1 | //===- FuncToLLVM.cpp - Func to LLVM dialect conversion -------------------===// |
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 file implements a pass to convert MLIR Func and builtin dialects |
10 | // into the LLVM IR dialect. |
11 | // |
12 | //===----------------------------------------------------------------------===// |
13 | |
14 | #include "mlir/Conversion/FuncToLLVM/ConvertFuncToLLVMPass.h" |
15 | |
16 | #include "mlir/Analysis/DataLayoutAnalysis.h" |
17 | #include "mlir/Conversion/ArithToLLVM/ArithToLLVM.h" |
18 | #include "mlir/Conversion/ControlFlowToLLVM/ControlFlowToLLVM.h" |
19 | #include "mlir/Conversion/ConvertToLLVM/ToLLVMInterface.h" |
20 | #include "mlir/Conversion/FuncToLLVM/ConvertFuncToLLVM.h" |
21 | #include "mlir/Conversion/LLVMCommon/ConversionTarget.h" |
22 | #include "mlir/Conversion/LLVMCommon/Pattern.h" |
23 | #include "mlir/Conversion/LLVMCommon/VectorPattern.h" |
24 | #include "mlir/Dialect/Func/IR/FuncOps.h" |
25 | #include "mlir/Dialect/LLVMIR/FunctionCallUtils.h" |
26 | #include "mlir/Dialect/LLVMIR/LLVMDialect.h" |
27 | #include "mlir/Dialect/LLVMIR/LLVMTypes.h" |
28 | #include "mlir/Dialect/Utils/StaticValueUtils.h" |
29 | #include "mlir/IR/Attributes.h" |
30 | #include "mlir/IR/Builders.h" |
31 | #include "mlir/IR/BuiltinAttributeInterfaces.h" |
32 | #include "mlir/IR/BuiltinAttributes.h" |
33 | #include "mlir/IR/BuiltinOps.h" |
34 | #include "mlir/IR/IRMapping.h" |
35 | #include "mlir/IR/PatternMatch.h" |
36 | #include "mlir/IR/SymbolTable.h" |
37 | #include "mlir/IR/TypeUtilities.h" |
38 | #include "mlir/Support/LogicalResult.h" |
39 | #include "mlir/Support/MathExtras.h" |
40 | #include "mlir/Transforms/DialectConversion.h" |
41 | #include "mlir/Transforms/Passes.h" |
42 | #include "llvm/ADT/SmallVector.h" |
43 | #include "llvm/ADT/TypeSwitch.h" |
44 | #include "llvm/IR/DerivedTypes.h" |
45 | #include "llvm/IR/IRBuilder.h" |
46 | #include "llvm/IR/Type.h" |
47 | #include "llvm/Support/Casting.h" |
48 | #include "llvm/Support/CommandLine.h" |
49 | #include "llvm/Support/FormatVariadic.h" |
50 | #include <algorithm> |
51 | #include <functional> |
52 | #include <optional> |
53 | |
54 | namespace mlir { |
55 | #define GEN_PASS_DEF_CONVERTFUNCTOLLVMPASS |
56 | #define GEN_PASS_DEF_SETLLVMMODULEDATALAYOUTPASS |
57 | #include "mlir/Conversion/Passes.h.inc" |
58 | } // namespace mlir |
59 | |
60 | using namespace mlir; |
61 | |
62 | #define PASS_NAME "convert-func-to-llvm" |
63 | |
64 | static constexpr StringRef varargsAttrName = "func.varargs" ; |
65 | static constexpr StringRef linkageAttrName = "llvm.linkage" ; |
66 | static constexpr StringRef barePtrAttrName = "llvm.bareptr" ; |
67 | |
68 | /// Return `true` if the `op` should use bare pointer calling convention. |
69 | static bool shouldUseBarePtrCallConv(Operation *op, |
70 | const LLVMTypeConverter *typeConverter) { |
71 | return (op && op->hasAttr(name: barePtrAttrName)) || |
72 | typeConverter->getOptions().useBarePtrCallConv; |
73 | } |
74 | |
75 | /// Only retain those attributes that are not constructed by |
76 | /// `LLVMFuncOp::build`. |
77 | static void filterFuncAttributes(FunctionOpInterface func, |
78 | SmallVectorImpl<NamedAttribute> &result) { |
79 | for (const NamedAttribute &attr : func->getDiscardableAttrs()) { |
80 | if (attr.getName() == linkageAttrName || |
81 | attr.getName() == varargsAttrName || |
82 | attr.getName() == LLVM::LLVMDialect::getReadnoneAttrName()) |
83 | continue; |
84 | result.push_back(attr); |
85 | } |
86 | } |
87 | |
88 | /// Propagate argument/results attributes. |
89 | static void propagateArgResAttrs(OpBuilder &builder, bool resultStructType, |
90 | FunctionOpInterface funcOp, |
91 | LLVM::LLVMFuncOp wrapperFuncOp) { |
92 | auto argAttrs = funcOp.getAllArgAttrs(); |
93 | if (!resultStructType) { |
94 | if (auto resAttrs = funcOp.getAllResultAttrs()) |
95 | wrapperFuncOp.setAllResultAttrs(resAttrs); |
96 | if (argAttrs) |
97 | wrapperFuncOp.setAllArgAttrs(argAttrs); |
98 | } else { |
99 | SmallVector<Attribute> argAttributes; |
100 | // Only modify the argument and result attributes when the result is now |
101 | // an argument. |
102 | if (argAttrs) { |
103 | argAttributes.push_back(builder.getDictionaryAttr({})); |
104 | argAttributes.append(argAttrs.begin(), argAttrs.end()); |
105 | wrapperFuncOp.setAllArgAttrs(argAttributes); |
106 | } |
107 | } |
108 | cast<FunctionOpInterface>(wrapperFuncOp.getOperation()) |
109 | .setVisibility(funcOp.getVisibility()); |
110 | } |
111 | |
112 | /// Creates an auxiliary function with pointer-to-memref-descriptor-struct |
113 | /// arguments instead of unpacked arguments. This function can be called from C |
114 | /// by passing a pointer to a C struct corresponding to a memref descriptor. |
115 | /// Similarly, returned memrefs are passed via pointers to a C struct that is |
116 | /// passed as additional argument. |
117 | /// Internally, the auxiliary function unpacks the descriptor into individual |
118 | /// components and forwards them to `newFuncOp` and forwards the results to |
119 | /// the extra arguments. |
120 | static void wrapForExternalCallers(OpBuilder &rewriter, Location loc, |
121 | const LLVMTypeConverter &typeConverter, |
122 | FunctionOpInterface funcOp, |
123 | LLVM::LLVMFuncOp newFuncOp) { |
124 | auto type = cast<FunctionType>(funcOp.getFunctionType()); |
125 | auto [wrapperFuncType, resultStructType] = |
126 | typeConverter.convertFunctionTypeCWrapper(type: type); |
127 | |
128 | SmallVector<NamedAttribute> attributes; |
129 | filterFuncAttributes(funcOp, attributes); |
130 | |
131 | auto wrapperFuncOp = rewriter.create<LLVM::LLVMFuncOp>( |
132 | loc, llvm::formatv("_mlir_ciface_{0}" , funcOp.getName()).str(), |
133 | wrapperFuncType, LLVM::Linkage::External, /*dsoLocal=*/false, |
134 | /*cconv=*/LLVM::CConv::C, /*comdat=*/nullptr, attributes); |
135 | propagateArgResAttrs(rewriter, !!resultStructType, funcOp, wrapperFuncOp); |
136 | |
137 | OpBuilder::InsertionGuard guard(rewriter); |
138 | rewriter.setInsertionPointToStart(wrapperFuncOp.addEntryBlock(rewriter)); |
139 | |
140 | SmallVector<Value, 8> args; |
141 | size_t argOffset = resultStructType ? 1 : 0; |
142 | for (auto [index, argType] : llvm::enumerate(type.getInputs())) { |
143 | Value arg = wrapperFuncOp.getArgument(index + argOffset); |
144 | if (auto memrefType = dyn_cast<MemRefType>(argType)) { |
145 | Value loaded = rewriter.create<LLVM::LoadOp>( |
146 | loc, typeConverter.convertType(memrefType), arg); |
147 | MemRefDescriptor::unpack(rewriter, loc, loaded, memrefType, args); |
148 | continue; |
149 | } |
150 | if (isa<UnrankedMemRefType>(argType)) { |
151 | Value loaded = rewriter.create<LLVM::LoadOp>( |
152 | loc, typeConverter.convertType(argType), arg); |
153 | UnrankedMemRefDescriptor::unpack(rewriter, loc, loaded, args); |
154 | continue; |
155 | } |
156 | |
157 | args.push_back(arg); |
158 | } |
159 | |
160 | auto call = rewriter.create<LLVM::CallOp>(loc, newFuncOp, args); |
161 | |
162 | if (resultStructType) { |
163 | rewriter.create<LLVM::StoreOp>(loc, call.getResult(), |
164 | wrapperFuncOp.getArgument(0)); |
165 | rewriter.create<LLVM::ReturnOp>(loc, ValueRange{}); |
166 | } else { |
167 | rewriter.create<LLVM::ReturnOp>(loc, call.getResults()); |
168 | } |
169 | } |
170 | |
171 | /// Creates an auxiliary function with pointer-to-memref-descriptor-struct |
172 | /// arguments instead of unpacked arguments. Creates a body for the (external) |
173 | /// `newFuncOp` that allocates a memref descriptor on stack, packs the |
174 | /// individual arguments into this descriptor and passes a pointer to it into |
175 | /// the auxiliary function. If the result of the function cannot be directly |
176 | /// returned, we write it to a special first argument that provides a pointer |
177 | /// to a corresponding struct. This auxiliary external function is now |
178 | /// compatible with functions defined in C using pointers to C structs |
179 | /// corresponding to a memref descriptor. |
180 | static void wrapExternalFunction(OpBuilder &builder, Location loc, |
181 | const LLVMTypeConverter &typeConverter, |
182 | FunctionOpInterface funcOp, |
183 | LLVM::LLVMFuncOp newFuncOp) { |
184 | OpBuilder::InsertionGuard guard(builder); |
185 | |
186 | auto [wrapperType, resultStructType] = |
187 | typeConverter.convertFunctionTypeCWrapper( |
188 | type: cast<FunctionType>(funcOp.getFunctionType())); |
189 | // This conversion can only fail if it could not convert one of the argument |
190 | // types. But since it has been applied to a non-wrapper function before, it |
191 | // should have failed earlier and not reach this point at all. |
192 | assert(wrapperType && "unexpected type conversion failure" ); |
193 | |
194 | SmallVector<NamedAttribute, 4> attributes; |
195 | filterFuncAttributes(funcOp, attributes); |
196 | |
197 | // Create the auxiliary function. |
198 | auto wrapperFunc = builder.create<LLVM::LLVMFuncOp>( |
199 | loc, llvm::formatv("_mlir_ciface_{0}" , funcOp.getName()).str(), |
200 | wrapperType, LLVM::Linkage::External, /*dsoLocal=*/false, |
201 | /*cconv=*/LLVM::CConv::C, /*comdat=*/nullptr, attributes); |
202 | propagateArgResAttrs(builder, !!resultStructType, funcOp, wrapperFunc); |
203 | |
204 | // The wrapper that we synthetize here should only be visible in this module. |
205 | newFuncOp.setLinkage(LLVM::Linkage::Private); |
206 | builder.setInsertionPointToStart(newFuncOp.addEntryBlock(builder)); |
207 | |
208 | // Get a ValueRange containing arguments. |
209 | FunctionType type = cast<FunctionType>(funcOp.getFunctionType()); |
210 | SmallVector<Value, 8> args; |
211 | args.reserve(N: type.getNumInputs()); |
212 | ValueRange wrapperArgsRange(newFuncOp.getArguments()); |
213 | |
214 | if (resultStructType) { |
215 | // Allocate the struct on the stack and pass the pointer. |
216 | Type resultType = cast<LLVM::LLVMFunctionType>(wrapperType).getParamType(0); |
217 | Value one = builder.create<LLVM::ConstantOp>( |
218 | loc, typeConverter.convertType(builder.getIndexType()), |
219 | builder.getIntegerAttr(builder.getIndexType(), 1)); |
220 | Value result = |
221 | builder.create<LLVM::AllocaOp>(loc, resultType, resultStructType, one); |
222 | args.push_back(Elt: result); |
223 | } |
224 | |
225 | // Iterate over the inputs of the original function and pack values into |
226 | // memref descriptors if the original type is a memref. |
227 | for (Type input : type.getInputs()) { |
228 | Value arg; |
229 | int numToDrop = 1; |
230 | auto memRefType = dyn_cast<MemRefType>(input); |
231 | auto unrankedMemRefType = dyn_cast<UnrankedMemRefType>(input); |
232 | if (memRefType || unrankedMemRefType) { |
233 | numToDrop = memRefType |
234 | ? MemRefDescriptor::getNumUnpackedValues(memRefType) |
235 | : UnrankedMemRefDescriptor::getNumUnpackedValues(); |
236 | Value packed = |
237 | memRefType |
238 | ? MemRefDescriptor::pack(builder, loc, typeConverter, memRefType, |
239 | wrapperArgsRange.take_front(numToDrop)) |
240 | : UnrankedMemRefDescriptor::pack( |
241 | builder, loc, typeConverter, unrankedMemRefType, |
242 | wrapperArgsRange.take_front(numToDrop)); |
243 | |
244 | auto ptrTy = LLVM::LLVMPointerType::get(builder.getContext()); |
245 | Value one = builder.create<LLVM::ConstantOp>( |
246 | loc, typeConverter.convertType(builder.getIndexType()), |
247 | builder.getIntegerAttr(builder.getIndexType(), 1)); |
248 | Value allocated = builder.create<LLVM::AllocaOp>( |
249 | loc, ptrTy, packed.getType(), one, /*alignment=*/0); |
250 | builder.create<LLVM::StoreOp>(loc, packed, allocated); |
251 | arg = allocated; |
252 | } else { |
253 | arg = wrapperArgsRange[0]; |
254 | } |
255 | |
256 | args.push_back(arg); |
257 | wrapperArgsRange = wrapperArgsRange.drop_front(numToDrop); |
258 | } |
259 | assert(wrapperArgsRange.empty() && "did not map some of the arguments" ); |
260 | |
261 | auto call = builder.create<LLVM::CallOp>(loc, wrapperFunc, args); |
262 | |
263 | if (resultStructType) { |
264 | Value result = |
265 | builder.create<LLVM::LoadOp>(loc, resultStructType, args.front()); |
266 | builder.create<LLVM::ReturnOp>(loc, result); |
267 | } else { |
268 | builder.create<LLVM::ReturnOp>(loc, call.getResults()); |
269 | } |
270 | } |
271 | |
272 | /// Modifies the body of the function to construct the `MemRefDescriptor` from |
273 | /// the bare pointer calling convention lowering of `memref` types. |
274 | static void modifyFuncOpToUseBarePtrCallingConv( |
275 | ConversionPatternRewriter &rewriter, Location loc, |
276 | const LLVMTypeConverter &typeConverter, LLVM::LLVMFuncOp funcOp, |
277 | TypeRange oldArgTypes) { |
278 | if (funcOp.getBody().empty()) |
279 | return; |
280 | |
281 | // Promote bare pointers from memref arguments to memref descriptors at the |
282 | // beginning of the function so that all the memrefs in the function have a |
283 | // uniform representation. |
284 | Block *entryBlock = &funcOp.getBody().front(); |
285 | auto blockArgs = entryBlock->getArguments(); |
286 | assert(blockArgs.size() == oldArgTypes.size() && |
287 | "The number of arguments and types doesn't match" ); |
288 | |
289 | OpBuilder::InsertionGuard guard(rewriter); |
290 | rewriter.setInsertionPointToStart(entryBlock); |
291 | for (auto it : llvm::zip(blockArgs, oldArgTypes)) { |
292 | BlockArgument arg = std::get<0>(it); |
293 | Type argTy = std::get<1>(it); |
294 | |
295 | // Unranked memrefs are not supported in the bare pointer calling |
296 | // convention. We should have bailed out before in the presence of |
297 | // unranked memrefs. |
298 | assert(!isa<UnrankedMemRefType>(argTy) && |
299 | "Unranked memref is not supported" ); |
300 | auto memrefTy = dyn_cast<MemRefType>(argTy); |
301 | if (!memrefTy) |
302 | continue; |
303 | |
304 | // Replace barePtr with a placeholder (undef), promote barePtr to a ranked |
305 | // or unranked memref descriptor and replace placeholder with the last |
306 | // instruction of the memref descriptor. |
307 | // TODO: The placeholder is needed to avoid replacing barePtr uses in the |
308 | // MemRef descriptor instructions. We may want to have a utility in the |
309 | // rewriter to properly handle this use case. |
310 | Location loc = funcOp.getLoc(); |
311 | auto placeholder = rewriter.create<LLVM::UndefOp>( |
312 | loc, typeConverter.convertType(memrefTy)); |
313 | rewriter.replaceUsesOfBlockArgument(arg, placeholder); |
314 | |
315 | Value desc = MemRefDescriptor::fromStaticShape(rewriter, loc, typeConverter, |
316 | memrefTy, arg); |
317 | rewriter.replaceOp(placeholder, {desc}); |
318 | } |
319 | } |
320 | |
321 | FailureOr<LLVM::LLVMFuncOp> |
322 | mlir::convertFuncOpToLLVMFuncOp(FunctionOpInterface funcOp, |
323 | ConversionPatternRewriter &rewriter, |
324 | const LLVMTypeConverter &converter) { |
325 | // Check the funcOp has `FunctionType`. |
326 | auto funcTy = dyn_cast<FunctionType>(funcOp.getFunctionType()); |
327 | if (!funcTy) |
328 | return rewriter.notifyMatchFailure( |
329 | funcOp, "Only support FunctionOpInterface with FunctionType" ); |
330 | |
331 | // Convert the original function arguments. They are converted using the |
332 | // LLVMTypeConverter provided to this legalization pattern. |
333 | auto varargsAttr = funcOp->getAttrOfType<BoolAttr>(varargsAttrName); |
334 | TypeConverter::SignatureConversion result(funcOp.getNumArguments()); |
335 | auto llvmType = converter.convertFunctionSignature( |
336 | funcTy: funcTy, isVariadic: varargsAttr && varargsAttr.getValue(), |
337 | useBarePtrCallConv: shouldUseBarePtrCallConv(funcOp, &converter), result); |
338 | if (!llvmType) |
339 | return rewriter.notifyMatchFailure(funcOp, "signature conversion failed" ); |
340 | |
341 | // Create an LLVM function, use external linkage by default until MLIR |
342 | // functions have linkage. |
343 | LLVM::Linkage linkage = LLVM::Linkage::External; |
344 | if (funcOp->hasAttr(linkageAttrName)) { |
345 | auto attr = |
346 | dyn_cast<mlir::LLVM::LinkageAttr>(funcOp->getAttr(linkageAttrName)); |
347 | if (!attr) { |
348 | funcOp->emitError() << "Contains " << linkageAttrName |
349 | << " attribute not of type LLVM::LinkageAttr" ; |
350 | return rewriter.notifyMatchFailure( |
351 | funcOp, "Contains linkage attribute not of type LLVM::LinkageAttr" ); |
352 | } |
353 | linkage = attr.getLinkage(); |
354 | } |
355 | |
356 | SmallVector<NamedAttribute, 4> attributes; |
357 | filterFuncAttributes(funcOp, attributes); |
358 | auto newFuncOp = rewriter.create<LLVM::LLVMFuncOp>( |
359 | funcOp.getLoc(), funcOp.getName(), llvmType, linkage, |
360 | /*dsoLocal=*/false, /*cconv=*/LLVM::CConv::C, /*comdat=*/nullptr, |
361 | attributes); |
362 | cast<FunctionOpInterface>(newFuncOp.getOperation()) |
363 | .setVisibility(funcOp.getVisibility()); |
364 | |
365 | // Create a memory effect attribute corresponding to readnone. |
366 | StringRef readnoneAttrName = LLVM::LLVMDialect::getReadnoneAttrName(); |
367 | if (funcOp->hasAttr(readnoneAttrName)) { |
368 | auto attr = funcOp->getAttrOfType<UnitAttr>(readnoneAttrName); |
369 | if (!attr) { |
370 | funcOp->emitError() << "Contains " << readnoneAttrName |
371 | << " attribute not of type UnitAttr" ; |
372 | return rewriter.notifyMatchFailure( |
373 | funcOp, "Contains readnone attribute not of type UnitAttr" ); |
374 | } |
375 | auto memoryAttr = LLVM::MemoryEffectsAttr::get( |
376 | rewriter.getContext(), |
377 | {LLVM::ModRefInfo::NoModRef, LLVM::ModRefInfo::NoModRef, |
378 | LLVM::ModRefInfo::NoModRef}); |
379 | newFuncOp.setMemoryAttr(memoryAttr); |
380 | } |
381 | |
382 | // Propagate argument/result attributes to all converted arguments/result |
383 | // obtained after converting a given original argument/result. |
384 | if (ArrayAttr resAttrDicts = funcOp.getAllResultAttrs()) { |
385 | assert(!resAttrDicts.empty() && "expected array to be non-empty" ); |
386 | if (funcOp.getNumResults() == 1) |
387 | newFuncOp.setAllResultAttrs(resAttrDicts); |
388 | } |
389 | if (ArrayAttr argAttrDicts = funcOp.getAllArgAttrs()) { |
390 | SmallVector<Attribute> newArgAttrs( |
391 | cast<LLVM::LLVMFunctionType>(llvmType).getNumParams()); |
392 | for (unsigned i = 0, e = funcOp.getNumArguments(); i < e; ++i) { |
393 | // Some LLVM IR attribute have a type attached to them. During FuncOp -> |
394 | // LLVMFuncOp conversion these types may have changed. Account for that |
395 | // change by converting attributes' types as well. |
396 | SmallVector<NamedAttribute, 4> convertedAttrs; |
397 | auto attrsDict = cast<DictionaryAttr>(argAttrDicts[i]); |
398 | convertedAttrs.reserve(N: attrsDict.size()); |
399 | for (const NamedAttribute &attr : attrsDict) { |
400 | const auto convert = [&](const NamedAttribute &attr) { |
401 | return TypeAttr::get(converter.convertType( |
402 | cast<TypeAttr>(attr.getValue()).getValue())); |
403 | }; |
404 | if (attr.getName().getValue() == |
405 | LLVM::LLVMDialect::getByValAttrName()) { |
406 | convertedAttrs.push_back(rewriter.getNamedAttr( |
407 | LLVM::LLVMDialect::getByValAttrName(), convert(attr))); |
408 | } else if (attr.getName().getValue() == |
409 | LLVM::LLVMDialect::getByRefAttrName()) { |
410 | convertedAttrs.push_back(rewriter.getNamedAttr( |
411 | LLVM::LLVMDialect::getByRefAttrName(), convert(attr))); |
412 | } else if (attr.getName().getValue() == |
413 | LLVM::LLVMDialect::getStructRetAttrName()) { |
414 | convertedAttrs.push_back(rewriter.getNamedAttr( |
415 | LLVM::LLVMDialect::getStructRetAttrName(), convert(attr))); |
416 | } else if (attr.getName().getValue() == |
417 | LLVM::LLVMDialect::getInAllocaAttrName()) { |
418 | convertedAttrs.push_back(rewriter.getNamedAttr( |
419 | LLVM::LLVMDialect::getInAllocaAttrName(), convert(attr))); |
420 | } else { |
421 | convertedAttrs.push_back(attr); |
422 | } |
423 | } |
424 | auto mapping = result.getInputMapping(input: i); |
425 | assert(mapping && "unexpected deletion of function argument" ); |
426 | // Only attach the new argument attributes if there is a one-to-one |
427 | // mapping from old to new types. Otherwise, attributes might be |
428 | // attached to types that they do not support. |
429 | if (mapping->size == 1) { |
430 | newArgAttrs[mapping->inputNo] = |
431 | DictionaryAttr::get(rewriter.getContext(), convertedAttrs); |
432 | continue; |
433 | } |
434 | // TODO: Implement custom handling for types that expand to multiple |
435 | // function arguments. |
436 | for (size_t j = 0; j < mapping->size; ++j) |
437 | newArgAttrs[mapping->inputNo + j] = |
438 | DictionaryAttr::get(rewriter.getContext(), {}); |
439 | } |
440 | if (!newArgAttrs.empty()) |
441 | newFuncOp.setAllArgAttrs(rewriter.getArrayAttr(newArgAttrs)); |
442 | } |
443 | |
444 | rewriter.inlineRegionBefore(funcOp.getFunctionBody(), newFuncOp.getBody(), |
445 | newFuncOp.end()); |
446 | if (failed(rewriter.convertRegionTypes(region: &newFuncOp.getBody(), converter, |
447 | entryConversion: &result))) { |
448 | return rewriter.notifyMatchFailure(funcOp, |
449 | "region types conversion failed" ); |
450 | } |
451 | |
452 | return newFuncOp; |
453 | } |
454 | |
455 | namespace { |
456 | |
457 | struct FuncOpConversionBase : public ConvertOpToLLVMPattern<func::FuncOp> { |
458 | protected: |
459 | using ConvertOpToLLVMPattern<func::FuncOp>::ConvertOpToLLVMPattern; |
460 | |
461 | // Convert input FuncOp to LLVMFuncOp by using the LLVMTypeConverter provided |
462 | // to this legalization pattern. |
463 | FailureOr<LLVM::LLVMFuncOp> |
464 | convertFuncOpToLLVMFuncOp(func::FuncOp funcOp, |
465 | ConversionPatternRewriter &rewriter) const { |
466 | return mlir::convertFuncOpToLLVMFuncOp( |
467 | cast<FunctionOpInterface>(funcOp.getOperation()), rewriter, |
468 | *getTypeConverter()); |
469 | } |
470 | }; |
471 | |
472 | /// FuncOp legalization pattern that converts MemRef arguments to pointers to |
473 | /// MemRef descriptors (LLVM struct data types) containing all the MemRef type |
474 | /// information. |
475 | struct FuncOpConversion : public FuncOpConversionBase { |
476 | FuncOpConversion(const LLVMTypeConverter &converter) |
477 | : FuncOpConversionBase(converter) {} |
478 | |
479 | LogicalResult |
480 | matchAndRewrite(func::FuncOp funcOp, OpAdaptor adaptor, |
481 | ConversionPatternRewriter &rewriter) const override { |
482 | FailureOr<LLVM::LLVMFuncOp> newFuncOp = |
483 | convertFuncOpToLLVMFuncOp(funcOp, rewriter); |
484 | if (failed(result: newFuncOp)) |
485 | return rewriter.notifyMatchFailure(funcOp, "Could not convert funcop" ); |
486 | |
487 | if (!shouldUseBarePtrCallConv(funcOp, this->getTypeConverter())) { |
488 | if (funcOp->getAttrOfType<UnitAttr>( |
489 | LLVM::LLVMDialect::getEmitCWrapperAttrName())) { |
490 | if (newFuncOp->isVarArg()) |
491 | return funcOp->emitError("C interface for variadic functions is not " |
492 | "supported yet." ); |
493 | |
494 | if (newFuncOp->isExternal()) |
495 | wrapExternalFunction(rewriter, funcOp->getLoc(), *getTypeConverter(), |
496 | funcOp, *newFuncOp); |
497 | else |
498 | wrapForExternalCallers(rewriter, funcOp->getLoc(), |
499 | *getTypeConverter(), funcOp, *newFuncOp); |
500 | } |
501 | } else { |
502 | modifyFuncOpToUseBarePtrCallingConv(rewriter, funcOp->getLoc(), |
503 | *getTypeConverter(), *newFuncOp, |
504 | funcOp.getFunctionType().getInputs()); |
505 | } |
506 | |
507 | rewriter.eraseOp(op: funcOp); |
508 | return success(); |
509 | } |
510 | }; |
511 | |
512 | struct ConstantOpLowering : public ConvertOpToLLVMPattern<func::ConstantOp> { |
513 | using ConvertOpToLLVMPattern<func::ConstantOp>::ConvertOpToLLVMPattern; |
514 | |
515 | LogicalResult |
516 | matchAndRewrite(func::ConstantOp op, OpAdaptor adaptor, |
517 | ConversionPatternRewriter &rewriter) const override { |
518 | auto type = typeConverter->convertType(op.getResult().getType()); |
519 | if (!type || !LLVM::isCompatibleType(type: type)) |
520 | return rewriter.notifyMatchFailure(op, "failed to convert result type" ); |
521 | |
522 | auto newOp = |
523 | rewriter.create<LLVM::AddressOfOp>(op.getLoc(), type, op.getValue()); |
524 | for (const NamedAttribute &attr : op->getAttrs()) { |
525 | if (attr.getName().strref() == "value" ) |
526 | continue; |
527 | newOp->setAttr(attr.getName(), attr.getValue()); |
528 | } |
529 | rewriter.replaceOp(op, newOp->getResults()); |
530 | return success(); |
531 | } |
532 | }; |
533 | |
534 | // A CallOp automatically promotes MemRefType to a sequence of alloca/store and |
535 | // passes the pointer to the MemRef across function boundaries. |
536 | template <typename CallOpType> |
537 | struct CallOpInterfaceLowering : public ConvertOpToLLVMPattern<CallOpType> { |
538 | using ConvertOpToLLVMPattern<CallOpType>::ConvertOpToLLVMPattern; |
539 | using Super = CallOpInterfaceLowering<CallOpType>; |
540 | using Base = ConvertOpToLLVMPattern<CallOpType>; |
541 | |
542 | LogicalResult matchAndRewriteImpl(CallOpType callOp, |
543 | typename CallOpType::Adaptor adaptor, |
544 | ConversionPatternRewriter &rewriter, |
545 | bool useBarePtrCallConv = false) const { |
546 | // Pack the result types into a struct. |
547 | Type packedResult = nullptr; |
548 | unsigned numResults = callOp.getNumResults(); |
549 | auto resultTypes = llvm::to_vector<4>(callOp.getResultTypes()); |
550 | |
551 | if (numResults != 0) { |
552 | if (!(packedResult = this->getTypeConverter()->packFunctionResults( |
553 | resultTypes, useBarePtrCallConv))) |
554 | return failure(); |
555 | } |
556 | |
557 | if (useBarePtrCallConv) { |
558 | for (auto it : callOp->getOperands()) { |
559 | Type operandType = it.getType(); |
560 | if (isa<UnrankedMemRefType>(Val: operandType)) { |
561 | // Unranked memref is not supported in the bare pointer calling |
562 | // convention. |
563 | return failure(); |
564 | } |
565 | } |
566 | } |
567 | auto promoted = this->getTypeConverter()->promoteOperands( |
568 | callOp.getLoc(), /*opOperands=*/callOp->getOperands(), |
569 | adaptor.getOperands(), rewriter, useBarePtrCallConv); |
570 | auto newOp = rewriter.create<LLVM::CallOp>( |
571 | callOp.getLoc(), packedResult ? TypeRange(packedResult) : TypeRange(), |
572 | promoted, callOp->getAttrs()); |
573 | |
574 | SmallVector<Value, 4> results; |
575 | if (numResults < 2) { |
576 | // If < 2 results, packing did not do anything and we can just return. |
577 | results.append(newOp.result_begin(), newOp.result_end()); |
578 | } else { |
579 | // Otherwise, it had been converted to an operation producing a structure. |
580 | // Extract individual results from the structure and return them as list. |
581 | results.reserve(N: numResults); |
582 | for (unsigned i = 0; i < numResults; ++i) { |
583 | results.push_back(rewriter.create<LLVM::ExtractValueOp>( |
584 | callOp.getLoc(), newOp->getResult(0), i)); |
585 | } |
586 | } |
587 | |
588 | if (useBarePtrCallConv) { |
589 | // For the bare-ptr calling convention, promote memref results to |
590 | // descriptors. |
591 | assert(results.size() == resultTypes.size() && |
592 | "The number of arguments and types doesn't match" ); |
593 | this->getTypeConverter()->promoteBarePtrsToDescriptors( |
594 | rewriter, callOp.getLoc(), resultTypes, results); |
595 | } else if (failed(this->copyUnrankedDescriptors(rewriter, callOp.getLoc(), |
596 | resultTypes, results, |
597 | /*toDynamic=*/false))) { |
598 | return failure(); |
599 | } |
600 | |
601 | rewriter.replaceOp(callOp, results); |
602 | return success(); |
603 | } |
604 | }; |
605 | |
606 | class CallOpLowering : public CallOpInterfaceLowering<func::CallOp> { |
607 | public: |
608 | CallOpLowering(const LLVMTypeConverter &typeConverter, |
609 | // Can be nullptr. |
610 | const SymbolTable *symbolTable, PatternBenefit benefit = 1) |
611 | : CallOpInterfaceLowering<func::CallOp>(typeConverter, benefit), |
612 | symbolTable(symbolTable) {} |
613 | |
614 | LogicalResult |
615 | matchAndRewrite(func::CallOp callOp, OpAdaptor adaptor, |
616 | ConversionPatternRewriter &rewriter) const override { |
617 | bool useBarePtrCallConv = false; |
618 | if (getTypeConverter()->getOptions().useBarePtrCallConv) { |
619 | useBarePtrCallConv = true; |
620 | } else if (symbolTable != nullptr) { |
621 | // Fast lookup. |
622 | Operation *callee = |
623 | symbolTable->lookup(callOp.getCalleeAttr().getValue()); |
624 | useBarePtrCallConv = |
625 | callee != nullptr && callee->hasAttr(name: barePtrAttrName); |
626 | } else { |
627 | // Warning: This is a linear lookup. |
628 | Operation *callee = |
629 | SymbolTable::lookupNearestSymbolFrom(callOp, callOp.getCalleeAttr()); |
630 | useBarePtrCallConv = |
631 | callee != nullptr && callee->hasAttr(name: barePtrAttrName); |
632 | } |
633 | return matchAndRewriteImpl(callOp, adaptor, rewriter, useBarePtrCallConv); |
634 | } |
635 | |
636 | private: |
637 | const SymbolTable *symbolTable = nullptr; |
638 | }; |
639 | |
640 | struct CallIndirectOpLowering |
641 | : public CallOpInterfaceLowering<func::CallIndirectOp> { |
642 | using Super::Super; |
643 | |
644 | LogicalResult |
645 | matchAndRewrite(func::CallIndirectOp callIndirectOp, OpAdaptor adaptor, |
646 | ConversionPatternRewriter &rewriter) const override { |
647 | return matchAndRewriteImpl(callIndirectOp, adaptor, rewriter); |
648 | } |
649 | }; |
650 | |
651 | struct UnrealizedConversionCastOpLowering |
652 | : public ConvertOpToLLVMPattern<UnrealizedConversionCastOp> { |
653 | using ConvertOpToLLVMPattern< |
654 | UnrealizedConversionCastOp>::ConvertOpToLLVMPattern; |
655 | |
656 | LogicalResult |
657 | matchAndRewrite(UnrealizedConversionCastOp op, OpAdaptor adaptor, |
658 | ConversionPatternRewriter &rewriter) const override { |
659 | SmallVector<Type> convertedTypes; |
660 | if (succeeded(typeConverter->convertTypes(op.getOutputs().getTypes(), |
661 | convertedTypes)) && |
662 | convertedTypes == adaptor.getInputs().getTypes()) { |
663 | rewriter.replaceOp(op, adaptor.getInputs()); |
664 | return success(); |
665 | } |
666 | |
667 | convertedTypes.clear(); |
668 | if (succeeded(typeConverter->convertTypes(adaptor.getInputs().getTypes(), |
669 | convertedTypes)) && |
670 | convertedTypes == op.getOutputs().getType()) { |
671 | rewriter.replaceOp(op, adaptor.getInputs()); |
672 | return success(); |
673 | } |
674 | return failure(); |
675 | } |
676 | }; |
677 | |
678 | // Special lowering pattern for `ReturnOps`. Unlike all other operations, |
679 | // `ReturnOp` interacts with the function signature and must have as many |
680 | // operands as the function has return values. Because in LLVM IR, functions |
681 | // can only return 0 or 1 value, we pack multiple values into a structure type. |
682 | // Emit `UndefOp` followed by `InsertValueOp`s to create such structure if |
683 | // necessary before returning it |
684 | struct ReturnOpLowering : public ConvertOpToLLVMPattern<func::ReturnOp> { |
685 | using ConvertOpToLLVMPattern<func::ReturnOp>::ConvertOpToLLVMPattern; |
686 | |
687 | LogicalResult |
688 | matchAndRewrite(func::ReturnOp op, OpAdaptor adaptor, |
689 | ConversionPatternRewriter &rewriter) const override { |
690 | Location loc = op.getLoc(); |
691 | unsigned numArguments = op.getNumOperands(); |
692 | SmallVector<Value, 4> updatedOperands; |
693 | |
694 | auto funcOp = op->getParentOfType<LLVM::LLVMFuncOp>(); |
695 | bool useBarePtrCallConv = |
696 | shouldUseBarePtrCallConv(funcOp, this->getTypeConverter()); |
697 | if (useBarePtrCallConv) { |
698 | // For the bare-ptr calling convention, extract the aligned pointer to |
699 | // be returned from the memref descriptor. |
700 | for (auto it : llvm::zip(op->getOperands(), adaptor.getOperands())) { |
701 | Type oldTy = std::get<0>(it).getType(); |
702 | Value newOperand = std::get<1>(it); |
703 | if (isa<MemRefType>(oldTy) && getTypeConverter()->canConvertToBarePtr( |
704 | cast<BaseMemRefType>(oldTy))) { |
705 | MemRefDescriptor memrefDesc(newOperand); |
706 | newOperand = memrefDesc.allocatedPtr(rewriter, loc); |
707 | } else if (isa<UnrankedMemRefType>(oldTy)) { |
708 | // Unranked memref is not supported in the bare pointer calling |
709 | // convention. |
710 | return failure(); |
711 | } |
712 | updatedOperands.push_back(newOperand); |
713 | } |
714 | } else { |
715 | updatedOperands = llvm::to_vector<4>(adaptor.getOperands()); |
716 | (void)copyUnrankedDescriptors(rewriter, loc, op.getOperands().getTypes(), |
717 | updatedOperands, |
718 | /*toDynamic=*/true); |
719 | } |
720 | |
721 | // If ReturnOp has 0 or 1 operand, create it and return immediately. |
722 | if (numArguments <= 1) { |
723 | rewriter.replaceOpWithNewOp<LLVM::ReturnOp>( |
724 | op, TypeRange(), updatedOperands, op->getAttrs()); |
725 | return success(); |
726 | } |
727 | |
728 | // Otherwise, we need to pack the arguments into an LLVM struct type before |
729 | // returning. |
730 | auto packedType = getTypeConverter()->packFunctionResults( |
731 | op.getOperandTypes(), useBarePtrCallConv); |
732 | if (!packedType) { |
733 | return rewriter.notifyMatchFailure(op, "could not convert result types" ); |
734 | } |
735 | |
736 | Value packed = rewriter.create<LLVM::UndefOp>(loc, packedType); |
737 | for (auto [idx, operand] : llvm::enumerate(First&: updatedOperands)) { |
738 | packed = rewriter.create<LLVM::InsertValueOp>(loc, packed, operand, idx); |
739 | } |
740 | rewriter.replaceOpWithNewOp<LLVM::ReturnOp>(op, TypeRange(), packed, |
741 | op->getAttrs()); |
742 | return success(); |
743 | } |
744 | }; |
745 | } // namespace |
746 | |
747 | void mlir::populateFuncToLLVMFuncOpConversionPattern( |
748 | LLVMTypeConverter &converter, RewritePatternSet &patterns) { |
749 | patterns.add<FuncOpConversion>(arg&: converter); |
750 | } |
751 | |
752 | void mlir::populateFuncToLLVMConversionPatterns( |
753 | LLVMTypeConverter &converter, RewritePatternSet &patterns, |
754 | const SymbolTable *symbolTable) { |
755 | populateFuncToLLVMFuncOpConversionPattern(converter, patterns); |
756 | patterns.add<CallIndirectOpLowering>(arg&: converter); |
757 | patterns.add<CallOpLowering>(arg&: converter, args&: symbolTable); |
758 | patterns.add<ConstantOpLowering>(arg&: converter); |
759 | patterns.add<ReturnOpLowering>(arg&: converter); |
760 | } |
761 | |
762 | namespace { |
763 | /// A pass converting Func operations into the LLVM IR dialect. |
764 | struct ConvertFuncToLLVMPass |
765 | : public impl::ConvertFuncToLLVMPassBase<ConvertFuncToLLVMPass> { |
766 | using Base::Base; |
767 | |
768 | /// Run the dialect converter on the module. |
769 | void runOnOperation() override { |
770 | ModuleOp m = getOperation(); |
771 | StringRef dataLayout; |
772 | auto dataLayoutAttr = dyn_cast_or_null<StringAttr>( |
773 | m->getAttr(LLVM::LLVMDialect::getDataLayoutAttrName())); |
774 | if (dataLayoutAttr) |
775 | dataLayout = dataLayoutAttr.getValue(); |
776 | |
777 | if (failed(LLVM::LLVMDialect::verifyDataLayoutString( |
778 | dataLayout, [this](const Twine &message) { |
779 | getOperation().emitError() << message.str(); |
780 | }))) { |
781 | signalPassFailure(); |
782 | return; |
783 | } |
784 | |
785 | const auto &dataLayoutAnalysis = getAnalysis<DataLayoutAnalysis>(); |
786 | |
787 | LowerToLLVMOptions options(&getContext(), |
788 | dataLayoutAnalysis.getAtOrAbove(m)); |
789 | options.useBarePtrCallConv = useBarePtrCallConv; |
790 | if (indexBitwidth != kDeriveIndexBitwidthFromDataLayout) |
791 | options.overrideIndexBitwidth(indexBitwidth); |
792 | options.dataLayout = llvm::DataLayout(dataLayout); |
793 | |
794 | LLVMTypeConverter typeConverter(&getContext(), options, |
795 | &dataLayoutAnalysis); |
796 | |
797 | std::optional<SymbolTable> optSymbolTable = std::nullopt; |
798 | const SymbolTable *symbolTable = nullptr; |
799 | if (!options.useBarePtrCallConv) { |
800 | optSymbolTable.emplace(m); |
801 | symbolTable = &optSymbolTable.value(); |
802 | } |
803 | |
804 | RewritePatternSet patterns(&getContext()); |
805 | populateFuncToLLVMConversionPatterns(converter&: typeConverter, patterns, symbolTable); |
806 | |
807 | // TODO(https://github.com/llvm/llvm-project/issues/70982): Remove these in |
808 | // favor of their dedicated conversion passes. |
809 | arith::populateArithToLLVMConversionPatterns(converter&: typeConverter, patterns); |
810 | cf::populateControlFlowToLLVMConversionPatterns(converter&: typeConverter, patterns); |
811 | |
812 | LLVMConversionTarget target(getContext()); |
813 | if (failed(applyPartialConversion(m, target, std::move(patterns)))) |
814 | signalPassFailure(); |
815 | } |
816 | }; |
817 | |
818 | struct SetLLVMModuleDataLayoutPass |
819 | : public impl::SetLLVMModuleDataLayoutPassBase< |
820 | SetLLVMModuleDataLayoutPass> { |
821 | using Base::Base; |
822 | |
823 | /// Run the dialect converter on the module. |
824 | void runOnOperation() override { |
825 | if (failed(LLVM::LLVMDialect::verifyDataLayoutString( |
826 | this->dataLayout, [this](const Twine &message) { |
827 | getOperation().emitError() << message.str(); |
828 | }))) { |
829 | signalPassFailure(); |
830 | return; |
831 | } |
832 | ModuleOp m = getOperation(); |
833 | m->setAttr(LLVM::LLVMDialect::getDataLayoutAttrName(), |
834 | StringAttr::get(m.getContext(), this->dataLayout)); |
835 | } |
836 | }; |
837 | } // namespace |
838 | |
839 | //===----------------------------------------------------------------------===// |
840 | // ConvertToLLVMPatternInterface implementation |
841 | //===----------------------------------------------------------------------===// |
842 | |
843 | namespace { |
844 | /// Implement the interface to convert Func to LLVM. |
845 | struct FuncToLLVMDialectInterface : public ConvertToLLVMPatternInterface { |
846 | using ConvertToLLVMPatternInterface::ConvertToLLVMPatternInterface; |
847 | /// Hook for derived dialect interface to provide conversion patterns |
848 | /// and mark dialect legal for the conversion target. |
849 | void populateConvertToLLVMConversionPatterns( |
850 | ConversionTarget &target, LLVMTypeConverter &typeConverter, |
851 | RewritePatternSet &patterns) const final { |
852 | populateFuncToLLVMConversionPatterns(converter&: typeConverter, patterns); |
853 | } |
854 | }; |
855 | } // namespace |
856 | |
857 | void mlir::registerConvertFuncToLLVMInterface(DialectRegistry ®istry) { |
858 | registry.addExtension(extensionFn: +[](MLIRContext *ctx, func::FuncDialect *dialect) { |
859 | dialect->addInterfaces<FuncToLLVMDialectInterface>(); |
860 | }); |
861 | } |
862 | |