On Thu, Mar 01, 2012 at 10:27:02AM +0000, Dave Martin wrote:

So, where there's a compelling reason to inline these things, we can use
the existing techniques if we're alert to the risks.  But in cases where
there isn't a compelling reason, aren't we just inviting fragility
unnecessarily?

The practical experience from the kernel suggests that there isn't a
problem - that's not to say that future versions of gcc won't become
a problem, and that the compiler guys may refuse to fix it.

I think it's a feature that we should be pressing gcc guys for - it's
fairly fundamental to any programming which requires interfaces that
require certain args in certain registers, or receive results in
certain registers.

The options over this are basically:
1. refusing to upgrade to any version of gcc which does not allow
   registers-in-asm
2. doing the store-to-memory reload-in-asm thing
3. hand-coding veneers for every call to marshall the registers

Each of those has its down sides, but I suspect with (1), it may be
possible to have enough people applying pressure to the compiler guys
that they finally see sense on this matter.