[llvm-commits] [llvm] r106953 - /llvm/trunk/test/Transforms/PartialSpecialize/two-specializations.ll

[Kenneth Uildriks]

On Sat, Jun 26, 2010 at 2:44 PM, Benjamin Kramer
<benny.kra at googlemail.com> wrote:
>
> On 26.06.2010, at 20:47, Kenneth Uildriks wrote:
>
> Hi Kenneth,
>
>
> this wont work, the second opt call doesn't read from stdin. Furthermore the test
> still passes without even applying partialspecialization, rendering it useless.

Sorry about the mixup.  My notion was to stop the ARM buildbot failure
as quickly as possible; also, there is another test in the folder that
checks for the specialization being applied, whereas the last version
I checked in simply tests that it is not being wrongly applied.
>
> IMHO the partial specialization pass should be fixed to produce deterministic
> output across all platforms. It probably depends on the order of pointers or
> the elements in a DenseMap somewhere.
>
Possibly, but I haven't been able to find it yet.  As far as I can
tell, the order of specialization depends on the order of the Module's
function iterator.

My ultimate goal is to have the specialization names depend on the
specialized constant values somehow, so that redundant specializations
across modules can be merged at link time.  That would also remove any
non-deterministic specialization names, and would probably be the
easiest way to do so.  So since I have to make it deterministic
anyway, I might as well press forward with that.