debug-reduction

Debugging Test-Case Reduction Problems

You are helping the user diagnose and fix problems with their test-case reduction. Something has gone wrong — the reduced output is unexpected, reduction is stuck, or the interestingness test isn't working properly.

Common Problems and Diagnosis

1. Reduced Output Triggers a Different Bug (Slippage)

Symptoms:

• The reduced test case looks nothing like expected
• The error message in the reduced output differs from the original
• The reduced test case is suspiciously small or degenerate

Diagnosis: The interestingness test is too broad. The reducer found a simpler input that satisfies the test but exercises a different code path.

Fixes:

• Match a more specific error message (include function names, error codes, source locations from the buggy tool's internal code — NOT from the test case being reduced)
• Add validity checks to reject degenerate inputs
• For C/C++: add UB checks, compile with warnings-as-errors
• If the original bug produces a stack trace, match distinctive frames from the trace

2. Reduced Output Is Trivially Small (Empty or Near-Empty)

Symptoms:

• Result is 0-1 bytes
• shrinkray warns about trivial results
• The "reduced" test case is meaningless

Diagnosis: The interestingness test accepts inputs that don't genuinely exhibit the bug. Often: empty input causes the tool to error, and the test matches any error.

Fixes:

• Add test -s "$1" || exit 1 to reject empty files
• Check for the specific error, not just any non-zero exit code
• Add a "not bogus" check: verify the input is valid according to some reference tool before checking for the bug

3. Reduction Gets Stuck (No Progress)

Symptoms:

• File size plateaus early and won't decrease further
• The reducer runs for a long time without improvement
• Result is much larger than expected

Diagnosis possibilities:

• Interestingness test is too restrictive: Over-constraining with grep checks prevents the reducer from removing code
• Format-specific constructs: Some file structures resist generic byte-level reduction
• Interdependent code: Removing any single piece breaks the bug, but removing multiple pieces together would work

Fixes:

• Remove unnecessary grep constraints from the interestingness test
• For shrinkray: the pass tiers handle this progressively — ensure reduction ran to completion
• Try manual reduction: look at the stuck result and try removing something by hand. If that works but the reducer couldn't do it, file a bug
• For C/C++: preprocess the file first (gcc -E file.c > file.i) to eliminate header dependencies
• Try a different reducer or combine reducers (run creduce then shrinkray, or vice versa)

4. Interestingness Test Is Flaky

Symptoms:

• shrinkray warns about inconsistent test results
• Reduction makes progress but then seems to "undo" improvements
• Different runs produce very different results

Diagnosis: The interestingness test gives different results on the same input. Common causes:

• Timing-dependent bugs (race conditions)
• ASLR affecting memory addresses
• Tool under test has internal non-determinism
• Test depends on external state (network, temp files from previous run)

Fixes:

• Run the test multiple times on the same input to confirm flakiness
• Disable ASLR: setarch $(uname -m) -R ./test.sh file.c
• Pin random seeds if the tool supports it
• Clean up temp files at the start of each test invocation
• Use --parallelism=1 with shrinkray to eliminate parallel interference
• For race conditions: add sleep or retry logic (but this slows reduction significantly)

5. Reduction Is Too Slow

Symptoms:

• Each test invocation takes many seconds
• Overall reduction takes hours or days
• CPU usage is low (waiting on I/O or timeouts)

Diagnosis: The interestingness test is slow, or timeouts are too generous.

Fixes:

• Reorder checks: Put the cheapest, most-likely-to-fail check first
• Reduce timeout: If the original test takes 1 second, a 5-second timeout is fine. Don't use the default 300s from creduce
• Speed up compilation: Use -S instead of -c, -Wfatal-errors, -w (suppress warnings when they don't matter)
• Avoid unnecessary work: Don't link if you only need to compile. Don't run if you only need to compile.
• Use shrinkray's parallelism: Ensure --parallelism is set to your core count (default)
• Profile the test: time ./test.sh file.c to see where time is spent

6. Reduced Output Has Undefined Behavior (C/C++)

Symptoms:

• Reduced test case compiles but behavior depends on compiler/flags
• Different compilers give different results on the reduced code
• AddressSanitizer or UBSan flag issues in the reduced code

Diagnosis: The reducer introduced UB during reduction. This is expected and nearly inevitable for C/C++ miscompilation bugs if the interestingness test doesn't guard against it.

Fixes:

• Add UB sanitizer checks to the interestingness test: -fsanitize=undefined
• Compile with strict warnings: -Wall -Wextra -pedantic -Werror
• Use both GCC and Clang for cross-checking
• Consider using heavyweight tools: Frama-C, tis-interpreter, or kcc/RV-Match
• Note: for crash bugs (ICE), UB in the test case is usually fine — the bug is that the compiler crashes, not that the code is wrong

7. Test Works Manually But Not Under the Reducer

Symptoms:

• ./test.sh file.c && echo interesting works fine manually
• But shrinkray says the initial test case is not interesting

Diagnosis: This is almost always a temporary directory problem. shrinkray runs the test in a fresh temp directory, not your working directory. Anything that depends on your local environment will break.

Fixes:

• Use $1 for the test case: The file argument is an absolute path and works from any directory. If the test references the file by a hardcoded relative path or basename, switch to $1.
• Use absolute paths for everything else: Tools, helper scripts, auxiliary data files, reference implementations — anything that isn't the test case itself must be an absolute path or resolved via SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)".
• Check PATH: Tools you have on PATH in your shell may not be on PATH in the reducer's environment. Use full paths to be safe.
• Missing dependencies: If your test script sources other files or uses tools not in the default PATH, they won't be available in the temp directory.
• Permissions: Ensure the test is executable (chmod +x test.sh).

Quick diagnostic: Run the test from a different directory to simulate the reducer's environment:

cd /tmp && /absolute/path/to/test.sh /absolute/path/to/original_file

If this fails but running from your project directory succeeds, the test has a directory dependency.

Diagnostic Steps

When the user reports a problem, work through these steps:

1. Read the interestingness test — most problems originate here
2. Check what reducer and options are being used — especially input type, timeout, parallelism
3. Test the interestingness test manually:
   • Run it on the original file (should exit 0)
   • Run it on an empty file (should exit non-zero — if it exits 0, the test is too permissive)
   • Run it on a known-good file (should exit non-zero)
   • Run it multiple times on the same file (should give consistent results)
4. Look at the reduced output — does it still trigger the original bug, or a different one?
5. Check for environment issues — absolute paths, permissions, temp directory behavior

When to Suggest Rewriting vs. Patching

• Slippage: Usually needs a rewrite of the bug-detection phase (more specific matching)
• Trivially small output: Add a guard (quick patch)
• Stuck reduction: Remove over-constraints (quick patch) or try a different approach
• Flaky tests: May need fundamental rethinking if the bug is inherently non-deterministic
• Slow tests: Reorder checks and optimize (incremental patching)

shrinkray Debugging Options

• Check --timeout setting (auto-calibrated by default — if the initial run is slow, the timeout may be too generous)
• Check --input-type matches how the test reads its input
• Use --volume=debug for detailed pass-by-pass progress
• Check .shrinkray/ history directory for intermediate results
• Use --also-interesting=101 in the test to record interesting-but-wrong variants
• Try --parallelism=1 to eliminate parallel-related issues
• Try --formatter=none if auto-formatting is interfering with reduction