1 //===- NeonEmitter.cpp - Generate arm_neon.h for use with clang -*- C++ -*-===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This tablegen backend is responsible for emitting arm_neon.h, which includes
11 // a declaration and definition of each function specified by the ARM NEON
12 // compiler interface. See ARM document DUI0348B.
14 // Each NEON instruction is implemented in terms of 1 or more functions which
15 // are suffixed with the element type of the input vectors. Functions may be
16 // implemented in terms of generic vector operations such as +, *, -, etc. or
17 // by calling a __builtin_-prefixed function which will be handled by clang's
20 // Additional validation code can be generated by this file when runHeader() is
21 // called, rather than the normal run() entry point.
23 // See also the documentation in include/clang/Basic/arm_neon.td.
25 //===----------------------------------------------------------------------===//
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.
143 bool Float, Signed, Immediate, Void, Poly, Constant, Pointer;
144 // ScalarForMangling and NoManglingQ are really not suited to live here as
145 // they are not related to the type. But they live in the TypeSpec (not the
146 // prototype), so this is really the only place to store them.
147 bool ScalarForMangling, NoManglingQ;
148 unsigned Bitwidth, ElementBitwidth, NumVectors;
152 : Float(false), Signed(false), Immediate(false), Void(true), Poly(false),
153 Constant(false), Pointer(false), ScalarForMangling(false),
154 NoManglingQ(false), Bitwidth(0), ElementBitwidth(0), NumVectors(0) {}
156 Type(TypeSpec TS, char CharMod)
157 : TS(std::move(TS)), Float(false), Signed(false), Immediate(false),
158 Void(false), Poly(false), Constant(false), Pointer(false),
159 ScalarForMangling(false), NoManglingQ(false), Bitwidth(0),
160 ElementBitwidth(0), NumVectors(0) {
161 applyModifier(CharMod);
164 /// Returns a type representing "void".
165 static Type getVoid() { return Type(); }
167 bool operator==(const Type &Other) const { return str() == Other.str(); }
168 bool operator!=(const Type &Other) const { return !operator==(Other); }
173 bool isScalarForMangling() const { return ScalarForMangling; }
174 bool noManglingQ() const { return NoManglingQ; }
176 bool isPointer() const { return Pointer; }
177 bool isFloating() const { return Float; }
178 bool isInteger() const { return !Float && !Poly; }
179 bool isSigned() const { return Signed; }
180 bool isImmediate() const { return Immediate; }
181 bool isScalar() const { return NumVectors == 0; }
182 bool isVector() const { return NumVectors > 0; }
183 bool isFloat() const { return Float && ElementBitwidth == 32; }
184 bool isDouble() const { return Float && ElementBitwidth == 64; }
185 bool isHalf() const { return Float && ElementBitwidth == 16; }
186 bool isPoly() const { return Poly; }
187 bool isChar() const { return ElementBitwidth == 8; }
188 bool isShort() const { return !Float && ElementBitwidth == 16; }
189 bool isInt() const { return !Float && ElementBitwidth == 32; }
190 bool isLong() const { return !Float && ElementBitwidth == 64; }
191 bool isVoid() const { return Void; }
192 unsigned getNumElements() const { return Bitwidth / ElementBitwidth; }
193 unsigned getSizeInBits() const { return Bitwidth; }
194 unsigned getElementSizeInBits() const { return ElementBitwidth; }
195 unsigned getNumVectors() const { return NumVectors; }
200 void makeUnsigned() { Signed = false; }
201 void makeSigned() { Signed = true; }
203 void makeInteger(unsigned ElemWidth, bool Sign) {
208 ElementBitwidth = ElemWidth;
211 void makeImmediate(unsigned ElemWidth) {
216 ElementBitwidth = ElemWidth;
220 Bitwidth = ElementBitwidth;
224 void makeOneVector() {
230 assert_with_loc(Bitwidth != 128, "Can't get bigger than 128!");
235 assert_with_loc(Bitwidth != 64, "Can't get smaller than 64!");
239 /// Return the C string representation of a type, which is the typename
240 /// defined in stdint.h or arm_neon.h.
241 std::string str() const;
243 /// Return the string representation of a type, which is an encoded
244 /// string for passing to the BUILTIN() macro in Builtins.def.
245 std::string builtin_str() const;
247 /// Return the value in NeonTypeFlags for this type.
248 unsigned getNeonEnum() const;
250 /// Parse a type from a stdint.h or arm_neon.h typedef name,
251 /// for example uint32x2_t or int64_t.
252 static Type fromTypedefName(StringRef Name);
255 /// Creates the type based on the typespec string in TS.
256 /// Sets "Quad" to true if the "Q" or "H" modifiers were
257 /// seen. This is needed by applyModifier as some modifiers
258 /// only take effect if the type size was changed by "Q" or "H".
259 void applyTypespec(bool &Quad);
260 /// Applies a prototype modifier to the type.
261 void applyModifier(char Mod);
264 //===----------------------------------------------------------------------===//
266 //===----------------------------------------------------------------------===//
268 /// A variable is a simple class that just has a type and a name.
274 Variable() : T(Type::getVoid()), N("") {}
275 Variable(Type T, std::string N) : T(std::move(T)), N(std::move(N)) {}
277 Type getType() const { return T; }
278 std::string getName() const { return "__" + N; }
281 //===----------------------------------------------------------------------===//
283 //===----------------------------------------------------------------------===//
285 /// The main grunt class. This represents an instantiation of an intrinsic with
286 /// a particular typespec and prototype.
288 friend class DagEmitter;
290 /// The Record this intrinsic was created from.
292 /// The unmangled name and prototype.
293 std::string Name, Proto;
294 /// The input and output typespecs. InTS == OutTS except when
295 /// CartesianProductOfTypes is 1 - this is the case for vreinterpret.
296 TypeSpec OutTS, InTS;
297 /// The base class kind. Most intrinsics use ClassS, which has full type
298 /// info for integers (s32/u32). Some use ClassI, which doesn't care about
299 /// signedness (i32), while some (ClassB) have no type at all, only a width
302 /// The list of DAGs for the body. May be empty, in which case we should
303 /// emit a builtin call.
305 /// The architectural #ifdef guard.
307 /// Set if the Unvailable bit is 1. This means we don't generate a body,
308 /// just an "unavailable" attribute on a declaration.
310 /// Is this intrinsic safe for big-endian? or does it need its arguments
314 /// The types of return value [0] and parameters [1..].
315 std::vector<Type> Types;
316 /// The local variables defined.
317 std::map<std::string, Variable> Variables;
318 /// NeededEarly - set if any other intrinsic depends on this intrinsic.
320 /// UseMacro - set if we should implement using a macro or unset for a
323 /// The set of intrinsics that this intrinsic uses/requires.
324 std::set<Intrinsic *> Dependencies;
325 /// The "base type", which is Type('d', OutTS). InBaseType is only
326 /// different if CartesianProductOfTypes = 1 (for vreinterpret).
327 Type BaseType, InBaseType;
328 /// The return variable.
330 /// A postfix to apply to every variable. Defaults to "".
331 std::string VariablePostfix;
333 NeonEmitter &Emitter;
334 std::stringstream OS;
337 Intrinsic(Record *R, StringRef Name, StringRef Proto, TypeSpec OutTS,
338 TypeSpec InTS, ClassKind CK, ListInit *Body, NeonEmitter &Emitter,
339 StringRef Guard, bool IsUnavailable, bool BigEndianSafe)
340 : R(R), Name(Name.str()), Proto(Proto.str()), OutTS(OutTS), InTS(InTS),
341 CK(CK), Body(Body), Guard(Guard.str()), IsUnavailable(IsUnavailable),
342 BigEndianSafe(BigEndianSafe), NeededEarly(false), UseMacro(false),
343 BaseType(OutTS, 'd'), InBaseType(InTS, 'd'), Emitter(Emitter) {
344 // If this builtin takes an immediate argument, we need to #define it rather
345 // than use a standard declaration, so that SemaChecking can range check
346 // the immediate passed by the user.
347 if (Proto.find('i') != std::string::npos)
350 // Pointer arguments need to use macros to avoid hiding aligned attributes
351 // from the pointer type.
352 if (Proto.find('p') != std::string::npos ||
353 Proto.find('c') != std::string::npos)
356 // It is not permitted to pass or return an __fp16 by value, so intrinsics
357 // taking a scalar float16_t must be implemented as macros.
358 if (OutTS.find('h') != std::string::npos &&
359 Proto.find('s') != std::string::npos)
362 // Modify the TypeSpec per-argument to get a concrete Type, and create
363 // known variables for each.
364 // Types[0] is the return value.
365 Types.emplace_back(OutTS, Proto[0]);
366 for (unsigned I = 1; I < Proto.size(); ++I)
367 Types.emplace_back(InTS, Proto[I]);
370 /// Get the Record that this intrinsic is based off.
371 Record *getRecord() const { return R; }
372 /// Get the set of Intrinsics that this intrinsic calls.
373 /// this is the set of immediate dependencies, NOT the
374 /// transitive closure.
375 const std::set<Intrinsic *> &getDependencies() const { return Dependencies; }
376 /// Get the architectural guard string (#ifdef).
377 std::string getGuard() const { return Guard; }
378 /// Get the non-mangled name.
379 std::string getName() const { return Name; }
381 /// Return true if the intrinsic takes an immediate operand.
382 bool hasImmediate() const {
383 return Proto.find('i') != std::string::npos;
386 /// Return the parameter index of the immediate operand.
387 unsigned getImmediateIdx() const {
388 assert(hasImmediate());
389 unsigned Idx = Proto.find('i');
390 assert(Idx > 0 && "Can't return an immediate!");
394 /// Return true if the intrinsic takes an splat operand.
395 bool hasSplat() const { return Proto.find('a') != std::string::npos; }
397 /// Return the parameter index of the splat operand.
398 unsigned getSplatIdx() const {
400 unsigned Idx = Proto.find('a');
401 assert(Idx > 0 && "Can't return a splat!");
405 unsigned getNumParams() const { return Proto.size() - 1; }
406 Type getReturnType() const { return Types[0]; }
407 Type getParamType(unsigned I) const { return Types[I + 1]; }
408 Type getBaseType() const { return BaseType; }
409 /// Return the raw prototype string.
410 std::string getProto() const { return Proto; }
412 /// Return true if the prototype has a scalar argument.
413 /// This does not return true for the "splat" code ('a').
414 bool protoHasScalar() const;
416 /// Return the index that parameter PIndex will sit at
417 /// in a generated function call. This is often just PIndex,
418 /// but may not be as things such as multiple-vector operands
419 /// and sret parameters need to be taken into accont.
420 unsigned getGeneratedParamIdx(unsigned PIndex) {
422 if (getReturnType().getNumVectors() > 1)
423 // Multiple vectors are passed as sret.
426 for (unsigned I = 0; I < PIndex; ++I)
427 Idx += std::max(1U, getParamType(I).getNumVectors());
432 bool hasBody() const { return Body && !Body->getValues().empty(); }
434 void setNeededEarly() { NeededEarly = true; }
436 bool operator<(const Intrinsic &Other) const {
437 // Sort lexicographically on a two-tuple (Guard, Name)
438 if (Guard != Other.Guard)
439 return Guard < Other.Guard;
440 return Name < Other.Name;
443 ClassKind getClassKind(bool UseClassBIfScalar = false) {
444 if (UseClassBIfScalar && !protoHasScalar())
449 /// Return the name, mangled with type information.
450 /// If ForceClassS is true, use ClassS (u32/s32) instead
451 /// of the intrinsic's own type class.
452 std::string getMangledName(bool ForceClassS = false) const;
453 /// Return the type code for a builtin function call.
454 std::string getInstTypeCode(Type T, ClassKind CK) const;
455 /// Return the type string for a BUILTIN() macro in Builtins.def.
456 std::string getBuiltinTypeStr();
458 /// Generate the intrinsic, returning code.
459 std::string generate();
460 /// Perform type checking and populate the dependency graph, but
461 /// don't generate code yet.
465 std::string mangleName(std::string Name, ClassKind CK) const;
467 void initVariables();
468 std::string replaceParamsIn(std::string S);
470 void emitBodyAsBuiltinCall();
472 void generateImpl(bool ReverseArguments,
473 StringRef NamePrefix, StringRef CallPrefix);
475 void emitBody(StringRef CallPrefix);
476 void emitShadowedArgs();
477 void emitArgumentReversal();
478 void emitReturnReversal();
479 void emitReverseVariable(Variable &Dest, Variable &Src);
481 void emitClosingBrace();
482 void emitOpeningBrace();
483 void emitPrototype(StringRef NamePrefix);
487 StringRef CallPrefix;
490 DagEmitter(Intrinsic &Intr, StringRef CallPrefix) :
491 Intr(Intr), CallPrefix(CallPrefix) {
493 std::pair<Type, std::string> emitDagArg(Init *Arg, std::string ArgName);
494 std::pair<Type, std::string> emitDagSaveTemp(DagInit *DI);
495 std::pair<Type, std::string> emitDagSplat(DagInit *DI);
496 std::pair<Type, std::string> emitDagDup(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 // runHeader - Emit all the __builtin prototypes used in arm_neon.h
556 void runHeader(raw_ostream &o);
558 // runTests - Emit tests for all the Neon intrinsics.
559 void runTests(raw_ostream &o);
562 } // end anonymous namespace
564 //===----------------------------------------------------------------------===//
565 // Type implementation
566 //===----------------------------------------------------------------------===//
568 std::string Type::str() const {
573 if (!Signed && isInteger())
583 S += utostr(ElementBitwidth);
585 S += "x" + utostr(getNumElements());
587 S += "x" + utostr(NumVectors);
598 std::string Type::builtin_str() const {
604 // All pointers are void pointers.
606 else if (isInteger())
607 switch (ElementBitwidth) {
608 case 8: S += "c"; break;
609 case 16: S += "s"; break;
610 case 32: S += "i"; break;
611 case 64: S += "Wi"; break;
612 case 128: S += "LLLi"; break;
613 default: llvm_unreachable("Unhandled case!");
616 switch (ElementBitwidth) {
617 case 16: S += "h"; break;
618 case 32: S += "f"; break;
619 case 64: S += "d"; break;
620 default: llvm_unreachable("Unhandled case!");
623 if (isChar() && !Pointer)
624 // Make chars explicitly signed.
626 else if (isInteger() && !Pointer && !Signed)
629 // Constant indices are "int", but have the "constant expression" modifier.
631 assert(isInteger() && isSigned());
636 if (Constant) S += "C";
637 if (Pointer) S += "*";
642 for (unsigned I = 0; I < NumVectors; ++I)
643 Ret += "V" + utostr(getNumElements()) + S;
648 unsigned Type::getNeonEnum() const {
650 switch (ElementBitwidth) {
651 case 8: Addend = 0; break;
652 case 16: Addend = 1; break;
653 case 32: Addend = 2; break;
654 case 64: Addend = 3; break;
655 case 128: Addend = 4; break;
656 default: llvm_unreachable("Unhandled element bitwidth!");
659 unsigned Base = (unsigned)NeonTypeFlags::Int8 + Addend;
661 // Adjustment needed because Poly32 doesn't exist.
664 Base = (unsigned)NeonTypeFlags::Poly8 + Addend;
667 assert(Addend != 0 && "Float8 doesn't exist!");
668 Base = (unsigned)NeonTypeFlags::Float16 + (Addend - 1);
672 Base |= (unsigned)NeonTypeFlags::QuadFlag;
673 if (isInteger() && !Signed)
674 Base |= (unsigned)NeonTypeFlags::UnsignedFlag;
679 Type Type::fromTypedefName(StringRef Name) {
685 if (Name.front() == 'u') {
687 Name = Name.drop_front();
692 if (Name.startswith("float")) {
694 Name = Name.drop_front(5);
695 } else if (Name.startswith("poly")) {
697 Name = Name.drop_front(4);
699 assert(Name.startswith("int"));
700 Name = Name.drop_front(3);
704 for (I = 0; I < Name.size(); ++I) {
705 if (!isdigit(Name[I]))
708 Name.substr(0, I).getAsInteger(10, T.ElementBitwidth);
709 Name = Name.drop_front(I);
711 T.Bitwidth = T.ElementBitwidth;
714 if (Name.front() == 'x') {
715 Name = Name.drop_front();
717 for (I = 0; I < Name.size(); ++I) {
718 if (!isdigit(Name[I]))
722 Name.substr(0, I).getAsInteger(10, NumLanes);
723 Name = Name.drop_front(I);
724 T.Bitwidth = T.ElementBitwidth * NumLanes;
729 if (Name.front() == 'x') {
730 Name = Name.drop_front();
732 for (I = 0; I < Name.size(); ++I) {
733 if (!isdigit(Name[I]))
736 Name.substr(0, I).getAsInteger(10, T.NumVectors);
737 Name = Name.drop_front(I);
740 assert(Name.startswith("_t") && "Malformed typedef!");
744 void Type::applyTypespec(bool &Quad) {
746 ScalarForMangling = false;
748 Poly = Float = false;
749 ElementBitwidth = ~0U;
756 ScalarForMangling = true;
778 ElementBitwidth = 16;
784 ElementBitwidth = 32;
790 ElementBitwidth = 64;
793 ElementBitwidth = 128;
794 // Poly doesn't have a 128x1 type.
799 llvm_unreachable("Unhandled type code!");
802 assert(ElementBitwidth != ~0U && "Bad element bitwidth!");
804 Bitwidth = Quad ? 128 : 64;
807 void Type::applyModifier(char Mod) {
808 bool AppliedQuad = false;
809 applyTypespec(AppliedQuad);
826 Bitwidth = ElementBitwidth;
833 Bitwidth = ElementBitwidth;
842 assert(!Poly && "'u' can't be used with poly types!");
846 Bitwidth = ElementBitwidth = 64;
851 Bitwidth = ElementBitwidth = 32;
857 ElementBitwidth = 32;
861 ElementBitwidth = 64;
865 ElementBitwidth = 16;
876 ElementBitwidth *= 2;
880 ElementBitwidth *= 2;
885 ElementBitwidth = Bitwidth = 32;
893 ElementBitwidth = Bitwidth = 64;
899 ElementBitwidth /= 2;
900 Bitwidth = ElementBitwidth;
904 ElementBitwidth *= 2;
905 Bitwidth = ElementBitwidth;
910 Bitwidth = ElementBitwidth;
921 Bitwidth = ElementBitwidth;
925 ElementBitwidth /= 2;
928 ElementBitwidth /= 2;
932 ElementBitwidth /= 2;
936 ElementBitwidth /= 2;
966 llvm_unreachable("Unhandled character!");
970 //===----------------------------------------------------------------------===//
971 // Intrinsic implementation
972 //===----------------------------------------------------------------------===//
974 std::string Intrinsic::getInstTypeCode(Type T, ClassKind CK) const {
975 char typeCode = '\0';
976 bool printNumber = true;
983 else if (T.isInteger())
984 typeCode = T.isSigned() ? 's' : 'u';
1004 if (typeCode != '\0')
1005 S.push_back(typeCode);
1007 S += utostr(T.getElementSizeInBits());
1012 static bool isFloatingPointProtoModifier(char Mod) {
1013 return Mod == 'F' || Mod == 'f' || Mod == 'H';
1016 std::string Intrinsic::getBuiltinTypeStr() {
1017 ClassKind LocalCK = getClassKind(true);
1020 Type RetT = getReturnType();
1021 if ((LocalCK == ClassI || LocalCK == ClassW) && RetT.isScalar() &&
1023 RetT.makeInteger(RetT.getElementSizeInBits(), false);
1025 // Since the return value must be one type, return a vector type of the
1026 // appropriate width which we will bitcast. An exception is made for
1027 // returning structs of 2, 3, or 4 vectors which are returned in a sret-like
1028 // fashion, storing them to a pointer arg.
1029 if (RetT.getNumVectors() > 1) {
1030 S += "vv*"; // void result with void* first argument
1033 RetT.makeInteger(RetT.getElementSizeInBits(), false);
1034 if (!RetT.isScalar() && !RetT.isSigned())
1037 bool ForcedVectorFloatingType = isFloatingPointProtoModifier(Proto[0]);
1038 if (LocalCK == ClassB && !RetT.isScalar() && !ForcedVectorFloatingType)
1039 // Cast to vector of 8-bit elements.
1040 RetT.makeInteger(8, true);
1042 S += RetT.builtin_str();
1045 for (unsigned I = 0; I < getNumParams(); ++I) {
1046 Type T = getParamType(I);
1048 T.makeInteger(T.getElementSizeInBits(), false);
1050 bool ForcedFloatingType = isFloatingPointProtoModifier(Proto[I + 1]);
1051 if (LocalCK == ClassB && !T.isScalar() && !ForcedFloatingType)
1052 T.makeInteger(8, true);
1053 // Halves always get converted to 8-bit elements.
1054 if (T.isHalf() && T.isVector() && !T.isScalarForMangling())
1055 T.makeInteger(8, true);
1057 if (LocalCK == ClassI)
1060 if (hasImmediate() && getImmediateIdx() == I)
1061 T.makeImmediate(32);
1063 S += T.builtin_str();
1066 // Extra constant integer to hold type class enum for this function, e.g. s8
1067 if (LocalCK == ClassB)
1073 std::string Intrinsic::getMangledName(bool ForceClassS) const {
1074 // Check if the prototype has a scalar operand with the type of the vector
1075 // elements. If not, bitcasting the args will take care of arg checking.
1076 // The actual signedness etc. will be taken care of with special enums.
1077 ClassKind LocalCK = CK;
1078 if (!protoHasScalar())
1081 return mangleName(Name, ForceClassS ? ClassS : LocalCK);
1084 std::string Intrinsic::mangleName(std::string Name, ClassKind LocalCK) const {
1085 std::string typeCode = getInstTypeCode(BaseType, LocalCK);
1086 std::string S = Name;
1088 if (Name == "vcvt_f16_f32" || Name == "vcvt_f32_f16" ||
1089 Name == "vcvt_f32_f64" || Name == "vcvt_f64_f32")
1092 if (!typeCode.empty()) {
1093 // If the name ends with _xN (N = 2,3,4), insert the typeCode before _xN.
1094 if (Name.size() >= 3 && isdigit(Name.back()) &&
1095 Name[Name.length() - 2] == 'x' && Name[Name.length() - 3] == '_')
1096 S.insert(S.length() - 3, "_" + typeCode);
1098 S += "_" + typeCode;
1101 if (BaseType != InBaseType) {
1102 // A reinterpret - out the input base type at the end.
1103 S += "_" + getInstTypeCode(InBaseType, LocalCK);
1106 if (LocalCK == ClassB)
1109 // Insert a 'q' before the first '_' character so that it ends up before
1110 // _lane or _n on vector-scalar operations.
1111 if (BaseType.getSizeInBits() == 128 && !BaseType.noManglingQ()) {
1112 size_t Pos = S.find('_');
1117 if (BaseType.isScalarForMangling()) {
1118 switch (BaseType.getElementSizeInBits()) {
1119 case 8: Suffix = 'b'; break;
1120 case 16: Suffix = 'h'; break;
1121 case 32: Suffix = 's'; break;
1122 case 64: Suffix = 'd'; break;
1123 default: llvm_unreachable("Bad suffix!");
1126 if (Suffix != '\0') {
1127 size_t Pos = S.find('_');
1128 S.insert(Pos, &Suffix, 1);
1134 std::string Intrinsic::replaceParamsIn(std::string S) {
1135 while (S.find('$') != std::string::npos) {
1136 size_t Pos = S.find('$');
1137 size_t End = Pos + 1;
1138 while (isalpha(S[End]))
1141 std::string VarName = S.substr(Pos + 1, End - Pos - 1);
1142 assert_with_loc(Variables.find(VarName) != Variables.end(),
1143 "Variable not defined!");
1144 S.replace(Pos, End - Pos, Variables.find(VarName)->second.getName());
1150 void Intrinsic::initVariables() {
1153 // Modify the TypeSpec per-argument to get a concrete Type, and create
1154 // known variables for each.
1155 for (unsigned I = 1; I < Proto.size(); ++I) {
1156 char NameC = '0' + (I - 1);
1157 std::string Name = "p";
1158 Name.push_back(NameC);
1160 Variables[Name] = Variable(Types[I], Name + VariablePostfix);
1162 RetVar = Variable(Types[0], "ret" + VariablePostfix);
1165 void Intrinsic::emitPrototype(StringRef NamePrefix) {
1169 OS << "__ai " << Types[0].str() << " ";
1171 OS << NamePrefix.str() << mangleName(Name, ClassS) << "(";
1173 for (unsigned I = 0; I < getNumParams(); ++I) {
1177 char NameC = '0' + I;
1178 std::string Name = "p";
1179 Name.push_back(NameC);
1180 assert(Variables.find(Name) != Variables.end());
1181 Variable &V = Variables[Name];
1184 OS << V.getType().str() << " ";
1191 void Intrinsic::emitOpeningBrace() {
1193 OS << " __extension__ ({";
1199 void Intrinsic::emitClosingBrace() {
1206 void Intrinsic::emitNewLine() {
1213 void Intrinsic::emitReverseVariable(Variable &Dest, Variable &Src) {
1214 if (Dest.getType().getNumVectors() > 1) {
1217 for (unsigned K = 0; K < Dest.getType().getNumVectors(); ++K) {
1218 OS << " " << Dest.getName() << ".val[" << K << "] = "
1219 << "__builtin_shufflevector("
1220 << Src.getName() << ".val[" << K << "], "
1221 << Src.getName() << ".val[" << K << "]";
1222 for (int J = Dest.getType().getNumElements() - 1; J >= 0; --J)
1228 OS << " " << Dest.getName()
1229 << " = __builtin_shufflevector(" << Src.getName() << ", " << Src.getName();
1230 for (int J = Dest.getType().getNumElements() - 1; J >= 0; --J)
1237 void Intrinsic::emitArgumentReversal() {
1241 // Reverse all vector arguments.
1242 for (unsigned I = 0; I < getNumParams(); ++I) {
1243 std::string Name = "p" + utostr(I);
1244 std::string NewName = "rev" + utostr(I);
1246 Variable &V = Variables[Name];
1247 Variable NewV(V.getType(), NewName + VariablePostfix);
1249 if (!NewV.getType().isVector() || NewV.getType().getNumElements() == 1)
1252 OS << " " << NewV.getType().str() << " " << NewV.getName() << ";";
1253 emitReverseVariable(NewV, V);
1258 void Intrinsic::emitReturnReversal() {
1261 if (!getReturnType().isVector() || getReturnType().isVoid() ||
1262 getReturnType().getNumElements() == 1)
1264 emitReverseVariable(RetVar, RetVar);
1267 void Intrinsic::emitShadowedArgs() {
1268 // Macro arguments are not type-checked like inline function arguments,
1269 // so assign them to local temporaries to get the right type checking.
1273 for (unsigned I = 0; I < getNumParams(); ++I) {
1274 // Do not create a temporary for an immediate argument.
1275 // That would defeat the whole point of using a macro!
1276 if (hasImmediate() && Proto[I+1] == 'i')
1278 // Do not create a temporary for pointer arguments. The input
1279 // pointer may have an alignment hint.
1280 if (getParamType(I).isPointer())
1283 std::string Name = "p" + utostr(I);
1285 assert(Variables.find(Name) != Variables.end());
1286 Variable &V = Variables[Name];
1288 std::string NewName = "s" + utostr(I);
1289 Variable V2(V.getType(), NewName + VariablePostfix);
1291 OS << " " << V2.getType().str() << " " << V2.getName() << " = "
1292 << V.getName() << ";";
1299 // We don't check 'a' in this function, because for builtin function the
1300 // argument matching to 'a' uses a vector type splatted from a scalar type.
1301 bool Intrinsic::protoHasScalar() const {
1302 return (Proto.find('s') != std::string::npos ||
1303 Proto.find('z') != std::string::npos ||
1304 Proto.find('r') != std::string::npos ||
1305 Proto.find('b') != std::string::npos ||
1306 Proto.find('$') != std::string::npos ||
1307 Proto.find('y') != std::string::npos ||
1308 Proto.find('o') != std::string::npos);
1311 void Intrinsic::emitBodyAsBuiltinCall() {
1314 // If this builtin returns a struct 2, 3, or 4 vectors, pass it as an implicit
1315 // sret-like argument.
1316 bool SRet = getReturnType().getNumVectors() >= 2;
1320 // Call the non-splat builtin: chop off the "_n" suffix from the name.
1321 assert(N.endswith("_n"));
1325 ClassKind LocalCK = CK;
1326 if (!protoHasScalar())
1329 if (!getReturnType().isVoid() && !SRet)
1330 S += "(" + RetVar.getType().str() + ") ";
1332 S += "__builtin_neon_" + mangleName(N, LocalCK) + "(";
1335 S += "&" + RetVar.getName() + ", ";
1337 for (unsigned I = 0; I < getNumParams(); ++I) {
1338 Variable &V = Variables["p" + utostr(I)];
1339 Type T = V.getType();
1341 // Handle multiple-vector values specially, emitting each subvector as an
1342 // argument to the builtin.
1343 if (T.getNumVectors() > 1) {
1344 // Check if an explicit cast is needed.
1346 if (T.isChar() || T.isPoly() || !T.isSigned()) {
1349 T2.makeInteger(8, /*Signed=*/true);
1350 Cast = "(" + T2.str() + ")";
1353 for (unsigned J = 0; J < T.getNumVectors(); ++J)
1354 S += Cast + V.getName() + ".val[" + utostr(J) + "], ";
1359 Type CastToType = T;
1360 if (hasSplat() && I == getSplatIdx()) {
1361 Arg = "(" + BaseType.str() + ") {";
1362 for (unsigned J = 0; J < BaseType.getNumElements(); ++J) {
1369 CastToType = BaseType;
1374 // Check if an explicit cast is needed.
1375 if (CastToType.isVector()) {
1376 CastToType.makeInteger(8, true);
1377 Arg = "(" + CastToType.str() + ")" + Arg;
1383 // Extra constant integer to hold type class enum for this function, e.g. s8
1384 if (getClassKind(true) == ClassB) {
1385 Type ThisTy = getReturnType();
1386 if (Proto[0] == 'v' || isFloatingPointProtoModifier(Proto[0]))
1387 ThisTy = getParamType(0);
1388 if (ThisTy.isPointer())
1389 ThisTy = getParamType(1);
1391 S += utostr(ThisTy.getNeonEnum());
1393 // Remove extraneous ", ".
1399 std::string RetExpr;
1400 if (!SRet && !RetVar.getType().isVoid())
1401 RetExpr = RetVar.getName() + " = ";
1403 OS << " " << RetExpr << S;
1407 void Intrinsic::emitBody(StringRef CallPrefix) {
1408 std::vector<std::string> Lines;
1410 assert(RetVar.getType() == Types[0]);
1411 // Create a return variable, if we're not void.
1412 if (!RetVar.getType().isVoid()) {
1413 OS << " " << RetVar.getType().str() << " " << RetVar.getName() << ";";
1417 if (!Body || Body->getValues().empty()) {
1418 // Nothing specific to output - must output a builtin.
1419 emitBodyAsBuiltinCall();
1423 // We have a list of "things to output". The last should be returned.
1424 for (auto *I : Body->getValues()) {
1425 if (StringInit *SI = dyn_cast<StringInit>(I)) {
1426 Lines.push_back(replaceParamsIn(SI->getAsString()));
1427 } else if (DagInit *DI = dyn_cast<DagInit>(I)) {
1428 DagEmitter DE(*this, CallPrefix);
1429 Lines.push_back(DE.emitDag(DI).second + ";");
1433 assert(!Lines.empty() && "Empty def?");
1434 if (!RetVar.getType().isVoid())
1435 Lines.back().insert(0, RetVar.getName() + " = ");
1437 for (auto &L : Lines) {
1443 void Intrinsic::emitReturn() {
1444 if (RetVar.getType().isVoid())
1447 OS << " " << RetVar.getName() << ";";
1449 OS << " return " << RetVar.getName() << ";";
1453 std::pair<Type, std::string> Intrinsic::DagEmitter::emitDag(DagInit *DI) {
1454 // At this point we should only be seeing a def.
1455 DefInit *DefI = cast<DefInit>(DI->getOperator());
1456 std::string Op = DefI->getAsString();
1458 if (Op == "cast" || Op == "bitcast")
1459 return emitDagCast(DI, Op == "bitcast");
1460 if (Op == "shuffle")
1461 return emitDagShuffle(DI);
1463 return emitDagDup(DI);
1465 return emitDagSplat(DI);
1466 if (Op == "save_temp")
1467 return emitDagSaveTemp(DI);
1469 return emitDagOp(DI);
1471 return emitDagCall(DI);
1472 if (Op == "name_replace")
1473 return emitDagNameReplace(DI);
1474 if (Op == "literal")
1475 return emitDagLiteral(DI);
1476 assert_with_loc(false, "Unknown operation!");
1477 return std::make_pair(Type::getVoid(), "");
1480 std::pair<Type, std::string> Intrinsic::DagEmitter::emitDagOp(DagInit *DI) {
1481 std::string Op = cast<StringInit>(DI->getArg(0))->getAsUnquotedString();
1482 if (DI->getNumArgs() == 2) {
1484 std::pair<Type, std::string> R =
1485 emitDagArg(DI->getArg(1), DI->getArgNameStr(1));
1486 return std::make_pair(R.first, Op + R.second);
1488 assert(DI->getNumArgs() == 3 && "Can only handle unary and binary ops!");
1489 std::pair<Type, std::string> R1 =
1490 emitDagArg(DI->getArg(1), DI->getArgNameStr(1));
1491 std::pair<Type, std::string> R2 =
1492 emitDagArg(DI->getArg(2), DI->getArgNameStr(2));
1493 assert_with_loc(R1.first == R2.first, "Argument type mismatch!");
1494 return std::make_pair(R1.first, R1.second + " " + Op + " " + R2.second);
1498 std::pair<Type, std::string> Intrinsic::DagEmitter::emitDagCall(DagInit *DI) {
1499 std::vector<Type> Types;
1500 std::vector<std::string> Values;
1501 for (unsigned I = 0; I < DI->getNumArgs() - 1; ++I) {
1502 std::pair<Type, std::string> R =
1503 emitDagArg(DI->getArg(I + 1), DI->getArgNameStr(I + 1));
1504 Types.push_back(R.first);
1505 Values.push_back(R.second);
1508 // Look up the called intrinsic.
1510 if (StringInit *SI = dyn_cast<StringInit>(DI->getArg(0)))
1511 N = SI->getAsUnquotedString();
1513 N = emitDagArg(DI->getArg(0), "").second;
1514 Intrinsic &Callee = Intr.Emitter.getIntrinsic(N, Types);
1516 // Make sure the callee is known as an early def.
1517 Callee.setNeededEarly();
1518 Intr.Dependencies.insert(&Callee);
1520 // Now create the call itself.
1521 std::string S = CallPrefix.str() + Callee.getMangledName(true) + "(";
1522 for (unsigned I = 0; I < DI->getNumArgs() - 1; ++I) {
1529 return std::make_pair(Callee.getReturnType(), S);
1532 std::pair<Type, std::string> Intrinsic::DagEmitter::emitDagCast(DagInit *DI,
1534 // (cast MOD* VAL) -> cast VAL to type given by MOD.
1535 std::pair<Type, std::string> R = emitDagArg(
1536 DI->getArg(DI->getNumArgs() - 1),
1537 DI->getArgNameStr(DI->getNumArgs() - 1));
1538 Type castToType = R.first;
1539 for (unsigned ArgIdx = 0; ArgIdx < DI->getNumArgs() - 1; ++ArgIdx) {
1541 // MOD can take several forms:
1542 // 1. $X - take the type of parameter / variable X.
1543 // 2. The value "R" - take the type of the return type.
1545 // 4. The value "U" or "S" to switch the signedness.
1546 // 5. The value "H" or "D" to half or double the bitwidth.
1547 // 6. The value "8" to convert to 8-bit (signed) integer lanes.
1548 if (!DI->getArgNameStr(ArgIdx).empty()) {
1549 assert_with_loc(Intr.Variables.find(DI->getArgNameStr(ArgIdx)) !=
1550 Intr.Variables.end(),
1551 "Variable not found");
1552 castToType = Intr.Variables[DI->getArgNameStr(ArgIdx)].getType();
1554 StringInit *SI = dyn_cast<StringInit>(DI->getArg(ArgIdx));
1555 assert_with_loc(SI, "Expected string type or $Name for cast type");
1557 if (SI->getAsUnquotedString() == "R") {
1558 castToType = Intr.getReturnType();
1559 } else if (SI->getAsUnquotedString() == "U") {
1560 castToType.makeUnsigned();
1561 } else if (SI->getAsUnquotedString() == "S") {
1562 castToType.makeSigned();
1563 } else if (SI->getAsUnquotedString() == "H") {
1564 castToType.halveLanes();
1565 } else if (SI->getAsUnquotedString() == "D") {
1566 castToType.doubleLanes();
1567 } else if (SI->getAsUnquotedString() == "8") {
1568 castToType.makeInteger(8, true);
1570 castToType = Type::fromTypedefName(SI->getAsUnquotedString());
1571 assert_with_loc(!castToType.isVoid(), "Unknown typedef");
1578 // Emit a reinterpret cast. The second operand must be an lvalue, so create
1580 std::string N = "reint";
1582 while (Intr.Variables.find(N) != Intr.Variables.end())
1583 N = "reint" + utostr(++I);
1584 Intr.Variables[N] = Variable(R.first, N + Intr.VariablePostfix);
1586 Intr.OS << R.first.str() << " " << Intr.Variables[N].getName() << " = "
1590 S = "*(" + castToType.str() + " *) &" + Intr.Variables[N].getName() + "";
1592 // Emit a normal (static) cast.
1593 S = "(" + castToType.str() + ")(" + R.second + ")";
1596 return std::make_pair(castToType, S);
1599 std::pair<Type, std::string> Intrinsic::DagEmitter::emitDagShuffle(DagInit *DI){
1600 // See the documentation in arm_neon.td for a description of these operators.
1601 class LowHalf : public SetTheory::Operator {
1603 void apply(SetTheory &ST, DagInit *Expr, SetTheory::RecSet &Elts,
1604 ArrayRef<SMLoc> Loc) override {
1605 SetTheory::RecSet Elts2;
1606 ST.evaluate(Expr->arg_begin(), Expr->arg_end(), Elts2, Loc);
1607 Elts.insert(Elts2.begin(), Elts2.begin() + (Elts2.size() / 2));
1611 class HighHalf : public SetTheory::Operator {
1613 void apply(SetTheory &ST, DagInit *Expr, SetTheory::RecSet &Elts,
1614 ArrayRef<SMLoc> Loc) override {
1615 SetTheory::RecSet Elts2;
1616 ST.evaluate(Expr->arg_begin(), Expr->arg_end(), Elts2, Loc);
1617 Elts.insert(Elts2.begin() + (Elts2.size() / 2), Elts2.end());
1621 class Rev : public SetTheory::Operator {
1622 unsigned ElementSize;
1625 Rev(unsigned ElementSize) : ElementSize(ElementSize) {}
1627 void apply(SetTheory &ST, DagInit *Expr, SetTheory::RecSet &Elts,
1628 ArrayRef<SMLoc> Loc) override {
1629 SetTheory::RecSet Elts2;
1630 ST.evaluate(Expr->arg_begin() + 1, Expr->arg_end(), Elts2, Loc);
1632 int64_t VectorSize = cast<IntInit>(Expr->getArg(0))->getValue();
1633 VectorSize /= ElementSize;
1635 std::vector<Record *> Revved;
1636 for (unsigned VI = 0; VI < Elts2.size(); VI += VectorSize) {
1637 for (int LI = VectorSize - 1; LI >= 0; --LI) {
1638 Revved.push_back(Elts2[VI + LI]);
1642 Elts.insert(Revved.begin(), Revved.end());
1646 class MaskExpander : public SetTheory::Expander {
1650 MaskExpander(unsigned N) : N(N) {}
1652 void expand(SetTheory &ST, Record *R, SetTheory::RecSet &Elts) override {
1653 unsigned Addend = 0;
1654 if (R->getName() == "mask0")
1656 else if (R->getName() == "mask1")
1660 for (unsigned I = 0; I < N; ++I)
1661 Elts.insert(R->getRecords().getDef("sv" + utostr(I + Addend)));
1665 // (shuffle arg1, arg2, sequence)
1666 std::pair<Type, std::string> Arg1 =
1667 emitDagArg(DI->getArg(0), DI->getArgNameStr(0));
1668 std::pair<Type, std::string> Arg2 =
1669 emitDagArg(DI->getArg(1), DI->getArgNameStr(1));
1670 assert_with_loc(Arg1.first == Arg2.first,
1671 "Different types in arguments to shuffle!");
1674 SetTheory::RecSet Elts;
1675 ST.addOperator("lowhalf", llvm::make_unique<LowHalf>());
1676 ST.addOperator("highhalf", llvm::make_unique<HighHalf>());
1677 ST.addOperator("rev",
1678 llvm::make_unique<Rev>(Arg1.first.getElementSizeInBits()));
1679 ST.addExpander("MaskExpand",
1680 llvm::make_unique<MaskExpander>(Arg1.first.getNumElements()));
1681 ST.evaluate(DI->getArg(2), Elts, None);
1683 std::string S = "__builtin_shufflevector(" + Arg1.second + ", " + Arg2.second;
1684 for (auto &E : Elts) {
1685 StringRef Name = E->getName();
1686 assert_with_loc(Name.startswith("sv"),
1687 "Incorrect element kind in shuffle mask!");
1688 S += ", " + Name.drop_front(2).str();
1692 // Recalculate the return type - the shuffle may have halved or doubled it.
1694 if (Elts.size() > T.getNumElements()) {
1696 Elts.size() == T.getNumElements() * 2,
1697 "Can only double or half the number of elements in a shuffle!");
1699 } else if (Elts.size() < T.getNumElements()) {
1701 Elts.size() == T.getNumElements() / 2,
1702 "Can only double or half the number of elements in a shuffle!");
1706 return std::make_pair(T, S);
1709 std::pair<Type, std::string> Intrinsic::DagEmitter::emitDagDup(DagInit *DI) {
1710 assert_with_loc(DI->getNumArgs() == 1, "dup() expects one argument");
1711 std::pair<Type, std::string> A = emitDagArg(DI->getArg(0),
1712 DI->getArgNameStr(0));
1713 assert_with_loc(A.first.isScalar(), "dup() expects a scalar argument");
1715 Type T = Intr.getBaseType();
1716 assert_with_loc(T.isVector(), "dup() used but default type is scalar!");
1717 std::string S = "(" + T.str() + ") {";
1718 for (unsigned I = 0; I < T.getNumElements(); ++I) {
1725 return std::make_pair(T, S);
1728 std::pair<Type, std::string> Intrinsic::DagEmitter::emitDagSplat(DagInit *DI) {
1729 assert_with_loc(DI->getNumArgs() == 2, "splat() expects two arguments");
1730 std::pair<Type, std::string> A = emitDagArg(DI->getArg(0),
1731 DI->getArgNameStr(0));
1732 std::pair<Type, std::string> B = emitDagArg(DI->getArg(1),
1733 DI->getArgNameStr(1));
1735 assert_with_loc(B.first.isScalar(),
1736 "splat() requires a scalar int as the second argument");
1738 std::string S = "__builtin_shufflevector(" + A.second + ", " + A.second;
1739 for (unsigned I = 0; I < Intr.getBaseType().getNumElements(); ++I) {
1740 S += ", " + B.second;
1744 return std::make_pair(Intr.getBaseType(), S);
1747 std::pair<Type, std::string> Intrinsic::DagEmitter::emitDagSaveTemp(DagInit *DI) {
1748 assert_with_loc(DI->getNumArgs() == 2, "save_temp() expects two arguments");
1749 std::pair<Type, std::string> A = emitDagArg(DI->getArg(1),
1750 DI->getArgNameStr(1));
1752 assert_with_loc(!A.first.isVoid(),
1753 "Argument to save_temp() must have non-void type!");
1755 std::string N = DI->getArgNameStr(0);
1756 assert_with_loc(!N.empty(),
1757 "save_temp() expects a name as the first argument");
1759 assert_with_loc(Intr.Variables.find(N) == Intr.Variables.end(),
1760 "Variable already defined!");
1761 Intr.Variables[N] = Variable(A.first, N + Intr.VariablePostfix);
1764 A.first.str() + " " + Intr.Variables[N].getName() + " = " + A.second;
1766 return std::make_pair(Type::getVoid(), S);
1769 std::pair<Type, std::string>
1770 Intrinsic::DagEmitter::emitDagNameReplace(DagInit *DI) {
1771 std::string S = Intr.Name;
1773 assert_with_loc(DI->getNumArgs() == 2, "name_replace requires 2 arguments!");
1774 std::string ToReplace = cast<StringInit>(DI->getArg(0))->getAsUnquotedString();
1775 std::string ReplaceWith = cast<StringInit>(DI->getArg(1))->getAsUnquotedString();
1777 size_t Idx = S.find(ToReplace);
1779 assert_with_loc(Idx != std::string::npos, "name should contain '" + ToReplace + "'!");
1780 S.replace(Idx, ToReplace.size(), ReplaceWith);
1782 return std::make_pair(Type::getVoid(), S);
1785 std::pair<Type, std::string> Intrinsic::DagEmitter::emitDagLiteral(DagInit *DI){
1786 std::string Ty = cast<StringInit>(DI->getArg(0))->getAsUnquotedString();
1787 std::string Value = cast<StringInit>(DI->getArg(1))->getAsUnquotedString();
1788 return std::make_pair(Type::fromTypedefName(Ty), Value);
1791 std::pair<Type, std::string>
1792 Intrinsic::DagEmitter::emitDagArg(Init *Arg, std::string ArgName) {
1793 if (!ArgName.empty()) {
1794 assert_with_loc(!Arg->isComplete(),
1795 "Arguments must either be DAGs or names, not both!");
1796 assert_with_loc(Intr.Variables.find(ArgName) != Intr.Variables.end(),
1797 "Variable not defined!");
1798 Variable &V = Intr.Variables[ArgName];
1799 return std::make_pair(V.getType(), V.getName());
1802 assert(Arg && "Neither ArgName nor Arg?!");
1803 DagInit *DI = dyn_cast<DagInit>(Arg);
1804 assert_with_loc(DI, "Arguments must either be DAGs or names!");
1809 std::string Intrinsic::generate() {
1810 // Little endian intrinsics are simple and don't require any argument
1812 OS << "#ifdef __LITTLE_ENDIAN__\n";
1814 generateImpl(false, "", "");
1818 // Big endian intrinsics are more complex. The user intended these
1819 // intrinsics to operate on a vector "as-if" loaded by (V)LDR,
1820 // but we load as-if (V)LD1. So we should swap all arguments and
1821 // swap the return value too.
1823 // If we call sub-intrinsics, we should call a version that does
1824 // not re-swap the arguments!
1825 generateImpl(true, "", "__noswap_");
1827 // If we're needed early, create a non-swapping variant for
1830 generateImpl(false, "__noswap_", "__noswap_");
1837 void Intrinsic::generateImpl(bool ReverseArguments,
1838 StringRef NamePrefix, StringRef CallPrefix) {
1841 // If we call a macro, our local variables may be corrupted due to
1842 // lack of proper lexical scoping. So, add a globally unique postfix
1843 // to every variable.
1845 // indexBody() should have set up the Dependencies set by now.
1846 for (auto *I : Dependencies)
1848 VariablePostfix = "_" + utostr(Emitter.getUniqueNumber());
1854 emitPrototype(NamePrefix);
1856 if (IsUnavailable) {
1857 OS << " __attribute__((unavailable));";
1861 if (ReverseArguments)
1862 emitArgumentReversal();
1863 emitBody(CallPrefix);
1864 if (ReverseArguments)
1865 emitReturnReversal();
1871 CurrentRecord = nullptr;
1874 void Intrinsic::indexBody() {
1881 CurrentRecord = nullptr;
1884 //===----------------------------------------------------------------------===//
1885 // NeonEmitter implementation
1886 //===----------------------------------------------------------------------===//
1888 Intrinsic &NeonEmitter::getIntrinsic(StringRef Name, ArrayRef<Type> Types) {
1889 // First, look up the name in the intrinsic map.
1890 assert_with_loc(IntrinsicMap.find(Name.str()) != IntrinsicMap.end(),
1891 ("Intrinsic '" + Name + "' not found!").str());
1892 auto &V = IntrinsicMap.find(Name.str())->second;
1893 std::vector<Intrinsic *> GoodVec;
1895 // Create a string to print if we end up failing.
1896 std::string ErrMsg = "looking up intrinsic '" + Name.str() + "(";
1897 for (unsigned I = 0; I < Types.size(); ++I) {
1900 ErrMsg += Types[I].str();
1903 ErrMsg += "Available overloads:\n";
1905 // Now, look through each intrinsic implementation and see if the types are
1908 ErrMsg += " - " + I.getReturnType().str() + " " + I.getMangledName();
1910 for (unsigned A = 0; A < I.getNumParams(); ++A) {
1913 ErrMsg += I.getParamType(A).str();
1917 if (I.getNumParams() != Types.size())
1921 for (unsigned Arg = 0; Arg < Types.size(); ++Arg) {
1922 if (I.getParamType(Arg) != Types[Arg]) {
1928 GoodVec.push_back(&I);
1931 assert_with_loc(!GoodVec.empty(),
1932 "No compatible intrinsic found - " + ErrMsg);
1933 assert_with_loc(GoodVec.size() == 1, "Multiple overloads found - " + ErrMsg);
1935 return *GoodVec.front();
1938 void NeonEmitter::createIntrinsic(Record *R,
1939 SmallVectorImpl<Intrinsic *> &Out) {
1940 std::string Name = R->getValueAsString("Name");
1941 std::string Proto = R->getValueAsString("Prototype");
1942 std::string Types = R->getValueAsString("Types");
1943 Record *OperationRec = R->getValueAsDef("Operation");
1944 bool CartesianProductOfTypes = R->getValueAsBit("CartesianProductOfTypes");
1945 bool BigEndianSafe = R->getValueAsBit("BigEndianSafe");
1946 std::string Guard = R->getValueAsString("ArchGuard");
1947 bool IsUnavailable = OperationRec->getValueAsBit("Unavailable");
1949 // Set the global current record. This allows assert_with_loc to produce
1950 // decent location information even when highly nested.
1953 ListInit *Body = OperationRec->getValueAsListInit("Ops");
1955 std::vector<TypeSpec> TypeSpecs = TypeSpec::fromTypeSpecs(Types);
1957 ClassKind CK = ClassNone;
1958 if (R->getSuperClasses().size() >= 2)
1959 CK = ClassMap[R->getSuperClasses()[1].first];
1961 std::vector<std::pair<TypeSpec, TypeSpec>> NewTypeSpecs;
1962 for (auto TS : TypeSpecs) {
1963 if (CartesianProductOfTypes) {
1964 Type DefaultT(TS, 'd');
1965 for (auto SrcTS : TypeSpecs) {
1966 Type DefaultSrcT(SrcTS, 'd');
1968 DefaultSrcT.getSizeInBits() != DefaultT.getSizeInBits())
1970 NewTypeSpecs.push_back(std::make_pair(TS, SrcTS));
1973 NewTypeSpecs.push_back(std::make_pair(TS, TS));
1977 std::sort(NewTypeSpecs.begin(), NewTypeSpecs.end());
1978 NewTypeSpecs.erase(std::unique(NewTypeSpecs.begin(), NewTypeSpecs.end()),
1979 NewTypeSpecs.end());
1980 auto &Entry = IntrinsicMap[Name];
1982 for (auto &I : NewTypeSpecs) {
1983 Entry.emplace_back(R, Name, Proto, I.first, I.second, CK, Body, *this,
1984 Guard, IsUnavailable, BigEndianSafe);
1985 Out.push_back(&Entry.back());
1988 CurrentRecord = nullptr;
1991 /// genBuiltinsDef: Generate the BuiltinsARM.def and BuiltinsAArch64.def
1992 /// declaration of builtins, checking for unique builtin declarations.
1993 void NeonEmitter::genBuiltinsDef(raw_ostream &OS,
1994 SmallVectorImpl<Intrinsic *> &Defs) {
1995 OS << "#ifdef GET_NEON_BUILTINS\n";
1997 // We only want to emit a builtin once, and we want to emit them in
1998 // alphabetical order, so use a std::set.
1999 std::set<std::string> Builtins;
2001 for (auto *Def : Defs) {
2004 // Functions with 'a' (the splat code) in the type prototype should not get
2005 // their own builtin as they use the non-splat variant.
2006 if (Def->hasSplat())
2009 std::string S = "BUILTIN(__builtin_neon_" + Def->getMangledName() + ", \"";
2011 S += Def->getBuiltinTypeStr();
2017 for (auto &S : Builtins)
2022 /// Generate the ARM and AArch64 overloaded type checking code for
2023 /// SemaChecking.cpp, checking for unique builtin declarations.
2024 void NeonEmitter::genOverloadTypeCheckCode(raw_ostream &OS,
2025 SmallVectorImpl<Intrinsic *> &Defs) {
2026 OS << "#ifdef GET_NEON_OVERLOAD_CHECK\n";
2028 // We record each overload check line before emitting because subsequent Inst
2029 // definitions may extend the number of permitted types (i.e. augment the
2030 // Mask). Use std::map to avoid sorting the table by hash number.
2031 struct OverloadInfo {
2035 OverloadInfo() : Mask(0ULL), PtrArgNum(0), HasConstPtr(false) {}
2037 std::map<std::string, OverloadInfo> OverloadMap;
2039 for (auto *Def : Defs) {
2040 // If the def has a body (that is, it has Operation DAGs), it won't call
2041 // __builtin_neon_* so we don't need to generate a definition for it.
2044 // Functions with 'a' (the splat code) in the type prototype should not get
2045 // their own builtin as they use the non-splat variant.
2046 if (Def->hasSplat())
2048 // Functions which have a scalar argument cannot be overloaded, no need to
2049 // check them if we are emitting the type checking code.
2050 if (Def->protoHasScalar())
2053 uint64_t Mask = 0ULL;
2054 Type Ty = Def->getReturnType();
2055 if (Def->getProto()[0] == 'v' ||
2056 isFloatingPointProtoModifier(Def->getProto()[0]))
2057 Ty = Def->getParamType(0);
2059 Ty = Def->getParamType(1);
2061 Mask |= 1ULL << Ty.getNeonEnum();
2063 // Check if the function has a pointer or const pointer argument.
2064 std::string Proto = Def->getProto();
2066 bool HasConstPtr = false;
2067 for (unsigned I = 0; I < Def->getNumParams(); ++I) {
2068 char ArgType = Proto[I + 1];
2069 if (ArgType == 'c') {
2074 if (ArgType == 'p') {
2079 // For sret builtins, adjust the pointer argument index.
2080 if (PtrArgNum >= 0 && Def->getReturnType().getNumVectors() > 1)
2083 std::string Name = Def->getName();
2084 // Omit type checking for the pointer arguments of vld1_lane, vld1_dup,
2085 // and vst1_lane intrinsics. Using a pointer to the vector element
2086 // type with one of those operations causes codegen to select an aligned
2087 // load/store instruction. If you want an unaligned operation,
2088 // the pointer argument needs to have less alignment than element type,
2089 // so just accept any pointer type.
2090 if (Name == "vld1_lane" || Name == "vld1_dup" || Name == "vst1_lane") {
2092 HasConstPtr = false;
2096 std::string Name = Def->getMangledName();
2097 OverloadMap.insert(std::make_pair(Name, OverloadInfo()));
2098 OverloadInfo &OI = OverloadMap[Name];
2100 OI.PtrArgNum |= PtrArgNum;
2101 OI.HasConstPtr = HasConstPtr;
2105 for (auto &I : OverloadMap) {
2106 OverloadInfo &OI = I.second;
2108 OS << "case NEON::BI__builtin_neon_" << I.first << ": ";
2109 OS << "mask = 0x" << Twine::utohexstr(OI.Mask) << "ULL";
2110 if (OI.PtrArgNum >= 0)
2111 OS << "; PtrArgNum = " << OI.PtrArgNum;
2113 OS << "; HasConstPtr = true";
2120 NeonEmitter::genIntrinsicRangeCheckCode(raw_ostream &OS,
2121 SmallVectorImpl<Intrinsic *> &Defs) {
2122 OS << "#ifdef GET_NEON_IMMEDIATE_CHECK\n";
2124 std::set<std::string> Emitted;
2126 for (auto *Def : Defs) {
2129 // Functions with 'a' (the splat code) in the type prototype should not get
2130 // their own builtin as they use the non-splat variant.
2131 if (Def->hasSplat())
2133 // Functions which do not have an immediate do not need to have range
2134 // checking code emitted.
2135 if (!Def->hasImmediate())
2137 if (Emitted.find(Def->getMangledName()) != Emitted.end())
2140 std::string LowerBound, UpperBound;
2142 Record *R = Def->getRecord();
2143 if (R->getValueAsBit("isVCVT_N")) {
2144 // VCVT between floating- and fixed-point values takes an immediate
2145 // in the range [1, 32) for f32 or [1, 64) for f64.
2147 if (Def->getBaseType().getElementSizeInBits() == 32)
2151 } else if (R->getValueAsBit("isScalarShift")) {
2152 // Right shifts have an 'r' in the name, left shifts do not. Convert
2153 // instructions have the same bounds and right shifts.
2154 if (Def->getName().find('r') != std::string::npos ||
2155 Def->getName().find("cvt") != std::string::npos)
2158 UpperBound = utostr(Def->getReturnType().getElementSizeInBits() - 1);
2159 } else if (R->getValueAsBit("isShift")) {
2160 // Builtins which are overloaded by type will need to have their upper
2161 // bound computed at Sema time based on the type constant.
2163 // Right shifts have an 'r' in the name, left shifts do not.
2164 if (Def->getName().find('r') != std::string::npos)
2166 UpperBound = "RFT(TV, true)";
2167 } else if (Def->getClassKind(true) == ClassB) {
2168 // ClassB intrinsics have a type (and hence lane number) that is only
2169 // known at runtime.
2170 if (R->getValueAsBit("isLaneQ"))
2171 UpperBound = "RFT(TV, false, true)";
2173 UpperBound = "RFT(TV, false, false)";
2175 // The immediate generally refers to a lane in the preceding argument.
2176 assert(Def->getImmediateIdx() > 0);
2177 Type T = Def->getParamType(Def->getImmediateIdx() - 1);
2178 UpperBound = utostr(T.getNumElements() - 1);
2181 // Calculate the index of the immediate that should be range checked.
2182 unsigned Idx = Def->getNumParams();
2183 if (Def->hasImmediate())
2184 Idx = Def->getGeneratedParamIdx(Def->getImmediateIdx());
2186 OS << "case NEON::BI__builtin_neon_" << Def->getMangledName() << ": "
2187 << "i = " << Idx << ";";
2188 if (!LowerBound.empty())
2189 OS << " l = " << LowerBound << ";";
2190 if (!UpperBound.empty())
2191 OS << " u = " << UpperBound << ";";
2194 Emitted.insert(Def->getMangledName());
2200 /// runHeader - Emit a file with sections defining:
2201 /// 1. the NEON section of BuiltinsARM.def and BuiltinsAArch64.def.
2202 /// 2. the SemaChecking code for the type overload checking.
2203 /// 3. the SemaChecking code for validation of intrinsic immediate arguments.
2204 void NeonEmitter::runHeader(raw_ostream &OS) {
2205 std::vector<Record *> RV = Records.getAllDerivedDefinitions("Inst");
2207 SmallVector<Intrinsic *, 128> Defs;
2209 createIntrinsic(R, Defs);
2211 // Generate shared BuiltinsXXX.def
2212 genBuiltinsDef(OS, Defs);
2214 // Generate ARM overloaded type checking code for SemaChecking.cpp
2215 genOverloadTypeCheckCode(OS, Defs);
2217 // Generate ARM range checking code for shift/lane immediates.
2218 genIntrinsicRangeCheckCode(OS, Defs);
2221 /// run - Read the records in arm_neon.td and output arm_neon.h. arm_neon.h
2222 /// is comprised of type definitions and function declarations.
2223 void NeonEmitter::run(raw_ostream &OS) {
2224 OS << "/*===---- arm_neon.h - ARM Neon intrinsics "
2225 "------------------------------"
2228 " * Permission is hereby granted, free of charge, to any person "
2231 " * of this software and associated documentation files (the "
2234 " * in the Software without restriction, including without limitation "
2237 " * to use, copy, modify, merge, publish, distribute, sublicense, "
2239 " * copies of the Software, and to permit persons to whom the Software "
2241 " * furnished to do so, subject to the following conditions:\n"
2243 " * The above copyright notice and this permission notice shall be "
2245 " * all copies or substantial portions of the Software.\n"
2247 " * THE SOFTWARE IS PROVIDED \"AS IS\", WITHOUT WARRANTY OF ANY KIND, "
2249 " * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF "
2250 "MERCHANTABILITY,\n"
2251 " * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT "
2253 " * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR "
2255 " * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, "
2257 " * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER "
2259 " * THE SOFTWARE.\n"
2261 " *===-----------------------------------------------------------------"
2266 OS << "#ifndef __ARM_NEON_H\n";
2267 OS << "#define __ARM_NEON_H\n\n";
2269 OS << "#if !defined(__ARM_NEON)\n";
2270 OS << "#error \"NEON support not enabled\"\n";
2273 OS << "#include <stdint.h>\n\n";
2275 // Emit NEON-specific scalar typedefs.
2276 OS << "typedef float float32_t;\n";
2277 OS << "typedef __fp16 float16_t;\n";
2279 OS << "#ifdef __aarch64__\n";
2280 OS << "typedef double float64_t;\n";
2283 // For now, signedness of polynomial types depends on target
2284 OS << "#ifdef __aarch64__\n";
2285 OS << "typedef uint8_t poly8_t;\n";
2286 OS << "typedef uint16_t poly16_t;\n";
2287 OS << "typedef uint64_t poly64_t;\n";
2288 OS << "typedef __uint128_t poly128_t;\n";
2290 OS << "typedef int8_t poly8_t;\n";
2291 OS << "typedef int16_t poly16_t;\n";
2294 // Emit Neon vector typedefs.
2295 std::string TypedefTypes(
2296 "cQcsQsiQilQlUcQUcUsQUsUiQUiUlQUlhQhfQfdQdPcQPcPsQPsPlQPl");
2297 std::vector<TypeSpec> TDTypeVec = TypeSpec::fromTypeSpecs(TypedefTypes);
2299 // Emit vector typedefs.
2300 bool InIfdef = false;
2301 for (auto &TS : TDTypeVec) {
2304 if (T.isDouble() || (T.isPoly() && T.isLong()))
2307 if (InIfdef && !IsA64) {
2311 if (!InIfdef && IsA64) {
2312 OS << "#ifdef __aarch64__\n";
2317 OS << "typedef __attribute__((neon_polyvector_type(";
2319 OS << "typedef __attribute__((neon_vector_type(";
2323 OS << T.getNumElements() << "))) ";
2325 OS << " " << T.str() << ";\n";
2331 // Emit struct typedefs.
2333 for (unsigned NumMembers = 2; NumMembers <= 4; ++NumMembers) {
2334 for (auto &TS : TDTypeVec) {
2337 if (T.isDouble() || (T.isPoly() && T.isLong()))
2340 if (InIfdef && !IsA64) {
2344 if (!InIfdef && IsA64) {
2345 OS << "#ifdef __aarch64__\n";
2349 char M = '2' + (NumMembers - 2);
2351 OS << "typedef struct " << VT.str() << " {\n";
2352 OS << " " << T.str() << " val";
2353 OS << "[" << NumMembers << "]";
2355 OS << VT.str() << ";\n";
2363 OS << "#define __ai static inline __attribute__((__always_inline__, "
2364 "__nodebug__))\n\n";
2366 SmallVector<Intrinsic *, 128> Defs;
2367 std::vector<Record *> RV = Records.getAllDerivedDefinitions("Inst");
2369 createIntrinsic(R, Defs);
2371 for (auto *I : Defs)
2375 Defs.begin(), Defs.end(),
2376 [](const Intrinsic *A, const Intrinsic *B) { return *A < *B; });
2378 // Only emit a def when its requirements have been met.
2379 // FIXME: This loop could be made faster, but it's fast enough for now.
2380 bool MadeProgress = true;
2381 std::string InGuard;
2382 while (!Defs.empty() && MadeProgress) {
2383 MadeProgress = false;
2385 for (SmallVector<Intrinsic *, 128>::iterator I = Defs.begin();
2386 I != Defs.end(); /*No step*/) {
2387 bool DependenciesSatisfied = true;
2388 for (auto *II : (*I)->getDependencies()) {
2389 if (std::find(Defs.begin(), Defs.end(), II) != Defs.end())
2390 DependenciesSatisfied = false;
2392 if (!DependenciesSatisfied) {
2393 // Try the next one.
2398 // Emit #endif/#if pair if needed.
2399 if ((*I)->getGuard() != InGuard) {
2400 if (!InGuard.empty())
2402 InGuard = (*I)->getGuard();
2403 if (!InGuard.empty())
2404 OS << "#if " << InGuard << "\n";
2407 // Actually generate the intrinsic code.
2408 OS << (*I)->generate();
2410 MadeProgress = true;
2414 assert(Defs.empty() && "Some requirements were not satisfied!");
2415 if (!InGuard.empty())
2419 OS << "#undef __ai\n\n";
2420 OS << "#endif /* __ARM_NEON_H */\n";
2425 void EmitNeon(RecordKeeper &Records, raw_ostream &OS) {
2426 NeonEmitter(Records).run(OS);
2429 void EmitNeonSema(RecordKeeper &Records, raw_ostream &OS) {
2430 NeonEmitter(Records).runHeader(OS);
2433 void EmitNeonTest(RecordKeeper &Records, raw_ostream &OS) {
2434 llvm_unreachable("Neon test generation no longer implemented!");
2437 } // end namespace clang