Hi everyone,
I've been looking at some node-level optimisations as of late, and I
came across a couple of potential ones while doing experiments with the
test on i38527, namely (R is an ordinal tempref, and n is an ordinal
constant):
add
|- add
| |- R
| |- R
|
|- R
Becoming:
mul
|- R
|- 3
And:
add
|- mul
| |- R
| |- n
|
|- R
Becoming:
mul
|- R
|- n+1
I've been having a couple of issues though that I haven't been able to
solve, the main thing is that taddnode.simplify doesn't get called to
transmute the nodes as desired because pass 1 is already complete for
most of the nodes by the time CSE is executed.
I've tried a few workarounds, but ran into problems:
- Once CSE is complete, calling simplify via foreachnodestatic with the
pre-process option is unacceptably slow.
- Attempting to selectively reset the pass 1 flags (in an attempt to
minimise the massive performance loss with manually calling simplify as
above) causes Internal Error 200405231 because it seems to upset
implicit finally blocks.
- Moving "add_entry_exit_code" to after "optcse" causes compilation of
case blocks to malfunction.
Can anyone give me advice on how to implement such a node-level
CSE-specific optimisation? Such a thing may be useful for pure functions
(as well as when writing things like "4 * n + 4 * n + 4 * n + 4 * n + 4
* n"!). Thanks in advance.
Gareth aka. Kit
--
This email has been checked for viruses by Avast antivirus software.
https://www.avast.com/antivirus
_______________________________________________
fpc-devel maillist - fpc-devel@lists.freepascal.org
https://lists.freepascal.org/cgi-bin/mailman/listinfo/fpc-devel
