[CodeGen] Fix AtomicExpand invalidation issue caused by r247429.
[oota-llvm.git] / lib / CodeGen / AtomicExpandPass.cpp
index ed4924e1ba075112dfcc9402ce38323470680c2e..73102ccfece1006a23f2708a5f14e403b3380c45 100644 (file)
@@ -8,13 +8,16 @@
 //===----------------------------------------------------------------------===//
 //
 // This file contains a pass (at IR level) to replace atomic instructions with
-// appropriate (intrinsic-based) ldrex/strex loops.
+// either (intrinsic-based) load-linked/store-conditional loops or
+// AtomicCmpXchg.
 //
 //===----------------------------------------------------------------------===//
 
+#include "llvm/CodeGen/AtomicExpandUtils.h"
 #include "llvm/CodeGen/Passes.h"
 #include "llvm/IR/Function.h"
 #include "llvm/IR/IRBuilder.h"
+#include "llvm/IR/InstIterator.h"
 #include "llvm/IR/Instructions.h"
 #include "llvm/IR/Intrinsics.h"
 #include "llvm/IR/Module.h"
@@ -30,20 +33,28 @@ using namespace llvm;
 namespace {
   class AtomicExpand: public FunctionPass {
     const TargetMachine *TM;
+    const TargetLowering *TLI;
   public:
     static char ID; // Pass identification, replacement for typeid
     explicit AtomicExpand(const TargetMachine *TM = nullptr)
-      : FunctionPass(ID), TM(TM) {
+      : FunctionPass(ID), TM(TM), TLI(nullptr) {
       initializeAtomicExpandPass(*PassRegistry::getPassRegistry());
     }
 
     bool runOnFunction(Function &F) override;
-    bool expandAtomicInsts(Function &F);
 
-    bool expandAtomicLoad(LoadInst *LI);
-    bool expandAtomicStore(StoreInst *LI);
-    bool expandAtomicRMW(AtomicRMWInst *AI);
+  private:
+    bool bracketInstWithFences(Instruction *I, AtomicOrdering Order,
+                               bool IsStore, bool IsLoad);
+    bool tryExpandAtomicLoad(LoadInst *LI);
+    bool expandAtomicLoadToLL(LoadInst *LI);
+    bool expandAtomicLoadToCmpXchg(LoadInst *LI);
+    bool expandAtomicStore(StoreInst *SI);
+    bool tryExpandAtomicRMW(AtomicRMWInst *AI);
+    bool expandAtomicRMWToLLSC(AtomicRMWInst *AI);
     bool expandAtomicCmpXchg(AtomicCmpXchgInst *CI);
+    bool isIdempotentRMW(AtomicRMWInst *AI);
+    bool simplifyIdempotentRMW(AtomicRMWInst *AI);
   };
 }
 
@@ -58,66 +69,129 @@ FunctionPass *llvm::createAtomicExpandPass(const TargetMachine *TM) {
 }
 
 bool AtomicExpand::runOnFunction(Function &F) {
-  if (!TM || !TM->getSubtargetImpl()->enableAtomicExpand())
+  if (!TM || !TM->getSubtargetImpl(F)->enableAtomicExpand())
     return false;
+  TLI = TM->getSubtargetImpl(F)->getTargetLowering();
 
   SmallVector<Instruction *, 1> AtomicInsts;
 
   // Changing control-flow while iterating through it is a bad idea, so gather a
   // list of all atomic instructions before we start.
-  for (BasicBlock &BB : F)
-    for (Instruction &Inst : BB) {
-      if (isa<AtomicRMWInst>(&Inst) || isa<AtomicCmpXchgInst>(&Inst) ||
-          (isa<LoadInst>(&Inst) && cast<LoadInst>(&Inst)->isAtomic()) ||
-          (isa<StoreInst>(&Inst) && cast<StoreInst>(&Inst)->isAtomic()))
-        AtomicInsts.push_back(&Inst);
-    }
+  for (inst_iterator I = inst_begin(F), E = inst_end(F); I != E; ++I) {
+    if (I->isAtomic())
+      AtomicInsts.push_back(&*I);
+  }
 
   bool MadeChange = false;
-  for (Instruction *Inst : AtomicInsts) {
-    if (!TM->getSubtargetImpl()->getTargetLowering()->shouldExpandAtomicInIR(
-            Inst))
-      continue;
+  for (auto I : AtomicInsts) {
+    auto LI = dyn_cast<LoadInst>(I);
+    auto SI = dyn_cast<StoreInst>(I);
+    auto RMWI = dyn_cast<AtomicRMWInst>(I);
+    auto CASI = dyn_cast<AtomicCmpXchgInst>(I);
+    assert((LI || SI || RMWI || CASI || isa<FenceInst>(I)) &&
+           "Unknown atomic instruction");
+
+    auto FenceOrdering = Monotonic;
+    bool IsStore, IsLoad;
+    if (TLI->getInsertFencesForAtomic()) {
+      if (LI && isAtLeastAcquire(LI->getOrdering())) {
+        FenceOrdering = LI->getOrdering();
+        LI->setOrdering(Monotonic);
+        IsStore = false;
+        IsLoad = true;
+      } else if (SI && isAtLeastRelease(SI->getOrdering())) {
+        FenceOrdering = SI->getOrdering();
+        SI->setOrdering(Monotonic);
+        IsStore = true;
+        IsLoad = false;
+      } else if (RMWI && (isAtLeastRelease(RMWI->getOrdering()) ||
+                          isAtLeastAcquire(RMWI->getOrdering()))) {
+        FenceOrdering = RMWI->getOrdering();
+        RMWI->setOrdering(Monotonic);
+        IsStore = IsLoad = true;
+      } else if (CASI && !TLI->shouldExpandAtomicCmpXchgInIR(CASI) &&
+                 (isAtLeastRelease(CASI->getSuccessOrdering()) ||
+                  isAtLeastAcquire(CASI->getSuccessOrdering()))) {
+        // If a compare and swap is lowered to LL/SC, we can do smarter fence
+        // insertion, with a stronger one on the success path than on the
+        // failure path. As a result, fence insertion is directly done by
+        // expandAtomicCmpXchg in that case.
+        FenceOrdering = CASI->getSuccessOrdering();
+        CASI->setSuccessOrdering(Monotonic);
+        CASI->setFailureOrdering(Monotonic);
+        IsStore = IsLoad = true;
+      }
+
+      if (FenceOrdering != Monotonic) {
+        MadeChange |= bracketInstWithFences(I, FenceOrdering, IsStore, IsLoad);
+      }
+    }
 
-    if (AtomicRMWInst *AI = dyn_cast<AtomicRMWInst>(Inst))
-      MadeChange |= expandAtomicRMW(AI);
-    else if (AtomicCmpXchgInst *CI = dyn_cast<AtomicCmpXchgInst>(Inst))
-      MadeChange |= expandAtomicCmpXchg(CI);
-    else if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
-      MadeChange |= expandAtomicLoad(LI);
-    else if (StoreInst *SI = dyn_cast<StoreInst>(Inst))
+    if (LI) {
+      MadeChange |= tryExpandAtomicLoad(LI);
+    } else if (SI && TLI->shouldExpandAtomicStoreInIR(SI)) {
       MadeChange |= expandAtomicStore(SI);
-    else
-      llvm_unreachable("Unknown atomic instruction");
+    } else if (RMWI) {
+      // There are two different ways of expanding RMW instructions:
+      // - into a load if it is idempotent
+      // - into a Cmpxchg/LL-SC loop otherwise
+      // we try them in that order.
+
+      if (isIdempotentRMW(RMWI) && simplifyIdempotentRMW(RMWI)) {
+        MadeChange = true;
+      } else {
+        MadeChange |= tryExpandAtomicRMW(RMWI);
+      }
+    } else if (CASI && TLI->shouldExpandAtomicCmpXchgInIR(CASI)) {
+      MadeChange |= expandAtomicCmpXchg(CASI);
+    }
   }
-
   return MadeChange;
 }
 
-bool AtomicExpand::expandAtomicLoad(LoadInst *LI) {
-  auto TLI = TM->getSubtargetImpl()->getTargetLowering();
-  // If getInsertFencesForAtomic() returns true, then the target does not want
-  // to deal with memory orders, and emitLeading/TrailingFence should take care
-  // of everything. Otherwise, emitLeading/TrailingFence are no-op and we
-  // should preserve the ordering.
-  AtomicOrdering MemOpOrder =
-      TLI->getInsertFencesForAtomic() ? Monotonic : LI->getOrdering();
-  IRBuilder<> Builder(LI);
+bool AtomicExpand::bracketInstWithFences(Instruction *I, AtomicOrdering Order,
+                                         bool IsStore, bool IsLoad) {
+  IRBuilder<> Builder(I);
+
+  auto LeadingFence = TLI->emitLeadingFence(Builder, Order, IsStore, IsLoad);
+
+  auto TrailingFence = TLI->emitTrailingFence(Builder, Order, IsStore, IsLoad);
+  // The trailing fence is emitted before the instruction instead of after
+  // because there is no easy way of setting Builder insertion point after
+  // an instruction. So we must erase it from the BB, and insert it back
+  // in the right place.
+  // We have a guard here because not every atomic operation generates a
+  // trailing fence.
+  if (TrailingFence) {
+    TrailingFence->removeFromParent();
+    TrailingFence->insertAfter(I);
+  }
 
-  // Note that although no fence is required before atomic load on ARM, it is
-  // required before SequentiallyConsistent loads for the recommended Power
-  // mapping (see http://www.cl.cam.ac.uk/~pes20/cpp/cpp0xmappings.html).
-  // So we let the target choose what to emit.
-  TLI->emitLeadingFence(Builder, LI->getOrdering(),
-                        /*IsStore=*/false, /*IsLoad=*/true);
+  return (LeadingFence || TrailingFence);
+}
 
-  // The only 64-bit load guaranteed to be single-copy atomic by ARM is
-  // an ldrexd (A3.5.3).
-  Value *Val =
-      TLI->emitLoadLinked(Builder, LI->getPointerOperand(), MemOpOrder);
+bool AtomicExpand::tryExpandAtomicLoad(LoadInst *LI) {
+  switch (TLI->shouldExpandAtomicLoadInIR(LI)) {
+  case TargetLoweringBase::AtomicExpansionKind::None:
+    return false;
+  case TargetLoweringBase::AtomicExpansionKind::LLSC: {
+    return expandAtomicLoadToLL(LI);
+  }
+  case TargetLoweringBase::AtomicExpansionKind::CmpXChg: {
+    return expandAtomicLoadToCmpXchg(LI);
+  }
+  }
+  llvm_unreachable("Unhandled case in tryExpandAtomicLoad");
+}
+
+bool AtomicExpand::expandAtomicLoadToLL(LoadInst *LI) {
+  IRBuilder<> Builder(LI);
 
-  TLI->emitTrailingFence(Builder, LI->getOrdering(),
-                         /*IsStore=*/false, /*IsLoad=*/true);
+  // On some architectures, load-linked instructions are atomic for larger
+  // sizes than normal loads. For example, the only 64-bit load guaranteed
+  // to be single-copy atomic by ARM is an ldrexd (A3.5.3).
+  Value *Val =
+      TLI->emitLoadLinked(Builder, LI->getPointerOperand(), LI->getOrdering());
 
   LI->replaceAllUsesWith(Val);
   LI->eraseFromParent();
@@ -125,10 +199,31 @@ bool AtomicExpand::expandAtomicLoad(LoadInst *LI) {
   return true;
 }
 
+bool AtomicExpand::expandAtomicLoadToCmpXchg(LoadInst *LI) {
+  IRBuilder<> Builder(LI);
+  AtomicOrdering Order = LI->getOrdering();
+  Value *Addr = LI->getPointerOperand();
+  Type *Ty = cast<PointerType>(Addr->getType())->getElementType();
+  Constant *DummyVal = Constant::getNullValue(Ty);
+
+  Value *Pair = Builder.CreateAtomicCmpXchg(
+      Addr, DummyVal, DummyVal, Order,
+      AtomicCmpXchgInst::getStrongestFailureOrdering(Order));
+  Value *Loaded = Builder.CreateExtractValue(Pair, 0, "loaded");
+
+  LI->replaceAllUsesWith(Loaded);
+  LI->eraseFromParent();
+
+  return true;
+}
+
 bool AtomicExpand::expandAtomicStore(StoreInst *SI) {
-  // The only atomic 64-bit store on ARM is an strexd that succeeds, which means
-  // we need a loop and the entire instruction is essentially an "atomicrmw
-  // xchg" that ignores the value loaded.
+  // This function is only called on atomic stores that are too large to be
+  // atomic if implemented as a native store. So we replace them by an
+  // atomic swap, that can be implemented for example as a ldrex/strex on ARM
+  // or lock cmpxchg8/16b on X86, as these are atomic for larger sizes.
+  // It is the responsibility of the target to only signal expansion via
+  // shouldExpandAtomicRMW in cases where this is required and possible.
   IRBuilder<> Builder(SI);
   AtomicRMWInst *AI =
       Builder.CreateAtomicRMW(AtomicRMWInst::Xchg, SI->getPointerOperand(),
@@ -136,22 +231,77 @@ bool AtomicExpand::expandAtomicStore(StoreInst *SI) {
   SI->eraseFromParent();
 
   // Now we have an appropriate swap instruction, lower it as usual.
-  return expandAtomicRMW(AI);
+  return tryExpandAtomicRMW(AI);
 }
 
-bool AtomicExpand::expandAtomicRMW(AtomicRMWInst *AI) {
-  auto TLI = TM->getSubtargetImpl()->getTargetLowering();
-  AtomicOrdering Order = AI->getOrdering();
+static void createCmpXchgInstFun(IRBuilder<> &Builder, Value *Addr,
+                                 Value *Loaded, Value *NewVal,
+                                 AtomicOrdering MemOpOrder,
+                                 Value *&Success, Value *&NewLoaded) {
+  Value* Pair = Builder.CreateAtomicCmpXchg(
+      Addr, Loaded, NewVal, MemOpOrder,
+      AtomicCmpXchgInst::getStrongestFailureOrdering(MemOpOrder));
+  Success = Builder.CreateExtractValue(Pair, 1, "success");
+  NewLoaded = Builder.CreateExtractValue(Pair, 0, "newloaded");
+}
+
+bool AtomicExpand::tryExpandAtomicRMW(AtomicRMWInst *AI) {
+  switch (TLI->shouldExpandAtomicRMWInIR(AI)) {
+  case TargetLoweringBase::AtomicExpansionKind::None:
+    return false;
+  case TargetLoweringBase::AtomicExpansionKind::LLSC: {
+    return expandAtomicRMWToLLSC(AI);
+  }
+  case TargetLoweringBase::AtomicExpansionKind::CmpXChg: {
+    return expandAtomicRMWToCmpXchg(AI, createCmpXchgInstFun);
+  }
+  }
+  llvm_unreachable("Unhandled case in tryExpandAtomicRMW");
+}
+
+/// Emit IR to implement the given atomicrmw operation on values in registers,
+/// returning the new value.
+static Value *performAtomicOp(AtomicRMWInst::BinOp Op, IRBuilder<> &Builder,
+                              Value *Loaded, Value *Inc) {
+  Value *NewVal;
+  switch (Op) {
+  case AtomicRMWInst::Xchg:
+    return Inc;
+  case AtomicRMWInst::Add:
+    return Builder.CreateAdd(Loaded, Inc, "new");
+  case AtomicRMWInst::Sub:
+    return Builder.CreateSub(Loaded, Inc, "new");
+  case AtomicRMWInst::And:
+    return Builder.CreateAnd(Loaded, Inc, "new");
+  case AtomicRMWInst::Nand:
+    return Builder.CreateNot(Builder.CreateAnd(Loaded, Inc), "new");
+  case AtomicRMWInst::Or:
+    return Builder.CreateOr(Loaded, Inc, "new");
+  case AtomicRMWInst::Xor:
+    return Builder.CreateXor(Loaded, Inc, "new");
+  case AtomicRMWInst::Max:
+    NewVal = Builder.CreateICmpSGT(Loaded, Inc);
+    return Builder.CreateSelect(NewVal, Loaded, Inc, "new");
+  case AtomicRMWInst::Min:
+    NewVal = Builder.CreateICmpSLE(Loaded, Inc);
+    return Builder.CreateSelect(NewVal, Loaded, Inc, "new");
+  case AtomicRMWInst::UMax:
+    NewVal = Builder.CreateICmpUGT(Loaded, Inc);
+    return Builder.CreateSelect(NewVal, Loaded, Inc, "new");
+  case AtomicRMWInst::UMin:
+    NewVal = Builder.CreateICmpULE(Loaded, Inc);
+    return Builder.CreateSelect(NewVal, Loaded, Inc, "new");
+  default:
+    llvm_unreachable("Unknown atomic op");
+  }
+}
+
+bool AtomicExpand::expandAtomicRMWToLLSC(AtomicRMWInst *AI) {
+  AtomicOrdering MemOpOrder = AI->getOrdering();
   Value *Addr = AI->getPointerOperand();
   BasicBlock *BB = AI->getParent();
   Function *F = BB->getParent();
   LLVMContext &Ctx = F->getContext();
-  // If getInsertFencesForAtomic() return true, then the target does not want to
-  // deal with memory orders, and emitLeading/TrailingFence should take care of
-  // everything. Otherwise, emitLeading/TrailingFence are no-op and we should
-  // preserve the ordering.
-  AtomicOrdering MemOpOrder =
-      TLI->getInsertFencesForAtomic() ? Monotonic : Order;
 
   // Given: atomicrmw some_op iN* %addr, iN %incr ordering
   //
@@ -178,56 +328,14 @@ bool AtomicExpand::expandAtomicRMW(AtomicRMWInst *AI) {
   // the branch entirely.
   std::prev(BB->end())->eraseFromParent();
   Builder.SetInsertPoint(BB);
-  TLI->emitLeadingFence(Builder, Order, /*IsStore=*/true, /*IsLoad=*/true);
   Builder.CreateBr(LoopBB);
 
   // Start the main loop block now that we've taken care of the preliminaries.
   Builder.SetInsertPoint(LoopBB);
   Value *Loaded = TLI->emitLoadLinked(Builder, Addr, MemOpOrder);
 
-  Value *NewVal;
-  switch (AI->getOperation()) {
-  case AtomicRMWInst::Xchg:
-    NewVal = AI->getValOperand();
-    break;
-  case AtomicRMWInst::Add:
-    NewVal = Builder.CreateAdd(Loaded, AI->getValOperand(), "new");
-    break;
-  case AtomicRMWInst::Sub:
-    NewVal = Builder.CreateSub(Loaded, AI->getValOperand(), "new");
-    break;
-  case AtomicRMWInst::And:
-    NewVal = Builder.CreateAnd(Loaded, AI->getValOperand(), "new");
-    break;
-  case AtomicRMWInst::Nand:
-    NewVal = Builder.CreateNot(Builder.CreateAnd(Loaded, AI->getValOperand()),
-                               "new");
-    break;
-  case AtomicRMWInst::Or:
-    NewVal = Builder.CreateOr(Loaded, AI->getValOperand(), "new");
-    break;
-  case AtomicRMWInst::Xor:
-    NewVal = Builder.CreateXor(Loaded, AI->getValOperand(), "new");
-    break;
-  case AtomicRMWInst::Max:
-    NewVal = Builder.CreateICmpSGT(Loaded, AI->getValOperand());
-    NewVal = Builder.CreateSelect(NewVal, Loaded, AI->getValOperand(), "new");
-    break;
-  case AtomicRMWInst::Min:
-    NewVal = Builder.CreateICmpSLE(Loaded, AI->getValOperand());
-    NewVal = Builder.CreateSelect(NewVal, Loaded, AI->getValOperand(), "new");
-    break;
-  case AtomicRMWInst::UMax:
-    NewVal = Builder.CreateICmpUGT(Loaded, AI->getValOperand());
-    NewVal = Builder.CreateSelect(NewVal, Loaded, AI->getValOperand(), "new");
-    break;
-  case AtomicRMWInst::UMin:
-    NewVal = Builder.CreateICmpULE(Loaded, AI->getValOperand());
-    NewVal = Builder.CreateSelect(NewVal, Loaded, AI->getValOperand(), "new");
-    break;
-  default:
-    llvm_unreachable("Unknown atomic op");
-  }
+  Value *NewVal =
+      performAtomicOp(AI->getOperation(), Builder, Loaded, AI->getValOperand());
 
   Value *StoreSuccess =
       TLI->emitStoreConditional(Builder, NewVal, Addr, MemOpOrder);
@@ -236,7 +344,6 @@ bool AtomicExpand::expandAtomicRMW(AtomicRMWInst *AI) {
   Builder.CreateCondBr(TryAgain, LoopBB, ExitBB);
 
   Builder.SetInsertPoint(ExitBB, ExitBB->begin());
-  TLI->emitTrailingFence(Builder, Order, /*IsStore=*/true, /*IsLoad=*/true);
 
   AI->replaceAllUsesWith(Loaded);
   AI->eraseFromParent();
@@ -245,17 +352,16 @@ bool AtomicExpand::expandAtomicRMW(AtomicRMWInst *AI) {
 }
 
 bool AtomicExpand::expandAtomicCmpXchg(AtomicCmpXchgInst *CI) {
-  auto TLI = TM->getSubtargetImpl()->getTargetLowering();
   AtomicOrdering SuccessOrder = CI->getSuccessOrdering();
   AtomicOrdering FailureOrder = CI->getFailureOrdering();
   Value *Addr = CI->getPointerOperand();
   BasicBlock *BB = CI->getParent();
   Function *F = BB->getParent();
   LLVMContext &Ctx = F->getContext();
-  // If getInsertFencesForAtomic() return true, then the target does not want to
-  // deal with memory orders, and emitLeading/TrailingFence should take care of
-  // everything. Otherwise, emitLeading/TrailingFence are no-op and we should
-  // preserve the ordering.
+  // If getInsertFencesForAtomic() returns true, then the target does not want
+  // to deal with memory orders, and emitLeading/TrailingFence should take care
+  // of everything. Otherwise, emitLeading/TrailingFence are no-op and we
+  // should preserve the ordering.
   AtomicOrdering MemOpOrder =
       TLI->getInsertFencesForAtomic() ? Monotonic : SuccessOrder;
 
@@ -308,7 +414,7 @@ bool AtomicExpand::expandAtomicCmpXchg(AtomicCmpXchgInst *CI) {
   Value *ShouldStore =
       Builder.CreateICmpEQ(Loaded, CI->getCompareOperand(), "should_store");
 
-  // If the the cmpxchg doesn't actually need any ordering when it fails, we can
+  // If the cmpxchg doesn't actually need any ordering when it fails, we can
   // jump straight past that fence instruction (if it exists).
   Builder.CreateCondBr(ShouldStore, TryStoreBB, FailureBB);
 
@@ -377,3 +483,100 @@ bool AtomicExpand::expandAtomicCmpXchg(AtomicCmpXchgInst *CI) {
   CI->eraseFromParent();
   return true;
 }
+
+bool AtomicExpand::isIdempotentRMW(AtomicRMWInst* RMWI) {
+  auto C = dyn_cast<ConstantInt>(RMWI->getValOperand());
+  if(!C)
+    return false;
+
+  AtomicRMWInst::BinOp Op = RMWI->getOperation();
+  switch(Op) {
+    case AtomicRMWInst::Add:
+    case AtomicRMWInst::Sub:
+    case AtomicRMWInst::Or:
+    case AtomicRMWInst::Xor:
+      return C->isZero();
+    case AtomicRMWInst::And:
+      return C->isMinusOne();
+    // FIXME: we could also treat Min/Max/UMin/UMax by the INT_MIN/INT_MAX/...
+    default:
+      return false;
+  }
+}
+
+bool AtomicExpand::simplifyIdempotentRMW(AtomicRMWInst* RMWI) {
+  if (auto ResultingLoad = TLI->lowerIdempotentRMWIntoFencedLoad(RMWI)) {
+    tryExpandAtomicLoad(ResultingLoad);
+    return true;
+  }
+  return false;
+}
+
+bool llvm::expandAtomicRMWToCmpXchg(AtomicRMWInst *AI,
+                                    CreateCmpXchgInstFun CreateCmpXchg) {
+  assert(AI);
+
+  AtomicOrdering MemOpOrder =
+      AI->getOrdering() == Unordered ? Monotonic : AI->getOrdering();
+  Value *Addr = AI->getPointerOperand();
+  BasicBlock *BB = AI->getParent();
+  Function *F = BB->getParent();
+  LLVMContext &Ctx = F->getContext();
+
+  // Given: atomicrmw some_op iN* %addr, iN %incr ordering
+  //
+  // The standard expansion we produce is:
+  //     [...]
+  //     %init_loaded = load atomic iN* %addr
+  //     br label %loop
+  // loop:
+  //     %loaded = phi iN [ %init_loaded, %entry ], [ %new_loaded, %loop ]
+  //     %new = some_op iN %loaded, %incr
+  //     %pair = cmpxchg iN* %addr, iN %loaded, iN %new
+  //     %new_loaded = extractvalue { iN, i1 } %pair, 0
+  //     %success = extractvalue { iN, i1 } %pair, 1
+  //     br i1 %success, label %atomicrmw.end, label %loop
+  // atomicrmw.end:
+  //     [...]
+  BasicBlock *ExitBB = BB->splitBasicBlock(AI, "atomicrmw.end");
+  BasicBlock *LoopBB = BasicBlock::Create(Ctx, "atomicrmw.start", F, ExitBB);
+
+  // This grabs the DebugLoc from AI.
+  IRBuilder<> Builder(AI);
+
+  // The split call above "helpfully" added a branch at the end of BB (to the
+  // wrong place), but we want a load. It's easiest to just remove
+  // the branch entirely.
+  std::prev(BB->end())->eraseFromParent();
+  Builder.SetInsertPoint(BB);
+  LoadInst *InitLoaded = Builder.CreateLoad(Addr);
+  // Atomics require at least natural alignment.
+  InitLoaded->setAlignment(AI->getType()->getPrimitiveSizeInBits() / 8);
+  Builder.CreateBr(LoopBB);
+
+  // Start the main loop block now that we've taken care of the preliminaries.
+  Builder.SetInsertPoint(LoopBB);
+  PHINode *Loaded = Builder.CreatePHI(AI->getType(), 2, "loaded");
+  Loaded->addIncoming(InitLoaded, BB);
+
+  Value *NewVal =
+      performAtomicOp(AI->getOperation(), Builder, Loaded, AI->getValOperand());
+
+  Value *NewLoaded = nullptr;
+  Value *Success = nullptr;
+
+  CreateCmpXchg(Builder, Addr, Loaded, NewVal, MemOpOrder,
+                Success, NewLoaded);
+  assert(Success && NewLoaded);
+
+  Loaded->addIncoming(NewLoaded, LoopBB);
+
+  Builder.CreateCondBr(Success, ExitBB, LoopBB);
+
+  Builder.SetInsertPoint(ExitBB, ExitBB->begin());
+
+  AI->replaceAllUsesWith(NewLoaded);
+  AI->eraseFromParent();
+
+  return true;
+}