Re: [PATCH] SPU: Use FRAME_GROWS_DOWNWARD

[<trevor_smigiel@playstation.sony.com>]

Stefan,

In the "stack_protect_set" and "stack_protect_test_si" patterns, the
type attribute should be multi1. This indicates that the first
instruction is in the odd pipeline, which allows slightly better
scheduling of these patterns.

In "stack_protect_test_si", shouldn't the "+r" of the first operand be
"=&r"?  Also, please change "register_operand" to "spu_reg_operand".
Based on the code, it is not necessary now, but it protects us from
possible issues in the future.

I'm a little suspicious of the memory predicates, constraints, and the
hard coded load/store instructions.  How do we know that the memory rtl
will always match the hard coded instructions?

How did you compare the test suite results?  Did you use
contrib/compare_tests?  Or just compare the numbers/statistics printed
at the end?  (Use compare_tests.)

Trevor


* Stefan Schulze Frielinghaus <XXSCHULZ@de.ibm.com> [2008-10-21 07:00]:
> 
> Hi,
> 
> to use the Stack Smashing Protector the frame needs to grow downwards.
> Attached patch accomplishes that for the Cell B.E. SPU architecture.
> 
> Tested against regression suite without any difference in pass/fail
> statistics (except the extra test passes for SSP)..
> 
> OK for mainline?
> (See attached file: spu.patch)
> Mit freundlichen Gruessen / Kind regards
> 
> Stefan Schulze Frielinghaus
> 
> IBM Systems & Technology Group, Systems Software Development
> SW Linux on Cell BE Development & Evaluation
> -------------------------------------------------------------------------------
> 
> IBM Deutschland
> Schoenaicher Str. 220
> 71032 Boeblingen
> Phone: ++49-(0)7031-16-2173
> E-Mail: xxschulz@de.ibm.com
> -------------------------------------------------------------------------------
> 
> IBM Deutschland Research & Development GmbH
> Vorsitzender des Aufsichtsrats: Martin Jetter
> Geschaeftsfuehrung: Herbert Kircher
> Sitz der Gesellschaft: Boeblingen
> Registergericht: Amtsgericht Stuttgart, HRB 243294