llvm-project

87e8e7d8 - [SPIRV] Implement lowering for llvm.matrix.transpose and llvm.matrix.multiply (#172050)

[SPIRV] Implement lowering for llvm.matrix.transpose and llvm.matrix.multiply (#172050) This patch implements the lowering for the llvm.matrix.transpose and llvm.matrix.multiply intrinsics in the SPIR-V backend. - llvm.matrix.transpose is lowered to a G_SHUFFLE_VECTOR with a mask calculated to transpose the elements. - llvm.matrix.multiply is lowered by decomposing the operation into dot products of rows and columns: - Rows and columns are extracted using G_UNMERGE_VALUES or shuffles. - Dot products are computed using OpDot for floating point vectors or standard arithmetic for scalars/integers. - The result is reconstructed using G_BUILD_VECTOR. This change also updates SPIRVPostLegalizer to improve type deduction for G_UNMERGE_VALUES, enabling correct type assignment for the intermediate virtual registers generated during lowering. New tests are added to verify support for various matrix sizes and element types (float and int).