> Xinliang David Li <davidxl at google.com> writes: >>> From: <betulb at codeaurora.org> >>> Date: Tue, Apr 7, 2015 at 12:44 PM >>> Subject: [LLVMdev] IC profiling infrastructure >>> To: llvmdev at cs.uiuc.edu >>> >>> >>> >>> Hi All, >>> >>> We had sent out an RFC in October on indirect call target profiling. >>> The >>> proposal was about profiling target addresses seen at indirect call >>> sites. >>> Using the profile data we're seeing up to %8 performance improvements >>> on >>> individual spec benchmarks where indirect call sites are present. We've >>> already started uploading our patches to the phabricator. I'm looking >>> forward to your reviews and comments on the code and ready to respond >>> to >>> your design related queries. >>> >>> There were few questions posted on the RFC that were not responded. >>> Here >>> are the much delayed comments. >>> >> >> Hi Betul, thank you for your patience. I have completed initial >> comparison with a few alternative value profile designs. My conclusion >> is that your proposed approach should well in practice. The study can >> be found here: >> https://docs.google.com/document/u/1/d/1k-_k_DLFBh8h3XMnPAi6za-XpmjOIPHX_x6UB6PULfw/pub > > Thanks for looking at this David. > > Betul: I also have some thoughts on the approach and implementation of > this, but haven't had a chance to go over it in detail. I hope to have > some feedback for you on all of this sometime next week, and I'll start > reviewing the individual patches after that.Hi All, I've posted three more patches yesterday. They might be missing some cosmetic fixes, but the support for profiling multiple value kinds have been added to the readers, writers and runtime. I'd appreciate your comments on the CL's. Thanks, -Betul> >>> 1) Added dependencies: Our implementation adds dependency on >>> calloc/free >>> as weâre generating/maintaining a linked list at run time. >> >> If it becomes a problem for some, there is a way to handle that -- but >> at a cost of more memory required (to be conservative). One of the >> good feature of using dynamic memory is that it allows counter array >> allocation on the fly which eliminates the need to allocate memory for >> lots of cold/unexecuted functions. >> >>> We also added >>> dependency on the usage of mutexes to prevent memory leaks in the case >>> multiple threads trying to insert a new target address for the same IC >>> site into the linked list. To least impact the performance we only >>> added >>> mutexes around the pointer assignment and kept any dynamic memory >>> allocation/free operations outside of the mutexed code. >> >> This (using mutexes) should be and can be avoided -- see the above >> report. >> >>> >>> 2) Indirect call data being present in sampling profile output: This is >>> unfortunately not helping in our case due to perf depending on lbr >>> support. To our knowledge lbr support is not present on ARM platforms. >>> >> >> yes. >> >>> 3) Losing profiling support on targets not supporting malloc/mutexes: >>> The >>> added dependency on calloc/free/mutexes may perhaps be eliminated >>> (although our current solution does not handle this) through having a >>> separate run time library for value profiling purposes. Instrumentation >>> can link in two run time libraries when value profiling (an instance of >>> it >>> being indirect call target profiling) is enabled on the command line. >> >> See above. >> >>> >>> 4) Performance of the instrumented code: Instrumentation with IC >>> profiling >>> patches resulted in 7% degradation across spec benchmarks at -O2. For >>> the >>> benchmarks that did not have any IC sites, no performance degradation >>> was >>> observed. This data is gathered using the ref data set for spec. >>> >> >> I'd like to make the runtime part of the change to be shared and used >> as a general purpose value profiler (not just indirect call >> promotion), but this can be done as a follow up. >> >> I will start with some reviews. Hopefully others will help with reviews >> too. >> >> thanks, >> >> David >> >> >> >>> Thanks, >>> -Betul Buyukkurt >>> >>> Qualcomm Innovation Center, Inc. >>> Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum, a >>> Linux >>> Foundation Collaborative Project >>> >>> >>> >>> _______________________________________________ >>> LLVM Developers mailing list >>> LLVMdev at cs.uiuc.edu http://llvm.cs.uiuc.edu >>> http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev >>> >> >> _______________________________________________ >> LLVM Developers mailing list >> LLVMdev at cs.uiuc.edu http://llvm.cs.uiuc.edu >> http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev >
I have sent my review comments. I think most of my high level concerns have been addressed (except for last few minor fix ups). Justin, do you have a chance to take a look? thanks, David On Wed, May 13, 2015 at 10:49 AM, Betul Buyukkurt <betulb at codeaurora.org> wrote:>> Xinliang David Li <davidxl at google.com> writes: >>>> From: <betulb at codeaurora.org> >>>> Date: Tue, Apr 7, 2015 at 12:44 PM >>>> Subject: [LLVMdev] IC profiling infrastructure >>>> To: llvmdev at cs.uiuc.edu >>>> >>>> >>>> >>>> Hi All, >>>> >>>> We had sent out an RFC in October on indirect call target profiling. >>>> The >>>> proposal was about profiling target addresses seen at indirect call >>>> sites. >>>> Using the profile data we're seeing up to %8 performance improvements >>>> on >>>> individual spec benchmarks where indirect call sites are present. We've >>>> already started uploading our patches to the phabricator. I'm looking >>>> forward to your reviews and comments on the code and ready to respond >>>> to >>>> your design related queries. >>>> >>>> There were few questions posted on the RFC that were not responded. >>>> Here >>>> are the much delayed comments. >>>> >>> >>> Hi Betul, thank you for your patience. I have completed initial >>> comparison with a few alternative value profile designs. My conclusion >>> is that your proposed approach should well in practice. The study can >>> be found here: >>> https://docs.google.com/document/u/1/d/1k-_k_DLFBh8h3XMnPAi6za-XpmjOIPHX_x6UB6PULfw/pub >> >> Thanks for looking at this David. >> >> Betul: I also have some thoughts on the approach and implementation of >> this, but haven't had a chance to go over it in detail. I hope to have >> some feedback for you on all of this sometime next week, and I'll start >> reviewing the individual patches after that. > > Hi All, > > I've posted three more patches yesterday. They might be missing some > cosmetic fixes, but the support for profiling multiple value kinds have > been added to the readers, writers and runtime. I'd appreciate your > comments on the CL's. > > Thanks, > -Betul > >> >>>> 1) Added dependencies: Our implementation adds dependency on >>>> calloc/free >>>> as we’re generating/maintaining a linked list at run time. >>> >>> If it becomes a problem for some, there is a way to handle that -- but >>> at a cost of more memory required (to be conservative). One of the >>> good feature of using dynamic memory is that it allows counter array >>> allocation on the fly which eliminates the need to allocate memory for >>> lots of cold/unexecuted functions. >>> >>>> We also added >>>> dependency on the usage of mutexes to prevent memory leaks in the case >>>> multiple threads trying to insert a new target address for the same IC >>>> site into the linked list. To least impact the performance we only >>>> added >>>> mutexes around the pointer assignment and kept any dynamic memory >>>> allocation/free operations outside of the mutexed code. >>> >>> This (using mutexes) should be and can be avoided -- see the above >>> report. >>> >>>> >>>> 2) Indirect call data being present in sampling profile output: This is >>>> unfortunately not helping in our case due to perf depending on lbr >>>> support. To our knowledge lbr support is not present on ARM platforms. >>>> >>> >>> yes. >>> >>>> 3) Losing profiling support on targets not supporting malloc/mutexes: >>>> The >>>> added dependency on calloc/free/mutexes may perhaps be eliminated >>>> (although our current solution does not handle this) through having a >>>> separate run time library for value profiling purposes. Instrumentation >>>> can link in two run time libraries when value profiling (an instance of >>>> it >>>> being indirect call target profiling) is enabled on the command line. >>> >>> See above. >>> >>>> >>>> 4) Performance of the instrumented code: Instrumentation with IC >>>> profiling >>>> patches resulted in 7% degradation across spec benchmarks at -O2. For >>>> the >>>> benchmarks that did not have any IC sites, no performance degradation >>>> was >>>> observed. This data is gathered using the ref data set for spec. >>>> >>> >>> I'd like to make the runtime part of the change to be shared and used >>> as a general purpose value profiler (not just indirect call >>> promotion), but this can be done as a follow up. >>> >>> I will start with some reviews. Hopefully others will help with reviews >>> too. >>> >>> thanks, >>> >>> David >>> >>> >>> >>>> Thanks, >>>> -Betul Buyukkurt >>>> >>>> Qualcomm Innovation Center, Inc. >>>> Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum, a >>>> Linux >>>> Foundation Collaborative Project >>>> >>>> >>>> >>>> _______________________________________________ >>>> LLVM Developers mailing list >>>> LLVMdev at cs.uiuc.edu http://llvm.cs.uiuc.edu >>>> http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev >>>> >>> >>> _______________________________________________ >>> LLVM Developers mailing list >>> LLVMdev at cs.uiuc.edu http://llvm.cs.uiuc.edu >>> http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev >> > >
Hi Betul, I've finally gotten around to going over this in detail - sorry for the delay, and thanks for working on this. I think that the general approach is a good one, and that this will end up working well. I have a couple of points on a high level, and I'll also send some review for the patches you've sent out. - The raw profile format does not need to be backwards compatible, that's only important for the indexed one. Instead of adding RawInstrValueProfReader, just change RawInstrProfReader and reject the data if the version is 1. Similarly, don't change the raw profile magic - just bump the version to 2. - We don't need to store the value profiling kind in the data at all. The frontend knows which invocations of the intrinsic are for each kind implicitly, much like it knows the difference between a counter for an "if" and a "for" apart implicitly. Just store one set of profiling data and let the frontend sort it out. - Given that, the instrprof libraries needn't know about kinds at all, that can be dealt with locally in clang's CodeGenPGO, where we need to know what to do with the data. I'd just drop the kind completely for now. It'll be easy to add later without affecting much. - We should be able to statically allocate the first entry for each callsite for __llvm_profile_value_data. It will always be there, and for call sites that always call the same target it's nice to avoid the allocation. - The changes in the llvm and compiler-rt repos need their own tests. - Please use clang-format on your changes. Several of the patches have unusual or non-llvm-style whitespace/wrapping/etc. Finally, the LLVM patch is quite large. We can probably make it a bit easier to stage and review by splitting out a couple of things and doing them first: 1. Changing the __llvm_profile_data variables so they're no longer const. 2. Adding the InstrProfRecord type and updating the unittests. Also, this type should just go in the InstrProf.h header, no need for its own. 3. Arguably, we could update the raw format and implement the RawInstrProfReader changes (but ignoring the new data) first. If you think this is worthwhile go ahead, but if it'll lead to too much extra busywork it probably isn't worth it. I'm sending some review on the patches themselves as well, but I expect those to mostly be on the mechanical aspects since the high level points are already written down here. Xinliang David Li <davidxl at google.com> writes:> I have sent my review comments. I think most of my high level concerns > have been addressed (except for last few minor fix ups). > > Justin, do you have a chance to take a look? > > thanks, > > David > > On Wed, May 13, 2015 at 10:49 AM, Betul Buyukkurt <betulb at codeaurora.org> wrote: >>> Xinliang David Li <davidxl at google.com> writes: >>>>> From: <betulb at codeaurora.org> >>>>> Date: Tue, Apr 7, 2015 at 12:44 PM >>>>> Subject: [LLVMdev] IC profiling infrastructure >>>>> To: llvmdev at cs.uiuc.edu >>>>> >>>>> >>>>> >>>>> Hi All, >>>>> >>>>> We had sent out an RFC in October on indirect call target profiling. >>>>> The >>>>> proposal was about profiling target addresses seen at indirect call >>>>> sites. >>>>> Using the profile data we're seeing up to %8 performance improvements >>>>> on >>>>> individual spec benchmarks where indirect call sites are present. We've >>>>> already started uploading our patches to the phabricator. I'm looking >>>>> forward to your reviews and comments on the code and ready to respond >>>>> to >>>>> your design related queries. >>>>> >>>>> There were few questions posted on the RFC that were not responded. >>>>> Here >>>>> are the much delayed comments. >>>>> >>>> >>>> Hi Betul, thank you for your patience. I have completed initial >>>> comparison with a few alternative value profile designs. My conclusion >>>> is that your proposed approach should well in practice. The study can >>>> be found here: >>>> https://docs.google.com/document/u/1/d/1k-_k_DLFBh8h3XMnPAi6za-XpmjOIPHX_x6UB6PULfw/pub >>> >>> Thanks for looking at this David. >>> >>> Betul: I also have some thoughts on the approach and implementation of >>> this, but haven't had a chance to go over it in detail. I hope to have >>> some feedback for you on all of this sometime next week, and I'll start >>> reviewing the individual patches after that. >> >> Hi All, >> >> I've posted three more patches yesterday. They might be missing some >> cosmetic fixes, but the support for profiling multiple value kinds have >> been added to the readers, writers and runtime. I'd appreciate your >> comments on the CL's. >> >> Thanks, >> -Betul >> >>> >>>>> 1) Added dependencies: Our implementation adds dependency on >>>>> calloc/free >>>>> as we’re generating/maintaining a linked list at run time. >>>> >>>> If it becomes a problem for some, there is a way to handle that -- but >>>> at a cost of more memory required (to be conservative). One of the >>>> good feature of using dynamic memory is that it allows counter array >>>> allocation on the fly which eliminates the need to allocate memory for >>>> lots of cold/unexecuted functions. >>>> >>>>> We also added >>>>> dependency on the usage of mutexes to prevent memory leaks in the case >>>>> multiple threads trying to insert a new target address for the same IC >>>>> site into the linked list. To least impact the performance we only >>>>> added >>>>> mutexes around the pointer assignment and kept any dynamic memory >>>>> allocation/free operations outside of the mutexed code. >>>> >>>> This (using mutexes) should be and can be avoided -- see the above >>>> report. >>>> >>>>> >>>>> 2) Indirect call data being present in sampling profile output: This is >>>>> unfortunately not helping in our case due to perf depending on lbr >>>>> support. To our knowledge lbr support is not present on ARM platforms. >>>>> >>>> >>>> yes. >>>> >>>>> 3) Losing profiling support on targets not supporting malloc/mutexes: >>>>> The >>>>> added dependency on calloc/free/mutexes may perhaps be eliminated >>>>> (although our current solution does not handle this) through having a >>>>> separate run time library for value profiling purposes. Instrumentation >>>>> can link in two run time libraries when value profiling (an instance of >>>>> it >>>>> being indirect call target profiling) is enabled on the command line. >>>> >>>> See above. >>>> >>>>> >>>>> 4) Performance of the instrumented code: Instrumentation with IC >>>>> profiling >>>>> patches resulted in 7% degradation across spec benchmarks at -O2. For >>>>> the >>>>> benchmarks that did not have any IC sites, no performance degradation >>>>> was >>>>> observed. This data is gathered using the ref data set for spec. >>>>> >>>> >>>> I'd like to make the runtime part of the change to be shared and used >>>> as a general purpose value profiler (not just indirect call >>>> promotion), but this can be done as a follow up. >>>> >>>> I will start with some reviews. Hopefully others will help with reviews >>>> too. >>>> >>>> thanks, >>>> >>>> David >>>> >>>> >>>> >>>>> Thanks, >>>>> -Betul Buyukkurt >>>>> >>>>> Qualcomm Innovation Center, Inc. >>>>> Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum, a >>>>> Linux >>>>> Foundation Collaborative Project >>>>> >>>>> >>>>> >>>>> _______________________________________________ >>>>> LLVM Developers mailing list >>>>> LLVMdev at cs.uiuc.edu http://llvm.cs.uiuc.edu >>>>> http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev >>>>> >>>> >>>> _______________________________________________ >>>> LLVM Developers mailing list >>>> LLVMdev at cs.uiuc.edu http://llvm.cs.uiuc.edu >>>> http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev >>> >> >>