2.6-2 FTBFS on sparc (bus error in test cases)

backtrace of on bus error (recreated the test-case by hand and ran within gdb):

Starting program: /tmp/patch/patch-2.6/src/patch -f --merge c < ab.diff
patching file c

Program received signal SIGBUS, Bus error.
0x00016cb4 in bestmatch (xoff=-7, xlim=5, yoff=-26610410486401528, ylim=8, min=1, max=2, py=0xffa177a0) at src/bestmatch.h:153
153 *py = ymax;
(gdb) bt
#0 0x00016cb4 in bestmatch (xoff=-7, xlim=5, yoff=-26610410486401528, ylim=8, min=1, max=2, py=0xffa177a0) at src/bestmatch.h:153
#1 0x00018568 in locate_merge (hunk=1, outstate=<value optimized out>, where=5, somefailed=<value optimized out>) at src/merge.c:105
#2 merge_hunk (hunk=1, outstate=<value optimized out>, where=5, somefailed=<value optimized out>) at src/merge.c:229
#3 0x0001ebb0 in main (argc=<value optimized out>, argv=<value optimized out>) at src/patch.c:289

hm, something is really strange here:

info registers
(gdb) info registers
g0 0x0 0
g1 0x0 0
g2 0xffffffff -1
g3 0xffffffff -1
g4 0x0 0
g5 0x38 56
g6 0xfeffffff -16777217
g7 0xf7f05270 -135245200
o0 0x0 0
o1 0x2 2
o2 0xffffffff -1
o3 0xffffffff -1
o4 0x2 2
o5 0xffd8f648 -2558392
sp 0xffd8f5a0 0xffd8f5a0
o7 0x16ce4 93412
l0 0x0 0
l1 0x1 1
l2 0x0 0
l3 0x8 8
l4 0xffffffff -1
l5 0xfffffff7 -9
l6 0x0 0
l7 0x8 8
i0 0xffffffff -1
i1 0xfffffff9 -7
i2 0x0 0
i3 0x5 5
i4 0xffd8f5f8 -2558472
i5 0xffd8f608 -2558456
fp 0xffd8f690 0xffd8f690
i7 0x18560 99680
y 0x0 0
psr 0xff800087 [ #0 #1 #2 S #23 #24 #25 #26 #27 #28 #29 #30 #31 ]
wim 0x0 0
tbr 0x0 0
pc 0x16cb4 0x16cb4 <bestmatch+948>
npc 0x16cb8 0x16cb8 <bestmatch+952>
fsr 0x0 [ ]
csr 0x0 0

corresponding code (in parts, bus error at last instruction):
0x00016c9c <bestmatch+924>: ld [ %fp + -44 ], %o4
0x00016ca0 <bestmatch+928>: ld [ %fp + -84 ], %g3
0x00016ca4 <bestmatch+932>: cmp %g3, 0
0x00016ca8 <bestmatch+936>: be,pn %icc, 0x16cb8 <bestmatch+952>
0x00016cac <bestmatch+940>: nop
0x00016cb0 <bestmatch+944>: ldd [ %fp + -48 ], %o2
0x00016cb4 <bestmatch+948>: std %o2, [ %g3 ]

hm, I've just tried a build with disabling stack protector, there the test cases work. Maybe that's related, but maybe it isn't.

bus error implies a misaligned memory access? It's certainly possible the stack protector could be doing that -- sparc is less well tested with it. If we could get a minimal reproducer, we should be able to raise it with upstream gcc.

hm, the more I look at the source in bestmatch.h, I think that this is not a gcc related problem, but rather an out of bounds problem in bestmatch.h (might be caused by yoff being very far off). I'll investigate further.

patch test-case that fails when called with
patch -f --merge somefile < ab.diff

yes, stack-protector did its job fine on sparc, it's indeed an out of bounds access.

*pfft*, I'm going in circles. the safety assertions from me are completely wrong, since fd points to the middle of V. Hence it's not an out-of-bounds access after all, and I'm back at the beginning with debugging.

causes build failure in fpc on sparc. see bug #2253

This bug was fixed in the package patch - 2.6-2ubuntu1

---------------
patch (2.6-2ubuntu1) lucid; urgency=low

  * Build with -fno-stack-protector on sparc to pass the testsuite.
    LP: #534459.
 -- Matthias Klose <email address hidden> Mon, 05 Apr 2010 03:18:36 +0200

sorry, didn't yet manage to find a minimal test-case that is anything less than the complete patch package.
However I found another workaround: Adding assertions on the (hopefully) correct bounds in bestmatch.h seems to somehow also produce a working patch binary, patch attached (but please double-check).

oh, crap, attached patch is in reverted form :(

closing the GCC task, sparc is gone