[lld][InstrProf] Sort startup functions for compression #107348
Merged
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
|
@llvm/pr-subscribers-lld @llvm/pr-subscribers-lld-macho Author: Ellis Hoag (ellishg) ChangesBefore this, startup functions were only ordered to reduce the number of page faults. This patch introduces I built a large binary without this flag (Base) and with this flag (Startup Compression). I also removed all startup hashes (Remove Startup Hashes) to see how ordering them for startup impacts the number of page faults. I recorded "Page Fault Curve Area" from the verbose logs and compressed the resulting binary using
Depends on #107347 Full diff: https://github.com/llvm/llvm-project/pull/107348.diff 7 Files Affected:
diff --git a/lld/MachO/BPSectionOrderer.cpp b/lld/MachO/BPSectionOrderer.cpp
index 568843d72bbb50..71758070ddc8ae 100644
--- a/lld/MachO/BPSectionOrderer.cpp
+++ b/lld/MachO/BPSectionOrderer.cpp
@@ -60,12 +60,13 @@ getRelocHash(const Reloc &reloc,
return getRelocHash(kind, sectionIdx.value_or(0), 0, reloc.addend);
}
-static void constructNodesForCompression(
- const SmallVector<const InputSection *> §ions,
+static SmallVector<
+ std::pair<unsigned, SmallVector<BPFunctionNode::UtilityNodeT>>>
+getUnsForCompression(
+ ArrayRef<const InputSection *> sections,
const DenseMap<const InputSection *, uint64_t> §ionToIdx,
- const SmallVector<unsigned> §ionIdxs,
- std::vector<BPFunctionNode> &nodes,
- DenseMap<unsigned, SmallVector<unsigned>> &duplicateSectionIdxs,
+ ArrayRef<unsigned> sectionIdxs,
+ DenseMap<unsigned, SmallVector<unsigned>> *duplicateSectionIdxs,
BPFunctionNode::UtilityNodeT &maxUN) {
TimeTraceScope timeScope("Build nodes for compression");
@@ -103,54 +104,59 @@ static void constructNodesForCompression(
for (auto hash : hashes)
++hashFrequency[hash];
- // Merge section that are nearly identical
- SmallVector<std::pair<unsigned, SmallVector<uint64_t>>> newSectionHashes;
- DenseMap<uint64_t, unsigned> wholeHashToSectionIdx;
- for (auto &[sectionIdx, hashes] : sectionHashes) {
- uint64_t wholeHash = 0;
- for (auto hash : hashes)
- if (hashFrequency[hash] > 5)
- wholeHash ^= hash;
- auto [it, wasInserted] =
- wholeHashToSectionIdx.insert(std::make_pair(wholeHash, sectionIdx));
- if (wasInserted) {
- newSectionHashes.emplace_back(sectionIdx, hashes);
- } else {
- duplicateSectionIdxs[it->getSecond()].push_back(sectionIdx);
+ if (duplicateSectionIdxs) {
+ // Merge section that are nearly identical
+ SmallVector<std::pair<unsigned, SmallVector<uint64_t>>> newSectionHashes;
+ DenseMap<uint64_t, unsigned> wholeHashToSectionIdx;
+ for (auto &[sectionIdx, hashes] : sectionHashes) {
+ uint64_t wholeHash = 0;
+ for (auto hash : hashes)
+ if (hashFrequency[hash] > 5)
+ wholeHash ^= hash;
+ auto [it, wasInserted] =
+ wholeHashToSectionIdx.insert(std::make_pair(wholeHash, sectionIdx));
+ if (wasInserted) {
+ newSectionHashes.emplace_back(sectionIdx, hashes);
+ } else {
+ (*duplicateSectionIdxs)[it->getSecond()].push_back(sectionIdx);
+ }
}
- }
- sectionHashes = newSectionHashes;
+ sectionHashes = newSectionHashes;
- // Recompute hash frequencies
- hashFrequency.clear();
- for (auto &[sectionIdx, hashes] : sectionHashes)
- for (auto hash : hashes)
- ++hashFrequency[hash];
+ // Recompute hash frequencies
+ hashFrequency.clear();
+ for (auto &[sectionIdx, hashes] : sectionHashes)
+ for (auto hash : hashes)
+ ++hashFrequency[hash];
+ }
// Filter rare and common hashes and assign each a unique utility node that
// doesn't conflict with the trace utility nodes
DenseMap<uint64_t, BPFunctionNode::UtilityNodeT> hashToUN;
for (auto &[hash, frequency] : hashFrequency) {
- if (frequency <= 1 || frequency * 2 > wholeHashToSectionIdx.size())
+ if (frequency <= 1 || frequency * 2 > sectionHashes.size())
continue;
hashToUN[hash] = ++maxUN;
}
- std::vector<BPFunctionNode::UtilityNodeT> uns;
+ SmallVector<std::pair<unsigned, SmallVector<BPFunctionNode::UtilityNodeT>>>
+ sectionUns;
for (auto &[sectionIdx, hashes] : sectionHashes) {
+ SmallVector<BPFunctionNode::UtilityNodeT> uns;
for (auto &hash : hashes) {
auto it = hashToUN.find(hash);
if (it != hashToUN.end())
uns.push_back(it->second);
}
- nodes.emplace_back(sectionIdx, uns);
- uns.clear();
+ sectionUns.emplace_back(sectionIdx, uns);
}
+ return sectionUns;
}
DenseMap<const InputSection *, size_t> lld::macho::runBalancedPartitioning(
size_t &highestAvailablePriority, StringRef profilePath,
- bool forFunctionCompression, bool forDataCompression, bool verbose) {
+ bool forFunctionCompression, bool forDataCompression,
+ bool compressionSortStartupFunctions, bool verbose) {
SmallVector<const InputSection *> sections;
DenseMap<const InputSection *, uint64_t> sectionToIdx;
@@ -185,10 +191,11 @@ DenseMap<const InputSection *, size_t> lld::macho::runBalancedPartitioning(
sectionIdxs.end());
}
- std::vector<BPFunctionNode> nodesForStartup;
BPFunctionNode::UtilityNodeT maxUN = 0;
DenseMap<unsigned, SmallVector<BPFunctionNode::UtilityNodeT>>
startupSectionIdxUNs;
+ // Used to define the initial order for startup functions.
+ DenseMap<unsigned, size_t> sectionIdxToTimestamp;
std::unique_ptr<InstrProfReader> reader;
if (!profilePath.empty()) {
auto fs = vfs::getRealFileSystem();
@@ -202,8 +209,6 @@ DenseMap<const InputSection *, size_t> lld::macho::runBalancedPartitioning(
}
auto &traces = reader->getTemporalProfTraces();
- // Used to define the initial order for startup functions.
- DenseMap<unsigned, size_t> sectionIdxToTimestamp;
DenseMap<unsigned, BPFunctionNode::UtilityNodeT> sectionIdxToFirstUN;
for (size_t traceIdx = 0; traceIdx < traces.size(); traceIdx++) {
uint64_t currentSize = 0, cutoffSize = 1;
@@ -245,15 +250,6 @@ DenseMap<const InputSection *, size_t> lld::macho::runBalancedPartitioning(
++maxUN;
sectionIdxToFirstUN.clear();
}
-
- // These uns should already be sorted without duplicates.
- for (auto &[sectionIdx, uns] : startupSectionIdxUNs)
- nodesForStartup.emplace_back(sectionIdx, uns);
-
- llvm::sort(nodesForStartup, [§ionIdxToTimestamp](auto &L, auto &R) {
- return std::make_pair(sectionIdxToTimestamp[L.Id], L.Id) <
- std::make_pair(sectionIdxToTimestamp[R.Id], R.Id);
- });
}
SmallVector<unsigned> sectionIdxsForFunctionCompression,
@@ -271,21 +267,45 @@ DenseMap<const InputSection *, size_t> lld::macho::runBalancedPartitioning(
}
}
- std::vector<BPFunctionNode> nodesForFunctionCompression,
- nodesForDataCompression;
- // Map a section index (to be ordered for compression) to a list of duplicate
- // section indices (not ordered for compression).
- DenseMap<unsigned, SmallVector<unsigned>> duplicateFunctionSectionIdxs,
- duplicateDataSectionIdxs;
- constructNodesForCompression(
+ SmallVector<unsigned> startupIdxs;
+ if (compressionSortStartupFunctions)
+ for (auto &[sectionIdx, uns] : startupSectionIdxUNs)
+ startupIdxs.push_back(sectionIdx);
+ auto unsForStartupFunctionCompression =
+ getUnsForCompression(sections, sectionToIdx, startupIdxs, nullptr, maxUN);
+
+ // Map a section index (order directly) to a list of duplicate section indices
+ // (not ordered directly).
+ DenseMap<unsigned, SmallVector<unsigned>> duplicateSectionIdxs;
+ auto unsForFunctionCompression = getUnsForCompression(
sections, sectionToIdx, sectionIdxsForFunctionCompression,
- nodesForFunctionCompression, duplicateFunctionSectionIdxs, maxUN);
- constructNodesForCompression(
+ &duplicateSectionIdxs, maxUN);
+ auto unsForDataCompression = getUnsForCompression(
sections, sectionToIdx, sectionIdxsForDataCompression,
- nodesForDataCompression, duplicateDataSectionIdxs, maxUN);
+ &duplicateSectionIdxs, maxUN);
- // Sort nodes by their Id (which is the section index) because the input
- // linker order tends to be not bad
+ for (auto &[sectionIdx, compressionUns] : unsForStartupFunctionCompression) {
+ auto &uns = startupSectionIdxUNs[sectionIdx];
+ uns.append(compressionUns);
+ llvm::sort(uns);
+ uns.erase(std::unique(uns.begin(), uns.end()), uns.end());
+ }
+ std::vector<BPFunctionNode> nodesForStartup, nodesForFunctionCompression,
+ nodesForDataCompression;
+ for (auto &[sectionIdx, uns] : startupSectionIdxUNs)
+ nodesForStartup.emplace_back(sectionIdx, uns);
+ for (auto &[sectionIdx, uns] : unsForFunctionCompression)
+ nodesForFunctionCompression.emplace_back(sectionIdx, uns);
+ for (auto &[sectionIdx, uns] : unsForDataCompression)
+ nodesForDataCompression.emplace_back(sectionIdx, uns);
+
+ // Use the first timestamp to define the initial order for startup nodes.
+ llvm::sort(nodesForStartup, [§ionIdxToTimestamp](auto &L, auto &R) {
+ return std::make_pair(sectionIdxToTimestamp[L.Id], L.Id) <
+ std::make_pair(sectionIdxToTimestamp[R.Id], R.Id);
+ });
+ // Sort compression nodes by their Id (which is the section index) because the
+ // input linker order tends to be not bad.
llvm::sort(nodesForFunctionCompression,
[](auto &L, auto &R) { return L.Id < R.Id; });
llvm::sort(nodesForDataCompression,
@@ -318,8 +338,8 @@ DenseMap<const InputSection *, size_t> lld::macho::runBalancedPartitioning(
if (orderedSections.insert(isec))
++numCodeCompressionSections;
- auto It = duplicateFunctionSectionIdxs.find(node.Id);
- if (It == duplicateFunctionSectionIdxs.end())
+ auto It = duplicateSectionIdxs.find(node.Id);
+ if (It == duplicateSectionIdxs.end())
continue;
for (auto dupSecIdx : It->getSecond()) {
const auto *dupIsec = sections[dupSecIdx];
@@ -332,8 +352,8 @@ DenseMap<const InputSection *, size_t> lld::macho::runBalancedPartitioning(
const auto *isec = sections[node.Id];
if (orderedSections.insert(isec))
++numDataCompressionSections;
- auto It = duplicateDataSectionIdxs.find(node.Id);
- if (It == duplicateDataSectionIdxs.end())
+ auto It = duplicateSectionIdxs.find(node.Id);
+ if (It == duplicateSectionIdxs.end())
continue;
for (auto dupSecIdx : It->getSecond()) {
const auto *dupIsec = sections[dupSecIdx];
diff --git a/lld/MachO/BPSectionOrderer.h b/lld/MachO/BPSectionOrderer.h
index 6f9eefd5d82beb..cefd0ceb10e561 100644
--- a/lld/MachO/BPSectionOrderer.h
+++ b/lld/MachO/BPSectionOrderer.h
@@ -30,7 +30,7 @@ llvm::DenseMap<const lld::macho::InputSection *, size_t>
runBalancedPartitioning(size_t &highestAvailablePriority,
llvm::StringRef profilePath,
bool forFunctionCompression, bool forDataCompression,
- bool verbose);
+ bool compressionSortStartupFunctions, bool verbose);
} // namespace lld::macho
diff --git a/lld/MachO/Config.h b/lld/MachO/Config.h
index 5beb0662ba7274..17259ee75aa4db 100644
--- a/lld/MachO/Config.h
+++ b/lld/MachO/Config.h
@@ -218,6 +218,7 @@ struct Configuration {
llvm::StringRef printSymbolOrder;
llvm::StringRef irpgoProfileSortProfilePath;
+ bool compressionSortStartupFunctions = false;
bool functionOrderForCompression = false;
bool dataOrderForCompression = false;
bool verboseBpSectionOrderer = false;
diff --git a/lld/MachO/Driver.cpp b/lld/MachO/Driver.cpp
index 6a1ff96ed65697..514ee1933f2d1a 100644
--- a/lld/MachO/Driver.cpp
+++ b/lld/MachO/Driver.cpp
@@ -1772,6 +1772,9 @@ bool link(ArrayRef<const char *> argsArr, llvm::raw_ostream &stdoutOS,
config->irpgoProfileSortProfilePath = arg->getValue();
IncompatWithCGSort(arg->getSpelling());
}
+ config->compressionSortStartupFunctions =
+ args.hasFlag(OPT_compression_sort_startup_functions,
+ OPT_no_compression_sort_startup_functions, false);
if (const Arg *arg = args.getLastArg(OPT_compression_sort)) {
StringRef compressionSortStr = arg->getValue();
if (compressionSortStr == "function") {
diff --git a/lld/MachO/Options.td b/lld/MachO/Options.td
index 9c9570cdbeb05c..4ac1e2080de5cc 100644
--- a/lld/MachO/Options.td
+++ b/lld/MachO/Options.td
@@ -131,6 +131,8 @@ def irpgo_profile_sort_eq: Joined<["--"], "irpgo-profile-sort=">,
Alias<!cast<Separate>(irpgo_profile_sort)>, MetaVarName<"<profile>">,
HelpText<"Read the IRPGO profile at <profile> to order sections to improve startup time">,
Group<grp_lld>;
+def compression_sort_startup_functions: Flag<["--"], "compression-sort-startup-functions">, HelpText<"TODO">, Group<grp_lld>;
+def no_compression_sort_startup_functions: Flag<["--"], "no-compression-sort-startup-functions">, HelpText<"TODO">, Group<grp_lld>;
def compression_sort: Joined<["--"], "compression-sort=">,
MetaVarName<"[none,function,data,both]">,
HelpText<"Order sections to improve compressed size">, Group<grp_lld>;
diff --git a/lld/MachO/SectionPriorities.cpp b/lld/MachO/SectionPriorities.cpp
index 69c301d8ff8a71..1e7fb5973b8086 100644
--- a/lld/MachO/SectionPriorities.cpp
+++ b/lld/MachO/SectionPriorities.cpp
@@ -359,6 +359,7 @@ macho::PriorityBuilder::buildInputSectionPriorities() {
sectionPriorities = runBalancedPartitioning(
highestAvailablePriority, config->irpgoProfileSortProfilePath,
config->functionOrderForCompression, config->dataOrderForCompression,
+ config->compressionSortStartupFunctions,
config->verboseBpSectionOrderer);
} else if (config->callGraphProfileSort) {
// Sort sections by the profile data provided by __LLVM,__cg_profile
diff --git a/lld/test/MachO/bp-section-orderer-stress.s b/lld/test/MachO/bp-section-orderer-stress.s
index fdc6a20e2655b9..986e2d8fd1069b 100644
--- a/lld/test/MachO/bp-section-orderer-stress.s
+++ b/lld/test/MachO/bp-section-orderer-stress.s
@@ -7,8 +7,8 @@
# RUN: llvm-mc -filetype=obj -triple=arm64-apple-darwin %t.s -o %t.o
# RUN: llvm-profdata merge %t.proftext -o %t.profdata
-# RUN: %lld -arch arm64 -lSystem -e _main --icf=all -o - %t.o --irpgo-profile-sort=%t.profdata --compression-sort=both | llvm-nm --numeric-sort --format=just-symbols - > %t.order1.txt
-# RUN: %lld -arch arm64 -lSystem -e _main --icf=all -o - %t.o --irpgo-profile-sort=%t.profdata --compression-sort=both | llvm-nm --numeric-sort --format=just-symbols - > %t.order2.txt
+# RUN: %lld -arch arm64 -lSystem -e _main --icf=all -o - %t.o --irpgo-profile-sort=%t.profdata --compression-sort-startup-functions --compression-sort=both | llvm-nm --numeric-sort --format=just-symbols - > %t.order1.txt
+# RUN: %lld -arch arm64 -lSystem -e _main --icf=all -o - %t.o --irpgo-profile-sort=%t.profdata --compression-sort-startup-functions --compression-sort=both | llvm-nm --numeric-sort --format=just-symbols - > %t.order2.txt
# RUN: diff %t.order1.txt %t.order2.txt
import random
|
ellishg
requested review from
thevinster,
MaskRay,
DataCorrupted and
kyulee-com
ellishg
force-pushed
the
bp-startup-compression
branch
from
a9d9531 to
e1f2853
Compare
|
This is a really interesting idea! It looks like you've added compression constraints (basically utility nodes based on content) on top of the existing startup performance constraints (utility nodes based on execution traces) for the startup symbols. You're trying to run BP to achieve both compressed size and startup performance at the same time. However, I'm wondering how much of an impact this will have depending on how many startup symbols there are compared to the total number of symbols. If there aren't many startup symbols, maybe shuffling them around won't really change performance or size much. But if there are a lot, this could potentially make startup slower while reducing the size more significantly. Could you share some info on how many startup symbols are involved in this experiment? Also, I'm a bit confused about what |
In these tests I ordered about 45K functions for startup and 2.1M functions for compression. However, I suspect the functions in the startup set are quite large because ordering them does impact compressed size. I did not measure the size of these sets in bytes.
I ignored the startup traces when ordering those startup symbols. The purpose was to show an upper bound for how bad page faults can get and how good compression can get. Basically, this shows that "Startup Compression" is halfway between ordering completely for startup and ordering completely for compression. diff --git a/lld/MachO/BPSectionOrderer.cpp b/lld/MachO/BPSectionOrderer.cpp
index 71758070ddc8..92032af67daa 100644
--- a/lld/MachO/BPSectionOrderer.cpp
+++ b/lld/MachO/BPSectionOrderer.cpp
@@ -286,6 +286,7 @@ DenseMap<const InputSection *, size_t> lld::macho::runBalancedPartitioning(
for (auto &[sectionIdx, compressionUns] : unsForStartupFunctionCompression) {
auto &uns = startupSectionIdxUNs[sectionIdx];
+ uns.clear();
uns.append(compressionUns);
llvm::sort(uns);
uns.erase(std::unique(uns.begin(), uns.end()), uns.end()); |
|
I built a large app and measured the number of page faults during startup on a local device for 5 trials. I built "Base" using
It seems that "Startup Compression" is not as performant as "Base", but not nearly as bad as "No PGO". Furthermore, "Startup Compression" has the best compression ratio of the three. I think this shows that this flag is useful and the user can decide if this tradeoff is worthwhile. |
ellishg
added a commit
that referenced
this pull request
Sep 5, 2024
Refactor some code in `BPSectionOrderer.cpp` in preparation for #107348. * Rename `constructNodesForCompression()` -> `getUnsForCompression()` and return a `SmallVector` directly rather than populating a vector alias * Pass `duplicateSectionIdxs` as a pointer to make it possible to skip finding (nearly) duplicate sections * Combine `duplicate{Function,Data}SectionIdxs` into one variable * Compute all `BPFunctionNode` vectors at the end (like `nodesForStartup`) There should be no functional change.
| | Page Fault Area | Compressed Size | | --------------------- | --------------- | --------------- | | Base | 1.045659e+09 | 25098246 | | Startup Compression | 1.430330e+09 | 24740192 | | Remove Startup Hashes | 2.272126e+09 | 24574929 |
ellishg
force-pushed
the
bp-startup-compression
branch
from
e1f2853 to
db36976
Compare
kyulee-com
approved these changes
Sep 5, 2024
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Before this, startup functions were only ordered to reduce the number of page faults. This patch introduces
--compression-sort-startup-functionswhich orders these function to also improve compressed size at the same time. Of course, this will regress some startup performance in favor of improved compressed size.I built a large binary without this flag (Base) and with this flag (Startup Compression). I also removed all startup hashes (Remove Startup Hashes) to see how ordering them for startup impacts the number of page faults. I recorded "Page Fault Curve Area" from the verbose logs and compressed the resulting binary using
zipfor "Compressed Size".Depends on #107347