Repository navigation

Mikeo/lowmem improvements - #136

Open

ounsworth wants to merge 22 commits into

bcgit:feature/simple-ciphersfrom

ounsworth:mikeo/lowmem_bench_tweaks

Commits on Sep 29, 2026
- Adjusted memory footprint benches
  
  ounsworth
  authored and
  dghgit
  committed
  305e7f1
- Re-ran mem benches to get new values.
  
  ounsworth
  authored and
  dghgit
  committed
  90afcc4
- Claude found some sloppy pass-by-values that resulted in 2 extra copies of the ct [u8]
  
  ounsworth
  authored and
  dghgit
  committed
  74595a9
- Claude found reconstruction of pk from sk lead to unintentional copies on the stack
  
  ounsworth
  authored and
  dghgit
  committed
  66893c9
- Claude optimized pass-by-reference function signatures
  
  ounsworth
  authored and
  dghgit
  committed
  5433b65
- Claude optimized placing a short-lived value into its own scope.
  
  ounsworth
  authored and
  dghgit
  committed
  9614bc0
- Claude optimized variable usage to avoid duplicate copies
  
  ounsworth
  authored and
  dghgit
  committed
  362ff35
- Claude found an insidious .try_into() that made a silent copy
  
  ounsworth
  authored and
  dghgit
  committed
  fbbea01
- Claude optimized the Hint, which is one bit per coeff -- instead of being held in a Polynomial [i32; 256], it's now bit-packed into a [u8; 32].
  
  ounsworth
  authored and
  dghgit
  committed
  f12e1a4
- Avoiding a .try_into() that makes a silent copy.
  
  ounsworth
  authored and
  dghgit
  committed
  d333337
- Optimized buffer reuse.
  
  ounsworth
  authored and
  dghgit
  committed
  cb48a99
- Misc cleanup.
  
  ounsworth
  authored and
  dghgit
  committed
  4831203
- Updated bench tables.
  
  ounsworth
  authored and
  dghgit
  committed
  37d0ecf
- Tweaked the SHAKE buffering to match the SHAKE internal block size.
  
  ounsworth
  authored and
  dghgit
  committed
  f587491
- Minor aesthetic cleanup of the new packed make_hint.
  
  ounsworth
  authored and
  dghgit
  committed
  1f08985
- Fable adjusted the memory benchmarking harnesses to account for how LLVM lays out stack frames and give slightly better numbers
  
  ounsworth
  authored and
  dghgit
  committed
  e25aaf8
- Fixed a bug in memory benches where mldsa verify had the public key and sig on the heap, thus under-reporting
  
  ounsworth
  authored and
  dghgit
  committed
  95ad2d1
- Fable optimized away some uneccessary copies in the non-lowmemory implemtations.
  
  ounsworth
  authored and
  dghgit
  committed
  0c52a6b
- Fable optimization to put #[inline(never)] guards around the A_hat expansion so that it doesn't get sucked into the caller's stack frame.
  
  ounsworth
  authored and
  dghgit
  committed
  ecc4466
- Saving the Fable lessons-learned from memory hygiene into a claude skill memory-hygiene-in-rust.
  
  ounsworth
  authored and
  dghgit
  committed
  c05869b
- QUALITY_AND_STYLE: drop the duplicated pointer to the memory-hygiene skill
  
  authored and
  dghgit
  committed
  2cb124f
- Claude pointed out comments that were now out of alignment with the code.
  
  ounsworth
  authored and
  dghgit
  committed
  681ead7

Navigation Menu

Repository navigation

Mikeo/lowmem improvements - #136

Mikeo/lowmem improvements#136
ounsworth wants to merge 22 commits into
bcgit:feature/simple-ciphersfrom
ounsworth:mikeo/lowmem_bench_tweaks

Commits on Sep 29, 2026

Adjusted memory footprint benches

Re-ran mem benches to get new values.

Claude found some sloppy pass-by-values that resulted in 2 extra copies of the ct [u8]

Claude found reconstruction of pk from sk lead to unintentional copies on the stack

Claude optimized pass-by-reference function signatures

Claude optimized placing a short-lived value into its own scope.

Claude optimized variable usage to avoid duplicate copies

Claude found an insidious .try_into() that made a silent copy

Claude optimized the Hint, which is one bit per coeff -- instead of being held in a Polynomial [i32; 256], it's now bit-packed into a [u8; 32].

Avoiding a .try_into() that makes a silent copy.

Optimized buffer reuse.

Misc cleanup.

Updated bench tables.

Tweaked the SHAKE buffering to match the SHAKE internal block size.

Minor aesthetic cleanup of the new packed make_hint.

Fable adjusted the memory benchmarking harnesses to account for how LLVM lays out stack frames and give slightly better numbers

Fixed a bug in memory benches where mldsa verify had the public key and sig on the heap, thus under-reporting

Fable optimized away some uneccessary copies in the non-lowmemory implemtations.

Fable optimization to put #[inline(never)] guards around the A_hat expansion so that it doesn't get sucked into the caller's stack frame.

Saving the Fable lessons-learned from memory hygiene into a claude skill memory-hygiene-in-rust.

QUALITY_AND_STYLE: drop the duplicated pointer to the memory-hygiene skill

Claude pointed out comments that were now out of alignment with the code.