On Wed, Oct 18, 2017 at 10:22 AM, Michal Karm <michal.baba...@gmail.com> wrote: > On 10/18/2017 03:39 PM, William A Rowe Jr wrote: >> In building your repro case, you'll need to toggle several flags to see how >> to >> trip this up; >> >> |- CMAKE_C_FLAGS_RELEASE:STRING=/MD /O2 /Ob2 /DNDEBUG + >> CMAKE_C_FLAGS_RELEASE:STRING=/O2 /Wall /Zi| >> I'd start with your problem case /O2 /Wall /Zi, and retest adding /Ob2, /MD >> and /NDEBUG in that order. Simply slice and dice until you find the flags >> which cause the negative min limit to fail ERANGE, but the positive max >> case to not toggle errno. >> >> Good luck > > The "problem" is the presence or lack thereof of /MD. > > Updated notes: https://github.com/Karm/_strtoi64_reproducer#conclusion > > I'm gonna try to fiddle with VS and compare asm for that part. > I don't understand why not having /MD is a problem for this single threaded > dead simple case. > > I'm also bewildered as to why my trimmed down example works > even without /MD...
There are four different representations of _strtoi64() in the different c libraries (/MD, /MT, their debug flavors) and then multiple representations depending on your choice of the optimization level combined with /Ob# (1 or 2) for inline expansion. One or more is clearly bugged. Another option is that the wrong errno is in use in your resulting testall.exe vs the libapr-1.dll, because of mismatched c libraries, but that doesn't explain why the negative test correctly toggles errno, while the positive test does not.
