Re: [Qemu-devel] [PATCH v3 7/9] cutils: Rewrite x86 buffer zero checking

[Paolo Bonzini]

On 13/09/2016 16:49, Paolo Bonzini wrote:
> 
> 
> On 13/09/2016 16:17, Paolo Bonzini wrote:
>>
>>
>> On 13/09/2016 15:26, Paolo Bonzini wrote:
>>> I need this on top to fix compilation with older compilers:
>>>
>>> diff --git a/util/bufferiszero.c b/util/bufferiszero.c
>>> index 3b39f82..1ce6b7a 100644
>>> --- a/util/bufferiszero.c
>>> +++ b/util/bufferiszero.c
>>> @@ -71,13 +71,13 @@ buffer_zero_int(const void *buf, size_t len)
>>>
>>>  #if defined(CONFIG_AVX2_OPT) || defined(__SSE2__)
>>>  #include <cpuid.h>
>>> -#include <x86intrin.h>
>>>
>>>  /* Note that we're going to check for LEN >= 64 for all of these.  */
>>>
>>>  #ifdef CONFIG_AVX2_OPT
>>>  #pragma GCC push_options
>>>  #pragma GCC target("avx2")
>>> +#include <immintrin.h>
>>>
>>>  static bool
>>>  buffer_zero_avx2(const void *buf, size_t len)
>>> @@ -181,6 +181,8 @@ buffer_zero_sse4(const void *buf, size_t len)
>>>  #pragma GCC target("sse2")
>>>  #endif /* CONFIG_AVX2_OPT */
>>>
>>> +#include <emmintrin.h>
>>> +
>>>  static bool
>>>  buffer_zero_sse2(const void *buf, size_t len)
>>>  {
>>>
>>
>> Nope, not enough on GCC 4.8:
>>
>> $ nm util/bufferiszero.o
>> 0000000000000000 b buffer_accel
>> 0000000000000430 T buffer_is_zero
>> 0000000000000240 t buffer_zero_avx
>> 0000000000000330 t buffer_zero_avx2
>> 0000000000000000 t buffer_zero_int
>> 00000000000000b0 t buffer_zero_sse2
>> 0000000000000190 t buffer_zero_sse4
>> 0000000000000008 b cpuid_cache
>>                  U _GLOBAL_OFFSET_TABLE_
>> 0000000000000000 t init_cpuid_cache
>>                  U _mm_cmpeq_epi8
>>                  U _mm_loadu_si128
>>                  U _mm_movemask_epi8
>>                  U _mm_setzero_si128
>>                  U _mm_testz_si128
>> 00000000000003d0 T test_buffer_is_zero_next_accel
>>
>> Since I have no idea what's going on, I'm dropping these the last three
>> patches.
> 
> I still have no idea what's going on, but dropping this patch is enough.

Ok, it's the AVX version, the failure somehow is related to vptest not
being in AVX1.

Paolo