How to bufferize scf.while?

MLIR
1

Assume you have the following code. Is it possible to bufferize it and then synthesize deallocation? Am I correct in saying that scf.while is not yet covered by --scf-bufferize? I have verified, and if I put scf.for instead of the scf.while, it works. Funny enough, scf.if is not handled, either.

scf.while (%x0=%c0):(tensor<256xi16>)->(tensor<256xi16>) {
    %cond = constant 1: i1
    scf.condition(%cond) %x0:tensor<256xi16>
} do {
^bb0(%x0:tensor<256xi16>):
    ... // some tensor operations that require deallocation
    scf.yield %x0:tensor<256xi16>
}
2

Correct, scf.while it is not currently handled by scf-bufferize. It should be easy to extend SCF/Transforms/StructuralTypeConversions.cpp by doing something similar to what we do for ForOp. You’ll also need a new pattern for scf.condition. This code is tested via scf-bufferize and you can add a test in SCF/bufferize.mlir. Patches welcome!

Also, scf.if should be handled by scf-bufferize. Do you have a test case that is broken for scf.if?

3

I have a piece of code with an scf.if embedded inside an scf.for, and the whole pipeline of bufferization breaks on the synthesis of the deallocation (I encode output arguments as input memrefs).

4

It would be good to file your failing test as a bug at https://bugs.llvm.org/