[WIP] PPC/PPC64: Fix ABI errors.

[redstar]

Issue #1652 seems to be caused by ABI errors in the bootstrap compiler.
This PR updates the PPC/PPC64 ABI to conform to clang/gcc output.

[redstar]

This is not yet perfect. It fixes some errors but not all. Especially it is not possible to compile LDC 1.1.0 with this version.

[kinke]

Is the omission of IntegerRewrite::isObsoleteFor() really necessary?

[redstar]

This is what clang seems to do. But I can check it.

[kinke]

IIRC, clang also does obsolete integer rewrites for System V. It's just that a LL struct with a single int/pointer is normally passed/returned identically to a rewritten integer. Omitting the rewrite in these cases just leads to a nicer and shorter IR and a shorter log file.

[redstar]

It seems to be different on PPC. From our test suite, using clang 3.7 on Linux/PPC64:

ldc_cabi2.cpp:

declare void @dretb1(%struct.B1* sret, i8) #0

With the obsolete integer rewrites in place I get for ldc_cabi1.d:

define void @dretb1(%ldc_cabi1.B1* noalias sret align 1 %.sret_arg, %ldc_cabi1.B1 %x_arg) comdat {

And without it:

define void @dretb1(%ldc_cabi1.B1* noalias sret align 1 %.sret_arg, i8 %x_arg) comdat {

I prefer to match clang as close as possible here.

[kinke]

As I said, as long as ldc_cabi1.B1 contains a single i8, there should be absolutely no difference (machine code wise) except for uglier IR.

I prefer to match clang as close as possible here.

Then we should do it consistently, either removing the obsoleteness check for all ABIs or doing it everywhere (as currently done for at least all 3 major ABIs I'm familiar with - x86 and x86_64). Needless to say, I prefer keeping it.