[LLVMdev] Optimization opportunity

[Chris Lattner]

On Sat, 28 Aug 2004, Jeff Cohen wrote:
> I succumbed to temptation and made the improvement.  Diffs are attached
> for X86ISelSimple.cpp and X86InstrBuilder.h.

The patches look great, that's a very elegant way to address the problem,
and also simplified some code, thanks!

> I determined that the reason two instructions are generated in the first
> place, instead of being folded immediately into one, is because locals
> do not have a physical offset assigned at that time.  There is a
> peephole optimization pass after the stack frame is finalized, but the
> problem with folding them there is that the physical registers have also
> been assigned, so register CX (in this case) would be wasted.

Yup.

> So I generalized the concept of a "full address".  There were four
> variables, base reg, scale, index reg and displacement, that were being
> passed around.  I converted them into a single struct and added a new
> type field that allows the base reg to instead be a frame index.  The
> extra lea instruction is now folded immediately, and the code for
> processing store instructions shrinks quite nicely.

Sounds good.

> This should be generalizable even further to handle global variables.  I
> didn't go that far, because it appeared there may be other work
> required.  It wasn't clear a displacement to a global was supported.

That would be cool as well.  A displacement from a global should be
supported.  Look at the code generated for:

int G;
void foo() { G = 4; }

... and you'll see what I mean.  :)

-Chris

-- 
http://llvm.org/
http://nondot.org/sabre/