Improve packed field decoding

[osa1]

• Inline _readPacked manually and _withLimit with a pragma to eliminate
  closure allocation and calls in packed decoding loops.

  • _readPacked is manually inlined as VM's inliner doesn't properly fold
    constants after the inlinings, see
    Closure and closure context field loads are not folded in AOT dart-lang/sdk#60068.

• Introduce PbList._addUnchecked to add to the list without checking the
  value for validity and list for mutability.

• When decoding a packed field, check the list mutability once, instead of for
  every element.

• When decoding a packed scalar field, don't check for value validity.

  For scalar fields we need to make sure the field value is not null, which is
  already guaranteed in the call sites as e.g. input.readDouble doesn't
  return nullable.

• Sprinkle a bunch of prefer-inlines to make sure VM will inline one liners.

VM benchmarks before:

protobuf_PackedInt32Decoding(RunTimeRaw): 25598.8125 us.
protobuf_PackedInt64Decoding(RunTimeRaw): 67932.43333333333 us.
protobuf_PackedUint32Decoding(RunTimeRaw): 24668.844444444443 us.
protobuf_PackedUint64Decoding(RunTimeRaw): 64615.066666666666 us.
protobuf_PackedSint32Decoding(RunTimeRaw): 26037.275 us.
protobuf_PackedSint64Decoding(RunTimeRaw): 100819.65 us.
protobuf_PackedBoolDecoding(RunTimeRaw): 34733.4 us.
protobuf_PackedEnumDecoding(RunTimeRaw): 48379.659999999996 us.

VM benchmarks after:

protobuf_PackedInt32Decoding(RunTimeRaw): 19653.9 us.
protobuf_PackedInt64Decoding(RunTimeRaw): 48627.9 us.
protobuf_PackedUint32Decoding(RunTimeRaw): 19279.29090909091 us.
protobuf_PackedUint64Decoding(RunTimeRaw): 50681.8 us.
protobuf_PackedSint32Decoding(RunTimeRaw): 20271.854545454546 us.
protobuf_PackedSint64Decoding(RunTimeRaw): 83777.8 us.
protobuf_PackedBoolDecoding(RunTimeRaw): 24850.555555555555 us.
protobuf_PackedEnumDecoding(RunTimeRaw): 45205.659999999996 us.

Wasm benchmarks before (-O2):

protobuf_PackedInt32Decoding(RunTimeRaw): 64220.0 us.
protobuf_PackedInt64Decoding(RunTimeRaw): 81033.33333333334 us.
protobuf_PackedUint32Decoding(RunTimeRaw): 60800.0 us.
protobuf_PackedUint64Decoding(RunTimeRaw): 82700.0 us.
protobuf_PackedSint32Decoding(RunTimeRaw): 72433.33333333334 us.
protobuf_PackedSint64Decoding(RunTimeRaw): 142150.0 us.
protobuf_PackedBoolDecoding(RunTimeRaw): 27775.0 us.
protobuf_PackedEnumDecoding(RunTimeRaw): 43980.0 us.

Wasm benchmarks after:

protobuf_PackedInt32Decoding(RunTimeRaw): 56050.0 us.
protobuf_PackedInt64Decoding(RunTimeRaw): 74633.33333333334 us.
protobuf_PackedUint32Decoding(RunTimeRaw): 56525.0 us.
protobuf_PackedUint64Decoding(RunTimeRaw): 69400.0 us.
protobuf_PackedSint32Decoding(RunTimeRaw): 51925.0 us.
protobuf_PackedSint64Decoding(RunTimeRaw): 116250.0 us.
protobuf_PackedBoolDecoding(RunTimeRaw): 18427.272727272728 us.
protobuf_PackedEnumDecoding(RunTimeRaw): 41600.0 us.

---

cl/755309114

[mkustermann]

It seems in dart2wasm we have

// before
protobuf_PackedInt64Decoding(RunTimeRaw): 77833.33333333334 us.

// after
protobuf_PackedInt64Decoding(RunTimeRaw): 90650.0 us.

i.e. a regression, is this expected?

[mkustermann]

lgtm with comments

[mkustermann]

It seems to be hand-inlined code, omitting the _check call. It seems not very nice we have to copy&paste this many times.

Why not pass in a bool whether to check or not, and force inline the function. E.g.

_addInternal(input.readBytes(), omitCheck: true);

@pragma('vm/dart2js/dart2wasm:prefer-inline')
_addInternal(E value, {bool omitCheck = false}) {
  _checkModifiable(value, 'add');
  if (!omitCheck) _check(value);
  _wrappedList.add(value);
}

[osa1]

This function wouldn't work below where we check for modifiable once before adding the elements, and then never check it while adding the elements.

We could have this version:

_addInternal(E value, {bool omitElementCheck = false, bool omitModifiableCheck = false}) {
  if (!omitModifiableCheck) _checkModifiable(value, 'add');
  if (!omitElementCheck) _check(value);
  _wrappedList.add(value);
}

But we would still need a _checkModifiable call in packed fields as we want to run it once.

Overall not sure if this version is better than the current..

[osa1]

It seems in dart2wasm we have

// before
protobuf_PackedInt64Decoding(RunTimeRaw): 77833.33333333334 us.

// after
protobuf_PackedInt64Decoding(RunTimeRaw): 90650.0 us.

i.e. a regression, is this expected?

Maybe my CPU became busy with other things while running it, when I run the benchmarks again I consistently get better numbers with this PR. I updated the PR description with more recent benchmark numbers from a run just now.

[osa1]

I tested this internally in cl/755309114. Merging ...