[llvm-dev] About OpenMP dialect in MLIR (original) (raw)

Michael Kruse via llvm-dev llvm-dev at lists.llvm.org
Fri Feb 14 12:23:30 PST 2020

Am Fr., 14. Feb. 2020 um 12:21 Uhr schrieb Vinay Madhusudan via llvm-dev <llvm-dev at lists.llvm.org>:

I would like to quote Chris here:

“if you ignore the engineering expense, it would clearly make sense to reimplement the mid-level LLVM optimizers on top of MLIR and replace include/llvm/IR with a dialect definition in MLIR instead.“ -- http://lists.llvm.org/pipermail/llvm-dev/2020-January/138341.html

IMHO, it's not just the engineering expense, but also additional overhead from having a more general data structure that clang does not need. In some sense, LLVM-IR has been designed to match the semantics of C, such that a more general representation makes less sense. There are still opportunities, e.g. representing C++ virtual methods instead lowering to a vtable lookup. This could make devirtualization easier.

However, it seems nobody is currently pushing for this change to happen, there is not even an RFC on whether the community wants this change. As such, I would not plan on using MLIR if your frontend language is C/C++/Objective-C.

Michael

More information about the llvm-dev mailing list