On Thu, Mar 10, 2011 at 3:59 PM, Richard Guenther <richard.guent...@gmail.com> wrote:
>> <richard.guent...@gmail.com> wrote: >> >>>> Using binutils-2.21, a couple of >>>> gcc.c-torture/execute/builtins/__-chk.c testcases fail on >>>> alphaev68-pc-linux-gnu (-lto) with: >>>> >>>> /usr/lib/gcc/alpha-unknown-linux-gnu/4.4.5/../../../../alpha-unknown-linux-gnu/bin/ld: >>>> Warning: alignment 8 of symbol `buf5' in >>>> /tmp/ccgnDykf.ltrans1.ltrans.o is smaller than 16 in >>>> /tmp/ccc3QsSw.o.ironly >>>> /usr/lib/gcc/alpha-unknown-linux-gnu/4.4.5/../../../../alpha-unknown-linux-gnu/bin/ld: >>>> Warning: alignment 8 of symbol `buf7' in >>>> /tmp/ccgnDykf.ltrans1.ltrans.o is smaller than 16 in >>>> /tmp/ccc3QsSw.o.ironly >>>> /usr/lib/gcc/alpha-unknown-linux-gnu/4.4.5/../../../../alpha-unknown-linux-gnu/bin/ld: >>>> Warning: alignment 8 of symbol `buf1' in >>>> /tmp/ccgnDykf.ltrans1.ltrans.o is smaller than 16 in >>>> /tmp/ccc3QsSw.o.ironly >>>> >>>> Attached patch fixes these failures. >>> >>> I think this needs more investigation as there are no conflicting >>> definitions of those vars that would warrant this kind of diagnostic from >>> ld. >>> >>> We probably bring the vars local by making them hidden and distribute >>> them to multiple ltrans units (with only one definiton obviously and >>> multiple >>> externs). >> >> Jan claims that this problem is a linker bug, so perhaps the patch is >> still suitable for the gcc testsuite. The patch doesn't alter tests in >> any way, while it avoids false positives due to linker bugs. > > Are you sure? Making the vars static enables folding the zero initialization. I was looking at other similar testcases (snprintf-chk.c, vsprintf-chk.c), where uninitialized buffer is declared as static (and it didn't fail lto tests). Anyway, let's ask the author of the test (CC'd). > I don't think we should make testsuite changes to paper over bugs > elsewhere. How far do we want to reach in case the test uncovers the problem in (sort of...) unrelated product? Uros.
