kaidokert / fixed-bigint-rs / 26866698639

98%

push

github

Apply black_box to remaining Ct mask-select sites (#120)

Same fix mechanism as PR #118 (`black_box` opacification to defeat
the LLVM XOR/AND-select → cmov rewrite), applied to four more sites
that have the same pattern:

- `const_leading_zeros_ct` — `total += (!decided) & v_lz` plus the
  `wrapping_neg`-derived decided flag.
- `const_trailing_zeros_ct` — mirror.
- `const_cmp_ct` — `result |= (!decided) & here` plus
  here-non-zero. Used by ct_gt/ct_lt/Ord::cmp under the Ct
  personality.
- `ct_checked_pow` inner loop — per-iteration bit-select for the
  square-and-multiply ladder.

These are defensive fixes for the XOR/AND-select pattern. They
prevent latent bugs at sites no fixture currently exposes — most
won't trip ctgrind today because the fixtures taint both operands
symmetrically, and the rewrite to a secret-flag cmov only fires
when one operand is non-tainted.

Of note: this PR is *necessary but not sufficient* to clear the
ct-ctgrind x86_64 failures on `leading_zeros` / `trailing_zeros`.
The remaining violations are upstream — Rust's `u32::leading_zeros()`
on baseline x86_64 (no `lzcnt`) lowers to `test/je/bsr`, with the
`je` handling `bsr`-undefined-on-zero. That's a real conditional
jump on a tainted flag inside the std intrinsic, not in our code.
Will be addressed by setting `-C target-cpu=x86-64-v2` on the
ct-ctgrind workflow's x86_64 row in a follow-up.

4 of 4 new or added lines in 1 file covered. (100.0%)

2860 of 2909 relevant lines covered (98.32%)

367.28 hits per line