Standard ML of New Jersey
Version 110.32, December 22, 2000

Home page


This version is intended for compiler hackers. The version ought to be quite stable, however we have not run our full regression testing.


This version is being used to check point a large number of changes to MLRISC, which now uses infinite precision integers on all architectures. We have run the Basis regressions tests so the changes ought to be fairly robust.

There are a number of other CM, Basis, and Tool related changes as well.



This change fixes two problems:

  1. There was a proliferation of constructors in MLRISC to handle integers of various sizes, e.g.
        LI of int   | LI32 of  | LI64 of
    which are all replaced with
        LI of

  2. There is the ever nagging question of cross compiling to 64bits.

Compilation Speed

The x86 compiling the compiler is 9% slower than 110.31, however, a lot of this, and more, can be recovered. The compiler presently uses infinite precision in the front end (AST creation), which is converted to 31 and 32 bit integers in the middle end (FLINT and CPS), and then converted back to infinite precision in MLRISC (with the current changes).

Infinite precision in the middle end will allow a uniform treatment of optimizations across words and integers of different widths/sizes. Currently, most optimizations are biased towards integers and words represented in 31 bits only.

What's next

Not in chronological order:
  1. Use infinite precision in the middle end to allow a cleaner and more uniform optimization module for arithmetic.

  2. Support for the generation of infinite precision values in the back end.

  3. Changes to support the single toplevel declaration:
       structure LargeInt = IntInf


Added the missing bit level functions to the Basis IntInf structure.


Improvements to the 'nowhere' tool to handle OR-patterns. The nowhere tool is a syntactic extension that allows for guarded patterns. See MLRISC/Tools/Doc/nowhere.tex.