On Fri, 23 Mar 2018 22:34:03 +0300 (MSK) Alexander Monakov <amonakov@xxxxxxxxx> wrote: AM> On Fri, 23 Mar 2018, Vadim Zeitlin wrote: AM> > Yet they definitely do. I may not have a SSCCE of it, but building 4 AM> > versions of the same object file/executable that differ only in the warning AM> > options used: AM> > AM> > 1. Both -Wall and -Woverloaded-virtual AM> > 2. Only -Wall AM> > 3. Only -Woverloaded-virtual AM> > 4. Neither AM> AM> Since -Wall enables many individual warning options (you can see the list AM> by supplying -Wall -Q --help=warnings on gcc command line), to investigate AM> further you can try to find a minimal subset of -Wall that still exhibits AM> the issue. Thanks for the idea and the hint (I didn't know about the -Q option)! I did this and it turns out the culprit is "-Wnonnull": using it with "-Woverloaded-virtual" changes the generated code, while using all the other options included in -Wall doesn't. Looking at gcc sources, this seems completely impossible as OPT_Woverloaded_virtual is only used once in gcc/cp/class.c and this clearly can't affect anything else. Yet somehow it does... AM> > Would you have any idea about how could this be explained? AM> AM> By all indications this is an unusual compiler bug. Please file a report AM> in GCC Bugzilla if you can, even if the testcase has a large line count. I made another attempt to use delta and it managed to produce a "just" 771 line file reproducing the problem, but I discovered another impossible thing while doing this: not only the generated code depends on the warnings options, but it also somehow depends on the name of the input file! Delta finished by producing a file tmp1/16795.cpp, yet when I tested it manually, the file didn't pass the check, i.e. compiling it with and without warning options produced the same object file. It looked like a bug in delta itself, so I've started debugging it and while doing it noticed that it always runs the test script with input.cpp as argument, and not the final file name. And, surprise, renaming the *same* file to input.cpp indeed makes the code generated from it when [not] using warning options different... This really looks like a bad bug due to using uninitialized memory somewhere in gcc itself. The only question is whether it's a problem of the particular gcc build I'm using or not. Anyhow, I've created https://gcc.gnu.org/bugzilla/show_bug.cgi?id=85091 now, please let me know if I can do anything else here. Thanks, VZ
Attachment:
pgp6LDyhmMMdL.pgp
Description: PGP signature
