[llvm-commits] Speeding up instruction selection

[Dan Gohman]

On Mar 31, 2008, at 5:09 AM, Roman Levenstein wrote:
> Hi Dan, Hi Evan,
>
> Here is the updated patch.
>
> 1) I introduced the SDOperandPtr class, as I suggested in my last
> email. This class is a sort of an "intelligent" pointer. With this
> class, there is no need to copy the SDOperand arrays into the
> temporary SmallVectors. Therefore there is no additional overhead.

I don't know about *no* additional overhead :-), but I don't
expect it'll be a big deal. I didn't expect the previous copying
solution to be a big deal either, but I think this new approach
is fine.

Please add a doxygen comment for the SDOperandPtr class to
document it.

In changes like this in SelectionDAG.h:

-                    const SDOperand *Ops, unsigned NumOps);
+                    const SDOperandPtr Ops, unsigned NumOps);

the const is now superfluous.

In the SDOperandPtr class,

+  SDOperandPtr(SDOperand * op_ptr) {
+    ptr = op_ptr;
+    object_size = sizeof(SDOperand);
+  }
+
+  SDOperandPtr(const SDOperand * op_ptr) {
+    ptr = op_ptr;
+    object_size = sizeof(SDOperand);
+  }

The first of these two constructors is unnecessary because
the second can be used in both cases.

>
> 2) Additionally, I also finally made SDOperand a member of the SDUse,
> instead of using it as a base class. Now we can be sure that no
> unintentional conversion of SDUse into SDOperand happens.

Ok.

>
>
> 3) I also changed the return types of some methods from SDUse to
> SDOperand, as it was suggested during the review.

Ok. Remember that Chris also suggested that many of these methods
can be changed to return an SDOperand by value instead of by
reference. If you decide to do that, please do it with a
separate commit.

With the issues mentioned above addressed, this patch looks good
to me.

Thanks,

Dan