dart-bytecode, arm64: +4.742% geomean
dart-bytecode-jit-unopt, arm64: +12.73% geomean
dart2js-compile, x64: +3.635% geomean
In the polymorphic and unlinked cases, call to a stub the does a linear scan against an ICData.
In the monomorphic case, call to a prologue of the expected target function that checks the expected receiver class. There is additional indirection in the JIT version compared to the AOT version to also tick a usage counter so the inliner can make good decisions.
In the megamorphic case, call to a stub that does a hash table lookup against a MegamorphicCache.
Megamorphic call sites face a loss of precision in usage counts. The call site count is not recorded and the usage counter of the target function is used as an approximation.
Monomorphic and megamorphic calls sites are reset to the polymorphic/unlinked state on hot reload.
Monomorphic and megamorphic calls sites do not check the stepping state, so they are reset to the polymorphic/unlinked state when stepping begins and disabled.
Back-edges now increment the usage counter in addition to checking it. This ensures function with loops containing monomorphic calls will eventually cross the optimization threshold.
Fixed backwards use of kMonomorphicEntryOffset and kPolymorphicEntryOffset.
Fixed C stack overflow when bouncing between the KBC interpreter and a simulator.
Bug: https://github.com/dart-lang/sdk/issues/26780
Bug: https://github.com/dart-lang/sdk/issues/36409
Bug: https://github.com/dart-lang/sdk/issues/36731
Change-Id: I78a49cccd962703a459288e71ce246ed845df474
Reviewed-on: https://dart-review.googlesource.com/c/sdk/+/102820
Commit-Queue: Ryan Macnak <rmacnak@google.com>
Reviewed-by: Alexander Markov <alexmarkov@google.com>
This reverts commit fde6a5917e.
Reason for revert: This commit looks to be causing new test failures on dart2js. While there are some odd results on some builders that look like infra failures, some builders (for example, dart2js-minified-strong-linux-x64-d8) show the new failing tests clearly.
Original change's description:
> [vm, compiler] Unoptimized megamorphic calls.
>
> When an instance call in unoptimized code creates more than FLAG_max_polymorphic_checks cases, switch the call to use a MegamorphicCache instead of ICData. The prevents unbounded collection of type feedback, and gives improvements on microbenchmarks in the 3-8% range for unoptimized code.
>
> It also leads to a loss of target frequency information for the optimizer, leading to different ordering for range checks in polymorphic inlining. This leads to changes on megamorphic microbenchmarks from -31% to +60%, weighted toward the negative end.
>
> In practice the frequency information seems unimportant, as dart2js has 4.01% geomean improvement.
>
> This is a step toward direct monomorphic calls in unoptimized code, which will also make use of the patching and type feedback extraction added here.
>
> Bug: https://github.com/dart-lang/sdk/issues/26780
> Bug: https://github.com/dart-lang/sdk/issues/36409
> Bug: https://github.com/dart-lang/sdk/issues/36731
> Change-Id: I29f53f23b6794c5f5f0db8b8184788cee16fd9c5
> Reviewed-on: https://dart-review.googlesource.com/c/sdk/+/99270
> Reviewed-by: Alexander Markov <alexmarkov@google.com>
TBR=rmacnak@google.com,alexmarkov@google.com,ajcbik@google.com
Change-Id: Icad46b93cdf8541a00563f49da6b4ac0a4df1ba1
No-Presubmit: true
No-Tree-Checks: true
No-Try: true
Bug: https://github.com/dart-lang/sdk/issues/26780, https://github.com/dart-lang/sdk/issues/36409, https://github.com/dart-lang/sdk/issues/36731
Reviewed-on: https://dart-review.googlesource.com/c/sdk/+/103440
Reviewed-by: Teagan Strickland <sstrickl@google.com>
Commit-Queue: Teagan Strickland <sstrickl@google.com>
When an instance call in unoptimized code creates more than FLAG_max_polymorphic_checks cases, switch the call to use a MegamorphicCache instead of ICData. The prevents unbounded collection of type feedback, and gives improvements on microbenchmarks in the 3-8% range for unoptimized code.
It also leads to a loss of target frequency information for the optimizer, leading to different ordering for range checks in polymorphic inlining. This leads to changes on megamorphic microbenchmarks from -31% to +60%, weighted toward the negative end.
In practice the frequency information seems unimportant, as dart2js has 4.01% geomean improvement.
This is a step toward direct monomorphic calls in unoptimized code, which will also make use of the patching and type feedback extraction added here.
Bug: https://github.com/dart-lang/sdk/issues/26780
Bug: https://github.com/dart-lang/sdk/issues/36409
Bug: https://github.com/dart-lang/sdk/issues/36731
Change-Id: I29f53f23b6794c5f5f0db8b8184788cee16fd9c5
Reviewed-on: https://dart-review.googlesource.com/c/sdk/+/99270
Reviewed-by: Alexander Markov <alexmarkov@google.com>
If the --use-bare-instructions flag is enabled we will:
* Make call sites load the target directly from the pool (instead of
the code object) - this saves one instruction (and an indirect load)
* Ensure the object pool will have direct entry addresses by:
- Letting the clustered snapshot reader change any StubCode::UnlinkedCall()
in the object pool by it's monomorphic entry
- Change the code patcher to patch SwitchableCalls by writing the
monomorphic entry into the pool (instead of the code object)
Issue https://github.com/dart-lang/sdk/issues/33274
Change-Id: I4e41fc8e4461bde477cc559a6a4fccaaf3a350b5
Reviewed-on: https://dart-review.googlesource.com/c/86160
Commit-Queue: Martin Kustermann <kustermann@google.com>
Reviewed-by: Ryan Macnak <rmacnak@google.com>
Reviewed-by: Vyacheslav Egorov <vegorov@google.com>
When throwing to a frame scheduled for lazy deopt, update the continuation pc for that frame to be the catch handler.
Weaken new assert that the deopt pc belongs to the frame's code as the deopt pc for the last eager deopt in a function lies outside the code, after the call to the deopt stub.
R=fschneider@google.com
Review URL: https://codereview.chromium.org/2392613002 .
The call sequence is very similar to a classic IC call, except the guarded class and the target are loaded indirectly from the constant pool instead of as immediates. In the monomorphic case, we call directly to the expected target with a class check in the callee. In the unlinked, polymorphic and megamorphic cases, we call a stub; these case are now call-through instead of call-and-return.
Every code, except stubs involved in switchable calls, includes the class check sequence at the beginning. So we now distinguish between a checked and an unchecked entry point. Generated code except the switchable call continues to use the unchecked entry point.
PC offsets are calculated relative to the beginning of the instruction stream, rather than either entry point.
BUG=
R=fschneider@google.com
Review URL: https://codereview.chromium.org/2226893002 .
Reserve first element in the Function's ic_data_array to hold the edge
counter array.
Until now we had a one-element array per edge counter. This reduces memory
used by edge counters.
This CL allows to optimize code without having to have the unoptimized
code present.
Also, save space in Instruction by making place_id_ and lifetime_position_
a union. place_id_ is exclusively needed by Load/StoreOptimizer,
lifetime_position by the FlowGraphAllocator.
BUG=
R=asiva@google.com
Review URL: https://codereview.chromium.org//1343383003 .
Instead of calling code object directly, call indirectly and
pass the code object in a register. The object pool is then loaded from
the code object. This is another preparation step for making generated code
relocatable.
All non-ia32 platforms:
No entry patching.
ARM:
PC marker (now code object) moves to the same place as on x64 (below saved PP, above saved FP).
R9 is now used as PP, R10 as CODE_REG.
BUG=
R=rmacnak@google.com
Review URL: https://codereview.chromium.org//1192103004 .
The first invocation of a native functions goes into LinkNativeCall which
determines the final entry point and patches the object pool entry.
When running precompiled code, this makes deserializing the object
pool entries for native functions easy, they all initially point to
a single entry (LinkNativeCall).
BUG=
R=rmacnak@google.com
Review URL: https://codereview.chromium.org//1294113004 .
This adds meta-information to object pool entries to allow storing
untagged immediates or code addresses (ExternalLabel) directly.
This eliminates the need to generate extra code to preserve the LSB
when storing immediates as smis (x64, arm64).
BUG=
Review URL: https://codereview.chromium.org//1175523002.
Instead of changing the generated code, change the target of calls
by changing the corresponding entry in the constant pool. Patchable
call sites can't share constant pool entries for now.
x64 and arm64 already had support for patching call targets in the
constant pool, but the debugger used to change the code instead.
This eliminates the fixed debugger stub entries in every constant pool.
R=hausner@google.com
Review URL: https://codereview.chromium.org//1137313002
Ryan Macnak