attachment set

Wed, 14 Aug 2013 07:24:29 GMT

attachment → boost_1_54_0-m68k-1.patch

explicitly pad arrays of VAR_EDITs on the stack to a multiple of sizeof(LIST*)

Wed, 14 Aug 2013 07:54:00 GMT

Originally reported at Debian <http://bugs.debian.org/cgi-bin/bugreport.cgi?bug=719484>.

Attempting to bootstrap boost_1_54_0 on m68k-linux fails early on with an assertion failure in tools/build/v2/engine/function.c:

../bootstrap/jam0 -f build.jam --toolset=gcc --toolset-root= clean jam0: function.c:233: check_alignment: Assertion `(size_t)s->data % sizeof( LIST * ) == 0' failed. Aborted

With additional debugging added to that file, I observed that the stack went out of alignment in a stack_allocate call from expand_modifiers line 857, which allocated a multiple of sizeof(VAR_EDIT). However, sizeof(VAR_EDIT) is NOT a multiple of sizeof(LIST*) on m68k-linux, so the assertion triggered.

The thing to know about m68k on Linux is that it has more relaxed alignment rules than most machines. In this case, even though VAR_EDIT contains some uint32_t-sized fields (pointers and ints) plus 6 plain char fields, its size ends up being k*sizeof(uint32_t) [for the embedded pointer and int fields] + 6, with no additional padding, which isn't a multiple of sizeof(LIST*).

Assuming the sizeof(LIST*) alignment assertion actually is needed, the obvious solution is to ensure that (a) the stack end is aligned when initially allocated, and (b) affected allocation sites pad their allocations to a multiple of the alignment requirement (and likewise for the deallocation sites).

During a complete build I couldn't find any other issues than the ones already mentioned (inital stack alignment, sizeof(VAR_EDIT)), so this patch only addresses those specific sites. There are other sites that allocate or deallocate data that isn't necessarily a multiple of sizeof(LIST*), but none of those triggered assertion failures during my build, so I'm leaving them alone.

component changed; owner set

Thu, 09 Oct 2014 06:15:38 GMT

owner set to Vladimir Prus
component Building Boost → build

status changed; resolution set

Steven Watanabe — Sat, 14 Mar 2015 02:19:37 GMT

status new → closed
resolution → fixed

I've fixed this by relaxing the alignment constraint to check for the alignment of LIST * instead of its size.

Boost C++ Libraries: Ticket #8993: boost_1_54_0 bootstrap broken on m68k-linux due to alignment mistakes

attachment set

component changed; owner set

status changed; resolution set