[llvm-commits] [llvm] r79553 - in /llvm/trunk: lib/Target/X86/X86ISelDAGToDAG.cpp test/CodeGen/X86/abi-isel.ll test/CodeGen/X86/codemodel.ll test/CodeGen/X86/rip-rel-address.ll

[Dan Gohman]

On Aug 20, 2009, at 2:30 PM, Chris Lattner wrote:


>
> On Aug 20, 2009, at 11:23 AM, Dan Gohman wrote:
>
>
>> Author: djg
>>
>> Date: Thu Aug 20 13:23:44 2009
>>
>> New Revision: 79553
>>
>>
>>
>> URL: http://llvm.org/viewvc/llvm-project?rev=79553&view=rev
>>
>> Log:
>>
>> Fix an x86 code size regression: prefer RIP-relative addressing
>>
>> over absolute addressing even in non-PIC mode (unless the address
>>
>> has an index or something else incompatible), because it has a
>>
>> smaller encoding.
>>
>
> Hi Dan,
>
> Are you sure sure that this is safe?  Unless you have a strong  
> reason not to, I'd prefer to hold this until after 2.6.

LLVM has done this for quite a while; it was only recently lost in the
AsmPrinter whirlwind. The x86-64 ABI document has a sentence dedicated
to explicitly permitting this. The ABI document also uses this in all
relevant examples. And, it's what GCC does. So yes, I'm fairly sure
it's safe.

Dan