Reapply: [MemCpyOpt] implement single BB stack-move optimization which unify the static unescaped allocas

khei4 · khei4 · commit ca68a7f956f2 · 2023-08-15T22:13:09.000+09:00
This reverts commit 2077180.
diff --git a/llvm/include/llvm/Transforms/Scalar/MemCpyOptimizer.h b/llvm/include/llvm/Transforms/Scalar/MemCpyOptimizer.h
@@ -20,6 +20,7 @@
 namespace llvm {
 
 class AAResults;
+class AllocaInst;
 class BatchAAResults;
 class AssumptionCache;
 class CallBase;
@@ -77,6 +78,9 @@ class MemCpyOptPass : public PassInfoMixin<MemCpyOptPass> {
   Instruction *tryMergingIntoMemset(Instruction *I, Value *StartPtr,
                                     Value *ByteVal);
   bool moveUp(StoreInst *SI, Instruction *P, const LoadInst *LI);
+  bool performStackMoveOptzn(Instruction *Load, Instruction *Store,
+                             AllocaInst *DestAlloca, AllocaInst *SrcAlloca,
+                             uint64_t Size, BatchAAResults &BAA);
 
   void eraseInstruction(Instruction *I);
   bool iterateOnFunction(Function &F);
diff --git a/llvm/lib/Transforms/Scalar/MemCpyOptimizer.cpp b/llvm/lib/Transforms/Scalar/MemCpyOptimizer.cpp
@@ -66,9 +66,10 @@ static cl::opt<bool> EnableMemCpyOptWithoutLibcalls(
 
 STATISTIC(NumMemCpyInstr, "Number of memcpy instructions deleted");
 STATISTIC(NumMemSetInfer, "Number of memsets inferred");
-STATISTIC(NumMoveToCpy,   "Number of memmoves converted to memcpy");
-STATISTIC(NumCpyToSet,    "Number of memcpys converted to memset");
-STATISTIC(NumCallSlot,    "Number of call slot optimizations performed");
+STATISTIC(NumMoveToCpy, "Number of memmoves converted to memcpy");
+STATISTIC(NumCpyToSet, "Number of memcpys converted to memset");
+STATISTIC(NumCallSlot, "Number of call slot optimizations performed");
+STATISTIC(NumStackMove, "Number of stack-move optimizations performed");
 
 namespace {
 
@@ -730,6 +731,23 @@ bool MemCpyOptPass::processStoreOfLoad(StoreInst *SI, LoadInst *LI,
     return true;
   }
 
+  // If this is a load-store pair from a stack slot to a stack slot, we
+  // might be able to perform the stack-move optimization just as we do for
+  // memcpys from an alloca to an alloca.
+  if (auto *DestAlloca = dyn_cast<AllocaInst>(SI->getPointerOperand())) {
+    if (auto *SrcAlloca = dyn_cast<AllocaInst>(LI->getPointerOperand())) {
+      if (performStackMoveOptzn(LI, SI, DestAlloca, SrcAlloca,
+                                DL.getTypeStoreSize(T), BAA)) {
+        // Avoid invalidating the iterator.
+        BBI = SI->getNextNonDebugInstruction()->getIterator();
+        eraseInstruction(SI);
+        eraseInstruction(LI);
+        ++NumMemCpyInstr;
+        return true;
+      }
+    }
+  }
+
   return false;
 }
 
@@ -1408,6 +1426,227 @@ bool MemCpyOptPass::performMemCpyToMemSetOptzn(MemCpyInst *MemCpy,
   return true;
 }
 
+// Attempts to optimize the pattern whereby memory is copied from an alloca to
+// another alloca, where the two allocas don't have conflicting mod/ref. If
+// successful, the two allocas can be merged into one and the transfer can be
+// deleted. This pattern is generated frequently in Rust, due to the ubiquity of
+// move operations in that language.
+//
+// Once we determine that the optimization is safe to perform, we replace all
+// uses of the destination alloca with the source alloca. We also "shrink wrap"
+// the lifetime markers of the single merged alloca to before the first use
+// and after the last use. Note that the "shrink wrapping" procedure is a safe
+// transformation only because we restrict the scope of this optimization to
+// allocas that aren't captured.
+bool MemCpyOptPass::performStackMoveOptzn(Instruction *Load, Instruction *Store,
+                                          AllocaInst *DestAlloca,
+                                          AllocaInst *SrcAlloca, uint64_t Size,
+                                          BatchAAResults &BAA) {
+  LLVM_DEBUG(dbgs() << "Stack Move: Attempting to optimize:\n"
+                    << *Store << "\n");
+
+  // Make sure the two allocas are in the same address space.
+  if (SrcAlloca->getAddressSpace() != DestAlloca->getAddressSpace()) {
+    LLVM_DEBUG(dbgs() << "Stack Move: Address space mismatch\n");
+    return false;
+  }
+
+  // 1. Check that copy is full. Calculate the static size of the allocas to be
+  // merged, bail out if we can't.
+  const DataLayout &DL = DestAlloca->getModule()->getDataLayout();
+  std::optional<TypeSize> SrcSize = SrcAlloca->getAllocationSize(DL);
+  if (!SrcSize || SrcSize->isScalable() || Size != SrcSize->getFixedValue()) {
+    LLVM_DEBUG(dbgs() << "Stack Move: Source alloca size mismatch\n");
+    return false;
+  }
+  std::optional<TypeSize> DestSize = DestAlloca->getAllocationSize(DL);
+  if (!DestSize || DestSize->isScalable() ||
+      Size != DestSize->getFixedValue()) {
+    LLVM_DEBUG(dbgs() << "Stack Move: Destination alloca size mismatch\n");
+    return false;
+  }
+
+  // 2-1. Check that src and dest are static allocas, which are not affected by
+  // stacksave/stackrestore.
+  if (!SrcAlloca->isStaticAlloca() || !DestAlloca->isStaticAlloca() ||
+      SrcAlloca->getParent() != Load->getParent() ||
+      SrcAlloca->getParent() != Store->getParent())
+    return false;
+
+  // 2-2. Check that src and dest are never captured, unescaped allocas. Also
+  // collect lifetime markers first/last users in order to shrink wrap the
+  // lifetimes, and instructions with noalias metadata to remove them.
+
+  SmallVector<Instruction *, 4> LifetimeMarkers;
+  Instruction *FirstUser = nullptr, *LastUser = nullptr;
+  SmallSet<Instruction *, 4> NoAliasInstrs;
+
+  // Recursively track the user and check whether modified alias exist.
+  auto IsDereferenceableOrNull = [](Value *V, const DataLayout &DL) -> bool {
+    bool CanBeNull, CanBeFreed;
+    return V->getPointerDereferenceableBytes(DL, CanBeNull, CanBeFreed);
+  };
+
+  auto CaptureTrackingWithModRef =
+      [&](Instruction *AI,
+          function_ref<bool(Instruction *)> ModRefCallback) -> bool {
+    SmallVector<Instruction *, 8> Worklist;
+    Worklist.push_back(AI);
+    unsigned MaxUsesToExplore = getDefaultMaxUsesToExploreForCaptureTracking();
+    Worklist.reserve(MaxUsesToExplore);
+    SmallSet<const Use *, 20> Visited;
+    while (!Worklist.empty()) {
+      Instruction *I = Worklist.back();
+      Worklist.pop_back();
+      for (const Use &U : I->uses()) {
+        if (Visited.size() >= MaxUsesToExplore) {
+          LLVM_DEBUG(
+              dbgs()
+              << "Stack Move: Exceeded max uses to see ModRef, bailing\n");
+          return false;
+        }
+        if (!Visited.insert(&U).second)
+          continue;
+        switch (DetermineUseCaptureKind(U, IsDereferenceableOrNull)) {
+        case UseCaptureKind::MAY_CAPTURE:
+          return false;
+        case UseCaptureKind::PASSTHROUGH:
+          // Instructions cannot have non-instruction users.
+          Worklist.push_back(cast<Instruction>(U.getUser()));
+          continue;
+        case UseCaptureKind::NO_CAPTURE: {
+          auto *UI = cast<Instruction>(U.getUser());
+          if (DestAlloca->getParent() != UI->getParent())
+            return false;
+          if (!FirstUser || UI->comesBefore(FirstUser))
+            FirstUser = UI;
+          if (!LastUser || LastUser->comesBefore(UI))
+            LastUser = UI;
+          if (UI->isLifetimeStartOrEnd()) {
+            // We note the locations of these intrinsic calls so that we can
+            // delete them later if the optimization succeeds, this is safe
+            // since both llvm.lifetime.start and llvm.lifetime.end intrinsics
+            // practically fill all the bytes of the alloca with an undefined
+            // value, although conceptually marked as alive/dead.
+            int64_t Size = cast<ConstantInt>(UI->getOperand(0))->getSExtValue();
+            if (Size < 0 || Size == DestSize) {
+              LifetimeMarkers.push_back(UI);
+              continue;
+            }
+          }
+          if (UI->hasMetadata(LLVMContext::MD_noalias))
+            NoAliasInstrs.insert(UI);
+          if (!ModRefCallback(UI))
+            return false;
+        }
+        }
+      }
+    }
+    return true;
+  };
+
+  // 3. Check that dest has no Mod/Ref, except full size lifetime intrinsics,
+  // from the alloca to the Store.
+  ModRefInfo DestModRef = ModRefInfo::NoModRef;
+  MemoryLocation DestLoc(DestAlloca, LocationSize::precise(Size));
+  auto DestModRefCallback = [&](Instruction *UI) -> bool {
+    // We don't care about the store itself.
+    if (UI == Store)
+      return true;
+    ModRefInfo Res = BAA.getModRefInfo(UI, DestLoc);
+    // FIXME: For multi-BB cases, we need to see reachability from it to
+    // store.
+    // Bailout if Dest may have any ModRef before Store.
+    if (UI->comesBefore(Store) && isModOrRefSet(Res))
+      return false;
+    DestModRef |= BAA.getModRefInfo(UI, DestLoc);
+
+    return true;
+  };
+
+  if (!CaptureTrackingWithModRef(DestAlloca, DestModRefCallback))
+    return false;
+
+  // 3. Check that, from after the Load to the end of the BB,
+  // 3-1. if the dest has any Mod, src has no Ref, and
+  // 3-2. if the dest has any Ref, src has no Mod except full-sized lifetimes.
+  MemoryLocation SrcLoc(SrcAlloca, LocationSize::precise(Size));
+
+  auto SrcModRefCallback = [&](Instruction *UI) -> bool {
+    // Any ModRef before Load doesn't matter, also Load and Store can be
+    // ignored.
+    if (UI->comesBefore(Load) || UI == Load || UI == Store)
+      return true;
+    ModRefInfo Res = BAA.getModRefInfo(UI, SrcLoc);
+    if ((isModSet(DestModRef) && isRefSet(Res)) ||
+        (isRefSet(DestModRef) && isModSet(Res)))
+      return false;
+
+    return true;
+  };
+
+  if (!CaptureTrackingWithModRef(SrcAlloca, SrcModRefCallback))
+    return false;
+
+  // We can do the transformation. First, align the allocas appropriately.
+  SrcAlloca->setAlignment(
+      std::max(SrcAlloca->getAlign(), DestAlloca->getAlign()));
+
+  // Merge the two allocas.
+  DestAlloca->replaceAllUsesWith(SrcAlloca);
+  eraseInstruction(DestAlloca);
+
+  // Drop metadata on the source alloca.
+  SrcAlloca->dropUnknownNonDebugMetadata();
+
+  // Do "shrink wrap" the lifetimes, if the original lifetime intrinsics exists.
+  if (!LifetimeMarkers.empty()) {
+    LLVMContext &C = SrcAlloca->getContext();
+    IRBuilder<> Builder(C);
+
+    ConstantInt *AllocaSize = ConstantInt::get(Type::getInt64Ty(C), Size);
+    // Create a new lifetime start marker before the first user of src or alloca
+    // users.
+    Builder.SetInsertPoint(FirstUser->getParent(), FirstUser->getIterator());
+    auto *Start = Builder.CreateLifetimeStart(SrcAlloca, AllocaSize);
+    auto *FirstMA = MSSA->getMemoryAccess(FirstUser);
+    auto *StartMA = MSSAU->createMemoryAccessBefore(
+        Start, FirstMA->getDefiningAccess(), FirstMA);
+    MSSAU->insertDef(cast<MemoryDef>(StartMA), /*RenameUses=*/true);
+
+    // Create a new lifetime end marker after the last user of src or alloca
+    // users.
+    // FIXME: If the last user is the terminator for the bb, we can insert
+    // lifetime.end marker to the immidiate post-dominator, but currently do
+    // nothing.
+    if (!LastUser->isTerminator()) {
+      Builder.SetInsertPoint(LastUser->getParent(), ++LastUser->getIterator());
+      auto *End = Builder.CreateLifetimeEnd(SrcAlloca, AllocaSize);
+      auto *LastMA = MSSA->getMemoryAccess(LastUser);
+      // FIXME: the second argument should be LastMA if LastMA is MemoryDef, but
+      // that's updated by insertDef.
+      auto *EndMA = MSSAU->createMemoryAccessAfter(
+          End, LastMA->getDefiningAccess(), LastMA);
+      MSSAU->insertDef(cast<MemoryDef>(EndMA), /*RenameUses=*/true);
+    }
+
+    // Remove all other lifetime markers.
+    for (Instruction *I : LifetimeMarkers)
+      eraseInstruction(I);
+  }
+
+  // As this transformation can cause memory accesses that didn't previously
+  // alias to begin to alias one another, we remove !noalias metadata from any
+  // uses of either alloca. This is conservative, but more precision doesn't
+  // seem worthwhile right now.
+  for (Instruction *I : NoAliasInstrs)
+    I->setMetadata(LLVMContext::MD_noalias, nullptr);
+
+  LLVM_DEBUG(dbgs() << "Stack Move: Performed staack-move optimization\n");
+  NumStackMove++;
+  return true;
+}
+
 /// Perform simplification of memcpy's.  If we have memcpy A
 /// which copies X to Y, and memcpy B which copies Y to Z, then we can rewrite
 /// B to be a memcpy from X to Z (or potentially a memmove, depending on
@@ -1464,13 +1703,14 @@ bool MemCpyOptPass::processMemCpy(MemCpyInst *M, BasicBlock::iterator &BBI) {
   MemoryAccess *SrcClobber = MSSA->getWalker()->getClobberingMemoryAccess(
       AnyClobber, MemoryLocation::getForSource(M), BAA);
 
-  // There are four possible optimizations we can do for memcpy:
+  // There are five possible optimizations we can do for memcpy:
   //   a) memcpy-memcpy xform which exposes redundance for DSE.
   //   b) call-memcpy xform for return slot optimization.
   //   c) memcpy from freshly alloca'd space or space that has just started
   //      its lifetime copies undefined data, and we can therefore eliminate
   //      the memcpy in favor of the data that was already at the destination.
   //   d) memcpy from a just-memset'd source can be turned into memset.
+  //   e) elimination of memcpy via stack-move optimization.
   if (auto *MD = dyn_cast<MemoryDef>(SrcClobber)) {
     if (Instruction *MI = MD->getMemoryInst()) {
       if (auto *CopySize = dyn_cast<ConstantInt>(M->getLength())) {
@@ -1489,7 +1729,8 @@ bool MemCpyOptPass::processMemCpy(MemCpyInst *M, BasicBlock::iterator &BBI) {
         }
       }
       if (auto *MDep = dyn_cast<MemCpyInst>(MI))
-        return processMemCpyMemCpyDependence(M, MDep, BAA);
+        if (processMemCpyMemCpyDependence(M, MDep, BAA))
+          return true;
       if (auto *MDep = dyn_cast<MemSetInst>(MI)) {
         if (performMemCpyToMemSetOptzn(M, MDep, BAA)) {
           LLVM_DEBUG(dbgs() << "Converted memcpy to memset\n");
@@ -1508,6 +1749,27 @@ bool MemCpyOptPass::processMemCpy(MemCpyInst *M, BasicBlock::iterator &BBI) {
     }
   }
 
+  // If the transfer is from a stack slot to a stack slot, then we may be able
+  // to perform the stack-move optimization. See the comments in
+  // performStackMoveOptzn() for more details.
+  auto *DestAlloca = dyn_cast<AllocaInst>(M->getDest());
+  if (!DestAlloca)
+    return false;
+  auto *SrcAlloca = dyn_cast<AllocaInst>(M->getSource());
+  if (!SrcAlloca)
+    return false;
+  ConstantInt *Len = dyn_cast<ConstantInt>(M->getLength());
+  if (Len == nullptr)
+    return false;
+  if (performStackMoveOptzn(M, M, DestAlloca, SrcAlloca, Len->getZExtValue(),
+                            BAA)) {
+    // Avoid invalidating the iterator.
+    BBI = M->getNextNonDebugInstruction()->getIterator();
+    eraseInstruction(M);
+    ++NumMemCpyInstr;
+    return true;
+  }
+
   return false;
 }
 
diff --git a/llvm/test/Transforms/MemCpyOpt/callslot.ll b/llvm/test/Transforms/MemCpyOpt/callslot.ll
@@ -56,11 +56,9 @@ define void @write_dest_between_call_and_memcpy() {
 
 define void @write_src_between_call_and_memcpy() {
 ; CHECK-LABEL: @write_src_between_call_and_memcpy(
-; CHECK-NEXT:    [[DEST:%.*]] = alloca [16 x i8], align 1
 ; CHECK-NEXT:    [[SRC:%.*]] = alloca [16 x i8], align 1
 ; CHECK-NEXT:    call void @llvm.memset.p0.i64(ptr [[SRC]], i8 0, i64 16, i1 false)
 ; CHECK-NEXT:    store i8 1, ptr [[SRC]], align 1
-; CHECK-NEXT:    call void @llvm.memcpy.p0.p0.i64(ptr [[DEST]], ptr [[SRC]], i64 16, i1 false)
 ; CHECK-NEXT:    ret void
 ;
   %dest = alloca [16 x i8]
diff --git a/llvm/test/Transforms/MemCpyOpt/lifetime-missing.ll b/llvm/test/Transforms/MemCpyOpt/lifetime-missing.ll
@@ -13,15 +13,14 @@ declare void @llvm.memset.p0.i64(ptr nocapture writeonly, i8, i64, i1 immarg)
 define void @test() {
 ; CHECK-LABEL: define void @test() {
 ; CHECK-NEXT:  entry:
-; CHECK-NEXT:    [[AGG_TMP3_SROA_35:%.*]] = alloca [20 x i8], align 4
 ; CHECK-NEXT:    [[AGG_TMP_SROA_14:%.*]] = alloca [20 x i8], align 4
 ; CHECK-NEXT:    [[AGG_TMP_SROA_14_128_SROA_IDX:%.*]] = getelementptr i8, ptr [[AGG_TMP_SROA_14]], i64 4
+; CHECK-NEXT:    call void @llvm.lifetime.start.p0(i64 20, ptr [[AGG_TMP_SROA_14]])
 ; CHECK-NEXT:    call void @llvm.memset.p0.i64(ptr [[AGG_TMP_SROA_14_128_SROA_IDX]], i8 0, i64 1, i1 false)
-; CHECK-NEXT:    call void @llvm.lifetime.start.p0(i64 20, ptr [[AGG_TMP3_SROA_35]])
-; CHECK-NEXT:    call void @llvm.memcpy.p0.p0.i64(ptr [[AGG_TMP3_SROA_35]], ptr [[AGG_TMP_SROA_14]], i64 20, i1 false)
-; CHECK-NEXT:    [[AGG_TMP3_SROA_35_128_SROA_IDX:%.*]] = getelementptr i8, ptr [[AGG_TMP3_SROA_35]], i64 4
-; CHECK-NEXT:    call void @llvm.memcpy.p0.p0.i64(ptr inttoptr (i64 4 to ptr), ptr [[AGG_TMP3_SROA_35_128_SROA_IDX]], i64 1, i1 false)
-; CHECK-NEXT:    call void @llvm.memcpy.p0.p0.i64(ptr null, ptr [[AGG_TMP3_SROA_35_128_SROA_IDX]], i64 1, i1 false)
+; CHECK-NEXT:    [[AGG_TMP3_SROA_35_128_SROA_IDX:%.*]] = getelementptr i8, ptr [[AGG_TMP_SROA_14]], i64 4
+; CHECK-NEXT:    call void @llvm.memset.p0.i64(ptr inttoptr (i64 4 to ptr), i8 0, i64 1, i1 false)
+; CHECK-NEXT:    call void @llvm.lifetime.end.p0(i64 20, ptr [[AGG_TMP_SROA_14]])
+; CHECK-NEXT:    call void @llvm.memset.p0.i64(ptr null, i8 0, i64 1, i1 false)
 ; CHECK-NEXT:    ret void
 ;
 entry:
diff --git a/llvm/test/Transforms/MemCpyOpt/stack-move.ll b/llvm/test/Transforms/MemCpyOpt/stack-move.ll
