[PATCH] IRBuilder: Allow retrieval of the inserted instructions

[Hal Finkel]

----- Original Message -----

> From: "Adam Nemet" <anemet at apple.com>
> To: "Hal Finkel" <hfinkel at anl.gov>
> Cc: reviews+D7421+public+86df0334cf13545b at reviews.llvm.org,
> "llvm-commits" <llvm-commits at cs.uiuc.edu>, "Nadav Rotem
> <nrotem at apple.com> (nrotem at apple.com)" <nrotem at apple.com>, "Arnold
> Schwaighofer" <aschwaighofer at apple.com>, chandlerc at gmail.com
> Sent: Wednesday, February 4, 2015 10:04:54 PM
> Subject: Re: [PATCH] IRBuilder: Allow retrieval of the inserted
> instructions

> > On Feb 4, 2015, at 6:14 PM, Hal Finkel < hfinkel at anl.gov > wrote:
> 

> > ----- Original Message -----
> 

> > > From: "Adam Nemet" < anemet at apple.com >
> > 
> 
> > > To: anemet at apple.com , hfinkel at anl.gov , nrotem at apple.com ,
> > > aschwaighofer at apple.com , chandlerc at gmail.com
> > 
> 
> > > Cc: llvm-commits at cs.uiuc.edu
> > 
> 
> > > Sent: Wednesday, February 4, 2015 5:58:45 PM
> > 
> 
> > > Subject: [PATCH] IRBuilder: Allow retrieval of the inserted
> > > instructions
> > 
> 

> > > Hi hfinkel, nadav, aschwaighofer, chandlerc,
> > 
> 

> > > This is an RFC.
> > 
> 

> > > When the Loop Vectorizer builds the instruction sequence for
> > > checks
> > 
> 
> > > it tries
> > 
> 
> > > to determine the first instruction that was emitted in the
> > > current
> > 
> 
> > > block.
> > 
> 
> > > This is then used to split the block.
> > 
> 

> > Silly question: Why don't we split the block first?
> 

> Because we may end up not needing any checks in which case we don’t
> split. The decision of whether checks are necessary is local to both
> addRuntimeCheck and addStrideCheck.
Okay, but an alternative design would be to always split early, and then let SimplifyCFG put them back together again should we not actually end up with any checks, right? I'm not necessarily advocating for this approach, and think that the marker code will be cleaner than the existing getFirstInst code, but splitting the block eagerly will also create a trivial way of marking that point, and I want to understand if there are any downsides to that approach before we invent something new. 

Thanks again, 
Hal 

> Adam

> > -Hal
> 

> > > It uses a custom solution for this implemented in the static
> > > function
> > 
> 
> > > getFirstInst. The pattern is something like:
> > 
> 

> > > Value *V = IRBuilder.CreateBlah(...);
> > 
> 
> > > FirstInst = getFirstInst(FirstInst, V, Loc)
> > 
> 
> > > Value *V2 = IRBuilder.CreateBlah(...);
> > 
> 
> > > FirstInst = getFirstInst(FirstInst, V2, Loc);
> > 
> 

> > > (Since CreateBlah may return a constant we may not generate the
> > > first
> > 
> 
> > > instruction for V.)
> > 
> 

-- 

Hal Finkel 
Assistant Computational Scientist 
Leadership Computing Facility 
Argonne National Laboratory 
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20150205/8b7aa769/attachment.html>