More compact snapshots.

More compact snapshots.

Uses a variable-length encoding for the w and x parts of the snapshot
byte code. Does more work at snapshot creation time to find a more
optimal choice of bytecodes to encode the snapshot.

Also arranges for the null, true, and false objects to be in fixed positions, relative to each other. This frees up 2 registers in the ARM interpreter, which we use for the program and the dispatch table.

R=sgjesse@google.com
BUG=

Committed: https://github.com/dartino/sdk/commit/880b7978e511193ae8162f437dfe1b6553a05629

[erikcorry, 2016-06-13 13:05:35]

Description was changed from

==========
More compact snapshots.

Uses a variable-length encoding for the w and x parts of the snapshot
byte code. Does more work at snapshot creation time to find a more
optimal choice of bytecodes to encode the snapshot.

R=sgjesse@google.com
BUG=
==========

to

==========
More compact snapshots.

Uses a variable-length encoding for the w and x parts of the snapshot
byte code. Does more work at snapshot creation time to find a more
optimal choice of bytecodes to encode the snapshot.

Also arranges for the null, true, and false objects to be in fixed positions,
relative to each other. This frees up 2 registers in the ARM interpreter, which
we use for the program and the dispatch table.

R=sgjesse@google.com
BUG=
==========

[Søren Gjesse, 2016-06-14 08:33:49]

LGTM

https://codereview.chromium.org/2067483002/diff/1/src/vm/snapshot.cc
File src/vm/snapshot.cc (right):

https://codereview.chromium.org/2067483002/diff/1/src/vm/snapshot.cc#newcode745
src/vm/snapshot.cc:745: // pass it keeps track of all possible decisions for the
last 4 pointers. When
4 -> kStatesLog2

https://codereview.chromium.org/2067483002/diff/1/src/vm/snapshot.cc#newcode751
src/vm/snapshot.cc:751: SnapshotOracle(bool smi_mode, PortableAddressMap* map,
SnapshotWriter* writer)
Would it be better to loose the smi_mode, and create two subclasses with
VisitBlock/VisitInteger/VisitLiteralInteger implementations?

https://codereview.chromium.org/2067483002/diff/1/src/vm/snapshot.cc#newcode753
src/vm/snapshot.cc:753: for (int i = 0; i < kStates; i++) {
Please mention somewhere that costs_ and regs_ hold three trees.

https://codereview.chromium.org/2067483002/diff/1/src/vm/snapshot.cc#newcode793
src/vm/snapshot.cc:793: // Add four bogus entries to the list of pointers to be
encoded. These
four -> final/enough?

https://codereview.chromium.org/2067483002/diff/1/src/vm/snapshot.cc#newcode825
src/vm/snapshot.cc:825: // Keep entries 0-7, and spread them out.
0-7 -> lower half / left side

https://codereview.chromium.org/2067483002/diff/1/src/vm/snapshot.cc#newcode833
src/vm/snapshot.cc:833: // Keep entries 8-15, and spread them out.
8-15 -> upper half / right side

[erikcorry, 2016-06-15 10:30:48]

https://codereview.chromium.org/2067483002/diff/1/src/vm/snapshot.cc
File src/vm/snapshot.cc (right):

https://codereview.chromium.org/2067483002/diff/1/src/vm/snapshot.cc#newcode745
src/vm/snapshot.cc:745: // pass it keeps track of all possible decisions for the
last 4 pointers. When
On 2016/06/14 08:33:49, Søren Gjesse wrote:
> 4 -> kStatesLog2

Done.

https://codereview.chromium.org/2067483002/diff/1/src/vm/snapshot.cc#newcode751
src/vm/snapshot.cc:751: SnapshotOracle(bool smi_mode, PortableAddressMap* map,
SnapshotWriter* writer)
On 2016/06/14 08:33:49, Søren Gjesse wrote:
> Would it be better to loose the smi_mode, and create two subclasses with
> VisitBlock/VisitInteger/VisitLiteralInteger implementations?

After discussion offline we agreed that this isn't a clear win, so I'll leave
it.

https://codereview.chromium.org/2067483002/diff/1/src/vm/snapshot.cc#newcode753
src/vm/snapshot.cc:753: for (int i = 0; i < kStates; i++) {
On 2016/06/14 08:33:49, Søren Gjesse wrote:
> Please mention somewhere that costs_ and regs_ hold three trees.

Done.

https://codereview.chromium.org/2067483002/diff/1/src/vm/snapshot.cc#newcode793
src/vm/snapshot.cc:793: // Add four bogus entries to the list of pointers to be
encoded. These
On 2016/06/14 08:33:49, Søren Gjesse wrote:
> four -> final/enough?

Done.

https://codereview.chromium.org/2067483002/diff/1/src/vm/snapshot.cc#newcode825
src/vm/snapshot.cc:825: // Keep entries 0-7, and spread them out.
On 2016/06/14 08:33:49, Søren Gjesse wrote:
> 0-7 -> lower half / left side

Done.

https://codereview.chromium.org/2067483002/diff/1/src/vm/snapshot.cc#newcode833
src/vm/snapshot.cc:833: // Keep entries 8-15, and spread them out.
On 2016/06/14 08:33:49, Søren Gjesse wrote:
> 8-15 -> upper half / right side

Done.

[erikcorry, 2016-06-15 11:49:36]

Description was changed from

==========
More compact snapshots.

Uses a variable-length encoding for the w and x parts of the snapshot
byte code. Does more work at snapshot creation time to find a more
optimal choice of bytecodes to encode the snapshot.

Also arranges for the null, true, and false objects to be in fixed positions,
relative to each other. This frees up 2 registers in the ARM interpreter, which
we use for the program and the dispatch table.

R=sgjesse@google.com
BUG=
==========

to

==========
More compact snapshots.

Uses a variable-length encoding for the w and x parts of the snapshot
byte code. Does more work at snapshot creation time to find a more
optimal choice of bytecodes to encode the snapshot.

Also arranges for the null, true, and false objects to be in fixed positions,
relative to each other. This frees up 2 registers in the ARM interpreter, which
we use for the program and the dispatch table.

R=sgjesse@google.com
BUG=

Committed:
https://github.com/dartino/sdk/commit/880b7978e511193ae8162f437dfe1b6553a05629
==========

[erikcorry, 2016-06-15 11:49:38]

Committed patchset #2 (id:20001) manually as
880b7978e511193ae8162f437dfe1b6553a05629 (presubmit successful).