// 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) {
usgn = true;
break;
case 's':
+ case 'a':
scal = true;
break;
case 'k':
return type;
}
+/// TypeString - for a modifier and type, generate the name of the typedef for
+/// that type. If generic is true, emit the generic vector type rather than
+/// the public NEON type. QUc -> uint8x8t_t / __neon_uint8x8_t.
static std::string TypeString(const char mod, StringRef typestr,
- bool ret = false) {
+ bool generic = false) {
bool quad = false;
bool poly = false;
bool usgn = false;
SmallString<128> s;
- if (ret)
+ if (generic)
s += "__neon_";
if (usgn)
return s.str();
}
+/// TypeString - 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;
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")
return s;
}
+static std::string Duplicate(unsigned nElts, StringRef typestr,
+ const std::string &a) {
+ std::string s;
+
+ s = "(__neon_" + TypeString('d', typestr) + "){ ";
+ for (unsigned i = 0; i != nElts; ++i) {
+ s += a;
+ if ((i + 1) < nElts)
+ s += ", ";
+ }
+ 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";
-
bool dummy, quad = false;
char type = ClassifyType(typestr, quad, dummy, dummy);
unsigned nElts = 0;
case 'h': nElts = 4; break;
case 'f': nElts = 2; break;
}
- nElts <<= quad;
+
+ std::string ts = TypeString(proto[0], typestr);
+ std::string s = ts + " r; r";
if (structTypes)
s += ".val";
case OpSub:
s += a + " - " + b;
break;
+ case OpMulN:
+ b = Duplicate(nElts << (int)quad, typestr, "b");
case OpMul:
s += a + " * " + b;
break;
+ case OpMlaN:
+ c = Duplicate(nElts << (int)quad, typestr, "c");
case OpMla:
s += a + " + ( " + b + " * " + c + " )";
break;
+ case OpMlsN:
+ c = Duplicate(nElts << (int)quad, typestr, "c");
case OpMls:
s += a + " - ( " + b + " * " + c + " )";
break;
s += "(__neon_int64x1_t)(((__neon_int64x2_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 << (int)quad, typestr, a);
+ break;
+ case OpSelect:
+ // ((0 & 1) | (~0 & 2))
+ ts = TypeString(proto[1], typestr);
+ s += "( " + a + " & (__neon_" + ts + ")" + b + ") | ";
+ s += "(~" + a + " & (__neon_" + ts + ")" + c + ")";
+ break;
+ case OpRev16:
+ s += "__builtin_shufflevector(" + a + ", " + a;
+ for (unsigned i = 2; i <= nElts << (int)quad; i += 2)
+ for (unsigned j = 0; j != 2; ++j)
+ s += ", " + utostr(i - j - 1);
+ s += ")";
+ break;
+ case OpRev32:
+ nElts >>= 1;
+ s += "__builtin_shufflevector(" + a + ", " + a;
+ for (unsigned i = nElts; i <= nElts << (1 + (int)quad); i += nElts)
+ for (unsigned j = 0; j != nElts; ++j)
+ s += ", " + utostr(i - j - 1);
+ s += ")";
+ break;
+ case OpRev64:
+ s += "__builtin_shufflevector(" + a + ", " + a;
+ for (unsigned i = nElts; i <= nElts << (int)quad; i += nElts)
+ for (unsigned j = 0; j != nElts; ++j)
+ s += ", " + utostr(i - j - 1);
+ s += ")";
break;
default:
throw "unknown OpKind!";
static std::string GenBuiltin(const std::string &name, const std::string &proto,
StringRef typestr, ClassKind ck,
bool structTypes = true) {
+ bool dummy, quad = 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;
+ }
+ if (quad) nElts <<= 1;
+
char arg = 'a';
std::string s;
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 += "(";
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')){
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;
}
- // Parenthesize the args from the macro.
- if (define)
- s.push_back('(');
- s.push_back(arg);
- if (define)
- s.push_back(')');
+ if (splat && (i + 1) == e)
+ s += Duplicate(nElts, typestr, args);
+ else
+ s += args;
if (structTypes && proto[i] != 's' && proto[i] != 'i' && proto[i] != 'l' &&
- proto[i] != 'p' && proto[i] != 'c') {
+ proto[i] != 'p' && proto[i] != 'c' && proto[i] != 'a') {
s += ".val";
}
if ((i + 1) < e)
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 << "#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;
+ }
+}
+
+/// 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.
+ namestr = MangleName(name, TypeVec[ti], ck);
+ if (EmittedMap.count(namestr))
+ continue;
+ EmittedMap[namestr] = OpNone;
+
+ unsigned immidx = 0;
+ 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";
}