On 02/03/18 21:57, Luc Van Oostenryck wrote: > On Fri, Mar 02, 2018 at 08:57:27PM +0000, Ramsay Jones wrote: >> >> I quickly tried this on cygwin (because I don't have LLVM installed >> on Linux at the moment). I have had, for a long time, three of >> the backend (ie. LLVM) tests failing on master. I have e-mailed the >> list about these failures before - but I had forgotten to e-mail >> about some 'workarounds' that I have found since then (and they may >> well be specific to cygwin). >> >> Anyway, without those 'workarounds' I see several more failures in >> the backend tests (but _not_ backend/call-variadic.c), but if I apply > > ... > >> Hope That Helps! > > Yes, it helps a lot to understand the problems you had. > Thanks. Note that I didn't get any unexpected failures - Chris mentioned some failures (on Linux?). > I've quickly checked and those workarounds doesn't have > a negative effect on my environments, so I'm fine with > them and to me they can even go into to current HEAD > (or upcoming release 0.5.2 if Chris is fine doing one). I don't think I would apply them as-is - I've been meaning to install LLVM on Linux again, so I can check these things out for myself. Also, I think these are platform specific (maybe!). ;-) > OTOH, I would prefer to better understand what they > do (OK, choosing non-default method for exeception handling > and not using JIT) and even more why they're only needed > for your env (cygwin). Are you using an old version of LLVM > or something relatively recent like 3.8, 3.9 or 4.0+? The LLVM version is 5.0.1. I think the exception-model is specific to the cygwin (Windows) platform. I chanced upon that solution by noting the errors related to some '.seh_xxx' asm directives (Structured Exception Handling) and skimming the 'llc --help' output for clues. [Perhaps that should be the default for the cygwin build of LLVM!] Forcing the use of the interpreter was just a wild guess that paid off! I tried several things, including using an intermediate file rather than piping into the stdin of LLI - don't quote me, but I think that may have worked as well (I don't remember). [It may have something to do with buffering on the pipe.] In other words, I didn't investigate too much, since the LLVM backend is not something I'm too interested in, except for a general desire to keep bugs out of the project. ATB, Ramsay Jones -- To unsubscribe from this list: send the line "unsubscribe linux-sparse" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
