Re: [AArch64], patch] PR71727 fix -mstrict-align

[Christophe Lyon <christophe dot lyon at linaro dot org>]

ping^2 ?


On 21 August 2017 at 15:04, Christophe Lyon <christophe.lyon@linaro.org> wrote:
> ping ?
> https://gcc.gnu.org/ml/gcc-patches/2017-07/msg01063.html
>
> Christophe
>
>
> On 18 July 2017 at 14:50, Christophe Lyon <christophe.lyon@linaro.org> wrote:
>> Hello,
>>
>> I've received a complaint that GCC for AArch64 would generate
>> vectorized code relying on unaligned memory accesses even when using
>> -mstrict-align. This is a problem for code where such accesses lead to
>> memory faults.
>>
>> A previous patch (r243333) introduced
>> aarch64_builtin_support_vector_misalignment, which rejects such
>> accesses when the element size is 64 bits, and accept them otherwise,
>> which I think it shouldn't. The testcase added at that time only used
>> 64 bits elements, and therefore didn't fully test the patch.
>>
>> The report I received is about vectorized accesses to an array of
>> unsigned chars, whose start address is not aligned on a 128 bits
>> boundary.
>>
>> The attached patch fixes the problem by making
>> aarch64_builtin_support_vector_misalignment always return false when
>> the misalignment is not known at compile time.
>>
>> I've also added a testcase, which tries to check if the array start
>> address alignment is checked (using %16, and-ing with #15), so that
>> loop peeling is performed *before* using vectorized accesses. Without
>> the patch, vectorized accesses are used at the beginning of the array,
>> and byte accesses are used for the remainder at the end, and there is
>> not such 'and wX,wX,15'.
>>
>> BTW, I'm not sure about the same hook for arm... it seems to me it has
>> a similar problem.
>>
>> OK?
>>
>> Thanks,
>>
>> Christophe