The priority-based inliner currenlty uses block count combined with callee entry count to drive callsite inlining. This doesn't work well with LTO where postlink inlining is driven by prelink-annotated block count which could be based on the merge of all context profiles. I'm fixing it by using callee profile entry count only which should be context-sensitive.
I'm seeing 0.2% perf improvment for one of our internal large benchmarks with probe-based non-CS profile.
Reviewed By: wenlei
Differential Revision: https://reviews.llvm.org/D120784
Trimming whitespace or carriage returns from symbols allows this code
to work on Windows and makes it match other places symbol lists are
handled.
Reviewed By: MaskRay
Differential Revision: https://reviews.llvm.org/D117570
This is an extension of **profi** post-processing step that rebalances counts
in CFGs that have basic blocks w/o probes (aka "unknown" blocks). Specifically,
the new version finds many more "unknown" subgraphs and marks more "unknown"
basic blocks as hot (which prevents unwanted optimization passes).
I see up to 0.5% perf on some (large) binaries, e.g., clang-10 and gcc-8.
The algorithm is still linear and yields no build time overhead.
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
This is a continuation of D109860 and D109903.
An important challenge for profile inference is caused by the fact that the
sample profile is collected on a fully optimized binary, while the block and
edge frequencies are consumed on an early stage of the compilation that operates
with a non-optimized IR. As a result, some of the basic blocks may not have
associated sample counts, and it is up to the algorithm to deduce missing
frequencies. The problem is illustrated in the figure where three basic
blocks are not present in the optimized binary and hence, receive no samples
during profiling.
We found that it is beneficial to treat all such blocks equally. Otherwise the
compiler may decide that some blocks are “cold” and apply undesirable
optimizations (e.g., hot-cold splitting) regressing the performance. Therefore,
we want to distribute the counts evenly along the blocks with missing samples.
This is achieved by a post-processing step that identifies "dangling" subgraphs
consisting of basic blocks with no sampled counts; once the subgraphs are
found, we rebalance the flow so as every branch probability is 50:50 within the
subgraphs.
Our experiments indicate up to 1% performance win using the optimization on
some binaries and a significant improvement in the quality of profile counts
(when compared to ground-truth instrumentation-based counts)
{F19093045}
Reviewed By: hoy
Differential Revision: https://reviews.llvm.org/D109980
This is a continuation of D109860.
Traditional flow-based algorithms cannot guarantee that the resulting edge
frequencies correspond to a *connected* flow in the control-flow graph. For
example, for an instance in the attached figure, a flow-based (or any other)
inference algorithm may produce an output in which the hot loop is disconnected
from the entry block (refer to the rightmost graph in the figure). Furthermore,
creating a connected minimum-cost maximum flow is a computationally NP-hard
problem. Hence, we apply a post-processing adjustments to the computed flow
by connecting all isolated flow components ("islands").
This feature helps to keep all blocks with sample counts connected and results
in significant performance wins for some binaries.
{F19077343}
Reviewed By: hoy
Differential Revision: https://reviews.llvm.org/D109903
The benefits of sampling-based PGO crucially depends on the quality of profile
data. This diff implements a flow-based algorithm, called profi, that helps to
overcome the inaccuracies in a profile after it is collected.
Profi is an extended and significantly re-engineered classic MCMF (min-cost
max-flow) approach suggested by Levin, Newman, and Haber [2008, Complementing
missing and inaccurate profiling using a minimum cost circulation algorithm]. It
models profile inference as an optimization problem on a control-flow graph with
the objectives and constraints capturing the desired properties of profile data.
Three important challenges that are being solved by profi:
- "fixing" errors in profiles caused by sampling;
- converting basic block counts to edge frequencies (branch probabilities);
- dealing with "dangling" blocks having no samples in the profile.
The main implementation (and required docs) are in SampleProfileInference.cpp.
The worst-time complexity is quadratic in the number of blocks in a function,
O(|V|^2). However a careful engineering and extensive evaluation shows that
the running time is (slightly) super-linear. In particular, instances with
1000 blocks are solved within 0.1 second.
The algorithm has been extensively tested internally on prod workloads,
significantly improving the quality of generated profile data and providing
speedups in the range from 0% to 5%. For "smaller" benchmarks (SPEC06/17), it
generally improves the performance (with a few outliers) but extra work in
the compiler might be needed to re-tune existing optimization passes relying on
profile counts.
UPD Dec 1st 2021:
- synced the declaration and definition of the option `SampleProfileUseProfi ` to use type `cl::opt<bool`;
- added `inline` for `SampleProfileInference<BT>::findUnlikelyJumps` and `SampleProfileInference<BT>::isExit` to avoid linking problems on windows.
Reviewed By: wenlei, hoy
Differential Revision: https://reviews.llvm.org/D109860
For nodes that are in a cycle of a profiled call graph, the current order the underlying scc_iter computes purely depends on how those nodes are reached from outside the SCC and inside the SCC, based on the Tarjan algorithm. This does not honor profile edge hotness, thus does not gurantee hot callsites to be inlined prior to cold callsites. To mitigate that, I'm adding an extra sorter on top of scc_iter to sort scc functions in the order of callsite hotness, instead of changing the internal of scc_iter.
Sorting on callsite hotness can be optimally based on detecting cycles on a directed call graph, i.e, to remove the coldest edge until a cycle is broken. However, detecting cycles isn't cheap. I'm using an MST-based approach which is faster and appear to deliver some performance wins.
Reviewed By: wenlei
Differential Revision: https://reviews.llvm.org/D114204
The benefits of sampling-based PGO crucially depends on the quality of profile
data. This diff implements a flow-based algorithm, called profi, that helps to
overcome the inaccuracies in a profile after it is collected.
Profi is an extended and significantly re-engineered classic MCMF (min-cost
max-flow) approach suggested by Levin, Newman, and Haber [2008, Complementing
missing and inaccurate profiling using a minimum cost circulation algorithm]. It
models profile inference as an optimization problem on a control-flow graph with
the objectives and constraints capturing the desired properties of profile data.
Three important challenges that are being solved by profi:
- "fixing" errors in profiles caused by sampling;
- converting basic block counts to edge frequencies (branch probabilities);
- dealing with "dangling" blocks having no samples in the profile.
The main implementation (and required docs) are in SampleProfileInference.cpp.
The worst-time complexity is quadratic in the number of blocks in a function,
O(|V|^2). However a careful engineering and extensive evaluation shows that
the running time is (slightly) super-linear. In particular, instances with
1000 blocks are solved within 0.1 second.
The algorithm has been extensively tested internally on prod workloads,
significantly improving the quality of generated profile data and providing
speedups in the range from 0% to 5%. For "smaller" benchmarks (SPEC06/17), it
generally improves the performance (with a few outliers) but extra work in
the compiler might be needed to re-tune existing optimization passes relying on
profile counts.
Reviewed By: wenlei, hoy
Differential Revision: https://reviews.llvm.org/D109860
The benefits of sampling-based PGO crucially depends on the quality of profile
data. This diff implements a flow-based algorithm, called profi, that helps to
overcome the inaccuracies in a profile after it is collected.
Profi is an extended and significantly re-engineered classic MCMF (min-cost
max-flow) approach suggested by Levin, Newman, and Haber [2008, Complementing
missing and inaccurate profiling using a minimum cost circulation algorithm]. It
models profile inference as an optimization problem on a control-flow graph with
the objectives and constraints capturing the desired properties of profile data.
Three important challenges that are being solved by profi:
- "fixing" errors in profiles caused by sampling;
- converting basic block counts to edge frequencies (branch probabilities);
- dealing with "dangling" blocks having no samples in the profile.
The main implementation (and required docs) are in SampleProfileInference.cpp.
The worst-time complexity is quadratic in the number of blocks in a function,
O(|V|^2). However a careful engineering and extensive evaluation shows that
the running time is (slightly) super-linear. In particular, instances with
1000 blocks are solved within 0.1 second.
The algorithm has been extensively tested internally on prod workloads,
significantly improving the quality of generated profile data and providing
speedups in the range from 0% to 5%. For "smaller" benchmarks (SPEC06/17), it
generally improves the performance (with a few outliers) but extra work in
the compiler might be needed to re-tune existing optimization passes relying on
profile counts.
Reviewed By: wenlei, hoy
Differential Revision: https://reviews.llvm.org/D109860
Adds the following switches:
1. --sample-profile-inline-replay-fallback/--cgscc-inline-replay-fallback: controls what the replay advisor does for inline sites that are not present in the replay. Options are:
1. Original: defers to original advisor
2. AlwaysInline: inline all sites not in replay
3. NeverInline: inline no sites not in replay
2. --sample-profile-inline-replay-format/--cgscc-inline-replay-format: controls what format should be generated to match against the replay remarks. Options are:
1. Line
2. LineColumn
3. LineDiscriminator
4. LineColumnDiscriminator
Adds support for negative inlining decisions. These are denoted by "will not be inlined into" as compared to the positive "inlined into" in the remarks.
All of these together with the previous `--sample-profile-inline-replay-scope/--cgscc-inline-replay-scope` allow tweaking in how to apply replay. In my testing, I'm using:
1. --sample-profile-inline-replay-scope/--cgscc-inline-replay-scope = Function to only replay on a function
2. --sample-profile-inline-replay-fallback/--cgscc-inline-replay-fallback = NeverInline since I'm feeding in only positive remarks to the replay system
3. --sample-profile-inline-replay-format/--cgscc-inline-replay-format = Line since I'm generating the remarks from DWARF information from GCC which can conflict quite heavily in column number compared to Clang
An alternative configuration could be to do Function, AlwaysInline, Line fallback with negative remarks which closer matches the final call-sites. Note that this can lead to unbounded inlining if a negative remark doesn't match/exist for one reason or another.
Updated various tests to cover the new switches and negative remarks
Testing:
ninja check-all
Reviewed By: wenlei, mtrofin
Differential Revision: https://reviews.llvm.org/D112040
Replay in sample profiling needs to be asked on candidates that may not have counts or below the threshold. If replay is in effect for a function make sure these are captured and also imported during thinLTO.
Testing:
ninja check-all
Reviewed By: wenlei
Differential Revision: https://reviews.llvm.org/D112033
The goal is to allow grafting an inline tree from Clang or GCC into a new compilation without affecting other functions. For GCC, we're doing this by extracting the inline tree from dwarf information and generating the equivalent remarks.
This allows easier side-by-side asm analysis and a trial way to see if a particular inlining setup provides benefits by itself.
Testing:
ninja check-all
Reviewed By: wenlei, mtrofin
Differential Revision: https://reviews.llvm.org/D110658
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
The change adds a switch to allow sample loader to use global pre-inliner's decision instead. The pre-inliner in llvm-profgen makes inline decision globally based on whole program profile and function byte size as cost proxy.
Since pre-inliner also adjusts/merges context profile based on its inline decision, honoring its inline decision in sample loader would lead to better post-inline profile quality especially for thinlto where cross module profile merging isn't possible without pre-inliner.
Minor fix in profile reader is also included. When pre-inliner is use, we now also turn off the default merging and trimming logic unless it's explicitly asked.
Differential Revision: https://reviews.llvm.org/D108677
Clang diagnostics refer to identifier names in quotes.
This patch makes inline remarks conform to the convention.
New behavior:
```
% clang -O2 -Rpass=inline -Rpass-missed=inline -S a.c
a.c:4:25: remark: 'foo' inlined into 'bar' with (cost=-30, threshold=337) at callsite bar:0:25; [-Rpass=inline]
int bar(int a) { return foo(a); }
^
```
Reviewed By: hoy
Differential Revision: https://reviews.llvm.org/D107791
This patch teaches the sample profile loader to merge function
attributes after inlining functions.
Without this patch, the compiler could inline a function requiring the
512-bit vector width into its caller without merging function
attributes, triggering a failure during instruction selection.
Differential Revision: https://reviews.llvm.org/D105729
Previously dangling samples were represented by INT64_MAX in sample profile while probes never executed were not reported. This was based on an observation that dangling probes were only at a smaller portion than zero-count probes. However, with compiler optimizations, dangling probes end up becoming at large portion of all probes in general and reporting them does not make sense from profile size point of view. This change flips sample reporting by reporting zero-count probes instead. This enabled dangling probe to be represented by none (missing entry in profile). This has a couple benefits:
1. Reducing sample profile size in optimize mode, even when the number of non-executed probes outperform the number of dangling probes, since INT64_MAX takes more space over 0 to encode.
2. Binary size savings. No need to encode dangling probe anymore, since missing probes are treated as dangling in the profile reader.
3. Reducing compiler work to track dangling probes. However, for probes that are real dead and removed, we still need the compiler to identify them so that they can be reported as zero-count, instead of mistreated as dangling probes.
4. Improving counts quality by respecting the counts already collected on the non-dangling copy of a probe. A probe, when duplicated, gets two copies at runtime. If one of them is dangling while the other is not, merging the two probes at profile generation time will cause the real samples collected on the non-dangling one to be discarded. Not reporting the dangling counterpart will keep the real samples.
5. Better readability.
6. Be consistent with non-CS dwarf line number based profile. Zero counts are trusted by the compiler counts inferencer while missing counts will be inferred by the compiler.
Note that the current patch does include any work for #3. There will be follow-up changes.
For #1, I've seen for a large Facebook service, the text profile is reduced by 7%. For extbinary profile, the size of LBRProfileSection is reduced by 35%.
For #4, I have seen general counts quality for SPEC2017 is improved by 10%.
Reviewed By: wenlei, wlei, wmi
Differential Revision: https://reviews.llvm.org/D104129
The current implementation for computing relative block frequencies does
not handle correctly control-flow graphs containing irreducible loops. This
results in suboptimally generated binaries, whose perf can be up to 5%
worse than optimal.
To resolve the problem, we apply a post-processing step, which iteratively
updates block frequencies based on the frequencies of their predesessors.
This corresponds to finding the stationary point of the Markov chain by
an iterative method aka "PageRank computation". The algorithm takes at
most O(|E| * IterativeBFIMaxIterations) steps but typically converges faster.
It is turned on by passing option `use-iterative-bfi-inference`
and applied only for functions containing profile data and irreducible loops.
Tested on SPEC06/17, where it is helping to get correct profile counts for one of
the binaries (403.gcc). In prod binaries, we've seen a speedup of up to 2%-5%
for binaries containing functions with hot irreducible loops.
Reviewed By: hoy, wenlei, davidxl
Differential Revision: https://reviews.llvm.org/D103289
This patch was split from https://reviews.llvm.org/D102246
[SampleFDO] New hierarchical discriminator for Flow Sensitive SampleFDO
This is mainly for ProfileData part of change. It will load
FS Profile when such profile is detected. For an extbinary format profile,
create_llvm_prof tool will add a flag to profile summary section.
For other format profiles, the users need to use an internal option
(-profile-isfs) to tell the compiler that the profile uses FS discriminators.
This patch also simplified the bit API used by FS discriminators.
Differential Revision: https://reviews.llvm.org/D103041
We see a regression related to low probe factor(0.01) which prevents some callsites being promoted in ICPPass and later cause the missing inline in CGSCC inliner. The root cause is due to redundant(the second) multiplication of the probe factor and this change try to fix it.
`Sum` does multiply a factor right after findCallSamples but later when using as the parameter in setProbeDistributionFactor, it multiplies one again.
This change could get ~2% perf back on mcf benchmark. In mcf, previously the corresponding factor is 1 and it's the recent feature introducing the <1 factor then trigger this bug.
Reviewed By: hoy, wenlei
Differential Revision: https://reviews.llvm.org/D99787
Currently prof metadata with branch counts is added only for BranchInst and SwitchInst, but not for IndirectBrInst. As a result, BPI/BFI make incorrect inferences for indirect branches, which can be very hot.
This diff adds metadata for IndirectBrInst, in addition to BranchInst and SwitchInst.
Reviewed By: wmi, wenlei
Differential Revision: https://reviews.llvm.org/D99550
value profile annotated after inlining.
In https://reviews.llvm.org/D96806 and https://reviews.llvm.org/D97350, we
use the magic number -1 in the value profile to avoid repeated indirect call
promotion to the same target for an indirect call. Function updateIDTMetaData
is used to mark an target as being promoted in the value profile with the
magic number. updateIDTMetaData is also used to update the value profile
when an indirect call is inlined and new inline instance profile should be
applied. For the second case, currently updateIDTMetaData mixes up the
existing value profile of the indirect call with the new profile, leading
to the problematic senario that a target count is larger than the total count
in the value profile.
The patch fixes the problem. When updateIDTMetaData is used to update the
value profile after inlining, all the values in the existing value profile
will be dropped except the values with the magic number counts.
Differential Revision: https://reviews.llvm.org/D98835
For ThinLTO's prelink compilation, we need to put external inline candidates into an import list attached to function's entry count metadata. This enables ThinLink to treat such cross module callee as hot in summary index, and later helps postlink to import them for profile guided cross module inlining.
For AutoFDO, the import list is retrieved by traversing the nested inlinee functions. For CSSPGO, since profile is flatterned, a few things need to happen for it to work:
- When loading input profile in extended binary format, we need to load all child context profile whose parent is in current module, so context trie for current module includes potential cross module inlinee.
- In order to make the above happen, we need to know whether input profile is CSSPGO profile before start reading function profile, hence a flag for profile summary section is added.
- When searching for cross module inline candidate, we need to walk through the context trie instead of nested inlinee profile (callsite sample of AutoFDO profile).
- Now that we have more accurate counts with CSSPGO, we swtiched to use entry count instead of total count to decided if an external callee is potentially beneficial to inline. This make it consistent with how we determine whether call tagert is potential inline candidate.
Differential Revision: https://reviews.llvm.org/D98590
now -funique-internal-linkage-name flag is available, and we want to flip
it on by default since it is beneficial to have separate sample profiles
for different internal symbols with the same name. As a preparation, we
want to avoid regression caused by the flip.
When we flip -funique-internal-linkage-name on, the profile is collected
from binary built without -funique-internal-linkage-name so it has no uniq
suffix, but the IR in the optimized build contains the suffix. This kind of
mismatch may introduce transient regression.
To avoid such mismatch, we introduce a NameTable section flag indicating
whether there is any name in the profile containing uniq suffix. Compiler
will decide whether to keep uniq suffix during name canonicalization
depending on the NameTable section flag. The flag is only available for
extbinary format. For other formats, by default compiler will keep uniq
suffix so they will only experience transient regression when
-funique-internal-linkage-name is just flipped.
Another type of regression is caused by places where we miss to call
getCanonicalFnName. Those places are fixed.
Differential Revision: https://reviews.llvm.org/D96932
sample loader pass.
In https://reviews.llvm.org/rG5fb65c02ca5e91e7e1a00e0efdb8edc899f3e4b9,
to prevent repeated indirect call promotion for the same indirect call
and the same target, we used zero-count value profile to indicate an
indirect call has been promoted for a certain target. We removed
PromotedInsns cache in the same patch. However, there was a problem in
that patch described below, and that problem led me to add PromotedInsns
back as a mitigation in
https://reviews.llvm.org/rG4ffad1fb489f691825d6c7d78e1626de142f26cf.
When we get value profile from metadata by calling getValueProfDataFromInst,
we need to specify the maximum possible number of values we expect to read.
We uses MaxNumPromotions in the last patch so the maximum number of value
information extracted from metadata is MaxNumPromotions. If we have many
values including zero-count values when we write the metadata, some of them
will be dropped when we read them because we only read MaxNumPromotions
values. It will allow repeated indirect call promotion again. We need to
make sure if there are values indicating promoted targets, those values need
to be saved in metadata with higher priority than other values.
The patch fixed that problem. We change to use -1 to represent the count
of a promoted target instead of 0 so it is easier to sort the values.
When we prepare to update the metadata in updateIDTMetaData, we will sort
the values in the descending count order and extract only MaxNumPromotions
values to write into metadata. Since -1 is the max uint64_t number, if we
have equal to or less than MaxNumPromotions of -1 count values, they will
all be kept in metadata. If we have more than MaxNumPromotions of -1 count
values, we will only save MaxNumPromotions such values maximally. In such
case, we have logic in place in doesHistoryAllowICP to guarantee no more
promotion in sample loader pass will happen for the indirect call, because
it has been promoted enough.
With this change, now we can remove PromotedInsns without problem.
Differential Revision: https://reviews.llvm.org/D97350
into profile symbol list.
When test is unrepresentative to production behavior, sample profile
collected from production can cause unexpected performance behavior
in test. To triage such issue, it is useful to have a cutoff flag
to control how many symbols will be included into profile symbol list
in order to do binary search.
Differential Revision: https://reviews.llvm.org/D97623
Found a problem in indirect call promotion in sample loader pass. Currently
if an indirect call is promoted for a target, and if the parent function is
inlined into some other function, the indirect call can be promoted for the
same target again. That is redundent which can harm performance and can cause
excessive compile time in some extreme case.
The patch fixes the issue. If a target is promoted for an indirect call, the
patch will write ICP metadata with the target call count being set to 0.
In the later ICP in sample profile loader, if it sees a target has 0 count
for an indirect call, it knows the target has been promoted and won't do
indirect call promotion for the indirect call.
The fix brings 0.1~0.2% performance on our search benchmark.
Differential Revision: https://reviews.llvm.org/D96806
Functions are currently processed by the sample profiler loader in a top-down order defined by the static call graph. The order is being adjusted to be a top-down order based on the input context-sensitive profile. One benefit is that the processing order of caller and callee in one SCC would follow the context order in the profile to favor more inlining. Another benefit is that the processing order of caller and callee through an indirect call (which is not on the static call graph) can be honored which in turn allows for more inlining.
The profile top-down order for SCC is also extended to support non-CS profiles.
Two switches `-mllvm -use-profile-indirect-call-edges` and `-mllvm -use-profile-top-down-order` are being introduced.
Reviewed By: wmi
Differential Revision: https://reviews.llvm.org/D95988
Sample re-annotation is required in LTO time to achieve a reasonable post-inline profile quality. However, we have seen that such LTO-time re-annotation degrades profile quality. This is mainly caused by preLTO code duplication that is done by passes such as loop unrolling, jump threading, indirect call promotion etc, where samples corresponding to a source location are aggregated multiple times due to the duplicates. In this change we are introducing a concept of distribution factor for pseudo probes so that samples can be distributed for duplicated probes scaled by a factor. We hope that optimizations duplicating code well-maintain the branch frequency information (BFI) based on which probe distribution factors are calculated. Distribution factors are updated at the end of preLTO pipeline to reflect an estimated portion of the real execution count.
This change also introduces a pseudo probe verifier that can be run after each IR passes to detect duplicated pseudo probes.
A saturated distribution factor stands for 1.0. A pesudo probe will carry a factor with the value ranged from 0.0 to 1.0. A 64-bit integral distribution factor field that represents [0.0, 1.0] is associated to each block probe. Unfortunately this cannot be done for callsite probes due to the size limitation of a 32-bit Dwarf discriminator. A 7-bit distribution factor is used instead.
Changes are also needed to the sample profile inliner to deal with prorated callsite counts. Call sites duplicated by PreLTO passes, when later on inlined in LTO time, should have the callees’s probe prorated based on the Prelink-computed distribution factors. The distribution factors should also be taken into account when computing hotness for inline candidates. Also, Indirect call promotion results in multiple callisites. The original samples should be distributed across them. This is fixed by adjusting the callisites' distribution factors.
Reviewed By: wmi
Differential Revision: https://reviews.llvm.org/D93264
This change implemented call site prioritized BFS profile guided inlining for sample profile loader. The new inlining strategy maximize the benefit of context-sensitive profile as mentioned in the follow up discussion of CSSPGO RFC. The change will not affect today's AutoFDO as it's opt-in. CSSPGO now defaults to the new FDO inliner, but can fall back to today's replay inliner using a switch (`-sample-profile-prioritized-inline=0`).
Motivation
With baseline AutoFDO, the inliner in sample profile loader only replays previous inlining, and the use of profile is only for pruning previous inlining that turned out to be cold. Due to the nature of replay, the FDO inliner is simple with hotness being the only decision factor. It has the following limitations that we're improving now for CSSPGO.
- It doesn't take inline candidate size into account. Since it's doing replay, the size growth is bounded by previous CGSCC inlining. With context-sensitive profile, FDO inliner is no longer limited by previous inlining, so we need to take size into account to avoid significant size bloat.
- The way it looks at hotness is not accurate. It uses total samples in an inlinee as proxy for hotness, while what really matters for an inline decision is the call site count. This is an unfortunate fall back because call site count and callee entry count are not reliable due to dwarf based correlation, especially for inlinees. Now paired with pseudo-probe, we have accurate call site count and callee's entry count, so we can use that to gauge hotness more accurately.
- It treats all call sites from a block as hot as long as there's one call site considered hot. This is normally true, but since total samples is used as hotness proxy, this transitiveness within block magnifies the inacurate hotness heuristic. With pseduo-probe and the change above, this is no longer an issue for CSSPGO.
New FDO Inliner
Putting all the requirement for CSSPGO together, we need a top-down call site prioritized BFS inliner. Here're reasons why each component is needed.
- Top-down: We need a top-down inliner to better leverage context-sensitive profile, so inlining is driven by accurate context profile, and post-inline is also accurate. This is already implemented in https://reviews.llvm.org/D70655.
- Size Cap: For top-down inliner, taking function size into account for inline decision alone isn't sufficient to control size growth. We also need to explicitly cap size growth because with top-down inlining, we can grow inliner size significantly with large number of smaller inlinees even if each individually passes the cost/size check.
- Prioritize call sites: With size cap, inlining order also becomes important, because if we stop inlining due to size budget limit, we'd want to use budget towards the most beneficial call sites.
- BFS inline: Same as call site prioritization, if we stop inlining due to size budget limit, we want a balanced inline tree, rather than going deep on one call path.
Note that the new inliner avoids repeatedly evaluating same set of call site, so it should help with compile time too. For this reason, we could transition today's FDO inliner to use a queue with equal priority to avoid wasted reevaluation of same call site (TODO).
Speculative indirect call promotion and inlining is also supported now with CSSPGO just like baseline AutoFDO.
Tunings and knobs
I created tuning knobs for size growth/cap control, and for hot threshold separate from CGSCC inliner. The default values are selected based on initial tuning with CSSPGO.
Results
Evaluated with an internal LLVM fork couple months ago, plus another change to adjust hot-threshold cutoff for context profile (will send up after this one), the new inliner show ~1% geomean perf win on spec2006 with CSSPGO, while reducing code size too. The measurement was done using train-train setup, MonoLTO w/ new pass manager and pseudo-probe. Note that this is just a starting point - we hope that the new inliner will open up more opportunity with CSSPGO, but it will certainly take more time and effort to make it fully calibrated and ready for bigger workloads (we're working on it).
Differential Revision: https://reviews.llvm.org/D94001
Fixing up a couple places where `getCallSiteIdentifier` is needed to support pseudo-probe-based callsites.
Also fixing an issue in the extbinary profile reader where the metadata section is not fully scanned based on the number of profiles loaded only for the current module.
Reviewed By: wmi, wenlei
Differential Revision: https://reviews.llvm.org/D95791
turning off SampleFDO silently.
Currently sample loader pass turns off SampleFDO optimization silently when
it sees error in reading the profile. This behavior will defeat the tests
which could have caught those bad/incompatible profile problems. This patch
change the behavior to report error.
Differential Revision: https://reviews.llvm.org/D95269
separate sections.
For ThinLTO, all the function profiles without context has been annotated to
outline functions if possible in prelink phase. In postlink phase, profile
annotation in postlink phase is only meaningful for function profile with
context. If the profile is large, it is better to split the profile into two
parts, one with context and one without, so the profile reading in postlink
phase only has to read the part with context. To have the profile splitting,
we extend the ExtBinary format to support different section arrangement. It
will be flexible to add other section layout in the future without the need
to create new class inheriting from ExtBinary class.
Differential Revision: https://reviews.llvm.org/D94435
This change modifies the source location formatting from:
LineNumber.Discriminator
to:
LineNumber:ColumnNumber.Discriminator
The motivation here is to enhance location information for inline replay that currently exists for the SampleProfile inliner. This will be leveraged further in inline replay for the CGSCC inliner in the related diff.
The ReplayInlineAdvisor is also modified to read the new format and now takes into account the callee for greater accuracy.
Testing:
ninja check-llvm
Reviewed By: mtrofin
Differential Revision: https://reviews.llvm.org/D94333
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
MD5 is used.
Currently during sample profile loading, NameTable has to be loaded entirely
up front before any name string is retrieved. That is because NameTable is
stored using ULEB128 encoding and cannot be directly accessed like an array.
However, if MD5 is used to represent name in the NameTable, it has fixed
length. If MD5 names are stored in uint64_t type instead of ULEB128, NameTable
can be accessed like an array then in many cases only part of the NameTable
has to be read. This is helpful for reducing compile time especially when
small source file is compiled. We find that after this change, the elapsed
time to build a large application distributively is reduced by 5% and the
accumulative cpu time used for building is also reduced by 5%. The size of
the profile is slightly reduced with this change by ~0.2%, and that also
indicates encoding MD5 in ULEB128 doesn't save the storage space.
Differential Revision: https://reviews.llvm.org/D92621
This change adds the context-senstive sample PGO infracture described in CSSPGO RFC (https://groups.google.com/g/llvm-dev/c/1p1rdYbL93s). It introduced an abstraction between input profile and profile loader that queries input profile for functions. Specifically, there's now the notion of base profile and context profile, and they are managed by the new SampleContextTracker for adjusting and merging profiles based on inline decisions. It works with top-down profiled guided inliner in profile loader (https://reviews.llvm.org/D70655) for better inlining with specialization and better post-inline profile fidelity. In the future, we can also expose this infrastructure to CGSCC inliner in order for it to take advantage of context-sensitive profile. This change is the consumption part of context-sensitive profile (The generation part is in this stack: https://reviews.llvm.org/D89707). We've seen good results internally in conjunction with Pseudo-probe (https://reviews.llvm.org/D86193). Pacthes for integration with Pseudo-probe coming up soon.
Currently the new infrastructure kick in when input profile contains the new context-sensitive profile; otherwise it's no-op and does not affect existing AutoFDO.
**Interface**
There're two sets of interfaces for query and tracking respectively exposed from SampleContextTracker. For query, now instead of simply getting a profile from input for a function, we can explicitly query base profile or context profile for given call path of a function. For tracking, there're separate APIs for marking context profile as inlined, or promoting and merging not inlined context profile.
- Query base profile (`getBaseSamplesFor`)
Base profile is the merged synthetic profile for function's CFG profile from any outstanding (not inlined) context. We can query base profile by function.
- Query context profile (`getContextSamplesFor`)
Context profile is a function's CFG profile for a given calling context. We can query context profile by context string.
- Track inlined context profile (`markContextSamplesInlined`)
When a function is inlined for given calling context, we need to mark the context profile for that context as inlined. This is to make sure we don't include inlined context profile when synthesizing base profile for that inlined function.
- Track not-inlined context profile (`promoteMergeContextSamplesTree`)
When a function is not inlined for given calling context, we need to promote the context profile tree so the not inlined context becomes top-level context. This preserve the sub-context under that function so later inline decision for that not inlined function will still have context profile for its call tree. Note that profile will be merged if needed when promoting a context profile tree if any of the node already exists at its promoted destination.
**Implementation**
Implementation-wise, `SampleContext` is created as abstraction for context. Currently it's a string for call path, and we can later optimize it to something more efficient, e.g. context id. Each `SampleContext` also has a `ContextState` indicating whether it's raw context profile from input, whether it's inlined or merged, whether it's synthetic profile created by compiler. Each `FunctionSamples` now has a `SampleContext` that tells whether it's base profile or context profile, and for context profile what is the context and state.
On top of the above context representation, a custom trie tree is implemented to track and manager context profiles. Specifically, `SampleContextTracker` is implemented that encapsulates a trie tree with `ContextTireNode` as node. Each node of the trie tree represents a frame in calling context, thus the path from root to a node represents a valid calling context. We also track `FunctionSamples` for each node, so this trie tree can serve efficient query for context profile. Accordingly, context profile tree promotion now becomes moving a subtree to be under the root of entire tree, and merge nodes for subtree if this move encounters existing nodes.
**Integration**
`SampleContextTracker` is now also integrated with AutoFDO, `SampleProfileReader` and `SampleProfileLoader`. When we detected input profile contains context-sensitive profile, `SampleContextTracker` will be used to track profiles, and all profile query will go to `SampleContextTracker` instead of `SampleProfileReader` automatically. Tracking APIs are called automatically for each inline decision from `SampleProfileLoader`.
Differential Revision: https://reviews.llvm.org/D90125
This is the #2 of 2 changes that make remarks hotness threshold option
available in more tools. The changes also allow the threshold to sync with
hotness threshold from profile summary with special value 'auto'.
This change expands remarks hotness threshold option
-fdiagnostics-hotness-threshold in clang and *-remarks-hotness-threshold in
other tools to utilize hotness threshold from profile summary.
Remarks hotness filtering relies on several driver options. Table below lists
how different options are correlated and affect final remarks outputs:
| profile | hotness | threshold | remarks printed |
|---------|---------|-----------|-----------------|
| No | No | No | All |
| No | No | Yes | None |
| No | Yes | No | All |
| No | Yes | Yes | None |
| Yes | No | No | All |
| Yes | No | Yes | None |
| Yes | Yes | No | All |
| Yes | Yes | Yes | >=threshold |
In the presence of profile summary, it is often more desirable to directly use
the hotness threshold from profile summary. The new argument value 'auto'
indicates threshold will be synced with hotness threshold from profile summary
during compilation. The "auto" threshold relies on the availability of profile
summary. In case of missing such information, no remarks will be generated.
Differential Revision: https://reviews.llvm.org/D85808
and indirect call promotion candidate.
Profile remapping is a feature to match a function in the module with its
profile in sample profile if the function name and the name in profile look
different but are equivalent using given remapping rules. This is a useful
feature to keep the performance stable by specifying some remapping rules
when sampleFDO targets are going through some large scale function signature
change.
However, currently profile remapping support is only valid for outline
function profile in SampleFDO. It cannot match a callee with an inline
instance profile if they have different but equivalent names. We found
that without the support for inline instance profile, remapping is less
effective for some large scale change.
To add that support, before any remapping lookup happens, all the names
in the profile will be inserted into remapper and the Key to the name
mapping will be recorded in a map called NameMap in the remapper. During
name lookup, a Key will be returned for the given name and it will be used
to extract an equivalent name in the profile from NameMap. So with the help
of the NameMap, we can translate any given name to an equivalent name in
the profile if it exists. Whenever we try to match a name in the module to
a name in the profile, we will try the match with the original name first,
and if it doesn't match, we will use the equivalent name got from remapper
to try the match for another time. In this way, the patch can enhance the
profile remapping support for searching inline instance and searching
indirect call promotion candidate.
In a planned large scale change of int64 type (long long) to int64_t (long),
we found the performance of a google internal benchmark degraded by 2% if
nothing was done. If existing profile remapping was enabled, the performance
degradation dropped to 1.2%. If the profile remapping with the current patch
was enabled, the performance degradation further dropped to 0.14% (Note the
experiment was done before searching indirect call promotion candidate was
added. We hope with the remapping support of searching indirect call promotion
candidate, the degradation can drop to 0% in the end. It will be evaluated
post commit).
Differential Revision: https://reviews.llvm.org/D86332
This change added a new inline advisor that takes optimization remarks from previous inlining as input, and provides the decision as advice so current inlining can replay inline decisions of a different compilation. Dwarf inline stack with line and discriminator is used as anchor for call sites including call context. The change can be useful for Inliner tuning as it provides a channel to allow external input for tweaking inline decisions. Existing alternatives like alwaysinline attribute is per-function, not per-callsite. Per-callsite inline intrinsic can be another solution (not yet existing), but it's intrusive to implement and also does not differentiate call context.
A switch -sample-profile-inline-replay=<inline_remarks_file> is added to hook up the new inline advisor with SampleProfileLoader's inline decision for replay. Since SampleProfileLoader does top-down inlining, inline decision can be specialized for each call context, hence we should be able to replay inlining accurately. However with a bottom-up inliner like CGSCC inlining, the replay can be limited due to lack of specialization for different call context. Apart from that limitation, the new inline advisor can still be used by regular CGSCC inliner later if needed for tuning purpose.
This is a resubmit of https://reviews.llvm.org/D83743
is enabled.
When -sample-profile-merge-inlinee is enabled, new FunctionSamples may be
created during profile merge without GUIDToFuncNameMap being initialized.
That will occasionally cause compiler crash. The patch fixes it.
Differential Revision: https://reviews.llvm.org/D84994
Summary:
This change added a new inline advisor that takes optimization remarks from previous inlining as input, and provides the decision as advice so current inlining can replay inline decisions of a different compilation. Dwarf inline stack with line and discriminator is used as anchor for call sites including call context. The change can be useful for Inliner tuning as it provides a channel to allow external input for tweaking inline decisions. Existing alternatives like alwaysinline attribute is per-function, not per-callsite. Per-callsite inline intrinsic can be another solution (not yet existing), but it's intrusive to implement and also does not differentiate call context.
A switch -sample-profile-inline-replay=<inline_remarks_file> is added to hook up the new inline advisor with SampleProfileLoader's inline decision for replay. Since SampleProfileLoader does top-down inlining, inline decision can be specialized for each call context, hence we should be able to replay inlining accurately. However with a bottom-up inliner like CGSCC inlining, the replay can be limited due to lack of specialization for different call context. Apart from that limitation, the new inline advisor can still be used by regular CGSCC inliner later if needed for tuning purpose.
Subscribers: mgorny, aprantl, hiraditya, llvm-commits
Tags: #llvm
Resubmit for https://reviews.llvm.org/D84086
Fangrui Song