Re: [PATCH] * config.sub: Consolidate OS version checking

[John Ericson]

On 05/24/18 01:00, Ben Elliston wrote:
> Applied, thanks.  I'm just waiting on the -sequent patch, John, and
> then I think I am all caught up.
>
> Cheers, Ben

Ah right. That's why the 6/6 didn't apply in the first place. I didn't 
realize that until now; since the ptx one deletes code, but revised 6/6 
actually technically subsumes it.

> Also, don't you mean "ptx" (not "ipx")?

Indeed

> Can you please add a test case or two to config-sub.data that shows 
the correct behaviour?

Well it depends what we want. Right now

> valid_cpu-valid_os

in the most general case normalizes to

> valid_cpu-unknown-valid_os

with a big exception of valid_cpu being x86 of some sort, in which case 
it normalized to

> valid_cpu-pc-valid_os

This matters because at the bottom "unknown" is defaulted to various 
actual vendors where unambigous, while "pc" isn't. This means we have

> $ ./config.sub x86_64-unknown-ptx
> x86_64-sequent-ptx

but

> $ ./config.sub x86_64-pc-ptx
> x86_64-pc-ptx

and

> $ ./config.sub x86_64-ptx
> x86_64-pc-ptx

There's a few ways forward. The simplest is just to get rid of the x86 
specical case --- with LLVM's influence on configs these days I don't 
think "unknown" would be as surprising as the comment in the code 
complains. Another solution is to across the board "refine" "unknown" 
and "pc" alike to more specific vendors. A third solution would be to 
default to unknown up above, but then refine to pc as a last resort at 
the bottom. The one complication with that is that *-unknown is a valid 
basic_machine pattern, which I think is a bit permissive, but that can 
be worked around with boolean flag variables.

For reference:

5308529936fc0c7bbbc8e04ad31974982481ecd0 commit that added the ptx exception

3836f7b0c4f252e56aa1124a9b235ba62c2a45da commit that added the *-unknown 
pattern