My opinions on tree-level and RTL-level optimization

[Roger Sayle]

On Mon, 18 Apr 2005, Paolo Bonzini wrote:
> Roger proposed lowering 64-bit arithmetic to 32-bit in tree-ssa!  How
> would you do it?  Take
>
>      long long a, b, c;
>      c = a + b;
>
> Would it be
>
>      c = ((int)a + (int)b)
>          + ((int) (a >> 32) + (int) (b >> 32)
>             + ((unsigned int) a < (unsigned int) b)) << 32;
>
> Or will you introduce new tree codes and uglifying tree-ssa?
> Seriously...


I think you may have misinterpreted or over-interpreted
what I meant by performing optimizations/lowering, earlier vs.
later.  When I suggested the i386.c backend should lower DImode
operations earlier, my intention was to lower during RTL expansion
instead of where it currently happens after reload.  Whilst doing
it in tree-ssa would technically also be "doing it earlier", this
isn't really what I had in mind.

The fact that DImode moves aren't lowered, and that the high and
low parts of a DImode register can't be independently live/dead
(for register allocation) is the reason for PR17236, and why GCC
generates poorer code than Intel on something as simple as "a * b".


Not that it would be impossible to do this with trees.  It's not
uncommon for representations to be lowered, for example high-level
trees to low-level trees.  RTL has always done this, where addressof
could be seen as a lowering pass, as can combine, reload, regstack
and the process of splitting,  where the RTL representation afterwards
is a more accurate description of the generated code than the RTL
representation before.

Roger
--