summaryrefslogtreecommitdiff
path: root/compiler/optimizing/stack_map_stream.h
AgeCommit message (Collapse)Author
2021-03-05Add code size to CodeInfo.David Srbecky
This is in preparation of removing it from OatQuickMethodHeader. Bug: 123510633 Test: m test-art-host-gtest Test: ./art/test.py -b -r --host Change-Id: I5c5adb4c040e329b81c1393aa1b80ee017729c8a
2019-12-03JIT baseline: trigger optimized compilation on hotness threshold.Nicolas Geoffray
- Add a new hotness count in the ProfilingInfo to not conflict with interpreter hotness which may use it for OSR. - Add a baseline flag in the OatQuickMethodHeader to identify baseline compiled methods. - Add a -Xusetieredjit flag to experiment and test. Bug: 119800099 Test: test.py with Xusetieredjit to true Change-Id: I8512853f869f1312e3edc60bf64413dee9143c52
2019-10-14Revert "Make compiler/optimizing/ symbols hidden."Vladimir Marko
This reverts commit e2727154f25e0db9a5bb92af494d8e47b181dfcf. Reason for revert: Breaks ASAN tests (ODR violation). Bug: 142365358 Change-Id: I38103d74a1297256c81d90872b6902ff1e9ef7a4
2019-10-14Make compiler/optimizing/ symbols hidden.Vladimir Marko
Make symbols in compiler/optimizing hidden by a namespace attribute. The unit intrinsic_objects.{h,cc} is excluded as it is needed by dex2oat. As the symbols are no longer exported, gtests are now linked with the static version of the libartd-compiler library. libart-compiler.so size: - before: arm: 2396152 arm64: 3345280 - after: arm: 2016176 (-371KiB, -15.9%) arm64: 2874480 (-460KiB, -14.1%) Test: m test-art-host-gtest Test: testrunner.py --host --optimizing --jit Bug: 142365358 Change-Id: I1fb04a33351f53f00b389a1642e81a68e40912a8
2019-07-15ART: Introduce stackmap with no DexRegInfo.Artem Serov
Some of safepoints don't need to have DexRegisterMap info; this will decrease the stackmap size. .oat file size reduction: - boot.oat: -233 kb (-5.4%) - boot-framework.oat: -704 kb (-4.9%) Test: 461-get-reference-vreg, 466-get-live-vreg. Test: 543-env-long-ref, 616-cha*. Test: test-art-target, +gc-stress. Change-Id: Idbad355770e30a30dcf14127642e03ee666878b8
2019-06-18Stack maps: Handle special cases using flags.David Srbecky
Keep the BitTable decoder simple (1+NumColumns varints). Move special case handling up to CodeInfo (empty/dedup). This speeds up CodeInfo by 5%, and maps startup by 0.05%. Change in size is negligible (the bits mostly just move). Test: test.py -b --host --64 --optimizing Change-Id: Ib6abe52f04384de9ffd7cfba04a3124b62f713ff
2019-05-26Optimize stack maps: add fast path for no inline info.David Srbecky
Consumers of CodeInfo can skip significant chunks of work if they can quickly determine that method has no inlining. Store this fact as a flag bit at the start of code info. This changes binary format and adds <0.1% to oat size. I added the extra flag field as the simplest solution for now, although I would like to use it for more things in the future. (e.g. store the special cases of empty/deduped tables in it) This improves app startup by 0.4% (maps,speed). PMD on golem seems to gets around 15% faster. Bug: 133257467 Test: ./art/test.py -b --host --64 Change-Id: Ia498a31bafc74b51cc95b8c70cf1da4b0e3d894e
2019-03-06Revert "Revert^2 "Add code size to CodeInfo""Nicolas Geoffray
This reverts commit e1412dacbf1d2a809bd1fca658cc8cb8f61f8ee6. Bug: 123510633 Bug: 127305289 Reason for revert: b/127305289 Change-Id: I54557b05a44777f1fa2c15bde4fa648980f42eed
2019-03-05Revert "Don't pack code size in CodeInfo."Nicolas Geoffray
This reverts commit e35ac04a1a9a22b1c4386b27f3a30cd840aa17b1. Bug: 123510633 Bug: 127305289 Reason for revert: b/127305289 Change-Id: I18c2d9291411b31641333c14c47da8c4fdf317f7
2019-02-27Don't pack code size in CodeInfo.David Srbecky
The unpacking is tricky for host tooling as we need to propagate ISA. This adds 0.05% to oat file size. Bug: 123510633 Change-Id: I5618db5e5dbe83d8a2bb89aef61cb0b10e336f40
2019-02-14Revert^2 "Add code size to CodeInfo"David Srbecky
This temporarily adds 0.25% to oat file size. The space will be reclaimed back in follow-up CL. This reverts commit 8f20a23a35fa6fbe4dcb4ff70268a24dc7fb2a24. Reason for revert: Reland as-is after CL/903819 Bug: 123510633 Test: DCHECK compare the two stored code sizes. Change-Id: Ia3ab31c208948f4996188764fcdcba13d9977d19
2019-02-12Revert "Add code size to CodeInfo"David Srbecky
This reverts commit 68efa7b1128486e08ae60cd27181645b27bbd2e4. Reason for revert: Breaks tests Change-Id: I28fb143990f58e0d5f0b106bea9d9a159f19297e
2019-02-11Add code size to CodeInfoDavid Srbecky
This temporarily adds 0.25% to oat file size. The space will be reclaimed back in follow-up CL. Bug: 123510633 Test: DCHECK compare the two stored code sizes. Change-Id: I15340824ca637fd075a4cef87771b06cb96bb9f4
2018-08-10Add helper method to iterate over BitTable fields in CodeInfo.David Srbecky
Avoid the repetitive code patterns and simplify code. Test: test-art-host-gtest-stack_map_test Test: checked output of oatdump Change-Id: I2354bc652837eb34efeecf4de56a027384544034
2018-08-03Simplify StackMapStream encoding.David Srbecky
Test: test-art-host-gtest-stack_map_test Change-Id: Ife021d03e4e486043ec609f9af8673ace7bde497
2018-08-02Merge "Reuse arena memory for each block in scheduler."Vladimir Marko
2018-08-02Reuse arena memory for each block in scheduler.Vladimir Marko
This reduces the peak memory used for large methods with multiple blocks to schedule. Compiling the aosp_taimen-userdebug boot image, the most memory hungry method BatteryStats.dumpLocked has the Scheduler memory allocations in ArenaStack hidden by the register allocator: - before: MEM: used: 8300224, allocated: 9175040, lost: 197360 Scheduler 8300224 - after: MEM: used: 5914296, allocated: 7864320, lost: 78200 SsaLiveness 5532840 RegAllocator 144968 RegAllocVldt 236488 The total arena memory used, including the ArenaAllocator not listed above, goes from 44333648 to 41950324 (-5.4%). (Measured with kArenaAllocatorCountAllocations=true, kArenaAllocatorPreciseTracking=false.) Also remove one unnecessary -Wframe-larger-than= workaround and add one workaround for large frame with the above arena alloc tracking flags. Test: m test-art-host-gtest Test: testrunner.py --host Bug: 34053922 Change-Id: I7fd8d90dcc13b184b1e5bd0bcac072388710a129
2018-08-02Encode frame info using varints.David Srbecky
This saves 0.3% of oat file size. Test: test-art-host-gtest-stack_map_test Change-Id: I85003946a9579f03cb1ed2b5e9b2c62b3efe6734
2018-08-01Move MethodInfo to CodeInfo.David Srbecky
There is no need to treat it specially any more, because of the de-duplication at BitTable level. This saves 0.6% of oat file size. Test: test-art-host-gtest Change-Id: Ife7927d736243879a41d6f325d49ebf6930a63f6
2018-07-13Remove unused InvokeInfo from stack maps and the runtime.David Srbecky
Test: test-art-host-gtest Change-Id: I5ce28973042f9241e72ceb52fc5db472ca571563
2018-06-26Add method frame info to CodeInfo.David Srbecky
The stored information will be used in follow-up CLs. This temporarily increases .oat file size by 0.7%. Test: test-art-host-gtest Change-Id: Ie7d898b06398ae44287bb1e8153861ab112a216c
2018-06-14Template BitTable based on the accessors.David Srbecky
Test: test-art-host-gtest-stack_map_test Test: test-art-host-gtest-bit_table_test Change-Id: I96c04e21864009b64cb3177a0e9f0f8782a9b10b
2018-06-14Merge "Move some helper methods to DexRegisterLocation."David Srbecky
2018-06-14Remove POD helper classes for BitTableBuilder.David Srbecky
Instead of declaring the classes explicitly and then casting, create generic BitTableBuilder::Entry class for that purpose. This removes the need to keep the POD helper classes in sync. Test: test-art-host-gtest-stack_map_test Test: test-art-host-gtest-bit_table_test Change-Id: I4c632313bafd3a4bc823648436a5310b6f2a1d13
2018-06-14Move some helper methods to DexRegisterLocation.David Srbecky
Test: test-art-host-gtest-stack_map_test Change-Id: I0abab008159db023d531df69214cd3bb8c0639bd
2018-06-13Add Kind column to stack maps.David Srbecky
Add 'Kind' column to stack maps which marks special stack map types, and use it at run-time to add extra sanity checks. It will also allow us to binary search the stack maps. The column increases .oat file by 0.2%. Test: test-art-host-gtest-stack_map_test Change-Id: I2a9143afa0e32bb06174604ca81a64c41fed232f
2018-06-07Delta-compress register maps in stack maps.David Srbecky
The register maps tend to be similar from stack map to stack map, so instead of encoding them again, store only the modified ones. The dex register bitmap stores the delta now - if register has been modified since the previous stack map, the bit will be set. The decoding logic scans backwards through stack maps until it eventfully finds the most recent value of each register. This CL saves ~2.5% of .oat file size (~10% of stackmap size). Due to the scan, this makes dex register decoding slower by factor of 2.5, but that still beats the old algorithm before refactoring. Test: test-art-host-gtest-stack_map_test Change-Id: Id5217a329eb757954e0c9447f38b05ec34118f84
2018-06-03De-duplicate inline info in stack maps.David Srbecky
This saves 0.3% of .oat file size. Test: test-art-host-gtest-stack_map_test Change-Id: Ic7d5addf04fb9b7a2f29a7d1d99ea93b39388fd2
2018-06-01Rewrite stackmap verification code.David Srbecky
The new version is more complicated but it gives much higher confidence about the correctness of the stackmap encoding. The old version was comparing the internal builder entries to the decoded information, which verified the bit-level manipulations, but it did not verify that we created the internal state correctly. The new version directly compares the parameters passed to the StackMapStream and the decoded values. This way, it really tests the whole system. It uses lambda captures to record the parameters. Test: test-art-host-gtest-stack_map_test Change-Id: Ib92819cc35ce0d790128392d303f6feabd7d9c74
2018-06-01Rewrite dex register map encoding in stackmaps.David Srbecky
Simplify code by encoding dex register maps using BitTables. The overall design is unchanged (bitmask+indices+catalogue). This CL saves ~0.4% of .oat file size. The dex register map decoding is factor of 3 faster now (based on the time to verify the register maps on Arm). This is not too surprising as the old version was O(n^2). It also reduces compiler arena memory usage by 11% since the BitTableBuilder is more memory efficient, we store less intermediate data, and we deduplicate most data on the fly. Test: test-art-host-gtest-stack_map_test Change-Id: Ib703a5ddf7f581280522d589e4a2bfebe53c26a9
2018-05-30Simplify DexRegisterMap API.David Srbecky
Store some of the needed decoding state explicitly to avoid passing it around all the time. The DexRegisterMap class is rewritten in next CL. Test: test-art-host-gtest-stack_map_test Change-Id: Ie268dff2a1c1da2e08f0e6799ae51c30e11f350b
2018-05-30Remove the CodeOffset helper class.David Srbecky
I need to reduce the StackMapEntry to a POD type so that it can be used in BitTableBuilder. Test: test-art-host-gtest-stack_map_test Change-Id: I5f9ad7fdc9c9405f22669a11aea14f925ef06ef7
2018-05-30Revert^2 "Optimize register mask and stack mask in stack maps."David Srbecky
This reverts commit 8b20b5c1f5b454b2f8b8bff492c88724b5002600. Reason for revert: Retry submit unmodified after fixing the test. Use BitTable to store the masks as well and move the deduplication responsibility to the BitTable builders. Don't generate entries for masks which are all zeros. This saves 0.2% of .oat file size on both Arm64 and Arm. Encode registers as (value+shift) due to tailing zeros. This saves 1.0% of .oat file size on Arm64 and 0.2% on Arm. Test: test-art-target-gtest-exception_test Test: test-art-host-gtest-bit_table_test Test: test-art-host-gtest-stack_map_test Change-Id: Ib643776dbec3f051cc29cd13ff39e453fab5fae9
2018-05-29Revert "Optimize register mask and stack mask in stack maps."Vladimir Marko
This reverts commit ffaf87a429766ed80e6afee5bebea93db539620b. Reason for revert: Breaks exception_test32 on target for CMS and heap poisoning configs. Change-Id: I127c17f693e28211a799f73a50e73105edee7e4c
2018-05-28Optimize register mask and stack mask in stack maps.David Srbecky
Use BitTable to store the masks as well and move the deduplication responsibility to the BitTable builders. Don't generate entries for masks which are all zeros. This saves 0.2% of .oat file size on both Arm64 and Arm. Encode registers as (value+shift) due to tailing zeros. This saves 1.0% of .oat file size on Arm64 and 0.2% on Arm. Test: test-art-host-gtest Change-Id: I636b7edd49e10e8afc9f2aa385b5980f7ee0e1f1
2018-05-21Rewrite stackmap encoding code.David Srbecky
Remove most of the code related to handling of bit encodings. The design is still same; the encodings are just more implicit. Most of the complexity is replaced with a single BitTable class, which is a generic purpose table of tightly bit-packed integers. It has its own header which stores the bit-encoding of columns, and that removes the need to handle the encodings explicitly. Other classes, like StackMap, are accessors into the BitTable, with named getter methods for the individual columns. This CL saves ~1% of .oat file size (~4% of stackmap size). Test: test-art-host-gtest Change-Id: I7e92683753b0cc376300e3b23d892feac3670890
2018-04-05Move remaining runtime/base stuff to libartbaseDavid Sehr
Move the remainder of the Arena stuff, plus dumpable and runtime/*memory_region* to libartbase. More preparation to build profiling library. Bug: 22322814 Test: make -j 50 checkbuild Change-Id: Iaf26d310c89bc58846553281576c18102f5e4122
2017-11-15Use intrinsic codegen for compiling intrinsic methods.Vladimir Marko
When compiling an intrinsic method, generate a graph that invokes the same method and try to compile it. If the call is actually intrinsified (or simplified to other HIR) and yields a leaf method, use the result of this compilation attempt, otherwise compile the actual code or JNI stub. Note that CodeGenerator::CreateThrowingSlowPathLocations() actually marks the locations as kNoCall if the throw is not in a catch block, thus considering some throwing methods (for example, String.charAt()) as leaf methods. We would ideally want to use the intrinsic codegen for all intrinsics that do not generate a slow-path call to the default implementation. Relying on the leaf method is suboptimal as we're missing out on methods that do other types of calls, for example runtime calls. This shall be fixed in a subsequent CL. Test: m test-art-host-gtest Test: testrunner.py --host --optimizing Bug: 67717501 Change-Id: I640fda7c22d4ff494b5ff77ebec3b7f5f75af652
2017-10-17Use ScopedArenaAllocator for code generation.Vladimir Marko
Reuse the memory previously allocated on the ArenaStack by optimization passes. This CL handles only the architecture-independent codegen and slow paths, architecture-dependent codegen allocations shall be moved to the ScopedArenaAllocator in a follow-up. Memory needed to compile the two most expensive methods for aosp_angler-userdebug boot image: BatteryStats.dumpCheckinLocked() : 19.6MiB -> 18.5MiB (-1189KiB) BatteryStats.dumpLocked(): 39.3MiB -> 37.0MiB (-2379KiB) Also move definitions of functions that use bit_vector-inl.h from bit_vector.h also to bit_vector-inl.h . Test: m test-art-host-gtest Test: testrunner.py --host --optimizing Bug: 64312607 Change-Id: I84688c3a5a95bf90f56bd3a150bc31fedc95f29c
2017-10-11Use ScopedArenaAllocator for building HGraph.Vladimir Marko
Memory needed to compile the two most expensive methods for aosp_angler-userdebug boot image: BatteryStats.dumpCheckinLocked() : 21.1MiB -> 20.2MiB BatteryStats.dumpLocked(): 42.0MiB -> 40.3MiB This is because all the memory previously used by the graph builder is reused by later passes. And finish the "arena"->"allocator" renaming; make renamed allocator pointers that are members of classes const when appropriate (and make a few more members around them const). Test: m test-art-host-gtest Test: testrunner.py --host Bug: 64312607 Change-Id: Ia50aafc80c05941ae5b96984ba4f31ed4c78255e
2017-09-18ART: Move kDexNoIndex to dex_file_types.hAndreas Gampe
Define the constant with the types to allow lowering the dependency on DexFile. Test: m Change-Id: I3c61421db45be96d2057e01b1a7825883d8bd178
2017-03-16Add method info to oat filesMathieu Chartier
The method info data is stored separately from the code info to reduce oat size by improving deduplication of stack maps. To reduce code size, this moves the invoke info and inline info method indices to this table. Oat size for a large app (arm64): 77746816 -> 74023552 (-4.8%) Average oat size reduction for golem (arm64): 2% Repurposed unused SrcMapElem deduping to be for MethodInfo. TODO: Delete SrcMapElem in a follow up CL. Bug: 36124906 Test: clean-oat-host && test-art-host-run-test Change-Id: I2241362e728389030b959f42161ce817cf6e2009
2017-02-15Add invoke infos to stack mapsMathieu Chartier
Invoke info records the invoke type and dex method index for invokes that may reach artQuickResolutionTrampoline. Having this information recorded allows the runtime to avoid reading the dex code and pulling in extra pages. Code size increase for a large app: 93886360 -> 95811480 (2.05% increase) 1/2 of the code size increase is from making less stack maps deduped. I suspect there is less deduping because of the invoke info method index. Merged disabled until we measure the RAM savings. Test: test-art-host, N6P boots Bug: 34109702 Change-Id: I6c5e4a60675a1d7c76dee0561a12909e4ab6d5d9
2017-02-13Deduplicate register maps for both inline and normal stack mapsMathieu Chartier
Before it only deduplicated the normal stack map dex register maps. Code size for a large app: 93341616 -> 92678040 (-0.7%) Added test. Bug: 34621054 Test: test-art-host Change-Id: I4fab4e40915bfa12cb978edbb3cbc19e2cf00954
2017-02-08Clean up code info table layoutMathieu Chartier
Previously: Table layout was computed multiple places like stack_map_stream, and getters. This made it difficult to add new stack map tables and made the code hard to understand. This change makes the table layout specified all inside of the code info. Updating the layout only requires changing ComputeTableOffsets. Changed the stack map inline info offset to be an index, so that it is not require the inline infos are directly after the dex register table. Oat file size for a large app: 94459576 -> 93882040 (-0.61%) Updated oatdump and fixed a bug that was incorrectly computing the register mask bytes. Bug: 34621054 Test: test-art-host Change-Id: I3a7f141e09d5a18bce2bc6c9439835244a22016e
2017-02-01Deduplicate register masksMathieu Chartier
Data is commonly shared between different stack maps. The register masks are stored after the stack masks. Oat size for a large app: 96722288 -> 94485872 (-2.31%) Average oat size reduction according to golem -3.193%. Bug: 34621054 Test: test-art-host Change-Id: I5eacf668992e866d11ddba0c01675038a16cdfb4
2017-01-31Deduplicate stack masksDavid Srbecky
The stack masks repeat often enough so that it is worth deduplicating them. Oat size for a large app: 98143600 -> 96722288 (-1.44%) Bug: 34621054 Test: test-art-host Change-Id: If73d51e46066357049d5be2e406ae9a32b7ff1f4
2017-01-20Compressed native PC for stack mapsMathieu Chartier
Compress native PC based on instruction alignment. This reduces the size of stack maps, boot.oat is 0.4% smaller for arm64. Test: test-art-host, test-art-target, N6P booting Change-Id: I2b70eecabda88b06fa80a85688fd992070d54278
2017-01-16Put inlined ArtMethod pointer in stack maps.Nicolas Geoffray
Currently done for JIT. Can be extended for AOT and inlined boot image methods. Also refactor the lookup of a inlined method at runtime to not rely on the dex cache, but look at the class loader tables. bug: 30933338 test: test-art-host, test-art-target Change-Id: I58bd4d763b82ab8ca3023742835ac388671d1794
2016-06-21Replace String.charAt() with HIR.Vladimir Marko
Replace String.charAt() with HArrayLength, HBoundsCheck and HArrayGet. This allows GVN on the HArrayLength and BCE on the HBoundsCheck as well as using the infrastructure for HArrayGet, i.e. better handling of constant indexes than the old intrinsic and using the HArm64IntermediateAddress. Bug: 28330359 Change-Id: I32bf1da7eeafe82537a60416abf6ac412baa80dc
generated by cgit v1.2.3 (git 2.25.1) at 2025-10-03 12:02:05 +0000