4 ms·
A fuzzer should be one of the last lines of defence against human errors. This was also caught by fuzzing sqlite, not the offending compiler clang. When all but
by globular-toast 6y ago
A fuzzer should be one of the last lines of defence against human errors. This was also caught by fuzzing sqlite, not the offending compiler clang. When all but the last lines of defence are breached it's absolutely cause for concern and shouldn't be brushed away like all is fine.
- jfkebwjsbx 6y agoTesting compiler outputs by testing popular libraries an apps generated with it is a fine and common approach. Expecting a compiler (and the hardware it runs on) to be perfect is a mistake.
- thechao 6y agoExcept for compcert, which is basically perfect.
- jraph 6y agoWell, not quite, because there are parts of CompCert that are not proven [1]: - Parsing, type-checking, and pre-simplifications - Assembling and linking And there have been bugs on edge cases there. Moreover, correctness is limited to what is specified. COQ does not guarantee that there are no missing specifications or theorems. The C standard is written in English, people had to manually encode the standard in COQ. [1] http://compcert.inria.fr/compcert-C.html http://compcert.inria.fr/compcert-C.html
- touisteur 6y agoAnd the performance of the generated binary might be underwhelming for some.
- cesarb 6y agoIt's all right if it's running the testcases of these popular libraries; however, it seems that this was not caught by the testcases, only by a fuzzer (which not only is not as deterministic, but also can be much more costly to run).