On Fri, 10 Jun 2011, Richard Henderson wrote:
> > + /* Do copy propagation */
> > + if (!(def->flags & (TCG_OPF_CALL_CLOBBER | TCG_OPF_SIDE_EFFECTS)))
> > {
>
> Why are you suppressing copy propagation in this way? I see no reason for it.
>
This was suggested by Aurelien Jarno. I do not know how safe it is to
propagate copies through such operations so I decided to be
conservative.
> > case INDEX_op_brcond_i64:
> > #endif
> > + memset(temps, 0, nb_temps * sizeof(struct tcg_temp_info));
>
> Perhaps to pedantic, but we can do better in extended basic blocks.
> A full flush of all temps is only needed at labels.
>
Not much better unfortunately. Globals got spilled at basic block end,
temps just die. The only things we can keep are locals but I have not
seen much of them in the intermediate representation.
----
Kirill Batuzov
