Hi again,
I have reworked and simplified this clean up/refactoring suggestion:
Updated high-level description:
http://cr.openjdk.java.net/~mgronlun/8140485/unified/unified_review/81
40485_updated.jpg
Updated webrev:
http://cr.openjdk.java.net/~mgronlun/8140485/unified/unified_review/we
brev02/
Updated patch:
http://cr.openjdk.java.net/~mgronlun/8140485/unified/unified_review/81
40485_open_unified_updated.patch
Thanks in advance
Markus
-----Original Message-----
From: Markus Gronlund
Sent: den 27 oktober 2015 13:21
To: hotspot-runtime-...@openjdk.java.net;
serviceability-...@openjdk.net
Subject: RFR(L): 8140485: Class load and creation clean up
Greetings,
I have spent some time refactoring and cleaning up parts of the class load and
creation sequence and code.
Bug: https://bugs.openjdk.java.net/browse/JDK-8140485
Overview:
There are quite a lot of changes here - hopefully it will not be a deterrent of taking this on- first of all,
a lot of changes are mechanical and applicable to "form" more than content- for example, variable,
parameter and member function "constness". Same applies to other "domino changes", for
example when moving an included header from a .hpp file (maybe replacing with a fwd decl instead).
I have tried to stage the changes for easier overview / review (including a
high level diagram). The patches/webrevs are split into a patch series of 7
patches (there is also a unified patch).
Use case:
My initial use case for starting all of this clean up / refactoring work is a
need to reuse mechanics of the class parsing and creation sequence.
Cleanup / refactoring objectives:
- Allow for class parsing implementation reuse
- Decouple class parsing and Klass creation
- Decouple JVMTI class file load hook from ClassFileParser
- Add compiler assisted read-only intent in order to assist future
refactoring attempts
- "componentify" ClassFileParser / apply to Klass/IK/AK creations
- Take advantage of some optimizations opportunities (constructors and
underlying "calloc" memory)
- Maintain class load performance
High level diagram explaining the updated suggested sequence (high level):
http://cr.openjdk.java.net/~mgronlun/8140485/8140485_Class_load_and_cr
eation_cleanup_high_level_sequence.jpg
Split webrevs/patches for easier overview / review (do note that not all of
these splits will compile individually - there is a unified patch link that
includes everything, pls see below):
Sequence: "first" (order of changes)
Name: "8140485_1_unified entry point"
Link:
http://cr.openjdk.java.net/~mgronlun/8140485/split/first/webrev01/
Summary:
Split out JVMTI agent CFLH from CFP - moved to
SD::create_vm_representation_prologue().
Channel all class loading via a unified entry point -
SD::create_vm_representation()
Split out ClassPathEntry into own header.
Creation of new class ClassPathFileStream (specialized ClassPathStream) to
retain ClassLoaderExt::Context functionality in the refactored code.
"Need verify" property is carried within the CFS itself.
Comment:
SystemDictionary::load_instance_class() will now have a call to
"load_class_from_classpath" - the relevance of this name might need to
be modified (modules etc?)
Sequence: "second"
Name: "8140485_2_read_only_class_file_stream"
Link:
http://cr.openjdk.java.net/~mgronlun/8140485/split/second/webrev01/
Summary:
The underlying bytestream (buffer) in a CFS representation should not be
altered (advertently or inadvertently) once created, made read-only.
Comment:
Modifications should work with copies.
Sequence: "third"
Name: "8140485_3_componentify_class_file_parser"
Link:
http://cr.openjdk.java.net/~mgronlun/8140485/split/third/webrev01/
Summary:
In order to allow for code refactoring of CFP - stack allocated variables are
modified into fields. The entire CFP acts as a generic RAII class for the
allocated structures (note CHECK macros).
I have not fulfilled refactoring of some of the longest methods since it's is
hard to really understand the impact of the inlining I guess we are after here..
The performance work done so far indicated improvements in still passing parameters
instead of reading the fields directly (so you might still see "fat" methods
taking a lot of parameters even though the fields are set).
"Consts everywhere.."
One of the real hard issues trying to refactor this code is the lack of
expressed read-only intent - it's hard to determine (both on reading but more
importantly when compiling) what can/should change.
Yes, I have entered the "const" rabbit hole related to this code (in some cases
it might even be termed gratuitous (event 'consting' function parameter pointers in some
cases for example)) - hopefully this will help in upcoming attempts to refactor/simplify
this code.
Accessors - CFP becomes a data container to be used in constructors
(see change seven). In addition, you will now have a possibility to
fetch data from a parser without creating a fullblown Klass
Comment:
I think the code could be improved and more refactoring could be
accomplished by introducing a Metaspace allocation RAII class - this
would allow the functions to decouple even more that is currently
(might be next steps)
Sequence: "fourth"
Name: "8140485_4_hide_impl_details_class_file_parser"
Link:
http://cr.openjdk.java.net/~mgronlun/8140485/split/fourth/webrev01/
Summary:
Reduce the visibility of CFP impl methods by internal linkage instead of
external linkage.
Comment:
In my opinion, we should attempt to break the pattern of having private functions
declared in header files when the private function does not need to reach inside
"this".
(that is, it can use an external interface alt. might not need access
at all (helpers))
Sequence: "fifth"
Name: "8140485_5_update_entry_point"
Link:
http://cr.openjdk.java.net/~mgronlun/8140485/split/fifth/webrev01/
Summary:
Update signatures. Remove the parameter passing where not needed
("parsed_name" reference for example)
Sequence: "sixth"
Name: "8140485_6_types_const_includes_repercussions"
Link:
http://cr.openjdk.java.net/~mgronlun/8140485/split/sixth/webrev01/
Summary:
Rippling effects of stronger constness.
Forward includes where can be applied. This revealed other code modules not
having includes -added.
Downcasting of const qualified types (applied Scott Meyers idiom of
dual const_cast additions (see Effective C++))
Sequence: "seventh"
Name: "8140485_7_applied_use_in_constructors_and_some_optimizations"
Link:
http://cr.openjdk.java.net/~mgronlun/8140485/split/seventh/webrev01/
Summary:
Take advantage of above modifications to gain direct upshots in other areas.
Use CFP as the data container when constructing Klass/IK/AK.
Optimizations: Reduce/optimize Klass/InstanceKlass/ArrayKlass/ConstantPool constructors
by using underlying invariant on "calloc" equivalent MetaspaceObj allocations.
Unified change/patch:
Webrev: http://cr.openjdk.java.net/~mgronlun/8140485/unified/webrev01/
Patch:
http://cr.openjdk.java.net/~mgronlun/8140485/unified/8140485_unified.p
atch
Summary:
Unified (folded) patch for easier application/imports, for
compilation/try out
Testing:
I have done targeted testing during development especially using the
"runtime/ParallelClassLoading" test suite in order to ensure
functional equivalency. JPRT testing. Performance tests (see below)
Performance:
Unstructured:
I have used Java Flight Recorder (JFR), which envelops the class loading
sequence with high resolution timers - I have used this information to ensure
performance has been stable regarding the changes.
Structured:
Startup/footprint-benchmarks (Linux32, Linux64, Windows32, Windows64,
Solaris SPARC-T)
It is hard to read anything decisive around these (the effects most
likely hides in noise), but at least determine that the changes do not
introduce anything statistically significant (+/-) (both time and
space)
Thanks in advance
Markus
