llvm-project/llvm/tools/llvm-profgen/ProfileGenerator.cpp

Ignoring revisions in .git-blame-ignore-revs. Click here to bypass and see the normal blame view.

1152 lines
43 KiB
C++
Raw Normal View History

[CSSPGO][llvm-profgen] Context-sensitive profile data generation This stack of changes introduces `llvm-profgen` utility which generates a profile data file from given perf script data files for sample-based PGO. It’s part of(not only) the CSSPGO work. Specifically to support context-sensitive with/without pseudo probe profile, it implements a series of functionalities including perf trace parsing, instruction symbolization, LBR stack/call frame stack unwinding, pseudo probe decoding, etc. Also high throughput is achieved by multiple levels of sample aggregation and compatible format with one stop is generated at the end. Please refer to: https://groups.google.com/g/llvm-dev/c/1p1rdYbL93s for the CSSPGO RFC. This change supports context-sensitive profile data generation into llvm-profgen. With simultaneous sampling for LBR and call stack, we can identify leaf of LBR sample with calling context from stack sample . During the process of deriving fall through path from LBR entries, we unwind LBR by replaying all the calls and returns (including implicit calls/returns due to inlining) backwards on top of the sampled call stack. Then the state of call stack as we unwind through LBR always represents the calling context of current fall through path. we have two types of virtual unwinding 1) LBR unwinding and 2) linear range unwinding. Specifically, for each LBR entry which can be classified into call, return, regular branch, LBR unwinding will replay the operation by pushing, popping or switching leaf frame towards the call stack and since the initial call stack is most recently sampled, the replay should be in anti-execution order, i.e. for the regular case, pop the call stack when LBR is call, push frame on call stack when LBR is return. After each LBR processed, it also needs to align with the next LBR by going through instructions from previous LBR's target to current LBR's source, which we named linear unwinding. As instruction from linear range can come from different function by inlining, linear unwinding will do the range splitting and record counters through the range with same inline context. With each fall through path from LBR unwinding, we aggregate each sample into counters by the calling context and eventually generate full context sensitive profile (without relying on inlining) to driver compiler's PGO/FDO. A breakdown of noteworthy changes: - Added `HybridSample` class as the abstraction perf sample including LBR stack and call stack * Extended `PerfReader` to implement auto-detect whether input perf script output contains CS profile, then do the parsing. Multiple `HybridSample` are extracted * Speed up by aggregating `HybridSample` into `AggregatedSamples` * Added VirtualUnwinder that consumes aggregated `HybridSample` and implements unwinding of calls, returns, and linear path that contains implicit call/return from inlining. Ranges and branches counters are aggregated by the calling context.
 Here calling context is string type, each context is a pair of function name and callsite location info, the whole context is like `main:1 @ foo:2 @ bar`. * Added PorfileGenerater that accumulates counters by ranges unfolding or branch target mapping, then generates context-sensitive function profile including function body, inferring callee's head sample, callsite target samples, eventually records into ProfileMap.
 * Leveraged LLVM build-in(`SampleProfWriter`) writer to support different serialization format with no stop - `getCanonicalFnName` for callee name and name from ELF section - Added regression test for both unwinding and profile generation Test Plan: ninja & ninja check-llvm Reviewed By: hoy, wenlei, wmi Differential Revision: https://reviews.llvm.org/D89723
2020-10-19 12:55:59 -07:00
//===-- ProfileGenerator.cpp - Profile Generator ---------------*- C++ -*-===//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
//
//===----------------------------------------------------------------------===//
#include "ProfileGenerator.h"
#include "ErrorHandling.h"
#include "PerfReader.h"
#include "ProfiledBinary.h"
#include "llvm/DebugInfo/Symbolize/SymbolizableModule.h"
#include "llvm/ProfileData/ProfileCommon.h"
#include <algorithm>
#include <float.h>
#include <unordered_set>
#include <utility>
[CSSPGO][llvm-profgen] Context-sensitive profile data generation This stack of changes introduces `llvm-profgen` utility which generates a profile data file from given perf script data files for sample-based PGO. It’s part of(not only) the CSSPGO work. Specifically to support context-sensitive with/without pseudo probe profile, it implements a series of functionalities including perf trace parsing, instruction symbolization, LBR stack/call frame stack unwinding, pseudo probe decoding, etc. Also high throughput is achieved by multiple levels of sample aggregation and compatible format with one stop is generated at the end. Please refer to: https://groups.google.com/g/llvm-dev/c/1p1rdYbL93s for the CSSPGO RFC. This change supports context-sensitive profile data generation into llvm-profgen. With simultaneous sampling for LBR and call stack, we can identify leaf of LBR sample with calling context from stack sample . During the process of deriving fall through path from LBR entries, we unwind LBR by replaying all the calls and returns (including implicit calls/returns due to inlining) backwards on top of the sampled call stack. Then the state of call stack as we unwind through LBR always represents the calling context of current fall through path. we have two types of virtual unwinding 1) LBR unwinding and 2) linear range unwinding. Specifically, for each LBR entry which can be classified into call, return, regular branch, LBR unwinding will replay the operation by pushing, popping or switching leaf frame towards the call stack and since the initial call stack is most recently sampled, the replay should be in anti-execution order, i.e. for the regular case, pop the call stack when LBR is call, push frame on call stack when LBR is return. After each LBR processed, it also needs to align with the next LBR by going through instructions from previous LBR's target to current LBR's source, which we named linear unwinding. As instruction from linear range can come from different function by inlining, linear unwinding will do the range splitting and record counters through the range with same inline context. With each fall through path from LBR unwinding, we aggregate each sample into counters by the calling context and eventually generate full context sensitive profile (without relying on inlining) to driver compiler's PGO/FDO. A breakdown of noteworthy changes: - Added `HybridSample` class as the abstraction perf sample including LBR stack and call stack * Extended `PerfReader` to implement auto-detect whether input perf script output contains CS profile, then do the parsing. Multiple `HybridSample` are extracted * Speed up by aggregating `HybridSample` into `AggregatedSamples` * Added VirtualUnwinder that consumes aggregated `HybridSample` and implements unwinding of calls, returns, and linear path that contains implicit call/return from inlining. Ranges and branches counters are aggregated by the calling context.
 Here calling context is string type, each context is a pair of function name and callsite location info, the whole context is like `main:1 @ foo:2 @ bar`. * Added PorfileGenerater that accumulates counters by ranges unfolding or branch target mapping, then generates context-sensitive function profile including function body, inferring callee's head sample, callsite target samples, eventually records into ProfileMap.
 * Leveraged LLVM build-in(`SampleProfWriter`) writer to support different serialization format with no stop - `getCanonicalFnName` for callee name and name from ELF section - Added regression test for both unwinding and profile generation Test Plan: ninja & ninja check-llvm Reviewed By: hoy, wenlei, wmi Differential Revision: https://reviews.llvm.org/D89723
2020-10-19 12:55:59 -07:00
[llvm-profgen] Support LBR only perf script This change aims at supporting LBR only sample perf script which is used for regular(Non-CS) profile generation. A LBR perf script includes a batch of LBR sample which starts with a frame pointer and a group of 32 LBR entries is followed. The FROM/TO LBR pair and the range between two consecutive entries (the former entry's TO and the latter entry's FROM) will be used to infer function profile info. An example of LBR perf script(created by `perf script -F ip,brstack -i perf.data`) ``` 40062f 0x40062f/0x4005b0/P/-/-/9 0x400645/0x4005ff/P/-/-/1 0x400637/0x400645/P/-/-/1 ... 4005d7 0x4005d7/0x4005e5/P/-/-/8 0x40062f/0x4005b0/P/-/-/6 0x400645/0x4005ff/P/-/-/1 ... ... ``` For implementation: - Extended a new child class `LBRPerfReader` for the sample parsing, reused all the functionalities in `extractLBRStack` except for an extension to parsing leading instruction pointer. - `HybridSample` is reused(just leave the call stack empty) and the parsed samples is still aggregated in `AggregatedSamples`. After that, range samples, branch sample, address samples are computed and recorded. - Reused `ContextSampleCounterMap` to store the raw profile, since it's no need to aggregation by context, here it just registered one sample counter with a fake context key. - Unified to use `show-raw-profile` instead of `show-unwinder-output` to dump the intermediate raw profile, see the comments of the format of the raw profile. For CS profile, it remains to output the unwinder output. Profile generation part will come soon. Differential Revision: https://reviews.llvm.org/D108153
2021-08-31 13:27:42 -07:00
cl::opt<std::string> OutputFilename("output", cl::value_desc("output"),
cl::Required,
cl::desc("Output profile file"));
static cl::alias OutputA("o", cl::desc("Alias for --output"),
cl::aliasopt(OutputFilename));
[CSSPGO][llvm-profgen] Context-sensitive profile data generation This stack of changes introduces `llvm-profgen` utility which generates a profile data file from given perf script data files for sample-based PGO. It’s part of(not only) the CSSPGO work. Specifically to support context-sensitive with/without pseudo probe profile, it implements a series of functionalities including perf trace parsing, instruction symbolization, LBR stack/call frame stack unwinding, pseudo probe decoding, etc. Also high throughput is achieved by multiple levels of sample aggregation and compatible format with one stop is generated at the end. Please refer to: https://groups.google.com/g/llvm-dev/c/1p1rdYbL93s for the CSSPGO RFC. This change supports context-sensitive profile data generation into llvm-profgen. With simultaneous sampling for LBR and call stack, we can identify leaf of LBR sample with calling context from stack sample . During the process of deriving fall through path from LBR entries, we unwind LBR by replaying all the calls and returns (including implicit calls/returns due to inlining) backwards on top of the sampled call stack. Then the state of call stack as we unwind through LBR always represents the calling context of current fall through path. we have two types of virtual unwinding 1) LBR unwinding and 2) linear range unwinding. Specifically, for each LBR entry which can be classified into call, return, regular branch, LBR unwinding will replay the operation by pushing, popping or switching leaf frame towards the call stack and since the initial call stack is most recently sampled, the replay should be in anti-execution order, i.e. for the regular case, pop the call stack when LBR is call, push frame on call stack when LBR is return. After each LBR processed, it also needs to align with the next LBR by going through instructions from previous LBR's target to current LBR's source, which we named linear unwinding. As instruction from linear range can come from different function by inlining, linear unwinding will do the range splitting and record counters through the range with same inline context. With each fall through path from LBR unwinding, we aggregate each sample into counters by the calling context and eventually generate full context sensitive profile (without relying on inlining) to driver compiler's PGO/FDO. A breakdown of noteworthy changes: - Added `HybridSample` class as the abstraction perf sample including LBR stack and call stack * Extended `PerfReader` to implement auto-detect whether input perf script output contains CS profile, then do the parsing. Multiple `HybridSample` are extracted * Speed up by aggregating `HybridSample` into `AggregatedSamples` * Added VirtualUnwinder that consumes aggregated `HybridSample` and implements unwinding of calls, returns, and linear path that contains implicit call/return from inlining. Ranges and branches counters are aggregated by the calling context.
 Here calling context is string type, each context is a pair of function name and callsite location info, the whole context is like `main:1 @ foo:2 @ bar`. * Added PorfileGenerater that accumulates counters by ranges unfolding or branch target mapping, then generates context-sensitive function profile including function body, inferring callee's head sample, callsite target samples, eventually records into ProfileMap.
 * Leveraged LLVM build-in(`SampleProfWriter`) writer to support different serialization format with no stop - `getCanonicalFnName` for callee name and name from ELF section - Added regression test for both unwinding and profile generation Test Plan: ninja & ninja check-llvm Reviewed By: hoy, wenlei, wmi Differential Revision: https://reviews.llvm.org/D89723
2020-10-19 12:55:59 -07:00
static cl::opt<SampleProfileFormat> OutputFormat(
"format", cl::desc("Format of output profile"), cl::init(SPF_Ext_Binary),
[CSSPGO][llvm-profgen] Context-sensitive profile data generation This stack of changes introduces `llvm-profgen` utility which generates a profile data file from given perf script data files for sample-based PGO. It’s part of(not only) the CSSPGO work. Specifically to support context-sensitive with/without pseudo probe profile, it implements a series of functionalities including perf trace parsing, instruction symbolization, LBR stack/call frame stack unwinding, pseudo probe decoding, etc. Also high throughput is achieved by multiple levels of sample aggregation and compatible format with one stop is generated at the end. Please refer to: https://groups.google.com/g/llvm-dev/c/1p1rdYbL93s for the CSSPGO RFC. This change supports context-sensitive profile data generation into llvm-profgen. With simultaneous sampling for LBR and call stack, we can identify leaf of LBR sample with calling context from stack sample . During the process of deriving fall through path from LBR entries, we unwind LBR by replaying all the calls and returns (including implicit calls/returns due to inlining) backwards on top of the sampled call stack. Then the state of call stack as we unwind through LBR always represents the calling context of current fall through path. we have two types of virtual unwinding 1) LBR unwinding and 2) linear range unwinding. Specifically, for each LBR entry which can be classified into call, return, regular branch, LBR unwinding will replay the operation by pushing, popping or switching leaf frame towards the call stack and since the initial call stack is most recently sampled, the replay should be in anti-execution order, i.e. for the regular case, pop the call stack when LBR is call, push frame on call stack when LBR is return. After each LBR processed, it also needs to align with the next LBR by going through instructions from previous LBR's target to current LBR's source, which we named linear unwinding. As instruction from linear range can come from different function by inlining, linear unwinding will do the range splitting and record counters through the range with same inline context. With each fall through path from LBR unwinding, we aggregate each sample into counters by the calling context and eventually generate full context sensitive profile (without relying on inlining) to driver compiler's PGO/FDO. A breakdown of noteworthy changes: - Added `HybridSample` class as the abstraction perf sample including LBR stack and call stack * Extended `PerfReader` to implement auto-detect whether input perf script output contains CS profile, then do the parsing. Multiple `HybridSample` are extracted * Speed up by aggregating `HybridSample` into `AggregatedSamples` * Added VirtualUnwinder that consumes aggregated `HybridSample` and implements unwinding of calls, returns, and linear path that contains implicit call/return from inlining. Ranges and branches counters are aggregated by the calling context.
 Here calling context is string type, each context is a pair of function name and callsite location info, the whole context is like `main:1 @ foo:2 @ bar`. * Added PorfileGenerater that accumulates counters by ranges unfolding or branch target mapping, then generates context-sensitive function profile including function body, inferring callee's head sample, callsite target samples, eventually records into ProfileMap.
 * Leveraged LLVM build-in(`SampleProfWriter`) writer to support different serialization format with no stop - `getCanonicalFnName` for callee name and name from ELF section - Added regression test for both unwinding and profile generation Test Plan: ninja & ninja check-llvm Reviewed By: hoy, wenlei, wmi Differential Revision: https://reviews.llvm.org/D89723
2020-10-19 12:55:59 -07:00
cl::values(
clEnumValN(SPF_Binary, "binary", "Binary encoding (default)"),
clEnumValN(SPF_Compact_Binary, "compbinary", "Compact binary encoding"),
clEnumValN(SPF_Ext_Binary, "extbinary", "Extensible binary encoding"),
clEnumValN(SPF_Text, "text", "Text encoding"),
clEnumValN(SPF_GCC, "gcc",
"GCC encoding (only meaningful for -sample)")));
cl::opt<bool> UseMD5(
"use-md5", cl::init(false), cl::Hidden,
cl::desc("Use md5 to represent function names in the output profile (only "
"meaningful for -extbinary)"));
static cl::opt<bool> PopulateProfileSymbolList(
"populate-profile-symbol-list", cl::init(false), cl::Hidden,
cl::desc("Populate profile symbol list (only meaningful for -extbinary)"));
static cl::opt<bool> FillZeroForAllFuncs(
"fill-zero-for-all-funcs", cl::init(false), cl::Hidden,
cl::desc("Attribute all functions' range with zero count "
"even it's not hit by any samples."));
static cl::opt<int32_t, true> RecursionCompression(
"compress-recursion",
cl::desc("Compressing recursion by deduplicating adjacent frame "
"sequences up to the specified size. -1 means no size limit."),
cl::Hidden,
cl::location(llvm::sampleprof::CSProfileGenerator::MaxCompressionSize));
static cl::opt<bool>
TrimColdProfile("trim-cold-profile", cl::init(false), cl::ZeroOrMore,
cl::desc("If the total count of the profile is smaller "
"than threshold, it will be trimmed."));
static cl::opt<bool> CSProfMergeColdContext(
"csprof-merge-cold-context", cl::init(true), cl::ZeroOrMore,
cl::desc("If the total count of context profile is smaller than "
"the threshold, it will be merged into context-less base "
"profile."));
static cl::opt<uint32_t> CSProfMaxColdContextDepth(
"csprof-max-cold-context-depth", cl::init(1), cl::ZeroOrMore,
cl::desc("Keep the last K contexts while merging cold profile. 1 means the "
"context-less base profile"));
static cl::opt<int, true> CSProfMaxContextDepth(
"csprof-max-context-depth", cl::ZeroOrMore,
cl::desc("Keep the last K contexts while merging profile. -1 means no "
"depth limit."),
cl::location(llvm::sampleprof::CSProfileGenerator::MaxContextDepth));
static cl::opt<double> HotFunctionDensityThreshold(
"hot-function-density-threshold", llvm::cl::init(1000),
llvm::cl::desc(
"specify density threshold for hot functions (default: 1000)"),
llvm::cl::Optional);
static cl::opt<bool> ShowDensity("show-density", llvm::cl::init(false),
llvm::cl::desc("show profile density details"),
llvm::cl::Optional);
static cl::opt<bool> UpdateTotalSamples(
"update-total-samples", llvm::cl::init(false),
llvm::cl::desc(
"Update total samples by accumulating all its body samples."),
llvm::cl::Optional);
extern cl::opt<int> ProfileSummaryCutoffHot;
[CSSPGO] Use nested context-sensitive profile. 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
2021-12-14 10:03:05 -08:00
static cl::opt<bool> GenCSNestedProfile(
"gen-cs-nested-profile", cl::Hidden, cl::init(true),
[CSSPGO] Use nested context-sensitive profile. 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
2021-12-14 10:03:05 -08:00
cl::desc("Generate nested function profiles for CSSPGO"));
[CSSPGO][llvm-profgen] Context-sensitive profile data generation This stack of changes introduces `llvm-profgen` utility which generates a profile data file from given perf script data files for sample-based PGO. It’s part of(not only) the CSSPGO work. Specifically to support context-sensitive with/without pseudo probe profile, it implements a series of functionalities including perf trace parsing, instruction symbolization, LBR stack/call frame stack unwinding, pseudo probe decoding, etc. Also high throughput is achieved by multiple levels of sample aggregation and compatible format with one stop is generated at the end. Please refer to: https://groups.google.com/g/llvm-dev/c/1p1rdYbL93s for the CSSPGO RFC. This change supports context-sensitive profile data generation into llvm-profgen. With simultaneous sampling for LBR and call stack, we can identify leaf of LBR sample with calling context from stack sample . During the process of deriving fall through path from LBR entries, we unwind LBR by replaying all the calls and returns (including implicit calls/returns due to inlining) backwards on top of the sampled call stack. Then the state of call stack as we unwind through LBR always represents the calling context of current fall through path. we have two types of virtual unwinding 1) LBR unwinding and 2) linear range unwinding. Specifically, for each LBR entry which can be classified into call, return, regular branch, LBR unwinding will replay the operation by pushing, popping or switching leaf frame towards the call stack and since the initial call stack is most recently sampled, the replay should be in anti-execution order, i.e. for the regular case, pop the call stack when LBR is call, push frame on call stack when LBR is return. After each LBR processed, it also needs to align with the next LBR by going through instructions from previous LBR's target to current LBR's source, which we named linear unwinding. As instruction from linear range can come from different function by inlining, linear unwinding will do the range splitting and record counters through the range with same inline context. With each fall through path from LBR unwinding, we aggregate each sample into counters by the calling context and eventually generate full context sensitive profile (without relying on inlining) to driver compiler's PGO/FDO. A breakdown of noteworthy changes: - Added `HybridSample` class as the abstraction perf sample including LBR stack and call stack * Extended `PerfReader` to implement auto-detect whether input perf script output contains CS profile, then do the parsing. Multiple `HybridSample` are extracted * Speed up by aggregating `HybridSample` into `AggregatedSamples` * Added VirtualUnwinder that consumes aggregated `HybridSample` and implements unwinding of calls, returns, and linear path that contains implicit call/return from inlining. Ranges and branches counters are aggregated by the calling context.
 Here calling context is string type, each context is a pair of function name and callsite location info, the whole context is like `main:1 @ foo:2 @ bar`. * Added PorfileGenerater that accumulates counters by ranges unfolding or branch target mapping, then generates context-sensitive function profile including function body, inferring callee's head sample, callsite target samples, eventually records into ProfileMap.
 * Leveraged LLVM build-in(`SampleProfWriter`) writer to support different serialization format with no stop - `getCanonicalFnName` for callee name and name from ELF section - Added regression test for both unwinding and profile generation Test Plan: ninja & ninja check-llvm Reviewed By: hoy, wenlei, wmi Differential Revision: https://reviews.llvm.org/D89723
2020-10-19 12:55:59 -07:00
using namespace llvm;
using namespace sampleprof;
namespace llvm {
namespace sampleprof {
// Initialize the MaxCompressionSize to -1 which means no size limit
int32_t CSProfileGenerator::MaxCompressionSize = -1;
int CSProfileGenerator::MaxContextDepth = -1;
bool ProfileGeneratorBase::UseFSDiscriminator = false;
std::unique_ptr<ProfileGeneratorBase>
ProfileGeneratorBase::create(ProfiledBinary *Binary,
const ContextSampleCounterMap *SampleCounters,
[CSSPGO] Use nested context-sensitive profile. 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
2021-12-14 10:03:05 -08:00
bool ProfileIsCSFlat) {
std::unique_ptr<ProfileGeneratorBase> Generator;
[CSSPGO] Use nested context-sensitive profile. 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
2021-12-14 10:03:05 -08:00
if (ProfileIsCSFlat) {
if (Binary->useFSDiscriminator())
exitWithError("FS discriminator is not supported in CS profile.");
Generator.reset(new CSProfileGenerator(Binary, SampleCounters));
[CSSPGO][llvm-profgen] Context-sensitive profile data generation This stack of changes introduces `llvm-profgen` utility which generates a profile data file from given perf script data files for sample-based PGO. It’s part of(not only) the CSSPGO work. Specifically to support context-sensitive with/without pseudo probe profile, it implements a series of functionalities including perf trace parsing, instruction symbolization, LBR stack/call frame stack unwinding, pseudo probe decoding, etc. Also high throughput is achieved by multiple levels of sample aggregation and compatible format with one stop is generated at the end. Please refer to: https://groups.google.com/g/llvm-dev/c/1p1rdYbL93s for the CSSPGO RFC. This change supports context-sensitive profile data generation into llvm-profgen. With simultaneous sampling for LBR and call stack, we can identify leaf of LBR sample with calling context from stack sample . During the process of deriving fall through path from LBR entries, we unwind LBR by replaying all the calls and returns (including implicit calls/returns due to inlining) backwards on top of the sampled call stack. Then the state of call stack as we unwind through LBR always represents the calling context of current fall through path. we have two types of virtual unwinding 1) LBR unwinding and 2) linear range unwinding. Specifically, for each LBR entry which can be classified into call, return, regular branch, LBR unwinding will replay the operation by pushing, popping or switching leaf frame towards the call stack and since the initial call stack is most recently sampled, the replay should be in anti-execution order, i.e. for the regular case, pop the call stack when LBR is call, push frame on call stack when LBR is return. After each LBR processed, it also needs to align with the next LBR by going through instructions from previous LBR's target to current LBR's source, which we named linear unwinding. As instruction from linear range can come from different function by inlining, linear unwinding will do the range splitting and record counters through the range with same inline context. With each fall through path from LBR unwinding, we aggregate each sample into counters by the calling context and eventually generate full context sensitive profile (without relying on inlining) to driver compiler's PGO/FDO. A breakdown of noteworthy changes: - Added `HybridSample` class as the abstraction perf sample including LBR stack and call stack * Extended `PerfReader` to implement auto-detect whether input perf script output contains CS profile, then do the parsing. Multiple `HybridSample` are extracted * Speed up by aggregating `HybridSample` into `AggregatedSamples` * Added VirtualUnwinder that consumes aggregated `HybridSample` and implements unwinding of calls, returns, and linear path that contains implicit call/return from inlining. Ranges and branches counters are aggregated by the calling context.
 Here calling context is string type, each context is a pair of function name and callsite location info, the whole context is like `main:1 @ foo:2 @ bar`. * Added PorfileGenerater that accumulates counters by ranges unfolding or branch target mapping, then generates context-sensitive function profile including function body, inferring callee's head sample, callsite target samples, eventually records into ProfileMap.
 * Leveraged LLVM build-in(`SampleProfWriter`) writer to support different serialization format with no stop - `getCanonicalFnName` for callee name and name from ELF section - Added regression test for both unwinding and profile generation Test Plan: ninja & ninja check-llvm Reviewed By: hoy, wenlei, wmi Differential Revision: https://reviews.llvm.org/D89723
2020-10-19 12:55:59 -07:00
} else {
Generator.reset(new ProfileGenerator(Binary, SampleCounters));
[CSSPGO][llvm-profgen] Context-sensitive profile data generation This stack of changes introduces `llvm-profgen` utility which generates a profile data file from given perf script data files for sample-based PGO. It’s part of(not only) the CSSPGO work. Specifically to support context-sensitive with/without pseudo probe profile, it implements a series of functionalities including perf trace parsing, instruction symbolization, LBR stack/call frame stack unwinding, pseudo probe decoding, etc. Also high throughput is achieved by multiple levels of sample aggregation and compatible format with one stop is generated at the end. Please refer to: https://groups.google.com/g/llvm-dev/c/1p1rdYbL93s for the CSSPGO RFC. This change supports context-sensitive profile data generation into llvm-profgen. With simultaneous sampling for LBR and call stack, we can identify leaf of LBR sample with calling context from stack sample . During the process of deriving fall through path from LBR entries, we unwind LBR by replaying all the calls and returns (including implicit calls/returns due to inlining) backwards on top of the sampled call stack. Then the state of call stack as we unwind through LBR always represents the calling context of current fall through path. we have two types of virtual unwinding 1) LBR unwinding and 2) linear range unwinding. Specifically, for each LBR entry which can be classified into call, return, regular branch, LBR unwinding will replay the operation by pushing, popping or switching leaf frame towards the call stack and since the initial call stack is most recently sampled, the replay should be in anti-execution order, i.e. for the regular case, pop the call stack when LBR is call, push frame on call stack when LBR is return. After each LBR processed, it also needs to align with the next LBR by going through instructions from previous LBR's target to current LBR's source, which we named linear unwinding. As instruction from linear range can come from different function by inlining, linear unwinding will do the range splitting and record counters through the range with same inline context. With each fall through path from LBR unwinding, we aggregate each sample into counters by the calling context and eventually generate full context sensitive profile (without relying on inlining) to driver compiler's PGO/FDO. A breakdown of noteworthy changes: - Added `HybridSample` class as the abstraction perf sample including LBR stack and call stack * Extended `PerfReader` to implement auto-detect whether input perf script output contains CS profile, then do the parsing. Multiple `HybridSample` are extracted * Speed up by aggregating `HybridSample` into `AggregatedSamples` * Added VirtualUnwinder that consumes aggregated `HybridSample` and implements unwinding of calls, returns, and linear path that contains implicit call/return from inlining. Ranges and branches counters are aggregated by the calling context.
 Here calling context is string type, each context is a pair of function name and callsite location info, the whole context is like `main:1 @ foo:2 @ bar`. * Added PorfileGenerater that accumulates counters by ranges unfolding or branch target mapping, then generates context-sensitive function profile including function body, inferring callee's head sample, callsite target samples, eventually records into ProfileMap.
 * Leveraged LLVM build-in(`SampleProfWriter`) writer to support different serialization format with no stop - `getCanonicalFnName` for callee name and name from ELF section - Added regression test for both unwinding and profile generation Test Plan: ninja & ninja check-llvm Reviewed By: hoy, wenlei, wmi Differential Revision: https://reviews.llvm.org/D89723
2020-10-19 12:55:59 -07:00
}
ProfileGeneratorBase::UseFSDiscriminator = Binary->useFSDiscriminator();
FunctionSamples::ProfileIsFS = Binary->useFSDiscriminator();
[CSSPGO][llvm-profgen] Context-sensitive profile data generation This stack of changes introduces `llvm-profgen` utility which generates a profile data file from given perf script data files for sample-based PGO. It’s part of(not only) the CSSPGO work. Specifically to support context-sensitive with/without pseudo probe profile, it implements a series of functionalities including perf trace parsing, instruction symbolization, LBR stack/call frame stack unwinding, pseudo probe decoding, etc. Also high throughput is achieved by multiple levels of sample aggregation and compatible format with one stop is generated at the end. Please refer to: https://groups.google.com/g/llvm-dev/c/1p1rdYbL93s for the CSSPGO RFC. This change supports context-sensitive profile data generation into llvm-profgen. With simultaneous sampling for LBR and call stack, we can identify leaf of LBR sample with calling context from stack sample . During the process of deriving fall through path from LBR entries, we unwind LBR by replaying all the calls and returns (including implicit calls/returns due to inlining) backwards on top of the sampled call stack. Then the state of call stack as we unwind through LBR always represents the calling context of current fall through path. we have two types of virtual unwinding 1) LBR unwinding and 2) linear range unwinding. Specifically, for each LBR entry which can be classified into call, return, regular branch, LBR unwinding will replay the operation by pushing, popping or switching leaf frame towards the call stack and since the initial call stack is most recently sampled, the replay should be in anti-execution order, i.e. for the regular case, pop the call stack when LBR is call, push frame on call stack when LBR is return. After each LBR processed, it also needs to align with the next LBR by going through instructions from previous LBR's target to current LBR's source, which we named linear unwinding. As instruction from linear range can come from different function by inlining, linear unwinding will do the range splitting and record counters through the range with same inline context. With each fall through path from LBR unwinding, we aggregate each sample into counters by the calling context and eventually generate full context sensitive profile (without relying on inlining) to driver compiler's PGO/FDO. A breakdown of noteworthy changes: - Added `HybridSample` class as the abstraction perf sample including LBR stack and call stack * Extended `PerfReader` to implement auto-detect whether input perf script output contains CS profile, then do the parsing. Multiple `HybridSample` are extracted * Speed up by aggregating `HybridSample` into `AggregatedSamples` * Added VirtualUnwinder that consumes aggregated `HybridSample` and implements unwinding of calls, returns, and linear path that contains implicit call/return from inlining. Ranges and branches counters are aggregated by the calling context.
 Here calling context is string type, each context is a pair of function name and callsite location info, the whole context is like `main:1 @ foo:2 @ bar`. * Added PorfileGenerater that accumulates counters by ranges unfolding or branch target mapping, then generates context-sensitive function profile including function body, inferring callee's head sample, callsite target samples, eventually records into ProfileMap.
 * Leveraged LLVM build-in(`SampleProfWriter`) writer to support different serialization format with no stop - `getCanonicalFnName` for callee name and name from ELF section - Added regression test for both unwinding and profile generation Test Plan: ninja & ninja check-llvm Reviewed By: hoy, wenlei, wmi Differential Revision: https://reviews.llvm.org/D89723
2020-10-19 12:55:59 -07:00
return Generator;
[CSSPGO][llvm-profgen] Context-sensitive profile data generation This stack of changes introduces `llvm-profgen` utility which generates a profile data file from given perf script data files for sample-based PGO. It’s part of(not only) the CSSPGO work. Specifically to support context-sensitive with/without pseudo probe profile, it implements a series of functionalities including perf trace parsing, instruction symbolization, LBR stack/call frame stack unwinding, pseudo probe decoding, etc. Also high throughput is achieved by multiple levels of sample aggregation and compatible format with one stop is generated at the end. Please refer to: https://groups.google.com/g/llvm-dev/c/1p1rdYbL93s for the CSSPGO RFC. This change supports context-sensitive profile data generation into llvm-profgen. With simultaneous sampling for LBR and call stack, we can identify leaf of LBR sample with calling context from stack sample . During the process of deriving fall through path from LBR entries, we unwind LBR by replaying all the calls and returns (including implicit calls/returns due to inlining) backwards on top of the sampled call stack. Then the state of call stack as we unwind through LBR always represents the calling context of current fall through path. we have two types of virtual unwinding 1) LBR unwinding and 2) linear range unwinding. Specifically, for each LBR entry which can be classified into call, return, regular branch, LBR unwinding will replay the operation by pushing, popping or switching leaf frame towards the call stack and since the initial call stack is most recently sampled, the replay should be in anti-execution order, i.e. for the regular case, pop the call stack when LBR is call, push frame on call stack when LBR is return. After each LBR processed, it also needs to align with the next LBR by going through instructions from previous LBR's target to current LBR's source, which we named linear unwinding. As instruction from linear range can come from different function by inlining, linear unwinding will do the range splitting and record counters through the range with same inline context. With each fall through path from LBR unwinding, we aggregate each sample into counters by the calling context and eventually generate full context sensitive profile (without relying on inlining) to driver compiler's PGO/FDO. A breakdown of noteworthy changes: - Added `HybridSample` class as the abstraction perf sample including LBR stack and call stack * Extended `PerfReader` to implement auto-detect whether input perf script output contains CS profile, then do the parsing. Multiple `HybridSample` are extracted * Speed up by aggregating `HybridSample` into `AggregatedSamples` * Added VirtualUnwinder that consumes aggregated `HybridSample` and implements unwinding of calls, returns, and linear path that contains implicit call/return from inlining. Ranges and branches counters are aggregated by the calling context.
 Here calling context is string type, each context is a pair of function name and callsite location info, the whole context is like `main:1 @ foo:2 @ bar`. * Added PorfileGenerater that accumulates counters by ranges unfolding or branch target mapping, then generates context-sensitive function profile including function body, inferring callee's head sample, callsite target samples, eventually records into ProfileMap.
 * Leveraged LLVM build-in(`SampleProfWriter`) writer to support different serialization format with no stop - `getCanonicalFnName` for callee name and name from ELF section - Added regression test for both unwinding and profile generation Test Plan: ninja & ninja check-llvm Reviewed By: hoy, wenlei, wmi Differential Revision: https://reviews.llvm.org/D89723
2020-10-19 12:55:59 -07:00
}
std::unique_ptr<ProfileGeneratorBase>
ProfileGeneratorBase::create(ProfiledBinary *Binary,
const SampleProfileMap &&Profiles,
bool ProfileIsCSFlat) {
std::unique_ptr<ProfileGeneratorBase> Generator;
if (ProfileIsCSFlat) {
if (Binary->useFSDiscriminator())
exitWithError("FS discriminator is not supported in CS profile.");
Generator.reset(new CSProfileGenerator(Binary, std::move(Profiles)));
} else {
Generator.reset(new ProfileGenerator(Binary, std::move(Profiles)));
}
ProfileGeneratorBase::UseFSDiscriminator = Binary->useFSDiscriminator();
FunctionSamples::ProfileIsFS = Binary->useFSDiscriminator();
return Generator;
}
void ProfileGeneratorBase::write(std::unique_ptr<SampleProfileWriter> Writer,
SampleProfileMap &ProfileMap) {
// Populate profile symbol list if extended binary format is used.
ProfileSymbolList SymbolList;
if (PopulateProfileSymbolList && OutputFormat == SPF_Ext_Binary) {
Binary->populateSymbolListFromDWARF(SymbolList);
Writer->setProfileSymbolList(&SymbolList);
}
if (std::error_code EC = Writer->write(ProfileMap))
exitWithError(std::move(EC));
}
void ProfileGeneratorBase::write() {
[CSSPGO][llvm-profgen] Context-sensitive profile data generation This stack of changes introduces `llvm-profgen` utility which generates a profile data file from given perf script data files for sample-based PGO. It’s part of(not only) the CSSPGO work. Specifically to support context-sensitive with/without pseudo probe profile, it implements a series of functionalities including perf trace parsing, instruction symbolization, LBR stack/call frame stack unwinding, pseudo probe decoding, etc. Also high throughput is achieved by multiple levels of sample aggregation and compatible format with one stop is generated at the end. Please refer to: https://groups.google.com/g/llvm-dev/c/1p1rdYbL93s for the CSSPGO RFC. This change supports context-sensitive profile data generation into llvm-profgen. With simultaneous sampling for LBR and call stack, we can identify leaf of LBR sample with calling context from stack sample . During the process of deriving fall through path from LBR entries, we unwind LBR by replaying all the calls and returns (including implicit calls/returns due to inlining) backwards on top of the sampled call stack. Then the state of call stack as we unwind through LBR always represents the calling context of current fall through path. we have two types of virtual unwinding 1) LBR unwinding and 2) linear range unwinding. Specifically, for each LBR entry which can be classified into call, return, regular branch, LBR unwinding will replay the operation by pushing, popping or switching leaf frame towards the call stack and since the initial call stack is most recently sampled, the replay should be in anti-execution order, i.e. for the regular case, pop the call stack when LBR is call, push frame on call stack when LBR is return. After each LBR processed, it also needs to align with the next LBR by going through instructions from previous LBR's target to current LBR's source, which we named linear unwinding. As instruction from linear range can come from different function by inlining, linear unwinding will do the range splitting and record counters through the range with same inline context. With each fall through path from LBR unwinding, we aggregate each sample into counters by the calling context and eventually generate full context sensitive profile (without relying on inlining) to driver compiler's PGO/FDO. A breakdown of noteworthy changes: - Added `HybridSample` class as the abstraction perf sample including LBR stack and call stack * Extended `PerfReader` to implement auto-detect whether input perf script output contains CS profile, then do the parsing. Multiple `HybridSample` are extracted * Speed up by aggregating `HybridSample` into `AggregatedSamples` * Added VirtualUnwinder that consumes aggregated `HybridSample` and implements unwinding of calls, returns, and linear path that contains implicit call/return from inlining. Ranges and branches counters are aggregated by the calling context.
 Here calling context is string type, each context is a pair of function name and callsite location info, the whole context is like `main:1 @ foo:2 @ bar`. * Added PorfileGenerater that accumulates counters by ranges unfolding or branch target mapping, then generates context-sensitive function profile including function body, inferring callee's head sample, callsite target samples, eventually records into ProfileMap.
 * Leveraged LLVM build-in(`SampleProfWriter`) writer to support different serialization format with no stop - `getCanonicalFnName` for callee name and name from ELF section - Added regression test for both unwinding and profile generation Test Plan: ninja & ninja check-llvm Reviewed By: hoy, wenlei, wmi Differential Revision: https://reviews.llvm.org/D89723
2020-10-19 12:55:59 -07:00
auto WriterOrErr = SampleProfileWriter::create(OutputFilename, OutputFormat);
if (std::error_code EC = WriterOrErr.getError())
exitWithError(EC, OutputFilename);
if (UseMD5) {
if (OutputFormat != SPF_Ext_Binary)
WithColor::warning() << "-use-md5 is ignored. Specify "
"--format=extbinary to enable it\n";
else
WriterOrErr.get()->setUseMD5();
}
write(std::move(WriterOrErr.get()), ProfileMap);
[CSSPGO][llvm-profgen] Context-sensitive profile data generation This stack of changes introduces `llvm-profgen` utility which generates a profile data file from given perf script data files for sample-based PGO. It’s part of(not only) the CSSPGO work. Specifically to support context-sensitive with/without pseudo probe profile, it implements a series of functionalities including perf trace parsing, instruction symbolization, LBR stack/call frame stack unwinding, pseudo probe decoding, etc. Also high throughput is achieved by multiple levels of sample aggregation and compatible format with one stop is generated at the end. Please refer to: https://groups.google.com/g/llvm-dev/c/1p1rdYbL93s for the CSSPGO RFC. This change supports context-sensitive profile data generation into llvm-profgen. With simultaneous sampling for LBR and call stack, we can identify leaf of LBR sample with calling context from stack sample . During the process of deriving fall through path from LBR entries, we unwind LBR by replaying all the calls and returns (including implicit calls/returns due to inlining) backwards on top of the sampled call stack. Then the state of call stack as we unwind through LBR always represents the calling context of current fall through path. we have two types of virtual unwinding 1) LBR unwinding and 2) linear range unwinding. Specifically, for each LBR entry which can be classified into call, return, regular branch, LBR unwinding will replay the operation by pushing, popping or switching leaf frame towards the call stack and since the initial call stack is most recently sampled, the replay should be in anti-execution order, i.e. for the regular case, pop the call stack when LBR is call, push frame on call stack when LBR is return. After each LBR processed, it also needs to align with the next LBR by going through instructions from previous LBR's target to current LBR's source, which we named linear unwinding. As instruction from linear range can come from different function by inlining, linear unwinding will do the range splitting and record counters through the range with same inline context. With each fall through path from LBR unwinding, we aggregate each sample into counters by the calling context and eventually generate full context sensitive profile (without relying on inlining) to driver compiler's PGO/FDO. A breakdown of noteworthy changes: - Added `HybridSample` class as the abstraction perf sample including LBR stack and call stack * Extended `PerfReader` to implement auto-detect whether input perf script output contains CS profile, then do the parsing. Multiple `HybridSample` are extracted * Speed up by aggregating `HybridSample` into `AggregatedSamples` * Added VirtualUnwinder that consumes aggregated `HybridSample` and implements unwinding of calls, returns, and linear path that contains implicit call/return from inlining. Ranges and branches counters are aggregated by the calling context.
 Here calling context is string type, each context is a pair of function name and callsite location info, the whole context is like `main:1 @ foo:2 @ bar`. * Added PorfileGenerater that accumulates counters by ranges unfolding or branch target mapping, then generates context-sensitive function profile including function body, inferring callee's head sample, callsite target samples, eventually records into ProfileMap.
 * Leveraged LLVM build-in(`SampleProfWriter`) writer to support different serialization format with no stop - `getCanonicalFnName` for callee name and name from ELF section - Added regression test for both unwinding and profile generation Test Plan: ninja & ninja check-llvm Reviewed By: hoy, wenlei, wmi Differential Revision: https://reviews.llvm.org/D89723
2020-10-19 12:55:59 -07:00
}
void ProfileGeneratorBase::showDensitySuggestion(double Density) {
if (Density == 0.0)
WithColor::warning() << "The --profile-summary-cutoff-hot option may be "
"set too low. Please check your command.\n";
else if (Density < HotFunctionDensityThreshold)
WithColor::warning()
<< "AutoFDO is estimated to optimize better with "
<< format("%.1f", HotFunctionDensityThreshold / Density)
<< "x more samples. Please consider increasing sampling rate or "
"profiling for longer duration to get more samples.\n";
if (ShowDensity)
outs() << "Minimum profile density for hot functions with top "
<< format("%.2f",
static_cast<double>(ProfileSummaryCutoffHot.getValue()) /
10000)
<< "% total samples: " << format("%.1f", Density) << "\n";
}
double ProfileGeneratorBase::calculateDensity(const SampleProfileMap &Profiles,
uint64_t HotCntThreshold) {
double Density = DBL_MAX;
std::vector<const FunctionSamples *> HotFuncs;
for (auto &I : Profiles) {
auto &FuncSamples = I.second;
if (FuncSamples.getTotalSamples() < HotCntThreshold)
continue;
HotFuncs.emplace_back(&FuncSamples);
}
for (auto *FuncSamples : HotFuncs) {
auto *Func = Binary->getBinaryFunction(FuncSamples->getName());
if (!Func)
continue;
uint64_t FuncSize = Func->getFuncSize();
if (FuncSize == 0)
continue;
Density =
std::min(Density, static_cast<double>(FuncSamples->getTotalSamples()) /
FuncSize);
}
return Density == DBL_MAX ? 0.0 : Density;
}
void ProfileGeneratorBase::findDisjointRanges(RangeSample &DisjointRanges,
const RangeSample &Ranges) {
[CSSPGO][llvm-profgen] Context-sensitive profile data generation This stack of changes introduces `llvm-profgen` utility which generates a profile data file from given perf script data files for sample-based PGO. It’s part of(not only) the CSSPGO work. Specifically to support context-sensitive with/without pseudo probe profile, it implements a series of functionalities including perf trace parsing, instruction symbolization, LBR stack/call frame stack unwinding, pseudo probe decoding, etc. Also high throughput is achieved by multiple levels of sample aggregation and compatible format with one stop is generated at the end. Please refer to: https://groups.google.com/g/llvm-dev/c/1p1rdYbL93s for the CSSPGO RFC. This change supports context-sensitive profile data generation into llvm-profgen. With simultaneous sampling for LBR and call stack, we can identify leaf of LBR sample with calling context from stack sample . During the process of deriving fall through path from LBR entries, we unwind LBR by replaying all the calls and returns (including implicit calls/returns due to inlining) backwards on top of the sampled call stack. Then the state of call stack as we unwind through LBR always represents the calling context of current fall through path. we have two types of virtual unwinding 1) LBR unwinding and 2) linear range unwinding. Specifically, for each LBR entry which can be classified into call, return, regular branch, LBR unwinding will replay the operation by pushing, popping or switching leaf frame towards the call stack and since the initial call stack is most recently sampled, the replay should be in anti-execution order, i.e. for the regular case, pop the call stack when LBR is call, push frame on call stack when LBR is return. After each LBR processed, it also needs to align with the next LBR by going through instructions from previous LBR's target to current LBR's source, which we named linear unwinding. As instruction from linear range can come from different function by inlining, linear unwinding will do the range splitting and record counters through the range with same inline context. With each fall through path from LBR unwinding, we aggregate each sample into counters by the calling context and eventually generate full context sensitive profile (without relying on inlining) to driver compiler's PGO/FDO. A breakdown of noteworthy changes: - Added `HybridSample` class as the abstraction perf sample including LBR stack and call stack * Extended `PerfReader` to implement auto-detect whether input perf script output contains CS profile, then do the parsing. Multiple `HybridSample` are extracted * Speed up by aggregating `HybridSample` into `AggregatedSamples` * Added VirtualUnwinder that consumes aggregated `HybridSample` and implements unwinding of calls, returns, and linear path that contains implicit call/return from inlining. Ranges and branches counters are aggregated by the calling context.
 Here calling context is string type, each context is a pair of function name and callsite location info, the whole context is like `main:1 @ foo:2 @ bar`. * Added PorfileGenerater that accumulates counters by ranges unfolding or branch target mapping, then generates context-sensitive function profile including function body, inferring callee's head sample, callsite target samples, eventually records into ProfileMap.
 * Leveraged LLVM build-in(`SampleProfWriter`) writer to support different serialization format with no stop - `getCanonicalFnName` for callee name and name from ELF section - Added regression test for both unwinding and profile generation Test Plan: ninja & ninja check-llvm Reviewed By: hoy, wenlei, wmi Differential Revision: https://reviews.llvm.org/D89723
2020-10-19 12:55:59 -07:00
/*
Regions may overlap with each other. Using the boundary info, find all
disjoint ranges and their sample count. BoundaryPoint contains the count
2021-01-11 09:08:39 -08:00
multiple samples begin/end at this points.
[CSSPGO][llvm-profgen] Context-sensitive profile data generation This stack of changes introduces `llvm-profgen` utility which generates a profile data file from given perf script data files for sample-based PGO. It’s part of(not only) the CSSPGO work. Specifically to support context-sensitive with/without pseudo probe profile, it implements a series of functionalities including perf trace parsing, instruction symbolization, LBR stack/call frame stack unwinding, pseudo probe decoding, etc. Also high throughput is achieved by multiple levels of sample aggregation and compatible format with one stop is generated at the end. Please refer to: https://groups.google.com/g/llvm-dev/c/1p1rdYbL93s for the CSSPGO RFC. This change supports context-sensitive profile data generation into llvm-profgen. With simultaneous sampling for LBR and call stack, we can identify leaf of LBR sample with calling context from stack sample . During the process of deriving fall through path from LBR entries, we unwind LBR by replaying all the calls and returns (including implicit calls/returns due to inlining) backwards on top of the sampled call stack. Then the state of call stack as we unwind through LBR always represents the calling context of current fall through path. we have two types of virtual unwinding 1) LBR unwinding and 2) linear range unwinding. Specifically, for each LBR entry which can be classified into call, return, regular branch, LBR unwinding will replay the operation by pushing, popping or switching leaf frame towards the call stack and since the initial call stack is most recently sampled, the replay should be in anti-execution order, i.e. for the regular case, pop the call stack when LBR is call, push frame on call stack when LBR is return. After each LBR processed, it also needs to align with the next LBR by going through instructions from previous LBR's target to current LBR's source, which we named linear unwinding. As instruction from linear range can come from different function by inlining, linear unwinding will do the range splitting and record counters through the range with same inline context. With each fall through path from LBR unwinding, we aggregate each sample into counters by the calling context and eventually generate full context sensitive profile (without relying on inlining) to driver compiler's PGO/FDO. A breakdown of noteworthy changes: - Added `HybridSample` class as the abstraction perf sample including LBR stack and call stack * Extended `PerfReader` to implement auto-detect whether input perf script output contains CS profile, then do the parsing. Multiple `HybridSample` are extracted * Speed up by aggregating `HybridSample` into `AggregatedSamples` * Added VirtualUnwinder that consumes aggregated `HybridSample` and implements unwinding of calls, returns, and linear path that contains implicit call/return from inlining. Ranges and branches counters are aggregated by the calling context.
 Here calling context is string type, each context is a pair of function name and callsite location info, the whole context is like `main:1 @ foo:2 @ bar`. * Added PorfileGenerater that accumulates counters by ranges unfolding or branch target mapping, then generates context-sensitive function profile including function body, inferring callee's head sample, callsite target samples, eventually records into ProfileMap.
 * Leveraged LLVM build-in(`SampleProfWriter`) writer to support different serialization format with no stop - `getCanonicalFnName` for callee name and name from ELF section - Added regression test for both unwinding and profile generation Test Plan: ninja & ninja check-llvm Reviewed By: hoy, wenlei, wmi Differential Revision: https://reviews.llvm.org/D89723
2020-10-19 12:55:59 -07:00
|<--100-->| Sample1
|<------200------>| Sample2
A B C
In the example above,
Sample1 begins at A, ends at B, its value is 100.
Sample2 beings at A, ends at C, its value is 200.
For A, BeginCount is the sum of sample begins at A, which is 300 and no
samples ends at A, so EndCount is 0.
Then boundary points A, B, and C with begin/end counts are:
A: (300, 0)
B: (0, 100)
C: (0, 200)
*/
struct BoundaryPoint {
// Sum of sample counts beginning at this point
uint64_t BeginCount = UINT64_MAX;
[CSSPGO][llvm-profgen] Context-sensitive profile data generation This stack of changes introduces `llvm-profgen` utility which generates a profile data file from given perf script data files for sample-based PGO. It’s part of(not only) the CSSPGO work. Specifically to support context-sensitive with/without pseudo probe profile, it implements a series of functionalities including perf trace parsing, instruction symbolization, LBR stack/call frame stack unwinding, pseudo probe decoding, etc. Also high throughput is achieved by multiple levels of sample aggregation and compatible format with one stop is generated at the end. Please refer to: https://groups.google.com/g/llvm-dev/c/1p1rdYbL93s for the CSSPGO RFC. This change supports context-sensitive profile data generation into llvm-profgen. With simultaneous sampling for LBR and call stack, we can identify leaf of LBR sample with calling context from stack sample . During the process of deriving fall through path from LBR entries, we unwind LBR by replaying all the calls and returns (including implicit calls/returns due to inlining) backwards on top of the sampled call stack. Then the state of call stack as we unwind through LBR always represents the calling context of current fall through path. we have two types of virtual unwinding 1) LBR unwinding and 2) linear range unwinding. Specifically, for each LBR entry which can be classified into call, return, regular branch, LBR unwinding will replay the operation by pushing, popping or switching leaf frame towards the call stack and since the initial call stack is most recently sampled, the replay should be in anti-execution order, i.e. for the regular case, pop the call stack when LBR is call, push frame on call stack when LBR is return. After each LBR processed, it also needs to align with the next LBR by going through instructions from previous LBR's target to current LBR's source, which we named linear unwinding. As instruction from linear range can come from different function by inlining, linear unwinding will do the range splitting and record counters through the range with same inline context. With each fall through path from LBR unwinding, we aggregate each sample into counters by the calling context and eventually generate full context sensitive profile (without relying on inlining) to driver compiler's PGO/FDO. A breakdown of noteworthy changes: - Added `HybridSample` class as the abstraction perf sample including LBR stack and call stack * Extended `PerfReader` to implement auto-detect whether input perf script output contains CS profile, then do the parsing. Multiple `HybridSample` are extracted * Speed up by aggregating `HybridSample` into `AggregatedSamples` * Added VirtualUnwinder that consumes aggregated `HybridSample` and implements unwinding of calls, returns, and linear path that contains implicit call/return from inlining. Ranges and branches counters are aggregated by the calling context.
 Here calling context is string type, each context is a pair of function name and callsite location info, the whole context is like `main:1 @ foo:2 @ bar`. * Added PorfileGenerater that accumulates counters by ranges unfolding or branch target mapping, then generates context-sensitive function profile including function body, inferring callee's head sample, callsite target samples, eventually records into ProfileMap.
 * Leveraged LLVM build-in(`SampleProfWriter`) writer to support different serialization format with no stop - `getCanonicalFnName` for callee name and name from ELF section - Added regression test for both unwinding and profile generation Test Plan: ninja & ninja check-llvm Reviewed By: hoy, wenlei, wmi Differential Revision: https://reviews.llvm.org/D89723
2020-10-19 12:55:59 -07:00
// Sum of sample counts ending at this point
uint64_t EndCount = UINT64_MAX;
// Is the begin point of a zero range.
bool IsZeroRangeBegin = false;
// Is the end point of a zero range.
bool IsZeroRangeEnd = false;
void addBeginCount(uint64_t Count) {
if (BeginCount == UINT64_MAX)
BeginCount = 0;
BeginCount += Count;
}
[CSSPGO][llvm-profgen] Context-sensitive profile data generation This stack of changes introduces `llvm-profgen` utility which generates a profile data file from given perf script data files for sample-based PGO. It’s part of(not only) the CSSPGO work. Specifically to support context-sensitive with/without pseudo probe profile, it implements a series of functionalities including perf trace parsing, instruction symbolization, LBR stack/call frame stack unwinding, pseudo probe decoding, etc. Also high throughput is achieved by multiple levels of sample aggregation and compatible format with one stop is generated at the end. Please refer to: https://groups.google.com/g/llvm-dev/c/1p1rdYbL93s for the CSSPGO RFC. This change supports context-sensitive profile data generation into llvm-profgen. With simultaneous sampling for LBR and call stack, we can identify leaf of LBR sample with calling context from stack sample . During the process of deriving fall through path from LBR entries, we unwind LBR by replaying all the calls and returns (including implicit calls/returns due to inlining) backwards on top of the sampled call stack. Then the state of call stack as we unwind through LBR always represents the calling context of current fall through path. we have two types of virtual unwinding 1) LBR unwinding and 2) linear range unwinding. Specifically, for each LBR entry which can be classified into call, return, regular branch, LBR unwinding will replay the operation by pushing, popping or switching leaf frame towards the call stack and since the initial call stack is most recently sampled, the replay should be in anti-execution order, i.e. for the regular case, pop the call stack when LBR is call, push frame on call stack when LBR is return. After each LBR processed, it also needs to align with the next LBR by going through instructions from previous LBR's target to current LBR's source, which we named linear unwinding. As instruction from linear range can come from different function by inlining, linear unwinding will do the range splitting and record counters through the range with same inline context. With each fall through path from LBR unwinding, we aggregate each sample into counters by the calling context and eventually generate full context sensitive profile (without relying on inlining) to driver compiler's PGO/FDO. A breakdown of noteworthy changes: - Added `HybridSample` class as the abstraction perf sample including LBR stack and call stack * Extended `PerfReader` to implement auto-detect whether input perf script output contains CS profile, then do the parsing. Multiple `HybridSample` are extracted * Speed up by aggregating `HybridSample` into `AggregatedSamples` * Added VirtualUnwinder that consumes aggregated `HybridSample` and implements unwinding of calls, returns, and linear path that contains implicit call/return from inlining. Ranges and branches counters are aggregated by the calling context.
 Here calling context is string type, each context is a pair of function name and callsite location info, the whole context is like `main:1 @ foo:2 @ bar`. * Added PorfileGenerater that accumulates counters by ranges unfolding or branch target mapping, then generates context-sensitive function profile including function body, inferring callee's head sample, callsite target samples, eventually records into ProfileMap.
 * Leveraged LLVM build-in(`SampleProfWriter`) writer to support different serialization format with no stop - `getCanonicalFnName` for callee name and name from ELF section - Added regression test for both unwinding and profile generation Test Plan: ninja & ninja check-llvm Reviewed By: hoy, wenlei, wmi Differential Revision: https://reviews.llvm.org/D89723
2020-10-19 12:55:59 -07:00
void addEndCount(uint64_t Count) {
if (EndCount == UINT64_MAX)
EndCount = 0;
EndCount += Count;
}
[CSSPGO][llvm-profgen] Context-sensitive profile data generation This stack of changes introduces `llvm-profgen` utility which generates a profile data file from given perf script data files for sample-based PGO. It’s part of(not only) the CSSPGO work. Specifically to support context-sensitive with/without pseudo probe profile, it implements a series of functionalities including perf trace parsing, instruction symbolization, LBR stack/call frame stack unwinding, pseudo probe decoding, etc. Also high throughput is achieved by multiple levels of sample aggregation and compatible format with one stop is generated at the end. Please refer to: https://groups.google.com/g/llvm-dev/c/1p1rdYbL93s for the CSSPGO RFC. This change supports context-sensitive profile data generation into llvm-profgen. With simultaneous sampling for LBR and call stack, we can identify leaf of LBR sample with calling context from stack sample . During the process of deriving fall through path from LBR entries, we unwind LBR by replaying all the calls and returns (including implicit calls/returns due to inlining) backwards on top of the sampled call stack. Then the state of call stack as we unwind through LBR always represents the calling context of current fall through path. we have two types of virtual unwinding 1) LBR unwinding and 2) linear range unwinding. Specifically, for each LBR entry which can be classified into call, return, regular branch, LBR unwinding will replay the operation by pushing, popping or switching leaf frame towards the call stack and since the initial call stack is most recently sampled, the replay should be in anti-execution order, i.e. for the regular case, pop the call stack when LBR is call, push frame on call stack when LBR is return. After each LBR processed, it also needs to align with the next LBR by going through instructions from previous LBR's target to current LBR's source, which we named linear unwinding. As instruction from linear range can come from different function by inlining, linear unwinding will do the range splitting and record counters through the range with same inline context. With each fall through path from LBR unwinding, we aggregate each sample into counters by the calling context and eventually generate full context sensitive profile (without relying on inlining) to driver compiler's PGO/FDO. A breakdown of noteworthy changes: - Added `HybridSample` class as the abstraction perf sample including LBR stack and call stack * Extended `PerfReader` to implement auto-detect whether input perf script output contains CS profile, then do the parsing. Multiple `HybridSample` are extracted * Speed up by aggregating `HybridSample` into `AggregatedSamples` * Added VirtualUnwinder that consumes aggregated `HybridSample` and implements unwinding of calls, returns, and linear path that contains implicit call/return from inlining. Ranges and branches counters are aggregated by the calling context.
 Here calling context is string type, each context is a pair of function name and callsite location info, the whole context is like `main:1 @ foo:2 @ bar`. * Added PorfileGenerater that accumulates counters by ranges unfolding or branch target mapping, then generates context-sensitive function profile including function body, inferring callee's head sample, callsite target samples, eventually records into ProfileMap.
 * Leveraged LLVM build-in(`SampleProfWriter`) writer to support different serialization format with no stop - `getCanonicalFnName` for callee name and name from ELF section - Added regression test for both unwinding and profile generation Test Plan: ninja & ninja check-llvm Reviewed By: hoy, wenlei, wmi Differential Revision: https://reviews.llvm.org/D89723
2020-10-19 12:55:59 -07:00
};
/*
For the above example. With boundary points, follwing logic finds two
disjoint region of
[A,B]: 300
[B+1,C]: 200
If there is a boundary point that both begin and end, the point itself
becomes a separate disjoint region. For example, if we have original
ranges of
|<--- 100 --->|
|<--- 200 --->|
A B C
there are three boundary points with their begin/end counts of
A: (100, 0)
B: (200, 100)
C: (0, 200)
the disjoint ranges would be
[A, B-1]: 100
[B, B]: 300
[B+1, C]: 200.
Example for zero value range:
|<--- 100 --->|
|<--- 200 --->|
|<--------------- 0 ----------------->|
A B C D E F
[A, B-1] : 0
[B, C] : 100
[C+1, D-1]: 0
[D, E] : 200
[E+1, F] : 0
[CSSPGO][llvm-profgen] Context-sensitive profile data generation This stack of changes introduces `llvm-profgen` utility which generates a profile data file from given perf script data files for sample-based PGO. It’s part of(not only) the CSSPGO work. Specifically to support context-sensitive with/without pseudo probe profile, it implements a series of functionalities including perf trace parsing, instruction symbolization, LBR stack/call frame stack unwinding, pseudo probe decoding, etc. Also high throughput is achieved by multiple levels of sample aggregation and compatible format with one stop is generated at the end. Please refer to: https://groups.google.com/g/llvm-dev/c/1p1rdYbL93s for the CSSPGO RFC. This change supports context-sensitive profile data generation into llvm-profgen. With simultaneous sampling for LBR and call stack, we can identify leaf of LBR sample with calling context from stack sample . During the process of deriving fall through path from LBR entries, we unwind LBR by replaying all the calls and returns (including implicit calls/returns due to inlining) backwards on top of the sampled call stack. Then the state of call stack as we unwind through LBR always represents the calling context of current fall through path. we have two types of virtual unwinding 1) LBR unwinding and 2) linear range unwinding. Specifically, for each LBR entry which can be classified into call, return, regular branch, LBR unwinding will replay the operation by pushing, popping or switching leaf frame towards the call stack and since the initial call stack is most recently sampled, the replay should be in anti-execution order, i.e. for the regular case, pop the call stack when LBR is call, push frame on call stack when LBR is return. After each LBR processed, it also needs to align with the next LBR by going through instructions from previous LBR's target to current LBR's source, which we named linear unwinding. As instruction from linear range can come from different function by inlining, linear unwinding will do the range splitting and record counters through the range with same inline context. With each fall through path from LBR unwinding, we aggregate each sample into counters by the calling context and eventually generate full context sensitive profile (without relying on inlining) to driver compiler's PGO/FDO. A breakdown of noteworthy changes: - Added `HybridSample` class as the abstraction perf sample including LBR stack and call stack * Extended `PerfReader` to implement auto-detect whether input perf script output contains CS profile, then do the parsing. Multiple `HybridSample` are extracted * Speed up by aggregating `HybridSample` into `AggregatedSamples` * Added VirtualUnwinder that consumes aggregated `HybridSample` and implements unwinding of calls, returns, and linear path that contains implicit call/return from inlining. Ranges and branches counters are aggregated by the calling context.
 Here calling context is string type, each context is a pair of function name and callsite location info, the whole context is like `main:1 @ foo:2 @ bar`. * Added PorfileGenerater that accumulates counters by ranges unfolding or branch target mapping, then generates context-sensitive function profile including function body, inferring callee's head sample, callsite target samples, eventually records into ProfileMap.
 * Leveraged LLVM build-in(`SampleProfWriter`) writer to support different serialization format with no stop - `getCanonicalFnName` for callee name and name from ELF section - Added regression test for both unwinding and profile generation Test Plan: ninja & ninja check-llvm Reviewed By: hoy, wenlei, wmi Differential Revision: https://reviews.llvm.org/D89723
2020-10-19 12:55:59 -07:00
*/
std::map<uint64_t, BoundaryPoint> Boundaries;
for (const auto &Item : Ranges) {
assert(Item.first.first <= Item.first.second &&
"Invalid instruction range");
auto &BeginPoint = Boundaries[Item.first.first];
auto &EndPoint = Boundaries[Item.first.second];
[CSSPGO][llvm-profgen] Context-sensitive profile data generation This stack of changes introduces `llvm-profgen` utility which generates a profile data file from given perf script data files for sample-based PGO. It’s part of(not only) the CSSPGO work. Specifically to support context-sensitive with/without pseudo probe profile, it implements a series of functionalities including perf trace parsing, instruction symbolization, LBR stack/call frame stack unwinding, pseudo probe decoding, etc. Also high throughput is achieved by multiple levels of sample aggregation and compatible format with one stop is generated at the end. Please refer to: https://groups.google.com/g/llvm-dev/c/1p1rdYbL93s for the CSSPGO RFC. This change supports context-sensitive profile data generation into llvm-profgen. With simultaneous sampling for LBR and call stack, we can identify leaf of LBR sample with calling context from stack sample . During the process of deriving fall through path from LBR entries, we unwind LBR by replaying all the calls and returns (including implicit calls/returns due to inlining) backwards on top of the sampled call stack. Then the state of call stack as we unwind through LBR always represents the calling context of current fall through path. we have two types of virtual unwinding 1) LBR unwinding and 2) linear range unwinding. Specifically, for each LBR entry which can be classified into call, return, regular branch, LBR unwinding will replay the operation by pushing, popping or switching leaf frame towards the call stack and since the initial call stack is most recently sampled, the replay should be in anti-execution order, i.e. for the regular case, pop the call stack when LBR is call, push frame on call stack when LBR is return. After each LBR processed, it also needs to align with the next LBR by going through instructions from previous LBR's target to current LBR's source, which we named linear unwinding. As instruction from linear range can come from different function by inlining, linear unwinding will do the range splitting and record counters through the range with same inline context. With each fall through path from LBR unwinding, we aggregate each sample into counters by the calling context and eventually generate full context sensitive profile (without relying on inlining) to driver compiler's PGO/FDO. A breakdown of noteworthy changes: - Added `HybridSample` class as the abstraction perf sample including LBR stack and call stack * Extended `PerfReader` to implement auto-detect whether input perf script output contains CS profile, then do the parsing. Multiple `HybridSample` are extracted * Speed up by aggregating `HybridSample` into `AggregatedSamples` * Added VirtualUnwinder that consumes aggregated `HybridSample` and implements unwinding of calls, returns, and linear path that contains implicit call/return from inlining. Ranges and branches counters are aggregated by the calling context.
 Here calling context is string type, each context is a pair of function name and callsite location info, the whole context is like `main:1 @ foo:2 @ bar`. * Added PorfileGenerater that accumulates counters by ranges unfolding or branch target mapping, then generates context-sensitive function profile including function body, inferring callee's head sample, callsite target samples, eventually records into ProfileMap.
 * Leveraged LLVM build-in(`SampleProfWriter`) writer to support different serialization format with no stop - `getCanonicalFnName` for callee name and name from ELF section - Added regression test for both unwinding and profile generation Test Plan: ninja & ninja check-llvm Reviewed By: hoy, wenlei, wmi Differential Revision: https://reviews.llvm.org/D89723
2020-10-19 12:55:59 -07:00
uint64_t Count = Item.second;
BeginPoint.addBeginCount(Count);
EndPoint.addEndCount(Count);
if (Count == 0) {
BeginPoint.IsZeroRangeBegin = true;
EndPoint.IsZeroRangeEnd = true;
}
[CSSPGO][llvm-profgen] Context-sensitive profile data generation This stack of changes introduces `llvm-profgen` utility which generates a profile data file from given perf script data files for sample-based PGO. It’s part of(not only) the CSSPGO work. Specifically to support context-sensitive with/without pseudo probe profile, it implements a series of functionalities including perf trace parsing, instruction symbolization, LBR stack/call frame stack unwinding, pseudo probe decoding, etc. Also high throughput is achieved by multiple levels of sample aggregation and compatible format with one stop is generated at the end. Please refer to: https://groups.google.com/g/llvm-dev/c/1p1rdYbL93s for the CSSPGO RFC. This change supports context-sensitive profile data generation into llvm-profgen. With simultaneous sampling for LBR and call stack, we can identify leaf of LBR sample with calling context from stack sample . During the process of deriving fall through path from LBR entries, we unwind LBR by replaying all the calls and returns (including implicit calls/returns due to inlining) backwards on top of the sampled call stack. Then the state of call stack as we unwind through LBR always represents the calling context of current fall through path. we have two types of virtual unwinding 1) LBR unwinding and 2) linear range unwinding. Specifically, for each LBR entry which can be classified into call, return, regular branch, LBR unwinding will replay the operation by pushing, popping or switching leaf frame towards the call stack and since the initial call stack is most recently sampled, the replay should be in anti-execution order, i.e. for the regular case, pop the call stack when LBR is call, push frame on call stack when LBR is return. After each LBR processed, it also needs to align with the next LBR by going through instructions from previous LBR's target to current LBR's source, which we named linear unwinding. As instruction from linear range can come from different function by inlining, linear unwinding will do the range splitting and record counters through the range with same inline context. With each fall through path from LBR unwinding, we aggregate each sample into counters by the calling context and eventually generate full context sensitive profile (without relying on inlining) to driver compiler's PGO/FDO. A breakdown of noteworthy changes: - Added `HybridSample` class as the abstraction perf sample including LBR stack and call stack * Extended `PerfReader` to implement auto-detect whether input perf script output contains CS profile, then do the parsing. Multiple `HybridSample` are extracted * Speed up by aggregating `HybridSample` into `AggregatedSamples` * Added VirtualUnwinder that consumes aggregated `HybridSample` and implements unwinding of calls, returns, and linear path that contains implicit call/return from inlining. Ranges and branches counters are aggregated by the calling context.
 Here calling context is string type, each context is a pair of function name and callsite location info, the whole context is like `main:1 @ foo:2 @ bar`. * Added PorfileGenerater that accumulates counters by ranges unfolding or branch target mapping, then generates context-sensitive function profile including function body, inferring callee's head sample, callsite target samples, eventually records into ProfileMap.
 * Leveraged LLVM build-in(`SampleProfWriter`) writer to support different serialization format with no stop - `getCanonicalFnName` for callee name and name from ELF section - Added regression test for both unwinding and profile generation Test Plan: ninja & ninja check-llvm Reviewed By: hoy, wenlei, wmi Differential Revision: https://reviews.llvm.org/D89723
2020-10-19 12:55:59 -07:00
}
// Use UINT64_MAX to indicate there is no existing range between BeginAddress
// and the next valid address
uint64_t BeginAddress = UINT64_MAX;
int ZeroRangeDepth = 0;
uint64_t Count = 0;
for (const auto &Item : Boundaries) {
[CSSPGO][llvm-profgen] Context-sensitive profile data generation This stack of changes introduces `llvm-profgen` utility which generates a profile data file from given perf script data files for sample-based PGO. It’s part of(not only) the CSSPGO work. Specifically to support context-sensitive with/without pseudo probe profile, it implements a series of functionalities including perf trace parsing, instruction symbolization, LBR stack/call frame stack unwinding, pseudo probe decoding, etc. Also high throughput is achieved by multiple levels of sample aggregation and compatible format with one stop is generated at the end. Please refer to: https://groups.google.com/g/llvm-dev/c/1p1rdYbL93s for the CSSPGO RFC. This change supports context-sensitive profile data generation into llvm-profgen. With simultaneous sampling for LBR and call stack, we can identify leaf of LBR sample with calling context from stack sample . During the process of deriving fall through path from LBR entries, we unwind LBR by replaying all the calls and returns (including implicit calls/returns due to inlining) backwards on top of the sampled call stack. Then the state of call stack as we unwind through LBR always represents the calling context of current fall through path. we have two types of virtual unwinding 1) LBR unwinding and 2) linear range unwinding. Specifically, for each LBR entry which can be classified into call, return, regular branch, LBR unwinding will replay the operation by pushing, popping or switching leaf frame towards the call stack and since the initial call stack is most recently sampled, the replay should be in anti-execution order, i.e. for the regular case, pop the call stack when LBR is call, push frame on call stack when LBR is return. After each LBR processed, it also needs to align with the next LBR by going through instructions from previous LBR's target to current LBR's source, which we named linear unwinding. As instruction from linear range can come from different function by inlining, linear unwinding will do the range splitting and record counters through the range with same inline context. With each fall through path from LBR unwinding, we aggregate each sample into counters by the calling context and eventually generate full context sensitive profile (without relying on inlining) to driver compiler's PGO/FDO. A breakdown of noteworthy changes: - Added `HybridSample` class as the abstraction perf sample including LBR stack and call stack * Extended `PerfReader` to implement auto-detect whether input perf script output contains CS profile, then do the parsing. Multiple `HybridSample` are extracted * Speed up by aggregating `HybridSample` into `AggregatedSamples` * Added VirtualUnwinder that consumes aggregated `HybridSample` and implements unwinding of calls, returns, and linear path that contains implicit call/return from inlining. Ranges and branches counters are aggregated by the calling context.
 Here calling context is string type, each context is a pair of function name and callsite location info, the whole context is like `main:1 @ foo:2 @ bar`. * Added PorfileGenerater that accumulates counters by ranges unfolding or branch target mapping, then generates context-sensitive function profile including function body, inferring callee's head sample, callsite target samples, eventually records into ProfileMap.
 * Leveraged LLVM build-in(`SampleProfWriter`) writer to support different serialization format with no stop - `getCanonicalFnName` for callee name and name from ELF section - Added regression test for both unwinding and profile generation Test Plan: ninja & ninja check-llvm Reviewed By: hoy, wenlei, wmi Differential Revision: https://reviews.llvm.org/D89723
2020-10-19 12:55:59 -07:00
uint64_t Address = Item.first;
const BoundaryPoint &Point = Item.second;
if (Point.BeginCount != UINT64_MAX) {
if (BeginAddress != UINT64_MAX)
[CSSPGO][llvm-profgen] Context-sensitive profile data generation This stack of changes introduces `llvm-profgen` utility which generates a profile data file from given perf script data files for sample-based PGO. It’s part of(not only) the CSSPGO work. Specifically to support context-sensitive with/without pseudo probe profile, it implements a series of functionalities including perf trace parsing, instruction symbolization, LBR stack/call frame stack unwinding, pseudo probe decoding, etc. Also high throughput is achieved by multiple levels of sample aggregation and compatible format with one stop is generated at the end. Please refer to: https://groups.google.com/g/llvm-dev/c/1p1rdYbL93s for the CSSPGO RFC. This change supports context-sensitive profile data generation into llvm-profgen. With simultaneous sampling for LBR and call stack, we can identify leaf of LBR sample with calling context from stack sample . During the process of deriving fall through path from LBR entries, we unwind LBR by replaying all the calls and returns (including implicit calls/returns due to inlining) backwards on top of the sampled call stack. Then the state of call stack as we unwind through LBR always represents the calling context of current fall through path. we have two types of virtual unwinding 1) LBR unwinding and 2) linear range unwinding. Specifically, for each LBR entry which can be classified into call, return, regular branch, LBR unwinding will replay the operation by pushing, popping or switching leaf frame towards the call stack and since the initial call stack is most recently sampled, the replay should be in anti-execution order, i.e. for the regular case, pop the call stack when LBR is call, push frame on call stack when LBR is return. After each LBR processed, it also needs to align with the next LBR by going through instructions from previous LBR's target to current LBR's source, which we named linear unwinding. As instruction from linear range can come from different function by inlining, linear unwinding will do the range splitting and record counters through the range with same inline context. With each fall through path from LBR unwinding, we aggregate each sample into counters by the calling context and eventually generate full context sensitive profile (without relying on inlining) to driver compiler's PGO/FDO. A breakdown of noteworthy changes: - Added `HybridSample` class as the abstraction perf sample including LBR stack and call stack * Extended `PerfReader` to implement auto-detect whether input perf script output contains CS profile, then do the parsing. Multiple `HybridSample` are extracted * Speed up by aggregating `HybridSample` into `AggregatedSamples` * Added VirtualUnwinder that consumes aggregated `HybridSample` and implements unwinding of calls, returns, and linear path that contains implicit call/return from inlining. Ranges and branches counters are aggregated by the calling context.
 Here calling context is string type, each context is a pair of function name and callsite location info, the whole context is like `main:1 @ foo:2 @ bar`. * Added PorfileGenerater that accumulates counters by ranges unfolding or branch target mapping, then generates context-sensitive function profile including function body, inferring callee's head sample, callsite target samples, eventually records into ProfileMap.
 * Leveraged LLVM build-in(`SampleProfWriter`) writer to support different serialization format with no stop - `getCanonicalFnName` for callee name and name from ELF section - Added regression test for both unwinding and profile generation Test Plan: ninja & ninja check-llvm Reviewed By: hoy, wenlei, wmi Differential Revision: https://reviews.llvm.org/D89723
2020-10-19 12:55:59 -07:00
DisjointRanges[{BeginAddress, Address - 1}] = Count;
Count += Point.BeginCount;
BeginAddress = Address;
ZeroRangeDepth += Point.IsZeroRangeBegin;
[CSSPGO][llvm-profgen] Context-sensitive profile data generation This stack of changes introduces `llvm-profgen` utility which generates a profile data file from given perf script data files for sample-based PGO. It’s part of(not only) the CSSPGO work. Specifically to support context-sensitive with/without pseudo probe profile, it implements a series of functionalities including perf trace parsing, instruction symbolization, LBR stack/call frame stack unwinding, pseudo probe decoding, etc. Also high throughput is achieved by multiple levels of sample aggregation and compatible format with one stop is generated at the end. Please refer to: https://groups.google.com/g/llvm-dev/c/1p1rdYbL93s for the CSSPGO RFC. This change supports context-sensitive profile data generation into llvm-profgen. With simultaneous sampling for LBR and call stack, we can identify leaf of LBR sample with calling context from stack sample . During the process of deriving fall through path from LBR entries, we unwind LBR by replaying all the calls and returns (including implicit calls/returns due to inlining) backwards on top of the sampled call stack. Then the state of call stack as we unwind through LBR always represents the calling context of current fall through path. we have two types of virtual unwinding 1) LBR unwinding and 2) linear range unwinding. Specifically, for each LBR entry which can be classified into call, return, regular branch, LBR unwinding will replay the operation by pushing, popping or switching leaf frame towards the call stack and since the initial call stack is most recently sampled, the replay should be in anti-execution order, i.e. for the regular case, pop the call stack when LBR is call, push frame on call stack when LBR is return. After each LBR processed, it also needs to align with the next LBR by going through instructions from previous LBR's target to current LBR's source, which we named linear unwinding. As instruction from linear range can come from different function by inlining, linear unwinding will do the range splitting and record counters through the range with same inline context. With each fall through path from LBR unwinding, we aggregate each sample into counters by the calling context and eventually generate full context sensitive profile (without relying on inlining) to driver compiler's PGO/FDO. A breakdown of noteworthy changes: - Added `HybridSample` class as the abstraction perf sample including LBR stack and call stack * Extended `PerfReader` to implement auto-detect whether input perf script output contains CS profile, then do the parsing. Multiple `HybridSample` are extracted * Speed up by aggregating `HybridSample` into `AggregatedSamples` * Added VirtualUnwinder that consumes aggregated `HybridSample` and implements unwinding of calls, returns, and linear path that contains implicit call/return from inlining. Ranges and branches counters are aggregated by the calling context.
 Here calling context is string type, each context is a pair of function name and callsite location info, the whole context is like `main:1 @ foo:2 @ bar`. * Added PorfileGenerater that accumulates counters by ranges unfolding or branch target mapping, then generates context-sensitive function profile including function body, inferring callee's head sample, callsite target samples, eventually records into ProfileMap.
 * Leveraged LLVM build-in(`SampleProfWriter`) writer to support different serialization format with no stop - `getCanonicalFnName` for callee name and name from ELF section - Added regression test for both unwinding and profile generation Test Plan: ninja & ninja check-llvm Reviewed By: hoy, wenlei, wmi Differential Revision: https://reviews.llvm.org/D89723
2020-10-19 12:55:59 -07:00
}
if (Point.EndCount != UINT64_MAX) {
assert((BeginAddress != UINT64_MAX) &&
"First boundary point cannot be 'end' point");
[CSSPGO][llvm-profgen] Context-sensitive profile data generation This stack of changes introduces `llvm-profgen` utility which generates a profile data file from given perf script data files for sample-based PGO. It’s part of(not only) the CSSPGO work. Specifically to support context-sensitive with/without pseudo probe profile, it implements a series of functionalities including perf trace parsing, instruction symbolization, LBR stack/call frame stack unwinding, pseudo probe decoding, etc. Also high throughput is achieved by multiple levels of sample aggregation and compatible format with one stop is generated at the end. Please refer to: https://groups.google.com/g/llvm-dev/c/1p1rdYbL93s for the CSSPGO RFC. This change supports context-sensitive profile data generation into llvm-profgen. With simultaneous sampling for LBR and call stack, we can identify leaf of LBR sample with calling context from stack sample . During the process of deriving fall through path from LBR entries, we unwind LBR by replaying all the calls and returns (including implicit calls/returns due to inlining) backwards on top of the sampled call stack. Then the state of call stack as we unwind through LBR always represents the calling context of current fall through path. we have two types of virtual unwinding 1) LBR unwinding and 2) linear range unwinding. Specifically, for each LBR entry which can be classified into call, return, regular branch, LBR unwinding will replay the operation by pushing, popping or switching leaf frame towards the call stack and since the initial call stack is most recently sampled, the replay should be in anti-execution order, i.e. for the regular case, pop the call stack when LBR is call, push frame on call stack when LBR is return. After each LBR processed, it also needs to align with the next LBR by going through instructions from previous LBR's target to current LBR's source, which we named linear unwinding. As instruction from linear range can come from different function by inlining, linear unwinding will do the range splitting and record counters through the range with same inline context. With each fall through path from LBR unwinding, we aggregate each sample into counters by the calling context and eventually generate full context sensitive profile (without relying on inlining) to driver compiler's PGO/FDO. A breakdown of noteworthy changes: - Added `HybridSample` class as the abstraction perf sample including LBR stack and call stack * Extended `PerfReader` to implement auto-detect whether input perf script output contains CS profile, then do the parsing. Multiple `HybridSample` are extracted * Speed up by aggregating `HybridSample` into `AggregatedSamples` * Added VirtualUnwinder that consumes aggregated `HybridSample` and implements unwinding of calls, returns, and linear path that contains implicit call/return from inlining. Ranges and branches counters are aggregated by the calling context.
 Here calling context is string type, each context is a pair of function name and callsite location info, the whole context is like `main:1 @ foo:2 @ bar`. * Added PorfileGenerater that accumulates counters by ranges unfolding or branch target mapping, then generates context-sensitive function profile including function body, inferring callee's head sample, callsite target samples, eventually records into ProfileMap.
 * Leveraged LLVM build-in(`SampleProfWriter`) writer to support different serialization format with no stop - `getCanonicalFnName` for callee name and name from ELF section - Added regression test for both unwinding and profile generation Test Plan: ninja & ninja check-llvm Reviewed By: hoy, wenlei, wmi Differential Revision: https://reviews.llvm.org/D89723
2020-10-19 12:55:59 -07:00
DisjointRanges[{BeginAddress, Address}] = Count;
assert(Count >= Point.EndCount && "Mismatched live ranges");
[CSSPGO][llvm-profgen] Context-sensitive profile data generation This stack of changes introduces `llvm-profgen` utility which generates a profile data file from given perf script data files for sample-based PGO. It’s part of(not only) the CSSPGO work. Specifically to support context-sensitive with/without pseudo probe profile, it implements a series of functionalities including perf trace parsing, instruction symbolization, LBR stack/call frame stack unwinding, pseudo probe decoding, etc. Also high throughput is achieved by multiple levels of sample aggregation and compatible format with one stop is generated at the end. Please refer to: https://groups.google.com/g/llvm-dev/c/1p1rdYbL93s for the CSSPGO RFC. This change supports context-sensitive profile data generation into llvm-profgen. With simultaneous sampling for LBR and call stack, we can identify leaf of LBR sample with calling context from stack sample . During the process of deriving fall through path from LBR entries, we unwind LBR by replaying all the calls and returns (including implicit calls/returns due to inlining) backwards on top of the sampled call stack. Then the state of call stack as we unwind through LBR always represents the calling context of current fall through path. we have two types of virtual unwinding 1) LBR unwinding and 2) linear range unwinding. Specifically, for each LBR entry which can be classified into call, return, regular branch, LBR unwinding will replay the operation by pushing, popping or switching leaf frame towards the call stack and since the initial call stack is most recently sampled, the replay should be in anti-execution order, i.e. for the regular case, pop the call stack when LBR is call, push frame on call stack when LBR is return. After each LBR processed, it also needs to align with the next LBR by going through instructions from previous LBR's target to current LBR's source, which we named linear unwinding. As instruction from linear range can come from different function by inlining, linear unwinding will do the range splitting and record counters through the range with same inline context. With each fall through path from LBR unwinding, we aggregate each sample into counters by the calling context and eventually generate full context sensitive profile (without relying on inlining) to driver compiler's PGO/FDO. A breakdown of noteworthy changes: - Added `HybridSample` class as the abstraction perf sample including LBR stack and call stack * Extended `PerfReader` to implement auto-detect whether input perf script output contains CS profile, then do the parsing. Multiple `HybridSample` are extracted * Speed up by aggregating `HybridSample` into `AggregatedSamples` * Added VirtualUnwinder that consumes aggregated `HybridSample` and implements unwinding of calls, returns, and linear path that contains implicit call/return from inlining. Ranges and branches counters are aggregated by the calling context.
 Here calling context is string type, each context is a pair of function name and callsite location info, the whole context is like `main:1 @ foo:2 @ bar`. * Added PorfileGenerater that accumulates counters by ranges unfolding or branch target mapping, then generates context-sensitive function profile including function body, inferring callee's head sample, callsite target samples, eventually records into ProfileMap.
 * Leveraged LLVM build-in(`SampleProfWriter`) writer to support different serialization format with no stop - `getCanonicalFnName` for callee name and name from ELF section - Added regression test for both unwinding and profile generation Test Plan: ninja & ninja check-llvm Reviewed By: hoy, wenlei, wmi Differential Revision: https://reviews.llvm.org/D89723
2020-10-19 12:55:59 -07:00
Count -= Point.EndCount;
BeginAddress = Address + 1;
ZeroRangeDepth -= Point.IsZeroRangeEnd;
// If the remaining count is zero and it's no longer in a zero range, this
// means we consume all the ranges before, thus mark BeginAddress as
// UINT64_MAX. e.g. supposing we have two non-overlapping ranges:
// [<---- 10 ---->]
// [<---- 20 ---->]
// A B C D
// The BeginAddress(B+1) will reset to invalid(UINT64_MAX), so we won't
// have the [B+1, C-1] zero range.
if (Count == 0 && ZeroRangeDepth == 0)
BeginAddress = UINT64_MAX;
[CSSPGO][llvm-profgen] Context-sensitive profile data generation This stack of changes introduces `llvm-profgen` utility which generates a profile data file from given perf script data files for sample-based PGO. It’s part of(not only) the CSSPGO work. Specifically to support context-sensitive with/without pseudo probe profile, it implements a series of functionalities including perf trace parsing, instruction symbolization, LBR stack/call frame stack unwinding, pseudo probe decoding, etc. Also high throughput is achieved by multiple levels of sample aggregation and compatible format with one stop is generated at the end. Please refer to: https://groups.google.com/g/llvm-dev/c/1p1rdYbL93s for the CSSPGO RFC. This change supports context-sensitive profile data generation into llvm-profgen. With simultaneous sampling for LBR and call stack, we can identify leaf of LBR sample with calling context from stack sample . During the process of deriving fall through path from LBR entries, we unwind LBR by replaying all the calls and returns (including implicit calls/returns due to inlining) backwards on top of the sampled call stack. Then the state of call stack as we unwind through LBR always represents the calling context of current fall through path. we have two types of virtual unwinding 1) LBR unwinding and 2) linear range unwinding. Specifically, for each LBR entry which can be classified into call, return, regular branch, LBR unwinding will replay the operation by pushing, popping or switching leaf frame towards the call stack and since the initial call stack is most recently sampled, the replay should be in anti-execution order, i.e. for the regular case, pop the call stack when LBR is call, push frame on call stack when LBR is return. After each LBR processed, it also needs to align with the next LBR by going through instructions from previous LBR's target to current LBR's source, which we named linear unwinding. As instruction from linear range can come from different function by inlining, linear unwinding will do the range splitting and record counters through the range with same inline context. With each fall through path from LBR unwinding, we aggregate each sample into counters by the calling context and eventually generate full context sensitive profile (without relying on inlining) to driver compiler's PGO/FDO. A breakdown of noteworthy changes: - Added `HybridSample` class as the abstraction perf sample including LBR stack and call stack * Extended `PerfReader` to implement auto-detect whether input perf script output contains CS profile, then do the parsing. Multiple `HybridSample` are extracted * Speed up by aggregating `HybridSample` into `AggregatedSamples` * Added VirtualUnwinder that consumes aggregated `HybridSample` and implements unwinding of calls, returns, and linear path that contains implicit call/return from inlining. Ranges and branches counters are aggregated by the calling context.
 Here calling context is string type, each context is a pair of function name and callsite location info, the whole context is like `main:1 @ foo:2 @ bar`. * Added PorfileGenerater that accumulates counters by ranges unfolding or branch target mapping, then generates context-sensitive function profile including function body, inferring callee's head sample, callsite target samples, eventually records into ProfileMap.
 * Leveraged LLVM build-in(`SampleProfWriter`) writer to support different serialization format with no stop - `getCanonicalFnName` for callee name and name from ELF section - Added regression test for both unwinding and profile generation Test Plan: ninja & ninja check-llvm Reviewed By: hoy, wenlei, wmi Differential Revision: https://reviews.llvm.org/D89723
2020-10-19 12:55:59 -07:00
}
}
}
void ProfileGeneratorBase::updateBodySamplesforFunctionProfile(
FunctionSamples &FunctionProfile, const SampleContextFrame &LeafLoc,
uint64_t Count) {
// Use the maximum count of samples with same line location
uint32_t Discriminator = getBaseDiscriminator(LeafLoc.Location.Discriminator);
// Use duplication factor to compensated for loop unroll/vectorization.
// Note that this is only needed when we're taking MAX of the counts at
// the location instead of SUM.
Count *= getDuplicationFactor(LeafLoc.Location.Discriminator);
ErrorOr<uint64_t> R =
FunctionProfile.findSamplesAt(LeafLoc.Location.LineOffset, Discriminator);
uint64_t PreviousCount = R ? R.get() : 0;
if (PreviousCount <= Count) {
FunctionProfile.addBodySamples(LeafLoc.Location.LineOffset, Discriminator,
Count - PreviousCount);
}
}
void ProfileGeneratorBase::updateTotalSamples() {
if (!UpdateTotalSamples)
return;
for (auto &Item : ProfileMap) {
FunctionSamples &FunctionProfile = Item.second;
FunctionProfile.updateTotalSamples();
}
}
void ProfileGeneratorBase::collectProfiledFunctions() {
std::unordered_set<const BinaryFunction *> ProfiledFunctions;
if (SampleCounters) {
// Go through all the stacks, ranges and branches in sample counters, use
// the start of the range to look up the function it belongs and record the
// function.
for (const auto &CI : *SampleCounters) {
if (const auto *CtxKey = dyn_cast<AddrBasedCtxKey>(CI.first.getPtr())) {
for (auto Addr : CtxKey->Context) {
if (FuncRange *FRange = Binary->findFuncRangeForOffset(
Binary->virtualAddrToOffset(Addr)))
ProfiledFunctions.insert(FRange->Func);
}
}
for (auto Item : CI.second.RangeCounter) {
uint64_t StartOffset = Item.first.first;
if (FuncRange *FRange = Binary->findFuncRangeForOffset(StartOffset))
ProfiledFunctions.insert(FRange->Func);
}
for (auto Item : CI.second.BranchCounter) {
uint64_t SourceOffset = Item.first.first;
uint64_t TargetOffset = Item.first.first;
if (FuncRange *FRange = Binary->findFuncRangeForOffset(SourceOffset))
ProfiledFunctions.insert(FRange->Func);
if (FuncRange *FRange = Binary->findFuncRangeForOffset(TargetOffset))
ProfiledFunctions.insert(FRange->Func);
}
}
} else {
// This is for the case the input is a llvm sample profile.
for (const auto &FS : ProfileMap) {
if (auto *Func = Binary->getBinaryFunction(FS.first.getName()))
ProfiledFunctions.insert(Func);
}
}
Binary->setProfiledFunctions(ProfiledFunctions);
}
FunctionSamples &
ProfileGenerator::getTopLevelFunctionProfile(StringRef FuncName) {
SampleContext Context(FuncName);
auto Ret = ProfileMap.emplace(Context, FunctionSamples());
if (Ret.second) {
FunctionSamples &FProfile = Ret.first->second;
FProfile.setContext(Context);
}
return Ret.first->second;
}
void ProfileGenerator::generateProfile() {
collectProfiledFunctions();
if (Binary->usePseudoProbes())
Binary->decodePseudoProbe();
if (SampleCounters) {
if (Binary->usePseudoProbes()) {
generateProbeBasedProfile();
} else {
generateLineNumBasedProfile();
}
}
postProcessProfiles();
}
void ProfileGenerator::postProcessProfiles() {
computeSummaryAndThreshold();
trimColdProfiles(ProfileMap, ColdCountThreshold);
calculateAndShowDensity(ProfileMap);
}
void ProfileGenerator::trimColdProfiles(const SampleProfileMap &Profiles,
uint64_t ColdCntThreshold) {
if (!TrimColdProfile)
return;
// Move cold profiles into a tmp container.
std::vector<SampleContext> ColdProfiles;
for (const auto &I : ProfileMap) {
if (I.second.getTotalSamples() < ColdCntThreshold)
ColdProfiles.emplace_back(I.first);
}
// Remove the cold profile from ProfileMap.
for (const auto &I : ColdProfiles)
ProfileMap.erase(I);
}
void ProfileGenerator::generateLineNumBasedProfile() {
assert(SampleCounters->size() == 1 &&
"Must have one entry for profile generation.");
const SampleCounter &SC = SampleCounters->begin()->second;
// Fill in function body samples
populateBodySamplesForAllFunctions(SC.RangeCounter);
// Fill in boundary sample counts as well as call site samples for calls
populateBoundarySamplesForAllFunctions(SC.BranchCounter);
updateTotalSamples();
}
void ProfileGenerator::generateProbeBasedProfile() {
assert(SampleCounters->size() == 1 &&
"Must have one entry for profile generation.");
// Enable pseudo probe functionalities in SampleProf
FunctionSamples::ProfileIsProbeBased = true;
const SampleCounter &SC = SampleCounters->begin()->second;
// Fill in function body samples
populateBodySamplesWithProbesForAllFunctions(SC.RangeCounter);
// Fill in boundary sample counts as well as call site samples for calls
populateBoundarySamplesWithProbesForAllFunctions(SC.BranchCounter);
updateTotalSamples();
}
void ProfileGenerator::populateBodySamplesWithProbesForAllFunctions(
const RangeSample &RangeCounter) {
ProbeCounterMap ProbeCounter;
// preprocessRangeCounter returns disjoint ranges, so no longer to redo it
// inside extractProbesFromRange.
extractProbesFromRange(preprocessRangeCounter(RangeCounter), ProbeCounter,
false);
for (const auto &PI : ProbeCounter) {
const MCDecodedPseudoProbe *Probe = PI.first;
uint64_t Count = PI.second;
SampleContextFrameVector FrameVec;
Binary->getInlineContextForProbe(Probe, FrameVec, true);
FunctionSamples &FunctionProfile =
getLeafProfileAndAddTotalSamples(FrameVec, Count);
FunctionProfile.addBodySamplesForProbe(Probe->getIndex(), Count);
if (Probe->isEntry())
FunctionProfile.addHeadSamples(Count);
}
}
void ProfileGenerator::populateBoundarySamplesWithProbesForAllFunctions(
const BranchSample &BranchCounters) {
for (const auto &Entry : BranchCounters) {
uint64_t SourceOffset = Entry.first.first;
uint64_t TargetOffset = Entry.first.second;
uint64_t Count = Entry.second;
assert(Count != 0 && "Unexpected zero weight branch");
StringRef CalleeName = getCalleeNameForOffset(TargetOffset);
if (CalleeName.size() == 0)
continue;
uint64_t SourceAddress = Binary->offsetToVirtualAddr(SourceOffset);
const MCDecodedPseudoProbe *CallProbe =
Binary->getCallProbeForAddr(SourceAddress);
if (CallProbe == nullptr)
continue;
// Record called target sample and its count.
SampleContextFrameVector FrameVec;
Binary->getInlineContextForProbe(CallProbe, FrameVec, true);
if (!FrameVec.empty()) {
FunctionSamples &FunctionProfile =
getLeafProfileAndAddTotalSamples(FrameVec, 0);
FunctionProfile.addCalledTargetSamples(
FrameVec.back().Location.LineOffset, 0, CalleeName, Count);
}
}
}
FunctionSamples &ProfileGenerator::getLeafProfileAndAddTotalSamples(
const SampleContextFrameVector &FrameVec, uint64_t Count) {
// Get top level profile
FunctionSamples *FunctionProfile =
&getTopLevelFunctionProfile(FrameVec[0].FuncName);
FunctionProfile->addTotalSamples(Count);
if (Binary->usePseudoProbes()) {
const auto *FuncDesc = Binary->getFuncDescForGUID(
Function::getGUID(FunctionProfile->getName()));
FunctionProfile->setFunctionHash(FuncDesc->FuncHash);
}
for (size_t I = 1; I < FrameVec.size(); I++) {
LineLocation Callsite(
FrameVec[I - 1].Location.LineOffset,
getBaseDiscriminator(FrameVec[I - 1].Location.Discriminator));
FunctionSamplesMap &SamplesMap =
FunctionProfile->functionSamplesAt(Callsite);
auto Ret =
SamplesMap.emplace(FrameVec[I].FuncName.str(), FunctionSamples());
if (Ret.second) {
SampleContext Context(FrameVec[I].FuncName);
Ret.first->second.setContext(Context);
}
FunctionProfile = &Ret.first->second;
FunctionProfile->addTotalSamples(Count);
if (Binary->usePseudoProbes()) {
const auto *FuncDesc = Binary->getFuncDescForGUID(
Function::getGUID(FunctionProfile->getName()));
FunctionProfile->setFunctionHash(FuncDesc->FuncHash);
}
}
return *FunctionProfile;
}
RangeSample
ProfileGenerator::preprocessRangeCounter(const RangeSample &RangeCounter) {
RangeSample Ranges(RangeCounter.begin(), RangeCounter.end());
if (FillZeroForAllFuncs) {
for (auto &FuncI : Binary->getAllBinaryFunctions()) {
for (auto &R : FuncI.second.Ranges) {
Ranges[{R.first, R.second - 1}] += 0;
}
}
} else {
// For each range, we search for all ranges of the function it belongs to
// and initialize it with zero count, so it remains zero if doesn't hit any
// samples. This is to be consistent with compiler that interpret zero count
// as unexecuted(cold).
for (const auto &I : RangeCounter) {
uint64_t StartOffset = I.first.first;
for (const auto &Range : Binary->getRangesForOffset(StartOffset))
Ranges[{Range.first, Range.second - 1}] += 0;
}
}
RangeSample DisjointRanges;
findDisjointRanges(DisjointRanges, Ranges);
return DisjointRanges;
}
void ProfileGenerator::populateBodySamplesForAllFunctions(
const RangeSample &RangeCounter) {
for (const auto &Range : preprocessRangeCounter(RangeCounter)) {
uint64_t RangeBegin = Binary->offsetToVirtualAddr(Range.first.first);
uint64_t RangeEnd = Binary->offsetToVirtualAddr(Range.first.second);
uint64_t Count = Range.second;
InstructionPointer IP(Binary, RangeBegin, true);
// Disjoint ranges may have range in the middle of two instr,
// e.g. If Instr1 at Addr1, and Instr2 at Addr2, disjoint range
// can be Addr1+1 to Addr2-1. We should ignore such range.
if (IP.Address > RangeEnd)
continue;
do {
uint64_t Offset = Binary->virtualAddrToOffset(IP.Address);
const SampleContextFrameVector &FrameVec =
Binary->getFrameLocationStack(Offset);
if (!FrameVec.empty()) {
// FIXME: As accumulating total count per instruction caused some
// regression, we changed to accumulate total count per byte as a
// workaround. Tuning hotness threshold on the compiler side might be
// necessary in the future.
FunctionSamples &FunctionProfile = getLeafProfileAndAddTotalSamples(
FrameVec, Count * Binary->getInstSize(Offset));
updateBodySamplesforFunctionProfile(FunctionProfile, FrameVec.back(),
Count);
}
} while (IP.advance() && IP.Address <= RangeEnd);
}
}
StringRef ProfileGeneratorBase::getCalleeNameForOffset(uint64_t TargetOffset) {
// Get the function range by branch target if it's a call branch.
auto *FRange = Binary->findFuncRangeForStartOffset(TargetOffset);
// We won't accumulate sample count for a range whose start is not the real
// function entry such as outlined function or inner labels.
if (!FRange || !FRange->IsFuncEntry)
return StringRef();
return FunctionSamples::getCanonicalFnName(FRange->getFuncName());
}
void ProfileGenerator::populateBoundarySamplesForAllFunctions(
const BranchSample &BranchCounters) {
for (const auto &Entry : BranchCounters) {
uint64_t SourceOffset = Entry.first.first;
uint64_t TargetOffset = Entry.first.second;
uint64_t Count = Entry.second;
assert(Count != 0 && "Unexpected zero weight branch");
StringRef CalleeName = getCalleeNameForOffset(TargetOffset);
if (CalleeName.size() == 0)
continue;
// Record called target sample and its count.
const SampleContextFrameVector &FrameVec =
Binary->getFrameLocationStack(SourceOffset);
if (!FrameVec.empty()) {
FunctionSamples &FunctionProfile =
getLeafProfileAndAddTotalSamples(FrameVec, 0);
FunctionProfile.addCalledTargetSamples(
FrameVec.back().Location.LineOffset,
getBaseDiscriminator(FrameVec.back().Location.Discriminator),
CalleeName, Count);
}
// Add head samples for callee.
FunctionSamples &CalleeProfile = getTopLevelFunctionProfile(CalleeName);
CalleeProfile.addHeadSamples(Count);
}
}
void ProfileGeneratorBase::calculateAndShowDensity(
const SampleProfileMap &Profiles) {
double Density = calculateDensity(Profiles, HotCountThreshold);
showDensitySuggestion(Density);
}
FunctionSamples &CSProfileGenerator::getFunctionProfileForContext(
const SampleContextFrameVector &Context, bool WasLeafInlined) {
auto I = ProfileMap.find(SampleContext(Context));
if (I == ProfileMap.end()) {
// Save the new context for future references.
SampleContextFrames NewContext = *Contexts.insert(Context).first;
SampleContext FContext(NewContext, RawContext);
auto Ret = ProfileMap.emplace(FContext, FunctionSamples());
if (WasLeafInlined)
FContext.setAttribute(ContextWasInlined);
[CSSPGO][llvm-profgen] Context-sensitive profile data generation This stack of changes introduces `llvm-profgen` utility which generates a profile data file from given perf script data files for sample-based PGO. It’s part of(not only) the CSSPGO work. Specifically to support context-sensitive with/without pseudo probe profile, it implements a series of functionalities including perf trace parsing, instruction symbolization, LBR stack/call frame stack unwinding, pseudo probe decoding, etc. Also high throughput is achieved by multiple levels of sample aggregation and compatible format with one stop is generated at the end. Please refer to: https://groups.google.com/g/llvm-dev/c/1p1rdYbL93s for the CSSPGO RFC. This change supports context-sensitive profile data generation into llvm-profgen. With simultaneous sampling for LBR and call stack, we can identify leaf of LBR sample with calling context from stack sample . During the process of deriving fall through path from LBR entries, we unwind LBR by replaying all the calls and returns (including implicit calls/returns due to inlining) backwards on top of the sampled call stack. Then the state of call stack as we unwind through LBR always represents the calling context of current fall through path. we have two types of virtual unwinding 1) LBR unwinding and 2) linear range unwinding. Specifically, for each LBR entry which can be classified into call, return, regular branch, LBR unwinding will replay the operation by pushing, popping or switching leaf frame towards the call stack and since the initial call stack is most recently sampled, the replay should be in anti-execution order, i.e. for the regular case, pop the call stack when LBR is call, push frame on call stack when LBR is return. After each LBR processed, it also needs to align with the next LBR by going through instructions from previous LBR's target to current LBR's source, which we named linear unwinding. As instruction from linear range can come from different function by inlining, linear unwinding will do the range splitting and record counters through the range with same inline context. With each fall through path from LBR unwinding, we aggregate each sample into counters by the calling context and eventually generate full context sensitive profile (without relying on inlining) to driver compiler's PGO/FDO. A breakdown of noteworthy changes: - Added `HybridSample` class as the abstraction perf sample including LBR stack and call stack * Extended `PerfReader` to implement auto-detect whether input perf script output contains CS profile, then do the parsing. Multiple `HybridSample` are extracted * Speed up by aggregating `HybridSample` into `AggregatedSamples` * Added VirtualUnwinder that consumes aggregated `HybridSample` and implements unwinding of calls, returns, and linear path that contains implicit call/return from inlining. Ranges and branches counters are aggregated by the calling context.
 Here calling context is string type, each context is a pair of function name and callsite location info, the whole context is like `main:1 @ foo:2 @ bar`. * Added PorfileGenerater that accumulates counters by ranges unfolding or branch target mapping, then generates context-sensitive function profile including function body, inferring callee's head sample, callsite target samples, eventually records into ProfileMap.
 * Leveraged LLVM build-in(`SampleProfWriter`) writer to support different serialization format with no stop - `getCanonicalFnName` for callee name and name from ELF section - Added regression test for both unwinding and profile generation Test Plan: ninja & ninja check-llvm Reviewed By: hoy, wenlei, wmi Differential Revision: https://reviews.llvm.org/D89723
2020-10-19 12:55:59 -07:00
FunctionSamples &FProfile = Ret.first->second;
FProfile.setContext(FContext);
return Ret.first->second;
} else {
// Update ContextWasInlined attribute for existing contexts.
// The current function can be called in two ways:
// - when processing a probe of the current frame
// - when processing the entry probe of an inlinee's frame, which
// is then used to update the callsite count of the current frame.
// The two can happen in any order, hence here we are making sure
// `ContextWasInlined` is always set as expected.
// TODO: Note that the former does not always happen if no probes of the
// current frame has samples, and if the latter happens, we could lose the
// attribute. This should be fixed.
if (WasLeafInlined)
I->second.getContext().setAttribute(ContextWasInlined);
[CSSPGO][llvm-profgen] Context-sensitive profile data generation This stack of changes introduces `llvm-profgen` utility which generates a profile data file from given perf script data files for sample-based PGO. It’s part of(not only) the CSSPGO work. Specifically to support context-sensitive with/without pseudo probe profile, it implements a series of functionalities including perf trace parsing, instruction symbolization, LBR stack/call frame stack unwinding, pseudo probe decoding, etc. Also high throughput is achieved by multiple levels of sample aggregation and compatible format with one stop is generated at the end. Please refer to: https://groups.google.com/g/llvm-dev/c/1p1rdYbL93s for the CSSPGO RFC. This change supports context-sensitive profile data generation into llvm-profgen. With simultaneous sampling for LBR and call stack, we can identify leaf of LBR sample with calling context from stack sample . During the process of deriving fall through path from LBR entries, we unwind LBR by replaying all the calls and returns (including implicit calls/returns due to inlining) backwards on top of the sampled call stack. Then the state of call stack as we unwind through LBR always represents the calling context of current fall through path. we have two types of virtual unwinding 1) LBR unwinding and 2) linear range unwinding. Specifically, for each LBR entry which can be classified into call, return, regular branch, LBR unwinding will replay the operation by pushing, popping or switching leaf frame towards the call stack and since the initial call stack is most recently sampled, the replay should be in anti-execution order, i.e. for the regular case, pop the call stack when LBR is call, push frame on call stack when LBR is return. After each LBR processed, it also needs to align with the next LBR by going through instructions from previous LBR's target to current LBR's source, which we named linear unwinding. As instruction from linear range can come from different function by inlining, linear unwinding will do the range splitting and record counters through the range with same inline context. With each fall through path from LBR unwinding, we aggregate each sample into counters by the calling context and eventually generate full context sensitive profile (without relying on inlining) to driver compiler's PGO/FDO. A breakdown of noteworthy changes: - Added `HybridSample` class as the abstraction perf sample including LBR stack and call stack * Extended `PerfReader` to implement auto-detect whether input perf script output contains CS profile, then do the parsing. Multiple `HybridSample` are extracted * Speed up by aggregating `HybridSample` into `AggregatedSamples` * Added VirtualUnwinder that consumes aggregated `HybridSample` and implements unwinding of calls, returns, and linear path that contains implicit call/return from inlining. Ranges and branches counters are aggregated by the calling context.
 Here calling context is string type, each context is a pair of function name and callsite location info, the whole context is like `main:1 @ foo:2 @ bar`. * Added PorfileGenerater that accumulates counters by ranges unfolding or branch target mapping, then generates context-sensitive function profile including function body, inferring callee's head sample, callsite target samples, eventually records into ProfileMap.
 * Leveraged LLVM build-in(`SampleProfWriter`) writer to support different serialization format with no stop - `getCanonicalFnName` for callee name and name from ELF section - Added regression test for both unwinding and profile generation Test Plan: ninja & ninja check-llvm Reviewed By: hoy, wenlei, wmi Differential Revision: https://reviews.llvm.org/D89723
2020-10-19 12:55:59 -07:00
}
return I->second;
[CSSPGO][llvm-profgen] Context-sensitive profile data generation This stack of changes introduces `llvm-profgen` utility which generates a profile data file from given perf script data files for sample-based PGO. It’s part of(not only) the CSSPGO work. Specifically to support context-sensitive with/without pseudo probe profile, it implements a series of functionalities including perf trace parsing, instruction symbolization, LBR stack/call frame stack unwinding, pseudo probe decoding, etc. Also high throughput is achieved by multiple levels of sample aggregation and compatible format with one stop is generated at the end. Please refer to: https://groups.google.com/g/llvm-dev/c/1p1rdYbL93s for the CSSPGO RFC. This change supports context-sensitive profile data generation into llvm-profgen. With simultaneous sampling for LBR and call stack, we can identify leaf of LBR sample with calling context from stack sample . During the process of deriving fall through path from LBR entries, we unwind LBR by replaying all the calls and returns (including implicit calls/returns due to inlining) backwards on top of the sampled call stack. Then the state of call stack as we unwind through LBR always represents the calling context of current fall through path. we have two types of virtual unwinding 1) LBR unwinding and 2) linear range unwinding. Specifically, for each LBR entry which can be classified into call, return, regular branch, LBR unwinding will replay the operation by pushing, popping or switching leaf frame towards the call stack and since the initial call stack is most recently sampled, the replay should be in anti-execution order, i.e. for the regular case, pop the call stack when LBR is call, push frame on call stack when LBR is return. After each LBR processed, it also needs to align with the next LBR by going through instructions from previous LBR's target to current LBR's source, which we named linear unwinding. As instruction from linear range can come from different function by inlining, linear unwinding will do the range splitting and record counters through the range with same inline context. With each fall through path from LBR unwinding, we aggregate each sample into counters by the calling context and eventually generate full context sensitive profile (without relying on inlining) to driver compiler's PGO/FDO. A breakdown of noteworthy changes: - Added `HybridSample` class as the abstraction perf sample including LBR stack and call stack * Extended `PerfReader` to implement auto-detect whether input perf script output contains CS profile, then do the parsing. Multiple `HybridSample` are extracted * Speed up by aggregating `HybridSample` into `AggregatedSamples` * Added VirtualUnwinder that consumes aggregated `HybridSample` and implements unwinding of calls, returns, and linear path that contains implicit call/return from inlining. Ranges and branches counters are aggregated by the calling context.
 Here calling context is string type, each context is a pair of function name and callsite location info, the whole context is like `main:1 @ foo:2 @ bar`. * Added PorfileGenerater that accumulates counters by ranges unfolding or branch target mapping, then generates context-sensitive function profile including function body, inferring callee's head sample, callsite target samples, eventually records into ProfileMap.
 * Leveraged LLVM build-in(`SampleProfWriter`) writer to support different serialization format with no stop - `getCanonicalFnName` for callee name and name from ELF section - Added regression test for both unwinding and profile generation Test Plan: ninja & ninja check-llvm Reviewed By: hoy, wenlei, wmi Differential Revision: https://reviews.llvm.org/D89723
2020-10-19 12:55:59 -07:00
}
void CSProfileGenerator::generateProfile() {
[CSSPGO] Use nested context-sensitive profile. 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
2021-12-14 10:03:05 -08:00
FunctionSamples::ProfileIsCSFlat = true;
collectProfiledFunctions();
if (Binary->usePseudoProbes())
Binary->decodePseudoProbe();
if (SampleCounters) {
if (Binary->usePseudoProbes()) {
generateProbeBasedProfile();
} else {
generateLineNumBasedProfile();
}
}
if (Binary->getTrackFuncContextSize())
computeSizeForProfiledFunctions();
postProcessProfiles();
}
void CSProfileGenerator::computeSizeForProfiledFunctions() {
std::unordered_set<const BinaryFunction *> ProfiledFunctions;
for (auto *Func : Binary->getProfiledFunctions())
Binary->computeInlinedContextSizeForFunc(Func);
// Flush the symbolizer to save memory.
Binary->flushSymbolizer();
}
void CSProfileGenerator::generateLineNumBasedProfile() {
for (const auto &CI : *SampleCounters) {
const auto *CtxKey = cast<StringBasedCtxKey>(CI.first.getPtr());
// Get or create function profile for the range
FunctionSamples &FunctionProfile =
[CSSPGO] Split context string to deduplicate function name used in the context. 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
2021-08-25 11:40:34 -07:00
getFunctionProfileForContext(CtxKey->Context, CtxKey->WasLeafInlined);
// Fill in function body samples
populateBodySamplesForFunction(FunctionProfile, CI.second.RangeCounter);
// Fill in boundary sample counts as well as call site samples for calls
populateBoundarySamplesForFunction(CtxKey->Context, FunctionProfile,
CI.second.BranchCounter);
}
// Fill in call site value sample for inlined calls and also use context to
// infer missing samples. Since we don't have call count for inlined
// functions, we estimate it from inlinee's profile using the entry of the
// body sample.
populateInferredFunctionSamples();
updateTotalSamples();
[CSSPGO][llvm-profgen] Context-sensitive profile data generation This stack of changes introduces `llvm-profgen` utility which generates a profile data file from given perf script data files for sample-based PGO. It’s part of(not only) the CSSPGO work. Specifically to support context-sensitive with/without pseudo probe profile, it implements a series of functionalities including perf trace parsing, instruction symbolization, LBR stack/call frame stack unwinding, pseudo probe decoding, etc. Also high throughput is achieved by multiple levels of sample aggregation and compatible format with one stop is generated at the end. Please refer to: https://groups.google.com/g/llvm-dev/c/1p1rdYbL93s for the CSSPGO RFC. This change supports context-sensitive profile data generation into llvm-profgen. With simultaneous sampling for LBR and call stack, we can identify leaf of LBR sample with calling context from stack sample . During the process of deriving fall through path from LBR entries, we unwind LBR by replaying all the calls and returns (including implicit calls/returns due to inlining) backwards on top of the sampled call stack. Then the state of call stack as we unwind through LBR always represents the calling context of current fall through path. we have two types of virtual unwinding 1) LBR unwinding and 2) linear range unwinding. Specifically, for each LBR entry which can be classified into call, return, regular branch, LBR unwinding will replay the operation by pushing, popping or switching leaf frame towards the call stack and since the initial call stack is most recently sampled, the replay should be in anti-execution order, i.e. for the regular case, pop the call stack when LBR is call, push frame on call stack when LBR is return. After each LBR processed, it also needs to align with the next LBR by going through instructions from previous LBR's target to current LBR's source, which we named linear unwinding. As instruction from linear range can come from different function by inlining, linear unwinding will do the range splitting and record counters through the range with same inline context. With each fall through path from LBR unwinding, we aggregate each sample into counters by the calling context and eventually generate full context sensitive profile (without relying on inlining) to driver compiler's PGO/FDO. A breakdown of noteworthy changes: - Added `HybridSample` class as the abstraction perf sample including LBR stack and call stack * Extended `PerfReader` to implement auto-detect whether input perf script output contains CS profile, then do the parsing. Multiple `HybridSample` are extracted * Speed up by aggregating `HybridSample` into `AggregatedSamples` * Added VirtualUnwinder that consumes aggregated `HybridSample` and implements unwinding of calls, returns, and linear path that contains implicit call/return from inlining. Ranges and branches counters are aggregated by the calling context.
 Here calling context is string type, each context is a pair of function name and callsite location info, the whole context is like `main:1 @ foo:2 @ bar`. * Added PorfileGenerater that accumulates counters by ranges unfolding or branch target mapping, then generates context-sensitive function profile including function body, inferring callee's head sample, callsite target samples, eventually records into ProfileMap.
 * Leveraged LLVM build-in(`SampleProfWriter`) writer to support different serialization format with no stop - `getCanonicalFnName` for callee name and name from ELF section - Added regression test for both unwinding and profile generation Test Plan: ninja & ninja check-llvm Reviewed By: hoy, wenlei, wmi Differential Revision: https://reviews.llvm.org/D89723
2020-10-19 12:55:59 -07:00
}
void CSProfileGenerator::populateBodySamplesForFunction(
FunctionSamples &FunctionProfile, const RangeSample &RangeCounter) {
[CSSPGO][llvm-profgen] Refactor to unify hashable interface for trace sample and context-sensitive counter As we plan to support both CSSPGO and AutoFDO for llvm-profgen, we will have different kinds of perf sample and different kinds of sample counter(cs/non-cs, with/without pseudo probe) which both need to do aggregation in hash map. This change implements the hashable interface(`Hashable`) and the unified base class for them to have better extensibility and reusability. Currently perf trace sample and sample counter with context implemented this `Hashable` and the class hierarchy is like: ``` | Hashable | PerfSample | HybridSample | LBRSample | ContextKey | StringBasedCtxKey | ProbeBasedCtxKey | CallsiteBasedCtxKey | ... ``` - Class specifying `Hashable` should implement `getHashCode` and `isEqual`. Here we make `getHashCode` a non-virtual function to avoid vtable overhead, so derived class should calculate and assign the base class's HashCode manually. This also provides the flexibility for calculating the hash code incrementally(like rolling hash) during frame stack unwinding - `isEqual` is a virtual function, which will have perf overhead. In the future, if we redesign a better hash function, then we can just skip this or switch to non-virtual function. - Added `PerfSample` and `ContextKey` as base class for perf sample and counter context key, leveraging llvm-style RTTI for this. - Added `StringBasedCtxKey` class extending `ContextKey` to use string as context id. - Refactor `AggregationCounter` to take all kinds of `PerfSample` as key - Refactor `ContextSampleCounter` to take all kinds of `ContextKey` as key - Other refactoring work: - Create a wrapper class `SampleCounter` to wrap `RangeCounter` and `BranchCounter` - Hoist `ContextId` and `FunctionProfile` out of `populateFunctionBodySamples` and `populateFunctionBoundarySamples` to reuse them in ProfileGenerator Differential Revision: https://reviews.llvm.org/D92584
2020-12-02 23:10:11 -08:00
// Compute disjoint ranges first, so we can use MAX
// for calculating count for each location.
RangeSample Ranges;
findDisjointRanges(Ranges, RangeCounter);
for (const auto &Range : Ranges) {
[CSSPGO][llvm-profgen] Refactor to unify hashable interface for trace sample and context-sensitive counter As we plan to support both CSSPGO and AutoFDO for llvm-profgen, we will have different kinds of perf sample and different kinds of sample counter(cs/non-cs, with/without pseudo probe) which both need to do aggregation in hash map. This change implements the hashable interface(`Hashable`) and the unified base class for them to have better extensibility and reusability. Currently perf trace sample and sample counter with context implemented this `Hashable` and the class hierarchy is like: ``` | Hashable | PerfSample | HybridSample | LBRSample | ContextKey | StringBasedCtxKey | ProbeBasedCtxKey | CallsiteBasedCtxKey | ... ``` - Class specifying `Hashable` should implement `getHashCode` and `isEqual`. Here we make `getHashCode` a non-virtual function to avoid vtable overhead, so derived class should calculate and assign the base class's HashCode manually. This also provides the flexibility for calculating the hash code incrementally(like rolling hash) during frame stack unwinding - `isEqual` is a virtual function, which will have perf overhead. In the future, if we redesign a better hash function, then we can just skip this or switch to non-virtual function. - Added `PerfSample` and `ContextKey` as base class for perf sample and counter context key, leveraging llvm-style RTTI for this. - Added `StringBasedCtxKey` class extending `ContextKey` to use string as context id. - Refactor `AggregationCounter` to take all kinds of `PerfSample` as key - Refactor `ContextSampleCounter` to take all kinds of `ContextKey` as key - Other refactoring work: - Create a wrapper class `SampleCounter` to wrap `RangeCounter` and `BranchCounter` - Hoist `ContextId` and `FunctionProfile` out of `populateFunctionBodySamples` and `populateFunctionBoundarySamples` to reuse them in ProfileGenerator Differential Revision: https://reviews.llvm.org/D92584
2020-12-02 23:10:11 -08:00
uint64_t RangeBegin = Binary->offsetToVirtualAddr(Range.first.first);
uint64_t RangeEnd = Binary->offsetToVirtualAddr(Range.first.second);
uint64_t Count = Range.second;
// Disjoint ranges have introduce zero-filled gap that
// doesn't belong to current context, filter them out.
if (Count == 0)
continue;
InstructionPointer IP(Binary, RangeBegin, true);
// Disjoint ranges may have range in the middle of two instr,
// e.g. If Instr1 at Addr1, and Instr2 at Addr2, disjoint range
// can be Addr1+1 to Addr2-1. We should ignore such range.
if (IP.Address > RangeEnd)
continue;
do {
[CSSPGO][llvm-profgen] Refactor to unify hashable interface for trace sample and context-sensitive counter As we plan to support both CSSPGO and AutoFDO for llvm-profgen, we will have different kinds of perf sample and different kinds of sample counter(cs/non-cs, with/without pseudo probe) which both need to do aggregation in hash map. This change implements the hashable interface(`Hashable`) and the unified base class for them to have better extensibility and reusability. Currently perf trace sample and sample counter with context implemented this `Hashable` and the class hierarchy is like: ``` | Hashable | PerfSample | HybridSample | LBRSample | ContextKey | StringBasedCtxKey | ProbeBasedCtxKey | CallsiteBasedCtxKey | ... ``` - Class specifying `Hashable` should implement `getHashCode` and `isEqual`. Here we make `getHashCode` a non-virtual function to avoid vtable overhead, so derived class should calculate and assign the base class's HashCode manually. This also provides the flexibility for calculating the hash code incrementally(like rolling hash) during frame stack unwinding - `isEqual` is a virtual function, which will have perf overhead. In the future, if we redesign a better hash function, then we can just skip this or switch to non-virtual function. - Added `PerfSample` and `ContextKey` as base class for perf sample and counter context key, leveraging llvm-style RTTI for this. - Added `StringBasedCtxKey` class extending `ContextKey` to use string as context id. - Refactor `AggregationCounter` to take all kinds of `PerfSample` as key - Refactor `ContextSampleCounter` to take all kinds of `ContextKey` as key - Other refactoring work: - Create a wrapper class `SampleCounter` to wrap `RangeCounter` and `BranchCounter` - Hoist `ContextId` and `FunctionProfile` out of `populateFunctionBodySamples` and `populateFunctionBoundarySamples` to reuse them in ProfileGenerator Differential Revision: https://reviews.llvm.org/D92584
2020-12-02 23:10:11 -08:00
uint64_t Offset = Binary->virtualAddrToOffset(IP.Address);
auto LeafLoc = Binary->getInlineLeafFrameLoc(Offset);
if (LeafLoc.hasValue()) {
// Recording body sample for this specific context
updateBodySamplesforFunctionProfile(FunctionProfile, *LeafLoc, Count);
FunctionProfile.addTotalSamples(Count);
}
} while (IP.advance() && IP.Address <= RangeEnd);
[CSSPGO][llvm-profgen] Context-sensitive profile data generation This stack of changes introduces `llvm-profgen` utility which generates a profile data file from given perf script data files for sample-based PGO. It’s part of(not only) the CSSPGO work. Specifically to support context-sensitive with/without pseudo probe profile, it implements a series of functionalities including perf trace parsing, instruction symbolization, LBR stack/call frame stack unwinding, pseudo probe decoding, etc. Also high throughput is achieved by multiple levels of sample aggregation and compatible format with one stop is generated at the end. Please refer to: https://groups.google.com/g/llvm-dev/c/1p1rdYbL93s for the CSSPGO RFC. This change supports context-sensitive profile data generation into llvm-profgen. With simultaneous sampling for LBR and call stack, we can identify leaf of LBR sample with calling context from stack sample . During the process of deriving fall through path from LBR entries, we unwind LBR by replaying all the calls and returns (including implicit calls/returns due to inlining) backwards on top of the sampled call stack. Then the state of call stack as we unwind through LBR always represents the calling context of current fall through path. we have two types of virtual unwinding 1) LBR unwinding and 2) linear range unwinding. Specifically, for each LBR entry which can be classified into call, return, regular branch, LBR unwinding will replay the operation by pushing, popping or switching leaf frame towards the call stack and since the initial call stack is most recently sampled, the replay should be in anti-execution order, i.e. for the regular case, pop the call stack when LBR is call, push frame on call stack when LBR is return. After each LBR processed, it also needs to align with the next LBR by going through instructions from previous LBR's target to current LBR's source, which we named linear unwinding. As instruction from linear range can come from different function by inlining, linear unwinding will do the range splitting and record counters through the range with same inline context. With each fall through path from LBR unwinding, we aggregate each sample into counters by the calling context and eventually generate full context sensitive profile (without relying on inlining) to driver compiler's PGO/FDO. A breakdown of noteworthy changes: - Added `HybridSample` class as the abstraction perf sample including LBR stack and call stack * Extended `PerfReader` to implement auto-detect whether input perf script output contains CS profile, then do the parsing. Multiple `HybridSample` are extracted * Speed up by aggregating `HybridSample` into `AggregatedSamples` * Added VirtualUnwinder that consumes aggregated `HybridSample` and implements unwinding of calls, returns, and linear path that contains implicit call/return from inlining. Ranges and branches counters are aggregated by the calling context.
 Here calling context is string type, each context is a pair of function name and callsite location info, the whole context is like `main:1 @ foo:2 @ bar`. * Added PorfileGenerater that accumulates counters by ranges unfolding or branch target mapping, then generates context-sensitive function profile including function body, inferring callee's head sample, callsite target samples, eventually records into ProfileMap.
 * Leveraged LLVM build-in(`SampleProfWriter`) writer to support different serialization format with no stop - `getCanonicalFnName` for callee name and name from ELF section - Added regression test for both unwinding and profile generation Test Plan: ninja & ninja check-llvm Reviewed By: hoy, wenlei, wmi Differential Revision: https://reviews.llvm.org/D89723
2020-10-19 12:55:59 -07:00
}
}
void CSProfileGenerator::populateBoundarySamplesForFunction(
[CSSPGO] Split context string to deduplicate function name used in the context. 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
2021-08-25 11:40:34 -07:00
SampleContextFrames ContextId, FunctionSamples &FunctionProfile,
const BranchSample &BranchCounters) {
[CSSPGO][llvm-profgen] Refactor to unify hashable interface for trace sample and context-sensitive counter As we plan to support both CSSPGO and AutoFDO for llvm-profgen, we will have different kinds of perf sample and different kinds of sample counter(cs/non-cs, with/without pseudo probe) which both need to do aggregation in hash map. This change implements the hashable interface(`Hashable`) and the unified base class for them to have better extensibility and reusability. Currently perf trace sample and sample counter with context implemented this `Hashable` and the class hierarchy is like: ``` | Hashable | PerfSample | HybridSample | LBRSample | ContextKey | StringBasedCtxKey | ProbeBasedCtxKey | CallsiteBasedCtxKey | ... ``` - Class specifying `Hashable` should implement `getHashCode` and `isEqual`. Here we make `getHashCode` a non-virtual function to avoid vtable overhead, so derived class should calculate and assign the base class's HashCode manually. This also provides the flexibility for calculating the hash code incrementally(like rolling hash) during frame stack unwinding - `isEqual` is a virtual function, which will have perf overhead. In the future, if we redesign a better hash function, then we can just skip this or switch to non-virtual function. - Added `PerfSample` and `ContextKey` as base class for perf sample and counter context key, leveraging llvm-style RTTI for this. - Added `StringBasedCtxKey` class extending `ContextKey` to use string as context id. - Refactor `AggregationCounter` to take all kinds of `PerfSample` as key - Refactor `ContextSampleCounter` to take all kinds of `ContextKey` as key - Other refactoring work: - Create a wrapper class `SampleCounter` to wrap `RangeCounter` and `BranchCounter` - Hoist `ContextId` and `FunctionProfile` out of `populateFunctionBodySamples` and `populateFunctionBoundarySamples` to reuse them in ProfileGenerator Differential Revision: https://reviews.llvm.org/D92584
2020-12-02 23:10:11 -08:00
for (const auto &Entry : BranchCounters) {
[CSSPGO][llvm-profgen] Refactor to unify hashable interface for trace sample and context-sensitive counter As we plan to support both CSSPGO and AutoFDO for llvm-profgen, we will have different kinds of perf sample and different kinds of sample counter(cs/non-cs, with/without pseudo probe) which both need to do aggregation in hash map. This change implements the hashable interface(`Hashable`) and the unified base class for them to have better extensibility and reusability. Currently perf trace sample and sample counter with context implemented this `Hashable` and the class hierarchy is like: ``` | Hashable | PerfSample | HybridSample | LBRSample | ContextKey | StringBasedCtxKey | ProbeBasedCtxKey | CallsiteBasedCtxKey | ... ``` - Class specifying `Hashable` should implement `getHashCode` and `isEqual`. Here we make `getHashCode` a non-virtual function to avoid vtable overhead, so derived class should calculate and assign the base class's HashCode manually. This also provides the flexibility for calculating the hash code incrementally(like rolling hash) during frame stack unwinding - `isEqual` is a virtual function, which will have perf overhead. In the future, if we redesign a better hash function, then we can just skip this or switch to non-virtual function. - Added `PerfSample` and `ContextKey` as base class for perf sample and counter context key, leveraging llvm-style RTTI for this. - Added `StringBasedCtxKey` class extending `ContextKey` to use string as context id. - Refactor `AggregationCounter` to take all kinds of `PerfSample` as key - Refactor `ContextSampleCounter` to take all kinds of `ContextKey` as key - Other refactoring work: - Create a wrapper class `SampleCounter` to wrap `RangeCounter` and `BranchCounter` - Hoist `ContextId` and `FunctionProfile` out of `populateFunctionBodySamples` and `populateFunctionBoundarySamples` to reuse them in ProfileGenerator Differential Revision: https://reviews.llvm.org/D92584
2020-12-02 23:10:11 -08:00
uint64_t SourceOffset = Entry.first.first;
uint64_t TargetOffset = Entry.first.second;
uint64_t Count = Entry.second;
assert(Count != 0 && "Unexpected zero weight branch");
StringRef CalleeName = getCalleeNameForOffset(TargetOffset);
[CSSPGO][llvm-profgen] Refactor to unify hashable interface for trace sample and context-sensitive counter As we plan to support both CSSPGO and AutoFDO for llvm-profgen, we will have different kinds of perf sample and different kinds of sample counter(cs/non-cs, with/without pseudo probe) which both need to do aggregation in hash map. This change implements the hashable interface(`Hashable`) and the unified base class for them to have better extensibility and reusability. Currently perf trace sample and sample counter with context implemented this `Hashable` and the class hierarchy is like: ``` | Hashable | PerfSample | HybridSample | LBRSample | ContextKey | StringBasedCtxKey | ProbeBasedCtxKey | CallsiteBasedCtxKey | ... ``` - Class specifying `Hashable` should implement `getHashCode` and `isEqual`. Here we make `getHashCode` a non-virtual function to avoid vtable overhead, so derived class should calculate and assign the base class's HashCode manually. This also provides the flexibility for calculating the hash code incrementally(like rolling hash) during frame stack unwinding - `isEqual` is a virtual function, which will have perf overhead. In the future, if we redesign a better hash function, then we can just skip this or switch to non-virtual function. - Added `PerfSample` and `ContextKey` as base class for perf sample and counter context key, leveraging llvm-style RTTI for this. - Added `StringBasedCtxKey` class extending `ContextKey` to use string as context id. - Refactor `AggregationCounter` to take all kinds of `PerfSample` as key - Refactor `ContextSampleCounter` to take all kinds of `ContextKey` as key - Other refactoring work: - Create a wrapper class `SampleCounter` to wrap `RangeCounter` and `BranchCounter` - Hoist `ContextId` and `FunctionProfile` out of `populateFunctionBodySamples` and `populateFunctionBoundarySamples` to reuse them in ProfileGenerator Differential Revision: https://reviews.llvm.org/D92584
2020-12-02 23:10:11 -08:00
if (CalleeName.size() == 0)
continue;
// Record called target sample and its count
auto LeafLoc = Binary->getInlineLeafFrameLoc(SourceOffset);
if (!LeafLoc.hasValue())
continue;
FunctionProfile.addCalledTargetSamples(
LeafLoc->Location.LineOffset,
getBaseDiscriminator(LeafLoc->Location.Discriminator), CalleeName,
Count);
[CSSPGO][llvm-profgen] Refactor to unify hashable interface for trace sample and context-sensitive counter As we plan to support both CSSPGO and AutoFDO for llvm-profgen, we will have different kinds of perf sample and different kinds of sample counter(cs/non-cs, with/without pseudo probe) which both need to do aggregation in hash map. This change implements the hashable interface(`Hashable`) and the unified base class for them to have better extensibility and reusability. Currently perf trace sample and sample counter with context implemented this `Hashable` and the class hierarchy is like: ``` | Hashable | PerfSample | HybridSample | LBRSample | ContextKey | StringBasedCtxKey | ProbeBasedCtxKey | CallsiteBasedCtxKey | ... ``` - Class specifying `Hashable` should implement `getHashCode` and `isEqual`. Here we make `getHashCode` a non-virtual function to avoid vtable overhead, so derived class should calculate and assign the base class's HashCode manually. This also provides the flexibility for calculating the hash code incrementally(like rolling hash) during frame stack unwinding - `isEqual` is a virtual function, which will have perf overhead. In the future, if we redesign a better hash function, then we can just skip this or switch to non-virtual function. - Added `PerfSample` and `ContextKey` as base class for perf sample and counter context key, leveraging llvm-style RTTI for this. - Added `StringBasedCtxKey` class extending `ContextKey` to use string as context id. - Refactor `AggregationCounter` to take all kinds of `PerfSample` as key - Refactor `ContextSampleCounter` to take all kinds of `ContextKey` as key - Other refactoring work: - Create a wrapper class `SampleCounter` to wrap `RangeCounter` and `BranchCounter` - Hoist `ContextId` and `FunctionProfile` out of `populateFunctionBodySamples` and `populateFunctionBoundarySamples` to reuse them in ProfileGenerator Differential Revision: https://reviews.llvm.org/D92584
2020-12-02 23:10:11 -08:00
// Record head sample for called target(callee)
[CSSPGO] Split context string to deduplicate function name used in the context. 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
2021-08-25 11:40:34 -07:00
SampleContextFrameVector CalleeCtx(ContextId.begin(), ContextId.end());
assert(CalleeCtx.back().FuncName == LeafLoc->FuncName &&
[CSSPGO] Split context string to deduplicate function name used in the context. 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
2021-08-25 11:40:34 -07:00
"Leaf function name doesn't match");
CalleeCtx.back() = *LeafLoc;
CalleeCtx.emplace_back(CalleeName, LineLocation(0, 0));
FunctionSamples &CalleeProfile = getFunctionProfileForContext(CalleeCtx);
[CSSPGO][llvm-profgen] Refactor to unify hashable interface for trace sample and context-sensitive counter As we plan to support both CSSPGO and AutoFDO for llvm-profgen, we will have different kinds of perf sample and different kinds of sample counter(cs/non-cs, with/without pseudo probe) which both need to do aggregation in hash map. This change implements the hashable interface(`Hashable`) and the unified base class for them to have better extensibility and reusability. Currently perf trace sample and sample counter with context implemented this `Hashable` and the class hierarchy is like: ``` | Hashable | PerfSample | HybridSample | LBRSample | ContextKey | StringBasedCtxKey | ProbeBasedCtxKey | CallsiteBasedCtxKey | ... ``` - Class specifying `Hashable` should implement `getHashCode` and `isEqual`. Here we make `getHashCode` a non-virtual function to avoid vtable overhead, so derived class should calculate and assign the base class's HashCode manually. This also provides the flexibility for calculating the hash code incrementally(like rolling hash) during frame stack unwinding - `isEqual` is a virtual function, which will have perf overhead. In the future, if we redesign a better hash function, then we can just skip this or switch to non-virtual function. - Added `PerfSample` and `ContextKey` as base class for perf sample and counter context key, leveraging llvm-style RTTI for this. - Added `StringBasedCtxKey` class extending `ContextKey` to use string as context id. - Refactor `AggregationCounter` to take all kinds of `PerfSample` as key - Refactor `ContextSampleCounter` to take all kinds of `ContextKey` as key - Other refactoring work: - Create a wrapper class `SampleCounter` to wrap `RangeCounter` and `BranchCounter` - Hoist `ContextId` and `FunctionProfile` out of `populateFunctionBodySamples` and `populateFunctionBoundarySamples` to reuse them in ProfileGenerator Differential Revision: https://reviews.llvm.org/D92584
2020-12-02 23:10:11 -08:00
CalleeProfile.addHeadSamples(Count);
[CSSPGO][llvm-profgen] Context-sensitive profile data generation This stack of changes introduces `llvm-profgen` utility which generates a profile data file from given perf script data files for sample-based PGO. It’s part of(not only) the CSSPGO work. Specifically to support context-sensitive with/without pseudo probe profile, it implements a series of functionalities including perf trace parsing, instruction symbolization, LBR stack/call frame stack unwinding, pseudo probe decoding, etc. Also high throughput is achieved by multiple levels of sample aggregation and compatible format with one stop is generated at the end. Please refer to: https://groups.google.com/g/llvm-dev/c/1p1rdYbL93s for the CSSPGO RFC. This change supports context-sensitive profile data generation into llvm-profgen. With simultaneous sampling for LBR and call stack, we can identify leaf of LBR sample with calling context from stack sample . During the process of deriving fall through path from LBR entries, we unwind LBR by replaying all the calls and returns (including implicit calls/returns due to inlining) backwards on top of the sampled call stack. Then the state of call stack as we unwind through LBR always represents the calling context of current fall through path. we have two types of virtual unwinding 1) LBR unwinding and 2) linear range unwinding. Specifically, for each LBR entry which can be classified into call, return, regular branch, LBR unwinding will replay the operation by pushing, popping or switching leaf frame towards the call stack and since the initial call stack is most recently sampled, the replay should be in anti-execution order, i.e. for the regular case, pop the call stack when LBR is call, push frame on call stack when LBR is return. After each LBR processed, it also needs to align with the next LBR by going through instructions from previous LBR's target to current LBR's source, which we named linear unwinding. As instruction from linear range can come from different function by inlining, linear unwinding will do the range splitting and record counters through the range with same inline context. With each fall through path from LBR unwinding, we aggregate each sample into counters by the calling context and eventually generate full context sensitive profile (without relying on inlining) to driver compiler's PGO/FDO. A breakdown of noteworthy changes: - Added `HybridSample` class as the abstraction perf sample including LBR stack and call stack * Extended `PerfReader` to implement auto-detect whether input perf script output contains CS profile, then do the parsing. Multiple `HybridSample` are extracted * Speed up by aggregating `HybridSample` into `AggregatedSamples` * Added VirtualUnwinder that consumes aggregated `HybridSample` and implements unwinding of calls, returns, and linear path that contains implicit call/return from inlining. Ranges and branches counters are aggregated by the calling context.
 Here calling context is string type, each context is a pair of function name and callsite location info, the whole context is like `main:1 @ foo:2 @ bar`. * Added PorfileGenerater that accumulates counters by ranges unfolding or branch target mapping, then generates context-sensitive function profile including function body, inferring callee's head sample, callsite target samples, eventually records into ProfileMap.
 * Leveraged LLVM build-in(`SampleProfWriter`) writer to support different serialization format with no stop - `getCanonicalFnName` for callee name and name from ELF section - Added regression test for both unwinding and profile generation Test Plan: ninja & ninja check-llvm Reviewed By: hoy, wenlei, wmi Differential Revision: https://reviews.llvm.org/D89723
2020-10-19 12:55:59 -07:00
}
}
[CSSPGO] Split context string to deduplicate function name used in the context. 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
2021-08-25 11:40:34 -07:00
static SampleContextFrame
getCallerContext(SampleContextFrames CalleeContext,
SampleContextFrameVector &CallerContext) {
assert(CalleeContext.size() > 1 && "Unexpected empty context");
CalleeContext = CalleeContext.drop_back();
CallerContext.assign(CalleeContext.begin(), CalleeContext.end());
SampleContextFrame CallerFrame = CallerContext.back();
CallerContext.back().Location = LineLocation(0, 0);
[CSSPGO] Split context string to deduplicate function name used in the context. 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
2021-08-25 11:40:34 -07:00
return CallerFrame;
[CSSPGO][llvm-profgen] Context-sensitive profile data generation This stack of changes introduces `llvm-profgen` utility which generates a profile data file from given perf script data files for sample-based PGO. It’s part of(not only) the CSSPGO work. Specifically to support context-sensitive with/without pseudo probe profile, it implements a series of functionalities including perf trace parsing, instruction symbolization, LBR stack/call frame stack unwinding, pseudo probe decoding, etc. Also high throughput is achieved by multiple levels of sample aggregation and compatible format with one stop is generated at the end. Please refer to: https://groups.google.com/g/llvm-dev/c/1p1rdYbL93s for the CSSPGO RFC. This change supports context-sensitive profile data generation into llvm-profgen. With simultaneous sampling for LBR and call stack, we can identify leaf of LBR sample with calling context from stack sample . During the process of deriving fall through path from LBR entries, we unwind LBR by replaying all the calls and returns (including implicit calls/returns due to inlining) backwards on top of the sampled call stack. Then the state of call stack as we unwind through LBR always represents the calling context of current fall through path. we have two types of virtual unwinding 1) LBR unwinding and 2) linear range unwinding. Specifically, for each LBR entry which can be classified into call, return, regular branch, LBR unwinding will replay the operation by pushing, popping or switching leaf frame towards the call stack and since the initial call stack is most recently sampled, the replay should be in anti-execution order, i.e. for the regular case, pop the call stack when LBR is call, push frame on call stack when LBR is return. After each LBR processed, it also needs to align with the next LBR by going through instructions from previous LBR's target to current LBR's source, which we named linear unwinding. As instruction from linear range can come from different function by inlining, linear unwinding will do the range splitting and record counters through the range with same inline context. With each fall through path from LBR unwinding, we aggregate each sample into counters by the calling context and eventually generate full context sensitive profile (without relying on inlining) to driver compiler's PGO/FDO. A breakdown of noteworthy changes: - Added `HybridSample` class as the abstraction perf sample including LBR stack and call stack * Extended `PerfReader` to implement auto-detect whether input perf script output contains CS profile, then do the parsing. Multiple `HybridSample` are extracted * Speed up by aggregating `HybridSample` into `AggregatedSamples` * Added VirtualUnwinder that consumes aggregated `HybridSample` and implements unwinding of calls, returns, and linear path that contains implicit call/return from inlining. Ranges and branches counters are aggregated by the calling context.
 Here calling context is string type, each context is a pair of function name and callsite location info, the whole context is like `main:1 @ foo:2 @ bar`. * Added PorfileGenerater that accumulates counters by ranges unfolding or branch target mapping, then generates context-sensitive function profile including function body, inferring callee's head sample, callsite target samples, eventually records into ProfileMap.
 * Leveraged LLVM build-in(`SampleProfWriter`) writer to support different serialization format with no stop - `getCanonicalFnName` for callee name and name from ELF section - Added regression test for both unwinding and profile generation Test Plan: ninja & ninja check-llvm Reviewed By: hoy, wenlei, wmi Differential Revision: https://reviews.llvm.org/D89723
2020-10-19 12:55:59 -07:00
}
void CSProfileGenerator::populateInferredFunctionSamples() {
for (const auto &Item : ProfileMap) {
[CSSPGO] Split context string to deduplicate function name used in the context. 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
2021-08-25 11:40:34 -07:00
const auto &CalleeContext = Item.first;
[CSSPGO][llvm-profgen] Context-sensitive profile data generation This stack of changes introduces `llvm-profgen` utility which generates a profile data file from given perf script data files for sample-based PGO. It’s part of(not only) the CSSPGO work. Specifically to support context-sensitive with/without pseudo probe profile, it implements a series of functionalities including perf trace parsing, instruction symbolization, LBR stack/call frame stack unwinding, pseudo probe decoding, etc. Also high throughput is achieved by multiple levels of sample aggregation and compatible format with one stop is generated at the end. Please refer to: https://groups.google.com/g/llvm-dev/c/1p1rdYbL93s for the CSSPGO RFC. This change supports context-sensitive profile data generation into llvm-profgen. With simultaneous sampling for LBR and call stack, we can identify leaf of LBR sample with calling context from stack sample . During the process of deriving fall through path from LBR entries, we unwind LBR by replaying all the calls and returns (including implicit calls/returns due to inlining) backwards on top of the sampled call stack. Then the state of call stack as we unwind through LBR always represents the calling context of current fall through path. we have two types of virtual unwinding 1) LBR unwinding and 2) linear range unwinding. Specifically, for each LBR entry which can be classified into call, return, regular branch, LBR unwinding will replay the operation by pushing, popping or switching leaf frame towards the call stack and since the initial call stack is most recently sampled, the replay should be in anti-execution order, i.e. for the regular case, pop the call stack when LBR is call, push frame on call stack when LBR is return. After each LBR processed, it also needs to align with the next LBR by going through instructions from previous LBR's target to current LBR's source, which we named linear unwinding. As instruction from linear range can come from different function by inlining, linear unwinding will do the range splitting and record counters through the range with same inline context. With each fall through path from LBR unwinding, we aggregate each sample into counters by the calling context and eventually generate full context sensitive profile (without relying on inlining) to driver compiler's PGO/FDO. A breakdown of noteworthy changes: - Added `HybridSample` class as the abstraction perf sample including LBR stack and call stack * Extended `PerfReader` to implement auto-detect whether input perf script output contains CS profile, then do the parsing. Multiple `HybridSample` are extracted * Speed up by aggregating `HybridSample` into `AggregatedSamples` * Added VirtualUnwinder that consumes aggregated `HybridSample` and implements unwinding of calls, returns, and linear path that contains implicit call/return from inlining. Ranges and branches counters are aggregated by the calling context.
 Here calling context is string type, each context is a pair of function name and callsite location info, the whole context is like `main:1 @ foo:2 @ bar`. * Added PorfileGenerater that accumulates counters by ranges unfolding or branch target mapping, then generates context-sensitive function profile including function body, inferring callee's head sample, callsite target samples, eventually records into ProfileMap.
 * Leveraged LLVM build-in(`SampleProfWriter`) writer to support different serialization format with no stop - `getCanonicalFnName` for callee name and name from ELF section - Added regression test for both unwinding and profile generation Test Plan: ninja & ninja check-llvm Reviewed By: hoy, wenlei, wmi Differential Revision: https://reviews.llvm.org/D89723
2020-10-19 12:55:59 -07:00
const FunctionSamples &CalleeProfile = Item.second;
// If we already have head sample counts, we must have value profile
// for call sites added already. Skip to avoid double counting.
if (CalleeProfile.getHeadSamples())
continue;
// If we don't have context, nothing to do for caller's call site.
// This could happen for entry point function.
[CSSPGO] Split context string to deduplicate function name used in the context. 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
2021-08-25 11:40:34 -07:00
if (CalleeContext.isBaseContext())
[CSSPGO][llvm-profgen] Context-sensitive profile data generation This stack of changes introduces `llvm-profgen` utility which generates a profile data file from given perf script data files for sample-based PGO. It’s part of(not only) the CSSPGO work. Specifically to support context-sensitive with/without pseudo probe profile, it implements a series of functionalities including perf trace parsing, instruction symbolization, LBR stack/call frame stack unwinding, pseudo probe decoding, etc. Also high throughput is achieved by multiple levels of sample aggregation and compatible format with one stop is generated at the end. Please refer to: https://groups.google.com/g/llvm-dev/c/1p1rdYbL93s for the CSSPGO RFC. This change supports context-sensitive profile data generation into llvm-profgen. With simultaneous sampling for LBR and call stack, we can identify leaf of LBR sample with calling context from stack sample . During the process of deriving fall through path from LBR entries, we unwind LBR by replaying all the calls and returns (including implicit calls/returns due to inlining) backwards on top of the sampled call stack. Then the state of call stack as we unwind through LBR always represents the calling context of current fall through path. we have two types of virtual unwinding 1) LBR unwinding and 2) linear range unwinding. Specifically, for each LBR entry which can be classified into call, return, regular branch, LBR unwinding will replay the operation by pushing, popping or switching leaf frame towards the call stack and since the initial call stack is most recently sampled, the replay should be in anti-execution order, i.e. for the regular case, pop the call stack when LBR is call, push frame on call stack when LBR is return. After each LBR processed, it also needs to align with the next LBR by going through instructions from previous LBR's target to current LBR's source, which we named linear unwinding. As instruction from linear range can come from different function by inlining, linear unwinding will do the range splitting and record counters through the range with same inline context. With each fall through path from LBR unwinding, we aggregate each sample into counters by the calling context and eventually generate full context sensitive profile (without relying on inlining) to driver compiler's PGO/FDO. A breakdown of noteworthy changes: - Added `HybridSample` class as the abstraction perf sample including LBR stack and call stack * Extended `PerfReader` to implement auto-detect whether input perf script output contains CS profile, then do the parsing. Multiple `HybridSample` are extracted * Speed up by aggregating `HybridSample` into `AggregatedSamples` * Added VirtualUnwinder that consumes aggregated `HybridSample` and implements unwinding of calls, returns, and linear path that contains implicit call/return from inlining. Ranges and branches counters are aggregated by the calling context.
 Here calling context is string type, each context is a pair of function name and callsite location info, the whole context is like `main:1 @ foo:2 @ bar`. * Added PorfileGenerater that accumulates counters by ranges unfolding or branch target mapping, then generates context-sensitive function profile including function body, inferring callee's head sample, callsite target samples, eventually records into ProfileMap.
 * Leveraged LLVM build-in(`SampleProfWriter`) writer to support different serialization format with no stop - `getCanonicalFnName` for callee name and name from ELF section - Added regression test for both unwinding and profile generation Test Plan: ninja & ninja check-llvm Reviewed By: hoy, wenlei, wmi Differential Revision: https://reviews.llvm.org/D89723
2020-10-19 12:55:59 -07:00
continue;
// Infer Caller's frame loc and context ID through string splitting
[CSSPGO] Split context string to deduplicate function name used in the context. 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
2021-08-25 11:40:34 -07:00
SampleContextFrameVector CallerContextId;
SampleContextFrame &&CallerLeafFrameLoc =
getCallerContext(CalleeContext.getContextFrames(), CallerContextId);
SampleContextFrames CallerContext(CallerContextId);
[CSSPGO][llvm-profgen] Context-sensitive profile data generation This stack of changes introduces `llvm-profgen` utility which generates a profile data file from given perf script data files for sample-based PGO. It’s part of(not only) the CSSPGO work. Specifically to support context-sensitive with/without pseudo probe profile, it implements a series of functionalities including perf trace parsing, instruction symbolization, LBR stack/call frame stack unwinding, pseudo probe decoding, etc. Also high throughput is achieved by multiple levels of sample aggregation and compatible format with one stop is generated at the end. Please refer to: https://groups.google.com/g/llvm-dev/c/1p1rdYbL93s for the CSSPGO RFC. This change supports context-sensitive profile data generation into llvm-profgen. With simultaneous sampling for LBR and call stack, we can identify leaf of LBR sample with calling context from stack sample . During the process of deriving fall through path from LBR entries, we unwind LBR by replaying all the calls and returns (including implicit calls/returns due to inlining) backwards on top of the sampled call stack. Then the state of call stack as we unwind through LBR always represents the calling context of current fall through path. we have two types of virtual unwinding 1) LBR unwinding and 2) linear range unwinding. Specifically, for each LBR entry which can be classified into call, return, regular branch, LBR unwinding will replay the operation by pushing, popping or switching leaf frame towards the call stack and since the initial call stack is most recently sampled, the replay should be in anti-execution order, i.e. for the regular case, pop the call stack when LBR is call, push frame on call stack when LBR is return. After each LBR processed, it also needs to align with the next LBR by going through instructions from previous LBR's target to current LBR's source, which we named linear unwinding. As instruction from linear range can come from different function by inlining, linear unwinding will do the range splitting and record counters through the range with same inline context. With each fall through path from LBR unwinding, we aggregate each sample into counters by the calling context and eventually generate full context sensitive profile (without relying on inlining) to driver compiler's PGO/FDO. A breakdown of noteworthy changes: - Added `HybridSample` class as the abstraction perf sample including LBR stack and call stack * Extended `PerfReader` to implement auto-detect whether input perf script output contains CS profile, then do the parsing. Multiple `HybridSample` are extracted * Speed up by aggregating `HybridSample` into `AggregatedSamples` * Added VirtualUnwinder that consumes aggregated `HybridSample` and implements unwinding of calls, returns, and linear path that contains implicit call/return from inlining. Ranges and branches counters are aggregated by the calling context.
 Here calling context is string type, each context is a pair of function name and callsite location info, the whole context is like `main:1 @ foo:2 @ bar`. * Added PorfileGenerater that accumulates counters by ranges unfolding or branch target mapping, then generates context-sensitive function profile including function body, inferring callee's head sample, callsite target samples, eventually records into ProfileMap.
 * Leveraged LLVM build-in(`SampleProfWriter`) writer to support different serialization format with no stop - `getCanonicalFnName` for callee name and name from ELF section - Added regression test for both unwinding and profile generation Test Plan: ninja & ninja check-llvm Reviewed By: hoy, wenlei, wmi Differential Revision: https://reviews.llvm.org/D89723
2020-10-19 12:55:59 -07:00
// It's possible that we haven't seen any sample directly in the caller,
// in which case CallerProfile will not exist. But we can't modify
// ProfileMap while iterating it.
// TODO: created function profile for those callers too
[CSSPGO] Split context string to deduplicate function name used in the context. 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
2021-08-25 11:40:34 -07:00
if (ProfileMap.find(CallerContext) == ProfileMap.end())
[CSSPGO][llvm-profgen] Context-sensitive profile data generation This stack of changes introduces `llvm-profgen` utility which generates a profile data file from given perf script data files for sample-based PGO. It’s part of(not only) the CSSPGO work. Specifically to support context-sensitive with/without pseudo probe profile, it implements a series of functionalities including perf trace parsing, instruction symbolization, LBR stack/call frame stack unwinding, pseudo probe decoding, etc. Also high throughput is achieved by multiple levels of sample aggregation and compatible format with one stop is generated at the end. Please refer to: https://groups.google.com/g/llvm-dev/c/1p1rdYbL93s for the CSSPGO RFC. This change supports context-sensitive profile data generation into llvm-profgen. With simultaneous sampling for LBR and call stack, we can identify leaf of LBR sample with calling context from stack sample . During the process of deriving fall through path from LBR entries, we unwind LBR by replaying all the calls and returns (including implicit calls/returns due to inlining) backwards on top of the sampled call stack. Then the state of call stack as we unwind through LBR always represents the calling context of current fall through path. we have two types of virtual unwinding 1) LBR unwinding and 2) linear range unwinding. Specifically, for each LBR entry which can be classified into call, return, regular branch, LBR unwinding will replay the operation by pushing, popping or switching leaf frame towards the call stack and since the initial call stack is most recently sampled, the replay should be in anti-execution order, i.e. for the regular case, pop the call stack when LBR is call, push frame on call stack when LBR is return. After each LBR processed, it also needs to align with the next LBR by going through instructions from previous LBR's target to current LBR's source, which we named linear unwinding. As instruction from linear range can come from different function by inlining, linear unwinding will do the range splitting and record counters through the range with same inline context. With each fall through path from LBR unwinding, we aggregate each sample into counters by the calling context and eventually generate full context sensitive profile (without relying on inlining) to driver compiler's PGO/FDO. A breakdown of noteworthy changes: - Added `HybridSample` class as the abstraction perf sample including LBR stack and call stack * Extended `PerfReader` to implement auto-detect whether input perf script output contains CS profile, then do the parsing. Multiple `HybridSample` are extracted * Speed up by aggregating `HybridSample` into `AggregatedSamples` * Added VirtualUnwinder that consumes aggregated `HybridSample` and implements unwinding of calls, returns, and linear path that contains implicit call/return from inlining. Ranges and branches counters are aggregated by the calling context.
 Here calling context is string type, each context is a pair of function name and callsite location info, the whole context is like `main:1 @ foo:2 @ bar`. * Added PorfileGenerater that accumulates counters by ranges unfolding or branch target mapping, then generates context-sensitive function profile including function body, inferring callee's head sample, callsite target samples, eventually records into ProfileMap.
 * Leveraged LLVM build-in(`SampleProfWriter`) writer to support different serialization format with no stop - `getCanonicalFnName` for callee name and name from ELF section - Added regression test for both unwinding and profile generation Test Plan: ninja & ninja check-llvm Reviewed By: hoy, wenlei, wmi Differential Revision: https://reviews.llvm.org/D89723
2020-10-19 12:55:59 -07:00
continue;
[CSSPGO] Split context string to deduplicate function name used in the context. 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
2021-08-25 11:40:34 -07:00
FunctionSamples &CallerProfile = ProfileMap[CallerContext];
[CSSPGO][llvm-profgen] Context-sensitive profile data generation This stack of changes introduces `llvm-profgen` utility which generates a profile data file from given perf script data files for sample-based PGO. It’s part of(not only) the CSSPGO work. Specifically to support context-sensitive with/without pseudo probe profile, it implements a series of functionalities including perf trace parsing, instruction symbolization, LBR stack/call frame stack unwinding, pseudo probe decoding, etc. Also high throughput is achieved by multiple levels of sample aggregation and compatible format with one stop is generated at the end. Please refer to: https://groups.google.com/g/llvm-dev/c/1p1rdYbL93s for the CSSPGO RFC. This change supports context-sensitive profile data generation into llvm-profgen. With simultaneous sampling for LBR and call stack, we can identify leaf of LBR sample with calling context from stack sample . During the process of deriving fall through path from LBR entries, we unwind LBR by replaying all the calls and returns (including implicit calls/returns due to inlining) backwards on top of the sampled call stack. Then the state of call stack as we unwind through LBR always represents the calling context of current fall through path. we have two types of virtual unwinding 1) LBR unwinding and 2) linear range unwinding. Specifically, for each LBR entry which can be classified into call, return, regular branch, LBR unwinding will replay the operation by pushing, popping or switching leaf frame towards the call stack and since the initial call stack is most recently sampled, the replay should be in anti-execution order, i.e. for the regular case, pop the call stack when LBR is call, push frame on call stack when LBR is return. After each LBR processed, it also needs to align with the next LBR by going through instructions from previous LBR's target to current LBR's source, which we named linear unwinding. As instruction from linear range can come from different function by inlining, linear unwinding will do the range splitting and record counters through the range with same inline context. With each fall through path from LBR unwinding, we aggregate each sample into counters by the calling context and eventually generate full context sensitive profile (without relying on inlining) to driver compiler's PGO/FDO. A breakdown of noteworthy changes: - Added `HybridSample` class as the abstraction perf sample including LBR stack and call stack * Extended `PerfReader` to implement auto-detect whether input perf script output contains CS profile, then do the parsing. Multiple `HybridSample` are extracted * Speed up by aggregating `HybridSample` into `AggregatedSamples` * Added VirtualUnwinder that consumes aggregated `HybridSample` and implements unwinding of calls, returns, and linear path that contains implicit call/return from inlining. Ranges and branches counters are aggregated by the calling context.
 Here calling context is string type, each context is a pair of function name and callsite location info, the whole context is like `main:1 @ foo:2 @ bar`. * Added PorfileGenerater that accumulates counters by ranges unfolding or branch target mapping, then generates context-sensitive function profile including function body, inferring callee's head sample, callsite target samples, eventually records into ProfileMap.
 * Leveraged LLVM build-in(`SampleProfWriter`) writer to support different serialization format with no stop - `getCanonicalFnName` for callee name and name from ELF section - Added regression test for both unwinding and profile generation Test Plan: ninja & ninja check-llvm Reviewed By: hoy, wenlei, wmi Differential Revision: https://reviews.llvm.org/D89723
2020-10-19 12:55:59 -07:00
// Since we don't have call count for inlined functions, we
// estimate it from inlinee's profile using entry body sample.
uint64_t EstimatedCallCount = CalleeProfile.getEntrySamples();
// If we don't have samples with location, use 1 to indicate live.
if (!EstimatedCallCount && !CalleeProfile.getBodySamples().size())
EstimatedCallCount = 1;
CallerProfile.addCalledTargetSamples(
CallerLeafFrameLoc.Location.LineOffset,
CallerLeafFrameLoc.Location.Discriminator,
[CSSPGO] Split context string to deduplicate function name used in the context. 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
2021-08-25 11:40:34 -07:00
CalleeProfile.getContext().getName(), EstimatedCallCount);
CallerProfile.addBodySamples(CallerLeafFrameLoc.Location.LineOffset,
CallerLeafFrameLoc.Location.Discriminator,
EstimatedCallCount);
CallerProfile.addTotalSamples(EstimatedCallCount);
[CSSPGO][llvm-profgen] Context-sensitive profile data generation This stack of changes introduces `llvm-profgen` utility which generates a profile data file from given perf script data files for sample-based PGO. It’s part of(not only) the CSSPGO work. Specifically to support context-sensitive with/without pseudo probe profile, it implements a series of functionalities including perf trace parsing, instruction symbolization, LBR stack/call frame stack unwinding, pseudo probe decoding, etc. Also high throughput is achieved by multiple levels of sample aggregation and compatible format with one stop is generated at the end. Please refer to: https://groups.google.com/g/llvm-dev/c/1p1rdYbL93s for the CSSPGO RFC. This change supports context-sensitive profile data generation into llvm-profgen. With simultaneous sampling for LBR and call stack, we can identify leaf of LBR sample with calling context from stack sample . During the process of deriving fall through path from LBR entries, we unwind LBR by replaying all the calls and returns (including implicit calls/returns due to inlining) backwards on top of the sampled call stack. Then the state of call stack as we unwind through LBR always represents the calling context of current fall through path. we have two types of virtual unwinding 1) LBR unwinding and 2) linear range unwinding. Specifically, for each LBR entry which can be classified into call, return, regular branch, LBR unwinding will replay the operation by pushing, popping or switching leaf frame towards the call stack and since the initial call stack is most recently sampled, the replay should be in anti-execution order, i.e. for the regular case, pop the call stack when LBR is call, push frame on call stack when LBR is return. After each LBR processed, it also needs to align with the next LBR by going through instructions from previous LBR's target to current LBR's source, which we named linear unwinding. As instruction from linear range can come from different function by inlining, linear unwinding will do the range splitting and record counters through the range with same inline context. With each fall through path from LBR unwinding, we aggregate each sample into counters by the calling context and eventually generate full context sensitive profile (without relying on inlining) to driver compiler's PGO/FDO. A breakdown of noteworthy changes: - Added `HybridSample` class as the abstraction perf sample including LBR stack and call stack * Extended `PerfReader` to implement auto-detect whether input perf script output contains CS profile, then do the parsing. Multiple `HybridSample` are extracted * Speed up by aggregating `HybridSample` into `AggregatedSamples` * Added VirtualUnwinder that consumes aggregated `HybridSample` and implements unwinding of calls, returns, and linear path that contains implicit call/return from inlining. Ranges and branches counters are aggregated by the calling context.
 Here calling context is string type, each context is a pair of function name and callsite location info, the whole context is like `main:1 @ foo:2 @ bar`. * Added PorfileGenerater that accumulates counters by ranges unfolding or branch target mapping, then generates context-sensitive function profile including function body, inferring callee's head sample, callsite target samples, eventually records into ProfileMap.
 * Leveraged LLVM build-in(`SampleProfWriter`) writer to support different serialization format with no stop - `getCanonicalFnName` for callee name and name from ELF section - Added regression test for both unwinding and profile generation Test Plan: ninja & ninja check-llvm Reviewed By: hoy, wenlei, wmi Differential Revision: https://reviews.llvm.org/D89723
2020-10-19 12:55:59 -07:00
}
}
[CSSPGO][llvm-profgen] Context-sensitive global pre-inliner This change sets up a framework in llvm-profgen to estimate inline decision and adjust context-sensitive profile based on that. We call it a global pre-inliner in llvm-profgen. It will serve two purposes: 1) Since context profile for not inlined context will be merged into base profile, if we estimate a context will not be inlined, we can merge the context profile in the output to save profile size. 2) For thinLTO, when a context involving functions from different modules is not inined, we can't merge functions profiles across modules, leading to suboptimal post-inline count quality. By estimating some inline decisions, we would be able to adjust/merge context profiles beforehand as a mitigation. Compiler inline heuristic uses inline cost which is not available in llvm-profgen. But since inline cost is closely related to size, we could get an estimate through function size from debug info. Because the size we have in llvm-profgen is the final size, it could also be more accurate than the inline cost estimation in the compiler. This change only has the framework, with a few TODOs left for follow up patches for a complete implementation: 1) We need to retrieve size for funciton//inlinee from debug info for inlining estimation. Currently we use number of samples in a profile as place holder for size estimation. 2) Currently the thresholds are using the values used by sample loader inliner. But they need to be tuned since the size here is fully optimized machine code size, instead of inline cost based on not yet fully optimized IR. Differential Revision: https://reviews.llvm.org/D99146
2021-03-05 07:50:36 -08:00
void CSProfileGenerator::postProcessProfiles() {
// Compute hot/cold threshold based on profile. This will be used for cold
// context profile merging/trimming.
computeSummaryAndThreshold();
// Run global pre-inliner to adjust/merge context profile based on estimated
// inline decisions.
if (EnableCSPreInliner) {
CSPreInliner(ProfileMap, *Binary, HotCountThreshold, ColdCountThreshold)
.run();
// Turn off the profile merger by default unless it is explicitly enabled.
if (!CSProfMergeColdContext.getNumOccurrences())
CSProfMergeColdContext = false;
}
[CSSPGO][llvm-profgen] Context-sensitive global pre-inliner This change sets up a framework in llvm-profgen to estimate inline decision and adjust context-sensitive profile based on that. We call it a global pre-inliner in llvm-profgen. It will serve two purposes: 1) Since context profile for not inlined context will be merged into base profile, if we estimate a context will not be inlined, we can merge the context profile in the output to save profile size. 2) For thinLTO, when a context involving functions from different modules is not inined, we can't merge functions profiles across modules, leading to suboptimal post-inline count quality. By estimating some inline decisions, we would be able to adjust/merge context profiles beforehand as a mitigation. Compiler inline heuristic uses inline cost which is not available in llvm-profgen. But since inline cost is closely related to size, we could get an estimate through function size from debug info. Because the size we have in llvm-profgen is the final size, it could also be more accurate than the inline cost estimation in the compiler. This change only has the framework, with a few TODOs left for follow up patches for a complete implementation: 1) We need to retrieve size for funciton//inlinee from debug info for inlining estimation. Currently we use number of samples in a profile as place holder for size estimation. 2) Currently the thresholds are using the values used by sample loader inliner. But they need to be tuned since the size here is fully optimized machine code size, instead of inline cost based on not yet fully optimized IR. Differential Revision: https://reviews.llvm.org/D99146
2021-03-05 07:50:36 -08:00
// Trim and merge cold context profile using cold threshold above.
if (TrimColdProfile || CSProfMergeColdContext) {
SampleContextTrimmer(ProfileMap)
.trimAndMergeColdContextProfiles(
HotCountThreshold, TrimColdProfile, CSProfMergeColdContext,
CSProfMaxColdContextDepth, EnableCSPreInliner);
}
// Merge function samples of CS profile to calculate profile density.
sampleprof::SampleProfileMap ContextLessProfiles;
for (const auto &I : ProfileMap) {
ContextLessProfiles[I.second.getName()].merge(I.second);
}
calculateAndShowDensity(ContextLessProfiles);
[CSSPGO] Use nested context-sensitive profile. 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
2021-12-14 10:03:05 -08:00
if (GenCSNestedProfile) {
CSProfileConverter CSConverter(ProfileMap);
CSConverter.convertProfiles();
FunctionSamples::ProfileIsCSFlat = false;
FunctionSamples::ProfileIsCSNested = EnableCSPreInliner;
}
[CSSPGO][llvm-profgen] Context-sensitive global pre-inliner This change sets up a framework in llvm-profgen to estimate inline decision and adjust context-sensitive profile based on that. We call it a global pre-inliner in llvm-profgen. It will serve two purposes: 1) Since context profile for not inlined context will be merged into base profile, if we estimate a context will not be inlined, we can merge the context profile in the output to save profile size. 2) For thinLTO, when a context involving functions from different modules is not inined, we can't merge functions profiles across modules, leading to suboptimal post-inline count quality. By estimating some inline decisions, we would be able to adjust/merge context profiles beforehand as a mitigation. Compiler inline heuristic uses inline cost which is not available in llvm-profgen. But since inline cost is closely related to size, we could get an estimate through function size from debug info. Because the size we have in llvm-profgen is the final size, it could also be more accurate than the inline cost estimation in the compiler. This change only has the framework, with a few TODOs left for follow up patches for a complete implementation: 1) We need to retrieve size for funciton//inlinee from debug info for inlining estimation. Currently we use number of samples in a profile as place holder for size estimation. 2) Currently the thresholds are using the values used by sample loader inliner. But they need to be tuned since the size here is fully optimized machine code size, instead of inline cost based on not yet fully optimized IR. Differential Revision: https://reviews.llvm.org/D99146
2021-03-05 07:50:36 -08:00
}
void ProfileGeneratorBase::computeSummaryAndThreshold() {
SampleProfileSummaryBuilder Builder(ProfileSummaryBuilder::DefaultCutoffs);
auto Summary = Builder.computeSummaryForProfiles(ProfileMap);
HotCountThreshold = ProfileSummaryBuilder::getHotCountThreshold(
(Summary->getDetailedSummary()));
ColdCountThreshold = ProfileSummaryBuilder::getColdCountThreshold(
(Summary->getDetailedSummary()));
}
void ProfileGeneratorBase::extractProbesFromRange(
const RangeSample &RangeCounter, ProbeCounterMap &ProbeCounter,
bool FindDisjointRanges) {
const RangeSample *PRanges = &RangeCounter;
RangeSample Ranges;
if (FindDisjointRanges) {
findDisjointRanges(Ranges, RangeCounter);
PRanges = &Ranges;
2021-01-11 09:08:39 -08:00
}
for (const auto &Range : *PRanges) {
2021-01-11 09:08:39 -08:00
uint64_t RangeBegin = Binary->offsetToVirtualAddr(Range.first.first);
uint64_t RangeEnd = Binary->offsetToVirtualAddr(Range.first.second);
uint64_t Count = Range.second;
InstructionPointer IP(Binary, RangeBegin, true);
// Disjoint ranges may have range in the middle of two instr,
// e.g. If Instr1 at Addr1, and Instr2 at Addr2, disjoint range
// can be Addr1+1 to Addr2-1. We should ignore such range.
if (IP.Address > RangeEnd)
continue;
do {
2021-01-11 09:08:39 -08:00
const AddressProbesMap &Address2ProbesMap =
Binary->getAddress2ProbesMap();
auto It = Address2ProbesMap.find(IP.Address);
if (It != Address2ProbesMap.end()) {
for (const auto &Probe : It->second) {
ProbeCounter[&Probe] += Count;
}
}
} while (IP.advance() && IP.Address <= RangeEnd);
2021-01-11 09:08:39 -08:00
}
}
static void
extractPrefixContextStack(SampleContextFrameVector &ContextStack,
const SmallVectorImpl<uint64_t> &Addresses,
ProfiledBinary *Binary) {
SmallVector<const MCDecodedPseudoProbe *, 16> Probes;
for (auto Addr : reverse(Addresses)) {
const MCDecodedPseudoProbe *CallProbe = Binary->getCallProbeForAddr(Addr);
// These could be the cases when a probe is not found at a calliste. Cutting
// off the context from here since the inliner will not know how to consume
// a context with unknown callsites.
// 1. for functions that are not sampled when
// --decode-probe-for-profiled-functions-only is on.
// 2. for a merged callsite. Callsite merging may cause the loss of original
// probe IDs.
// 3. for an external callsite.
if (!CallProbe)
break;
Probes.push_back(CallProbe);
}
std::reverse(Probes.begin(), Probes.end());
// Extract context stack for reusing, leaf context stack will be added
// compressed while looking up function profile.
for (const auto *P : Probes) {
Binary->getInlineContextForProbe(P, ContextStack, true);
}
}
void CSProfileGenerator::generateProbeBasedProfile() {
// Enable pseudo probe functionalities in SampleProf
FunctionSamples::ProfileIsProbeBased = true;
for (const auto &CI : *SampleCounters) {
const AddrBasedCtxKey *CtxKey =
dyn_cast<AddrBasedCtxKey>(CI.first.getPtr());
SampleContextFrameVector ContextStack;
extractPrefixContextStack(ContextStack, CtxKey->Context, Binary);
// Fill in function body samples from probes, also infer caller's samples
// from callee's probe
populateBodySamplesWithProbes(CI.second.RangeCounter, ContextStack);
// Fill in boundary samples for a call probe
populateBoundarySamplesWithProbes(CI.second.BranchCounter, ContextStack);
}
}
void CSProfileGenerator::populateBodySamplesWithProbes(
[CSSPGO] Split context string to deduplicate function name used in the context. 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
2021-08-25 11:40:34 -07:00
const RangeSample &RangeCounter, SampleContextFrames ContextStack) {
2021-01-11 09:08:39 -08:00
ProbeCounterMap ProbeCounter;
// Extract the top frame probes by looking up each address among the range in
// the Address2ProbeMap
extractProbesFromRange(RangeCounter, ProbeCounter);
std::unordered_map<MCDecodedPseudoProbeInlineTree *,
std::unordered_set<FunctionSamples *>>
FrameSamples;
for (const auto &PI : ProbeCounter) {
const MCDecodedPseudoProbe *Probe = PI.first;
2021-01-11 09:08:39 -08:00
uint64_t Count = PI.second;
// Disjoint ranges have introduce zero-filled gap that
// doesn't belong to current context, filter them out.
if (!Probe->isBlock() || Count == 0)
continue;
2021-01-11 09:08:39 -08:00
FunctionSamples &FunctionProfile =
[CSSPGO] Split context string to deduplicate function name used in the context. 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
2021-08-25 11:40:34 -07:00
getFunctionProfileForLeafProbe(ContextStack, Probe);
// Record the current frame and FunctionProfile whenever samples are
// collected for non-danglie probes. This is for reporting all of the
// zero count probes of the frame later.
FrameSamples[Probe->getInlineTreeNode()].insert(&FunctionProfile);
FunctionProfile.addBodySamplesForProbe(Probe->getIndex(), Count);
2021-01-11 09:08:39 -08:00
FunctionProfile.addTotalSamples(Count);
if (Probe->isEntry()) {
FunctionProfile.addHeadSamples(Count);
// Look up for the caller's function profile
const auto *InlinerDesc = Binary->getInlinerDescForProbe(Probe);
SampleContextFrames CalleeContextId =
FunctionProfile.getContext().getContextFrames();
if (InlinerDesc != nullptr && CalleeContextId.size() > 1) {
2021-01-11 09:08:39 -08:00
// Since the context id will be compressed, we have to use callee's
// context id to infer caller's context id to ensure they share the
// same context prefix.
[CSSPGO] Split context string to deduplicate function name used in the context. 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
2021-08-25 11:40:34 -07:00
SampleContextFrameVector CallerContextId;
SampleContextFrame &&CallerLeafFrameLoc =
2021-01-11 09:08:39 -08:00
getCallerContext(CalleeContextId, CallerContextId);
uint64_t CallerIndex = CallerLeafFrameLoc.Location.LineOffset;
2021-01-11 09:08:39 -08:00
assert(CallerIndex &&
"Inferred caller's location index shouldn't be zero!");
FunctionSamples &CallerProfile =
getFunctionProfileForContext(CallerContextId);
2021-01-11 09:08:39 -08:00
CallerProfile.setFunctionHash(InlinerDesc->FuncHash);
CallerProfile.addBodySamples(CallerIndex, 0, Count);
CallerProfile.addTotalSamples(Count);
CallerProfile.addCalledTargetSamples(
[CSSPGO] Split context string to deduplicate function name used in the context. 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
2021-08-25 11:40:34 -07:00
CallerIndex, 0, FunctionProfile.getContext().getName(), Count);
2021-01-11 09:08:39 -08:00
}
}
}
// Assign zero count for remaining probes without sample hits to
// differentiate from probes optimized away, of which the counts are unknown
// and will be inferred by the compiler.
for (auto &I : FrameSamples) {
for (auto *FunctionProfile : I.second) {
for (auto *Probe : I.first->getProbes()) {
FunctionProfile->addBodySamplesForProbe(Probe->getIndex(), 0);
}
}
2021-01-11 09:08:39 -08:00
}
}
void CSProfileGenerator::populateBoundarySamplesWithProbes(
[CSSPGO] Split context string to deduplicate function name used in the context. 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
2021-08-25 11:40:34 -07:00
const BranchSample &BranchCounter, SampleContextFrames ContextStack) {
for (const auto &BI : BranchCounter) {
2021-01-11 09:08:39 -08:00
uint64_t SourceOffset = BI.first.first;
uint64_t TargetOffset = BI.first.second;
uint64_t Count = BI.second;
uint64_t SourceAddress = Binary->offsetToVirtualAddr(SourceOffset);
const MCDecodedPseudoProbe *CallProbe =
Binary->getCallProbeForAddr(SourceAddress);
2021-01-11 09:08:39 -08:00
if (CallProbe == nullptr)
continue;
FunctionSamples &FunctionProfile =
[CSSPGO] Split context string to deduplicate function name used in the context. 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
2021-08-25 11:40:34 -07:00
getFunctionProfileForLeafProbe(ContextStack, CallProbe);
FunctionProfile.addBodySamples(CallProbe->getIndex(), 0, Count);
2021-01-11 09:08:39 -08:00
FunctionProfile.addTotalSamples(Count);
StringRef CalleeName = getCalleeNameForOffset(TargetOffset);
2021-01-11 09:08:39 -08:00
if (CalleeName.size() == 0)
continue;
FunctionProfile.addCalledTargetSamples(CallProbe->getIndex(), 0, CalleeName,
2021-01-11 09:08:39 -08:00
Count);
}
}
FunctionSamples &CSProfileGenerator::getFunctionProfileForLeafProbe(
[CSSPGO] Split context string to deduplicate function name used in the context. 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
2021-08-25 11:40:34 -07:00
SampleContextFrames ContextStack, const MCDecodedPseudoProbe *LeafProbe) {
// Explicitly copy the context for appending the leaf context
SampleContextFrameVector NewContextStack(ContextStack.begin(),
ContextStack.end());
Binary->getInlineContextForProbe(LeafProbe, NewContextStack, true);
2021-01-11 09:08:39 -08:00
// For leaf inlined context with the top frame, we should strip off the top
// frame's probe id, like:
// Inlined stack: [foo:1, bar:2], the ContextId will be "foo:1 @ bar"
[CSSPGO] Split context string to deduplicate function name used in the context. 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
2021-08-25 11:40:34 -07:00
auto LeafFrame = NewContextStack.back();
LeafFrame.Location = LineLocation(0, 0);
[CSSPGO] Split context string to deduplicate function name used in the context. 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
2021-08-25 11:40:34 -07:00
NewContextStack.pop_back();
// Compress the context string except for the leaf frame
CSProfileGenerator::compressRecursionContext(NewContextStack);
CSProfileGenerator::trimContext(NewContextStack);
NewContextStack.push_back(LeafFrame);
const auto *FuncDesc = Binary->getFuncDescForGUID(LeafProbe->getGuid());
bool WasLeafInlined = LeafProbe->getInlineTreeNode()->hasInlineSite();
[CSSPGO] Split context string to deduplicate function name used in the context. 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
2021-08-25 11:40:34 -07:00
FunctionSamples &FunctionProile =
getFunctionProfileForContext(NewContextStack, WasLeafInlined);
[CSSPGO] Split context string to deduplicate function name used in the context. 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
2021-08-25 11:40:34 -07:00
FunctionProile.setFunctionHash(FuncDesc->FuncHash);
return FunctionProile;
2021-01-11 09:08:39 -08:00
}
[CSSPGO][llvm-profgen] Context-sensitive profile data generation This stack of changes introduces `llvm-profgen` utility which generates a profile data file from given perf script data files for sample-based PGO. It’s part of(not only) the CSSPGO work. Specifically to support context-sensitive with/without pseudo probe profile, it implements a series of functionalities including perf trace parsing, instruction symbolization, LBR stack/call frame stack unwinding, pseudo probe decoding, etc. Also high throughput is achieved by multiple levels of sample aggregation and compatible format with one stop is generated at the end. Please refer to: https://groups.google.com/g/llvm-dev/c/1p1rdYbL93s for the CSSPGO RFC. This change supports context-sensitive profile data generation into llvm-profgen. With simultaneous sampling for LBR and call stack, we can identify leaf of LBR sample with calling context from stack sample . During the process of deriving fall through path from LBR entries, we unwind LBR by replaying all the calls and returns (including implicit calls/returns due to inlining) backwards on top of the sampled call stack. Then the state of call stack as we unwind through LBR always represents the calling context of current fall through path. we have two types of virtual unwinding 1) LBR unwinding and 2) linear range unwinding. Specifically, for each LBR entry which can be classified into call, return, regular branch, LBR unwinding will replay the operation by pushing, popping or switching leaf frame towards the call stack and since the initial call stack is most recently sampled, the replay should be in anti-execution order, i.e. for the regular case, pop the call stack when LBR is call, push frame on call stack when LBR is return. After each LBR processed, it also needs to align with the next LBR by going through instructions from previous LBR's target to current LBR's source, which we named linear unwinding. As instruction from linear range can come from different function by inlining, linear unwinding will do the range splitting and record counters through the range with same inline context. With each fall through path from LBR unwinding, we aggregate each sample into counters by the calling context and eventually generate full context sensitive profile (without relying on inlining) to driver compiler's PGO/FDO. A breakdown of noteworthy changes: - Added `HybridSample` class as the abstraction perf sample including LBR stack and call stack * Extended `PerfReader` to implement auto-detect whether input perf script output contains CS profile, then do the parsing. Multiple `HybridSample` are extracted * Speed up by aggregating `HybridSample` into `AggregatedSamples` * Added VirtualUnwinder that consumes aggregated `HybridSample` and implements unwinding of calls, returns, and linear path that contains implicit call/return from inlining. Ranges and branches counters are aggregated by the calling context.
 Here calling context is string type, each context is a pair of function name and callsite location info, the whole context is like `main:1 @ foo:2 @ bar`. * Added PorfileGenerater that accumulates counters by ranges unfolding or branch target mapping, then generates context-sensitive function profile including function body, inferring callee's head sample, callsite target samples, eventually records into ProfileMap.
 * Leveraged LLVM build-in(`SampleProfWriter`) writer to support different serialization format with no stop - `getCanonicalFnName` for callee name and name from ELF section - Added regression test for both unwinding and profile generation Test Plan: ninja & ninja check-llvm Reviewed By: hoy, wenlei, wmi Differential Revision: https://reviews.llvm.org/D89723
2020-10-19 12:55:59 -07:00
} // end namespace sampleprof
} // end namespace llvm