[InstrProfiling] Lightweight Instrumentation

RFC: Lightweight Instrumentation

Hi all,

Our team at Facebook would like to propose a lightweight variant of IR instrumentation PGO for use in the mobile space. IRPGO is a proven technology in LLVM that can boost performance for server workloads. However, the larger binary resulting from instrumentation significantly limits its use for mobile applications. In this proposal, we introduce a few changes to IRPGO to reduce the instrumented binary size, making it suitable for PGO on mobile devices.

This proposal is driven by the same need behind the earlier MIP (machine IR profile) prototype. But unlike MIP where there is significant divergence from IRPGO, this proposed lightweight instrumentation fits into the existing IRPGO framework with a few extensions to achieve a smaller instrumented binary.

We’d like to share the new design and results from our prototype and get feedback.

Best,
Ellis, Kyungwoo, and Wenlei

Motivation

In the mobile space, profile guided optimization can also have an outsized impact on performance just like PGO for server workloads, but conventional instrumentation comes with a large binary size and code size increase as high as 50%, which limits its use for mobile application for two reasons:

  • Mobile applications are very sensitive to total binary size as larger binaries take longer to download and use more space on devices. There could be a hard size limit for over-the-air (OTA) updates for this reason.
  • When code (.text) size increases, it takes longer for applications to start up and could also degrade runtime performance due to more page faults on devices with limited RAM.
    Reducing the size overhead from instrumentation would make IRPGO usable for mobile applications so we could send instrumented binaries through OTA updates in production environments, collect representative production profiles, and apply PGO.

OverviewThe size overhead from IRPGO mainly comes from two things: 1) metadata for mapping raw counts back to IR/CFG, which has to stay with the binary. 2) the increased .text size due to insertion of instrumented code and less effective optimization after instrumentation. Two extensions are proposed to reduce the size overhead from each of the above:

  • We allow the use of debug info / dwarf as alternative metadata for mapping counts to IR, aka profile correlation. Debug info is extractable from the binary, therefore such metadata doesn’t need to be shipped to mobile devices. Debug info has been used extensively for sampling based PGO in LLVM, so it has reasonable quality to support profile correlation.
  • We add the flexibility to allow coarse grained instrumentation that: 1) only insert probes at function entry instead of each block (or blocks decided by MST placement); 2) optional coverage mode using one byte booleans in addition to today’s counting mode using 8 byte counters.
    The extensions offer a spectrum of trade-off choices from the most accurate PGO to something very lightweight that can be used in mobile space. With debug info extracted and using function entry coverage mode, the size increase can be reduced from close to 50% down to below 5% (measured with clang self-build PGO).

Extractable MetadataWith today’s IRPGO, the instrumentation runtime dumps out a profraw profile at the end of training. The runtime creates a header and appends data from the __llvm_prf_data, __llvm_prf_cnts, and __llvm_prf_names sections to create a profraw profile. The __llvm_prf_data section contains references to each function’s profile data (in __llvm_prf_cnts) and name (in __llvm_prf_names) so they are needed to correlate profile data to the functions they instrument.

Some kind of metadata to correlate counts back to IR (specifically CFG blocks) is unavoidable. One way to reduce binary size is to make such metadata extractable so they don’t have to be shipped to mobile devices. We could make __llvm_prf_data and __llvm_prf_names extractable, but the cost will be non-trivial and it will be a breaking change. On the other hand, debug info is extractable from binary and it already does a very good job of maintaining mapping between address and source location / symbols. Sample PGO depends entirely on debug info for profile correlation. So we picked debug info as the alternative for extractable metadata.

In our proposed instrumentation, we create a special global struct, e.g., __profc__Z3foov, to hold counters for a particular function. The __llvm_prf_cnts data section holds all of these structs and serves as placeholder for raw profile counters. In our final instrumented binary, we only have probe instructions and raw profile data without any instrumentation metadata, i.e., there are no __llvm_prf_names or __llvm_prf_data sections but we still have a __llvm_prf_cnts section. At runtime, we dump the __llvm_prf_cnts section to a file without any processing after profiling. To differentiate from IRPGO, the output from runtime is called proflite and we can add another VARIANT_MASK_ flag to the Version field of the profile header. At llvm-profdata post-processing time, we use debug info to correlate our raw profile data as follows. First we identify an instrumented function and look for its special global struct that holds counters (__profc__Z3foov) in the debug info. The debug info can tell us the address of that symbol in the binary and we can compute its offset from the __llvm_prof_cnts section. Then we can use that offset to read the function entry and block counters from the proflite file. Finally we populate profdata output for each function following the existing format.

Value profile is not going to be supported with extractable metadata right now, though we believe it can also be added following a similar scheme.

To improve debug info quality for profile correlation, -fdebug-info-for-profiling from AutoFDO can be used. Additionally, we could also use pseudo-probe from CSSPGO as the alternative metadata which is also fully extractable.

We propose a new flag -fprofile-generate-correlate=[profdata|debug-info|pseudo-probe] to choose what metadata to use for profile correlation. Either we correlate with today’s IRPGO metadata and keep them in their own sections (__llvm_prf_data and __llvm_prf_names), with debug info, or with pseudo-probe.

Coarse-grained InstrumentationIn addition to reducing metadata size ( __llvm_prf_names and __llvm_prf_data), we can also tune down .text size and __llvm_prf_cnts size. We do this by 1) only instrumenting function entries instead of each block and 2) lowering precision by tracking single byte coverage data rather than 8 byte counters. This is a trade-off between profile quality and binary size.

Function profile vs block profile and counting mode vs coverage mode can all be selected independently using our proposed flag -fprofile-generate-mode=[func-cov|block-cov|func-cnt|block-cnt], and they can work with both extractable metadata as well as IRPGO‘s correlation method. func-cov and block-cov use single byte booleans for coverage data while func-cnt and block-cnt use 8 byte counters. block-cnt represents today’s IRPGO which is the default.

When using a profile generated from modes other than block-cnt, additional profile inference is needed before the counts can be consumed by optimizations. Such inference is done during profile loading and so it’s transparent to optimizations.

  • For block coverage mode, we will use coverage info to seed block count inference, and leverage static branch probability at the same time to produce a CFG profile that honors zero count blocks and converts live block coverage data into synthetic counts.
  • For function count mode, we will derived a CFG profile entirely from static branch probability, then scale the CFG profile based on function entry count.
  • Function coverage mode is handled similar to function count mode. For covered/live functions, we will derived a CFG profile entirely from static branch probability first, then scale that CFG profile by a constant.
    Experiments showed that even with coarse-grained function entry profiles, mobile application can still benefit from PGO. But the smaller binary make it possible for mobile to use PGO.

WorkflowSince these are extensions that share the same underlying PGO framework, the workflow for lightweight PGO is very similar to existing IRPGO.

The diagram below has the PGO workflow today (shown in red) in comparison with the workflow for lightweight instrumentation (shown in green). We first create an instrumentation build that produces a raw profile at runtime. Then we use the llvm-profdata tool to convert that raw profile to a profile that the compiler can consume in the PGO build. The main difference for lightweight instrumentation is that we create an instrumentation build with debug info and we use that debug info to create our final profile.

## Prototype & ResultsWe have a proof of concept using dwarf as the extractable metadata and single byte function coverage instrumentation. We measured code size by building Clang with and without instrumentation using -Oz and no value profiling. Our lightweight instrumented Clang binary is only +4 MB (+3.48%) larger than a non-instrumented binary. We compare this with today’s PGO instrumentation Clang binary which is +54 MB (+46.96%) larger. If we used debug info to correlate normal instrumentation (without value profiling) instead of just function coverage then we would expect to see an overhead of +43.2 MB (+37.5%). We don’t have performance data on clang experiments using the prototype since not all components are implemented. However, an alternative implementation earlier (similar to MIP) delivered good performance boost for mobile applications.

Hi Ellis,

I support this proposal – we’ve implemented some of these in our downstream compiler in order to support code coverage with embedded use cases. See the lightning talk slides from last year’s developers’ meeting: https://llvm.org/devmtg/2020-09/slides/PhippsAlan_EmbeddedCodeCoverage_LLVM_Conf_Talk_final.pdf

Specifically, we keep all of the mapping data in the binary file and limit allocatable memory to raw profile counters only. We extend llvm-profdata to extract data from the binary file and combine it with the raw profile data when producing an indexed profile.

And we very recently implemented function-entry-only coverage as well as support for variable size counters. We would like to upstream some or all of this in the future, so perhaps we can work with you on doing that. However, none of our work is productized for general PGO – only code coverage (so no support for reading data back in, for example).

Alan Phipps

MCU Compiler Team

Texas Instruments, Inc.

By the way, it is easier to view this RFC in google groups at https://groups.google.com/g/llvm-dev/c/r03Z6JoN7d4.

Hi Ellis, thanks for the proposal. Improving the usability of Instrumentation PGO is indeed very important.

From the results data below, if I understand correctly, the main savings are from supporting the coverage mode (using boolean counters), right? If we only enable that, the meta data based IRPGO clang size will be 10 MB larger (__llvm_prf_names are strippible or easily doable).

About __llvm_prof_data – it also serves the purpose of detecting CFG mismatch (with cfg hashing). On the other hand, about half of the size is used for value profiling purposes, so for coverage mode when value profile is not needed, its size can be cut in half – leaving the total overhead to be roughly 7 MiB, very close to the debug info based matching scheme.

I support the proposal related to different profiling modes (entry only, boolean counter). I suggest having those features upstreamed. In addition, changes that can reduce existing IRPGO size (e.g, strippable name section, reduced __llvm_prof_data) are also very welcome. After those are done, we will have a better idea if the size is still an issue (with a better comparison with the debug info based method).

thanks,

David

(re-post due to size limit)

Hi Ellis, thanks for the proposal. Improving the usability of Instrumentation PGO is indeed very important.

From the results data below, if I understand correctly, the main savings are from supporting the coverage mode (using boolean counters), right? If we only enable that, the meta data based IRPGO clang size will be 10 MB larger (__llvm_prf_names are strippible or easily doable).

About __llvm_prof_data – it also serves the purpose of detecting CFG mismatch (with cfg hashing). On the other hand, about half of the size is used for value profiling purposes, so for coverage mode when value profile is not needed, its size can be cut in half – leaving the total overhead to be roughly 7 MiB, very close to the debug info based matching scheme.

I support the proposal related to different profiling modes (entry only, boolean counter). I suggest having those features upstreamed. In addition, changes that can reduce existing IRPGO size (e.g, strippable name section, reduced __llvm_prof_data) are also very welcome. After those are done, we will have a better idea if the size is still an issue (with a better comparison with the debug info based method).

thanks,

David

Thanks for the feedback, David.

You’re right that most of the savings comes from coarse-grained instrumentation. However, the situation we’re facing for mobile (and also embedded systems) comes with very tight size constraints. Some components are already built with -Oz, and we’re constantly on the lookout for extra MiB to save so more “features” can get into the components.

For clang self-build example, 7M overhead is much better than 50M+, and 50M->7M indeed look close to 50M->4M as improvements. But comparing to non-PGO, this is still +7M vs +4M. The extra 3M is considered quite significant, and could potentially be a deal breaker for some cases.

In short, this is “close” as you mentioned, but not good enough still. Using dwarf as metadata also has a few benefits over tweaking existing metadata to be extractable: it’s less intrusive, and it’s also a more standardized metadata comparing to PGO’s own metadata.

Thanks,

Wenlei

Hi David,

Thanks for the feedback! I’m happy to see interest in improving instrumentation size in upstream LLVM.

As for detecting CFG mismatches with __llvm_prf_data, we can easily store the cfg hash in Dwarf using the DW_TAG_LLVM_annotation attribute which can store any link-time constant. This value would be included final default.profdata profile without any format change, so the compile would consume it in the same way.

Ellis

Thanks for the feedback, David.

You’re right that most of the savings comes from coarse-grained instrumentation. However, the situation we’re facing for mobile (and also embedded systems) comes with very tight size constraints. Some components are already built with -Oz, and we’re constantly on the lookout for extra MiB to save so more “features” can get into the components.

For clang self-build example, 7M overhead is much better than 50M+, and 50M->7M indeed look close to 50M->4M as improvements. But comparing to non-PGO, this is still +7M vs +4M. The extra 3M is considered quite significant, and could potentially be a deal breaker for some cases.

In short, this is “close” as you mentioned, but not good enough still. Using dwarf as metadata also has a few benefits over tweaking existing metadata to be extractable: it’s less intrusive, and it’s also a more standardized metadata comparing to PGO’s own metadata.

While dwarf is a standard way of program annotation, using it for instrumentation PGO does mean an additional dependency (instead of being self contained).

This proposal requires debug_type info to be emitted, right? What is the object size and compile time overhead? If this can be trimmed, it is a reasonable way to emit the profile data mapping information at compile time.

David

While dwarf is a standard way of program annotation, using it for instrumentation PGO does mean an additional dependency (instead of being self contained).

I think your point on having debug info as dependency is fair. But given most release builds need to generate debug info anyways, this dependency seems to be a minor downside. Hence the tradeoff between extra dependency and minimum binary size can be worthwhile. In the end, this does not “regress” an existing feature by adding extra dependency – i.e. it doesn’t affect today’s IRPGO; and the actual dependency, the debug info is already commonly used. Hope this is an acceptable trade-off.

This proposal requires debug_type info to be emitted, right? What is the object size and compile time overhead? If this can be trimmed, it is a reasonable way to emit the profile data mapping information at compile time.

We don’t really require debug_type, so yes it can be trimmed as a dependency. However, in practice whether we should trim it also depends on whether it makes sense from dwarf perspective, and if it’s worth a new mode/variant for dwarf info. I tend it view the trimming of debug_type as orthogonal to this proposal – as long as we keep the lightweight PGO technically independent of debug types, the trimming can be done later if it’s appropriate from debug info perspective and when the actual use case arises.

Thanks,

Wenlei

While dwarf is a standard way of program annotation, using it for instrumentation PGO does mean an additional dependency (instead of being self contained).

I think your point on having debug info as dependency is fair. But given most release builds need to generate debug info anyways, this dependency seems to be a minor downside. Hence the tradeoff between extra dependency and minimum binary size can be worthwhile. In the end, this does not “regress” an existing feature by adding extra dependency – i.e. it doesn’t affect today’s IRPGO; and the actual dependency, the debug info is already commonly used. Hope this is an acceptable trade-off.

This proposal requires debug_type info to be emitted, right? What is the object size and compile time overhead? If this can be trimmed, it is a reasonable way to emit the profile data mapping information at compile time.

We don’t really require debug_type, so yes it can be trimmed as a dependency. However, in practice whether we should trim it also depends on whether it makes sense from dwarf perspective, and if it’s worth a new mode/variant for dwarf info. I tend it view the trimming of debug_type as orthogonal to this proposal – as long as we keep the lightweight PGO technically independent of debug types, the trimming can be done later if it’s appropriate from debug info perspective and when the actual use case arises.

Thanks. +Vedant and +Petr for their opinion on using debug info for matching purposes.

David

From our perspective, using debug info wouldn’t be an issue since we always include it in every build.

The issue that hasn’t been brought up yet is that the proposed solution only covers DWARF, but the IR instrumentation also supports Windows which uses PDB. More generally, the existing implementation is independent of the debugging format being used. I’m not familiar with PDB to say whether the proposed solution could be extended to also support Windows, but if not then we’ll have to support both the use of __llvm_prf_data and __llvm_prf_names sections as well as debug info which seems like an extra complexity.

I agree with David that making __llvm_prf_names strippable and cutting down the size of __llvm_prf_data would be a worthwhile effort independently of the proposed changes.

While I am not familiar with the PDB format, most of the changes are independent of the debug info format type. For the instrumentation counters, we emit a !DIGlobalVariable() in the LLVM IR which is emitted as Dwarf or CodeView automatically. That being said, I was planning on first extending the llvm-profdata tool to handle Dwarf and then I can add support for CodeView if needed.

ok thanks. SGTM.

Since this work has landed and is now on LLVM 14, I would like to give a quick followup describing how to use Lightweight Instrumentation and it’s results.

I added the llvm flag -debug-info-correlate with -fprofile-generate to create an instrumenting build that uses external debug info instead of metadata that remains in the binary. This is implemented in ⚙ D115693 [Try2][InstrProf] Attach debug info to counters and ⚙ D115915 [Try2][InstrProf] Add Correlator class to read debug info. Note that this is not yet compatible with value profiling so you’ll need to use -disable-vp=true.

clang -fprofile-generate -mllvm -debug-info-correlate -mllvm --disable-vp=true main.cpp
./a.out
llvm-profdata merge -o default.profile --debug-info=a.out default.proflite

Then I added the llvm flag -pgo-function-entry-coverage in ⚙ D116180 [InstrProf] Add single byte coverage mode to instrument function coverage by using a single coverage byte for each function rather than using 64 bit counters for block counts. In practice, this is useful to detect dead code.

With these two additions, I was able to create a clang binary with function coverage instrumentation that only 7M (4.9%) larger than the non-instrumented binary. In contrast, the default instrumentation without value profiling has an overhead of 31M (21.7%).

Let me know if you have any questions!

4 Likes