| Summary: | cloop overflow test is broken in recent clang | ||
|---|---|---|---|
| Product: | WebKit | Reporter: | Richard O'Grady <rjogrady> |
| Component: | JavaScriptCore | Assignee: | Nobody <webkit-unassigned> |
| Status: | NEW --- | ||
| Severity: | Normal | ||
| Priority: | P2 | ||
| Version: | 528+ (Nightly build) | ||
| Hardware: | Unspecified | ||
| OS: | Unspecified | ||
cloop.rb's cloopAddOverflowTest() uses this check: "((#{signBit}(b) != #{signBit}(a)) && (#{signBit}(b-a) == #{signBit}(a)))" to detect overflow, and similarly for subtraction. In clang r221045, an optimization was introduced which breaks this for the case where we are incrementing a value by 1. Repro on Linux (with clang r221045 or above): $ CC=clang $ CXX=clang $ build-jsc --cloop --release --gtk Then execute this in jsc: for (var i = 0x7fffffff; i < 0x80000000; i++) {print(i);} Notice that i wraps around to become negative. Using safeAdd() and safeSub() from CheckedArithmetic.h avoids the problem.