Currently serializing the ast for kernel is done in two passe:
1) Scan the program to find and index all strings. These are then
sorted based on frequency and assigned an id. All string-
references are refering to that id. As small numbers use less
space in the binary than big numbers, sorting the numbers by
frequency saves a certain amount of space.
In addition the string indexing is "hijacked" for the
"LimitedBinaryPrinter" to also perform some CanonicalName
re-indexing.
2) We then serialize the entire thing.
This CL gets rid of a pass by not indexing the strings up-front.
Whenever it is asked to serialize a string it adds it to the index
(if not already there). The serialization is otherwise the same.
This means that:
1) Strings are not sorted by frequency, i.e. the binary output size
can by bigger (numbers below).
2) The stringindex and canonical names are moved to the end of the
binary instead of the front. As we still need it up front for
deserialization some additional data is added to the
ProgramIndex.
3) The "hijacking" done in "LimitedBinaryPrinter" is replaced by
an alternative.
4) We don't spend time on walking the tree twice.
The cost is the binary size. Compiling helloworld with fasta,
as well as looking at outline.dill, platform.dill and
vmservice_io.dill reveals these numbers:
* helloworld.dill is 0.657248732% bigger (26573 bytes)
* outline.dill is 1.686911399% bigger (9395 bytes)
* platform.dill is 0.657062238% bigger (26565 bytes)
* vmservice_io.dill is 0.44991899% bigger (19147 bytes)
The cost does thus not appear to be very big.
The gain is the serialization time.
From 20 runs of an instrumented VM/serialization, running numbers
through calculations stolens from ministat
(https://www.freebsd.org/cgi/man.cgi?query=ministat) reveals the
following:
* Serialization time: -21.69% +/- 1.44%
* Total time spend in relevant parts of bootstrap_nocore.cc,
dart_api_impl.cc (Dart_LoadKernel), bootstrap_nocore.cc,
dart_api_impl.cc (LoadKernelProgram) as well as serialization:
-14.01% +/- 1.58%
From 5 runs of
"time python tools/test.py -m release -cdartk language -j6"
(again run through ministat calculations):
* real: -4.18% +/- 0.5%
* user: -4.2% +/- 0.29%
* sys: No difference at 95%
* user+sys: -3.3% +/- 0.36%
Change-Id: I1c220eac083496994f0a9f1e2a2445b3707c9a93
Reviewed-on: https://dart-review.googlesource.com/2880
Reviewed-by: Samir Jindel <sjindel@google.com>
New folder structure (nested under vm/):
- compiler/
- jit/ - JIT specific code
- aot/ - AOT specific code
- backend/ - all middle-end and back-end code (IL, flow graph)
- assembler/ - assemblers and disassemblers
- frontend/ - front ends (AST -> IL, Kernel -> IL)
compiler/README.md would be the documentation root for the compiler
pipeline
Bug: https://github.com/dart-lang/sdk/issues/30575
Change-Id: I2dfd9688793bff737f7632ddc77fca766875ce36
Reviewed-on: https://dart-review.googlesource.com/2940
Reviewed-by: Vyacheslav Egorov <vegorov@google.com>
Commit-Queue: Vyacheslav Egorov <vegorov@google.com>
This CL copies the kernel bodies for all functions and
fields into the VM heap. The function bodies in the VM
heap are then used when compiling the flowgraphs.
This theoretically means that the malloc'd data can be
freed and that snapshotting from kernel could possibly
work, though it hasn't been tested.
R=kmillikin@google.com
Review-Url: https://codereview.chromium.org/2972343002 .
This CL allows for streaming big parts of the binary,
i.e. without using the AST nodes.
It is thus a stepping-stone in getting rid of the AST nodes in the VM.
Generally, all Expressions except "FunctionExpression",
and all Statements except "FunctionDeclaration" can be streamed.
There are currently not streamed because they create new functions,
which has a pointer to an AstNode (which we don't have when streaming).
Once we no longer need AstNodes at all these can be streamed as well.
This is, I think, mostly a matter of streaming the ScopeBuilder as well,
something that is not currently done.
The way the streaming is build, one has to stream an entire subtree.
That means, that if an expression (or statement), A, that is generally
streamable contains an expression or a statement, B, that is not streamable,
A cannot be streamed.
The way this is build is by marking AstNodes as streamable or not
("cannot_stream_" field). That way we know up front whether we can stream
a subtree or not.
The streaming is done via "kernel_binary_flowgraph".
In this file there are many obvious comments, e.g.
```
TokenPosition position = ReadPosition(); // read position.
```
This has been done in an attempt to add a comment to everything that
reads from the binary to make it stand out more.
All changes from kernel_to_il up to and including May 2nd 2017
should be included.
R=kmillikin@google.com
Review-Url: https://codereview.chromium.org/2854393002 .
In the VM's Kernel representation, introduce wrapper classes for
string and name indexes so it is obvious which one is which. For
convenience there is an implicit conversion so that they can each be
used where an int is allowed. However, there is no implicit
conversion _to_ either of these types.
BUG=
R=vegorov@google.com
Review-Url: https://codereview.chromium.org/2860823002 .
The canonical name table is copied into a typed data array in the VM's
heap. The encoding is the same as in the binary except that the
integer indexes are fixed-size.
Canonical names are now integer indexes instead of objects allocated
in the C++ heap.
BUG=
R=jensj@google.com, vegorov@google.com
Review-Url: https://codereview.chromium.org/2853423002 .
Copy the Kernel string offsets into a uint32 array in the VM's heap.
This avoids allocating small string objects with new and avoids having
a table of the canonical strings.
Instead of an offset and a size, strings are now represented as
indexes into the string table in the heap. The start offset of string
N is found at byte offset N*4 because it is a uint32, and the end
offset is found at byte offset (N+1)*4. The strings themselves are
just integer indexes instead of pointers.
In the stream flow graph builder, string access is all random access.
R=jensj@google.com, vegorov@google.com
Review-Url: https://codereview.chromium.org/2852943003 .
- Instead of a pointer to new'd memory, Kernel strings now have an
offset from the start of the string data.
- When the streaming reader encounters the string data it records the
offset from the start of the binary. This offset is stored in the
Kernel Program and is used to compute the offset for strings.
- When a KernelReader is constructed, the string data is copied into a
Uint8 array in the VM's heap.
- A pointer to the string data is put into every Kernel script so it
can be used for constructing VM strings at compile time.
The source table does not use Kernel strings any more because those
strings are not found in the raw string data. Instead, the source
table uses new'd buffers for strings (but this will be cleaned up
separately).
R=jensj@google.com
Review-Url: https://codereview.chromium.org/2820363002 .
Jens Johansen