Integer Overflow Checking Cost

TLDR

• Integer overflow checks cost ~3-5% on typical workloads, but clang’s diagnostic-enabled sanitizer causes mis-optimizations that inflate overhead to 28% on bzip2 compression.

Key Takeaways

• Theoretical worst-case: 2x penalty per add/sub; across SPECint workload mix (40% load/store, 10% branches, 50% other), that translates to ~3% total slowdown.
• bzip2 benchmarks: -fsanitize=signed-integer-overflow,unsigned-integer-overflow caused 28% slower compression, 9% slower decompression; -fsanitize-undefined-trap-on-error caused nearly zero overhead.
• The gap is a compiler bug: clang chose %ebx as destination, forcing a callee-save push/pop and blocking register allocation optimizations, not a fundamental hardware limit.
• Simple sum loops hit 4-6x slowdown under fsanitize because the sanitizer prevents SSE vectorization.
• clang 3.8+ and gcc 5+ fix the register allocation issue; gcc’s -ftrapv only checks signed overflow and has been broken since 2008.

Hacker News Comment Review

• No substantive HN discussion yet; the resubmission was flagged as timely given recent Linux kernel integer overflow work (lwn.net/Articles/1065889).

Original | Discuss on HN