================
Comment at: lib/CodeGen/TargetInfo.cpp:3481
@@ -3457,1 +3480,3 @@
+ // the stack.
+ AllocatedVFP = 17;
return;
----------------
Oliver Stannard wrote:
> Renato Golin wrote:
> > Oliver Stannard wrote:
> > > Renato Golin wrote:
> > > > Why are you saturating? Is there any place that especially check
> > > > against 17? Or are all of them like this, >=16?
> > > On line 3336, we check whether AllocatedVFP > 16, as we need to know
> > > whether some CPRCs have been allocated on the stack.
> > Shouldn't we just change that check, then? I think it makes more sense.
> This can't be done just by changing the check.
>
> Without the line "AllocatedVFP = 17;", AllocatedVFPs will be set to 16 for
> the following two cases:
> * 8 doubles - exactly fills the VFP registers
> * 9 doubles - fills the VFP registers with no gaps, plus one double on the
> stack
>
> However, we need to differentiate between the two cases, as in the first we
> can split a struct between the GPRs and stack, but in the second case we
> cannot. (AAPCS rule C.5)
>
> Because of this, I am using (AllocatedVFPs > NumVFPs) to signify that at
> least one CPRC has been allocated to the stack, and the same with
> AllocatedGPRs. Once AllocatedVFPs > NumVFPs, we don't actually need to track
> the exact number of bytes used on the stack, only that this is non-zero.
Right, ok, makes sense. LGTM.
http://llvm-reviews.chandlerc.com/D2726
_______________________________________________
cfe-commits mailing list
[email protected]
http://lists.cs.uiuc.edu/mailman/listinfo/cfe-commits
