Re: [Patch] 1/3 Add new builtin __builtin_flush_icache().

[Mark Mitchell <mark@codesourcery.com>]

David Daney wrote:

> My new implementation has three cases for __builtin___clear_cache():
> 
> 1) No "clear_cache" insn is supplied but the target does supply a
> definition for CLEAR_INSN_CACHE.  In this case (arm, m68k IIRC) the new
> builtin will expand to a call to the existing __clear_cache in libgcc.
> 
> 2) A "clear_cache" insn is supplied.  For this case, the builtin expands
> to said insn.  The mips portion of the patch currently has the only
> implemtation of "clear_cache".
> 
> 3) Neither "clear_cache" or CLEAR_INSN_CACHE is defined.  This case
> covers most GCC targets including but not limited to i386 and x86_64. 
> In this case the builtin expands to nothing, so it is a nop.

I think one could make an argument that for case 3)
__builtin___clear_cache should still expant to __clear_cache (just as
with case 1).  That would allow you to provide a new libgcc.so in future
and fix binaries that were trying to clear caches, but failing.

But, I don't think that's a very good argument.  It means that on a
system with no cache, you now have to explicit define a clear_cache
instruction (expanding to nothing) in order to say "no, there really is
no cache to clear here".

> 2007-07-08  David Daney  <ddaney@avtrex.com>
> 
>    * builtins.def (BUILT_IN_CLEAR_CACHE): New builtin.
>    * builtins.c (expand_builtin___clear_cache): New function.
>    (expand_builtin): Call expand_builtin___clear_cache for
>    BUILT_IN_CLEAR_CACHE case.
>    * doc/extend.texi (__builtin___clear_cache): Document new builtin.
>    * testsuite/gcc.dg/builtins-64.c: New test.

:REVIEWMAIL: OK

Thanks,

-- 
Mark Mitchell
CodeSourcery
mark@codesourcery.com
(650) 331-3385 x713