[Bug libgcj/40860] [4.4/4.5 regression] regressions in libjava testsuite on arm-linux

["mikpe at it dot uu dot se" <gcc-bugzilla at gcc dot gnu dot org>]

------- Comment #26 from mikpe at it dot uu dot se  2010-03-20 22:36 -------
Created an attachment (id=20147)
 --> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=20147&action=view)
test program to take a stack trace using _Unwind_ API

I'm attaching a test program which sets up a predictable call chain, and then
uses _Unwind_Backtrace() and _Unwind_GetRegionStart() to get the entry points
of the functions in the call chain into an array. It prints the corresponding
function names and verifies that the function entry points are correct.

With gcc-4.4.3 and binutils-2.19.1 on armv5tel-unknown-linux-gnueabi this works
fine:

> gcc -funwind-tables -fno-unit-at-a-time -O -o stacktrace-2.19.1 stacktrace.c
> ./stacktrace-2.19.1 
bar
f2
main

With the same gcc binary but binutils-20100316 the program fails:

> PATH=/tmp/binutils-2.20.51-install/bin:$PATH gcc -funwind-tables -fno-unit-at-a-time -O stacktrace.c -o stacktrace-2.20.51
> ./stacktrace-2.20.51 
bar
f1
f1
Abort

The reason for the failure is that ld merged the .ARM.exidx entries for f2 and
main with the one for f1, which breaks _Unwind_GetRegionStart(). First we see
that f1, f2, and main are adjacent in memory:

> nm -v stacktrace-2.20.51
...
00008588 T tracefn
000085bc T bar
0000865c T f1
0000866c T f2
0000867c T main
00008698 T __libc_csu_fini
...

With binutils-2.19.1 they each get their own .ARM.exidx entry:

> /tmp/binutils-2.20.51-install/bin/readelf --unwind stacktrace-2.19.1 

Unwind table index '.ARM.exidx' at offset 0x778 contains 7 entries:

0x845c <_start>: 0x1 [cantunwind]

0x8508 <printf_fn>: 0x80a8b0b0
  Compact model 0
  0xa8      pop {r4r14}
  0xb0      finish
  0xb0      finish

0x8588 <tracefn>: 0x80aab0b0
  Compact model 0
  0xaa      pop {r4, r5, r6r14}
  0xb0      finish
  0xb0      finish

0x85bc <bar>: 0x8003a8b0
  Compact model 0
  0x03      vsp = vsp + 16
  0xa8      pop {r4r14}
  0xb0      finish

0x865c <f1>: 0x80a8b0b0
  Compact model 0
  0xa8      pop {r4r14}
  0xb0      finish
  0xb0      finish

0x866c <f2>: 0x80a8b0b0
  Compact model 0
  0xa8      pop {r4r14}
  0xb0      finish
  0xb0      finish

0x867c <main>: 0x80a8b0b0
  Compact model 0
  0xa8      pop {r4r14}
  0xb0      finish
  0xb0      finish

But with binutils-20100316 f2 and main don't have their own entries, so PCs in
them are reported as belonging to f1 instead:

> /tmp/binutils-2.20.51-install/bin/readelf --unwind stacktrace-2.20.51 

Unwind table index '.ARM.exidx' at offset 0x778 contains 6 entries:

0x845c <_start>: 0x1 [cantunwind]

0x8508 <printf_fn>: 0x80a8b0b0
  Compact model 0
  0xa8      pop {r4r14}
  0xb0      finish
  0xb0      finish

0x8588 <tracefn>: 0x80aab0b0
  Compact model 0
  0xaa      pop {r4, r5, r6r14}
  0xb0      finish
  0xb0      finish

0x85bc <bar>: 0x8003a8b0
  Compact model 0
  0x03      vsp = vsp + 16
  0xa8      pop {r4r14}
  0xb0      finish

0x865c <f1>: 0x80a8b0b0
  Compact model 0
  0xa8      pop {r4r14}
  0xb0      finish
  0xb0      finish

0x8698 <__libc_csu_fini>: 0x1 [cantunwind]

This test program also works on i686-linux, sparc64-linux -m32/-m64, and
powerpc64-linux -m32.

I understand the benefits of shrinking .ARM.exidx tables, but it does break
parts of gcc's <unwind.h> implementation (and thus libjava), so should be
opt-in via an explicit option and not done by default.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=40860