// a declaration and definition of each function specified by the ARM NEON
// compiler interface. See ARM document DUI0348B.
//
+// Each NEON instruction is implemented in terms of 1 or more functions which
+// are suffixed with the element type of the input vectors. Functions may be
+// implemented in terms of generic vector operations such as +, *, -, etc. or
+// by calling a __builtin_-prefixed function which will be handled by clang's
+// CodeGen library.
+//
+// Additional validation code can be generated by this file when runHeader() is
+// called, rather than the normal run() entry point.
+//
//===----------------------------------------------------------------------===//
#include "NeonEmitter.h"
using namespace llvm;
+/// ParseTypes - break down a string such as "fQf" into a vector of StringRefs,
+/// which each StringRef representing a single type declared in the string.
+/// for "fQf" we would end up with 2 StringRefs, "f", and "Qf", representing
+/// 2xfloat and 4xfloat respectively.
static void ParseTypes(Record *r, std::string &s,
SmallVectorImpl<StringRef> &TV) {
const char *data = s.data();
}
}
+/// Widen - Convert a type code into the next wider type. char -> short,
+/// short -> int, etc.
static char Widen(const char t) {
switch (t) {
case 'c':
return '\0';
}
+/// Narrow - Convert a type code into the next smaller type. short -> char,
+/// float -> half float, etc.
static char Narrow(const char t) {
switch (t) {
case 's':
return '\0';
}
+/// For a particular StringRef, return the base type code, and whether it has
+/// the quad-vector, polynomial, or unsigned modifiers set.
static char ClassifyType(StringRef ty, bool &quad, bool &poly, bool &usgn) {
unsigned off = 0;
return ty[off];
}
+/// ModType - Transform a type code and its modifiers based on a mod code. The
+/// mod code definitions may be found at the top of arm_neon.td.
static char ModType(const char mod, char type, bool &quad, bool &poly,
bool &usgn, bool &scal, bool &cnst, bool &pntr) {
switch (mod) {
type = 'f';
usgn = false;
break;
+ case 'g':
+ quad = false;
+ break;
case 'w':
type = Widen(type);
quad = true;
usgn = true;
break;
case 's':
+ case 'a':
scal = true;
break;
case 'k':
case 'c':
cnst = true;
case 'p':
- usgn = false;
- poly = false;
pntr = true;
scal = true;
break;
return type;
}
-static std::string TypeString(const char mod, StringRef typestr,
- bool ret = false) {
+/// TypeString - for a modifier and type, generate the name of the typedef for
+/// that type. QUc -> uint8x8_t.
+static std::string TypeString(const char mod, StringRef typestr) {
bool quad = false;
bool poly = false;
bool usgn = false;
SmallString<128> s;
- if (ret)
- s += "__neon_";
-
if (usgn)
s.push_back('u');
return s.str();
}
+/// BuiltinTypeString - for a modifier and type, generate the clang
+/// BuiltinsARM.def prototype code for the function. See the top of clang's
+/// Builtins.def for a description of the type strings.
static std::string BuiltinTypeString(const char mod, StringRef typestr,
ClassKind ck, bool ret) {
bool quad = false;
// Based on the modifying character, change the type and width if necessary.
type = ModType(mod, type, quad, poly, usgn, scal, cnst, pntr);
- if (pntr)
+ if (pntr) {
+ usgn = false;
+ poly = false;
type = 'v';
-
+ }
if (type == 'h') {
type = 's';
usgn = true;
}
// Since the return value must be one type, return a vector type of the
- // appropriate width which we will bitcast.
+ // appropriate width which we will bitcast. An exception is made for
+ // returning structs of 2, 3, or 4 vectors which are returned in a sret-like
+ // fashion, storing them to a pointer arg.
if (ret) {
- if (mod == '2')
- return quad ? "V32c" : "V16c";
- if (mod == '3')
- return quad ? "V48c" : "V24c";
- if (mod == '4')
- return quad ? "V64c" : "V32c";
- if (mod == 'f')
+ if (mod == '2' || mod == '3' || mod == '4')
+ return "vv*";
+ if (mod == 'f' || (ck != ClassB && type == 'f'))
return quad ? "V4f" : "V2f";
- if (mod == 'x' || mod == 'u')
+ if (ck != ClassB && type == 's')
+ return quad ? "V8s" : "V4s";
+ if (ck != ClassB && type == 'i')
return quad ? "V4i" : "V2i";
+ if (ck != ClassB && type == 'l')
+ return quad ? "V2LLi" : "V1LLi";
return quad ? "V16c" : "V8c";
}
return quad ? "V16cV16cV16c" : "V8cV8cV8c";
if (mod == '4')
return quad ? "V16cV16cV16cV16c" : "V8cV8cV8cV8c";
- if (mod == 'f')
- return quad ? "V4f" : "V2f";
+ if (mod == 'f' || (ck != ClassB && type == 'f'))
+ return quad ? "V4f" : "V2f";
+ if (ck != ClassB && type == 's')
+ return quad ? "V8s" : "V4s";
+ if (ck != ClassB && type == 'i')
+ return quad ? "V4i" : "V2i";
+ if (ck != ClassB && type == 'l')
+ return quad ? "V2LLi" : "V1LLi";
+
return quad ? "V16c" : "V8c";
}
-// Turn "vst2_lane" into "vst2q_lane_f32", etc.
+/// MangleName - Append a type or width suffix to a base neon function name,
+/// and insert a 'q' in the appropriate location if the operation works on
+/// 128b rather than 64b. E.g. turn "vst2_lane" into "vst2q_lane_f32", etc.
static std::string MangleName(const std::string &name, StringRef typestr,
ClassKind ck) {
if (name == "vcvt_f32_f16")
// Generate the string "(argtype a, argtype b, ...)"
static std::string GenArgs(const std::string &proto, StringRef typestr) {
+ bool define = proto.find('i') != std::string::npos;
char arg = 'a';
std::string s;
s += "(";
for (unsigned i = 1, e = proto.size(); i != e; ++i, ++arg) {
- s += TypeString(proto[i], typestr);
- s.push_back(' ');
+ if (!define) {
+ s += TypeString(proto[i], typestr);
+ s.push_back(' ');
+ }
s.push_back(arg);
if ((i + 1) < e)
s += ", ";
return s;
}
-// Generate the definition for this intrinsic, e.g. "a + b" for OpAdd.
-// If structTypes is true, the NEON types are structs of vector types rather
-// than vector types, and the call becomes "a.val + b.val"
-static std::string GenOpString(OpKind op, const std::string &proto,
- StringRef typestr, bool structTypes = true) {
- std::string ts = TypeString(proto[0], typestr);
- std::string s = ts + " r; r";
+static std::string Duplicate(unsigned nElts, StringRef typestr,
+ const std::string &a) {
+ std::string s;
+
+ s = "(" + TypeString('d', typestr) + "){ ";
+ for (unsigned i = 0; i != nElts; ++i) {
+ s += a;
+ if ((i + 1) < nElts)
+ s += ", ";
+ }
+ s += " }";
+
+ return s;
+}
- bool quad, dummy;
+static unsigned GetNumElements(StringRef typestr, bool &quad) {
+ quad = false;
+ bool dummy = false;
char type = ClassifyType(typestr, quad, dummy, dummy);
unsigned nElts = 0;
switch (type) {
- case 'c': nElts = 8; break;
- case 's': nElts = 4; break;
- case 'i': nElts = 2; break;
- case 'l': nElts = 1; break;
- case 'h': nElts = 4; break;
- case 'f': nElts = 2; break;
+ case 'c': nElts = 8; break;
+ case 's': nElts = 4; break;
+ case 'i': nElts = 2; break;
+ case 'l': nElts = 1; break;
+ case 'h': nElts = 4; break;
+ case 'f': nElts = 2; break;
+ default:
+ throw "unhandled type!";
+ break;
}
- nElts <<= quad;
+ if (quad) nElts <<= 1;
+ return nElts;
+}
+
+// Generate the definition for this intrinsic, e.g. "a + b" for OpAdd.
+static std::string GenOpString(OpKind op, const std::string &proto,
+ StringRef typestr) {
+ bool quad;
+ unsigned nElts = GetNumElements(typestr, quad);
- if (structTypes)
- s += ".val";
+ std::string ts = TypeString(proto[0], typestr);
+ std::string s;
+ if (op == OpHi || op == OpLo) {
+ s = "union { " + ts + " r; double d; } u; u.d";
+ } else {
+ s = ts + " r; r";
+ }
s += " = ";
- std::string a, b, c;
- if (proto.size() > 1)
- a = (structTypes && proto[1] != 'l' && proto[1] != 's') ? "a.val" : "a";
- b = structTypes ? "b.val" : "b";
- c = structTypes ? "c.val" : "c";
-
switch(op) {
case OpAdd:
- s += a + " + " + b;
+ s += "a + b";
break;
case OpSub:
- s += a + " - " + b;
+ s += "a - b";
+ break;
+ case OpMulN:
+ s += "a * " + Duplicate(nElts, typestr, "b");
break;
case OpMul:
- s += a + " * " + b;
+ s += "a * b";
+ break;
+ case OpMlaN:
+ s += "a + (b * " + Duplicate(nElts, typestr, "c") + ")";
break;
case OpMla:
- s += a + " + ( " + b + " * " + c + " )";
+ s += "a + (b * c)";
+ break;
+ case OpMlsN:
+ s += "a - (b * " + Duplicate(nElts, typestr, "c") + ")";
break;
case OpMls:
- s += a + " - ( " + b + " * " + c + " )";
+ s += "a - (b * c)";
break;
case OpEq:
- s += "(__neon_" + ts + ")(" + a + " == " + b + ")";
+ s += "(" + ts + ")(a == b)";
break;
case OpGe:
- s += "(__neon_" + ts + ")(" + a + " >= " + b + ")";
+ s += "(" + ts + ")(a >= b)";
break;
case OpLe:
- s += "(__neon_" + ts + ")(" + a + " <= " + b + ")";
+ s += "(" + ts + ")(a <= b)";
break;
case OpGt:
- s += "(__neon_" + ts + ")(" + a + " > " + b + ")";
+ s += "(" + ts + ")(a > b)";
break;
case OpLt:
- s += "(__neon_" + ts + ")(" + a + " < " + b + ")";
+ s += "(" + ts + ")(a < b)";
break;
case OpNeg:
- s += " -" + a;
+ s += " -a";
break;
case OpNot:
- s += " ~" + a;
+ s += " ~a";
break;
case OpAnd:
- s += a + " & " + b;
+ s += "a & b";
break;
case OpOr:
- s += a + " | " + b;
+ s += "a | b";
break;
case OpXor:
- s += a + " ^ " + b;
+ s += "a ^ b";
break;
case OpAndNot:
- s += a + " & ~" + b;
+ s += "a & ~b";
break;
case OpOrNot:
- s += a + " | ~" + b;
+ s += "a | ~b";
break;
case OpCast:
- s += "(__neon_" + ts + ")" + a;
+ s += "(" + ts + ")a";
break;
case OpConcat:
- s += "__builtin_shufflevector((__neon_int64x1_t)" + a;
- s += ", (__neon_int64x1_t)" + b + ", 0, 1)";
+ s += "__builtin_shufflevector((int64x1_t)a";
+ s += ", (int64x1_t)b, 0, 1)";
+ break;
+ case OpHi:
+ s += "(((float64x2_t)a)[1])";
+ break;
+ case OpLo:
+ s += "(((float64x2_t)a)[0])";
break;
case OpDup:
- s += "(__neon_" + ts + "){ ";
- for (unsigned i = 0; i != nElts; ++i) {
- s += a;
- if ((i + 1) < nElts)
- s += ", ";
- }
- s += " }";
+ s += Duplicate(nElts, typestr, "a");
+ break;
+ case OpSelect:
+ // ((0 & 1) | (~0 & 2))
+ ts = TypeString(proto[1], typestr);
+ s += "(a & (" + ts + ")b) | ";
+ s += "(~a & (" + ts + ")c)";
+ break;
+ case OpRev16:
+ s += "__builtin_shufflevector(a, a";
+ for (unsigned i = 2; i <= nElts; i += 2)
+ for (unsigned j = 0; j != 2; ++j)
+ s += ", " + utostr(i - j - 1);
+ s += ")";
break;
+ case OpRev32: {
+ unsigned WordElts = nElts >> (1 + (int)quad);
+ s += "__builtin_shufflevector(a, a";
+ for (unsigned i = WordElts; i <= nElts; i += WordElts)
+ for (unsigned j = 0; j != WordElts; ++j)
+ s += ", " + utostr(i - j - 1);
+ s += ")";
+ break;
+ }
+ case OpRev64: {
+ unsigned DblWordElts = nElts >> (int)quad;
+ s += "__builtin_shufflevector(a, a";
+ for (unsigned i = DblWordElts; i <= nElts; i += DblWordElts)
+ for (unsigned j = 0; j != DblWordElts; ++j)
+ s += ", " + utostr(i - j - 1);
+ s += ")";
+ break;
+ }
default:
throw "unknown OpKind!";
break;
}
- s += "; return r;";
+ if (op == OpHi || op == OpLo)
+ s += "; return u.r;";
+ else
+ s += "; return r;";
return s;
}
bool cnst = false;
bool pntr = false;
- // base type to get the type string for.
+ // Base type to get the type string for.
char type = ClassifyType(typestr, quad, poly, usgn);
// Based on the modifying character, change the type and width if necessary.
type = ModType(mod, type, quad, poly, usgn, scal, cnst, pntr);
-
+
if (usgn)
ret |= 0x08;
- if (quad)
+ if (quad && proto[1] != 'g')
ret |= 0x10;
switch (type) {
}
// Generate the definition for this intrinsic, e.g. __builtin_neon_cls(a)
-// If structTypes is true, the NEON types are structs of vector types rather
-// than vector types, and the call becomes __builtin_neon_cls(a.val)
static std::string GenBuiltin(const std::string &name, const std::string &proto,
- StringRef typestr, ClassKind ck,
- bool structTypes = true) {
+ StringRef typestr, ClassKind ck) {
+ bool quad;
+ unsigned nElts = GetNumElements(typestr, quad);
char arg = 'a';
std::string s;
- bool unioning = (proto[0] == '2' || proto[0] == '3' || proto[0] == '4');
+ // If this builtin returns a struct 2, 3, or 4 vectors, pass it as an implicit
+ // sret-like argument.
+ bool sret = (proto[0] == '2' || proto[0] == '3' || proto[0] == '4');
+
+ // If this builtin takes an immediate argument, we need to #define it rather
+ // than use a standard declaration, so that SemaChecking can range check
+ // the immediate passed by the user.
+ bool define = proto.find('i') != std::string::npos;
// If all types are the same size, bitcasting the args will take care
// of arg checking. The actual signedness etc. will be taken care of with
ck = ClassB;
if (proto[0] != 'v') {
- if (unioning) {
- s += "union { ";
- s += TypeString(proto[0], typestr, true) + " val; ";
- s += TypeString(proto[0], typestr, false) + " s; ";
- s += "} r;";
+ std::string ts = TypeString(proto[0], typestr);
+
+ if (define) {
+ if (sret)
+ s += "({ " + ts + " r; ";
+ else if (proto[0] != 's')
+ s += "(" + ts + ")";
+ } else if (sret) {
+ s += ts + " r; ";
} else {
- s += TypeString(proto[0], typestr);
+ s += ts + " r; r = ";
}
-
- s += " r; r";
- if (structTypes && proto[0] != 's' && proto[0] != 'i' && proto[0] != 'l')
- s += ".val";
-
- s += " = ";
- }
+ }
+
+ bool splat = proto.find('a') != std::string::npos;
s += "__builtin_neon_";
- s += MangleName(name, typestr, ck);
+ if (splat) {
+ std::string vname(name, 0, name.size()-2);
+ s += MangleName(vname, typestr, ck);
+ } else {
+ s += MangleName(name, typestr, ck);
+ }
s += "(";
+
+ // Pass the address of the return variable as the first argument to sret-like
+ // builtins.
+ if (sret)
+ s += "&r, ";
for (unsigned i = 1, e = proto.size(); i != e; ++i, ++arg) {
+ std::string args = std::string(&arg, 1);
+ if (define)
+ args = "(" + args + ")";
+
// Handle multiple-vector values specially, emitting each subvector as an
// argument to the __builtin.
- if (structTypes && (proto[i] == '2' || proto[i] == '3' || proto[i] == '4')){
+ if (proto[i] == '2' || proto[i] == '3' || proto[i] == '4') {
for (unsigned vi = 0, ve = proto[i] - '0'; vi != ve; ++vi) {
- s.push_back(arg);
- s += ".val[" + utostr(vi) + "]";
+ s += args + ".val[" + utostr(vi) + "]";
if ((vi + 1) < ve)
s += ", ";
}
continue;
}
- s.push_back(arg);
-
- if (structTypes && proto[i] != 's' && proto[i] != 'i' && proto[i] != 'l' &&
- proto[i] != 'p' && proto[i] != 'c') {
- s += ".val";
- }
+ if (splat && (i + 1) == e)
+ s += Duplicate(nElts, typestr, args);
+ else
+ s += args;
if ((i + 1) < e)
s += ", ";
}
if (ck == ClassB)
s += ", " + utostr(GetNeonEnum(proto, typestr));
- s += ");";
+ if (define)
+ s += ")";
+ else
+ s += ");";
if (proto[0] != 'v') {
- if (unioning)
- s += " return r.s;";
- else
+ if (define) {
+ if (sret)
+ s += "; r; })";
+ } else {
s += " return r;";
+ }
}
return s;
}
return s;
}
+/// run - Read the records in arm_neon.td and output arm_neon.h. arm_neon.h
+/// is comprised of type definitions and function declarations.
void NeonEmitter::run(raw_ostream &OS) {
EmitSourceFileHeader("ARM NEON Header", OS);
OS << "#include <stdint.h>\n\n";
// Emit NEON-specific scalar typedefs.
- // FIXME: probably need to do something better for polynomial types.
- // FIXME: is this the correct thing to do for float16?
OS << "typedef float float32_t;\n";
- OS << "typedef uint8_t poly8_t;\n";
- OS << "typedef uint16_t poly16_t;\n";
+ OS << "typedef int8_t poly8_t;\n";
+ OS << "typedef int16_t poly16_t;\n";
OS << "typedef uint16_t float16_t;\n";
// Emit Neon vector typedefs.
ParseTypes(0, TypedefTypes, TDTypeVec);
// Emit vector typedefs.
- for (unsigned v = 1; v != 5; ++v) {
- for (unsigned i = 0, e = TDTypeVec.size(); i != e; ++i) {
- bool dummy, quad = false;
- (void) ClassifyType(TDTypeVec[i], quad, dummy, dummy);
- OS << "typedef __attribute__(( __vector_size__(";
-
- OS << utostr(8*v*(quad ? 2 : 1)) << ") )) ";
- if (!quad)
- OS << " ";
+ for (unsigned i = 0, e = TDTypeVec.size(); i != e; ++i) {
+ bool dummy, quad = false, poly = false;
+ (void) ClassifyType(TDTypeVec[i], quad, poly, dummy);
+ if (poly)
+ OS << "typedef __attribute__((neon_polyvector_type(";
+ else
+ OS << "typedef __attribute__((neon_vector_type(";
- OS << TypeString('s', TDTypeVec[i]);
- OS << " __neon_";
+ unsigned nElts = GetNumElements(TDTypeVec[i], quad);
+ OS << utostr(nElts) << "))) ";
+ if (nElts < 10)
+ OS << " ";
- char t = (v == 1) ? 'd' : '0' + v;
- OS << TypeString(t, TDTypeVec[i]) << ";\n";
- }
+ OS << TypeString('s', TDTypeVec[i]);
+ OS << " " << TypeString('d', TDTypeVec[i]) << ";\n";
}
OS << "\n";
+ OS << "typedef __attribute__((__vector_size__(8))) "
+ "double float64x1_t;\n";
+ OS << "typedef __attribute__((__vector_size__(16))) "
+ "double float64x2_t;\n";
+ OS << "\n";
// Emit struct typedefs.
- for (unsigned vi = 1; vi != 5; ++vi) {
+ for (unsigned vi = 2; vi != 5; ++vi) {
for (unsigned i = 0, e = TDTypeVec.size(); i != e; ++i) {
std::string ts = TypeString('d', TDTypeVec[i]);
- std::string vs = (vi > 1) ? TypeString('0' + vi, TDTypeVec[i]) : ts;
- OS << "typedef struct __" << vs << " {\n";
- OS << " __neon_" << ts << " val";
- if (vi > 1)
- OS << "[" << utostr(vi) << "]";
- OS << ";\n} " << vs << ";\n\n";
+ std::string vs = TypeString('0' + vi, TDTypeVec[i]);
+ OS << "typedef struct " << vs << " {\n";
+ OS << " " << ts << " val";
+ OS << "[" << utostr(vi) << "]";
+ OS << ";\n} ";
+ OS << vs << ";\n\n";
}
}
OpKind k = OpMap[R->getValueAsDef("Operand")->getName()];
+ bool define = Proto.find('i') != std::string::npos;
+
for (unsigned ti = 0, te = TypeVec.size(); ti != te; ++ti) {
assert(!Proto.empty() && "");
// static always inline + return type
- OS << "__ai " << TypeString(Proto[0], TypeVec[ti]);
+ if (define)
+ OS << "#define";
+ else
+ OS << "__ai " << TypeString(Proto[0], TypeVec[ti]);
// Function name with type suffix
OS << " " << MangleName(name, TypeVec[ti], ClassS);
OS << GenArgs(Proto, TypeVec[ti]);
// Definition.
- OS << " { ";
+ if (define)
+ OS << " ";
+ else
+ OS << " { ";
if (k != OpNone) {
OS << GenOpString(k, Proto, TypeVec[ti]);
throw TGError(R->getLoc(), "Builtin has no class kind");
OS << GenBuiltin(name, Proto, TypeVec[ti], ck);
}
-
- OS << " }\n";
+ if (!define)
+ OS << " }";
+ OS << "\n";
}
OS << "\n";
}
OS << "#endif /* __ARM_NEON_H */\n";
}
+static unsigned RangeFromType(StringRef typestr) {
+ // base type to get the type string for.
+ bool quad = false, dummy = false;
+ char type = ClassifyType(typestr, quad, dummy, dummy);
+
+ switch (type) {
+ case 'c':
+ return (8 << (int)quad) - 1;
+ case 'h':
+ case 's':
+ return (4 << (int)quad) - 1;
+ case 'f':
+ case 'i':
+ return (2 << (int)quad) - 1;
+ case 'l':
+ return (1 << (int)quad) - 1;
+ default:
+ throw "unhandled type!";
+ break;
+ }
+ assert(0 && "unreachable");
+ return 0;
+}
+
+/// runHeader - Emit a file with sections defining:
+/// 1. the NEON section of BuiltinsARM.def.
+/// 2. the SemaChecking code for the type overload checking.
+/// 3. the SemaChecking code for validation of intrinsic immedate arguments.
void NeonEmitter::runHeader(raw_ostream &OS) {
std::vector<Record*> RV = Records.getAllDerivedDefinitions("Inst");
StringMap<OpKind> EmittedMap;
+ // Generate BuiltinsARM.def for NEON
+ OS << "#ifdef GET_NEON_BUILTINS\n";
for (unsigned i = 0, e = RV.size(); i != e; ++i) {
Record *R = RV[i];
-
OpKind k = OpMap[R->getValueAsDef("Operand")->getName()];
if (k != OpNone)
continue;
-
- std::string name = LowercaseString(R->getName());
+
std::string Proto = R->getValueAsString("Prototype");
+
+ // Functions with 'a' (the splat code) in the type prototype should not get
+ // their own builtin as they use the non-splat variant.
+ if (Proto.find('a') != std::string::npos)
+ continue;
+
std::string Types = R->getValueAsString("Types");
-
SmallVector<StringRef, 16> TypeVec;
ParseTypes(R, Types, TypeVec);
-
+
if (R->getSuperClasses().size() < 2)
throw TGError(R->getLoc(), "Builtin has no class kind");
+ std::string name = LowercaseString(R->getName());
ClassKind ck = ClassMap[R->getSuperClasses()[1]];
for (unsigned ti = 0, te = TypeVec.size(); ti != te; ++ti) {
+ // Generate the BuiltinsARM.def declaration for this builtin, ensuring
+ // that each unique BUILTIN() macro appears only once in the output
+ // stream.
std::string bd = GenBuiltinDef(name, Proto, TypeVec[ti], ck);
if (EmittedMap.count(bd))
continue;
OS << bd << "\n";
}
}
+ OS << "#endif\n\n";
+
+ // Generate the overloaded type checking code for SemaChecking.cpp
+ OS << "#ifdef GET_NEON_OVERLOAD_CHECK\n";
+ for (unsigned i = 0, e = RV.size(); i != e; ++i) {
+ Record *R = RV[i];
+ OpKind k = OpMap[R->getValueAsDef("Operand")->getName()];
+ if (k != OpNone)
+ continue;
+
+ std::string Proto = R->getValueAsString("Prototype");
+ std::string Types = R->getValueAsString("Types");
+ std::string name = LowercaseString(R->getName());
+
+ // Functions with 'a' (the splat code) in the type prototype should not get
+ // their own builtin as they use the non-splat variant.
+ if (Proto.find('a') != std::string::npos)
+ continue;
+
+ // Functions which have a scalar argument cannot be overloaded, no need to
+ // check them if we are emitting the type checking code.
+ if (Proto.find('s') != std::string::npos)
+ continue;
+
+ SmallVector<StringRef, 16> TypeVec;
+ ParseTypes(R, Types, TypeVec);
+
+ if (R->getSuperClasses().size() < 2)
+ throw TGError(R->getLoc(), "Builtin has no class kind");
+
+ int si = -1, qi = -1;
+ unsigned mask = 0, qmask = 0;
+ for (unsigned ti = 0, te = TypeVec.size(); ti != te; ++ti) {
+ // Generate the switch case(s) for this builtin for the type validation.
+ bool quad = false, poly = false, usgn = false;
+ (void) ClassifyType(TypeVec[ti], quad, poly, usgn);
+
+ if (quad) {
+ qi = ti;
+ qmask |= 1 << GetNeonEnum(Proto, TypeVec[ti]);
+ } else {
+ si = ti;
+ mask |= 1 << GetNeonEnum(Proto, TypeVec[ti]);
+ }
+ }
+ if (mask)
+ OS << "case ARM::BI__builtin_neon_"
+ << MangleName(name, TypeVec[si], ClassB)
+ << ": mask = " << "0x" << utohexstr(mask) << "; break;\n";
+ if (qmask)
+ OS << "case ARM::BI__builtin_neon_"
+ << MangleName(name, TypeVec[qi], ClassB)
+ << ": mask = " << "0x" << utohexstr(qmask) << "; break;\n";
+ }
+ OS << "#endif\n\n";
+
+ // Generate the intrinsic range checking code for shift/lane immediates.
+ OS << "#ifdef GET_NEON_IMMEDIATE_CHECK\n";
+ for (unsigned i = 0, e = RV.size(); i != e; ++i) {
+ Record *R = RV[i];
+
+ OpKind k = OpMap[R->getValueAsDef("Operand")->getName()];
+ if (k != OpNone)
+ continue;
+
+ std::string name = LowercaseString(R->getName());
+ std::string Proto = R->getValueAsString("Prototype");
+ std::string Types = R->getValueAsString("Types");
+
+ // Functions with 'a' (the splat code) in the type prototype should not get
+ // their own builtin as they use the non-splat variant.
+ if (Proto.find('a') != std::string::npos)
+ continue;
+
+ // Functions which do not have an immediate do not need to have range
+ // checking code emitted.
+ if (Proto.find('i') == std::string::npos)
+ continue;
+
+ SmallVector<StringRef, 16> TypeVec;
+ ParseTypes(R, Types, TypeVec);
+
+ if (R->getSuperClasses().size() < 2)
+ throw TGError(R->getLoc(), "Builtin has no class kind");
+
+ ClassKind ck = ClassMap[R->getSuperClasses()[1]];
+
+ for (unsigned ti = 0, te = TypeVec.size(); ti != te; ++ti) {
+ std::string namestr, shiftstr, rangestr;
+
+ // Builtins which are overloaded by type will need to have their upper
+ // bound computed at Sema time based on the type constant.
+ if (Proto.find('s') == std::string::npos) {
+ ck = ClassB;
+ if (R->getValueAsBit("isShift")) {
+ shiftstr = ", true";
+
+ // Right shifts have an 'r' in the name, left shifts do not.
+ if (name.find('r') != std::string::npos)
+ rangestr = "l = 1; ";
+ }
+ rangestr += "u = RFT(TV" + shiftstr + ")";
+ } else {
+ rangestr = "u = " + utostr(RangeFromType(TypeVec[ti]));
+ }
+ // Make sure cases appear only once by uniquing them in a string map.
+ namestr = MangleName(name, TypeVec[ti], ck);
+ if (EmittedMap.count(namestr))
+ continue;
+ EmittedMap[namestr] = OpNone;
+
+ // Calculate the index of the immediate that should be range checked.
+ unsigned immidx = 0;
+
+ // Builtins that return a struct of multiple vectors have an extra
+ // leading arg for the struct return.
+ if (Proto[0] == '2' || Proto[0] == '3' || Proto[0] == '4')
+ ++immidx;
+
+ // Add one to the index for each argument until we reach the immediate
+ // to be checked. Structs of vectors are passed as multiple arguments.
+ for (unsigned ii = 1, ie = Proto.size(); ii != ie; ++ii) {
+ switch (Proto[ii]) {
+ default: immidx += 1; break;
+ case '2': immidx += 2; break;
+ case '3': immidx += 3; break;
+ case '4': immidx += 4; break;
+ case 'i': ie = ii + 1; break;
+ }
+ }
+ OS << "case ARM::BI__builtin_neon_" << MangleName(name, TypeVec[ti], ck)
+ << ": i = " << immidx << "; " << rangestr << "; break;\n";
+ }
+ }
+ OS << "#endif\n\n";
}
projects / oota-llvm.git / blobdiff