Insight into the DMD back-end

[Don]

Stephan wrote:
> On 07.12.2010 06:57, %u wrote:
>> Yay for more comments like this one.
>>
>> Don 2010-12-06 11:53:27 PST
>> -- 
>> Bearophile -- That's an interesting [automatic fuzzy testing]link. 
>> Currently,
>> DMD back-end bugs are
>> being found at the rate of about 3 per year. So yes, fuzzy testing of 
>> DMC could
>> probably flush out some backend bugs a bit faster.
>> -------------------
>>
>> Here's what's happening. First, in this code:
>>
>>      for (int i = 0; i<  10; i++) {
>>          foo(i * 5 - 6);
>>      }
>> it sees that i and 10 are always>=0, so the signed comparison "i<  10" is
>> replaced with an unsigned one. (This happens in the backend in 
>> constprop() ).
>> Then, while dealing with loop invariants, it rewrites the loop into:
>>
>> for (int _i2 = -6; _i2<  10*5 - 6; _i2 += 5)
>> {
>>    foo(_i2);
>> }
>>
>> Fine. Except that it had changed the comparison into an unsigned one!
>> Particularly interesting is the case where the call is foo(i*5-50);
>> Then, the loop becomes:
>> for (int _i2 = -50; _i2<  0; _i2 += 5)
>>
>> Since an unsigned value is NEVER less than zero, it just drops the loop
>> completely!
>>
>> Nasty.
>> -- 
>> http://d.puremagic.com/issues/show_bug.cgi?id=5294
> 
> What i think is more disturbing is Walters response:
> 
>> I'm not sure how to fix that one yet, but it has been there for 25 years
>> now, so I am not sure it is urgent!
> 
> I often ran into this strange behaviour when using -O optimization 
> without knowing where it came from and it is so disturbing when i think 
> of people newly getting interested in D making the experience when 
> trying to compare it with C/C++ and then finding out the optimization 
> makes strange things. I think out of a image perspective such bugs must 
> be high priority, ESPECIALLY if it lies there for 25years already.

Yes, such bugs are given maximum priority. The only question is, should 
we delay the next release until this one is fixed?

BTW, I said there are about 3 of these bugs per year.
Here's the ones from the past two years, together with the time elapsed 
between reporting and fix:
2697 [6 months]
3521 [3 weeks]
3558 [10 weeks]
3633 [1 week]
3736 [2 weeks]
4443 [3 weeks]

Other wrong-code bugs were in the front end, or the glue layer. True 
back-end bugs are very rare, and are always very serious (largely 
because when you encounter them, they are almost impossible to recognize).