llvm-project/mlir/lib/Transforms/Utils/GreedyPatternRewriteDriver.cpp
Chris Lattner 967d934180 Fix two issues:
1) We incorrectly reassociated non-reassociative operations like subi, causing
    miscompilations.
 2) When constant folding, we didn't add users of the new constant back to the
    worklist for reprocessing, causing us to miss some cases (pointed out by
    Uday).

The code for tensorflow/mlir#2 is gross, but I'll add the new APIs in a followup patch.

PiperOrigin-RevId: 218803984
2019-03-29 13:40:35 -07:00

365 lines
13 KiB
C++

//===- GreedyPatternRewriteDriver.cpp - A greedy rewriter -----------------===//
//
// Copyright 2019 The MLIR Authors.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
// =============================================================================
//
// This file implements mlir::applyPatternsGreedily.
//
//===----------------------------------------------------------------------===//
#include "mlir/IR/Builders.h"
#include "mlir/IR/BuiltinOps.h"
#include "mlir/IR/PatternMatch.h"
#include "llvm/ADT/DenseMap.h"
using namespace mlir;
namespace {
class WorklistRewriter;
/// This is a worklist-driven driver for the PatternMatcher, which repeatedly
/// applies the locally optimal patterns in a roughly "bottom up" way.
class GreedyPatternRewriteDriver {
public:
explicit GreedyPatternRewriteDriver(OwningPatternList &&patterns)
: matcher(std::move(patterns)) {
worklist.reserve(64);
}
void simplifyFunction(Function *currentFunction, WorklistRewriter &rewriter);
void addToWorklist(Operation *op) {
worklistMap[op] = worklist.size();
worklist.push_back(op);
}
Operation *popFromWorklist() {
auto *op = worklist.back();
worklist.pop_back();
// This operation is no longer in the worklist, keep worklistMap up to date.
if (op)
worklistMap.erase(op);
return op;
}
/// If the specified operation is in the worklist, remove it. If not, this is
/// a no-op.
void removeFromWorklist(Operation *op) {
auto it = worklistMap.find(op);
if (it != worklistMap.end()) {
assert(worklist[it->second] == op && "malformed worklist data structure");
worklist[it->second] = nullptr;
}
}
private:
/// The low-level pattern matcher.
PatternMatcher matcher;
/// The worklist for this transformation keeps track of the operations that
/// need to be revisited, plus their index in the worklist. This allows us to
/// efficiently remove operations from the worklist when they are removed even
/// if they aren't the root of a pattern.
std::vector<Operation *> worklist;
DenseMap<Operation *, unsigned> worklistMap;
/// As part of canonicalization, we move constants to the top of the entry
/// block of the current function and de-duplicate them. This keeps track of
/// constants we have done this for.
DenseMap<std::pair<Attribute, Type *>, Operation *> uniquedConstants;
};
}; // end anonymous namespace
/// This is a listener object that updates our worklists and other data
/// structures in response to operations being added and removed.
namespace {
class WorklistRewriter : public PatternRewriter {
public:
WorklistRewriter(GreedyPatternRewriteDriver &driver, MLIRContext *context)
: PatternRewriter(context), driver(driver) {}
virtual void setInsertionPoint(Operation *op) = 0;
// If an operation is about to be removed, make sure it is not in our
// worklist anymore because we'd get dangling references to it.
void notifyOperationRemoved(Operation *op) override {
driver.removeFromWorklist(op);
}
GreedyPatternRewriteDriver &driver;
};
} // end anonymous namespace
void GreedyPatternRewriteDriver::simplifyFunction(Function *currentFunction,
WorklistRewriter &rewriter) {
// These are scratch vectors used in the constant folding loop below.
SmallVector<Attribute, 8> operandConstants, resultConstants;
while (!worklist.empty()) {
auto *op = popFromWorklist();
// Nulls get added to the worklist when operations are removed, ignore them.
if (op == nullptr)
continue;
// If we have a constant op, unique it into the entry block.
if (auto constant = op->dyn_cast<ConstantOp>()) {
// If this constant is dead, remove it, being careful to keep
// uniquedConstants up to date.
if (constant->use_empty()) {
auto it =
uniquedConstants.find({constant->getValue(), constant->getType()});
if (it != uniquedConstants.end() && it->second == op)
uniquedConstants.erase(it);
constant->erase();
continue;
}
// Check to see if we already have a constant with this type and value:
auto &entry = uniquedConstants[std::make_pair(constant->getValue(),
constant->getType())];
if (entry) {
// If this constant is already our uniqued one, then leave it alone.
if (entry == op)
continue;
// Otherwise replace this redundant constant with the uniqued one. We
// know this is safe because we move constants to the top of the
// function when they are uniqued, so we know they dominate all uses.
constant->replaceAllUsesWith(entry->getResult(0));
constant->erase();
continue;
}
// If we have no entry, then we should unique this constant as the
// canonical version. To ensure safe dominance, move the operation to the
// top of the function.
entry = op;
// TODO: If we make terminators into Operations then we could turn this
// into a nice Operation::moveBefore(Operation*) method. We just need the
// guarantee that a block is non-empty.
if (auto *cfgFunc = dyn_cast<CFGFunction>(currentFunction)) {
auto &entryBB = cfgFunc->front();
cast<OperationInst>(op)->moveBefore(&entryBB, entryBB.begin());
} else {
auto *mlFunc = cast<MLFunction>(currentFunction);
cast<OperationStmt>(op)->moveBefore(mlFunc, mlFunc->begin());
}
continue;
}
// If the operation has no side effects, and no users, then it is trivially
// dead - remove it.
if (op->hasNoSideEffect() && op->use_empty()) {
op->erase();
continue;
}
// Check to see if any operands to the instruction is constant and whether
// the operation knows how to constant fold itself.
operandConstants.clear();
for (auto *operand : op->getOperands()) {
Attribute operandCst;
if (auto *operandOp = operand->getDefiningOperation()) {
if (auto operandConstantOp = operandOp->dyn_cast<ConstantOp>())
operandCst = operandConstantOp->getValue();
}
operandConstants.push_back(operandCst);
}
// If constant folding was successful, create the result constants, RAUW the
// operation and remove it.
resultConstants.clear();
if (!op->constantFold(operandConstants, resultConstants)) {
rewriter.setInsertionPoint(op);
for (unsigned i = 0, e = op->getNumResults(); i != e; ++i) {
auto *res = op->getResult(i);
if (res->use_empty()) // ignore dead uses.
continue;
// If we already have a canonicalized version of this constant, just
// reuse it. Otherwise create a new one.
SSAValue *cstValue;
auto it = uniquedConstants.find({resultConstants[i], res->getType()});
if (it != uniquedConstants.end())
cstValue = it->second->getResult(0);
else
cstValue = rewriter.create<ConstantOp>(
op->getLoc(), resultConstants[i], res->getType());
// Add all the users of the result to the worklist so we make sure to
// revisit them.
//
// TODO: This is super gross. SSAValue use iterators should have an
// "owner" that can be downcasted to operation and other things. This
// will require a rejiggering of the class hierarchies.
if (auto *stmt = dyn_cast<OperationStmt>(op)) {
// TODO: Add a result->getUsers() iterator.
for (auto &operand : stmt->getResult(i)->getUses()) {
if (auto *op = dyn_cast<OperationStmt>(operand.getOwner()))
addToWorklist(op);
}
} else {
auto *inst = cast<OperationInst>(op);
// TODO: Add a result->getUsers() iterator.
for (auto &operand : inst->getResult(i)->getUses()) {
if (auto *op = dyn_cast<OperationInst>(operand.getOwner()))
addToWorklist(op);
}
}
res->replaceAllUsesWith(cstValue);
}
assert(op->hasNoSideEffect() && "Constant folded op with side effects?");
op->erase();
continue;
}
// If this is a commutative binary operation with a constant on the left
// side move it to the right side.
if (operandConstants.size() == 2 && operandConstants[0] &&
!operandConstants[1] && op->isCommutative()) {
auto *newLHS = op->getOperand(1);
op->setOperand(1, op->getOperand(0));
op->setOperand(0, newLHS);
}
// Check to see if we have any patterns that match this node.
auto match = matcher.findMatch(op);
if (!match.first)
continue;
// Make sure that any new operations are inserted at this point.
rewriter.setInsertionPoint(op);
match.first->rewrite(op, std::move(match.second), rewriter);
}
uniquedConstants.clear();
}
static void processMLFunction(MLFunction *fn, OwningPatternList &&patterns) {
class MLFuncRewriter : public WorklistRewriter {
public:
MLFuncRewriter(GreedyPatternRewriteDriver &driver, MLFuncBuilder &builder)
: WorklistRewriter(driver, builder.getContext()), builder(builder) {}
// Implement the hook for creating operations, and make sure that newly
// created ops are added to the worklist for processing.
Operation *createOperation(const OperationState &state) override {
auto *result = builder.createOperation(state);
driver.addToWorklist(result);
return result;
}
// When the root of a pattern is about to be replaced, it can trigger
// simplifications to its users - make sure to add them to the worklist
// before the root is changed.
void notifyRootReplaced(Operation *op) override {
auto *opStmt = cast<OperationStmt>(op);
for (auto *result : opStmt->getResults())
// TODO: Add a result->getUsers() iterator.
for (auto &user : result->getUses()) {
if (auto *op = dyn_cast<OperationStmt>(user.getOwner()))
driver.addToWorklist(op);
}
// TODO: Walk the operand list dropping them as we go. If any of them
// drop to zero uses, then add them to the worklist to allow them to be
// deleted as dead.
}
void setInsertionPoint(Operation *op) override {
// Any new operations should be added before this statement.
builder.setInsertionPoint(cast<OperationStmt>(op));
}
private:
MLFuncBuilder &builder;
};
GreedyPatternRewriteDriver driver(std::move(patterns));
fn->walk([&](OperationStmt *stmt) { driver.addToWorklist(stmt); });
MLFuncBuilder mlBuilder(fn);
MLFuncRewriter rewriter(driver, mlBuilder);
driver.simplifyFunction(fn, rewriter);
}
static void processCFGFunction(CFGFunction *fn, OwningPatternList &&patterns) {
class CFGFuncRewriter : public WorklistRewriter {
public:
CFGFuncRewriter(GreedyPatternRewriteDriver &driver, CFGFuncBuilder &builder)
: WorklistRewriter(driver, builder.getContext()), builder(builder) {}
// Implement the hook for creating operations, and make sure that newly
// created ops are added to the worklist for processing.
Operation *createOperation(const OperationState &state) override {
auto *result = builder.createOperation(state);
driver.addToWorklist(result);
return result;
}
// When the root of a pattern is about to be replaced, it can trigger
// simplifications to its users - make sure to add them to the worklist
// before the root is changed.
void notifyRootReplaced(Operation *op) override {
auto *opStmt = cast<OperationInst>(op);
for (auto *result : opStmt->getResults())
// TODO: Add a result->getUsers() iterator.
for (auto &user : result->getUses()) {
if (auto *op = dyn_cast<OperationInst>(user.getOwner()))
driver.addToWorklist(op);
}
// TODO: Walk the operand list dropping them as we go. If any of them
// drop to zero uses, then add them to the worklist to allow them to be
// deleted as dead.
}
void setInsertionPoint(Operation *op) override {
// Any new operations should be added before this instruction.
builder.setInsertionPoint(cast<OperationInst>(op));
}
private:
CFGFuncBuilder &builder;
};
GreedyPatternRewriteDriver driver(std::move(patterns));
for (auto &bb : *fn)
for (auto &op : bb)
driver.addToWorklist(&op);
CFGFuncBuilder cfgBuilder(fn);
CFGFuncRewriter rewriter(driver, cfgBuilder);
driver.simplifyFunction(fn, rewriter);
}
/// Rewrite the specified function by repeatedly applying the highest benefit
/// patterns in a greedy work-list driven manner.
///
void mlir::applyPatternsGreedily(Function *fn, OwningPatternList &&patterns) {
if (auto *cfg = dyn_cast<CFGFunction>(fn)) {
processCFGFunction(cfg, std::move(patterns));
} else {
processMLFunction(cast<MLFunction>(fn), std::move(patterns));
}
}