 823   return getI8Imm(X86::getExtractVEXTRACT128Immediate(N), SDLoc(N));
 824 }]>;
 825
 826 // INSERT_get_vinsert128_imm xform function: convert insert_subvector index to
 827 // VINSERTF128/VINSERTI128 imm.
 828 def INSERT_get_vinsert128_imm : SDNodeXForm<insert_subvector, [{
 829   return getI8Imm(X86::getInsertVINSERT128Immediate(N), SDLoc(N));
 830 }]>;
 831
 832 // EXTRACT_get_vextract256_imm xform function: convert extract_subvector index
 833 // to VEXTRACTF64x4 imm.
 834 def EXTRACT_get_vextract256_imm : SDNodeXForm<extract_subvector, [{
 835   return getI8Imm(X86::getExtractVEXTRACT256Immediate(N), SDLoc(N));
 836 }]>;
 837
 838 // INSERT_get_vinsert256_imm xform function: convert insert_subvector index to
 839 // VINSERTF64x4 imm.
 840 def INSERT_get_vinsert256_imm : SDNodeXForm<insert_subvector, [{
 841   return getI8Imm(X86::getInsertVINSERT256Immediate(N), SDLoc(N));
 842 }]>;
 843
 844 def vextract128_extract : PatFrag<(ops node:$bigvec, node:$index),
 845                                    (extract_subvector node:$bigvec,
 846                                                       node:$index), [{
 847   return X86::isVEXTRACT128Index(N);
 848 }], EXTRACT_get_vextract128_imm>;
 849
 850 def vinsert128_insert : PatFrag<(ops node:$bigvec, node:$smallvec,
 851                                       node:$index),
 852                                  (insert_subvector node:$bigvec, node:$smallvec,
 853                                                    node:$index), [{
 854   return X86::isVINSERT128Index(N);
 855 }], INSERT_get_vinsert128_imm>;
 856
 857
 858 def vextract256_extract : PatFrag<(ops node:$bigvec, node:$index),
 859                                    (extract_subvector node:$bigvec,
 860                                                       node:$index), [{
 861   return X86::isVEXTRACT256Index(N);
 862 }], EXTRACT_get_vextract256_imm>;
 863
 864 def vinsert256_insert : PatFrag<(ops node:$bigvec, node:$smallvec,
 865                                       node:$index),
 866                                  (insert_subvector node:$bigvec, node:$smallvec,
 867                                                    node:$index), [{
 868   return X86::isVINSERT256Index(N);
 869 }], INSERT_get_vinsert256_imm>;
 870
 871 def masked_load_aligned128 : PatFrag<(ops node:$src1, node:$src2, node:$src3),
 872                          (masked_load node:$src1, node:$src2, node:$src3), [{
 873   if (auto *Load = dyn_cast<MaskedLoadSDNode>(N))
 874     return Load->getAlignment() >= 16;
 875   return false;
 876 }]>;
 877
 878 def masked_load_aligned256 : PatFrag<(ops node:$src1, node:$src2, node:$src3),
 879                          (masked_load node:$src1, node:$src2, node:$src3), [{
 880   if (auto *Load = dyn_cast<MaskedLoadSDNode>(N))
 881     return Load->getAlignment() >= 32;
 882   return false;
 883 }]>;
 884
 885 def masked_load_aligned512 : PatFrag<(ops node:$src1, node:$src2, node:$src3),
 886                          (masked_load node:$src1, node:$src2, node:$src3), [{
 887   if (auto *Load = dyn_cast<MaskedLoadSDNode>(N))
 888     return Load->getAlignment() >= 64;
 889   return false;
 890 }]>;
 891
 892 def masked_load_unaligned : PatFrag<(ops node:$src1, node:$src2, node:$src3),
 893                          (masked_load node:$src1, node:$src2, node:$src3), [{
 894   return isa<MaskedLoadSDNode>(N);
 895 }]>;
 896
 897 // masked store fragments.
 898 // X86mstore can't be implemented in core DAG files because some targets
 899 // doesn't support vector type ( llvm-tblgen will fail)
 900 def X86mstore : PatFrag<(ops node:$src1, node:$src2, node:$src3),
 901                         (masked_store node:$src1, node:$src2, node:$src3), [{
 902   return !cast<MaskedStoreSDNode>(N)->isTruncatingStore();
 903 }]>;
 904
 905 def masked_store_aligned128 : PatFrag<(ops node:$src1, node:$src2, node:$src3),
 906                          (X86mstore node:$src1, node:$src2, node:$src3), [{
 907   if (auto *Store = dyn_cast<MaskedStoreSDNode>(N))
 908     return Store->getAlignment() >= 16;
 909   return false;
 910 }]>;
 911
 912 def masked_store_aligned256 : PatFrag<(ops node:$src1, node:$src2, node:$src3),
 913                          (X86mstore node:$src1, node:$src2, node:$src3), [{
 914   if (auto *Store = dyn_cast<MaskedStoreSDNode>(N))
 915     return Store->getAlignment() >= 32;
 916   return false;
 917 }]>;
 918
 919 def masked_store_aligned512 : PatFrag<(ops node:$src1, node:$src2, node:$src3),
 920                          (X86mstore node:$src1, node:$src2, node:$src3), [{
 921   if (auto *Store = dyn_cast<MaskedStoreSDNode>(N))
 922     return Store->getAlignment() >= 64;
 923   return false;
 924 }]>;
 925
 926 def masked_store_unaligned : PatFrag<(ops node:$src1, node:$src2, node:$src3),
 927                          (X86mstore node:$src1, node:$src2, node:$src3), [{
 928   return isa<MaskedStoreSDNode>(N);
 929 }]>;
 930
 931 // masked truncstore fragments
 932 // X86mtruncstore can't be implemented in core DAG files because some targets
 933 // doesn't support vector type ( llvm-tblgen will fail)
 934 def X86mtruncstore : PatFrag<(ops node:$src1, node:$src2, node:$src3),
 935                              (masked_store node:$src1, node:$src2, node:$src3), [{
 936     return cast<MaskedStoreSDNode>(N)->isTruncatingStore();
 937 }]>;
 938 def masked_truncstorevi8 :
 939   PatFrag<(ops node:$src1, node:$src2, node:$src3),
 940           (X86mtruncstore node:$src1, node:$src2, node:$src3), [{
 941   return cast<MaskedStoreSDNode>(N)->getMemoryVT().getScalarType() == MVT::i8;
 942 }]>;
 943 def masked_truncstorevi16 :
 944   PatFrag<(ops node:$src1, node:$src2, node:$src3),
 945           (X86mtruncstore node:$src1, node:$src2, node:$src3), [{
 946   return cast<MaskedStoreSDNode>(N)->getMemoryVT().getScalarType() == MVT::i16;
 947 }]>;
 948 def masked_truncstorevi32 :
 949   PatFrag<(ops node:$src1, node:$src2, node:$src3),
 950           (X86mtruncstore node:$src1, node:$src2, node:$src3), [{
 951   return cast<MaskedStoreSDNode>(N)->getMemoryVT().getScalarType() == MVT::i32;
 952 }]>;