1 //===- NeonEmitter.cpp - Generate arm_neon.h for use with clang -*- C++ -*-===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
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.
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
20 // Additional validation code can be generated by this file when runHeader() is
21 // called, rather than the normal run() entry point.
23 //===----------------------------------------------------------------------===//
25 #include "NeonEmitter.h"
26 #include "llvm/ADT/SmallString.h"
27 #include "llvm/ADT/SmallVector.h"
28 #include "llvm/ADT/StringExtras.h"
33 /// ParseTypes - break down a string such as "fQf" into a vector of StringRefs,
34 /// which each StringRef representing a single type declared in the string.
35 /// for "fQf" we would end up with 2 StringRefs, "f", and "Qf", representing
36 /// 2xfloat and 4xfloat respectively.
37 static void ParseTypes(Record *r, std::string &s,
38 SmallVectorImpl<StringRef> &TV) {
39 const char *data = s.data();
42 for (unsigned i = 0, e = s.size(); i != e; ++i, ++len) {
43 if (data[len] == 'P' || data[len] == 'Q' || data[len] == 'U')
55 throw TGError(r->getLoc(),
56 "Unexpected letter: " + std::string(data + len, 1));
59 TV.push_back(StringRef(data, len + 1));
65 /// Widen - Convert a type code into the next wider type. char -> short,
66 /// short -> int, etc.
67 static char Widen(const char t) {
75 default: throw "unhandled type in widen!";
80 /// Narrow - Convert a type code into the next smaller type. short -> char,
81 /// float -> half float, etc.
82 static char Narrow(const char t) {
92 default: throw "unhandled type in narrow!";
97 /// For a particular StringRef, return the base type code, and whether it has
98 /// the quad-vector, polynomial, or unsigned modifiers set.
99 static char ClassifyType(StringRef ty, bool &quad, bool &poly, bool &usgn) {
103 if (ty[off] == 'Q') {
109 if (ty[off] == 'P') {
114 // remember unsigned.
115 if (ty[off] == 'U') {
120 // base type to get the type string for.
124 /// ModType - Transform a type code and its modifiers based on a mod code. The
125 /// mod code definitions may be found at the top of arm_neon.td.
126 static char ModType(const char mod, char type, bool &quad, bool &poly,
127 bool &usgn, bool &scal, bool &cnst, bool &pntr) {
200 /// TypeString - for a modifier and type, generate the name of the typedef for
201 /// that type. QUc -> uint8x8_t.
202 static std::string TypeString(const char mod, StringRef typestr) {
215 // base type to get the type string for.
216 char type = ClassifyType(typestr, quad, poly, usgn);
218 // Based on the modifying character, change the type and width if necessary.
219 type = ModType(mod, type, quad, poly, usgn, scal, cnst, pntr);
228 s += poly ? "poly8" : "int8";
231 s += quad ? "x16" : "x8";
234 s += poly ? "poly16" : "int16";
237 s += quad ? "x8" : "x4";
243 s += quad ? "x4" : "x2";
249 s += quad ? "x2" : "x1";
255 s += quad ? "x8" : "x4";
261 s += quad ? "x4" : "x2";
264 throw "unhandled type!";
275 // Append _t, finishing the type string typedef type.
287 /// BuiltinTypeString - for a modifier and type, generate the clang
288 /// BuiltinsARM.def prototype code for the function. See the top of clang's
289 /// Builtins.def for a description of the type strings.
290 static std::string BuiltinTypeString(const char mod, StringRef typestr,
291 ClassKind ck, bool ret) {
304 // base type to get the type string for.
305 char type = ClassifyType(typestr, quad, poly, usgn);
307 // Based on the modifying character, change the type and width if necessary.
308 type = ModType(mod, type, quad, poly, usgn, scal, cnst, pntr);
310 // All pointers are void* pointers. Change type to 'v' now.
316 // Treat half-float ('h') types as unsigned short ('s') types.
321 usgn = usgn | poly | ((ck == ClassI || ck == ClassW) && scal && type != 'f');
328 else if (type == 'c')
329 s.push_back('S'); // make chars explicitly signed
331 if (type == 'l') // 64-bit long
343 // Since the return value must be one type, return a vector type of the
344 // appropriate width which we will bitcast. An exception is made for
345 // returning structs of 2, 3, or 4 vectors which are returned in a sret-like
346 // fashion, storing them to a pointer arg.
348 if (mod >= '2' && mod <= '4')
349 return "vv*"; // void result with void* first argument
350 if (mod == 'f' || (ck != ClassB && type == 'f'))
351 return quad ? "V4f" : "V2f";
352 if (ck != ClassB && type == 's')
353 return quad ? "V8s" : "V4s";
354 if (ck != ClassB && type == 'i')
355 return quad ? "V4i" : "V2i";
356 if (ck != ClassB && type == 'l')
357 return quad ? "V2LLi" : "V1LLi";
359 return quad ? "V16Sc" : "V8Sc";
362 // Non-return array types are passed as individual vectors.
364 return quad ? "V16ScV16Sc" : "V8ScV8Sc";
366 return quad ? "V16ScV16ScV16Sc" : "V8ScV8ScV8Sc";
368 return quad ? "V16ScV16ScV16ScV16Sc" : "V8ScV8ScV8ScV8Sc";
370 if (mod == 'f' || (ck != ClassB && type == 'f'))
371 return quad ? "V4f" : "V2f";
372 if (ck != ClassB && type == 's')
373 return quad ? "V8s" : "V4s";
374 if (ck != ClassB && type == 'i')
375 return quad ? "V4i" : "V2i";
376 if (ck != ClassB && type == 'l')
377 return quad ? "V2LLi" : "V1LLi";
379 return quad ? "V16Sc" : "V8Sc";
382 /// MangleName - Append a type or width suffix to a base neon function name,
383 /// and insert a 'q' in the appropriate location if the operation works on
384 /// 128b rather than 64b. E.g. turn "vst2_lane" into "vst2q_lane_f32", etc.
385 static std::string MangleName(const std::string &name, StringRef typestr,
387 if (name == "vcvt_f32_f16")
393 char type = ClassifyType(typestr, quad, poly, usgn);
395 std::string s = name;
400 case ClassS: s += poly ? "_p8" : usgn ? "_u8" : "_s8"; break;
401 case ClassI: s += "_i8"; break;
402 case ClassW: s += "_8"; break;
408 case ClassS: s += poly ? "_p16" : usgn ? "_u16" : "_s16"; break;
409 case ClassI: s += "_i16"; break;
410 case ClassW: s += "_16"; break;
416 case ClassS: s += usgn ? "_u32" : "_s32"; break;
417 case ClassI: s += "_i32"; break;
418 case ClassW: s += "_32"; break;
424 case ClassS: s += usgn ? "_u64" : "_s64"; break;
425 case ClassI: s += "_i64"; break;
426 case ClassW: s += "_64"; break;
433 case ClassI: s += "_f16"; break;
434 case ClassW: s += "_16"; break;
441 case ClassI: s += "_f32"; break;
442 case ClassW: s += "_32"; break;
447 throw "unhandled type!";
453 // Insert a 'q' before the first '_' character so that it ends up before
454 // _lane or _n on vector-scalar operations.
456 size_t pos = s.find('_');
457 s = s.insert(pos, "q");
462 // Generate the string "(argtype a, argtype b, ...)"
463 static std::string GenArgs(const std::string &proto, StringRef typestr) {
464 bool define = proto.find('i') != std::string::npos;
470 for (unsigned i = 1, e = proto.size(); i != e; ++i, ++arg) {
472 // Immediate macro arguments are used directly instead of being assigned
473 // to local temporaries; prepend an underscore prefix to make their
474 // names consistent with the local temporaries.
478 s += TypeString(proto[i], typestr) + " __";
489 // Macro arguments are not type-checked like inline function arguments, so
490 // assign them to local temporaries to get the right type checking.
491 static std::string GenMacroLocals(const std::string &proto, StringRef typestr) {
495 for (unsigned i = 1, e = proto.size(); i != e; ++i, ++arg) {
496 // Do not create a temporary for an immediate argument.
497 // That would defeat the whole point of using a macro!
498 if (proto[i] == 'i') continue;
500 s += TypeString(proto[i], typestr) + " __";
511 // Use the vmovl builtin to sign-extend or zero-extend a vector.
512 static std::string Extend(const std::string &proto, StringRef typestr,
513 const std::string &a) {
515 s = MangleName("vmovl", typestr, ClassS);
520 static std::string Duplicate(unsigned nElts, StringRef typestr,
521 const std::string &a) {
524 s = "(" + TypeString('d', typestr) + "){ ";
525 for (unsigned i = 0; i != nElts; ++i) {
535 static std::string SplatLane(unsigned nElts, const std::string &vec,
536 const std::string &lane) {
537 std::string s = "__builtin_shufflevector(" + vec + ", " + vec;
538 for (unsigned i = 0; i < nElts; ++i)
544 static unsigned GetNumElements(StringRef typestr, bool &quad) {
547 char type = ClassifyType(typestr, quad, dummy, dummy);
550 case 'c': nElts = 8; break;
551 case 's': nElts = 4; break;
552 case 'i': nElts = 2; break;
553 case 'l': nElts = 1; break;
554 case 'h': nElts = 4; break;
555 case 'f': nElts = 2; break;
557 throw "unhandled type!";
560 if (quad) nElts <<= 1;
564 // Generate the definition for this intrinsic, e.g. "a + b" for OpAdd.
565 static std::string GenOpString(OpKind op, const std::string &proto,
568 unsigned nElts = GetNumElements(typestr, quad);
570 // If this builtin takes an immediate argument, we need to #define it rather
571 // than use a standard declaration, so that SemaChecking can range check
572 // the immediate passed by the user.
573 bool define = proto.find('i') != std::string::npos;
575 std::string ts = TypeString(proto[0], typestr);
577 if (op == OpHi || op == OpLo) {
578 s = "union { " + ts + " r; double d; } u; u.d = ";
579 } else if (!define) {
591 s += "__a * " + Duplicate(nElts, typestr, "__b") + ";";
594 s += "__a * " + SplatLane(nElts, "__b", "__c") + ";";
600 s += Extend(proto, typestr, "__a") + " * " +
601 Extend(proto, typestr,
602 Duplicate(nElts << (int)quad, typestr, "__b")) + ";";
605 s += Extend(proto, typestr, "__a") + " * " +
606 Extend(proto, typestr, SplatLane(nElts, "__b", "__c")) + ";";
609 s += Extend(proto, typestr, "__a") + " * " +
610 Extend(proto, typestr, "__b") + ";";
613 s += "__a + (__b * " + Duplicate(nElts, typestr, "__c") + ");";
616 s += "__a + (__b * " + SplatLane(nElts, "__c", "__d") + ");";
619 s += "__a + (__b * __c);";
622 s += "__a + (" + Extend(proto, typestr, "__b") + " * " +
623 Extend(proto, typestr, Duplicate(nElts, typestr, "__c")) + ");";
626 s += "__a + (" + Extend(proto, typestr, "__b") + " * " +
627 Extend(proto, typestr, SplatLane(nElts, "__c", "__d")) + ");";
630 s += "__a + (" + Extend(proto, typestr, "__b") + " * " +
631 Extend(proto, typestr, "__c") + ");";
634 s += "__a - (__b * " + Duplicate(nElts, typestr, "__c") + ");";
637 s += "__a - (__b * " + SplatLane(nElts, "__c", "__d") + ");";
640 s += "__a - (__b * __c);";
643 s += "__a - (" + Extend(proto, typestr, "__b") + " * " +
644 Extend(proto, typestr, Duplicate(nElts, typestr, "__c")) + ");";
647 s += "__a - (" + Extend(proto, typestr, "__b") + " * " +
648 Extend(proto, typestr, SplatLane(nElts, "__c", "__d")) + ");";
651 s += "__a - (" + Extend(proto, typestr, "__b") + " * " +
652 Extend(proto, typestr, "__c") + ");";
655 s += "(" + ts + ")(__a == __b);";
658 s += "(" + ts + ")(__a >= __b);";
661 s += "(" + ts + ")(__a <= __b);";
664 s += "(" + ts + ")(__a > __b);";
667 s += "(" + ts + ")(__a < __b);";
691 s += "(" + ts + ")__a;";
694 s += "(" + ts + ")__builtin_shufflevector((int64x1_t)__a";
695 s += ", (int64x1_t)__b, 0, 1);";
698 s += "(((float64x2_t)__a)[1]);";
701 s += "(((float64x2_t)__a)[0]);";
704 s += Duplicate(nElts, typestr, "__a") + ";";
707 s += SplatLane(nElts, "__a", "__b") + ";";
710 // ((0 & 1) | (~0 & 2))
712 ts = TypeString(proto[1], typestr);
713 s += "((__a & (" + ts + ")__b) | ";
714 s += "(~__a & (" + ts + ")__c));";
717 s += "__builtin_shufflevector(__a, __a";
718 for (unsigned i = 2; i <= nElts; i += 2)
719 for (unsigned j = 0; j != 2; ++j)
720 s += ", " + utostr(i - j - 1);
724 unsigned WordElts = nElts >> (1 + (int)quad);
725 s += "__builtin_shufflevector(__a, __a";
726 for (unsigned i = WordElts; i <= nElts; i += WordElts)
727 for (unsigned j = 0; j != WordElts; ++j)
728 s += ", " + utostr(i - j - 1);
733 unsigned DblWordElts = nElts >> (int)quad;
734 s += "__builtin_shufflevector(__a, __a";
735 for (unsigned i = DblWordElts; i <= nElts; i += DblWordElts)
736 for (unsigned j = 0; j != DblWordElts; ++j)
737 s += ", " + utostr(i - j - 1);
742 throw "unknown OpKind!";
745 if (op == OpHi || op == OpLo) {
753 static unsigned GetNeonEnum(const std::string &proto, StringRef typestr) {
754 unsigned mod = proto[0];
757 if (mod == 'v' || mod == 'f')
767 // Base type to get the type string for.
768 char type = ClassifyType(typestr, quad, poly, usgn);
770 // Based on the modifying character, change the type and width if necessary.
771 type = ModType(mod, type, quad, poly, usgn, scal, cnst, pntr);
775 if (quad && proto[1] != 'g')
798 throw "unhandled type!";
804 // Generate the definition for this intrinsic, e.g. __builtin_neon_cls(a)
805 static std::string GenBuiltin(const std::string &name, const std::string &proto,
806 StringRef typestr, ClassKind ck) {
809 // If this builtin returns a struct 2, 3, or 4 vectors, pass it as an implicit
810 // sret-like argument.
811 bool sret = (proto[0] >= '2' && proto[0] <= '4');
813 // If this builtin takes an immediate argument, we need to #define it rather
814 // than use a standard declaration, so that SemaChecking can range check
815 // the immediate passed by the user.
816 bool define = proto.find('i') != std::string::npos;
818 // Check if the prototype has a scalar operand with the type of the vector
819 // elements. If not, bitcasting the args will take care of arg checking.
820 // The actual signedness etc. will be taken care of with special enums.
821 if (proto.find('s') == std::string::npos)
824 if (proto[0] != 'v') {
825 std::string ts = TypeString(proto[0], typestr);
835 s += "return (" + ts + ")";
839 bool splat = proto.find('a') != std::string::npos;
841 s += "__builtin_neon_";
843 // Call the non-splat builtin: chop off the "_n" suffix from the name.
844 std::string vname(name, 0, name.size()-2);
845 s += MangleName(vname, typestr, ck);
847 s += MangleName(name, typestr, ck);
851 // Pass the address of the return variable as the first argument to sret-like
857 for (unsigned i = 1, e = proto.size(); i != e; ++i, ++arg) {
858 std::string args = std::string(&arg, 1);
860 // Use the local temporaries instead of the macro arguments.
863 bool argQuad = false;
864 bool argPoly = false;
865 bool argUsgn = false;
866 bool argScalar = false;
868 char argType = ClassifyType(typestr, argQuad, argPoly, argUsgn);
869 argType = ModType(proto[i], argType, argQuad, argPoly, argUsgn, argScalar,
872 // Handle multiple-vector values specially, emitting each subvector as an
873 // argument to the __builtin.
874 if (proto[i] >= '2' && proto[i] <= '4') {
875 // Check if an explicit cast is needed.
876 if (argType != 'c' || argPoly || argUsgn)
877 args = (argQuad ? "(int8x16_t)" : "(int8x8_t)") + args;
879 for (unsigned vi = 0, ve = proto[i] - '0'; vi != ve; ++vi) {
880 s += args + ".val[" + utostr(vi) + "]";
890 if (splat && (i + 1) == e)
891 args = Duplicate(GetNumElements(typestr, argQuad), typestr, args);
893 // Check if an explicit cast is needed.
894 if ((splat || !argScalar) &&
895 ((ck == ClassB && argType != 'c') || argPoly || argUsgn)) {
896 std::string argTypeStr = "c";
898 argTypeStr = argType;
900 argTypeStr = "Q" + argTypeStr;
901 args = "(" + TypeString('d', argTypeStr) + ")" + args;
909 // Extra constant integer to hold type class enum for this function, e.g. s8
911 s += ", " + utostr(GetNeonEnum(proto, typestr));
915 if (proto[0] != 'v' && sret) {
924 static std::string GenBuiltinDef(const std::string &name,
925 const std::string &proto,
926 StringRef typestr, ClassKind ck) {
927 std::string s("BUILTIN(__builtin_neon_");
929 // If all types are the same size, bitcasting the args will take care
930 // of arg checking. The actual signedness etc. will be taken care of with
932 if (proto.find('s') == std::string::npos)
935 s += MangleName(name, typestr, ck);
938 for (unsigned i = 0, e = proto.size(); i != e; ++i)
939 s += BuiltinTypeString(proto[i], typestr, ck, i == 0);
941 // Extra constant integer to hold type class enum for this function, e.g. s8
949 static std::string GenIntrinsic(const std::string &name,
950 const std::string &proto,
951 StringRef outTypeStr, StringRef inTypeStr,
952 OpKind kind, ClassKind classKind) {
953 assert(!proto.empty() && "");
954 bool define = proto.find('i') != std::string::npos;
957 // static always inline + return type
961 s += "__ai " + TypeString(proto[0], outTypeStr) + " ";
963 // Function name with type suffix
964 std::string mangledName = MangleName(name, outTypeStr, ClassS);
965 if (outTypeStr != inTypeStr) {
966 // If the input type is different (e.g., for vreinterpret), append a suffix
967 // for the input type. String off a "Q" (quad) prefix so that MangleName
968 // does not insert another "q" in the name.
969 unsigned typeStrOff = (inTypeStr[0] == 'Q' ? 1 : 0);
970 StringRef inTypeNoQuad = inTypeStr.substr(typeStrOff);
971 mangledName = MangleName(mangledName, inTypeNoQuad, ClassS);
975 // Function arguments
976 s += GenArgs(proto, inTypeStr);
980 s += " __extension__ ({ \\\n ";
981 s += GenMacroLocals(proto, inTypeStr);
987 s += GenOpString(kind, proto, outTypeStr);
989 s += GenBuiltin(name, proto, outTypeStr, classKind);
998 /// run - Read the records in arm_neon.td and output arm_neon.h. arm_neon.h
999 /// is comprised of type definitions and function declarations.
1000 void NeonEmitter::run(raw_ostream &OS) {
1001 EmitSourceFileHeader("ARM NEON Header", OS);
1003 // FIXME: emit license into file?
1005 OS << "#ifndef __ARM_NEON_H\n";
1006 OS << "#define __ARM_NEON_H\n\n";
1008 OS << "#ifndef __ARM_NEON__\n";
1009 OS << "#error \"NEON support not enabled\"\n";
1012 OS << "#include <stdint.h>\n\n";
1014 // Emit NEON-specific scalar typedefs.
1015 OS << "typedef float float32_t;\n";
1016 OS << "typedef int8_t poly8_t;\n";
1017 OS << "typedef int16_t poly16_t;\n";
1018 OS << "typedef uint16_t float16_t;\n";
1020 // Emit Neon vector typedefs.
1021 std::string TypedefTypes("cQcsQsiQilQlUcQUcUsQUsUiQUiUlQUlhQhfQfPcQPcPsQPs");
1022 SmallVector<StringRef, 24> TDTypeVec;
1023 ParseTypes(0, TypedefTypes, TDTypeVec);
1025 // Emit vector typedefs.
1026 for (unsigned i = 0, e = TDTypeVec.size(); i != e; ++i) {
1027 bool dummy, quad = false, poly = false;
1028 (void) ClassifyType(TDTypeVec[i], quad, poly, dummy);
1030 OS << "typedef __attribute__((neon_polyvector_type(";
1032 OS << "typedef __attribute__((neon_vector_type(";
1034 unsigned nElts = GetNumElements(TDTypeVec[i], quad);
1035 OS << utostr(nElts) << "))) ";
1039 OS << TypeString('s', TDTypeVec[i]);
1040 OS << " " << TypeString('d', TDTypeVec[i]) << ";\n";
1043 OS << "typedef __attribute__((__vector_size__(8))) "
1044 "double float64x1_t;\n";
1045 OS << "typedef __attribute__((__vector_size__(16))) "
1046 "double float64x2_t;\n";
1049 // Emit struct typedefs.
1050 for (unsigned vi = 2; vi != 5; ++vi) {
1051 for (unsigned i = 0, e = TDTypeVec.size(); i != e; ++i) {
1052 std::string ts = TypeString('d', TDTypeVec[i]);
1053 std::string vs = TypeString('0' + vi, TDTypeVec[i]);
1054 OS << "typedef struct " << vs << " {\n";
1055 OS << " " << ts << " val";
1056 OS << "[" << utostr(vi) << "]";
1058 OS << vs << ";\n\n";
1062 OS << "#define __ai static __attribute__((__always_inline__))\n\n";
1064 std::vector<Record*> RV = Records.getAllDerivedDefinitions("Inst");
1066 // Emit vmovl intrinsics first so they can be used by other intrinsics.
1067 emitIntrinsic(OS, Records.getDef("VMOVL"));
1069 // Unique the return+pattern types, and assign them.
1070 for (unsigned i = 0, e = RV.size(); i != e; ++i) {
1072 if (R->getName() != "VMOVL")
1073 emitIntrinsic(OS, R);
1076 OS << "#undef __ai\n\n";
1077 OS << "#endif /* __ARM_NEON_H */\n";
1080 /// emitIntrinsic - Write out the arm_neon.h header file definitions for the
1081 /// intrinsics specified by record R.
1082 void NeonEmitter::emitIntrinsic(raw_ostream &OS, Record *R) {
1083 std::string name = R->getValueAsString("Name");
1084 std::string Proto = R->getValueAsString("Prototype");
1085 std::string Types = R->getValueAsString("Types");
1087 SmallVector<StringRef, 16> TypeVec;
1088 ParseTypes(R, Types, TypeVec);
1090 OpKind kind = OpMap[R->getValueAsDef("Operand")->getName()];
1092 ClassKind classKind = ClassNone;
1093 if (R->getSuperClasses().size() >= 2)
1094 classKind = ClassMap[R->getSuperClasses()[1]];
1095 if (classKind == ClassNone && kind == OpNone)
1096 throw TGError(R->getLoc(), "Builtin has no class kind");
1098 for (unsigned ti = 0, te = TypeVec.size(); ti != te; ++ti) {
1099 if (kind == OpReinterpret) {
1100 bool outQuad = false;
1102 (void)ClassifyType(TypeVec[ti], outQuad, dummy, dummy);
1103 for (unsigned srcti = 0, srcte = TypeVec.size();
1104 srcti != srcte; ++srcti) {
1105 bool inQuad = false;
1106 (void)ClassifyType(TypeVec[srcti], inQuad, dummy, dummy);
1107 if (srcti == ti || inQuad != outQuad)
1109 OS << GenIntrinsic(name, Proto, TypeVec[ti], TypeVec[srcti],
1113 OS << GenIntrinsic(name, Proto, TypeVec[ti], TypeVec[ti],
1120 static unsigned RangeFromType(StringRef typestr) {
1121 // base type to get the type string for.
1122 bool quad = false, dummy = false;
1123 char type = ClassifyType(typestr, quad, dummy, dummy);
1127 return (8 << (int)quad) - 1;
1130 return (4 << (int)quad) - 1;
1133 return (2 << (int)quad) - 1;
1135 return (1 << (int)quad) - 1;
1137 throw "unhandled type!";
1140 assert(0 && "unreachable");
1144 /// runHeader - Emit a file with sections defining:
1145 /// 1. the NEON section of BuiltinsARM.def.
1146 /// 2. the SemaChecking code for the type overload checking.
1147 /// 3. the SemaChecking code for validation of intrinsic immedate arguments.
1148 void NeonEmitter::runHeader(raw_ostream &OS) {
1149 std::vector<Record*> RV = Records.getAllDerivedDefinitions("Inst");
1151 StringMap<OpKind> EmittedMap;
1153 // Generate BuiltinsARM.def for NEON
1154 OS << "#ifdef GET_NEON_BUILTINS\n";
1155 for (unsigned i = 0, e = RV.size(); i != e; ++i) {
1157 OpKind k = OpMap[R->getValueAsDef("Operand")->getName()];
1161 std::string Proto = R->getValueAsString("Prototype");
1163 // Functions with 'a' (the splat code) in the type prototype should not get
1164 // their own builtin as they use the non-splat variant.
1165 if (Proto.find('a') != std::string::npos)
1168 std::string Types = R->getValueAsString("Types");
1169 SmallVector<StringRef, 16> TypeVec;
1170 ParseTypes(R, Types, TypeVec);
1172 if (R->getSuperClasses().size() < 2)
1173 throw TGError(R->getLoc(), "Builtin has no class kind");
1175 std::string name = R->getValueAsString("Name");
1176 ClassKind ck = ClassMap[R->getSuperClasses()[1]];
1178 for (unsigned ti = 0, te = TypeVec.size(); ti != te; ++ti) {
1179 // Generate the BuiltinsARM.def declaration for this builtin, ensuring
1180 // that each unique BUILTIN() macro appears only once in the output
1182 std::string bd = GenBuiltinDef(name, Proto, TypeVec[ti], ck);
1183 if (EmittedMap.count(bd))
1186 EmittedMap[bd] = OpNone;
1192 // Generate the overloaded type checking code for SemaChecking.cpp
1193 OS << "#ifdef GET_NEON_OVERLOAD_CHECK\n";
1194 for (unsigned i = 0, e = RV.size(); i != e; ++i) {
1196 OpKind k = OpMap[R->getValueAsDef("Operand")->getName()];
1200 std::string Proto = R->getValueAsString("Prototype");
1201 std::string Types = R->getValueAsString("Types");
1202 std::string name = R->getValueAsString("Name");
1204 // Functions with 'a' (the splat code) in the type prototype should not get
1205 // their own builtin as they use the non-splat variant.
1206 if (Proto.find('a') != std::string::npos)
1209 // Functions which have a scalar argument cannot be overloaded, no need to
1210 // check them if we are emitting the type checking code.
1211 if (Proto.find('s') != std::string::npos)
1214 SmallVector<StringRef, 16> TypeVec;
1215 ParseTypes(R, Types, TypeVec);
1217 if (R->getSuperClasses().size() < 2)
1218 throw TGError(R->getLoc(), "Builtin has no class kind");
1220 int si = -1, qi = -1;
1221 unsigned mask = 0, qmask = 0;
1222 for (unsigned ti = 0, te = TypeVec.size(); ti != te; ++ti) {
1223 // Generate the switch case(s) for this builtin for the type validation.
1224 bool quad = false, poly = false, usgn = false;
1225 (void) ClassifyType(TypeVec[ti], quad, poly, usgn);
1229 qmask |= 1 << GetNeonEnum(Proto, TypeVec[ti]);
1232 mask |= 1 << GetNeonEnum(Proto, TypeVec[ti]);
1236 OS << "case ARM::BI__builtin_neon_"
1237 << MangleName(name, TypeVec[si], ClassB)
1238 << ": mask = " << "0x" << utohexstr(mask) << "; break;\n";
1240 OS << "case ARM::BI__builtin_neon_"
1241 << MangleName(name, TypeVec[qi], ClassB)
1242 << ": mask = " << "0x" << utohexstr(qmask) << "; break;\n";
1246 // Generate the intrinsic range checking code for shift/lane immediates.
1247 OS << "#ifdef GET_NEON_IMMEDIATE_CHECK\n";
1248 for (unsigned i = 0, e = RV.size(); i != e; ++i) {
1251 OpKind k = OpMap[R->getValueAsDef("Operand")->getName()];
1255 std::string name = R->getValueAsString("Name");
1256 std::string Proto = R->getValueAsString("Prototype");
1257 std::string Types = R->getValueAsString("Types");
1259 // Functions with 'a' (the splat code) in the type prototype should not get
1260 // their own builtin as they use the non-splat variant.
1261 if (Proto.find('a') != std::string::npos)
1264 // Functions which do not have an immediate do not need to have range
1265 // checking code emitted.
1266 if (Proto.find('i') == std::string::npos)
1269 SmallVector<StringRef, 16> TypeVec;
1270 ParseTypes(R, Types, TypeVec);
1272 if (R->getSuperClasses().size() < 2)
1273 throw TGError(R->getLoc(), "Builtin has no class kind");
1275 ClassKind ck = ClassMap[R->getSuperClasses()[1]];
1277 for (unsigned ti = 0, te = TypeVec.size(); ti != te; ++ti) {
1278 std::string namestr, shiftstr, rangestr;
1280 // Builtins which are overloaded by type will need to have their upper
1281 // bound computed at Sema time based on the type constant.
1282 if (Proto.find('s') == std::string::npos) {
1284 if (R->getValueAsBit("isShift")) {
1285 shiftstr = ", true";
1287 // Right shifts have an 'r' in the name, left shifts do not.
1288 if (name.find('r') != std::string::npos)
1289 rangestr = "l = 1; ";
1291 rangestr += "u = RFT(TV" + shiftstr + ")";
1293 rangestr = "u = " + utostr(RangeFromType(TypeVec[ti]));
1295 // Make sure cases appear only once by uniquing them in a string map.
1296 namestr = MangleName(name, TypeVec[ti], ck);
1297 if (EmittedMap.count(namestr))
1299 EmittedMap[namestr] = OpNone;
1301 // Calculate the index of the immediate that should be range checked.
1302 unsigned immidx = 0;
1304 // Builtins that return a struct of multiple vectors have an extra
1305 // leading arg for the struct return.
1306 if (Proto[0] >= '2' && Proto[0] <= '4')
1309 // Add one to the index for each argument until we reach the immediate
1310 // to be checked. Structs of vectors are passed as multiple arguments.
1311 for (unsigned ii = 1, ie = Proto.size(); ii != ie; ++ii) {
1312 switch (Proto[ii]) {
1313 default: immidx += 1; break;
1314 case '2': immidx += 2; break;
1315 case '3': immidx += 3; break;
1316 case '4': immidx += 4; break;
1317 case 'i': ie = ii + 1; break;
1320 OS << "case ARM::BI__builtin_neon_" << MangleName(name, TypeVec[ti], ck)
1321 << ": i = " << immidx << "; " << rangestr << "; break;\n";