llvm-project

c5ba2e4d - [MLIR][XeGPU] Hoist i4 extensions above lane shuffles (#234760)

[MLIR][XeGPU] Hoist i4 extensions above lane shuffles (#234760) Hoist the i4-to-i8 portion of integer extensions through single-use lane-shuffle and vector data-movement chains before XeGPU-to-XeVM conversion. Legalize narrow casts into byte-level operations in a separate preprocessing stage so the existing bitcast-shuffle lowering can handle the widened fragments without repacking. Cover signed and unsigned extensions, supported fragment widths, layout-distribution shuffles, and unsupported consumers. Scope preprocessing to casts and generated operations to preserve existing conversion diagnostics.