On 11/23/2017 07:33 AM, Jonathan Wakely wrote: > On 22 Nov 2017 11:41 p.m., "Mike Gulick" wrote: >> I have tried looking at this in gdb for quite a few days, but I haven't found >> any definitive answers. The one possible clue is that I have noticed that the >> source_location numbers in use are close to LINE_MAP_MAX_LOCATION_WITH_COLS >> (0x60000000). E.g. I am seeing source locations in the debugger of 1610043911 >> (0x5FF75207). In other examples, I am seeing source locations just over this >> limit. >> >> I would greatly appreciate suggestions for debugging this issue further. I have >> been looking at it for over a week and am hitting a bit of a wall. > > Are there any odd characters in the headers, specifically > carriage-returns (i.e. DOS line-endings) or maybe some Unicode space > characters that aren't whitespace in ASCII? > There may be, but unfortunately this doesn't look to be relevant. See my next comment. > It's certainly possible you've hit some edge case and the linemap is > being corrupted. If you add the -v option to your gcc commands you'll > see the cc1plus command that's being run, and then you can run that > command under valgrind to see if there are any errors shown. > Alternatively, rebuild GCC after configuring it with the > --enable-checking=valgrind option, which will build GCC with valgrind > debugging built-in. That build of GCC will run very slowly, but using > it on your source code will check for any uninitialized reads/writes > or other memory bugs that could be the cause of what you're seeing. > Thanks for the suggestion. I did initially try creating a valgrind build of gcc 7.2 and ran it under valgrind as documented in the gcc wiki, but it did not report any issues. I still haven't had any success trying to track down the source of the issue in the debugger. However I did have success in creating a simple reproduction case. This bug *is* dependent on crossing over LINE_MAP_MAX_LOCATION_WITH_COLS, so the reproducer includes a simple gcc plugin that sets the line map's highest_location field before the preprocessor runs. I have created bug 83173 and have attached the reproduction source code to that bug. -Mike
