On 20 October 2010 21:32, Matthias Klose <d...@debian.org> wrote:
> On 21.10.2010 01:43, Nick Lewycky wrote:
>
>> Package: g++-4.4
>> Version: 4.4.5-4
>> Severity: normal
>>
>>
>> This version of gcc has a bug where code that does a shift-right
>> by literal one
>> may be miscompiled, even at -O0. This is breaking my build of llvm and
>> even
>> a
>> gcc bootstrap.
>>
> >
>
>> Fortunately, there's an unintrusive patch for the problem:
>>
>> http://gcc.gnu.org/ml/gcc-patches/2010-09/msg01070.html
>>
>> which has already been accepted into gcc head, but I didn't see it get
>> applied
>> to the release branches. Please add this patch as a debian-local change.
>>
>
> when does this trigger? the testcase succeeds in the debian builds og
> gcc-4.4 and gcc-4.5, and llvm-2.7 did build too.
I'm using llvm from svn. There's a "make check" failure in
test/Assembler/alignstack.ll due to this code in
lib/Bitcode/Reader/BitcodeReader.cpp:1192 does:
bool IsAlignStack = Record[0] >> 1;
where Record[0] is an unsigned long & equal to 3. The resulting IsAlignStack
is set to false. While that one instance could be fixed by replacing it with
& 2, there's a bunch of other places in the code which also use >> 1 to
actually mean shift right, and the clang that's built out of this doesn't
work very well at all.
Note that due to the nature of the bug, small changes in the code could
affect register layout such that this problem does not manifest. It's
entirely possible that a build of llvm 2.8 might not show the same testcase
failure but still have problems in other cases.
Nick
