utils/TableGen/NeonEmitter.cpp

   1 //===- NeonEmitter.cpp - Generate arm_neon.h for use with clang -*- C++ -*-===//
   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 tablegen backend is responsible for emitting arm_neon.h, which includes
  11 // a declaration and definition of each function specified by the ARM NEON
  12 // compiler interface.  See ARM document DUI0348B.
  13 //
  14 // Each NEON instruction is implemented in terms of 1 or more functions which
  15 // are suffixed with the element type of the input vectors.  Functions may be
  16 // implemented in terms of generic vector operations such as +, *, -, etc. or
  17 // by calling a __builtin_-prefixed function which will be handled by clang's
  18 // CodeGen library.
  19 //
  20 // Additional validation code can be generated by this file when runHeader() is
  21 // called, rather than the normal run() entry point.  A complete set of tests
  22 // for Neon intrinsics can be generated by calling the runTests() entry point.
  23 //
  24 //===----------------------------------------------------------------------===//
  25
  26 #include "NeonEmitter.h"
  27 #include "llvm/ADT/SmallString.h"
  28 #include "llvm/ADT/SmallVector.h"
  29 #include "llvm/ADT/StringExtras.h"
  30 #include <string>
  31
  32 using namespace llvm;
  33
  34 /// ParseTypes - break down a string such as "fQf" into a vector of StringRefs,
  35 /// which each StringRef representing a single type declared in the string.
  36 /// for "fQf" we would end up with 2 StringRefs, "f", and "Qf", representing
  37 /// 2xfloat and 4xfloat respectively.
  38 static void ParseTypes(Record *r, std::string &s,
  39                        SmallVectorImpl<StringRef> &TV) {
  40   const char *data = s.data();
  41   int len = 0;
  42
  43   for (unsigned i = 0, e = s.size(); i != e; ++i, ++len) {
  44     if (data[len] == 'P' || data[len] == 'Q' || data[len] == 'U')
  45       continue;
  46
  47     switch (data[len]) {
  48       case 'c':
  49       case 's':
  50       case 'i':
  51       case 'l':
  52       case 'h':
  53       case 'f':
  54         break;
  55       default:
  56         throw TGError(r->getLoc(),
  57                       "Unexpected letter: " + std::string(data + len, 1));
  58         break;
  59     }
  60     TV.push_back(StringRef(data, len + 1));
  61     data += len + 1;
  62     len = -1;
  63   }
  64 }
  65
  66 /// Widen - Convert a type code into the next wider type.  char -> short,
  67 /// short -> int, etc.
  68 static char Widen(const char t) {
  69   switch (t) {
  70     case 'c':
  71       return 's';
  72     case 's':
  73       return 'i';
  74     case 'i':
  75       return 'l';
  76     case 'h':
  77       return 'f';
  78     default: throw "unhandled type in widen!";
  79   }
  80   return '\0';
  81 }
  82
  83 /// Narrow - Convert a type code into the next smaller type.  short -> char,
  84 /// float -> half float, etc.
  85 static char Narrow(const char t) {
  86   switch (t) {
  87     case 's':
  88       return 'c';
  89     case 'i':
  90       return 's';
  91     case 'l':
  92       return 'i';
  93     case 'f':
  94       return 'h';
  95     default: throw "unhandled type in narrow!";
  96   }
  97   return '\0';
  98 }
  99
 100 /// For a particular StringRef, return the base type code, and whether it has
 101 /// the quad-vector, polynomial, or unsigned modifiers set.
 102 static char ClassifyType(StringRef ty, bool &quad, bool &poly, bool &usgn) {
 103   unsigned off = 0;
 104
 105   // remember quad.
 106   if (ty[off] == 'Q') {
 107     quad = true;
 108     ++off;
 109   }
 110
 111   // remember poly.
 112   if (ty[off] == 'P') {
 113     poly = true;
 114     ++off;
 115   }
 116
 117   // remember unsigned.
 118   if (ty[off] == 'U') {
 119     usgn = true;
 120     ++off;
 121   }
 122
 123   // base type to get the type string for.
 124   return ty[off];
 125 }
 126
 127 /// ModType - Transform a type code and its modifiers based on a mod code. The
 128 /// mod code definitions may be found at the top of arm_neon.td.
 129 static char ModType(const char mod, char type, bool &quad, bool &poly,
 130                     bool &usgn, bool &scal, bool &cnst, bool &pntr) {
 131   switch (mod) {
 132     case 't':
 133       if (poly) {
 134         poly = false;
 135         usgn = true;
 136       }
 137       break;
 138     case 'u':
 139       usgn = true;
 140       poly = false;
 141       if (type == 'f')
 142         type = 'i';
 143       break;
 144     case 'x':
 145       usgn = false;
 146       poly = false;
 147       if (type == 'f')
 148         type = 'i';
 149       break;
 150     case 'f':
 151       if (type == 'h')
 152         quad = true;
 153       type = 'f';
 154       usgn = false;
 155       break;
 156     case 'g':
 157       quad = false;
 158       break;
 159     case 'w':
 160       type = Widen(type);
 161       quad = true;
 162       break;
 163     case 'n':
 164       type = Widen(type);
 165       break;
 166     case 'i':
 167       type = 'i';
 168       scal = true;
 169       break;
 170     case 'l':
 171       type = 'l';
 172       scal = true;
 173       usgn = true;
 174       break;
 175     case 's':
 176     case 'a':
 177       scal = true;
 178       break;
 179     case 'k':
 180       quad = true;
 181       break;
 182     case 'c':
 183       cnst = true;
 184     case 'p':
 185       pntr = true;
 186       scal = true;
 187       break;
 188     case 'h':
 189       type = Narrow(type);
 190       if (type == 'h')
 191         quad = false;
 192       break;
 193     case 'e':
 194       type = Narrow(type);
 195       usgn = true;
 196       break;
 197     default:
 198       break;
 199   }
 200   return type;
 201 }
 202
 203 /// TypeString - for a modifier and type, generate the name of the typedef for
 204 /// that type.  QUc -> uint8x8_t.
 205 static std::string TypeString(const char mod, StringRef typestr) {
 206   bool quad = false;
 207   bool poly = false;
 208   bool usgn = false;
 209   bool scal = false;
 210   bool cnst = false;
 211   bool pntr = false;
 212
 213   if (mod == 'v')
 214     return "void";
 215   if (mod == 'i')
 216     return "int";
 217
 218   // base type to get the type string for.
 219   char type = ClassifyType(typestr, quad, poly, usgn);
 220
 221   // Based on the modifying character, change the type and width if necessary.
 222   type = ModType(mod, type, quad, poly, usgn, scal, cnst, pntr);
 223
 224   SmallString<128> s;
 225
 226   if (usgn)
 227     s.push_back('u');
 228
 229   switch (type) {
 230     case 'c':
 231       s += poly ? "poly8" : "int8";
 232       if (scal)
 233         break;
 234       s += quad ? "x16" : "x8";
 235       break;
 236     case 's':
 237       s += poly ? "poly16" : "int16";
 238       if (scal)
 239         break;
 240       s += quad ? "x8" : "x4";
 241       break;
 242     case 'i':
 243       s += "int32";
 244       if (scal)
 245         break;
 246       s += quad ? "x4" : "x2";
 247       break;
 248     case 'l':
 249       s += "int64";
 250       if (scal)
 251         break;
 252       s += quad ? "x2" : "x1";
 253       break;
 254     case 'h':
 255       s += "float16";
 256       if (scal)
 257         break;
 258       s += quad ? "x8" : "x4";
 259       break;
 260     case 'f':
 261       s += "float32";
 262       if (scal)
 263         break;
 264       s += quad ? "x4" : "x2";
 265       break;
 266     default:
 267       throw "unhandled type!";
 268       break;
 269   }
 270
 271   if (mod == '2')
 272     s += "x2";
 273   if (mod == '3')
 274     s += "x3";
 275   if (mod == '4')
 276     s += "x4";
 277
 278   // Append _t, finishing the type string typedef type.
 279   s += "_t";
 280
 281   if (cnst)
 282     s += " const";
 283
 284   if (pntr)
 285     s += " *";
 286
 287   return s.str();
 288 }
 289
 290 /// BuiltinTypeString - for a modifier and type, generate the clang
 291 /// BuiltinsARM.def prototype code for the function.  See the top of clang's
 292 /// Builtins.def for a description of the type strings.
 293 static std::string BuiltinTypeString(const char mod, StringRef typestr,
 294                                      ClassKind ck, bool ret) {
 295   bool quad = false;
 296   bool poly = false;
 297   bool usgn = false;
 298   bool scal = false;
 299   bool cnst = false;
 300   bool pntr = false;
 301
 302   if (mod == 'v')
 303     return "v"; // void
 304   if (mod == 'i')
 305     return "i"; // int
 306
 307   // base type to get the type string for.
 308   char type = ClassifyType(typestr, quad, poly, usgn);
 309
 310   // Based on the modifying character, change the type and width if necessary.
 311   type = ModType(mod, type, quad, poly, usgn, scal, cnst, pntr);
 312
 313   // All pointers are void* pointers.  Change type to 'v' now.
 314   if (pntr) {
 315     usgn = false;
 316     poly = false;
 317     type = 'v';
 318   }
 319   // Treat half-float ('h') types as unsigned short ('s') types.
 320   if (type == 'h') {
 321     type = 's';
 322     usgn = true;
 323   }
 324   usgn = usgn | poly | ((ck == ClassI || ck == ClassW) && scal && type != 'f');
 325
 326   if (scal) {
 327     SmallString<128> s;
 328
 329     if (usgn)
 330       s.push_back('U');
 331     else if (type == 'c')
 332       s.push_back('S'); // make chars explicitly signed
 333
 334     if (type == 'l') // 64-bit long
 335       s += "LLi";
 336     else
 337       s.push_back(type);
 338
 339     if (cnst)
 340       s.push_back('C');
 341     if (pntr)
 342       s.push_back('*');
 343     return s.str();
 344   }
 345
 346   // Since the return value must be one type, return a vector type of the
 347   // appropriate width which we will bitcast.  An exception is made for
 348   // returning structs of 2, 3, or 4 vectors which are returned in a sret-like
 349   // fashion, storing them to a pointer arg.
 350   if (ret) {
 351     if (mod >= '2' && mod <= '4')
 352       return "vv*"; // void result with void* first argument
 353     if (mod == 'f' || (ck != ClassB && type == 'f'))
 354       return quad ? "V4f" : "V2f";
 355     if (ck != ClassB && type == 's')
 356       return quad ? "V8s" : "V4s";
 357     if (ck != ClassB && type == 'i')
 358       return quad ? "V4i" : "V2i";
 359     if (ck != ClassB && type == 'l')
 360       return quad ? "V2LLi" : "V1LLi";
 361
 362     return quad ? "V16Sc" : "V8Sc";
 363   }
 364
 365   // Non-return array types are passed as individual vectors.
 366   if (mod == '2')
 367     return quad ? "V16ScV16Sc" : "V8ScV8Sc";
 368   if (mod == '3')
 369     return quad ? "V16ScV16ScV16Sc" : "V8ScV8ScV8Sc";
 370   if (mod == '4')
 371     return quad ? "V16ScV16ScV16ScV16Sc" : "V8ScV8ScV8ScV8Sc";
 372
 373   if (mod == 'f' || (ck != ClassB && type == 'f'))
 374     return quad ? "V4f" : "V2f";
 375   if (ck != ClassB && type == 's')
 376     return quad ? "V8s" : "V4s";
 377   if (ck != ClassB && type == 'i')
 378     return quad ? "V4i" : "V2i";
 379   if (ck != ClassB && type == 'l')
 380     return quad ? "V2LLi" : "V1LLi";
 381
 382   return quad ? "V16Sc" : "V8Sc";
 383 }
 384
 385 /// MangleName - Append a type or width suffix to a base neon function name,
 386 /// and insert a 'q' in the appropriate location if the operation works on
 387 /// 128b rather than 64b.   E.g. turn "vst2_lane" into "vst2q_lane_f32", etc.
 388 static std::string MangleName(const std::string &name, StringRef typestr,
 389                               ClassKind ck) {
 390   if (name == "vcvt_f32_f16")
 391     return name;
 392
 393   bool quad = false;
 394   bool poly = false;
 395   bool usgn = false;
 396   char type = ClassifyType(typestr, quad, poly, usgn);
 397
 398   std::string s = name;
 399
 400   switch (type) {
 401   case 'c':
 402     switch (ck) {
 403     case ClassS: s += poly ? "_p8" : usgn ? "_u8" : "_s8"; break;
 404     case ClassI: s += "_i8"; break;
 405     case ClassW: s += "_8"; break;
 406     default: break;
 407     }
 408     break;
 409   case 's':
 410     switch (ck) {
 411     case ClassS: s += poly ? "_p16" : usgn ? "_u16" : "_s16"; break;
 412     case ClassI: s += "_i16"; break;
 413     case ClassW: s += "_16"; break;
 414     default: break;
 415     }
 416     break;
 417   case 'i':
 418     switch (ck) {
 419     case ClassS: s += usgn ? "_u32" : "_s32"; break;
 420     case ClassI: s += "_i32"; break;
 421     case ClassW: s += "_32"; break;
 422     default: break;
 423     }
 424     break;
 425   case 'l':
 426     switch (ck) {
 427     case ClassS: s += usgn ? "_u64" : "_s64"; break;
 428     case ClassI: s += "_i64"; break;
 429     case ClassW: s += "_64"; break;
 430     default: break;
 431     }
 432     break;
 433   case 'h':
 434     switch (ck) {
 435     case ClassS:
 436     case ClassI: s += "_f16"; break;
 437     case ClassW: s += "_16"; break;
 438     default: break;
 439     }
 440     break;
 441   case 'f':
 442     switch (ck) {
 443     case ClassS:
 444     case ClassI: s += "_f32"; break;
 445     case ClassW: s += "_32"; break;
 446     default: break;
 447     }
 448     break;
 449   default:
 450     throw "unhandled type!";
 451     break;
 452   }
 453   if (ck == ClassB)
 454     s += "_v";
 455
 456   // Insert a 'q' before the first '_' character so that it ends up before
 457   // _lane or _n on vector-scalar operations.
 458   if (quad) {
 459     size_t pos = s.find('_');
 460     s = s.insert(pos, "q");
 461   }
 462   return s;
 463 }
 464
 465 // Generate the string "(argtype a, argtype b, ...)"
 466 static std::string GenArgs(const std::string &proto, StringRef typestr) {
 467   bool define = proto.find('i') != std::string::npos;
 468   char arg = 'a';
 469
 470   std::string s;
 471   s += "(";
 472
 473   for (unsigned i = 1, e = proto.size(); i != e; ++i, ++arg) {
 474     if (define) {
 475       // Immediate macro arguments are used directly instead of being assigned
 476       // to local temporaries; prepend an underscore prefix to make their
 477       // names consistent with the local temporaries.
 478       if (proto[i] == 'i')
 479         s += "__";
 480     } else {
 481       s += TypeString(proto[i], typestr) + " __";
 482     }
 483     s.push_back(arg);
 484     if ((i + 1) < e)
 485       s += ", ";
 486   }
 487
 488   s += ")";
 489   return s;
 490 }
 491
 492 // Macro arguments are not type-checked like inline function arguments, so
 493 // assign them to local temporaries to get the right type checking.
 494 static std::string GenMacroLocals(const std::string &proto, StringRef typestr) {
 495   char arg = 'a';
 496   std::string s;
 497
 498   for (unsigned i = 1, e = proto.size(); i != e; ++i, ++arg) {
 499     // Do not create a temporary for an immediate argument.
 500     // That would defeat the whole point of using a macro!
 501     if (proto[i] == 'i') continue;
 502
 503     s += TypeString(proto[i], typestr) + " __";
 504     s.push_back(arg);
 505     s += " = (";
 506     s.push_back(arg);
 507     s += "); ";
 508   }
 509
 510   s += "\\\n  ";
 511   return s;
 512 }
 513
 514 // Use the vmovl builtin to sign-extend or zero-extend a vector.
 515 static std::string Extend(StringRef typestr, const std::string &a) {
 516   std::string s;
 517   s = MangleName("vmovl", typestr, ClassS);
 518   s += "(" + a + ")";
 519   return s;
 520 }
 521
 522 static std::string Duplicate(unsigned nElts, StringRef typestr,
 523                              const std::string &a) {
 524   std::string s;
 525
 526   s = "(" + TypeString('d', typestr) + "){ ";
 527   for (unsigned i = 0; i != nElts; ++i) {
 528     s += a;
 529     if ((i + 1) < nElts)
 530       s += ", ";
 531   }
 532   s += " }";
 533
 534   return s;
 535 }
 536
 537 static std::string SplatLane(unsigned nElts, const std::string &vec,
 538                              const std::string &lane) {
 539   std::string s = "__builtin_shufflevector(" + vec + ", " + vec;
 540   for (unsigned i = 0; i < nElts; ++i)
 541     s += ", " + lane;
 542   s += ")";
 543   return s;
 544 }
 545
 546 static unsigned GetNumElements(StringRef typestr, bool &quad) {
 547   quad = false;
 548   bool dummy = false;
 549   char type = ClassifyType(typestr, quad, dummy, dummy);
 550   unsigned nElts = 0;
 551   switch (type) {
 552   case 'c': nElts = 8; break;
 553   case 's': nElts = 4; break;
 554   case 'i': nElts = 2; break;
 555   case 'l': nElts = 1; break;
 556   case 'h': nElts = 4; break;
 557   case 'f': nElts = 2; break;
 558   default:
 559     throw "unhandled type!";
 560     break;
 561   }
 562   if (quad) nElts <<= 1;
 563   return nElts;
 564 }
 565
 566 // Generate the definition for this intrinsic, e.g. "a + b" for OpAdd.
 567 static std::string GenOpString(OpKind op, const std::string &proto,
 568                                StringRef typestr) {
 569   bool quad;
 570   unsigned nElts = GetNumElements(typestr, quad);
 571
 572   // If this builtin takes an immediate argument, we need to #define it rather
 573   // than use a standard declaration, so that SemaChecking can range check
 574   // the immediate passed by the user.
 575   bool define = proto.find('i') != std::string::npos;
 576
 577   std::string ts = TypeString(proto[0], typestr);
 578   std::string s;
 579   if (!define) {
 580     s = "return ";
 581   }
 582
 583   switch(op) {
 584   case OpAdd:
 585     s += "__a + __b;";
 586     break;
 587   case OpAddl:
 588     s += Extend(typestr, "__a") + " + " + Extend(typestr, "__b") + ";";
 589     break;
 590   case OpAddw:
 591     s += "__a + " + Extend(typestr, "__b") + ";";
 592     break;
 593   case OpSub:
 594     s += "__a - __b;";
 595     break;
 596   case OpSubl:
 597     s += Extend(typestr, "__a") + " - " + Extend(typestr, "__b") + ";";
 598     break;
 599   case OpSubw:
 600     s += "__a - " + Extend(typestr, "__b") + ";";
 601     break;
 602   case OpMulN:
 603     s += "__a * " + Duplicate(nElts, typestr, "__b") + ";";
 604     break;
 605   case OpMulLane:
 606     s += "__a * " + SplatLane(nElts, "__b", "__c") + ";";
 607     break;
 608   case OpMul:
 609     s += "__a * __b;";
 610     break;
 611   case OpMullLane:
 612     s += MangleName("vmull", typestr, ClassS) + "(__a, " +
 613       SplatLane(nElts, "__b", "__c") + ");";
 614     break;
 615   case OpMlaN:
 616     s += "__a + (__b * " + Duplicate(nElts, typestr, "__c") + ");";
 617     break;
 618   case OpMlaLane:
 619     s += "__a + (__b * " + SplatLane(nElts, "__c", "__d") + ");";
 620     break;
 621   case OpMla:
 622     s += "__a + (__b * __c);";
 623     break;
 624   case OpMlalN:
 625     s += "__a + " + MangleName("vmull", typestr, ClassS) + "(__b, " +
 626       Duplicate(nElts, typestr, "__c") + ");";
 627     break;
 628   case OpMlalLane:
 629     s += "__a + " + MangleName("vmull", typestr, ClassS) + "(__b, " +
 630       SplatLane(nElts, "__c", "__d") + ");";
 631     break;
 632   case OpMlal:
 633     s += "__a + " + MangleName("vmull", typestr, ClassS) + "(__b, __c);";
 634     break;
 635   case OpMlsN:
 636     s += "__a - (__b * " + Duplicate(nElts, typestr, "__c") + ");";
 637     break;
 638   case OpMlsLane:
 639     s += "__a - (__b * " + SplatLane(nElts, "__c", "__d") + ");";
 640     break;
 641   case OpMls:
 642     s += "__a - (__b * __c);";
 643     break;
 644   case OpMlslN:
 645     s += "__a - " + MangleName("vmull", typestr, ClassS) + "(__b, " +
 646       Duplicate(nElts, typestr, "__c") + ");";
 647     break;
 648   case OpMlslLane:
 649     s += "__a - " + MangleName("vmull", typestr, ClassS) + "(__b, " +
 650       SplatLane(nElts, "__c", "__d") + ");";
 651     break;
 652   case OpMlsl:
 653     s += "__a - " + MangleName("vmull", typestr, ClassS) + "(__b, __c);";
 654     break;
 655   case OpQDMullLane:
 656     s += MangleName("vqdmull", typestr, ClassS) + "(__a, " +
 657       SplatLane(nElts, "__b", "__c") + ");";
 658     break;
 659   case OpQDMlalLane:
 660     s += MangleName("vqdmlal", typestr, ClassS) + "(__a, __b, " +
 661       SplatLane(nElts, "__c", "__d") + ");";
 662     break;
 663   case OpQDMlslLane:
 664     s += MangleName("vqdmlsl", typestr, ClassS) + "(__a, __b, " +
 665       SplatLane(nElts, "__c", "__d") + ");";
 666     break;
 667   case OpQDMulhLane:
 668     s += MangleName("vqdmulh", typestr, ClassS) + "(__a, " +
 669       SplatLane(nElts, "__b", "__c") + ");";
 670     break;
 671   case OpQRDMulhLane:
 672     s += MangleName("vqrdmulh", typestr, ClassS) + "(__a, " +
 673       SplatLane(nElts, "__b", "__c") + ");";
 674     break;
 675   case OpEq:
 676     s += "(" + ts + ")(__a == __b);";
 677     break;
 678   case OpGe:
 679     s += "(" + ts + ")(__a >= __b);";
 680     break;
 681   case OpLe:
 682     s += "(" + ts + ")(__a <= __b);";
 683     break;
 684   case OpGt:
 685     s += "(" + ts + ")(__a > __b);";
 686     break;
 687   case OpLt:
 688     s += "(" + ts + ")(__a < __b);";
 689     break;
 690   case OpNeg:
 691     s += " -__a;";
 692     break;
 693   case OpNot:
 694     s += " ~__a;";
 695     break;
 696   case OpAnd:
 697     s += "__a & __b;";
 698     break;
 699   case OpOr:
 700     s += "__a | __b;";
 701     break;
 702   case OpXor:
 703     s += "__a ^ __b;";
 704     break;
 705   case OpAndNot:
 706     s += "__a & ~__b;";
 707     break;
 708   case OpOrNot:
 709     s += "__a | ~__b;";
 710     break;
 711   case OpCast:
 712     s += "(" + ts + ")__a;";
 713     break;
 714   case OpConcat:
 715     s += "(" + ts + ")__builtin_shufflevector((int64x1_t)__a";
 716     s += ", (int64x1_t)__b, 0, 1);";
 717     break;
 718   case OpHi:
 719     s += "(" + ts +
 720       ")__builtin_shufflevector((int64x2_t)__a, (int64x2_t)__a, 1);";
 721     break;
 722   case OpLo:
 723     s += "(" + ts +
 724       ")__builtin_shufflevector((int64x2_t)__a, (int64x2_t)__a, 0);";
 725     break;
 726   case OpDup:
 727     s += Duplicate(nElts, typestr, "__a") + ";";
 728     break;
 729   case OpDupLane:
 730     s += SplatLane(nElts, "__a", "__b") + ";";
 731     break;
 732   case OpSelect:
 733     // ((0 & 1) | (~0 & 2))
 734     s += "(" + ts + ")";
 735     ts = TypeString(proto[1], typestr);
 736     s += "((__a & (" + ts + ")__b) | ";
 737     s += "(~__a & (" + ts + ")__c));";
 738     break;
 739   case OpRev16:
 740     s += "__builtin_shufflevector(__a, __a";
 741     for (unsigned i = 2; i <= nElts; i += 2)
 742       for (unsigned j = 0; j != 2; ++j)
 743         s += ", " + utostr(i - j - 1);
 744     s += ");";
 745     break;
 746   case OpRev32: {
 747     unsigned WordElts = nElts >> (1 + (int)quad);
 748     s += "__builtin_shufflevector(__a, __a";
 749     for (unsigned i = WordElts; i <= nElts; i += WordElts)
 750       for (unsigned j = 0; j != WordElts; ++j)
 751         s += ", " + utostr(i - j - 1);
 752     s += ");";
 753     break;
 754   }
 755   case OpRev64: {
 756     unsigned DblWordElts = nElts >> (int)quad;
 757     s += "__builtin_shufflevector(__a, __a";
 758     for (unsigned i = DblWordElts; i <= nElts; i += DblWordElts)
 759       for (unsigned j = 0; j != DblWordElts; ++j)
 760         s += ", " + utostr(i - j - 1);
 761     s += ");";
 762     break;
 763   }
 764   case OpAbdl: {
 765     std::string abd = MangleName("vabd", typestr, ClassS) + "(__a, __b)";
 766     if (typestr[0] != 'U') {
 767       // vabd results are always unsigned and must be zero-extended.
 768       std::string utype = "U" + typestr.str();
 769       s += "(" + TypeString(proto[0], typestr) + ")";
 770       abd = "(" + TypeString('d', utype) + ")" + abd;
 771       s += Extend(utype, abd) + ";";
 772     } else {
 773       s += Extend(typestr, abd) + ";";
 774     }
 775     break;
 776   }
 777   case OpAba:
 778     s += "__a + " + MangleName("vabd", typestr, ClassS) + "(__b, __c);";
 779     break;
 780   case OpAbal: {
 781     s += "__a + ";
 782     std::string abd = MangleName("vabd", typestr, ClassS) + "(__b, __c)";
 783     if (typestr[0] != 'U') {
 784       // vabd results are always unsigned and must be zero-extended.
 785       std::string utype = "U" + typestr.str();
 786       s += "(" + TypeString(proto[0], typestr) + ")";
 787       abd = "(" + TypeString('d', utype) + ")" + abd;
 788       s += Extend(utype, abd) + ";";
 789     } else {
 790       s += Extend(typestr, abd) + ";";
 791     }
 792     break;
 793   }
 794   default:
 795     throw "unknown OpKind!";
 796     break;
 797   }
 798   return s;
 799 }
 800
 801 static unsigned GetNeonEnum(const std::string &proto, StringRef typestr) {
 802   unsigned mod = proto[0];
 803   unsigned ret = 0;
 804
 805   if (mod == 'v' || mod == 'f')
 806     mod = proto[1];
 807
 808   bool quad = false;
 809   bool poly = false;
 810   bool usgn = false;
 811   bool scal = false;
 812   bool cnst = false;
 813   bool pntr = false;
 814
 815   // Base type to get the type string for.
 816   char type = ClassifyType(typestr, quad, poly, usgn);
 817
 818   // Based on the modifying character, change the type and width if necessary.
 819   type = ModType(mod, type, quad, poly, usgn, scal, cnst, pntr);
 820
 821   if (usgn)
 822     ret |= 0x08;
 823   if (quad && proto[1] != 'g')
 824     ret |= 0x10;
 825
 826   switch (type) {
 827     case 'c':
 828       ret |= poly ? 5 : 0;
 829       break;
 830     case 's':
 831       ret |= poly ? 6 : 1;
 832       break;
 833     case 'i':
 834       ret |= 2;
 835       break;
 836     case 'l':
 837       ret |= 3;
 838       break;
 839     case 'h':
 840       ret |= 7;
 841       break;
 842     case 'f':
 843       ret |= 4;
 844       break;
 845     default:
 846       throw "unhandled type!";
 847       break;
 848   }
 849   return ret;
 850 }
 851
 852 // Generate the definition for this intrinsic, e.g. __builtin_neon_cls(a)
 853 static std::string GenBuiltin(const std::string &name, const std::string &proto,
 854                               StringRef typestr, ClassKind ck) {
 855   std::string s;
 856
 857   // If this builtin returns a struct 2, 3, or 4 vectors, pass it as an implicit
 858   // sret-like argument.
 859   bool sret = (proto[0] >= '2' && proto[0] <= '4');
 860
 861   // If this builtin takes an immediate argument, we need to #define it rather
 862   // than use a standard declaration, so that SemaChecking can range check
 863   // the immediate passed by the user.
 864   bool define = proto.find('i') != std::string::npos;
 865
 866   // Check if the prototype has a scalar operand with the type of the vector
 867   // elements.  If not, bitcasting the args will take care of arg checking.
 868   // The actual signedness etc. will be taken care of with special enums.
 869   if (proto.find('s') == std::string::npos)
 870     ck = ClassB;
 871
 872   if (proto[0] != 'v') {
 873     std::string ts = TypeString(proto[0], typestr);
 874
 875     if (define) {
 876       if (sret)
 877         s += ts + " r; ";
 878       else
 879         s += "(" + ts + ")";
 880     } else if (sret) {
 881       s += ts + " r; ";
 882     } else {
 883       s += "return (" + ts + ")";
 884     }
 885   }
 886
 887   bool splat = proto.find('a') != std::string::npos;
 888
 889   s += "__builtin_neon_";
 890   if (splat) {
 891     // Call the non-splat builtin: chop off the "_n" suffix from the name.
 892     std::string vname(name, 0, name.size()-2);
 893     s += MangleName(vname, typestr, ck);
 894   } else {
 895     s += MangleName(name, typestr, ck);
 896   }
 897   s += "(";
 898
 899   // Pass the address of the return variable as the first argument to sret-like
 900   // builtins.
 901   if (sret)
 902     s += "&r, ";
 903
 904   char arg = 'a';
 905   for (unsigned i = 1, e = proto.size(); i != e; ++i, ++arg) {
 906     std::string args = std::string(&arg, 1);
 907
 908     // Use the local temporaries instead of the macro arguments.
 909     args = "__" + args;
 910
 911     bool argQuad = false;
 912     bool argPoly = false;
 913     bool argUsgn = false;
 914     bool argScalar = false;
 915     bool dummy = false;
 916     char argType = ClassifyType(typestr, argQuad, argPoly, argUsgn);
 917     argType = ModType(proto[i], argType, argQuad, argPoly, argUsgn, argScalar,
 918                       dummy, dummy);
 919
 920     // Handle multiple-vector values specially, emitting each subvector as an
 921     // argument to the __builtin.
 922     if (proto[i] >= '2' && proto[i] <= '4') {
 923       // Check if an explicit cast is needed.
 924       if (argType != 'c' || argPoly || argUsgn)
 925         args = (argQuad ? "(int8x16_t)" : "(int8x8_t)") + args;
 926
 927       for (unsigned vi = 0, ve = proto[i] - '0'; vi != ve; ++vi) {
 928         s += args + ".val[" + utostr(vi) + "]";
 929         if ((vi + 1) < ve)
 930           s += ", ";
 931       }
 932       if ((i + 1) < e)
 933         s += ", ";
 934
 935       continue;
 936     }
 937
 938     if (splat && (i + 1) == e)
 939       args = Duplicate(GetNumElements(typestr, argQuad), typestr, args);
 940
 941     // Check if an explicit cast is needed.
 942     if ((splat || !argScalar) &&
 943         ((ck == ClassB && argType != 'c') || argPoly || argUsgn)) {
 944       std::string argTypeStr = "c";
 945       if (ck != ClassB)
 946         argTypeStr = argType;
 947       if (argQuad)
 948         argTypeStr = "Q" + argTypeStr;
 949       args = "(" + TypeString('d', argTypeStr) + ")" + args;
 950     }
 951
 952     s += args;
 953     if ((i + 1) < e)
 954       s += ", ";
 955   }
 956
 957   // Extra constant integer to hold type class enum for this function, e.g. s8
 958   if (ck == ClassB)
 959     s += ", " + utostr(GetNeonEnum(proto, typestr));
 960
 961   s += ");";
 962
 963   if (proto[0] != 'v' && sret) {
 964     if (define)
 965       s += " r;";
 966     else
 967       s += " return r;";
 968   }
 969   return s;
 970 }
 971
 972 static std::string GenBuiltinDef(const std::string &name,
 973                                  const std::string &proto,
 974                                  StringRef typestr, ClassKind ck) {
 975   std::string s("BUILTIN(__builtin_neon_");
 976
 977   // If all types are the same size, bitcasting the args will take care
 978   // of arg checking.  The actual signedness etc. will be taken care of with
 979   // special enums.
 980   if (proto.find('s') == std::string::npos)
 981     ck = ClassB;
 982
 983   s += MangleName(name, typestr, ck);
 984   s += ", \"";
 985
 986   for (unsigned i = 0, e = proto.size(); i != e; ++i)
 987     s += BuiltinTypeString(proto[i], typestr, ck, i == 0);
 988
 989   // Extra constant integer to hold type class enum for this function, e.g. s8
 990   if (ck == ClassB)
 991     s += "i";
 992
 993   s += "\", \"n\")";
 994   return s;
 995 }
 996
 997 static std::string GenIntrinsic(const std::string &name,
 998                                 const std::string &proto,
 999                                 StringRef outTypeStr, StringRef inTypeStr,
1000                                 OpKind kind, ClassKind classKind) {
1001   assert(!proto.empty() && "");
1002   bool define = proto.find('i') != std::string::npos;
1003   std::string s;
1004
1005   // static always inline + return type
1006   if (define)
1007     s += "#define ";
1008   else
1009     s += "__ai " + TypeString(proto[0], outTypeStr) + " ";
1010
1011   // Function name with type suffix
1012   std::string mangledName = MangleName(name, outTypeStr, ClassS);
1013   if (outTypeStr != inTypeStr) {
1014     // If the input type is different (e.g., for vreinterpret), append a suffix
1015     // for the input type.  String off a "Q" (quad) prefix so that MangleName
1016     // does not insert another "q" in the name.
1017     unsigned typeStrOff = (inTypeStr[0] == 'Q' ? 1 : 0);
1018     StringRef inTypeNoQuad = inTypeStr.substr(typeStrOff);
1019     mangledName = MangleName(mangledName, inTypeNoQuad, ClassS);
1020   }
1021   s += mangledName;
1022
1023   // Function arguments
1024   s += GenArgs(proto, inTypeStr);
1025
1026   // Definition.
1027   if (define) {
1028     s += " __extension__ ({ \\\n  ";
1029     s += GenMacroLocals(proto, inTypeStr);
1030   } else {
1031     s += " { \\\n  ";
1032   }
1033
1034   if (kind != OpNone)
1035     s += GenOpString(kind, proto, outTypeStr);
1036   else
1037     s += GenBuiltin(name, proto, outTypeStr, classKind);
1038   if (define)
1039     s += " })";
1040   else
1041     s += " }";
1042   s += "\n";
1043   return s;
1044 }
1045
1046 /// run - Read the records in arm_neon.td and output arm_neon.h.  arm_neon.h
1047 /// is comprised of type definitions and function declarations.
1048 void NeonEmitter::run(raw_ostream &OS) {
1049   OS <<
1050     "/*===---- arm_neon.h - ARM Neon intrinsics ------------------------------"
1051     "---===\n"
1052     " *\n"
1053     " * Permission is hereby granted, free of charge, to any person obtaining "
1054     "a copy\n"
1055     " * of this software and associated documentation files (the \"Software\"),"
1056     " to deal\n"
1057     " * in the Software without restriction, including without limitation the "
1058     "rights\n"
1059     " * to use, copy, modify, merge, publish, distribute, sublicense, "
1060     "and/or sell\n"
1061     " * copies of the Software, and to permit persons to whom the Software is\n"
1062     " * furnished to do so, subject to the following conditions:\n"
1063     " *\n"
1064     " * The above copyright notice and this permission notice shall be "
1065     "included in\n"
1066     " * all copies or substantial portions of the Software.\n"
1067     " *\n"
1068     " * THE SOFTWARE IS PROVIDED \"AS IS\", WITHOUT WARRANTY OF ANY KIND, "
1069     "EXPRESS OR\n"
1070     " * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF "
1071     "MERCHANTABILITY,\n"
1072     " * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT "
1073     "SHALL THE\n"
1074     " * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR "
1075     "OTHER\n"
1076     " * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, "
1077     "ARISING FROM,\n"
1078     " * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER "
1079     "DEALINGS IN\n"
1080     " * THE SOFTWARE.\n"
1081     " *\n"
1082     " *===--------------------------------------------------------------------"
1083     "---===\n"
1084     " */\n\n";
1085
1086   OS << "#ifndef __ARM_NEON_H\n";
1087   OS << "#define __ARM_NEON_H\n\n";
1088
1089   OS << "#ifndef __ARM_NEON__\n";
1090   OS << "#error \"NEON support not enabled\"\n";
1091   OS << "#endif\n\n";
1092
1093   OS << "#include <stdint.h>\n\n";
1094
1095   // Emit NEON-specific scalar typedefs.
1096   OS << "typedef float float32_t;\n";
1097   OS << "typedef int8_t poly8_t;\n";
1098   OS << "typedef int16_t poly16_t;\n";
1099   OS << "typedef uint16_t float16_t;\n";
1100
1101   // Emit Neon vector typedefs.
1102   std::string TypedefTypes("cQcsQsiQilQlUcQUcUsQUsUiQUiUlQUlhQhfQfPcQPcPsQPs");
1103   SmallVector<StringRef, 24> TDTypeVec;
1104   ParseTypes(0, TypedefTypes, TDTypeVec);
1105
1106   // Emit vector typedefs.
1107   for (unsigned i = 0, e = TDTypeVec.size(); i != e; ++i) {
1108     bool dummy, quad = false, poly = false;
1109     (void) ClassifyType(TDTypeVec[i], quad, poly, dummy);
1110     if (poly)
1111       OS << "typedef __attribute__((neon_polyvector_type(";
1112     else
1113       OS << "typedef __attribute__((neon_vector_type(";
1114
1115     unsigned nElts = GetNumElements(TDTypeVec[i], quad);
1116     OS << utostr(nElts) << "))) ";
1117     if (nElts < 10)
1118       OS << " ";
1119
1120     OS << TypeString('s', TDTypeVec[i]);
1121     OS << " " << TypeString('d', TDTypeVec[i]) << ";\n";
1122   }
1123   OS << "\n";
1124
1125   // Emit struct typedefs.
1126   for (unsigned vi = 2; vi != 5; ++vi) {
1127     for (unsigned i = 0, e = TDTypeVec.size(); i != e; ++i) {
1128       std::string ts = TypeString('d', TDTypeVec[i]);
1129       std::string vs = TypeString('0' + vi, TDTypeVec[i]);
1130       OS << "typedef struct " << vs << " {\n";
1131       OS << "  " << ts << " val";
1132       OS << "[" << utostr(vi) << "]";
1133       OS << ";\n} ";
1134       OS << vs << ";\n\n";
1135     }
1136   }
1137
1138   OS << "#define __ai static __attribute__((__always_inline__))\n\n";
1139
1140   std::vector<Record*> RV = Records.getAllDerivedDefinitions("Inst");
1141
1142   // Emit vmovl, vmull and vabd intrinsics first so they can be used by other
1143   // intrinsics.  (Some of the saturating multiply instructions are also
1144   // used to implement the corresponding "_lane" variants, but tablegen
1145   // sorts the records into alphabetical order so that the "_lane" variants
1146   // come after the intrinsics they use.)
1147   emitIntrinsic(OS, Records.getDef("VMOVL"));
1148   emitIntrinsic(OS, Records.getDef("VMULL"));
1149   emitIntrinsic(OS, Records.getDef("VABD"));
1150
1151   for (unsigned i = 0, e = RV.size(); i != e; ++i) {
1152     Record *R = RV[i];
1153     if (R->getName() != "VMOVL" &&
1154         R->getName() != "VMULL" &&
1155         R->getName() != "VABD")
1156       emitIntrinsic(OS, R);
1157   }
1158
1159   OS << "#undef __ai\n\n";
1160   OS << "#endif /* __ARM_NEON_H */\n";
1161 }
1162
1163 /// emitIntrinsic - Write out the arm_neon.h header file definitions for the
1164 /// intrinsics specified by record R.
1165 void NeonEmitter::emitIntrinsic(raw_ostream &OS, Record *R) {
1166   std::string name = R->getValueAsString("Name");
1167   std::string Proto = R->getValueAsString("Prototype");
1168   std::string Types = R->getValueAsString("Types");
1169
1170   SmallVector<StringRef, 16> TypeVec;
1171   ParseTypes(R, Types, TypeVec);
1172
1173   OpKind kind = OpMap[R->getValueAsDef("Operand")->getName()];
1174
1175   ClassKind classKind = ClassNone;
1176   if (R->getSuperClasses().size() >= 2)
1177     classKind = ClassMap[R->getSuperClasses()[1]];
1178   if (classKind == ClassNone && kind == OpNone)
1179     throw TGError(R->getLoc(), "Builtin has no class kind");
1180
1181   for (unsigned ti = 0, te = TypeVec.size(); ti != te; ++ti) {
1182     if (kind == OpReinterpret) {
1183       bool outQuad = false;
1184       bool dummy = false;
1185       (void)ClassifyType(TypeVec[ti], outQuad, dummy, dummy);
1186       for (unsigned srcti = 0, srcte = TypeVec.size();
1187            srcti != srcte; ++srcti) {
1188         bool inQuad = false;
1189         (void)ClassifyType(TypeVec[srcti], inQuad, dummy, dummy);
1190         if (srcti == ti || inQuad != outQuad)
1191           continue;
1192         OS << GenIntrinsic(name, Proto, TypeVec[ti], TypeVec[srcti],
1193                            OpCast, ClassS);
1194       }
1195     } else {
1196       OS << GenIntrinsic(name, Proto, TypeVec[ti], TypeVec[ti],
1197                          kind, classKind);
1198     }
1199   }
1200   OS << "\n";
1201 }
1202
1203 static unsigned RangeFromType(const char mod, StringRef typestr) {
1204   // base type to get the type string for.
1205   bool quad = false, dummy = false;
1206   char type = ClassifyType(typestr, quad, dummy, dummy);
1207   type = ModType(mod, type, quad, dummy, dummy, dummy, dummy, dummy);
1208
1209   switch (type) {
1210     case 'c':
1211       return (8 << (int)quad) - 1;
1212     case 'h':
1213     case 's':
1214       return (4 << (int)quad) - 1;
1215     case 'f':
1216     case 'i':
1217       return (2 << (int)quad) - 1;
1218     case 'l':
1219       return (1 << (int)quad) - 1;
1220     default:
1221       throw "unhandled type!";
1222       break;
1223   }
1224   assert(0 && "unreachable");
1225   return 0;
1226 }
1227
1228 /// runHeader - Emit a file with sections defining:
1229 /// 1. the NEON section of BuiltinsARM.def.
1230 /// 2. the SemaChecking code for the type overload checking.
1231 /// 3. the SemaChecking code for validation of intrinsic immedate arguments.
1232 void NeonEmitter::runHeader(raw_ostream &OS) {
1233   std::vector<Record*> RV = Records.getAllDerivedDefinitions("Inst");
1234
1235   StringMap<OpKind> EmittedMap;
1236
1237   // Generate BuiltinsARM.def for NEON
1238   OS << "#ifdef GET_NEON_BUILTINS\n";
1239   for (unsigned i = 0, e = RV.size(); i != e; ++i) {
1240     Record *R = RV[i];
1241     OpKind k = OpMap[R->getValueAsDef("Operand")->getName()];
1242     if (k != OpNone)
1243       continue;
1244
1245     std::string Proto = R->getValueAsString("Prototype");
1246
1247     // Functions with 'a' (the splat code) in the type prototype should not get
1248     // their own builtin as they use the non-splat variant.
1249     if (Proto.find('a') != std::string::npos)
1250       continue;
1251
1252     std::string Types = R->getValueAsString("Types");
1253     SmallVector<StringRef, 16> TypeVec;
1254     ParseTypes(R, Types, TypeVec);
1255
1256     if (R->getSuperClasses().size() < 2)
1257       throw TGError(R->getLoc(), "Builtin has no class kind");
1258
1259     std::string name = R->getValueAsString("Name");
1260     ClassKind ck = ClassMap[R->getSuperClasses()[1]];
1261
1262     for (unsigned ti = 0, te = TypeVec.size(); ti != te; ++ti) {
1263       // Generate the BuiltinsARM.def declaration for this builtin, ensuring
1264       // that each unique BUILTIN() macro appears only once in the output
1265       // stream.
1266       std::string bd = GenBuiltinDef(name, Proto, TypeVec[ti], ck);
1267       if (EmittedMap.count(bd))
1268         continue;
1269
1270       EmittedMap[bd] = OpNone;
1271       OS << bd << "\n";
1272     }
1273   }
1274   OS << "#endif\n\n";
1275
1276   // Generate the overloaded type checking code for SemaChecking.cpp
1277   OS << "#ifdef GET_NEON_OVERLOAD_CHECK\n";
1278   for (unsigned i = 0, e = RV.size(); i != e; ++i) {
1279     Record *R = RV[i];
1280     OpKind k = OpMap[R->getValueAsDef("Operand")->getName()];
1281     if (k != OpNone)
1282       continue;
1283
1284     std::string Proto = R->getValueAsString("Prototype");
1285     std::string Types = R->getValueAsString("Types");
1286     std::string name = R->getValueAsString("Name");
1287
1288     // Functions with 'a' (the splat code) in the type prototype should not get
1289     // their own builtin as they use the non-splat variant.
1290     if (Proto.find('a') != std::string::npos)
1291       continue;
1292
1293     // Functions which have a scalar argument cannot be overloaded, no need to
1294     // check them if we are emitting the type checking code.
1295     if (Proto.find('s') != std::string::npos)
1296       continue;
1297
1298     SmallVector<StringRef, 16> TypeVec;
1299     ParseTypes(R, Types, TypeVec);
1300
1301     if (R->getSuperClasses().size() < 2)
1302       throw TGError(R->getLoc(), "Builtin has no class kind");
1303
1304     int si = -1, qi = -1;
1305     unsigned mask = 0, qmask = 0;
1306     for (unsigned ti = 0, te = TypeVec.size(); ti != te; ++ti) {
1307       // Generate the switch case(s) for this builtin for the type validation.
1308       bool quad = false, poly = false, usgn = false;
1309       (void) ClassifyType(TypeVec[ti], quad, poly, usgn);
1310
1311       if (quad) {
1312         qi = ti;
1313         qmask |= 1 << GetNeonEnum(Proto, TypeVec[ti]);
1314       } else {
1315         si = ti;
1316         mask |= 1 << GetNeonEnum(Proto, TypeVec[ti]);
1317       }
1318     }
1319     if (mask)
1320       OS << "case ARM::BI__builtin_neon_"
1321          << MangleName(name, TypeVec[si], ClassB)
1322          << ": mask = " << "0x" << utohexstr(mask) << "; break;\n";
1323     if (qmask)
1324       OS << "case ARM::BI__builtin_neon_"
1325          << MangleName(name, TypeVec[qi], ClassB)
1326          << ": mask = " << "0x" << utohexstr(qmask) << "; break;\n";
1327   }
1328   OS << "#endif\n\n";
1329
1330   // Generate the intrinsic range checking code for shift/lane immediates.
1331   OS << "#ifdef GET_NEON_IMMEDIATE_CHECK\n";
1332   for (unsigned i = 0, e = RV.size(); i != e; ++i) {
1333     Record *R = RV[i];
1334
1335     OpKind k = OpMap[R->getValueAsDef("Operand")->getName()];
1336     if (k != OpNone)
1337       continue;
1338
1339     std::string name = R->getValueAsString("Name");
1340     std::string Proto = R->getValueAsString("Prototype");
1341     std::string Types = R->getValueAsString("Types");
1342
1343     // Functions with 'a' (the splat code) in the type prototype should not get
1344     // their own builtin as they use the non-splat variant.
1345     if (Proto.find('a') != std::string::npos)
1346       continue;
1347
1348     // Functions which do not have an immediate do not need to have range
1349     // checking code emitted.
1350     size_t immPos = Proto.find('i');
1351     if (immPos == std::string::npos)
1352       continue;
1353
1354     SmallVector<StringRef, 16> TypeVec;
1355     ParseTypes(R, Types, TypeVec);
1356
1357     if (R->getSuperClasses().size() < 2)
1358       throw TGError(R->getLoc(), "Builtin has no class kind");
1359
1360     ClassKind ck = ClassMap[R->getSuperClasses()[1]];
1361
1362     for (unsigned ti = 0, te = TypeVec.size(); ti != te; ++ti) {
1363       std::string namestr, shiftstr, rangestr;
1364
1365       // Builtins which are overloaded by type will need to have their upper
1366       // bound computed at Sema time based on the type constant.
1367       if (Proto.find('s') == std::string::npos) {
1368         ck = ClassB;
1369         if (R->getValueAsBit("isShift")) {
1370           shiftstr = ", true";
1371
1372           // Right shifts have an 'r' in the name, left shifts do not.
1373           if (name.find('r') != std::string::npos)
1374             rangestr = "l = 1; ";
1375         }
1376         rangestr += "u = RFT(TV" + shiftstr + ")";
1377       } else {
1378         // The immediate generally refers to a lane in the preceding argument.
1379         assert(immPos > 0 && "unexpected immediate operand");
1380         rangestr = "u = " + utostr(RangeFromType(Proto[immPos-1], TypeVec[ti]));
1381       }
1382       // Make sure cases appear only once by uniquing them in a string map.
1383       namestr = MangleName(name, TypeVec[ti], ck);
1384       if (EmittedMap.count(namestr))
1385         continue;
1386       EmittedMap[namestr] = OpNone;
1387
1388       // Calculate the index of the immediate that should be range checked.
1389       unsigned immidx = 0;
1390
1391       // Builtins that return a struct of multiple vectors have an extra
1392       // leading arg for the struct return.
1393       if (Proto[0] >= '2' && Proto[0] <= '4')
1394         ++immidx;
1395
1396       // Add one to the index for each argument until we reach the immediate
1397       // to be checked.  Structs of vectors are passed as multiple arguments.
1398       for (unsigned ii = 1, ie = Proto.size(); ii != ie; ++ii) {
1399         switch (Proto[ii]) {
1400           default:  immidx += 1; break;
1401           case '2': immidx += 2; break;
1402           case '3': immidx += 3; break;
1403           case '4': immidx += 4; break;
1404           case 'i': ie = ii + 1; break;
1405         }
1406       }
1407       OS << "case ARM::BI__builtin_neon_" << MangleName(name, TypeVec[ti], ck)
1408          << ": i = " << immidx << "; " << rangestr << "; break;\n";
1409     }
1410   }
1411   OS << "#endif\n\n";
1412 }
1413
1414 /// GenTest - Write out a test for the intrinsic specified by the name and
1415 /// type strings, including the embedded patterns for FileCheck to match.
1416 static std::string GenTest(const std::string &name,
1417                            const std::string &proto,
1418                            StringRef outTypeStr, StringRef inTypeStr,
1419                            bool isShift) {
1420   assert(!proto.empty() && "");
1421   std::string s;
1422
1423   // Function name with type suffix
1424   std::string mangledName = MangleName(name, outTypeStr, ClassS);
1425   if (outTypeStr != inTypeStr) {
1426     // If the input type is different (e.g., for vreinterpret), append a suffix
1427     // for the input type.  String off a "Q" (quad) prefix so that MangleName
1428     // does not insert another "q" in the name.
1429     unsigned typeStrOff = (inTypeStr[0] == 'Q' ? 1 : 0);
1430     StringRef inTypeNoQuad = inTypeStr.substr(typeStrOff);
1431     mangledName = MangleName(mangledName, inTypeNoQuad, ClassS);
1432   }
1433
1434   // Emit the FileCheck patterns.
1435   s += "// CHECK: test_" + mangledName + "\n";
1436   // s += "// CHECK: \n"; // FIXME: + expected instruction opcode.
1437
1438   // Emit the start of the test function.
1439   s += TypeString(proto[0], outTypeStr) + " test_" + mangledName + "(";
1440   char arg = 'a';
1441   std::string comma;
1442   for (unsigned i = 1, e = proto.size(); i != e; ++i, ++arg) {
1443     // Do not create arguments for values that must be immediate constants.
1444     if (proto[i] == 'i')
1445       continue;
1446     s += comma + TypeString(proto[i], inTypeStr) + " ";
1447     s.push_back(arg);
1448     comma = ", ";
1449   }
1450   s += ") { \\\n  ";
1451
1452   if (proto[0] != 'v')
1453     s += "return ";
1454   s += mangledName + "(";
1455   arg = 'a';
1456   for (unsigned i = 1, e = proto.size(); i != e; ++i, ++arg) {
1457     if (proto[i] == 'i') {
1458       // For immediate operands, test the maximum value.
1459       if (isShift)
1460         s += "1"; // FIXME
1461       else
1462         // The immediate generally refers to a lane in the preceding argument.
1463         s += utostr(RangeFromType(proto[i-1], inTypeStr));
1464     } else {
1465       s.push_back(arg);
1466     }
1467     if ((i + 1) < e)
1468       s += ", ";
1469   }
1470   s += ");\n}\n\n";
1471   return s;
1472 }
1473
1474 /// runTests - Write out a complete set of tests for all of the Neon
1475 /// intrinsics.
1476 void NeonEmitter::runTests(raw_ostream &OS) {
1477   OS <<
1478     "// RUN: %clang_cc1 -triple thumbv7-apple-darwin \\\n"
1479     "// RUN:  -target-cpu cortex-a9 -ffreestanding -S -o - %s | FileCheck %s\n"
1480     "\n"
1481     "#include <arm_neon.h>\n"
1482     "\n";
1483
1484   std::vector<Record*> RV = Records.getAllDerivedDefinitions("Inst");
1485   for (unsigned i = 0, e = RV.size(); i != e; ++i) {
1486     Record *R = RV[i];
1487     std::string name = R->getValueAsString("Name");
1488     std::string Proto = R->getValueAsString("Prototype");
1489     std::string Types = R->getValueAsString("Types");
1490     bool isShift = R->getValueAsBit("isShift");
1491
1492     SmallVector<StringRef, 16> TypeVec;
1493     ParseTypes(R, Types, TypeVec);
1494
1495     OpKind kind = OpMap[R->getValueAsDef("Operand")->getName()];
1496     for (unsigned ti = 0, te = TypeVec.size(); ti != te; ++ti) {
1497       if (kind == OpReinterpret) {
1498         bool outQuad = false;
1499         bool dummy = false;
1500         (void)ClassifyType(TypeVec[ti], outQuad, dummy, dummy);
1501         for (unsigned srcti = 0, srcte = TypeVec.size();
1502              srcti != srcte; ++srcti) {
1503           bool inQuad = false;
1504           (void)ClassifyType(TypeVec[srcti], inQuad, dummy, dummy);
1505           if (srcti == ti || inQuad != outQuad)
1506             continue;
1507           OS << GenTest(name, Proto, TypeVec[ti], TypeVec[srcti], isShift);
1508         }
1509       } else {
1510         OS << GenTest(name, Proto, TypeVec[ti], TypeVec[ti], isShift);
1511       }
1512     }
1513     OS << "\n";
1514   }
1515 }
1516