/Users/buildslave/jenkins/workspace/coverage/llvm-project/clang/utils/TableGen/MveEmitter.cpp
Line | Count | Source (jump to first uncovered line) |
1 | | //===- MveEmitter.cpp - Generate arm_mve.h for use with clang -*- C++ -*-=====// |
2 | | // |
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 |
6 | | // |
7 | | //===----------------------------------------------------------------------===// |
8 | | // |
9 | | // This set of linked tablegen backends is responsible for emitting the bits |
10 | | // and pieces that implement <arm_mve.h>, which is defined by the ACLE standard |
11 | | // and provides a set of types and functions for (more or less) direct access |
12 | | // to the MVE instruction set, including the scalar shifts as well as the |
13 | | // vector instructions. |
14 | | // |
15 | | // MVE's standard intrinsic functions are unusual in that they have a system of |
16 | | // polymorphism. For example, the function vaddq() can behave like vaddq_u16(), |
17 | | // vaddq_f32(), vaddq_s8(), etc., depending on the types of the vector |
18 | | // arguments you give it. |
19 | | // |
20 | | // This constrains the implementation strategies. The usual approach to making |
21 | | // the user-facing functions polymorphic would be to either use |
22 | | // __attribute__((overloadable)) to make a set of vaddq() functions that are |
23 | | // all inline wrappers on the underlying clang builtins, or to define a single |
24 | | // vaddq() macro which expands to an instance of _Generic. |
25 | | // |
26 | | // The inline-wrappers approach would work fine for most intrinsics, except for |
27 | | // the ones that take an argument required to be a compile-time constant, |
28 | | // because if you wrap an inline function around a call to a builtin, the |
29 | | // constant nature of the argument is not passed through. |
30 | | // |
31 | | // The _Generic approach can be made to work with enough effort, but it takes a |
32 | | // lot of machinery, because of the design feature of _Generic that even the |
33 | | // untaken branches are required to pass all front-end validity checks such as |
34 | | // type-correctness. You can work around that by nesting further _Generics all |
35 | | // over the place to coerce things to the right type in untaken branches, but |
36 | | // what you get out is complicated, hard to guarantee its correctness, and |
37 | | // worst of all, gives _completely unreadable_ error messages if the user gets |
38 | | // the types wrong for an intrinsic call. |
39 | | // |
40 | | // Therefore, my strategy is to introduce a new __attribute__ that allows a |
41 | | // function to be mapped to a clang builtin even though it doesn't have the |
42 | | // same name, and then declare all the user-facing MVE function names with that |
43 | | // attribute, mapping each one directly to the clang builtin. And the |
44 | | // polymorphic ones have __attribute__((overloadable)) as well. So once the |
45 | | // compiler has resolved the overload, it knows the internal builtin ID of the |
46 | | // selected function, and can check the immediate arguments against that; and |
47 | | // if the user gets the types wrong in a call to a polymorphic intrinsic, they |
48 | | // get a completely clear error message showing all the declarations of that |
49 | | // function in the header file and explaining why each one doesn't fit their |
50 | | // call. |
51 | | // |
52 | | // The downside of this is that if every clang builtin has to correspond |
53 | | // exactly to a user-facing ACLE intrinsic, then you can't save work in the |
54 | | // frontend by doing it in the header file: CGBuiltin.cpp has to do the entire |
55 | | // job of converting an ACLE intrinsic call into LLVM IR. So the Tablegen |
56 | | // description for an MVE intrinsic has to contain a full description of the |
57 | | // sequence of IRBuilder calls that clang will need to make. |
58 | | // |
59 | | //===----------------------------------------------------------------------===// |
60 | | |
61 | | #include "llvm/ADT/APInt.h" |
62 | | #include "llvm/ADT/StringRef.h" |
63 | | #include "llvm/ADT/StringSwitch.h" |
64 | | #include "llvm/Support/Casting.h" |
65 | | #include "llvm/Support/raw_ostream.h" |
66 | | #include "llvm/TableGen/Error.h" |
67 | | #include "llvm/TableGen/Record.h" |
68 | | #include "llvm/TableGen/StringToOffsetTable.h" |
69 | | #include <cassert> |
70 | | #include <cstddef> |
71 | | #include <cstdint> |
72 | | #include <list> |
73 | | #include <map> |
74 | | #include <memory> |
75 | | #include <set> |
76 | | #include <string> |
77 | | #include <vector> |
78 | | |
79 | | using namespace llvm; |
80 | | |
81 | | namespace { |
82 | | |
83 | | class EmitterBase; |
84 | | class Result; |
85 | | |
86 | | // ----------------------------------------------------------------------------- |
87 | | // A system of classes to represent all the types we'll need to deal with in |
88 | | // the prototypes of intrinsics. |
89 | | // |
90 | | // Query methods include finding out the C name of a type; the "LLVM name" in |
91 | | // the sense of a C++ code snippet that can be used in the codegen function; |
92 | | // the suffix that represents the type in the ACLE intrinsic naming scheme |
93 | | // (e.g. 's32' represents int32_t in intrinsics such as vaddq_s32); whether the |
94 | | // type is floating-point related (hence should be under #ifdef in the MVE |
95 | | // header so that it isn't included in integer-only MVE mode); and the type's |
96 | | // size in bits. Not all subtypes support all these queries. |
97 | | |
98 | | class Type { |
99 | | public: |
100 | | enum class TypeKind { |
101 | | // Void appears as a return type (for store intrinsics, which are pure |
102 | | // side-effect). It's also used as the parameter type in the Tablegen |
103 | | // when an intrinsic doesn't need to come in various suffixed forms like |
104 | | // vfooq_s8,vfooq_u16,vfooq_f32. |
105 | | Void, |
106 | | |
107 | | // Scalar is used for ordinary int and float types of all sizes. |
108 | | Scalar, |
109 | | |
110 | | // Vector is used for anything that occupies exactly one MVE vector |
111 | | // register, i.e. {uint,int,float}NxM_t. |
112 | | Vector, |
113 | | |
114 | | // MultiVector is used for the {uint,int,float}NxMxK_t types used by the |
115 | | // interleaving load/store intrinsics v{ld,st}{2,4}q. |
116 | | MultiVector, |
117 | | |
118 | | // Predicate is used by all the predicated intrinsics. Its C |
119 | | // representation is mve_pred16_t (which is just an alias for uint16_t). |
120 | | // But we give more detail here, by indicating that a given predicate |
121 | | // instruction is logically regarded as a vector of i1 containing the |
122 | | // same number of lanes as the input vector type. So our Predicate type |
123 | | // comes with a lane count, which we use to decide which kind of <n x i1> |
124 | | // we'll invoke the pred_i2v IR intrinsic to translate it into. |
125 | | Predicate, |
126 | | |
127 | | // Pointer is used for pointer types (obviously), and comes with a flag |
128 | | // indicating whether it's a pointer to a const or mutable instance of |
129 | | // the pointee type. |
130 | | Pointer, |
131 | | }; |
132 | | |
133 | | private: |
134 | | const TypeKind TKind; |
135 | | |
136 | | protected: |
137 | 0 | Type(TypeKind K) : TKind(K) {} |
138 | | |
139 | | public: |
140 | 0 | TypeKind typeKind() const { return TKind; } |
141 | 0 | virtual ~Type() = default; |
142 | | virtual bool requiresFloat() const = 0; |
143 | | virtual bool requiresMVE() const = 0; |
144 | | virtual unsigned sizeInBits() const = 0; |
145 | | virtual std::string cName() const = 0; |
146 | 0 | virtual std::string llvmName() const { |
147 | 0 | PrintFatalError("no LLVM type name available for type " + cName()); |
148 | 0 | } |
149 | 0 | virtual std::string acleSuffix(std::string) const { |
150 | 0 | PrintFatalError("no ACLE suffix available for this type"); |
151 | 0 | } |
152 | | }; |
153 | | |
154 | | enum class ScalarTypeKind { SignedInt, UnsignedInt, Float }; |
155 | 0 | inline std::string toLetter(ScalarTypeKind kind) { |
156 | 0 | switch (kind) { |
157 | 0 | case ScalarTypeKind::SignedInt: |
158 | 0 | return "s"; |
159 | 0 | case ScalarTypeKind::UnsignedInt: |
160 | 0 | return "u"; |
161 | 0 | case ScalarTypeKind::Float: |
162 | 0 | return "f"; |
163 | 0 | } |
164 | 0 | llvm_unreachable("Unhandled ScalarTypeKind enum"); |
165 | 0 | } |
166 | 0 | inline std::string toCPrefix(ScalarTypeKind kind) { |
167 | 0 | switch (kind) { |
168 | 0 | case ScalarTypeKind::SignedInt: |
169 | 0 | return "int"; |
170 | 0 | case ScalarTypeKind::UnsignedInt: |
171 | 0 | return "uint"; |
172 | 0 | case ScalarTypeKind::Float: |
173 | 0 | return "float"; |
174 | 0 | } |
175 | 0 | llvm_unreachable("Unhandled ScalarTypeKind enum"); |
176 | 0 | } |
177 | | |
178 | | class VoidType : public Type { |
179 | | public: |
180 | 0 | VoidType() : Type(TypeKind::Void) {} |
181 | 0 | unsigned sizeInBits() const override { return 0; } |
182 | 0 | bool requiresFloat() const override { return false; } |
183 | 0 | bool requiresMVE() const override { return false; } |
184 | 0 | std::string cName() const override { return "void"; } |
185 | | |
186 | 0 | static bool classof(const Type *T) { return T->typeKind() == TypeKind::Void; } |
187 | 0 | std::string acleSuffix(std::string) const override { return ""; } |
188 | | }; |
189 | | |
190 | | class PointerType : public Type { |
191 | | const Type *Pointee; |
192 | | bool Const; |
193 | | |
194 | | public: |
195 | | PointerType(const Type *Pointee, bool Const) |
196 | 0 | : Type(TypeKind::Pointer), Pointee(Pointee), Const(Const) {} |
197 | 0 | unsigned sizeInBits() const override { return 32; } |
198 | 0 | bool requiresFloat() const override { return Pointee->requiresFloat(); } |
199 | 0 | bool requiresMVE() const override { return Pointee->requiresMVE(); } |
200 | 0 | std::string cName() const override { |
201 | 0 | std::string Name = Pointee->cName(); |
202 | | |
203 | | // The syntax for a pointer in C is different when the pointee is |
204 | | // itself a pointer. The MVE intrinsics don't contain any double |
205 | | // pointers, so we don't need to worry about that wrinkle. |
206 | 0 | assert(!isa<PointerType>(Pointee) && "Pointer to pointer not supported"); |
207 | | |
208 | 0 | if (Const) |
209 | 0 | Name = "const " + Name; |
210 | 0 | return Name + " *"; |
211 | 0 | } |
212 | 0 | std::string llvmName() const override { |
213 | 0 | return "llvm::PointerType::getUnqual(" + Pointee->llvmName() + ")"; |
214 | 0 | } |
215 | 0 | const Type *getPointeeType() const { return Pointee; } |
216 | | |
217 | 0 | static bool classof(const Type *T) { |
218 | 0 | return T->typeKind() == TypeKind::Pointer; |
219 | 0 | } |
220 | | }; |
221 | | |
222 | | // Base class for all the types that have a name of the form |
223 | | // [prefix][numbers]_t, like int32_t, uint16x8_t, float32x4x2_t. |
224 | | // |
225 | | // For this sub-hierarchy we invent a cNameBase() method which returns the |
226 | | // whole name except for the trailing "_t", so that Vector and MultiVector can |
227 | | // append an extra "x2" or whatever to their element type's cNameBase(). Then |
228 | | // the main cName() query method puts "_t" on the end for the final type name. |
229 | | |
230 | | class CRegularNamedType : public Type { |
231 | | using Type::Type; |
232 | | virtual std::string cNameBase() const = 0; |
233 | | |
234 | | public: |
235 | 0 | std::string cName() const override { return cNameBase() + "_t"; } |
236 | | }; |
237 | | |
238 | | class ScalarType : public CRegularNamedType { |
239 | | ScalarTypeKind Kind; |
240 | | unsigned Bits; |
241 | | std::string NameOverride; |
242 | | |
243 | | public: |
244 | 0 | ScalarType(const Record *Record) : CRegularNamedType(TypeKind::Scalar) { |
245 | 0 | Kind = StringSwitch<ScalarTypeKind>(Record->getValueAsString("kind")) |
246 | 0 | .Case("s", ScalarTypeKind::SignedInt) |
247 | 0 | .Case("u", ScalarTypeKind::UnsignedInt) |
248 | 0 | .Case("f", ScalarTypeKind::Float); |
249 | 0 | Bits = Record->getValueAsInt("size"); |
250 | 0 | NameOverride = std::string(Record->getValueAsString("nameOverride")); |
251 | 0 | } |
252 | 0 | unsigned sizeInBits() const override { return Bits; } |
253 | 0 | ScalarTypeKind kind() const { return Kind; } |
254 | 0 | std::string suffix() const { return toLetter(Kind) + utostr(Bits); } |
255 | 0 | std::string cNameBase() const override { |
256 | 0 | return toCPrefix(Kind) + utostr(Bits); |
257 | 0 | } |
258 | 0 | std::string cName() const override { |
259 | 0 | if (NameOverride.empty()) |
260 | 0 | return CRegularNamedType::cName(); |
261 | 0 | return NameOverride; |
262 | 0 | } |
263 | 0 | std::string llvmName() const override { |
264 | 0 | if (Kind == ScalarTypeKind::Float) { |
265 | 0 | if (Bits == 16) |
266 | 0 | return "HalfTy"; |
267 | 0 | if (Bits == 32) |
268 | 0 | return "FloatTy"; |
269 | 0 | if (Bits == 64) |
270 | 0 | return "DoubleTy"; |
271 | 0 | PrintFatalError("bad size for floating type"); |
272 | 0 | } |
273 | 0 | return "Int" + utostr(Bits) + "Ty"; |
274 | 0 | } |
275 | 0 | std::string acleSuffix(std::string overrideLetter) const override { |
276 | 0 | return "_" + (overrideLetter.size() ? overrideLetter : toLetter(Kind)) |
277 | 0 | + utostr(Bits); |
278 | 0 | } |
279 | 0 | bool isInteger() const { return Kind != ScalarTypeKind::Float; } |
280 | 0 | bool requiresFloat() const override { return !isInteger(); } |
281 | 0 | bool requiresMVE() const override { return false; } |
282 | 0 | bool hasNonstandardName() const { return !NameOverride.empty(); } |
283 | | |
284 | 0 | static bool classof(const Type *T) { |
285 | 0 | return T->typeKind() == TypeKind::Scalar; |
286 | 0 | } |
287 | | }; |
288 | | |
289 | | class VectorType : public CRegularNamedType { |
290 | | const ScalarType *Element; |
291 | | unsigned Lanes; |
292 | | |
293 | | public: |
294 | | VectorType(const ScalarType *Element, unsigned Lanes) |
295 | 0 | : CRegularNamedType(TypeKind::Vector), Element(Element), Lanes(Lanes) {} |
296 | 0 | unsigned sizeInBits() const override { return Lanes * Element->sizeInBits(); } |
297 | 0 | unsigned lanes() const { return Lanes; } |
298 | 0 | bool requiresFloat() const override { return Element->requiresFloat(); } |
299 | 0 | bool requiresMVE() const override { return true; } |
300 | 0 | std::string cNameBase() const override { |
301 | 0 | return Element->cNameBase() + "x" + utostr(Lanes); |
302 | 0 | } |
303 | 0 | std::string llvmName() const override { |
304 | 0 | return "llvm::FixedVectorType::get(" + Element->llvmName() + ", " + |
305 | 0 | utostr(Lanes) + ")"; |
306 | 0 | } |
307 | | |
308 | 0 | static bool classof(const Type *T) { |
309 | 0 | return T->typeKind() == TypeKind::Vector; |
310 | 0 | } |
311 | | }; |
312 | | |
313 | | class MultiVectorType : public CRegularNamedType { |
314 | | const VectorType *Element; |
315 | | unsigned Registers; |
316 | | |
317 | | public: |
318 | | MultiVectorType(unsigned Registers, const VectorType *Element) |
319 | | : CRegularNamedType(TypeKind::MultiVector), Element(Element), |
320 | 0 | Registers(Registers) {} |
321 | 0 | unsigned sizeInBits() const override { |
322 | 0 | return Registers * Element->sizeInBits(); |
323 | 0 | } |
324 | 0 | unsigned registers() const { return Registers; } |
325 | 0 | bool requiresFloat() const override { return Element->requiresFloat(); } |
326 | 0 | bool requiresMVE() const override { return true; } |
327 | 0 | std::string cNameBase() const override { |
328 | 0 | return Element->cNameBase() + "x" + utostr(Registers); |
329 | 0 | } |
330 | | |
331 | | // MultiVectorType doesn't override llvmName, because we don't expect to do |
332 | | // automatic code generation for the MVE intrinsics that use it: the {vld2, |
333 | | // vld4, vst2, vst4} family are the only ones that use these types, so it was |
334 | | // easier to hand-write the codegen for dealing with these structs than to |
335 | | // build in lots of extra automatic machinery that would only be used once. |
336 | | |
337 | 0 | static bool classof(const Type *T) { |
338 | 0 | return T->typeKind() == TypeKind::MultiVector; |
339 | 0 | } |
340 | | }; |
341 | | |
342 | | class PredicateType : public CRegularNamedType { |
343 | | unsigned Lanes; |
344 | | |
345 | | public: |
346 | | PredicateType(unsigned Lanes) |
347 | 0 | : CRegularNamedType(TypeKind::Predicate), Lanes(Lanes) {} |
348 | 0 | unsigned sizeInBits() const override { return 16; } |
349 | 0 | std::string cNameBase() const override { return "mve_pred16"; } |
350 | 0 | bool requiresFloat() const override { return false; }; |
351 | 0 | bool requiresMVE() const override { return true; } |
352 | 0 | std::string llvmName() const override { |
353 | 0 | return "llvm::FixedVectorType::get(Builder.getInt1Ty(), " + utostr(Lanes) + |
354 | 0 | ")"; |
355 | 0 | } |
356 | | |
357 | 0 | static bool classof(const Type *T) { |
358 | 0 | return T->typeKind() == TypeKind::Predicate; |
359 | 0 | } |
360 | | }; |
361 | | |
362 | | // ----------------------------------------------------------------------------- |
363 | | // Class to facilitate merging together the code generation for many intrinsics |
364 | | // by means of varying a few constant or type parameters. |
365 | | // |
366 | | // Most obviously, the intrinsics in a single parametrised family will have |
367 | | // code generation sequences that only differ in a type or two, e.g. vaddq_s8 |
368 | | // and vaddq_u16 will look the same apart from putting a different vector type |
369 | | // in the call to CGM.getIntrinsic(). But also, completely different intrinsics |
370 | | // will often code-generate in the same way, with only a different choice of |
371 | | // _which_ IR intrinsic they lower to (e.g. vaddq_m_s8 and vmulq_m_s8), but |
372 | | // marshalling the arguments and return values of the IR intrinsic in exactly |
373 | | // the same way. And others might differ only in some other kind of constant, |
374 | | // such as a lane index. |
375 | | // |
376 | | // So, when we generate the IR-building code for all these intrinsics, we keep |
377 | | // track of every value that could possibly be pulled out of the code and |
378 | | // stored ahead of time in a local variable. Then we group together intrinsics |
379 | | // by textual equivalence of the code that would result if _all_ those |
380 | | // parameters were stored in local variables. That gives us maximal sets that |
381 | | // can be implemented by a single piece of IR-building code by changing |
382 | | // parameter values ahead of time. |
383 | | // |
384 | | // After we've done that, we do a second pass in which we only allocate _some_ |
385 | | // of the parameters into local variables, by tracking which ones have the same |
386 | | // values as each other (so that a single variable can be reused) and which |
387 | | // ones are the same across the whole set (so that no variable is needed at |
388 | | // all). |
389 | | // |
390 | | // Hence the class below. Its allocParam method is invoked during code |
391 | | // generation by every method of a Result subclass (see below) that wants to |
392 | | // give it the opportunity to pull something out into a switchable parameter. |
393 | | // It returns a variable name for the parameter, or (if it's being used in the |
394 | | // second pass once we've decided that some parameters don't need to be stored |
395 | | // in variables after all) it might just return the input expression unchanged. |
396 | | |
397 | | struct CodeGenParamAllocator { |
398 | | // Accumulated during code generation |
399 | | std::vector<std::string> *ParamTypes = nullptr; |
400 | | std::vector<std::string> *ParamValues = nullptr; |
401 | | |
402 | | // Provided ahead of time in pass 2, to indicate which parameters are being |
403 | | // assigned to what. This vector contains an entry for each call to |
404 | | // allocParam expected during code gen (which we counted up in pass 1), and |
405 | | // indicates the number of the parameter variable that should be returned, or |
406 | | // -1 if this call shouldn't allocate a parameter variable at all. |
407 | | // |
408 | | // We rely on the recursive code generation working identically in passes 1 |
409 | | // and 2, so that the same list of calls to allocParam happen in the same |
410 | | // order. That guarantees that the parameter numbers recorded in pass 1 will |
411 | | // match the entries in this vector that store what EmitterBase::EmitBuiltinCG |
412 | | // decided to do about each one in pass 2. |
413 | | std::vector<int> *ParamNumberMap = nullptr; |
414 | | |
415 | | // Internally track how many things we've allocated |
416 | | unsigned nparams = 0; |
417 | | |
418 | 0 | std::string allocParam(StringRef Type, StringRef Value) { |
419 | 0 | unsigned ParamNumber; |
420 | |
|
421 | 0 | if (!ParamNumberMap) { |
422 | | // In pass 1, unconditionally assign a new parameter variable to every |
423 | | // value we're asked to process. |
424 | 0 | ParamNumber = nparams++; |
425 | 0 | } else { |
426 | | // In pass 2, consult the map provided by the caller to find out which |
427 | | // variable we should be keeping things in. |
428 | 0 | int MapValue = (*ParamNumberMap)[nparams++]; |
429 | 0 | if (MapValue < 0) |
430 | 0 | return std::string(Value); |
431 | 0 | ParamNumber = MapValue; |
432 | 0 | } |
433 | | |
434 | | // If we've allocated a new parameter variable for the first time, store |
435 | | // its type and value to be retrieved after codegen. |
436 | 0 | if (ParamTypes && ParamTypes->size() == ParamNumber) |
437 | 0 | ParamTypes->push_back(std::string(Type)); |
438 | 0 | if (ParamValues && ParamValues->size() == ParamNumber) |
439 | 0 | ParamValues->push_back(std::string(Value)); |
440 | | |
441 | | // Unimaginative naming scheme for parameter variables. |
442 | 0 | return "Param" + utostr(ParamNumber); |
443 | 0 | } |
444 | | }; |
445 | | |
446 | | // ----------------------------------------------------------------------------- |
447 | | // System of classes that represent all the intermediate values used during |
448 | | // code-generation for an intrinsic. |
449 | | // |
450 | | // The base class 'Result' can represent a value of the LLVM type 'Value', or |
451 | | // sometimes 'Address' (for loads/stores, including an alignment requirement). |
452 | | // |
453 | | // In the case where the Tablegen provides a value in the codegen dag as a |
454 | | // plain integer literal, the Result object we construct here will be one that |
455 | | // returns true from hasIntegerConstantValue(). This allows the generated C++ |
456 | | // code to use the constant directly in contexts which can take a literal |
457 | | // integer, such as Builder.CreateExtractValue(thing, 1), without going to the |
458 | | // effort of calling llvm::ConstantInt::get() and then pulling the constant |
459 | | // back out of the resulting llvm:Value later. |
460 | | |
461 | | class Result { |
462 | | public: |
463 | | // Convenient shorthand for the pointer type we'll be using everywhere. |
464 | | using Ptr = std::shared_ptr<Result>; |
465 | | |
466 | | private: |
467 | | Ptr Predecessor; |
468 | | std::string VarName; |
469 | | bool VarNameUsed = false; |
470 | | unsigned Visited = 0; |
471 | | |
472 | | public: |
473 | 0 | virtual ~Result() = default; |
474 | | using Scope = std::map<std::string, Ptr>; |
475 | | virtual void genCode(raw_ostream &OS, CodeGenParamAllocator &) const = 0; |
476 | 0 | virtual bool hasIntegerConstantValue() const { return false; } |
477 | 0 | virtual uint32_t integerConstantValue() const { return 0; } |
478 | 0 | virtual bool hasIntegerValue() const { return false; } |
479 | 0 | virtual std::string getIntegerValue(const std::string &) { |
480 | 0 | llvm_unreachable("non-working Result::getIntegerValue called"); |
481 | 0 | } |
482 | 0 | virtual std::string typeName() const { return "Value *"; } |
483 | | |
484 | | // Mostly, when a code-generation operation has a dependency on prior |
485 | | // operations, it's because it uses the output values of those operations as |
486 | | // inputs. But there's one exception, which is the use of 'seq' in Tablegen |
487 | | // to indicate that operations have to be performed in sequence regardless of |
488 | | // whether they use each others' output values. |
489 | | // |
490 | | // So, the actual generation of code is done by depth-first search, using the |
491 | | // prerequisites() method to get a list of all the other Results that have to |
492 | | // be computed before this one. That method divides into the 'predecessor', |
493 | | // set by setPredecessor() while processing a 'seq' dag node, and the list |
494 | | // returned by 'morePrerequisites', which each subclass implements to return |
495 | | // a list of the Results it uses as input to whatever its own computation is |
496 | | // doing. |
497 | | |
498 | 0 | virtual void morePrerequisites(std::vector<Ptr> &output) const {} |
499 | 0 | std::vector<Ptr> prerequisites() const { |
500 | 0 | std::vector<Ptr> ToRet; |
501 | 0 | if (Predecessor) |
502 | 0 | ToRet.push_back(Predecessor); |
503 | 0 | morePrerequisites(ToRet); |
504 | 0 | return ToRet; |
505 | 0 | } |
506 | | |
507 | 0 | void setPredecessor(Ptr p) { |
508 | | // If the user has nested one 'seq' node inside another, and this |
509 | | // method is called on the return value of the inner 'seq' (i.e. |
510 | | // the final item inside it), then we can't link _this_ node to p, |
511 | | // because it already has a predecessor. Instead, walk the chain |
512 | | // until we find the first item in the inner seq, and link that to |
513 | | // p, so that nesting seqs has the obvious effect of linking |
514 | | // everything together into one long sequential chain. |
515 | 0 | Result *r = this; |
516 | 0 | while (r->Predecessor) |
517 | 0 | r = r->Predecessor.get(); |
518 | 0 | r->Predecessor = p; |
519 | 0 | } |
520 | | |
521 | | // Each Result will be assigned a variable name in the output code, but not |
522 | | // all those variable names will actually be used (e.g. the return value of |
523 | | // Builder.CreateStore has void type, so nobody will want to refer to it). To |
524 | | // prevent annoying compiler warnings, we track whether each Result's |
525 | | // variable name was ever actually mentioned in subsequent statements, so |
526 | | // that it can be left out of the final generated code. |
527 | 0 | std::string varname() { |
528 | 0 | VarNameUsed = true; |
529 | 0 | return VarName; |
530 | 0 | } |
531 | 0 | void setVarname(const StringRef s) { VarName = std::string(s); } |
532 | 0 | bool varnameUsed() const { return VarNameUsed; } |
533 | | |
534 | | // Emit code to generate this result as a Value *. |
535 | 0 | virtual std::string asValue() { |
536 | 0 | return varname(); |
537 | 0 | } |
538 | | |
539 | | // Code generation happens in multiple passes. This method tracks whether a |
540 | | // Result has yet been visited in a given pass, without the need for a |
541 | | // tedious loop in between passes that goes through and resets a 'visited' |
542 | | // flag back to false: you just set Pass=1 the first time round, and Pass=2 |
543 | | // the second time. |
544 | 0 | bool needsVisiting(unsigned Pass) { |
545 | 0 | bool ToRet = Visited < Pass; |
546 | 0 | Visited = Pass; |
547 | 0 | return ToRet; |
548 | 0 | } |
549 | | }; |
550 | | |
551 | | // Result subclass that retrieves one of the arguments to the clang builtin |
552 | | // function. In cases where the argument has pointer type, we call |
553 | | // EmitPointerWithAlignment and store the result in a variable of type Address, |
554 | | // so that load and store IR nodes can know the right alignment. Otherwise, we |
555 | | // call EmitScalarExpr. |
556 | | // |
557 | | // There are aggregate parameters in the MVE intrinsics API, but we don't deal |
558 | | // with them in this Tablegen back end: they only arise in the vld2q/vld4q and |
559 | | // vst2q/vst4q family, which is few enough that we just write the code by hand |
560 | | // for those in CGBuiltin.cpp. |
561 | | class BuiltinArgResult : public Result { |
562 | | public: |
563 | | unsigned ArgNum; |
564 | | bool AddressType; |
565 | | bool Immediate; |
566 | | BuiltinArgResult(unsigned ArgNum, bool AddressType, bool Immediate) |
567 | 0 | : ArgNum(ArgNum), AddressType(AddressType), Immediate(Immediate) {} |
568 | 0 | void genCode(raw_ostream &OS, CodeGenParamAllocator &) const override { |
569 | 0 | OS << (AddressType ? "EmitPointerWithAlignment" : "EmitScalarExpr") |
570 | 0 | << "(E->getArg(" << ArgNum << "))"; |
571 | 0 | } |
572 | 0 | std::string typeName() const override { |
573 | 0 | return AddressType ? "Address" : Result::typeName(); |
574 | 0 | } |
575 | | // Emit code to generate this result as a Value *. |
576 | 0 | std::string asValue() override { |
577 | 0 | if (AddressType) |
578 | 0 | return "(" + varname() + ".getPointer())"; |
579 | 0 | return Result::asValue(); |
580 | 0 | } |
581 | 0 | bool hasIntegerValue() const override { return Immediate; } |
582 | 0 | std::string getIntegerValue(const std::string &IntType) override { |
583 | 0 | return "GetIntegerConstantValue<" + IntType + ">(E->getArg(" + |
584 | 0 | utostr(ArgNum) + "), getContext())"; |
585 | 0 | } |
586 | | }; |
587 | | |
588 | | // Result subclass for an integer literal appearing in Tablegen. This may need |
589 | | // to be turned into an llvm::Result by means of llvm::ConstantInt::get(), or |
590 | | // it may be used directly as an integer, depending on which IRBuilder method |
591 | | // it's being passed to. |
592 | | class IntLiteralResult : public Result { |
593 | | public: |
594 | | const ScalarType *IntegerType; |
595 | | uint32_t IntegerValue; |
596 | | IntLiteralResult(const ScalarType *IntegerType, uint32_t IntegerValue) |
597 | 0 | : IntegerType(IntegerType), IntegerValue(IntegerValue) {} |
598 | | void genCode(raw_ostream &OS, |
599 | 0 | CodeGenParamAllocator &ParamAlloc) const override { |
600 | 0 | OS << "llvm::ConstantInt::get(" |
601 | 0 | << ParamAlloc.allocParam("llvm::Type *", IntegerType->llvmName()) |
602 | 0 | << ", "; |
603 | 0 | OS << ParamAlloc.allocParam(IntegerType->cName(), utostr(IntegerValue)) |
604 | 0 | << ")"; |
605 | 0 | } |
606 | 0 | bool hasIntegerConstantValue() const override { return true; } |
607 | 0 | uint32_t integerConstantValue() const override { return IntegerValue; } |
608 | | }; |
609 | | |
610 | | // Result subclass representing a cast between different integer types. We use |
611 | | // our own ScalarType abstraction as the representation of the target type, |
612 | | // which gives both size and signedness. |
613 | | class IntCastResult : public Result { |
614 | | public: |
615 | | const ScalarType *IntegerType; |
616 | | Ptr V; |
617 | | IntCastResult(const ScalarType *IntegerType, Ptr V) |
618 | 0 | : IntegerType(IntegerType), V(V) {} |
619 | | void genCode(raw_ostream &OS, |
620 | 0 | CodeGenParamAllocator &ParamAlloc) const override { |
621 | 0 | OS << "Builder.CreateIntCast(" << V->varname() << ", " |
622 | 0 | << ParamAlloc.allocParam("llvm::Type *", IntegerType->llvmName()) << ", " |
623 | 0 | << ParamAlloc.allocParam("bool", |
624 | 0 | IntegerType->kind() == ScalarTypeKind::SignedInt |
625 | 0 | ? "true" |
626 | 0 | : "false") |
627 | 0 | << ")"; |
628 | 0 | } |
629 | 0 | void morePrerequisites(std::vector<Ptr> &output) const override { |
630 | 0 | output.push_back(V); |
631 | 0 | } |
632 | | }; |
633 | | |
634 | | // Result subclass representing a cast between different pointer types. |
635 | | class PointerCastResult : public Result { |
636 | | public: |
637 | | const PointerType *PtrType; |
638 | | Ptr V; |
639 | | PointerCastResult(const PointerType *PtrType, Ptr V) |
640 | 0 | : PtrType(PtrType), V(V) {} |
641 | | void genCode(raw_ostream &OS, |
642 | 0 | CodeGenParamAllocator &ParamAlloc) const override { |
643 | 0 | OS << "Builder.CreatePointerCast(" << V->asValue() << ", " |
644 | 0 | << ParamAlloc.allocParam("llvm::Type *", PtrType->llvmName()) << ")"; |
645 | 0 | } |
646 | 0 | void morePrerequisites(std::vector<Ptr> &output) const override { |
647 | 0 | output.push_back(V); |
648 | 0 | } |
649 | | }; |
650 | | |
651 | | // Result subclass representing a call to an IRBuilder method. Each IRBuilder |
652 | | // method we want to use will have a Tablegen record giving the method name and |
653 | | // describing any important details of how to call it, such as whether a |
654 | | // particular argument should be an integer constant instead of an llvm::Value. |
655 | | class IRBuilderResult : public Result { |
656 | | public: |
657 | | StringRef CallPrefix; |
658 | | std::vector<Ptr> Args; |
659 | | std::set<unsigned> AddressArgs; |
660 | | std::map<unsigned, std::string> IntegerArgs; |
661 | | IRBuilderResult(StringRef CallPrefix, std::vector<Ptr> Args, |
662 | | std::set<unsigned> AddressArgs, |
663 | | std::map<unsigned, std::string> IntegerArgs) |
664 | | : CallPrefix(CallPrefix), Args(Args), AddressArgs(AddressArgs), |
665 | 0 | IntegerArgs(IntegerArgs) {} |
666 | | void genCode(raw_ostream &OS, |
667 | 0 | CodeGenParamAllocator &ParamAlloc) const override { |
668 | 0 | OS << CallPrefix; |
669 | 0 | const char *Sep = ""; |
670 | 0 | for (unsigned i = 0, e = Args.size(); i < e; ++i) { |
671 | 0 | Ptr Arg = Args[i]; |
672 | 0 | auto it = IntegerArgs.find(i); |
673 | |
|
674 | 0 | OS << Sep; |
675 | 0 | Sep = ", "; |
676 | |
|
677 | 0 | if (it != IntegerArgs.end()) { |
678 | 0 | if (Arg->hasIntegerConstantValue()) |
679 | 0 | OS << "static_cast<" << it->second << ">(" |
680 | 0 | << ParamAlloc.allocParam(it->second, |
681 | 0 | utostr(Arg->integerConstantValue())) |
682 | 0 | << ")"; |
683 | 0 | else if (Arg->hasIntegerValue()) |
684 | 0 | OS << ParamAlloc.allocParam(it->second, |
685 | 0 | Arg->getIntegerValue(it->second)); |
686 | 0 | } else { |
687 | 0 | OS << Arg->varname(); |
688 | 0 | } |
689 | 0 | } |
690 | 0 | OS << ")"; |
691 | 0 | } |
692 | 0 | void morePrerequisites(std::vector<Ptr> &output) const override { |
693 | 0 | for (unsigned i = 0, e = Args.size(); i < e; ++i) { |
694 | 0 | Ptr Arg = Args[i]; |
695 | 0 | if (IntegerArgs.find(i) != IntegerArgs.end()) |
696 | 0 | continue; |
697 | 0 | output.push_back(Arg); |
698 | 0 | } |
699 | 0 | } |
700 | | }; |
701 | | |
702 | | // Result subclass representing making an Address out of a Value. |
703 | | class AddressResult : public Result { |
704 | | public: |
705 | | Ptr Arg; |
706 | | const Type *Ty; |
707 | | unsigned Align; |
708 | | AddressResult(Ptr Arg, const Type *Ty, unsigned Align) |
709 | 0 | : Arg(Arg), Ty(Ty), Align(Align) {} |
710 | | void genCode(raw_ostream &OS, |
711 | 0 | CodeGenParamAllocator &ParamAlloc) const override { |
712 | 0 | OS << "Address(" << Arg->varname() << ", " << Ty->llvmName() |
713 | 0 | << ", CharUnits::fromQuantity(" << Align << "))"; |
714 | 0 | } |
715 | 0 | std::string typeName() const override { |
716 | 0 | return "Address"; |
717 | 0 | } |
718 | 0 | void morePrerequisites(std::vector<Ptr> &output) const override { |
719 | 0 | output.push_back(Arg); |
720 | 0 | } |
721 | | }; |
722 | | |
723 | | // Result subclass representing a call to an IR intrinsic, which we first have |
724 | | // to look up using an Intrinsic::ID constant and an array of types. |
725 | | class IRIntrinsicResult : public Result { |
726 | | public: |
727 | | std::string IntrinsicID; |
728 | | std::vector<const Type *> ParamTypes; |
729 | | std::vector<Ptr> Args; |
730 | | IRIntrinsicResult(StringRef IntrinsicID, std::vector<const Type *> ParamTypes, |
731 | | std::vector<Ptr> Args) |
732 | | : IntrinsicID(std::string(IntrinsicID)), ParamTypes(ParamTypes), |
733 | 0 | Args(Args) {} |
734 | | void genCode(raw_ostream &OS, |
735 | 0 | CodeGenParamAllocator &ParamAlloc) const override { |
736 | 0 | std::string IntNo = ParamAlloc.allocParam( |
737 | 0 | "Intrinsic::ID", "Intrinsic::" + IntrinsicID); |
738 | 0 | OS << "Builder.CreateCall(CGM.getIntrinsic(" << IntNo; |
739 | 0 | if (!ParamTypes.empty()) { |
740 | 0 | OS << ", {"; |
741 | 0 | const char *Sep = ""; |
742 | 0 | for (auto T : ParamTypes) { |
743 | 0 | OS << Sep << ParamAlloc.allocParam("llvm::Type *", T->llvmName()); |
744 | 0 | Sep = ", "; |
745 | 0 | } |
746 | 0 | OS << "}"; |
747 | 0 | } |
748 | 0 | OS << "), {"; |
749 | 0 | const char *Sep = ""; |
750 | 0 | for (auto Arg : Args) { |
751 | 0 | OS << Sep << Arg->asValue(); |
752 | 0 | Sep = ", "; |
753 | 0 | } |
754 | 0 | OS << "})"; |
755 | 0 | } |
756 | 0 | void morePrerequisites(std::vector<Ptr> &output) const override { |
757 | 0 | output.insert(output.end(), Args.begin(), Args.end()); |
758 | 0 | } |
759 | | }; |
760 | | |
761 | | // Result subclass that specifies a type, for use in IRBuilder operations such |
762 | | // as CreateBitCast that take a type argument. |
763 | | class TypeResult : public Result { |
764 | | public: |
765 | | const Type *T; |
766 | 0 | TypeResult(const Type *T) : T(T) {} |
767 | 0 | void genCode(raw_ostream &OS, CodeGenParamAllocator &) const override { |
768 | 0 | OS << T->llvmName(); |
769 | 0 | } |
770 | 0 | std::string typeName() const override { |
771 | 0 | return "llvm::Type *"; |
772 | 0 | } |
773 | | }; |
774 | | |
775 | | // ----------------------------------------------------------------------------- |
776 | | // Class that describes a single ACLE intrinsic. |
777 | | // |
778 | | // A Tablegen record will typically describe more than one ACLE intrinsic, by |
779 | | // means of setting the 'list<Type> Params' field to a list of multiple |
780 | | // parameter types, so as to define vaddq_{s8,u8,...,f16,f32} all in one go. |
781 | | // We'll end up with one instance of ACLEIntrinsic for *each* parameter type, |
782 | | // rather than a single one for all of them. Hence, the constructor takes both |
783 | | // a Tablegen record and the current value of the parameter type. |
784 | | |
785 | | class ACLEIntrinsic { |
786 | | // Structure documenting that one of the intrinsic's arguments is required to |
787 | | // be a compile-time constant integer, and what constraints there are on its |
788 | | // value. Used when generating Sema checking code. |
789 | | struct ImmediateArg { |
790 | | enum class BoundsType { ExplicitRange, UInt }; |
791 | | BoundsType boundsType; |
792 | | int64_t i1, i2; |
793 | | StringRef ExtraCheckType, ExtraCheckArgs; |
794 | | const Type *ArgType; |
795 | | }; |
796 | | |
797 | | // For polymorphic intrinsics, FullName is the explicit name that uniquely |
798 | | // identifies this variant of the intrinsic, and ShortName is the name it |
799 | | // shares with at least one other intrinsic. |
800 | | std::string ShortName, FullName; |
801 | | |
802 | | // Name of the architecture extension, used in the Clang builtin name |
803 | | StringRef BuiltinExtension; |
804 | | |
805 | | // A very small number of intrinsics _only_ have a polymorphic |
806 | | // variant (vuninitializedq taking an unevaluated argument). |
807 | | bool PolymorphicOnly; |
808 | | |
809 | | // Another rarely-used flag indicating that the builtin doesn't |
810 | | // evaluate its argument(s) at all. |
811 | | bool NonEvaluating; |
812 | | |
813 | | // True if the intrinsic needs only the C header part (no codegen, semantic |
814 | | // checks, etc). Used for redeclaring MVE intrinsics in the arm_cde.h header. |
815 | | bool HeaderOnly; |
816 | | |
817 | | const Type *ReturnType; |
818 | | std::vector<const Type *> ArgTypes; |
819 | | std::map<unsigned, ImmediateArg> ImmediateArgs; |
820 | | Result::Ptr Code; |
821 | | |
822 | | std::map<std::string, std::string> CustomCodeGenArgs; |
823 | | |
824 | | // Recursive function that does the internals of code generation. |
825 | | void genCodeDfs(Result::Ptr V, std::list<Result::Ptr> &Used, |
826 | 0 | unsigned Pass) const { |
827 | 0 | if (!V->needsVisiting(Pass)) |
828 | 0 | return; |
829 | | |
830 | 0 | for (Result::Ptr W : V->prerequisites()) |
831 | 0 | genCodeDfs(W, Used, Pass); |
832 | |
|
833 | 0 | Used.push_back(V); |
834 | 0 | } |
835 | | |
836 | | public: |
837 | 0 | const std::string &shortName() const { return ShortName; } |
838 | 0 | const std::string &fullName() const { return FullName; } |
839 | 0 | StringRef builtinExtension() const { return BuiltinExtension; } |
840 | 0 | const Type *returnType() const { return ReturnType; } |
841 | 0 | const std::vector<const Type *> &argTypes() const { return ArgTypes; } |
842 | 0 | bool requiresFloat() const { |
843 | 0 | if (ReturnType->requiresFloat()) |
844 | 0 | return true; |
845 | 0 | for (const Type *T : ArgTypes) |
846 | 0 | if (T->requiresFloat()) |
847 | 0 | return true; |
848 | 0 | return false; |
849 | 0 | } |
850 | 0 | bool requiresMVE() const { |
851 | 0 | return ReturnType->requiresMVE() || |
852 | 0 | any_of(ArgTypes, [](const Type *T) { return T->requiresMVE(); }); |
853 | 0 | } |
854 | 0 | bool polymorphic() const { return ShortName != FullName; } |
855 | 0 | bool polymorphicOnly() const { return PolymorphicOnly; } |
856 | 0 | bool nonEvaluating() const { return NonEvaluating; } |
857 | 0 | bool headerOnly() const { return HeaderOnly; } |
858 | | |
859 | | // External entry point for code generation, called from EmitterBase. |
860 | | void genCode(raw_ostream &OS, CodeGenParamAllocator &ParamAlloc, |
861 | 0 | unsigned Pass) const { |
862 | 0 | assert(!headerOnly() && "Called genCode for header-only intrinsic"); |
863 | 0 | if (!hasCode()) { |
864 | 0 | for (auto kv : CustomCodeGenArgs) |
865 | 0 | OS << " " << kv.first << " = " << kv.second << ";\n"; |
866 | 0 | OS << " break; // custom code gen\n"; |
867 | 0 | return; |
868 | 0 | } |
869 | 0 | std::list<Result::Ptr> Used; |
870 | 0 | genCodeDfs(Code, Used, Pass); |
871 | |
|
872 | 0 | unsigned varindex = 0; |
873 | 0 | for (Result::Ptr V : Used) |
874 | 0 | if (V->varnameUsed()) |
875 | 0 | V->setVarname("Val" + utostr(varindex++)); |
876 | |
|
877 | 0 | for (Result::Ptr V : Used) { |
878 | 0 | OS << " "; |
879 | 0 | if (V == Used.back()) { |
880 | 0 | assert(!V->varnameUsed()); |
881 | 0 | OS << "return "; // FIXME: what if the top-level thing is void? |
882 | 0 | } else if (V->varnameUsed()) { |
883 | 0 | std::string Type = V->typeName(); |
884 | 0 | OS << V->typeName(); |
885 | 0 | if (!StringRef(Type).endswith("*")) |
886 | 0 | OS << " "; |
887 | 0 | OS << V->varname() << " = "; |
888 | 0 | } |
889 | 0 | V->genCode(OS, ParamAlloc); |
890 | 0 | OS << ";\n"; |
891 | 0 | } |
892 | 0 | } |
893 | 0 | bool hasCode() const { return Code != nullptr; } |
894 | | |
895 | 0 | static std::string signedHexLiteral(const llvm::APInt &iOrig) { |
896 | 0 | llvm::APInt i = iOrig.trunc(64); |
897 | 0 | SmallString<40> s; |
898 | 0 | i.toString(s, 16, true, true); |
899 | 0 | return std::string(s.str()); |
900 | 0 | } |
901 | | |
902 | 0 | std::string genSema() const { |
903 | 0 | assert(!headerOnly() && "Called genSema for header-only intrinsic"); |
904 | 0 | std::vector<std::string> SemaChecks; |
905 | |
|
906 | 0 | for (const auto &kv : ImmediateArgs) { |
907 | 0 | const ImmediateArg &IA = kv.second; |
908 | |
|
909 | 0 | llvm::APInt lo(128, 0), hi(128, 0); |
910 | 0 | switch (IA.boundsType) { |
911 | 0 | case ImmediateArg::BoundsType::ExplicitRange: |
912 | 0 | lo = IA.i1; |
913 | 0 | hi = IA.i2; |
914 | 0 | break; |
915 | 0 | case ImmediateArg::BoundsType::UInt: |
916 | 0 | lo = 0; |
917 | 0 | hi = llvm::APInt::getMaxValue(IA.i1).zext(128); |
918 | 0 | break; |
919 | 0 | } |
920 | | |
921 | 0 | std::string Index = utostr(kv.first); |
922 | | |
923 | | // Emit a range check if the legal range of values for the |
924 | | // immediate is smaller than the _possible_ range of values for |
925 | | // its type. |
926 | 0 | unsigned ArgTypeBits = IA.ArgType->sizeInBits(); |
927 | 0 | llvm::APInt ArgTypeRange = llvm::APInt::getMaxValue(ArgTypeBits).zext(128); |
928 | 0 | llvm::APInt ActualRange = (hi-lo).trunc(64).sext(128); |
929 | 0 | if (ActualRange.ult(ArgTypeRange)) |
930 | 0 | SemaChecks.push_back("SemaBuiltinConstantArgRange(TheCall, " + Index + |
931 | 0 | ", " + signedHexLiteral(lo) + ", " + |
932 | 0 | signedHexLiteral(hi) + ")"); |
933 | |
|
934 | 0 | if (!IA.ExtraCheckType.empty()) { |
935 | 0 | std::string Suffix; |
936 | 0 | if (!IA.ExtraCheckArgs.empty()) { |
937 | 0 | std::string tmp; |
938 | 0 | StringRef Arg = IA.ExtraCheckArgs; |
939 | 0 | if (Arg == "!lanesize") { |
940 | 0 | tmp = utostr(IA.ArgType->sizeInBits()); |
941 | 0 | Arg = tmp; |
942 | 0 | } |
943 | 0 | Suffix = (Twine(", ") + Arg).str(); |
944 | 0 | } |
945 | 0 | SemaChecks.push_back((Twine("SemaBuiltinConstantArg") + |
946 | 0 | IA.ExtraCheckType + "(TheCall, " + Index + |
947 | 0 | Suffix + ")") |
948 | 0 | .str()); |
949 | 0 | } |
950 | |
|
951 | 0 | assert(!SemaChecks.empty()); |
952 | 0 | } |
953 | 0 | if (SemaChecks.empty()) |
954 | 0 | return ""; |
955 | 0 | return join(std::begin(SemaChecks), std::end(SemaChecks), |
956 | 0 | " ||\n ") + |
957 | 0 | ";\n"; |
958 | 0 | } |
959 | | |
960 | | ACLEIntrinsic(EmitterBase &ME, Record *R, const Type *Param); |
961 | | }; |
962 | | |
963 | | // ----------------------------------------------------------------------------- |
964 | | // The top-level class that holds all the state from analyzing the entire |
965 | | // Tablegen input. |
966 | | |
967 | | class EmitterBase { |
968 | | protected: |
969 | | // EmitterBase holds a collection of all the types we've instantiated. |
970 | | VoidType Void; |
971 | | std::map<std::string, std::unique_ptr<ScalarType>> ScalarTypes; |
972 | | std::map<std::tuple<ScalarTypeKind, unsigned, unsigned>, |
973 | | std::unique_ptr<VectorType>> |
974 | | VectorTypes; |
975 | | std::map<std::pair<std::string, unsigned>, std::unique_ptr<MultiVectorType>> |
976 | | MultiVectorTypes; |
977 | | std::map<unsigned, std::unique_ptr<PredicateType>> PredicateTypes; |
978 | | std::map<std::string, std::unique_ptr<PointerType>> PointerTypes; |
979 | | |
980 | | // And all the ACLEIntrinsic instances we've created. |
981 | | std::map<std::string, std::unique_ptr<ACLEIntrinsic>> ACLEIntrinsics; |
982 | | |
983 | | public: |
984 | | // Methods to create a Type object, or return the right existing one from the |
985 | | // maps stored in this object. |
986 | 0 | const VoidType *getVoidType() { return &Void; } |
987 | 0 | const ScalarType *getScalarType(StringRef Name) { |
988 | 0 | return ScalarTypes[std::string(Name)].get(); |
989 | 0 | } |
990 | 0 | const ScalarType *getScalarType(Record *R) { |
991 | 0 | return getScalarType(R->getName()); |
992 | 0 | } |
993 | 0 | const VectorType *getVectorType(const ScalarType *ST, unsigned Lanes) { |
994 | 0 | std::tuple<ScalarTypeKind, unsigned, unsigned> key(ST->kind(), |
995 | 0 | ST->sizeInBits(), Lanes); |
996 | 0 | if (VectorTypes.find(key) == VectorTypes.end()) |
997 | 0 | VectorTypes[key] = std::make_unique<VectorType>(ST, Lanes); |
998 | 0 | return VectorTypes[key].get(); |
999 | 0 | } |
1000 | 0 | const VectorType *getVectorType(const ScalarType *ST) { |
1001 | 0 | return getVectorType(ST, 128 / ST->sizeInBits()); |
1002 | 0 | } |
1003 | | const MultiVectorType *getMultiVectorType(unsigned Registers, |
1004 | 0 | const VectorType *VT) { |
1005 | 0 | std::pair<std::string, unsigned> key(VT->cNameBase(), Registers); |
1006 | 0 | if (MultiVectorTypes.find(key) == MultiVectorTypes.end()) |
1007 | 0 | MultiVectorTypes[key] = std::make_unique<MultiVectorType>(Registers, VT); |
1008 | 0 | return MultiVectorTypes[key].get(); |
1009 | 0 | } |
1010 | 0 | const PredicateType *getPredicateType(unsigned Lanes) { |
1011 | 0 | unsigned key = Lanes; |
1012 | 0 | if (PredicateTypes.find(key) == PredicateTypes.end()) |
1013 | 0 | PredicateTypes[key] = std::make_unique<PredicateType>(Lanes); |
1014 | 0 | return PredicateTypes[key].get(); |
1015 | 0 | } |
1016 | 0 | const PointerType *getPointerType(const Type *T, bool Const) { |
1017 | 0 | PointerType PT(T, Const); |
1018 | 0 | std::string key = PT.cName(); |
1019 | 0 | if (PointerTypes.find(key) == PointerTypes.end()) |
1020 | 0 | PointerTypes[key] = std::make_unique<PointerType>(PT); |
1021 | 0 | return PointerTypes[key].get(); |
1022 | 0 | } |
1023 | | |
1024 | | // Methods to construct a type from various pieces of Tablegen. These are |
1025 | | // always called in the context of setting up a particular ACLEIntrinsic, so |
1026 | | // there's always an ambient parameter type (because we're iterating through |
1027 | | // the Params list in the Tablegen record for the intrinsic), which is used |
1028 | | // to expand Tablegen classes like 'Vector' which mean something different in |
1029 | | // each member of a parametric family. |
1030 | | const Type *getType(Record *R, const Type *Param); |
1031 | | const Type *getType(DagInit *D, const Type *Param); |
1032 | | const Type *getType(Init *I, const Type *Param); |
1033 | | |
1034 | | // Functions that translate the Tablegen representation of an intrinsic's |
1035 | | // code generation into a collection of Value objects (which will then be |
1036 | | // reprocessed to read out the actual C++ code included by CGBuiltin.cpp). |
1037 | | Result::Ptr getCodeForDag(DagInit *D, const Result::Scope &Scope, |
1038 | | const Type *Param); |
1039 | | Result::Ptr getCodeForDagArg(DagInit *D, unsigned ArgNum, |
1040 | | const Result::Scope &Scope, const Type *Param); |
1041 | | Result::Ptr getCodeForArg(unsigned ArgNum, const Type *ArgType, bool Promote, |
1042 | | bool Immediate); |
1043 | | |
1044 | | void GroupSemaChecks(std::map<std::string, std::set<std::string>> &Checks); |
1045 | | |
1046 | | // Constructor and top-level functions. |
1047 | | |
1048 | | EmitterBase(RecordKeeper &Records); |
1049 | 0 | virtual ~EmitterBase() = default; |
1050 | | |
1051 | | virtual void EmitHeader(raw_ostream &OS) = 0; |
1052 | | virtual void EmitBuiltinDef(raw_ostream &OS) = 0; |
1053 | | virtual void EmitBuiltinSema(raw_ostream &OS) = 0; |
1054 | | void EmitBuiltinCG(raw_ostream &OS); |
1055 | | void EmitBuiltinAliases(raw_ostream &OS); |
1056 | | }; |
1057 | | |
1058 | 0 | const Type *EmitterBase::getType(Init *I, const Type *Param) { |
1059 | 0 | if (auto Dag = dyn_cast<DagInit>(I)) |
1060 | 0 | return getType(Dag, Param); |
1061 | 0 | if (auto Def = dyn_cast<DefInit>(I)) |
1062 | 0 | return getType(Def->getDef(), Param); |
1063 | | |
1064 | 0 | PrintFatalError("Could not convert this value into a type"); |
1065 | 0 | } |
1066 | | |
1067 | 0 | const Type *EmitterBase::getType(Record *R, const Type *Param) { |
1068 | | // Pass to a subfield of any wrapper records. We don't expect more than one |
1069 | | // of these: immediate operands are used as plain numbers rather than as |
1070 | | // llvm::Value, so it's meaningless to promote their type anyway. |
1071 | 0 | if (R->isSubClassOf("Immediate")) |
1072 | 0 | R = R->getValueAsDef("type"); |
1073 | 0 | else if (R->isSubClassOf("unpromoted")) |
1074 | 0 | R = R->getValueAsDef("underlying_type"); |
1075 | |
|
1076 | 0 | if (R->getName() == "Void") |
1077 | 0 | return getVoidType(); |
1078 | 0 | if (R->isSubClassOf("PrimitiveType")) |
1079 | 0 | return getScalarType(R); |
1080 | 0 | if (R->isSubClassOf("ComplexType")) |
1081 | 0 | return getType(R->getValueAsDag("spec"), Param); |
1082 | | |
1083 | 0 | PrintFatalError(R->getLoc(), "Could not convert this record into a type"); |
1084 | 0 | } |
1085 | | |
1086 | 0 | const Type *EmitterBase::getType(DagInit *D, const Type *Param) { |
1087 | | // The meat of the getType system: types in the Tablegen are represented by a |
1088 | | // dag whose operators select sub-cases of this function. |
1089 | |
|
1090 | 0 | Record *Op = cast<DefInit>(D->getOperator())->getDef(); |
1091 | 0 | if (!Op->isSubClassOf("ComplexTypeOp")) |
1092 | 0 | PrintFatalError( |
1093 | 0 | "Expected ComplexTypeOp as dag operator in type expression"); |
1094 | |
|
1095 | 0 | if (Op->getName() == "CTO_Parameter") { |
1096 | 0 | if (isa<VoidType>(Param)) |
1097 | 0 | PrintFatalError("Parametric type in unparametrised context"); |
1098 | 0 | return Param; |
1099 | 0 | } |
1100 | | |
1101 | 0 | if (Op->getName() == "CTO_Vec") { |
1102 | 0 | const Type *Element = getType(D->getArg(0), Param); |
1103 | 0 | if (D->getNumArgs() == 1) { |
1104 | 0 | return getVectorType(cast<ScalarType>(Element)); |
1105 | 0 | } else { |
1106 | 0 | const Type *ExistingVector = getType(D->getArg(1), Param); |
1107 | 0 | return getVectorType(cast<ScalarType>(Element), |
1108 | 0 | cast<VectorType>(ExistingVector)->lanes()); |
1109 | 0 | } |
1110 | 0 | } |
1111 | | |
1112 | 0 | if (Op->getName() == "CTO_Pred") { |
1113 | 0 | const Type *Element = getType(D->getArg(0), Param); |
1114 | 0 | return getPredicateType(128 / Element->sizeInBits()); |
1115 | 0 | } |
1116 | | |
1117 | 0 | if (Op->isSubClassOf("CTO_Tuple")) { |
1118 | 0 | unsigned Registers = Op->getValueAsInt("n"); |
1119 | 0 | const Type *Element = getType(D->getArg(0), Param); |
1120 | 0 | return getMultiVectorType(Registers, cast<VectorType>(Element)); |
1121 | 0 | } |
1122 | | |
1123 | 0 | if (Op->isSubClassOf("CTO_Pointer")) { |
1124 | 0 | const Type *Pointee = getType(D->getArg(0), Param); |
1125 | 0 | return getPointerType(Pointee, Op->getValueAsBit("const")); |
1126 | 0 | } |
1127 | | |
1128 | 0 | if (Op->getName() == "CTO_CopyKind") { |
1129 | 0 | const ScalarType *STSize = cast<ScalarType>(getType(D->getArg(0), Param)); |
1130 | 0 | const ScalarType *STKind = cast<ScalarType>(getType(D->getArg(1), Param)); |
1131 | 0 | for (const auto &kv : ScalarTypes) { |
1132 | 0 | const ScalarType *RT = kv.second.get(); |
1133 | 0 | if (RT->kind() == STKind->kind() && RT->sizeInBits() == STSize->sizeInBits()) |
1134 | 0 | return RT; |
1135 | 0 | } |
1136 | 0 | PrintFatalError("Cannot find a type to satisfy CopyKind"); |
1137 | 0 | } |
1138 | | |
1139 | 0 | if (Op->isSubClassOf("CTO_ScaleSize")) { |
1140 | 0 | const ScalarType *STKind = cast<ScalarType>(getType(D->getArg(0), Param)); |
1141 | 0 | int Num = Op->getValueAsInt("num"), Denom = Op->getValueAsInt("denom"); |
1142 | 0 | unsigned DesiredSize = STKind->sizeInBits() * Num / Denom; |
1143 | 0 | for (const auto &kv : ScalarTypes) { |
1144 | 0 | const ScalarType *RT = kv.second.get(); |
1145 | 0 | if (RT->kind() == STKind->kind() && RT->sizeInBits() == DesiredSize) |
1146 | 0 | return RT; |
1147 | 0 | } |
1148 | 0 | PrintFatalError("Cannot find a type to satisfy ScaleSize"); |
1149 | 0 | } |
1150 | | |
1151 | 0 | PrintFatalError("Bad operator in type dag expression"); |
1152 | 0 | } |
1153 | | |
1154 | | Result::Ptr EmitterBase::getCodeForDag(DagInit *D, const Result::Scope &Scope, |
1155 | 0 | const Type *Param) { |
1156 | 0 | Record *Op = cast<DefInit>(D->getOperator())->getDef(); |
1157 | |
|
1158 | 0 | if (Op->getName() == "seq") { |
1159 | 0 | Result::Scope SubScope = Scope; |
1160 | 0 | Result::Ptr PrevV = nullptr; |
1161 | 0 | for (unsigned i = 0, e = D->getNumArgs(); i < e; ++i) { |
1162 | | // We don't use getCodeForDagArg here, because the argument name |
1163 | | // has different semantics in a seq |
1164 | 0 | Result::Ptr V = |
1165 | 0 | getCodeForDag(cast<DagInit>(D->getArg(i)), SubScope, Param); |
1166 | 0 | StringRef ArgName = D->getArgNameStr(i); |
1167 | 0 | if (!ArgName.empty()) |
1168 | 0 | SubScope[std::string(ArgName)] = V; |
1169 | 0 | if (PrevV) |
1170 | 0 | V->setPredecessor(PrevV); |
1171 | 0 | PrevV = V; |
1172 | 0 | } |
1173 | 0 | return PrevV; |
1174 | 0 | } else if (Op->isSubClassOf("Type")) { |
1175 | 0 | if (D->getNumArgs() != 1) |
1176 | 0 | PrintFatalError("Type casts should have exactly one argument"); |
1177 | 0 | const Type *CastType = getType(Op, Param); |
1178 | 0 | Result::Ptr Arg = getCodeForDagArg(D, 0, Scope, Param); |
1179 | 0 | if (const auto *ST = dyn_cast<ScalarType>(CastType)) { |
1180 | 0 | if (!ST->requiresFloat()) { |
1181 | 0 | if (Arg->hasIntegerConstantValue()) |
1182 | 0 | return std::make_shared<IntLiteralResult>( |
1183 | 0 | ST, Arg->integerConstantValue()); |
1184 | 0 | else |
1185 | 0 | return std::make_shared<IntCastResult>(ST, Arg); |
1186 | 0 | } |
1187 | 0 | } else if (const auto *PT = dyn_cast<PointerType>(CastType)) { |
1188 | 0 | return std::make_shared<PointerCastResult>(PT, Arg); |
1189 | 0 | } |
1190 | 0 | PrintFatalError("Unsupported type cast"); |
1191 | 0 | } else if (Op->getName() == "address") { |
1192 | 0 | if (D->getNumArgs() != 2) |
1193 | 0 | PrintFatalError("'address' should have two arguments"); |
1194 | 0 | Result::Ptr Arg = getCodeForDagArg(D, 0, Scope, Param); |
1195 | |
|
1196 | 0 | const Type *Ty = nullptr; |
1197 | 0 | if (auto *DI = dyn_cast<DagInit>(D->getArg(0))) |
1198 | 0 | if (auto *PTy = dyn_cast<PointerType>(getType(DI->getOperator(), Param))) |
1199 | 0 | Ty = PTy->getPointeeType(); |
1200 | 0 | if (!Ty) |
1201 | 0 | PrintFatalError("'address' pointer argument should be a pointer"); |
1202 | |
|
1203 | 0 | unsigned Alignment; |
1204 | 0 | if (auto *II = dyn_cast<IntInit>(D->getArg(1))) { |
1205 | 0 | Alignment = II->getValue(); |
1206 | 0 | } else { |
1207 | 0 | PrintFatalError("'address' alignment argument should be an integer"); |
1208 | 0 | } |
1209 | 0 | return std::make_shared<AddressResult>(Arg, Ty, Alignment); |
1210 | 0 | } else if (Op->getName() == "unsignedflag") { |
1211 | 0 | if (D->getNumArgs() != 1) |
1212 | 0 | PrintFatalError("unsignedflag should have exactly one argument"); |
1213 | 0 | Record *TypeRec = cast<DefInit>(D->getArg(0))->getDef(); |
1214 | 0 | if (!TypeRec->isSubClassOf("Type")) |
1215 | 0 | PrintFatalError("unsignedflag's argument should be a type"); |
1216 | 0 | if (const auto *ST = dyn_cast<ScalarType>(getType(TypeRec, Param))) { |
1217 | 0 | return std::make_shared<IntLiteralResult>( |
1218 | 0 | getScalarType("u32"), ST->kind() == ScalarTypeKind::UnsignedInt); |
1219 | 0 | } else { |
1220 | 0 | PrintFatalError("unsignedflag's argument should be a scalar type"); |
1221 | 0 | } |
1222 | 0 | } else if (Op->getName() == "bitsize") { |
1223 | 0 | if (D->getNumArgs() != 1) |
1224 | 0 | PrintFatalError("bitsize should have exactly one argument"); |
1225 | 0 | Record *TypeRec = cast<DefInit>(D->getArg(0))->getDef(); |
1226 | 0 | if (!TypeRec->isSubClassOf("Type")) |
1227 | 0 | PrintFatalError("bitsize's argument should be a type"); |
1228 | 0 | if (const auto *ST = dyn_cast<ScalarType>(getType(TypeRec, Param))) { |
1229 | 0 | return std::make_shared<IntLiteralResult>(getScalarType("u32"), |
1230 | 0 | ST->sizeInBits()); |
1231 | 0 | } else { |
1232 | 0 | PrintFatalError("bitsize's argument should be a scalar type"); |
1233 | 0 | } |
1234 | 0 | } else { |
1235 | 0 | std::vector<Result::Ptr> Args; |
1236 | 0 | for (unsigned i = 0, e = D->getNumArgs(); i < e; ++i) |
1237 | 0 | Args.push_back(getCodeForDagArg(D, i, Scope, Param)); |
1238 | 0 | if (Op->isSubClassOf("IRBuilderBase")) { |
1239 | 0 | std::set<unsigned> AddressArgs; |
1240 | 0 | std::map<unsigned, std::string> IntegerArgs; |
1241 | 0 | for (Record *sp : Op->getValueAsListOfDefs("special_params")) { |
1242 | 0 | unsigned Index = sp->getValueAsInt("index"); |
1243 | 0 | if (sp->isSubClassOf("IRBuilderAddrParam")) { |
1244 | 0 | AddressArgs.insert(Index); |
1245 | 0 | } else if (sp->isSubClassOf("IRBuilderIntParam")) { |
1246 | 0 | IntegerArgs[Index] = std::string(sp->getValueAsString("type")); |
1247 | 0 | } |
1248 | 0 | } |
1249 | 0 | return std::make_shared<IRBuilderResult>(Op->getValueAsString("prefix"), |
1250 | 0 | Args, AddressArgs, IntegerArgs); |
1251 | 0 | } else if (Op->isSubClassOf("IRIntBase")) { |
1252 | 0 | std::vector<const Type *> ParamTypes; |
1253 | 0 | for (Record *RParam : Op->getValueAsListOfDefs("params")) |
1254 | 0 | ParamTypes.push_back(getType(RParam, Param)); |
1255 | 0 | std::string IntName = std::string(Op->getValueAsString("intname")); |
1256 | 0 | if (Op->getValueAsBit("appendKind")) |
1257 | 0 | IntName += "_" + toLetter(cast<ScalarType>(Param)->kind()); |
1258 | 0 | return std::make_shared<IRIntrinsicResult>(IntName, ParamTypes, Args); |
1259 | 0 | } else { |
1260 | 0 | PrintFatalError("Unsupported dag node " + Op->getName()); |
1261 | 0 | } |
1262 | 0 | } |
1263 | 0 | } |
1264 | | |
1265 | | Result::Ptr EmitterBase::getCodeForDagArg(DagInit *D, unsigned ArgNum, |
1266 | | const Result::Scope &Scope, |
1267 | 0 | const Type *Param) { |
1268 | 0 | Init *Arg = D->getArg(ArgNum); |
1269 | 0 | StringRef Name = D->getArgNameStr(ArgNum); |
1270 | |
|
1271 | 0 | if (!Name.empty()) { |
1272 | 0 | if (!isa<UnsetInit>(Arg)) |
1273 | 0 | PrintFatalError( |
1274 | 0 | "dag operator argument should not have both a value and a name"); |
1275 | 0 | auto it = Scope.find(std::string(Name)); |
1276 | 0 | if (it == Scope.end()) |
1277 | 0 | PrintFatalError("unrecognized variable name '" + Name + "'"); |
1278 | 0 | return it->second; |
1279 | 0 | } |
1280 | | |
1281 | | // Sometimes the Arg is a bit. Prior to multiclass template argument |
1282 | | // checking, integers would sneak through the bit declaration, |
1283 | | // but now they really are bits. |
1284 | 0 | if (auto *BI = dyn_cast<BitInit>(Arg)) |
1285 | 0 | return std::make_shared<IntLiteralResult>(getScalarType("u32"), |
1286 | 0 | BI->getValue()); |
1287 | | |
1288 | 0 | if (auto *II = dyn_cast<IntInit>(Arg)) |
1289 | 0 | return std::make_shared<IntLiteralResult>(getScalarType("u32"), |
1290 | 0 | II->getValue()); |
1291 | | |
1292 | 0 | if (auto *DI = dyn_cast<DagInit>(Arg)) |
1293 | 0 | return getCodeForDag(DI, Scope, Param); |
1294 | | |
1295 | 0 | if (auto *DI = dyn_cast<DefInit>(Arg)) { |
1296 | 0 | Record *Rec = DI->getDef(); |
1297 | 0 | if (Rec->isSubClassOf("Type")) { |
1298 | 0 | const Type *T = getType(Rec, Param); |
1299 | 0 | return std::make_shared<TypeResult>(T); |
1300 | 0 | } |
1301 | 0 | } |
1302 | | |
1303 | 0 | PrintError("bad DAG argument type for code generation"); |
1304 | 0 | PrintNote("DAG: " + D->getAsString()); |
1305 | 0 | if (TypedInit *Typed = dyn_cast<TypedInit>(Arg)) |
1306 | 0 | PrintNote("argument type: " + Typed->getType()->getAsString()); |
1307 | 0 | PrintFatalNote("argument number " + Twine(ArgNum) + ": " + Arg->getAsString()); |
1308 | 0 | } |
1309 | | |
1310 | | Result::Ptr EmitterBase::getCodeForArg(unsigned ArgNum, const Type *ArgType, |
1311 | 0 | bool Promote, bool Immediate) { |
1312 | 0 | Result::Ptr V = std::make_shared<BuiltinArgResult>( |
1313 | 0 | ArgNum, isa<PointerType>(ArgType), Immediate); |
1314 | |
|
1315 | 0 | if (Promote) { |
1316 | 0 | if (const auto *ST = dyn_cast<ScalarType>(ArgType)) { |
1317 | 0 | if (ST->isInteger() && ST->sizeInBits() < 32) |
1318 | 0 | V = std::make_shared<IntCastResult>(getScalarType("u32"), V); |
1319 | 0 | } else if (const auto *PT = dyn_cast<PredicateType>(ArgType)) { |
1320 | 0 | V = std::make_shared<IntCastResult>(getScalarType("u32"), V); |
1321 | 0 | V = std::make_shared<IRIntrinsicResult>("arm_mve_pred_i2v", |
1322 | 0 | std::vector<const Type *>{PT}, |
1323 | 0 | std::vector<Result::Ptr>{V}); |
1324 | 0 | } |
1325 | 0 | } |
1326 | |
|
1327 | 0 | return V; |
1328 | 0 | } |
1329 | | |
1330 | | ACLEIntrinsic::ACLEIntrinsic(EmitterBase &ME, Record *R, const Type *Param) |
1331 | 0 | : ReturnType(ME.getType(R->getValueAsDef("ret"), Param)) { |
1332 | | // Derive the intrinsic's full name, by taking the name of the |
1333 | | // Tablegen record (or override) and appending the suffix from its |
1334 | | // parameter type. (If the intrinsic is unparametrised, its |
1335 | | // parameter type will be given as Void, which returns the empty |
1336 | | // string for acleSuffix.) |
1337 | 0 | StringRef BaseName = |
1338 | 0 | (R->isSubClassOf("NameOverride") ? R->getValueAsString("basename") |
1339 | 0 | : R->getName()); |
1340 | 0 | StringRef overrideLetter = R->getValueAsString("overrideKindLetter"); |
1341 | 0 | FullName = |
1342 | 0 | (Twine(BaseName) + Param->acleSuffix(std::string(overrideLetter))).str(); |
1343 | | |
1344 | | // Derive the intrinsic's polymorphic name, by removing components from the |
1345 | | // full name as specified by its 'pnt' member ('polymorphic name type'), |
1346 | | // which indicates how many type suffixes to remove, and any other piece of |
1347 | | // the name that should be removed. |
1348 | 0 | Record *PolymorphicNameType = R->getValueAsDef("pnt"); |
1349 | 0 | SmallVector<StringRef, 8> NameParts; |
1350 | 0 | StringRef(FullName).split(NameParts, '_'); |
1351 | 0 | for (unsigned i = 0, e = PolymorphicNameType->getValueAsInt( |
1352 | 0 | "NumTypeSuffixesToDiscard"); |
1353 | 0 | i < e; ++i) |
1354 | 0 | NameParts.pop_back(); |
1355 | 0 | if (!PolymorphicNameType->isValueUnset("ExtraSuffixToDiscard")) { |
1356 | 0 | StringRef ExtraSuffix = |
1357 | 0 | PolymorphicNameType->getValueAsString("ExtraSuffixToDiscard"); |
1358 | 0 | auto it = NameParts.end(); |
1359 | 0 | while (it != NameParts.begin()) { |
1360 | 0 | --it; |
1361 | 0 | if (*it == ExtraSuffix) { |
1362 | 0 | NameParts.erase(it); |
1363 | 0 | break; |
1364 | 0 | } |
1365 | 0 | } |
1366 | 0 | } |
1367 | 0 | ShortName = join(std::begin(NameParts), std::end(NameParts), "_"); |
1368 | |
|
1369 | 0 | BuiltinExtension = R->getValueAsString("builtinExtension"); |
1370 | |
|
1371 | 0 | PolymorphicOnly = R->getValueAsBit("polymorphicOnly"); |
1372 | 0 | NonEvaluating = R->getValueAsBit("nonEvaluating"); |
1373 | 0 | HeaderOnly = R->getValueAsBit("headerOnly"); |
1374 | | |
1375 | | // Process the intrinsic's argument list. |
1376 | 0 | DagInit *ArgsDag = R->getValueAsDag("args"); |
1377 | 0 | Result::Scope Scope; |
1378 | 0 | for (unsigned i = 0, e = ArgsDag->getNumArgs(); i < e; ++i) { |
1379 | 0 | Init *TypeInit = ArgsDag->getArg(i); |
1380 | |
|
1381 | 0 | bool Promote = true; |
1382 | 0 | if (auto TypeDI = dyn_cast<DefInit>(TypeInit)) |
1383 | 0 | if (TypeDI->getDef()->isSubClassOf("unpromoted")) |
1384 | 0 | Promote = false; |
1385 | | |
1386 | | // Work out the type of the argument, for use in the function prototype in |
1387 | | // the header file. |
1388 | 0 | const Type *ArgType = ME.getType(TypeInit, Param); |
1389 | 0 | ArgTypes.push_back(ArgType); |
1390 | | |
1391 | | // If the argument is a subclass of Immediate, record the details about |
1392 | | // what values it can take, for Sema checking. |
1393 | 0 | bool Immediate = false; |
1394 | 0 | if (auto TypeDI = dyn_cast<DefInit>(TypeInit)) { |
1395 | 0 | Record *TypeRec = TypeDI->getDef(); |
1396 | 0 | if (TypeRec->isSubClassOf("Immediate")) { |
1397 | 0 | Immediate = true; |
1398 | |
|
1399 | 0 | Record *Bounds = TypeRec->getValueAsDef("bounds"); |
1400 | 0 | ImmediateArg &IA = ImmediateArgs[i]; |
1401 | 0 | if (Bounds->isSubClassOf("IB_ConstRange")) { |
1402 | 0 | IA.boundsType = ImmediateArg::BoundsType::ExplicitRange; |
1403 | 0 | IA.i1 = Bounds->getValueAsInt("lo"); |
1404 | 0 | IA.i2 = Bounds->getValueAsInt("hi"); |
1405 | 0 | } else if (Bounds->getName() == "IB_UEltValue") { |
1406 | 0 | IA.boundsType = ImmediateArg::BoundsType::UInt; |
1407 | 0 | IA.i1 = Param->sizeInBits(); |
1408 | 0 | } else if (Bounds->getName() == "IB_LaneIndex") { |
1409 | 0 | IA.boundsType = ImmediateArg::BoundsType::ExplicitRange; |
1410 | 0 | IA.i1 = 0; |
1411 | 0 | IA.i2 = 128 / Param->sizeInBits() - 1; |
1412 | 0 | } else if (Bounds->isSubClassOf("IB_EltBit")) { |
1413 | 0 | IA.boundsType = ImmediateArg::BoundsType::ExplicitRange; |
1414 | 0 | IA.i1 = Bounds->getValueAsInt("base"); |
1415 | 0 | const Type *T = ME.getType(Bounds->getValueAsDef("type"), Param); |
1416 | 0 | IA.i2 = IA.i1 + T->sizeInBits() - 1; |
1417 | 0 | } else { |
1418 | 0 | PrintFatalError("unrecognised ImmediateBounds subclass"); |
1419 | 0 | } |
1420 | |
|
1421 | 0 | IA.ArgType = ArgType; |
1422 | |
|
1423 | 0 | if (!TypeRec->isValueUnset("extra")) { |
1424 | 0 | IA.ExtraCheckType = TypeRec->getValueAsString("extra"); |
1425 | 0 | if (!TypeRec->isValueUnset("extraarg")) |
1426 | 0 | IA.ExtraCheckArgs = TypeRec->getValueAsString("extraarg"); |
1427 | 0 | } |
1428 | 0 | } |
1429 | 0 | } |
1430 | | |
1431 | | // The argument will usually have a name in the arguments dag, which goes |
1432 | | // into the variable-name scope that the code gen will refer to. |
1433 | 0 | StringRef ArgName = ArgsDag->getArgNameStr(i); |
1434 | 0 | if (!ArgName.empty()) |
1435 | 0 | Scope[std::string(ArgName)] = |
1436 | 0 | ME.getCodeForArg(i, ArgType, Promote, Immediate); |
1437 | 0 | } |
1438 | | |
1439 | | // Finally, go through the codegen dag and translate it into a Result object |
1440 | | // (with an arbitrary DAG of depended-on Results hanging off it). |
1441 | 0 | DagInit *CodeDag = R->getValueAsDag("codegen"); |
1442 | 0 | Record *MainOp = cast<DefInit>(CodeDag->getOperator())->getDef(); |
1443 | 0 | if (MainOp->isSubClassOf("CustomCodegen")) { |
1444 | | // Or, if it's the special case of CustomCodegen, just accumulate |
1445 | | // a list of parameters we're going to assign to variables before |
1446 | | // breaking from the loop. |
1447 | 0 | CustomCodeGenArgs["CustomCodeGenType"] = |
1448 | 0 | (Twine("CustomCodeGen::") + MainOp->getValueAsString("type")).str(); |
1449 | 0 | for (unsigned i = 0, e = CodeDag->getNumArgs(); i < e; ++i) { |
1450 | 0 | StringRef Name = CodeDag->getArgNameStr(i); |
1451 | 0 | if (Name.empty()) { |
1452 | 0 | PrintFatalError("Operands to CustomCodegen should have names"); |
1453 | 0 | } else if (auto *II = dyn_cast<IntInit>(CodeDag->getArg(i))) { |
1454 | 0 | CustomCodeGenArgs[std::string(Name)] = itostr(II->getValue()); |
1455 | 0 | } else if (auto *SI = dyn_cast<StringInit>(CodeDag->getArg(i))) { |
1456 | 0 | CustomCodeGenArgs[std::string(Name)] = std::string(SI->getValue()); |
1457 | 0 | } else { |
1458 | 0 | PrintFatalError("Operands to CustomCodegen should be integers"); |
1459 | 0 | } |
1460 | 0 | } |
1461 | 0 | } else { |
1462 | 0 | Code = ME.getCodeForDag(CodeDag, Scope, Param); |
1463 | 0 | } |
1464 | 0 | } |
1465 | | |
1466 | 0 | EmitterBase::EmitterBase(RecordKeeper &Records) { |
1467 | | // Construct the whole EmitterBase. |
1468 | | |
1469 | | // First, look up all the instances of PrimitiveType. This gives us the list |
1470 | | // of vector typedefs we have to put in arm_mve.h, and also allows us to |
1471 | | // collect all the useful ScalarType instances into a big list so that we can |
1472 | | // use it for operations such as 'find the unsigned version of this signed |
1473 | | // integer type'. |
1474 | 0 | for (Record *R : Records.getAllDerivedDefinitions("PrimitiveType")) |
1475 | 0 | ScalarTypes[std::string(R->getName())] = std::make_unique<ScalarType>(R); |
1476 | | |
1477 | | // Now go through the instances of Intrinsic, and for each one, iterate |
1478 | | // through its list of type parameters making an ACLEIntrinsic for each one. |
1479 | 0 | for (Record *R : Records.getAllDerivedDefinitions("Intrinsic")) { |
1480 | 0 | for (Record *RParam : R->getValueAsListOfDefs("params")) { |
1481 | 0 | const Type *Param = getType(RParam, getVoidType()); |
1482 | 0 | auto Intrinsic = std::make_unique<ACLEIntrinsic>(*this, R, Param); |
1483 | 0 | ACLEIntrinsics[Intrinsic->fullName()] = std::move(Intrinsic); |
1484 | 0 | } |
1485 | 0 | } |
1486 | 0 | } |
1487 | | |
1488 | | /// A wrapper on raw_string_ostream that contains its own buffer rather than |
1489 | | /// having to point it at one elsewhere. (In other words, it works just like |
1490 | | /// std::ostringstream; also, this makes it convenient to declare a whole array |
1491 | | /// of them at once.) |
1492 | | /// |
1493 | | /// We have to set this up using multiple inheritance, to ensure that the |
1494 | | /// string member has been constructed before raw_string_ostream's constructor |
1495 | | /// is given a pointer to it. |
1496 | | class string_holder { |
1497 | | protected: |
1498 | | std::string S; |
1499 | | }; |
1500 | | class raw_self_contained_string_ostream : private string_holder, |
1501 | | public raw_string_ostream { |
1502 | | public: |
1503 | 0 | raw_self_contained_string_ostream() : raw_string_ostream(S) {} |
1504 | | }; |
1505 | | |
1506 | | const char LLVMLicenseHeader[] = |
1507 | | " *\n" |
1508 | | " *\n" |
1509 | | " * Part of the LLVM Project, under the Apache License v2.0 with LLVM" |
1510 | | " Exceptions.\n" |
1511 | | " * See https://llvm.org/LICENSE.txt for license information.\n" |
1512 | | " * SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception\n" |
1513 | | " *\n" |
1514 | | " *===-----------------------------------------------------------------" |
1515 | | "------===\n" |
1516 | | " */\n" |
1517 | | "\n"; |
1518 | | |
1519 | | // Machinery for the grouping of intrinsics by similar codegen. |
1520 | | // |
1521 | | // The general setup is that 'MergeableGroup' stores the things that a set of |
1522 | | // similarly shaped intrinsics have in common: the text of their code |
1523 | | // generation, and the number and type of their parameter variables. |
1524 | | // MergeableGroup is the key in a std::map whose value is a set of |
1525 | | // OutputIntrinsic, which stores the ways in which a particular intrinsic |
1526 | | // specializes the MergeableGroup's generic description: the function name and |
1527 | | // the _values_ of the parameter variables. |
1528 | | |
1529 | | struct ComparableStringVector : std::vector<std::string> { |
1530 | | // Infrastructure: a derived class of vector<string> which comes with an |
1531 | | // ordering, so that it can be used as a key in maps and an element in sets. |
1532 | | // There's no requirement on the ordering beyond being deterministic. |
1533 | 0 | bool operator<(const ComparableStringVector &rhs) const { |
1534 | 0 | if (size() != rhs.size()) |
1535 | 0 | return size() < rhs.size(); |
1536 | 0 | for (size_t i = 0, e = size(); i < e; ++i) |
1537 | 0 | if ((*this)[i] != rhs[i]) |
1538 | 0 | return (*this)[i] < rhs[i]; |
1539 | 0 | return false; |
1540 | 0 | } |
1541 | | }; |
1542 | | |
1543 | | struct OutputIntrinsic { |
1544 | | const ACLEIntrinsic *Int; |
1545 | | std::string Name; |
1546 | | ComparableStringVector ParamValues; |
1547 | 0 | bool operator<(const OutputIntrinsic &rhs) const { |
1548 | 0 | if (Name != rhs.Name) |
1549 | 0 | return Name < rhs.Name; |
1550 | 0 | return ParamValues < rhs.ParamValues; |
1551 | 0 | } |
1552 | | }; |
1553 | | struct MergeableGroup { |
1554 | | std::string Code; |
1555 | | ComparableStringVector ParamTypes; |
1556 | 0 | bool operator<(const MergeableGroup &rhs) const { |
1557 | 0 | if (Code != rhs.Code) |
1558 | 0 | return Code < rhs.Code; |
1559 | 0 | return ParamTypes < rhs.ParamTypes; |
1560 | 0 | } |
1561 | | }; |
1562 | | |
1563 | 0 | void EmitterBase::EmitBuiltinCG(raw_ostream &OS) { |
1564 | | // Pass 1: generate code for all the intrinsics as if every type or constant |
1565 | | // that can possibly be abstracted out into a parameter variable will be. |
1566 | | // This identifies the sets of intrinsics we'll group together into a single |
1567 | | // piece of code generation. |
1568 | |
|
1569 | 0 | std::map<MergeableGroup, std::set<OutputIntrinsic>> MergeableGroupsPrelim; |
1570 | |
|
1571 | 0 | for (const auto &kv : ACLEIntrinsics) { |
1572 | 0 | const ACLEIntrinsic &Int = *kv.second; |
1573 | 0 | if (Int.headerOnly()) |
1574 | 0 | continue; |
1575 | | |
1576 | 0 | MergeableGroup MG; |
1577 | 0 | OutputIntrinsic OI; |
1578 | |
|
1579 | 0 | OI.Int = ∬ |
1580 | 0 | OI.Name = Int.fullName(); |
1581 | 0 | CodeGenParamAllocator ParamAllocPrelim{&MG.ParamTypes, &OI.ParamValues}; |
1582 | 0 | raw_string_ostream OS(MG.Code); |
1583 | 0 | Int.genCode(OS, ParamAllocPrelim, 1); |
1584 | 0 | OS.flush(); |
1585 | |
|
1586 | 0 | MergeableGroupsPrelim[MG].insert(OI); |
1587 | 0 | } |
1588 | | |
1589 | | // Pass 2: for each of those groups, optimize the parameter variable set by |
1590 | | // eliminating 'parameters' that are the same for all intrinsics in the |
1591 | | // group, and merging together pairs of parameter variables that take the |
1592 | | // same values as each other for all intrinsics in the group. |
1593 | |
|
1594 | 0 | std::map<MergeableGroup, std::set<OutputIntrinsic>> MergeableGroups; |
1595 | |
|
1596 | 0 | for (const auto &kv : MergeableGroupsPrelim) { |
1597 | 0 | const MergeableGroup &MG = kv.first; |
1598 | 0 | std::vector<int> ParamNumbers; |
1599 | 0 | std::map<ComparableStringVector, int> ParamNumberMap; |
1600 | | |
1601 | | // Loop over the parameters for this group. |
1602 | 0 | for (size_t i = 0, e = MG.ParamTypes.size(); i < e; ++i) { |
1603 | | // Is this parameter the same for all intrinsics in the group? |
1604 | 0 | const OutputIntrinsic &OI_first = *kv.second.begin(); |
1605 | 0 | bool Constant = all_of(kv.second, [&](const OutputIntrinsic &OI) { |
1606 | 0 | return OI.ParamValues[i] == OI_first.ParamValues[i]; |
1607 | 0 | }); |
1608 | | |
1609 | | // If so, record it as -1, meaning 'no parameter variable needed'. Then |
1610 | | // the corresponding call to allocParam in pass 2 will not generate a |
1611 | | // variable at all, and just use the value inline. |
1612 | 0 | if (Constant) { |
1613 | 0 | ParamNumbers.push_back(-1); |
1614 | 0 | continue; |
1615 | 0 | } |
1616 | | |
1617 | | // Otherwise, make a list of the values this parameter takes for each |
1618 | | // intrinsic, and see if that value vector matches anything we already |
1619 | | // have. We also record the parameter type, so that we don't accidentally |
1620 | | // match up two parameter variables with different types. (Not that |
1621 | | // there's much chance of them having textually equivalent values, but in |
1622 | | // _principle_ it could happen.) |
1623 | 0 | ComparableStringVector key; |
1624 | 0 | key.push_back(MG.ParamTypes[i]); |
1625 | 0 | for (const auto &OI : kv.second) |
1626 | 0 | key.push_back(OI.ParamValues[i]); |
1627 | |
|
1628 | 0 | auto Found = ParamNumberMap.find(key); |
1629 | 0 | if (Found != ParamNumberMap.end()) { |
1630 | | // Yes, an existing parameter variable can be reused for this. |
1631 | 0 | ParamNumbers.push_back(Found->second); |
1632 | 0 | continue; |
1633 | 0 | } |
1634 | | |
1635 | | // No, we need a new parameter variable. |
1636 | 0 | int ExistingIndex = ParamNumberMap.size(); |
1637 | 0 | ParamNumberMap[key] = ExistingIndex; |
1638 | 0 | ParamNumbers.push_back(ExistingIndex); |
1639 | 0 | } |
1640 | | |
1641 | | // Now we're ready to do the pass 2 code generation, which will emit the |
1642 | | // reduced set of parameter variables we've just worked out. |
1643 | |
|
1644 | 0 | for (const auto &OI_prelim : kv.second) { |
1645 | 0 | const ACLEIntrinsic *Int = OI_prelim.Int; |
1646 | |
|
1647 | 0 | MergeableGroup MG; |
1648 | 0 | OutputIntrinsic OI; |
1649 | |
|
1650 | 0 | OI.Int = OI_prelim.Int; |
1651 | 0 | OI.Name = OI_prelim.Name; |
1652 | 0 | CodeGenParamAllocator ParamAlloc{&MG.ParamTypes, &OI.ParamValues, |
1653 | 0 | &ParamNumbers}; |
1654 | 0 | raw_string_ostream OS(MG.Code); |
1655 | 0 | Int->genCode(OS, ParamAlloc, 2); |
1656 | 0 | OS.flush(); |
1657 | |
|
1658 | 0 | MergeableGroups[MG].insert(OI); |
1659 | 0 | } |
1660 | 0 | } |
1661 | | |
1662 | | // Output the actual C++ code. |
1663 | |
|
1664 | 0 | for (const auto &kv : MergeableGroups) { |
1665 | 0 | const MergeableGroup &MG = kv.first; |
1666 | | |
1667 | | // List of case statements in the main switch on BuiltinID, and an open |
1668 | | // brace. |
1669 | 0 | const char *prefix = ""; |
1670 | 0 | for (const auto &OI : kv.second) { |
1671 | 0 | OS << prefix << "case ARM::BI__builtin_arm_" << OI.Int->builtinExtension() |
1672 | 0 | << "_" << OI.Name << ":"; |
1673 | |
|
1674 | 0 | prefix = "\n"; |
1675 | 0 | } |
1676 | 0 | OS << " {\n"; |
1677 | |
|
1678 | 0 | if (!MG.ParamTypes.empty()) { |
1679 | | // If we've got some parameter variables, then emit their declarations... |
1680 | 0 | for (size_t i = 0, e = MG.ParamTypes.size(); i < e; ++i) { |
1681 | 0 | StringRef Type = MG.ParamTypes[i]; |
1682 | 0 | OS << " " << Type; |
1683 | 0 | if (!Type.endswith("*")) |
1684 | 0 | OS << " "; |
1685 | 0 | OS << " Param" << utostr(i) << ";\n"; |
1686 | 0 | } |
1687 | | |
1688 | | // ... and an inner switch on BuiltinID that will fill them in with each |
1689 | | // individual intrinsic's values. |
1690 | 0 | OS << " switch (BuiltinID) {\n"; |
1691 | 0 | for (const auto &OI : kv.second) { |
1692 | 0 | OS << " case ARM::BI__builtin_arm_" << OI.Int->builtinExtension() |
1693 | 0 | << "_" << OI.Name << ":\n"; |
1694 | 0 | for (size_t i = 0, e = MG.ParamTypes.size(); i < e; ++i) |
1695 | 0 | OS << " Param" << utostr(i) << " = " << OI.ParamValues[i] << ";\n"; |
1696 | 0 | OS << " break;\n"; |
1697 | 0 | } |
1698 | 0 | OS << " }\n"; |
1699 | 0 | } |
1700 | | |
1701 | | // And finally, output the code, and close the outer pair of braces. (The |
1702 | | // code will always end with a 'return' statement, so we need not insert a |
1703 | | // 'break' here.) |
1704 | 0 | OS << MG.Code << "}\n"; |
1705 | 0 | } |
1706 | 0 | } |
1707 | | |
1708 | 0 | void EmitterBase::EmitBuiltinAliases(raw_ostream &OS) { |
1709 | | // Build a sorted table of: |
1710 | | // - intrinsic id number |
1711 | | // - full name |
1712 | | // - polymorphic name or -1 |
1713 | 0 | StringToOffsetTable StringTable; |
1714 | 0 | OS << "static const IntrinToName MapData[] = {\n"; |
1715 | 0 | for (const auto &kv : ACLEIntrinsics) { |
1716 | 0 | const ACLEIntrinsic &Int = *kv.second; |
1717 | 0 | if (Int.headerOnly()) |
1718 | 0 | continue; |
1719 | 0 | int32_t ShortNameOffset = |
1720 | 0 | Int.polymorphic() ? StringTable.GetOrAddStringOffset(Int.shortName()) |
1721 | 0 | : -1; |
1722 | 0 | OS << " { ARM::BI__builtin_arm_" << Int.builtinExtension() << "_" |
1723 | 0 | << Int.fullName() << ", " |
1724 | 0 | << StringTable.GetOrAddStringOffset(Int.fullName()) << ", " |
1725 | 0 | << ShortNameOffset << "},\n"; |
1726 | 0 | } |
1727 | 0 | OS << "};\n\n"; |
1728 | |
|
1729 | 0 | OS << "ArrayRef<IntrinToName> Map(MapData);\n\n"; |
1730 | |
|
1731 | 0 | OS << "static const char IntrinNames[] = {\n"; |
1732 | 0 | StringTable.EmitString(OS); |
1733 | 0 | OS << "};\n\n"; |
1734 | 0 | } |
1735 | | |
1736 | | void EmitterBase::GroupSemaChecks( |
1737 | 0 | std::map<std::string, std::set<std::string>> &Checks) { |
1738 | 0 | for (const auto &kv : ACLEIntrinsics) { |
1739 | 0 | const ACLEIntrinsic &Int = *kv.second; |
1740 | 0 | if (Int.headerOnly()) |
1741 | 0 | continue; |
1742 | 0 | std::string Check = Int.genSema(); |
1743 | 0 | if (!Check.empty()) |
1744 | 0 | Checks[Check].insert(Int.fullName()); |
1745 | 0 | } |
1746 | 0 | } |
1747 | | |
1748 | | // ----------------------------------------------------------------------------- |
1749 | | // The class used for generating arm_mve.h and related Clang bits |
1750 | | // |
1751 | | |
1752 | | class MveEmitter : public EmitterBase { |
1753 | | public: |
1754 | 0 | MveEmitter(RecordKeeper &Records) : EmitterBase(Records){}; |
1755 | | void EmitHeader(raw_ostream &OS) override; |
1756 | | void EmitBuiltinDef(raw_ostream &OS) override; |
1757 | | void EmitBuiltinSema(raw_ostream &OS) override; |
1758 | | }; |
1759 | | |
1760 | 0 | void MveEmitter::EmitHeader(raw_ostream &OS) { |
1761 | | // Accumulate pieces of the header file that will be enabled under various |
1762 | | // different combinations of #ifdef. The index into parts[] is made up of |
1763 | | // the following bit flags. |
1764 | 0 | constexpr unsigned Float = 1; |
1765 | 0 | constexpr unsigned UseUserNamespace = 2; |
1766 | |
|
1767 | 0 | constexpr unsigned NumParts = 4; |
1768 | 0 | raw_self_contained_string_ostream parts[NumParts]; |
1769 | | |
1770 | | // Write typedefs for all the required vector types, and a few scalar |
1771 | | // types that don't already have the name we want them to have. |
1772 | |
|
1773 | 0 | parts[0] << "typedef uint16_t mve_pred16_t;\n"; |
1774 | 0 | parts[Float] << "typedef __fp16 float16_t;\n" |
1775 | 0 | "typedef float float32_t;\n"; |
1776 | 0 | for (const auto &kv : ScalarTypes) { |
1777 | 0 | const ScalarType *ST = kv.second.get(); |
1778 | 0 | if (ST->hasNonstandardName()) |
1779 | 0 | continue; |
1780 | 0 | raw_ostream &OS = parts[ST->requiresFloat() ? Float : 0]; |
1781 | 0 | const VectorType *VT = getVectorType(ST); |
1782 | |
|
1783 | 0 | OS << "typedef __attribute__((__neon_vector_type__(" << VT->lanes() |
1784 | 0 | << "), __clang_arm_mve_strict_polymorphism)) " << ST->cName() << " " |
1785 | 0 | << VT->cName() << ";\n"; |
1786 | | |
1787 | | // Every vector type also comes with a pair of multi-vector types for |
1788 | | // the VLD2 and VLD4 instructions. |
1789 | 0 | for (unsigned n = 2; n <= 4; n += 2) { |
1790 | 0 | const MultiVectorType *MT = getMultiVectorType(n, VT); |
1791 | 0 | OS << "typedef struct { " << VT->cName() << " val[" << n << "]; } " |
1792 | 0 | << MT->cName() << ";\n"; |
1793 | 0 | } |
1794 | 0 | } |
1795 | 0 | parts[0] << "\n"; |
1796 | 0 | parts[Float] << "\n"; |
1797 | | |
1798 | | // Write declarations for all the intrinsics. |
1799 | |
|
1800 | 0 | for (const auto &kv : ACLEIntrinsics) { |
1801 | 0 | const ACLEIntrinsic &Int = *kv.second; |
1802 | | |
1803 | | // We generate each intrinsic twice, under its full unambiguous |
1804 | | // name and its shorter polymorphic name (if the latter exists). |
1805 | 0 | for (bool Polymorphic : {false, true}) { |
1806 | 0 | if (Polymorphic && !Int.polymorphic()) |
1807 | 0 | continue; |
1808 | 0 | if (!Polymorphic && Int.polymorphicOnly()) |
1809 | 0 | continue; |
1810 | | |
1811 | | // We also generate each intrinsic under a name like __arm_vfooq |
1812 | | // (which is in C language implementation namespace, so it's |
1813 | | // safe to define in any conforming user program) and a shorter |
1814 | | // one like vfooq (which is in user namespace, so a user might |
1815 | | // reasonably have used it for something already). If so, they |
1816 | | // can #define __ARM_MVE_PRESERVE_USER_NAMESPACE before |
1817 | | // including the header, which will suppress the shorter names |
1818 | | // and leave only the implementation-namespace ones. Then they |
1819 | | // have to write __arm_vfooq everywhere, of course. |
1820 | | |
1821 | 0 | for (bool UserNamespace : {false, true}) { |
1822 | 0 | raw_ostream &OS = parts[(Int.requiresFloat() ? Float : 0) | |
1823 | 0 | (UserNamespace ? UseUserNamespace : 0)]; |
1824 | | |
1825 | | // Make the name of the function in this declaration. |
1826 | |
|
1827 | 0 | std::string FunctionName = |
1828 | 0 | Polymorphic ? Int.shortName() : Int.fullName(); |
1829 | 0 | if (!UserNamespace) |
1830 | 0 | FunctionName = "__arm_" + FunctionName; |
1831 | | |
1832 | | // Make strings for the types involved in the function's |
1833 | | // prototype. |
1834 | |
|
1835 | 0 | std::string RetTypeName = Int.returnType()->cName(); |
1836 | 0 | if (!StringRef(RetTypeName).endswith("*")) |
1837 | 0 | RetTypeName += " "; |
1838 | |
|
1839 | 0 | std::vector<std::string> ArgTypeNames; |
1840 | 0 | for (const Type *ArgTypePtr : Int.argTypes()) |
1841 | 0 | ArgTypeNames.push_back(ArgTypePtr->cName()); |
1842 | 0 | std::string ArgTypesString = |
1843 | 0 | join(std::begin(ArgTypeNames), std::end(ArgTypeNames), ", "); |
1844 | | |
1845 | | // Emit the actual declaration. All these functions are |
1846 | | // declared 'static inline' without a body, which is fine |
1847 | | // provided clang recognizes them as builtins, and has the |
1848 | | // effect that this type signature is used in place of the one |
1849 | | // that Builtins.def didn't provide. That's how we can get |
1850 | | // structure types that weren't defined until this header was |
1851 | | // included to be part of the type signature of a builtin that |
1852 | | // was known to clang already. |
1853 | | // |
1854 | | // The declarations use __attribute__(__clang_arm_builtin_alias), |
1855 | | // so that each function declared will be recognized as the |
1856 | | // appropriate MVE builtin in spite of its user-facing name. |
1857 | | // |
1858 | | // (That's better than making them all wrapper functions, |
1859 | | // partly because it avoids any compiler error message citing |
1860 | | // the wrapper function definition instead of the user's code, |
1861 | | // and mostly because some MVE intrinsics have arguments |
1862 | | // required to be compile-time constants, and that property |
1863 | | // can't be propagated through a wrapper function. It can be |
1864 | | // propagated through a macro, but macros can't be overloaded |
1865 | | // on argument types very easily - you have to use _Generic, |
1866 | | // which makes error messages very confusing when the user |
1867 | | // gets it wrong.) |
1868 | | // |
1869 | | // Finally, the polymorphic versions of the intrinsics are |
1870 | | // also defined with __attribute__(overloadable), so that when |
1871 | | // the same name is defined with several type signatures, the |
1872 | | // right thing happens. Each one of the overloaded |
1873 | | // declarations is given a different builtin id, which |
1874 | | // has exactly the effect we want: first clang resolves the |
1875 | | // overload to the right function, then it knows which builtin |
1876 | | // it's referring to, and then the Sema checking for that |
1877 | | // builtin can check further things like the constant |
1878 | | // arguments. |
1879 | | // |
1880 | | // One more subtlety is the newline just before the return |
1881 | | // type name. That's a cosmetic tweak to make the error |
1882 | | // messages legible if the user gets the types wrong in a call |
1883 | | // to a polymorphic function: this way, clang will print just |
1884 | | // the _final_ line of each declaration in the header, to show |
1885 | | // the type signatures that would have been legal. So all the |
1886 | | // confusing machinery with __attribute__ is left out of the |
1887 | | // error message, and the user sees something that's more or |
1888 | | // less self-documenting: "here's a list of actually readable |
1889 | | // type signatures for vfooq(), and here's why each one didn't |
1890 | | // match your call". |
1891 | |
|
1892 | 0 | OS << "static __inline__ __attribute__((" |
1893 | 0 | << (Polymorphic ? "__overloadable__, " : "") |
1894 | 0 | << "__clang_arm_builtin_alias(__builtin_arm_mve_" << Int.fullName() |
1895 | 0 | << ")))\n" |
1896 | 0 | << RetTypeName << FunctionName << "(" << ArgTypesString << ");\n"; |
1897 | 0 | } |
1898 | 0 | } |
1899 | 0 | } |
1900 | 0 | for (auto &part : parts) |
1901 | 0 | part << "\n"; |
1902 | | |
1903 | | // Now we've finished accumulating bits and pieces into the parts[] array. |
1904 | | // Put it all together to write the final output file. |
1905 | |
|
1906 | 0 | OS << "/*===---- arm_mve.h - ARM MVE intrinsics " |
1907 | 0 | "-----------------------------------===\n" |
1908 | 0 | << LLVMLicenseHeader |
1909 | 0 | << "#ifndef __ARM_MVE_H\n" |
1910 | 0 | "#define __ARM_MVE_H\n" |
1911 | 0 | "\n" |
1912 | 0 | "#if !__ARM_FEATURE_MVE\n" |
1913 | 0 | "#error \"MVE support not enabled\"\n" |
1914 | 0 | "#endif\n" |
1915 | 0 | "\n" |
1916 | 0 | "#include <stdint.h>\n" |
1917 | 0 | "\n" |
1918 | 0 | "#ifdef __cplusplus\n" |
1919 | 0 | "extern \"C\" {\n" |
1920 | 0 | "#endif\n" |
1921 | 0 | "\n"; |
1922 | |
|
1923 | 0 | for (size_t i = 0; i < NumParts; ++i) { |
1924 | 0 | std::vector<std::string> conditions; |
1925 | 0 | if (i & Float) |
1926 | 0 | conditions.push_back("(__ARM_FEATURE_MVE & 2)"); |
1927 | 0 | if (i & UseUserNamespace) |
1928 | 0 | conditions.push_back("(!defined __ARM_MVE_PRESERVE_USER_NAMESPACE)"); |
1929 | |
|
1930 | 0 | std::string condition = |
1931 | 0 | join(std::begin(conditions), std::end(conditions), " && "); |
1932 | 0 | if (!condition.empty()) |
1933 | 0 | OS << "#if " << condition << "\n\n"; |
1934 | 0 | OS << parts[i].str(); |
1935 | 0 | if (!condition.empty()) |
1936 | 0 | OS << "#endif /* " << condition << " */\n\n"; |
1937 | 0 | } |
1938 | |
|
1939 | 0 | OS << "#ifdef __cplusplus\n" |
1940 | 0 | "} /* extern \"C\" */\n" |
1941 | 0 | "#endif\n" |
1942 | 0 | "\n" |
1943 | 0 | "#endif /* __ARM_MVE_H */\n"; |
1944 | 0 | } |
1945 | | |
1946 | 0 | void MveEmitter::EmitBuiltinDef(raw_ostream &OS) { |
1947 | 0 | for (const auto &kv : ACLEIntrinsics) { |
1948 | 0 | const ACLEIntrinsic &Int = *kv.second; |
1949 | 0 | OS << "BUILTIN(__builtin_arm_mve_" << Int.fullName() |
1950 | 0 | << ", \"\", \"n\")\n"; |
1951 | 0 | } |
1952 | |
|
1953 | 0 | std::set<std::string> ShortNamesSeen; |
1954 | |
|
1955 | 0 | for (const auto &kv : ACLEIntrinsics) { |
1956 | 0 | const ACLEIntrinsic &Int = *kv.second; |
1957 | 0 | if (Int.polymorphic()) { |
1958 | 0 | StringRef Name = Int.shortName(); |
1959 | 0 | if (ShortNamesSeen.find(std::string(Name)) == ShortNamesSeen.end()) { |
1960 | 0 | OS << "BUILTIN(__builtin_arm_mve_" << Name << ", \"vi.\", \"nt"; |
1961 | 0 | if (Int.nonEvaluating()) |
1962 | 0 | OS << "u"; // indicate that this builtin doesn't evaluate its args |
1963 | 0 | OS << "\")\n"; |
1964 | 0 | ShortNamesSeen.insert(std::string(Name)); |
1965 | 0 | } |
1966 | 0 | } |
1967 | 0 | } |
1968 | 0 | } |
1969 | | |
1970 | 0 | void MveEmitter::EmitBuiltinSema(raw_ostream &OS) { |
1971 | 0 | std::map<std::string, std::set<std::string>> Checks; |
1972 | 0 | GroupSemaChecks(Checks); |
1973 | |
|
1974 | 0 | for (const auto &kv : Checks) { |
1975 | 0 | for (StringRef Name : kv.second) |
1976 | 0 | OS << "case ARM::BI__builtin_arm_mve_" << Name << ":\n"; |
1977 | 0 | OS << " return " << kv.first; |
1978 | 0 | } |
1979 | 0 | } |
1980 | | |
1981 | | // ----------------------------------------------------------------------------- |
1982 | | // Class that describes an ACLE intrinsic implemented as a macro. |
1983 | | // |
1984 | | // This class is used when the intrinsic is polymorphic in 2 or 3 types, but we |
1985 | | // want to avoid a combinatorial explosion by reinterpreting the arguments to |
1986 | | // fixed types. |
1987 | | |
1988 | | class FunctionMacro { |
1989 | | std::vector<StringRef> Params; |
1990 | | StringRef Definition; |
1991 | | |
1992 | | public: |
1993 | | FunctionMacro(const Record &R); |
1994 | | |
1995 | 0 | const std::vector<StringRef> &getParams() const { return Params; } |
1996 | 0 | StringRef getDefinition() const { return Definition; } |
1997 | | }; |
1998 | | |
1999 | 0 | FunctionMacro::FunctionMacro(const Record &R) { |
2000 | 0 | Params = R.getValueAsListOfStrings("params"); |
2001 | 0 | Definition = R.getValueAsString("definition"); |
2002 | 0 | } |
2003 | | |
2004 | | // ----------------------------------------------------------------------------- |
2005 | | // The class used for generating arm_cde.h and related Clang bits |
2006 | | // |
2007 | | |
2008 | | class CdeEmitter : public EmitterBase { |
2009 | | std::map<StringRef, FunctionMacro> FunctionMacros; |
2010 | | |
2011 | | public: |
2012 | | CdeEmitter(RecordKeeper &Records); |
2013 | | void EmitHeader(raw_ostream &OS) override; |
2014 | | void EmitBuiltinDef(raw_ostream &OS) override; |
2015 | | void EmitBuiltinSema(raw_ostream &OS) override; |
2016 | | }; |
2017 | | |
2018 | 0 | CdeEmitter::CdeEmitter(RecordKeeper &Records) : EmitterBase(Records) { |
2019 | 0 | for (Record *R : Records.getAllDerivedDefinitions("FunctionMacro")) |
2020 | 0 | FunctionMacros.emplace(R->getName(), FunctionMacro(*R)); |
2021 | 0 | } |
2022 | | |
2023 | 0 | void CdeEmitter::EmitHeader(raw_ostream &OS) { |
2024 | | // Accumulate pieces of the header file that will be enabled under various |
2025 | | // different combinations of #ifdef. The index into parts[] is one of the |
2026 | | // following: |
2027 | 0 | constexpr unsigned None = 0; |
2028 | 0 | constexpr unsigned MVE = 1; |
2029 | 0 | constexpr unsigned MVEFloat = 2; |
2030 | |
|
2031 | 0 | constexpr unsigned NumParts = 3; |
2032 | 0 | raw_self_contained_string_ostream parts[NumParts]; |
2033 | | |
2034 | | // Write typedefs for all the required vector types, and a few scalar |
2035 | | // types that don't already have the name we want them to have. |
2036 | |
|
2037 | 0 | parts[MVE] << "typedef uint16_t mve_pred16_t;\n"; |
2038 | 0 | parts[MVEFloat] << "typedef __fp16 float16_t;\n" |
2039 | 0 | "typedef float float32_t;\n"; |
2040 | 0 | for (const auto &kv : ScalarTypes) { |
2041 | 0 | const ScalarType *ST = kv.second.get(); |
2042 | 0 | if (ST->hasNonstandardName()) |
2043 | 0 | continue; |
2044 | | // We don't have float64x2_t |
2045 | 0 | if (ST->kind() == ScalarTypeKind::Float && ST->sizeInBits() == 64) |
2046 | 0 | continue; |
2047 | 0 | raw_ostream &OS = parts[ST->requiresFloat() ? MVEFloat : MVE]; |
2048 | 0 | const VectorType *VT = getVectorType(ST); |
2049 | |
|
2050 | 0 | OS << "typedef __attribute__((__neon_vector_type__(" << VT->lanes() |
2051 | 0 | << "), __clang_arm_mve_strict_polymorphism)) " << ST->cName() << " " |
2052 | 0 | << VT->cName() << ";\n"; |
2053 | 0 | } |
2054 | 0 | parts[MVE] << "\n"; |
2055 | 0 | parts[MVEFloat] << "\n"; |
2056 | | |
2057 | | // Write declarations for all the intrinsics. |
2058 | |
|
2059 | 0 | for (const auto &kv : ACLEIntrinsics) { |
2060 | 0 | const ACLEIntrinsic &Int = *kv.second; |
2061 | | |
2062 | | // We generate each intrinsic twice, under its full unambiguous |
2063 | | // name and its shorter polymorphic name (if the latter exists). |
2064 | 0 | for (bool Polymorphic : {false, true}) { |
2065 | 0 | if (Polymorphic && !Int.polymorphic()) |
2066 | 0 | continue; |
2067 | 0 | if (!Polymorphic && Int.polymorphicOnly()) |
2068 | 0 | continue; |
2069 | | |
2070 | 0 | raw_ostream &OS = |
2071 | 0 | parts[Int.requiresFloat() ? MVEFloat |
2072 | 0 | : Int.requiresMVE() ? MVE : None]; |
2073 | | |
2074 | | // Make the name of the function in this declaration. |
2075 | 0 | std::string FunctionName = |
2076 | 0 | "__arm_" + (Polymorphic ? Int.shortName() : Int.fullName()); |
2077 | | |
2078 | | // Make strings for the types involved in the function's |
2079 | | // prototype. |
2080 | 0 | std::string RetTypeName = Int.returnType()->cName(); |
2081 | 0 | if (!StringRef(RetTypeName).endswith("*")) |
2082 | 0 | RetTypeName += " "; |
2083 | |
|
2084 | 0 | std::vector<std::string> ArgTypeNames; |
2085 | 0 | for (const Type *ArgTypePtr : Int.argTypes()) |
2086 | 0 | ArgTypeNames.push_back(ArgTypePtr->cName()); |
2087 | 0 | std::string ArgTypesString = |
2088 | 0 | join(std::begin(ArgTypeNames), std::end(ArgTypeNames), ", "); |
2089 | | |
2090 | | // Emit the actual declaration. See MveEmitter::EmitHeader for detailed |
2091 | | // comments |
2092 | 0 | OS << "static __inline__ __attribute__((" |
2093 | 0 | << (Polymorphic ? "__overloadable__, " : "") |
2094 | 0 | << "__clang_arm_builtin_alias(__builtin_arm_" << Int.builtinExtension() |
2095 | 0 | << "_" << Int.fullName() << ")))\n" |
2096 | 0 | << RetTypeName << FunctionName << "(" << ArgTypesString << ");\n"; |
2097 | 0 | } |
2098 | 0 | } |
2099 | |
|
2100 | 0 | for (const auto &kv : FunctionMacros) { |
2101 | 0 | StringRef Name = kv.first; |
2102 | 0 | const FunctionMacro &FM = kv.second; |
2103 | |
|
2104 | 0 | raw_ostream &OS = parts[MVE]; |
2105 | 0 | OS << "#define " |
2106 | 0 | << "__arm_" << Name << "(" << join(FM.getParams(), ", ") << ") " |
2107 | 0 | << FM.getDefinition() << "\n"; |
2108 | 0 | } |
2109 | |
|
2110 | 0 | for (auto &part : parts) |
2111 | 0 | part << "\n"; |
2112 | | |
2113 | | // Now we've finished accumulating bits and pieces into the parts[] array. |
2114 | | // Put it all together to write the final output file. |
2115 | |
|
2116 | 0 | OS << "/*===---- arm_cde.h - ARM CDE intrinsics " |
2117 | 0 | "-----------------------------------===\n" |
2118 | 0 | << LLVMLicenseHeader |
2119 | 0 | << "#ifndef __ARM_CDE_H\n" |
2120 | 0 | "#define __ARM_CDE_H\n" |
2121 | 0 | "\n" |
2122 | 0 | "#if !__ARM_FEATURE_CDE\n" |
2123 | 0 | "#error \"CDE support not enabled\"\n" |
2124 | 0 | "#endif\n" |
2125 | 0 | "\n" |
2126 | 0 | "#include <stdint.h>\n" |
2127 | 0 | "\n" |
2128 | 0 | "#ifdef __cplusplus\n" |
2129 | 0 | "extern \"C\" {\n" |
2130 | 0 | "#endif\n" |
2131 | 0 | "\n"; |
2132 | |
|
2133 | 0 | for (size_t i = 0; i < NumParts; ++i) { |
2134 | 0 | std::string condition; |
2135 | 0 | if (i == MVEFloat) |
2136 | 0 | condition = "__ARM_FEATURE_MVE & 2"; |
2137 | 0 | else if (i == MVE) |
2138 | 0 | condition = "__ARM_FEATURE_MVE"; |
2139 | |
|
2140 | 0 | if (!condition.empty()) |
2141 | 0 | OS << "#if " << condition << "\n\n"; |
2142 | 0 | OS << parts[i].str(); |
2143 | 0 | if (!condition.empty()) |
2144 | 0 | OS << "#endif /* " << condition << " */\n\n"; |
2145 | 0 | } |
2146 | |
|
2147 | 0 | OS << "#ifdef __cplusplus\n" |
2148 | 0 | "} /* extern \"C\" */\n" |
2149 | 0 | "#endif\n" |
2150 | 0 | "\n" |
2151 | 0 | "#endif /* __ARM_CDE_H */\n"; |
2152 | 0 | } |
2153 | | |
2154 | 0 | void CdeEmitter::EmitBuiltinDef(raw_ostream &OS) { |
2155 | 0 | for (const auto &kv : ACLEIntrinsics) { |
2156 | 0 | if (kv.second->headerOnly()) |
2157 | 0 | continue; |
2158 | 0 | const ACLEIntrinsic &Int = *kv.second; |
2159 | 0 | OS << "BUILTIN(__builtin_arm_cde_" << Int.fullName() |
2160 | 0 | << ", \"\", \"ncU\")\n"; |
2161 | 0 | } |
2162 | 0 | } |
2163 | | |
2164 | 0 | void CdeEmitter::EmitBuiltinSema(raw_ostream &OS) { |
2165 | 0 | std::map<std::string, std::set<std::string>> Checks; |
2166 | 0 | GroupSemaChecks(Checks); |
2167 | |
|
2168 | 0 | for (const auto &kv : Checks) { |
2169 | 0 | for (StringRef Name : kv.second) |
2170 | 0 | OS << "case ARM::BI__builtin_arm_cde_" << Name << ":\n"; |
2171 | 0 | OS << " Err = " << kv.first << " break;\n"; |
2172 | 0 | } |
2173 | 0 | } |
2174 | | |
2175 | | } // namespace |
2176 | | |
2177 | | namespace clang { |
2178 | | |
2179 | | // MVE |
2180 | | |
2181 | 0 | void EmitMveHeader(RecordKeeper &Records, raw_ostream &OS) { |
2182 | 0 | MveEmitter(Records).EmitHeader(OS); |
2183 | 0 | } |
2184 | | |
2185 | 0 | void EmitMveBuiltinDef(RecordKeeper &Records, raw_ostream &OS) { |
2186 | 0 | MveEmitter(Records).EmitBuiltinDef(OS); |
2187 | 0 | } |
2188 | | |
2189 | 0 | void EmitMveBuiltinSema(RecordKeeper &Records, raw_ostream &OS) { |
2190 | 0 | MveEmitter(Records).EmitBuiltinSema(OS); |
2191 | 0 | } |
2192 | | |
2193 | 0 | void EmitMveBuiltinCG(RecordKeeper &Records, raw_ostream &OS) { |
2194 | 0 | MveEmitter(Records).EmitBuiltinCG(OS); |
2195 | 0 | } |
2196 | | |
2197 | 0 | void EmitMveBuiltinAliases(RecordKeeper &Records, raw_ostream &OS) { |
2198 | 0 | MveEmitter(Records).EmitBuiltinAliases(OS); |
2199 | 0 | } |
2200 | | |
2201 | | // CDE |
2202 | | |
2203 | 0 | void EmitCdeHeader(RecordKeeper &Records, raw_ostream &OS) { |
2204 | 0 | CdeEmitter(Records).EmitHeader(OS); |
2205 | 0 | } |
2206 | | |
2207 | 0 | void EmitCdeBuiltinDef(RecordKeeper &Records, raw_ostream &OS) { |
2208 | 0 | CdeEmitter(Records).EmitBuiltinDef(OS); |
2209 | 0 | } |
2210 | | |
2211 | 0 | void EmitCdeBuiltinSema(RecordKeeper &Records, raw_ostream &OS) { |
2212 | 0 | CdeEmitter(Records).EmitBuiltinSema(OS); |
2213 | 0 | } |
2214 | | |
2215 | 0 | void EmitCdeBuiltinCG(RecordKeeper &Records, raw_ostream &OS) { |
2216 | 0 | CdeEmitter(Records).EmitBuiltinCG(OS); |
2217 | 0 | } |
2218 | | |
2219 | 0 | void EmitCdeBuiltinAliases(RecordKeeper &Records, raw_ostream &OS) { |
2220 | 0 | CdeEmitter(Records).EmitBuiltinAliases(OS); |
2221 | 0 | } |
2222 | | |
2223 | | } // end namespace clang |