Use #102774 to allocate storage for decoded probes (`PseudoProbeVec`)
and function records (`InlineTreeVec`).
Leverage that to also shrink sizes of `MCDecodedPseudoProbe`:
- Drop Guid since it's accessible via `InlineTree`.
`MCDecodedPseudoProbeInlineTree`:
- Keep track of probes and inlinees using `ArrayRef`s now that probes
and function records belonging to the same function are allocated
contiguously.
This reduces peak RSS from 13.7 GiB to 9.7 GiB and pseudo probe parsing
time (as part of perf2bolt) from 15.3s to 9.6s for a large binary with
400MiB .pseudo_probe section containing 43M probes and 25M function
records.
Depends on:
#102774#102787#102788
Reviewers: maksfb, rafaelauler, dcci, ayermolo, wlei-llvm
Reviewed By: wlei-llvm
Pull Request: https://github.com/llvm/llvm-project/pull/102789
This change introduces two options which may be used to create profiles
of arbitrary PMU events.
1. `--leading-ip-only` provides a simple sample-IP-based profile mode.
This is not useful for building a profile of execution frequency, but it
is useful for building new types of profiles.
For example, to build a profile of unpredictable branches:
perf record -b -e branch-misses:upp -o perf.data ... llvm-profgen
--perfdata perf.data --leading-ip-only ...
2. `--perf-event=event` enables the creation of a profile concerned with
a specific event or set of events. The names given should match the
"event" field as emitted by perf-script(1).
This option has two spellings: `--perf-event` and `--perf-events`. The
plural spelling accepts a comma-separated list. The singular spelling
appends a single event name to the set of events which will be used.
This is meant to accommodate event names containing commas.
Combined, these options allow generating multiple kinds of profiles from
a single `perf record` collection. For example, to generate both
execution frequency and branch mispredict profiles:
perf record -c 1000003 -b -e
br_inst_retired.near_taken:upp,br_misp_retired.all_branches:upp ...
llvm-profgen --output execution.prof
--perf-event=br_inst_retired.near_taken:upp ...
llvm-profgen --leading-ip-only --output unpredictable.prof
--perf-event=br_misp_retired.all_branches:upp ...
These additions are in support of more general HWPGO[^1], allowing
feedback from a wider range of hardware events.
[^1]:
https://llvm.org/devmtg/2024-04/slides/TechnicalTalks/Xiao-EnablingHW-BasedPGO.pdf
---------
Co-authored-by: Tim Creech <tcreech@tcreech.com>
Also some control flow simplifications.
Notably, this doesn't address `sampleprof_error`. I *think* the style
there tries to match `std::error_category`.
Also left `hash_value` as-is, because it matches what we do in Hashing.h
The profile density feature(the amount of samples in the profile
relative to the program size) is used to identify insufficient sample
issue and provide hints for user to increase sample count. A low-density
profile can be inaccurate due to statistical noise, which can hurt FDO
performance.
This change introduces two improvements to the current density work.
1. The density calculation/definition is changed. Previously, the
density of a profile was calculated as the minimum density for all warm
functions (a function was considered warm if its total samples were
within the top N percent of the profile). However, there is a problem
that a high total sample profile can have a very low density, which
makes the density value unstable.
- Instead, we want to find a density number such that if a function's
density is below this value, it is considered low-density function. We
consider the whole profile is bad if a group of low-density functions
have the sum of samples that exceeds N percent cut-off of the total
samples.
- In implementation, we sort the function profiles by density, iterate
them in descending order and keep accumulating the body samples until
the sum exceeds the (100% - N) percentage of the total_samples, the
profile-density is the last(minimum) function-density of processed
functions. We introduce the a flag(`--profile-density-threshold`) for
this percentage threshold.
2. The density is now calculated based on final(compiler used) profiles
instead of merged context-less profiles.
For the built-in local initialization function(`__cxx_global_var_init`,
`__tls_init` prefix), there could be multiple versions of the functions
in the final binary, e.g. `__cxx_global_var_init`, which is a wrapper of
global variable ctors, the compiler could assign suffixes like
`__cxx_global_var_init.N` for different ctors.
However, in the profile generation, we call `getCanonicalFnName` to
canonicalize the names which strip the suffixes. Therefore, samples from
different functions queries the same profile(only
`__cxx_global_var_init`) and the counts are merged. As the functions are
essentially different, entries of the merged profile are ambiguous. In
sample loading, for each version of this function, the IR from one
version would be attributed towards a merged entries, which is
inaccurate, especially for fuzzy profile matching, it gets multiple
callsites(from different function) but using to match one callsite,
which mislead the matching and report a lot of false positives.
Hence, we want to filter them out from the profile map during the
profile generation time. The profiles are all cold functions, it won't
have perf impact.
This is phase 2 of the MD5 refactoring on Sample Profile following
https://reviews.llvm.org/D147740
In previous implementation, when a MD5 Sample Profile is read, the
reader first converts the MD5 values to strings, and then create a
StringRef as if the numerical strings are regular function names, and
later on IPO transformation passes perform string comparison over these
numerical strings for profile matching. This is inefficient since it
causes many small heap allocations.
In this patch I created a class `ProfileFuncRef` that is similar to
`StringRef` but it can represent a hash value directly without any
conversion, and it will be more efficient (I will attach some benchmark
results later) when being used in associative containers.
ProfileFuncRef guarantees the same function name in string form or in
MD5 form has the same hash value, which also fix a few issue in IPO
passes where function matching/lookup only check for function name
string, while returns a no-match if the profile is MD5.
When testing on an internal large profile (> 1 GB, with more than 10
million functions), the full profile load time is reduced from 28 sec to
25 sec in average, and reading function offset table from 0.78s to 0.7s
This is phase 1 of multiple planned improvements on the sample profile loader. The major change is to use MD5 hash code ((instead of the function itself) as the key to look up the function offset table and the profiles, which significantly reduce the time it takes to construct the map.
The optimization is based on the fact that many practical sample profiles are using MD5 values for function names to reduce profile size, so we shouldn't need to convert the MD5 to a string and then to a SampleContext and use it as the map's key, because it's extremely slow.
Several changes to note:
(1) For non-CS SampleContext, if it is already MD5 string, the hash value will be its integral value, instead of hashing the MD5 again. In phase 2 this is going to be optimized further using a union to represent MD5 function (without converting it to string) and regular function names.
(2) The SampleProfileMap is a wrapper to *map<uint64_t, FunctionSamples>, while providing interface allowing using SampleContext as key, so that existing code still work. It will check for MD5 collision (unlikely but not too unlikely, since we only takes the lower 64 bits) and handle it to at least guarantee compilation correctness (conflicting old profile is dropped, instead of returning an old profile with inconsistent context). Other code should not try to use MD5 as key to access the map directly, because it will not be able to handle MD5 collision at all. (see exception at (5) )
(3) Any SampleProfileMap::emplace() followed by SampleContext assignment if newly inserted, should be replaced with SampleProfileMap::Create(), which does the same thing.
(4) Previously we ensure an invariant that in SampleProfileMap, the key is equal to the Context of the value, for profile map that is eventually being used for output (as in llvm-profdata/llvm-profgen). Since the key became MD5 hash, only the value keeps the context now, in several places where an intermediate SampleProfileMap is created, each new FunctionSample's context is set immediately after insertion, which is necessary to "remember" the context otherwise irretrievable.
(5) When reading a profile, we cache the MD5 values of all functions, because they are used at least twice (one to index into FuncOffsetTable, the other into SampleProfileMap, more if there are additional sections), in this case the SampleProfileMap is directly accessed with MD5 value so that we don't recalculate it each time (expensive)
Performance impact:
When reading a ~1GB extbinary profile (fixed length MD5, not compressed) with 10 million function names and 2.5 million top level functions (non CS functions, each function has varying nesting level from 0 to 20), this patch improves the function offset table loading time by 20%, and improves full profile read by 5%.
Reviewed By: davidxl, snehasish
Differential Revision: https://reviews.llvm.org/D147740
This is phase 1 of multiple planned improvements on the sample profile loader. The major change is to use MD5 hash code ((instead of the function itself) as the key to look up the function offset table and the profiles, which significantly reduce the time it takes to construct the map.
The optimization is based on the fact that many practical sample profiles are using MD5 values for function names to reduce profile size, so we shouldn't need to convert the MD5 to a string and then to a SampleContext and use it as the map's key, because it's extremely slow.
Several changes to note:
(1) For non-CS SampleContext, if it is already MD5 string, the hash value will be its integral value, instead of hashing the MD5 again. In phase 2 this is going to be optimized further using a union to represent MD5 function (without converting it to string) and regular function names.
(2) The SampleProfileMap is a wrapper to *map<uint64_t, FunctionSamples>, while providing interface allowing using SampleContext as key, so that existing code still work. It will check for MD5 collision (unlikely but not too unlikely, since we only takes the lower 64 bits) and handle it to at least guarantee compilation correctness (conflicting old profile is dropped, instead of returning an old profile with inconsistent context). Other code should not try to use MD5 as key to access the map directly, because it will not be able to handle MD5 collision at all. (see exception at (5) )
(3) Any SampleProfileMap::emplace() followed by SampleContext assignment if newly inserted, should be replaced with SampleProfileMap::Create(), which does the same thing.
(4) Previously we ensure an invariant that in SampleProfileMap, the key is equal to the Context of the value, for profile map that is eventually being used for output (as in llvm-profdata/llvm-profgen). Since the key became MD5 hash, only the value keeps the context now, in several places where an intermediate SampleProfileMap is created, each new FunctionSample's context is set immediately after insertion, which is necessary to "remember" the context otherwise irretrievable.
(5) When reading a profile, we cache the MD5 values of all functions, because they are used at least twice (one to index into FuncOffsetTable, the other into SampleProfileMap, more if there are additional sections), in this case the SampleProfileMap is directly accessed with MD5 value so that we don't recalculate it each time (expensive)
Performance impact:
When reading a ~1GB extbinary profile (fixed length MD5, not compressed) with 10 million function names and 2.5 million top level functions (non CS functions, each function has varying nesting level from 0 to 20), this patch improves the function offset table loading time by 20%, and improves full profile read by 5%.
Reviewed By: davidxl, snehasish
Differential Revision: https://reviews.llvm.org/D147740
This is phase 1 of multiple planned improvements on the sample profile loader. The major change is to use MD5 hash code ((instead of the function itself) as the key to look up the function offset table and the profiles, which significantly reduce the time it takes to construct the map.
The optimization is based on the fact that many practical sample profiles are using MD5 values for function names to reduce profile size, so we shouldn't need to convert the MD5 to a string and then to a SampleContext and use it as the map's key, because it's extremely slow.
Several changes to note:
(1) For non-CS SampleContext, if it is already MD5 string, the hash value will be its integral value, instead of hashing the MD5 again. In phase 2 this is going to be optimized further using a union to represent MD5 function (without converting it to string) and regular function names.
(2) The SampleProfileMap is a wrapper to *map<uint64_t, FunctionSamples>, while providing interface allowing using SampleContext as key, so that existing code still work. It will check for MD5 collision (unlikely but not too unlikely, since we only takes the lower 64 bits) and handle it to at least guarantee compilation correctness (conflicting old profile is dropped, instead of returning an old profile with inconsistent context). Other code should not try to use MD5 as key to access the map directly, because it will not be able to handle MD5 collision at all. (see exception at (5) )
(3) Any SampleProfileMap::emplace() followed by SampleContext assignment if newly inserted, should be replaced with SampleProfileMap::Create(), which does the same thing.
(4) Previously we ensure an invariant that in SampleProfileMap, the key is equal to the Context of the value, for profile map that is eventually being used for output (as in llvm-profdata/llvm-profgen). Since the key became MD5 hash, only the value keeps the context now, in several places where an intermediate SampleProfileMap is created, each new FunctionSample's context is set immediately after insertion, which is necessary to "remember" the context otherwise irretrievable.
(5) When reading a profile, we cache the MD5 values of all functions, because they are used at least twice (one to index into FuncOffsetTable, the other into SampleProfileMap, more if there are additional sections), in this case the SampleProfileMap is directly accessed with MD5 value so that we don't recalculate it each time (expensive)
Performance impact:
When reading a ~1GB extbinary profile (fixed length MD5, not compressed) with 10 million function names and 2.5 million top level functions (non CS functions, each function has varying nesting level from 0 to 20), this patch improves the function offset table loading time by 20%, and improves full profile read by 5%.
Reviewed By: davidxl, snehasish
Differential Revision: https://reviews.llvm.org/D147740
This is phase 1 of multiple planned improvements on the sample profile loader. The major change is to use MD5 hash code ((instead of the function itself) as the key to look up the function offset table and the profiles, which significantly reduce the time it takes to construct the map.
The optimization is based on the fact that many practical sample profiles are using MD5 values for function names to reduce profile size, so we shouldn't need to convert the MD5 to a string and then to a SampleContext and use it as the map's key, because it's extremely slow.
Several changes to note:
(1) For non-CS SampleContext, if it is already MD5 string, the hash value will be its integral value, instead of hashing the MD5 again. In phase 2 this is going to be optimized further using a union to represent MD5 function (without converting it to string) and regular function names.
(2) The SampleProfileMap is a wrapper to *map<uint64_t, FunctionSamples>, while providing interface allowing using SampleContext as key, so that existing code still work. It will check for MD5 collision (unlikely but not too unlikely, since we only takes the lower 64 bits) and handle it to at least guarantee compilation correctness (conflicting old profile is dropped, instead of returning an old profile with inconsistent context). Other code should not try to use MD5 as key to access the map directly, because it will not be able to handle MD5 collision at all. (see exception at (5) )
(3) Any SampleProfileMap::emplace() followed by SampleContext assignment if newly inserted, should be replaced with SampleProfileMap::Create(), which does the same thing.
(4) Previously we ensure an invariant that in SampleProfileMap, the key is equal to the Context of the value, for profile map that is eventually being used for output (as in llvm-profdata/llvm-profgen). Since the key became MD5 hash, only the value keeps the context now, in several places where an intermediate SampleProfileMap is created, each new FunctionSample's context is set immediately after insertion, which is necessary to "remember" the context otherwise irretrievable.
(5) When reading a profile, we cache the MD5 values of all functions, because they are used at least twice (one to index into FuncOffsetTable, the other into SampleProfileMap, more if there are additional sections), in this case the SampleProfileMap is directly accessed with MD5 value so that we don't recalculate it each time (expensive)
Performance impact:
When reading a ~1GB extbinary profile (fixed length MD5, not compressed) with 10 million function names and 2.5 million top level functions (non CS functions, each function has varying nesting level from 0 to 20), this patch improves the function offset table loading time by 20%, and improves full profile read by 5%.
Reviewed By: davidxl, snehasish
Differential Revision: https://reviews.llvm.org/D147740
This change enables generating pseudo-probe-based FS-AFDO profiles. The change is straightforward based-on previous change {D147651} by just injecting FS-discriminators into various profile generation spot.
Reviewed By: wenlei
Differential Revision: https://reviews.llvm.org/D147957
For profile staleness report, before it only counts for the top-level function samples in the nested profile, the samples in the inlinees are ignored. This could affect the quality of the metrics when there are heavily inlined functions. This change adds a feature to flatten the nested profile and we're changing to use flatten profile as the input for stale profile detection and matching.
Example for profile flattening:
```
Original profile:
_Z3bazi:20301:1000
1: 1000
3: 2000
5: inline1:1600
1: 600
3: inline2:500
1: 500
Flattened profile:
_Z3bazi:18701:1000
1: 1000
3: 2000
5: 600 inline1:600
inline1:1100:600
1: 600
3: 500 inline2: 500
inline2:500:500
1: 500
```
This feature could be useful for offline analysis, like understanding the hotness of each individual function. So I'm adding the support to `llvm-profdata merge` under `--gen-flattened-profile`.
Reviewed By: hoy, wenlei
Differential Revision: https://reviews.llvm.org/D146452
This change introduces a missing frame inferrer aiming at fixing missing frames. It current only handles missing frames due to the compiler tail call elimination (TCE) but could also be extended to supporting other scenarios like frame pointer omission. When a tail called function is sampled, the caller frame will be missing from the call chain because the caller frame is reused for the callee frame. While TCE is beneficial to both perf and reducing stack overflow, a workaround being made in this change aims to find back the missing frames as much as possible.
The idea behind this work is to build a dynamic call graph that consists of only tail call edges constructed from LBR samples and DFS-search for a unique path for a given source frame and target frame on the graph. The unique path will be used to fill in the missing frames between the source and target. Note that only a unique path counts. Multiple paths are treated unreachable since we don't want to overcount for any particular possible path.
A switch --infer-missing-frame is introduced and defaults to be on.
Some testing results:
- 0.4% perf win according to three internal benchmarks.
- About 2/3 of the missing tail call frames can be recovered, according to an internal benchmark.
- 10% more profile generation time.
Reviewed By: wenlei
Differential Revision: https://reviews.llvm.org/D139367
As titled. The change should have minimal impact since the targets of branch samples are mostly covered by range samples.
Reviewed By: wenlei, wlei
Differential Revision: https://reviews.llvm.org/D137203
Currently pseudo probe encoding for a function is like:
- For the first probe, a relocation from it to its physical position in the code body
- For subsequent probes, an incremental offset from the current probe to the previous probe
The relocation could potentially cause relocation overflow during link time. I'm now replacing it with an offset from the first probe to the function start address.
A source function could be lowered into multiple binary functions due to outlining (e.g, coro-split). Since those binary function have independent link-time layout, to really avoid relocations from .pseudo_probe sections to .text sections, the offset to replace with should really be the offset from the probe's enclosing binary function, rather than from the entry of the source function. This requires some changes to previous section-based emission scheme which now switches to be function-based. The assembly form of pseudo probe directive is also changed correspondingly, i.e, reflecting the binary function name.
Most of the source functions end up with only one binary function. For those don't, a sentinel probe is emitted for each of the binary functions with a different name from the source. The sentinel probe indicates the binary function name to differentiate subsequent probes from the ones from a different binary function. For examples, given source function
```
Foo() {
…
Probe 1
…
Probe 2
}
```
If it is transformed into two binary functions:
```
Foo:
…
Foo.outlined:
…
```
The encoding for the two binary functions will be separate:
```
GUID of Foo
Probe 1
GUID of Foo
Sentinel probe of Foo.outlined
Probe 2
```
Then probe1 will be decoded against binary `Foo`'s address, and Probe 2 will be decoded against `Foo.outlined`. The sentinel probe of `Foo.outlined` makes sure there's not accidental relocation from `Foo.outlined`'s probes to `Foo`'s entry address.
On the BOLT side, to be minimal intrusive, the pseudo probe re-encoding sticks with the old encoding format. This is fine since unlike linker, Bolt processes the pseudo probe section as a whole and it is free from relocation overflow issues.
The change is downwards compatible as long as there's no mixed use of the old encoding and the new encoding.
Reviewed By: wenlei, maksfb
Differential Revision: https://reviews.llvm.org/D135912
Differential Revision: https://reviews.llvm.org/D135914
Differential Revision: https://reviews.llvm.org/D136394
In `computeInlinedContextSizeForRange`, the offset of range is only used one time, there is no need to cache the frame location stack.
Measured on one internal service binary, this can save 2GB memory usage and reduce a small run time (avoid one hash search).
Reviewed By: hoy, wenlei
Differential Revision: https://reviews.llvm.org/D128859
This is to fix two issues related with loading address:
1) When multiple MMAPs occur and their loading address are different, before it only used the first MMap as base address, all perf address after it used the wrong base address.
2) For pseudo probe profile, the address is always based on preferred loading address. If the base address is not equal to the preferred loading address, the pseudo probe address query will be wrong.
Solution: Instead of converting the address to offset lazily, right now all the address after parsing are converted on the fly based on preferred loading address in the parsing time. There is no "offset" used in profile generator any more.
Reviewed By: hoy, wenlei
Differential Revision: https://reviews.llvm.org/D126827
The name `getEntrySamples` was misleading for 2 reasons. One, it's
close in name to `Function::getEntryCount`, but the equivalent here is
`getHeadSamples`; second, as opposed to the other get* APIs in
`FunctionSamples`, it performs an estimate/heuristic rather than just
retrieving raw data (or a non-heuristic derivate off that data, like
`getMaxCountInside`)
The new name should more clearly communicate its intent; and, being
close (in name) to `getHeadSamples`, it should allow the reader discover
the relation between them.
Also updated the doc comments for both `getHeadSamples[Estimate]` so a
reader may better understand the relation between them.
Differential Revision: https://reviews.llvm.org/D130281
This is the followup patch to https://reviews.llvm.org/D125246 for the `SampleContextTracker` part. Before the promotion and merging of the context is based on the SampleContext(the array of frame), this causes a lot of cost to the memory. This patch detaches the tracker from using the array ref instead to use the context trie itself. This can save a lot of memory usage and benefit both the compiler's CS inliner and llvm-profgen's pre-inliner.
One structure needs to be specially treated is the `FuncToCtxtProfiles`, this is used to get all the functionSamples for one function to do the merging and promoting. Before it search each functions' context and traverse the trie to get the node of the context. Now we don't have the context inside the profile, instead we directly use an auxiliary map `ProfileToNodeMap` for profile , it initialize to create the FunctionSamples to TrieNode relations and keep updating it during promoting and merging the node.
Moreover, I was expecting the results before and after remain the same, but I found that the order of FuncToCtxtProfiles matter and affect the results. This can happen on recursive context case, but the difference should be small. Now we don't have the context, so I just used a vector for the order, the result is still deterministic.
Measured on one huge size(12GB) profile from one of our internal service. The profile similarity difference is 99.999%, and the running time is improved by 3X(debug mode) and the memory is reduced from 170GB to 90GB.
Reviewed By: hoy, wenlei
Differential Revision: https://reviews.llvm.org/D127031
Our investigation showed ProfileMap's key is the bottleneck of the memory consumption for CS profile generation on some large services. This patch tries to optimize it by storing the CS function samples using the context trie tree structure instead of the context frame array ref. Parts of code in `ContextTrieNode` are reused.
Our experiment on one internal service showed that the context key's memory can be reduced from 80GB to 300MB.
To be compatible with non-CS profiles, the profile writer still needs to use ProfileMap as input, so rebuild the ProfileMap using the context trie in `postProcessProfiles`.
The optimization is not complete yet, next step is to reimplement Pre-inliner or profile trimmer, after that, ProfileMap should be small to be written.
Reviewed By: hoy, wenlei
Differential Revision: https://reviews.llvm.org/D125246
Some cl::ZeroOrMore were added to avoid the `may only occur zero or one times!`
error. More were added due to cargo cult. Since the error has been removed,
cl::ZeroOrMore is unneeded.
Also remove cl::init(false) while touching the lines.
Current profile generation caculcates callsite body samples and call target samples separately. The former is done based on LBR range samples while the latter is done based on branch samples. Note that there's a subtle difference. LBR ranges is formed from two consecutive branch samples. Therefore the last entry in a LBR record will not be counted towards body samples while there's still a chance for it to be counted towards call targets if it is a function call. I'm making sense of the call body samples by updating it to the aggregation of call targets.
Reviewed By: wenlei
Differential Revision: https://reviews.llvm.org/D122609
The per-callsite size threshold used today to drive preinline decision is based on hotness/coldness cutoff. The default setup is for callsites with a sample count above the hotness cutoff (99%), a 1500 size threshold is used. Any callsite below 99.99% coldness cutoff uses a zero threshold. This has a couple issues:
1. While both cutoffs and size thoresholds are configurable, different applications may need different setups, making a universal setup impractical.
2. The callsites between hotness cutoff and coldness cutoff are not considered as inline candidates, which could be a missing opportunity.
3. Hot callsites always use the same threshold. In reality we may want a bigger threshold for hotter callsites.
In this change we are introducing a linear threshold regardless of hot/cold cutoffs. Given a sample space, a threshold is computed for a callsite based on the position of that callsite sample in the whole space. With that we no longer need to define what's hot or cold. Callsites with different hotness will get a different threshold. This should overcome the above three issues.
I have seen good results with a universal default setup for two of our internal services.
For one service, 0.2% to 0.5% perf improvement over a baseline with a previous default setup, on-par code size.
For the second service, 0.5% to 0.8% perf improvement over a baseline with a previous default setup, 0.2% code size increase; on-par performance and code size with a baseline that is with a carefully tuned cutoff to cover enough hot functions.
Reviewed By: wenlei
Differential Revision: https://reviews.llvm.org/D125023
To be more clear and definitive, I'm renaming `ProfileIsCSFlat` back to `ProfileIsCS` which stands for full context-sensitive flat profiles. `ProfileIsCSNested` is now renamed to `ProfileIsPreInlined` and is extended to be applicable for CS flat profiles too. More specifically, `ProfileIsPreInlined` is for any kind of profiles (flat or nested) that contain 'ShouldBeInlined' contexts. The flag is encoded in the profile summary section for extbinary profiles and is computed on-the-fly for text profiles.
Reviewed By: wenlei
Differential Revision: https://reviews.llvm.org/D122602
This patch is fixing two issues for both CS and non-CS.
1) For external-call-internal, the head samples of the the internal function should be recorded.
2) avoid ignoring LBR after meeting the interrupt branch for CS profile
LBR parser is shared between CS and non-CS, we found it's error-prone while dealing with artificial branch inside LBR parser. Since artificial branch is mainly used for CS profile unwinding, this patch tries to simplify LBR parser by decoupling artificial branch code from it, the concept of artificial branch is removed and split into two transitional branches(internal-to-external, external-to-internal). Then we leave all the processing of external branch to unwinder.
Specifically for unwinder, remembering that we introduce external frame in https://reviews.llvm.org/D115550. We can just take external address as a regular address and reuse current unwind function(unwindCall, unwindReturn). For a normal case, the external frame will match an external LBR, and it will be filtered out by `unwindLinear` without losing any context.
The data also shows that the interrupt or standalone LBR pattern(unpaired case) does exist, we choose to handle it by clearing the call stack and keeping unwinding. Here we leverage checking in `unwindLinear`, because a standalone LBR, no matter its type, since it doesn’t have other part to pair, it will eventually cause a wrong linear range, like [external, internal], [internal, external]. Then set the state to invalid there.
Reviewed By: hoy, wenlei
Differential Revision: https://reviews.llvm.org/D118177
A context can be created by invoking the `getFunctionProfileForContext` function in two ways:
- by using a probe and its calling context.
- by removing the leaf frame from an existing contexts. The first way is used when generating a function profile for a given LBR range, and the input `WasLeafInlined` is computed depending on the actually probe in the LBR range. The second way is used when using the entry count of an inlinee function profile to update its inliner callsite count, so `WasLeafInlined` is unknown for the inliner frame.
The two invocations can happen in different order on different platforms, since the lbr ranges are stored in an unordered_map, and we are making sure `ContextWasInlined` is always set correctly.
This should fix the random test failure introduced by D121655
Reviewed By: wenlei
Differential Revision: https://reviews.llvm.org/D122844
Sometimes we would like to run post-processing repeatedly on the original sample profile for tuning. In order to avoid regenerating the original profile from scratch every time, this change adds the support of reading in the original profile (called symbolized profile) and running the post-processor on it.
Reviewed By: wenlei
Differential Revision: https://reviews.llvm.org/D121655
Complete pseudo probes decoding can result in large memory usage. In practice only a small porting of the decoded probes are used in profile generation. I'm changing the full decoding mode to be decoding for profiled functions only, though we still do a full scan of the .pseudoprobe section due to a missing table-of-content but we don't have to build the in-memory data structure for functions not sampled.
To build the in-memory data structure for profiled functions only, I'm rewriting the previous non-recursive probe decoding logic to be recursive. This is easy to read and maintain.
I also have to change the previous representation of unsymbolized context from probe-based stack to address-based stack since the profiled functions are unknown yet by the time of virtual unwinding. The address-based stack will be converted to probe-based stack after virtual unwinding and on-demand probe decoding.
I'm seeing 20GB memory is saved for one of our internal large service.
Reviewed By: wenlei
Differential Revision: https://reviews.llvm.org/D121643
CS nested profile has a benefit over the CS flat profile that is to speed up the build while achieve an on-par performance. I'm turning it on by default for CSSPGO.
Reviewed By: wenlei
Differential Revision: https://reviews.llvm.org/D121142
I'm bring up the support of pseudo-probe-based non-CS profile generation. The approach is quite similar to generating dwarf-based non-CS profile. The main difference is for a given linear instruction range, instead of each disassembled instruction, pseudo probes that are covered by the range are processed. The pseudo probe extraction code is shared with CS probe profile generation.
I'm seeing 0.7% performance win for one of our internal large benchmark compared to using non-CS dwarf-based profile, and 0.5% win for another large benchmark when combined with profi.
Reviewed By: wenlei
Differential Revision: https://reviews.llvm.org/D120335
Tracking optimized-away inlinees based on all probes in a binary is expansive in terms of memory usage I'm making the tracking on-demand based on profiled functions only. This saves about 10% memory overall for a medium-sized benchmark.
Before:
note: After parsePerfTraces
note: Thu Jan 27 18:42:09 2022
note: VM: 8.68 GB RSS: 8.39 GB
note: After computeSizeForProfiledFunctions
note: Thu Jan 27 18:42:41 2022
note: **VM: 10.63 GB RSS: 10.20 GB**
note: After generateProbeBasedProfile
note: Thu Jan 27 18:45:49 2022
note: VM: 25.00 GB RSS: 24.95 GB
note: After postProcessProfiles
note: Thu Jan 27 18:49:29 2022
note: VM: 26.34 GB RSS: 26.27 GB
After:
note: After parsePerfTraces
note: Fri Jan 28 12:04:49 2022
note: VM: 8.68 GB RSS: 7.65 GB
note: After computeSizeForProfiledFunctions
note: Fri Jan 28 12:05:26 2022
note: **VM: 8.68 GB RSS: 8.42 GB**
note: After generateProbeBasedProfile
note: Fri Jan 28 12:08:03 2022
note: VM: 22.93 GB RSS: 22.89 GB
note: After postProcessProfiles
note: Fri Jan 28 12:11:30 2022
note: VM: 24.27 GB RSS: 24.22 GB
This should be a no-diff change in terms of profile quality.
Reviewed By: wenlei
Differential Revision: https://reviews.llvm.org/D118515
Error out instead of siliently generate empty profile when trying to generate AutoFDO profile with probe binary.
Differential Revision: https://reviews.llvm.org/D116508
CSSPGO currently employs a flat profile format for context-sensitive profiles. Such a flat profile allows for precisely manipulating contexts that is either inlined or not inlined. This is a benefit over the nested profile format used by non-CS AutoFDO. A downside of this is the longer build time due to parsing the indexing the full CS contexts.
For a CS flat profile, though only the context profiles relevant to a module are loaded when that module is compiled, the cost to figure out what profiles are relevant is noticeably high when there're many contexts, since the sample reader will need to scan all context strings anyway. On the contrary, a nested function profile has its related inline subcontexts isolated from other unrelated contexts. Therefore when compiling a set of functions, unrelated contexts will never need to be scanned.
In this change we are exploring using nested profile format for CSSPGO. This is expected to work based on an assumption that with a preinliner-computed profile all contexts are precomputed and expected to be inlined by the compiler. Contexts not expected to be inlined will be cut off and returned to corresponding base profiles (for top-level outlined functions). This naturally forms a nested profile where all nested contexts are expected to be inlined. The compiler will less likely optimize on derived contexts that are not precomputed.
A CS-nested profile will look exactly the same with regular nested profile except that each nested profile can come with an attributes. With pseudo probes, a nested profile shown as below can also have a CFG checksum.
```
main:1968679:12
2: 24
3: 28 _Z5funcAi:18
3.1: 28 _Z5funcBi:30
3: _Z5funcAi:1467398
0: 10
1: 10 _Z8funcLeafi:11
3: 24
1: _Z8funcLeafi:1467299
0: 6
1: 6
3: 287884
4: 287864 _Z3fibi:315608
15: 23
!CFGChecksum: 138828622701
!Attributes: 2
!CFGChecksum: 281479271677951
!Attributes: 2
```
Specific work included in this change:
- A recursive profile converter to convert CS flat profile to nested profile.
- Extend function checksum and attribute metadata to be stored in nested way for text profile and extbinary profile.
- Unifiy sample loader inliner path for CS and preinlined nested profile.
- Changes in the sample loader to support probe-based nested profile.
I've seen promising results regarding build time. A nested profile can result in a 20% shorter build time than a CS flat profile while keep an on-par performance. This is with -duplicate-contexts-into-base=1.
Test Plan:
Reviewed By: wenlei
Differential Revision: https://reviews.llvm.org/D115205
Since total sample and body sample are used to compute hotness threshold in compiler, we found in some services changing the total samples computation will cause noticeable regression. Hence, here we will revert the changes and just keep all total samples number identical to the old tool.
Three changes in this diff:
1. Revert previous diff(https://reviews.llvm.org/D112672: [llvm-profgen] Update total samples by accumulating all its body samples) and put it under a switch.
2. Keep the negative line number. Although compiler doesn't consume the count but it will be used to compute hot threshold.
3. Change to accumulate total samples per byte instead of per instruction.
Reviewed By: hoy, wenlei
Differential Revision: https://reviews.llvm.org/D115013
This change allows to trim the profile if it's considered to be cold for baseline AutoFDO. We reuse the cold threshold from `ProfileSummaryBuilder::getColdCountThreshold(..)` which can be set by percent(--profile-summary-cutoff-cold) or by value(--profile-summary-cold-count).
Reviewed By: hoy, wenlei
Differential Revision: https://reviews.llvm.org/D113785
Due to the debug info merging, there may have some contexts with zero probe id, we should truncate the context to avoid misleading pre-inliner.
Reviewed By: hoy, wenlei
Differential Revision: https://reviews.llvm.org/D114284
