Re: [PATCH 2/6] FRV: Fix FRV arch compile errors [try #3]

[David Howells <dhowells@redhat.com>]

Andrew Morton <akpm@osdl.org> wrote:

> - The __init-style tags on declarations don't actually do anything and
>   the compiler doesn't check for consistency with the definition - it's
>   best to just omit it from the declaration.

Well, you're wrong.  They *do* do something.  They stop the compiler using the
register-relative addressing reserved for small data.  If this isn't in there,
then the linker will spit out a relocation error.

On fixed-size instruction architectures, it takes several instructions to
dereference an absolute address, so you try and squeeze all your small data
into a section of its own, plonk a register in the middle of it, and use
indirect-addressing relative to that register.  This saves you two or more
instructions and a register on each normal global variable access.

For instance, on FRV, it may change:

	sethi.p		%hi(nr_kernel_pages),gr4
	setlo		%lo(nr_kernel_pages),gr4
	ld		@(gr4,gr0),gr5

Into:

	ld		@(gr16,%gprel(nr_kernel_pages)),gr5

Small data is 1-byte, 2-byte, 4-byte or 8-byte values that aren't arrays or
const.

Setting a section marker on the variable *declaration* disables the
gp-relative addressing and forces the longer absolute addressing.

> - Setting nr_kernel_pages to be unloaded at free_initmem() seems risky.

That's nothing to do with my patch.

> - nr_kernel_pages is actually __meminitdata.

Okay, I'll fix my patch to be that instead.

David