[PATCH] Enable SLP-vectorization of intrinsics.

[Hal Finkel]

----- Original Message -----
> From: "Raul Silvera" <rsilvera at google.com>
> To: "Hal Finkel" <hfinkel at anl.gov>
> Cc: "Nadav Rotem" <nrotem at apple.com>, reviews+D2535+public+071e8fd9238b048b at llvm-reviews.chandlerc.com, "Chandler
> Carruth" <chandlerc at gmail.com>, "llvm-commits" <llvm-commits at cs.uiuc.edu>
> Sent: Tuesday, January 21, 2014 8:45:22 PM
> Subject: Re: [PATCH] Enable SLP-vectorization of intrinsics.
> 
> 
> 
> I had a closer look trying to get a case to work vectorizing sqrt.
> Unfortunately we fail to SLP vectorize even simple cases because the
> sqrt intrinsic (as well as the other math-related intrinsics) are
> marked as read-only, creating dependences with user stores.
> 
> 
> I found the code in include/llvm/IR/Intrinsics.td where these
> attributes are set, and there is a comment about them being treated
> conservatively to model their dependence to hardware FP rounding
> modes. This could be addressed by having a separate attribute to
> model FP hardware settings (think of it as sqrt reading a specific
> system memory location instead of reading all of user memory).
> 
> 
> Has something like that been considered in the past, or is there a
> different long-term strategy for dealing with these implicit
> dependences?

To be honest, I'm not exactly sure why we both modeling this at all. Clang/LLVM have a stated policy of not supporting floating-point environment access (#pragma STDC FENV_ACCESS ON is not supported -- Clang will complain if you try). Our modeling of these things is explicitly broken in other respects.

 -Hal

> 
> 
> If I manually remove the attribute we do SLP-vectorize a simple case
> and seem to get the expected speedups from using vector sqrt
> instructions.
> 
> 
> 
> On Fri, Jan 17, 2014 at 12:41 PM, Hal Finkel < hfinkel at anl.gov >
> wrote:
> 
> 
> 
> ----- Original Message -----
> > From: "Nadav Rotem" < nrotem at apple.com >
> 
> > To: "Hal Finkel" < hfinkel at anl.gov >
> > Cc:
> > reviews+D2535+public+071e8fd9238b048b at llvm-reviews.chandlerc.com ,
> > chandlerc at gmail.com , llvm-commits at cs.uiuc.edu ,
> > rsilvera at google.com
> > Sent: Friday, January 17, 2014 2:18:53 PM
> > Subject: Re: [PATCH] Enable SLP-vectorization of intrinsics.
> > 
> 
> > I think that we are converting some functions to llvm intrinsics.
> > Right?
> 
> Clang produces intrinsics for sqrt (only in fast-math mode), pow and
> fma. I think that's it, and until you get to CodeGen, I don't think
> we normally convert anything else at the LLVM level.
> 
> -Hal
> 
> 
> 
> > 
> > On Jan 17, 2014, at 12:11 PM, Hal Finkel < hfinkel at anl.gov > wrote:
> > 
> > > ----- Original Message -----
> > >> From: "Nadav Rotem" < nrotem at apple.com >
> > >> To: nrotem at apple.com , rsilvera at google.com
> > >> Cc: chandlerc at gmail.com , llvm-commits at cs.uiuc.edu
> > >> Sent: Friday, January 17, 2014 2:06:55 PM
> > >> Subject: Re: [PATCH] Enable SLP-vectorization of intrinsics.
> > >> 
> > >> 
> > >> The code looks great. I was wondering if you had a chance to run
> > >> some performance tests on it. Does it catch anything in the LLVM
> > >> test suite?
> > > 
> > > Why do you expect it to catch anything in the test-suite until we
> > > get the function calls too (just as we do in the loop
> > > vectorizer)?
> > > 
> > > -Hal
> > > 
> > >> 
> > >> http://llvm-reviews.chandlerc.com/D2535
> > >> _______________________________________________
> > >> llvm-commits mailing list
> > >> llvm-commits at cs.uiuc.edu
> > >> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
> > >> 
> > > 
> > > --
> > > Hal Finkel
> > > Assistant Computational Scientist
> > > Leadership Computing Facility
> > > Argonne National Laboratory
> > 
> > 
> 
> --
> Hal Finkel
> Assistant Computational Scientist
> Leadership Computing Facility
> Argonne National Laboratory
> 
> 
> 
> 
> --
> 
> 
> Raúl E. Silvera
> 
> 

-- 
Hal Finkel
Assistant Computational Scientist
Leadership Computing Facility
Argonne National Laboratory