[Cbe-oss-dev] elfspe & argument passing

[Jeremy Kerr]

Ulrich,

> Link with -mstd-main.

Thanks! I'll give that a go.

> >  * why does it need to be done in crt0 (hence causing an extra DMA)
> > anyway? This could done in libspe instead, where we set up the
> > original registers. We'd just need to copy the argv region to the
> > top of local store, and offset the stack pointer appropriately.
>
> For a long time, crt0 would ignore the incoming stack pointer
> and always reset $1 to the top of local store.

So this problem is fixed now? Is it reasonable to assume that a user is 
using a fixed crt0?

> Also, when we discussed this, it was considered more efficient
> to copy the arguments in via DMA.

This is for contexts that are context-switched out, so the writes to the 
mem file are backed by PPE memory. The entire PPE LS buffer will be 
DMA-ed to the SPE when the context is scheduled, so we might as well 
include argv[] there, and save an extra DMA.

This also means that we can simplify crt0 a little (doesn't need to do 
the DMA for argv), and maybe even remove the need for -mstd-main ?

Of course, this doesn't apply to running contexts, where the writes to 
the mem file are backed to the real LS, and would possibly be slower 
than the extra DMA. But if you're calling spe_program_run() on an 
already-running context, slower writes will be the least of your 
problems :)

Cheers,


Jeremy