[Jikesrvm-researchers] i2f implementation on PPC

[Eliot M. <mo...@cs...>]

>>>>> "Feng" == Feng QIAN <fen...@ma...> writes:

    Feng> Hi,
    Feng> The implementation of i2f in the PPC baseline compiler seems not correct:

    Feng>   /**
    Feng>     * Emit code to implement the i2f bytecode
    Feng>     */
    Feng>    protected final void emit_i2f() {
    Feng>      asm.emitLWZ    (T0,  0, SP);               // T0 is X (an int)
    Feng>      asm.emitCMPI  (T0,  0);                   // is X < 0
    Feng>      asm.emitLFDtoc(F0, VM_Entrypoints.IEEEmagicField.getOffset(), T1);  // F0 
    Feng> is MAGIC
    Feng>      asm.emitSTFD  (F0, -4, SP);               // MAGIC on stack 

    Feng>      asm.emitSTW   (T0,  0, SP);               // if 0 <= X, MAGIC + X
    Feng>      VM_ForwardReference fr = asm.emitForwardBC(GE);
    Feng>      asm.emitLWZ    (T0, -4, SP);               // T0 is top of MAGIC
    Feng>      asm.emitADDI   (T0, -1, T0);               // decrement top of MAGIC
    Feng>      asm.emitSTW   (T0, -4, SP);               // MAGIC + X is on stack
    Feng>      fr.resolve(asm);
    Feng>      asm.emitLFD   (F1, -4, SP);               // F1 is MAGIC + X
    Feng>      asm.emitFSUB    (F1, F1, F0);               // F1 is X
    Feng>      asm.emitSTFS  (F1,  0, SP);               // float(X) is on stack
    Feng>    }


    Feng> it uses one more stack slot than it is allowed. For some methods, actually it 
    Feng> overwrites the compiled method id, that's the source of problem Chris and I 
    Feng> encountered in previous days.

    Feng> Can we use S0 to hold temporary value instead of use SP-4 ?

I see the problem, but you cannot move values directly between GP and FP
registers: they have to flow through memory. And here, we are dealing with
a double, which clearly takes 8 bytes, whereas the input and output are
each 4 bytes long. A static location won't do (could be accessed by
multiple threads at same time), so we probably need to use a field in the
VM_Processor (fast and simple alternative to the more obvious notion of a
thread local varable) to hold the scratch floating point number. This is
not hard to recode, just a bit annoying (and makes VM_Processor less cleab
-- but I think there are other cases a bit like this).

Other ideas?

-- Eliot