Re: [RFC] New intermediate languages

[Michael Meissner <meissner at cygnus dot com>]

On Thu, Feb 01, 2001 at 01:55:01AM -0500, Diego Novillo wrote:
> These are some of the general properties of MRTL.
> 
> - Reduced target dependencies. The RTL patterns generated at this
>   stage are fixed and have well-known semantics. Some
>   architectural features will likely creep in, like word size,
>   number of physical registers and whatnot.
> 
> - No information loss in the tree->MRTL transition. After the
>   MRTL form is transformed it will not be possible to go back
>   to the tree information.

Obviously we do have to keep the part of the tree interface that the debug
support uses in MRTL and RTL.

> - Control flow representation is similar to RTL. No structured
>   control-flow, everything is represented with jumps and
>   compare/branch.
> 
> - Arrays are lowered to pointer arithmetic and structures
>   flattened. No libcalls are generated for floating point and/or
>   integer arithmetic.
> 
> These patterns would be generated directly from stmt.c/expr.c.
> The plan is to write two passes tree->MRTL and MRTL->RTL which
> should not change existing behaviour. Once this is done, I would
> start building a CFG and an SSA form for MRTL (this can be mostly
> ripped off the current implementation). At that point we can
> start porting scalar transformations to the new form.
> 
> 
> Data types
> ----------
> 
> Use the same modes defined in RTL.
> 
> 
> Control Statements
> ------------------
> 
> (insn)
> (mcall (addr) (arg1) (arg2) ... (argN))
>     This is a new pattern for function calls. Alternatively we
>     could define an abstract stack machine and expand calls using
>     the existing code. I'd rather see a more compact
>     representation, though.

This probably  should be:

	(mcall (addr) [(arg1) (arg2) (arg3) ...])

in order to handle the variable number of arguments.

> (jump_insn (set (pc) (expr)))
>     Standard jump. expr can be either a label_ref or an
>     if_then_else for unconditional and conditional jumps.
>     
> (return)
> (return X)
>     Standard return with and without a value.

The second needs a new RTL, since other than vectors all RTL have a fixed
number of arguments.

> (code_label)
> (barrier)
> (note NOTE_*)
> 
> 
> Expressions
> -----------
> 
> Constants:
>   (const_int i)
>   (const_double:m addr i0 i1)
>   (const_string str)

We really should come up with a replacement for const_double that can be used
for all constants.

-- 
Michael Meissner, Red Hat, Inc.  (GCC group)
PMB 198, 174 Littleton Road #3, Westford, Massachusetts 01886, USA
Work:	  meissner@redhat.com		phone: +1 978-486-9304
Non-work: meissner@spectacle-pond.org	fax:   +1 978-692-4482