diff --git a/docs/ReleaseNotes.md b/docs/ReleaseNotes.md
index 375d68d2fa..0582747ab7 100644
--- a/docs/ReleaseNotes.md
+++ b/docs/ReleaseNotes.md
@@ -24,6 +24,8 @@ line upon naming the release. Refer to previous for appropriate section names.
 
 #### Bug Fixes
 
+- Fixed an optimizer crash when scalarizing an out-of-bounds vector access
+  [#8940](https://github.com/microsoft/DirectXShaderCompiler/issues/8940).
 - Fixed derivative operations being moved into divergent control flow, which
   could produce incorrect results
   [#8001](https://github.com/microsoft/DirectXShaderCompiler/issues/8001).
diff --git a/lib/Transforms/Scalar/ScalarReplAggregatesHLSL.cpp b/lib/Transforms/Scalar/ScalarReplAggregatesHLSL.cpp
index d80a678651..11421f5067 100644
--- a/lib/Transforms/Scalar/ScalarReplAggregatesHLSL.cpp
+++ b/lib/Transforms/Scalar/ScalarReplAggregatesHLSL.cpp
@@ -1323,6 +1323,10 @@ void MarkUnsafe(AllocaInfo &I, Instruction *User) {
   DEBUG(dbgs() << "  Transformation preventing inst: " << *User << '\n');
 }
 
+static bool isValidVectorIndex(const ConstantInt *Index, unsigned NumElements) {
+  return !Index->isNegative() && Index->getValue().ult(NumElements);
+}
+
 /// isSafeGEP - Check if a GEP instruction can be handled for scalar
 /// replacement.  It is safe when all the indices are constant, in-bounds
 /// references, and when the resulting offset corresponds to an element within
@@ -1363,6 +1367,8 @@ void isSafeGEP(GetElementPtrInst *GEPI, uint64_t &Offset, AllocaInfo &Info) {
     }
     // Allow dynamic indexing
     ConstantInt *IdxVal = dyn_cast<ConstantInt>(GEPIt.getOperand());
+    if (GEPIt->isVectorTy() && IdxVal && !isValidVectorIndex(IdxVal, arraySize))
+      return MarkUnsafe(Info, GEPI);
     if (!IdxVal) {
       // for dynamic index, use array size - 1 to check the offset
       *indicesIt = Constant::getIntegerValue(
@@ -1386,6 +1392,8 @@ void isSafeGEP(GetElementPtrInst *GEPI, uint64_t &Offset, AllocaInfo &Info) {
     }
     // Allow dynamic indexing
     ConstantInt *IdxVal = dyn_cast<ConstantInt>(GEPIt.getOperand());
+    if (GEPIt->isVectorTy() && IdxVal && !isValidVectorIndex(IdxVal, arraySize))
+      return MarkUnsafe(Info, GEPI);
     if (!IdxVal) {
       // for dynamic index, use array size - 1 to check the offset
       *indicesIt = Constant::getIntegerValue(
@@ -1690,7 +1698,10 @@ bool hasDynamicVectorIndexing(Value *V) {
       for (; GEPIt != E; ++GEPIt) {
         if (isa<VectorType>(*GEPIt)) {
           Value *VecIdx = GEPIt.getOperand();
-          if (!isa<ConstantInt>(VecIdx))
+          ConstantInt *ConstantVecIdx = dyn_cast<ConstantInt>(VecIdx);
+          if (!ConstantVecIdx ||
+              !isValidVectorIndex(ConstantVecIdx,
+                                  GEPIt->getVectorNumElements()))
             return true;
         }
       }
diff --git a/tools/clang/test/DXC/Passes/ScalarReplHLSL/vector-index-bounds.ll b/tools/clang/test/DXC/Passes/ScalarReplHLSL/vector-index-bounds.ll
new file mode 100644
index 0000000000..82fcea760a
--- /dev/null
+++ b/tools/clang/test/DXC/Passes/ScalarReplHLSL/vector-index-bounds.ll
@@ -0,0 +1,77 @@
+; RUN: %dxopt %s -hlsl-passes-resume -scalarrepl-param-hlsl -S | FileCheck %s
+
+; Scalar replacement splits an array of vectors into one scalar array per
+; vector lane. Verify that invalid local lane indices are not scalarized.
+; Constant global GEPs are first canonicalized across the entire aggregate,
+; turning lane 7 into array index 1, lane 3, which can be scalarized safely.
+
+; CHECK-DAG: %valid.3 = alloca [2 x float]
+; CHECK-DAG: %nested = alloca [2 x %struct.S]
+; CHECK-DAG: %out_of_bounds = alloca [2 x <4 x float>]
+; CHECK-DAG: %negative = alloca [2 x <4 x float>]
+; CHECK-DAG: @global_out_of_bounds.3 = internal global [2 x float] zeroinitializer
+; CHECK: getelementptr inbounds [2 x <4 x float>], [2 x <4 x float>]* %out_of_bounds, i32 0, i32 0, i32 7
+; CHECK: getelementptr inbounds [2 x <4 x float>], [2 x <4 x float>]* %negative, i32 0, i32 0, i32 -1
+; CHECK: getelementptr inbounds [2 x float], [2 x float]* %valid.3, i32 0, i32 0
+; CHECK: getelementptr [2 x %struct.S], [2 x %struct.S]* %nested, i32 0, i32 0, i32 0, i32 7
+; CHECK: getelementptr inbounds ([2 x float], [2 x float]* @global_out_of_bounds.3, i32 0, i64 1)
+
+target datalayout = "e-m:e-p:32:32-i1:32-i8:32-i16:32-i32:32-i64:64-f16:32-f32:32-f64:64-n8:16:32:64"
+target triple = "dxil-ms-dx"
+
+%ConstantBuffer = type opaque
+%struct.S = type { <4 x float> }
+
+@"$Globals" = external constant %ConstantBuffer
+@global_out_of_bounds = internal global [2 x <4 x float>] zeroinitializer
+
+define <4 x float> @main() {
+entry:
+  %out_of_bounds = alloca [2 x <4 x float>]
+  %out_of_bounds.vector = getelementptr [2 x <4 x float>], [2 x <4 x float>]* %out_of_bounds, i32 0, i32 0
+  %out_of_bounds.element = getelementptr <4 x float>, <4 x float>* %out_of_bounds.vector, i32 0, i32 7
+  store float 9.000000e+00, float* %out_of_bounds.element
+
+  %negative = alloca [2 x <4 x float>]
+  %negative.vector = getelementptr [2 x <4 x float>], [2 x <4 x float>]* %negative, i32 0, i32 0
+  %negative.element = getelementptr <4 x float>, <4 x float>* %negative.vector, i32 0, i32 -1
+  store float 9.000000e+00, float* %negative.element
+
+  %valid = alloca [2 x <4 x float>]
+  %valid.vector = getelementptr [2 x <4 x float>], [2 x <4 x float>]* %valid, i32 0, i32 0
+  %valid.element = getelementptr <4 x float>, <4 x float>* %valid.vector, i32 0, i32 3
+  store float 9.000000e+00, float* %valid.element
+
+  %nested = alloca [2 x %struct.S]
+  %nested.element = getelementptr [2 x %struct.S], [2 x %struct.S]* %nested, i32 0, i32 0, i32 0, i32 7
+  store float 9.000000e+00, float* %nested.element
+
+  store float 9.000000e+00, float* getelementptr inbounds ([2 x <4 x float>], [2 x <4 x float>]* @global_out_of_bounds, i32 0, i32 0, i32 7)
+
+  ret <4 x float> zeroinitializer
+}
+
+!pauseresume = !{!0}
+!dx.version = !{!1}
+!dx.valver = !{!2}
+!dx.shaderModel = !{!3}
+!dx.typeAnnotations = !{!4}
+!dx.entryPoints = !{!8}
+!dx.fnprops = !{!12}
+!dx.options = !{!13, !14}
+
+!0 = !{!"hlsl-hlemit", !"hlsl-hlensure"}
+!1 = !{i32 1, i32 6}
+!2 = !{i32 1, i32 10}
+!3 = !{!"ps", i32 6, i32 6}
+!4 = !{i32 1, <4 x float> ()* @main, !5}
+!5 = !{!6}
+!6 = !{i32 1, !7, !7}
+!7 = !{}
+!8 = !{<4 x float> ()* @main, !"main", null, !9, null}
+!9 = !{null, null, !10, null}
+!10 = !{!11}
+!11 = !{i32 0, %ConstantBuffer* @"$Globals", !"$Globals", i32 0, i32 -1, i32 1, i32 0, null}
+!12 = !{<4 x float> ()* @main, i32 0, i1 false}
+!13 = !{i32 64}
+!14 = !{i32 -1}