Darwin X86_64_RELOC_UNSIGNED loses addend for symbol alias

[llvmbot]

Bugzilla Link	19203
Version	trunk
OS	MacOS X
Reporter	LLVM Bugzilla Contributor
CC	@atamazov,@bcardosolopes,@MaskRay,@lhames

Extended Description

$ cat test1.s
.data
a = b + 4
.long a
Sandman:llvm respindola$ cat test2.s
.data
a = b + 8
.long a
$ llvm-mc -filetype=obj test1.s -o test1.o -triple=x86_64-apple-darwin
$ llvm-mc -filetype=obj test2.s -o test2.o -triple=x86_64-apple-darwin
$ cmp test1.o test2.o

We should probably include the 4 and 8 in the relocation or produce an error if that is not possible.

[lhames]

The relocations are both for 'a', rather than 'b'. Looks like MC isn't checking the aliases before spitting out the relocations. I'll look into it.

[lhames]

Well... this is pretty awful.

.data
a = 4
.long a

Generates a data section containing one long initialized to four, as you'd expect. As Rafael noted, adding a symbol to the RHS like this:

.data
a = b + 4
.long a

Generates a data section with an external relocation for 'a' (not 'b').

This "feature" goes back at least to the cctools implementation of as.

My gut reaction is that we should either add a warning, or fix this behavior, or add a warning, wait a while for people to fix their horrible code, then fix this behavior.

[lhames]

Not to be fixed for now, since this matches historical behavior. I'm going to leave this open though, since we may want to revisit this in the future.

[lhames]

mentioned in issue llvm/llvm-bugzilla-archive#25381

[lhames]

mentioned in issue llvm/llvm-bugzilla-archive#25381