re-implement crc32c merge algorithm with SIMD

[lihuiba]

to reduce memory & cache footprint, as well as lines of code.

[Coldwings]

using

constexpr auto k = crc_merge_table_pclmulqdq[(blksz/8 - 1)*2];
__m128i c0 = {(long)crc}, c1 = {(long)crc1};
auto t = SSE::pclmulqdq<0x00>(c0, &k) ^ SSE::pclmulqdq<0x10>(c1, &k);

make k as a constexpr uint64_t, not a reference,
should be better showing k will be a compile time const and crc32_merge_table_pclmulqdq will be drop by optimization.

[Coldwings]

blksz_8 could also be constexpr

[Coldwings]

using

constexpr auto k = crc_merge_table_pclmulqdq[(blksz/8 - 1)*2];
__m128i c0 = {(long)crc}, c1 = {(long)crc1};
auto t = SSE::pclmulqdq<0x00>(c0, &k) ^ SSE::pclmulqdq<0x10>(c1, &k);

make k as a constexpr uint64_t, not a reference, should be better showing k will be a compile time const and crc32_merge_table_pclmulqdq will be drop by optimization.

Seems crc_merge_table_pclmulqdq is a uint64_t table, but actually used as a uint128 table.
so constexpr is not a available choice.

[lihuiba]

It needs the address of the elements in the table (as uint64_t*), and loads 2 uint64_ts from that address, so we can not take the address of local variable k.