[Attributor][FIX] Ensure we use the right AAExecutionDomain

Before we might have ended up queriying the AAExecutionDomain of a
different function, which resulted in wrong optimistic results.

Partially fixes #60425

(cherry picked from commit 18a2975)

Author: jdoerfert

llvm/lib/Transforms/IPO/AttributorAttributes.cpp

@@ -1043,12 +1043,14 @@ struct AAPointerInfoImpl
     const auto &NoSyncAA = A.getAAFor<AANoSync>(
         QueryingAA, IRPosition::function(Scope), DepClassTy::OPTIONAL);
     const auto *ExecDomainAA = A.lookupAAFor<AAExecutionDomain>(
-        IRPosition::function(Scope), &QueryingAA, DepClassTy::OPTIONAL);
+        IRPosition::function(Scope), &QueryingAA, DepClassTy::NONE);
     bool AllInSameNoSyncFn = NoSyncAA.isAssumedNoSync();
     bool InstIsExecutedByInitialThreadOnly =
         ExecDomainAA && ExecDomainAA->isExecutedByInitialThreadOnly(I);
     bool InstIsExecutedInAlignedRegion =
         ExecDomainAA && ExecDomainAA->isExecutedInAlignedRegion(A, I);
+    if (InstIsExecutedInAlignedRegion || InstIsExecutedByInitialThreadOnly)
+      A.recordDependence(*ExecDomainAA, QueryingAA, DepClassTy::OPTIONAL);
 
     InformationCache &InfoCache = A.getInfoCache();
     bool IsThreadLocalObj =
@@ -1063,14 +1065,24 @@ struct AAPointerInfoImpl
     auto CanIgnoreThreadingForInst = [&](const Instruction &I) -> bool {
       if (IsThreadLocalObj || AllInSameNoSyncFn)
         return true;
-      if (!ExecDomainAA)
+      const auto *FnExecDomainAA =
+          I.getFunction() == &Scope
+              ? ExecDomainAA
+              : A.lookupAAFor<AAExecutionDomain>(
+                    IRPosition::function(*I.getFunction()), &QueryingAA,
+                    DepClassTy::NONE);
+      if (!FnExecDomainAA)
         return false;
       if (InstIsExecutedInAlignedRegion ||
-          ExecDomainAA->isExecutedInAlignedRegion(A, I))
+          FnExecDomainAA->isExecutedInAlignedRegion(A, I)) {
+        A.recordDependence(*FnExecDomainAA, QueryingAA, DepClassTy::OPTIONAL);
         return true;
+      }
       if (InstIsExecutedByInitialThreadOnly &&
-          ExecDomainAA->isExecutedByInitialThreadOnly(I))
+          FnExecDomainAA->isExecutedByInitialThreadOnly(I)) {
+        A.recordDependence(*FnExecDomainAA, QueryingAA, DepClassTy::OPTIONAL);
         return true;
+      }
       return false;
     };
 

llvm/lib/Transforms/IPO/OpenMPOpt.cpp

@@ -2680,6 +2680,8 @@ struct AAExecutionDomainFunction : public AAExecutionDomain {
 
   bool isExecutedInAlignedRegion(Attributor &A,
                                  const Instruction &I) const override {
+    assert(I.getFunction() == getAnchorScope() &&
+           "Instruction is out of scope!");
     if (!isValidState() || isa<CallBase>(I))
       return false;
 

llvm/test/Transforms/Attributor/reduced/aa_execution_domain_wrong_fn.ll

@@ -0,0 +1,46 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --function-signature --scrub-attributes --check-attributes --check-globals --include-generated-funcs
+; RUN: opt -passes=openmp-opt -S < %s | FileCheck %s --check-prefixes=CHECK
+
+%"struct.ompx::state::TeamStateTy" = type { %"struct.ompx::state::ICVStateTy", i32, i32, ptr }
+%"struct.ompx::state::ICVStateTy" = type { i32, i32, i32, i32, i32, i32 }
+
+@_ZN4ompx5state9TeamStateE = internal addrspace(3) global %"struct.ompx::state::TeamStateTy" undef
+
+define weak_odr amdgpu_kernel void @__omp_offloading_16_1d1156__Z38test_target_teams_distribute__parallelv_l16() {
+  %1 = tail call i32 @__kmpc_target_init(ptr null, i8 0, i1 false)
+  ret void
+}
+
+define internal i32 @__kmpc_target_init(ptr %0, i8 %1, i1 %2) {
+  store <2 x i32> zeroinitializer, ptr addrspace(3) @_ZN4ompx5state9TeamStateE, align 16
+  %4 = call i1 @__kmpc_kernel_parallel()
+  ret i32 0
+}
+
+define internal i1 @__kmpc_kernel_parallel() {
+  %1 = load ptr, ptr addrspace(3) @_ZN4ompx5state9TeamStateE, align 8
+  ret i1 false
+}
+
+!llvm.module.flags = !{!0}
+
+!0 = !{i32 7, !"openmp", i32 50}
+;.
+; CHECK: @[[_ZN4OMPX5STATE9TEAMSTATEE:[a-zA-Z0-9_$"\\.-]+]] = internal addrspace(3) global %"struct.ompx::state::TeamStateTy" undef
+;.
+; CHECK-LABEL: define {{[^@]+}}@__omp_offloading_16_1d1156__Z38test_target_teams_distribute__parallelv_l16() {
+; CHECK-NEXT:    [[TMP1:%.*]] = tail call i32 @__kmpc_target_init(ptr null, i8 0, i1 false)
+; CHECK-NEXT:    ret void
+;
+;
+; CHECK: Function Attrs: norecurse nosync nounwind memory(write)
+; CHECK-LABEL: define {{[^@]+}}@__kmpc_target_init
+; CHECK-SAME: (ptr [[TMP0:%.*]], i8 [[TMP1:%.*]], i1 [[TMP2:%.*]]) #[[ATTR0:[0-9]+]] {
+; CHECK-NEXT:    ret i32 0
+;
+;.
+; CHECK: attributes #[[ATTR0]] = { norecurse nosync nounwind memory(write) }
+; CHECK: attributes #[[ATTR1:[0-9]+]] = { nosync nounwind }
+;.
+; CHECK: [[META0:![0-9]+]] = !{i32 7, !"openmp", i32 50}
+;.