Batuhan Taskaya <[email protected]> added the comment:
There are multiple issues with doing 'static' loop invariant code motion in
python, but the most obvious one is how would you detect if anything in the
body is related to the loop or not? The trivial way that the compiled languages
do this optimization is generating some sort of data dependency graph, and
seeing whether anything in an expression refers to the loop target though in
Python you can't know for sure unless you know every possible type in the loop
body and the type of the loop target which is something nearly impossible to
infer (due to the dynamic nature of python). You could theoretically do this
only if the whole loop only uses simple values:
for x in [1, 2, 3, 4, 5]:
a = 5
if a > 3:
a += 2
continue
and in that case, it is better to just optimize the whole loop away but since
this kind of code doesn't exist at all in the public, there is no need to make
the compiler more complicated.
Another option is generating multiple code objects with guards, which would
require to do first some sort of partial type inference and generate the code
for the possible types. It would require a substantial amount of changes on
both the compiler and the interpreter which is something I'd say can be
deferred for later if we implement a tracing JIT. With traces, it should be way
easier to do this sort of trivial compiler optimization (loop invariants,
constant propagation, GVN, etc.) since we would know the types, and possible
code paths.
----------
nosy: +BTaskaya
_______________________________________
Python tracker <[email protected]>
<https://bugs.python.org/issue44505>
_______________________________________
_______________________________________________
Python-bugs-list mailing list
Unsubscribe:
https://mail.python.org/mailman/options/python-bugs-list/archive%40mail-archive.com
