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 function looks up MD5ToNameMap to return a name for a given MD5.
https://github.com/llvm/llvm-project/pull/66825 adds MD5 of global
variable names into this map. So rename methods and update comments
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
Emit warnings when `InstrProfCorrelator` finds problems with debug info for lightweight instrumentation profile correlation. To prevent excessive printing, only emit the first 5 warnings.
In addition, remove a diagnostic about missing debug info in `InstrProfiling.cpp`. Some compiler-generated functions, e.g., `__clang_call_terminate`, does not emit debug info and will fail a build if `-Werror` is used. This warning is not actionable by the user and I have not seen non-compiler-generated functions fail this test.
Reviewed By: smeenai
Differential Revision: https://reviews.llvm.org/D156006
Prior to this diff, names in the `__llvm_prf_names` section had the format `[<filepath>:]<function-name>`, e.g., `main.cpp:foo`, `bar`. `<filepath>` is used to discriminate between possibly identical function names when linkage is local and `<function-name>` simply comes from `F.getName()`. This has two problems:
* `:` is commonly found in Objective-C functions so that names like `main.mm:-[C foo::]` and `-[C bar::]` are difficult to parse
* `<function-name>` might be different from the linkage name, so it cannot be used to pass a function order to the linker via `-symbol-ordering-file` or `-order_file` (see https://discourse.llvm.org/t/rfc-temporal-profiling-extension-for-irpgo/68068)
Instead, this diff changes the format to `[<filepath>;]<linkage-name>`, e.g., `main.cpp;_foo`, `_bar`. The hope is that `;` won't realistically be found in either `<filepath>` or `<linkage-name>`.
To prevent invalidating all prior IRPGO profiles, we also lookup the prior name format when a record is not found (see `InstrProfSymtab::create()`, `readMemprof()`, and `getInstrProfRecord()`). It seems that Swift and Clang FE-PGO rely on the original `getPGOFuncName()`, so we cannot simply replace it.
Reviewed By: MaskRay
Differential Revision: https://reviews.llvm.org/D156569
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 makes the new `order` subcommand part of the help.
As a side effect, also make llvm::map_range compatible with plain
arrays.
Differential Revision: https://reviews.llvm.org/D153303
In [0] we described an algorithm called //BalancedPartitioning// (bp) to consume function traces [1] and compute a function order that reduces the number of page faults during startup.
This patch adds the `order` command to the `llvm-profdata` tool which uses bp to output a function order that can be passed to the linker via `--symbol-ordering-file=`.
Special thanks to Sergey Pupyrev and Julian Mestre for designing this balanced partitioning algorithm.
[0] https://discourse.llvm.org/t/rfc-temporal-profiling-extension-for-irpgo/68068
[1] https://reviews.llvm.org/D147287
Reviewed By: spupyrev
Differential Revision: https://reviews.llvm.org/D147812
The `llvm-profdata --version` output now looks like:
llvm-profdata
LLVM (http://llvm.org/):
LLVM version 17.0.0git
Optimized build with assertions.
This makes llvm-profdata more consistent with other tools.
Reviewed By: simon_tatham
Differential Revision: https://reviews.llvm.org/D150964
ExtBinary is compatible to, and more superior than Binary format, which is the current default output format. In the long run we are looking to only support ExtBinary format and Text format (for visual inspection), and drop Binary format as well. Since Binary format was the default, we expect many users are still using it, so let's change the default output format first, and hopefully the usage decreases over time
Reviewed By: davidxl, hoy
Differential Revision: https://reviews.llvm.org/D149700
When a user uses a mismatched clang + llvm-profdata, they didn't get a very
informative error message. It would just say "unsupported version".
As a result, users are often confused as to what they are supposed to do and
tend to assume that it's a bug in the profiling runtime.
This patch improves the error message by:
- Adding a new class of error (`raw_profile_version_mismatch`) to make it clear
that, specifically, the *raw profile* version is unsupported because of a
tool mismatch.
- Adding an error message that tells the user which raw profile version was
encountered, which version was expected, and instructs them to align their
tool versions.
To support this, this patch also updates `InstrProfError::take` to also
propagate the optional error message.
Differential Revision: https://reviews.llvm.org/D149361
As discussed in [0], add a `weight` field to temporal profiling traces found in profiles. This allows users to use the `--weighted-input=` flag in the `llvm-profdata merge` command to weight traces from different scenarios differently.
Note that this is a breaking change, but since [1] landed very recently and there is no way to "use" this trace data, there should be no users of this feature. We believe it is acceptable to land this change without bumping the profile format version.
[0] https://reviews.llvm.org/D147812#4259507
[1] https://reviews.llvm.org/D147287
Reviewed By: snehasish
Differential Revision: https://reviews.llvm.org/D148150
As described in [0], this extends IRPGO to support //Temporal Profiling//.
When `-pgo-temporal-instrumentation` is used we add the `llvm.instrprof.timestamp()` intrinsic to the entry of functions which in turn gets lowered to a call to the compiler-rt function `INSTR_PROF_PROFILE_SET_TIMESTAMP()`. A new field in the `llvm_prf_cnts` section stores each function's timestamp. Then in `llvm-profdata merge` we convert these function timestamps into a //trace// and add it to the indexed profile.
Since these traces could significantly increase the profile size, we've added `-max-temporal-profile-trace-length` and `-temporal-profile-trace-reservoir-size` to limit the length of a trace and the number of traces in a profile, respectively.
In a future diff we plan to use these traces to construct an optimized function order to reduce the number of page faults during startup.
Special thanks to Julian Mestre for helping with reservoir sampling.
[0] https://discourse.llvm.org/t/rfc-temporal-profiling-extension-for-irpgo/68068
Reviewed By: snehasish
Differential Revision: https://reviews.llvm.org/D147287
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
D139603 (add option to llvm-profdata to reduce output profile size) contains test cases that are not cross-platform. Moving those tests to unit test and making sure the feature is callable from llvm library
Reviewed By: snehasish
Differential Revision: https://reviews.llvm.org/D141446
D139603 (add option to llvm-profdata to reduce output profile size) contains test cases that are not cross-platform. Moving those tests to unit test and making sure the feature is callable from llvm library
Reviewed By: snehasish
Differential Revision: https://reviews.llvm.org/D141446
Make the access to profile data going through virtual file system so the
inputs can be remapped. In the context of the caching, it can make sure
we capture the inputs and provided an immutable input as profile data.
Reviewed By: akyrtzi, benlangmuir
Differential Revision: https://reviews.llvm.org/D139052
D139603 (add option to llvm-profdata to reduce output profile size) contains test cases that are not cross-platform. Moving those tests to unit test and making sure the feature is callable from llvm library
Reviewed By: snehasish
Differential Revision: https://reviews.llvm.org/D141446
Allow user to specify `--output-size-limit=n` to cap the size of generated profile to be strictly under n. Functions with the lowest total sample count are dropped first if necessary. Due to using a heuristic, excessive functions may be dropped to satisfy the size requirement
Reviewed By: snehasish
Differential Revision: https://reviews.llvm.org/D139603
This patch adds support for including binary ids in an indexed profile.
It adds a new field into the header that points to the offset of the
binary id section. The binary id section consists of a size of the
section, and a list of binary ids (if they are present) that consist
of two parts: length and data.
This patch guarantees that indexed profile is backwards compatible
after adding binary ids.
Differential Revision: https://reviews.llvm.org/D135929
This patch adds support for including binary ids in an indexed profile.
It adds a new field into the header that points to the offset of the
binary id section. The binary id section consists of a size of the
section, and a list of binary ids (if they are present) that consist
of two parts: length and data.
This patch guarantees that indexed profile is backwards compatible
after adding binary ids.
Differential Revision: https://reviews.llvm.org/D135929
When merging profiles with multiple threads, the `mergeWriterContexts()` function is used to merge profile data between writers. This must be in sync with `loadInput()` which merges profiles to a single writer. This diff merges the profile kind correctly in `mergeWriterContexts()` to fix a subtle bug.
Reviewed By: phosek
Differential Revision: https://reviews.llvm.org/D139755
Adding a switch to drop profile symbol list during merging AutoFDO profiles. This is needed to minimize the impact on default profiles when the profile symbol list is enabled for the source input profiles. The symbol list is quite large and could potentially slow down the compiler.
Reviewed By: davidxl, wenlei
Differential Revision: https://reviews.llvm.org/D139486
We need to flatten the SampleFDO profile in profile supplementation
because the InstrFDO profile does not have inlined callsite counters.
Without flattening profile, FDO optimizations are not stable:
we will not supplement the second generation profile when the modified
functions are all inlined.
This patch fixes this issue: we will flatten the profile for functions
that appears in FDO profile.
Note that we only need to find the hot/warm functions in SampleFDO
profile, so we will not perform a full flatten. We will use
a DFS traversal to compute the accumulated entry count and max bodycount.
This is much cheaper than full flattening.
Differential Revision: https://reviews.llvm.org/D138893
In https://reviews.llvm.org/D135127 we created the show flag
`--output-format` which was confusing because it behaved differently
than the same flag in the merge command. So, rename the flag to
`--show-format`. This also allows us to add the `text` option to mean
"normal text output" rather than "text-encoded profiles" like it does
for the merge command.
Reviewed By: wenlei
Differential Revision: https://reviews.llvm.org/D135467
Change the behavior of the `llvm-profdata show --debug-info=` command to dump a YAML file when using debug info correlation since it provides more information in a parseable format.
Reviewed By: yozhu, phosek
Differential Revision: https://reviews.llvm.org/D134770
Add `--output-format` option for the `llvm-profdata show` command to select the type of output. The existing `--text` flag is used to emit text encoded profiles. To avoid confusion, `--output-format=text-encoding` indicates that the output will be profiles encoded in the text format, and `--output-format=text` indicates the default text output that doesn't necessarily represent a profile.
`--output-format=json` is an alias for `--json` and `--output-format=yaml` will be used in D134770.
Reviewed By: phosek
Differential Revision: https://reviews.llvm.org/D135127
The llvm-driver, enabled with LLVM_TOOL_LLVM_DRIVER_BUILD combines many llvm executables
into one to save overall toolchain size. This patch adds a few more llvm tools to the
llvm-driver.
Differential Revision: https://reviews.llvm.org/D135281
Interestingly, MathExtras.h doesn't use <cmath> declaration, so move it out of
that header and include it when needed.
No functional change intended, but there's no longer a transitive include
fromMathExtras.h to cmath.
Current implementation promotes a non-cold function in the SampleFDO profile
into a hot function in the FDO profile. This is too aggressive. This patch
promotes a hot functions in the SampleFDO profile into a hot function, and a
warm function in SampleFDO into a warm function in FDO.
Differential Revision: https://reviews.llvm.org/D132601
This patch has the following changes:
(1) Handling of internal linkage functions (static functions)
Static functions in FDO have a prefix of source file name, while they do not
have one in SampleFDO. Current implementation does not handle this and we are
not updating the profile for static functions. This patch fixes this.
(2) Handling of -funique-internal-linakge-symbols
Again this is for the internal linkage functions. Option
-funique-internal-linakge-symbols can now be applied to both FDO and SampleFDO
compilation. When it is used, it demangles internal linkage function names and
adds a hash value as the postfix.
When both SampleFDO and FDO profiles use this option, or both
not use this option, changes in (1) should handle this.
Here we also handle when the SampleFDO profile using this option while FDO
profile not using this option, or vice versa.
There is one case where this patch won't work: If one of the profiles used
mangled name and the other does not. For example, if the SampleFDO profile
uses clang c-compiler and without -funique-internal-linakge-symbols, while
the FDO profile uses -funique-internal-linakge-symbols. The SampleFDO profile
contains unmangled names while the FDO profile contains mangled names. If
both profiles use c++ compiler, this won't happen. We think this use case
is rare and does not justify the effort to fix.
Differential Revision: https://reviews.llvm.org/D132600
1) We now use the count size in FDO as the main factor to deal with pre-inliner.
Currently we use the number of sample records in the SampleFDO profile. But
that only counts the top-level body sample records (not including the nested
call-sites). We are seeing some big functions not being updated because of
this. I think using the count size in FDO profile is more reasonable to judge if
the function is likely to be inlined to the callers in pre-inliner.
(2) We use getMaxCount in SampleFDO rather the HeadSample to determine if
if the function is hot in SampleFDO. This is in-sync with the logic
in the compiler (also HeadSample can be 0).
Differential Revision: https://reviews.llvm.org/D132602
