[DTrace-devel] [PATCH 01/10] Reduce register pressure in substr()

[Kris Van Hees]

On Sat, Mar 19, 2022 at 06:25:24PM -0400, Eugene Loh via DTrace-devel wrote:
> On 3/18/22 6:54 PM, Kris Van Hees wrote:
> 
> > On Fri, Mar 18, 2022 at 04:30:13PM -0400, Eugene Loh via DTrace-devel wrote:
> > > On 3/18/22 3:04 PM, Kris Van Hees via DTrace-devel wrote:
> > > > diff --git a/libdtrace/dt_cg.c b/libdtrace/dt_cg.c
> > > > +	emit(dlp,  BPF_MOV_REG(BPF_REG_2, str->dn_reg));
> > > >    	dt_regset_free(drp, str->dn_reg);
> > > >    	dt_cg_tstring_free(yypcb, str);
> > > > -	emit(dlp, BPF_MOV_REG(BPF_REG_3, idx->dn_reg));
> > > > +	emit(dlp,  BPF_MOV_REG(BPF_REG_3, idx->dn_reg));
> > > >    	dt_regset_free(drp, idx->dn_reg);
> > > >    	if (cnt != NULL) {
> > > >    		emit(dlp, BPF_MOV_REG(BPF_REG_4, cnt->dn_reg));
> > > > @@ -4030,8 +4024,17 @@ dt_cg_subr_substr(dt_node_t *dnp, dt_irlist_t *dlp, dt_regset_t *drp)
> > > >    	idp = dt_dlib_get_func(yypcb->pcb_hdl, "dt_substr");
> > > >    	assert(idp != NULL);
> > > >    	emite(dlp,  BPF_CALL_FUNC(idp->di_id), idp);
> > > > -	dt_regset_free_args(drp);
> > > > +
> > > > +	/*
> > > > +	 * Allocate the result register, and assign the result to it..
> > > > +	 */
> > > > +	dnp->dn_reg = dt_regset_alloc(drp);
> > > > +	if (dnp->dn_reg == -1)
> > > > +		longjmp(yypcb->pcb_jmpbuf, EDT_NOREG);
> > > > +
> > > > +	emit(dlp, BPF_MOV_REG(dnp->dn_reg, BPF_REG_0));
> > > >    	dt_regset_free(drp, BPF_REG_0);
> > > > +	dt_regset_free_args(drp);
> > > I think the dt_regset_free_args(drp) call should remain tight up against the
> > > BPF_CALL_FUNC().  After all, as soon as that function call is made, the regs
> > > are toast.  Or, say dnp->dn_reg is assigned one of %r1-%r5 and therefore
> > > spills that register.  (Does that cause the BPF verifier to complain?)  Then
> > > we free_args() and fill the register, overwriting what was just there.
> > We should be avoiding crossing lines with alloc/free of registers,
> Why?  This isn't a stack/LIFO situation.
> 
> Further, we do this all the time.  E.g., we alloc regs for the func call,
> then we alloc dnp->dn_reg, then we free the func-call regs, and later we
> will free dnp->dn_reg.  We do this sort of thing routinely.

Yes, and I intend to fix those things once I have a better register spilling
mechanism finalized.

> Also, what about the scenario I raised?  E.g., let's say dnp->dn_reg is
> assigned %r5.  After all, the situation is that we are facing register
> pressure.  So:
> 
> *)  alloc %r1-%r5
> *)  alloc %r0
> *)  call func (%r0 is given a value while %r1-%r5 are clobbered)
> *)  alloc dnp->dn_reg, getting %r5, whose clobbered value is spilled to
> stack
> *)  %r5 = %r0
> *)  free %r0
> *)  free %r1-%r5, filling %r5 (dnp->dn_reg) with the clobbered value
> > so if we
> > alloc the args first and then %r0, we should free %r0 first, and then the args.

We do not actually have a valid solution for this scenario in the more general
case.  So, while a vlaid issue, I am not going to address it because the whole
register spilling support is fragile at best and is in dire need of reworking.
I started that but we shouldn't be holding up these new features until that is
done.

Sadly, we are going to live with the register spilling issues for a little
while longer.  But with the implementation of a push/pop stack facility, we
should be able to do better.