Index: docs/BitSets.rst =================================================================== --- docs/BitSets.rst +++ docs/BitSets.rst @@ -11,16 +11,27 @@ To use the mechanism, a client creates a global metadata node named ``llvm.bitsets``. Each element is a metadata node with three elements: the first is a metadata string containing an identifier for the bitset, -the second is a global variable and the third is a byte offset into the -global variable. +the second is either a global variable or a function and the third is a +byte offset into the global (generally zero for functions). Each bitset must +exclusively contain either global variables or functions. This will cause a link-time optimization pass to generate bitsets from the -memory addresses referenced from the elements of the bitset metadata. The pass -will lay out the referenced globals consecutively, so their definitions must -be available at LTO time. The `GlobalLayoutBuilder`_ class is responsible for -laying out the globals efficiently to minimize the sizes of the underlying -bitsets. An intrinsic, :ref:`llvm.bitset.test `, generates code -to test whether a given pointer is a member of a bitset. +memory addresses referenced from the elements of the bitset metadata. The +pass will lay out referenced global variables consecutively, so their +definitions must be available at LTO time, but functions in bit sets may +refer to external entities. + +Note that if an externally defined function is a member of a bitset, there +is no guarantee that its identity within the module will be the same as its +identity outside of the module, as the former will be the jump table entry +if a jump table is necessary. However, the identity of functions defined +within the module will be consistent within and outside of the module, +as the jump table entry will always act as the function entry point. + +The `GlobalLayoutBuilder`_ class is responsible for laying out the globals +efficiently to minimize the sizes of the underlying bitsets. An intrinsic, +:ref:`llvm.bitset.test `, generates code to test whether a +given pointer is a member of a bitset. :Example: Index: include/llvm/Transforms/IPO/LowerBitSets.h =================================================================== --- include/llvm/Transforms/IPO/LowerBitSets.h +++ include/llvm/Transforms/IPO/LowerBitSets.h @@ -26,7 +26,7 @@ namespace llvm { class DataLayout; -class GlobalVariable; +class GlobalObject; class Value; class raw_ostream; @@ -56,7 +56,7 @@ bool containsGlobalOffset(uint64_t Offset) const; bool containsValue(const DataLayout &DL, - const DenseMap &GlobalLayout, + const DenseMap &GlobalLayout, Value *V, uint64_t COffset = 0) const; void print(raw_ostream &OS) const; Index: lib/Transforms/IPO/LowerBitSets.cpp =================================================================== --- lib/Transforms/IPO/LowerBitSets.cpp +++ lib/Transforms/IPO/LowerBitSets.cpp @@ -19,6 +19,8 @@ #include "llvm/ADT/Triple.h" #include "llvm/IR/Constant.h" #include "llvm/IR/Constants.h" +#include "llvm/IR/Function.h" +#include "llvm/IR/GlobalObject.h" #include "llvm/IR/GlobalVariable.h" #include "llvm/IR/IRBuilder.h" #include "llvm/IR/Instructions.h" @@ -61,9 +63,9 @@ bool BitSetInfo::containsValue( const DataLayout &DL, - const DenseMap &GlobalLayout, Value *V, + const DenseMap &GlobalLayout, Value *V, uint64_t COffset) const { - if (auto GV = dyn_cast(V)) { + if (auto GV = dyn_cast(V)) { auto I = GlobalLayout.find(GV); if (I == GlobalLayout.end()) return false; @@ -211,6 +213,7 @@ Module *M; bool LinkerSubsectionsViaSymbols; + Triple::ArchType Arch; IntegerType *Int1Ty; IntegerType *Int8Ty; IntegerType *Int32Ty; @@ -228,17 +231,25 @@ BitSetInfo buildBitSet(MDString *BitSet, - const DenseMap &GlobalLayout); + const DenseMap &GlobalLayout); ByteArrayInfo *createByteArray(BitSetInfo &BSI); void allocateByteArrays(); Value *createBitSetTest(IRBuilder<> &B, BitSetInfo &BSI, ByteArrayInfo *&BAI, Value *BitOffset); + void lowerBitSetCalls(ArrayRef BitSets, + Constant *CombinedGlobalIntAddr, + const DenseMap &GlobalLayout); Value * lowerBitSetCall(CallInst *CI, BitSetInfo &BSI, ByteArrayInfo *&BAI, - GlobalVariable *CombinedGlobal, - const DenseMap &GlobalLayout); - void buildBitSetsFromGlobals(const std::vector &BitSets, - const std::vector &Globals); + Constant *CombinedGlobal, + const DenseMap &GlobalLayout); + void buildBitSetsFromGlobalVariables(ArrayRef BitSets, + ArrayRef Globals); + unsigned getJumpTableEntrySize(); + Constant *createJumpTableEntry(Function *Src, Function *Dest, + unsigned Distance); + void buildBitSetsFromFunctions(ArrayRef BitSets, + ArrayRef Globals); bool buildBitSets(); bool eraseBitSetMetadata(); @@ -262,6 +273,7 @@ Triple TargetTriple(M->getTargetTriple()); LinkerSubsectionsViaSymbols = TargetTriple.isMacOSX(); + Arch = TargetTriple.getArch(); Int1Ty = Type::getInt1Ty(M->getContext()); Int8Ty = Type::getInt8Ty(M->getContext()); @@ -281,7 +293,7 @@ /// GlobalLayout. BitSetInfo LowerBitSets::buildBitSet( MDString *BitSet, - const DenseMap &GlobalLayout) { + const DenseMap &GlobalLayout) { BitSetBuilder BSB; // Compute the byte offset of each element of this bitset. @@ -289,8 +301,11 @@ for (MDNode *Op : BitSetNM->operands()) { if (Op->getOperand(0) != BitSet || !Op->getOperand(1)) continue; - auto OpGlobal = dyn_cast( - cast(Op->getOperand(1))->getValue()); + Constant *OpConst = + cast(Op->getOperand(1))->getValue(); + if (auto GA = dyn_cast(OpConst)) + OpConst = GA->getAliasee(); + auto OpGlobal = dyn_cast(OpConst); if (!OpGlobal) continue; uint64_t Offset = @@ -439,17 +454,16 @@ /// replace the call with. Value *LowerBitSets::lowerBitSetCall( CallInst *CI, BitSetInfo &BSI, ByteArrayInfo *&BAI, - GlobalVariable *CombinedGlobal, - const DenseMap &GlobalLayout) { + Constant *CombinedGlobalIntAddr, + const DenseMap &GlobalLayout) { Value *Ptr = CI->getArgOperand(0); const DataLayout &DL = M->getDataLayout(); if (BSI.containsValue(DL, GlobalLayout, Ptr)) - return ConstantInt::getTrue(CombinedGlobal->getParent()->getContext()); + return ConstantInt::getTrue(M->getContext()); - Constant *GlobalAsInt = ConstantExpr::getPtrToInt(CombinedGlobal, IntPtrTy); Constant *OffsetedGlobalAsInt = ConstantExpr::getAdd( - GlobalAsInt, ConstantInt::get(IntPtrTy, BSI.ByteOffset)); + CombinedGlobalIntAddr, ConstantInt::get(IntPtrTy, BSI.ByteOffset)); BasicBlock *InitialBB = CI->getParent(); @@ -508,9 +522,8 @@ /// Given a disjoint set of bitsets and globals, layout the globals, build the /// bit sets and lower the llvm.bitset.test calls. -void LowerBitSets::buildBitSetsFromGlobals( - const std::vector &BitSets, - const std::vector &Globals) { +void LowerBitSets::buildBitSetsFromGlobalVariables( + ArrayRef BitSets, ArrayRef Globals) { // Build a new global with the combined contents of the referenced globals. std::vector GlobalInits; const DataLayout &DL = M->getDataLayout(); @@ -541,11 +554,74 @@ DL.getStructLayout(cast(NewInit->getType())); // Compute the offsets of the original globals within the new global. - DenseMap GlobalLayout; + DenseMap GlobalLayout; for (unsigned I = 0; I != Globals.size(); ++I) // Multiply by 2 to account for padding elements. GlobalLayout[Globals[I]] = CombinedGlobalLayout->getElementOffset(I * 2); + Constant *GlobalAsInt = ConstantExpr::getPtrToInt(CombinedGlobal, IntPtrTy); + lowerBitSetCalls(BitSets, GlobalAsInt, GlobalLayout); + + // Build aliases pointing to offsets into the combined global for each + // global from which we built the combined global, and replace references + // to the original globals with references to the aliases. + for (unsigned I = 0; I != Globals.size(); ++I) { + // Multiply by 2 to account for padding elements. + Constant *CombinedGlobalIdxs[] = {ConstantInt::get(Int32Ty, 0), + ConstantInt::get(Int32Ty, I * 2)}; + Constant *CombinedGlobalElemPtr = ConstantExpr::getGetElementPtr( + NewInit->getType(), CombinedGlobal, CombinedGlobalIdxs); + if (LinkerSubsectionsViaSymbols) { + Globals[I]->replaceAllUsesWith(CombinedGlobalElemPtr); + } else { + GlobalAlias *GAlias = + GlobalAlias::create(Globals[I]->getType(), Globals[I]->getLinkage(), + "", CombinedGlobalElemPtr, M); + GAlias->takeName(Globals[I]); + Globals[I]->replaceAllUsesWith(GAlias); + } + Globals[I]->eraseFromParent(); + } +} + +unsigned LowerBitSets::getJumpTableEntrySize() { + if (Arch != Triple::x86 && Arch != Triple::x86_64) + report_fatal_error("Unsupported architecture for jump tables"); + + return 8; +} + +// Create a constant representing a jump table entry for the target. This +// consists of an instruction sequence containing a relative branch to Dest. The +// constant will be laid out at address Src-(Len*Distance) where Len is the +// target-specific jump table entry size. +Constant *LowerBitSets::createJumpTableEntry(Function *Src, Function *Dest, + unsigned Distance) { + if (Arch != Triple::x86 && Arch != Triple::x86_64) + report_fatal_error("Unsupported architecture for jump tables"); + + ConstantInt *Jmp = ConstantInt::get(Int8Ty, 0xe9); + + // Build a constant representing the displacement between the constant's + // address and Dest. This will resolve to a PC32 relocation referring to Dest. + Constant *DestInt = ConstantExpr::getPtrToInt(Dest, Int64Ty); + Constant *SrcInt = ConstantExpr::getPtrToInt(Src, Int64Ty); + Constant *Disp = ConstantExpr::getSub(DestInt, SrcInt); + ConstantInt *DispOffset = ConstantInt::get(Int64Ty, Distance * 8 - 5); + Constant *OffsetedDisp = ConstantExpr::getAdd(Disp, DispOffset); + OffsetedDisp = ConstantExpr::getTrunc(OffsetedDisp, Int32Ty); + + ConstantInt *Int3 = ConstantInt::get(Int8Ty, 0xcc); + + Constant *Fields[] = { + Jmp, OffsetedDisp, Int3, Int3, Int3, + }; + return ConstantStruct::getAnon(Fields, /*Packed=*/true); +} + +void LowerBitSets::lowerBitSetCalls( + ArrayRef BitSets, Constant *CombinedGlobalIntAddr, + const DenseMap &GlobalLayout) { // For each bitset in this disjoint set... for (MDString *BS : BitSets) { // Build the bitset. @@ -560,22 +636,66 @@ // Lower each call to llvm.bitset.test for this bitset. for (CallInst *CI : BitSetTestCallSites[BS]) { ++NumBitSetCallsLowered; - Value *Lowered = lowerBitSetCall(CI, BSI, BAI, CombinedGlobal, GlobalLayout); + Value *Lowered = + lowerBitSetCall(CI, BSI, BAI, CombinedGlobalIntAddr, GlobalLayout); CI->replaceAllUsesWith(Lowered); CI->eraseFromParent(); } } +} - // Build aliases pointing to offsets into the combined global for each - // global from which we built the combined global, and replace references - // to the original globals with references to the aliases. - for (unsigned I = 0; I != Globals.size(); ++I) { - // Multiply by 2 to account for padding elements. - Constant *CombinedGlobalIdxs[] = {ConstantInt::get(Int32Ty, 0), - ConstantInt::get(Int32Ty, I * 2)}; - Constant *CombinedGlobalElemPtr = ConstantExpr::getGetElementPtr( - NewInit->getType(), CombinedGlobal, CombinedGlobalIdxs); - if (LinkerSubsectionsViaSymbols) { +void LowerBitSets::buildBitSetsFromFunctions( + ArrayRef BitSets, ArrayRef Globals) { + // This simple layout is based on the regular layout of jump tables. + DenseMap GlobalLayout; + unsigned EntrySize = getJumpTableEntrySize(); + for (unsigned I = 0; I != Globals.size(); ++I) + GlobalLayout[Globals[I]] = I * EntrySize; + + std::vector NewGlobals; + // If our last function is an external reference, we won't have a function to + // hang our jump table onto. To handle this special case, we create a dummy + // function and use that as the last function. + if (Globals.back()->isDeclarationForLinker()) { + NewGlobals.insert(NewGlobals.end(), Globals.begin(), Globals.end()); + Function *NewFn = Function::Create( + FunctionType::get(Type::getVoidTy(M->getContext()), /*isVarArg=*/false), + Function::PrivateLinkage, "", M); + BasicBlock *BB = BasicBlock::Create(M->getContext(), "", NewFn); + IRBuilder<> IRB(BB); + IRB.CreateUnreachable(); + NewGlobals.push_back(NewFn); + Globals = NewGlobals; + } + + Function *LastFn = Globals.back(); + + // Create the alias for the last function first, so that any references to + // LastFn created below use the real LastFn rather than being RAUW'd. + if (!LinkerSubsectionsViaSymbols) { + GlobalAlias *GAlias = + GlobalAlias::create(LastFn->getType(), LastFn->getLinkage(), "", + nullptr, M); + GAlias->takeName(LastFn); + LastFn->replaceAllUsesWith(GAlias); + GAlias->setAliasee(LastFn); + } + LastFn->setLinkage(GlobalValue::PrivateLinkage); + + Constant *JumpTableBegin = ConstantExpr::getSub( + ConstantExpr::getPtrToInt(LastFn, Int64Ty), + ConstantInt::get(Int64Ty, EntrySize * (Globals.size() - 1))); + lowerBitSetCalls(BitSets, JumpTableBegin, GlobalLayout); + + // Build aliases pointing to offsets into the jump table, and replace + // references to the original functions with references to the aliases. + for (unsigned I = 0; I != Globals.size()-1; ++I) { + Constant *CombinedGlobalElemPtr = ConstantExpr::getIntToPtr( + ConstantExpr::getSub( + ConstantExpr::getPtrToInt(LastFn, Int64Ty), + ConstantInt::get(Int64Ty, EntrySize * (Globals.size() - 1 - I))), + Globals[I]->getType()); + if (LinkerSubsectionsViaSymbols || Globals[I]->isDeclarationForLinker()) { Globals[I]->replaceAllUsesWith(CombinedGlobalElemPtr); } else { GlobalAlias *GAlias = @@ -584,8 +704,21 @@ GAlias->takeName(Globals[I]); Globals[I]->replaceAllUsesWith(GAlias); } - Globals[I]->eraseFromParent(); + if (!Globals[I]->isDeclarationForLinker()) + Globals[I]->setLinkage(GlobalValue::PrivateLinkage); } + + // Modify the prefix data of the last function in Globals to contain jump table + // entries for every other element. + std::vector JumpTableEntries; + for (unsigned I = 0; I != Globals.size()-1; ++I) { + JumpTableEntries.push_back(createJumpTableEntry(LastFn, Globals[I], + Globals.size()-1-I)); + } + if (!JumpTableEntries.empty()) + LastFn->setPrefixData(ConstantArray::get( + ArrayType::get(JumpTableEntries[0]->getType(), JumpTableEntries.size()), + JumpTableEntries)); } /// Lower all bit sets in this module. @@ -598,7 +731,7 @@ // Equivalence class set containing bitsets and the globals they reference. // This is used to partition the set of bitsets in the module into disjoint // sets. - typedef EquivalenceClasses> + typedef EquivalenceClasses> GlobalClassesTy; GlobalClassesTy GlobalClasses; @@ -643,9 +776,11 @@ auto OpConstMD = dyn_cast(Op->getOperand(1)); if (!OpConstMD) report_fatal_error("Bit set element must be a constant"); - auto OpGlobal = dyn_cast(OpConstMD->getValue()); + auto OpGlobal = dyn_cast(OpConstMD->getValue()); if (!OpGlobal) continue; + if (isa(OpGlobal) && OpGlobal->isDeclarationForLinker()) + report_fatal_error("Bit set global var element must be a definition"); auto OffsetConstMD = dyn_cast(Op->getOperand(2)); if (!OffsetConstMD) @@ -673,17 +808,17 @@ // Build the list of bitsets and referenced globals in this disjoint set. std::vector BitSets; - std::vector Globals; + std::vector Globals; llvm::DenseMap BitSetIndices; - llvm::DenseMap GlobalIndices; + llvm::DenseMap GlobalIndices; for (GlobalClassesTy::member_iterator MI = GlobalClasses.member_begin(I); MI != GlobalClasses.member_end(); ++MI) { if ((*MI).is()) { BitSetIndices[MI->get()] = BitSets.size(); BitSets.push_back(MI->get()); } else { - GlobalIndices[MI->get()] = Globals.size(); - Globals.push_back(MI->get()); + GlobalIndices[MI->get()] = Globals.size(); + Globals.push_back(MI->get()); } } @@ -699,7 +834,7 @@ if (I == BitSetIndices.end()) continue; - auto OpGlobal = dyn_cast( + auto OpGlobal = dyn_cast( cast(Op->getOperand(1))->getValue()); if (!OpGlobal) continue; @@ -723,7 +858,7 @@ GLB.addFragment(MemSet); // Build a vector of globals with the computed layout. - std::vector OrderedGlobals(Globals.size()); + std::vector OrderedGlobals(Globals.size()); auto OGI = OrderedGlobals.begin(); for (auto &&F : GLB.Fragments) for (auto &&Offset : F) @@ -735,7 +870,19 @@ }); // Build the bitsets from this disjoint set. - buildBitSetsFromGlobals(BitSets, OrderedGlobals); + if (OrderedGlobals.empty() || isa(OrderedGlobals[0])) + buildBitSetsFromGlobalVariables( + BitSets, + ArrayRef( + reinterpret_cast(OrderedGlobals.data()), + reinterpret_cast(OrderedGlobals.data() + + OrderedGlobals.size()))); + else + buildBitSetsFromFunctions( + BitSets, ArrayRef( + reinterpret_cast(OrderedGlobals.data()), + reinterpret_cast(OrderedGlobals.data() + + OrderedGlobals.size()))); } allocateByteArrays(); Index: test/Transforms/LowerBitSets/function-ext.ll =================================================================== --- /dev/null +++ test/Transforms/LowerBitSets/function-ext.ll @@ -0,0 +1,20 @@ +; RUN: opt -S -lowerbitsets < %s | FileCheck %s + +target triple = "x86_64-unknown-linux-gnu" + +declare void @foo() + +define i1 @bar(i8* %ptr) { + ; CHECK: icmp eq i64 {{.*}}, sub (i64 ptrtoint (void ()* @[[DUMMYNAME:.*]] to i64), i64 8) + %p = call i1 @llvm.bitset.test(i8* %ptr, metadata !"void") + ret i1 %p +} + +; CHECK: define private void @[[DUMMYNAME]]() prefix [1 x <{ i8, i32, i8, i8, i8 }>] [<{ i8, i32, i8, i8, i8 }> <{ i8 -23, i32 trunc (i64 add (i64 sub (i64 ptrtoint (void ()* @foo to i64), i64 ptrtoint (void ()* @[[DUMMYNAME]] to i64)), i64 3) to i32), i8 -52, i8 -52, i8 -52 }>] +; CHECK-NEXT: unreachable + +declare i1 @llvm.bitset.test(i8* %ptr, metadata %bitset) nounwind readnone + +!0 = !{!"void", void ()* @foo, i64 0} + +!llvm.bitsets = !{!0} Index: test/Transforms/LowerBitSets/function.ll =================================================================== --- /dev/null +++ test/Transforms/LowerBitSets/function.ll @@ -0,0 +1,30 @@ +; RUN: opt -S -lowerbitsets < %s | FileCheck %s + +target triple = "x86_64-unknown-linux-gnu" +target datalayout = "e-p:64:64" + +; CHECK: @g = alias void ()* @[[GNAME:.*]] +; CHECK: @f = alias inttoptr (i64 sub (i64 ptrtoint (void ()* @[[GNAME]] to i64), i64 8) to void ()*) + +; CHECK: define private void @[[FNAME:.*]]() { +define void @f() { + ret void +} + +; CHECK: define private void @[[GNAME]]() prefix [1 x <{ i8, i32, i8, i8, i8 }>] [<{ i8, i32, i8, i8, i8 }> <{ i8 -23, i32 trunc (i64 add (i64 sub (i64 ptrtoint (void ()* @[[FNAME]] to i64), i64 ptrtoint (void ()* @[[GNAME]] to i64)), i64 3) to i32), i8 -52, i8 -52, i8 -52 }>] { +define void @g() { + ret void +} + +!0 = !{!"bitset1", void ()* @f, i32 0} +!1 = !{!"bitset1", void ()* @g, i32 0} + +!llvm.bitsets = !{ !0, !1 } + +declare i1 @llvm.bitset.test(i8* %ptr, metadata %bitset) nounwind readnone + +define i1 @foo(i8* %p) { + ; CHECK: sub i64 %1, sub (i64 ptrtoint (void ()* @[[GNAME]] to i64), i64 8) + %x = call i1 @llvm.bitset.test(i8* %p, metadata !"bitset1") + ret i1 %x +} Index: test/Transforms/LowerBitSets/function1.ll =================================================================== --- /dev/null +++ test/Transforms/LowerBitSets/function1.ll @@ -0,0 +1,23 @@ +; RUN: opt -S -lowerbitsets < %s | FileCheck %s + +target triple = "x86_64-unknown-linux-gnu" +target datalayout = "e-p:64:64" + +; CHECK: @f = alias void ()* @[[FNAME:.*]] + +; CHECK: define private void @[[FNAME]]() { +define void @f() { + ret void +} + +!0 = !{!"bitset1", void ()* @f, i32 0} + +!llvm.bitsets = !{ !0 } + +declare i1 @llvm.bitset.test(i8* %ptr, metadata %bitset) nounwind readnone + +define i1 @foo(i8* %p) { + ; CHECK: icmp eq i64 %{{.*}}, ptrtoint (void ()* @[[FNAME]] to i64) + %x = call i1 @llvm.bitset.test(i8* %p, metadata !"bitset1") + ret i1 %x +}