Re: [PATCH]: R10000 Needs LL/SC Workaround in Gcc

[Kumba <kumba@gentoo.org>]

Richard Sandiford wrote:
> 
> As Maciej said, this should really be controlled by an -mfix-r10000
> command-line option, not by the _MIPS_ARCH_* macro.  (In this context,
> _MIPS_ARCH_* is a property of the compiler that you're using to build
> gcc itself.)
> 
> There are two ways we could handle this:
> 
>   - Make -mfix-r10000 require -mbranch-likely.  (It mustn't _imply_
>     -mbranch-likely.  It should simply check that -mbranch-likely is
>     already in effect.)
> 
>   - Make -mfix-r10000 insert nops when -mbranch-likely is not in effect.

Does using -mbranch-likely change the output of those specific asm commands that 
my original patch was altering?  Or will -mfix-r10000 need to not only check the 
status of -mbranch-likely and set it if not set, but also need to modify the 
referenced beq/beqzl sets in mips.h?

If so, I assume a test for both TARGET_FIX_R10000 and TARGET_BRANCHLIKELY would 
be needed, and then if TARGET_BRANCHLIKELY doesn't exist, but TARGET_FIX_R10000 
is, insert 28 nops before beq.  Sound correct?


On setting -mbranch-likely, I found what I think is the appropriate section in 
mips.c around Line 13810:

   /* If neither -mbranch-likely nor -mno-branch-likely was given
      on the command line, set MASK_BRANCHLIKELY based on the target
      architecture and tuning flags.  Annulled delay slots are a
      size win, so we only consider the processor-specific tuning
      for !optimize_size.  */
   if ((target_flags_explicit & MASK_BRANCHLIKELY) == 0)
     {
       if (ISA_HAS_BRANCHLIKELY
           && (optimize_size
               || (mips_tune_info->tune_flags & PTF_AVOID_BRANCHLIKELY) == 0))
         target_flags |= MASK_BRANCHLIKELY;
       else
         target_flags &= ~MASK_BRANCHLIKELY;
     }
   else if (TARGET_BRANCHLIKELY && !ISA_HAS_BRANCHLIKELY)
     warning (0, "the %qs architecture does not support branch-likely"
              " instructions", mips_arch_info->name);

I'm kind of thinking that the -mfix-r10000 setting to include -mbranch-likely 
would fit here (Assuming this is what can enable/disable that option via 
MASK_BRANCHLIKELY), but if I'm reading it right, optimizing for size disables 
brach-likely instructions.  Shouldn't -mfix-r10000 override that?

Would an equivalent conditional like this be close?:

       if (ISA_HAS_BRANCHLIKELY
           && ((optimize_size || TARGET_FIX_R10000)
               || (mips_tune_info->tune_flags & PTF_AVOID_BRANCHLIKELY) == 0))


Also, does anyone have a copy of the R10000 Silicon Errata documentation kicking 
around?  Thiemo brought up a point that we may need ssnop instead of nop, but 
I'd need to check the errata for that, and that doesn't seem to exist anywhere 
anymore.  I found an old link to it on MIPS' site, but nothing else.  I've only 
got Vr10000 manuals from SGI and NEC, and they don't seem to cover 
revision-specific errata any.

Thanks!

-- 
Joshua Kinard
Gentoo/MIPS
kumba@gentoo.org

"The past tempts us, the present confuses us, the future frightens us.  And our 
lives slip away, moment by moment, lost in that vast, terrible in-between."

--Emperor Turhan, Centauri Republic