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 "llvm/ADT/ArrayRef.h"
27 #include "llvm/ADT/DenseMap.h"
28 #include "llvm/ADT/None.h"
29 #include "llvm/ADT/SmallVector.h"
30 #include "llvm/ADT/STLExtras.h"
31 #include "llvm/ADT/StringExtras.h"
32 #include "llvm/ADT/StringRef.h"
33 #include "llvm/Support/Casting.h"
34 #include "llvm/Support/ErrorHandling.h"
35 #include "llvm/Support/raw_ostream.h"
36 #include "llvm/TableGen/Error.h"
37 #include "llvm/TableGen/Record.h"
38 #include "llvm/TableGen/SetTheory.h"
56 // While globals are generally bad, this one allows us to perform assertions
57 // liberally and somehow still trace them back to the def they indirectly
59 static Record *CurrentRecord = nullptr;
60 static void assert_with_loc(bool Assertion, const std::string &Str) {
63 PrintFatalError(CurrentRecord->getLoc(), Str);
71 ClassI, // generic integer instruction, e.g., "i8" suffix
72 ClassS, // signed/unsigned/poly, e.g., "s8", "u8" or "p8" suffix
73 ClassW, // width-specific instruction, e.g., "8" suffix
74 ClassB, // bitcast arguments with enum argument to specify type
75 ClassL, // Logical instructions which are op instructions
76 // but we need to not emit any suffix for in our
78 ClassNoTest // Instructions which we do not test since they are
79 // not TRUE instructions.
82 /// NeonTypeFlags - Flags to identify the types for overloaded Neon
83 /// builtins. These must be kept in sync with the flags in
84 /// include/clang/Basic/TargetBuiltins.h.
85 namespace NeonTypeFlags {
87 enum { EltTypeMask = 0xf, UnsignedFlag = 0x10, QuadFlag = 0x20 };
103 } // end namespace NeonTypeFlags
107 //===----------------------------------------------------------------------===//
109 //===----------------------------------------------------------------------===//
111 /// A TypeSpec is just a simple wrapper around a string, but gets its own type
112 /// for strong typing purposes.
114 /// A TypeSpec can be used to create a type.
115 class TypeSpec : public std::string {
117 static std::vector<TypeSpec> fromTypeSpecs(StringRef Str) {
118 std::vector<TypeSpec> Ret;
120 for (char I : Str.str()) {
123 Ret.push_back(TypeSpec(Acc));
133 //===----------------------------------------------------------------------===//
135 //===----------------------------------------------------------------------===//
137 /// A Type. Not much more to say here.
142 bool Float, Signed, Immediate, Void, Poly, Constant, Pointer;
143 // ScalarForMangling and NoManglingQ are really not suited to live here as
144 // they are not related to the type. But they live in the TypeSpec (not the
145 // prototype), so this is really the only place to store them.
146 bool ScalarForMangling, NoManglingQ;
147 unsigned Bitwidth, ElementBitwidth, NumVectors;
151 : Float(false), Signed(false), Immediate(false), Void(true), Poly(false),
152 Constant(false), Pointer(false), ScalarForMangling(false),
153 NoManglingQ(false), Bitwidth(0), ElementBitwidth(0), NumVectors(0) {}
155 Type(TypeSpec TS, char CharMod)
156 : TS(std::move(TS)), Float(false), Signed(false), Immediate(false),
157 Void(false), Poly(false), Constant(false), Pointer(false),
158 ScalarForMangling(false), NoManglingQ(false), Bitwidth(0),
159 ElementBitwidth(0), NumVectors(0) {
160 applyModifier(CharMod);
163 /// Returns a type representing "void".
164 static Type getVoid() { return Type(); }
166 bool operator==(const Type &Other) const { return str() == Other.str(); }
167 bool operator!=(const Type &Other) const { return !operator==(Other); }
172 bool isScalarForMangling() const { return ScalarForMangling; }
173 bool noManglingQ() const { return NoManglingQ; }
175 bool isPointer() const { return Pointer; }
176 bool isFloating() const { return Float; }
177 bool isInteger() const { return !Float && !Poly; }
178 bool isSigned() const { return Signed; }
179 bool isImmediate() const { return Immediate; }
180 bool isScalar() const { return NumVectors == 0; }
181 bool isVector() const { return NumVectors > 0; }
182 bool isFloat() const { return Float && ElementBitwidth == 32; }
183 bool isDouble() const { return Float && ElementBitwidth == 64; }
184 bool isHalf() const { return Float && ElementBitwidth == 16; }
185 bool isPoly() const { return Poly; }
186 bool isChar() const { return ElementBitwidth == 8; }
187 bool isShort() const { return !Float && ElementBitwidth == 16; }
188 bool isInt() const { return !Float && ElementBitwidth == 32; }
189 bool isLong() const { return !Float && ElementBitwidth == 64; }
190 bool isVoid() const { return Void; }
191 unsigned getNumElements() const { return Bitwidth / ElementBitwidth; }
192 unsigned getSizeInBits() const { return Bitwidth; }
193 unsigned getElementSizeInBits() const { return ElementBitwidth; }
194 unsigned getNumVectors() const { return NumVectors; }
199 void makeUnsigned() { Signed = false; }
200 void makeSigned() { Signed = true; }
202 void makeInteger(unsigned ElemWidth, bool Sign) {
207 ElementBitwidth = ElemWidth;
210 void makeImmediate(unsigned ElemWidth) {
215 ElementBitwidth = ElemWidth;
219 Bitwidth = ElementBitwidth;
223 void makeOneVector() {
229 assert_with_loc(Bitwidth != 128, "Can't get bigger than 128!");
234 assert_with_loc(Bitwidth != 64, "Can't get smaller than 64!");
238 /// Return the C string representation of a type, which is the typename
239 /// defined in stdint.h or arm_neon.h.
240 std::string str() const;
242 /// Return the string representation of a type, which is an encoded
243 /// string for passing to the BUILTIN() macro in Builtins.def.
244 std::string builtin_str() const;
246 /// Return the value in NeonTypeFlags for this type.
247 unsigned getNeonEnum() const;
249 /// Parse a type from a stdint.h or arm_neon.h typedef name,
250 /// for example uint32x2_t or int64_t.
251 static Type fromTypedefName(StringRef Name);
254 /// Creates the type based on the typespec string in TS.
255 /// Sets "Quad" to true if the "Q" or "H" modifiers were
256 /// seen. This is needed by applyModifier as some modifiers
257 /// only take effect if the type size was changed by "Q" or "H".
258 void applyTypespec(bool &Quad);
259 /// Applies a prototype modifier to the type.
260 void applyModifier(char Mod);
263 //===----------------------------------------------------------------------===//
265 //===----------------------------------------------------------------------===//
267 /// A variable is a simple class that just has a type and a name.
273 Variable() : T(Type::getVoid()), N("") {}
274 Variable(Type T, std::string N) : T(std::move(T)), N(std::move(N)) {}
276 Type getType() const { return T; }
277 std::string getName() const { return "__" + N; }
280 //===----------------------------------------------------------------------===//
282 //===----------------------------------------------------------------------===//
284 /// The main grunt class. This represents an instantiation of an intrinsic with
285 /// a particular typespec and prototype.
287 friend class DagEmitter;
289 /// The Record this intrinsic was created from.
291 /// The unmangled name and prototype.
292 std::string Name, Proto;
293 /// The input and output typespecs. InTS == OutTS except when
294 /// CartesianProductOfTypes is 1 - this is the case for vreinterpret.
295 TypeSpec OutTS, InTS;
296 /// The base class kind. Most intrinsics use ClassS, which has full type
297 /// info for integers (s32/u32). Some use ClassI, which doesn't care about
298 /// signedness (i32), while some (ClassB) have no type at all, only a width
301 /// The list of DAGs for the body. May be empty, in which case we should
302 /// emit a builtin call.
304 /// The architectural #ifdef guard.
306 /// Set if the Unavailable bit is 1. This means we don't generate a body,
307 /// just an "unavailable" attribute on a declaration.
309 /// Is this intrinsic safe for big-endian? or does it need its arguments
313 /// The types of return value [0] and parameters [1..].
314 std::vector<Type> Types;
315 /// The local variables defined.
316 std::map<std::string, Variable> Variables;
317 /// NeededEarly - set if any other intrinsic depends on this intrinsic.
319 /// UseMacro - set if we should implement using a macro or unset for a
322 /// The set of intrinsics that this intrinsic uses/requires.
323 std::set<Intrinsic *> Dependencies;
324 /// The "base type", which is Type('d', OutTS). InBaseType is only
325 /// different if CartesianProductOfTypes = 1 (for vreinterpret).
326 Type BaseType, InBaseType;
327 /// The return variable.
329 /// A postfix to apply to every variable. Defaults to "".
330 std::string VariablePostfix;
332 NeonEmitter &Emitter;
333 std::stringstream OS;
336 Intrinsic(Record *R, StringRef Name, StringRef Proto, TypeSpec OutTS,
337 TypeSpec InTS, ClassKind CK, ListInit *Body, NeonEmitter &Emitter,
338 StringRef Guard, bool IsUnavailable, bool BigEndianSafe)
339 : R(R), Name(Name.str()), Proto(Proto.str()), OutTS(OutTS), InTS(InTS),
340 CK(CK), Body(Body), Guard(Guard.str()), IsUnavailable(IsUnavailable),
341 BigEndianSafe(BigEndianSafe), NeededEarly(false), UseMacro(false),
342 BaseType(OutTS, 'd'), InBaseType(InTS, 'd'), Emitter(Emitter) {
343 // If this builtin takes an immediate argument, we need to #define it rather
344 // than use a standard declaration, so that SemaChecking can range check
345 // the immediate passed by the user.
346 if (Proto.find('i') != std::string::npos)
349 // Pointer arguments need to use macros to avoid hiding aligned attributes
350 // from the pointer type.
351 if (Proto.find('p') != std::string::npos ||
352 Proto.find('c') != std::string::npos)
355 // It is not permitted to pass or return an __fp16 by value, so intrinsics
356 // taking a scalar float16_t must be implemented as macros.
357 if (OutTS.find('h') != std::string::npos &&
358 Proto.find('s') != std::string::npos)
361 // Modify the TypeSpec per-argument to get a concrete Type, and create
362 // known variables for each.
363 // Types[0] is the return value.
364 Types.emplace_back(OutTS, Proto[0]);
365 for (unsigned I = 1; I < Proto.size(); ++I)
366 Types.emplace_back(InTS, Proto[I]);
369 /// Get the Record that this intrinsic is based off.
370 Record *getRecord() const { return R; }
371 /// Get the set of Intrinsics that this intrinsic calls.
372 /// this is the set of immediate dependencies, NOT the
373 /// transitive closure.
374 const std::set<Intrinsic *> &getDependencies() const { return Dependencies; }
375 /// Get the architectural guard string (#ifdef).
376 std::string getGuard() const { return Guard; }
377 /// Get the non-mangled name.
378 std::string getName() const { return Name; }
380 /// Return true if the intrinsic takes an immediate operand.
381 bool hasImmediate() const {
382 return Proto.find('i') != std::string::npos;
385 /// Return the parameter index of the immediate operand.
386 unsigned getImmediateIdx() const {
387 assert(hasImmediate());
388 unsigned Idx = Proto.find('i');
389 assert(Idx > 0 && "Can't return an immediate!");
393 /// Return true if the intrinsic takes an splat operand.
394 bool hasSplat() const { return Proto.find('a') != std::string::npos; }
396 /// Return the parameter index of the splat operand.
397 unsigned getSplatIdx() const {
399 unsigned Idx = Proto.find('a');
400 assert(Idx > 0 && "Can't return a splat!");
404 unsigned getNumParams() const { return Proto.size() - 1; }
405 Type getReturnType() const { return Types[0]; }
406 Type getParamType(unsigned I) const { return Types[I + 1]; }
407 Type getBaseType() const { return BaseType; }
408 /// Return the raw prototype string.
409 std::string getProto() const { return Proto; }
411 /// Return true if the prototype has a scalar argument.
412 /// This does not return true for the "splat" code ('a').
413 bool protoHasScalar() const;
415 /// Return the index that parameter PIndex will sit at
416 /// in a generated function call. This is often just PIndex,
417 /// but may not be as things such as multiple-vector operands
418 /// and sret parameters need to be taken into accont.
419 unsigned getGeneratedParamIdx(unsigned PIndex) {
421 if (getReturnType().getNumVectors() > 1)
422 // Multiple vectors are passed as sret.
425 for (unsigned I = 0; I < PIndex; ++I)
426 Idx += std::max(1U, getParamType(I).getNumVectors());
431 bool hasBody() const { return Body && !Body->getValues().empty(); }
433 void setNeededEarly() { NeededEarly = true; }
435 bool operator<(const Intrinsic &Other) const {
436 // Sort lexicographically on a two-tuple (Guard, Name)
437 if (Guard != Other.Guard)
438 return Guard < Other.Guard;
439 return Name < Other.Name;
442 ClassKind getClassKind(bool UseClassBIfScalar = false) {
443 if (UseClassBIfScalar && !protoHasScalar())
448 /// Return the name, mangled with type information.
449 /// If ForceClassS is true, use ClassS (u32/s32) instead
450 /// of the intrinsic's own type class.
451 std::string getMangledName(bool ForceClassS = false) const;
452 /// Return the type code for a builtin function call.
453 std::string getInstTypeCode(Type T, ClassKind CK) const;
454 /// Return the type string for a BUILTIN() macro in Builtins.def.
455 std::string getBuiltinTypeStr();
457 /// Generate the intrinsic, returning code.
458 std::string generate();
459 /// Perform type checking and populate the dependency graph, but
460 /// don't generate code yet.
464 std::string mangleName(std::string Name, ClassKind CK) const;
466 void initVariables();
467 std::string replaceParamsIn(std::string S);
469 void emitBodyAsBuiltinCall();
471 void generateImpl(bool ReverseArguments,
472 StringRef NamePrefix, StringRef CallPrefix);
474 void emitBody(StringRef CallPrefix);
475 void emitShadowedArgs();
476 void emitArgumentReversal();
477 void emitReturnReversal();
478 void emitReverseVariable(Variable &Dest, Variable &Src);
480 void emitClosingBrace();
481 void emitOpeningBrace();
482 void emitPrototype(StringRef NamePrefix);
486 StringRef CallPrefix;
489 DagEmitter(Intrinsic &Intr, StringRef CallPrefix) :
490 Intr(Intr), CallPrefix(CallPrefix) {
492 std::pair<Type, std::string> emitDagArg(Init *Arg, std::string ArgName);
493 std::pair<Type, std::string> emitDagSaveTemp(DagInit *DI);
494 std::pair<Type, std::string> emitDagSplat(DagInit *DI);
495 std::pair<Type, std::string> emitDagDup(DagInit *DI);
496 std::pair<Type, std::string> emitDagDupTyped(DagInit *DI);
497 std::pair<Type, std::string> emitDagShuffle(DagInit *DI);
498 std::pair<Type, std::string> emitDagCast(DagInit *DI, bool IsBitCast);
499 std::pair<Type, std::string> emitDagCall(DagInit *DI);
500 std::pair<Type, std::string> emitDagNameReplace(DagInit *DI);
501 std::pair<Type, std::string> emitDagLiteral(DagInit *DI);
502 std::pair<Type, std::string> emitDagOp(DagInit *DI);
503 std::pair<Type, std::string> emitDag(DagInit *DI);
507 //===----------------------------------------------------------------------===//
509 //===----------------------------------------------------------------------===//
512 RecordKeeper &Records;
513 DenseMap<Record *, ClassKind> ClassMap;
514 std::map<std::string, std::deque<Intrinsic>> IntrinsicMap;
515 unsigned UniqueNumber;
517 void createIntrinsic(Record *R, SmallVectorImpl<Intrinsic *> &Out);
518 void genBuiltinsDef(raw_ostream &OS, SmallVectorImpl<Intrinsic *> &Defs);
519 void genOverloadTypeCheckCode(raw_ostream &OS,
520 SmallVectorImpl<Intrinsic *> &Defs);
521 void genIntrinsicRangeCheckCode(raw_ostream &OS,
522 SmallVectorImpl<Intrinsic *> &Defs);
525 /// Called by Intrinsic - this attempts to get an intrinsic that takes
526 /// the given types as arguments.
527 Intrinsic &getIntrinsic(StringRef Name, ArrayRef<Type> Types);
529 /// Called by Intrinsic - returns a globally-unique number.
530 unsigned getUniqueNumber() { return UniqueNumber++; }
532 NeonEmitter(RecordKeeper &R) : Records(R), UniqueNumber(0) {
533 Record *SI = R.getClass("SInst");
534 Record *II = R.getClass("IInst");
535 Record *WI = R.getClass("WInst");
536 Record *SOpI = R.getClass("SOpInst");
537 Record *IOpI = R.getClass("IOpInst");
538 Record *WOpI = R.getClass("WOpInst");
539 Record *LOpI = R.getClass("LOpInst");
540 Record *NoTestOpI = R.getClass("NoTestOpInst");
542 ClassMap[SI] = ClassS;
543 ClassMap[II] = ClassI;
544 ClassMap[WI] = ClassW;
545 ClassMap[SOpI] = ClassS;
546 ClassMap[IOpI] = ClassI;
547 ClassMap[WOpI] = ClassW;
548 ClassMap[LOpI] = ClassL;
549 ClassMap[NoTestOpI] = ClassNoTest;
552 // run - Emit arm_neon.h.inc
553 void run(raw_ostream &o);
555 // runFP16 - Emit arm_fp16.h.inc
556 void runFP16(raw_ostream &o);
558 // runHeader - Emit all the __builtin prototypes used in arm_neon.h
560 void runHeader(raw_ostream &o);
562 // runTests - Emit tests for all the Neon intrinsics.
563 void runTests(raw_ostream &o);
566 } // end anonymous namespace
568 //===----------------------------------------------------------------------===//
569 // Type implementation
570 //===----------------------------------------------------------------------===//
572 std::string Type::str() const {
577 if (!Signed && isInteger())
587 S += utostr(ElementBitwidth);
589 S += "x" + utostr(getNumElements());
591 S += "x" + utostr(NumVectors);
602 std::string Type::builtin_str() const {
608 // All pointers are void pointers.
610 else if (isInteger())
611 switch (ElementBitwidth) {
612 case 8: S += "c"; break;
613 case 16: S += "s"; break;
614 case 32: S += "i"; break;
615 case 64: S += "Wi"; break;
616 case 128: S += "LLLi"; break;
617 default: llvm_unreachable("Unhandled case!");
620 switch (ElementBitwidth) {
621 case 16: S += "h"; break;
622 case 32: S += "f"; break;
623 case 64: S += "d"; break;
624 default: llvm_unreachable("Unhandled case!");
627 if (isChar() && !Pointer)
628 // Make chars explicitly signed.
630 else if (isInteger() && !Pointer && !Signed)
633 // Constant indices are "int", but have the "constant expression" modifier.
635 assert(isInteger() && isSigned());
640 if (Constant) S += "C";
641 if (Pointer) S += "*";
646 for (unsigned I = 0; I < NumVectors; ++I)
647 Ret += "V" + utostr(getNumElements()) + S;
652 unsigned Type::getNeonEnum() const {
654 switch (ElementBitwidth) {
655 case 8: Addend = 0; break;
656 case 16: Addend = 1; break;
657 case 32: Addend = 2; break;
658 case 64: Addend = 3; break;
659 case 128: Addend = 4; break;
660 default: llvm_unreachable("Unhandled element bitwidth!");
663 unsigned Base = (unsigned)NeonTypeFlags::Int8 + Addend;
665 // Adjustment needed because Poly32 doesn't exist.
668 Base = (unsigned)NeonTypeFlags::Poly8 + Addend;
671 assert(Addend != 0 && "Float8 doesn't exist!");
672 Base = (unsigned)NeonTypeFlags::Float16 + (Addend - 1);
676 Base |= (unsigned)NeonTypeFlags::QuadFlag;
677 if (isInteger() && !Signed)
678 Base |= (unsigned)NeonTypeFlags::UnsignedFlag;
683 Type Type::fromTypedefName(StringRef Name) {
689 if (Name.front() == 'u') {
691 Name = Name.drop_front();
696 if (Name.startswith("float")) {
698 Name = Name.drop_front(5);
699 } else if (Name.startswith("poly")) {
701 Name = Name.drop_front(4);
703 assert(Name.startswith("int"));
704 Name = Name.drop_front(3);
708 for (I = 0; I < Name.size(); ++I) {
709 if (!isdigit(Name[I]))
712 Name.substr(0, I).getAsInteger(10, T.ElementBitwidth);
713 Name = Name.drop_front(I);
715 T.Bitwidth = T.ElementBitwidth;
718 if (Name.front() == 'x') {
719 Name = Name.drop_front();
721 for (I = 0; I < Name.size(); ++I) {
722 if (!isdigit(Name[I]))
726 Name.substr(0, I).getAsInteger(10, NumLanes);
727 Name = Name.drop_front(I);
728 T.Bitwidth = T.ElementBitwidth * NumLanes;
733 if (Name.front() == 'x') {
734 Name = Name.drop_front();
736 for (I = 0; I < Name.size(); ++I) {
737 if (!isdigit(Name[I]))
740 Name.substr(0, I).getAsInteger(10, T.NumVectors);
741 Name = Name.drop_front(I);
744 assert(Name.startswith("_t") && "Malformed typedef!");
748 void Type::applyTypespec(bool &Quad) {
750 ScalarForMangling = false;
752 Poly = Float = false;
753 ElementBitwidth = ~0U;
760 ScalarForMangling = true;
782 ElementBitwidth = 16;
788 ElementBitwidth = 32;
794 ElementBitwidth = 64;
797 ElementBitwidth = 128;
798 // Poly doesn't have a 128x1 type.
803 llvm_unreachable("Unhandled type code!");
806 assert(ElementBitwidth != ~0U && "Bad element bitwidth!");
808 Bitwidth = Quad ? 128 : 64;
811 void Type::applyModifier(char Mod) {
812 bool AppliedQuad = false;
813 applyTypespec(AppliedQuad);
830 Bitwidth = ElementBitwidth;
837 Bitwidth = ElementBitwidth;
846 assert(!Poly && "'u' can't be used with poly types!");
850 Bitwidth = ElementBitwidth = 64;
855 Bitwidth = ElementBitwidth = 32;
860 Bitwidth = ElementBitwidth = 16;
865 Bitwidth = ElementBitwidth = 32;
871 Bitwidth = ElementBitwidth = 64;
877 Bitwidth = ElementBitwidth = 32;
883 Bitwidth = ElementBitwidth = 64;
890 ElementBitwidth = 32;
894 ElementBitwidth = 64;
898 ElementBitwidth = 16;
904 ElementBitwidth = 16;
910 ElementBitwidth = 16;
921 ElementBitwidth *= 2;
925 ElementBitwidth *= 2;
930 ElementBitwidth = Bitwidth = 32;
938 ElementBitwidth = Bitwidth = 64;
944 ElementBitwidth /= 2;
945 Bitwidth = ElementBitwidth;
949 ElementBitwidth *= 2;
950 Bitwidth = ElementBitwidth;
955 Bitwidth = ElementBitwidth;
966 Bitwidth = ElementBitwidth;
970 ElementBitwidth /= 2;
973 ElementBitwidth /= 2;
977 ElementBitwidth /= 2;
981 ElementBitwidth /= 2;
1013 ElementBitwidth = 8;
1016 ElementBitwidth = 8;
1021 ElementBitwidth = 8;
1024 llvm_unreachable("Unhandled character!");
1028 //===----------------------------------------------------------------------===//
1029 // Intrinsic implementation
1030 //===----------------------------------------------------------------------===//
1032 std::string Intrinsic::getInstTypeCode(Type T, ClassKind CK) const {
1033 char typeCode = '\0';
1034 bool printNumber = true;
1041 else if (T.isInteger())
1042 typeCode = T.isSigned() ? 's' : 'u';
1062 if (typeCode != '\0')
1063 S.push_back(typeCode);
1065 S += utostr(T.getElementSizeInBits());
1070 static bool isFloatingPointProtoModifier(char Mod) {
1071 return Mod == 'F' || Mod == 'f' || Mod == 'H' || Mod == 'Y' || Mod == 'I';
1074 std::string Intrinsic::getBuiltinTypeStr() {
1075 ClassKind LocalCK = getClassKind(true);
1078 Type RetT = getReturnType();
1079 if ((LocalCK == ClassI || LocalCK == ClassW) && RetT.isScalar() &&
1081 RetT.makeInteger(RetT.getElementSizeInBits(), false);
1083 // Since the return value must be one type, return a vector type of the
1084 // appropriate width which we will bitcast. An exception is made for
1085 // returning structs of 2, 3, or 4 vectors which are returned in a sret-like
1086 // fashion, storing them to a pointer arg.
1087 if (RetT.getNumVectors() > 1) {
1088 S += "vv*"; // void result with void* first argument
1091 RetT.makeInteger(RetT.getElementSizeInBits(), false);
1092 if (!RetT.isScalar() && !RetT.isSigned())
1095 bool ForcedVectorFloatingType = isFloatingPointProtoModifier(Proto[0]);
1096 if (LocalCK == ClassB && !RetT.isScalar() && !ForcedVectorFloatingType)
1097 // Cast to vector of 8-bit elements.
1098 RetT.makeInteger(8, true);
1100 S += RetT.builtin_str();
1103 for (unsigned I = 0; I < getNumParams(); ++I) {
1104 Type T = getParamType(I);
1106 T.makeInteger(T.getElementSizeInBits(), false);
1108 bool ForcedFloatingType = isFloatingPointProtoModifier(Proto[I + 1]);
1109 if (LocalCK == ClassB && !T.isScalar() && !ForcedFloatingType)
1110 T.makeInteger(8, true);
1111 // Halves always get converted to 8-bit elements.
1112 if (T.isHalf() && T.isVector() && !T.isScalarForMangling())
1113 T.makeInteger(8, true);
1115 if (LocalCK == ClassI)
1118 if (hasImmediate() && getImmediateIdx() == I)
1119 T.makeImmediate(32);
1121 S += T.builtin_str();
1124 // Extra constant integer to hold type class enum for this function, e.g. s8
1125 if (LocalCK == ClassB)
1131 std::string Intrinsic::getMangledName(bool ForceClassS) const {
1132 // Check if the prototype has a scalar operand with the type of the vector
1133 // elements. If not, bitcasting the args will take care of arg checking.
1134 // The actual signedness etc. will be taken care of with special enums.
1135 ClassKind LocalCK = CK;
1136 if (!protoHasScalar())
1139 return mangleName(Name, ForceClassS ? ClassS : LocalCK);
1142 std::string Intrinsic::mangleName(std::string Name, ClassKind LocalCK) const {
1143 std::string typeCode = getInstTypeCode(BaseType, LocalCK);
1144 std::string S = Name;
1146 if (Name == "vcvt_f16_f32" || Name == "vcvt_f32_f16" ||
1147 Name == "vcvt_f32_f64" || Name == "vcvt_f64_f32")
1150 if (!typeCode.empty()) {
1151 // If the name ends with _xN (N = 2,3,4), insert the typeCode before _xN.
1152 if (Name.size() >= 3 && isdigit(Name.back()) &&
1153 Name[Name.length() - 2] == 'x' && Name[Name.length() - 3] == '_')
1154 S.insert(S.length() - 3, "_" + typeCode);
1156 S += "_" + typeCode;
1159 if (BaseType != InBaseType) {
1160 // A reinterpret - out the input base type at the end.
1161 S += "_" + getInstTypeCode(InBaseType, LocalCK);
1164 if (LocalCK == ClassB)
1167 // Insert a 'q' before the first '_' character so that it ends up before
1168 // _lane or _n on vector-scalar operations.
1169 if (BaseType.getSizeInBits() == 128 && !BaseType.noManglingQ()) {
1170 size_t Pos = S.find('_');
1175 if (BaseType.isScalarForMangling()) {
1176 switch (BaseType.getElementSizeInBits()) {
1177 case 8: Suffix = 'b'; break;
1178 case 16: Suffix = 'h'; break;
1179 case 32: Suffix = 's'; break;
1180 case 64: Suffix = 'd'; break;
1181 default: llvm_unreachable("Bad suffix!");
1184 if (Suffix != '\0') {
1185 size_t Pos = S.find('_');
1186 S.insert(Pos, &Suffix, 1);
1192 std::string Intrinsic::replaceParamsIn(std::string S) {
1193 while (S.find('$') != std::string::npos) {
1194 size_t Pos = S.find('$');
1195 size_t End = Pos + 1;
1196 while (isalpha(S[End]))
1199 std::string VarName = S.substr(Pos + 1, End - Pos - 1);
1200 assert_with_loc(Variables.find(VarName) != Variables.end(),
1201 "Variable not defined!");
1202 S.replace(Pos, End - Pos, Variables.find(VarName)->second.getName());
1208 void Intrinsic::initVariables() {
1211 // Modify the TypeSpec per-argument to get a concrete Type, and create
1212 // known variables for each.
1213 for (unsigned I = 1; I < Proto.size(); ++I) {
1214 char NameC = '0' + (I - 1);
1215 std::string Name = "p";
1216 Name.push_back(NameC);
1218 Variables[Name] = Variable(Types[I], Name + VariablePostfix);
1220 RetVar = Variable(Types[0], "ret" + VariablePostfix);
1223 void Intrinsic::emitPrototype(StringRef NamePrefix) {
1227 OS << "__ai " << Types[0].str() << " ";
1229 OS << NamePrefix.str() << mangleName(Name, ClassS) << "(";
1231 for (unsigned I = 0; I < getNumParams(); ++I) {
1235 char NameC = '0' + I;
1236 std::string Name = "p";
1237 Name.push_back(NameC);
1238 assert(Variables.find(Name) != Variables.end());
1239 Variable &V = Variables[Name];
1242 OS << V.getType().str() << " ";
1249 void Intrinsic::emitOpeningBrace() {
1251 OS << " __extension__ ({";
1257 void Intrinsic::emitClosingBrace() {
1264 void Intrinsic::emitNewLine() {
1271 void Intrinsic::emitReverseVariable(Variable &Dest, Variable &Src) {
1272 if (Dest.getType().getNumVectors() > 1) {
1275 for (unsigned K = 0; K < Dest.getType().getNumVectors(); ++K) {
1276 OS << " " << Dest.getName() << ".val[" << K << "] = "
1277 << "__builtin_shufflevector("
1278 << Src.getName() << ".val[" << K << "], "
1279 << Src.getName() << ".val[" << K << "]";
1280 for (int J = Dest.getType().getNumElements() - 1; J >= 0; --J)
1286 OS << " " << Dest.getName()
1287 << " = __builtin_shufflevector(" << Src.getName() << ", " << Src.getName();
1288 for (int J = Dest.getType().getNumElements() - 1; J >= 0; --J)
1295 void Intrinsic::emitArgumentReversal() {
1299 // Reverse all vector arguments.
1300 for (unsigned I = 0; I < getNumParams(); ++I) {
1301 std::string Name = "p" + utostr(I);
1302 std::string NewName = "rev" + utostr(I);
1304 Variable &V = Variables[Name];
1305 Variable NewV(V.getType(), NewName + VariablePostfix);
1307 if (!NewV.getType().isVector() || NewV.getType().getNumElements() == 1)
1310 OS << " " << NewV.getType().str() << " " << NewV.getName() << ";";
1311 emitReverseVariable(NewV, V);
1316 void Intrinsic::emitReturnReversal() {
1319 if (!getReturnType().isVector() || getReturnType().isVoid() ||
1320 getReturnType().getNumElements() == 1)
1322 emitReverseVariable(RetVar, RetVar);
1325 void Intrinsic::emitShadowedArgs() {
1326 // Macro arguments are not type-checked like inline function arguments,
1327 // so assign them to local temporaries to get the right type checking.
1331 for (unsigned I = 0; I < getNumParams(); ++I) {
1332 // Do not create a temporary for an immediate argument.
1333 // That would defeat the whole point of using a macro!
1334 if (hasImmediate() && Proto[I+1] == 'i')
1336 // Do not create a temporary for pointer arguments. The input
1337 // pointer may have an alignment hint.
1338 if (getParamType(I).isPointer())
1341 std::string Name = "p" + utostr(I);
1343 assert(Variables.find(Name) != Variables.end());
1344 Variable &V = Variables[Name];
1346 std::string NewName = "s" + utostr(I);
1347 Variable V2(V.getType(), NewName + VariablePostfix);
1349 OS << " " << V2.getType().str() << " " << V2.getName() << " = "
1350 << V.getName() << ";";
1357 // We don't check 'a' in this function, because for builtin function the
1358 // argument matching to 'a' uses a vector type splatted from a scalar type.
1359 bool Intrinsic::protoHasScalar() const {
1360 return (Proto.find('s') != std::string::npos ||
1361 Proto.find('z') != std::string::npos ||
1362 Proto.find('r') != std::string::npos ||
1363 Proto.find('b') != std::string::npos ||
1364 Proto.find('$') != std::string::npos ||
1365 Proto.find('y') != std::string::npos ||
1366 Proto.find('o') != std::string::npos);
1369 void Intrinsic::emitBodyAsBuiltinCall() {
1372 // If this builtin returns a struct 2, 3, or 4 vectors, pass it as an implicit
1373 // sret-like argument.
1374 bool SRet = getReturnType().getNumVectors() >= 2;
1378 // Call the non-splat builtin: chop off the "_n" suffix from the name.
1379 assert(N.endswith("_n"));
1383 ClassKind LocalCK = CK;
1384 if (!protoHasScalar())
1387 if (!getReturnType().isVoid() && !SRet)
1388 S += "(" + RetVar.getType().str() + ") ";
1390 S += "__builtin_neon_" + mangleName(N, LocalCK) + "(";
1393 S += "&" + RetVar.getName() + ", ";
1395 for (unsigned I = 0; I < getNumParams(); ++I) {
1396 Variable &V = Variables["p" + utostr(I)];
1397 Type T = V.getType();
1399 // Handle multiple-vector values specially, emitting each subvector as an
1400 // argument to the builtin.
1401 if (T.getNumVectors() > 1) {
1402 // Check if an explicit cast is needed.
1404 if (T.isChar() || T.isPoly() || !T.isSigned()) {
1407 T2.makeInteger(8, /*Signed=*/true);
1408 Cast = "(" + T2.str() + ")";
1411 for (unsigned J = 0; J < T.getNumVectors(); ++J)
1412 S += Cast + V.getName() + ".val[" + utostr(J) + "], ";
1417 Type CastToType = T;
1418 if (hasSplat() && I == getSplatIdx()) {
1419 Arg = "(" + BaseType.str() + ") {";
1420 for (unsigned J = 0; J < BaseType.getNumElements(); ++J) {
1427 CastToType = BaseType;
1432 // Check if an explicit cast is needed.
1433 if (CastToType.isVector()) {
1434 CastToType.makeInteger(8, true);
1435 Arg = "(" + CastToType.str() + ")" + Arg;
1441 // Extra constant integer to hold type class enum for this function, e.g. s8
1442 if (getClassKind(true) == ClassB) {
1443 Type ThisTy = getReturnType();
1444 if (Proto[0] == 'v' || isFloatingPointProtoModifier(Proto[0]))
1445 ThisTy = getParamType(0);
1446 if (ThisTy.isPointer())
1447 ThisTy = getParamType(1);
1449 S += utostr(ThisTy.getNeonEnum());
1451 // Remove extraneous ", ".
1457 std::string RetExpr;
1458 if (!SRet && !RetVar.getType().isVoid())
1459 RetExpr = RetVar.getName() + " = ";
1461 OS << " " << RetExpr << S;
1465 void Intrinsic::emitBody(StringRef CallPrefix) {
1466 std::vector<std::string> Lines;
1468 assert(RetVar.getType() == Types[0]);
1469 // Create a return variable, if we're not void.
1470 if (!RetVar.getType().isVoid()) {
1471 OS << " " << RetVar.getType().str() << " " << RetVar.getName() << ";";
1475 if (!Body || Body->getValues().empty()) {
1476 // Nothing specific to output - must output a builtin.
1477 emitBodyAsBuiltinCall();
1481 // We have a list of "things to output". The last should be returned.
1482 for (auto *I : Body->getValues()) {
1483 if (StringInit *SI = dyn_cast<StringInit>(I)) {
1484 Lines.push_back(replaceParamsIn(SI->getAsString()));
1485 } else if (DagInit *DI = dyn_cast<DagInit>(I)) {
1486 DagEmitter DE(*this, CallPrefix);
1487 Lines.push_back(DE.emitDag(DI).second + ";");
1491 assert(!Lines.empty() && "Empty def?");
1492 if (!RetVar.getType().isVoid())
1493 Lines.back().insert(0, RetVar.getName() + " = ");
1495 for (auto &L : Lines) {
1501 void Intrinsic::emitReturn() {
1502 if (RetVar.getType().isVoid())
1505 OS << " " << RetVar.getName() << ";";
1507 OS << " return " << RetVar.getName() << ";";
1511 std::pair<Type, std::string> Intrinsic::DagEmitter::emitDag(DagInit *DI) {
1512 // At this point we should only be seeing a def.
1513 DefInit *DefI = cast<DefInit>(DI->getOperator());
1514 std::string Op = DefI->getAsString();
1516 if (Op == "cast" || Op == "bitcast")
1517 return emitDagCast(DI, Op == "bitcast");
1518 if (Op == "shuffle")
1519 return emitDagShuffle(DI);
1521 return emitDagDup(DI);
1522 if (Op == "dup_typed")
1523 return emitDagDupTyped(DI);
1525 return emitDagSplat(DI);
1526 if (Op == "save_temp")
1527 return emitDagSaveTemp(DI);
1529 return emitDagOp(DI);
1531 return emitDagCall(DI);
1532 if (Op == "name_replace")
1533 return emitDagNameReplace(DI);
1534 if (Op == "literal")
1535 return emitDagLiteral(DI);
1536 assert_with_loc(false, "Unknown operation!");
1537 return std::make_pair(Type::getVoid(), "");
1540 std::pair<Type, std::string> Intrinsic::DagEmitter::emitDagOp(DagInit *DI) {
1541 std::string Op = cast<StringInit>(DI->getArg(0))->getAsUnquotedString();
1542 if (DI->getNumArgs() == 2) {
1544 std::pair<Type, std::string> R =
1545 emitDagArg(DI->getArg(1), DI->getArgNameStr(1));
1546 return std::make_pair(R.first, Op + R.second);
1548 assert(DI->getNumArgs() == 3 && "Can only handle unary and binary ops!");
1549 std::pair<Type, std::string> R1 =
1550 emitDagArg(DI->getArg(1), DI->getArgNameStr(1));
1551 std::pair<Type, std::string> R2 =
1552 emitDagArg(DI->getArg(2), DI->getArgNameStr(2));
1553 assert_with_loc(R1.first == R2.first, "Argument type mismatch!");
1554 return std::make_pair(R1.first, R1.second + " " + Op + " " + R2.second);
1558 std::pair<Type, std::string> Intrinsic::DagEmitter::emitDagCall(DagInit *DI) {
1559 std::vector<Type> Types;
1560 std::vector<std::string> Values;
1561 for (unsigned I = 0; I < DI->getNumArgs() - 1; ++I) {
1562 std::pair<Type, std::string> R =
1563 emitDagArg(DI->getArg(I + 1), DI->getArgNameStr(I + 1));
1564 Types.push_back(R.first);
1565 Values.push_back(R.second);
1568 // Look up the called intrinsic.
1570 if (StringInit *SI = dyn_cast<StringInit>(DI->getArg(0)))
1571 N = SI->getAsUnquotedString();
1573 N = emitDagArg(DI->getArg(0), "").second;
1574 Intrinsic &Callee = Intr.Emitter.getIntrinsic(N, Types);
1576 // Make sure the callee is known as an early def.
1577 Callee.setNeededEarly();
1578 Intr.Dependencies.insert(&Callee);
1580 // Now create the call itself.
1581 std::string S = CallPrefix.str() + Callee.getMangledName(true) + "(";
1582 for (unsigned I = 0; I < DI->getNumArgs() - 1; ++I) {
1589 return std::make_pair(Callee.getReturnType(), S);
1592 std::pair<Type, std::string> Intrinsic::DagEmitter::emitDagCast(DagInit *DI,
1594 // (cast MOD* VAL) -> cast VAL to type given by MOD.
1595 std::pair<Type, std::string> R = emitDagArg(
1596 DI->getArg(DI->getNumArgs() - 1),
1597 DI->getArgNameStr(DI->getNumArgs() - 1));
1598 Type castToType = R.first;
1599 for (unsigned ArgIdx = 0; ArgIdx < DI->getNumArgs() - 1; ++ArgIdx) {
1601 // MOD can take several forms:
1602 // 1. $X - take the type of parameter / variable X.
1603 // 2. The value "R" - take the type of the return type.
1605 // 4. The value "U" or "S" to switch the signedness.
1606 // 5. The value "H" or "D" to half or double the bitwidth.
1607 // 6. The value "8" to convert to 8-bit (signed) integer lanes.
1608 if (!DI->getArgNameStr(ArgIdx).empty()) {
1609 assert_with_loc(Intr.Variables.find(DI->getArgNameStr(ArgIdx)) !=
1610 Intr.Variables.end(),
1611 "Variable not found");
1612 castToType = Intr.Variables[DI->getArgNameStr(ArgIdx)].getType();
1614 StringInit *SI = dyn_cast<StringInit>(DI->getArg(ArgIdx));
1615 assert_with_loc(SI, "Expected string type or $Name for cast type");
1617 if (SI->getAsUnquotedString() == "R") {
1618 castToType = Intr.getReturnType();
1619 } else if (SI->getAsUnquotedString() == "U") {
1620 castToType.makeUnsigned();
1621 } else if (SI->getAsUnquotedString() == "S") {
1622 castToType.makeSigned();
1623 } else if (SI->getAsUnquotedString() == "H") {
1624 castToType.halveLanes();
1625 } else if (SI->getAsUnquotedString() == "D") {
1626 castToType.doubleLanes();
1627 } else if (SI->getAsUnquotedString() == "8") {
1628 castToType.makeInteger(8, true);
1630 castToType = Type::fromTypedefName(SI->getAsUnquotedString());
1631 assert_with_loc(!castToType.isVoid(), "Unknown typedef");
1638 // Emit a reinterpret cast. The second operand must be an lvalue, so create
1640 std::string N = "reint";
1642 while (Intr.Variables.find(N) != Intr.Variables.end())
1643 N = "reint" + utostr(++I);
1644 Intr.Variables[N] = Variable(R.first, N + Intr.VariablePostfix);
1646 Intr.OS << R.first.str() << " " << Intr.Variables[N].getName() << " = "
1650 S = "*(" + castToType.str() + " *) &" + Intr.Variables[N].getName() + "";
1652 // Emit a normal (static) cast.
1653 S = "(" + castToType.str() + ")(" + R.second + ")";
1656 return std::make_pair(castToType, S);
1659 std::pair<Type, std::string> Intrinsic::DagEmitter::emitDagShuffle(DagInit *DI){
1660 // See the documentation in arm_neon.td for a description of these operators.
1661 class LowHalf : public SetTheory::Operator {
1663 void apply(SetTheory &ST, DagInit *Expr, SetTheory::RecSet &Elts,
1664 ArrayRef<SMLoc> Loc) override {
1665 SetTheory::RecSet Elts2;
1666 ST.evaluate(Expr->arg_begin(), Expr->arg_end(), Elts2, Loc);
1667 Elts.insert(Elts2.begin(), Elts2.begin() + (Elts2.size() / 2));
1671 class HighHalf : public SetTheory::Operator {
1673 void apply(SetTheory &ST, DagInit *Expr, SetTheory::RecSet &Elts,
1674 ArrayRef<SMLoc> Loc) override {
1675 SetTheory::RecSet Elts2;
1676 ST.evaluate(Expr->arg_begin(), Expr->arg_end(), Elts2, Loc);
1677 Elts.insert(Elts2.begin() + (Elts2.size() / 2), Elts2.end());
1681 class Rev : public SetTheory::Operator {
1682 unsigned ElementSize;
1685 Rev(unsigned ElementSize) : ElementSize(ElementSize) {}
1687 void apply(SetTheory &ST, DagInit *Expr, SetTheory::RecSet &Elts,
1688 ArrayRef<SMLoc> Loc) override {
1689 SetTheory::RecSet Elts2;
1690 ST.evaluate(Expr->arg_begin() + 1, Expr->arg_end(), Elts2, Loc);
1692 int64_t VectorSize = cast<IntInit>(Expr->getArg(0))->getValue();
1693 VectorSize /= ElementSize;
1695 std::vector<Record *> Revved;
1696 for (unsigned VI = 0; VI < Elts2.size(); VI += VectorSize) {
1697 for (int LI = VectorSize - 1; LI >= 0; --LI) {
1698 Revved.push_back(Elts2[VI + LI]);
1702 Elts.insert(Revved.begin(), Revved.end());
1706 class MaskExpander : public SetTheory::Expander {
1710 MaskExpander(unsigned N) : N(N) {}
1712 void expand(SetTheory &ST, Record *R, SetTheory::RecSet &Elts) override {
1713 unsigned Addend = 0;
1714 if (R->getName() == "mask0")
1716 else if (R->getName() == "mask1")
1720 for (unsigned I = 0; I < N; ++I)
1721 Elts.insert(R->getRecords().getDef("sv" + utostr(I + Addend)));
1725 // (shuffle arg1, arg2, sequence)
1726 std::pair<Type, std::string> Arg1 =
1727 emitDagArg(DI->getArg(0), DI->getArgNameStr(0));
1728 std::pair<Type, std::string> Arg2 =
1729 emitDagArg(DI->getArg(1), DI->getArgNameStr(1));
1730 assert_with_loc(Arg1.first == Arg2.first,
1731 "Different types in arguments to shuffle!");
1734 SetTheory::RecSet Elts;
1735 ST.addOperator("lowhalf", llvm::make_unique<LowHalf>());
1736 ST.addOperator("highhalf", llvm::make_unique<HighHalf>());
1737 ST.addOperator("rev",
1738 llvm::make_unique<Rev>(Arg1.first.getElementSizeInBits()));
1739 ST.addExpander("MaskExpand",
1740 llvm::make_unique<MaskExpander>(Arg1.first.getNumElements()));
1741 ST.evaluate(DI->getArg(2), Elts, None);
1743 std::string S = "__builtin_shufflevector(" + Arg1.second + ", " + Arg2.second;
1744 for (auto &E : Elts) {
1745 StringRef Name = E->getName();
1746 assert_with_loc(Name.startswith("sv"),
1747 "Incorrect element kind in shuffle mask!");
1748 S += ", " + Name.drop_front(2).str();
1752 // Recalculate the return type - the shuffle may have halved or doubled it.
1754 if (Elts.size() > T.getNumElements()) {
1756 Elts.size() == T.getNumElements() * 2,
1757 "Can only double or half the number of elements in a shuffle!");
1759 } else if (Elts.size() < T.getNumElements()) {
1761 Elts.size() == T.getNumElements() / 2,
1762 "Can only double or half the number of elements in a shuffle!");
1766 return std::make_pair(T, S);
1769 std::pair<Type, std::string> Intrinsic::DagEmitter::emitDagDup(DagInit *DI) {
1770 assert_with_loc(DI->getNumArgs() == 1, "dup() expects one argument");
1771 std::pair<Type, std::string> A = emitDagArg(DI->getArg(0),
1772 DI->getArgNameStr(0));
1773 assert_with_loc(A.first.isScalar(), "dup() expects a scalar argument");
1775 Type T = Intr.getBaseType();
1776 assert_with_loc(T.isVector(), "dup() used but default type is scalar!");
1777 std::string S = "(" + T.str() + ") {";
1778 for (unsigned I = 0; I < T.getNumElements(); ++I) {
1785 return std::make_pair(T, S);
1788 std::pair<Type, std::string> Intrinsic::DagEmitter::emitDagDupTyped(DagInit *DI) {
1789 assert_with_loc(DI->getNumArgs() == 2, "dup_typed() expects two arguments");
1790 std::pair<Type, std::string> A = emitDagArg(DI->getArg(0),
1791 DI->getArgNameStr(0));
1792 std::pair<Type, std::string> B = emitDagArg(DI->getArg(1),
1793 DI->getArgNameStr(1));
1794 assert_with_loc(B.first.isScalar(),
1795 "dup_typed() requires a scalar as the second argument");
1798 assert_with_loc(T.isVector(), "dup_typed() used but target type is scalar!");
1799 std::string S = "(" + T.str() + ") {";
1800 for (unsigned I = 0; I < T.getNumElements(); ++I) {
1807 return std::make_pair(T, S);
1810 std::pair<Type, std::string> Intrinsic::DagEmitter::emitDagSplat(DagInit *DI) {
1811 assert_with_loc(DI->getNumArgs() == 2, "splat() expects two arguments");
1812 std::pair<Type, std::string> A = emitDagArg(DI->getArg(0),
1813 DI->getArgNameStr(0));
1814 std::pair<Type, std::string> B = emitDagArg(DI->getArg(1),
1815 DI->getArgNameStr(1));
1817 assert_with_loc(B.first.isScalar(),
1818 "splat() requires a scalar int as the second argument");
1820 std::string S = "__builtin_shufflevector(" + A.second + ", " + A.second;
1821 for (unsigned I = 0; I < Intr.getBaseType().getNumElements(); ++I) {
1822 S += ", " + B.second;
1826 return std::make_pair(Intr.getBaseType(), S);
1829 std::pair<Type, std::string> Intrinsic::DagEmitter::emitDagSaveTemp(DagInit *DI) {
1830 assert_with_loc(DI->getNumArgs() == 2, "save_temp() expects two arguments");
1831 std::pair<Type, std::string> A = emitDagArg(DI->getArg(1),
1832 DI->getArgNameStr(1));
1834 assert_with_loc(!A.first.isVoid(),
1835 "Argument to save_temp() must have non-void type!");
1837 std::string N = DI->getArgNameStr(0);
1838 assert_with_loc(!N.empty(),
1839 "save_temp() expects a name as the first argument");
1841 assert_with_loc(Intr.Variables.find(N) == Intr.Variables.end(),
1842 "Variable already defined!");
1843 Intr.Variables[N] = Variable(A.first, N + Intr.VariablePostfix);
1846 A.first.str() + " " + Intr.Variables[N].getName() + " = " + A.second;
1848 return std::make_pair(Type::getVoid(), S);
1851 std::pair<Type, std::string>
1852 Intrinsic::DagEmitter::emitDagNameReplace(DagInit *DI) {
1853 std::string S = Intr.Name;
1855 assert_with_loc(DI->getNumArgs() == 2, "name_replace requires 2 arguments!");
1856 std::string ToReplace = cast<StringInit>(DI->getArg(0))->getAsUnquotedString();
1857 std::string ReplaceWith = cast<StringInit>(DI->getArg(1))->getAsUnquotedString();
1859 size_t Idx = S.find(ToReplace);
1861 assert_with_loc(Idx != std::string::npos, "name should contain '" + ToReplace + "'!");
1862 S.replace(Idx, ToReplace.size(), ReplaceWith);
1864 return std::make_pair(Type::getVoid(), S);
1867 std::pair<Type, std::string> Intrinsic::DagEmitter::emitDagLiteral(DagInit *DI){
1868 std::string Ty = cast<StringInit>(DI->getArg(0))->getAsUnquotedString();
1869 std::string Value = cast<StringInit>(DI->getArg(1))->getAsUnquotedString();
1870 return std::make_pair(Type::fromTypedefName(Ty), Value);
1873 std::pair<Type, std::string>
1874 Intrinsic::DagEmitter::emitDagArg(Init *Arg, std::string ArgName) {
1875 if (!ArgName.empty()) {
1876 assert_with_loc(!Arg->isComplete(),
1877 "Arguments must either be DAGs or names, not both!");
1878 assert_with_loc(Intr.Variables.find(ArgName) != Intr.Variables.end(),
1879 "Variable not defined!");
1880 Variable &V = Intr.Variables[ArgName];
1881 return std::make_pair(V.getType(), V.getName());
1884 assert(Arg && "Neither ArgName nor Arg?!");
1885 DagInit *DI = dyn_cast<DagInit>(Arg);
1886 assert_with_loc(DI, "Arguments must either be DAGs or names!");
1891 std::string Intrinsic::generate() {
1892 // Little endian intrinsics are simple and don't require any argument
1894 OS << "#ifdef __LITTLE_ENDIAN__\n";
1896 generateImpl(false, "", "");
1900 // Big endian intrinsics are more complex. The user intended these
1901 // intrinsics to operate on a vector "as-if" loaded by (V)LDR,
1902 // but we load as-if (V)LD1. So we should swap all arguments and
1903 // swap the return value too.
1905 // If we call sub-intrinsics, we should call a version that does
1906 // not re-swap the arguments!
1907 generateImpl(true, "", "__noswap_");
1909 // If we're needed early, create a non-swapping variant for
1912 generateImpl(false, "__noswap_", "__noswap_");
1919 void Intrinsic::generateImpl(bool ReverseArguments,
1920 StringRef NamePrefix, StringRef CallPrefix) {
1923 // If we call a macro, our local variables may be corrupted due to
1924 // lack of proper lexical scoping. So, add a globally unique postfix
1925 // to every variable.
1927 // indexBody() should have set up the Dependencies set by now.
1928 for (auto *I : Dependencies)
1930 VariablePostfix = "_" + utostr(Emitter.getUniqueNumber());
1936 emitPrototype(NamePrefix);
1938 if (IsUnavailable) {
1939 OS << " __attribute__((unavailable));";
1943 if (ReverseArguments)
1944 emitArgumentReversal();
1945 emitBody(CallPrefix);
1946 if (ReverseArguments)
1947 emitReturnReversal();
1953 CurrentRecord = nullptr;
1956 void Intrinsic::indexBody() {
1963 CurrentRecord = nullptr;
1966 //===----------------------------------------------------------------------===//
1967 // NeonEmitter implementation
1968 //===----------------------------------------------------------------------===//
1970 Intrinsic &NeonEmitter::getIntrinsic(StringRef Name, ArrayRef<Type> Types) {
1971 // First, look up the name in the intrinsic map.
1972 assert_with_loc(IntrinsicMap.find(Name.str()) != IntrinsicMap.end(),
1973 ("Intrinsic '" + Name + "' not found!").str());
1974 auto &V = IntrinsicMap.find(Name.str())->second;
1975 std::vector<Intrinsic *> GoodVec;
1977 // Create a string to print if we end up failing.
1978 std::string ErrMsg = "looking up intrinsic '" + Name.str() + "(";
1979 for (unsigned I = 0; I < Types.size(); ++I) {
1982 ErrMsg += Types[I].str();
1985 ErrMsg += "Available overloads:\n";
1987 // Now, look through each intrinsic implementation and see if the types are
1990 ErrMsg += " - " + I.getReturnType().str() + " " + I.getMangledName();
1992 for (unsigned A = 0; A < I.getNumParams(); ++A) {
1995 ErrMsg += I.getParamType(A).str();
1999 if (I.getNumParams() != Types.size())
2003 for (unsigned Arg = 0; Arg < Types.size(); ++Arg) {
2004 if (I.getParamType(Arg) != Types[Arg]) {
2010 GoodVec.push_back(&I);
2013 assert_with_loc(!GoodVec.empty(),
2014 "No compatible intrinsic found - " + ErrMsg);
2015 assert_with_loc(GoodVec.size() == 1, "Multiple overloads found - " + ErrMsg);
2017 return *GoodVec.front();
2020 void NeonEmitter::createIntrinsic(Record *R,
2021 SmallVectorImpl<Intrinsic *> &Out) {
2022 std::string Name = R->getValueAsString("Name");
2023 std::string Proto = R->getValueAsString("Prototype");
2024 std::string Types = R->getValueAsString("Types");
2025 Record *OperationRec = R->getValueAsDef("Operation");
2026 bool CartesianProductOfTypes = R->getValueAsBit("CartesianProductOfTypes");
2027 bool BigEndianSafe = R->getValueAsBit("BigEndianSafe");
2028 std::string Guard = R->getValueAsString("ArchGuard");
2029 bool IsUnavailable = OperationRec->getValueAsBit("Unavailable");
2031 // Set the global current record. This allows assert_with_loc to produce
2032 // decent location information even when highly nested.
2035 ListInit *Body = OperationRec->getValueAsListInit("Ops");
2037 std::vector<TypeSpec> TypeSpecs = TypeSpec::fromTypeSpecs(Types);
2039 ClassKind CK = ClassNone;
2040 if (R->getSuperClasses().size() >= 2)
2041 CK = ClassMap[R->getSuperClasses()[1].first];
2043 std::vector<std::pair<TypeSpec, TypeSpec>> NewTypeSpecs;
2044 for (auto TS : TypeSpecs) {
2045 if (CartesianProductOfTypes) {
2046 Type DefaultT(TS, 'd');
2047 for (auto SrcTS : TypeSpecs) {
2048 Type DefaultSrcT(SrcTS, 'd');
2050 DefaultSrcT.getSizeInBits() != DefaultT.getSizeInBits())
2052 NewTypeSpecs.push_back(std::make_pair(TS, SrcTS));
2055 NewTypeSpecs.push_back(std::make_pair(TS, TS));
2059 llvm::sort(NewTypeSpecs);
2060 NewTypeSpecs.erase(std::unique(NewTypeSpecs.begin(), NewTypeSpecs.end()),
2061 NewTypeSpecs.end());
2062 auto &Entry = IntrinsicMap[Name];
2064 for (auto &I : NewTypeSpecs) {
2065 Entry.emplace_back(R, Name, Proto, I.first, I.second, CK, Body, *this,
2066 Guard, IsUnavailable, BigEndianSafe);
2067 Out.push_back(&Entry.back());
2070 CurrentRecord = nullptr;
2073 /// genBuiltinsDef: Generate the BuiltinsARM.def and BuiltinsAArch64.def
2074 /// declaration of builtins, checking for unique builtin declarations.
2075 void NeonEmitter::genBuiltinsDef(raw_ostream &OS,
2076 SmallVectorImpl<Intrinsic *> &Defs) {
2077 OS << "#ifdef GET_NEON_BUILTINS\n";
2079 // We only want to emit a builtin once, and we want to emit them in
2080 // alphabetical order, so use a std::set.
2081 std::set<std::string> Builtins;
2083 for (auto *Def : Defs) {
2086 // Functions with 'a' (the splat code) in the type prototype should not get
2087 // their own builtin as they use the non-splat variant.
2088 if (Def->hasSplat())
2091 std::string S = "BUILTIN(__builtin_neon_" + Def->getMangledName() + ", \"";
2093 S += Def->getBuiltinTypeStr();
2099 for (auto &S : Builtins)
2104 /// Generate the ARM and AArch64 overloaded type checking code for
2105 /// SemaChecking.cpp, checking for unique builtin declarations.
2106 void NeonEmitter::genOverloadTypeCheckCode(raw_ostream &OS,
2107 SmallVectorImpl<Intrinsic *> &Defs) {
2108 OS << "#ifdef GET_NEON_OVERLOAD_CHECK\n";
2110 // We record each overload check line before emitting because subsequent Inst
2111 // definitions may extend the number of permitted types (i.e. augment the
2112 // Mask). Use std::map to avoid sorting the table by hash number.
2113 struct OverloadInfo {
2117 OverloadInfo() : Mask(0ULL), PtrArgNum(0), HasConstPtr(false) {}
2119 std::map<std::string, OverloadInfo> OverloadMap;
2121 for (auto *Def : Defs) {
2122 // If the def has a body (that is, it has Operation DAGs), it won't call
2123 // __builtin_neon_* so we don't need to generate a definition for it.
2126 // Functions with 'a' (the splat code) in the type prototype should not get
2127 // their own builtin as they use the non-splat variant.
2128 if (Def->hasSplat())
2130 // Functions which have a scalar argument cannot be overloaded, no need to
2131 // check them if we are emitting the type checking code.
2132 if (Def->protoHasScalar())
2135 uint64_t Mask = 0ULL;
2136 Type Ty = Def->getReturnType();
2137 if (Def->getProto()[0] == 'v' ||
2138 isFloatingPointProtoModifier(Def->getProto()[0]))
2139 Ty = Def->getParamType(0);
2141 Ty = Def->getParamType(1);
2143 Mask |= 1ULL << Ty.getNeonEnum();
2145 // Check if the function has a pointer or const pointer argument.
2146 std::string Proto = Def->getProto();
2148 bool HasConstPtr = false;
2149 for (unsigned I = 0; I < Def->getNumParams(); ++I) {
2150 char ArgType = Proto[I + 1];
2151 if (ArgType == 'c') {
2156 if (ArgType == 'p') {
2161 // For sret builtins, adjust the pointer argument index.
2162 if (PtrArgNum >= 0 && Def->getReturnType().getNumVectors() > 1)
2165 std::string Name = Def->getName();
2166 // Omit type checking for the pointer arguments of vld1_lane, vld1_dup,
2167 // and vst1_lane intrinsics. Using a pointer to the vector element
2168 // type with one of those operations causes codegen to select an aligned
2169 // load/store instruction. If you want an unaligned operation,
2170 // the pointer argument needs to have less alignment than element type,
2171 // so just accept any pointer type.
2172 if (Name == "vld1_lane" || Name == "vld1_dup" || Name == "vst1_lane") {
2174 HasConstPtr = false;
2178 std::string Name = Def->getMangledName();
2179 OverloadMap.insert(std::make_pair(Name, OverloadInfo()));
2180 OverloadInfo &OI = OverloadMap[Name];
2182 OI.PtrArgNum |= PtrArgNum;
2183 OI.HasConstPtr = HasConstPtr;
2187 for (auto &I : OverloadMap) {
2188 OverloadInfo &OI = I.second;
2190 OS << "case NEON::BI__builtin_neon_" << I.first << ": ";
2191 OS << "mask = 0x" << Twine::utohexstr(OI.Mask) << "ULL";
2192 if (OI.PtrArgNum >= 0)
2193 OS << "; PtrArgNum = " << OI.PtrArgNum;
2195 OS << "; HasConstPtr = true";
2201 void NeonEmitter::genIntrinsicRangeCheckCode(raw_ostream &OS,
2202 SmallVectorImpl<Intrinsic *> &Defs) {
2203 OS << "#ifdef GET_NEON_IMMEDIATE_CHECK\n";
2205 std::set<std::string> Emitted;
2207 for (auto *Def : Defs) {
2210 // Functions with 'a' (the splat code) in the type prototype should not get
2211 // their own builtin as they use the non-splat variant.
2212 if (Def->hasSplat())
2214 // Functions which do not have an immediate do not need to have range
2215 // checking code emitted.
2216 if (!Def->hasImmediate())
2218 if (Emitted.find(Def->getMangledName()) != Emitted.end())
2221 std::string LowerBound, UpperBound;
2223 Record *R = Def->getRecord();
2224 if (R->getValueAsBit("isVCVT_N")) {
2225 // VCVT between floating- and fixed-point values takes an immediate
2226 // in the range [1, 32) for f32 or [1, 64) for f64 or [1, 16) for f16.
2228 if (Def->getBaseType().getElementSizeInBits() == 16 ||
2229 Def->getName().find('h') != std::string::npos)
2230 // VCVTh operating on FP16 intrinsics in range [1, 16)
2232 else if (Def->getBaseType().getElementSizeInBits() == 32)
2236 } else if (R->getValueAsBit("isScalarShift")) {
2237 // Right shifts have an 'r' in the name, left shifts do not. Convert
2238 // instructions have the same bounds and right shifts.
2239 if (Def->getName().find('r') != std::string::npos ||
2240 Def->getName().find("cvt") != std::string::npos)
2243 UpperBound = utostr(Def->getReturnType().getElementSizeInBits() - 1);
2244 } else if (R->getValueAsBit("isShift")) {
2245 // Builtins which are overloaded by type will need to have their upper
2246 // bound computed at Sema time based on the type constant.
2248 // Right shifts have an 'r' in the name, left shifts do not.
2249 if (Def->getName().find('r') != std::string::npos)
2251 UpperBound = "RFT(TV, true)";
2252 } else if (Def->getClassKind(true) == ClassB) {
2253 // ClassB intrinsics have a type (and hence lane number) that is only
2254 // known at runtime.
2255 if (R->getValueAsBit("isLaneQ"))
2256 UpperBound = "RFT(TV, false, true)";
2258 UpperBound = "RFT(TV, false, false)";
2260 // The immediate generally refers to a lane in the preceding argument.
2261 assert(Def->getImmediateIdx() > 0);
2262 Type T = Def->getParamType(Def->getImmediateIdx() - 1);
2263 UpperBound = utostr(T.getNumElements() - 1);
2266 // Calculate the index of the immediate that should be range checked.
2267 unsigned Idx = Def->getNumParams();
2268 if (Def->hasImmediate())
2269 Idx = Def->getGeneratedParamIdx(Def->getImmediateIdx());
2271 OS << "case NEON::BI__builtin_neon_" << Def->getMangledName() << ": "
2272 << "i = " << Idx << ";";
2273 if (!LowerBound.empty())
2274 OS << " l = " << LowerBound << ";";
2275 if (!UpperBound.empty())
2276 OS << " u = " << UpperBound << ";";
2279 Emitted.insert(Def->getMangledName());
2285 /// runHeader - Emit a file with sections defining:
2286 /// 1. the NEON section of BuiltinsARM.def and BuiltinsAArch64.def.
2287 /// 2. the SemaChecking code for the type overload checking.
2288 /// 3. the SemaChecking code for validation of intrinsic immediate arguments.
2289 void NeonEmitter::runHeader(raw_ostream &OS) {
2290 std::vector<Record *> RV = Records.getAllDerivedDefinitions("Inst");
2292 SmallVector<Intrinsic *, 128> Defs;
2294 createIntrinsic(R, Defs);
2296 // Generate shared BuiltinsXXX.def
2297 genBuiltinsDef(OS, Defs);
2299 // Generate ARM overloaded type checking code for SemaChecking.cpp
2300 genOverloadTypeCheckCode(OS, Defs);
2302 // Generate ARM range checking code for shift/lane immediates.
2303 genIntrinsicRangeCheckCode(OS, Defs);
2306 /// run - Read the records in arm_neon.td and output arm_neon.h. arm_neon.h
2307 /// is comprised of type definitions and function declarations.
2308 void NeonEmitter::run(raw_ostream &OS) {
2309 OS << "/*===---- arm_neon.h - ARM Neon intrinsics "
2310 "------------------------------"
2313 " * Permission is hereby granted, free of charge, to any person "
2316 " * of this software and associated documentation files (the "
2319 " * in the Software without restriction, including without limitation "
2322 " * to use, copy, modify, merge, publish, distribute, sublicense, "
2324 " * copies of the Software, and to permit persons to whom the Software "
2326 " * furnished to do so, subject to the following conditions:\n"
2328 " * The above copyright notice and this permission notice shall be "
2330 " * all copies or substantial portions of the Software.\n"
2332 " * THE SOFTWARE IS PROVIDED \"AS IS\", WITHOUT WARRANTY OF ANY KIND, "
2334 " * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF "
2335 "MERCHANTABILITY,\n"
2336 " * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT "
2338 " * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR "
2340 " * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, "
2342 " * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER "
2344 " * THE SOFTWARE.\n"
2346 " *===-----------------------------------------------------------------"
2351 OS << "#ifndef __ARM_NEON_H\n";
2352 OS << "#define __ARM_NEON_H\n\n";
2354 OS << "#if !defined(__ARM_NEON)\n";
2355 OS << "#error \"NEON support not enabled\"\n";
2358 OS << "#include <stdint.h>\n\n";
2360 // Emit NEON-specific scalar typedefs.
2361 OS << "typedef float float32_t;\n";
2362 OS << "typedef __fp16 float16_t;\n";
2364 OS << "#ifdef __aarch64__\n";
2365 OS << "typedef double float64_t;\n";
2368 // For now, signedness of polynomial types depends on target
2369 OS << "#ifdef __aarch64__\n";
2370 OS << "typedef uint8_t poly8_t;\n";
2371 OS << "typedef uint16_t poly16_t;\n";
2372 OS << "typedef uint64_t poly64_t;\n";
2373 OS << "typedef __uint128_t poly128_t;\n";
2375 OS << "typedef int8_t poly8_t;\n";
2376 OS << "typedef int16_t poly16_t;\n";
2379 // Emit Neon vector typedefs.
2380 std::string TypedefTypes(
2381 "cQcsQsiQilQlUcQUcUsQUsUiQUiUlQUlhQhfQfdQdPcQPcPsQPsPlQPl");
2382 std::vector<TypeSpec> TDTypeVec = TypeSpec::fromTypeSpecs(TypedefTypes);
2384 // Emit vector typedefs.
2385 bool InIfdef = false;
2386 for (auto &TS : TDTypeVec) {
2389 if (T.isDouble() || (T.isPoly() && T.isLong()))
2392 if (InIfdef && !IsA64) {
2396 if (!InIfdef && IsA64) {
2397 OS << "#ifdef __aarch64__\n";
2402 OS << "typedef __attribute__((neon_polyvector_type(";
2404 OS << "typedef __attribute__((neon_vector_type(";
2408 OS << T.getNumElements() << "))) ";
2410 OS << " " << T.str() << ";\n";
2416 // Emit struct typedefs.
2418 for (unsigned NumMembers = 2; NumMembers <= 4; ++NumMembers) {
2419 for (auto &TS : TDTypeVec) {
2422 if (T.isDouble() || (T.isPoly() && T.isLong()))
2425 if (InIfdef && !IsA64) {
2429 if (!InIfdef && IsA64) {
2430 OS << "#ifdef __aarch64__\n";
2434 char M = '2' + (NumMembers - 2);
2436 OS << "typedef struct " << VT.str() << " {\n";
2437 OS << " " << T.str() << " val";
2438 OS << "[" << NumMembers << "]";
2440 OS << VT.str() << ";\n";
2448 OS << "#define __ai static __inline__ __attribute__((__always_inline__, "
2449 "__nodebug__))\n\n";
2451 SmallVector<Intrinsic *, 128> Defs;
2452 std::vector<Record *> RV = Records.getAllDerivedDefinitions("Inst");
2454 createIntrinsic(R, Defs);
2456 for (auto *I : Defs)
2459 llvm::stable_sort(Defs, llvm::less_ptr<Intrinsic>());
2461 // Only emit a def when its requirements have been met.
2462 // FIXME: This loop could be made faster, but it's fast enough for now.
2463 bool MadeProgress = true;
2464 std::string InGuard;
2465 while (!Defs.empty() && MadeProgress) {
2466 MadeProgress = false;
2468 for (SmallVector<Intrinsic *, 128>::iterator I = Defs.begin();
2469 I != Defs.end(); /*No step*/) {
2470 bool DependenciesSatisfied = true;
2471 for (auto *II : (*I)->getDependencies()) {
2472 if (llvm::is_contained(Defs, II))
2473 DependenciesSatisfied = false;
2475 if (!DependenciesSatisfied) {
2476 // Try the next one.
2481 // Emit #endif/#if pair if needed.
2482 if ((*I)->getGuard() != InGuard) {
2483 if (!InGuard.empty())
2485 InGuard = (*I)->getGuard();
2486 if (!InGuard.empty())
2487 OS << "#if " << InGuard << "\n";
2490 // Actually generate the intrinsic code.
2491 OS << (*I)->generate();
2493 MadeProgress = true;
2497 assert(Defs.empty() && "Some requirements were not satisfied!");
2498 if (!InGuard.empty())
2502 OS << "#undef __ai\n\n";
2503 OS << "#endif /* __ARM_NEON_H */\n";
2506 /// run - Read the records in arm_fp16.td and output arm_fp16.h. arm_fp16.h
2507 /// is comprised of type definitions and function declarations.
2508 void NeonEmitter::runFP16(raw_ostream &OS) {
2509 OS << "/*===---- arm_fp16.h - ARM FP16 intrinsics "
2510 "------------------------------"
2513 " * Permission is hereby granted, free of charge, to any person "
2514 "obtaining a copy\n"
2515 " * of this software and associated documentation files (the "
2516 "\"Software\"), to deal\n"
2517 " * in the Software without restriction, including without limitation "
2519 " * to use, copy, modify, merge, publish, distribute, sublicense, "
2521 " * copies of the Software, and to permit persons to whom the Software "
2523 " * furnished to do so, subject to the following conditions:\n"
2525 " * The above copyright notice and this permission notice shall be "
2527 " * all copies or substantial portions of the Software.\n"
2529 " * THE SOFTWARE IS PROVIDED \"AS IS\", WITHOUT WARRANTY OF ANY KIND, "
2531 " * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF "
2532 "MERCHANTABILITY,\n"
2533 " * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT "
2535 " * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR "
2537 " * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, "
2539 " * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER "
2541 " * THE SOFTWARE.\n"
2543 " *===-----------------------------------------------------------------"
2548 OS << "#ifndef __ARM_FP16_H\n";
2549 OS << "#define __ARM_FP16_H\n\n";
2551 OS << "#include <stdint.h>\n\n";
2553 OS << "typedef __fp16 float16_t;\n";
2555 OS << "#define __ai static __inline__ __attribute__((__always_inline__, "
2556 "__nodebug__))\n\n";
2558 SmallVector<Intrinsic *, 128> Defs;
2559 std::vector<Record *> RV = Records.getAllDerivedDefinitions("Inst");
2561 createIntrinsic(R, Defs);
2563 for (auto *I : Defs)
2566 llvm::stable_sort(Defs, llvm::less_ptr<Intrinsic>());
2568 // Only emit a def when its requirements have been met.
2569 // FIXME: This loop could be made faster, but it's fast enough for now.
2570 bool MadeProgress = true;
2571 std::string InGuard;
2572 while (!Defs.empty() && MadeProgress) {
2573 MadeProgress = false;
2575 for (SmallVector<Intrinsic *, 128>::iterator I = Defs.begin();
2576 I != Defs.end(); /*No step*/) {
2577 bool DependenciesSatisfied = true;
2578 for (auto *II : (*I)->getDependencies()) {
2579 if (llvm::is_contained(Defs, II))
2580 DependenciesSatisfied = false;
2582 if (!DependenciesSatisfied) {
2583 // Try the next one.
2588 // Emit #endif/#if pair if needed.
2589 if ((*I)->getGuard() != InGuard) {
2590 if (!InGuard.empty())
2592 InGuard = (*I)->getGuard();
2593 if (!InGuard.empty())
2594 OS << "#if " << InGuard << "\n";
2597 // Actually generate the intrinsic code.
2598 OS << (*I)->generate();
2600 MadeProgress = true;
2604 assert(Defs.empty() && "Some requirements were not satisfied!");
2605 if (!InGuard.empty())
2609 OS << "#undef __ai\n\n";
2610 OS << "#endif /* __ARM_FP16_H */\n";
2615 void EmitNeon(RecordKeeper &Records, raw_ostream &OS) {
2616 NeonEmitter(Records).run(OS);
2619 void EmitFP16(RecordKeeper &Records, raw_ostream &OS) {
2620 NeonEmitter(Records).runFP16(OS);
2623 void EmitNeonSema(RecordKeeper &Records, raw_ostream &OS) {
2624 NeonEmitter(Records).runHeader(OS);
2627 void EmitNeonTest(RecordKeeper &Records, raw_ostream &OS) {
2628 llvm_unreachable("Neon test generation no longer implemented!");
2631 } // end namespace clang