bootstrap comparison failure with bootstrap-lto

[Matt]

On Fri, 18 Nov 2011, Ian Lance Taylor wrote:

> Matt <matt@use.net> writes:
>
>> Should I file a bug about the multiple runs being necessary? I poked
>> around a bit and it seemed that different -j options results in a
>> different number of tries. That being said, even -j1 required multiple
>> tries. (This may be a symptom of one of the issues that causes my
>> builds to fail spectacularly every blue moon when doing -jN>6.)
>
> Sure, go ahead and file a bug.  It would be nice to get these things
> cleaned up.

Done. http://gcc.gnu.org/bugzilla/show_bug.cgi?id=51232

  I added you on the cc for it, hope you don't mind.


> Great.  I think that now we've cleared through the weeds and gotten to
> the interesting part: why does it crash?  I think the failing test is
> gcc_AC_INITFINI_ARRAY in gcc/acinclude.m4.  It is intended to see
> whether the host linker correctly supports a mix of .init_array and
> .ctors sections with priorities.  This test is intended to be more or
> less independent of the compiler, and is intended to only test the
> linker.  Both the host gcc used to build stage1 and the stage1 gcc
> itself are presumably using the same linker (right?).  So, why would the
> test pass with the host gcc and fail with the stage1 gcc?
>
> If you can trace the points at which the static variable "count"
> changes, that may help us pin down what is going wrong.

Pin down, indeed :) When I try to watch or print 'count', I get this from 
GDB:
Reading symbols from /home/matt/src/gcc-trunk/obj/conftest...done.
(gdb) watch count
Watchpoint 1: count
(gdb) run
Starting program: /home/matt/src/gcc-trunk/obj/conftest
Error evaluating expression for watchpoint 1
value has been optimized out
Watchpoint 1 deleted.

So, it looks like an optimization bug of a sort to my naive eyes. As such, 
I started trying different combinations on the compile line, the original 
was:
/home/matt/src/gcc-trunk/obj/./prev-gcc/xgcc 
-B/home/matt/src/gcc-trunk/obj/./prev-gcc/ 
-B/home/matt/x86_64-unknown-linux-gnu/bin/ 
-B/home/matt/x86_64-unknown-linux-gnu/bin/ 
-B/home/matt/x86_64-unknown-linux-gnu/lib/ -isystem 
/home/matt/x86_64-unknown-linux-gnu/include -isystem 
/home/matt/x86_64-unknown-linux-gnu/sys-include  -o conftest -g -O2 
-flto=jobserver -frandom-seed=1  -static-libstdc++ -static-libgcc ~/finitest.c


If I take out the -flto=jobserver, the test no longer crashes and the 
variable is no longer optimized out incorrectly. Trying other -flto 
variants all elicited the original problem. When I tested compilation 
using the trunk *or* the system compiler (GCC 4.6.1-based) with 
-O2 -flto, I was able to repeat the problem exactly.

Is this a wrong-code bug, and/or should that test never be 
compiled with -flto?


--
tangled strands of DNA explain the way that I behave.
http://www.clock.org/~matt