How to handle subreg(mem(X)) after reload?
James E Wilson
wilson@specifixinc.com
Mon Jan 17 21:45:00 GMT 2005
Robert Baruch wrote:
> without a hard register. For strict base pointers, does that mean that
> there must exist, at a minimum, an addressing mode where base + offset
> is valid?
No. IA-64 for instance has no base+offset addressing modes. I believe
there is at least one other existing target with the same limitation.
GO_IF_LEGITIMATE_ADDRESS defines the addressing modes. The base
registers are used by reload. If we have an address that needs
reloading, we will generate a reload into a base register. So base
registers is just the set of registers that can be used to hold
addresses for load/store instructions.
The only potential problem here is that gcc assumes that addressing is
orthogonal, i.e. add address registers support all addressing modes. If
your target doesn't work this way, then a gcc port will be tricky, and
will probably require defining lots of extra constraints and special
patterns.
> 1. Does that mean FSR should be the only allowable base register?
I don't know enough about your target. Any register which can be used
to hold an address for load/store, and which supports all addressing
modes accepted by GO_IF_LEGITIMATE_ADDRESS, can be a base register.
> 2. Or is FSR disqualified as a base register because there is no
> addressing mode for FSR + offset?
No. If GO_IF_LEGITIMATE_ADDRESS does not include base+offset, then gcc
won't generate it.
> 3. Will I run into trouble if I have only one allowable strict base register?
That will make reload's job hard. If you have 3-operand instructions,
you may run into trouble if both input operands need reloads from
memory. If you have only 2-operand instructions, then it will probably
work out OK, as input operand reloads come before the instruction and
output operand reloads can be put after the instruction, so you may not
need two base registers at the same time.
You will just have to try it and see.
> 4. Or should I allow a few other registers and fix up the code to only
> use FSR later?
Don't know enough at this time.
--
Jim Wilson, GNU Tools Support, http://www.SpecifixInc.com
More information about the Gcc
mailing list