Could assertThrown be made safe/trusted?

[Timon Gehr]

On 01/17/2012 11:31 PM, Jonathan M Davis wrote:
> On Tuesday, January 17, 2012 14:14:36 H. S. Teoh wrote:
>> On Tue, Jan 17, 2012 at 05:04:18PM -0500, Jonathan M Davis wrote:
>> [...]
>>
>>> Attribute inferrence is a big step forward in making as much as
>>> possible @safe and pure, but there's still plenty to do there.
>>
>> [...]
>>
>> Funny you should mention that, I was just starting to wonder if I should
>> start littering my code with 'pure', and whether it's possible to make
>> the compiler infer it for me. From the little that I know, it seems that
>> in most cases 'pure' can be automatically inferred. The compiler already
>> distinguishes between weakly pure and strongly pure internally, so why
>> not take it all the way? Not sure how this will affect inter-module
>> analysis, though.
>>
>> But since this is apparently not yet implemented, just what *is*
>> implemented currently when you specify 'pure'? Common subexpression
>> factorization? Hoisting? Not (yet) memoization, apparently.
>
> _pure_ is implemented. It's @safe that isn't fully implemented. pure, @safe,
> and nothrow are inferred for templated functions when they're instantiated so
> that they can be pure (or @safe or nothrow) based on the code that's generated
> rather than always forcing it to be one or the other, since that would be far
> too restrictive. But that's completely unnecessary for normal functions. You
> _do_ need to mark those pure, @safe, or nothrow yourself.
>
> If attributes were inferred for normal functions, the compiler would always
> have to have the full source of every function. And even then, it might be an
> instance of the halting problem. Every function is and must be pure (or @safe
> or nothrow) or not when it's declared, and that's part of its signature, so it
> can be known even when the full source isn't. Inference works with templates
> only because they're generating code, and the compiler needs their full source
> anyway.
>
> - Jonathan M Davis

I think he is interested in the state of implementation of specific 
compiler _optimisations_ that make use of function purity in order to 
prove their correctness. IIRC ldc has CSE for pure functions, but I 
don't know exactly.