[Mlir-commits] [mlir] [MLIR][LINALG] Bubble up extract slice above Linalg Op - address multi-user case (PR #79078)

[llvmlistbot at llvm.org]

MaheshRavishankar wrote:

> > can have undesired consequence if applied generally without any user control
> 
> Could you please elaborate. I dont really see the difference between one `extract_slice` user and say two `extract_slice`users using two different part of the result of linalg.op. ... other than re-computation in case the two extracts intersection is non empty set. The null check seems not possible for dynamic shapes. Would you consider the optimization ok for static shapes although that does not really help.

The redundant computation is exactly what I am referring to. In such cases the introduction of redundant computation must be opt in and be controllable from the caller. 

https://github.com/llvm/llvm-project/pull/79078