https://gcc.gnu.org/bugzilla/show_bug.cgi?id=92538
Bug ID: 92538
Summary: Proposal for IPA init() constant propagation
Product: gcc
Version: 10.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: ipa
Assignee: unassigned at gcc dot gnu.org
Reporter: erick.oc...@theobroma-systems.com
CC: marxin at gcc dot gnu.org
Target Milestone: ---
Created attachment 47278
--> https://gcc.gnu.org/bugzilla/attachment.cgi?id=47278&action=edit
Patch
Hello,
we've been working on an lto optimization and would like to receive some
feedback on this candidate patch.
At the moment the patch compiles correctly when applied to master.
We have some initial test cases in a separate repository and have compiled and
ran a small subset of CPU 2017 benchmarks with comparable results.
The proposed transformation (ipa-initcall-cp) is a variant of interprocedural
constant propagation.
ip-initcall-cp collects all variables with static lifetime that contain only a
single write (like in the cases of initialization functions) and propagates it
to reads across the whole link unit.
In order to run, apply the patch and compile with `-lto -finitcall-cp`.
In order for this transformation to be sound
* writes that can be reached from a function pointer,
* writes that can be reached from a function with outside visibility, and
* writes that may happen after any read
are not taken into account.
In order to determine that no read happens before the write, we have to:
* find the main function
* find the function and basic block of the write
* for each read in another function
* make sure that call to write dominates all callees of the read function
* for each read in the same function
* make sure that write dominates read
Some initial drawbacks:
* We've noticed that we have to disable ipa-cp in order for ipa-initcall-cp to
run successfully.
This is most likely due to some issue with clones and we will need to make some
design changes.
The function materialize all clones fails after ipa-initcall-cp if ipa-cp is
not commented out.
Suggestions are welcomed.
* At the moment, I still need to clean the code a bit, since it doesn't pass
the standards.
* I still need to add tests using the testsuite as opposed to running them
myself.
Some future work:
* At the moment, ipa-initcall-cp only propagates values from a single write.
However, we could conceivably improve this work to propagate the first n-writes
and specialize functions based on the constants.
