| 9bdf108885a27ba05fae8501725649574d7c491b |
|
21-Jan-2016 |
Vladimir Marko <vmarko@google.com> |
Revert "Revert "Write dex files to oat file early.""
This reverts commit 919f5536182890d2e03f59b961acf8f7c836ff61.
Fix linker error (Mac build):
Replace inline definition of art::ZipArchive::~ZipArchive()
with an out-of-line definition in zip_archive.cc to avoid
direct reference to CloseArchive() from libart-compiler due
to inlining. Note that libart is linked against -lziparchive
but libart-compiler is not.
Change-Id: I92620ea0200282ca7ba9b7f61a592cb6468d90d8
|
| 919f5536182890d2e03f59b961acf8f7c836ff61 |
|
20-Jan-2016 |
Vladimir Marko <vmarko@google.com> |
Revert "Write dex files to oat file early."
This reverts commit 625a64aad13905d8a2454bf3cc0e874487b110d5.
Breaks the Mac build:
Undefined symbols for architecture i386:
"_CloseArchive", referenced from:
... in oat_writer.o
ld: symbol(s) not found for architecture i386
Change-Id: I21608bc51437834e1e6abde9bcbe5e7d9998197e
|
| 625a64aad13905d8a2454bf3cc0e874487b110d5 |
|
26-Nov-2015 |
Vladimir Marko <vmarko@google.com> |
Write dex files to oat file early.
Write dex files to oat file before we actually open and
verify them. Instead, open and verify the copies from the
oat file and use these. This way, in the most common case
of zipped dex files, we have mmapped dex files instead of
inflated dex files. That reduces the number of dirty pages
used by dex2oat.
Reading /proc/self/statm after we write the oat file for
a compilation of a certain large app on Nexus 5 AOSP build
with -j1, three attempts before and after this CL gave
before: 346061 189462 6269 26 0 140723 0
346189 189450 6269 26 0 140851 0
346061 189463 6269 26 0 140723 0
after: 346186 185808 23040 27 0 140468 0
346186 185819 23040 27 0 140468 0
346186 185822 23040 27 0 140468 0
These values are in pages (4KiB), so while the "size"
(=VmSize) is essentially unchanged, the "resident" (=VmRSS)
is over 14MiB less and the "shared" (i.e. backed by a file)
is 65.5MiB more. That is, the amount of dirty non-pageable
memory used is reduced by about 80MiB.
The oat file format has changed slightly, the class offset
table has been moved from the OatDexFile to its own section.
This actually fixes the alignment of these offsets as they
could have been unaligned previously, yet accessed as normal
with significant performance impact if the kernel has to
emulate the unaligned access (say, mips).
Change-Id: I0f4799bb1f1ca28e3533156a3494f55345c3e10a
|
| d9786b0e5be23ea0258405165098b4216579209c |
|
14-Oct-2015 |
Artem Udovichenko <artem.u@samsung.com> |
Implementation of fast lookup table to search class_def by descriptor
Lookup table is a hash table which built at compile time and stored
into oat file. At runtime the table is restored and used in the
method DexFile::FindClassDef(const char*) to perform fast search of
the class_def_idx by class descriptor. Advantages of the lookup table
over the HashSet (runtime/base/hash_set.h) are:
1. Lookup table is built at compile time and uses read-only memory at
runtime
2. Lookup table uses less memory then DexFile::Index (less by 80% for
/system/framework/framework.jar on Nexus5)
3. Lookup table does less string comparisons compared with HashSet
(less by 70% for zygote process on Nexus5)
The disadvantage of the lookup table is it increased boot.oat size by
0.2% on Nexus5 and application .oat file by 0.3% in average on Nexus5.
mathieuc changes:
Create lookup table in dex2oat to speed up compilation. Clean up code
to follow style guide and use less static functions. Added
performance measurements.
Compile ~100 APKs 5 times with filter interpret-only:
Before:
real 1m8.989s
user 0m59.318s
sys 0m7.773s
After:
real 1m1.493s
user 0m52.055s
sys 0m7.581s
App launch (AOSP N5 maps, average of 45 runs):
Before: 966.84ms
After: 923.733ms
Launch speedup is 4.7%
Memory usage compared to HashSet index on 50 various APK:
32 bit: HashSet ~625694b vs TypeLookupTable ~404268b
64 bit: HashSet ~1251390b vs TypeLookupTable ~404268b
Bug: 10921004
Bug: 20269715
Change-Id: I7246c1d9ad9fe81fe5c5907a4bf70396d8f9242a
|