RLE: better handling of ref_element/tail_addr [immutable]

eeckstein · eeckstein · commit f97876c9e713 · 2021-11-29T09:41:05.000+01:00
Rerun RLE with cutting off the base address of loads at `ref_element/tail_addr [immutable]`. This increases the chance of catching loads of immutable COW class properties or elements.
diff --git a/include/swift/SIL/InstructionUtils.h b/include/swift/SIL/InstructionUtils.h
@@ -26,8 +26,6 @@ namespace swift {
 /// nothing left to strip.
 SILValue getUnderlyingObject(SILValue V);
 
-SILValue getUnderlyingObjectStopAtMarkDependence(SILValue V);
-
 SILValue stripSinglePredecessorArgs(SILValue V);
 
 /// Return the underlying SILValue after stripping off all casts from the
diff --git a/include/swift/SILOptimizer/Utils/LoadStoreOptUtils.h b/include/swift/SILOptimizer/Utils/LoadStoreOptUtils.h
@@ -379,21 +379,37 @@ class LSLocation : public LSBase {
   static void reduce(LSLocation Base, SILModule *Mod,
                      TypeExpansionContext context, LSLocationList &Locs);
 
+  /// Gets the base address for `v`.
+  /// If `stopAtImmutable` is true, the base address is only calculated up to
+  /// a `ref_element_addr [immutable]` or a `ref_tail_addr [immutable]`.
+  /// Return the base address and true if such an immutable class projection
+  /// is found.
+  static std::pair<SILValue, bool>
+  getBaseAddressOrObject(SILValue v, bool stopAtImmutable);
+
   /// Enumerate the given Mem LSLocation.
-  static void enumerateLSLocation(TypeExpansionContext context, SILModule *M,
+  /// If `stopAtImmutable` is true, the base address is only calculated up to
+  /// a `ref_element_addr [immutable]` or a `ref_tail_addr [immutable]`.
+  /// Returns true if it's an immutable location.
+  static bool enumerateLSLocation(TypeExpansionContext context, SILModule *M,
                                   SILValue Mem,
                                   std::vector<LSLocation> &LSLocationVault,
                                   LSLocationIndexMap &LocToBit,
                                   LSLocationBaseMap &BaseToLoc,
-                                  TypeExpansionAnalysis *TE);
+                                  TypeExpansionAnalysis *TE,
+                                  bool stopAtImmutable);
 
   /// Enumerate all the locations in the function.
+  /// If `stopAtImmutable` is true, the base addresses are only calculated up to
+  /// a `ref_element_addr [immutable]` or a `ref_tail_addr [immutable]`.
   static void enumerateLSLocations(SILFunction &F,
                                    std::vector<LSLocation> &LSLocationVault,
                                    LSLocationIndexMap &LocToBit,
                                    LSLocationBaseMap &BaseToLoc,
                                    TypeExpansionAnalysis *TE,
-                                   std::pair<int, int> &LSCount);
+                                   bool stopAtImmutable,
+                                   int &numLoads, int &numStores,
+                                   bool &immutableLoadsFound);
 };
 
 static inline llvm::hash_code hash_value(const LSLocation &L) {
diff --git a/lib/SIL/Utils/InstructionUtils.cpp b/lib/SIL/Utils/InstructionUtils.cpp
@@ -62,18 +62,6 @@ SILValue swift::getUnderlyingObject(SILValue v) {
   }
 }
 
-SILValue swift::getUnderlyingObjectStopAtMarkDependence(SILValue v) {
-  while (true) {
-    SILValue v2 = stripCastsWithoutMarkDependence(v);
-    v2 = stripAddressProjections(v2);
-    v2 = stripIndexingInsts(v2);
-    v2 = lookThroughOwnershipInsts(v2);
-    if (v2 == v)
-      return v2;
-    v = v2;
-  }
-}
-
 /// Return the underlying SILValue after stripping off identity SILArguments if
 /// we belong to a BB with one predecessor.
 SILValue swift::stripSinglePredecessorArgs(SILValue V) {
diff --git a/lib/SILOptimizer/Transforms/DeadStoreElimination.cpp b/lib/SILOptimizer/Transforms/DeadStoreElimination.cpp
@@ -1182,14 +1182,17 @@ void DSEContext::runIterativeDSE() {
 }
 
 bool DSEContext::run() {
-  std::pair<int, int> LSCount = std::make_pair(0, 0);
+  int numLoads = 0, numStores = 0;
+  bool immutableLoadsFound = false;
   // Walk over the function and find all the locations accessed by
   // this function.
   LSLocation::enumerateLSLocations(*F, LocationVault, LocToBitIndex,
-                                   BaseToLocIndex, TE, LSCount);
+                                   BaseToLocIndex, TE,
+                                   /*stopAtImmutable*/ false,
+                                   numLoads, numStores, immutableLoadsFound);
 
   // Check how to optimize this function.
-  ProcessKind Kind = getProcessFunctionKind(LSCount.second);
+  ProcessKind Kind = getProcessFunctionKind(numStores);
   
   // We do not optimize this function at all.
   if (Kind == ProcessKind::ProcessNone)
diff --git a/lib/SILOptimizer/Transforms/RedundantLoadElimination.cpp b/lib/SILOptimizer/Transforms/RedundantLoadElimination.cpp
@@ -486,14 +486,21 @@ class RLEContext {
   /// If set, RLE ignores loads from that array type.
   NominalTypeDecl *ArrayType;
 
+  /// Se to true if loads with a `ref_element_addr [immutable]` or
+  /// `ref_tail_addr [immutable]` base address are found.
+  bool immutableLoadsFound = false;
+
+  /// Only optimize loads with a base address of  `ref_element_addr [immutable]`
+  /// `ref_tail_addr [immutable]`.
+  bool onlyImmutableLoads;
+
 #ifndef NDEBUG
   SILPrintContext printCtx;
 #endif
 
 public:
-  RLEContext(SILFunction *F, SILPassManager *PM, AliasAnalysis *AA,
-             TypeExpansionAnalysis *TE, PostOrderFunctionInfo *PO,
-             EpilogueARCFunctionInfo *EAFI, bool disableArrayLoads);
+  RLEContext(SILFunction *F, SILPassManager *PM,
+             bool disableArrayLoads, bool onlyImmutableLoads);
 
   RLEContext(const RLEContext &) = delete;
   RLEContext(RLEContext &&) = delete;
@@ -504,6 +511,8 @@ class RLEContext {
   /// Entry point to redundant load elimination.
   bool run();
 
+  bool shouldOptimizeImmutableLoads() const { return immutableLoadsFound; }
+
   SILFunction *getFunction() const { return Fn; }
 
   /// Use a set of ad hoc rules to tell whether we should run a pessimistic
@@ -570,6 +579,11 @@ class RLEContext {
           LI->getType().getNominalOrBoundGenericNominal() != ArrayType) {
         return LI;
       }
+      if (onlyImmutableLoads &&
+          !LSLocation::getBaseAddressOrObject(LI->getOperand(),
+                                             /*stopAtImmutable*/ true).second) {
+        return nullptr;
+      }
     }
     return nullptr;
   }
@@ -1200,14 +1214,17 @@ void BlockState::dump(RLEContext &Ctx) {
 //                          RLEContext Implementation
 //===----------------------------------------------------------------------===//
 
-RLEContext::RLEContext(SILFunction *F, SILPassManager *PM, AliasAnalysis *AA,
-                       TypeExpansionAnalysis *TE, PostOrderFunctionInfo *PO,
-                       EpilogueARCFunctionInfo *EAFI, bool disableArrayLoads)
-    : Fn(F), PM(PM), AA(AA), TE(TE), PO(PO), EAFI(EAFI), BBToLocState(F),
-      BBWithLoads(F),
+RLEContext::RLEContext(SILFunction *F, SILPassManager *PM,
+                       bool disableArrayLoads, bool onlyImmutableLoads)
+    : Fn(F), PM(PM), AA(PM->getAnalysis<AliasAnalysis>(F)),
+      TE(PM->getAnalysis<TypeExpansionAnalysis>()),
+      PO(PM->getAnalysis<PostOrderAnalysis>()->get(F)),
+      EAFI(PM->getAnalysis<EpilogueARCAnalysis>()->get(F)),
+      BBToLocState(F), BBWithLoads(F),
       ArrayType(disableArrayLoads
                     ? F->getModule().getASTContext().getArrayDecl()
-                    : nullptr)
+                    : nullptr),
+      onlyImmutableLoads(onlyImmutableLoads)
 #ifndef NDEBUG
       ,
       printCtx(llvm::dbgs(), /*Verbose=*/false, /*Sorted=*/true)
@@ -1567,14 +1584,15 @@ bool RLEContext::run() {
   // Phase 4. we perform the redundant load elimination.
   // Walk over the function and find all the locations accessed by
   // this function.
-  std::pair<int, int> LSCount = std::make_pair(0, 0); 
+  int numLoads = 0, numStores = 0;
   LSLocation::enumerateLSLocations(*Fn, LocationVault,
                                    LocToBitIndex,
                                    BaseToLocIndex, TE,
-                                   LSCount);
+                                   /*stopAtImmutable*/ onlyImmutableLoads,
+                                   numLoads, numStores, immutableLoadsFound);
 
   // Check how to optimize this function.
-  ProcessKind Kind = getProcessFunctionKind(LSCount.first, LSCount.second);
+  ProcessKind Kind = getProcessFunctionKind(numLoads, numStores);
   
   // We do not optimize this function at all.
   if (Kind == ProcessKind::ProcessNone)
@@ -1681,15 +1699,21 @@ class RedundantLoadElimination : public SILFunctionTransform {
     LLVM_DEBUG(llvm::dbgs() << "*** RLE on function: " << F->getName()
                             << " ***\n");
 
-    auto *AA = PM->getAnalysis<AliasAnalysis>(F);
-    auto *TE = PM->getAnalysis<TypeExpansionAnalysis>();
-    auto *PO = PM->getAnalysis<PostOrderAnalysis>()->get(F);
-    auto *EAFI = PM->getAnalysis<EpilogueARCAnalysis>()->get(F);
-
-    RLEContext RLE(F, PM, AA, TE, PO, EAFI, disableArrayLoads);
+    RLEContext RLE(F, PM, disableArrayLoads,
+                   /*onlyImmutableLoads*/ false);
     if (RLE.run()) {
       invalidateAnalysis(SILAnalysis::InvalidationKind::Instructions);
     }
+    if (RLE.shouldOptimizeImmutableLoads()) {
+      /// Re-running RLE with cutting base addresses off at
+      /// `ref_element_addr [immutable]` or `ref_tail_addr [immutable]` can
+      /// expose additional opportunities.
+      RLEContext RLE2(F, PM, disableArrayLoads,
+                      /*onlyImmutableLoads*/ true);
+      if (RLE2.run()) {
+        invalidateAnalysis(SILAnalysis::InvalidationKind::Instructions);
+      }
+    }
   }
 };
 
diff --git a/lib/SILOptimizer/Utils/LoadStoreOptUtils.cpp b/lib/SILOptimizer/Utils/LoadStoreOptUtils.cpp
@@ -245,28 +245,60 @@ void LSLocation::reduce(LSLocation Base, SILModule *M,
     replaceSubLocations(Base, M, context, Locs, SubLocations);
 }
 
-void LSLocation::enumerateLSLocation(TypeExpansionContext context, SILModule *M,
+std::pair<SILValue, bool>
+LSLocation::getBaseAddressOrObject(SILValue v, bool stopAtImmutable) {
+  bool isImmutable = false;
+  while (true) {
+    if (auto *rea = dyn_cast<RefElementAddrInst>(v)) {
+      if (rea->isImmutable()) {
+        isImmutable = true;
+        if (stopAtImmutable)
+          return {v, true};
+      }
+    }
+    if (auto *rta = dyn_cast<RefTailAddrInst>(v)) {
+      if (rta->isImmutable()) {
+        isImmutable = true;
+        if (stopAtImmutable)
+          return {v, true};
+      }
+    }
+    SILValue v2 = stripCastsWithoutMarkDependence(v);
+    v2 = stripSinglePredecessorArgs(v2);
+    if (Projection::isAddressProjection(v2))
+      v2 = cast<SingleValueInstruction>(v2)->getOperand(0);
+    v2 = stripIndexingInsts(v2);
+    v2 = lookThroughOwnershipInsts(v2);
+    if (v2 == v)
+      return {v2, isImmutable};
+    v = v2;
+  }
+}
+
+bool LSLocation::enumerateLSLocation(TypeExpansionContext context, SILModule *M,
                                      SILValue Mem,
                                      std::vector<LSLocation> &Locations,
                                      LSLocationIndexMap &IndexMap,
                                      LSLocationBaseMap &BaseMap,
-                                     TypeExpansionAnalysis *TypeCache) {
+                                     TypeExpansionAnalysis *TypeCache,
+                                     bool stopAtImmutable) {
   // We have processed this SILValue before.
   if (BaseMap.find(Mem) != BaseMap.end())
-    return;
+    return false;
 
   // Construct a Location to represent the memory written by this instruction.
   // ProjectionPath currently does not handle mark_dependence so stop our
   // underlying object search at these instructions.
   // We still get a benefit if we cse mark_dependence instructions and then
   // merge loads from them.
-  SILValue UO = getUnderlyingObjectStopAtMarkDependence(Mem);
+  auto baseAndImmutable = getBaseAddressOrObject(Mem, stopAtImmutable);
+  SILValue UO = baseAndImmutable.first;
   LSLocation L(UO, ProjectionPath::getProjectionPath(UO, Mem));
 
   // If we can't figure out the Base or Projection Path for the memory location,
   // simply ignore it for now.
   if (!L.isValid())
-    return;
+    return false;
 
   // Record the SILValue to location mapping.
   BaseMap[Mem] = L; 
@@ -281,6 +313,7 @@ void LSLocation::enumerateLSLocation(TypeExpansionContext context, SILModule *M,
     IndexMap[Loc] = Locations.size();
     Locations.push_back(Loc);
   }
+  return baseAndImmutable.first;
 }
 
 void
@@ -289,22 +322,26 @@ LSLocation::enumerateLSLocations(SILFunction &F,
                                  LSLocationIndexMap &IndexMap,
                                  LSLocationBaseMap &BaseMap,
                                  TypeExpansionAnalysis *TypeCache,
-                                 std::pair<int, int> &LSCount) {
+                                 bool stopAtImmutable,
+                                 int &numLoads, int &numStores,
+                                 bool &immutableLoadsFound) {
   // Enumerate all locations accessed by the loads or stores.
   for (auto &B : F) {
     for (auto &I : B) {
       if (auto *LI = dyn_cast<LoadInst>(&I)) {
-        enumerateLSLocation(F.getTypeExpansionContext(), &I.getModule(),
+        if (enumerateLSLocation(F.getTypeExpansionContext(), &I.getModule(),
                             LI->getOperand(), Locations, IndexMap, BaseMap,
-                            TypeCache);
-        ++LSCount.first;
+                            TypeCache, stopAtImmutable)) {
+          immutableLoadsFound = true;
+        }
+        ++numLoads;
         continue;
       }
       if (auto *SI = dyn_cast<StoreInst>(&I)) {
         enumerateLSLocation(F.getTypeExpansionContext(), &I.getModule(),
                             SI->getDest(), Locations, IndexMap, BaseMap,
-                            TypeCache);
-        ++LSCount.second;
+                            TypeCache, stopAtImmutable);
+        ++numStores;
         continue;
       }
     }
diff --git a/test/SILOptimizer/redundant_load_elim_ossa.sil b/test/SILOptimizer/redundant_load_elim_ossa.sil
@@ -163,6 +163,22 @@ bb0(%0 : @owned $AB):
   return %5 : $Int
 }
 
+// CHECK-LABEL: sil [ossa] @forward_load_of_immutable_class_property
+// CHECK:     [[L:%[0-9]+]] = load
+// CHECK:     apply %{{[0-9]+}}([[L]])
+// CHECK-NOT: load
+// CHECK:     return [[L]]
+// CHECK-LABEL: } // end sil function 'forward_load_of_immutable_class_property'
+sil [ossa] @forward_load_of_immutable_class_property : $@convention(thin) (@guaranteed AB) -> Int {
+bb0(%0 : @guaranteed $AB):
+  %1 = ref_element_addr [immutable] %0 : $AB, #AB.value
+  %2 = load [trivial] %1 : $*Int
+  %3 = function_ref @use_Int : $@convention(thin) (Int) -> ()
+  apply %3(%2) : $@convention(thin) (Int) -> ()
+  %5 = load [trivial] %1 : $*Int
+  return %5 : $Int
+}
+
 // CHECK-LABEL: sil hidden [ossa] @load_forward_across_end_cow_mutation :
 // CHECK-NOT: = load
 // CHECK-LABEL: } // end sil function 'load_forward_across_end_cow_mutation'
