constant-time-analysis

Constant-Time Analysis

Compile the code, inspect the emitted assembly or bytecode for variable-time instructions, then decide which of the flagged operations actually touch secrets. The compilation step is mechanical; the triage step is the work.

When to Use

• Implementing or reviewing a signature, encryption, KEM, or key derivation routine
• Code applies / or % to a value derived from a key, plaintext, nonce, or token
• The user mentions "constant-time", "timing attack", "side-channel", or "KyberSlash"
• Reviewing functions named sign, verify, encrypt, decrypt, derive_key

When NOT to Use

• Measuring timing variance on a running binary — use the constant-time-testing skill from the testing-handbook-skills plugin, which covers dudect and statistical approaches and may not be installed. This skill inspects compiler output statically and never executes the code under test.
• Non-cryptographic code, or crypto code where every input is public
• High-level API usage where a vetted library owns the constant-time guarantees
• Cache and other microarchitectural side channels — the assembly view cannot see them

Language Routing

Read the guide for the target language before interpreting any findings; each one lists that language's dangerous instructions and the idiomatic constant-time replacements.

Running the Analyzer

The analyzer takes one file and detects the language from its extension. Always pass --warnings:

uv run {baseDir}/ct_analyzer/analyzer.py --warnings <source_file>

Without it the analyzer reports only error-severity findings, which means division, modulo and weak RNG. Four detector families are warning severity and stay silent: secret-dependent branches, early-exit comparison (memcmp, strcmp, .equals, ==), table lookups indexed by a secret, and variable-time encoding. Early-exit comparison of an authentication tag is the most common timing bug in real code — Lucky Thirteen was exactly that — so a default run is quiet about the finding you are most likely to have.

Narrow a large file to the routines that handle secrets with a regex, for example --func 'sign|verify'.

Run natively compiled code (C, C++, Go, Rust, Swift) at more than one --arch and --opt-level. Division timing and branch lowering are architecture- and optimization-dependent: x86_64 IDIV and arm64 SDIV differ, and a cmov at -O2 can become a branch at -O0. A single clean run proves one configuration safe, not the code.

How --arch crosses depends on the toolchain. clang crosses with --target and needs no second compiler, but any source that includes libc headers also needs that target's C library headers — libc6-dev-riscv64-cross and friends — or it fails with bits/libc-header-start.h file not found. Go cross-builds through GOARCH, though go tool objdump has no riscv64 disassembler. A GNU cross toolchain is a separate binary, so gcc needs it named explicitly — --compiler x86_64-linux-gnu-gcc, --compiler riscv64-linux-gnu-gcc — and nothing is substituted for you, so the report always names the binary that ran. rustc needs the target's standard library (rustup target add), and Swift on Linux targets only the host. Compare against the toolchain that builds your product, not whichever cross build a distribution packages.

Re-run the whole sweep on the fix, across compilers, targets and every level including Os and Oz. Any fix that works by handing the compiler a constant divisor to strength-reduce is a fix only where the compiler chooses to cooperate, and that choice varies more than it looks. Replacing key_coef / (2 * gamma2) with a #defined divisor still emits a real divide here:

Strength reduction is an optimizer courtesy, not a language guarantee. Prefer an explicit multiply-shift, and verify it against the original expression over the full input range rather than on sampled values — an off-by-a-power-of-two reciprocal matches for millions of inputs before it diverges.

Java, Kotlin, and C# compile to JVM/CIL bytecode. The analyzer reads that bytecode, so --arch and --opt-level do not apply and the JIT may still introduce variable-time native code the analyzer cannot see.

Per-language coverage limits

Coverage is not uniform, and the gaps change what a clean report means:

Since findings and silence both depend on the configuration, say which compiler, architecture, and optimization level produced a result when reporting it.

To sweep a directory, loop in the shell — the analyzer is a deterministic script, one invocation per file:

for f in src/crypto/*.c; do uv run {baseDir}/ct_analyzer/analyzer.py --warnings --json "$f"; done

Prerequisites

On a "toolchain not found" error, see references/vm-compiled.md for JVM and .NET installation, macOS keg-only PATH configuration, and troubleshooting.

Interpreting Results

PASSED — no error-severity finding for the configuration you ran. Warnings do not affect it, so Result: PASSED alongside Warnings: 6 is normal and is not a clean result. Read the warning list before concluding anything.

FAILED — dangerous instructions found, reported per function:

[ERROR] SDIV
  Function: decompose_vulnerable
  Reason: SDIV has early termination optimization; execution time depends on operand values

Triaging Findings

The analyzer has no data flow analysis. It flags every dangerous instruction regardless of whether a secret reaches it, so a FAILED report is a worklist, not a verdict. Reporting the raw output as a set of vulnerabilities is the primary failure mode of this skill.

For each flagged instruction, read the source and answer one question: does an operand depend on secret data? Trace from the instruction's function back to the caller's inputs, then classify:

// FALSE POSITIVE: operands are a buffer length, already public from the ciphertext size
int num_blocks = data_len / 16;

// TRUE POSITIVE: dividend is a private-key coefficient; IDIV/SDIV leaks its magnitude
int32_t q = secret_coef / GAMMA2;

State the verdict and the data flow that justifies it for every flagged item. A finding you cannot trace to a secret is not a finding; say so explicitly rather than dropping it silently.

{baseDir}/ct_analyzer/tests/triage_samples/ holds a known-answer case per language: each fixture pairs a true positive with a false positive that the analyzer reports identically, and expectations.json records which is which and why. triage_c.c is the shortest example — the analyzer flags the division in both ct_high_bits and ct_block_count, and correct triage confirms the first and clears the second.

Weak-RNG and encoding findings ask a different question. For Math.random, mt_rand, random.randint, System.Random and base64_encode, no operand is secret, so "does an o

Content truncated.