1 //===- NeonEmitter.cpp - Generate arm_neon.h for use with clang -*- C++ -*-===//
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
7 //===----------------------------------------------------------------------===//
9 // This tablegen backend is responsible for emitting arm_neon.h, which includes
10 // a declaration and definition of each function specified by the ARM NEON
11 // compiler interface. See ARM document DUI0348B.
13 // Each NEON instruction is implemented in terms of 1 or more functions which
14 // are suffixed with the element type of the input vectors. Functions may be
15 // implemented in terms of generic vector operations such as +, *, -, etc. or
16 // by calling a __builtin_-prefixed function which will be handled by clang's
19 // Additional validation code can be generated by this file when runHeader() is
20 // called, rather than the normal run() entry point.
22 // See also the documentation in include/clang/Basic/arm_neon.td.
24 //===----------------------------------------------------------------------===//
26 #include "TableGenBackends.h"
27 #include "llvm/ADT/ArrayRef.h"
28 #include "llvm/ADT/DenseMap.h"
29 #include "llvm/ADT/None.h"
30 #include "llvm/ADT/SmallVector.h"
31 #include "llvm/ADT/STLExtras.h"
32 #include "llvm/ADT/StringExtras.h"
33 #include "llvm/ADT/StringRef.h"
34 #include "llvm/Support/Casting.h"
35 #include "llvm/Support/ErrorHandling.h"
36 #include "llvm/Support/raw_ostream.h"
37 #include "llvm/TableGen/Error.h"
38 #include "llvm/TableGen/Record.h"
39 #include "llvm/TableGen/SetTheory.h"
57 // While globals are generally bad, this one allows us to perform assertions
58 // liberally and somehow still trace them back to the def they indirectly
60 static Record *CurrentRecord = nullptr;
61 static void assert_with_loc(bool Assertion, const std::string &Str) {
64 PrintFatalError(CurrentRecord->getLoc(), Str);
72 ClassI, // generic integer instruction, e.g., "i8" suffix
73 ClassS, // signed/unsigned/poly, e.g., "s8", "u8" or "p8" suffix
74 ClassW, // width-specific instruction, e.g., "8" suffix
75 ClassB, // bitcast arguments with enum argument to specify type
76 ClassL, // Logical instructions which are op instructions
77 // but we need to not emit any suffix for in our
79 ClassNoTest // Instructions which we do not test since they are
80 // not TRUE instructions.
83 /// NeonTypeFlags - Flags to identify the types for overloaded Neon
84 /// builtins. These must be kept in sync with the flags in
85 /// include/clang/Basic/TargetBuiltins.h.
86 namespace NeonTypeFlags {
88 enum { EltTypeMask = 0xf, UnsignedFlag = 0x10, QuadFlag = 0x20 };
104 } // end namespace NeonTypeFlags
108 //===----------------------------------------------------------------------===//
110 //===----------------------------------------------------------------------===//
112 /// A TypeSpec is just a simple wrapper around a string, but gets its own type
113 /// for strong typing purposes.
115 /// A TypeSpec can be used to create a type.
116 class TypeSpec : public std::string {
118 static std::vector<TypeSpec> fromTypeSpecs(StringRef Str) {
119 std::vector<TypeSpec> Ret;
121 for (char I : Str.str()) {
124 Ret.push_back(TypeSpec(Acc));
134 //===----------------------------------------------------------------------===//
136 //===----------------------------------------------------------------------===//
138 /// A Type. Not much more to say here.
151 bool Immediate, Constant, Pointer;
152 // ScalarForMangling and NoManglingQ are really not suited to live here as
153 // they are not related to the type. But they live in the TypeSpec (not the
154 // prototype), so this is really the only place to store them.
155 bool ScalarForMangling, NoManglingQ;
156 unsigned Bitwidth, ElementBitwidth, NumVectors;
160 : Kind(Void), Immediate(false), Constant(false),
161 Pointer(false), ScalarForMangling(false), NoManglingQ(false),
162 Bitwidth(0), ElementBitwidth(0), NumVectors(0) {}
164 Type(TypeSpec TS, StringRef CharMods)
165 : TS(std::move(TS)), Kind(Void), Immediate(false),
166 Constant(false), Pointer(false), ScalarForMangling(false),
167 NoManglingQ(false), Bitwidth(0), ElementBitwidth(0), NumVectors(0) {
168 applyModifiers(CharMods);
171 /// Returns a type representing "void".
172 static Type getVoid() { return Type(); }
174 bool operator==(const Type &Other) const { return str() == Other.str(); }
175 bool operator!=(const Type &Other) const { return !operator==(Other); }
180 bool isScalarForMangling() const { return ScalarForMangling; }
181 bool noManglingQ() const { return NoManglingQ; }
183 bool isPointer() const { return Pointer; }
184 bool isValue() const { return !isVoid() && !isPointer(); }
185 bool isScalar() const { return isValue() && NumVectors == 0; }
186 bool isVector() const { return isValue() && NumVectors > 0; }
187 bool isConstPointer() const { return Constant; }
188 bool isFloating() const { return Kind == Float; }
189 bool isInteger() const { return Kind == SInt || Kind == UInt; }
190 bool isPoly() const { return Kind == Poly; }
191 bool isSigned() const { return Kind == SInt; }
192 bool isImmediate() const { return Immediate; }
193 bool isFloat() const { return isFloating() && ElementBitwidth == 32; }
194 bool isDouble() const { return isFloating() && ElementBitwidth == 64; }
195 bool isHalf() const { return isFloating() && ElementBitwidth == 16; }
196 bool isChar() const { return ElementBitwidth == 8; }
197 bool isShort() const { return isInteger() && ElementBitwidth == 16; }
198 bool isInt() const { return isInteger() && ElementBitwidth == 32; }
199 bool isLong() const { return isInteger() && ElementBitwidth == 64; }
200 bool isVoid() const { return Kind == Void; }
201 unsigned getNumElements() const { return Bitwidth / ElementBitwidth; }
202 unsigned getSizeInBits() const { return Bitwidth; }
203 unsigned getElementSizeInBits() const { return ElementBitwidth; }
204 unsigned getNumVectors() const { return NumVectors; }
209 void makeUnsigned() {
210 assert(!isVoid() && "not a potentially signed type");
214 assert(!isVoid() && "not a potentially signed type");
218 void makeInteger(unsigned ElemWidth, bool Sign) {
219 assert(!isVoid() && "converting void to int probably not useful");
220 Kind = Sign ? SInt : UInt;
222 ElementBitwidth = ElemWidth;
225 void makeImmediate(unsigned ElemWidth) {
228 ElementBitwidth = ElemWidth;
232 Bitwidth = ElementBitwidth;
236 void makeOneVector() {
242 assert_with_loc(Bitwidth != 128, "Can't get bigger than 128!");
247 assert_with_loc(Bitwidth != 64, "Can't get smaller than 64!");
251 /// Return the C string representation of a type, which is the typename
252 /// defined in stdint.h or arm_neon.h.
253 std::string str() const;
255 /// Return the string representation of a type, which is an encoded
256 /// string for passing to the BUILTIN() macro in Builtins.def.
257 std::string builtin_str() const;
259 /// Return the value in NeonTypeFlags for this type.
260 unsigned getNeonEnum() const;
262 /// Parse a type from a stdint.h or arm_neon.h typedef name,
263 /// for example uint32x2_t or int64_t.
264 static Type fromTypedefName(StringRef Name);
267 /// Creates the type based on the typespec string in TS.
268 /// Sets "Quad" to true if the "Q" or "H" modifiers were
269 /// seen. This is needed by applyModifier as some modifiers
270 /// only take effect if the type size was changed by "Q" or "H".
271 void applyTypespec(bool &Quad);
272 /// Applies prototype modifiers to the type.
273 void applyModifiers(StringRef Mods);
276 //===----------------------------------------------------------------------===//
278 //===----------------------------------------------------------------------===//
280 /// A variable is a simple class that just has a type and a name.
286 Variable() : T(Type::getVoid()), N("") {}
287 Variable(Type T, std::string N) : T(std::move(T)), N(std::move(N)) {}
289 Type getType() const { return T; }
290 std::string getName() const { return "__" + N; }
293 //===----------------------------------------------------------------------===//
295 //===----------------------------------------------------------------------===//
297 /// The main grunt class. This represents an instantiation of an intrinsic with
298 /// a particular typespec and prototype.
300 friend class DagEmitter;
302 /// The Record this intrinsic was created from.
304 /// The unmangled name.
306 /// The input and output typespecs. InTS == OutTS except when
307 /// CartesianProductOfTypes is 1 - this is the case for vreinterpret.
308 TypeSpec OutTS, InTS;
309 /// The base class kind. Most intrinsics use ClassS, which has full type
310 /// info for integers (s32/u32). Some use ClassI, which doesn't care about
311 /// signedness (i32), while some (ClassB) have no type at all, only a width
314 /// The list of DAGs for the body. May be empty, in which case we should
315 /// emit a builtin call.
317 /// The architectural #ifdef guard.
319 /// Set if the Unavailable bit is 1. This means we don't generate a body,
320 /// just an "unavailable" attribute on a declaration.
322 /// Is this intrinsic safe for big-endian? or does it need its arguments
326 /// The types of return value [0] and parameters [1..].
327 std::vector<Type> Types;
328 /// The index of the key type passed to CGBuiltin.cpp for polymorphic calls.
329 int PolymorphicKeyType;
330 /// The local variables defined.
331 std::map<std::string, Variable> Variables;
332 /// NeededEarly - set if any other intrinsic depends on this intrinsic.
334 /// UseMacro - set if we should implement using a macro or unset for a
337 /// The set of intrinsics that this intrinsic uses/requires.
338 std::set<Intrinsic *> Dependencies;
339 /// The "base type", which is Type('d', OutTS). InBaseType is only
340 /// different if CartesianProductOfTypes = 1 (for vreinterpret).
341 Type BaseType, InBaseType;
342 /// The return variable.
344 /// A postfix to apply to every variable. Defaults to "".
345 std::string VariablePostfix;
347 NeonEmitter &Emitter;
348 std::stringstream OS;
350 bool isBigEndianSafe() const {
354 for (const auto &T : Types){
355 if (T.isVector() && T.getNumElements() > 1)
362 Intrinsic(Record *R, StringRef Name, StringRef Proto, TypeSpec OutTS,
363 TypeSpec InTS, ClassKind CK, ListInit *Body, NeonEmitter &Emitter,
364 StringRef Guard, bool IsUnavailable, bool BigEndianSafe)
365 : R(R), Name(Name.str()), OutTS(OutTS), InTS(InTS), CK(CK), Body(Body),
366 Guard(Guard.str()), IsUnavailable(IsUnavailable),
367 BigEndianSafe(BigEndianSafe), PolymorphicKeyType(0), NeededEarly(false),
368 UseMacro(false), BaseType(OutTS, "."), InBaseType(InTS, "."),
370 // Modify the TypeSpec per-argument to get a concrete Type, and create
371 // known variables for each.
372 // Types[0] is the return value.
374 Types.emplace_back(OutTS, getNextModifiers(Proto, Pos));
375 StringRef Mods = getNextModifiers(Proto, Pos);
376 while (!Mods.empty()) {
377 Types.emplace_back(InTS, Mods);
378 if (Mods.find("!") != StringRef::npos)
379 PolymorphicKeyType = Types.size() - 1;
381 Mods = getNextModifiers(Proto, Pos);
384 for (auto Type : Types) {
385 // If this builtin takes an immediate argument, we need to #define it rather
386 // than use a standard declaration, so that SemaChecking can range check
387 // the immediate passed by the user.
389 // Pointer arguments need to use macros to avoid hiding aligned attributes
390 // from the pointer type.
392 // It is not permitted to pass or return an __fp16 by value, so intrinsics
393 // taking a scalar float16_t must be implemented as macros.
394 if (Type.isImmediate() || Type.isPointer() ||
395 (Type.isScalar() && Type.isHalf()))
400 /// Get the Record that this intrinsic is based off.
401 Record *getRecord() const { return R; }
402 /// Get the set of Intrinsics that this intrinsic calls.
403 /// this is the set of immediate dependencies, NOT the
404 /// transitive closure.
405 const std::set<Intrinsic *> &getDependencies() const { return Dependencies; }
406 /// Get the architectural guard string (#ifdef).
407 std::string getGuard() const { return Guard; }
408 /// Get the non-mangled name.
409 std::string getName() const { return Name; }
411 /// Return true if the intrinsic takes an immediate operand.
412 bool hasImmediate() const {
413 return std::any_of(Types.begin(), Types.end(),
414 [](const Type &T) { return T.isImmediate(); });
417 /// Return the parameter index of the immediate operand.
418 unsigned getImmediateIdx() const {
419 for (unsigned Idx = 0; Idx < Types.size(); ++Idx)
420 if (Types[Idx].isImmediate())
422 llvm_unreachable("Intrinsic has no immediate");
426 unsigned getNumParams() const { return Types.size() - 1; }
427 Type getReturnType() const { return Types[0]; }
428 Type getParamType(unsigned I) const { return Types[I + 1]; }
429 Type getBaseType() const { return BaseType; }
430 Type getPolymorphicKeyType() const { return Types[PolymorphicKeyType]; }
432 /// Return true if the prototype has a scalar argument.
433 bool protoHasScalar() const;
435 /// Return the index that parameter PIndex will sit at
436 /// in a generated function call. This is often just PIndex,
437 /// but may not be as things such as multiple-vector operands
438 /// and sret parameters need to be taken into accont.
439 unsigned getGeneratedParamIdx(unsigned PIndex) {
441 if (getReturnType().getNumVectors() > 1)
442 // Multiple vectors are passed as sret.
445 for (unsigned I = 0; I < PIndex; ++I)
446 Idx += std::max(1U, getParamType(I).getNumVectors());
451 bool hasBody() const { return Body && !Body->getValues().empty(); }
453 void setNeededEarly() { NeededEarly = true; }
455 bool operator<(const Intrinsic &Other) const {
456 // Sort lexicographically on a two-tuple (Guard, Name)
457 if (Guard != Other.Guard)
458 return Guard < Other.Guard;
459 return Name < Other.Name;
462 ClassKind getClassKind(bool UseClassBIfScalar = false) {
463 if (UseClassBIfScalar && !protoHasScalar())
468 /// Return the name, mangled with type information.
469 /// If ForceClassS is true, use ClassS (u32/s32) instead
470 /// of the intrinsic's own type class.
471 std::string getMangledName(bool ForceClassS = false) const;
472 /// Return the type code for a builtin function call.
473 std::string getInstTypeCode(Type T, ClassKind CK) const;
474 /// Return the type string for a BUILTIN() macro in Builtins.def.
475 std::string getBuiltinTypeStr();
477 /// Generate the intrinsic, returning code.
478 std::string generate();
479 /// Perform type checking and populate the dependency graph, but
480 /// don't generate code yet.
484 StringRef getNextModifiers(StringRef Proto, unsigned &Pos) const;
486 std::string mangleName(std::string Name, ClassKind CK) const;
488 void initVariables();
489 std::string replaceParamsIn(std::string S);
491 void emitBodyAsBuiltinCall();
493 void generateImpl(bool ReverseArguments,
494 StringRef NamePrefix, StringRef CallPrefix);
496 void emitBody(StringRef CallPrefix);
497 void emitShadowedArgs();
498 void emitArgumentReversal();
499 void emitReturnReversal();
500 void emitReverseVariable(Variable &Dest, Variable &Src);
502 void emitClosingBrace();
503 void emitOpeningBrace();
504 void emitPrototype(StringRef NamePrefix);
508 StringRef CallPrefix;
511 DagEmitter(Intrinsic &Intr, StringRef CallPrefix) :
512 Intr(Intr), CallPrefix(CallPrefix) {
514 std::pair<Type, std::string> emitDagArg(Init *Arg, std::string ArgName);
515 std::pair<Type, std::string> emitDagSaveTemp(DagInit *DI);
516 std::pair<Type, std::string> emitDagSplat(DagInit *DI);
517 std::pair<Type, std::string> emitDagDup(DagInit *DI);
518 std::pair<Type, std::string> emitDagDupTyped(DagInit *DI);
519 std::pair<Type, std::string> emitDagShuffle(DagInit *DI);
520 std::pair<Type, std::string> emitDagCast(DagInit *DI, bool IsBitCast);
521 std::pair<Type, std::string> emitDagCall(DagInit *DI);
522 std::pair<Type, std::string> emitDagNameReplace(DagInit *DI);
523 std::pair<Type, std::string> emitDagLiteral(DagInit *DI);
524 std::pair<Type, std::string> emitDagOp(DagInit *DI);
525 std::pair<Type, std::string> emitDag(DagInit *DI);
529 //===----------------------------------------------------------------------===//
531 //===----------------------------------------------------------------------===//
534 RecordKeeper &Records;
535 DenseMap<Record *, ClassKind> ClassMap;
536 std::map<std::string, std::deque<Intrinsic>> IntrinsicMap;
537 unsigned UniqueNumber;
539 void createIntrinsic(Record *R, SmallVectorImpl<Intrinsic *> &Out);
540 void genBuiltinsDef(raw_ostream &OS, SmallVectorImpl<Intrinsic *> &Defs);
541 void genOverloadTypeCheckCode(raw_ostream &OS,
542 SmallVectorImpl<Intrinsic *> &Defs);
543 void genIntrinsicRangeCheckCode(raw_ostream &OS,
544 SmallVectorImpl<Intrinsic *> &Defs);
547 /// Called by Intrinsic - this attempts to get an intrinsic that takes
548 /// the given types as arguments.
549 Intrinsic &getIntrinsic(StringRef Name, ArrayRef<Type> Types);
551 /// Called by Intrinsic - returns a globally-unique number.
552 unsigned getUniqueNumber() { return UniqueNumber++; }
554 NeonEmitter(RecordKeeper &R) : Records(R), UniqueNumber(0) {
555 Record *SI = R.getClass("SInst");
556 Record *II = R.getClass("IInst");
557 Record *WI = R.getClass("WInst");
558 Record *SOpI = R.getClass("SOpInst");
559 Record *IOpI = R.getClass("IOpInst");
560 Record *WOpI = R.getClass("WOpInst");
561 Record *LOpI = R.getClass("LOpInst");
562 Record *NoTestOpI = R.getClass("NoTestOpInst");
564 ClassMap[SI] = ClassS;
565 ClassMap[II] = ClassI;
566 ClassMap[WI] = ClassW;
567 ClassMap[SOpI] = ClassS;
568 ClassMap[IOpI] = ClassI;
569 ClassMap[WOpI] = ClassW;
570 ClassMap[LOpI] = ClassL;
571 ClassMap[NoTestOpI] = ClassNoTest;
574 // run - Emit arm_neon.h.inc
575 void run(raw_ostream &o);
577 // runFP16 - Emit arm_fp16.h.inc
578 void runFP16(raw_ostream &o);
580 // runHeader - Emit all the __builtin prototypes used in arm_neon.h
582 void runHeader(raw_ostream &o);
584 // runTests - Emit tests for all the Neon intrinsics.
585 void runTests(raw_ostream &o);
588 } // end anonymous namespace
590 //===----------------------------------------------------------------------===//
591 // Type implementation
592 //===----------------------------------------------------------------------===//
594 std::string Type::str() const {
599 if (isInteger() && !isSigned())
604 else if (isFloating())
609 S += utostr(ElementBitwidth);
611 S += "x" + utostr(getNumElements());
613 S += "x" + utostr(NumVectors);
624 std::string Type::builtin_str() const {
630 // All pointers are void pointers.
632 if (isConstPointer())
636 } else if (isInteger())
637 switch (ElementBitwidth) {
638 case 8: S += "c"; break;
639 case 16: S += "s"; break;
640 case 32: S += "i"; break;
641 case 64: S += "Wi"; break;
642 case 128: S += "LLLi"; break;
643 default: llvm_unreachable("Unhandled case!");
646 switch (ElementBitwidth) {
647 case 16: S += "h"; break;
648 case 32: S += "f"; break;
649 case 64: S += "d"; break;
650 default: llvm_unreachable("Unhandled case!");
653 // FIXME: NECESSARY???????????????????????????????????????????????????????????????????????
654 if (isChar() && !isPointer() && isSigned())
655 // Make chars explicitly signed.
657 else if (isInteger() && !isSigned())
660 // Constant indices are "int", but have the "constant expression" modifier.
662 assert(isInteger() && isSigned());
670 for (unsigned I = 0; I < NumVectors; ++I)
671 Ret += "V" + utostr(getNumElements()) + S;
676 unsigned Type::getNeonEnum() const {
678 switch (ElementBitwidth) {
679 case 8: Addend = 0; break;
680 case 16: Addend = 1; break;
681 case 32: Addend = 2; break;
682 case 64: Addend = 3; break;
683 case 128: Addend = 4; break;
684 default: llvm_unreachable("Unhandled element bitwidth!");
687 unsigned Base = (unsigned)NeonTypeFlags::Int8 + Addend;
689 // Adjustment needed because Poly32 doesn't exist.
692 Base = (unsigned)NeonTypeFlags::Poly8 + Addend;
695 assert(Addend != 0 && "Float8 doesn't exist!");
696 Base = (unsigned)NeonTypeFlags::Float16 + (Addend - 1);
700 Base |= (unsigned)NeonTypeFlags::QuadFlag;
701 if (isInteger() && !isSigned())
702 Base |= (unsigned)NeonTypeFlags::UnsignedFlag;
707 Type Type::fromTypedefName(StringRef Name) {
711 if (Name.front() == 'u') {
713 Name = Name.drop_front();
716 if (Name.startswith("float")) {
718 Name = Name.drop_front(5);
719 } else if (Name.startswith("poly")) {
721 Name = Name.drop_front(4);
723 assert(Name.startswith("int"));
724 Name = Name.drop_front(3);
728 for (I = 0; I < Name.size(); ++I) {
729 if (!isdigit(Name[I]))
732 Name.substr(0, I).getAsInteger(10, T.ElementBitwidth);
733 Name = Name.drop_front(I);
735 T.Bitwidth = T.ElementBitwidth;
738 if (Name.front() == 'x') {
739 Name = Name.drop_front();
741 for (I = 0; I < Name.size(); ++I) {
742 if (!isdigit(Name[I]))
746 Name.substr(0, I).getAsInteger(10, NumLanes);
747 Name = Name.drop_front(I);
748 T.Bitwidth = T.ElementBitwidth * NumLanes;
753 if (Name.front() == 'x') {
754 Name = Name.drop_front();
756 for (I = 0; I < Name.size(); ++I) {
757 if (!isdigit(Name[I]))
760 Name.substr(0, I).getAsInteger(10, T.NumVectors);
761 Name = Name.drop_front(I);
764 assert(Name.startswith("_t") && "Malformed typedef!");
768 void Type::applyTypespec(bool &Quad) {
770 ScalarForMangling = false;
772 ElementBitwidth = ~0U;
778 ScalarForMangling = true;
800 ElementBitwidth = 16;
806 ElementBitwidth = 32;
812 ElementBitwidth = 64;
815 ElementBitwidth = 128;
816 // Poly doesn't have a 128x1 type.
821 llvm_unreachable("Unhandled type code!");
824 assert(ElementBitwidth != ~0U && "Bad element bitwidth!");
826 Bitwidth = Quad ? 128 : 64;
829 void Type::applyModifiers(StringRef Mods) {
830 bool AppliedQuad = false;
831 applyTypespec(AppliedQuad);
833 for (char Mod : Mods) {
853 assert(ElementBitwidth < 128);
854 ElementBitwidth *= 2;
857 assert(ElementBitwidth > 8);
858 ElementBitwidth /= 2;
886 ElementBitwidth = Bitwidth = 32;
895 // Key type, handled elsewhere.
898 llvm_unreachable("Unhandled character!");
903 //===----------------------------------------------------------------------===//
904 // Intrinsic implementation
905 //===----------------------------------------------------------------------===//
907 StringRef Intrinsic::getNextModifiers(StringRef Proto, unsigned &Pos) const {
908 if (Proto.size() == Pos)
910 else if (Proto[Pos] != '(')
911 return Proto.substr(Pos++, 1);
913 size_t Start = Pos + 1;
914 size_t End = Proto.find(')', Start);
915 assert_with_loc(End != StringRef::npos, "unmatched modifier group paren");
917 return Proto.slice(Start, End);
920 std::string Intrinsic::getInstTypeCode(Type T, ClassKind CK) const {
921 char typeCode = '\0';
922 bool printNumber = true;
929 else if (T.isInteger())
930 typeCode = T.isSigned() ? 's' : 'u';
950 if (typeCode != '\0')
951 S.push_back(typeCode);
953 S += utostr(T.getElementSizeInBits());
958 std::string Intrinsic::getBuiltinTypeStr() {
959 ClassKind LocalCK = getClassKind(true);
962 Type RetT = getReturnType();
963 if ((LocalCK == ClassI || LocalCK == ClassW) && RetT.isScalar() &&
965 RetT.makeInteger(RetT.getElementSizeInBits(), false);
967 // Since the return value must be one type, return a vector type of the
968 // appropriate width which we will bitcast. An exception is made for
969 // returning structs of 2, 3, or 4 vectors which are returned in a sret-like
970 // fashion, storing them to a pointer arg.
971 if (RetT.getNumVectors() > 1) {
972 S += "vv*"; // void result with void* first argument
975 RetT.makeInteger(RetT.getElementSizeInBits(), false);
976 if (!RetT.isScalar() && RetT.isInteger() && !RetT.isSigned())
979 if (LocalCK == ClassB && RetT.isValue() && !RetT.isScalar())
980 // Cast to vector of 8-bit elements.
981 RetT.makeInteger(8, true);
983 S += RetT.builtin_str();
986 for (unsigned I = 0; I < getNumParams(); ++I) {
987 Type T = getParamType(I);
989 T.makeInteger(T.getElementSizeInBits(), false);
991 if (LocalCK == ClassB && !T.isScalar())
992 T.makeInteger(8, true);
993 // Halves always get converted to 8-bit elements.
994 if (T.isHalf() && T.isVector() && !T.isScalarForMangling())
995 T.makeInteger(8, true);
997 if (LocalCK == ClassI && T.isInteger())
1000 if (hasImmediate() && getImmediateIdx() == I)
1001 T.makeImmediate(32);
1003 S += T.builtin_str();
1006 // Extra constant integer to hold type class enum for this function, e.g. s8
1007 if (LocalCK == ClassB)
1013 std::string Intrinsic::getMangledName(bool ForceClassS) const {
1014 // Check if the prototype has a scalar operand with the type of the vector
1015 // elements. If not, bitcasting the args will take care of arg checking.
1016 // The actual signedness etc. will be taken care of with special enums.
1017 ClassKind LocalCK = CK;
1018 if (!protoHasScalar())
1021 return mangleName(Name, ForceClassS ? ClassS : LocalCK);
1024 std::string Intrinsic::mangleName(std::string Name, ClassKind LocalCK) const {
1025 std::string typeCode = getInstTypeCode(BaseType, LocalCK);
1026 std::string S = Name;
1028 if (Name == "vcvt_f16_f32" || Name == "vcvt_f32_f16" ||
1029 Name == "vcvt_f32_f64" || Name == "vcvt_f64_f32")
1032 if (!typeCode.empty()) {
1033 // If the name ends with _xN (N = 2,3,4), insert the typeCode before _xN.
1034 if (Name.size() >= 3 && isdigit(Name.back()) &&
1035 Name[Name.length() - 2] == 'x' && Name[Name.length() - 3] == '_')
1036 S.insert(S.length() - 3, "_" + typeCode);
1038 S += "_" + typeCode;
1041 if (BaseType != InBaseType) {
1042 // A reinterpret - out the input base type at the end.
1043 S += "_" + getInstTypeCode(InBaseType, LocalCK);
1046 if (LocalCK == ClassB)
1049 // Insert a 'q' before the first '_' character so that it ends up before
1050 // _lane or _n on vector-scalar operations.
1051 if (BaseType.getSizeInBits() == 128 && !BaseType.noManglingQ()) {
1052 size_t Pos = S.find('_');
1057 if (BaseType.isScalarForMangling()) {
1058 switch (BaseType.getElementSizeInBits()) {
1059 case 8: Suffix = 'b'; break;
1060 case 16: Suffix = 'h'; break;
1061 case 32: Suffix = 's'; break;
1062 case 64: Suffix = 'd'; break;
1063 default: llvm_unreachable("Bad suffix!");
1066 if (Suffix != '\0') {
1067 size_t Pos = S.find('_');
1068 S.insert(Pos, &Suffix, 1);
1074 std::string Intrinsic::replaceParamsIn(std::string S) {
1075 while (S.find('$') != std::string::npos) {
1076 size_t Pos = S.find('$');
1077 size_t End = Pos + 1;
1078 while (isalpha(S[End]))
1081 std::string VarName = S.substr(Pos + 1, End - Pos - 1);
1082 assert_with_loc(Variables.find(VarName) != Variables.end(),
1083 "Variable not defined!");
1084 S.replace(Pos, End - Pos, Variables.find(VarName)->second.getName());
1090 void Intrinsic::initVariables() {
1093 // Modify the TypeSpec per-argument to get a concrete Type, and create
1094 // known variables for each.
1095 for (unsigned I = 1; I < Types.size(); ++I) {
1096 char NameC = '0' + (I - 1);
1097 std::string Name = "p";
1098 Name.push_back(NameC);
1100 Variables[Name] = Variable(Types[I], Name + VariablePostfix);
1102 RetVar = Variable(Types[0], "ret" + VariablePostfix);
1105 void Intrinsic::emitPrototype(StringRef NamePrefix) {
1109 OS << "__ai " << Types[0].str() << " ";
1111 OS << NamePrefix.str() << mangleName(Name, ClassS) << "(";
1113 for (unsigned I = 0; I < getNumParams(); ++I) {
1117 char NameC = '0' + I;
1118 std::string Name = "p";
1119 Name.push_back(NameC);
1120 assert(Variables.find(Name) != Variables.end());
1121 Variable &V = Variables[Name];
1124 OS << V.getType().str() << " ";
1131 void Intrinsic::emitOpeningBrace() {
1133 OS << " __extension__ ({";
1139 void Intrinsic::emitClosingBrace() {
1146 void Intrinsic::emitNewLine() {
1153 void Intrinsic::emitReverseVariable(Variable &Dest, Variable &Src) {
1154 if (Dest.getType().getNumVectors() > 1) {
1157 for (unsigned K = 0; K < Dest.getType().getNumVectors(); ++K) {
1158 OS << " " << Dest.getName() << ".val[" << K << "] = "
1159 << "__builtin_shufflevector("
1160 << Src.getName() << ".val[" << K << "], "
1161 << Src.getName() << ".val[" << K << "]";
1162 for (int J = Dest.getType().getNumElements() - 1; J >= 0; --J)
1168 OS << " " << Dest.getName()
1169 << " = __builtin_shufflevector(" << Src.getName() << ", " << Src.getName();
1170 for (int J = Dest.getType().getNumElements() - 1; J >= 0; --J)
1177 void Intrinsic::emitArgumentReversal() {
1178 if (isBigEndianSafe())
1181 // Reverse all vector arguments.
1182 for (unsigned I = 0; I < getNumParams(); ++I) {
1183 std::string Name = "p" + utostr(I);
1184 std::string NewName = "rev" + utostr(I);
1186 Variable &V = Variables[Name];
1187 Variable NewV(V.getType(), NewName + VariablePostfix);
1189 if (!NewV.getType().isVector() || NewV.getType().getNumElements() == 1)
1192 OS << " " << NewV.getType().str() << " " << NewV.getName() << ";";
1193 emitReverseVariable(NewV, V);
1198 void Intrinsic::emitReturnReversal() {
1199 if (isBigEndianSafe())
1201 if (!getReturnType().isVector() || getReturnType().isVoid() ||
1202 getReturnType().getNumElements() == 1)
1204 emitReverseVariable(RetVar, RetVar);
1207 void Intrinsic::emitShadowedArgs() {
1208 // Macro arguments are not type-checked like inline function arguments,
1209 // so assign them to local temporaries to get the right type checking.
1213 for (unsigned I = 0; I < getNumParams(); ++I) {
1214 // Do not create a temporary for an immediate argument.
1215 // That would defeat the whole point of using a macro!
1216 if (getParamType(I).isImmediate())
1218 // Do not create a temporary for pointer arguments. The input
1219 // pointer may have an alignment hint.
1220 if (getParamType(I).isPointer())
1223 std::string Name = "p" + utostr(I);
1225 assert(Variables.find(Name) != Variables.end());
1226 Variable &V = Variables[Name];
1228 std::string NewName = "s" + utostr(I);
1229 Variable V2(V.getType(), NewName + VariablePostfix);
1231 OS << " " << V2.getType().str() << " " << V2.getName() << " = "
1232 << V.getName() << ";";
1239 bool Intrinsic::protoHasScalar() const {
1240 return std::any_of(Types.begin(), Types.end(), [](const Type &T) {
1241 return T.isScalar() && !T.isImmediate();
1245 void Intrinsic::emitBodyAsBuiltinCall() {
1248 // If this builtin returns a struct 2, 3, or 4 vectors, pass it as an implicit
1249 // sret-like argument.
1250 bool SRet = getReturnType().getNumVectors() >= 2;
1253 ClassKind LocalCK = CK;
1254 if (!protoHasScalar())
1257 if (!getReturnType().isVoid() && !SRet)
1258 S += "(" + RetVar.getType().str() + ") ";
1260 S += "__builtin_neon_" + mangleName(N, LocalCK) + "(";
1263 S += "&" + RetVar.getName() + ", ";
1265 for (unsigned I = 0; I < getNumParams(); ++I) {
1266 Variable &V = Variables["p" + utostr(I)];
1267 Type T = V.getType();
1269 // Handle multiple-vector values specially, emitting each subvector as an
1270 // argument to the builtin.
1271 if (T.getNumVectors() > 1) {
1272 // Check if an explicit cast is needed.
1274 if (LocalCK == ClassB) {
1277 T2.makeInteger(8, /*Signed=*/true);
1278 Cast = "(" + T2.str() + ")";
1281 for (unsigned J = 0; J < T.getNumVectors(); ++J)
1282 S += Cast + V.getName() + ".val[" + utostr(J) + "], ";
1286 std::string Arg = V.getName();
1287 Type CastToType = T;
1289 // Check if an explicit cast is needed.
1290 if (CastToType.isVector() &&
1291 (LocalCK == ClassB || (T.isHalf() && !T.isScalarForMangling()))) {
1292 CastToType.makeInteger(8, true);
1293 Arg = "(" + CastToType.str() + ")" + Arg;
1294 } else if (CastToType.isVector() && LocalCK == ClassI) {
1295 if (CastToType.isInteger())
1296 CastToType.makeSigned();
1297 Arg = "(" + CastToType.str() + ")" + Arg;
1303 // Extra constant integer to hold type class enum for this function, e.g. s8
1304 if (getClassKind(true) == ClassB) {
1305 S += utostr(getPolymorphicKeyType().getNeonEnum());
1307 // Remove extraneous ", ".
1313 std::string RetExpr;
1314 if (!SRet && !RetVar.getType().isVoid())
1315 RetExpr = RetVar.getName() + " = ";
1317 OS << " " << RetExpr << S;
1321 void Intrinsic::emitBody(StringRef CallPrefix) {
1322 std::vector<std::string> Lines;
1324 assert(RetVar.getType() == Types[0]);
1325 // Create a return variable, if we're not void.
1326 if (!RetVar.getType().isVoid()) {
1327 OS << " " << RetVar.getType().str() << " " << RetVar.getName() << ";";
1331 if (!Body || Body->getValues().empty()) {
1332 // Nothing specific to output - must output a builtin.
1333 emitBodyAsBuiltinCall();
1337 // We have a list of "things to output". The last should be returned.
1338 for (auto *I : Body->getValues()) {
1339 if (StringInit *SI = dyn_cast<StringInit>(I)) {
1340 Lines.push_back(replaceParamsIn(SI->getAsString()));
1341 } else if (DagInit *DI = dyn_cast<DagInit>(I)) {
1342 DagEmitter DE(*this, CallPrefix);
1343 Lines.push_back(DE.emitDag(DI).second + ";");
1347 assert(!Lines.empty() && "Empty def?");
1348 if (!RetVar.getType().isVoid())
1349 Lines.back().insert(0, RetVar.getName() + " = ");
1351 for (auto &L : Lines) {
1357 void Intrinsic::emitReturn() {
1358 if (RetVar.getType().isVoid())
1361 OS << " " << RetVar.getName() << ";";
1363 OS << " return " << RetVar.getName() << ";";
1367 std::pair<Type, std::string> Intrinsic::DagEmitter::emitDag(DagInit *DI) {
1368 // At this point we should only be seeing a def.
1369 DefInit *DefI = cast<DefInit>(DI->getOperator());
1370 std::string Op = DefI->getAsString();
1372 if (Op == "cast" || Op == "bitcast")
1373 return emitDagCast(DI, Op == "bitcast");
1374 if (Op == "shuffle")
1375 return emitDagShuffle(DI);
1377 return emitDagDup(DI);
1378 if (Op == "dup_typed")
1379 return emitDagDupTyped(DI);
1381 return emitDagSplat(DI);
1382 if (Op == "save_temp")
1383 return emitDagSaveTemp(DI);
1385 return emitDagOp(DI);
1387 return emitDagCall(DI);
1388 if (Op == "name_replace")
1389 return emitDagNameReplace(DI);
1390 if (Op == "literal")
1391 return emitDagLiteral(DI);
1392 assert_with_loc(false, "Unknown operation!");
1393 return std::make_pair(Type::getVoid(), "");
1396 std::pair<Type, std::string> Intrinsic::DagEmitter::emitDagOp(DagInit *DI) {
1397 std::string Op = cast<StringInit>(DI->getArg(0))->getAsUnquotedString();
1398 if (DI->getNumArgs() == 2) {
1400 std::pair<Type, std::string> R =
1401 emitDagArg(DI->getArg(1), DI->getArgNameStr(1));
1402 return std::make_pair(R.first, Op + R.second);
1404 assert(DI->getNumArgs() == 3 && "Can only handle unary and binary ops!");
1405 std::pair<Type, std::string> R1 =
1406 emitDagArg(DI->getArg(1), DI->getArgNameStr(1));
1407 std::pair<Type, std::string> R2 =
1408 emitDagArg(DI->getArg(2), DI->getArgNameStr(2));
1409 assert_with_loc(R1.first == R2.first, "Argument type mismatch!");
1410 return std::make_pair(R1.first, R1.second + " " + Op + " " + R2.second);
1414 std::pair<Type, std::string> Intrinsic::DagEmitter::emitDagCall(DagInit *DI) {
1415 std::vector<Type> Types;
1416 std::vector<std::string> Values;
1417 for (unsigned I = 0; I < DI->getNumArgs() - 1; ++I) {
1418 std::pair<Type, std::string> R =
1419 emitDagArg(DI->getArg(I + 1), DI->getArgNameStr(I + 1));
1420 Types.push_back(R.first);
1421 Values.push_back(R.second);
1424 // Look up the called intrinsic.
1426 if (StringInit *SI = dyn_cast<StringInit>(DI->getArg(0)))
1427 N = SI->getAsUnquotedString();
1429 N = emitDagArg(DI->getArg(0), "").second;
1430 Intrinsic &Callee = Intr.Emitter.getIntrinsic(N, Types);
1432 // Make sure the callee is known as an early def.
1433 Callee.setNeededEarly();
1434 Intr.Dependencies.insert(&Callee);
1436 // Now create the call itself.
1438 if (!Callee.isBigEndianSafe())
1439 S += CallPrefix.str();
1440 S += Callee.getMangledName(true) + "(";
1441 for (unsigned I = 0; I < DI->getNumArgs() - 1; ++I) {
1448 return std::make_pair(Callee.getReturnType(), S);
1451 std::pair<Type, std::string> Intrinsic::DagEmitter::emitDagCast(DagInit *DI,
1453 // (cast MOD* VAL) -> cast VAL to type given by MOD.
1454 std::pair<Type, std::string> R = emitDagArg(
1455 DI->getArg(DI->getNumArgs() - 1),
1456 DI->getArgNameStr(DI->getNumArgs() - 1));
1457 Type castToType = R.first;
1458 for (unsigned ArgIdx = 0; ArgIdx < DI->getNumArgs() - 1; ++ArgIdx) {
1460 // MOD can take several forms:
1461 // 1. $X - take the type of parameter / variable X.
1462 // 2. The value "R" - take the type of the return type.
1464 // 4. The value "U" or "S" to switch the signedness.
1465 // 5. The value "H" or "D" to half or double the bitwidth.
1466 // 6. The value "8" to convert to 8-bit (signed) integer lanes.
1467 if (!DI->getArgNameStr(ArgIdx).empty()) {
1468 assert_with_loc(Intr.Variables.find(DI->getArgNameStr(ArgIdx)) !=
1469 Intr.Variables.end(),
1470 "Variable not found");
1471 castToType = Intr.Variables[DI->getArgNameStr(ArgIdx)].getType();
1473 StringInit *SI = dyn_cast<StringInit>(DI->getArg(ArgIdx));
1474 assert_with_loc(SI, "Expected string type or $Name for cast type");
1476 if (SI->getAsUnquotedString() == "R") {
1477 castToType = Intr.getReturnType();
1478 } else if (SI->getAsUnquotedString() == "U") {
1479 castToType.makeUnsigned();
1480 } else if (SI->getAsUnquotedString() == "S") {
1481 castToType.makeSigned();
1482 } else if (SI->getAsUnquotedString() == "H") {
1483 castToType.halveLanes();
1484 } else if (SI->getAsUnquotedString() == "D") {
1485 castToType.doubleLanes();
1486 } else if (SI->getAsUnquotedString() == "8") {
1487 castToType.makeInteger(8, true);
1489 castToType = Type::fromTypedefName(SI->getAsUnquotedString());
1490 assert_with_loc(!castToType.isVoid(), "Unknown typedef");
1497 // Emit a reinterpret cast. The second operand must be an lvalue, so create
1499 std::string N = "reint";
1501 while (Intr.Variables.find(N) != Intr.Variables.end())
1502 N = "reint" + utostr(++I);
1503 Intr.Variables[N] = Variable(R.first, N + Intr.VariablePostfix);
1505 Intr.OS << R.first.str() << " " << Intr.Variables[N].getName() << " = "
1509 S = "*(" + castToType.str() + " *) &" + Intr.Variables[N].getName() + "";
1511 // Emit a normal (static) cast.
1512 S = "(" + castToType.str() + ")(" + R.second + ")";
1515 return std::make_pair(castToType, S);
1518 std::pair<Type, std::string> Intrinsic::DagEmitter::emitDagShuffle(DagInit *DI){
1519 // See the documentation in arm_neon.td for a description of these operators.
1520 class LowHalf : public SetTheory::Operator {
1522 void apply(SetTheory &ST, DagInit *Expr, SetTheory::RecSet &Elts,
1523 ArrayRef<SMLoc> Loc) override {
1524 SetTheory::RecSet Elts2;
1525 ST.evaluate(Expr->arg_begin(), Expr->arg_end(), Elts2, Loc);
1526 Elts.insert(Elts2.begin(), Elts2.begin() + (Elts2.size() / 2));
1530 class HighHalf : public SetTheory::Operator {
1532 void apply(SetTheory &ST, DagInit *Expr, SetTheory::RecSet &Elts,
1533 ArrayRef<SMLoc> Loc) override {
1534 SetTheory::RecSet Elts2;
1535 ST.evaluate(Expr->arg_begin(), Expr->arg_end(), Elts2, Loc);
1536 Elts.insert(Elts2.begin() + (Elts2.size() / 2), Elts2.end());
1540 class Rev : public SetTheory::Operator {
1541 unsigned ElementSize;
1544 Rev(unsigned ElementSize) : ElementSize(ElementSize) {}
1546 void apply(SetTheory &ST, DagInit *Expr, SetTheory::RecSet &Elts,
1547 ArrayRef<SMLoc> Loc) override {
1548 SetTheory::RecSet Elts2;
1549 ST.evaluate(Expr->arg_begin() + 1, Expr->arg_end(), Elts2, Loc);
1551 int64_t VectorSize = cast<IntInit>(Expr->getArg(0))->getValue();
1552 VectorSize /= ElementSize;
1554 std::vector<Record *> Revved;
1555 for (unsigned VI = 0; VI < Elts2.size(); VI += VectorSize) {
1556 for (int LI = VectorSize - 1; LI >= 0; --LI) {
1557 Revved.push_back(Elts2[VI + LI]);
1561 Elts.insert(Revved.begin(), Revved.end());
1565 class MaskExpander : public SetTheory::Expander {
1569 MaskExpander(unsigned N) : N(N) {}
1571 void expand(SetTheory &ST, Record *R, SetTheory::RecSet &Elts) override {
1572 unsigned Addend = 0;
1573 if (R->getName() == "mask0")
1575 else if (R->getName() == "mask1")
1579 for (unsigned I = 0; I < N; ++I)
1580 Elts.insert(R->getRecords().getDef("sv" + utostr(I + Addend)));
1584 // (shuffle arg1, arg2, sequence)
1585 std::pair<Type, std::string> Arg1 =
1586 emitDagArg(DI->getArg(0), DI->getArgNameStr(0));
1587 std::pair<Type, std::string> Arg2 =
1588 emitDagArg(DI->getArg(1), DI->getArgNameStr(1));
1589 assert_with_loc(Arg1.first == Arg2.first,
1590 "Different types in arguments to shuffle!");
1593 SetTheory::RecSet Elts;
1594 ST.addOperator("lowhalf", std::make_unique<LowHalf>());
1595 ST.addOperator("highhalf", std::make_unique<HighHalf>());
1596 ST.addOperator("rev",
1597 std::make_unique<Rev>(Arg1.first.getElementSizeInBits()));
1598 ST.addExpander("MaskExpand",
1599 std::make_unique<MaskExpander>(Arg1.first.getNumElements()));
1600 ST.evaluate(DI->getArg(2), Elts, None);
1602 std::string S = "__builtin_shufflevector(" + Arg1.second + ", " + Arg2.second;
1603 for (auto &E : Elts) {
1604 StringRef Name = E->getName();
1605 assert_with_loc(Name.startswith("sv"),
1606 "Incorrect element kind in shuffle mask!");
1607 S += ", " + Name.drop_front(2).str();
1611 // Recalculate the return type - the shuffle may have halved or doubled it.
1613 if (Elts.size() > T.getNumElements()) {
1615 Elts.size() == T.getNumElements() * 2,
1616 "Can only double or half the number of elements in a shuffle!");
1618 } else if (Elts.size() < T.getNumElements()) {
1620 Elts.size() == T.getNumElements() / 2,
1621 "Can only double or half the number of elements in a shuffle!");
1625 return std::make_pair(T, S);
1628 std::pair<Type, std::string> Intrinsic::DagEmitter::emitDagDup(DagInit *DI) {
1629 assert_with_loc(DI->getNumArgs() == 1, "dup() expects one argument");
1630 std::pair<Type, std::string> A = emitDagArg(DI->getArg(0),
1631 DI->getArgNameStr(0));
1632 assert_with_loc(A.first.isScalar(), "dup() expects a scalar argument");
1634 Type T = Intr.getBaseType();
1635 assert_with_loc(T.isVector(), "dup() used but default type is scalar!");
1636 std::string S = "(" + T.str() + ") {";
1637 for (unsigned I = 0; I < T.getNumElements(); ++I) {
1644 return std::make_pair(T, S);
1647 std::pair<Type, std::string> Intrinsic::DagEmitter::emitDagDupTyped(DagInit *DI) {
1648 assert_with_loc(DI->getNumArgs() == 2, "dup_typed() expects two arguments");
1649 std::pair<Type, std::string> A = emitDagArg(DI->getArg(0),
1650 DI->getArgNameStr(0));
1651 std::pair<Type, std::string> B = emitDagArg(DI->getArg(1),
1652 DI->getArgNameStr(1));
1653 assert_with_loc(B.first.isScalar(),
1654 "dup_typed() requires a scalar as the second argument");
1657 assert_with_loc(T.isVector(), "dup_typed() used but target type is scalar!");
1658 std::string S = "(" + T.str() + ") {";
1659 for (unsigned I = 0; I < T.getNumElements(); ++I) {
1666 return std::make_pair(T, S);
1669 std::pair<Type, std::string> Intrinsic::DagEmitter::emitDagSplat(DagInit *DI) {
1670 assert_with_loc(DI->getNumArgs() == 2, "splat() expects two arguments");
1671 std::pair<Type, std::string> A = emitDagArg(DI->getArg(0),
1672 DI->getArgNameStr(0));
1673 std::pair<Type, std::string> B = emitDagArg(DI->getArg(1),
1674 DI->getArgNameStr(1));
1676 assert_with_loc(B.first.isScalar(),
1677 "splat() requires a scalar int as the second argument");
1679 std::string S = "__builtin_shufflevector(" + A.second + ", " + A.second;
1680 for (unsigned I = 0; I < Intr.getBaseType().getNumElements(); ++I) {
1681 S += ", " + B.second;
1685 return std::make_pair(Intr.getBaseType(), S);
1688 std::pair<Type, std::string> Intrinsic::DagEmitter::emitDagSaveTemp(DagInit *DI) {
1689 assert_with_loc(DI->getNumArgs() == 2, "save_temp() expects two arguments");
1690 std::pair<Type, std::string> A = emitDagArg(DI->getArg(1),
1691 DI->getArgNameStr(1));
1693 assert_with_loc(!A.first.isVoid(),
1694 "Argument to save_temp() must have non-void type!");
1696 std::string N = DI->getArgNameStr(0);
1697 assert_with_loc(!N.empty(),
1698 "save_temp() expects a name as the first argument");
1700 assert_with_loc(Intr.Variables.find(N) == Intr.Variables.end(),
1701 "Variable already defined!");
1702 Intr.Variables[N] = Variable(A.first, N + Intr.VariablePostfix);
1705 A.first.str() + " " + Intr.Variables[N].getName() + " = " + A.second;
1707 return std::make_pair(Type::getVoid(), S);
1710 std::pair<Type, std::string>
1711 Intrinsic::DagEmitter::emitDagNameReplace(DagInit *DI) {
1712 std::string S = Intr.Name;
1714 assert_with_loc(DI->getNumArgs() == 2, "name_replace requires 2 arguments!");
1715 std::string ToReplace = cast<StringInit>(DI->getArg(0))->getAsUnquotedString();
1716 std::string ReplaceWith = cast<StringInit>(DI->getArg(1))->getAsUnquotedString();
1718 size_t Idx = S.find(ToReplace);
1720 assert_with_loc(Idx != std::string::npos, "name should contain '" + ToReplace + "'!");
1721 S.replace(Idx, ToReplace.size(), ReplaceWith);
1723 return std::make_pair(Type::getVoid(), S);
1726 std::pair<Type, std::string> Intrinsic::DagEmitter::emitDagLiteral(DagInit *DI){
1727 std::string Ty = cast<StringInit>(DI->getArg(0))->getAsUnquotedString();
1728 std::string Value = cast<StringInit>(DI->getArg(1))->getAsUnquotedString();
1729 return std::make_pair(Type::fromTypedefName(Ty), Value);
1732 std::pair<Type, std::string>
1733 Intrinsic::DagEmitter::emitDagArg(Init *Arg, std::string ArgName) {
1734 if (!ArgName.empty()) {
1735 assert_with_loc(!Arg->isComplete(),
1736 "Arguments must either be DAGs or names, not both!");
1737 assert_with_loc(Intr.Variables.find(ArgName) != Intr.Variables.end(),
1738 "Variable not defined!");
1739 Variable &V = Intr.Variables[ArgName];
1740 return std::make_pair(V.getType(), V.getName());
1743 assert(Arg && "Neither ArgName nor Arg?!");
1744 DagInit *DI = dyn_cast<DagInit>(Arg);
1745 assert_with_loc(DI, "Arguments must either be DAGs or names!");
1750 std::string Intrinsic::generate() {
1751 // Avoid duplicated code for big and little endian
1752 if (isBigEndianSafe()) {
1753 generateImpl(false, "", "");
1756 // Little endian intrinsics are simple and don't require any argument
1758 OS << "#ifdef __LITTLE_ENDIAN__\n";
1760 generateImpl(false, "", "");
1764 // Big endian intrinsics are more complex. The user intended these
1765 // intrinsics to operate on a vector "as-if" loaded by (V)LDR,
1766 // but we load as-if (V)LD1. So we should swap all arguments and
1767 // swap the return value too.
1769 // If we call sub-intrinsics, we should call a version that does
1770 // not re-swap the arguments!
1771 generateImpl(true, "", "__noswap_");
1773 // If we're needed early, create a non-swapping variant for
1776 generateImpl(false, "__noswap_", "__noswap_");
1783 void Intrinsic::generateImpl(bool ReverseArguments,
1784 StringRef NamePrefix, StringRef CallPrefix) {
1787 // If we call a macro, our local variables may be corrupted due to
1788 // lack of proper lexical scoping. So, add a globally unique postfix
1789 // to every variable.
1791 // indexBody() should have set up the Dependencies set by now.
1792 for (auto *I : Dependencies)
1794 VariablePostfix = "_" + utostr(Emitter.getUniqueNumber());
1800 emitPrototype(NamePrefix);
1802 if (IsUnavailable) {
1803 OS << " __attribute__((unavailable));";
1807 if (ReverseArguments)
1808 emitArgumentReversal();
1809 emitBody(CallPrefix);
1810 if (ReverseArguments)
1811 emitReturnReversal();
1817 CurrentRecord = nullptr;
1820 void Intrinsic::indexBody() {
1827 CurrentRecord = nullptr;
1830 //===----------------------------------------------------------------------===//
1831 // NeonEmitter implementation
1832 //===----------------------------------------------------------------------===//
1834 Intrinsic &NeonEmitter::getIntrinsic(StringRef Name, ArrayRef<Type> Types) {
1835 // First, look up the name in the intrinsic map.
1836 assert_with_loc(IntrinsicMap.find(Name.str()) != IntrinsicMap.end(),
1837 ("Intrinsic '" + Name + "' not found!").str());
1838 auto &V = IntrinsicMap.find(Name.str())->second;
1839 std::vector<Intrinsic *> GoodVec;
1841 // Create a string to print if we end up failing.
1842 std::string ErrMsg = "looking up intrinsic '" + Name.str() + "(";
1843 for (unsigned I = 0; I < Types.size(); ++I) {
1846 ErrMsg += Types[I].str();
1849 ErrMsg += "Available overloads:\n";
1851 // Now, look through each intrinsic implementation and see if the types are
1854 ErrMsg += " - " + I.getReturnType().str() + " " + I.getMangledName();
1856 for (unsigned A = 0; A < I.getNumParams(); ++A) {
1859 ErrMsg += I.getParamType(A).str();
1863 if (I.getNumParams() != Types.size())
1867 for (unsigned Arg = 0; Arg < Types.size(); ++Arg) {
1868 if (I.getParamType(Arg) != Types[Arg]) {
1874 GoodVec.push_back(&I);
1877 assert_with_loc(!GoodVec.empty(),
1878 "No compatible intrinsic found - " + ErrMsg);
1879 assert_with_loc(GoodVec.size() == 1, "Multiple overloads found - " + ErrMsg);
1881 return *GoodVec.front();
1884 void NeonEmitter::createIntrinsic(Record *R,
1885 SmallVectorImpl<Intrinsic *> &Out) {
1886 std::string Name = R->getValueAsString("Name");
1887 std::string Proto = R->getValueAsString("Prototype");
1888 std::string Types = R->getValueAsString("Types");
1889 Record *OperationRec = R->getValueAsDef("Operation");
1890 bool CartesianProductOfTypes = R->getValueAsBit("CartesianProductOfTypes");
1891 bool BigEndianSafe = R->getValueAsBit("BigEndianSafe");
1892 std::string Guard = R->getValueAsString("ArchGuard");
1893 bool IsUnavailable = OperationRec->getValueAsBit("Unavailable");
1895 // Set the global current record. This allows assert_with_loc to produce
1896 // decent location information even when highly nested.
1899 ListInit *Body = OperationRec->getValueAsListInit("Ops");
1901 std::vector<TypeSpec> TypeSpecs = TypeSpec::fromTypeSpecs(Types);
1903 ClassKind CK = ClassNone;
1904 if (R->getSuperClasses().size() >= 2)
1905 CK = ClassMap[R->getSuperClasses()[1].first];
1907 std::vector<std::pair<TypeSpec, TypeSpec>> NewTypeSpecs;
1908 for (auto TS : TypeSpecs) {
1909 if (CartesianProductOfTypes) {
1910 Type DefaultT(TS, ".");
1911 for (auto SrcTS : TypeSpecs) {
1912 Type DefaultSrcT(SrcTS, ".");
1914 DefaultSrcT.getSizeInBits() != DefaultT.getSizeInBits())
1916 NewTypeSpecs.push_back(std::make_pair(TS, SrcTS));
1919 NewTypeSpecs.push_back(std::make_pair(TS, TS));
1923 llvm::sort(NewTypeSpecs);
1924 NewTypeSpecs.erase(std::unique(NewTypeSpecs.begin(), NewTypeSpecs.end()),
1925 NewTypeSpecs.end());
1926 auto &Entry = IntrinsicMap[Name];
1928 for (auto &I : NewTypeSpecs) {
1929 Entry.emplace_back(R, Name, Proto, I.first, I.second, CK, Body, *this,
1930 Guard, IsUnavailable, BigEndianSafe);
1931 Out.push_back(&Entry.back());
1934 CurrentRecord = nullptr;
1937 /// genBuiltinsDef: Generate the BuiltinsARM.def and BuiltinsAArch64.def
1938 /// declaration of builtins, checking for unique builtin declarations.
1939 void NeonEmitter::genBuiltinsDef(raw_ostream &OS,
1940 SmallVectorImpl<Intrinsic *> &Defs) {
1941 OS << "#ifdef GET_NEON_BUILTINS\n";
1943 // We only want to emit a builtin once, and we want to emit them in
1944 // alphabetical order, so use a std::set.
1945 std::set<std::string> Builtins;
1947 for (auto *Def : Defs) {
1951 std::string S = "BUILTIN(__builtin_neon_" + Def->getMangledName() + ", \"";
1953 S += Def->getBuiltinTypeStr();
1959 for (auto &S : Builtins)
1964 /// Generate the ARM and AArch64 overloaded type checking code for
1965 /// SemaChecking.cpp, checking for unique builtin declarations.
1966 void NeonEmitter::genOverloadTypeCheckCode(raw_ostream &OS,
1967 SmallVectorImpl<Intrinsic *> &Defs) {
1968 OS << "#ifdef GET_NEON_OVERLOAD_CHECK\n";
1970 // We record each overload check line before emitting because subsequent Inst
1971 // definitions may extend the number of permitted types (i.e. augment the
1972 // Mask). Use std::map to avoid sorting the table by hash number.
1973 struct OverloadInfo {
1977 OverloadInfo() : Mask(0ULL), PtrArgNum(0), HasConstPtr(false) {}
1979 std::map<std::string, OverloadInfo> OverloadMap;
1981 for (auto *Def : Defs) {
1982 // If the def has a body (that is, it has Operation DAGs), it won't call
1983 // __builtin_neon_* so we don't need to generate a definition for it.
1986 // Functions which have a scalar argument cannot be overloaded, no need to
1987 // check them if we are emitting the type checking code.
1988 if (Def->protoHasScalar())
1991 uint64_t Mask = 0ULL;
1992 Mask |= 1ULL << Def->getPolymorphicKeyType().getNeonEnum();
1994 // Check if the function has a pointer or const pointer argument.
1996 bool HasConstPtr = false;
1997 for (unsigned I = 0; I < Def->getNumParams(); ++I) {
1998 const auto &Type = Def->getParamType(I);
1999 if (Type.isPointer()) {
2001 HasConstPtr = Type.isConstPointer();
2005 // For sret builtins, adjust the pointer argument index.
2006 if (PtrArgNum >= 0 && Def->getReturnType().getNumVectors() > 1)
2009 std::string Name = Def->getName();
2010 // Omit type checking for the pointer arguments of vld1_lane, vld1_dup,
2011 // and vst1_lane intrinsics. Using a pointer to the vector element
2012 // type with one of those operations causes codegen to select an aligned
2013 // load/store instruction. If you want an unaligned operation,
2014 // the pointer argument needs to have less alignment than element type,
2015 // so just accept any pointer type.
2016 if (Name == "vld1_lane" || Name == "vld1_dup" || Name == "vst1_lane") {
2018 HasConstPtr = false;
2022 std::string Name = Def->getMangledName();
2023 OverloadMap.insert(std::make_pair(Name, OverloadInfo()));
2024 OverloadInfo &OI = OverloadMap[Name];
2026 OI.PtrArgNum |= PtrArgNum;
2027 OI.HasConstPtr = HasConstPtr;
2031 for (auto &I : OverloadMap) {
2032 OverloadInfo &OI = I.second;
2034 OS << "case NEON::BI__builtin_neon_" << I.first << ": ";
2035 OS << "mask = 0x" << Twine::utohexstr(OI.Mask) << "ULL";
2036 if (OI.PtrArgNum >= 0)
2037 OS << "; PtrArgNum = " << OI.PtrArgNum;
2039 OS << "; HasConstPtr = true";
2045 void NeonEmitter::genIntrinsicRangeCheckCode(raw_ostream &OS,
2046 SmallVectorImpl<Intrinsic *> &Defs) {
2047 OS << "#ifdef GET_NEON_IMMEDIATE_CHECK\n";
2049 std::set<std::string> Emitted;
2051 for (auto *Def : Defs) {
2054 // Functions which do not have an immediate do not need to have range
2055 // checking code emitted.
2056 if (!Def->hasImmediate())
2058 if (Emitted.find(Def->getMangledName()) != Emitted.end())
2061 std::string LowerBound, UpperBound;
2063 Record *R = Def->getRecord();
2064 if (R->getValueAsBit("isVCVT_N")) {
2065 // VCVT between floating- and fixed-point values takes an immediate
2066 // in the range [1, 32) for f32 or [1, 64) for f64 or [1, 16) for f16.
2068 if (Def->getBaseType().getElementSizeInBits() == 16 ||
2069 Def->getName().find('h') != std::string::npos)
2070 // VCVTh operating on FP16 intrinsics in range [1, 16)
2072 else if (Def->getBaseType().getElementSizeInBits() == 32)
2076 } else if (R->getValueAsBit("isScalarShift")) {
2077 // Right shifts have an 'r' in the name, left shifts do not. Convert
2078 // instructions have the same bounds and right shifts.
2079 if (Def->getName().find('r') != std::string::npos ||
2080 Def->getName().find("cvt") != std::string::npos)
2083 UpperBound = utostr(Def->getReturnType().getElementSizeInBits() - 1);
2084 } else if (R->getValueAsBit("isShift")) {
2085 // Builtins which are overloaded by type will need to have their upper
2086 // bound computed at Sema time based on the type constant.
2088 // Right shifts have an 'r' in the name, left shifts do not.
2089 if (Def->getName().find('r') != std::string::npos)
2091 UpperBound = "RFT(TV, true)";
2092 } else if (Def->getClassKind(true) == ClassB) {
2093 // ClassB intrinsics have a type (and hence lane number) that is only
2094 // known at runtime.
2095 if (R->getValueAsBit("isLaneQ"))
2096 UpperBound = "RFT(TV, false, true)";
2098 UpperBound = "RFT(TV, false, false)";
2100 // The immediate generally refers to a lane in the preceding argument.
2101 assert(Def->getImmediateIdx() > 0);
2102 Type T = Def->getParamType(Def->getImmediateIdx() - 1);
2103 UpperBound = utostr(T.getNumElements() - 1);
2106 // Calculate the index of the immediate that should be range checked.
2107 unsigned Idx = Def->getNumParams();
2108 if (Def->hasImmediate())
2109 Idx = Def->getGeneratedParamIdx(Def->getImmediateIdx());
2111 OS << "case NEON::BI__builtin_neon_" << Def->getMangledName() << ": "
2112 << "i = " << Idx << ";";
2113 if (!LowerBound.empty())
2114 OS << " l = " << LowerBound << ";";
2115 if (!UpperBound.empty())
2116 OS << " u = " << UpperBound << ";";
2119 Emitted.insert(Def->getMangledName());
2125 /// runHeader - Emit a file with sections defining:
2126 /// 1. the NEON section of BuiltinsARM.def and BuiltinsAArch64.def.
2127 /// 2. the SemaChecking code for the type overload checking.
2128 /// 3. the SemaChecking code for validation of intrinsic immediate arguments.
2129 void NeonEmitter::runHeader(raw_ostream &OS) {
2130 std::vector<Record *> RV = Records.getAllDerivedDefinitions("Inst");
2132 SmallVector<Intrinsic *, 128> Defs;
2134 createIntrinsic(R, Defs);
2136 // Generate shared BuiltinsXXX.def
2137 genBuiltinsDef(OS, Defs);
2139 // Generate ARM overloaded type checking code for SemaChecking.cpp
2140 genOverloadTypeCheckCode(OS, Defs);
2142 // Generate ARM range checking code for shift/lane immediates.
2143 genIntrinsicRangeCheckCode(OS, Defs);
2146 /// run - Read the records in arm_neon.td and output arm_neon.h. arm_neon.h
2147 /// is comprised of type definitions and function declarations.
2148 void NeonEmitter::run(raw_ostream &OS) {
2149 OS << "/*===---- arm_neon.h - ARM Neon intrinsics "
2150 "------------------------------"
2153 " * Permission is hereby granted, free of charge, to any person "
2156 " * of this software and associated documentation files (the "
2159 " * in the Software without restriction, including without limitation "
2162 " * to use, copy, modify, merge, publish, distribute, sublicense, "
2164 " * copies of the Software, and to permit persons to whom the Software "
2166 " * furnished to do so, subject to the following conditions:\n"
2168 " * The above copyright notice and this permission notice shall be "
2170 " * all copies or substantial portions of the Software.\n"
2172 " * THE SOFTWARE IS PROVIDED \"AS IS\", WITHOUT WARRANTY OF ANY KIND, "
2174 " * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF "
2175 "MERCHANTABILITY,\n"
2176 " * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT "
2178 " * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR "
2180 " * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, "
2182 " * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER "
2184 " * THE SOFTWARE.\n"
2186 " *===-----------------------------------------------------------------"
2191 OS << "#ifndef __ARM_NEON_H\n";
2192 OS << "#define __ARM_NEON_H\n\n";
2194 OS << "#if !defined(__ARM_NEON)\n";
2195 OS << "#error \"NEON support not enabled\"\n";
2198 OS << "#include <stdint.h>\n\n";
2200 // Emit NEON-specific scalar typedefs.
2201 OS << "typedef float float32_t;\n";
2202 OS << "typedef __fp16 float16_t;\n";
2204 OS << "#ifdef __aarch64__\n";
2205 OS << "typedef double float64_t;\n";
2208 // For now, signedness of polynomial types depends on target
2209 OS << "#ifdef __aarch64__\n";
2210 OS << "typedef uint8_t poly8_t;\n";
2211 OS << "typedef uint16_t poly16_t;\n";
2212 OS << "typedef uint64_t poly64_t;\n";
2213 OS << "typedef __uint128_t poly128_t;\n";
2215 OS << "typedef int8_t poly8_t;\n";
2216 OS << "typedef int16_t poly16_t;\n";
2219 // Emit Neon vector typedefs.
2220 std::string TypedefTypes(
2221 "cQcsQsiQilQlUcQUcUsQUsUiQUiUlQUlhQhfQfdQdPcQPcPsQPsPlQPl");
2222 std::vector<TypeSpec> TDTypeVec = TypeSpec::fromTypeSpecs(TypedefTypes);
2224 // Emit vector typedefs.
2225 bool InIfdef = false;
2226 for (auto &TS : TDTypeVec) {
2229 if (T.isDouble() || (T.isPoly() && T.getElementSizeInBits() == 64))
2232 if (InIfdef && !IsA64) {
2236 if (!InIfdef && IsA64) {
2237 OS << "#ifdef __aarch64__\n";
2242 OS << "typedef __attribute__((neon_polyvector_type(";
2244 OS << "typedef __attribute__((neon_vector_type(";
2248 OS << T.getNumElements() << "))) ";
2250 OS << " " << T.str() << ";\n";
2256 // Emit struct typedefs.
2258 for (unsigned NumMembers = 2; NumMembers <= 4; ++NumMembers) {
2259 for (auto &TS : TDTypeVec) {
2262 if (T.isDouble() || (T.isPoly() && T.getElementSizeInBits() == 64))
2265 if (InIfdef && !IsA64) {
2269 if (!InIfdef && IsA64) {
2270 OS << "#ifdef __aarch64__\n";
2274 const char Mods[] = { static_cast<char>('2' + (NumMembers - 2)), 0};
2276 OS << "typedef struct " << VT.str() << " {\n";
2277 OS << " " << T.str() << " val";
2278 OS << "[" << NumMembers << "]";
2280 OS << VT.str() << ";\n";
2288 OS << "#define __ai static __inline__ __attribute__((__always_inline__, "
2289 "__nodebug__))\n\n";
2291 SmallVector<Intrinsic *, 128> Defs;
2292 std::vector<Record *> RV = Records.getAllDerivedDefinitions("Inst");
2294 createIntrinsic(R, Defs);
2296 for (auto *I : Defs)
2299 llvm::stable_sort(Defs, llvm::deref<std::less<>>());
2301 // Only emit a def when its requirements have been met.
2302 // FIXME: This loop could be made faster, but it's fast enough for now.
2303 bool MadeProgress = true;
2304 std::string InGuard;
2305 while (!Defs.empty() && MadeProgress) {
2306 MadeProgress = false;
2308 for (SmallVector<Intrinsic *, 128>::iterator I = Defs.begin();
2309 I != Defs.end(); /*No step*/) {
2310 bool DependenciesSatisfied = true;
2311 for (auto *II : (*I)->getDependencies()) {
2312 if (llvm::is_contained(Defs, II))
2313 DependenciesSatisfied = false;
2315 if (!DependenciesSatisfied) {
2316 // Try the next one.
2321 // Emit #endif/#if pair if needed.
2322 if ((*I)->getGuard() != InGuard) {
2323 if (!InGuard.empty())
2325 InGuard = (*I)->getGuard();
2326 if (!InGuard.empty())
2327 OS << "#if " << InGuard << "\n";
2330 // Actually generate the intrinsic code.
2331 OS << (*I)->generate();
2333 MadeProgress = true;
2337 assert(Defs.empty() && "Some requirements were not satisfied!");
2338 if (!InGuard.empty())
2342 OS << "#undef __ai\n\n";
2343 OS << "#endif /* __ARM_NEON_H */\n";
2346 /// run - Read the records in arm_fp16.td and output arm_fp16.h. arm_fp16.h
2347 /// is comprised of type definitions and function declarations.
2348 void NeonEmitter::runFP16(raw_ostream &OS) {
2349 OS << "/*===---- arm_fp16.h - ARM FP16 intrinsics "
2350 "------------------------------"
2353 " * Permission is hereby granted, free of charge, to any person "
2354 "obtaining a copy\n"
2355 " * of this software and associated documentation files (the "
2356 "\"Software\"), to deal\n"
2357 " * in the Software without restriction, including without limitation "
2359 " * to use, copy, modify, merge, publish, distribute, sublicense, "
2361 " * copies of the Software, and to permit persons to whom the Software "
2363 " * furnished to do so, subject to the following conditions:\n"
2365 " * The above copyright notice and this permission notice shall be "
2367 " * all copies or substantial portions of the Software.\n"
2369 " * THE SOFTWARE IS PROVIDED \"AS IS\", WITHOUT WARRANTY OF ANY KIND, "
2371 " * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF "
2372 "MERCHANTABILITY,\n"
2373 " * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT "
2375 " * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR "
2377 " * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, "
2379 " * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER "
2381 " * THE SOFTWARE.\n"
2383 " *===-----------------------------------------------------------------"
2388 OS << "#ifndef __ARM_FP16_H\n";
2389 OS << "#define __ARM_FP16_H\n\n";
2391 OS << "#include <stdint.h>\n\n";
2393 OS << "typedef __fp16 float16_t;\n";
2395 OS << "#define __ai static __inline__ __attribute__((__always_inline__, "
2396 "__nodebug__))\n\n";
2398 SmallVector<Intrinsic *, 128> Defs;
2399 std::vector<Record *> RV = Records.getAllDerivedDefinitions("Inst");
2401 createIntrinsic(R, Defs);
2403 for (auto *I : Defs)
2406 llvm::stable_sort(Defs, llvm::deref<std::less<>>());
2408 // Only emit a def when its requirements have been met.
2409 // FIXME: This loop could be made faster, but it's fast enough for now.
2410 bool MadeProgress = true;
2411 std::string InGuard;
2412 while (!Defs.empty() && MadeProgress) {
2413 MadeProgress = false;
2415 for (SmallVector<Intrinsic *, 128>::iterator I = Defs.begin();
2416 I != Defs.end(); /*No step*/) {
2417 bool DependenciesSatisfied = true;
2418 for (auto *II : (*I)->getDependencies()) {
2419 if (llvm::is_contained(Defs, II))
2420 DependenciesSatisfied = false;
2422 if (!DependenciesSatisfied) {
2423 // Try the next one.
2428 // Emit #endif/#if pair if needed.
2429 if ((*I)->getGuard() != InGuard) {
2430 if (!InGuard.empty())
2432 InGuard = (*I)->getGuard();
2433 if (!InGuard.empty())
2434 OS << "#if " << InGuard << "\n";
2437 // Actually generate the intrinsic code.
2438 OS << (*I)->generate();
2440 MadeProgress = true;
2444 assert(Defs.empty() && "Some requirements were not satisfied!");
2445 if (!InGuard.empty())
2449 OS << "#undef __ai\n\n";
2450 OS << "#endif /* __ARM_FP16_H */\n";
2453 void clang::EmitNeon(RecordKeeper &Records, raw_ostream &OS) {
2454 NeonEmitter(Records).run(OS);
2457 void clang::EmitFP16(RecordKeeper &Records, raw_ostream &OS) {
2458 NeonEmitter(Records).runFP16(OS);
2461 void clang::EmitNeonSema(RecordKeeper &Records, raw_ostream &OS) {
2462 NeonEmitter(Records).runHeader(OS);
2465 void clang::EmitNeonTest(RecordKeeper &Records, raw_ostream &OS) {
2466 llvm_unreachable("Neon test generation no longer implemented!");