[llvm-commits] PR6059 patch

[Rafael Espindola]

> Yes, but you also wrote (in the message that started this thread):
>> Attached is a potential fix for PR6059. One way to look at it is that
>> it is a partial revert of Bob's previous patch to fix PR5406.

In that it would get the alignment back, but not the i64 :-)

>> The ABI is defined from C/C++ to assembly. The generated IL is an
>> implementation detail. If the FE uses i32, the BE can do anything with
>> i64. I am fine with adding padding, not adding padding, forcing it to
>> memory, asserting or making me coffee :-)
>
> It has been my impression that in cases where it makes sense, LLVM IR should follow the ABI.  In this case, if there is a call that is marked with the AAPCS ABI, and it has i64 arguments, then those i64 values are expected to be passed as specified by AAPCS.

Lets let this rest a bit and look at it after the refactoring. We
should have a better view then.

> No.  Sorry.

NP. I know qemu-ppc worked at one time. Now I am getting segmentation
faults on any program that is not statically linked. I will try with
older ld.so and qemu.

Cheers,
-- 
Rafael Ávila de Espíndola