Google Git
Sign in
llvm / llvm-archive / 9ee4ded126524dc671b998a7c57f4d2433ee9d20 / . / safecode / tools / clang / utils / TableGen / NeonEmitter.cpp
blob: 34b955e8e9d280d917ecf683999dcd944361df69 [file] [log] [blame]
//===- NeonEmitter.cpp - Generate arm_neon.h for use with clang -*- C++ -*-===//
//
// The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
// This tablegen backend is responsible for emitting arm_neon.h, which includes
// 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. A complete set of tests
// for Neon intrinsics can be generated by calling the runTests() entry point.
//
//===----------------------------------------------------------------------===//
#include "llvm/ADT/DenseMap.h"
#include "llvm/ADT/SmallString.h"
#include "llvm/ADT/SmallVector.h"
#include "llvm/ADT/StringExtras.h"
#include "llvm/ADT/StringMap.h"
#include "llvm/Support/ErrorHandling.h"
#include "llvm/TableGen/Error.h"
#include "llvm/TableGen/Record.h"
#include "llvm/TableGen/TableGenBackend.h"
#include <string>
using namespace llvm;
enum OpKind {
OpNone,
OpUnavailable,
OpAdd,
OpAddl,
OpAddw,
OpSub,
OpSubl,
OpSubw,
OpMul,
OpMla,
OpMlal,
OpMls,
OpMlsl,
OpMulN,
OpMlaN,
OpMlsN,
OpMlalN,
OpMlslN,
OpMulLane,
OpMullLane,
OpMlaLane,
OpMlsLane,
OpMlalLane,
OpMlslLane,
OpQDMullLane,
OpQDMlalLane,
OpQDMlslLane,
OpQDMulhLane,
OpQRDMulhLane,
OpEq,
OpGe,
OpLe,
OpGt,
OpLt,
OpNeg,
OpNot,
OpAnd,
OpOr,
OpXor,
OpAndNot,
OpOrNot,
OpCast,
OpConcat,
OpDup,
OpDupLane,
OpHi,
OpLo,
OpSelect,
OpRev16,
OpRev32,
OpRev64,
OpReinterpret,
OpAbdl,
OpAba,
OpAbal
};
enum ClassKind {
ClassNone,
ClassI, // generic integer instruction, e.g., "i8" suffix
ClassS, // signed/unsigned/poly, e.g., "s8", "u8" or "p8" suffix
ClassW, // width-specific instruction, e.g., "8" suffix
ClassB, // bitcast arguments with enum argument to specify type
ClassL, // Logical instructions which are op instructions
// but we need to not emit any suffix for in our
// tests.
ClassNoTest // Instructions which we do not test since they are
// not TRUE instructions.
};
/// NeonTypeFlags - Flags to identify the types for overloaded Neon
/// builtins. These must be kept in sync with the flags in
/// include/clang/Basic/TargetBuiltins.h.
namespace {
class NeonTypeFlags {
enum {
EltTypeMask = 0xf,
UnsignedFlag = 0x10,
QuadFlag = 0x20
};
uint32_t Flags;
public:
enum EltType {
Int8,
Int16,
Int32,
Int64,
Poly8,
Poly16,
Float16,
Float32
};
NeonTypeFlags(unsigned F) : Flags(F) {}
NeonTypeFlags(EltType ET, bool IsUnsigned, bool IsQuad) : Flags(ET) {
if (IsUnsigned)
Flags |= UnsignedFlag;
if (IsQuad)
Flags |= QuadFlag;
}
uint32_t getFlags() const { return Flags; }
};
} // end anonymous namespace
namespace {
class NeonEmitter {
RecordKeeper &Records;
StringMap<OpKind> OpMap;
DenseMap<Record*, ClassKind> ClassMap;
public:
NeonEmitter(RecordKeeper &R) : Records(R) {
OpMap["OP_NONE"] = OpNone;
OpMap["OP_UNAVAILABLE"] = OpUnavailable;
OpMap["OP_ADD"] = OpAdd;
OpMap["OP_ADDL"] = OpAddl;
OpMap["OP_ADDW"] = OpAddw;
OpMap["OP_SUB"] = OpSub;
OpMap["OP_SUBL"] = OpSubl;
OpMap["OP_SUBW"] = OpSubw;
OpMap["OP_MUL"] = OpMul;
OpMap["OP_MLA"] = OpMla;
OpMap["OP_MLAL"] = OpMlal;
OpMap["OP_MLS"] = OpMls;
OpMap["OP_MLSL"] = OpMlsl;
OpMap["OP_MUL_N"] = OpMulN;
OpMap["OP_MLA_N"] = OpMlaN;
OpMap["OP_MLS_N"] = OpMlsN;
OpMap["OP_MLAL_N"] = OpMlalN;
OpMap["OP_MLSL_N"] = OpMlslN;
OpMap["OP_MUL_LN"]= OpMulLane;
OpMap["OP_MULL_LN"] = OpMullLane;
OpMap["OP_MLA_LN"]= OpMlaLane;
OpMap["OP_MLS_LN"]= OpMlsLane;
OpMap["OP_MLAL_LN"] = OpMlalLane;
OpMap["OP_MLSL_LN"] = OpMlslLane;
OpMap["OP_QDMULL_LN"] = OpQDMullLane;
OpMap["OP_QDMLAL_LN"] = OpQDMlalLane;
OpMap["OP_QDMLSL_LN"] = OpQDMlslLane;
OpMap["OP_QDMULH_LN"] = OpQDMulhLane;
OpMap["OP_QRDMULH_LN"] = OpQRDMulhLane;
OpMap["OP_EQ"] = OpEq;
OpMap["OP_GE"] = OpGe;
OpMap["OP_LE"] = OpLe;
OpMap["OP_GT"] = OpGt;
OpMap["OP_LT"] = OpLt;
OpMap["OP_NEG"] = OpNeg;
OpMap["OP_NOT"] = OpNot;
OpMap["OP_AND"] = OpAnd;
OpMap["OP_OR"] = OpOr;
OpMap["OP_XOR"] = OpXor;
OpMap["OP_ANDN"] = OpAndNot;
OpMap["OP_ORN"] = OpOrNot;
OpMap["OP_CAST"] = OpCast;
OpMap["OP_CONC"] = OpConcat;
OpMap["OP_HI"] = OpHi;
OpMap["OP_LO"] = OpLo;
OpMap["OP_DUP"] = OpDup;
OpMap["OP_DUP_LN"] = OpDupLane;
OpMap["OP_SEL"] = OpSelect;
OpMap["OP_REV16"] = OpRev16;
OpMap["OP_REV32"] = OpRev32;
OpMap["OP_REV64"] = OpRev64;
OpMap["OP_REINT"] = OpReinterpret;
OpMap["OP_ABDL"] = OpAbdl;
OpMap["OP_ABA"] = OpAba;
OpMap["OP_ABAL"] = OpAbal;
Record *SI = R.getClass("SInst");
Record *II = R.getClass("IInst");
Record *WI = R.getClass("WInst");
Record *SOpI = R.getClass("SOpInst");
Record *IOpI = R.getClass("IOpInst");
Record *WOpI = R.getClass("WOpInst");
Record *LOpI = R.getClass("LOpInst");
Record *NoTestOpI = R.getClass("NoTestOpInst");
ClassMap[SI] = ClassS;
ClassMap[II] = ClassI;
ClassMap[WI] = ClassW;
ClassMap[SOpI] = ClassS;
ClassMap[IOpI] = ClassI;
ClassMap[WOpI] = ClassW;
ClassMap[LOpI] = ClassL;
ClassMap[NoTestOpI] = ClassNoTest;
}
// run - Emit arm_neon.h.inc
void run(raw_ostream &o);
// runHeader - Emit all the __builtin prototypes used in arm_neon.h
void runHeader(raw_ostream &o);
// runTests - Emit tests for all the Neon intrinsics.
void runTests(raw_ostream &o);
private:
void emitIntrinsic(raw_ostream &OS, Record *R);
};
} // end anonymous namespace
/// 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();
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':
case 'i':
case 'l':
case 'h':
case 'f':
break;
default:
PrintFatalError(r->getLoc(),
"Unexpected letter: " + std::string(data + len, 1));
}
TV.push_back(StringRef(data, len + 1));
data += len + 1;
len = -1;
}
}
/// 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 's';
case 's':
return 'i';
case 'i':
return 'l';
case 'h':
return 'f';
default:
PrintFatalError("unhandled type in widen!");
}
}
/// 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 'c';
case 'i':
return 's';
case 'l':
return 'i';
case 'f':
return 'h';
default:
PrintFatalError("unhandled type in narrow!");
}
}
/// 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;
// 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];
}
/// 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) {
case 't':
if (poly) {
poly = false;
usgn = true;
}
break;
case 'u':
usgn = true;
poly = false;
if (type == 'f')
type = 'i';
break;
case 'x':
usgn = false;
poly = false;
if (type == 'f')
type = 'i';
break;
case 'f':
if (type == 'h')
quad = true;
type = 'f';
usgn = false;
break;
case 'g':
quad = false;
break;
case 'w':
type = Widen(type);
quad = true;
break;
case 'n':
type = Widen(type);
break;
case 'i':
type = 'i';
scal = true;
break;
case 'l':
type = 'l';
scal = true;
usgn = true;
break;
case 's':
case 'a':
scal = true;
break;
case 'k':
quad = true;
break;
case 'c':
cnst = true;
case 'p':
pntr = true;
scal = true;
break;
case 'h':
type = Narrow(type);
if (type == 'h')
quad = false;
break;
case 'e':
type = Narrow(type);
usgn = true;
break;
default:
break;
}
return type;
}
/// 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;
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";
if (scal)
break;
s += quad ? "x16" : "x8";
break;
case 's':
s += poly ? "poly16" : "int16";
if (scal)
break;
s += quad ? "x8" : "x4";
break;
case 'i':
s += "int32";
if (scal)
break;
s += quad ? "x4" : "x2";
break;
case 'l':
s += "int64";
if (scal)
break;
s += quad ? "x2" : "x1";
break;
case 'h':
s += "float16";
if (scal)
break;
s += quad ? "x8" : "x4";
break;
case 'f':
s += "float32";
if (scal)
break;
s += quad ? "x4" : "x2";
break;
default:
PrintFatalError("unhandled type!");
}
if (mod == '2')
s += "x2";
if (mod == '3')
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();
}
/// 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;
bool poly = false;
bool usgn = false;
bool scal = false;
bool cnst = false;
bool pntr = false;
if (mod == 'v')
return "v"; // void
if (mod == '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;
}
usgn = usgn | poly | ((ck == ClassI || ck == ClassW) && scal && type != 'f');
if (scal) {
SmallString<128> s;
if (usgn)
s.push_back('U');
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)
s.push_back('*');
return s.str();
}
// Since the return value must be one type, return a vector type of the
// 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' && 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 ? "V8s" : "V4s";
if (ck != ClassB && type == 'i')
return quad ? "V4i" : "V2i";
if (ck != ClassB && type == 'l')
return quad ? "V2LLi" : "V1LLi";
return quad ? "V16Sc" : "V8Sc";
}
// Non-return array types are passed as individual vectors.
if (mod == '2')
return quad ? "V16ScV16Sc" : "V8ScV8Sc";
if (mod == '3')
return quad ? "V16ScV16ScV16Sc" : "V8ScV8ScV8Sc";
if (mod == '4')
return quad ? "V16ScV16ScV16ScV16Sc" : "V8ScV8ScV8ScV8Sc";
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 ? "V16Sc" : "V8Sc";
}
/// InstructionTypeCode - Computes the ARM argument character code and
/// quad status for a specific type string and ClassKind.
static void InstructionTypeCode(const StringRef &typeStr,
const ClassKind ck,
bool &quad,
std::string &typeCode) {
bool poly = false;
bool usgn = false;
char type = ClassifyType(typeStr, quad, poly, usgn);
switch (type) {
case 'c':
switch (ck) {
case ClassS: typeCode = poly ? "p8" : usgn ? "u8" : "s8"; break;
case ClassI: typeCode = "i8"; break;
case ClassW: typeCode = "8"; break;
default: break;
}
break;
case 's':
switch (ck) {
case ClassS: typeCode = poly ? "p16" : usgn ? "u16" : "s16"; break;
case ClassI: typeCode = "i16"; break;
case ClassW: typeCode = "16"; break;
default: break;
}
break;
case 'i':
switch (ck) {
case ClassS: typeCode = usgn ? "u32" : "s32"; break;
case ClassI: typeCode = "i32"; break;
case ClassW: typeCode = "32"; break;
default: break;
}
break;
case 'l':
switch (ck) {
case ClassS: typeCode = usgn ? "u64" : "s64"; break;
case ClassI: typeCode = "i64"; break;
case ClassW: typeCode = "64"; break;
default: break;
}
break;
case 'h':
switch (ck) {
case ClassS:
case ClassI: typeCode = "f16"; break;
case ClassW: typeCode = "16"; break;
default: break;
}
break;
case 'f':
switch (ck) {
case ClassS:
case ClassI: typeCode = "f32"; break;
case ClassW: typeCode = "32"; break;
default: break;
}
break;
default:
PrintFatalError("unhandled type!");
}
}
/// 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;
std::string typeCode = "";
InstructionTypeCode(typestr, ck, quad, typeCode);
std::string s = name;
if (typeCode.size() > 0) {
s += "_" + typeCode;
}
if (ck == ClassB)
s += "_v";
// 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('_');
s = s.insert(pos, "q");
}
return s;
}
static void PreprocessInstruction(const StringRef &Name,
const std::string &InstName,
std::string &Prefix,
bool &HasNPostfix,
bool &HasLanePostfix,
bool &HasDupPostfix,
bool &IsSpecialVCvt,
size_t &TBNumber) {
// All of our instruction name fields from arm_neon.td are of the form
// <instructionname>_...
// Thus we grab our instruction name via computation of said Prefix.
const size_t PrefixEnd = Name.find_first_of('_');
// If InstName is passed in, we use that instead of our name Prefix.
Prefix = InstName.size() == 0? Name.slice(0, PrefixEnd).str() : InstName;
const StringRef Postfix = Name.slice(PrefixEnd, Name.size());
HasNPostfix = Postfix.count("_n");
HasLanePostfix = Postfix.count("_lane");
HasDupPostfix = Postfix.count("_dup");
IsSpecialVCvt = Postfix.size() != 0 && Name.count("vcvt");
if (InstName.compare("vtbl") == 0 ||
InstName.compare("vtbx") == 0) {
// If we have a vtblN/vtbxN instruction, use the instruction's ASCII
// encoding to get its true value.
TBNumber = Name[Name.size()-1] - 48;
}
}
/// GenerateRegisterCheckPatternsForLoadStores - Given a bunch of data we have
/// extracted, generate a FileCheck pattern for a Load Or Store
static void
GenerateRegisterCheckPatternForLoadStores(const StringRef &NameRef,
const std::string& OutTypeCode,
const bool &IsQuad,
const bool &HasDupPostfix,
const bool &HasLanePostfix,
const size_t Count,
std::string &RegisterSuffix) {
const bool IsLDSTOne = NameRef.count("vld1") || NameRef.count("vst1");
// If N == 3 || N == 4 and we are dealing with a quad instruction, Clang
// will output a series of v{ld,st}1s, so we have to handle it specially.
if ((Count == 3 || Count == 4) && IsQuad) {
RegisterSuffix += "{";
for (size_t i = 0; i < Count; i++) {
RegisterSuffix += "d{{[0-9]+}}";
if (HasDupPostfix) {
RegisterSuffix += "[]";
}
if (HasLanePostfix) {
RegisterSuffix += "[{{[0-9]+}}]";
}
if (i < Count-1) {
RegisterSuffix += ", ";
}
}
RegisterSuffix += "}";
} else {
// Handle normal loads and stores.
RegisterSuffix += "{";
for (size_t i = 0; i < Count; i++) {
RegisterSuffix += "d{{[0-9]+}}";
if (HasDupPostfix) {
RegisterSuffix += "[]";
}
if (HasLanePostfix) {
RegisterSuffix += "[{{[0-9]+}}]";
}
if (IsQuad && !HasLanePostfix) {
RegisterSuffix += ", d{{[0-9]+}}";
if (HasDupPostfix) {
RegisterSuffix += "[]";
}
}
if (i < Count-1) {
RegisterSuffix += ", ";
}
}
RegisterSuffix += "}, [r{{[0-9]+}}";
// We only include the alignment hint if we have a vld1.*64 or
// a dup/lane instruction.
if (IsLDSTOne) {
if ((HasLanePostfix || HasDupPostfix) && OutTypeCode != "8") {
RegisterSuffix += ", :" + OutTypeCode;
} else if (OutTypeCode == "64") {
RegisterSuffix += ", :64";
}
}
RegisterSuffix += "]";
}
}
static bool HasNPostfixAndScalarArgs(const StringRef &NameRef,
const bool &HasNPostfix) {
return (NameRef.count("vmla") ||
NameRef.count("vmlal") ||
NameRef.count("vmlsl") ||
NameRef.count("vmull") ||
NameRef.count("vqdmlal") ||
NameRef.count("vqdmlsl") ||
NameRef.count("vqdmulh") ||
NameRef.count("vqdmull") ||
NameRef.count("vqrdmulh")) && HasNPostfix;
}
static bool IsFiveOperandLaneAccumulator(const StringRef &NameRef,
const bool &HasLanePostfix) {
return (NameRef.count("vmla") ||
NameRef.count("vmls") ||
NameRef.count("vmlal") ||
NameRef.count("vmlsl") ||
(NameRef.count("vmul") && NameRef.size() == 3)||
NameRef.count("vqdmlal") ||
NameRef.count("vqdmlsl") ||
NameRef.count("vqdmulh") ||
NameRef.count("vqrdmulh")) && HasLanePostfix;
}
static bool IsSpecialLaneMultiply(const StringRef &NameRef,
const bool &HasLanePostfix,
const bool &IsQuad) {
const bool IsVMulOrMulh = (NameRef.count("vmul") || NameRef.count("mulh"))
&& IsQuad;
const bool IsVMull = NameRef.count("mull") && !IsQuad;
return (IsVMulOrMulh || IsVMull) && HasLanePostfix;
}
static void NormalizeProtoForRegisterPatternCreation(const std::string &Name,
const std::string &Proto,
const bool &HasNPostfix,
const bool &IsQuad,
const bool &HasLanePostfix,
const bool &HasDupPostfix,
std::string &NormedProto) {
// Handle generic case.
const StringRef NameRef(Name);
for (size_t i = 0, end = Proto.size(); i < end; i++) {
switch (Proto[i]) {
case 'u':
case 'f':
case 'd':
case 's':
case 'x':
case 't':
case 'n':
NormedProto += IsQuad? 'q' : 'd';
break;
case 'w':
case 'k':
NormedProto += 'q';
break;
case 'g':
case 'h':
case 'e':
NormedProto += 'd';
break;
case 'i':
NormedProto += HasLanePostfix? 'a' : 'i';
break;
case 'a':
if (HasLanePostfix) {
NormedProto += 'a';
} else if (HasNPostfixAndScalarArgs(NameRef, HasNPostfix)) {
NormedProto += IsQuad? 'q' : 'd';
} else {
NormedProto += 'i';
}
break;
}
}
// Handle Special Cases.
const bool IsNotVExt = !NameRef.count("vext");
const bool IsVPADAL = NameRef.count("vpadal");
const bool Is5OpLaneAccum = IsFiveOperandLaneAccumulator(NameRef,
HasLanePostfix);
const bool IsSpecialLaneMul = IsSpecialLaneMultiply(NameRef, HasLanePostfix,
IsQuad);
if (IsSpecialLaneMul) {
// If
NormedProto[2] = NormedProto[3];
NormedProto.erase(3);
} else if (NormedProto.size() == 4 &&
NormedProto[0] == NormedProto[1] &&
IsNotVExt) {
// If NormedProto.size() == 4 and the first two proto characters are the
// same, ignore the first.
NormedProto = NormedProto.substr(1, 3);
} else if (Is5OpLaneAccum) {
// If we have a 5 op lane accumulator operation, we take characters 1,2,4
std::string tmp = NormedProto.substr(1,2);
tmp += NormedProto[4];
NormedProto = tmp;
} else if (IsVPADAL) {
// If we have VPADAL, ignore the first character.
NormedProto = NormedProto.substr(0, 2);
} else if (NameRef.count("vdup") && NormedProto.size() > 2) {
// If our instruction is a dup instruction, keep only the first and
// last characters.
std::string tmp = "";
tmp += NormedProto[0];
tmp += NormedProto[NormedProto.size()-1];
NormedProto = tmp;
}
}
/// GenerateRegisterCheckPatterns - Given a bunch of data we have
/// extracted, generate a FileCheck pattern to check that an
/// instruction's arguments are correct.
static void GenerateRegisterCheckPattern(const std::string &Name,
const std::string &Proto,
const std::string &OutTypeCode,
const bool &HasNPostfix,
const bool &IsQuad,
const bool &HasLanePostfix,
const bool &HasDupPostfix,
const size_t &TBNumber,
std::string &RegisterSuffix) {
RegisterSuffix = "";
const StringRef NameRef(Name);
const StringRef ProtoRef(Proto);
if ((NameRef.count("vdup") || NameRef.count("vmov")) && HasNPostfix) {
return;
}
const bool IsLoadStore = NameRef.count("vld") || NameRef.count("vst");
const bool IsTBXOrTBL = NameRef.count("vtbl") || NameRef.count("vtbx");
if (IsLoadStore) {
// Grab N value from v{ld,st}N using its ascii representation.
const size_t Count = NameRef[3] - 48;
GenerateRegisterCheckPatternForLoadStores(NameRef, OutTypeCode, IsQuad,
HasDupPostfix, HasLanePostfix,
Count, RegisterSuffix);
} else if (IsTBXOrTBL) {
RegisterSuffix += "d{{[0-9]+}}, {";
for (size_t i = 0; i < TBNumber-1; i++) {
RegisterSuffix += "d{{[0-9]+}}, ";
}
RegisterSuffix += "d{{[0-9]+}}}, d{{[0-9]+}}";
} else {
// Handle a normal instruction.
if (NameRef.count("vget") || NameRef.count("vset"))
return;
// We first normalize our proto, since we only need to emit 4
// different types of checks, yet have more than 4 proto types
// that map onto those 4 patterns.
std::string NormalizedProto("");
NormalizeProtoForRegisterPatternCreation(Name, Proto, HasNPostfix, IsQuad,
HasLanePostfix, HasDupPostfix,
NormalizedProto);
for (size_t i = 0, end = NormalizedProto.size(); i < end; i++) {
const char &c = NormalizedProto[i];
switch (c) {
case 'q':
RegisterSuffix += "q{{[0-9]+}}, ";
break;
case 'd':
RegisterSuffix += "d{{[0-9]+}}, ";
break;
case 'i':
RegisterSuffix += "#{{[0-9]+}}, ";
break;
case 'a':
RegisterSuffix += "d{{[0-9]+}}[{{[0-9]}}], ";
break;
}
}
// Remove extra ", ".
RegisterSuffix = RegisterSuffix.substr(0, RegisterSuffix.size()-2);
}
}
/// GenerateChecksForIntrinsic - Given a specific instruction name +
/// typestr + class kind, generate the proper set of FileCheck
/// Patterns to check for. We could just return a string, but instead
/// use a vector since it provides us with the extra flexibility of
/// emitting multiple checks, which comes in handy for certain cases
/// like mla where we want to check for 2 different instructions.
static void GenerateChecksForIntrinsic(const std::string &Name,
const std::string &Proto,
StringRef &OutTypeStr,
StringRef &InTypeStr,
ClassKind Ck,
const std::string &InstName,
bool IsHiddenLOp,
std::vector<std::string>& Result) {
// If Ck is a ClassNoTest instruction, just return so no test is
// emitted.
if(Ck == ClassNoTest)
return;
if (Name == "vcvt_f32_f16") {
Result.push_back("vcvt.f32.f16");
return;
}
// Now we preprocess our instruction given the data we have to get the
// data that we need.
// Create a StringRef for String Manipulation of our Name.
const StringRef NameRef(Name);
// Instruction Prefix.
std::string Prefix;
// The type code for our out type string.
std::string OutTypeCode;
// To handle our different cases, we need to check for different postfixes.
// Is our instruction a quad instruction.
bool IsQuad = false;
// Our instruction is of the form <instructionname>_n.
bool HasNPostfix = false;
// Our instruction is of the form <instructionname>_lane.
bool HasLanePostfix = false;
// Our instruction is of the form <instructionname>_dup.
bool HasDupPostfix = false;
// Our instruction is a vcvt instruction which requires special handling.
bool IsSpecialVCvt = false;
// If we have a vtbxN or vtblN instruction, this is set to N.
size_t TBNumber = -1;
// Register Suffix
std::string RegisterSuffix;
PreprocessInstruction(NameRef, InstName, Prefix,
HasNPostfix, HasLanePostfix, HasDupPostfix,
IsSpecialVCvt, TBNumber);
InstructionTypeCode(OutTypeStr, Ck, IsQuad, OutTypeCode);
GenerateRegisterCheckPattern(Name, Proto, OutTypeCode, HasNPostfix, IsQuad,
HasLanePostfix, HasDupPostfix, TBNumber,
RegisterSuffix);
// In the following section, we handle a bunch of special cases. You can tell
// a special case by the fact we are returning early.
// If our instruction is a logical instruction without postfix or a
// hidden LOp just return the current Prefix.
if (Ck == ClassL || IsHiddenLOp) {
Result.push_back(Prefix + " " + RegisterSuffix);
return;
}
// If we have a vmov, due to the many different cases, some of which
// vary within the different intrinsics generated for a single
// instruction type, just output a vmov. (e.g. given an instruction
// A, A.u32 might be vmov and A.u8 might be vmov.8).
//
// FIXME: Maybe something can be done about this. The two cases that we care
// about are vmov as an LType and vmov as a WType.
if (Prefix == "vmov") {
Result.push_back(Prefix + " " + RegisterSuffix);
return;
}
// In the following section, we handle special cases.
if (OutTypeCode == "64") {
// If we have a 64 bit vdup/vext and are handling an uint64x1_t
// type, the intrinsic will be optimized away, so just return
// nothing. On the other hand if we are handling an uint64x2_t
// (i.e. quad instruction), vdup/vmov instructions should be
// emitted.
if (Prefix == "vdup" || Prefix == "vext") {
if (IsQuad) {
Result.push_back("{{vmov|vdup}}");
}
return;
}
// v{st,ld}{2,3,4}_{u,s}64 emit v{st,ld}1.64 instructions with
// multiple register operands.
bool MultiLoadPrefix = Prefix == "vld2" || Prefix == "vld3"
|| Prefix == "vld4";
bool MultiStorePrefix = Prefix == "vst2" || Prefix == "vst3"
|| Prefix == "vst4";
if (MultiLoadPrefix || MultiStorePrefix) {
Result.push_back(NameRef.slice(0, 3).str() + "1.64");
return;
}
// v{st,ld}1_{lane,dup}_{u64,s64} use vldr/vstr/vmov/str instead of
// emitting said instructions. So return a check for
// vldr/vstr/vmov/str instead.
if (HasLanePostfix || HasDupPostfix) {
if (Prefix == "vst1") {
Result.push_back("{{str|vstr|vmov}}");
return;
} else if (Prefix == "vld1") {
Result.push_back("{{ldr|vldr|vmov}}");
return;
}
}
}
// vzip.32/vuzp.32 are the same instruction as vtrn.32 and are
// sometimes disassembled as vtrn.32. We use a regex to handle both
// cases.
if ((Prefix == "vzip" || Prefix == "vuzp") && OutTypeCode == "32") {
Result.push_back("{{vtrn|" + Prefix + "}}.32 " + RegisterSuffix);
return;
}
// Currently on most ARM processors, we do not use vmla/vmls for
// quad floating point operations. Instead we output vmul + vadd. So
// check if we have one of those instructions and just output a
// check for vmul.
if (OutTypeCode == "f32") {
if (Prefix == "vmls") {
Result.push_back("vmul." + OutTypeCode + " " + RegisterSuffix);
Result.push_back("vsub." + OutTypeCode);
return;
} else if (Prefix == "vmla") {
Result.push_back("vmul." + OutTypeCode + " " + RegisterSuffix);
Result.push_back("vadd." + OutTypeCode);
return;
}
}
// If we have vcvt, get the input type from the instruction name
// (which should be of the form instname_inputtype) and append it
// before the output type.
if (Prefix == "vcvt") {
const std::string inTypeCode = NameRef.substr(NameRef.find_last_of("_")+1);
Prefix += "." + inTypeCode;
}
// Append output type code to get our final mangled instruction.
Prefix += "." + OutTypeCode;
Result.push_back(Prefix + " " + RegisterSuffix);
}
/// 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) {
// True for constant ints (i), pointers (p) and const pointers (c).
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 = UseMacro(proto);
char arg = 'a';
std::string s;
s += "(";
for (unsigned i = 1, e = proto.size(); i != e; ++i, ++arg) {
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;
}
// 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 (MacroArgUsedDirectly(proto, i))
continue;
generatedLocal = true;
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;
if ((i + 1) < nElts)
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;
}
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;
default:
PrintFatalError("unhandled type!");
}
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);
bool define = UseMacro(proto);
std::string ts = TypeString(proto[0], typestr);
std::string s;
if (!define) {
s = "return ";
}
switch(op) {
case OpAdd:
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;";
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") + ";";
break;
case OpMulLane:
s += "__a * " + SplatLane(nElts, "__b", "__c") + ";";
break;
case OpMul:
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") + ");";
break;
case OpMlaLane:
s += "__a + (__b * " + SplatLane(nElts, "__c", "__d") + ");";
break;
case OpMla:
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") + ");";
break;
case OpMlsLane:
s += "__a - (__b * " + SplatLane(nElts, "__c", "__d") + ");";
break;
case OpMls:
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);";
break;
case OpGe:
s += "(" + ts + ")(__a >= __b);";
break;
case OpLe:
s += "(" + ts + ")(__a <= __b);";
break;
case OpGt:
s += "(" + ts + ")(__a > __b);";
break;
case OpLt:
s += "(" + ts + ")(__a < __b);";
break;
case OpNeg:
s += " -__a;";
break;
case OpNot:
s += " ~__a;";
break;
case OpAnd:
s += "__a & __b;";
break;
case OpOr:
s += "__a | __b;";
break;
case OpXor:
s += "__a ^ __b;";
break;
case OpAndNot:
s += "__a & ~__b;";
break;
case OpOrNot:
s += "__a | ~__b;";
break;
case OpCast:
s += "(" + ts + ")__a;";
break;
case OpConcat:
s += "(" + ts + ")__builtin_shufflevector((int64x1_t)__a";
s += ", (int64x1_t)__b, 0, 1);";
break;
case OpHi:
s += "(" + ts +
")__builtin_shufflevector((int64x2_t)__a, (int64x2_t)__a, 1);";
break;
case OpLo:
s += "(" + ts +
")__builtin_shufflevector((int64x2_t)__a, (int64x2_t)__a, 0);";
break;
case OpDup:
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));";
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;
}
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:
PrintFatalError("unknown OpKind!");
}
return s;
}
static unsigned GetNeonEnum(const std::string &proto, StringRef typestr) {
unsigned mod = proto[0];
if (mod == 'v' || mod == 'f')
mod = proto[1];
bool quad = false;
bool poly = false;
bool usgn = false;
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);
NeonTypeFlags::EltType ET;
switch (type) {
case 'c':
ET = poly ? NeonTypeFlags::Poly8 : NeonTypeFlags::Int8;
break;
case 's':
ET = poly ? NeonTypeFlags::Poly16 : NeonTypeFlags::Int16;
break;
case 'i':
ET = NeonTypeFlags::Int32;
break;
case 'l':
ET = NeonTypeFlags::Int64;
break;
case 'h':
ET = NeonTypeFlags::Float16;
break;
case 'f':
ET = NeonTypeFlags::Float32;
break;
default:
PrintFatalError("unhandled type!");
}
NeonTypeFlags Flags(ET, usgn, quad && proto[1] != 'g');
return Flags.getFlags();
}
// 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) {
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] <= '4');
bool define = UseMacro(proto);
// 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
s += "(" + ts + ")";
} else if (sret) {
s += ts + " r; ";
} else {
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 {
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, ";
char arg = 'a';
for (unsigned i = 1, e = proto.size(); i != e; ++i, ++arg) {
std::string args = std::string(&arg, 1);
// 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] <= '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)
s += ", ";
}
if ((i + 1) < e)
s += ", ";
continue;
}
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));
s += ");";
if (proto[0] != 'v' && sret) {
if (define)
s += " r;";
else
s += " return r;";
}
return s;
}
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
// 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) && kind != OpUnavailable;
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 if (kind == OpUnavailable) {
s += " __attribute__((unavailable));\n";
return s;
} 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) {
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 << "#include <stdint.h>\n\n";
// Emit NEON-specific scalar typedefs.
OS << "typedef float float32_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.
std::string TypedefTypes("cQcsQsiQilQlUcQUcUsQUsUiQUiUlQUlhQhfQfPcQPcPsQPs");
SmallVector<StringRef, 24> TDTypeVec;
ParseTypes(0, TypedefTypes, TDTypeVec);
// Emit vector typedefs.
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(";
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";
// Emit struct typedefs.
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 = TypeString('0' + vi, TDTypeVec[i]);
OS << "typedef struct " << vs << " {\n";
OS << " " << ts << " val";
OS << "[" << utostr(vi) << "]";
OS << ";\n} ";
OS << vs << ";\n\n";
}
}
OS<<"#define __ai static inline __attribute__((__always_inline__, __nodebug__))\n\n";
std::vector<Record*> RV = Records.getAllDerivedDefinitions("Inst");
// 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];
if (R->getName() != "VMOVL" &&
R->getName() != "VMULL" &&
R->getName() != "VABD")
emitIntrinsic(OS, R);
}
OS << "#undef __ai\n\n";
OS << "#endif /* __ARM_NEON_H */\n";
}
/// 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)
PrintFatalError(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;
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:
PrintFatalError("unhandled type!");
}
}
/// 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 immediate 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 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)
PrintFatalError(R->getLoc(), "Builtin has no class kind");
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
// stream.
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) {
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 = 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)
PrintFatalError(R->getLoc(), "Builtin has no class kind");
int si = -1, qi = -1;
uint64_t 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 |= 1ULL << GetNeonEnum(Proto, TypeVec[ti]);
} else {
si = ti;
mask |= 1ULL << GetNeonEnum(Proto, TypeVec[ti]);
}
}
// Check if the builtin function has a pointer or const pointer argument.
int PtrArgNum = -1;
bool HasConstPtr = false;
for (unsigned arg = 1, arge = Proto.size(); arg != arge; ++arg) {
char ArgType = Proto[arg];
if (ArgType == 'c') {
HasConstPtr = true;
PtrArgNum = arg - 1;
break;
}
if (ArgType == 'p') {
PtrArgNum = arg - 1;
break;
}
}
// For sret builtins, adjust the pointer argument index.
if (PtrArgNum >= 0 && (Proto[0] >= '2' && Proto[0] <= '4'))
PtrArgNum += 1;
// Omit type checking for the pointer arguments of vld1_lane, vld1_dup,
// and vst1_lane intrinsics. Using a pointer to the vector element
// type with one of those operations causes codegen to select an aligned
// load/store instruction. If you want an unaligned operation,
// the pointer argument needs to have less alignment than element type,
// so just accept any pointer type.
if (name == "vld1_lane" || name == "vld1_dup" || name == "vst1_lane") {
PtrArgNum = -1;
HasConstPtr = false;
}
if (mask) {
OS << "case ARM::BI__builtin_neon_"
<< MangleName(name, TypeVec[si], ClassB)
<< ": mask = " << "0x" << utohexstr(mask) << "ULL";
if (PtrArgNum >= 0)
OS << "; PtrArgNum = " << PtrArgNum;
if (HasConstPtr)
OS << "; HasConstPtr = true";
OS << "; break;\n";
}
if (qmask) {
OS << "case ARM::BI__builtin_neon_"
<< MangleName(name, TypeVec[qi], ClassB)
<< ": mask = " << "0x" << utohexstr(qmask) << "ULL";
if (PtrArgNum >= 0)
OS << "; PtrArgNum = " << PtrArgNum;
if (HasConstPtr)
OS << "; HasConstPtr = true";
OS << "; 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 = 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.
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)
PrintFatalError(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;
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 {
// 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);
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] <= '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";
}
/// 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, bool isHiddenLOp,
ClassKind ck, const std::string &InstName) {
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);
}
std::vector<std::string> FileCheckPatterns;
GenerateChecksForIntrinsic(name, proto, outTypeStr, inTypeStr, ck, InstName,
isHiddenLOp, FileCheckPatterns);
// Emit the FileCheck patterns.
s += "// CHECK: test_" + mangledName + "\n";
// If for any reason we do not want to emit a check, mangledInst
// will be the empty string.
if (FileCheckPatterns.size()) {
for (std::vector<std::string>::const_iterator i = FileCheckPatterns.begin(),
e = FileCheckPatterns.end();
i != e;
++i) {
s += "// CHECK: " + *i + "\n";
}
}
// 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 thumbv7s-apple-darwin -target-abi apcs-gnu\\\n"
"// RUN: -target-cpu swift -ffreestanding -Os -S -o - %s\\\n"
"// RUN: | 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");
std::string InstName = R->getValueAsString("InstName");
bool isHiddenLOp = R->getValueAsBit("isHiddenLInst");
SmallVector<StringRef, 16> TypeVec;
ParseTypes(R, Types, TypeVec);
ClassKind ck = ClassMap[R->getSuperClasses()[1]];
OpKind kind = OpMap[R->getValueAsDef("Operand")->getName()];
if (kind == OpUnavailable)
continue;
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, isHiddenLOp, ck, InstName);
}
} else {
OS << GenTest(name, Proto, TypeVec[ti], TypeVec[ti],
isShift, isHiddenLOp, ck, InstName);
}
}
OS << "\n";
}
}
namespace clang {
void EmitNeon(RecordKeeper &Records, raw_ostream &OS) {
NeonEmitter(Records).run(OS);
}
void EmitNeonSema(RecordKeeper &Records, raw_ostream &OS) {
NeonEmitter(Records).runHeader(OS);
}
void EmitNeonTest(RecordKeeper &Records, raw_ostream &OS) {
NeonEmitter(Records).runTests(OS);
}
} // End namespace clang