[llvm] r216678 - Do not introduce new shuffle patterns after operation legalization if SHUFFLE_VECTOR

[Owen Anderson]

> On Aug 28, 2014, at 11:08 AM, Chandler Carruth <chandlerc at google.com> wrote:
> 
> 
> On Thu, Aug 28, 2014 at 11:04 AM, Chandler Carruth <chandlerc at google.com <mailto:chandlerc at google.com>> wrote:
> On Thu, Aug 28, 2014 at 10:49 AM, Owen Anderson <resistor at mac.com <mailto:resistor at mac.com>> wrote:
> Do not introduce new shuffle patterns after operation legalization if SHUFFLE_VECTOR
> was marked custom.  The target independent DAG combine has no way to know if
> the shuffles it is introducing are ones that the target could support or not.
> 
> Test case? This seems somewhat arbitrary otherwise...
> 
> (To be completely clear, in case this email wasn't, I understand that currently not all backends support a custom lowering of all shuffles.... but the x86 backend actually does support it (for legal types) and so it seems quite likely that someone could accidentally change this back unless we have more checking in the backend to back up the assertion that we can't form an arbitrary shuffle here. Essentially, I'm fine with the patch, but I'm worried that I or someone else could easily re-break this in the future unless we have some test coverage.)

The only testcase I have for this requires a not-in-tree target.

—Owen
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20140828/00e7688d/attachment.html>