Index: llvm/lib/Transforms/Scalar/Scalarizer.cpp
===================================================================
--- llvm/lib/Transforms/Scalar/Scalarizer.cpp
+++ llvm/lib/Transforms/Scalar/Scalarizer.cpp
@@ -762,6 +762,12 @@
       Res[I] = Builder.CreateBitCast(Op0[I], DstVT->getElementType(),
                                      BCI.getName() + ".i" + Twine(I));
   } else if (DstNumElems > SrcNumElems) {
+    // Abort and do nothing if we don't have the special case
+    // <M x t1> -> <N*M x t2>.
+    unsigned N = DstNumElems / SrcNumElems;
+    if (DstNumElems != SrcNumElems * N)
+      return false;
+
     // <M x t1> -> <N*M x t2>.  Convert each t1 to <N x t2> and copy the
     // individual elements to the destination.
     unsigned FanOut = DstNumElems / SrcNumElems;
@@ -781,6 +787,12 @@
         Res[ResI++] = Mid[MidI];
     }
   } else {
+    // Abort and do nothing if we don't have the special case
+    // <N*M x t1> -> <M x t2>.
+    unsigned N = SrcNumElems / DstNumElems;
+    if (SrcNumElems != DstNumElems * N)
+      return false;
+
     // <N*M x t1> -> <M x t2>.  Convert each group of <N x t1> into a t2.
     unsigned FanIn = SrcNumElems / DstNumElems;
     auto *MidTy = FixedVectorType::get(SrcVT->getElementType(), FanIn);
Index: llvm/test/Transforms/Scalarizer/bitcast.ll
===================================================================
--- /dev/null
+++ llvm/test/Transforms/Scalarizer/bitcast.ll
@@ -0,0 +1,104 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 2
+; RUN: opt -passes=scalarizer %s -o - -S | FileCheck %s
+
+; REQUIRES: asserts
+
+; Scalarize bitcasts of vectors to the same, larger and smaller number of
+; elements.
+; In some cases we can indeed scalarize, but in some cases we just leave the
+; bitcast be. In any case we should not crash, which used to happen for the
+; 3 to 4 and 4 to 3 cases.
+
+@g = global [12 x i8] zeroinitializer
+
+; Same number of elements
+
+define void @bitcast_3_to_3_elements() {
+; CHECK-LABEL: define void @bitcast_3_to_3_elements() {
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[VEC:%.*]] = load <3 x i32>, ptr @g, align 16
+; CHECK-NEXT:    [[VEC_I0:%.*]] = extractelement <3 x i32> [[VEC]], i64 0
+; CHECK-NEXT:    [[BC_I0:%.*]] = bitcast i32 [[VEC_I0]] to float
+; CHECK-NEXT:    [[VEC_I1:%.*]] = extractelement <3 x i32> [[VEC]], i64 1
+; CHECK-NEXT:    [[BC_I1:%.*]] = bitcast i32 [[VEC_I1]] to float
+; CHECK-NEXT:    [[VEC_I2:%.*]] = extractelement <3 x i32> [[VEC]], i64 2
+; CHECK-NEXT:    [[BC_I2:%.*]] = bitcast i32 [[VEC_I2]] to float
+; CHECK-NEXT:    ret void
+;
+entry:
+  %vec = load <3 x i32>, ptr @g
+  %bc = bitcast <3 x i32> %vec to <3 x float>
+  ret void
+}
+
+; Larger number of elements
+
+define void @bitcast_2_to_4_elements() {
+; CHECK-LABEL: define void @bitcast_2_to_4_elements() {
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[VEC:%.*]] = load <2 x i48>, ptr @g, align 16
+; CHECK-NEXT:    [[VEC_I0:%.*]] = extractelement <2 x i48> [[VEC]], i64 0
+; CHECK-NEXT:    [[VEC_I0_CAST:%.*]] = bitcast i48 [[VEC_I0]] to <2 x i24>
+; CHECK-NEXT:    [[VEC_I0_CAST_I0:%.*]] = extractelement <2 x i24> [[VEC_I0_CAST]], i64 0
+; CHECK-NEXT:    [[VEC_I0_CAST_I1:%.*]] = extractelement <2 x i24> [[VEC_I0_CAST]], i64 1
+; CHECK-NEXT:    [[VEC_I1:%.*]] = extractelement <2 x i48> [[VEC]], i64 1
+; CHECK-NEXT:    [[VEC_I1_CAST:%.*]] = bitcast i48 [[VEC_I1]] to <2 x i24>
+; CHECK-NEXT:    [[VEC_I1_CAST_I0:%.*]] = extractelement <2 x i24> [[VEC_I1_CAST]], i64 0
+; CHECK-NEXT:    [[VEC_I1_CAST_I1:%.*]] = extractelement <2 x i24> [[VEC_I1_CAST]], i64 1
+; CHECK-NEXT:    ret void
+;
+entry:
+  %vec = load <2 x i48>, ptr @g
+  %bc = bitcast <2 x i48> %vec to <4 x i24>
+  ret void
+}
+
+define void @bitcast_3_to_4_elements() {
+; CHECK-LABEL: define void @bitcast_3_to_4_elements() {
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[VEC:%.*]] = load <3 x i32>, ptr @g, align 16
+; CHECK-NEXT:    [[BC:%.*]] = bitcast <3 x i32> [[VEC]] to <4 x i24>
+; CHECK-NEXT:    ret void
+;
+entry:
+  %vec = load <3 x i32>, ptr @g
+  %bc = bitcast <3 x i32> %vec to <4x i24>
+  ret void
+}
+
+; Smaller number of elements
+
+define void @bitcast_4_to_2_elements() {
+; CHECK-LABEL: define void @bitcast_4_to_2_elements() {
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[VEC:%.*]] = load <4 x i24>, ptr @g, align 16
+; CHECK-NEXT:    [[VEC_I0:%.*]] = extractelement <4 x i24> [[VEC]], i64 0
+; CHECK-NEXT:    [[BC_I0_UPTO0:%.*]] = insertelement <2 x i24> poison, i24 [[VEC_I0]], i64 0
+; CHECK-NEXT:    [[VEC_I1:%.*]] = extractelement <4 x i24> [[VEC]], i64 1
+; CHECK-NEXT:    [[BC_I0_UPTO1:%.*]] = insertelement <2 x i24> [[BC_I0_UPTO0]], i24 [[VEC_I1]], i64 1
+; CHECK-NEXT:    [[BC_I0:%.*]] = bitcast <2 x i24> [[BC_I0_UPTO1]] to i48
+; CHECK-NEXT:    [[VEC_I2:%.*]] = extractelement <4 x i24> [[VEC]], i64 2
+; CHECK-NEXT:    [[BC_I1_UPTO0:%.*]] = insertelement <2 x i24> poison, i24 [[VEC_I2]], i64 0
+; CHECK-NEXT:    [[VEC_I3:%.*]] = extractelement <4 x i24> [[VEC]], i64 3
+; CHECK-NEXT:    [[BC_I1_UPTO1:%.*]] = insertelement <2 x i24> [[BC_I1_UPTO0]], i24 [[VEC_I3]], i64 1
+; CHECK-NEXT:    [[BC_I1:%.*]] = bitcast <2 x i24> [[BC_I1_UPTO1]] to i48
+; CHECK-NEXT:    ret void
+;
+entry:
+  %vec = load <4 x i24>, ptr @g
+  %bc = bitcast <4 x i24> %vec to <2 x i48>
+  ret void
+}
+
+define void @bitcast_4_to_3_elements() {
+; CHECK-LABEL: define void @bitcast_4_to_3_elements() {
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[VEC:%.*]] = load <4 x i24>, ptr @g, align 16
+; CHECK-NEXT:    [[BC:%.*]] = bitcast <4 x i24> [[VEC]] to <3 x i32>
+; CHECK-NEXT:    ret void
+;
+entry:
+  %vec = load <4 x i24>, ptr @g
+  %bc = bitcast <4 x i24> %vec to <3 x i32>
+  ret void
+}