[llvm] [BOLT] Refactor SplitFunctions for Function Reuse (PR #73078)
via llvm-commits
llvm-commits at lists.llvm.org
Wed Nov 22 09:08:58 PST 2023
https://github.com/ShatianWang updated https://github.com/llvm/llvm-project/pull/73078
>From 5ee5d6577a008249c50221e11b0a20a0315af9a5 Mon Sep 17 00:00:00 2001
From: Shatian Wang <shatian at meta.com>
Date: Thu, 2 Nov 2023 12:26:49 -0700
Subject: [PATCH 1/2] [BOLT] Extend calculateEmittedSize for Block Size
Calculation
This commit modifies BinaryContext::calculateEmittedSize to update the
BinaryBasicBlock::OutputAddressRange for each basic block in the input
BF. The modification is done in place, where BB.OutputAddressRange.second
less BB.OutputAddressRange.first now gives the emitted size of the basic
block.
---
bolt/include/bolt/Core/BinaryContext.h | 3 +++
bolt/lib/Core/BinaryContext.cpp | 35 +++++++++++++++++++++-----
2 files changed, 32 insertions(+), 6 deletions(-)
diff --git a/bolt/include/bolt/Core/BinaryContext.h b/bolt/include/bolt/Core/BinaryContext.h
index ad1bf2baaeb5b1e..17e55a673e8b489 100644
--- a/bolt/include/bolt/Core/BinaryContext.h
+++ b/bolt/include/bolt/Core/BinaryContext.h
@@ -1230,6 +1230,9 @@ class BinaryContext {
///
/// Return the pair where the first size is for the main part, and the second
/// size is for the cold one.
+ /// Modify BinaryBasicBlock::OutputAddressRange for each basic block in the
+ /// function in place so that BB.OutputAddressRange.second less
+ /// BB.OutputAddressRange.first gives the emitted size of BB.
std::pair<size_t, size_t> calculateEmittedSize(BinaryFunction &BF,
bool FixBranches = true);
diff --git a/bolt/lib/Core/BinaryContext.cpp b/bolt/lib/Core/BinaryContext.cpp
index 06b68765909d20e..d04f00efd27ce04 100644
--- a/bolt/lib/Core/BinaryContext.cpp
+++ b/bolt/lib/Core/BinaryContext.cpp
@@ -2331,14 +2331,37 @@ BinaryContext::calculateEmittedSize(BinaryFunction &BF, bool FixBranches) {
MCAsmLayout Layout(Assembler);
Assembler.layout(Layout);
+ // Obtain fragment sizes.
+ std::vector<uint64_t> FragmentSizes;
+ // Main fragment size.
const uint64_t HotSize =
Layout.getSymbolOffset(*EndLabel) - Layout.getSymbolOffset(*StartLabel);
- const uint64_t ColdSize =
- std::accumulate(SplitLabels.begin(), SplitLabels.end(), 0ULL,
- [&](const uint64_t Accu, const LabelRange &Labels) {
- return Accu + Layout.getSymbolOffset(*Labels.second) -
- Layout.getSymbolOffset(*Labels.first);
- });
+ FragmentSizes.push_back(HotSize);
+ // Split fragment sizes.
+ uint64_t ColdSize = 0;
+ for (const auto &Labels : SplitLabels) {
+ uint64_t Size = Layout.getSymbolOffset(*Labels.second) -
+ Layout.getSymbolOffset(*Labels.first);
+ FragmentSizes.push_back(Size);
+ ColdSize += Size;
+ }
+
+ // Populate new start and end offsets of each basic block.
+ BinaryBasicBlock *PrevBB = nullptr;
+ uint64_t FragmentIndex = 0;
+ for (FunctionFragment &FF : BF.getLayout().fragments()) {
+ for (BinaryBasicBlock *BB : FF) {
+ const uint64_t BBStartOffset = Layout.getSymbolOffset(*(BB->getLabel()));
+ BB->setOutputStartAddress(BBStartOffset);
+ if (PrevBB)
+ PrevBB->setOutputEndAddress(BBStartOffset);
+ PrevBB = BB;
+ }
+ if (PrevBB)
+ PrevBB->setOutputEndAddress(FragmentSizes[FragmentIndex]);
+ FragmentIndex++;
+ PrevBB = nullptr;
+ }
// Clean-up the effect of the code emission.
for (const MCSymbol &Symbol : Assembler.symbols()) {
>From 61de29546b253441f8a2eb9e2fea67718c342093 Mon Sep 17 00:00:00 2001
From: Shatian Wang <shatian at meta.com>
Date: Wed, 8 Nov 2023 08:20:11 -0800
Subject: [PATCH 2/2] [BOLT] Refactor SplitFunctions for Function Reuse
This commit updates SplitFunctions.h and SplitFunctions.cpp to enable
the reuse of createEHTrampolines, mergeEHTrampolines, hasFullProfile,
and allBlocksCold by a distinct function splitting pass (CDSplit).
---
bolt/include/bolt/Core/BinaryFunction.h | 14 ++++++++++
bolt/include/bolt/Passes/SplitFunctions.h | 32 +++++++++++------------
bolt/lib/Passes/SplitFunctions.cpp | 18 +++----------
3 files changed, 33 insertions(+), 31 deletions(-)
diff --git a/bolt/include/bolt/Core/BinaryFunction.h b/bolt/include/bolt/Core/BinaryFunction.h
index 72c360ca0c2db66..3723cccc50f040c 100644
--- a/bolt/include/bolt/Core/BinaryFunction.h
+++ b/bolt/include/bolt/Core/BinaryFunction.h
@@ -1272,6 +1272,20 @@ class BinaryFunction {
/// otherwise processed.
bool isPseudo() const { return IsPseudo; }
+ /// Return true if every block in the function has a valid execution count.
+ bool hasFullProfile() const {
+ return llvm::all_of(blocks(), [](const BinaryBasicBlock &BB) {
+ return BB.getExecutionCount() != BinaryBasicBlock::COUNT_NO_PROFILE;
+ });
+ }
+
+ /// Return true if every block in the function has a zero execution count.
+ bool allBlocksCold() const {
+ return llvm::all_of(blocks(), [](const BinaryBasicBlock &BB) {
+ return BB.getExecutionCount() == 0;
+ });
+ }
+
/// Return true if the function contains explicit or implicit indirect branch
/// to its split fragments, e.g., split jump table, landing pad in split
/// fragment.
diff --git a/bolt/include/bolt/Passes/SplitFunctions.h b/bolt/include/bolt/Passes/SplitFunctions.h
index 4058f3317dfbdbb..91b6d5518eaab26 100644
--- a/bolt/include/bolt/Passes/SplitFunctions.h
+++ b/bolt/include/bolt/Passes/SplitFunctions.h
@@ -50,6 +50,19 @@ class SplitFunctions : public BinaryFunctionPass {
/// Split function body into fragments.
void splitFunction(BinaryFunction &Function, SplitStrategy &Strategy);
+ std::atomic<uint64_t> SplitBytesHot{0ull};
+ std::atomic<uint64_t> SplitBytesCold{0ull};
+
+public:
+ explicit SplitFunctions(const cl::opt<bool> &PrintPass)
+ : BinaryFunctionPass(PrintPass) {}
+
+ bool shouldOptimize(const BinaryFunction &BF) const override;
+
+ const char *getName() const override { return "split-functions"; }
+
+ void runOnFunctions(BinaryContext &BC) override;
+
struct TrampolineKey {
FragmentNum SourceFN = FragmentNum::main();
const MCSymbol *Target = nullptr;
@@ -81,27 +94,14 @@ class SplitFunctions : public BinaryFunctionPass {
/// corresponding thrower block. The trampoline landing pad, when created,
/// will redirect the execution to the real landing pad in a different
/// fragment.
- TrampolineSetType createEHTrampolines(BinaryFunction &Function) const;
+ static TrampolineSetType createEHTrampolines(BinaryFunction &Function);
/// Merge trampolines into \p Layout without trampolines. The merge will place
/// a trampoline immediately before its destination. Used to revert the effect
/// of trampolines after createEHTrampolines().
- BasicBlockOrderType
+ static BasicBlockOrderType
mergeEHTrampolines(BinaryFunction &BF, BasicBlockOrderType &Layout,
- const TrampolineSetType &Trampolines) const;
-
- std::atomic<uint64_t> SplitBytesHot{0ull};
- std::atomic<uint64_t> SplitBytesCold{0ull};
-
-public:
- explicit SplitFunctions(const cl::opt<bool> &PrintPass)
- : BinaryFunctionPass(PrintPass) {}
-
- bool shouldOptimize(const BinaryFunction &BF) const override;
-
- const char *getName() const override { return "split-functions"; }
-
- void runOnFunctions(BinaryContext &BC) override;
+ const TrampolineSetType &Trampolines);
};
} // namespace bolt
diff --git a/bolt/lib/Passes/SplitFunctions.cpp b/bolt/lib/Passes/SplitFunctions.cpp
index 34973cecdf49161..223f8d17367845d 100644
--- a/bolt/lib/Passes/SplitFunctions.cpp
+++ b/bolt/lib/Passes/SplitFunctions.cpp
@@ -109,21 +109,9 @@ static cl::opt<SplitFunctionsStrategy> SplitStrategy(
} // namespace opts
namespace {
-bool hasFullProfile(const BinaryFunction &BF) {
- return llvm::all_of(BF.blocks(), [](const BinaryBasicBlock &BB) {
- return BB.getExecutionCount() != BinaryBasicBlock::COUNT_NO_PROFILE;
- });
-}
-
-bool allBlocksCold(const BinaryFunction &BF) {
- return llvm::all_of(BF.blocks(), [](const BinaryBasicBlock &BB) {
- return BB.getExecutionCount() == 0;
- });
-}
-
struct SplitProfile2 final : public SplitStrategy {
bool canSplit(const BinaryFunction &BF) override {
- return BF.hasValidProfile() && hasFullProfile(BF) && !allBlocksCold(BF);
+ return BF.hasValidProfile() && BF.hasFullProfile() && !BF.allBlocksCold();
}
bool keepEmpty() override { return false; }
@@ -434,7 +422,7 @@ void SplitFunctions::splitFunction(BinaryFunction &BF, SplitStrategy &S) {
}
SplitFunctions::TrampolineSetType
-SplitFunctions::createEHTrampolines(BinaryFunction &BF) const {
+SplitFunctions::createEHTrampolines(BinaryFunction &BF) {
const auto &MIB = BF.getBinaryContext().MIB;
// Map real landing pads to the corresponding trampolines.
@@ -501,7 +489,7 @@ SplitFunctions::createEHTrampolines(BinaryFunction &BF) const {
SplitFunctions::BasicBlockOrderType SplitFunctions::mergeEHTrampolines(
BinaryFunction &BF, SplitFunctions::BasicBlockOrderType &Layout,
- const SplitFunctions::TrampolineSetType &Trampolines) const {
+ const SplitFunctions::TrampolineSetType &Trampolines) {
DenseMap<const MCSymbol *, SmallVector<const MCSymbol *, 0>>
IncomingTrampolines;
for (const auto &Entry : Trampolines) {
More information about the llvm-commits
mailing list