On 07/24/2014 07:40 PM, Marat Zakirov wrote:
>
> On 07/24/2014 04:27 PM, Marat Zakirov wrote:
>>
>> On 07/23/2014 06:23 PM, Marat Zakirov wrote:
>>> Hi there!
>>>
>>> I made a patch which fixes regressions on ARM platforms with 
>>> disabled unaligned accesses. The problem is that 
>>> 'arm_vect_no_misalign' predicate do not check 'unaligned_access' 
>>> global variable to determine whether unaligned access to vector are 
>>> allowed. This leads to spurious vect.exp test fails when GCC is 
>>> configured --with-specs=%{!munaligned-access:-mno-unaligned-access}.
>>>
>>> Attached patch fixes ARM predicate and several tests to correctly 
>>> handle the issue.
>>>
>>> The following targets were reg. tested for multiple targets (ARM, 
>>> Thumb-1, Thumb-2, x86, x86_64) with and without 
>>> -mno-unaligned-access.  Analysis showed patch affects only vect.exp 
>>> tests so only vect.exp was tested.
>>>
>>> For x86, x86_64,  ARM without -mno-unaligned-access, Thumb-2 without 
>>> -mno-unaligned-access and Thumb-1 no regressions occured. For 
>>> ARM/Thumb2 with -mno-unaligned-access patch fixed most of failures 
>>> but triggered some problems (see attached log) for current vect.exp 
>>> tests:
>>> 1) https://gcc.gnu.org/bugzilla/show_bug.cgi?id=61887
>>> 2) Some XPASS'es due to unexpected loop versioning (e.g. 
>>> gcc.dg/vect/pr33804.c).
>>> 3) After predicate fix some passing tests which require unaligned 
>>> vector support become NA (this was expected).
>>>
>> Here is new version of patch and regression log. On the current trunk 
>> results are slightly different due to patches for Richard Biener (no 
>> UNRESOLVED fails) but some  PASS->XPASS regressions still remain (see 
>> attachment):
>>
>> PASS->XPASS: gcc.dg/vect/if-cvt-stores-vect-ifcvt-18.c 
>> scan-tree-dump-times vect "vectorized 1 loops" 1
>> PASS->XPASS: gcc.dg/vect/pr33804.c -flto -ffat-lto-objects 
>> scan-tree-dump-times vect "vectorized 1 loops" 1
>> etc.
>>
>> These XPASS'es are due to code versioning: current GCC creates 2 
>> versions of loop: aligned and misaligned. It's look like they are 
>> slightly out of date at lest for ARM.
>>
>
> On 07/24/2014 06:50 PM, Ramana Radhakrishnan wrote:
>> This is redundant.
>>
>>> -           || (defined(__ARMEL__) \
>>> +           || (defined(__ARM_FEATURE_UNALIGNED) \
>>> +               && defined(__ARMEL__) \
>>>                  && (!defined(__thumb__) || defined(__thumb2__)))
>> As is this line.
>>
>> I think you can restrict the check to defined(__ARM_FEATURE_UNALIGNED)
>> && defined(__ARMEL__)
>>
>>   __ARM_FEATURE_UNALIGNED should tell you whether unaligned access is
>> allowed or not, therefore you should no longer require any specific
>> "architectural" checks.
>>
>>
>>>          #error FOO
>>>          #endif
>>>
>> I'm not sure about the original intent of the tests right now.
>>
>> Ramana
>>
>
> Thank you Ramana!
>
> --Marat