release/20.x: [LAA] Consider accessed addrspace when mapping underlying obj to access. (#129087)

[fhahn]

Backport 275baed

In some cases, it is possible for the same underlying object to be accessed via pointers to different address spaces. This could lead to pointers from different address spaces ending up in the same dependency set, which isn't allowed (and triggers an assertion).

Update the mapping from underlying object -> last access to also include the accessing address space.

Fixes #124759.

PR: #129087

[llvmbot]

@llvm/pr-subscribers-llvm-analysis

Author: Florian Hahn (fhahn)

Changes

Backport 275baed

In some cases, it is possible for the same underlying object to be accessed via pointers to different address spaces. This could lead to pointers from different address spaces ending up in the same dependency set, which isn't allowed (and triggers an assertion).

Update the mapping from underlying object -> last access to also include the accessing address space.

Fixes #124759.

PR: #129087

---

Full diff: https://github.com/llvm/llvm-project/pull/129317.diff

2 Files Affected:

• (modified) llvm/lib/Analysis/LoopAccessAnalysis.cpp (+8-4)
• (added) llvm/test/Analysis/LoopAccessAnalysis/underlying-object-different-address-spaces.ll (+39)

diff --git a/llvm/lib/Analysis/LoopAccessAnalysis.cpp b/llvm/lib/Analysis/LoopAccessAnalysis.cpp
index 697b40403902c..dcd8a910d0f49 100644
--- a/llvm/lib/Analysis/LoopAccessAnalysis.cpp
+++ b/llvm/lib/Analysis/LoopAccessAnalysis.cpp
@@ -1283,8 +1283,10 @@ void AccessAnalysis::processMemAccesses() {
 
     bool SetHasWrite = false;
 
-    // Map of pointers to last access encountered.
-    typedef DenseMap<const Value*, MemAccessInfo> UnderlyingObjToAccessMap;
+    // Map of (pointer to underlying objects, accessed address space) to last
+    // access encountered.
+    typedef DenseMap<std::pair<const Value *, unsigned>, MemAccessInfo>
+        UnderlyingObjToAccessMap;
     UnderlyingObjToAccessMap ObjToLastAccess;
 
     // Set of access to check after all writes have been processed.
@@ -1364,12 +1366,14 @@ void AccessAnalysis::processMemAccesses() {
                     UnderlyingObj->getType()->getPointerAddressSpace()))
               continue;
 
+            unsigned AccessAS = cast<PointerType>(Ptr->getType())->getAddressSpace();
             UnderlyingObjToAccessMap::iterator Prev =
-                ObjToLastAccess.find(UnderlyingObj);
+                ObjToLastAccess.find({UnderlyingObj,AccessAS 
+                 });
             if (Prev != ObjToLastAccess.end())
               DepCands.unionSets(Access, Prev->second);
 
-            ObjToLastAccess[UnderlyingObj] = Access;
+            ObjToLastAccess[{UnderlyingObj, AccessAS}] = Access;
             LLVM_DEBUG(dbgs() << "  " << *UnderlyingObj << "\n");
           }
         }
diff --git a/llvm/test/Analysis/LoopAccessAnalysis/underlying-object-different-address-spaces.ll b/llvm/test/Analysis/LoopAccessAnalysis/underlying-object-different-address-spaces.ll
new file mode 100644
index 0000000000000..adf73c091be00
--- /dev/null
+++ b/llvm/test/Analysis/LoopAccessAnalysis/underlying-object-different-address-spaces.ll
@@ -0,0 +1,39 @@
+; NOTE: Assertions have been autogenerated by utils/update_analyze_test_checks.py UTC_ARGS: --version 5
+; RUN: opt -passes='print<access-info>' -disable-output %s 2>&1 | FileCheck %s
+
+target datalayout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-i128:128-f80:128-n8:16:32:64-S128"
+
+; Test case for https://github.com/llvm/llvm-project/issues/124759. The same
+; underlying object is access through pointers with different address spaces.
+define void @same_underlying_object_different_address_spaces(ptr %dst1.as1, ptr %dst2.as1) {
+; CHECK-LABEL: 'same_underlying_object_different_address_spaces'
+; CHECK-NEXT:    loop:
+; CHECK-NEXT:      Report: cannot identify array bounds
+; CHECK-NEXT:      Dependences:
+; CHECK-NEXT:      Run-time memory checks:
+; CHECK-NEXT:      Grouped accesses:
+; CHECK-EMPTY:
+; CHECK-NEXT:      Non vectorizable stores to invariant address were not found in loop.
+; CHECK-NEXT:      SCEV assumptions:
+; CHECK-EMPTY:
+; CHECK-NEXT:      Expressions re-written:
+;
+entry:
+  %alloc = alloca i8, i64 0, align 128
+  %as3 = addrspacecast ptr %alloc to ptr addrspace(3)
+  %as4 = addrspacecast ptr %alloc to ptr addrspace(4)
+  br label %loop
+
+loop:
+  %iv = phi i64 [ 0, %entry ], [ %iv.next, %loop ]
+  store i32 0, ptr addrspace(4) %as4, align 4
+  store i32 0, ptr %dst1.as1, align 4
+  %l = load i64, ptr addrspace(3) %as3, align 4
+  store i64 %l, ptr %dst2.as1, align 4
+  %iv.next = add i64 %iv, 1
+  %c = icmp eq i64 %iv.next, 100
+  br i1 %c, label %loop, label %exit
+
+exit:
+  ret void
+}

[Meinersbur]

LGTM

[tstellar]

@fhahn Can you rebase this branch?

[tstellar]

Merged: 1515c4a