I'm not sure about a good short description for this, so please bear with me. NetBSD is currently in a process of switching to gcc-4.5.3 as its system compiler. I have built a NetBSD release with the new gcc and, as a test, I decided to build NetBSD/landisk (SuperH) on a landisk machine itself (with the newly compiled system) to verify bootstrap/self-hosting. While it takes several days, it's an interesting exercise and last time I did it I actually found a sequence point bug in out lint... But I digressed. The build failed with: # compile libstdc++-v3/fstream-inst.o /usr/nb/tools/bin/shle--netbsdelf-c++ -frandom-seed=4cd80fc7 -Os -freorder-blocks -Werror -fno-implicit-templates -fdiagnostics-show-location=once --sysroot=/usr/nb/distrib/landisk -I/usr/src/external/gpl3/gcc/dist/gcc -I/usr/src/external/gpl3/gcc/dist/include -I/usr/src/external/gpl3/gcc/dist/libstdc++-v3/libsupc++ -I/usr/src/external/gpl3/gcc/lib/libstdc++-v3/../libstdc++-v3_4/arch/sh3el -I. -DHAVE_STDLIB_H -DHAVE_STRING_H -I/usr/src/external/gpl3/gcc/dist/libstdc++-v3/include -I/usr/src/external/gpl3/gcc/lib/libstdc++-v3/arch/sh3el -c -Wno-stack-protector -std=gnu++0x /usr/src/external/gpl3/gcc/dist/libstdc++-v3/src/fstream-inst.cc -o fstream-inst.o /usr/src/external/gpl3/gcc/dist/libstdc++-v3/src/fstream-inst.cc: In member function 'void std::basic_fstream<char>::_ZTv0_n12_NSt13basic_fstreamIcSt11char_traitsIcEED1Ev()': /usr/src/external/gpl3/gcc/dist/libstdc++-v3/src/fstream-inst.cc:46:1: internal compiler error: Bus error That SIGBUS happens in incr_ticks_for_insn at resource.c:1275 doing bb_ticks[b]++; and bb_ticks is an address aligned as x*4+2. Since it's not properly aligned for int*, there's SIGBUS triggered. bb_ticks is allocated at resource.c:1201 as: bb_ticks = XCNEWVEC (int, last_basic_block); and for the faulty allocation the last_basic_block is 0. Now, xcalloc in libiberty checks that if (nelem == 0 || elsize == 0) nelem = elsize = 1; and so it ends up with calloc(1, 1). NetBSD's jemalloc rounds allocation size=1 up to 2 and allocates memory at x*4+2 that IS properly aligned for size=2, but IS NOT properly aligned for int *bb_ticks. Since SuperH is strict about alignment, eventual invalid access in incr_ticks_for_insn causes SIGBUS and exposes the bug. The compiler under investigation in the above was build on SuperH to run on SuperH targeting SuperH. To verify this bug I did a cross-build of NetBSD/landisk on an amd64 RedHat machine. Then I manually ran the command that failed in my landisk build using valgrind. So in this verification I used compiler built on amd64, running on amd64 targeting SuperH. Valgrind complained about: ==12675== Invalid read of size 4 ==12675== at 0x743398: incr_ticks_for_insn (in netbsd/tools/libexec/gcc/shle--netbsdelf/4.5.3/cc1plus) ==12675== by 0x741636: fill_simple_delay_slots (in netbsd/tools/libexec/gcc/shle--netbsdelf/4.5.3/cc1plus) ==12675== by 0x741FA9: dbr_schedule (in netbsd/tools/libexec/gcc/shle--netbsdelf/4.5.3/cc1plus) ==12675== by 0x8D9BC7: sh_output_mi_thunk (in netbsd/tools/libexec/gcc/shle--netbsdelf/4.5.3/cc1plus) ==12675== by 0x8DF98E: cgraph_expand_function (in netbsd/tools/libexec/gcc/shle--netbsdelf/4.5.3/cc1plus) ==12675== by 0x8E11A4: cgraph_optimize (in netbsd/tools/libexec/gcc/shle--netbsdelf/4.5.3/cc1plus) ==12675== by 0x8E1A54: cgraph_finalize_compilation_unit (in netbsd/tools/libexec/gcc/shle--netbsdelf/4.5.3/cc1plus) ==12675== by 0x4D2BF4: cp_write_global_declarations (in netbsd/tools/libexec/gcc/shle--netbsdelf/4.5.3/cc1plus) ==12675== by 0x78D905: do_compile (in netbsd/tools/libexec/gcc/shle--netbsdelf/4.5.3/cc1plus) ==12675== by 0x78E04D: toplev_main (in netbsd/tools/libexec/gcc/shle--netbsdelf/4.5.3/cc1plus) ==12675== by 0x35C801D973: (below main) (in /lib64/libc-2.5.so) ==12675== Address 0x4f22524 is 3 bytes after a block of size 1 alloc'd ==12675== at 0x4A0516B: calloc (vg_replace_malloc.c:418) ==12675== by 0xA2C628: xcalloc (in netbsd/tools/libexec/gcc/shle--netbsdelf/4.5.3/cc1plus) ==12675== by 0x74412B: init_resource_info (in netbsd/tools/libexec/gcc/shle--netbsdelf/4.5.3/cc1plus) ==12675== by 0x741F62: dbr_schedule (in netbsd/tools/libexec/gcc/shle--netbsdelf/4.5.3/cc1plus) ==12675== by 0x8D9BC7: sh_output_mi_thunk (in netbsd/tools/libexec/gcc/shle--netbsdelf/4.5.3/cc1plus) ==12675== by 0x8DF98E: cgraph_expand_function (in netbsd/tools/libexec/gcc/shle--netbsdelf/4.5.3a/cc1plus) ==12675== by 0x8E11A4: cgraph_optimize (in netbsd/tools/libexec/gcc/shle--netbsdelf/4.5.3/cc1plus) ==12675== by 0x8E1A54: cgraph_finalize_compilation_unit (in netbsd/tools/libexec/gcc/shle--netbsdelf/4.5.3/cc1plus) ==12675== by 0x4D2BF4: cp_write_global_declarations (in netbsd/tools/libexec/gcc/shle--netbsdelf/4.5.3/cc1plus) ==12675== by 0x78D905: do_compile (in netbsd/tools/libexec/gcc/shle--netbsdelf/4.5.3/cc1plus) ==12675== by 0x78E04D: toplev_main (in netbsd/tools/libexec/gcc/shle--netbsdelf/4.5.3/cc1plus) ==12675== by 0x35C801D973: (below main) (in /lib64/libc-2.5.so) thus confirming the bug.
Created attachment 24995 [details] Preprocessed source that triggers the bug
The command that fails: .../libexec/gcc/shle--netbsdelf/4.5.3/cc1plus -fpreprocessed fstream-inst.i -quiet -dumpbase fstream-inst.i -auxbase-strip fstream-inst.o -Os -Werror -Wno-stack-protector -std=gnu++0x -frandom-seed=4cd80fc7 -freorder-blocks -fno-implicit-templates -fdiagnostics-show-location=once -o /tmp/cc83k8TV.s
last_basic_block should never be 0, it should at least be NUM_FIXED_BLOCKS.
The SH target is the only target which does: if (optimize > 0 && flag_delayed_branch) dbr_schedule (insns); in its thunks.
I've added gcc_assert (last_basic_block >= NUM_FIXED_BLOCKS) line to init_resource_info and confirmed that trunk and all released branches fail with the testcase given in #1 for sh4-unknown-linux-gnu. Perhaps if (optimize > 0 && flag_delayed_branch) dbr_schedule (insns); in sh.c:sh_output_mi_thunk might not be a big deal. I'm testing a patch which simply removes these lines.
Author: kkojima Date: Wed Aug 17 22:49:18 2011 New Revision: 177839 URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=177839 Log: PR target/50068 * config/sh/sh.c (sh_output_mi_thunk): Don't call dbr_schedule. Modified: trunk/gcc/ChangeLog trunk/gcc/config/sh/sh.c
This fixes the problem. Thanks.
Fixed.