This adds patterns for v16i16's vecreduce, using all the existing code to go via an i32 VADDV/VMLAV and truncating the result. Differential Revision: https://reviews.llvm.org/D85452
733 KiB
733 KiB