diff --git a/llvm/lib/CodeGen/GlobalISel/GISelValueTracking.cpp b/llvm/lib/CodeGen/GlobalISel/GISelValueTracking.cpp
index c1fb8b6d78ff8..ecba323f8d6bf 100644
--- a/llvm/lib/CodeGen/GlobalISel/GISelValueTracking.cpp
+++ b/llvm/lib/CodeGen/GlobalISel/GISelValueTracking.cpp
@@ -247,6 +247,7 @@ void GISelValueTracking::computeKnownBitsImpl(Register R, KnownBits &Known,
     for (unsigned Idx = 1; Idx < MI.getNumOperands(); Idx += 2) {
       const MachineOperand &Src = MI.getOperand(Idx);
       Register SrcReg = Src.getReg();
+      LLT SrcTy = MRI.getType(SrcReg);
       // Look through trivial copies and phis but don't look through trivial
       // copies or phis of the form `%1:(s32) = OP %0:gpr32`, known-bits
       // analysis is currently unable to determine the bit width of a
@@ -255,9 +256,15 @@ void GISelValueTracking::computeKnownBitsImpl(Register R, KnownBits &Known,
       // We can't use NoSubRegister by name as it's defined by each target but
       // it's always defined to be 0 by tablegen.
       if (SrcReg.isVirtual() && Src.getSubReg() == 0 /*NoSubRegister*/ &&
-          MRI.getType(SrcReg).isValid()) {
+          SrcTy.isValid()) {
+        // In case we're forwarding from a vector register to a non-vector
+        // register we need to update the demanded elements to reflect this
+        // before recursing.
+        APInt NowDemandedElts = SrcTy.isFixedVector() && !DstTy.isFixedVector()
+                                    ? APInt::getAllOnes(SrcTy.getNumElements())
+                                    : DemandedElts; // Known to be APInt(1, 1)
         // For COPYs we don't do anything, don't increase the depth.
-        computeKnownBitsImpl(SrcReg, Known2, DemandedElts,
+        computeKnownBitsImpl(SrcReg, Known2, NowDemandedElts,
                              Depth + (Opcode != TargetOpcode::COPY));
         Known2 = Known2.anyextOrTrunc(BitWidth);
         Known = Known.intersectWith(Known2);
diff --git a/llvm/test/CodeGen/AArch64/GlobalISel/knownbits-copy-vector-crash.ll b/llvm/test/CodeGen/AArch64/GlobalISel/knownbits-copy-vector-crash.ll
new file mode 100644
index 0000000000000..f15253682c336
--- /dev/null
+++ b/llvm/test/CodeGen/AArch64/GlobalISel/knownbits-copy-vector-crash.ll
@@ -0,0 +1,56 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 6
+; RUN: llc -global-isel -o - %s | FileCheck %s
+
+target triple = "aarch64-unknown-unknown"
+
+; Check we don't crash here when computing known bits.
+
+define <4 x i32> @test(<8 x i16> %in, i1 %continue) {
+; CHECK-LABEL: test:
+; CHECK:       // %bb.0: // %entry
+; CHECK-NEXT:    sub sp, sp, #16
+; CHECK-NEXT:    .cfi_def_cfa_offset 16
+; CHECK-NEXT:    mov w12, wzr
+; CHECK-NEXT:    mov x8, sp
+; CHECK-NEXT:    mov w9, #2 // =0x2
+; CHECK-NEXT:    mov w10, #0 // =0x0
+; CHECK-NEXT:  .LBB0_1: // %loop
+; CHECK-NEXT:    // =>This Inner Loop Header: Depth=1
+; CHECK-NEXT:    mov w11, w12
+; CHECK-NEXT:    mov w12, w12
+; CHECK-NEXT:    str q0, [sp]
+; CHECK-NEXT:    and x12, x12, #0x7
+; CHECK-NEXT:    umull x12, w12, w9
+; CHECK-NEXT:    ldrb w12, [x8, x12]
+; CHECK-NEXT:    cmp w12, #0
+; CHECK-NEXT:    cset w12, eq
+; CHECK-NEXT:    fmov s1, w12
+; CHECK-NEXT:    mov v1.b[1], w10
+; CHECK-NEXT:    mov v1.b[2], w10
+; CHECK-NEXT:    mov v1.b[3], w10
+; CHECK-NEXT:    fmov w12, s1
+; CHECK-NEXT:    tbz w0, #0, .LBB0_1
+; CHECK-NEXT:  // %bb.2: // %exit
+; CHECK-NEXT:    fmov s0, w11
+; CHECK-NEXT:    mov v0.s[1], wzr
+; CHECK-NEXT:    mov v0.s[2], wzr
+; CHECK-NEXT:    mov v0.s[3], wzr
+; CHECK-NEXT:    add sp, sp, #16
+; CHECK-NEXT:    ret
+entry:
+  br label %loop
+
+exit:
+  %result = insertelement <4 x i32> zeroinitializer, i32 %index, i64 0
+  ret <4 x i32> %result
+
+loop:
+  %index = phi i32 [ 0, %entry ], [ %insert.bitcast, %loop ]
+  %extracted = extractelement <8 x i16> %in, i32 %index
+  %masked = and i16 %extracted, 255
+  %maskedIsZero = icmp eq i16 %masked, 0
+  %maskedIsZero.zext = zext i1 %maskedIsZero to i8
+  %insert = insertelement <4 x i8> zeroinitializer, i8 %maskedIsZero.zext, i64 0
+  %insert.bitcast = bitcast <4 x i8> %insert to i32
+  br i1 %continue, label %exit, label %loop
+}