[llvm-dev] [RFC] Enhance Partial Inliner by using a general outlining scheme for cold blocks

Xinliang David Li via llvm-dev llvm-dev at lists.llvm.org
Tue Aug 15 14:36:03 PDT 2017


Hi Graham, Making partial inlining more general is something worth doing.
Regarding your implementation plan, I have some suggestions here:

*) Function outlining introduces additional runtime cost: passing of live
in values, returning of live out values (via memory), glue code in the
caller to handle regions without a single exit block etc.  The cost
analysis needs to factor in those carefully
*) Remove the limitation that there is only *one* outlined routine.
Instead, the algorithm can compute multiple single-entry/single exit or
single entry/multiple exit regions (cold ones) in the routine, and outline
each region into its own function. The benefit include
   1) simplify the design and implementation and most of the existing code
can be reused;
   2) provide more flexibility to allow most effective outlining;
   3) reduced runtime overhead of making calls to the outline functions.

thanks,

David

On Tue, Aug 15, 2017 at 11:22 AM, Graham Yiu via llvm-dev <
llvm-dev at lists.llvm.org> wrote:

> Hello,
>
> My team and I are looking to do some enhancements in the partial inliner
> in opt. Would appreciate any feedback that folks might have.
>
> # Partial Inlining in LLVM opt
>
> ## Summary
>
> ### Background
>
> Currently, the partial inliner searches the first few blocks of the callee
> and looks for a branch to the return block (ie. early return). If found, it
> attempts to outline the rest of the slow (or heavy) code so the inliner
> will be able to inline the fast (or light) code. If no early returns are
> found, the partial inliner will give up. As far as I can tell,
> BlockFrequency and BranchProbability information is only used when
> attempting to inline the early return code, and not used to determine
> whether to outline the slow code.
>
> ### Proposed changes
>
> In addition to looking for early returns, we should utilize profile
> information to outline blocks that are considered cold. If we can
> sufficiently reduce the size of the original function via this type of
> outlining, inlining should be able to inline the rest of the hot code.
>
> ## Details
>
> With the presence of profile information, we have a view of what code is
> infrequently executed and make better decisions on what to outline. Early
> return blocks that are infrequently executed should still be included as
> candidates for outlining, but will be treated just like any other cold
> blocks. Without profiling information, however, we should remain
> conservative and only partial inline in the presence of an early return in
> the first few blocks of a function (ie. peel the early return out of the
> function).
>
> To find cold regions to outline, we will traverse the CFG to find edges
> deemed 'cold' and look at the blocks dominated by the successor node. If,
> for some reason, that block has more than one predecessor, then we will
> skip this candidate as there should be a node that dominates this successor
> that has a single entry point. The last node in the dominance vector should
> also have a single successor. If it does not, then further investigation of
> the CFG is necessary to see when/how this situation occurs.
>
> We will need several heuristics to make sure we only outline in cases
> where we are confident it will result in a performance gain. Things such as
> threshold on when a branch is considered cold, the minimum number of times
> the predecessor node has to be executed in order for an edge to be
> considered (confidence factor), and the minimum size of the region to be
> outlined (can use inlining cost analysis like we currently do) will require
> some level of tuning.
>
> Similar to the current implementation, we will attempt to inline the
> leftover (hot) parts of the code, and if for some reason we cannot then we
> discard the modified function and its outlined code.
>
> ### Code changes
>
> The current Partial Inlining code first clones the function of interest
> and looks for a single set of blocks to outline. It then creates a function
> with the set the blocks, and saves the outlined function and outline
> callsite information as part of the function cloning container. In order to
> outline multiple regions of the function, we will need to change these
> containers to keep track of a list of regions to outline. We will also need
> to update the cost analysis to take into account multiple outlined
> functions.
>
> When a ProfileSummary is available, then we should skip the code that
> looks for early returns and go into new code that looks for cold regions to
> outline. When ProfileSummary is not available, then we can fall back to the
> existing code and look for early returns only.
>
> ### Tuning
>
> - The outlining heuristics will need to determine if a set of cold blocks
> is large enough to warrant the overhead of a function call. We also don't
> want the inliner to attempt to inline the outlined code later.
> - The threshold for determining whether a block is cold will also need to
> be tuned. In the case that profiling information is not accurate, we will
> pay the price of the additional call overhead for executing cold code.
> - The confidence factor, which can be viewed as the minimum number of
> times the predecessor has to be executed in order for an edge to be
> considered cold, should also be taken into account to avoid outlining code
> paths we have little information on.
>
> Graham Yiu
> LLVM Compiler Development
> IBM Toronto Software Lab
> Office: (905) 413-4077 C2-407/8200/Markham
> Email: gyiu at ca.ibm.com
>
> _______________________________________________
> LLVM Developers mailing list
> llvm-dev at lists.llvm.org
> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20170815/2bd300a3/attachment.html>


More information about the llvm-dev mailing list