Re: [PATCH 08/16] btrfs: move incompat and compat flag helpers to fs.c

[David Sterba <dsterba@suse.cz>]

On Tue, Oct 11, 2022 at 12:33:27PM +0200, David Sterba wrote:
> On Wed, Sep 14, 2022 at 01:18:13PM -0400, Josef Bacik wrote:
> > These helpers use functions not defined in fs.h, move them to fs.c to
> > keep the header clean.
> 
> This is one of the examples where functions are switched from inline to
> out-of-line, so I'll reply for all of them.
> 
> > +bool __btrfs_fs_incompat(struct btrfs_fs_info *fs_info, u64 flag)
> > +{
> > +	struct btrfs_super_block *disk_super;
> > +	disk_super = fs_info->super_copy;
> > +	return !!(btrfs_super_incompat_flags(disk_super) & flag);
> > +}
> 
> This needs a separate function and unlike for the inline it needs a
> call which has some impact on the instruction flow. This should be
> evaluated if the un-inlined function is on a hot path or not.
> 
> In this case I think the incompat bits should be inline as it's in many
> places part of switching access to some structures or lookup.
> 
> At minimum some numbers and assembly examination should be done before
> ding that. The header cleanup part is nice but if the changes have
> performance implications it's suddenly not so easy.
> 
> I'll do some analysis to get an idea how bad/good it is.

Not that good. When inlined it's basically

	mov memory, rax
	mov memory(rax), rax
	test 0x1, rax

while uninlined it's

	mov memory, rdi
	mov memory, rsi
	call
	test rax, rax

the function itself is 4 instructions but with some additional ones
compared to the inlined version (2).