Index: llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
===================================================================
--- llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -12861,6 +12861,20 @@
   bool STCoversLD =
       (Offset >= 0) &&
       (Offset * 8 + LDMemType.getSizeInBits() <= STMemType.getSizeInBits());
+
+  auto ReplaceLd = [&](LoadSDNode *LD, SDValue Val, SDValue Chain) -> SDValue {
+    if (LD->isIndexed()) {
+      bool IsSub = (LD->getAddressingMode() == ISD::PRE_DEC ||
+                    LD->getAddressingMode() == ISD::POST_DEC);
+      unsigned Opc = IsSub ? ISD::SUB : ISD::ADD;
+      SDValue Idx = DAG.getNode(Opc, SDLoc(LD), LD->getOperand(1).getValueType(),
+                             LD->getOperand(1), LD->getOperand(2));
+      SDValue Ops[] = {Val, Idx, Chain};
+      return CombineTo(LD, Ops, 3);
+    }
+    return CombineTo(LD, Val, Chain);
+  };
+
   if (!STCoversLD)
     return SDValue();
 
@@ -12868,7 +12882,7 @@
   if (Offset == 0 && LDType == STType && STMemType == LDMemType) {
     // Simple case: Direct non-truncating forwarding
     if (LDType.getSizeInBits() == LDMemType.getSizeInBits())
-      return CombineTo(LD, ST->getValue(), Chain);
+      return ReplaceLd(LD, ST->getValue(), Chain);
     // Can we model the truncate and extension with an and mask?
     if (STType.isInteger() && LDMemType.isInteger() && !STType.isVector() &&
         !LDMemType.isVector() && LD->getExtensionType() != ISD::SEXTLOAD) {
@@ -12878,7 +12892,7 @@
                                                STMemType.getSizeInBits()),
                           SDLoc(ST), STType);
       auto Val = DAG.getNode(ISD::AND, SDLoc(LD), LDType, ST->getValue(), Mask);
-      return CombineTo(LD, Val, Chain);
+      return ReplaceLd(LD, Val, Chain);
     }
   }
 
@@ -12903,7 +12917,7 @@
     }
     if (!extendLoadedValueToExtension(LD, Val))
       continue;
-    return CombineTo(LD, Val, Chain);
+    return ReplaceLd(LD, Val, Chain);
   } while (false);
 
   // On failure, cleanup dead nodes we may have created.
Index: llvm/test/CodeGen/ARM/pr39571.ll
===================================================================
--- /dev/null
+++ llvm/test/CodeGen/ARM/pr39571.ll
@@ -0,0 +1,33 @@
+; RUN: llc < %s -mtriple armv4t-unknown-linux-gnueabi -mattr=+strict-align
+
+; Avoid crash from forwarding indexed-loads back to store.
+%struct.anon = type { %struct.ma*, %struct.mb }
+%struct.ma = type { i8 }
+%struct.mb = type { i8, i8 }
+%struct.anon.0 = type { %struct.anon.1 }
+%struct.anon.1 = type { %struct.ds }
+%struct.ds = type <{ i8, %union.ie }>
+%union.ie = type { %struct.ib }
+%struct.ib = type { i8, i8, i16 }
+
+@a = common dso_local local_unnamed_addr global %struct.anon* null, align 4
+@b = common dso_local local_unnamed_addr global %struct.anon.0 zeroinitializer, align 1
+
+; Function Attrs: norecurse nounwind
+define dso_local void @func() local_unnamed_addr {
+entry:
+  %0 = load %struct.anon*, %struct.anon** @a, align 4
+  %ad = getelementptr inbounds %struct.anon, %struct.anon* %0, i32 0, i32 0
+  %1 = load %struct.ma*, %struct.ma** %ad, align 4
+  %c.sroa.0.0..sroa_idx = getelementptr inbounds %struct.ma, %struct.ma* %1, i32 0, i32 0
+  %c.sroa.0.0.copyload = load i8, i8* %c.sroa.0.0..sroa_idx, align 1
+  %cb = getelementptr inbounds %struct.anon, %struct.anon* %0, i32 0, i32 1
+  %band = getelementptr inbounds %struct.anon, %struct.anon* %0, i32 0, i32 1, i32 1
+  store i8 %c.sroa.0.0.copyload, i8* %band, align 4
+  store i8 6, i8* getelementptr inbounds (%struct.anon.0, %struct.anon.0* @b, i32 0, i32 0, i32 0, i32 1, i32 0, i32 0), align 1
+  store i8 2, i8* getelementptr inbounds (%struct.anon.0, %struct.anon.0* @b, i32 0, i32 0, i32 0, i32 1, i32 0, i32 1), align 1
+  %2 = bitcast %struct.mb* %cb to i32*
+  %3 = load i32, i32* bitcast (i8* getelementptr inbounds (%struct.anon.0, %struct.anon.0* @b, i32 0, i32 0, i32 0, i32 1, i32 0, i32 0) to i32*), align 1
+  store i32 %3, i32* %2, align 1
+  ret void
+}