On 2 April 2014 13:08, Anthony Green <gr...@moxielogic.com> wrote: > I though the answer was to simply add something like this... > > (define_insn "zero_extendqisi" > [(set (match_operand:SI 0 "register_operand" "=r") > (zero_extend:SI (match_operand:QI 1 "register_operand" "r")))] > "" > "; ZERO EXTEND (comment for debugging)")
That pattern is obviously not outputting valid code. You should make this a define_insn_and_split, with an r/r alternative that is split (after reload) as necesary into shifts, and an m/r alternative that outputs a load. sprinkle with rtx_cost adjustments as necessary. > But nothing changes in the example above. LOAD_EXTEND_OP can also avoid some unnecesary expansions. ALthough we still have a long-standing issue of unnecessary extensions for narrow integer types passed in/out of functions, and loaded from volatile memory.
