[SelectionDAG] Mark frame index as "aliased" at argument copy elison (#89712)

This is a fix for miscompiles reported in
  https://github.com/llvm/llvm-project/issues/89060

After argument copy elison the IR value for the eliminated alloca
is aliasing with the fixed stack object. This patch is making sure
that we mark the fixed stack object as being aliased with IR values
to avoid that for example schedulers are reordering accesses to
the fixed stack object. This could otherwise happen when there is a
mix of MemOperands refering the shared fixed stack slow via both
the IR value for the elided alloca, and via a fixed stack pseudo
source value (as would be the case when lowering the arguments).

(cherry picked from commit d8b253be56b3e9073b3e59123cf2da0bcde20c63)

diff --git a/llvm/include/llvm/CodeGen/MachineFrameInfo.h b/llvm/include/llvm/CodeGen/MachineFrameInfo.h
index 7d11d63d4066f4a500b899e092abaf8bd22e0247..c35faac09c4d939d5bad91b928d59a4a173242e9 100644 (file)
--- a/llvm/include/llvm/CodeGen/MachineFrameInfo.h
+++ b/llvm/include/llvm/CodeGen/MachineFrameInfo.h
@@ -697,6 +697,13 @@ public:
     return Objects[ObjectIdx+NumFixedObjects].isAliased;
   }
 
+  /// Set "maybe pointed to by an LLVM IR value" for an object.
+  void setIsAliasedObjectIndex(int ObjectIdx, bool IsAliased) {
+    assert(unsigned(ObjectIdx+NumFixedObjects) < Objects.size() &&
+           "Invalid Object Idx!");
+    Objects[ObjectIdx+NumFixedObjects].isAliased = IsAliased;
+  }
+
   /// Returns true if the specified index corresponds to an immutable object.
   bool isImmutableObjectIndex(int ObjectIdx) const {
     // Tail calling functions can clobber their function arguments.

diff --git a/llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp b/llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp
index 5ce1013f30fd1b8e12f9f369b97a85e70c6e26f2..7406a8ac1611de0fab36cd87e11063eacc660062 100644 (file)
--- a/llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp
@@ -10888,7 +10888,7 @@ static void tryToElideArgumentCopy(
   }
 
   // Perform the elision. Delete the old stack object and replace its only use
-  // in the variable info map. Mark the stack object as mutable.
+  // in the variable info map. Mark the stack object as mutable and aliased.
   LLVM_DEBUG({
     dbgs() << "Eliding argument copy from " << Arg << " to " << *AI << '\n'
            << "  Replacing frame index " << OldIndex << " with " << FixedIndex
@@ -10896,6 +10896,7 @@ static void tryToElideArgumentCopy(
   });
   MFI.RemoveStackObject(OldIndex);
   MFI.setIsImmutableObjectIndex(FixedIndex, false);
+  MFI.setIsAliasedObjectIndex(FixedIndex, true);
   AllocaIndex = FixedIndex;
   ArgCopyElisionFrameIndexMap.insert({OldIndex, FixedIndex});
   for (SDValue ArgVal : ArgVals)

diff --git a/llvm/test/CodeGen/Hexagon/arg-copy-elison.ll b/llvm/test/CodeGen/Hexagon/arg-copy-elison.ll
new file mode 100644 (file)
index 0000000..f0c30c3
--- /dev/null
+++ b/llvm/test/CodeGen/Hexagon/arg-copy-elison.ll
@@ -0,0 +1,39 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 4
+; RUN: llc -mtriple hexagon-- -o - %s | FileCheck %s
+
+; Reproducer for https://github.com/llvm/llvm-project/issues/89060
+;
+; Problem was a bug in argument copy elison. Given that the %alloca is
+; eliminated, the same frame index will be used for accessing %alloca and %a
+; on the fixed stack. Care must be taken when setting up
+; MachinePointerInfo/MemOperands for those accesses to either make sure that
+; we always refer to the fixed stack slot the same way (not using the
+; ir.alloca name), or make sure that we still detect that they alias each
+; other if using different kinds of MemOperands to identify the same fixed
+; stack entry.
+;
+define i32 @f(i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32, i32 %q1, i32 %a, i32 %q2) {
+; CHECK-LABEL: f:
+; CHECK:         .cfi_startproc
+; CHECK-NEXT:  // %bb.0:
+; CHECK-NEXT:    {
+; CHECK-NEXT:     r0 = memw(r29+#36)
+; CHECK-NEXT:     r1 = memw(r29+#28)
+; CHECK-NEXT:    }
+; CHECK-NEXT:    {
+; CHECK-NEXT:     r0 = sub(r1,r0)
+; CHECK-NEXT:     r2 = memw(r29+#32)
+; CHECK-NEXT:     memw(r29+#32) = ##666
+; CHECK-NEXT:    }
+; CHECK-NEXT:    {
+; CHECK-NEXT:     r0 = xor(r0,r2)
+; CHECK-NEXT:     jumpr r31
+; CHECK-NEXT:    }
+  %alloca = alloca i32
+  store i32 %a, ptr %alloca     ; Should be elided.
+  store i32 666, ptr %alloca
+  %x = sub i32 %q1, %q2
+  %y = xor i32 %x, %a           ; Results in a load of %a from fixed stack.
+                                ; Using same frame index as elided %alloca.
+  ret i32 %y
+}