llvm dev - Oct 2021 - scratch register for spill/reload

I have a target architecture with a register class that requires a GPR as a
scratch register to implement spill and reload operations. I have experimented
with just using a pseudo instruction then scavenging a GPR on demand when that
pseudo is lowered but that is not robust. Is there a protocol to allow a point
GPR to be allocated when a spill or reload is generated?
I tried to add a suitable virtual register as an implicit def, but this failed,
at least in the fast register allocator, because the generated instruction is
apparently not visited.
My fall back was to add an implicit definition of an ABI defined scratch
physical register. That generated incorrect code for some tests with the default
allocator.
How have other targets managed this problem?
Thanks
david


-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20211009/4e96cb50/attachment.html>

On 10/9/21 11:00 AM, David Callahan via llvm-dev wrote:
> I have a target architecture with a register class that requires a GPR as a
scratch register to implement spill and reload operations. I have experimented
with just using a pseudo instruction then scavenging a GPR on demand when that
pseudo is lowered but that is not robust. Is there a protocol to allow a point
GPR to be allocated when a spill or reload is generated?  
> 
> I tried to add a suitable virtual register as an implicit def, but this
failed, at least in the fast register allocator, because the generated
instruction is apparently not visited. 
> 
> My fall back was to add an implicit definition of an ABI defined scratch
physical register. That generated incorrect code for some tests with the default
allocator. 
> 
> How have other targets managed this problem? 
I can't answer how the ppc backend solves this, but the condition registers
and link
register on ppc/ppc64 have no load and store registers, so to spill them, you
first
have to copy them to/from a GPR and spill them that way.  If this is similar to
your
register class limitation, you could see how the ppc backend does this.

Peter