[PATCH] D28404: IRGen: Add optnone attribute on function during O0

[Robinson, Paul via cfe-commits]

(Re-add cfe-commits; otherwise same)

> -----Original Message-----
> From: cfe-commits [mailto:cfe-commits-bounces at lists.llvm.org] On Behalf Of
> Duncan P. N. Exon Smith via cfe-commits
> Sent: Monday, January 09, 2017 4:10 PM
> To: reviews+D28404+public+53e0f4655ef79c80 at reviews.llvm.org
> Cc: nhaehnle at gmail.com; wei.ding2 at amd.com; jholewinski at nvidia.com; Richard
> Smith; cfe-commits; Peter Collingbourne
> Subject: Re: [PATCH] D28404: IRGen: Add optnone attribute on function
> during O0
> 
> This seems like a massive rehash of a discussion Peter Collingbourne and I
> had about passing -O0 to the linker for -flto=full.  I had previously
> thought of LTO as "link time optimization", but in practice it's useful
> for (and required for correctness of some) non-optimization IR passes.
> 
> In other words, the basic question seems to be: "Should LTO support non-
> optimization use cases?"  I tend (now) to think it should -- having
> "optimization" in its name is an historical artifact -- because adding
> another way to run IR passes at link-time seems redundant.  Whereas, Paul,
> it seems like you disagree?

I am persuaded that there are non-optimization-based uses for clumps of
bitcode modules linked together.  (We could redefine the TLA if we like;
LTO = Link Time Operation?)

I am equally convinced that we have no good story for propagating a
variety of optimization- and codegen-related options to the top-level
LTO processor.  This is most especially true when different CUs might
reasonably have different options.  -O0 is the example at hand, but 
this problem seems to keep coming up and we keep hacking in ways to get 
the thing we think we need in the moment.

> 
> (Also, this discussion seems higher level than just the patch at hand...
> maybe llvm-dev would be more appropriate?)

I'd be fine with that.
--paulr
> 
> > On 2017-Jan-09, at 16:03, Paul Robinson via Phabricator
> <reviews at reviews.llvm.org> wrote:
> >
> > probinson added a comment.
> >
> > In https://reviews.llvm.org/D28404#640588, @mehdi_amini wrote:
> >
> >> Actually, as mentioned before, I could be fine with making `O0`
> incompatible with LTO, however security features like CFI (or other sort
> of whole-program analyses/instrumentations) requires LTO.
> >
> >
> > Well, "requires LTO" is overstating the case, AFAICT from the link you
> gave me.  Doesn't depend on //optimization// at all.  It depends on some
> interprocedural analyses given some particular scope/visibility boundary,
> which it is convenient to define as a set of linked bitcode modules, that
> by some happy chance is the same set of linked bitcode modules that LTO
> will operate on.
> >
> > If it's important to support combining a bitcode version of my-
> application with your-bitcode-library for this CFI or whatever, and you
> also want to let me have my-application be unoptimized while your-bitcode-
> library gets optimized, NOW we have a use-case.  (Maybe that's what you
> had in mind earlier, but for some reason I wasn't able to extract that out
> of any prior comments.  No matter.)
> >
> > I'm now thinking along the lines of a `-foptimize-off` flag (bikesheds
> welcome) which would set the default for the pragma to 'off'.  How is that
> different than what you wanted for `-O0`?  It is defined in terms of an
> existing pragma, which is WAY easier to explain and WAY easier to
> implement.  And, it still lets us say that `-c -O0 -flto` is a mistake, if
> that seems like a useful thing to say.
> >
> > Does that seem reasonable?  Fit your understanding of the needs?
> >
> >
> > https://reviews.llvm.org/D28404
> >
> >
> >
> 
> _______________________________________________
> cfe-commits mailing list
> cfe-commits at lists.llvm.org
> http://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits