This is the mail archive of the mailing list for the GCC project.

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: [tree-profiling-branch PATCH IPCP extensions + Function cloning

> > What makes templates so special so you want to avoid the versioning
> > here?
> Many libstdc++ tests failed when we versioned templates.
> I'll continue to investigate this, in the meantime I want to avoid it.

I really don't quite like to add extra interface just to paper around
the bug.  In the case you are not shooting to enable the pass by default
yet, I would preffer to simply leave it breaking libstdc++ until it gets
fixed (I think it is fine given the current relatively experimental
nature of branch) or perhaps I can help here?
> >> bootstrap failed until this change was done...
> >>
> > Hmm, this sounds like latent inliner (or at least copy_body) bug.  Would
> > you mind to explain what went wrong without this hunk?
> When copying the edge between the EXIT block and its predecessor, the 
> flags of the original edge
> are not copied, but it is assigned with FALLTHROUGH flag. 
> It appears that there are some functions in which this particular edge has 
> a different flag. Copying the
> cfg of such a function, in such manner, causes ICE for the copied version. 
> (The cfg verification fails 
> if I remember correctly.)
> I can try and recreate this case if you wish. 

This sounds like fine explanation to me actually.  I would guess that we
didn't hit the disaster here as for inlining the extra edges don't
matter and we re-create them in RTL passes at least.  Please make the
chunk to happen unconditionally and kill the old code creating exit edge

> Razya

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]