Switch to a better memmove

[stellar-aria]

This changes the (extremely naive) memmove implementation out for one that progressively copies in larger blocks until it reaches 8 quadwords.

[github-actions]

Test Results

106 tests  ±0   106 ✅ ±0   0s ⏱️ -1s
 16 suites ±0     0 💤 ±0 
 16 files   ±0     0 ❌ ±0 

Results for commit 19c9875. ± Comparison against base commit 6eab902.

♻️ This comment has been updated with latest results.

[stellar-aria]

Had to check for correctness after realizing that building for Release disabled my test assertions in the benchmark/test. New implementation passed.

[stellar-aria]

Apparently the multi-load/multi-store intrinsics are new to GCC 14.2 :(

[nikodemus]

Idle thought: there's not a ton of memmove use-sites, I wonder if we could guarantee a healthy (like quadword) alignment for most?

[stellar-aria]

Idle thought: there's not a ton of memmove use-sites, I wonder if we could guarantee a healthy (like quadword) alignment for most?

We probably could, but better to just have an agnostic algorithm that's really fast anyways like this one and not accidentally get caught up when something in std:: uses it.

[m-m-adams]

lgtm but also we should remember to update these when gcc is updated, this generates worse assembly than the vld1q_u8_x4 intrinsic

[stellar-aria]

lgtm but also we should remember to update these when gcc is updated, this generates worse assembly than the vld1q_u8_x4 intrinsic

Considering we'll likely get 14.2 this week, I'll probably just rip 'em out before the merge even happens. They'll collide and error otherwise

[stellar-aria]

Currently blocked by xpack-dev-tools/arm-none-eabi-gcc-xpack#38