firstVariableNumber initialization for LinearScan

[Doug Simon]

Hi Daniel,

On Mar 23, 2014, at 11:08 PM, D.Sturm <D.Sturm42 at gmail.com> wrote:

> The constructor for LinearScan uses "this.firstVariableNumber =
> registers.length;" to assign the first non-register id it can use.
> 
> That's a problem if the register numbering does have jumps in it, which
> happens to be true for the Aarch64 HotSpot runtime.

Doesn’t the com.oracle.graal.api.code.Architecture.registers for Aarch64 then have to contain dummy entries for the register numbers for which there no real registers? It would need to so as to conform with the specification of Architecture.getRegisters():

    /**
     * Gets an array of all available registers on this architecture. The index of each register in
     * this array is equal to its {@linkplain Register#number number}.
     */
    public Register[] getRegisters() {
        return registers.clone();
    }

This specification is what allows "this.firstVariableNumber = registers.length;” to be correct for all architectures. I wouldn’t be surprised if this property is assumed in other places as well.

So, you’d either need to use the dummy entries I mention above or use the translation table you mention below.

> Could we replace it with something like
>        int maxRegNumber = -1;
>        for (Register reg : this.registers) {
>            maxRegNumber = Math.max(maxRegNumber, reg.number);
>        }
>        this.firstVariableNumber = maxRegNumber + 1;
> 
> Otherwise I'd have to introduce a sequential numbering system for Graal and
> translate register numbers when transitioning from Graal to HotSpot.

-Doug