Re: [RFC] asm goto

[Richard Henderson <rth at redhat dot com>]

On 07/27/2009 03:29 PM, Adam Nemet wrote:
> Isn't the prediction logic a little too subtle here?  I.e. the fact that
> in
>
> // bb0
> if (foo)
>    goto L1;
>
> if (0)
> {
>    L1:
>    // bb1
> }
>
> // bb2
> L2:
>
> We assume that bb0->bb2 is more likely than bb0->bb1?  Is this BTW a
> heuristic we already have?  PRED_GOTO maybe?

A heuristic that minimizes the number of jumps will order the blocks 
properly, given that it can neither remove nor invert the asm goto jump.

That said, I have an additional plan for

  L1: __attribute__((cold))

that will affect the prediction logic such that the BB beginning with L1 
should be put into the .text.unlikely section (if supported), and edges 
to this label should be considered very unlikely.

But let's keep the two issues separate here.



r~