Expanding an ADDSI3 into 2 x ADDHI3 does not work

["mhauff at bigpond dot net dot au" <mhauff at bigpond dot net dot au>]

I have a fictitious machine which has a word size of 8-bits but can handle 16-bit adds and 16-bit mov's.  I am trying to build the most efficient support for handling an addsi3 insn.  My problem is that if I try to split up the addsi3 insn into a couple of addhi3 insns (using a define_expand template) the compiler appears to ignore this declaration and proceeds to implement addsi3 as a bunch of addqi's along with some carry propogation rtx's.  i.e. the compiler defaults to the word size of the machine and I can't seem to override this.
I could allow it to go and create its big long list of addqi's etc and then use some insn combining method such as a peephole optimizer but this seems really inefficient to me - especially when I can explicitly state how a larger insn should be split.

If I use the following addsi3 template:
(define_insn "addsi3"
 [(set (match_operand:SI 0 "general_operand" "=g")
       (plus:SI (match_operand:SI 1 "general_operand" "g")
                (match_operand:SI 2 "general_operand" "g")))]
  ""
  "addsi3 %1 %2 %0	;(%1 plus %2)->%0" )

I can observe addsi being used in the assembly output of my test case.

If I use:
(define_expand "addsi"
 [(set (match_operand:SI 0 "general_operand" "=g")
       (plus:SI (match_operand:SI 1 "general_operand" "g")
                (match_operand:SI 2 "general_operand" "g")))]
  ""
  "{
  emit_insn (gen_addhi3 (custom_subword(operands[0], 0, SImode),
								 custom_subword(operands[1], 0, SImode),
								 custom_subword(operands[2], 0, SImode)));
	emit_insn (gen_addhi3 (custom_subword(operands[0], 1, SImode),
								 custom_subword(operands[1], 1, SImode),
								 custom_subword(operands[2], 1, SImode)));
	DONE; 
   }" )

the output becomes a mess of addqi, cmpqi, and branches.

Any help would be great.
Thanks
Marty