[llvm-commits] [llvm] r76895 - /llvm/trunk/lib/Target/X86/X86RegisterInfo.cpp

[Duncan Sands]

Evan Cheng wrote:
> Any idea what's the real problem? This patch is just a hack.

The problem is definitely coming from libgcc.  If I get rid of the hack
(i.e. reapply Bill's patch) then everything works fine if I build with
llvm-gcc but don't link with the newly built libgcc_s.so.  Contrarywise,
if I build a problematic program using the system Ada compiler, but link
with the newly built llvm-gcc version of libgcc, then the program blows
up doing exception handling.

I don't have time to look into this further right now, but the following
tests from  SingleSource/Regression/C++/EH pass when linked with libgcc
built without Bill's patch, but fail when linked with libgcc built with
Bill's patch:

ctor_dtor_count
ctor_dtor_count-2
exception_spec_test
function_try_block
simple_rethrow
throw_rethrow_test

Perhaps the LLVM lowering of the unwinder intrinsics breaks with Bill's
patch applied?

Ciao,

Duncan.