mimeset crash kills Haiku build

[mbrumbelow]

Kernel: x86_64 hrev55625 CPU: 8 Processors Intel I5 Memory: 16GB Storage: 500GB Compiler: GCC 11.2.0_2021_07_28-3 x86_64-unknown-haiku

Building Haiku fails when mimeset is called. mimeset crashes and user terminates the process which terminates the build.

[mbrumbelow]

Crash Report

[waddlesplash]

This doesn't seem to reproduce with -O0 -g, so it is probably related to some optimization somehow.

[mbrumbelow]

Where should those be put so I can test the build out on my other machines?

[waddlesplash]

Add them to SubDirC[++]Flags in src/build/libroot/Jamfile.

[waddlesplash]

Ah, I missed that it only happens during package creation. Indeed I can still reproduce it with debug mode.

[waddlesplash]

Adding some extra error checks to libroot_build's dup() seems to fix the crash, but then mimeset seems to be stuck in a loop and never completes. So the problem appears to be elsewhere.

[waddlesplash]

The problem is really in the Storage Kit, it appears. With -O0 or -O1 in libbe_build storage Jamfile, mimeset works. With -O2, it loops forever (runs out of FDs and then just keeps going.)

[waddlesplash]

I've narrowed it down to Directory.cpp. Adding this line to storage/Jamfile avoids the problem:

C++FLAGS on [ FGristFiles Directory.o ] += -O1 ;

Changing it to O2 triggers the problem.

[mbrumbelow]

Could this be because of a change in the compiler from GCC 8 to 11?

[waddlesplash]

Yes.

The culprit function appears to be BDirectory::GetNextRef(entry_ref* ref). Using pragmas to build just this function at -O1 solves the problem.

[waddlesplash]

Fixed in hrev55647.