#include "mlir/Dialect/Bufferization/Transforms/OneShotModuleBufferize.h"
#include "mlir/Dialect/Bufferization/IR/BufferizableOpInterface.h"
#include "mlir/Dialect/Bufferization/IR/Bufferization.h"
#include "mlir/Dialect/Bufferization/Transforms/Bufferize.h"
#include "mlir/Dialect/Bufferization/Transforms/FuncBufferizableOpInterfaceImpl.h"
#include "mlir/Dialect/Bufferization/Transforms/OneShotAnalysis.h"
#include "mlir/Dialect/Bufferization/Transforms/Transforms.h"
#include "mlir/Dialect/Func/IR/FuncOps.h"
#include "mlir/Dialect/MemRef/IR/MemRef.h"
#include "mlir/IR/BuiltinTypes.h"
#include "mlir/IR/Operation.h"
using namespace mlir;
using namespace mlir::bufferization;
using namespace mlir::bufferization::func_ext;
using FuncCallerMap = DenseMap<func::FuncOp, DenseSet<Operation *>>;
static FuncAnalysisState &
getOrCreateFuncAnalysisState(OneShotAnalysisState &state) {
auto *result = state.getExtension<FuncAnalysisState>();
if (result)
return *result;
return state.addExtension<FuncAnalysisState>();
}
static func::ReturnOp getAssumedUniqueReturnOp(func::FuncOp funcOp) {
func::ReturnOp returnOp;
for (Block &b : funcOp.getBody()) {
if (auto candidateOp = dyn_cast<func::ReturnOp>(b.getTerminator())) {
if (returnOp)
return nullptr;
returnOp = candidateOp;
}
}
return returnOp;
}
namespace {
static void annotateEquivalentReturnBbArg(OpOperand &returnVal,
BlockArgument bbArg) {
const char *kEquivalentArgsAttr = "__equivalent_func_args__";
Operation *op = returnVal.getOwner();
SmallVector<int64_t> equivBbArgs;
if (op->hasAttr(kEquivalentArgsAttr)) {
auto attr = cast<ArrayAttr>(op->getAttr(kEquivalentArgsAttr));
equivBbArgs = llvm::to_vector<4>(llvm::map_range(attr, [](Attribute a) {
return cast<IntegerAttr>(a).getValue().getSExtValue();
}));
} else {
equivBbArgs.append(op->getNumOperands(), -1);
}
equivBbArgs[returnVal.getOperandNumber()] = bbArg.getArgNumber();
OpBuilder b(op->getContext());
op->setAttr(kEquivalentArgsAttr, b.getI64ArrayAttr(equivBbArgs));
}
static LogicalResult
aliasingFuncOpBBArgsAnalysis(FuncOp funcOp, OneShotAnalysisState &state,
FuncAnalysisState &funcState) {
if (funcOp.getBody().empty()) {
FunctionType type = funcOp.getFunctionType();
for (const auto &inputIt : llvm::enumerate(type.getInputs())) {
if (!isa<TensorType>(inputIt.value()))
continue;
for (const auto &resultIt : llvm::enumerate(type.getResults())) {
if (!isa<TensorType>(resultIt.value()))
continue;
int64_t returnIdx = resultIt.index();
int64_t bbArgIdx = inputIt.index();
funcState.aliasingReturnVals[funcOp][bbArgIdx].push_back(returnIdx);
}
}
return success();
}
func::ReturnOp returnOp = getAssumedUniqueReturnOp(funcOp);
assert(returnOp && "expected func with single return op");
for (OpOperand &returnVal : returnOp->getOpOperands())
if (isa<RankedTensorType>(returnVal.get().getType()))
for (BlockArgument bbArg : funcOp.getArguments())
if (isa<RankedTensorType>(bbArg.getType())) {
int64_t returnIdx = returnVal.getOperandNumber();
int64_t bbArgIdx = bbArg.getArgNumber();
if (state.areEquivalentBufferizedValues(returnVal.get(), bbArg)) {
funcState.equivalentFuncArgs[funcOp][returnIdx] = bbArgIdx;
if (state.getOptions().testAnalysisOnly)
annotateEquivalentReturnBbArg(returnVal, bbArg);
}
if (state.areAliasingBufferizedValues(returnVal.get(), bbArg))
funcState.aliasingReturnVals[funcOp][bbArgIdx].push_back(returnIdx);
}
return success();
}
static void annotateFuncArgAccess(func::FuncOp funcOp, int64_t idx, bool isRead,
bool isWritten) {
OpBuilder b(funcOp.getContext());
Attribute accessType;
if (isRead && isWritten) {
accessType = b.getStringAttr("read-write");
} else if (isRead) {
accessType = b.getStringAttr("read");
} else if (isWritten) {
accessType = b.getStringAttr("write");
} else {
accessType = b.getStringAttr("none");
}
funcOp.setArgAttr(idx, BufferizationDialect::kBufferAccessAttrName,
accessType);
}
static LogicalResult
funcOpBbArgReadWriteAnalysis(FuncOp funcOp, OneShotAnalysisState &state,
FuncAnalysisState &funcState) {
for (int64_t idx = 0, e = funcOp.getFunctionType().getNumInputs(); idx < e;
++idx) {
if (!isa<TensorType>(funcOp.getFunctionType().getInput(idx)))
continue;
bool isRead;
bool isWritten;
if (auto accessAttr = funcOp.getArgAttrOfType<StringAttr>(
idx, BufferizationDialect::kBufferAccessAttrName)) {
StringRef str = accessAttr.getValue();
isRead = str == "read" || str == "read-write";
isWritten = str == "write" || str == "read-write";
} else if (funcOp.getBody().empty()) {
isRead = true;
isWritten = true;
} else {
BlockArgument bbArg = funcOp.getArgument(idx);
isRead = state.isValueRead(bbArg);
isWritten = state.isValueWritten(bbArg);
}
if (state.getOptions().testAnalysisOnly)
annotateFuncArgAccess(funcOp, idx, isRead, isWritten);
if (isRead)
funcState.readBbArgs[funcOp].insert(idx);
if (isWritten)
funcState.writtenBbArgs[funcOp].insert(idx);
}
return success();
}
}
static void removeBufferizationAttributes(BlockArgument bbArg) {
auto funcOp = cast<func::FuncOp>(bbArg.getOwner()->getParentOp());
funcOp.removeArgAttr(bbArg.getArgNumber(),
BufferizationDialect::kBufferLayoutAttrName);
funcOp.removeArgAttr(bbArg.getArgNumber(),
BufferizationDialect::kWritableAttrName);
}
static func::FuncOp getCalledFunction(func::CallOp callOp) {
SymbolRefAttr sym =
llvm::dyn_cast_if_present<SymbolRefAttr>(callOp.getCallableForCallee());
if (!sym)
return nullptr;
return dyn_cast_or_null<func::FuncOp>(
SymbolTable::lookupNearestSymbolFrom(callOp, sym));
}
static void equivalenceAnalysis(func::FuncOp funcOp,
OneShotAnalysisState &state,
FuncAnalysisState &funcState) {
funcOp->walk([&](func::CallOp callOp) {
func::FuncOp calledFunction = getCalledFunction(callOp);
assert(calledFunction && "could not retrieved called func::FuncOp");
if (!funcState.equivalentFuncArgs.count(calledFunction))
return WalkResult::skip();
for (auto it : funcState.equivalentFuncArgs[calledFunction]) {
int64_t returnIdx = it.first;
int64_t bbargIdx = it.second;
if (!state.isInPlace(callOp->getOpOperand(bbargIdx)))
continue;
Value returnVal = callOp.getResult(returnIdx);
Value argVal = callOp->getOperand(bbargIdx);
state.unionEquivalenceClasses(returnVal, argVal);
}
return WalkResult::advance();
});
}
static bool hasTensorSignature(func::FuncOp funcOp) {
return llvm::any_of(funcOp.getFunctionType().getInputs(),
llvm::IsaPred<TensorType>) ||
llvm::any_of(funcOp.getFunctionType().getResults(),
llvm::IsaPred<TensorType>);
}
static LogicalResult
getFuncOpsOrderedByCalls(ModuleOp moduleOp,
SmallVectorImpl<func::FuncOp> &orderedFuncOps,
FuncCallerMap &callerMap) {
DenseMap<func::FuncOp, DenseSet<func::FuncOp>> calledBy;
DenseMap<func::FuncOp, unsigned> numberCallOpsContainedInFuncOp;
WalkResult res = moduleOp.walk([&](func::FuncOp funcOp) -> WalkResult {
if (!funcOp.getBody().empty()) {
func::ReturnOp returnOp = getAssumedUniqueReturnOp(funcOp);
if (!returnOp)
return funcOp->emitError()
<< "cannot bufferize a FuncOp with tensors and "
"without a unique ReturnOp";
}
numberCallOpsContainedInFuncOp[funcOp] = 0;
return funcOp.walk([&](func::CallOp callOp) -> WalkResult {
func::FuncOp calledFunction = getCalledFunction(callOp);
assert(calledFunction && "could not retrieved called func::FuncOp");
if (!hasTensorSignature(calledFunction))
return WalkResult::skip();
callerMap[calledFunction].insert(callOp);
if (calledBy[calledFunction].insert(funcOp).second) {
numberCallOpsContainedInFuncOp[funcOp]++;
}
return WalkResult::advance();
});
});
if (res.wasInterrupted())
return failure();
while (!numberCallOpsContainedInFuncOp.empty()) {
auto it = llvm::find_if(numberCallOpsContainedInFuncOp,
[](auto entry) { return entry.getSecond() == 0; });
if (it == numberCallOpsContainedInFuncOp.end())
return moduleOp.emitOpError(
"expected callgraph to be free of circular dependencies.");
orderedFuncOps.push_back(it->getFirst());
for (auto callee : calledBy[it->getFirst()])
numberCallOpsContainedInFuncOp[callee]--;
numberCallOpsContainedInFuncOp.erase(it);
}
return success();
}
static void foldMemRefCasts(func::FuncOp funcOp) {
if (funcOp.getBody().empty())
return;
func::ReturnOp returnOp = getAssumedUniqueReturnOp(funcOp);
SmallVector<Type> resultTypes;
for (OpOperand &operand : returnOp->getOpOperands()) {
if (auto castOp = operand.get().getDefiningOp<memref::CastOp>()) {
operand.set(castOp.getSource());
resultTypes.push_back(castOp.getSource().getType());
} else {
resultTypes.push_back(operand.get().getType());
}
}
auto newFuncType = FunctionType::get(
funcOp.getContext(), funcOp.getFunctionType().getInputs(), resultTypes);
funcOp.setType(newFuncType);
}
LogicalResult
mlir::bufferization::analyzeModuleOp(ModuleOp moduleOp,
OneShotAnalysisState &state,
BufferizationStatistics *statistics) {
assert(state.getOptions().bufferizeFunctionBoundaries &&
"expected that function boundary bufferization is activated");
FuncAnalysisState &funcState = getOrCreateFuncAnalysisState(state);
SmallVector<func::FuncOp> orderedFuncOps;
FuncCallerMap callerMap;
if (failed(getFuncOpsOrderedByCalls(moduleOp, orderedFuncOps, callerMap)))
return failure();
for (func::FuncOp funcOp : orderedFuncOps) {
if (!state.getOptions().isOpAllowed(funcOp))
continue;
funcState.startFunctionAnalysis(funcOp);
equivalenceAnalysis(funcOp, state, funcState);
if (failed(analyzeOp(funcOp, state, statistics)))
return failure();
if (failed(aliasingFuncOpBBArgsAnalysis(funcOp, state, funcState)) ||
failed(funcOpBbArgReadWriteAnalysis(funcOp, state, funcState)))
return failure();
funcState.analyzedFuncOps[funcOp] = FuncOpAnalysisState::Analyzed;
}
return success();
}
void mlir::bufferization::removeBufferizationAttributesInModule(
ModuleOp moduleOp) {
moduleOp.walk([&](func::FuncOp op) {
for (BlockArgument bbArg : op.getArguments())
removeBufferizationAttributes(bbArg);
});
}
LogicalResult mlir::bufferization::bufferizeModuleOp(
ModuleOp moduleOp, const OneShotBufferizationOptions &options,
BufferizationStatistics *statistics) {
assert(options.bufferizeFunctionBoundaries &&
"expected that function boundary bufferization is activated");
IRRewriter rewriter(moduleOp.getContext());
SmallVector<func::FuncOp> orderedFuncOps;
FuncCallerMap callerMap;
if (failed(getFuncOpsOrderedByCalls(moduleOp, orderedFuncOps, callerMap)))
return failure();
for (func::FuncOp funcOp : orderedFuncOps) {
if (llvm::is_contained(options.noAnalysisFuncFilter, funcOp.getSymName())) {
OneShotBufferizationOptions updatedOptions = options;
updatedOptions.copyBeforeWrite = true;
if (failed(bufferizeOp(funcOp, updatedOptions, statistics)))
return failure();
} else {
if (failed(bufferizeOp(funcOp, options, statistics)))
return failure();
}
if (options.inferFunctionResultLayout)
foldMemRefCasts(funcOp);
}
for (Operation &op : llvm::make_early_inc_range(moduleOp.getOps())) {
if (isa<func::FuncOp>(&op))
continue;
if (failed(bufferizeOp(&op, options, statistics)))
return failure();
}
removeBufferizationAttributesInModule(moduleOp);
return success();
}
LogicalResult mlir::bufferization::runOneShotModuleBufferize(
ModuleOp moduleOp, const OneShotBufferizationOptions &options,
BufferizationStatistics *statistics) {
assert(options.bufferizeFunctionBoundaries &&
"expected that function boundary bufferization is activated");
assert(!(options.copyBeforeWrite && options.testAnalysisOnly) &&
"invalid combination of bufferization flags");
if (!options.copyBeforeWrite) {
if (options.noAnalysisFuncFilter.empty()) {
if (failed(insertTensorCopies(moduleOp, options, statistics)))
return failure();
} else {
OpFilter::Entry::FilterFn analysisFilterFn = [=](Operation *op) {
auto func = dyn_cast<func::FuncOp>(op);
if (!func)
func = op->getParentOfType<func::FuncOp>();
if (func)
return llvm::is_contained(options.noAnalysisFuncFilter,
func.getSymName());
return false;
};
OneShotBufferizationOptions updatedOptions(options);
updatedOptions.opFilter.denyOperation(analysisFilterFn);
if (failed(insertTensorCopies(moduleOp, updatedOptions, statistics)))
return failure();
}
}
if (options.testAnalysisOnly)
return success();
if (failed(bufferizeModuleOp(moduleOp, options, statistics)))
return failure();
return success();
}