On C/C++ undefined behaviours

[Tomek Sowiński]

Dnia 21-08-2010 o 05:36:28 SK <sk at metrokings.com> napisał(a):

> Then to make this more concrete, what if D had an option to suspend
> compilation after the front-end finished?  The resulting executable
> contains the abstract RTL blobs and the compiler backend, which
> finishes the job for the specific platform on which the executable is
> launched.  The final binary is cached for subsequent launches.  You
> get good machine independence and the approach provides performance
> wins for operations like vectorizing where you don't know in advance
> what kind of SSE support you'll find.
>
> Hypothetically, why not?
>
> I dug around and found this:
> http://www.itjungle.com/tfh/tfh082007-printer01.html
>
>> From the article.
> Without getting too technical, here's what happens on the OS/400 and
> i5/OS platform when you create applications, which explains the
> problem customers ran into in 1995 and which IBM wants them to avoid
> in 2008. A programmer writes an application in say, RPG. They run it
> through a compiler, either using the Original Program Model (OPM) or
> the Integrated Language Environment (ILE) compilers, and the code
> compiles so they can run it. Or, rather, that is what it looks like to
> the programmer. What is really happening is that this application is
> compiled into an intermediate stage, which some IBMers have called RPG
> templates (in the case of RPG applications). These templates have a
> property called observability, which in essence means they are
> compiled to the TIMI layer. These intermediate templates are then used
> by the TIMI layer on an actual piece of hardware with a specific
> processor and instruction set to compile the application to run on
> that specific processor. TIMI compiles these RPG templates down to
> actual compiled code behind the scenes the first time an application
> runs, and because the code was originally compiled to the TIMI layer,
> there is no need to change the source code. Only the object code
> changes, which end users never had access to anyway because only TIMI
> can reach down there. This is the brilliant way that IBM has preserved
> customers' vast investments in RPG, COBOL, and other applications over
> the years

Interesting. Hypothesising, wouldn't it be better if compiling to metal  
happened already during installation? On the first run the compiler still  
has to hurry (1st impression counts), but downloading over network is so  
slow that while the user is watching the progress bar there'd be plenty of  
time for compiling the intermediate representation and expensive  
optimizations without even being noticed by the user.


Tomek