faster CRC32 implementation

[MartinNowak]

• use slicing by 8 algorithm with bigger precomputed tables
• roughly 4x faster

[WalterWaldron]

This suggests that this implementation is endian sensitive.
The link contains the endian aware modification and mentions that reordering the lookups could yield further performance improvement.

[MartinNowak]

Nope, the implementation also works for big endian, b/c it assembles the uint's from byte-wise reads instead of relying on unaligned hardware loads. Still did the reordering of the operations which indeed provided a noticeable speedup.

[WalterWaldron]

Ok, I hadn't digested that hasUnalignedReads was only to force that optimization in DMD rather than to provide a per-architecture switch.

LGTM on the basis of matching other slicing by 8 implementations.

[MartinNowak]

After your comment, I was actually a bit unsure whether genTables is endian correct, so I build a gdc cross-compiler and tested on my MIPS router. Renamed the enum to make clear that this can only be done on LE architectures.

[DmitryOlshansky]

LGTM, also love the CTFE construction of the table.

[DmitryOlshansky]

LGTM

[UplinkCoder]

Auto-merge toggled on

[bgaff]

Dumb question here but why can't sse4 crc32 numonics be used when available? I apologize of the answer is obvious

[kubo39]

@bgaff SSE4.2 crc32 instruction is for castagnoli polynomial, not IEEE.