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
Add a check to detect that the profiled binary was build with position
independent code. Add a test with a pie binary to which can be reused
later when support is added. Also clean up the error messages with
trailing colons.
Reviewed By: tejohnson
Differential Revision: https://reviews.llvm.org/D128564
To ease profile annotation, each of the callsites in a function can be
annotated with profile data - "IR metadata format for MemProf" [1]. This
patch extends the on-disk serialized record format to store the debug
information for allocation callsites incl inline frames. This change is
incompatible with the existing format i.e. indexed profiles must be
regenerated, raw profiles are unaffected.
[1] https://groups.google.com/g/llvm-dev/c/aWHsdMxKAfE/m/WtEmRqyhAgAJ
Reviewed By: tejohnson
Differential Revision: https://reviews.llvm.org/D121179
We add to ensure that we are observing the correct callstack order in memprof
during symbolization. There was some confusion whether the order of
DIFrame objects were reversed but in reality the leaf function is at
index 0 so no code changes are required.
Differential Revision: https://reviews.llvm.org/D121759
This patch adds support for optional memory profile information to be
included with and indexed profile. The indexed profile header adds a new
field which points to the offset of the memory profile section (if
present) in the indexed profile. For users who do not utilize this
feature the only overhead is a 64-bit offset in the header.
The memory profile section contains (1) profile metadata describing the
information recorded for each entry (2) an on-disk hashtable containing
the profile records indexed via llvm::md5(function_name). We chose to
introduce a separate hash table instead of the existing one since the
indexing for the instrumented fdo hash table is based on a CFG hash
which itself is perturbed by memprof instrumentation.
This commit also includes the changes reviewed separately in D120093.
Differential Revision: https://reviews.llvm.org/D120103
This reverts commit 85355a560a.
This patch adds support for optional memory profile information to be
included with and indexed profile. The indexed profile header adds a new
field which points to the offset of the memory profile section (if
present) in the indexed profile. For users who do not utilize this
feature the only overhead is a 64-bit offset in the header.
The memory profile section contains (1) profile metadata describing the
information recorded for each entry (2) an on-disk hashtable containing
the profile records indexed via llvm::md5(function_name). We chose to
introduce a separate hash table instead of the existing one since the
indexing for the instrumented fdo hash table is based on a CFG hash
which itself is perturbed by memprof instrumentation.
Differential Revision: https://reviews.llvm.org/D118653
This reverts commit 0f73fb18ca.
Use llvm/Profile/MIBEntryDef.inc instead of relative path.
Generated the raw profile data with `-mllvm
-enable-name-compression=false` so that builbots where the reader is
built without zlib do not fail.
Also updated the test build instructions.
This patch adds support for optional memory profile information to be
included with and indexed profile. The indexed profile header adds a new
field which points to the offset of the memory profile section (if
present) in the indexed profile. For users who do not utilize this
feature the only overhead is a 64-bit offset in the header.
The memory profile section contains (1) profile metadata describing the
information recorded for each entry (2) an on-disk hashtable containing
the profile records indexed via llvm::md5(function_name). We chose to
introduce a separate hash table instead of the existing one since the
indexing for the instrumented fdo hash table is based on a CFG hash
which itself is perturbed by memprof instrumentation.
Differential Revision: https://reviews.llvm.org/D118653
This change extends the RawMemProfReader to read all the sections of the
raw profile and symbolize the virtual addresses recorded as part of the
callstack for each allocation. For now the symbolization is used to
display the contents of the profile with llvm-profdata.
Differential Revision: https://reviews.llvm.org/D116784
Use the llvm flag `-pgo-function-entry-coverage` to create single byte "counters" to track functions coverage. This mode has significantly less size overhead in both code and data because
* We mark a function as "covered" with a store instead of an increment which generally requires fewer assembly instructions
* We use a single byte per function rather than 8 bytes per block
The trade off of course is that this mode only tells you if a function has been covered. This is useful, for example, to detect dead code.
When combined with debug info correlation [0] we are able to create an instrumented Clang binary that is only 150M (the vanilla Clang binary is 143M). That is an overhead of 7M (4.9%) compared to the default instrumentation (without value profiling) which has an overhead of 31M (21.7%).
[0] https://groups.google.com/g/llvm-dev/c/r03Z6JoN7d4
Reviewed By: kyulee
Differential Revision: https://reviews.llvm.org/D116180
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
The first 8b of each raw profile section need to be aligned to 8b since
the first item in each section is a u64 count of the number of items in
the section.
Summary of changes:
* Assert alignment when reading counts.
* Update test to check alignment, relax some size checks to allow padding.
* Update raw binary inputs for llvm-profdata tests.
Differential Revision: https://reviews.llvm.org/D114826
This commit adds initial support to llvm-profdata to read and print
summaries of raw memprof profiles.
Summary of changes:
* Refactor shared defs to MemProfData.inc
* Extend show_main to display memprof profile summaries.
* Add a simple raw memprof profile reader.
* Add a couple of tests to tools/llvm-profdata.
Differential Revision: https://reviews.llvm.org/D114286
This is to account for the change that made CountersPtr in __profd_
relative which landed in a1532ed275.
That change hasn't updated the raw profile version, and while the
profile layout stayed the same, profiles generated by tip-of-tree
LLVM are incompatible with 13.x tooling.
Differential Revision: https://reviews.llvm.org/D111123
format.
Currently when we add a new section in the profile format and generate a profile
containing the new section, older compiler which reads the new profile will
issue an error. The forward incompatibility can cause unnecessary churn when
extending the profile. This patch removes the incompatibility when adding a new
section for extbinary format.
Differential Revision: https://reviews.llvm.org/D109398
On some platform (eg: AIX), diff will complain about newline.
diff: Missing newline at the end of file
.../llvm/test/tools/llvm-profdata/Inputs/cs-sample.proftext.
Currently context strings contain a lot of duplicated function names and that significantly increase the profile size. This change split the context into a series of {name, offset, discriminator} tuples so function names used in the context can be replaced by the index into the name table and that significantly reduce the size consumed by context.
A follow-up improvement made in the compiler and profiling tools is to avoid reconstructing full context strings which is time- and memory- consuming. Instead a context vector of `StringRef` is adopted to represent the full context in all scenarios. As a result, the previous prevalent profile map which was implemented as a `StringRef` is now engineered as an unordered map keyed by `SampleContext`. `SampleContext` is reshaped to using an `ArrayRef` to represent a full context for CS profile. For non-CS profile, it falls back to use `StringRef` to represent a contextless function name. Both the `ArrayRef` and `StringRef` objects are underpinned by real array and string objects that are stored in producer buffers. For compiler, they are maintained by the sample reader. For llvm-profgen, they are maintained in `ProfiledBinary` and `ProfileGenerator`. Full context strings can be generated only in those cases of debugging and printing.
When it comes to profile format, nothing has changed to the text format, though internally CS context is implemented as a vector. Extbinary format is only changed for CS profile, with an additional `SecCSNameTable` section which stores all full contexts logically in the form of `vector<int>`, which each element as an offset points to `SecNameTable`. All occurrences of contexts elsewhere are redirected to using the offset of `SecCSNameTable`.
Testing
This is no-diff change in terms of code quality and profile content (for text profile).
For our internal large service (aka ads), the profile generation is cut to half, with a 20x smaller string-based extbinary format generated.
The compile time of ads is dropped by 25%.
Differential Revision: https://reviews.llvm.org/D107299
Sample profiles are stored in a string map which is basically an unordered map. Printing out profiles by simply walking the string map doesn't enforce an order. I'm sorting the map in the decreasing order of total samples to enable a more stable dump, which is good for comparing two dumps.
Reviewed By: wenlei, wlei
Differential Revision: https://reviews.llvm.org/D108147
This fixes support for merging profiles which broke as a consequence
of e50a38840d. The issue was missing
adjustment in merge logic to account for the binary IDs which are
now included in the raw profile just after header.
In addition, this change also:
* Includes the version in module signature that's used for merging
to avoid accidental attempts to merge incompatible profiles.
* Moves the binary IDs size field after version field in the header
as was suggested in the review.
Differential Revision: https://reviews.llvm.org/D107143
This fixes support for merging profiles which broke as a consequence
of e50a38840d. The issue was missing
adjustment in merge logic to account for the binary IDs which are
now included in the raw profile just after header.
In addition, this change also:
* Includes the version in module signature that's used for merging
to avoid accidental attempts to merge incompatible profiles.
* Moves the binary IDs size field after version field in the header
as was suggested in the review.
Differential Revision: https://reviews.llvm.org/D107143
Change `CountersPtr` in `__profd_` to a label difference, which is a link-time
constant. On ELF, when linking a shared object, this requires that `__profc_` is
either private or linkonce/linkonce_odr hidden. On COFF, we need D104564 so that
`.quad a-b` (64-bit label difference) can lower to a 32-bit PC-relative relocation.
```
# ELF: R_X86_64_PC64 (PC-relative)
.quad .L__profc_foo-.L__profd_foo
# Mach-O: a pair of 8-byte X86_64_RELOC_UNSIGNED and X86_64_RELOC_SUBTRACTOR
.quad l___profc_foo-l___profd_foo
# COFF: we actually use IMAGE_REL_AMD64_REL32/IMAGE_REL_ARM64_REL32 so
# the high 32-bit value is zero even if .L__profc_foo < .L__profd_foo
# As compensation, we truncate CountersDelta in the header so that
# __llvm_profile_merge_from_buffer and llvm-profdata reader keep working.
.quad .L__profc_foo-.L__profd_foo
```
(Note: link.exe sorts `.lprfc` before `.lprfd` even if the object writer
has `.lprfd` before `.lprfc`, so we cannot work around by reordering
`.lprfc` and `.lprfd`.)
With this change, a stage 2 (`-DLLVM_TARGETS_TO_BUILD=X86 -DLLVM_BUILD_INSTRUMENTED=IR`)
`ld -pie` linked clang is 1.74% smaller due to fewer R_X86_64_RELATIVE relocations.
```
% readelf -r pie | awk '$3~/R.*/{s[$3]++} END {for (k in s) print k, s[k]}'
R_X86_64_JUMP_SLO 331
R_X86_64_TPOFF64 2
R_X86_64_RELATIVE 476059 # was: 607712
R_X86_64_64 2616
R_X86_64_GLOB_DAT 31
```
The absolute function address (used by llvm-profdata to collect indirect call
targets) can be converted to relative as well, but is not done in this patch.
Differential Revision: https://reviews.llvm.org/D104556
This patch was split from https://reviews.llvm.org/D102246
[SampleFDO] New hierarchical discriminator for Flow Sensitive SampleFDO
This is for llvm-profdata part of change. It sets the bit masks for the
profile reader in llvm-profdata. Also add an internal option
"-fs-discriminator-pass" for show and merge command to process the profile
offline.
This patch also moved setDiscriminatorMaskedBitFrom() to
SampleProfileReader::create() to simplify the interface.
Differential Revision: https://reviews.llvm.org/D103550
Encountered a crash while running a debug build, where this code path would be taken due to a mismatch in profile coverage data versions. Without consuming the error, an assert would be triggered inside the destructor of Error.
Differential Revision: https://reviews.llvm.org/D99457
This changes adds attribute field for metadata of context profile. Currently we have an inline attribute that indicates whether the leaf frame corresponding to a context profile was inlined in previous build.
This will be used to help estimating inlining and be taken into account when trimming context. Changes for that in llvm-profgen will follow. It will also help tuning.
Differential Revision: https://reviews.llvm.org/D98823
Dangling probes are the probes associated to an empty block. This usually happens when all real instructions are optimized away from the block. There is a problem with dangling probes during the offline counts processing. The way the sample profiler works is that samples collected on the first physical instruction following a probe will be counted towards the probe. This logically equals to treating the instruction next to a probe as if it is from the same block of the probe. In the dangling probe case, the real instruction following a dangling probe actually starts a new block, and samples collected on the new block may cause issues when counted towards the empty block.
To mitigate this issue, we first try to move around a dangling probe inside its owning block. If there are still native instructions preceding the probe in the same block, we can then use them as a place holder to collect samples for the probe. A pass is added to walk each block backwards looking for probes not followed by any real instruction and moving them before the first real instruction. This is done right before the object emission.
If we are unlucky to find such in-block preceding instructions for a probe, the solution we are taking is to tag such probe as dangling so that the samples reported for them will not be trusted by the compiler. We leave it up to the counts inference algorithm to get such probes a reasonable count. The number `UINT64_MAX` is used to mark sample count as collected for a dangling probe.
Reviewed By: wmi
Differential Revision: https://reviews.llvm.org/D95962
This change brings up support of context-sensitive profiles in the format of extended binary. Existing sample profile reader/writer/merger code is being tweaked to reflect the fact of bracketed input contexts, like (`[...]`). The paired brackets are also needed in extbinary profiles because we don't yet have an otherwise good way to tell calling contexts apart from regular function names since the context delimiter `@` can somehow serve as a part of the C++ mangled names.
Reviewed By: wmi, wenlei
Differential Revision: https://reviews.llvm.org/D95547
This change enables pseudo-probe-based sample counts to be consumed by the sample profile loader under the regular `-fprofile-sample-use` switch with minimal adjustments to the existing sample file formats. After the counts are imported, a probe helper, aka, a `PseudoProbeManager` object, is automatically launched to verify the CFG checksum of every function in the current compilation against the corresponding checksum from the profile. Mismatched checksums will cause a function profile to be slipped. A `SampleProfileProber` pass is scheduled before any of the `SampleProfileLoader` instances so that the CFG checksums as well as probe mappings are available during the profile loading time. The `PseudoProbeManager` object is set up right after the profile reading is done. In the future a CFG-based fuzzy matching could be done in `PseudoProbeManager`.
Samples will be applied only to pseudo probe instructions as well as probed callsites once the checksum verification goes through. Those instructions are processed in the same way that regular instructions would be processed in the line-number-based scenario. In other words, a function is processed in a regular way as if it was reduced to just containing pseudo probes (block probes and callsites).
**Adjustment to profile format **
A CFG checksum field is being added to the existing AutoFDO profile formats. So far only the text format and the extended binary format are supported. For the text format, a new line like
```
!CFGChecksum: 12345
```
is added to the end of the body sample lines. For the extended binary profile format, we introduce a metadata section to store the checksum map from function names to their CFG checksums.
Differential Revision: https://reviews.llvm.org/D92347
Implemented the `llvm-profdata overlap` feature for sample profiles. It reports weighted //similarity// and unweighted //overlap// metrics at program and function level for two input profiles. Similarity metrics are symmetric with regards to the order of two input profiles. By default, the tool only reports program-level summary. Users can look into function-level details via additional options `--function`, `--similarity-cutoff`, and `--value-cutoff`.
The similarity metrics are designed as follows:
* Program-level summary
* Whole program profile similarity is an aggregate over function-level similarity `FS`: `PS = sum(FS(A) * avg_weight(A))` for all function `A`.
* Whole program sample overlap: `PSO = common_samples / total_samples`.
* Function overlap: `FO = #common_function / #total_function`.
* Hot-function overlap: `HFO = #common_hot_function / #total_hot_function`.
* Hot-block overlap: `HBO = #common_hot_block / #total_hot_block`.
* Function-level details
* Function-level similarity is an aggregate over line/block-level similarities `BS` of all sample lines/blocks in the function, weighted by the closeness of the function's weights in two profiles: `FS = sum(BS(i)) * (1 - weight_distance(A))`.
* Function-level sample overlap: `FSO = common_samples / total_samples` for samples in the function.
Reviewed By: wenlei, hoyFB, wmi
Differential Revision: https://reviews.llvm.org/D83852
PGO profile is usually more precise than sample profile. However, PGO profile
needs to be collected from loadtest and loadtest may not be representative
enough to the production workload. Sample profile collected from production
can be used as a supplement -- for functions cold in loadtest but warm/hot
in production, we can scale up the related function in PGO profile if the
function is warm or hot in sample profile.
The implementation contains changes in compiler side and llvm-profdata side.
Given an instr profile and a sample profile, for a function cold in PGO
profile but warm/hot in sample profile, llvm-profdata will either mark
all the counters in the profile to be -1 or scale up the max count in the
function to be above hot threshold, depending on the zero counter ratio in
the profile. The assumption is if there are too many counters being zero
in the function profile, the profile is more likely to cause harm than good,
then llvm-profdata will mark all the counters to be -1 indicating the
function is hot but the profile is unaccountable. In compiler side, if a
function profile with all -1 counters is seen, the function entry count will
be set to be above hot threshold but its internal profile will be dropped.
In the long run, it may be useful to let compiler support using PGO profile
and sample profile at the same time, but that requires more careful design
and more substantial changes to make two profiles work seamlessly. The patch
here serves as a simple intermediate solution.
Differential Revision: https://reviews.llvm.org/D81981
This patch includes the supporting code that enables always
instrumenting the function entry block by default.
This patch will NOT the default behavior.
It adds a variant bit in the profile version, adds new directives in
text profile format, and changes llvm-profdata tool accordingly.
This patch is a split of D83024 (https://reviews.llvm.org/D83024)
Many test changes from D83024 are also included.
Differential Revision: https://reviews.llvm.org/D84261
Summary:
Add the --hot-func-list feature to llvm-profdata show for sample profiles. This feature prints a list of hot functions whose max sample count are above the 99% threshold, with their numbers of total samples, total samples percentage, max samples, entry samples, and their function names.
Test Plan:
Reviewers: wenlei, hoyFB
Reviewed By: wenlei, hoyFB
Subscribers: hoyFB, wenlei, weihe, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D82355
Summary: Add the --hot-func-list feature to llvm-profdata show for sample profiles. This feature prints a list of hot functions whose max sample count are above the 99% threshold, with their numbers of total samples, total samples percentage, max samples, entry samples, and their function names.
Reviewers: wmi, hoyFB, wenlei
Reviewed By: wmi
Subscribers: hoyFB, wenlei, llvm-commits, weihe
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D81800
Add support for continuously syncing profile counter updates to a file.
The motivation for this is that programs do not always exit cleanly. On
iOS, for example, programs are usually killed via a signal from the OS.
Running atexit() handlers after catching a signal is unreliable, so some
method for progressively writing out profile data is necessary.
The approach taken here is to mmap() the `__llvm_prf_cnts` section onto
a raw profile. To do this, the linker must page-align the counter and
data sections, and the runtime must ensure that counters are mapped to a
page-aligned offset within a raw profile.
Continuous mode is (for the moment) incompatible with the online merging
mode. This limitation is lifted in https://reviews.llvm.org/D69586.
Continuous mode is also (for the moment) incompatible with value
profiling, as I'm not sure whether there is interest in this and the
implementation may be tricky.
As I have not been able to test extensively on non-Darwin platforms,
only Darwin support is included for the moment. However, continuous mode
may "just work" without modification on Linux and some UNIX-likes. AIUI
the default value for the GNU linker's `--section-alignment` flag is set
to the page size on many systems. This appears to be true for LLD as
well, as its `no_nmagic` option is on by default. Continuous mode will
not "just work" on Fuchsia or Windows, as it's not possible to mmap() a
section on these platforms. There is a proposal to add a layer of
indirection to the profile instrumentation to support these platforms.
rdar://54210980
Differential Revision: https://reviews.llvm.org/D68351
Rong Xu