Re: [Patch] MIPS: Supply correct call clobbered information for global pointer.

[David Daney <ddaney at avtrex dot com>]

Richard Sandiford wrote:
> David Daney <ddaney@avtrex.com> writes:
> > This patch supplies correct call clobbered information for the global
> > pointer for the o32 and o64 ABIs.
>
> Thanks for the patch.  Unfortunately, although this is clearly correct
> information at the ABI level, I don't think it's correct at the RTL level.
> Until the call patterns are split -- if they ever are -- the patterns
> include code to restore $gp.  This means that $gp is preserved across
> the unsplit call instructions, and that the current non-definition
> of PIC_OFFSET_TABLE_REG_CALL_CLOBBERED is correct at that point.
>
> The post-split calls explicitly clobber $gp, so I think we are already
> providing correct information as far as the call instructions themselves
> are concerned.  I'm not claiming I can construct a test that wouldn't
> work after the patch (to be honest, I haven't even tried).  But it seems
> dangerous in principle.
>
> I've got a series of patches that I hope to submit before 4.4 is out.
> As it happens, one of those patches included a fix for the same problem,
> so I've split it out here.  It brings about a 0.75% improvement in the
> CSiBE score for little-endian MIPS64 code at -Os.
>
> If you're happy with this approach, I can commit the patch separately
> from the other stuff.  (I'll test it on its own overnight.)

Eliminating the dead code is my main concern. I have not tested your 
patch, but assume that it is as effective as mine in eliminating the 
dead restore of $gp.

One thing that concerns me is that it feels like we are lying to 
data-flow.  In df-scan.c (df_get_exit_block_use_set) 
PIC_OFFSET_TABLE_REG_CALL_CLOBBERED is used to check if $gp is used in 
the exit block, which it is not in TARGET_OLDABI.

Since you are the maintainer and have a preference for your patch, you 
should commit it.  I have no objection to it.


David Daney