[Mlir-commits] [mlir] [mlir] Fix consumer fusion for producer with multiple results (PR #125915)

Prashant Kumar llvmlistbot at llvm.org
Sat Mar 8 04:57:12 PST 2025


pashu123 wrote:

> I was trying to reason through this and I dont think this actually works. This is pretty fragile and can produce incorrect code. Lets chat offline about this and get back. We probably have to drop this change, and try to do something a bit more constrained in the downstream use of consumer fusion.

Hi @MaheshRavishankar from the chat I inferred that if the operands of the consumer have different iteration order we can just bail out from the fusion. Shall I add that check?

https://github.com/llvm/llvm-project/pull/125915


More information about the Mlir-commits mailing list