Jakub Kuderski
Jakub Kuderski
Would linking against the iree compiler lib instead of llvmsupport fix it?
I wonder if we are even fixing the right problem and if linking against the iree compiler is the right thing to do instead of fighting hidden/duplicate symbols etc. I...
@benvanik do you see any outstanding issues?
Oh man, the tuner will need some updating after this lands...
> > Oh man, the tuner will need some updating after this lands... > > Oh, I would have thought this doesn't break the tuner. Does the tuner use the...
> I'm wondering if it is more appropriate to label ops more carefully with specific `MemoryEffects`? Perhaps `NoMemoryEffect` would imply that it also doesn't allocate memory for a new output...
cc: @krzysz00 @MaheshRavishankar Feel free to comment / edit if I missed something that we discussed when we talked about it at the end of Feb.
I've learned that there's a dedicated op for alignment: https://mlir.llvm.org/docs/Dialects/MemRef/#memrefassume_alignment-memrefassumealignmentop So there's probably no need to go and add `alignment` attributes to all the memory access ops?
Well, it depends if you can infer the alignment from the indexing math. But +1 that all this seems fragile.
First PR in the series: * https://github.com/llvm/llvm-project/pull/144344