Re: Query with legitimate_constant_p

[Dan Towner <dant at picochip dot com>]

Geoff Keating wrote:

> 
> > D.Towner wrote:
> 
> > >  I am writing a new port of GCC. I want to prevent symbol_refs from being
> > > used as immediate operands, so I am specifically rejecting these in
> > > legitimate_constant_p.
> >
> > You don't want to do this...
> >
> > > However, this also means that the addresses of
> > > symbol_refs can't be computed (e.g, loading the address into a  register),
> >
> > because clearly you _do_ use SYMBOL_REF as an immediate operand.
> >
> > > where previously a `mov??3' instruction was used to move the  symbol_ref
> > > address into a register. This no longer works, as the mov  instruction can't
> > > accept an immediate symbol_ref any more. How do I fix  this?
> >
> > Instead of doing that, use `n' constraints, and suitable predicates
> > where appropriate, to prevent GCC generating symbol references as
> > immediate operands where they won't work.
 
That doesn't work though. Consider the following:
 
   reg0 := symbol_ref("s")
   reg1 := mem(reg0 + 2)
 
Because the symbol_ref is a valid constant, gcc propagates the constant
into the memory load:
 
   reg1 := mem(symbol_ref("s") + 2)
 
I don't allow this to be generated directly, but the constant
propagation causes it to be generated, without appearing to recheck the
constraints
on the instruction. Thus, doesn't this mean that symbol_refs must fail
legitimate_constant_p in order to prevent them being propagated into
instructions?

=============================================================================
Daniel Towner
picoChip Designs Ltd., Riverside Buildings, 108, Walcot Street, BATH,
BA1 5BG
dant@picochip.com
07786 702589