proposed syntax change

[Robert Jacques]

On Fri, 07 Aug 2009 01:22:31 -0700, Yigal Chripun <yigal100 at gmail.com>  
wrote:

> Robert Jacques wrote:
>> On Thu, 06 Aug 2009 13:34:21 -0700, Yigal Chripun <yigal100 at gmail.com>  
>> wrote:
>>
>>> Ary Borenszweig wrote:
>>>> Adam D. Ruppe wrote:
>>>>> On Thu, Aug 06, 2009 at 03:06:49PM -0400, Paul D. Anderson wrote:
>>>>>> Oh wait...I think "//" is used elsewhere.
>>>>>
>>>>> Is this a joke?
>>>>  No. When porting C, C++, Java or C# code just search "//" and  
>>>> replace it with "--".
>>>>  Oh wait... I think "--" is used elsewhere.
>>>
>>> Why would I as a user want to have two ops that do the same thing?!
>>> Python's solution for this is wrong IMHO since the problem is not with  
>>> the op itself but rather with the way C handles numbers.
>>>
>>> 5 / 2 should always be the more precise FP division unless the  
>>> compiler knows or is instructed to do otherwise.
>>>
>>> int a = 5 / 2; // compiler knows to use integer division
>>  No, it doesn't. (i.e. Welcome to the limitations of a context-free  
>> grammar) The right hand of the expression has to be evaluated before  
>> the left, or otherwise function overloads, etc, don't work, so there's  
>> no way for the compiler to know the type of the expected result when  
>> 5/2 is evaluated.
>>
>>> auto b = 5 / 2; // b is double, FP division
>>>
>>> auto c = cast(int)(5/2)); // compiler instructed to use integer  
>>> division
>>  No, the cast would apply to the result of (5/2), not the '/' operator.
>>
>>> auto d = floor(5 / 2); // FP division floored by a function to int
>>  floor returns a real, not an int. I think you were looking for  
>> roundTo!int or roundTo!int(floor(5/2))
>>
>>> auto f = std.math.div(5, 2); // intristic that does integer division
>>>
>>> what's the rationale of not doing the above, besides C compatibility?
>>  The rationale is that integer division is very common and is usually  
>> assigned back into an int, and not a real. The only issue is literals,  
>> which were going to be handled with polysemous values (but that got  
>> dropped). But 2/5.0 is really not that much overhead.
>
> As you noted yourself, polysemous types help solve this and also the  
> compiler can have special handling (peep hole optimizations) for some of  
> the above cases.
> in case D roundTo sounds indeed better (and since in the above it's  
> compile time, it should be possible to optimize it.
>
> you've ignored case f which seems to me the most important: instead of  
> currently 5/2 == 2 there should be a div intristic function such that  
> div(5, 2) == 2  and that intristic will be the appropriate ASM  
> instruction for integer division.

No, I didn't. I just didn't have anything to say.

> In general, you can always round down a float but you can't get the  
> correct double out of the rounded int, so my question is how much a  
> performance hit is it to use the FP as default and only use the div  
> intristic where performance is really an issue?

Now that I think about it, turning integer division into a function is a  
really bad idea: 1) there's function call overhead and 2) even if inlined,  
register values and the stack still have to be manipulated (as function  
call syntax dictates where the inputs and outputs are located) Today, DMD  
can choose any set of register inputs and outputs.

As for the int-float conversion, apperently float-int is somewhat slow and  
destroys the floating point pipeline
 From http://mega-nerd.com/FPcast/
> The instruction which causes the real damage in this block is fldcw,  
> (FPU load control word) on lines 8 and 11. Whenever the FPU encounters  
> this instruction it flushes its pipeline and loads the control word  
> before continuing operation. The FPUs of modern CPUs like the Pentium  
> III, Pentium IV and AMD Athlons rely on deep pipelines to achieve higher  
> peak performance. Unfortunately certain pieces of C code can reduce the  
> floating point performance of the CPU to level of a non-pipelined FPU.

> So why is the fldcw instruction used? Unfortunately, it is required to  
> make the calculation meet the ISO C Standard which specifies that  
> casting from floating point to integer is a truncation operation.  
> However, if the fistpl instruction was executed without changing the  
> mode of the FPU, the value would have been rounded instead of truncated.  
> The standard rounding mode is required for all normal operations like  
> addition, subtraction, multiplication etc while truncation mode is  
> required for the float to int cast. Hence if a block of code contains a  
> float to int cast, the FPU will spend a large amount of its time  
> switching between the two modes.