Index: include/llvm/Target/TargetLowering.h
===================================================================
--- include/llvm/Target/TargetLowering.h
+++ include/llvm/Target/TargetLowering.h
@@ -271,6 +271,15 @@
     return true;
   }
 
+  /// isStoreBitCastBeneficial() - Mirror of isLoadBitCastBeneficial(). Return
+  /// true if the following transform is beneficial.
+  ///
+  /// (store (y (conv x)), y*)) -> (store x, (x*))
+  virtual bool isStoreBitCastBeneficial(EVT StoreVT, EVT BitcastVT) const {
+    // Default to the same logic as stores.
+    return isLoadBitCastBeneficial(StoreVT, BitcastVT);
+  }
+
   /// Return true if it is expected to be cheaper to do a store of a non-zero
   /// vector constant with the given size and type for the address space than to
   /// store the individual scalar element constants.
Index: lib/CodeGen/SelectionDAG/DAGCombiner.cpp
===================================================================
--- lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -11911,17 +11911,21 @@
   // resultant store does not need a higher alignment than the original.
   if (Value.getOpcode() == ISD::BITCAST && !ST->isTruncatingStore() &&
       ST->isUnindexed()) {
-    unsigned OrigAlign = ST->getAlignment();
     EVT SVT = Value.getOperand(0).getValueType();
-    unsigned Align = DAG.getDataLayout().getABITypeAlignment(
-        SVT.getTypeForEVT(*DAG.getContext()));
-    if (Align <= OrigAlign &&
-        ((!LegalOperations && !ST->isVolatile()) ||
-         TLI.isOperationLegalOrCustom(ISD::STORE, SVT)))
-      return DAG.getStore(Chain, SDLoc(N), Value.getOperand(0),
-                          Ptr, ST->getPointerInfo(), ST->isVolatile(),
-                          ST->isNonTemporal(), OrigAlign,
-                          ST->getAAInfo());
+    if (((!LegalOperations && !ST->isVolatile()) ||
+         TLI.isOperationLegalOrCustom(ISD::STORE, SVT)) &&
+        TLI.isStoreBitCastBeneficial(Value.getValueType(), SVT)) {
+      unsigned OrigAlign = ST->getAlignment();
+      bool Fast = false;
+      if (TLI.allowsMemoryAccess(*DAG.getContext(), DAG.getDataLayout(), SVT,
+                                 ST->getAddressSpace(), OrigAlign, &Fast) &&
+          Fast) {
+        return DAG.getStore(Chain, SDLoc(N), Value.getOperand(0),
+                            Ptr, ST->getPointerInfo(), ST->isVolatile(),
+                            ST->isNonTemporal(), OrigAlign,
+                            ST->getAAInfo());
+      }
+    }
   }
 
   // Turn 'store undef, Ptr' -> nothing.
Index: test/CodeGen/AMDGPU/reduce-store-width-alignment.ll
===================================================================
--- /dev/null
+++ test/CodeGen/AMDGPU/reduce-store-width-alignment.ll
@@ -0,0 +1,53 @@
+; RUN: llc -march=amdgcn -verify-machineinstrs < %s | FileCheck -check-prefix=GCN %s
+
+; GCN-LABEL: {{^}}store_v2i32_as_v4i16_align_4:
+; GCN: s_load_dwordx2
+; GCN: ds_write2_b32 v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+}} offset1:1{{$}}
+define void @store_v2i32_as_v4i16_align_4(<4 x i16> addrspace(3)* align 4 %out, <2 x i32> %x) #0 {
+  %x.bc = bitcast <2 x i32> %x to <4 x i16>
+  store <4 x i16> %x.bc, <4 x i16> addrspace(3)* %out, align 4
+  ret void
+}
+
+; GCN-LABEL: {{^}}store_v4i32_as_v8i16_align_4:
+; GCN: s_load_dwordx4
+; GCN-DAG: ds_write2_b32 v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+}} offset0:2 offset1:3
+; GCN-DAG: ds_write2_b32 v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+}} offset1:1{{$}}
+define void @store_v4i32_as_v8i16_align_4(<8 x i16> addrspace(3)* align 4 %out, <4 x i32> %x) #0 {
+  %x.bc = bitcast <4 x i32> %x to <8 x i16>
+  store <8 x i16> %x.bc, <8 x i16> addrspace(3)* %out, align 4
+  ret void
+}
+
+; GCN-LABEL: {{^}}store_v2i32_as_i64_align_4:
+; GCN: s_load_dwordx2
+; GCN: ds_write2_b32 v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+}} offset1:1{{$}}
+define void @store_v2i32_as_i64_align_4(<4 x i16> addrspace(3)* align 4 %out, <2 x i32> %x) #0 {
+  %x.bc = bitcast <2 x i32> %x to <4 x i16>
+  store <4 x i16> %x.bc, <4 x i16> addrspace(3)* %out, align 4
+  ret void
+}
+
+; GCN-LABEL: {{^}}store_v4i32_as_v2i64_align_4:
+; GCN: s_load_dwordx4
+; GCN-DAG: ds_write2_b32 v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+}} offset0:2 offset1:3
+; GCN-DAG: ds_write2_b32 v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+}} offset1:1{{$}}
+define void @store_v4i32_as_v2i64_align_4(<2 x i64> addrspace(3)* align 4 %out, <4 x i32> %x) #0 {
+  %x.bc = bitcast <4 x i32> %x to <2 x i64>
+  store <2 x i64> %x.bc, <2 x i64> addrspace(3)* %out, align 4
+  ret void
+}
+
+; GCN-LABEL: {{^}}store_v4i16_as_v2i32_align_4:
+; GCN: buffer_load_ushort
+; GCN: buffer_load_ushort
+; GCN: buffer_load_ushort
+; GCN: buffer_load_ushort
+; GCN: ds_write2_b32 v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+}} offset1:1{{$}}
+define void @store_v4i16_as_v2i32_align_4(<2 x i32> addrspace(3)* align 4 %out, <4 x i16> %x) #0 {
+  %x.bc = bitcast <4 x i16> %x to <2 x i32>
+  store <2 x i32> %x.bc, <2 x i32> addrspace(3)* %out, align 4
+  ret void
+}
+
+attributes #0 = { nounwind }
Index: test/CodeGen/ARM/vector-store.ll
===================================================================
--- test/CodeGen/ARM/vector-store.ll
+++ test/CodeGen/ARM/vector-store.ll
@@ -12,12 +12,14 @@
 }
 
 define void @store_v8i8_update(<8 x i8>** %ptr, <8 x i8> %val) {
-;CHECK-LABEL: store_v8i8_update:
-;CHECK: vst1.8 {{{d[0-9]+}}}, [{{r[0-9]+}}]!
+; CHECK-LABEL: store_v8i8_update:
+; CHECK: str r{{[0-9]+}}, [{{r[0-9]+}}]
+; CHECK: str r{{[0-9]+}}, [{{r[0-9]+}}, #4]
+; CHECK: add.w r{{[0-9]+}}, r{{[0-9]+}}, #8
 	%A = load <8 x i8>*, <8 x i8>** %ptr
 	store  <8 x i8> %val, <8 x i8>* %A, align 1
 	%inc = getelementptr <8 x i8>, <8 x i8>* %A, i38 1
-        store <8 x i8>* %inc, <8 x i8>** %ptr
+    store <8 x i8>* %inc, <8 x i8>** %ptr
 	ret void
 }
 
@@ -30,8 +32,10 @@
 }
 
 define void @store_v4i16_update(<4 x i16>** %ptr, <4 x i16> %val) {
-;CHECK-LABEL: store_v4i16_update:
-;CHECK: vst1.8 {{{d[0-9]+}}}, [{{r[0-9]+}}]!
+; CHECK-LABEL: store_v4i16_update:
+; CHECK: str r{{[0-9]+}}, [{{r[0-9]+}}]
+; CHECK: str r{{[0-9]+}}, [{{r[0-9]+}}, #4]
+; CHECK: add.w r{{[0-9]+}}, r{{[0-9]+}}, #8
 	%A = load <4 x i16>*, <4 x i16>** %ptr
 	store  <4 x i16> %val, <4 x i16>* %A, align 1
 	%inc = getelementptr <4 x i16>, <4 x i16>* %A, i34 1
@@ -48,8 +52,11 @@
 }
 
 define void @store_v2i32_update(<2 x i32>** %ptr, <2 x i32> %val) {
-;CHECK-LABEL: store_v2i32_update:
-;CHECK: vst1.8 {{{d[0-9]+}}}, [{{r[0-9]+}}]!
+; CHECK-LABEL: store_v2i32_update:
+; CHECK: str r{{[0-9]+}}, [{{r[0-9]+}}]
+; CHECK: str r{{[0-9]+}}, [{{r[0-9]+}}, #4]
+; CHECK: add.w r{{[0-9]+}}, r{{[0-9]+}}, #8
+
 	%A = load <2 x i32>*, <2 x i32>** %ptr
 	store  <2 x i32> %val, <2 x i32>* %A, align 1
 	%inc = getelementptr <2 x i32>, <2 x i32>* %A, i32 1
@@ -66,13 +73,15 @@
 }
 
 define void @store_v2f32_update(<2 x float>** %ptr, <2 x float> %val) {
-;CHECK-LABEL: store_v2f32_update:
-;CHECK: vst1.8 {{{d[0-9]+}}}, [{{r[0-9]+}}]!
-	%A = load <2 x float>*, <2 x float>** %ptr
-	store  <2 x float> %val, <2 x float>* %A, align 1
-	%inc = getelementptr <2 x float>, <2 x float>* %A, i32 1
-        store <2 x float>* %inc, <2 x float>** %ptr
-	ret void
+; CHECK-LABEL: store_v2f32_update:
+; CHECK: str r{{[0-9]+}}, [{{r[0-9]+}}]
+; CHECK: str r{{[0-9]+}}, [{{r[0-9]+}}, #4]
+; CHECK: add.w r{{[0-9]+}}, r{{[0-9]+}}, #8
+    %A = load <2 x float>*, <2 x float>** %ptr
+    store  <2 x float> %val, <2 x float>* %A, align 1
+    %inc = getelementptr <2 x float>, <2 x float>* %A, i32 1
+    store <2 x float>* %inc, <2 x float>** %ptr
+    ret void
 }
 
 define void @store_v1i64(<1 x i64>** %ptr, <1 x i64> %val) {
@@ -84,12 +93,15 @@
 }
 
 define void @store_v1i64_update(<1 x i64>** %ptr, <1 x i64> %val) {
-;CHECK-LABEL: store_v1i64_update:
-;CHECK: vst1.8 {{{d[0-9]+}}}, [{{r[0-9]+}}]!
+; CHECK-LABEL: store_v1i64_update:
+; CHECK: str r{{[0-9]+}}, [{{r[0-9]+}}]
+; CHECK: str r{{[0-9]+}}, [{{r[0-9]+}}, #4]
+; CHECK: add.w r{{[0-9]+}}, r{{[0-9]+}}, #8
+
 	%A = load <1 x i64>*, <1 x i64>** %ptr
 	store  <1 x i64> %val, <1 x i64>* %A, align 1
 	%inc = getelementptr <1 x i64>, <1 x i64>* %A, i31 1
-        store <1 x i64>* %inc, <1 x i64>** %ptr
+    store <1 x i64>* %inc, <1 x i64>** %ptr
 	ret void
 }
 
Index: test/CodeGen/X86/avx-vextractf128.ll
===================================================================
--- test/CodeGen/X86/avx-vextractf128.ll
+++ test/CodeGen/X86/avx-vextractf128.ll
@@ -119,7 +119,7 @@
 define void @t9(i64* %p) {
 ; CHECK-LABEL: t9:
 ; CHECK:       ## BB#0:
-; CHECK-NEXT:    vxorps %xmm0, %xmm0, %xmm0
+; CHECK-NEXT:    vxorps %ymm0, %ymm0, %ymm0
 ; CHECK-NEXT:    vmovups %ymm0, (%rdi)
 ; CHECK-NEXT:    vzeroupper
 ; CHECK-NEXT:    retq