[LAA] Drop alias scope metadata that is not valid across iterations

[nikic]

LAA currently adds memory locations with their original AATags to AST. However, scoped alias AATags may be valid only within one loop iteration, while LAA reasons across iterations.

Fix this by determining which alias scopes are defined inside the loop, and drop AATags that reference these scopes.

Fixes #79137.

[llvmbot]

@llvm/pr-subscribers-llvm-transforms

@llvm/pr-subscribers-llvm-analysis

Author: Nikita Popov (nikic)

Changes

LAA currently adds memory locations with their original AATags to AST. However, scoped alias AATags may be valid only within one loop iteration, while LAA reasons across iterations.

Fix this by determining which alias scopes are defined inside the loop, and drop AATags that reference these scopes.

Fixes #79137.

---

Full diff: https://github.com/llvm/llvm-project/pull/79161.diff

2 Files Affected:

• (modified) llvm/lib/Analysis/LoopAccessAnalysis.cpp (+46-7)
• (modified) llvm/test/Analysis/LoopAccessAnalysis/noalias-scope-decl.ll (+10-1)

diff --git a/llvm/lib/Analysis/LoopAccessAnalysis.cpp b/llvm/lib/Analysis/LoopAccessAnalysis.cpp
index 7e67c90152829dc..dd6b88fee415a7a 100644
--- a/llvm/lib/Analysis/LoopAccessAnalysis.cpp
+++ b/llvm/lib/Analysis/LoopAccessAnalysis.cpp
@@ -657,16 +657,18 @@ class AccessAnalysis {
 
   AccessAnalysis(Loop *TheLoop, AAResults *AA, LoopInfo *LI,
                  MemoryDepChecker::DepCandidates &DA,
-                 PredicatedScalarEvolution &PSE)
-      : TheLoop(TheLoop), BAA(*AA), AST(BAA), LI(LI), DepCands(DA), PSE(PSE) {
+                 PredicatedScalarEvolution &PSE,
+                 SmallPtrSetImpl<MDNode *> &LoopAliasScopes)
+      : TheLoop(TheLoop), BAA(*AA), AST(BAA), LI(LI), DepCands(DA), PSE(PSE),
+        LoopAliasScopes(LoopAliasScopes) {
     // We're analyzing dependences across loop iterations.
     BAA.enableCrossIterationMode();
   }
 
   /// Register a load  and whether it is only read from.
   void addLoad(MemoryLocation &Loc, Type *AccessTy, bool IsReadOnly) {
-    Value *Ptr = const_cast<Value*>(Loc.Ptr);
-    AST.add(Loc.getWithNewSize(LocationSize::beforeOrAfterPointer()));
+    Value *Ptr = const_cast<Value *>(Loc.Ptr);
+    AST.add(adjustLoc(Loc));
     Accesses[MemAccessInfo(Ptr, false)].insert(AccessTy);
     if (IsReadOnly)
       ReadOnlyPtr.insert(Ptr);
@@ -674,8 +676,8 @@ class AccessAnalysis {
 
   /// Register a store.
   void addStore(MemoryLocation &Loc, Type *AccessTy) {
-    Value *Ptr = const_cast<Value*>(Loc.Ptr);
-    AST.add(Loc.getWithNewSize(LocationSize::beforeOrAfterPointer()));
+    Value *Ptr = const_cast<Value *>(Loc.Ptr);
+    AST.add(adjustLoc(Loc));
     Accesses[MemAccessInfo(Ptr, true)].insert(AccessTy);
   }
 
@@ -731,6 +733,32 @@ class AccessAnalysis {
 private:
   typedef MapVector<MemAccessInfo, SmallSetVector<Type *, 1>> PtrAccessMap;
 
+  /// Adjust the MemoryLocation so that it represents accesses to this
+  /// location across all iterations, rather than a single one.
+  MemoryLocation adjustLoc(MemoryLocation Loc) const {
+    // The accessed location varies within the loop, but remains within the
+    // underlying object.
+    Loc.Size = LocationSize::beforeOrAfterPointer();
+    Loc.AATags.Scope = adjustAliasScopeList(Loc.AATags.Scope);
+    Loc.AATags.NoAlias = adjustAliasScopeList(Loc.AATags.NoAlias);
+    return Loc;
+  }
+
+  /// Drop alias scopes that are only valid within a single loop iteration.
+  MDNode *adjustAliasScopeList(MDNode *ScopeList) const {
+    if (!ScopeList)
+      return nullptr;
+
+    // For the sake of simplicity, drop the whole scope list if any scope is
+    // iteration-local.
+    if (any_of(ScopeList->operands(), [&](Metadata *Scope) {
+          return LoopAliasScopes.contains(cast<MDNode>(Scope));
+        }))
+      return nullptr;
+
+    return ScopeList;
+  }
+
   /// Go over all memory access and check whether runtime pointer checks
   /// are needed and build sets of dependency check candidates.
   void processMemAccesses();
@@ -775,6 +803,10 @@ class AccessAnalysis {
   PredicatedScalarEvolution &PSE;
 
   DenseMap<Value *, SmallVector<const Value *, 16>> UnderlyingObjects;
+
+  /// Alias scopes that are declared inside the loop, and as such not valid
+  /// across iterations.
+  SmallPtrSetImpl<MDNode *> &LoopAliasScopes;
 };
 
 } // end anonymous namespace
@@ -2283,6 +2315,7 @@ void LoopAccessInfo::analyzeLoop(AAResults *AA, LoopInfo *LI,
   // Holds the Load and Store instructions.
   SmallVector<LoadInst *, 16> Loads;
   SmallVector<StoreInst *, 16> Stores;
+  SmallPtrSet<MDNode *, 8> LoopAliasScopes;
 
   // Holds all the different accesses in the loop.
   unsigned NumReads = 0;
@@ -2326,6 +2359,11 @@ void LoopAccessInfo::analyzeLoop(AAResults *AA, LoopInfo *LI,
       if (HasComplexMemInst)
         continue;
 
+      // Record alias scopes defined inside the loop.
+      if (auto *Decl = dyn_cast<NoAliasScopeDeclInst>(&I))
+        for (Metadata *Op : Decl->getScopeList()->operands())
+          LoopAliasScopes.insert(cast<MDNode>(Op));
+
       // Many math library functions read the rounding mode. We will only
       // vectorize a loop if it contains known function calls that don't set
       // the flag. Therefore, it is safe to ignore this read from memory.
@@ -2407,7 +2445,8 @@ void LoopAccessInfo::analyzeLoop(AAResults *AA, LoopInfo *LI,
   }
 
   MemoryDepChecker::DepCandidates DependentAccesses;
-  AccessAnalysis Accesses(TheLoop, AA, LI, DependentAccesses, *PSE);
+  AccessAnalysis Accesses(TheLoop, AA, LI, DependentAccesses, *PSE,
+                          LoopAliasScopes);
 
   // Holds the analyzed pointers. We don't want to call getUnderlyingObjects
   // multiple times on the same object. If the ptr is accessed twice, once
diff --git a/llvm/test/Analysis/LoopAccessAnalysis/noalias-scope-decl.ll b/llvm/test/Analysis/LoopAccessAnalysis/noalias-scope-decl.ll
index 98bb5f99a40a1e2..fb296f5089422db 100644
--- a/llvm/test/Analysis/LoopAccessAnalysis/noalias-scope-decl.ll
+++ b/llvm/test/Analysis/LoopAccessAnalysis/noalias-scope-decl.ll
@@ -7,8 +7,17 @@
 define void @test_scope_in_loop(ptr %arg, i64 %num) {
 ; CHECK-LABEL: 'test_scope_in_loop'
 ; CHECK-NEXT:    loop:
-; CHECK-NEXT:      Memory dependences are safe
+; CHECK-NEXT:      Report: unsafe dependent memory operations in loop. Use #pragma clang loop distribute(enable) to allow loop distribution to attempt to isolate the offending operations into a separate loop
+; CHECK-NEXT:  Backward loop carried data dependence.
 ; CHECK-NEXT:      Dependences:
+; CHECK-NEXT:        Backward:
+; CHECK-NEXT:            %load.prev = load i8, ptr %prev.ptr, align 1, !alias.scope !0, !noalias !3 ->
+; CHECK-NEXT:            store i8 %add, ptr %cur.ptr, align 1, !alias.scope !3
+; CHECK-EMPTY:
+; CHECK-NEXT:        Forward:
+; CHECK-NEXT:            %load.cur = load i8, ptr %cur.ptr, align 1, !alias.scope !3 ->
+; CHECK-NEXT:            store i8 %add, ptr %cur.ptr, align 1, !alias.scope !3
+; CHECK-EMPTY:
 ; CHECK-NEXT:      Run-time memory checks:
 ; CHECK-NEXT:      Grouped accesses:
 ; CHECK-EMPTY:

[nikic]

I've also added a phase ordering test for this, to make sure it doesn't break once we move to the full restrict representation.

[fhahn]

LGTM, thanks!

[mati865]

Should it be backported to release 18 as a miscompilation fix?