[PATCH] D69498: IR: Invert convergent attribute handling

[Mehdi AMINI via Phabricator via llvm-commits]

mehdi_amini added a comment.

In D69498#1735763 <https://reviews.llvm.org/D69498#1735763>, @arsenm wrote:

> In D69498#1731265 <https://reviews.llvm.org/D69498#1731265>, @mehdi_amini wrote:
>
> > Just thought about a slight variation on this: what about adding a flag on the datalayout (or a module flag) but not use it in the transformations/analyses, instead use it only when creating Function by always setting the `convergent` attribute when the flag is present? This would require to always have a Module passed to the Function constructor though (the C API already does, we would just need to update the C++ users).
> >
> > So creating a Function in a Module with this flag would have the convergent attribute set on creation (can be unset explicitly).
>
>
> This sounds almost too low level. The bitcode loading of a function shouldn’t be adding attributes to non-intrinsics, otherwise it’s changing the meaning of the program and not preserving the input


You're right, but I thought loading a function from bitcode would be 1) create the function (it may get a default convergent attribute added) and 2) set the exact attributes (which would override the default)


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D69498/new/

https://reviews.llvm.org/D69498