[llvm] r268452 - PM: Port LoopRotation to the new loop pass manager

[Justin Bogner via llvm-commits]

Sean Silva <chisophugis at gmail.com> writes:
> On Tue, May 3, 2016 at 3:02 PM, Justin Bogner via llvm-commits <
> llvm-commits at lists.llvm.org> wrote:
>
>> Author: bogner
>> Date: Tue May  3 17:02:31 2016
>> New Revision: 268452
>>
>> URL: http://llvm.org/viewvc/llvm-project?rev=268452&view=rev
>> Log:
>> PM: Port LoopRotation to the new loop pass manager
 ...
>> --- llvm/trunk/test/Transforms/LoopRotate/basic.ll (original)
>> +++ llvm/trunk/test/Transforms/LoopRotate/basic.ll Tue May  3 17:02:31 2016
>> @@ -1,4 +1,6 @@
>>  ; RUN: opt -S -loop-rotate < %s | FileCheck %s
>> +; RUN: opt -S
>> -passes='require<loops>,require<targetir>,require<assumptions>,loop(rotate)'
>> < %s | FileCheck %s
>>
>
> Sorry if this is a stupid question, but why do we need to explicitly
> "require" the passes when the loop-rotate already declares the dependency?
> (I feel like there's some part of the bigger picture that I'm missing here)

To be clear, there are no dependencies in the new pass manager - passes
request analysis results from an analysis manager and they're calculated
or returned from the cache as appropriate. Naturally, the next question
one would ask is "okay then, why aren't these calculated on demand when
loop-rotate asks for them?".

The mechanical answer is that loop-rotate uses the `getCachedResult`
API, not the `getResult` one - the difference being that
`getCachedResult` only returns results if they're already calculated,
and null otherwise. Hence, without the require<> in the test, there are
no results. This of course leads to "why does loop-rotate use
getCachedResult?"

It has to. These analyses are function analyses, and a loop pass isn't
allowed to cause a function analysis to run, much like a function pass
isn't allowed to cause a module analysis to be run. They have to stick
to their own level. This enforces correct layering and acts as a
safeguard against accidentally doing extra work.

That said, the "require<loops>" is redundant, since a LoopPassManager
can't operate without having calculated that. There may also be an
argument to be made that the LPM should implicitly calculate some other
analyses, but we'd have to be careful not to do too much.