[PATCH 21/25] GCN Back-end (part 2/2).

[Jeff Law]

On 11/12/18 10:52 AM, Andrew Stubbs wrote:
> On 12/11/2018 17:20, Segher Boessenkool wrote:
>> If you don't want useless USEs deleted, use UNSPEC_VOLATILE instead?
>> Or actually use the register, i.e. as input to an actually needed
>> instruction.
> 
> They're not useless. If we want to do scalar operations in vector
> registers (and we often do, on this target), then we need to write a "1"
> into the EXEC (vector mask) register.
Presumably you're setting up active lanes or some such.  This may
ultimately be better modeled by ignoring the problem until much later in
the pipeline.

Shortly before assembly output you run a little LCM-like pass to find
optimal points to insert the assignment to the vector register.  It's a
lot like the mode switching stuff or zeroing the upper halves of the AVX
regsiters to avoid partial register stalls.  THe local properties are
different, but these all feel like the same class of problem.

> 
> Unless we want to rewrite all scalar operations in terms of vec_merge
> then there's no way to "actually use the register".
I think you need to correctly model it.  If you lie to the compiler
about what's going on, you're going to run into problems.
> 
> I might investigate putting the USE inside an UNSPEC_VOLATILE. That
> would have the advantage of letting combine run again. This feels like a
> future project I'd rather not have block the port submission though.
The gcn_legitimate_combined_insn code isn't really acceptable though.
You need a cleaner solution here.

> 
> If there are two instructions that both have an UNSPEC_VOLATILE, will
> combine coalesce them into one in the combined pattern?
I think you can put a different constant on each.

jeff