commit ea72d000f9ed3e11fb3192177aaf2ecd11bee4e8
author Max191 <44243577+Max191@users.noreply.github.com> Thu Jan 29 12:00:23 2026 -0500
committer GitHub <noreply@github.com> Thu Jan 29 12:00:23 2026 -0500
tree 65875f3573e83254bc9919e8178e6abada2ffcf0
parent 34d2446a50a94668ef878f49d76844980e6385cc

[GPU] Don't swap expand with slice in the same block (#23267)

Prevents swapping tensor.extract_slice ops with tensor.expand_shape ops
when they are in the same block during GPUFuseAndHoistParallelLoops.
This pattern is intended to be a tiling fusion pattern for
tensor.expand_shape operations, which always begin with the ops in
different blocks. When the ops are in the same block, this does not open
up more fusions, and instead moves the extract_slice further compute ops
that produce loads. This can block later optimizations that fold the
slice into loading operations like vector.transfer_read.

This restriction is done through a control function to maintain the
ability to swap same-block slice + expand ops, in case it is needed for
other use cases.

ci-extra: test_torch

---------

Signed-off-by: Max Dawkins <max.dawkins@gmail.com>