[Sbcl-devel] Peephole Optimizing

[DDL <dd...@ve...>]

Hi, long time lurker, first time poster.

I was poking around the Python's back end (!!) to see where a peephole =
optimizer could go.  As occasional comments in the mail archives have =
noted, there seems to be no place in the compiler that the full set of a =
function's assembly instructions are collected together.  The only =
apparent way to add a peephole optimizer to the assembler is by either a =
massive rewrite to the way machine instructions are emitted by vops, or =
some bad kludge that tries to capture the assembly instructions as the =
vops are processed (and keep syncronized with the backpatches and =
fixups).

Is there another option, such as optimizing the code completely =
separately from the compiler?  For example take a compiled function, use =
the disassembler to re-create the assembly code, then peephole process =
the assembly code and make any resulting changes to the machine code =
(either in parallel or queue up changes).    You would also need to then =
update the labels, offsets and any padding for the machine code.  =
Finally, bind the new machine instruction set to the proper slot of the =
symbol.

Another approach might be to read an existing fasl file in a streaming =
fashion, and optimize function code before writing it out to a new fasl =
file.

My question is does any of these alternatives make sense?  In some ways =
it would be a nice tool separate from the compiler that can allow =
experimenting and benchmarking certain code optimizations (of the =
peephole variety, not the stuff Python already excels at) to find low =
hanging fruit.

DDL