Skip to content

Latest commit

 

History

History
85 lines (69 loc) · 4.67 KB

CodeMetaData.md

File metadata and controls

85 lines (69 loc) · 4.67 KB
Raw

Code Metadata and Code Metadata Manager in the Eclipse OMR compiler

Introduction

This document explains the concepts and implementation of runtime code metadata and code metadata manager in the Eclipse OMR compiler.

Code Metadata

Code metadata contains runtime metadata information about a method. Metadata may include the starting address of allocated code memory and its size, the entry point address of compiled code (when it is invoked from either an interpreter or another piece of compiled code), the address where the compiled code ends, the compilation hotness level, a list of runtime assumptions made during the compilation, and any other information required by the runtime.

An instance of code metadata will be created for each method after it is compiled. The code metadata instance will exist for the duration of the method's runtime until it is unloaded or invalidated. Storage and lookup of code metadata instances are carried out by the code metadata manager.

In the Eclipse OMR compiler, an instance of code metadata is uniquely identified by its start and end PC addresses. Two instances of code metadata with PC address ranges that overlap are not allowed.

Code Metadata Manager

The code metadata manager is a singleton class created in the runtime to manage code metadata instances. The manager class constructor allocates an AVL tree in persistent memory to facilitate storage and lookup of code metadata instances created by the compiler. Each tree node in the AVL tree is a hash table representing a piece of code cache segment. A node is created and added to the AVL tree when a new code cache segment is allocated. For Eclipse OpenJ9, code cache segment size is configured to be 2MB. Entries in each hash table correspond to equal chunks of the associated code cache. For Eclipse OpenJ9, each hash table has 4096 entries and each entry is responsible for 512 bytes of the 2MB code cache. A hash table entry points to a metadata pointer or a list of metadata pointers whose compiled code occupies a part of the chunk or the entire chunk of the code cache.

For efficient insertion, deletion, and lookup of a particular metadata pointer, please note that to add or remove nodes from the AVL tree requires exclusive VM access, which means threads with VM access have to be stopped before the operation can be carried out. However, to add an entry to a hash table or to look up a metadata pointer given a PC address, only VM access is required, which means the operation can be carried out concurrently with other threads holding VM access. Removing a metadata pointer from a hash table in Eclipse OpenJ9 is done at the end of garbage collection when a method is unloaded or recompiled and reclaimed, and exclusive VM access is already obtained.

The following is the procedure for adding a metadata pointer with the start and end PC addresses of its compiled code:

  1. Find out which code cache the PC address range belongs to, i.e. which node in the AVL tree contains the PC address range;
  2. Calculate which chunks of the code cache correspond to the PC address range, i.e. what entries in the hash table will be updated;
  3. Add the metadata pointer to each hash table entry found in step #2. If the hash table entry has one metadata pointer, a list is created and prepended with the new one and the existing metadata pointer (with its low bit set) will terminate the list. If the entry is already a list of metadata pointers, either add the new metadata to the list if there is enough space or allocate space to copy the existing list and then add the new metadata pointer.