Value range propagation. about sdc HOT 3 OPEN

@bhelyer @deadalnix I can see some code for this in libd/src/d/semantic/vrp.d . Is this based on Jason Patterson's method or it is dominator tree based approach ? In which files this is used for simplification ? Is CTFE a probable candidate ?

from sdc.

I'm not sure what you mean. VRP is used when casting from an integral type to a smaller integral type. It is allowed in D if VRP can prove it is safe. It is called from caster.d as a fallback when regular integral cast fails.

vrp.d is very ad hoc right now, and could win from some improvement.

VRP here doesn't need to take into account flow analysis (that would make the front too slow) and is just exploring the expression tree and gathering possible value ranges.

VRP is also used by the optimizer, and take flow analysis into account, but this is the optimizer's job and is done by LLVM, not the frontend.

from sdc.

@deadalnix Ok I understand what you mean by VRP is used for check before casting.
As you pointed out correctly , compiler optimization VRP and for that matter I am referring very old but useful work http://www.lighterra.com/papers/valuerangeprop/Patterson1995-ValueRangeProp.ps.gz
It is more aggressive than constant propagation. LLVM has implementation for that but not in upstream compiler , it maintained separately by some research group ( as I recall ).
For the purpose of SDC what are some performance improvement ? Do you have any other reference implementation ( any other language) to compare with ?

from sdc.