On Sun, Nov 28, 2010 at 2:37 PM, Chris Lattner <clattner at apple.com> wrote:> On Nov 23, 2010, at 5:07 PM, Xinliang David Li wrote: > > Hi, I browsed the LLVM inliner implementation, and it seems there is room > for improvement. (I have not read it too carefully, so correct me if what I > observed is wrong). > > > > First the good side of the inliner -- the function level summary and > inline cost estimation is more elaborate and complete than gcc. For > instance, it considers callsite arguments and the effects of optimization > enabled by inlining. > > Yep, as others pointed out, this is intended to interact closely with the > per-function optimizations that get mixed in due to the inliner being a > callgraphscc pass. This is actually a really important property of the > inliner. If you have a function foo that calls a leaf function bar, the > sequence of optimization is: > > 1. Run the inliner on bar (noop, since it has no call sites) > 2. Run the per-function passes on bar. This generally shrinks it, and > prevents "abstraction penalty" from making bar look too big to inline. > 3. Run the inliner on foo. Since foo calls bar, we consider inlining bar > into foo and do so if profitable. > 4. Run the per-function passes on foo. If bar got inlined, this means that > we're running the per-function passes over the inlined contents of bar > again. >On-the-fly clean up (optimization) while doing bottom up inlining is nice as you described. Many other compilers chose not to do this way due to scalability concerns (with IPO) -- this can make the IPO the biggest bottom neck in terms of compile time (as it is serialized). Memory many not be a big issue for LLVM as I can see the good locality in pass manager. (Just curious, what is biggest application LLVM can build with IPO?)> > In a traditional optimizer like GCC's, you end up with problems where you > have to set a high inline threshold due to inlining-before-optimizing > causing "abstraction penalty problems". An early inliner is a hack that > tries to address this.It is a hack in some sense (but a common practice) -- but enables other flexibilities.> Another problem with this approach from the compile time perspective is > that you end up repeating work multiple times. For example, if there is a > common subexpression in a small function, you end up inlining it into many > places, then having to eliminate the common subexpression in each copy. >Early inlining + scalar opt can do the same, right?> > The LLVM inliner avoids these problems, but (as you point out) this really > does force it to being a bottom-up inliner. This means that the bottom-up > inliner needs to make decisions in strange ways in some cases: for example > if qux called foo, and foo were static, then (when processing foo) we may > decide not to inline bar into foo because it would be more profitable to > inline foo into qux. > > > Now more to the weakness of the inliner: > > > > 1) It is bottom up. The inlining is not done in the order based on the > priority of the callsites. It may leave important callsites (at top of the > cg) unlined due to higher cost after inline cost update. It also eliminates > the possibility of inline specialization. To change this, the inliner pass > may not use the pass manager infrastructure . (I noticed a hack in the > inliner to workaround the problem -- for static functions avoid inlining its > callees if it causes it to become too big ..) > > This is true, but I don't think it's a really large problem in practice. > We don't have a "global inline threshold limit" (which I've never > understood, except as a hack to prevent run-away inlining) so not visiting > in priority order shouldn't prevent high-priority-but-processed-late > candidates from being inlined. >global threshold can be used to control the unnecessary size growth. In some cases, the size increase may also cause increase in icache footprint leading to poor performance. In fact, with IPO/CMO, icache footprint can be modeled in some way and be used as one kind of global limit.> > The only potential issue I'm aware of is if we have A->B->C and we decide > to inline C into B when it would be more profitable to inline B into A and > leave C out of line. This can be handled with a heuristic like the one > above. > > > 2) There seems to be only one inliner pass. For calls to small > functions, it is better to perform early inlining as one of the local (per > function) optimizations followed by scalar opt clean up. This will sharpen > the summary information. (Note the inline summary update does not consider > the possible cleanup) > > Yep. This is a feature :) > > > 3) recursive inlining is not supported > > This is a policy decision. It's not clear whether it is really a good > idea, though I have seen some bugzilla or something about it. I agree that > it should be revisited. > > > 4) function with indirect branch is not inlined. What source construct > does indirect branch instr correspond to ? variable jump? > > See: > http://blog.llvm.org/2010/01/address-of-label-and-indirect-branches.html > > for more details. > > > 6) There is one heuristc used in inline-cost computation seems wrong: > > > > // Calls usually take a long time, so they make the inlining gain > smaller. > > InlineCost += CalleeFI->Metrics.NumCalls * > InlineConstants::CallPenalty; > > > > Does it try to block inlining of callees with lots of calls? Note > inlining such a function only increase static call counts. > > I think that this is a heuristic that Jakob came up with, but I think it's > a good one, also discussed elsewhere on the thread. > > When talking about inlining and tuning thresholds and heuristics, it is a > good idea to quantify what the expected or possible wins of inlining a > function are. Some of the ones I'm aware of: > > 1. In some cases, inlining shrinks code. > > 2. Inlining a function exposes optimization opportunities on the inlined > code, because constant propagation and other simplifications can take place. > > 3. Inlining a function exposes optimizations in the caller because > address-taken values can be promoted to registers. > > 4. Inlining a function can improve optimization in a caller because > interprocedural side-effect analysis isn't needed. For example, load/call > dependence may not be precise. This is something we should continue to > improve in the optimizer though. > > 5. Inlining code with indirect call sites and switches can improve branch > prediction if some callers of the function are biased differently than other > callers. This is pretty hard to predict without profile info though. > >Besides -- 1) reducing call overhead; 2) scheduling freedom; 3) enabling optimizations across inline instances of callee(s); 4) sharpening local analysis (mainly aliasing) results -- such as points to, malloc etc. It may also lose aliasing assertion (such as restrict aliasing) if not done properly.> > The "punish functions containing lots of calls" is based on the assumption > that functions which are mostly calls (again, this decision happens after > the callee has been inlined and simplified) aren't themselves doing much > work. >My point is that using static count of callsites as a indicator for this can be misleading. All the calls may be calls to cold external functions for instance. Thanks, David> > -Chris > >-------------- next part -------------- An HTML attachment was scrubbed... URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20101128/0dff7413/attachment.html>
> On-the-fly clean up (optimization) while doing bottom up inlining is nice as > you described. Many other compilers chose not to do this way due to > scalability concerns (with IPO) -- this can make the IPO the biggest bottom > neck in terms of compile time (as it is serialized). Memory many not be a > big issue for LLVM as I can see the good locality in pass manager. (Just > curious, what is biggest application LLVM can build with IPO?)I am not sure what is the biggest one, but the one I tried was clang itself with LTO: http://lists.cs.uiuc.edu/pipermail/llvmdev/2010-October/035584.html Cheers, Rafael
On Nov 28, 2010, at 11:39 PM, Xinliang David Li wrote:> 1. Run the inliner on bar (noop, since it has no call sites) > 2. Run the per-function passes on bar. This generally shrinks it, and prevents "abstraction penalty" from making bar look too big to inline. > 3. Run the inliner on foo. Since foo calls bar, we consider inlining bar into foo and do so if profitable. > 4. Run the per-function passes on foo. If bar got inlined, this means that we're running the per-function passes over the inlined contents of bar again. > > On-the-fly clean up (optimization) while doing bottom up inlining is nice as you described. Many other compilers chose not to do this way due to scalability concerns (with IPO) -- this can make the IPO the biggest bottom neck in terms of compile time (as it is serialized). Memory many not be a big issue for LLVM as I can see the good locality in pass manager. (Just curious, what is biggest application LLVM can build with IPO?)I don't really know, and I agree with you that LLVM's LTO isn't very scalable (it currently loads all the IR into memory). I haven't thought a lot about this, but I'd tackle that problem in three stages: 1. Our LTO model runs optimizations at both compile and link time, the compile-time optimizations should work as they do now IMO. This is controversial though, because doing so could cause (e.g.) an inlining to happen "early" that would be seen as a bad idea with full LTO information. The advantage of doing compile-time optimizations is that it both shrinks the IR, and speeds up an incremental rebuild by avoiding having to do simple optimizations again. 2. At LTO time, the bottom-up processing of the callgraph is still goodness and presents good locality (unless you have very very large SCC's). The tweak that we'd have to implement is lazy deserialization (already implemented) and reserialization to disk (which is missing). With this, you get much better memory footprint than "hold everything in memory at once". 3. To support multiple cores/machines, you break the callgraph SCC DAG into parallel chunks that can be farmed out. There is a lot of parallelism in a DAG. I don't know of anyone planning on working on LTO at the moment though.> In a traditional optimizer like GCC's, you end up with problems where you have to set a high inline threshold due to inlining-before-optimizing causing "abstraction penalty problems". An early inliner is a hack that tries to address this. > > It is a hack in some sense (but a common practice) -- but enables other flexibilities.The hack I'm referring to is the "raise the inline threshold". If the inliner has any language specificity to its inline threshold, I consider it a hack. There is no reason the inliner should have to know if it's building C or C++ code. It should be guided based on the structure of the code.> Another problem with this approach from the compile time perspective is that you end up repeating work multiple times. For example, if there is a common subexpression in a small function, you end up inlining it into many places, then having to eliminate the common subexpression in each copy. > > Early inlining + scalar opt can do the same, right?In some cases, but not in general, because you run into phase ordering problems.> This is true, but I don't think it's a really large problem in practice. We don't have a "global inline threshold limit" (which I've never understood, except as a hack to prevent run-away inlining) so not visiting in priority order shouldn't prevent high-priority-but-processed-late candidates from being inlined. > > global threshold can be used to control the unnecessary size growth. In some cases, the size increase may also cause increase in icache footprint leading to poor performance. In fact, with IPO/CMO, icache footprint can be modeled in some way and be used as one kind of global limit.I understand that, but that implies that you have some model for code locality. Setting a global code growth limit is (in my opinion) a hack unless you are aiming for the whole program to fit in the icache (which I don't think anyone tries to do :). With any other limit that is higher than your icache size, you are basically picking an *arbitrary* limit that is not based on the machine model or the instruction locality of the program.> The "punish functions containing lots of calls" is based on the assumption that functions which are mostly calls (again, this decision happens after the callee has been inlined and simplified) aren't themselves doing much work. > > My point is that using static count of callsites as a indicator for this can be misleading. All the calls may be calls to cold external functions for instance.Absolutely true. It may also be completely wrong for some functions. It's a heuristic :) -Chris -------------- next part -------------- An HTML attachment was scrubbed... URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20101129/afd85ace/attachment.html>
On Mon, Nov 29, 2010 at 10:56 AM, Chris Lattner <clattner at apple.com> wrote:> On Nov 28, 2010, at 11:39 PM, Xinliang David Li wrote: > > 1. Run the inliner on bar (noop, since it has no call sites) >> 2. Run the per-function passes on bar. This generally shrinks it, and >> prevents "abstraction penalty" from making bar look too big to inline. >> 3. Run the inliner on foo. Since foo calls bar, we consider inlining bar >> into foo and do so if profitable. >> 4. Run the per-function passes on foo. If bar got inlined, this means >> that we're running the per-function passes over the inlined contents of bar >> again. >> > > On-the-fly clean up (optimization) while doing bottom up inlining is nice > as you described. Many other compilers chose not to do this way due to > scalability concerns (with IPO) -- this can make the IPO the biggest bottom > neck in terms of compile time (as it is serialized). Memory many not be a > big issue for LLVM as I can see the good locality in pass manager. (Just > curious, what is biggest application LLVM can build with IPO?) > > > I don't really know, and I agree with you that LLVM's LTO isn't very > scalable (it currently loads all the IR into memory). I haven't thought a > lot about this, but I'd tackle that problem in three stages: > > 1. Our LTO model runs optimizations at both compile and link time, the > compile-time optimizations should work as they do now IMO. This is > controversial though, because doing so could cause (e.g.) an inlining to > happen "early" that would be seen as a bad idea with full LTO information. > The advantage of doing compile-time optimizations is that it both shrinks > the IR, and speeds up an incremental rebuild by avoiding having to do simple > optimizations again. > > 2. At LTO time, the bottom-up processing of the callgraph is still goodness > and presents good locality (unless you have very very large SCC's). The > tweak that we'd have to implement is lazy deserialization (already > implemented) and reserialization to disk (which is missing). With this, you > get much better memory footprint than "hold everything in memory at once". >IR is just one memory consumer. In LTO, there are also global data structures : global symtab, global type table, call graph, points-to graph, mod-ref info etc. In a compiler I worked with before, serialization is done on points-to graph and mod-ref info after the info is mapped from a global view to a per TU local view, and IR for each TU is mapped/unmapped on demand. The type/symbol info per TU is in different segment from the the code segment.> > 3. To support multiple cores/machines, you break the callgraph SCC DAG into > parallel chunks that can be farmed out. There is a lot of parallelism in a > DAG. > >Parallelism distributed across machines can be tricky -- involving lots of overhead such as rpc and data passing. You may also be surprised with side effects due to the parallelism using multi-core -- thrashing due to memory contention -- some per function level pass may use lots of memory for temporary data structure. It won't scale for the compiler workload -- i.e. get 2x speedup using 8 core.> > I understand that, but that implies that you have some model for code > locality. Setting a global code growth limit is (in my opinion) a hack > unless you are aiming for the whole program to fit in the icache (which I > don't think anyone tries to do :). > >Yes, global growth limit may be good for size control, but is a hack for control icache footprint. However, as I mentioned, the bottom up inline scheme make it impossible to use any heuristics involving 'global limit' which can be more complicated and fancier than the simple growth limit. For instance, there is no restriction that only one global limit can be used --- the compiler can partition the call graph into multiple locality regions, and set icache limit for each region. The inlining order can be done on a region by region basis. For each region, the region limit is applied and the priority queue must be used. Thanks, David> With any other limit that is higher than your icache size, you are > basically picking an *arbitrary* limit that is not based on the machine > model or the instruction locality of the program. > > The "punish functions containing lots of calls" is based on the assumption >> that functions which are mostly calls (again, this decision happens after >> the callee has been inlined and simplified) aren't themselves doing much >> work. >> > > My point is that using static count of callsites as a indicator for this > can be misleading. All the calls may be calls to cold external functions for > instance. > > > Absolutely true. It may also be completely wrong for some functions. It's > a heuristic :) > > -Chris > >-------------- next part -------------- An HTML attachment was scrubbed... URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20101130/c90f134a/attachment.html>