[cfe-commits] r63413 - /cfe/trunk/lib/CodeGen/CGCall.cpp

[Chris Lattner]

On Jan 30, 2009, at 11:46 AM, Douglas Gregor wrote:

>
> On Jan 30, 2009, at 11:38 AM, Daniel Dunbar wrote:
>
>> Author: ddunbar
>> Date: Fri Jan 30 13:38:39 2009
>> New Revision: 63413
>>
>> URL: http://llvm.org/viewvc/llvm-project?rev=63413&view=rev
>> Log:
>> x86_64 ABI: Pass <1 x double> in memory. This is arguably wrong, but
>> matches gcc 4.2 (not llvm-gcc).
>
> How frightening. Did you check any other compilers, e.g., Intel?
>
> I suggest that you put some kind of marker in the source for this kind
> of issue, e.g., FIXME or BUGCOMPAT or something like it, so we know
> where we've intentionally deviated from the specification.

The issue here is that we haven't even tried to make llvm-gcc  
compatible with "extended vector types", i.e. those that don't map  
natively onto a hardware vector.  These are *extremely* uncommon in  
practice.  llvm-gcc should be eventually fixed, so agreeing with  
gcc-4.2 is the right way to go.

-Chris