Interesting: http://blog.regehr.org/archives/1384 -- Andrei

On 4/12/2016 5:06 AM, Andrei Alexandrescu wrote:
> Interesting: http://blog.regehr.org/archives/1384 -- Andrei

Curiously never mentioned is the following optimization:

    return a+b*2+27;

becomes:

    LEA EAX,27[ESI][EDI*2]

To overflow check:

    ADD EDI,EDI
    JO overflow
    ADD EDI,27
    JO overflow
    MOV EAX,ESI
    ADD EAX,EDI
    JO overflow

I don't see efficiency there, even with the JO's being free.

On Wednesday, 13 April 2016 at 22:13:27 UTC, Walter Bright wrote:
> On 4/12/2016 5:06 AM, Andrei Alexandrescu wrote:
>> Interesting: http://blog.regehr.org/archives/1384 -- Andrei
>
> Curiously never mentioned is the following optimization:
>
>     return a+b*2+27;
>
> becomes:
>
>     LEA EAX,27[ESI][EDI*2]
>
> To overflow check:
>
>     ADD EDI,EDI
>     JO overflow
>     ADD EDI,27
>     JO overflow
>     MOV EAX,ESI
>     ADD EAX,EDI
>     JO overflow
>
> I don't see efficiency there, even with the JO's being free.

It is clearly not as optimal, but still pretty good. The article doesn't pretend it all come for free, just that it comes for much cheaper than before.

On Thursday, 14 April 2016 at 02:55:01 UTC, deadalnix wrote:
> On Wednesday, 13 April 2016 at 22:13:27 UTC, Walter Bright wrote:
>> On 4/12/2016 5:06 AM, Andrei Alexandrescu wrote:
>>> Interesting: http://blog.regehr.org/archives/1384 -- Andrei
>>
>> Curiously never mentioned is the following optimization:
>>
>>     return a+b*2+27;
>>
>> becomes:
>>
>>     LEA EAX,27[ESI][EDI*2]
>>
>> To overflow check:
>>
>>     ADD EDI,EDI
>>     JO overflow
>>     ADD EDI,27
>>     JO overflow
>>     MOV EAX,ESI
>>     ADD EAX,EDI
>>     JO overflow
>>
>> I don't see efficiency there, even with the JO's being free.
>
> It is clearly not as optimal, but still pretty good. The article doesn't pretend it all come for free, just that it comes for much cheaper than before.

Also, just checked, on sandy bridge, the LEA has 3clock latency (but start earlier in the pipeline) and the add 1, so it is not as bad as it looks (it is still bad).

On 4/13/2016 8:00 PM, deadalnix wrote:
> Also, just checked, on sandy bridge, the LEA has 3clock latency (but start
> earlier in the pipeline) and the add 1, so it is not as bad as it looks (it is
> still bad).

The size is larger, too (not so cache friendly). Integer arithmetic is sort of the bread and butter of computer programs, how much slowdown will people accept?

On Thursday, 14 April 2016 at 04:52:12 UTC, Walter Bright wrote:
> On 4/13/2016 8:00 PM, deadalnix wrote:
>> Also, just checked, on sandy bridge, the LEA has 3clock latency (but start
>> earlier in the pipeline) and the add 1, so it is not as bad as it looks (it is
>> still bad).
>
> The size is larger, too (not so cache friendly). Integer arithmetic is sort of the bread and butter of computer programs, how much slowdown will people accept?

Maybe not everywhere, but some security related code and/or tool like ubsan surely can benefit from this.