[llvm-commits] [PATCH/RFC] Loop Data Software Prefetching

Hal Finkel hfinkel at anl.gov
Thu Jan 31 06:59:48 PST 2013


----- Original Message -----
> From: "Hal Finkel" <hfinkel at anl.gov>
> To: "Nadav Rotem" <nrotem at apple.com>
> Cc: "l >> Commit Messages and Patches for LLVM" <llvm-commits at cs.uiuc.edu>
> Sent: Thursday, January 31, 2013 1:49:06 AM
> Subject: Re: [llvm-commits] [PATCH/RFC] Loop Data Software Prefetching
> 
> ----- Original Message -----
> > From: "Nadav Rotem" <nrotem at apple.com>
> > To: "Hal Finkel" <hfinkel at anl.gov>
> > Cc: "Evan Cheng" <evan.cheng at apple.com>, "l >> Commit Messages and
> > Patches for LLVM" <llvm-commits at cs.uiuc.edu>
> > Sent: Thursday, January 31, 2013 1:25:35 AM
> > Subject: Re: [llvm-commits] [PATCH/RFC] Loop Data Software
> > Prefetching
> > 
> > > 
> > > To be honest, I think that this is kind of silly. Some
> > > target-provided parameters are fixed properties of the target,
> > > and
> > > so no user-level customization needs to be provided. Other
> > > things,
> > > however, like whether to use software prefetching, the
> > > prefetching
> > > distance, etc. are just heuristic inputs, and we need to provide
> > > the ability for the user to override them.
> > 
> > In that case it needs to be a command line flag of the prefetcher
> > pass. Just like the loop-unrolling pass exposes the unrolling
> > threshold.
> 
> You might be right that this is a non-problem, but it seems fairly
> cumbersome. For prefetching distance, unrolling threshold, etc. we
> could move the default value into NOTTI, make the default value of
> the command line parameter 0 (or -1 or something else invalid), and
> then use the TTI value only if the command-line value is invalid.
> But how will this work for boolean parameters? Maybe we could make
> it possible to determine if a command-line parameter was actually
> provided or if it has its default value.

Alternatively, we could provide the ability for a target to change the default value of an option (by string, so there are no linking problems).

 -Hal

> 
> > We don't want to override every property of the target
> > machine using command line flags.
> > 
> 
> Agreed.
> 
> Thanks again,
> Hal
> _______________________________________________
> llvm-commits mailing list
> llvm-commits at cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
> 



More information about the llvm-commits mailing list