Skip to content

Opt memcpy memmove#682

Merged
mateoconlechuga merged 4 commits intomasterfrom
opt_memcpy_memmove
Jan 29, 2026
Merged

Opt memcpy memmove#682
mateoconlechuga merged 4 commits intomasterfrom
opt_memcpy_memmove

Conversation

@ZERICO2005
Copy link
Contributor

@ZERICO2005 ZERICO2005 commented Jan 5, 2026 •
edited
Loading

makes memcpy 1F faster and memmove 2F faster in the common case.

The timings can be adjusted slightly for memmove (src != dst version), depending on which is faster/better on average

    ; src >= dst | LDIR | 30F + 15R + 1 ; jr false
    ; src <  dst | LDDR | 34F + 12R + 2 ; jr true

    ; src >= dst | LDIR | 31F + 15R + 2 ; jr true
    ; src <  dst | LDDR | 33F + 12R + 1 ; jr false

@ZERICO2005 ZERICO2005 marked this pull request as ready for review January 6, 2026 18:19
@mateoconlechuga
Copy link
Member

mateoconlechuga commented Jan 6, 2026

Looks good, let's remove the PREFER_OS_LIBC option and just use these directly. I assume these optimizations are valid @calc84maniac?

calc84maniac
calc84maniac reviewed Jan 7, 2026
View reviewed changes
@calc84maniac
Copy link
Contributor

calc84maniac commented Jan 7, 2026

If the two memmove implementations are now identical aside from a ret z, maybe put only that in the .if 0 to keep it more maintainable (or remove the one we're not using entirely)

@ZERICO2005
Copy link
Contributor Author

ZERICO2005 commented Jan 7, 2026 •
edited
Loading

They differ slightly. The src == dst version has to have to jr to LDDR and fall-through to LDIR to keep timings equal. Whereas the src != dst version can either have the LDIR or LDDR path be the faster path (by 1F + 1)

The LDIR path has an extra 3R and the LDDR path has an extra 3F, so I wasn't sure which would be better to allocate the 1F + 1 savings to

@mateoconlechuga
Copy link
Member

mateoconlechuga commented Jan 16, 2026

Perhaps we should just remove the preference for the OS versions of these and just use them directly.

@ZERICO2005
Copy link
Contributor Author

ZERICO2005 commented Jan 16, 2026

I will need to rebase some changes before this can be merged

@mateoconlechuga mateoconlechuga merged commit c01531e into master Jan 29, 2026
9 checks passed
@mateoconlechuga mateoconlechuga deleted the opt_memcpy_memmove branch January 29, 2026 06:01
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

@calc84maniac calc84maniac calc84maniac left review comments

Assignees

No one assigned

Labels

libc optimization

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

3 participants

@ZERICO2005 @mateoconlechuga @calc84maniac