chromium/v8/src/compiler/revectorizer.h

// Copyright 2022 the V8 project authors. All rights reserved.
// Use of this source code is governed by a BSD-style license that can be
// found in the LICENSE file.

#ifndef V8_COMPILER_REVECTORIZER_H_
#define V8_COMPILER_REVECTORIZER_H_

// Revectorizer is an optimization to promote pairs of simd128 nodes to new
// simd256 nodes accelerated by wider vector available from hardware e.g. the
// YMM registers from AVX2 instruction set when possible and beneficial. The
// main algorithm is based on the Superword Level Parallel (SLP) vectorization
// technique.

#include <vector>

#include "src/base/small-vector.h"
#include "src/compiler/graph.h"
#include "src/compiler/linear-scheduler.h"
#include "src/compiler/machine-graph.h"
#include "src/compiler/machine-operator.h"
#include "src/compiler/node-marker.h"
#include "src/compiler/node-properties.h"
#include "src/compiler/node.h"
#include "src/compiler/schedule.h"
#include "src/zone/zone-containers.h"

namespace v8 {
namespace internal {
namespace compiler {

class SourcePositionTable;

struct V8_EXPORT_PRIVATE MemoryOffsetComparer { … };

StoreNodeSet;

// A PackNode consists of a fixed number of isomorphic simd128 nodes which can
// execute in parallel and convert to a 256-bit simd node later. The nodes in a
// PackNode must satisfy that they can be scheduled in the same basic block and
// are mutually independent.
class PackNode final : public NON_EXPORTED_BASE(ZoneObject) { … };

// An auxillary tree structure with a set of PackNodes based on the Superword
// Level Parallelism (SLP) vectorization technique. The BuildTree method will
// start from a selected root, e.g. a group of consecutive stores, and extend
// through value inputs to create new PackNodes if the inputs are valid, or
// conclude that the current PackNode is a leaf and terminate the tree.
// Below is an example of SLPTree where loads and stores in each PackNode are
// all consecutive.
// [Load0, Load1]  [Load2, Load3]
//           \       /
//          [Add0, Add1]
//                |
//         [Store0, Store1]
class SLPTree : public NON_EXPORTED_BASE(ZoneObject) { … };

// The Revectorizer pass will firstly collect seeds with valid group of
// consecutive stores as the root to build the SLPTree. If the SLPTree is built
// successfully, it will estimate the cost of the 256-bit transformation for
// each PackNode and conduct the final revectorization if benefitial.
class V8_EXPORT_PRIVATE Revectorizer final
    : public NON_EXPORTED_BASE(ZoneObject) { … };

}  // namespace compiler
}  // namespace internal
}  // namespace v8

#endif  // V8_COMPILER_REVECTORIZER_H_