[LLVMdev] Function Pass Manager

[Ivan Llopard]

Hi Jhon,

Finally, I've managed to run my pass as LoopPass avoiding this problem.
Thanks for the detailed explanation, I clearly misunderstood the 
functionality of the pass manager.

Ivan

On 05/06/2012 17:54, John Criswell wrote:
> On 6/5/12 10:39 AM, Ralf Karrenberg wrote:
>> Hi John,
>>
>> On 6/5/12 4:31 PM, John Criswell wrote:
>>> On 4/12/12 3:32 AM, Ivan Llopard wrote:
>>>> Hi again,
>>>>
>>>> I come back to this issue with an example. It's a pass which does
>>>> nothing but throw the 'Unable to schedule' error.
>>>>
>>>> namespace {
>>>> struct MyPass : public FunctionPass {
>>>> static char ID; // Pass identification, replacement for typeid
>>>> MyPass() : FunctionPass(ID) {
>>>> initializeMyPassPass(*PassRegistry::getPassRegistry());
>>>> }
>>>> virtual void getAnalysisUsage(AnalysisUsage&AU) const {
>>>> AU.addRequiredID(LoopSimplifyID);
>>>> AU.addPreservedID(LoopSimplifyID);
>>>> }
>>>> virtual bool runOnFunction(Function&F);
>>>> };
>>>> }
>>>
>>> Is LoopSimplifyID an analysis pass or a transform (optimization) pass?
>>
>> It is a transform pass (it can alter the CFG).
>>
>>> You cannot reliably use the addRequired() method to force a transform
>>> pass to run before another pass; there are cases in which the
>>> PassManager cannot schedule the required passes. You must simply ensure
>>> that LoopSimplify is executed before your pass.
>>
>> I am afraid I do not understand this. What is the purpose of
>> addRequired() if not ensuring that the pass is run before execution of
>> my own pass?
>
> The addRequired method is designed to tell the PassManager which
> *analysis* passes to run before your pass. This allows the PassManager
> to optimize the pass schedule so that analysis passes are only re-run
> when needed. Analysis passes don't modify the IR so PassManager can
> always find a way to schedule the passes.
>
> The reason why you can't require a transform pass is because you can get
> impossible-to-schedule pass schedules (or, at least, schedules that
> PassManager isn't smart enough to schedule because it assumes that all
> required passes are analysis passes) . For example, assume Pass C
> requires Passes A and B. Passes A and B invalidate all other passes
> because they are transform passes. There's no possible schedule for
> these passes.
>
>>
>> I would be perfectly happy with *any* solution that allows me to
>> ensure that LoopSimplify is executed before my FunctionPass.
>
> Sadly, there isn't such a feature. It is the job of the person (or
> program) putting together the pass schedule to ensure that transform
> passes are run in the correct order relative to other transform passes.
> Only analysis passes are scheduled automatically by PassManager.
>
> What you could do to catch bugs is to write an analysis pass that
> verifies that the loops have the structure that you need; this pass
> would assert out if the loops aren't in canonical form. Your pass could
> require this new analysis pass. If someone forgets to run the pass that
> puts loops into canonical form, this analysis pass would catch the error.
>
> -- John T.
>
>>
>> Thanks,
>> Ralf
>