When recalculating the number of target eden size, correctly consider the amount of existing eden regions in the target calculation.
Reviewed-by: tschatzl
Push the remembered sets of large objects with few referenced into the dirty card queue at the beginning of the evacuation so that they may end up with zero remembered set entries at the end of the collection, and are potentially reclaimed. Also improve timing measurements of the early reclaim mechanism, and shorten flag names.
Reviewed-by: brutisso, jmasa, dfazunen
Some verification code iterated over the heap using the region mapping array. This is not allowed. Changed to use the regular iteration method with closure.
Reviewed-by: jwilhelm, brutisso
Evaluate and improve object copy time by micro-optimizations and splitting out slow and fast paths aggressively.
Co-authored-by: Tony Printezis <tprintezis@twitter.com>
Reviewed-by: kbarrett, mgerdin, jmasa
Changes in JDK-8038423 always initialize (zero out) virtual memory used for auxiliary data structures. This causes a footprint regression for G1 in startup benchmarks. This is because they do not touch that memory at all, so the operating system does not actually commit these pages. The fix is to, if the initialization value of the data structures matches the default value of just committed memory (=0), do not do anything.
Reviewed-by: jwilhelm, brutisso
Splitting out a g1 allocator class to simply specialized allocators which can associate each allocation with a given context.
Reviewed-by: mgerdin, brutisso
Replace G1CodeRootSet with a Hashtable based implementation, merge Code Root Migration phase into Code Root Scanning
Reviewed-by: jmasa, brutisso, tschatzl
Allow G1 to decommit memory of arbitrary regions within the heap and their associated auxiliary data structures card table, BOT, hot card cache, and mark bitmaps.
Reviewed-by: mgerdin, brutisso, jwilhelm
Let HeapRegionSeq manage the heap region and auxiliary data to decrease the amount of responsibilities of G1CollectedHeap, and encapsulate this work from other code.
Reviewed-by: jwilhelm, jmasa, mgerdin, brutisso
Refactor preparation for compaction during Full GC so that it lazily initializes the first compaction point. This also avoids problems later when the first region may not be committed. Also reviewed by K. Barrett.
Reviewed-by: brutisso
G1 does not retire allocation buffers after reference processing work when -XX:+ParallelRefProcEnabled is enabled. This causes wrong calculation of PLAB sizes, as the amount of space wasted is not updated correctly.
Reviewed-by: brutisso
Change 8035330: Remove G1ParScanPartialArrayClosure and G1ParScanHeapEvacClosure broke the debug build on AIX. The method do_oop_partial_array() is added in a header, but requires the inline function par_write_ref() through several inlined calls. In some cpp files, like arguments.cpp, par_write_ref() is not defined as the corresponding inline header and is not included. The AIX debug VM does not start because of the missing symbol. This change solves this by cleaning up include dependencies.
Reviewed-by: tschatzl, stefank
The free list is being maintained in a sorted fashion and old and humongous regions are allocated from the bottom of the heap while young regions are allocated at the top.
Co-authored-by: Staffan Friberg <staffan.friberg@oracle.com>
Reviewed-by: tschatzl, mgerdin
Mentioned closures are actually wrapped methods. This adds confusion to readers, and in this case also increases code size as G1ParScanHeapEvacClosure is part of the oop_oop_iterate() methods. Move them into G1ParScanThreadState as methods.
Reviewed-by: stefank
Change the code cache remembered sets data structure from a GrowableArray to a chunked list of nmethods. This makes the data structure more amenable to parallelization, and decreases freeing time.
Reviewed-by: mgerdin, brutisso
Revert JDK-6976350 because it does not improve PLAB fragmentation. To the contrary, it tends to increase the amount of wasted space with many threads.
Reviewed-by: brutisso
