Attached is the second set of patches for the tracked resource memory management system.
Features: Separate memory pools for buffers, normal strings, constant strings; the latter is never compacted Each pool has its own alignment; these are currently set to 16, 4 and 2 respectively; suggestions welcome The current default allocations for these pools are: 16K, 32K and 8K respectively. Separate header pools for normal vs constant strings; the latter is never subjected to dead object detection String and Buffer processing combined as much as possible, since a string is really just a specialised buffer Most of the compile-time constants have been moved to variables; this will allow for dynamic changes to parameters API changes: I decided not to add new functions for allocating constant string headers and buffers. Instead, the constant flag must be set in the flags passed to string_make; these flags are now passed to new_string_header, and it will use the constant pool if requested. Currently, this only affects packfile.c, the patch for which is included herein. Note that changing the value of the constant flag during the life of a string could be detrimental to something. Other changes: A new opcode 'stringinfo' has been added to core.ops to allow access to the innards of a STRING, the list of arguments can be found in string.h. This currently serves very little purpose, but it will be useful for writing tests for COW, when this is added (shortly?) The code for 'open' in io.ops has been amended to use string_to_cstring instead of its own null-termination code; this code has mortality issues if termination of the mode string triggers GC; I have not bothered to handle that, as the correct solution is to pass Parrot strings directly to PIO_open. Still to do: The current handling of the free pool buffers in compact_buffer_pool is messy, and introduces another place that code needs to be added if a new resource type is added. I am currently thinking along the lines of some sort of 'system buffer' list, containing pointers to any buffer headers that are not in the normal arenas, but still need to be included in the collection process. COW will be added within a day or two; initially, just for full strings, as this doesn't need changes to anything outside string/resources String mortality still needs further work (e.g. the temporary transcoded strings created in string_concat, etc) Statistics collection may be moved to the pool level rather than the interpreter level, for possible use in dynamic tuning Documentation! Performance: This version is slightly faster in some cases; however, this is due to the increased memory allocation of the multiple pools, rather than any inherent improvements. -- Peter Gibbs EmKel Systems
phase2.patch
Description: Binary data
