1 //===- OneShotAnalysis.cpp - One-Shot (Single Pass) Analysis --------------===//
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 // One-Shot Analysis analyzes function bodies. Function boundaries (FuncOp
10 // bbArgs, CallOps, ReturnOps) are treated as "unknown" ops.
11 // ModuleBufferization.cpp is an extension of One-Shot Analysis for simple
12 // call graphs.
13 //
14 // One-Shot Bufferize consists of two phases.
15 //
16 // 1. Analyze ops to decide which OpResults can bufferize inplace, i.e., without
17 // inserting buffer copies. The analysis queries op bufferization semantics
18 // via `BufferizableOpInterface`.
19 // 2. Bufferize ops by calling `BufferizableOpInterface::bufferize`. This
20 // function does not generate buffer copies for OpResults that were decided
21 // to bufferize inplace during the analysis phase.
22 //
23 // This file contains only the analysis. The actual bufferization is implemented
24 // via `bufferizeOp` (Bufferize.h). For convenience, this file also contains a
25 // helper function `runOneShotBufferize` that analyzes an op (and its nested
26 // ops) and then bufferizes it.
27 //
28 // Inplace bufferization decisions are passed from the analysis to the
29 // bufferization phase via `AnalysisState` and `BufferizationAliasInfo`.
30 // They can be printed for debugging purposes with `testAnalysisOnly`.
31 //
32 // Ops that do not implement `BufferizableOpInterface` can be analyzed but are
33 // treated conservatively. E.g., the analysis has to assume that their tensor
34 // OpOperands bufferize to memory writes. While such ops can be analyzed, they
35 // are not bufferized and remain in the IR. to_tensor and to_memref ops are
36 // inserted at the bufferization boundary.
37 //
38 // This analysis caters to high-performance codegen where buffer reuse is deemed
39 // critical: the analysis should fail if the bufferized form of the function
40 // needs to return a buffer, unless `allowReturnAllocs` is enabled.
41
42 #include "mlir/Dialect/Bufferization/Transforms/OneShotAnalysis.h"
43
44 #include <random>
45
46 #include "mlir/Dialect/Bufferization/IR/BufferizableOpInterface.h"
47 #include "mlir/Dialect/Bufferization/IR/Bufferization.h"
48 #include "mlir/Dialect/Bufferization/Transforms/Bufferize.h"
49 #include "mlir/Dialect/Bufferization/Transforms/TensorCopyInsertion.h"
50 #include "mlir/Dialect/Func/IR/FuncOps.h"
51 #include "mlir/Dialect/MemRef/IR/MemRef.h"
52 #include "mlir/IR/AsmState.h"
53 #include "mlir/IR/Dominance.h"
54 #include "mlir/IR/Operation.h"
55 #include "mlir/IR/TypeUtilities.h"
56 #include "mlir/Interfaces/ControlFlowInterfaces.h"
57 #include "llvm/ADT/DenseSet.h"
58 #include "llvm/ADT/SetVector.h"
59
60 using namespace mlir;
61 using namespace mlir::bufferization;
62
isaTensor(Type t)63 static bool isaTensor(Type t) { return t.isa<TensorType>(); }
64
65 //===----------------------------------------------------------------------===//
66 // Bufferization-specific attribute manipulation.
67 // These are for testing and debugging only. Bufferization information is
68 // stored in BufferizationAliasInfo. When run with `testAnalysisOnly`, the IR
69 // is annotated with the results of the analysis (copied from
70 // BufferizationAliasInfo), so that they can be checked in tests.
71 //===----------------------------------------------------------------------===//
72
73 /// Attribute marker to specify op results that can be bufferized inPlace.
74 constexpr StringLiteral kInPlaceResultsAttrName = "__inplace_operands_attr__";
75
76 /// Mark whether OpOperand will be bufferized inplace.
setInPlaceOpOperand(OpOperand & opOperand,bool inPlace)77 static void setInPlaceOpOperand(OpOperand &opOperand, bool inPlace) {
78 Operation *op = opOperand.getOwner();
79 auto attr =
80 op->getAttr(kInPlaceResultsAttrName).dyn_cast_or_null<ArrayAttr>();
81 SmallVector<StringRef> inPlaceVector;
82 if (attr) {
83 inPlaceVector = SmallVector<StringRef>(
84 llvm::to_vector<4>(attr.getAsValueRange<StringAttr>()));
85 } else {
86 inPlaceVector = SmallVector<StringRef>(op->getNumOperands(), "none");
87 for (OpOperand &opOperand : op->getOpOperands())
88 if (opOperand.get().getType().isa<TensorType>())
89 inPlaceVector[opOperand.getOperandNumber()] = "false";
90 }
91
92 inPlaceVector[opOperand.getOperandNumber()] = inPlace ? "true" : "false";
93 op->setAttr(kInPlaceResultsAttrName,
94 OpBuilder(op).getStrArrayAttr(inPlaceVector));
95 }
96
97 //===----------------------------------------------------------------------===//
98 // BufferizationAliasInfo
99 //===----------------------------------------------------------------------===//
100
BufferizationAliasInfo(Operation * rootOp)101 BufferizationAliasInfo::BufferizationAliasInfo(Operation *rootOp) {
102 rootOp->walk([&](Operation *op) {
103 for (Value v : op->getResults())
104 if (v.getType().isa<TensorType>())
105 createAliasInfoEntry(v);
106 for (Region &r : op->getRegions())
107 for (Block &b : r.getBlocks())
108 for (auto bbArg : b.getArguments())
109 if (bbArg.getType().isa<TensorType>())
110 createAliasInfoEntry(bbArg);
111 });
112 }
113
114 /// Add a new entry for `v` in the `aliasInfo` and `equivalentInfo`. In the
115 /// beginning the alias and equivalence sets only contain `v` itself.
createAliasInfoEntry(Value v)116 void BufferizationAliasInfo::createAliasInfoEntry(Value v) {
117 aliasInfo.insert(v);
118 equivalentInfo.insert(v);
119 }
120
121 /// Insert an info entry for `newValue` and merge its alias set with that of
122 /// `alias`.
insertNewBufferAlias(Value newValue,Value alias)123 void BufferizationAliasInfo::insertNewBufferAlias(Value newValue, Value alias) {
124 createAliasInfoEntry(newValue);
125 aliasInfo.unionSets(newValue, alias);
126 }
127
128 /// Insert an info entry for `newValue` and merge its alias set with that of
129 /// `alias`. Additionally, merge their equivalence classes.
insertNewBufferEquivalence(Value newValue,Value alias)130 void BufferizationAliasInfo::insertNewBufferEquivalence(Value newValue,
131 Value alias) {
132 insertNewBufferAlias(newValue, alias);
133 equivalentInfo.unionSets(newValue, alias);
134 }
135
136 /// Return `true` if a value was marked as in-place bufferized.
isInPlace(OpOperand & operand) const137 bool BufferizationAliasInfo::isInPlace(OpOperand &operand) const {
138 return inplaceBufferized.contains(&operand);
139 }
140
141 /// Set the inPlace bufferization spec to true.
bufferizeInPlace(OpOperand & operand,AnalysisState & state)142 void BufferizationAliasInfo::bufferizeInPlace(OpOperand &operand,
143 AnalysisState &state) {
144 markInPlace(operand);
145 for (OpResult result : state.getAliasingOpResult(operand))
146 aliasInfo.unionSets(result, operand.get());
147 }
148
149 /// Set the inPlace bufferization spec to false.
bufferizeOutOfPlace(OpOperand & operand)150 void BufferizationAliasInfo::bufferizeOutOfPlace(OpOperand &operand) {
151 assert(!inplaceBufferized.contains(&operand) &&
152 "OpOperand was already decided to bufferize inplace");
153 }
154
155 /// Apply `fun` to all the members of the equivalence class of `v`.
applyOnEquivalenceClass(Value v,function_ref<void (Value)> fun) const156 void BufferizationAliasInfo::applyOnEquivalenceClass(
157 Value v, function_ref<void(Value)> fun) const {
158 auto leaderIt = equivalentInfo.findLeader(v);
159 for (auto mit = leaderIt, meit = equivalentInfo.member_end(); mit != meit;
160 ++mit) {
161 fun(*mit);
162 }
163 }
164
165 /// Apply `fun` to all aliases of `v`.
applyOnAliases(Value v,function_ref<void (Value)> fun) const166 void BufferizationAliasInfo::applyOnAliases(
167 Value v, function_ref<void(Value)> fun) const {
168 auto leaderIt = aliasInfo.findLeader(v);
169 for (auto mit = leaderIt, meit = aliasInfo.member_end(); mit != meit; ++mit) {
170 fun(*mit);
171 }
172 }
173
174 BufferizationAliasInfo::EquivalenceClassRangeType
getAliases(Value v) const175 BufferizationAliasInfo::getAliases(Value v) const {
176 DenseSet<Value> res;
177 auto it = aliasInfo.findValue(aliasInfo.getLeaderValue(v));
178 for (auto mit = aliasInfo.member_begin(it), meit = aliasInfo.member_end();
179 mit != meit; ++mit) {
180 res.insert(static_cast<Value>(*mit));
181 }
182 return BufferizationAliasInfo::EquivalenceClassRangeType(
183 aliasInfo.member_begin(it), aliasInfo.member_end());
184 }
185
186 //===----------------------------------------------------------------------===//
187 // OneShotAnalysisState
188 //===----------------------------------------------------------------------===//
189
OneShotAnalysisState(Operation * op,const OneShotBufferizationOptions & options)190 OneShotAnalysisState::OneShotAnalysisState(
191 Operation *op, const OneShotBufferizationOptions &options)
192 : AnalysisState(options), aliasInfo(op) {
193 // Set up alias sets for OpResults that must bufferize in-place. This should
194 // be done before making any other bufferization decisions.
195 op->walk([&](BufferizableOpInterface bufferizableOp) {
196 if (!options.isOpAllowed(bufferizableOp))
197 return WalkResult::skip();
198 for (OpOperand &opOperand : bufferizableOp->getOpOperands()) {
199 if (opOperand.get().getType().isa<TensorType>())
200 if (bufferizableOp.mustBufferizeInPlace(opOperand, *this)) {
201 for (OpResult opResult :
202 bufferizableOp.getAliasingOpResult(opOperand, *this))
203 aliasInfo.unionAliasSets(opOperand.get(), opResult);
204 aliasInfo.markInPlace(opOperand);
205 }
206 }
207 return WalkResult::advance();
208 });
209 }
210
isInPlace(OpOperand & opOperand) const211 bool OneShotAnalysisState::isInPlace(OpOperand &opOperand) const {
212 return aliasInfo.isInPlace(opOperand);
213 }
214
areEquivalentBufferizedValues(Value v1,Value v2) const215 bool OneShotAnalysisState::areEquivalentBufferizedValues(Value v1,
216 Value v2) const {
217 return aliasInfo.areEquivalentBufferizedValues(v1, v2);
218 }
219
areAliasingBufferizedValues(Value v1,Value v2) const220 bool OneShotAnalysisState::areAliasingBufferizedValues(Value v1,
221 Value v2) const {
222 return aliasInfo.areAliasingBufferizedValues(v1, v2);
223 }
224
225 // Gather yielded tensors in `yieldedTensors` by querying all aliases. This is
226 // to ensure that such information is available during bufferization time.
227 // Alias information can no longer be queried through BufferizationAliasInfo
228 // once we have started modifying the IR.
gatherYieldedTensors(Operation * op)229 void OneShotAnalysisState::gatherYieldedTensors(Operation *op) {
230 op->walk([&](Operation *returnOp) {
231 if (!isRegionReturnLike(returnOp) || !getOptions().isOpAllowed(returnOp))
232 return WalkResult::advance();
233
234 for (OpOperand &returnValOperand : returnOp->getOpOperands()) {
235 Value returnVal = returnValOperand.get();
236 // Skip non-tensor values.
237 if (!returnVal.getType().isa<TensorType>())
238 continue;
239
240 // Add all aliases of the returned value. But only the ones that are in
241 // the same block.
242 aliasInfo.applyOnAliases(returnVal, [&](Value v) {
243 if (auto bbArg = v.dyn_cast<BlockArgument>()) {
244 if (bbArg.getOwner()->getParentOp() == returnOp->getParentOp())
245 yieldedTensors.insert(bbArg);
246 return;
247 }
248 Operation *definingOp = v.getDefiningOp();
249 if (definingOp->getParentOp() == returnOp->getParentOp())
250 yieldedTensors.insert(v);
251 });
252 }
253
254 return WalkResult::advance();
255 });
256 }
257
gatherUndefinedTensorUses(Operation * op)258 void OneShotAnalysisState::gatherUndefinedTensorUses(Operation *op) {
259 op->walk([&](Operation *op) {
260 // Skip unknown ops.
261 auto bufferizableOp = getOptions().dynCastBufferizableOp(op);
262 if (!bufferizableOp)
263 return WalkResult::skip();
264
265 // Check all tensor OpResults.
266 for (OpResult opResult : op->getOpResults()) {
267 if (!opResult.getType().isa<TensorType>())
268 continue;
269
270 // If there is no preceding memory write, the tensor contents are
271 // undefined.
272 // Note: If `findLastPrecedingWrite` reaches the end of the reverse SSA
273 // use-def chain, it returns that value, regardless of whether it is a
274 // memory write or not.
275 SetVector<Value> lastWrites = findLastPrecedingWrite(opResult);
276 bool isUndefined = llvm::none_of(lastWrites, [&](Value lastWrite) {
277 if (auto bufferizableOp = getOptions().dynCastBufferizableOp(lastWrite))
278 return bufferizableOp.isMemoryWrite(lastWrite.cast<OpResult>(),
279 *this);
280 return true;
281 });
282 if (isUndefined)
283 for (OpOperand &use : opResult.getUses())
284 undefinedTensorUses.insert(&use);
285 }
286
287 return WalkResult::advance();
288 });
289 }
290
hasUndefinedContents(OpOperand * opOperand) const291 bool OneShotAnalysisState::hasUndefinedContents(OpOperand *opOperand) const {
292 return undefinedTensorUses.contains(opOperand);
293 }
294
isTensorYielded(Value tensor) const295 bool OneShotAnalysisState::isTensorYielded(Value tensor) const {
296 return yieldedTensors.contains(tensor);
297 }
298
isValueWritten(Value value) const299 bool OneShotAnalysisState::isValueWritten(Value value) const {
300 bool isWritten = false;
301 aliasInfo.applyOnAliases(value, [&](Value val) {
302 for (OpOperand &use : val.getUses())
303 if (isInPlace(use) && bufferizesToMemoryWrite(use))
304 isWritten = true;
305 });
306 return isWritten;
307 }
308
isWritable(Value value) const309 bool OneShotAnalysisState::isWritable(Value value) const {
310 // TODO: Out-of-place bufferized value could be considered writable.
311 if (auto bufferizableOp = getOptions().dynCastBufferizableOp(value))
312 return bufferizableOp.isWritable(value, *this);
313
314 // Query BufferizableOpInterface to see if the BlockArgument is writable.
315 if (auto bbArg = value.dyn_cast<BlockArgument>())
316 if (auto bufferizableOp =
317 getOptions().dynCastBufferizableOp(bbArg.getOwner()->getParentOp()))
318 return bufferizableOp.isWritable(bbArg, *this);
319
320 // Not a bufferizable op: The conservative answer is "not writable".
321 return false;
322 }
323
324 //===----------------------------------------------------------------------===//
325 // Bufferization-specific alias analysis.
326 //===----------------------------------------------------------------------===//
327
328 /// Return true if opOperand has been decided to bufferize in-place.
isInplaceMemoryWrite(OpOperand & opOperand,const BufferizationAliasInfo & aliasInfo,const AnalysisState & state)329 static bool isInplaceMemoryWrite(OpOperand &opOperand,
330 const BufferizationAliasInfo &aliasInfo,
331 const AnalysisState &state) {
332 // OpOperands that do not bufferize to a memory write do not write in-place.
333 if (!state.bufferizesToMemoryWrite(opOperand))
334 return false;
335 // Check current bufferization decisions.
336 return aliasInfo.isInPlace(opOperand);
337 }
338
339 /// Return true if `a` happens before `b`, i.e., `a` or one of its ancestors
340 /// properly dominates `b` and `b` is not inside `a`.
happensBefore(Operation * a,Operation * b,const DominanceInfo & domInfo)341 static bool happensBefore(Operation *a, Operation *b,
342 const DominanceInfo &domInfo) {
343 do {
344 // TODO: Instead of isProperAncestor + properlyDominates, we should use
345 // properlyDominatesImpl(a, b, /*enclosingOpOk=*/false)
346 if (a->isProperAncestor(b))
347 return false;
348 if (domInfo.properlyDominates(a, b))
349 return true;
350 } while ((a = a->getParentOp()));
351 return false;
352 }
353
354 /// For each given value, find the closest enclosing repetitive region. If this
355 /// is the same region for each value, return it. Otherwise return None.
356 /// Note: If there is no enclosing repetitive region, return nullptr.
357 static Optional<Region *>
getCommonEnclosingRepetitiveRegion(ArrayRef<Value> values)358 getCommonEnclosingRepetitiveRegion(ArrayRef<Value> values) {
359 if (values.empty())
360 return None;
361 Region *r = getEnclosingRepetitiveRegion(values.front());
362 for (Value value : values.drop_front())
363 if (getEnclosingRepetitiveRegion(value) != r)
364 return None;
365 return r;
366 }
367
368 /// Return `true` if the given tensor value is a memory write. Most values are
369 /// tensor writes, but ops that define a tensor SSA value without specifying its
370 /// contents (e.g., alloc_tensor) are not.
isMemoryWrite(Value value,const AnalysisState & state)371 static bool isMemoryWrite(Value value, const AnalysisState &state) {
372 auto opResult = value.dyn_cast<OpResult>();
373 if (!opResult)
374 return true;
375 auto bufferizableOp = state.getOptions().dynCastBufferizableOp(value);
376 if (!bufferizableOp)
377 return true;
378 return bufferizableOp.isMemoryWrite(opResult, state);
379 }
380
381 /// Annotate IR with details about the detected RaW conflict.
annotateConflict(OpOperand * uRead,OpOperand * uConflictingWrite,Value lastWrite)382 static void annotateConflict(OpOperand *uRead, OpOperand *uConflictingWrite,
383 Value lastWrite) {
384 static uint64_t counter = 0;
385 Operation *readingOp = uRead->getOwner();
386 Operation *conflictingWritingOp = uConflictingWrite->getOwner();
387
388 OpBuilder b(conflictingWritingOp->getContext());
389 std::string id = "C_" + std::to_string(counter++);
390
391 std::string conflictingWriteAttr =
392 id +
393 "[CONFL-WRITE: " + std::to_string(uConflictingWrite->getOperandNumber()) +
394 "]";
395 conflictingWritingOp->setAttr(conflictingWriteAttr, b.getUnitAttr());
396
397 std::string readAttr =
398 id + "[READ: " + std::to_string(uRead->getOperandNumber()) + "]";
399 readingOp->setAttr(readAttr, b.getUnitAttr());
400
401 if (auto opResult = lastWrite.dyn_cast<OpResult>()) {
402 std::string lastWriteAttr = id + "[LAST-WRITE: result " +
403 std::to_string(opResult.getResultNumber()) +
404 "]";
405 opResult.getDefiningOp()->setAttr(lastWriteAttr, b.getUnitAttr());
406 } else {
407 auto bbArg = lastWrite.cast<BlockArgument>();
408 std::string lastWriteAttr =
409 id + "[LAST-WRITE: bbArg " + std::to_string(bbArg.getArgNumber()) + "]";
410 bbArg.getOwner()->getParentOp()->setAttr(lastWriteAttr, b.getUnitAttr());
411 }
412 }
413
414 /// Given sets of uses and writes, return true if there is a RaW conflict under
415 /// the assumption that all given reads/writes alias the same buffer and that
416 /// all given writes bufferize inplace.
417 ///
418 /// A conflict is: According to SSA use-def chains, a read R is supposed to read
419 /// the result of a write W1. But because of bufferization decisions, R actually
420 /// reads another write W2.
hasReadAfterWriteInterference(const DenseSet<OpOperand * > & usesRead,const DenseSet<OpOperand * > & usesWrite,const DominanceInfo & domInfo,AnalysisState & state,const BufferizationAliasInfo & aliasInfo)421 static bool hasReadAfterWriteInterference(
422 const DenseSet<OpOperand *> &usesRead,
423 const DenseSet<OpOperand *> &usesWrite, const DominanceInfo &domInfo,
424 AnalysisState &state, const BufferizationAliasInfo &aliasInfo) {
425 const BufferizationOptions &options = state.getOptions();
426
427 // Gather all written aliases. Skip over aliases that are not actual writes.
428 SmallVector<Value> writtenAliases;
429 for (OpOperand *uWrite : usesWrite)
430 if (isMemoryWrite(uWrite->get(), state))
431 writtenAliases.push_back(uWrite->get());
432 // Find the inner-most enclosing repetitive region of each alias. If this is
433 // the same region for every alias, save it in `repetitiveRegionOfWrites`.
434 Optional<Region *> repetitiveRegionOfWrites =
435 getCommonEnclosingRepetitiveRegion(writtenAliases);
436
437 for (OpOperand *uRead : usesRead) {
438 Operation *readingOp = uRead->getOwner();
439
440 // Find most recent writes of uRead by following the SSA use-def chain.
441 // E.g.:
442 //
443 // %0 = "writing_op"(%t) : tensor<?x32> -> tensor<?xf32>
444 // %1 = "aliasing_op"(%0) : tensor<?x32> -> tensor<?xf32>
445 // %2 = "reading_op"(%1) : : tensor<?x32> -> not_a_tensor_type
446 //
447 // In the above example, if uRead is the OpOperand of reading_op, lastWrite
448 // is %0. Note that operations that create an alias but do not write (such
449 // as ExtractSliceOp) are skipped.
450 SetVector<Value> lastWrites = state.findLastPrecedingWrite(uRead->get());
451
452 // Look for conflicting memory writes. Potential conflicts are writes to an
453 // alias that have been decided to bufferize inplace.
454 for (OpOperand *uConflictingWrite : usesWrite) {
455 // Throughout this loop, check for multiple requirements that have to be
456 // met for uConflictingWrite to be an actual conflict.
457 Operation *conflictingWritingOp = uConflictingWrite->getOwner();
458
459 // Check if conflictingWritingOp is in the same repetitive region as all
460 // written aliases. If this is not the case, there is no meaningful
461 // `happensBefore` relationship because conflictingWritingOp may be
462 // executed multiple times. E.g.:
463 //
464 // %0 = ... : tensor<?xf32>
465 // scf.for ... {
466 // "reading_op"(%0) : tensor<?xf32>
467 // %1 = "writing_op"(%0) : tensor<?xf32> -> tensor<?xf32>
468 // ...
469 // }
470 //
471 // In the above example, reading_op happens before writing_op according to
472 // op dominance. However, both ops may happen multiple times; in
473 // particular, the second execution of reading_op happens after the first
474 // execution of writing_op. This is problematic if the tensor they operate
475 // on (%0) is defined outside of the loop.
476 //
477 // Counter example:
478 //
479 // scf.for ... {
480 // %0 = ... : tensor<?xf32>
481 // "reading_op"(%0) : tensor<?xf32>
482 // %1 = "writing_op"(%0) : tensor<?xf32> -> tensor<?xf32>
483 // ...
484 // }
485 //
486 // In this example, %0 is in the same repetitive region as
487 // conflictingWritingOp, so op dominance can be used to compute the
488 // `happensBefore` relationship.
489 //
490 // Note: iter_args of loops are not aliases of their respective block
491 // arguments, so op domanice can be used when analyzing ops that operate
492 // on them.
493 //
494 // Note: If `writtenAliases` is empty, there are no memory writes outside
495 // of the repetitive region of conflictingWritingOp, which means that all
496 // relevant aliases are inside the same repetitive region.
497 bool canUseOpDominance =
498 writtenAliases.empty() ||
499 repetitiveRegionOfWrites ==
500 getEnclosingRepetitiveRegion(conflictingWritingOp);
501
502 // No conflict if the readingOp dominates conflictingWritingOp, i.e., the
503 // write is not visible when reading.
504 //
505 // Note: If ops are executed multiple times (e.g., because they are inside
506 // a loop), there may be no meaningful `happensBefore` relationship.
507 if (canUseOpDominance &&
508 happensBefore(readingOp, conflictingWritingOp, domInfo))
509 continue;
510
511 // No conflict if the reading use equals the use of the conflicting write.
512 // A use cannot conflict with itself.
513 //
514 // Note: Just being the same op is not enough. It has to be the same use.
515 // Note: If the op is executed multiple times (e.g., because it is inside
516 // a loop), it may be conflicting with itself.
517 if (canUseOpDominance && uConflictingWrite == uRead)
518 continue;
519
520 // No conflict if the op interface says so.
521 if (auto bufferizableOp = options.dynCastBufferizableOp(readingOp))
522 if (bufferizableOp.isNotConflicting(uRead, uConflictingWrite, state))
523 continue;
524
525 if (conflictingWritingOp != readingOp)
526 if (auto bufferizableOp =
527 options.dynCastBufferizableOp(conflictingWritingOp))
528 if (bufferizableOp.isNotConflicting(uRead, uConflictingWrite, state))
529 continue;
530
531 // Ops are not conflicting if they are in mutually exclusive regions.
532 //
533 // Note: If ops are executed multiple times (e.g., because they are inside
534 // a loop), mutually exclusive regions may be executed multiple
535 // times.
536 if (canUseOpDominance &&
537 insideMutuallyExclusiveRegions(readingOp, conflictingWritingOp))
538 continue;
539
540 // Check all possible last writes.
541 for (Value lastWrite : lastWrites) {
542 // No conflict if the conflicting write happens before the last
543 // write.
544 if (Operation *writingOp = lastWrite.getDefiningOp()) {
545 if (happensBefore(conflictingWritingOp, writingOp, domInfo))
546 // conflictingWritingOp happens before writingOp. No conflict.
547 continue;
548 // No conflict if conflictingWritingOp is contained in writingOp.
549 if (writingOp->isProperAncestor(conflictingWritingOp))
550 continue;
551 } else {
552 auto bbArg = lastWrite.cast<BlockArgument>();
553 Block *block = bbArg.getOwner();
554 if (!block->findAncestorOpInBlock(*conflictingWritingOp))
555 // conflictingWritingOp happens outside of the block. No
556 // conflict.
557 continue;
558 }
559
560 // No conflict if the conflicting write and the last write are the same
561 // use.
562 SmallVector<OpResult> aliasingOpResult =
563 state.getAliasingOpResult(*uConflictingWrite);
564 if (aliasingOpResult.size() == 1 && aliasingOpResult[0] == lastWrite)
565 continue;
566
567 // All requirements are met. Conflict found!
568
569 if (options.printConflicts)
570 annotateConflict(uRead, uConflictingWrite, lastWrite);
571
572 return true;
573 }
574 }
575 }
576
577 return false;
578 }
579
580 // Helper function to iterate on aliases of `root` and capture the writes.
getAliasingInplaceWrites(DenseSet<OpOperand * > & res,Value root,const BufferizationAliasInfo & aliasInfo,const AnalysisState & state)581 static void getAliasingInplaceWrites(DenseSet<OpOperand *> &res, Value root,
582 const BufferizationAliasInfo &aliasInfo,
583 const AnalysisState &state) {
584 aliasInfo.applyOnAliases(root, [&](Value alias) {
585 for (auto &use : alias.getUses())
586 // Inplace write to a value that aliases root.
587 if (isInplaceMemoryWrite(use, aliasInfo, state))
588 res.insert(&use);
589 });
590 }
591
592 // Helper function to iterate on aliases of `root` and capture the reads.
getAliasingReads(DenseSet<OpOperand * > & res,Value root,const BufferizationAliasInfo & aliasInfo,const AnalysisState & state)593 static void getAliasingReads(DenseSet<OpOperand *> &res, Value root,
594 const BufferizationAliasInfo &aliasInfo,
595 const AnalysisState &state) {
596 aliasInfo.applyOnAliases(root, [&](Value alias) {
597 for (auto &use : alias.getUses())
598 // Read to a value that aliases root.
599 if (state.bufferizesToMemoryRead(use))
600 res.insert(&use);
601 });
602 }
603
604 /// Return true if bufferizing `operand` inplace would create a conflict. A read
605 /// R and a write W of the same alias set is a conflict if inplace bufferization
606 /// of W changes the value read by R to a value different from the one that
607 /// would be expected by tracing back R's origin through SSA use-def chains.
608 /// A conflict can only be introduced by a new alias and/or an inplace
609 /// bufferization decision.
610 ///
611 /// Example:
612 /// %0 = tensor.extract_slice %t[...][...][1, 1] {inplace?}
613 /// %1 = vector.transfer_write %v1, %t {inplace} : vector<5xf32>, tensor<?xf32>
614 /// %e = tensor.extract_slice %1
615 /// %2 = vector.transfer_write %v2, %0 {inplace} : vector<6xf32>, tensor<?xf32>
616 /// %3 = vector.transfer_read %e, %cst : tensor<?xf32>, vector<7xf32>
617 ///
618 /// In the above example, the two TransferWriteOps have already been decided to
619 /// bufferize inplace. Bufferizing the ExtractSliceOp inplace would create a
620 /// conflict because:
621 /// * According to SSA use-def chains, we expect to read the result of %1.
622 /// * However, adding an alias {%0, %t} would mean that the second
623 /// TransferWriteOp overwrites the first one. Therefore, the TransferReadOp
624 /// would no longer be reading the result of %1.
625 ///
626 /// If `checkConsistencyOnly` is true, this function checks if there is a
627 /// read-after-write conflict without bufferizing `operand` inplace. This would
628 /// indicate a problem with the current inplace bufferization decisions.
629 ///
630 /// Note: If `checkConsistencyOnly`, this function may be called with a null
631 /// OpResult. In that case, only the consistency of bufferization decisions
632 /// involving aliases of the given OpOperand are checked.
wouldCreateReadAfterWriteInterference(OpOperand & operand,const DominanceInfo & domInfo,AnalysisState & state,const BufferizationAliasInfo & aliasInfo,bool checkConsistencyOnly=false)633 static bool wouldCreateReadAfterWriteInterference(
634 OpOperand &operand, const DominanceInfo &domInfo, AnalysisState &state,
635 const BufferizationAliasInfo &aliasInfo,
636 bool checkConsistencyOnly = false) {
637 // Collect reads and writes of all aliases of OpOperand and OpResult.
638 DenseSet<OpOperand *> usesRead, usesWrite;
639 getAliasingReads(usesRead, operand.get(), aliasInfo, state);
640 getAliasingInplaceWrites(usesWrite, operand.get(), aliasInfo, state);
641 for (OpResult result : state.getAliasingOpResult(operand)) {
642 getAliasingReads(usesRead, result, aliasInfo, state);
643 getAliasingInplaceWrites(usesWrite, result, aliasInfo, state);
644 }
645 if (!checkConsistencyOnly && state.bufferizesToMemoryWrite(operand))
646 usesWrite.insert(&operand);
647
648 return hasReadAfterWriteInterference(usesRead, usesWrite, domInfo, state,
649 aliasInfo);
650 }
651
652 /// Check the reverse SSA use-def chain (following aliasing OpOperands) for
653 /// non-writable tensor values. Stop searching when an out-of-place bufferized
654 /// OpOperand was found (or when the OpOperand was not bufferized yet).
655 /// `currentOpOperand` is assumed to be in-place, even if that decision was not
656 /// materialized in `aliasInfo` yet.
657 static bool
hasPrecedingAliasingNonWritableTensor(Value value,OpOperand * currentOpOperand,const BufferizationAliasInfo & aliasInfo,const OneShotAnalysisState & state)658 hasPrecedingAliasingNonWritableTensor(Value value, OpOperand *currentOpOperand,
659 const BufferizationAliasInfo &aliasInfo,
660 const OneShotAnalysisState &state) {
661 SmallVector<Value> worklist;
662 worklist.push_back(value);
663 while (!worklist.empty()) {
664 Value nextVal = worklist.pop_back_val();
665 if (!state.isWritable(nextVal))
666 return true;
667
668 // If `nextVal` is not a BlockArgument: End of use-def chain reached.
669 auto opResult = nextVal.dyn_cast<OpResult>();
670 if (!opResult)
671 continue;
672
673 // Follow reverse SSA use-def chain.
674 SmallVector<OpOperand *> aliasingOpOperands =
675 state.getAliasingOpOperand(opResult);
676 for (OpOperand *opOperand : aliasingOpOperands)
677 if (aliasInfo.isInPlace(*opOperand) || currentOpOperand == opOperand)
678 worklist.push_back(opOperand->get());
679 }
680 return false;
681 }
682
683 /// Return true if bufferizing `operand` inplace would create a write to a
684 /// non-writable buffer.
wouldCreateWriteToNonWritableBuffer(OpOperand & operand,const BufferizationAliasInfo & aliasInfo,OneShotAnalysisState & state,bool checkConsistencyOnly=false)685 static bool wouldCreateWriteToNonWritableBuffer(
686 OpOperand &operand, const BufferizationAliasInfo &aliasInfo,
687 OneShotAnalysisState &state, bool checkConsistencyOnly = false) {
688 // Collect writes of all aliases of OpOperand and OpResult.
689 DenseSet<OpOperand *> usesWrite;
690 getAliasingInplaceWrites(usesWrite, operand.get(), aliasInfo, state);
691 for (OpResult result : state.getAliasingOpResult(operand)) {
692 getAliasingInplaceWrites(usesWrite, result, aliasInfo, state);
693 }
694 if (!checkConsistencyOnly && state.bufferizesToMemoryWrite(operand))
695 usesWrite.insert(&operand);
696
697 // Assuming that `operand` bufferizes in-place: For each write (to each
698 // alias), check if there is a non-writable tensor in the reverse SSA use-def
699 // chain.
700 for (OpOperand *uWrite : usesWrite)
701 if (hasPrecedingAliasingNonWritableTensor(uWrite->get(), &operand,
702 aliasInfo, state))
703 return true;
704
705 return false;
706 }
707
708 //===----------------------------------------------------------------------===//
709 // Bufferization analyses.
710 //===----------------------------------------------------------------------===//
711
712 /// Determine if `operand` can be bufferized in-place.
bufferizableInPlaceAnalysisImpl(OpOperand & operand,BufferizationAliasInfo & aliasInfo,OneShotAnalysisState & state,const DominanceInfo & domInfo)713 static LogicalResult bufferizableInPlaceAnalysisImpl(
714 OpOperand &operand, BufferizationAliasInfo &aliasInfo,
715 OneShotAnalysisState &state, const DominanceInfo &domInfo) {
716 bool foundInterference =
717 wouldCreateWriteToNonWritableBuffer(operand, aliasInfo, state) ||
718 wouldCreateReadAfterWriteInterference(operand, domInfo, state, aliasInfo);
719
720 if (foundInterference)
721 aliasInfo.bufferizeOutOfPlace(operand);
722 else
723 aliasInfo.bufferizeInPlace(operand, state);
724
725 return success();
726 }
727
728 /// Analyze the `ops` to determine which OpOperands are inplaceable. Walk ops in
729 /// reverse and bufferize ops greedily. This is a good starter heuristic.
730 ///
731 /// Even if an op does not read or write, it may still create an alias when
732 /// bufferized in-place. An example of such ops is tensor.extract_slice.
733 ///
734 /// Rationale for bufferizing `%1 = tensor.extract_slice %0[...]` inplace:
735 ///
736 /// When bufferized out of place, an ExtractSliceOp lowers to alloc + copy. This
737 /// cannot change the flow of information for either the source or the
738 /// result buffers.
739 ///
740 /// When bufferized inplace, an ExtractSliceOp does not by itself create any
741 /// read or write from memory. Instead, it has the effect of merging the alias
742 /// sets of the source and the result buffers.
743 ///
744 /// An analysis is required to ensure inplace bufferization would not result in
745 /// RaW dependence violations.
inPlaceAnalysis(SmallVector<Operation * > & ops,BufferizationAliasInfo & aliasInfo,OneShotAnalysisState & state,const DominanceInfo & domInfo,unsigned analysisFuzzerSeed=0)746 static LogicalResult inPlaceAnalysis(SmallVector<Operation *> &ops,
747 BufferizationAliasInfo &aliasInfo,
748 OneShotAnalysisState &state,
749 const DominanceInfo &domInfo,
750 unsigned analysisFuzzerSeed = 0) {
751 if (analysisFuzzerSeed) {
752 // This is a fuzzer. For testing purposes only. Randomize the order in which
753 // operations are analyzed. The bufferization quality is likely worse, but
754 // we want to make sure that no assertions are triggered anywhere.
755 std::mt19937 g(analysisFuzzerSeed);
756 llvm::shuffle(ops.begin(), ops.end(), g);
757 }
758
759 // Walk ops in reverse for better interference analysis.
760 for (Operation *op : reverse(ops))
761 for (OpOperand &opOperand : op->getOpOperands())
762 if (opOperand.get().getType().isa<TensorType>())
763 if (auto bufferizableOp = state.getOptions().dynCastBufferizableOp(op))
764 if (failed(bufferizableInPlaceAnalysisImpl(opOperand, aliasInfo,
765 state, domInfo)))
766 return failure();
767
768 return success();
769 }
770
771 /// Return true if the given op has a tensor result or a tensor operand.
hasTensorSemantics(Operation * op)772 static bool hasTensorSemantics(Operation *op) {
773 bool hasTensorResult = any_of(op->getResultTypes(), isaTensor);
774 bool hasTensorOperand = any_of(op->getOperandTypes(), isaTensor);
775 return hasTensorResult || hasTensorOperand;
776 }
777
778 /// Analyze all ops that are contained in `op`.
inPlaceAnalysis(Operation * op,BufferizationAliasInfo & aliasInfo,OneShotAnalysisState & state,const DominanceInfo & domInfo,unsigned analysisFuzzerSeed=0)779 static LogicalResult inPlaceAnalysis(Operation *op,
780 BufferizationAliasInfo &aliasInfo,
781 OneShotAnalysisState &state,
782 const DominanceInfo &domInfo,
783 unsigned analysisFuzzerSeed = 0) {
784 // Collect ops so we can build our own reverse traversal.
785 SmallVector<Operation *> ops;
786 op->walk([&](Operation *op) {
787 // No tensors => no buffers.
788 if (!hasTensorSemantics(op))
789 return;
790 ops.push_back(op);
791 });
792
793 return inPlaceAnalysis(ops, aliasInfo, state, domInfo, analysisFuzzerSeed);
794 }
795
796 /// Analyze equivalence of tied OpResult/OpOperand pairs of the given ops.
equivalenceAnalysis(SmallVector<Operation * > & ops,BufferizationAliasInfo & aliasInfo,AnalysisState & state)797 static void equivalenceAnalysis(SmallVector<Operation *> &ops,
798 BufferizationAliasInfo &aliasInfo,
799 AnalysisState &state) {
800 for (Operation *op : ops)
801 if (auto bufferizableOp = state.getOptions().dynCastBufferizableOp(op))
802 for (OpResult opResult : op->getOpResults())
803 if (opResult.getType().isa<TensorType>())
804 for (OpOperand *opOperand :
805 bufferizableOp.getAliasingOpOperand(opResult, state))
806 if (state.isInPlace(*opOperand))
807 if (bufferizableOp.bufferRelation(opResult, state) ==
808 BufferRelation::Equivalent)
809 aliasInfo.unionEquivalenceClasses(opResult, opOperand->get());
810 }
811
812 /// Analyze equivalence of tied OpResult/OpOperand pairs of all ops contained
813 /// in `op`.
equivalenceAnalysis(Operation * op,BufferizationAliasInfo & aliasInfo,AnalysisState & state)814 static void equivalenceAnalysis(Operation *op,
815 BufferizationAliasInfo &aliasInfo,
816 AnalysisState &state) {
817 // Traverse ops in PostOrder: Nested ops first, then enclosing ops.
818 SmallVector<Operation *> ops;
819 op->walk<WalkOrder::PostOrder>([&](Operation *op) {
820 // No tensors => no buffers.
821 if (none_of(op->getResultTypes(), isaTensor))
822 return;
823 ops.push_back(op);
824 });
825
826 equivalenceAnalysis(ops, aliasInfo, state);
827 }
828
829 /// Assert that the current bufferization decisions are consistent.
830 static LogicalResult
checkAliasInfoConsistency(Operation * op,const DominanceInfo & domInfo,AnalysisState & state,const BufferizationAliasInfo & aliasInfo)831 checkAliasInfoConsistency(Operation *op, const DominanceInfo &domInfo,
832 AnalysisState &state,
833 const BufferizationAliasInfo &aliasInfo) {
834 const BufferizationOptions &options = state.getOptions();
835 Operation *inconsistentOp = nullptr;
836 WalkResult walkResult = op->walk([&](Operation *op) {
837 if (auto bufferizableOp = options.dynCastBufferizableOp(op))
838 for (OpOperand &opOperand : op->getOpOperands())
839 if (opOperand.get().getType().isa<TensorType>()) {
840 if (wouldCreateReadAfterWriteInterference(
841 opOperand, domInfo, state, aliasInfo,
842 /*checkConsistencyOnly=*/true)) {
843 // This error can happen if certain "mustBufferizeInPlace" interface
844 // methods are implemented incorrectly, such that the IR already has
845 // a RaW conflict before making any bufferization decisions.
846 inconsistentOp = op;
847 return WalkResult::interrupt();
848 }
849 }
850 return WalkResult::advance();
851 });
852
853 if (walkResult.wasInterrupted())
854 return inconsistentOp->emitError("input IR has RaW conflict");
855 return success();
856 }
857
858 /// Annotate the IR with the result of the analysis. For testing/debugging only.
859 static void
annotateOpsWithBufferizationMarkers(Operation * op,const BufferizationAliasInfo & aliasInfo,AnalysisState & state)860 annotateOpsWithBufferizationMarkers(Operation *op,
861 const BufferizationAliasInfo &aliasInfo,
862 AnalysisState &state) {
863 op->walk([&](Operation *op) {
864 if (auto bufferizableOp = state.getOptions().dynCastBufferizableOp(op))
865 for (OpOperand &opOperand : op->getOpOperands())
866 if (opOperand.get().getType().isa<TensorType>())
867 setInPlaceOpOperand(opOperand, aliasInfo.isInPlace(opOperand));
868 });
869 }
870
871 /// Assert that IR is in destination-passing style. I.e., every value that is
872 /// returned or yielded from a block is:
873 /// * aliasing a bbArg of that block or a parent block, or
874 /// * aliasing an OpResult of a op in a parent block.
875 ///
876 /// Example:
877 /// ```
878 /// %0 = "some_op" : tensor<?xf32>
879 /// %1 = scf.if %c -> (tensor<?xf32>) {
880 /// scf.yield %0 : tensor<?xf32>
881 /// } else {
882 /// %t = linalg.alloc_tensor : tensor<?xf32>
883 /// scf.yield %t : tensor<?xf32>
884 /// }
885 /// ```
886 /// In the above example, the first scf.yield op satifies destination-passing
887 /// style because the yielded value %0 is defined in the parent block. The
888 /// second scf.yield op does not satisfy destination-passing style because the
889 /// yielded value %t is defined in the same block as the scf.yield op.
890 // TODO: The current implementation checks for equivalent values instead of
891 // aliasing values, which is stricter than needed. We can currently not check
892 // for aliasing values because the analysis is a maybe-alias analysis and we
893 // need a must-alias analysis here.
894 static LogicalResult
assertDestinationPassingStyle(Operation * op,AnalysisState & state,BufferizationAliasInfo & aliasInfo,SmallVector<Operation * > & newOps)895 assertDestinationPassingStyle(Operation *op, AnalysisState &state,
896 BufferizationAliasInfo &aliasInfo,
897 SmallVector<Operation *> &newOps) {
898 LogicalResult status = success();
899 DominanceInfo domInfo(op);
900 op->walk([&](Operation *returnOp) {
901 if (!isRegionReturnLike(returnOp) ||
902 !state.getOptions().isOpAllowed(returnOp))
903 return WalkResult::advance();
904
905 for (OpOperand &returnValOperand : returnOp->getOpOperands()) {
906 Value returnVal = returnValOperand.get();
907 // Skip non-tensor values.
908 if (!returnVal.getType().isa<TensorType>())
909 continue;
910
911 bool foundEquivValue = false;
912 aliasInfo.applyOnEquivalenceClass(returnVal, [&](Value equivVal) {
913 if (auto bbArg = equivVal.dyn_cast<BlockArgument>()) {
914 Operation *definingOp = bbArg.getOwner()->getParentOp();
915 if (definingOp->isProperAncestor(returnOp))
916 foundEquivValue = true;
917 return;
918 }
919
920 Operation *definingOp = equivVal.getDefiningOp();
921 if (definingOp->getBlock()->findAncestorOpInBlock(
922 *returnOp->getParentOp()))
923 // Skip ops that happen after `returnOp` and parent ops.
924 if (happensBefore(definingOp, returnOp, domInfo))
925 foundEquivValue = true;
926 });
927
928 if (!foundEquivValue)
929 status =
930 returnOp->emitError()
931 << "operand #" << returnValOperand.getOperandNumber()
932 << " of ReturnLike op does not satisfy destination passing style";
933 }
934
935 return WalkResult::advance();
936 });
937
938 return status;
939 }
940
analyzeOp(Operation * op,OneShotAnalysisState & state)941 LogicalResult bufferization::analyzeOp(Operation *op,
942 OneShotAnalysisState &state) {
943 DominanceInfo domInfo(op);
944 BufferizationAliasInfo &aliasInfo = state.getAliasInfo();
945 const auto &options =
946 static_cast<const OneShotBufferizationOptions &>(state.getOptions());
947
948 // Catch incorrect API usage.
949 assert((state.hasDialectState(func::FuncDialect::getDialectNamespace()) ||
950 !options.bufferizeFunctionBoundaries) &&
951 "must use ModuleBufferize to bufferize function boundaries");
952
953 if (failed(checkAliasInfoConsistency(op, domInfo, state, aliasInfo)))
954 return failure();
955
956 // If the analysis fails, just return.
957 if (failed(inPlaceAnalysis(op, aliasInfo, state, domInfo,
958 options.analysisFuzzerSeed)))
959 return failure();
960 equivalenceAnalysis(op, aliasInfo, state);
961
962 bool failedAnalysis = false;
963 if (!options.allowReturnAllocs) {
964 SmallVector<Operation *> newOps;
965 failedAnalysis |=
966 failed(assertDestinationPassingStyle(op, state, aliasInfo, newOps));
967 }
968
969 // Gather some extra analysis data.
970 state.gatherYieldedTensors(op);
971 state.gatherUndefinedTensorUses(op);
972
973 // Analysis verification: After setting up alias/equivalence sets, each op
974 // can check for expected invariants/limitations and fail the analysis if
975 // necessary.
976 op->walk([&](Operation *op) {
977 if (BufferizableOpInterface bufferizableOp =
978 options.dynCastBufferizableOp(op))
979 failedAnalysis |= failed(bufferizableOp.verifyAnalysis(state));
980 });
981
982 // Annotate operations if we only want to report the analysis.
983 if (options.testAnalysisOnly)
984 annotateOpsWithBufferizationMarkers(op, aliasInfo, state);
985
986 return success(!failedAnalysis);
987 }
988
989 LogicalResult
runOneShotBufferize(Operation * op,const OneShotBufferizationOptions & options)990 bufferization::runOneShotBufferize(Operation *op,
991 const OneShotBufferizationOptions &options) {
992 OneShotAnalysisState state(op, options);
993 if (failed(insertTensorCopies(op, options)))
994 return failure();
995 if (options.testAnalysisOnly)
996 return success();
997 return bufferizeOp(op, options, /*copyBeforeWrite=*/false);
998 }
999