You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Hi, Daniel Berlin pointed me out to this so I decided to give it a try :)
I have a branch where I updated this to 4.0, but here are a few notes: master...dcci:range
I currently stripped the time stats from the pass as TimeValue doesn't exist anymore in LLVM, so that should use something like std::chrono. I'm not particularly interested in them as I generally run passes under a debugger, but I guess those should be reintegrated at some point.
LLVM trunk has a new pass that builds e-SSA, PredicateInfo, so I assume the e-SSA construction could be replaced with the LLVM builtin one.
We have an implementation of Nuutila SCC walker in LLVM now as part of NewGVN. This is private right now, but could be moved to llvm/Transforms/Utils and re-used here.
The (still unimplemented) improvements in 2) and 3) should slim down pass quite a bit, I guess.
Eventually I'd like this work to get to LLVM (maybe building a VRP transform on top of it).
The text was updated successfully, but these errors were encountered:
Hi, Daniel Berlin pointed me out to this so I decided to give it a try :)
I have a branch where I updated this to 4.0, but here are a few notes:
master...dcci:range
I currently stripped the time stats from the pass as
TimeValue
doesn't exist anymore in LLVM, so that should use something likestd::chrono
. I'm not particularly interested in them as I generally run passes under a debugger, but I guess those should be reintegrated at some point.LLVM trunk has a new pass that builds e-SSA,
PredicateInfo
, so I assume the e-SSA construction could be replaced with the LLVM builtin one.We have an implementation of Nuutila SCC walker in LLVM now as part of NewGVN. This is private right now, but could be moved to
llvm/Transforms/Utils
and re-used here.The (still unimplemented) improvements in 2) and 3) should slim down pass quite a bit, I guess.
Eventually I'd like this work to get to LLVM (maybe building a VRP transform on top of it).
The text was updated successfully, but these errors were encountered: