Program extraction using native integers/words (not bignums) from Isabelle theory -

this question comes in context isabelle used formal software development in mind more pure maths theorization in mind (and standalone developer's context).

seems @ best, sml programs generated isabelle theory, use sml's intinf.int, not native integer type, int.int; if code_target_int, code_binary_nat or code_target_nat used. investigation of these theories sources seems confirm it's can do. native platform integers may required multiple reasons, including efficiency , case sml imperative program optionally translated imperative language subset (ex. c or ada), relevant when theory relies on imperative_hol theory. codegen.pdf document comes isabelle distribution, did not it, except in suggesting first of options below.

options may be:

• not using isabelle's int , nat , re‑create new numeric type scratch, use code_printing commands (with type_constructor , constant) give native platform representation , operations (implies inclusion of range limitations in way in theory) : must tedious, although unlikely error‑prone hope, due formal environment. note seems feasible isabelle's own int , nat… makes code generation fails, , nothing tells constants missing in code_printing command.
• if sml program compiled directly (ex. mlton), tweak sml environment replacement intinf structure : may unsafe or not feasible, , still requires embed range limitations in theory, previous options may better one.
• touch generated program change intinf int : easy, safe? (at least, intinf implements same signature int do, may it's safe). above, requires specifies bounds in theory in way, it's ok this.
• dive isabelle internals : surely unreasonable, worse second option.
• there exist word theory, according readings, it's seems not suited purpose.

are other known options not listed here? comments on listed options?

if there no ready‑to‑cook solutions (i feel there no @ time), hints or tracks best known? (ex. links documents, mentions of concepts).

update

points #2 , #3 of list, may ok (if is) if there single integer type. if program use more one, it's not applicable.

directly generating native words isabelle int unsound, because formalisation not take overflow account exists in reality.

it looks afp entry native_word want, though: http://afp.sourceforge.net/entries/native_word.shtml