[llvm-commits] [llvm] r132735 - in /llvm/trunk: lib/CodeGen/TargetLoweringObjectFileImpl.cpp test/CodeGen/X86/eh_frame.ll

[Duncan Sands]

Hi Rafael, I didn't add support for unnamed_addr to dragonegg yet, so I guess
this testcase currently passes :)

Ciao, Duncan.

> On 2011-06-17 18:22, Rafael Ávila de Espíndola wrote:
>> I have reduced it to
>>
>> @__FRAME_END__ = internal unnamed_addr constant [1 x i32]
>> zeroinitializer, section ".eh_frame", align 4
>>
>>
>> It works if you remove the unnamed_addr. I will debug it over the weekend.
>
> So, turns out this is a bug in llvm-gcc. Given the test
>
> static const int foo[]  __attribute__ ((used, section("foobar"))) = { 0 };
>
> clang produces
>
> @foo = internal constant [1 x i32] zeroinitializer, section "foobar",
> align 4
>
> llvm-gcc produces
>
> @foo = internal unnamed_addr constant [1 x i32] zeroinitializer, section
> "foobar"
>
> The presence of unnamed_addr allows llvm to merge the constant or, as is
> happening in this case, mark the section as mergeable.
>
> llvm-gcc is really old and reaching the end of its support. Is there any
> reason why you cannot use clang or dragonegg?
>
> Cheers,
> Rafael