[Vm-dev] bitShift: and runtime sign discussion

[Igor Stasenko]

On 23 December 2012 18:05, David T. Lewis <lewis at mail.msen.com> wrote:
>
> On Sun, Dec 23, 2012 at 02:14:37PM +0100, Nicolas Cellier wrote:
>>
>> As you may know, the CCodeGenerator is clever enough to
>> #generateBitShift:on:indent: with proper direction << or >> when the
>> shift is literal.
>> When the shift in (expr bitShift: shift) is a variable or another
>> expression, it will produce a runtime test (shift <0) ? expr>>shift :
>> expr <<shift;
>> This is good for an arbitrary shift, but more than often we know the
>> sign and direction of the shift in advance, and this runtime test is
>> void.
>> We must be aware that aggressive inlining will spread these useless
>> code all around.
>> Well, we might ignore impact on performance but we shall better not
>> ignore warnings generated by C compiler (or XCode analyzer) if we want
>> the code to be portable to other compilers (including future versions
>> of gcc).
>> ...and the eventual warnings will be spreaded too
>>
>> Maybe we could arrange for CCodeGenerator to recognize more cases when
>> the shift is unsigned, but
>> 1) the CCodeGenerator is yet unable to infer expression type (except
>> some hack for declared variable)
>> 2) we almost never use unsigned declaration in vm source anyway.
>>
>> In the mean time, I suggest we use << and >> directlty in smalltalk
>> code (slang) when we know the direction in advance.
>
> IMO it is better to write these expressions directly, as you have done
> in your examples. There are many places in the VM code where types are
> not carefully declared, and I suspect that trying to make the code
> generator be more clever would only lead to unintended side effects.
> Better to just write these explicitly using << and >> when you know
> that it is safe to do so.
>
+1

-- 
Best regards,
Igor Stasenko.