[llvm] r231186 - RegisterCoalescer: Gracefully continue if subrange merging fails.

[Hal Finkel]

----- Original Message -----
> From: "Chandler Carruth" <chandlerc at google.com>
> To: "Quentin Colombet" <qcolombet at apple.com>
> Cc: "Commit Messages and Patches for LLVM" <llvm-commits at cs.uiuc.edu>, "Matthias Braun" <matze at braunis.de>
> Sent: Tuesday, March 3, 2015 7:54:30 PM
> Subject: Re: [llvm] r231186 - RegisterCoalescer: Gracefully continue if	subrange merging fails.
> 
> 
> 
> Thanks both Quentin and Matthias.
> 
> 
> Two completely meta points:
> 
> 
> 1) It'd be useful to mention this in a commit that obviously has no
> test case. Saves a round trip checking whether it just was
> accidentally left off the commit.
> 
> 
> 2) This makes me want to encourage thinking about ways to unittest
> components of CodeGen in the future. Historically we haven't looked
> at that, but if we need to support and debug out-of-tree targets
> with interesting properties like this it may became important to
> grow that ability. Its an important use case for these libraries to
> work even with wildly different targets, and important that folks
> can check that they're handling edge cases without access to all the
> weird targets in the world. =]
> 

I'd like to encourage this independent of the use case of out-of-tree targets. We have a serious problem with testing of CodeGen components, generally; we currently construct IR-level tests, but this is very fragile. Many of the IR-level CodeGen tests, especially "bug-triggering" regression tests, don't currently test the logic they were originally designed to cover.

 -Hal

> 
> Anyways, thanks for explaining.
> 
> 
> On Tue, Mar 3, 2015 at 5:29 PM, Quentin Colombet <
> qcolombet at apple.com > wrote:
> 
> 
> 
> With the in-tree target this not exposed.
> 
> 
> Arguably we could create a fake target with such problem, but I do
> not think this is worth the complexity.
> 
> 
> Q.
> 
> 
> 
> 
> 
> On Mar 3, 2015, at 5:21 PM, Chandler Carruth < chandlerc at google.com >
> wrote:
> 
> 
> 
> 
> 
> 
> 
> On Tue, Mar 3, 2015 at 4:43 PM, Matthias Braun < matze at braunis.de >
> wrote:
> 
> 
> 
> Author: matze
> Date: Tue Mar 3 18:43:50 2015
> New Revision: 231186
> 
> URL: http://llvm.org/viewvc/llvm-project?rev=231186&view=rev
> Log:
> RegisterCoalescer: Gracefully continue if subrange merging fails.
> 
> There is a known bug where the register coalescer fails to merge
> subranges when multiple ranges end up in the "overflow" bit 32 of the
> lanemasks. A proper fix for this is complicated so for now this is a
> workaround which lets the register coalescer drop the subregister
> liveness information (we just loose some precision by that) and
> continue.
> 
> Modified:
> llvm/trunk/lib/CodeGen/RegisterCoalescer.cpp
> No test case? This seems like something really important to cover...
> _______________________________________________
> llvm-commits mailing list
> llvm-commits at cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
> 
> 
> _______________________________________________
> 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