Compiler bug?

[Steven Schveighoffer]

"Steven Schveighoffer" wrote
> "Jarrett Billingsley"wrote
>> "Steven Schveighoffer" wrote
>>>I think I found a bug in the compiler.  This bug seems to have affected 
>>>the two versions I tried of dmd (1.024 and 1.033), so it's been in there 
>>>a while.
>>>
>>> However, it's a very weird bug.
>>>
>>> I have a function that looks like this (for those who are interested, 
>>> this is from tango.io.Path):
>>>
>>> Time modified (char[] name)
>>> {
>>>        return timeStamps(name).modified;
>>> }
>>>
>>> the timeStamps function returns a collection of 3 Time structs, which 
>>> indicate the modified, created, and accessed time for a file named 
>>> 'name'.
>>>
>>> However, the result is corrupt.  I know this because I'm printing the 
>>> time in the timeStamps function, and it is different than the time 
>>> returned.
>>>
>>> In examining the entire flow of the code, I can find nothing wrong.
>>>
>>> This only happens with the -inline switch of dmd.
>>>
>>> If I rewrite the function in the following way, the problem goes 
>>> away(!):
>>>
>>> Time modified (char[] name)
>>> {
>>>  auto ts = timeStamps(name);
>>>        return ts.modified;
>>> }
>>>
>>> Now, I wanted to create a minimal case (without all of Tango that this 
>>> program uses), but I'm having difficulty.  So I wanted to inspect the 
>>> assembly, but I'm on windows, and there is no obj2asm in the dmd 
>>> directory. So I can't figure out what the compiler is doing differently. 
>>> I'm not sure where to begin.  Walter, is this enough info for you to 
>>> look into this?  I can send you the object file and source file if you 
>>> want to see how dmd generated the code.
>>
>> You have to buy the EUP to get the windows version of obj2asm, lol.
>
> sweet deal :)  So if Walter wants me to debug this, maybe he can send me a 
> copy of obj2asm for free :)
>
>>
>> Another way to disassemble is to compile in debug info with -g, and step 
>> through in a debugger, like WinDbg (shudder), ddbg (OK), or VS6 (yay). 
>> I'm not sure how -g behaves with -inline though.
>
> Yeah, I had lots of trouble with the debugger.  But the printout alone is 
> proof that it is failing.
>
>>
>> But if it makes you feel any better (?), this is definitely a 
>> g.  -inline should *never* change the behavior of code.  Bugzilla it if 
>> it hasn't been already.
>
> OK, so should I just attach all tango source used to the bug report? :P
>
> I suppose I could try and whittle it down.
>
> I'm just really surprised this is the first case.  It makes me think that 
> there is something really 'moons-are-aligned-properly' about the way 
> everything is compiled.
>
> -Steve

Got to a minimal Phobos-based case:

http://d.puremagic.com/issues/show_bug.cgi?id=2232

-Steve