[issue4753] Faster opcode dispatch on gcc

[Paolo 'Blaisorblade' Giarrusso]

Paolo 'Blaisorblade' Giarrusso <p.giarrusso at gmail.com> added the comment:

> We would have to change opcode.h for this to be truely useful (in
order to re-use OPCODE_LIST()).

Yep.

> I think that should be the subject of a separate bug entry for code
reorganization.

Agreed, I'll maybe try to find time for it.

> Thanks for all the explanation and pointers!
You're welcome, thanks to you for writing the patch! And 
> About register allocation, I wonder if the "temporary variables" u,v,w
could be declared separately in each opcode rather than at the top of
the eval function, so that the compiler doesn't try to store their
values between two opcodes.

I didn't look at the output, but that shouldn't be needed with decent
optimizers, since they are local variables, so the compiler has a full
picture of their usage (this does not happen with the content of the
heap, where the frame object may lie).

I think that change would just save some compilation time for dataflow
analysis, maybe :-). Or could make clearer which variables is used
where, but that is a matter of taste (what's there is fine for me).

I just studied liveness analysis in compilers, and it computes whether a
variable is live before and after each statement; if the value of a
variable is not used in some piece of code until the next write to the
variable, it is considered dead in that piece of code, and that variable
does not take space; since u, v, w are either unused or are written to
before usage in all opcodes, they're dead at the beginning of each
opcode, so they're also dead just before dispatch.


The only important thing is that the content of the jump table are known
to the compiler and that the compiler makes use of that. Simply passing
a non-const jump table to some function defined elsewhere (which could
in theory modify it) would make the output much worse.

_______________________________________
Python tracker <report at bugs.python.org>
<http://bugs.python.org/issue4753>
_______________________________________