[llvm-commits] [patch] minor instcombine+bitcast improvement

[Chris Lattner]

On Apr 6, 2007, at 10:51 AM, Reid Spencer wrote:
> On Fri, 2007-04-06 at 10:45 -0700, Chris Lattner wrote:
>> On Apr 6, 2007, at 9:01 AM, Dan Gohman wrote:
>>> This patch makes instcombine convert this:
>>>      %t = getelementptr float* cast ([2 x i32]* %str to float*),
>>> i32 %V
>>>
>>> into this:
>>>
>>>      %t = getelementptr float* cast ([2 x i32]* %str to [2 x float]
>>> *), i32 0, i32 %V
>>>
>>> instead of this:
>>>
>>>      %s = getelementptr [2 x i32]* %str, i32 0, i32 %V
>>>      %t = bitcast i32* %s to float*
>>>
>>> I'm not yet familiar with instcombine's canonicalization style; is
>>> this
>>> a desireable change? It doesn't cause any regressions for me.
>>
>> I don't have a strong preference either way.  Is there a reason that
>> you prefer this style to the previous one?
>
> That was my take too.
>
> I don't think this affects code gen at all so the only question is,  
> what
> else does it improve? It doesn't improve readability in my opinion but
> that's not a high priority.  Using the cast constant expression  
> instead
> of the cast instruction might be smaller but I'm not sure. So, I have
> the same question: "What's the benefit to this transform?"

The transform in CVS allows us to turn:

%t = getelementptr float* cast ([2 x i32]* %str to float*), i32 %V
%u = bitcast float* to i32*

into this:

      %s = getelementptr [2 x i32]* %str, i32 0, i32 %V
      %t = bitcast i32* %s to float*
      %u = bitcast float* to i32*

into this:

      %u = getelementptr [2 x i32]* %str, i32 0, i32 %V

-Chris