[llvm-dev] Loop Invariants Detection questions

[Friedman, Eli via llvm-dev]

On 1/18/2017 2:56 AM, Thomas RUBIANO wrote:
> Ty Eli for your answer.
>
> On Tue, Jan 17, 2017 at 8:11 PM, Friedman, Eli 
> <efriedma at codeaurora.org <mailto:efriedma at codeaurora.org>> wrote:
>
>     On 1/17/2017 7:12 AM, Thomas RUBIANO via llvm-dev wrote:
>
>         Hi all!
>
>         I'm new here, and would like to implement my own Loop
>         Invariant Detection adding some more information on
>         Quasi-Invariants.
>
>         First, is there anything about Quasi-Invariants detection in
>         LLVM I would missed?
>
>         I've seen LICM using LoopInfo::isLoopInvariant for finding
>         invariants.
>         It seems that this method considers a Value invariant if:
>         - it's an Instruction not presents in the current loop (what
>         does it mean? There is no dependence analysis on In and Out
>         "variables" of all instructions in the loop?)
>
>
>     isLoopInvariant just checks whether the definition of a value is
>     an instruction inside the loop.
>
>
> Ok, the term "definition" makes it clear. Do, if here it's:
> %1 = load i32, i32* %fact, align 4
> %mul = mul nsw i32 %1, %2
>
>
> load i32, i32* %fact, align 4 is the def of %1 and it's inside the 
> loop then it's not invariant…

Exactly.  If LICM can prove the load is in fact invariant, it will move 
it out of the loop.

>
>
>         - this Value is not an Instruction (then a Constant I guess…).
>
>
>     Or an function argument, or a few other obscure things which don't
>     really matter in this context.
>
>         I've seen LoopAccessAnalysis using it too. What does this
>         analysis do exactly on loop invariant address?
>
>
>     ScalarEvolution::isLoopInvariant works on SCEV expressions instead
>     of Values, but it's essentially the same thing.
>
>
> What can offer this SCEV expression more than Values?

See the comment at the beginning of lib/Analysis/ScalarEvolution.cpp for 
a brief description and some references.

>
>
>         Also DependenceAnalysis seems to give dependence information
>         on memory refs. But it seems to not be used by LICM…
>
>
>     Yes, the current LICM uses alias analysis in a more direct manner
>     (look for AliasSetTracker).
>
>
>         Also MemoryDependenceAnalysis "determines, for a given memory
>         operation, what preceding memory operations it depends on".
>
>         My question is: Where is really done this dependence analysis.
>         The one which figures out which Instructions depends on others?
>
>         Simply if I have:
>         %0 = load i32, i32* %coucou, align 4
>         %1 = load i32, i32* %fact, align 4
>         %2 = load i32, i32* %i, align 4
>         %mul = mul nsw i32 %1, %2
>
>         mul instruction will depends on the two precedents loads
>         because it uses their results %1 and %2 but not the first one.
>
>         I guess this is done somewhere, and there is a special
>         representation of this information but I didn't find, I'm a
>         bit lost ^^'
>
>
>     If you call "operands()" on the Instruction, it will return %1 and
>     %2.  Please keep in mind that LLVM IR is SSA.
>
>
> Yes but if I understand well, the AliasSetTracker can tell me which 
> load it corresponds to…
> How can I use this AliasSetTracker to disambiguate the registers?
> Here it's just having the correspondence %1 → %fact and %2 → %i
>
> It would be just perfect for me to have the "real" In and Out of each 
> Instruction.
> Maybe I should work on another level or with another object 
> representation?
>
> Ty again :)

Can you show me the complete function you're looking at?  Have you run 
mem2reg on your IR?

-Eli

-- 
Employee of Qualcomm Innovation Center, Inc.
Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum, a Linux Foundation Collaborative Project

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20170118/e0e4d456/attachment.html>