Re: [PATCH] MIPS: Opcode membership proposal

[David Daney <ddaney@caviumnetworks.com>]

On 12/18/2010 03:42 AM, Maciej W. Rozycki wrote:
> Hi,
>
>   While working on the microMIPS change I have stumbled across the GAS hack
> to support ISA subsetting for the Octeon processors, used to exclude some
> of the CP0 operations.  I decided to implement this properly, based on my
> previous experiments in this area.
>
>   Here's my proposal.  It adds an "exclusions" field to the mips_opcode
> structure.  The new field defines a set of CPU membership flags to exclude
> an opcode for just like they are used with the "membership" field,
> overriding the membership information in the latter field.  Currently only
> CPU_OCTEON has been made to work like this and flags may be added in the
> future as a need arises.  An ISA flag could be handled in a similar way,
> although one is unlikely to be ever needed here.
>
>   The advantage of this approach compared to the current one is not only
> GAS correctly excludes the unsupported opcodes, but `objdump' does as
> well.  And it scales easily, requiring no changes outside opcodes.
> Finally, it makes it clear which opcodes are really unsupported on Octeon
> -- with the current approach it is a bit buried down within code.
>
>   As I've thought adding a trailing zero to initialise the new field
> explicitly throughout the opcode tables would be unnecessary obfuscation
> I've decided to initialise it to a non-zero value for opcodes that require
> it only and allow the source files defining the MIPS opcode tables to have
> uninitialised structure members with the -Wno-missing-field-initializers
> GCC option.  I think it is a good compromise -- these tables are validated
> at run time too and the files contain no other code.
>
>   I have decided to retain the CP pinfo flag defined with the change that
> introduced the current Octeon opcode exclusion code.  The flag is not used
> anymore, but the effort previously taken to annotate the relevant opcodes
> may be useful for something else in the future.
>
>   No regressions for mips-sde-elf nor mips-linux-gnu.  Applies on top of
> the microMIPS change.
>

For the record, this seems like a sane approach.  I have not gone 
through the patch with a fine toothed comb, but it seems like an 
improvement over the current approach.

Andrew Pinski may want to take a look, but from my point of view, if 
Richard likes it, I think it should be committed.

Thanks,
David Daney