[Buildroot] Analysis of build results for 2017-02-26: librsvg failure

[Thomas Petazzoni <thomas.petazzoni@free-electrons.com>]

Hello,

On Tue, 28 Feb 2017 10:05:42 +0100, Thomas Petazzoni wrote:

> No, that's only the end of the strace output. I've posted the full
> strace output at:
> 
>   http://code.bulix.org/elused-120071
> 
> File descriptor 3 is:
> 
> open("/lib/x86_64-linux-gnu/libm.so.6", O_RDONLY|O_CLOEXEC) = 3
> 
> Which looks correct to me, when running a host binary on the host machine.

Ok, so I went through the next stage: running the program under gdb, so
now I know exactly the instruction that faulted, in which function, etc.

Program received signal SIGILL, Illegal instruction.
0x00007ffff4869f02 in have_feature ()
   from target:/home/thomas/projets/buildroot/output/host/usr/x86_64-buildroot-linux-gnu/sysroot/usr/lib/libpixman-1.so.0
(gdb) bt
#0  0x00007ffff4869f02 in have_feature ()
   from target:/home/thomas/projets/buildroot/output/host/usr/x86_64-buildroot-linux-gnu/sysroot/usr/lib/libpixman-1.so.0

So we're crashing in libpixman, in the have_feature() function, whose
goal is precisely to determine the capabilities of the CPU we are
running on.

The faulty instruction is:

(gdb) disassemble
Dump of assembler code for function have_feature:
   0x00007ffff4869ee1 <+0>:	push   %r13
   0x00007ffff4869ee3 <+2>:	push   %r12
   0x00007ffff4869ee5 <+4>:	mov    %edi,%r12d
   0x00007ffff4869ee8 <+7>:	push   %rbp
   0x00007ffff4869ee9 <+8>:	push   %rbx
   0x00007ffff4869eea <+9>:	sub    $0x18,%rsp
   0x00007ffff4869eee <+13>:	cmpl   $0x0,0x25e85f(%rip)        # 0x7ffff4ac8754 <initialized.4467>
   0x00007ffff4869ef5 <+20>:	jne    0x7ffff4869fc6 <have_feature+229>
   0x00007ffff4869efb <+26>:	mov    $0x1,%eax
   0x00007ffff4869f00 <+31>:	cpuid  
=> 0x00007ffff4869f02 <+33>:	bextr  $0x10f,%edx,%ebp
   0x00007ffff4869f0b <+42>:	shl    $0x4,%ebp
   0x00007ffff4869f0e <+45>:	mov    %ebp,%eax
   0x00007ffff4869f10 <+47>:	or     $0x1,%eax

This bextr instruction is indeed a somewhat new instruction: it was
introduced in the Haswell micro-architecture, but my i7-4600U is a
Haswell one. I'm not a big expert in Intel architecture, but I believe
this instruction should therefore be valid on my CPU.

What is very weird is that if I run this program outside of the build
process, it runs fine. It's only when it's run within the overall build
process that it crashes on this instruction. To get the above gdb
stuff, I had to hack the package Makefile to run the program under
gdbserver, and connect to it.

Looking at the pixman code, I don't see any place where a bextr
instruction is explicitly used, so it seems like it's generated by gcc.

Another weird thing: this only happens if the target is x86-64, with a
specific toolchain. Now, the question is: why a host package crash can
be related to the target architecture/toolchain...

/me calling Mulder and Scully :)

Thomas
-- 
Thomas Petazzoni, CTO, Free Electrons
Embedded Linux and Kernel engineering
http://free-electrons.com