Reload bug

[Eric Botcazou]

> > What you have done is to teach the reload pass to silently transform any
> > non-accessible part of a multi-word register into the low part.
>
> In the case the subreg is undable (ie it is in the SSE), or can this
> happen in the integer register too?

Only when the logic should have triggered, i.e with MMX and SSE regs but not 
with general regs since it doesn't trigger for them.

> Or we can backport the CANNOT_CHANGE_MODE_PATCH.

Big patch.

> I would be happy about any alternative for 3.2

Well, there aren't too many possibilities:
- either we prevent such invalid subregs from being created in the first 
place (but it appears that doing so brings about collateral damages in the 
form of (subreg:SI xmm:DI 0) being now rejected?),
- or we restore the machinery that was able to handle them.

-- 
Eric Botcazou