//===----------------------------------------------------------------------===//
//
// This tablegen backend is responsible for emitting arm_neon.h, which includes
-// a declaration and definition of each function specified by the ARM NEON
+// 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
+// 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.
+// called, rather than the normal run() entry point. A complete set of tests
+// for Neon intrinsics can be generated by calling the runTests() entry point.
//
//===----------------------------------------------------------------------===//
SmallVectorImpl<StringRef> &TV) {
const char *data = s.data();
int len = 0;
-
+
for (unsigned i = 0, e = s.size(); i != e; ++i, ++len) {
if (data[len] == 'P' || data[len] == 'Q' || data[len] == 'U')
continue;
-
+
switch (data[len]) {
case 'c':
case 's':
return 'i';
case 'i':
return 'l';
+ case 'h':
+ return 'f';
default: throw "unhandled type in widen!";
}
return '\0';
return 'i';
case 'f':
return 'h';
- default: throw "unhandled type in widen!";
+ default: throw "unhandled type in narrow!";
}
return '\0';
}
/// the quad-vector, polynomial, or unsigned modifiers set.
static char ClassifyType(StringRef ty, bool &quad, bool &poly, bool &usgn) {
unsigned off = 0;
-
+
// remember quad.
if (ty[off] == 'Q') {
quad = true;
++off;
}
-
+
// remember poly.
if (ty[off] == 'P') {
poly = true;
++off;
}
-
+
// remember unsigned.
if (ty[off] == 'U') {
usgn = true;
++off;
}
-
+
// base type to get the type string for.
return ty[off];
}
case 'n':
type = Widen(type);
break;
+ case 'i':
+ type = 'i';
+ scal = true;
+ break;
case 'l':
type = 'l';
scal = true;
bool scal = false;
bool cnst = false;
bool pntr = false;
-
+
if (mod == 'v')
return "void";
if (mod == 'i')
return "int";
-
+
// 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);
-
+
SmallString<128> s;
-
+
if (usgn)
s.push_back('u');
-
+
switch (type) {
case 'c':
s += poly ? "poly8" : "int8";
s += "x3";
if (mod == '4')
s += "x4";
-
+
// Append _t, finishing the type string typedef type.
s += "_t";
-
+
if (cnst)
s += " const";
-
+
if (pntr)
s += " *";
-
+
return s.str();
}
bool scal = false;
bool cnst = false;
bool pntr = false;
-
+
if (mod == 'v')
- return "v";
+ return "v"; // void
if (mod == 'i')
- return "i";
-
+ return "i"; // int
+
// 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);
+ // All pointers are void* pointers. Change type to 'v' now.
if (pntr) {
usgn = false;
poly = false;
type = 'v';
}
+ // Treat half-float ('h') types as unsigned short ('s') types.
if (type == 'h') {
type = 's';
usgn = true;
if (usgn)
s.push_back('U');
-
- if (type == 'l')
+ else if (type == 'c')
+ s.push_back('S'); // make chars explicitly signed
+
+ if (type == 'l') // 64-bit long
s += "LLi";
else
s.push_back(type);
-
+
if (cnst)
s.push_back('C');
if (pntr)
// 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' || mod == '3' || mod == '4')
- return "vv*";
+ if (mod >= '2' && mod <= '4')
+ return "vv*"; // void result with void* first argument
if (mod == 'f' || (ck != ClassB && type == 'f'))
return quad ? "V4f" : "V2f";
if (ck != ClassB && type == 's')
return quad ? "V4i" : "V2i";
if (ck != ClassB && type == 'l')
return quad ? "V2LLi" : "V1LLi";
-
- return quad ? "V16c" : "V8c";
- }
+
+ return quad ? "V16Sc" : "V8Sc";
+ }
// Non-return array types are passed as individual vectors.
if (mod == '2')
- return quad ? "V16cV16c" : "V8cV8c";
+ return quad ? "V16ScV16Sc" : "V8ScV8Sc";
if (mod == '3')
- return quad ? "V16cV16cV16c" : "V8cV8cV8c";
+ return quad ? "V16ScV16ScV16Sc" : "V8ScV8ScV8Sc";
if (mod == '4')
- return quad ? "V16cV16cV16cV16c" : "V8cV8cV8cV8c";
+ return quad ? "V16ScV16ScV16ScV16Sc" : "V8ScV8ScV8ScV8Sc";
if (mod == 'f' || (ck != ClassB && type == 'f'))
return quad ? "V4f" : "V2f";
return quad ? "V4i" : "V2i";
if (ck != ClassB && type == 'l')
return quad ? "V2LLi" : "V1LLi";
-
- return quad ? "V16c" : "V8c";
+
+ return quad ? "V16Sc" : "V8Sc";
}
-/// MangleName - Append a type or width suffix to a base neon function name,
+/// 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 name;
-
+
bool quad = false;
bool poly = false;
bool usgn = false;
char type = ClassifyType(typestr, quad, poly, usgn);
std::string s = name;
-
+
switch (type) {
case 'c':
switch (ck) {
}
if (ck == ClassB)
s += "_v";
-
- // Insert a 'q' before the first '_' character so that it ends up before
+
+ // Insert a 'q' before the first '_' character so that it ends up before
// _lane or _n on vector-scalar operations.
if (quad) {
size_t pos = s.find('_');
return s;
}
+/// UseMacro - Examine the prototype string to determine if the intrinsic
+/// should be defined as a preprocessor macro instead of an inline function.
+static bool UseMacro(const std::string &proto) {
+ // 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.
+ if (proto.find('i') != std::string::npos)
+ return true;
+
+ // Pointer arguments need to use macros to avoid hiding aligned attributes
+ // from the pointer type.
+ if (proto.find('p') != std::string::npos ||
+ proto.find('c') != std::string::npos)
+ return true;
+
+ return false;
+}
+
+/// MacroArgUsedDirectly - Return true if argument i for an intrinsic that is
+/// defined as a macro should be accessed directly instead of being first
+/// assigned to a local temporary.
+static bool MacroArgUsedDirectly(const std::string &proto, unsigned i) {
+ return (proto[i] == 'i' || proto[i] == 'p' || proto[i] == 'c');
+}
+
// 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;
+ bool define = UseMacro(proto);
char arg = 'a';
-
+
std::string s;
s += "(";
-
+
for (unsigned i = 1, e = proto.size(); i != e; ++i, ++arg) {
- if (!define) {
- s += TypeString(proto[i], typestr);
- s.push_back(' ');
+ if (define) {
+ // Some macro arguments are used directly instead of being assigned
+ // to local temporaries; prepend an underscore prefix to make their
+ // names consistent with the local temporaries.
+ if (MacroArgUsedDirectly(proto, i))
+ s += "__";
+ } else {
+ s += TypeString(proto[i], typestr) + " __";
}
s.push_back(arg);
if ((i + 1) < e)
s += ", ";
}
-
+
s += ")";
return s;
}
-static std::string Duplicate(unsigned nElts, StringRef typestr,
+// Macro arguments are not type-checked like inline function arguments, so
+// assign them to local temporaries to get the right type checking.
+static std::string GenMacroLocals(const std::string &proto, StringRef typestr) {
+ char arg = 'a';
+ std::string s;
+ bool generatedLocal = false;
+
+ for (unsigned i = 1, e = proto.size(); i != e; ++i, ++arg) {
+ // Do not create a temporary for an immediate argument.
+ // That would defeat the whole point of using a macro!
+ if (proto[i] == 'i')
+ continue;
+ generatedLocal = true;
+
+ // For other (non-immediate) arguments that are used directly, a local
+ // temporary is still needed to get the correct type checking, even though
+ // that temporary is not used for anything.
+ if (MacroArgUsedDirectly(proto, i)) {
+ s += TypeString(proto[i], typestr) + " __";
+ s.push_back(arg);
+ s += "_ = (__";
+ s.push_back(arg);
+ s += "); (void)__";
+ s.push_back(arg);
+ s += "_; ";
+ continue;
+ }
+
+ s += TypeString(proto[i], typestr) + " __";
+ s.push_back(arg);
+ s += " = (";
+ s.push_back(arg);
+ s += "); ";
+ }
+
+ if (generatedLocal)
+ s += "\\\n ";
+ return s;
+}
+
+// Use the vmovl builtin to sign-extend or zero-extend a vector.
+static std::string Extend(StringRef typestr, const std::string &a) {
+ std::string s;
+ s = MangleName("vmovl", typestr, ClassS);
+ s += "(" + a + ")";
+ return s;
+}
+
+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;
s += ", ";
}
s += " }";
-
+
+ return s;
+}
+
+static std::string SplatLane(unsigned nElts, const std::string &vec,
+ const std::string &lane) {
+ std::string s = "__builtin_shufflevector(" + vec + ", " + vec;
+ for (unsigned i = 0; i < nElts; ++i)
+ s += ", " + lane;
+ s += ")";
return s;
}
StringRef typestr) {
bool quad;
unsigned nElts = GetNumElements(typestr, quad);
-
+ bool define = UseMacro(proto);
+
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";
+ if (!define) {
+ s = "return ";
}
-
- s += " = ";
switch(op) {
case OpAdd:
- s += "a + b";
+ s += "__a + __b;";
+ break;
+ case OpAddl:
+ s += Extend(typestr, "__a") + " + " + Extend(typestr, "__b") + ";";
+ break;
+ case OpAddw:
+ s += "__a + " + Extend(typestr, "__b") + ";";
break;
case OpSub:
- s += "a - b";
+ s += "__a - __b;";
+ break;
+ case OpSubl:
+ s += Extend(typestr, "__a") + " - " + Extend(typestr, "__b") + ";";
+ break;
+ case OpSubw:
+ s += "__a - " + Extend(typestr, "__b") + ";";
break;
case OpMulN:
- s += "a * " + Duplicate(nElts, typestr, "b");
+ s += "__a * " + Duplicate(nElts, typestr, "__b") + ";";
+ break;
+ case OpMulLane:
+ s += "__a * " + SplatLane(nElts, "__b", "__c") + ";";
break;
case OpMul:
- s += "a * b";
+ s += "__a * __b;";
+ break;
+ case OpMullLane:
+ s += MangleName("vmull", typestr, ClassS) + "(__a, " +
+ SplatLane(nElts, "__b", "__c") + ");";
break;
case OpMlaN:
- s += "a + (b * " + Duplicate(nElts, typestr, "c") + ")";
+ s += "__a + (__b * " + Duplicate(nElts, typestr, "__c") + ");";
+ break;
+ case OpMlaLane:
+ s += "__a + (__b * " + SplatLane(nElts, "__c", "__d") + ");";
break;
case OpMla:
- s += "a + (b * c)";
+ s += "__a + (__b * __c);";
+ break;
+ case OpMlalN:
+ s += "__a + " + MangleName("vmull", typestr, ClassS) + "(__b, " +
+ Duplicate(nElts, typestr, "__c") + ");";
+ break;
+ case OpMlalLane:
+ s += "__a + " + MangleName("vmull", typestr, ClassS) + "(__b, " +
+ SplatLane(nElts, "__c", "__d") + ");";
+ break;
+ case OpMlal:
+ s += "__a + " + MangleName("vmull", typestr, ClassS) + "(__b, __c);";
break;
case OpMlsN:
- s += "a - (b * " + Duplicate(nElts, typestr, "c") + ")";
+ s += "__a - (__b * " + Duplicate(nElts, typestr, "__c") + ");";
+ break;
+ case OpMlsLane:
+ s += "__a - (__b * " + SplatLane(nElts, "__c", "__d") + ");";
break;
case OpMls:
- s += "a - (b * c)";
+ s += "__a - (__b * __c);";
+ break;
+ case OpMlslN:
+ s += "__a - " + MangleName("vmull", typestr, ClassS) + "(__b, " +
+ Duplicate(nElts, typestr, "__c") + ");";
+ break;
+ case OpMlslLane:
+ s += "__a - " + MangleName("vmull", typestr, ClassS) + "(__b, " +
+ SplatLane(nElts, "__c", "__d") + ");";
+ break;
+ case OpMlsl:
+ s += "__a - " + MangleName("vmull", typestr, ClassS) + "(__b, __c);";
+ break;
+ case OpQDMullLane:
+ s += MangleName("vqdmull", typestr, ClassS) + "(__a, " +
+ SplatLane(nElts, "__b", "__c") + ");";
+ break;
+ case OpQDMlalLane:
+ s += MangleName("vqdmlal", typestr, ClassS) + "(__a, __b, " +
+ SplatLane(nElts, "__c", "__d") + ");";
+ break;
+ case OpQDMlslLane:
+ s += MangleName("vqdmlsl", typestr, ClassS) + "(__a, __b, " +
+ SplatLane(nElts, "__c", "__d") + ");";
+ break;
+ case OpQDMulhLane:
+ s += MangleName("vqdmulh", typestr, ClassS) + "(__a, " +
+ SplatLane(nElts, "__b", "__c") + ");";
+ break;
+ case OpQRDMulhLane:
+ s += MangleName("vqrdmulh", typestr, ClassS) + "(__a, " +
+ SplatLane(nElts, "__b", "__c") + ");";
break;
case OpEq:
- s += "(" + ts + ")(a == b)";
+ s += "(" + ts + ")(__a == __b);";
break;
case OpGe:
- s += "(" + ts + ")(a >= b)";
+ s += "(" + ts + ")(__a >= __b);";
break;
case OpLe:
- s += "(" + ts + ")(a <= b)";
+ s += "(" + ts + ")(__a <= __b);";
break;
case OpGt:
- s += "(" + ts + ")(a > b)";
+ s += "(" + ts + ")(__a > __b);";
break;
case OpLt:
- s += "(" + 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 += "(" + ts + ")a";
+ s += "(" + ts + ")__a;";
break;
case OpConcat:
- s += "__builtin_shufflevector((int64x1_t)a";
- s += ", (int64x1_t)b, 0, 1)";
+ s += "(" + ts + ")__builtin_shufflevector((int64x1_t)__a";
+ s += ", (int64x1_t)__b, 0, 1);";
break;
case OpHi:
- s += "(((float64x2_t)a)[1])";
+ s += "(" + ts +
+ ")__builtin_shufflevector((int64x2_t)__a, (int64x2_t)__a, 1);";
break;
case OpLo:
- s += "(((float64x2_t)a)[0])";
+ s += "(" + ts +
+ ")__builtin_shufflevector((int64x2_t)__a, (int64x2_t)__a, 0);";
break;
case OpDup:
- s += Duplicate(nElts, typestr, "a");
+ s += Duplicate(nElts, typestr, "__a") + ";";
+ break;
+ case OpDupLane:
+ s += SplatLane(nElts, "__a", "__b") + ";";
break;
case OpSelect:
// ((0 & 1) | (~0 & 2))
+ s += "(" + ts + ")";
ts = TypeString(proto[1], typestr);
- s += "(a & (" + ts + ")b) | ";
- s += "(~a & (" + ts + ")c)";
+ s += "((__a & (" + ts + ")__b) | ";
+ s += "(~__a & (" + ts + ")__c));";
break;
case OpRev16:
- s += "__builtin_shufflevector(a, a";
+ 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 += ")";
+ s += ");";
break;
case OpRev32: {
unsigned WordElts = nElts >> (1 + (int)quad);
- s += "__builtin_shufflevector(a, a";
+ 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 += ")";
+ s += ");";
break;
}
case OpRev64: {
unsigned DblWordElts = nElts >> (int)quad;
- s += "__builtin_shufflevector(a, a";
+ 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 += ")";
+ s += ");";
+ break;
+ }
+ case OpAbdl: {
+ std::string abd = MangleName("vabd", typestr, ClassS) + "(__a, __b)";
+ if (typestr[0] != 'U') {
+ // vabd results are always unsigned and must be zero-extended.
+ std::string utype = "U" + typestr.str();
+ s += "(" + TypeString(proto[0], typestr) + ")";
+ abd = "(" + TypeString('d', utype) + ")" + abd;
+ s += Extend(utype, abd) + ";";
+ } else {
+ s += Extend(typestr, abd) + ";";
+ }
+ break;
+ }
+ case OpAba:
+ s += "__a + " + MangleName("vabd", typestr, ClassS) + "(__b, __c);";
+ break;
+ case OpAbal: {
+ s += "__a + ";
+ std::string abd = MangleName("vabd", typestr, ClassS) + "(__b, __c)";
+ if (typestr[0] != 'U') {
+ // vabd results are always unsigned and must be zero-extended.
+ std::string utype = "U" + typestr.str();
+ s += "(" + TypeString(proto[0], typestr) + ")";
+ abd = "(" + TypeString('d', utype) + ")" + abd;
+ s += Extend(utype, abd) + ";";
+ } else {
+ s += Extend(typestr, abd) + ";";
+ }
break;
}
default:
throw "unknown OpKind!";
break;
}
- if (op == OpHi || op == OpLo)
- s += "; return u.r;";
- else
- s += "; return r;";
return s;
}
bool scal = false;
bool cnst = false;
bool pntr = false;
-
+
// 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);
ret |= 0x08;
if (quad && proto[1] != 'g')
ret |= 0x10;
-
+
switch (type) {
- case 'c':
+ case 'c':
ret |= poly ? 5 : 0;
break;
case 's':
// Generate the definition for this intrinsic, e.g. __builtin_neon_cls(a)
static std::string GenBuiltin(const std::string &name, const std::string &proto,
StringRef typestr, ClassKind ck) {
- bool quad;
- unsigned nElts = GetNumElements(typestr, quad);
- char arg = 'a';
std::string s;
// 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');
+ bool sret = (proto[0] >= '2' && 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;
+ bool define = UseMacro(proto);
- // 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
- // special enums.
+ // Check if the prototype has a scalar operand with the type of the vector
+ // elements. If not, bitcasting the args will take care of arg checking.
+ // The actual signedness etc. will be taken care of with special enums.
if (proto.find('s') == std::string::npos)
ck = ClassB;
if (proto[0] != 'v') {
std::string ts = TypeString(proto[0], typestr);
-
+
if (define) {
if (sret)
- s += "({ " + ts + " r; ";
- else if (proto[0] != 's')
+ s += ts + " r; ";
+ else
s += "(" + ts + ")";
} else if (sret) {
s += ts + " r; ";
} else {
- s += ts + " r; r = ";
+ s += "return (" + ts + ")";
}
}
-
+
bool splat = proto.find('a') != std::string::npos;
-
+
s += "__builtin_neon_";
if (splat) {
+ // Call the non-splat builtin: chop off the "_n" suffix from the name.
std::string vname(name, 0, name.size()-2);
s += MangleName(vname, typestr, ck);
} else {
// builtins.
if (sret)
s += "&r, ";
-
+
+ char arg = 'a';
for (unsigned i = 1, e = proto.size(); i != e; ++i, ++arg) {
std::string args = std::string(&arg, 1);
- if (define)
- args = "(" + args + ")";
-
+
+ // Use the local temporaries instead of the macro arguments.
+ args = "__" + args;
+
+ bool argQuad = false;
+ bool argPoly = false;
+ bool argUsgn = false;
+ bool argScalar = false;
+ bool dummy = false;
+ char argType = ClassifyType(typestr, argQuad, argPoly, argUsgn);
+ argType = ModType(proto[i], argType, argQuad, argPoly, argUsgn, argScalar,
+ dummy, dummy);
+
// Handle multiple-vector values specially, emitting each subvector as an
// argument to the __builtin.
- if (proto[i] == '2' || proto[i] == '3' || proto[i] == '4') {
+ if (proto[i] >= '2' && proto[i] <= '4') {
+ // Check if an explicit cast is needed.
+ if (argType != 'c' || argPoly || argUsgn)
+ args = (argQuad ? "(int8x16_t)" : "(int8x8_t)") + args;
+
for (unsigned vi = 0, ve = proto[i] - '0'; vi != ve; ++vi) {
s += args + ".val[" + utostr(vi) + "]";
if ((vi + 1) < ve)
continue;
}
-
- if (splat && (i + 1) == e)
- s += Duplicate(nElts, typestr, args);
- else
- s += args;
+
+ if (splat && (i + 1) == e)
+ args = Duplicate(GetNumElements(typestr, argQuad), typestr, args);
+
+ // Check if an explicit cast is needed.
+ if ((splat || !argScalar) &&
+ ((ck == ClassB && argType != 'c') || argPoly || argUsgn)) {
+ std::string argTypeStr = "c";
+ if (ck != ClassB)
+ argTypeStr = argType;
+ if (argQuad)
+ argTypeStr = "Q" + argTypeStr;
+ args = "(" + TypeString('d', argTypeStr) + ")" + args;
+ }
+
+ s += args;
if ((i + 1) < e)
s += ", ";
}
-
+
// Extra constant integer to hold type class enum for this function, e.g. s8
if (ck == ClassB)
s += ", " + utostr(GetNeonEnum(proto, typestr));
-
- if (define)
- s += ")";
- else
- s += ");";
- if (proto[0] != 'v') {
- if (define) {
- if (sret)
- s += "; r; })";
- } else {
+ s += ");";
+
+ if (proto[0] != 'v' && sret) {
+ if (define)
+ s += " r;";
+ else
s += " return r;";
- }
}
return s;
}
-static std::string GenBuiltinDef(const std::string &name,
+static std::string GenBuiltinDef(const std::string &name,
const std::string &proto,
StringRef typestr, ClassKind ck) {
std::string s("BUILTIN(__builtin_neon_");
- // If all types are the same size, bitcasting the args will take care
+ // 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
// special enums.
if (proto.find('s') == std::string::npos)
ck = ClassB;
-
+
s += MangleName(name, typestr, ck);
s += ", \"";
-
+
for (unsigned i = 0, e = proto.size(); i != e; ++i)
s += BuiltinTypeString(proto[i], typestr, ck, i == 0);
// Extra constant integer to hold type class enum for this function, e.g. s8
if (ck == ClassB)
s += "i";
-
+
s += "\", \"n\")";
return s;
}
+static std::string GenIntrinsic(const std::string &name,
+ const std::string &proto,
+ StringRef outTypeStr, StringRef inTypeStr,
+ OpKind kind, ClassKind classKind) {
+ assert(!proto.empty() && "");
+ bool define = UseMacro(proto);
+ std::string s;
+
+ // static always inline + return type
+ if (define)
+ s += "#define ";
+ else
+ s += "__ai " + TypeString(proto[0], outTypeStr) + " ";
+
+ // Function name with type suffix
+ std::string mangledName = MangleName(name, outTypeStr, ClassS);
+ if (outTypeStr != inTypeStr) {
+ // If the input type is different (e.g., for vreinterpret), append a suffix
+ // for the input type. String off a "Q" (quad) prefix so that MangleName
+ // does not insert another "q" in the name.
+ unsigned typeStrOff = (inTypeStr[0] == 'Q' ? 1 : 0);
+ StringRef inTypeNoQuad = inTypeStr.substr(typeStrOff);
+ mangledName = MangleName(mangledName, inTypeNoQuad, ClassS);
+ }
+ s += mangledName;
+
+ // Function arguments
+ s += GenArgs(proto, inTypeStr);
+
+ // Definition.
+ if (define) {
+ s += " __extension__ ({ \\\n ";
+ s += GenMacroLocals(proto, inTypeStr);
+ } else {
+ s += " { \\\n ";
+ }
+
+ if (kind != OpNone)
+ s += GenOpString(kind, proto, outTypeStr);
+ else
+ s += GenBuiltin(name, proto, outTypeStr, classKind);
+ if (define)
+ s += " })";
+ else
+ s += " }";
+ s += "\n";
+ 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);
-
- // FIXME: emit license into file?
-
+ OS <<
+ "/*===---- arm_neon.h - ARM Neon intrinsics ------------------------------"
+ "---===\n"
+ " *\n"
+ " * Permission is hereby granted, free of charge, to any person obtaining "
+ "a copy\n"
+ " * of this software and associated documentation files (the \"Software\"),"
+ " to deal\n"
+ " * in the Software without restriction, including without limitation the "
+ "rights\n"
+ " * to use, copy, modify, merge, publish, distribute, sublicense, "
+ "and/or sell\n"
+ " * copies of the Software, and to permit persons to whom the Software is\n"
+ " * furnished to do so, subject to the following conditions:\n"
+ " *\n"
+ " * The above copyright notice and this permission notice shall be "
+ "included in\n"
+ " * all copies or substantial portions of the Software.\n"
+ " *\n"
+ " * THE SOFTWARE IS PROVIDED \"AS IS\", WITHOUT WARRANTY OF ANY KIND, "
+ "EXPRESS OR\n"
+ " * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF "
+ "MERCHANTABILITY,\n"
+ " * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT "
+ "SHALL THE\n"
+ " * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR "
+ "OTHER\n"
+ " * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, "
+ "ARISING FROM,\n"
+ " * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER "
+ "DEALINGS IN\n"
+ " * THE SOFTWARE.\n"
+ " *\n"
+ " *===--------------------------------------------------------------------"
+ "---===\n"
+ " */\n\n";
+
OS << "#ifndef __ARM_NEON_H\n";
OS << "#define __ARM_NEON_H\n\n";
-
+
OS << "#ifndef __ARM_NEON__\n";
OS << "#error \"NEON support not enabled\"\n";
OS << "#endif\n\n";
OS << "typedef __attribute__((neon_polyvector_type(";
else
OS << "typedef __attribute__((neon_vector_type(";
-
+
unsigned nElts = GetNumElements(TDTypeVec[i], quad);
OS << utostr(nElts) << "))) ";
if (nElts < 10)
OS << " ";
-
+
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 = 2; vi != 5; ++vi) {
OS << vs << ";\n\n";
}
}
-
+
OS << "#define __ai static __attribute__((__always_inline__))\n\n";
std::vector<Record*> RV = Records.getAllDerivedDefinitions("Inst");
-
- // Unique the return+pattern types, and assign them.
+
+ // Emit vmovl, vmull and vabd intrinsics first so they can be used by other
+ // intrinsics. (Some of the saturating multiply instructions are also
+ // used to implement the corresponding "_lane" variants, but tablegen
+ // sorts the records into alphabetical order so that the "_lane" variants
+ // come after the intrinsics they use.)
+ emitIntrinsic(OS, Records.getDef("VMOVL"));
+ emitIntrinsic(OS, Records.getDef("VMULL"));
+ emitIntrinsic(OS, Records.getDef("VABD"));
+
for (unsigned i = 0, e = RV.size(); i != e; ++i) {
Record *R = RV[i];
- std::string name = LowercaseString(R->getName());
- std::string Proto = R->getValueAsString("Prototype");
- std::string Types = R->getValueAsString("Types");
-
- SmallVector<StringRef, 16> TypeVec;
- ParseTypes(R, Types, TypeVec);
-
- 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
- if (define)
- OS << "#define";
- else
- OS << "__ai " << TypeString(Proto[0], TypeVec[ti]);
-
- // Function name with type suffix
- OS << " " << MangleName(name, TypeVec[ti], ClassS);
-
- // Function arguments
- OS << GenArgs(Proto, TypeVec[ti]);
-
- // Definition.
- if (define)
- OS << " ";
- else
- OS << " { ";
-
- if (k != OpNone) {
- OS << GenOpString(k, Proto, TypeVec[ti]);
- } else {
- if (R->getSuperClasses().size() < 2)
- throw TGError(R->getLoc(), "Builtin has no class kind");
-
- ClassKind ck = ClassMap[R->getSuperClasses()[1]];
-
- if (ck == ClassNone)
- throw TGError(R->getLoc(), "Builtin has no class kind");
- OS << GenBuiltin(name, Proto, TypeVec[ti], ck);
- }
- if (!define)
- OS << " }";
- OS << "\n";
- }
- OS << "\n";
+ if (R->getName() != "VMOVL" &&
+ R->getName() != "VMULL" &&
+ R->getName() != "VABD")
+ emitIntrinsic(OS, R);
}
+
OS << "#undef __ai\n\n";
OS << "#endif /* __ARM_NEON_H */\n";
}
-static unsigned RangeFromType(StringRef typestr) {
+/// emitIntrinsic - Write out the arm_neon.h header file definitions for the
+/// intrinsics specified by record R.
+void NeonEmitter::emitIntrinsic(raw_ostream &OS, Record *R) {
+ std::string name = R->getValueAsString("Name");
+ std::string Proto = R->getValueAsString("Prototype");
+ std::string Types = R->getValueAsString("Types");
+
+ SmallVector<StringRef, 16> TypeVec;
+ ParseTypes(R, Types, TypeVec);
+
+ OpKind kind = OpMap[R->getValueAsDef("Operand")->getName()];
+
+ ClassKind classKind = ClassNone;
+ if (R->getSuperClasses().size() >= 2)
+ classKind = ClassMap[R->getSuperClasses()[1]];
+ if (classKind == ClassNone && kind == OpNone)
+ throw TGError(R->getLoc(), "Builtin has no class kind");
+
+ for (unsigned ti = 0, te = TypeVec.size(); ti != te; ++ti) {
+ if (kind == OpReinterpret) {
+ bool outQuad = false;
+ bool dummy = false;
+ (void)ClassifyType(TypeVec[ti], outQuad, dummy, dummy);
+ for (unsigned srcti = 0, srcte = TypeVec.size();
+ srcti != srcte; ++srcti) {
+ bool inQuad = false;
+ (void)ClassifyType(TypeVec[srcti], inQuad, dummy, dummy);
+ if (srcti == ti || inQuad != outQuad)
+ continue;
+ OS << GenIntrinsic(name, Proto, TypeVec[ti], TypeVec[srcti],
+ OpCast, ClassS);
+ }
+ } else {
+ OS << GenIntrinsic(name, Proto, TypeVec[ti], TypeVec[ti],
+ kind, classKind);
+ }
+ }
+ OS << "\n";
+}
+
+static unsigned RangeFromType(const char mod, StringRef typestr) {
// base type to get the type string for.
bool quad = false, dummy = false;
char type = ClassifyType(typestr, quad, dummy, dummy);
-
+ type = ModType(mod, type, quad, dummy, dummy, dummy, dummy, dummy);
+
switch (type) {
case 'c':
return (8 << (int)quad) - 1;
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) {
continue;
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());
+
+ std::string name = R->getValueAsString("Name");
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
std::string bd = GenBuiltinDef(name, Proto, TypeVec[ti], ck);
if (EmittedMap.count(bd))
continue;
-
+
EmittedMap[bd] = OpNone;
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) {
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());
-
+ std::string name = R->getValueAsString("Name");
+
// 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]);
}
}
if (mask)
- OS << "case ARM::BI__builtin_neon_"
- << MangleName(name, TypeVec[si], ClassB)
- << ": mask = " << "0x" << utohexstr(mask) << "; break;\n";
+ 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 << "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 name = R->getValueAsString("Name");
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)
+ size_t immPos = Proto.find('i');
+ if (immPos == 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) {
+
+ if (R->getValueAsBit("isVCVT_N")) {
+ // VCVT between floating- and fixed-point values takes an immediate
+ // in the range 1 to 32.
+ ck = ClassB;
+ rangestr = "l = 1; u = 31"; // upper bound = l + u
+ } else if (Proto.find('s') == std::string::npos) {
+ // Builtins which are overloaded by type will need to have their upper
+ // bound computed at Sema time based on the type constant.
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]));
+ // The immediate generally refers to a lane in the preceding argument.
+ assert(immPos > 0 && "unexpected immediate operand");
+ rangestr = "u = " + utostr(RangeFromType(Proto[immPos-1], TypeVec[ti]));
}
// Make sure cases appear only once by uniquing them in a string map.
namestr = MangleName(name, TypeVec[ti], ck);
// 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')
+ if (Proto[0] >= '2' && Proto[0] <= '4')
++immidx;
-
- // Add one to the index for each argument until we reach the immediate
+
+ // 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]) {
case 'i': ie = ii + 1; break;
}
}
- OS << "case ARM::BI__builtin_neon_" << MangleName(name, TypeVec[ti], ck)
+ OS << "case ARM::BI__builtin_neon_" << MangleName(name, TypeVec[ti], ck)
<< ": i = " << immidx << "; " << rangestr << "; break;\n";
}
}
OS << "#endif\n\n";
}
+
+/// GenTest - Write out a test for the intrinsic specified by the name and
+/// type strings, including the embedded patterns for FileCheck to match.
+static std::string GenTest(const std::string &name,
+ const std::string &proto,
+ StringRef outTypeStr, StringRef inTypeStr,
+ bool isShift) {
+ assert(!proto.empty() && "");
+ std::string s;
+
+ // Function name with type suffix
+ std::string mangledName = MangleName(name, outTypeStr, ClassS);
+ if (outTypeStr != inTypeStr) {
+ // If the input type is different (e.g., for vreinterpret), append a suffix
+ // for the input type. String off a "Q" (quad) prefix so that MangleName
+ // does not insert another "q" in the name.
+ unsigned typeStrOff = (inTypeStr[0] == 'Q' ? 1 : 0);
+ StringRef inTypeNoQuad = inTypeStr.substr(typeStrOff);
+ mangledName = MangleName(mangledName, inTypeNoQuad, ClassS);
+ }
+
+ // Emit the FileCheck patterns.
+ s += "// CHECK: test_" + mangledName + "\n";
+ // s += "// CHECK: \n"; // FIXME: + expected instruction opcode.
+
+ // Emit the start of the test function.
+ s += TypeString(proto[0], outTypeStr) + " test_" + mangledName + "(";
+ char arg = 'a';
+ std::string comma;
+ for (unsigned i = 1, e = proto.size(); i != e; ++i, ++arg) {
+ // Do not create arguments for values that must be immediate constants.
+ if (proto[i] == 'i')
+ continue;
+ s += comma + TypeString(proto[i], inTypeStr) + " ";
+ s.push_back(arg);
+ comma = ", ";
+ }
+ s += ") { \\\n ";
+
+ if (proto[0] != 'v')
+ s += "return ";
+ s += mangledName + "(";
+ arg = 'a';
+ for (unsigned i = 1, e = proto.size(); i != e; ++i, ++arg) {
+ if (proto[i] == 'i') {
+ // For immediate operands, test the maximum value.
+ if (isShift)
+ s += "1"; // FIXME
+ else
+ // The immediate generally refers to a lane in the preceding argument.
+ s += utostr(RangeFromType(proto[i-1], inTypeStr));
+ } else {
+ s.push_back(arg);
+ }
+ if ((i + 1) < e)
+ s += ", ";
+ }
+ s += ");\n}\n\n";
+ return s;
+}
+
+/// runTests - Write out a complete set of tests for all of the Neon
+/// intrinsics.
+void NeonEmitter::runTests(raw_ostream &OS) {
+ OS <<
+ "// RUN: %clang_cc1 -triple thumbv7-apple-darwin \\\n"
+ "// RUN: -target-cpu cortex-a9 -ffreestanding -S -o - %s | FileCheck %s\n"
+ "\n"
+ "#include <arm_neon.h>\n"
+ "\n";
+
+ std::vector<Record*> RV = Records.getAllDerivedDefinitions("Inst");
+ for (unsigned i = 0, e = RV.size(); i != e; ++i) {
+ Record *R = RV[i];
+ std::string name = R->getValueAsString("Name");
+ std::string Proto = R->getValueAsString("Prototype");
+ std::string Types = R->getValueAsString("Types");
+ bool isShift = R->getValueAsBit("isShift");
+
+ SmallVector<StringRef, 16> TypeVec;
+ ParseTypes(R, Types, TypeVec);
+
+ OpKind kind = OpMap[R->getValueAsDef("Operand")->getName()];
+ for (unsigned ti = 0, te = TypeVec.size(); ti != te; ++ti) {
+ if (kind == OpReinterpret) {
+ bool outQuad = false;
+ bool dummy = false;
+ (void)ClassifyType(TypeVec[ti], outQuad, dummy, dummy);
+ for (unsigned srcti = 0, srcte = TypeVec.size();
+ srcti != srcte; ++srcti) {
+ bool inQuad = false;
+ (void)ClassifyType(TypeVec[srcti], inQuad, dummy, dummy);
+ if (srcti == ti || inQuad != outQuad)
+ continue;
+ OS << GenTest(name, Proto, TypeVec[ti], TypeVec[srcti], isShift);
+ }
+ } else {
+ OS << GenTest(name, Proto, TypeVec[ti], TypeVec[ti], isShift);
+ }
+ }
+ OS << "\n";
+ }
+}
+
projects / oota-llvm.git / blobdiff