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. A complete set of tests
22 // for Neon intrinsics can be generated by calling the runTests() entry point.
24 //===----------------------------------------------------------------------===//
26 #include "NeonEmitter.h"
27 #include "llvm/ADT/SmallString.h"
28 #include "llvm/ADT/SmallVector.h"
29 #include "llvm/ADT/StringExtras.h"
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();
43 for (unsigned i = 0, e = s.size(); i != e; ++i, ++len) {
44 if (data[len] == 'P' || data[len] == 'Q' || data[len] == 'U')
56 throw TGError(r->getLoc(),
57 "Unexpected letter: " + std::string(data + len, 1));
60 TV.push_back(StringRef(data, len + 1));
66 /// Widen - Convert a type code into the next wider type. char -> short,
67 /// short -> int, etc.
68 static char Widen(const char t) {
78 default: throw "unhandled type in widen!";
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) {
95 default: throw "unhandled type in narrow!";
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) {
106 if (ty[off] == 'Q') {
112 if (ty[off] == 'P') {
117 // remember unsigned.
118 if (ty[off] == 'U') {
123 // base type to get the type string for.
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) {
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) {
218 // base type to get the type string for.
219 char type = ClassifyType(typestr, quad, poly, usgn);
221 // Based on the modifying character, change the type and width if necessary.
222 type = ModType(mod, type, quad, poly, usgn, scal, cnst, pntr);
231 s += poly ? "poly8" : "int8";
234 s += quad ? "x16" : "x8";
237 s += poly ? "poly16" : "int16";
240 s += quad ? "x8" : "x4";
246 s += quad ? "x4" : "x2";
252 s += quad ? "x2" : "x1";
258 s += quad ? "x8" : "x4";
264 s += quad ? "x4" : "x2";
267 throw "unhandled type!";
278 // Append _t, finishing the type string typedef type.
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) {
307 // base type to get the type string for.
308 char type = ClassifyType(typestr, quad, poly, usgn);
310 // Based on the modifying character, change the type and width if necessary.
311 type = ModType(mod, type, quad, poly, usgn, scal, cnst, pntr);
313 // All pointers are void* pointers. Change type to 'v' now.
319 // Treat half-float ('h') types as unsigned short ('s') types.
324 usgn = usgn | poly | ((ck == ClassI || ck == ClassW) && scal && type != 'f');
331 else if (type == 'c')
332 s.push_back('S'); // make chars explicitly signed
334 if (type == 'l') // 64-bit long
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.
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";
362 return quad ? "V16Sc" : "V8Sc";
365 // Non-return array types are passed as individual vectors.
367 return quad ? "V16ScV16Sc" : "V8ScV8Sc";
369 return quad ? "V16ScV16ScV16Sc" : "V8ScV8ScV8Sc";
371 return quad ? "V16ScV16ScV16ScV16Sc" : "V8ScV8ScV8ScV8Sc";
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";
382 return quad ? "V16Sc" : "V8Sc";
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,
390 if (name == "vcvt_f32_f16")
396 char type = ClassifyType(typestr, quad, poly, usgn);
398 std::string s = name;
403 case ClassS: s += poly ? "_p8" : usgn ? "_u8" : "_s8"; break;
404 case ClassI: s += "_i8"; break;
405 case ClassW: s += "_8"; break;
411 case ClassS: s += poly ? "_p16" : usgn ? "_u16" : "_s16"; break;
412 case ClassI: s += "_i16"; break;
413 case ClassW: s += "_16"; break;
419 case ClassS: s += usgn ? "_u32" : "_s32"; break;
420 case ClassI: s += "_i32"; break;
421 case ClassW: s += "_32"; break;
427 case ClassS: s += usgn ? "_u64" : "_s64"; break;
428 case ClassI: s += "_i64"; break;
429 case ClassW: s += "_64"; break;
436 case ClassI: s += "_f16"; break;
437 case ClassW: s += "_16"; break;
444 case ClassI: s += "_f32"; break;
445 case ClassW: s += "_32"; break;
450 throw "unhandled type!";
456 // Insert a 'q' before the first '_' character so that it ends up before
457 // _lane or _n on vector-scalar operations.
459 size_t pos = s.find('_');
460 s = s.insert(pos, "q");
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;
473 for (unsigned i = 1, e = proto.size(); i != e; ++i, ++arg) {
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.
481 s += TypeString(proto[i], typestr) + " __";
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) {
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;
503 s += TypeString(proto[i], typestr) + " __";
514 // Use the vmovl builtin to sign-extend or zero-extend a vector.
515 static std::string Extend(StringRef typestr, const std::string &a) {
517 s = MangleName("vmovl", typestr, ClassS);
522 static std::string Duplicate(unsigned nElts, StringRef typestr,
523 const std::string &a) {
526 s = "(" + TypeString('d', typestr) + "){ ";
527 for (unsigned i = 0; i != nElts; ++i) {
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)
546 static unsigned GetNumElements(StringRef typestr, bool &quad) {
549 char type = ClassifyType(typestr, quad, dummy, dummy);
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;
559 throw "unhandled type!";
562 if (quad) nElts <<= 1;
566 // Generate the definition for this intrinsic, e.g. "a + b" for OpAdd.
567 static std::string GenOpString(OpKind op, const std::string &proto,
570 unsigned nElts = GetNumElements(typestr, quad);
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;
577 std::string ts = TypeString(proto[0], typestr);
588 s += Extend(typestr, "__a") + " + " + Extend(typestr, "__b") + ";";
591 s += "__a + " + Extend(typestr, "__b") + ";";
597 s += Extend(typestr, "__a") + " - " + Extend(typestr, "__b") + ";";
600 s += "__a - " + Extend(typestr, "__b") + ";";
603 s += "__a * " + Duplicate(nElts, typestr, "__b") + ";";
606 s += "__a * " + SplatLane(nElts, "__b", "__c") + ";";
612 s += Extend(typestr, "__a") + " * " +
613 Extend(typestr, Duplicate(nElts << (int)quad, typestr, "__b")) + ";";
616 s += Extend(typestr, "__a") + " * " +
617 Extend(typestr, SplatLane(nElts, "__b", "__c")) + ";";
620 s += Extend(typestr, "__a") + " * " + Extend(typestr, "__b") + ";";
623 s += "__a + (__b * " + Duplicate(nElts, typestr, "__c") + ");";
626 s += "__a + (__b * " + SplatLane(nElts, "__c", "__d") + ");";
629 s += "__a + (__b * __c);";
632 s += "__a + (" + Extend(typestr, "__b") + " * " +
633 Extend(typestr, Duplicate(nElts, typestr, "__c")) + ");";
636 s += "__a + (" + Extend(typestr, "__b") + " * " +
637 Extend(typestr, SplatLane(nElts, "__c", "__d")) + ");";
640 s += "__a + (" + Extend(typestr, "__b") + " * " +
641 Extend(typestr, "__c") + ");";
644 s += "__a - (__b * " + Duplicate(nElts, typestr, "__c") + ");";
647 s += "__a - (__b * " + SplatLane(nElts, "__c", "__d") + ");";
650 s += "__a - (__b * __c);";
653 s += "__a - (" + Extend(typestr, "__b") + " * " +
654 Extend(typestr, Duplicate(nElts, typestr, "__c")) + ");";
657 s += "__a - (" + Extend(typestr, "__b") + " * " +
658 Extend(typestr, SplatLane(nElts, "__c", "__d")) + ");";
661 s += "__a - (" + Extend(typestr, "__b") + " * " +
662 Extend(typestr, "__c") + ");";
665 s += MangleName("vqdmull", typestr, ClassS) + "(__a, " +
666 SplatLane(nElts, "__b", "__c") + ");";
669 s += MangleName("vqdmlal", typestr, ClassS) + "(__a, __b, " +
670 SplatLane(nElts, "__c", "__d") + ");";
673 s += MangleName("vqdmlsl", typestr, ClassS) + "(__a, __b, " +
674 SplatLane(nElts, "__c", "__d") + ");";
677 s += MangleName("vqdmulh", typestr, ClassS) + "(__a, " +
678 SplatLane(nElts, "__b", "__c") + ");";
681 s += MangleName("vqrdmulh", typestr, ClassS) + "(__a, " +
682 SplatLane(nElts, "__b", "__c") + ");";
685 s += "(" + ts + ")(__a == __b);";
688 s += "(" + ts + ")(__a >= __b);";
691 s += "(" + ts + ")(__a <= __b);";
694 s += "(" + ts + ")(__a > __b);";
697 s += "(" + ts + ")(__a < __b);";
721 s += "(" + ts + ")__a;";
724 s += "(" + ts + ")__builtin_shufflevector((int64x1_t)__a";
725 s += ", (int64x1_t)__b, 0, 1);";
729 ")__builtin_shufflevector((int64x2_t)__a, (int64x2_t)__a, 1);";
733 ")__builtin_shufflevector((int64x2_t)__a, (int64x2_t)__a, 0);";
736 s += Duplicate(nElts, typestr, "__a") + ";";
739 s += SplatLane(nElts, "__a", "__b") + ";";
742 // ((0 & 1) | (~0 & 2))
744 ts = TypeString(proto[1], typestr);
745 s += "((__a & (" + ts + ")__b) | ";
746 s += "(~__a & (" + ts + ")__c));";
749 s += "__builtin_shufflevector(__a, __a";
750 for (unsigned i = 2; i <= nElts; i += 2)
751 for (unsigned j = 0; j != 2; ++j)
752 s += ", " + utostr(i - j - 1);
756 unsigned WordElts = nElts >> (1 + (int)quad);
757 s += "__builtin_shufflevector(__a, __a";
758 for (unsigned i = WordElts; i <= nElts; i += WordElts)
759 for (unsigned j = 0; j != WordElts; ++j)
760 s += ", " + utostr(i - j - 1);
765 unsigned DblWordElts = nElts >> (int)quad;
766 s += "__builtin_shufflevector(__a, __a";
767 for (unsigned i = DblWordElts; i <= nElts; i += DblWordElts)
768 for (unsigned j = 0; j != DblWordElts; ++j)
769 s += ", " + utostr(i - j - 1);
774 std::string abd = MangleName("vabd", typestr, ClassS) + "(__a, __b)";
775 if (typestr[0] != 'U') {
776 // vabd results are always unsigned and must be zero-extended.
777 std::string utype = "U" + typestr.str();
778 s += "(" + TypeString(proto[0], typestr) + ")";
779 abd = "(" + TypeString('d', utype) + ")" + abd;
780 s += Extend(utype, abd) + ";";
782 s += Extend(typestr, abd) + ";";
787 s += "__a + " + MangleName("vabd", typestr, ClassS) + "(__b, __c);";
791 std::string abd = MangleName("vabd", typestr, ClassS) + "(__b, __c)";
792 if (typestr[0] != 'U') {
793 // vabd results are always unsigned and must be zero-extended.
794 std::string utype = "U" + typestr.str();
795 s += "(" + TypeString(proto[0], typestr) + ")";
796 abd = "(" + TypeString('d', utype) + ")" + abd;
797 s += Extend(utype, abd) + ";";
799 s += Extend(typestr, abd) + ";";
804 throw "unknown OpKind!";
810 static unsigned GetNeonEnum(const std::string &proto, StringRef typestr) {
811 unsigned mod = proto[0];
814 if (mod == 'v' || mod == 'f')
824 // Base type to get the type string for.
825 char type = ClassifyType(typestr, quad, poly, usgn);
827 // Based on the modifying character, change the type and width if necessary.
828 type = ModType(mod, type, quad, poly, usgn, scal, cnst, pntr);
832 if (quad && proto[1] != 'g')
855 throw "unhandled type!";
861 // Generate the definition for this intrinsic, e.g. __builtin_neon_cls(a)
862 static std::string GenBuiltin(const std::string &name, const std::string &proto,
863 StringRef typestr, ClassKind ck) {
866 // If this builtin returns a struct 2, 3, or 4 vectors, pass it as an implicit
867 // sret-like argument.
868 bool sret = (proto[0] >= '2' && proto[0] <= '4');
870 // If this builtin takes an immediate argument, we need to #define it rather
871 // than use a standard declaration, so that SemaChecking can range check
872 // the immediate passed by the user.
873 bool define = proto.find('i') != std::string::npos;
875 // Check if the prototype has a scalar operand with the type of the vector
876 // elements. If not, bitcasting the args will take care of arg checking.
877 // The actual signedness etc. will be taken care of with special enums.
878 if (proto.find('s') == std::string::npos)
881 if (proto[0] != 'v') {
882 std::string ts = TypeString(proto[0], typestr);
892 s += "return (" + ts + ")";
896 bool splat = proto.find('a') != std::string::npos;
898 s += "__builtin_neon_";
900 // Call the non-splat builtin: chop off the "_n" suffix from the name.
901 std::string vname(name, 0, name.size()-2);
902 s += MangleName(vname, typestr, ck);
904 s += MangleName(name, typestr, ck);
908 // Pass the address of the return variable as the first argument to sret-like
914 for (unsigned i = 1, e = proto.size(); i != e; ++i, ++arg) {
915 std::string args = std::string(&arg, 1);
917 // Use the local temporaries instead of the macro arguments.
920 bool argQuad = false;
921 bool argPoly = false;
922 bool argUsgn = false;
923 bool argScalar = false;
925 char argType = ClassifyType(typestr, argQuad, argPoly, argUsgn);
926 argType = ModType(proto[i], argType, argQuad, argPoly, argUsgn, argScalar,
929 // Handle multiple-vector values specially, emitting each subvector as an
930 // argument to the __builtin.
931 if (proto[i] >= '2' && proto[i] <= '4') {
932 // Check if an explicit cast is needed.
933 if (argType != 'c' || argPoly || argUsgn)
934 args = (argQuad ? "(int8x16_t)" : "(int8x8_t)") + args;
936 for (unsigned vi = 0, ve = proto[i] - '0'; vi != ve; ++vi) {
937 s += args + ".val[" + utostr(vi) + "]";
947 if (splat && (i + 1) == e)
948 args = Duplicate(GetNumElements(typestr, argQuad), typestr, args);
950 // Check if an explicit cast is needed.
951 if ((splat || !argScalar) &&
952 ((ck == ClassB && argType != 'c') || argPoly || argUsgn)) {
953 std::string argTypeStr = "c";
955 argTypeStr = argType;
957 argTypeStr = "Q" + argTypeStr;
958 args = "(" + TypeString('d', argTypeStr) + ")" + args;
966 // Extra constant integer to hold type class enum for this function, e.g. s8
968 s += ", " + utostr(GetNeonEnum(proto, typestr));
972 if (proto[0] != 'v' && sret) {
981 static std::string GenBuiltinDef(const std::string &name,
982 const std::string &proto,
983 StringRef typestr, ClassKind ck) {
984 std::string s("BUILTIN(__builtin_neon_");
986 // If all types are the same size, bitcasting the args will take care
987 // of arg checking. The actual signedness etc. will be taken care of with
989 if (proto.find('s') == std::string::npos)
992 s += MangleName(name, typestr, ck);
995 for (unsigned i = 0, e = proto.size(); i != e; ++i)
996 s += BuiltinTypeString(proto[i], typestr, ck, i == 0);
998 // Extra constant integer to hold type class enum for this function, e.g. s8
1006 static std::string GenIntrinsic(const std::string &name,
1007 const std::string &proto,
1008 StringRef outTypeStr, StringRef inTypeStr,
1009 OpKind kind, ClassKind classKind) {
1010 assert(!proto.empty() && "");
1011 bool define = proto.find('i') != std::string::npos;
1014 // static always inline + return type
1018 s += "__ai " + TypeString(proto[0], outTypeStr) + " ";
1020 // Function name with type suffix
1021 std::string mangledName = MangleName(name, outTypeStr, ClassS);
1022 if (outTypeStr != inTypeStr) {
1023 // If the input type is different (e.g., for vreinterpret), append a suffix
1024 // for the input type. String off a "Q" (quad) prefix so that MangleName
1025 // does not insert another "q" in the name.
1026 unsigned typeStrOff = (inTypeStr[0] == 'Q' ? 1 : 0);
1027 StringRef inTypeNoQuad = inTypeStr.substr(typeStrOff);
1028 mangledName = MangleName(mangledName, inTypeNoQuad, ClassS);
1032 // Function arguments
1033 s += GenArgs(proto, inTypeStr);
1037 s += " __extension__ ({ \\\n ";
1038 s += GenMacroLocals(proto, inTypeStr);
1044 s += GenOpString(kind, proto, outTypeStr);
1046 s += GenBuiltin(name, proto, outTypeStr, classKind);
1055 /// run - Read the records in arm_neon.td and output arm_neon.h. arm_neon.h
1056 /// is comprised of type definitions and function declarations.
1057 void NeonEmitter::run(raw_ostream &OS) {
1059 "/*===---- arm_neon.h - ARM Neon intrinsics ------------------------------"
1062 " * Permission is hereby granted, free of charge, to any person obtaining "
1064 " * of this software and associated documentation files (the \"Software\"),"
1066 " * in the Software without restriction, including without limitation the "
1068 " * to use, copy, modify, merge, publish, distribute, sublicense, "
1070 " * copies of the Software, and to permit persons to whom the Software is\n"
1071 " * furnished to do so, subject to the following conditions:\n"
1073 " * The above copyright notice and this permission notice shall be "
1075 " * all copies or substantial portions of the Software.\n"
1077 " * THE SOFTWARE IS PROVIDED \"AS IS\", WITHOUT WARRANTY OF ANY KIND, "
1079 " * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF "
1080 "MERCHANTABILITY,\n"
1081 " * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT "
1083 " * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR "
1085 " * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, "
1087 " * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER "
1089 " * THE SOFTWARE.\n"
1091 " *===--------------------------------------------------------------------"
1095 OS << "#ifndef __ARM_NEON_H\n";
1096 OS << "#define __ARM_NEON_H\n\n";
1098 OS << "#ifndef __ARM_NEON__\n";
1099 OS << "#error \"NEON support not enabled\"\n";
1102 OS << "#include <stdint.h>\n\n";
1104 // Emit NEON-specific scalar typedefs.
1105 OS << "typedef float float32_t;\n";
1106 OS << "typedef int8_t poly8_t;\n";
1107 OS << "typedef int16_t poly16_t;\n";
1108 OS << "typedef uint16_t float16_t;\n";
1110 // Emit Neon vector typedefs.
1111 std::string TypedefTypes("cQcsQsiQilQlUcQUcUsQUsUiQUiUlQUlhQhfQfPcQPcPsQPs");
1112 SmallVector<StringRef, 24> TDTypeVec;
1113 ParseTypes(0, TypedefTypes, TDTypeVec);
1115 // Emit vector typedefs.
1116 for (unsigned i = 0, e = TDTypeVec.size(); i != e; ++i) {
1117 bool dummy, quad = false, poly = false;
1118 (void) ClassifyType(TDTypeVec[i], quad, poly, dummy);
1120 OS << "typedef __attribute__((neon_polyvector_type(";
1122 OS << "typedef __attribute__((neon_vector_type(";
1124 unsigned nElts = GetNumElements(TDTypeVec[i], quad);
1125 OS << utostr(nElts) << "))) ";
1129 OS << TypeString('s', TDTypeVec[i]);
1130 OS << " " << TypeString('d', TDTypeVec[i]) << ";\n";
1134 // Emit struct typedefs.
1135 for (unsigned vi = 2; vi != 5; ++vi) {
1136 for (unsigned i = 0, e = TDTypeVec.size(); i != e; ++i) {
1137 std::string ts = TypeString('d', TDTypeVec[i]);
1138 std::string vs = TypeString('0' + vi, TDTypeVec[i]);
1139 OS << "typedef struct " << vs << " {\n";
1140 OS << " " << ts << " val";
1141 OS << "[" << utostr(vi) << "]";
1143 OS << vs << ";\n\n";
1147 OS << "#define __ai static __attribute__((__always_inline__))\n\n";
1149 std::vector<Record*> RV = Records.getAllDerivedDefinitions("Inst");
1151 // Emit vmovl and vabd intrinsics first so they can be used by other
1152 // intrinsics. (Some of the saturating multiply instructions are also
1153 // used to implement the corresponding "_lane" variants, but tablegen
1154 // sorts the records into alphabetical order so that the "_lane" variants
1155 // come after the intrinsics they use.)
1156 emitIntrinsic(OS, Records.getDef("VMOVL"));
1157 emitIntrinsic(OS, Records.getDef("VABD"));
1159 for (unsigned i = 0, e = RV.size(); i != e; ++i) {
1161 if (R->getName() != "VMOVL" && R->getName() != "VABD")
1162 emitIntrinsic(OS, R);
1165 OS << "#undef __ai\n\n";
1166 OS << "#endif /* __ARM_NEON_H */\n";
1169 /// emitIntrinsic - Write out the arm_neon.h header file definitions for the
1170 /// intrinsics specified by record R.
1171 void NeonEmitter::emitIntrinsic(raw_ostream &OS, Record *R) {
1172 std::string name = R->getValueAsString("Name");
1173 std::string Proto = R->getValueAsString("Prototype");
1174 std::string Types = R->getValueAsString("Types");
1176 SmallVector<StringRef, 16> TypeVec;
1177 ParseTypes(R, Types, TypeVec);
1179 OpKind kind = OpMap[R->getValueAsDef("Operand")->getName()];
1181 ClassKind classKind = ClassNone;
1182 if (R->getSuperClasses().size() >= 2)
1183 classKind = ClassMap[R->getSuperClasses()[1]];
1184 if (classKind == ClassNone && kind == OpNone)
1185 throw TGError(R->getLoc(), "Builtin has no class kind");
1187 for (unsigned ti = 0, te = TypeVec.size(); ti != te; ++ti) {
1188 if (kind == OpReinterpret) {
1189 bool outQuad = false;
1191 (void)ClassifyType(TypeVec[ti], outQuad, dummy, dummy);
1192 for (unsigned srcti = 0, srcte = TypeVec.size();
1193 srcti != srcte; ++srcti) {
1194 bool inQuad = false;
1195 (void)ClassifyType(TypeVec[srcti], inQuad, dummy, dummy);
1196 if (srcti == ti || inQuad != outQuad)
1198 OS << GenIntrinsic(name, Proto, TypeVec[ti], TypeVec[srcti],
1202 OS << GenIntrinsic(name, Proto, TypeVec[ti], TypeVec[ti],
1209 static unsigned RangeFromType(const char mod, StringRef typestr) {
1210 // base type to get the type string for.
1211 bool quad = false, dummy = false;
1212 char type = ClassifyType(typestr, quad, dummy, dummy);
1213 type = ModType(mod, type, quad, dummy, dummy, dummy, dummy, dummy);
1217 return (8 << (int)quad) - 1;
1220 return (4 << (int)quad) - 1;
1223 return (2 << (int)quad) - 1;
1225 return (1 << (int)quad) - 1;
1227 throw "unhandled type!";
1230 assert(0 && "unreachable");
1234 /// runHeader - Emit a file with sections defining:
1235 /// 1. the NEON section of BuiltinsARM.def.
1236 /// 2. the SemaChecking code for the type overload checking.
1237 /// 3. the SemaChecking code for validation of intrinsic immedate arguments.
1238 void NeonEmitter::runHeader(raw_ostream &OS) {
1239 std::vector<Record*> RV = Records.getAllDerivedDefinitions("Inst");
1241 StringMap<OpKind> EmittedMap;
1243 // Generate BuiltinsARM.def for NEON
1244 OS << "#ifdef GET_NEON_BUILTINS\n";
1245 for (unsigned i = 0, e = RV.size(); i != e; ++i) {
1247 OpKind k = OpMap[R->getValueAsDef("Operand")->getName()];
1251 std::string Proto = R->getValueAsString("Prototype");
1253 // Functions with 'a' (the splat code) in the type prototype should not get
1254 // their own builtin as they use the non-splat variant.
1255 if (Proto.find('a') != std::string::npos)
1258 std::string Types = R->getValueAsString("Types");
1259 SmallVector<StringRef, 16> TypeVec;
1260 ParseTypes(R, Types, TypeVec);
1262 if (R->getSuperClasses().size() < 2)
1263 throw TGError(R->getLoc(), "Builtin has no class kind");
1265 std::string name = R->getValueAsString("Name");
1266 ClassKind ck = ClassMap[R->getSuperClasses()[1]];
1268 for (unsigned ti = 0, te = TypeVec.size(); ti != te; ++ti) {
1269 // Generate the BuiltinsARM.def declaration for this builtin, ensuring
1270 // that each unique BUILTIN() macro appears only once in the output
1272 std::string bd = GenBuiltinDef(name, Proto, TypeVec[ti], ck);
1273 if (EmittedMap.count(bd))
1276 EmittedMap[bd] = OpNone;
1282 // Generate the overloaded type checking code for SemaChecking.cpp
1283 OS << "#ifdef GET_NEON_OVERLOAD_CHECK\n";
1284 for (unsigned i = 0, e = RV.size(); i != e; ++i) {
1286 OpKind k = OpMap[R->getValueAsDef("Operand")->getName()];
1290 std::string Proto = R->getValueAsString("Prototype");
1291 std::string Types = R->getValueAsString("Types");
1292 std::string name = R->getValueAsString("Name");
1294 // Functions with 'a' (the splat code) in the type prototype should not get
1295 // their own builtin as they use the non-splat variant.
1296 if (Proto.find('a') != std::string::npos)
1299 // Functions which have a scalar argument cannot be overloaded, no need to
1300 // check them if we are emitting the type checking code.
1301 if (Proto.find('s') != std::string::npos)
1304 SmallVector<StringRef, 16> TypeVec;
1305 ParseTypes(R, Types, TypeVec);
1307 if (R->getSuperClasses().size() < 2)
1308 throw TGError(R->getLoc(), "Builtin has no class kind");
1310 int si = -1, qi = -1;
1311 unsigned mask = 0, qmask = 0;
1312 for (unsigned ti = 0, te = TypeVec.size(); ti != te; ++ti) {
1313 // Generate the switch case(s) for this builtin for the type validation.
1314 bool quad = false, poly = false, usgn = false;
1315 (void) ClassifyType(TypeVec[ti], quad, poly, usgn);
1319 qmask |= 1 << GetNeonEnum(Proto, TypeVec[ti]);
1322 mask |= 1 << GetNeonEnum(Proto, TypeVec[ti]);
1326 OS << "case ARM::BI__builtin_neon_"
1327 << MangleName(name, TypeVec[si], ClassB)
1328 << ": mask = " << "0x" << utohexstr(mask) << "; break;\n";
1330 OS << "case ARM::BI__builtin_neon_"
1331 << MangleName(name, TypeVec[qi], ClassB)
1332 << ": mask = " << "0x" << utohexstr(qmask) << "; break;\n";
1336 // Generate the intrinsic range checking code for shift/lane immediates.
1337 OS << "#ifdef GET_NEON_IMMEDIATE_CHECK\n";
1338 for (unsigned i = 0, e = RV.size(); i != e; ++i) {
1341 OpKind k = OpMap[R->getValueAsDef("Operand")->getName()];
1345 std::string name = R->getValueAsString("Name");
1346 std::string Proto = R->getValueAsString("Prototype");
1347 std::string Types = R->getValueAsString("Types");
1349 // Functions with 'a' (the splat code) in the type prototype should not get
1350 // their own builtin as they use the non-splat variant.
1351 if (Proto.find('a') != std::string::npos)
1354 // Functions which do not have an immediate do not need to have range
1355 // checking code emitted.
1356 size_t immPos = Proto.find('i');
1357 if (immPos == std::string::npos)
1360 SmallVector<StringRef, 16> TypeVec;
1361 ParseTypes(R, Types, TypeVec);
1363 if (R->getSuperClasses().size() < 2)
1364 throw TGError(R->getLoc(), "Builtin has no class kind");
1366 ClassKind ck = ClassMap[R->getSuperClasses()[1]];
1368 for (unsigned ti = 0, te = TypeVec.size(); ti != te; ++ti) {
1369 std::string namestr, shiftstr, rangestr;
1371 // Builtins which are overloaded by type will need to have their upper
1372 // bound computed at Sema time based on the type constant.
1373 if (Proto.find('s') == std::string::npos) {
1375 if (R->getValueAsBit("isShift")) {
1376 shiftstr = ", true";
1378 // Right shifts have an 'r' in the name, left shifts do not.
1379 if (name.find('r') != std::string::npos)
1380 rangestr = "l = 1; ";
1382 rangestr += "u = RFT(TV" + shiftstr + ")";
1384 // The immediate generally refers to a lane in the preceding argument.
1385 assert(immPos > 0 && "unexpected immediate operand");
1386 rangestr = "u = " + utostr(RangeFromType(Proto[immPos-1], TypeVec[ti]));
1388 // Make sure cases appear only once by uniquing them in a string map.
1389 namestr = MangleName(name, TypeVec[ti], ck);
1390 if (EmittedMap.count(namestr))
1392 EmittedMap[namestr] = OpNone;
1394 // Calculate the index of the immediate that should be range checked.
1395 unsigned immidx = 0;
1397 // Builtins that return a struct of multiple vectors have an extra
1398 // leading arg for the struct return.
1399 if (Proto[0] >= '2' && Proto[0] <= '4')
1402 // Add one to the index for each argument until we reach the immediate
1403 // to be checked. Structs of vectors are passed as multiple arguments.
1404 for (unsigned ii = 1, ie = Proto.size(); ii != ie; ++ii) {
1405 switch (Proto[ii]) {
1406 default: immidx += 1; break;
1407 case '2': immidx += 2; break;
1408 case '3': immidx += 3; break;
1409 case '4': immidx += 4; break;
1410 case 'i': ie = ii + 1; break;
1413 OS << "case ARM::BI__builtin_neon_" << MangleName(name, TypeVec[ti], ck)
1414 << ": i = " << immidx << "; " << rangestr << "; break;\n";
1420 /// GenTest - Write out a test for the intrinsic specified by the name and
1421 /// type strings, including the embedded patterns for FileCheck to match.
1422 static std::string GenTest(const std::string &name,
1423 const std::string &proto,
1424 StringRef outTypeStr, StringRef inTypeStr,
1426 assert(!proto.empty() && "");
1429 // Function name with type suffix
1430 std::string mangledName = MangleName(name, outTypeStr, ClassS);
1431 if (outTypeStr != inTypeStr) {
1432 // If the input type is different (e.g., for vreinterpret), append a suffix
1433 // for the input type. String off a "Q" (quad) prefix so that MangleName
1434 // does not insert another "q" in the name.
1435 unsigned typeStrOff = (inTypeStr[0] == 'Q' ? 1 : 0);
1436 StringRef inTypeNoQuad = inTypeStr.substr(typeStrOff);
1437 mangledName = MangleName(mangledName, inTypeNoQuad, ClassS);
1440 // Emit the FileCheck patterns.
1441 s += "// CHECK: test_" + mangledName + "\n";
1442 // s += "// CHECK: \n"; // FIXME: + expected instruction opcode.
1444 // Emit the start of the test function.
1445 s += TypeString(proto[0], outTypeStr) + " test_" + mangledName + "(";
1448 for (unsigned i = 1, e = proto.size(); i != e; ++i, ++arg) {
1449 // Do not create arguments for values that must be immediate constants.
1450 if (proto[i] == 'i')
1452 s += comma + TypeString(proto[i], inTypeStr) + " ";
1458 if (proto[0] != 'v')
1460 s += mangledName + "(";
1462 for (unsigned i = 1, e = proto.size(); i != e; ++i, ++arg) {
1463 if (proto[i] == 'i') {
1464 // For immediate operands, test the maximum value.
1468 // The immediate generally refers to a lane in the preceding argument.
1469 s += utostr(RangeFromType(proto[i-1], inTypeStr));
1480 /// runTests - Write out a complete set of tests for all of the Neon
1482 void NeonEmitter::runTests(raw_ostream &OS) {
1484 "// RUN: %clang_cc1 -triple thumbv7-apple-darwin \\\n"
1485 "// RUN: -target-cpu cortex-a9 -ffreestanding -S -o - %s | FileCheck %s\n"
1487 "#include <arm_neon.h>\n"
1490 std::vector<Record*> RV = Records.getAllDerivedDefinitions("Inst");
1491 for (unsigned i = 0, e = RV.size(); i != e; ++i) {
1493 std::string name = R->getValueAsString("Name");
1494 std::string Proto = R->getValueAsString("Prototype");
1495 std::string Types = R->getValueAsString("Types");
1496 bool isShift = R->getValueAsBit("isShift");
1498 SmallVector<StringRef, 16> TypeVec;
1499 ParseTypes(R, Types, TypeVec);
1501 OpKind kind = OpMap[R->getValueAsDef("Operand")->getName()];
1502 for (unsigned ti = 0, te = TypeVec.size(); ti != te; ++ti) {
1503 if (kind == OpReinterpret) {
1504 bool outQuad = false;
1506 (void)ClassifyType(TypeVec[ti], outQuad, dummy, dummy);
1507 for (unsigned srcti = 0, srcte = TypeVec.size();
1508 srcti != srcte; ++srcti) {
1509 bool inQuad = false;
1510 (void)ClassifyType(TypeVec[srcti], inQuad, dummy, dummy);
1511 if (srcti == ti || inQuad != outQuad)
1513 OS << GenTest(name, Proto, TypeVec[ti], TypeVec[srcti], isShift);
1516 OS << GenTest(name, Proto, TypeVec[ti], TypeVec[ti], isShift);