New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

Byte code size and serialized byte code size optimizations (15-20% serialized byte code size reduction) #3594

Merged

chakrabot merged 1 commit into chakra-core:master from curtisman:bytecode

Sep 1, 2017

Contributor

curtisman commented Aug 28, 2017

Byte code size reductions:

Variable slot index size in layout

Serialized byte code size reductions:

Move has loopHeaderArray and has asmJsInfo into the bit field (instead of a byte)
Don't emit field that has non-zero default value.
Use the defined fields to indicate existing of more field to reduce serialized byte code size
Use constant type to indicate whether it is property string or not instead of an extra boolean
Add int8 and int16 constant type for small int encoding in the constant table
Separate out the byte offset from the character offset in the LineOffsetCache so that we can omit them if the source is purely ASCII-7
-Reduce by half the memory used by the LineOffsetCache for ASCII-7 sources, and it reduces the serialized byte code size as well.
Remove unused sourceSpans from byte code serialization.
Serialize the bit flag with constant encoding (since the bit flag always as a large value)

Also:

Clean up unused Reg2WithICIndex and REg1Int2 layout
Move LineOffsetCache to Runtime/Base instead of Common/DataStructures, and de-templatize it as we only use Recycler as the allocator anyway.

msftclas added the cla-not-required label

curtisman force-pushed the bytecode branch 3 times, most recently from c45931a to 31cc8d6 Compare

August 28, 2017 20:36

curtisman requested a review from pleath

August 28, 2017 20:36

dilijev reviewed

View reviewed changes

lib/Runtime/Language/InterpreterStackFrame.cpp

    
            @@ -8889,7 +8889,7 @@ const byte * InterpreterStackFrame::OP_ProfiledLoopBodyStart(const byte * ip)
          
                      return JavascriptOperators::OP_LdModuleSlot(playout->SlotIndex1, playout->SlotIndex2, scriptContext);

                  }

                  inline void InterpreterStackFrame::OP_StModuleSlot(Var instance, int32 slotIndex1, int32 slotIndex2, Var value)

                  inline void InterpreterStackFrame::OP_StModuleSlot(Var instance, uint32 slotIndex1, uint32 slotIndex2, Var value)

Contributor

dilijev Aug 31, 2017 •

edited

Loading

Var instance [](start = 55, length = 12)

instance param unused?

Contributor Author

curtisman Aug 31, 2017

Yeah.. it is borrowing the SET_ELEM_ENVSLOTNonVar in the interpreter dispatch, which to clean up will require more change. I will do that separately.

dilijev reviewed

View reviewed changes

lib/Common/DataStructures/LineOffsetCache.h

-              namespace JsUtil
-              {
-                  template <typename TAllocator>
-                  class LineOffsetCache

Contributor

dilijev Aug 31, 2017

LineOffsetCache [](start = 10, length = 15)

Any changes to this code besides splitting into .h and .cpp under lib/runtime/base and removing the JsUtil namespace?

Contributor Author

curtisman Aug 31, 2017

Yes. Splitting the data from charOffset/byteOffset pair to separate list, so that the byteOffset don't need to be stored if it is ASCII-7 (which charOffset == byteOffset)

Contributor

dilijev commented Aug 31, 2017

#undef DECLARE_SERIALIZABLE_ACCESSOR_FIELD

Move #undef DECLARE_SERIALIZABLE_ACCESSOR_FIELD_NO_CHECK after this line?

Refers to: lib/Runtime/SerializableFunctionFields.h:112 in 31cc8d6. [](commit_id = 31cc8d6, deletion_comment = False)

dilijev reviewed

View reviewed changes

lib/Runtime/ByteCode/ByteCodeSerializer.cpp

@@ @@ -64,7 +64,7 @@ namespace Js @@
                   const int magicStartOfPropIdsOfFormals = *(int*)"pif[";
                   const int magicEndOfPropIdsOfFormals = *(int*)"]pif";
                   const int magicStartOfSlotIdToNestedIndexArray = *(int*)"sni[";
-                  const int magicEndOfSlotIdToNestedIndexArray = *(int*)"]sni"
+                  const int magicEndOfSlotIdToNestedIndexArray = *(int*)"]sni";

Contributor

dilijev Aug 31, 2017

; [](start = 64, length = 1)

wait... how was this not an error?

Contributor Author

curtisman Aug 31, 2017

The code is only enabled explicitly to debug issues. No official build has this code enabled.

dilijev reviewed

View reviewed changes

lib/Runtime/ByteCode/ByteCodeSerializer.cpp Outdated

		bool has_slotIdInCachedScopeToNestedIndexArray : 1;
		bool has_debuggerScopeSlotArray : 1;

Contributor

dilijev Aug 31, 2017

nit: blank line

Contributor Author

curtisman Aug 31, 2017

removed

dilijev reviewed

View reviewed changes

lib/Runtime/ByteCode/ByteCodeSerializer.cpp Outdated

+                  ctStringTemplateCallsite = 10,
+                  ctPropertyString16 = 11,
+                  ctInt16 = 12,
+                  ctInt8 = 13,

Contributor

dilijev Aug 31, 2017

why not group the three int size constant types?

Contributor Author

curtisman Aug 31, 2017

I guess since the byte code is going to be updated already, I might as well?

Contributor Author

curtisman Aug 31, 2017

Sigh... I will have to renumber them all

dilijev reviewed

View reviewed changes

lib/Runtime/ByteCode/ByteCodeSerializer.cpp

@@ @@ -1784,45 +1813,35 @@ class ByteCodeBufferBuilder @@
                       return sizeof(serialization_alignment TStructType);
                   }
-                  uint32 AddPropertyIdOfFormals(BufferBuilderList & builder, FunctionBody * function)
+                  uint32 AddPropertyIdOfFormals(BufferBuilderList & builder, PropertyIdArray * propIds, FunctionBody * function)

Contributor

dilijev Aug 31, 2017

propIds [](start = 81, length = 7)

Assert(propIds != nullptr) ?

Contributor Author

curtisman Aug 31, 2017

Doesn't feel like the assert serves much value.....

dilijev reviewed

View reviewed changes

lib/Runtime/ByteCode/ByteCodeSerializer.cpp

@@ @@ -2365,6 +2413,11 @@ class ByteCodeBufferReader @@
                       return ReadConstantSizedInt32(buffer, remainingBytes, value);
                   }
+                  const byte * ReadConstantSizedUInt32(const byte * buffer, uint * value)
+                  {
+                      return ReadConstantSizedInt32(buffer, (int *)value);

Contributor

dilijev Aug 31, 2017

Int32 [](start = 32, length = 5)

mismatch of ReadUInt32 (name) and ReadInt32 (returned)

Contributor Author

curtisman Aug 31, 2017

The uint return is in out param.

pleath approved these changes

View reviewed changes

tcare approved these changes

View reviewed changes

Contributor

tcare left a comment


          Byte code size and serialized byte code size optimizations (15-20% se…

6eb4457

…rialized byte code size reduction)

Byte code size reductions:
- Variable slot index size in layout

Serialized byte code size reductions:
- Move has loopHeaderArray and has asmJsInfo into the bit field (instead of a byte)
- Don't emit field that has non-zero default value.
- Use the defined fields to indicate existing of more field to reduce serialized byte code size
- Use constant type to indicate whether it is property string or not instead of an extra boolean
- Add int8 and int16 constant type for small int encoding in the constant table
- Separate out the byte offset from the character offset in the LineOffsetCache so that we can omit them if the source is purely ASCII-7
    -Reduce by half the memory used by the LineOffsetCache for ASCII-7 sources, and it reduces the serialized byte code size as well.
- Remove unused sourceSpans from byte code serialization.
- Serialize the bit flag with constant encoding (since the bit flag always as a large value)

Also:
- Clean up unused Reg2WithICIndex and REg1Int2 layout
- Move LineOffsetCache to Runtime/Base instead of Common/DataStructures, and de-templatize it as we only use Recycler as the allocator anyway.

curtisman force-pushed the bytecode branch from 31cc8d6 to 6eb4457 Compare

August 31, 2017 22:16

chakrabot merged commit 6eb4457 into chakra-core:master

chakrabot pushed a commit that referenced this pull request


          [MERGE #3594 @curtisman] Byte code size and serialized byte code size…

f7b3d6a

… optimizations (15-20% serialized byte code size reduction)

Merge pull request #3594 from curtisman:bytecode

Byte code size reductions:
- Variable slot index size in layout

Serialized byte code size reductions:
- Move has loopHeaderArray and has asmJsInfo into the bit field (instead of a byte)
- Don't emit field that has non-zero default value.
- Use the defined fields to indicate existing of more field to reduce serialized byte code size
- Use constant type to indicate whether it is property string or not instead of an extra boolean
- Add int8 and int16 constant type for small int encoding in the constant table
- Separate out the byte offset from the character offset in the LineOffsetCache so that we can omit them if the source is purely ASCII-7
    -Reduce by half the memory used by the LineOffsetCache for ASCII-7 sources, and it reduces the serialized byte code size as well.
- Remove unused sourceSpans from byte code serialization.
- Serialize the bit flag with constant encoding (since the bit flag always as a large value)

Also:
- Clean up unused Reg2WithICIndex and REg1Int2 layout
- Move LineOffsetCache to Runtime/Base instead of Common/DataStructures, and de-templatize it as we only use Recycler as the allocator anyway.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet