[LLVMdev] 64 bit MRV problem; Missed optimizations.

[Frits van Bommel]

On Tue, Mar 8, 2011 at 8:37 AM, Duncan Sands <baldrick at free.fr> wrote:
>> I was tracking the issue discussed earlier, and I was wondering if a bug for the
>> missed optimizations, was ever filed, and if it has been fixed since? If so in
>> which llvm version, and more specifically which optimization pass.
>>
>> http://lists.cs.uiuc.edu/pipermail/llvmdev/2010-January/028877.html
>
> this doesn't really sound like an LLVM optimizer issue, it looks more like an
> issue with the llvm-g++ front-end to LLVM.  In order to conform to the platform
> ABI (x86-64 in this case), the front-end has to carefully arrange how parameters
> are passed to functions and return values handled.  This can result in nasty
> code that picks parameters apart and puts one bit in a float, another in an
> int etc.  It's often hard for the optimizers to do much about this - so the
> front-end needs to carefully do things in such a way as to help the optimizers
> as much as possible while maintaining ABI conformance.  I don't think this will
> ever be improved in llvm-g++ (which is now deprecated).  Hopefully clang does a
> better job.  I plan to rewrite the ABI stuff completely in dragonegg, so it may
> end up being fixed there one day.

When I wrote some ABI stuff for LDC, I noticed that a double is passed
the same way as <2 x float> on x86-64. My guess is that if what you're
passing is actually two floats instead of a double (and even if you're
passing a union that might be either) then passing it as <2 x float>
would result in much nicer code since LLVM should be able to figure
out how to take vectors apart if that's needed (or how to bitcast if
you want the double member of the union).