[vm/ffi] Enable inlining idempotent force-optimized functions

[dcharkes]

The dart:ffi memory operations are on par with dart:typed_data in AoT, but roughly 4x slower in JiT on x64/ia32/arm64/arm.

To get JiT in line, the indexed load/store should be made inlineable in unoptimized code. This is possible, even with code motion, because the index load/store are idempotent operations.

x64 measurements:

$ dart benchmarks/FfiMemory/dart/FfiMemory.dart 
FfiMemory.PointerInt8(RunTime): 72.17852760736196 us.

$ dart ../../golem/golem4/benchmarks/TypedData/dart/TypedData.dart 
TypedData.Int8ListBench(RunTime): 15.412078084904714 us.

$ DART_CONFIGURATION=ReleaseX64 pkg/vm/tool/precompiler2 benchmarks/FfiMemory/dart/FfiMemory.dart bench.dill
$ DART_CONFIGURATION=ReleaseX64 pkg/vm/tool/dart_precompiled_runtime2 bench.dill
FfiMemory.PointerInt8(RunTime): 19.55705275509705 us.

$ DART_CONFIGURATION=ReleaseX64 pkg/vm/tool/precompiler2 ../../golem/golem4/benchmarks/TypedData/dart/TypedData.dart bench.dill
$ DART_CONFIGURATION=ReleaseX64 pkg/vm/tool/dart_precompiled_runtime2 bench.dill
TypedData.Int8ListBench(RunTime): 19.509307113036012 us.

[mkustermann]

We use force-optimized functions outside the FFI. It's a general optimization we'd like to do. Removinglibrary-ffi label.

[alexmarkov]

Also see https://dart-review.googlesource.com/c/sdk/+/210648 for the discussion of inlining force-optimized functions.

In 243320 we're using Location::Any() to force handle inputs into FFI calls to be placed on the stack.
But once we start inlining the trampolines we also have to handle the case where inputs can be constants.

[zanderso]

@rmacnak-google @a-siva We discussed today that this optimization could also reduce the binary size overhead. We saw a ~50KB compressed increase in Flutter app size when switching the Engine to FFI natives.

[mkustermann]

There will be some other things needed before @FfiNative will benefit from inlining force-optimized functions. @dcharkes might look into those things.