[llvm-commits] [llvm] r126964 - in /llvm/trunk: lib/CodeGen/SelectionDAG/LegalizeIntegerTypes.cpp lib/CodeGen/SelectionDAG/LegalizeTypes.h test/CodeGen/X86/umulo-64.ll

[Eric Christopher]

On Mar 3, 2011, at 4:57 PM, Eli Friedman wrote:

> On Thu, Mar 3, 2011 at 4:21 PM, Eric Christopher <echristo at apple.com> wrote:
>> 
>> On Mar 3, 2011, at 4:16 PM, Eric Christopher wrote:
>> 
>>> Here's a simple patch that fixes the problem in something resembling correctness, but uses divide to do it - so yeah, performance?
>>> 
>>> Thoughts?
>> 
>> Uh, other than divide by zero.
>> 
>> *fixes*
>> 
>> -eric
> 
> It should work.  The only other approach I can think of involves
> essentially taking your previous approach, but doubling the width of
> the multiplication.  There isn't any code in LLVM at the moment to
> turn a  64x64->128 multiply into 32x32->64 multiplies, but it is
> feasible.

Yeah. That's what I was originally going for, but completely missed it.

Oddly I can't find a processor that actually dies a horrible death if I divide by zero.

-eric