On Wed, 2008-05-14 at 00:06 +0400, Ivan Kokshaysky wrote:

On Tue, May 13, 2008 at 09:59:17AM -0500, James Bottomley wrote:

quoted

But if that's true then the whole basis for our section based discards
is bogus, because the statement generating the jump table could be
inside the actual function body instead of being inlined, so if it's
failing on x86 as well, we likely need it fixed there too.

Correct.

quoted

quoted

It would be good, but in the meantime the 'noinline' fix seems to be
a most sensible option...

We can certainly add it as a short term option.  However, given the push
in certain quarters to make even more use of sections as a means of
discarding code, we're going to have to ask someone to fix the compiler.

Then maybe building with -fno-jump-tables option would be better solution.
I don't think that it will have a noticeable impact on code size or
performance, especially on x86.

So we need something like this (assuming gcc eventually gets fixed,
say gcc 4.5 ;)

KBUILD_CFLAGS	+= $(call cc-ifversion, -lt, 0405, \
			$(call cc-option, -fno-jump-tables))

This is way beyond a SCSI issue.  Time to take it to linux-arch I
think ... do you want to do the honours?

James