Re: [PATCH 1/3] Add ARM tune file overhaul based largely on work from Mark Hatle

[Mark Hatle <mark.hatle@windriver.com>]

On 7/27/11 8:33 AM, Richard Purdie wrote:
> On Wed, 2011-07-27 at 13:17 +0100, Phil Blundell wrote:
>> On Tue, 2011-07-26 at 13:44 +0100, Richard Purdie wrote:
>>> +TARGET_FPU = "${@d.getVar('ARMPKGSFX_FPU', True).strip('-') or 'soft'}"
>>
>> This seems a bit backwards.  Shouldn't TARGET_FPU be the primary
>> variable and then the package suffix be computed from that, rather than
>> vice versa?
> 
> It's been "fun" to use the rather limited constructs we have in these
> variables to construct the end result. I suspect this way around, it was
> the easiest way to get the right variables in the right places.
> 
>>> +ARMPKGSFX_THUMB .= "${@bb.utils.contains("TUNE_FEATURES", [ "armv4", "thumb" ], "t", "", d)}"
>>> +ARMPKGSFX_THUMB .= "${@bb.utils.contains("TUNE_FEATURES", [ "armv5", "thumb" ], "t", "", d)}"
>>> +ARMPKGSFX_THUMB .= "${@bb.utils.contains("TUNE_FEATURES", [ "armv6", "thumb" ], "t2", "", d)}"
>>> +ARMPKGSFX_THUMB .= "${@bb.utils.contains("TUNE_FEATURES", [ "armv7", "thumb" ], "t2", "", d)}"
>>
>> This is wrong: ARMv6 doesn't imply Thumb-2.
> 
> Ah, yes. I'll fix this.

Are you sure?  I thought ARMv6 -was- the first to support Thumb-2.  And armv5/4
were thumb(1).  Note, not all ARMv6 processors contain thumb support.

>>> +# Whether to compile with code to allow interworking between the two
>>> +# instruction sets. This allows thumb code to be executed on a primarily
>>> +# arm system and vice versa. It is strongly recommended that DISTROs not
>>> +# turn this off - the actual cost is very small.
>>> +TUNEVALID[no-thumb-interwork] = "Disable mixing of thumb and ARM functions"
>>> +TUNE_CCARGS += "${@bb.utils.contains("TUNE_FEATURES", "no-thumb-interwork", "-mno-thumb-interwork", "-mthumb-interwork", d)}"
>>> +OVERRIDES .= "${@bb.utils.contains("TUNE_FEATURES", "no-thumb-interwork", ":thumb-interwork", "", d)}"
>>
>> This is only relevant for v4t, I guess.  Interworking is basically
>> always on for v5 and later and (CeSI aside) it's impossible on v4, so
>> hardly anybody is going to be flipping this switch.  I'm not sure it
>> really merits an OVERRIDE.
> 
> I'd be happy to remove this option if there are no objections. It was
> left for compatibility with the existing tune-thumb file but as you say,
> it likely doesn't make much sense.
> 
>>> --- a/meta/conf/machine/include/tune-xscale.inc
>>> +++ b/meta/conf/machine/include/tune-xscale.inc
>>> @@ -1,11 +1,17 @@
>>> -require conf/machine/include/arm/arch-arm.inc
>>> +DEFAULTTUNE ?= "xscale"
>>>  
>>> -INHERIT += "siteinfo"
>>> +require conf/machine/include/arm/arch-armv5-dsp.inc
>>>  
>>> -TUNE_CCARGS = "-march=armv5te -mtune=xscale"
>>> -TARGET_CC_KERNEL_ARCH = "-march=armv5te -mtune=xscale"
>>> -TUNE_PKGARCH = "${@['armv5teb', 'armv5te'][bb.data.getVar('SITEINFO_ENDIANESS', d, 1) == 'le']}"
>>> -PACKAGE_EXTRA_ARCHS = "${@['armeb armv4b armv4tb armv5teb', 'arm armv4 armv4t armv5te'][bb.data.getVar('SITEINFO_ENDIANESS', d, 1) == 'le']}"
>>> +TUNEVALID[xscale] = "Enable PXA255/PXA26x Xscale specific processor optimizations"
>>> +TUNE_CCARGS += "${@bb.utils.contains("TUNE_FEATURES", "xscale", "-mtune=xscale", "", d)}"
>>> +
>>> +AVAILTUNES += "xscale"
>>> +TUNE_FEATURES_tune-xscale = "${TUNE_FEATURES_tune-armv5te} xscale"
>>> +PACKAGE_EXTRA_ARCHS_tune-xscale = "${PACKAGE_EXTRA_ARCHS_tune-armv5te}"
>>> +
>>> +AVAILTUNES += "xscale-be"
>>> +TUNE_FEATURES_tune-xscale = "${TUNE_FEATURES_tune-armv5teb} xscale"
>>> +PACKAGE_EXTRA_ARCHS_tune-xscale = "${PACKAGE_EXTRA_ARCHS_tune-armv5teb}"
>>
>> I guess that should be "_tune-xscale-be".
> 
> Yes, I'll fix, well spotted.

In the original work I did it was just tune-xscale (no be).  AFAIK there is no
little endian version of xscale.  (there was an iwmmxt that was similar to
xscale, but was slightly different.)

>> All in all it seems as though there's an awful lot of expanded cross
>> products in this set of patches and I can't help wondering whether a lot
>> of this stuff would be better computed programmatically.  I wouldn't be
>> at all surprised if there are other copy-and-paste errors like the
>> xscale one lurking in that mass of overrides, but it's very hard to spot
>> them by eye.  It seems particularly unfortunate that everything has to
>> be written out twice, once for big-endian and once for little-endian,
>> given that endianness is almost entirely orthogonal to all the other
>> "tuning" parameters.
> 
> At least three of us have now done a pass over this so hopefully we've
> spotted the major ones but I agree its less than ideal.
> 
> The alternative is to post process the variables somehow, or turn it all
> into anonymous python (which from a .conf file is not as easy as it
> sounds).
> 
> I do like the fact that it allows some standardisation of the options
> available in a given tune file as whether or not big endian was even
> possible was previously hit and miss.
> 
> So I think its an improvement but likely not the finished end result.
> 
> Cheers,
> 
> Richard
> 
> 
> 
> _______________________________________________
> Openembedded-core mailing list
> Openembedded-core@lists.openembedded.org
> http://lists.linuxtogo.org/cgi-bin/mailman/listinfo/openembedded-core