simdjson uses MMX but doesn't clear x87 state

[alexey-milovidov]

milovidov@example:~/ClickHouse/contrib/simdjson$ grep -r '__m64' .
./include/simdjson/numberparsing.h:  __m64 val;
./include/simdjson/numberparsing.h:  __m64 base = _mm_sub_pi8(val,_mm_set1_pi8('0'));
./include/simdjson/numberparsing.h:  __m64 basecmp = _mm_subs_pu8(base,_mm_set1_pi8(9));
./singleheader/simdjson.h:  __m64 val;
./singleheader/simdjson.h:  __m64 base = _mm_sub_pi8(val,_mm_set1_pi8('0'));
./singleheader/simdjson.h:  __m64 basecmp = _mm_subs_pu8(base,_mm_set1_pi8(9));
milovidov@example:~/ClickHouse/contrib/simdjson$ grep -r 'm_empty' .
milovidov@example:~/ClickHouse/contrib/simdjson$

This lead to errors in any subsequent usage of long double type.
Explanation: https://github.com/yandex/ClickHouse/blob/master/dbms/src/Compression/LZ4_decompress_faster.cpp#L78

  * But if we write code this way, it will lead to
  *  extremely weird and extremely non obvious
  *  effects in completely unrelated parts of code.
  *
  * Because using MMX registers alters the mode of operation of x87 FPU,
  *  and then operations with FPU become broken.

[alexey-milovidov]

Solution 1: simply remove MMX code (see the attached PR).
Solution 2: issue _mm_empty() call after parsing.

[lemire]

Fascinating.

[lemire]

Issue fixed by PR (which removes offending code).

Performance impact is negligible. Alexey Milovidov will be added to our list of contributors.