[lucy-dev] Interesting (short) piece on code generation

[Kurt Starsinic <ks...@gmail.com>]

Hi all,

I thought this was an interesting tidbit on code optimization and
architecture-specific limitations. This is the kind of topic that has come
up a lot in the book club, but it's generally useful to know for writing
optimized code:

http://stackoverflow.com/questions/15349959/gcc-refuses-to-emit-long-calls-for-operator-new-delete-on-powerpc?newsletter=1&nlcode=75484%7c33c5

The takeaway for me is, if there's a compiler flag that you don't
understand (in this case, -mlongcall), then there's probably a situation
where you wished you'd supplied it. :^)

- Kurt

Re: [lucy-dev] Interesting (short) piece on code generation

[we...@ctosonline.org]

On Mar 12, 2013, at 6:43 PM, Marvin Humphrey wrote:
> My rule of thumb: It's _never_ a compiler bug.
> 
> But I've been wrong about that before. ;)

:-)

http://perl5.git.perl.org/perl.git/commitdiff/ef5eb418

Re: [lucy-dev] Interesting (short) piece on code generation

[Marvin Humphrey <ma...@rectangular.com>]

On Tue, Mar 12, 2013 at 2:10 PM, Kurt Starsinic <ks...@gmail.com> wrote:
> I thought this was an interesting tidbit on code optimization and
> architecture-specific limitations. This is the kind of topic that has come
> up a lot in the book club, but it's generally useful to know for writing
> optimized code:
>
> http://stackoverflow.com/questions/15349959/gcc-refuses-to-emit-long-calls-for-operator-new-delete-on-powerpc?newsletter=1&nlcode=75484%7c33c5

So, this kind of problem only occurs when generating very large objects --
specifically, very large `.text` sections within objects (which, despite the
name, are actually the machine code).  Some processors like to use narrower
width integers for calculating jump addresses -- in this case 24 bits.  When
performing a jump from an address on one end of the object to an address on
the other end, if you've got really big object code, the distance between the
pointers might be bigger than 2^24 -- and then you're screwed because the
24-bit integer overflows.

"You can't get there from here!"

> The takeaway for me is, if there's a compiler flag that you don't
> understand (in this case, -mlongcall), then there's probably a situation
> where you wished you'd supplied it. :^)

Except that setting that flag didn't seem to help the poster.  The Stack
Overflow denizens are claiming a compiler bug!

My rule of thumb: It's _never_ a compiler bug.

But I've been wrong about that before. ;)

Marvin Humphrey