tree ec565d4fee1c1710516642776624e2d4859b584a
parent 403de00f4369dcfc3e5cf82299c18ec16559d405
author Connal de Souza <connaldesouza@google.com> 1788374059 -0700
committer Copybara-Service <copybara-worker@google.com> 1788374129 -0700

Merge CRC+memcpy functionality into Extend function.

This improves performance through improved instruction level parallelism.  We mostly do the copying separately, but in lockstep with the crc streams. We don't reuse the loads between the crc calculation and the copy, since the loads are to different register types, except for VPCLMUL streams. We also force AVX loads and stores for platforms that support them, even if the rest of the binary is built without AVX support. Also explicitly disallow PCLMUL streams on platforms that do not support AVX, so we can split out the AVX instructions to an implementation only for large copies that support AVX (or are not x86).

PiperOrigin-RevId: 975239746
Change-Id: I40610e23dc81045f0511e08f39ce11dfe4447804