llvm-project

106be215 - [VectorCombine] Fold bitcast(bitreverse(bitcast(x))) to fshl+bitreverse (#214362)

[VectorCombine] Fold bitcast(bitreverse(bitcast(x))) to fshl+bitreverse (#214362) Extend foldBitOrderReverseAndSwap to handle <2 x iN> element bitreverse patterns. Previously the fold only handled <N x i8> vectors using bswap to reorder octets. For 2-element vectors, the element swap is a rotate by half the bitwidth, which can be expressed as fshl(x, x, bitwidth/2). The transform rewrites: bitcast(bitreverse(bitcast(x))) -> bitreverse(fshl(x, x, N/2)) when the cost model indicates the scalar form is cheaper. For <2 x i8> we still prefer to use bswap, and we do not even consider to check if fshl is cheaper than bswap.