[llvm-testresults] Geryon-X86-64 i386 nightly tester results

[Dan Gohman]

On Apr 8, 2009, at 11:52 AM, Dale Johannesen wrote:

>
> On Apr 8, 2009, at 11:45 AMPDT, Dan Gohman wrote:
>
>>
>> On Apr 8, 2009, at 2:39 AM, Apache wrote:
>>>
>>> multisource/Benchmarks/Trimaran/netbench-url/netbench-url: -22.11%
>>> (2.85 => 3.48)
>>
>> This is reproducible, though it all comes down to this one
>> change in one hot 14-instruction loop:
>>
>> -    movl %r13d, %ecx
>> -    addl %ebp, ecx
>> +    leal (%rbp,%r13), %ecx
>>
>> This is exactly what the two-address lowering heuristics
>> are designed to do; the change is that they're no longer
>> missing this case. Fixing this would require some new
>> heuristics that can decide when mov+add is really better
>> than lea.
>
> I believe this is replacing a 32-bit add with a 64-bit add.  Would  
> it help to write the lea as
>   leal (%ebp, %r13d), %ecx
> ?  We ought to be able to do that when the result is 32 bits.


It doesn't appear to make a significant difference in this testcase.

Dan