13:10
So something that would be needed is a way to store jump locations within a
program. So do I make JITMethodProgram immutable or mutable? Once it is
passed to the native code generator it will not be touched again. Changing the
state could be used by the writer (such as making floating point operations
software). Alternatively, programs can be divided into secitons and used
that way, containing microops. Then any code that can jump to other areas would
do so via labels and such. There could never be a jump in the middle of a block
so that essentially each group becomes a basic block in a program.
14:31
Technically for a program, the number of locals and stack items used is pointless in most cases.
14:34
Although the large number of jump targets, I would need to parse the stack map table for optimal results (since that knows the type of elements on the stack).
14:35
However, if I defer handling code and read the exception and the stack map table, when parsing the byte code I will know the types used and all of the entry points in the code. Then in that case, I will not need micro-operations at all. I can have explicit code generation without building a temporary program. Then I can also just do a linear pass through the byte code also.
14:37
The only allocations needed would be for the stack map table information, the exception table, and the extra bytes that the byte code uses. This would be likely the fastest means of code generation. Then the only optimization that can be performed for the most part reliably would be caching of stack values. However reading the other tables first gives me the information needed where I can easily determine if a value is used or not. I can then take the approach I have thought up of before, where any write to a local is stored to the stack while anything on the stack is purely temporary. There would just need to be a temporary storage area in the event an exception occurs. However stack temporaries only need to be managed in a single basic block for the most part. Jumps to exceptions are ignored. Transition to another basic block just needs a translocation of entries if required (in the event of a jump back and a stack position was deallocated to the stack but at that given point it is allocated).
14:42
So in short, skipping the code for now, will produce a bit faster of a generator without going into lots of allocations to setup a program and then performing allocations. I have previously written compact stack map table parsers and state which I can use.
17:21
And I decided to split __ClassDecoder__ up, so now all of the resulting split
off code is in their own classes and much smaller. In the end some memory could
be saved in a way.
20:19
Oh yes, read does not have to read all the bytes on the input, what I need to
do is use readFully.