[flang-commits] [PATCH] D134285: [flang][RFC] Adding higher level FIR ops to ease expression lowering

[Jean Perier via Phabricator via flang-commits]

jeanPerier added inline comments.


================
Comment at: flang/docs/HighLevelFIR.md:960
+
+It has not been decided if these new operations should be added to FIR or if
+this should be an independent dialect.
----------------
mehdi_amini wrote:
> jeanPerier wrote:
> > tschuett wrote:
> > > Isn't this point about the expressiveness of MLIR? What are the engineering costs of adding a FIRX dialect for the higher ops?
> > > 
> > > What would be the benefit of having a separate dialect?
> > The benefit of having a separate dialect is to strongly split the high level ops that require information about Fortran variables to be retrievable in the IR (via the fir.def/fir.ref) and the current operations that are lower level and do not require such information.
> > After the translation pass of variable related operation, this dialect would be illegal.
> > 
> > There is a precedent in FIR with the fir::cg dialect ops (https://github.com/llvm/llvm-project/blob/main/flang/include/flang/Optimizer/CodeGen/CGOps.td) that helps simplifying the addressing and emboxing operation before codegen.
> > 
> > I think the engineering cost are mostly about having different td files, headers and .cpp for these new ops and types, but also to register the dialect in the passes that will work with it. That last point may be a bit more annoying (it does not matter with the fir::cg dialect because codegen is the only pass meant to be run with this dialect).
> > 
> > I do not have a strong opinion here.
> >  to register the dialect in the passes that will work with it.
> 
> To be clear: registration is only ever needed in a pass that produces entities from a dialect (ops, attributes, types) when this dialect isn't in the input already. So a pass lowering from a high-level dialect to a lower-level dialect only defines the lower-level dialects to be registered, no need to do anything for the high-level dialect.
> Similarly passes that transform within the same dialect don't need to declare anything at all.
Thanks for clarifying this point, then I do not see big engineering costs to splitting high level ops depending on the Fortran variable and expression concepts and low level operations operating on memory and simple (integer/floating point/complex) SSA values.

Maybe the new fir.declare op should still belong to FIR directly though. I think it would be valuable to keep it until as late as possible (it is not doing anything other than bookkeeping some Fortran level information about  memory storages, which can be useful until the end).


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D134285/new/

https://reviews.llvm.org/D134285