Index: include/llvm/CodeGen/GlobalISel/LegalizerHelper.h
===================================================================
--- include/llvm/CodeGen/GlobalISel/LegalizerHelper.h
+++ include/llvm/CodeGen/GlobalISel/LegalizerHelper.h
@@ -171,6 +171,8 @@
 
   LegalizeResult narrowScalarMul(MachineInstr &MI, unsigned TypeIdx, LLT Ty);
 
+  LegalizeResult narrowScalarSelect(MachineInstr &MI, unsigned TypeIdx, LLT Ty);
+
   LegalizeResult lowerBitCount(MachineInstr &MI, unsigned TypeIdx, LLT Ty);
 
   MachineRegisterInfo &MRI;
Index: lib/CodeGen/GlobalISel/LegalizerHelper.cpp
===================================================================
--- lib/CodeGen/GlobalISel/LegalizerHelper.cpp
+++ lib/CodeGen/GlobalISel/LegalizerHelper.cpp
@@ -139,6 +139,11 @@
   if (!LeftoverTy.isValid()) {
     assert(LeftoverRegs.empty());
 
+    if (!ResultTy.isVector()) {
+      MIRBuilder.buildMerge(DstReg, PartRegs);
+      return;
+    }
+
     if (PartTy.isVector())
       MIRBuilder.buildConcatVectors(DstReg, PartRegs);
     else
@@ -808,6 +813,8 @@
     MI.eraseFromParent();
     return Legalized;
   }
+  case TargetOpcode::G_SELECT:
+    return narrowScalarSelect(MI, TypeIdx, NarrowTy);
   case TargetOpcode::G_AND:
   case TargetOpcode::G_OR:
   case TargetOpcode::G_XOR: {
@@ -2221,6 +2228,52 @@
   return Legalized;
 }
 
+LegalizerHelper::LegalizeResult
+LegalizerHelper::narrowScalarSelect(MachineInstr &MI, unsigned TypeIdx,
+                                    LLT NarrowTy) {
+  if (TypeIdx != 0)
+    return UnableToLegalize;
+
+  unsigned CondReg = MI.getOperand(1).getReg();
+  LLT CondTy = MRI.getType(CondReg);
+  if (CondTy.isVector()) // TODO: Handle vselect
+    return UnableToLegalize;
+
+  unsigned DstReg = MI.getOperand(0).getReg();
+  LLT DstTy = MRI.getType(DstReg);
+
+  SmallVector<unsigned, 4> DstRegs, DstLeftoverRegs;
+  SmallVector<unsigned, 4> Src1Regs, Src1LeftoverRegs;
+  SmallVector<unsigned, 4> Src2Regs, Src2LeftoverRegs;
+  LLT LeftoverTy;
+  if (!extractParts(MI.getOperand(2).getReg(), DstTy, NarrowTy, LeftoverTy,
+                    Src1Regs, Src1LeftoverRegs))
+    return UnableToLegalize;
+
+  LLT Unused;
+  if (!extractParts(MI.getOperand(3).getReg(), DstTy, NarrowTy, Unused,
+                    Src2Regs, Src2LeftoverRegs))
+    llvm_unreachable("inconsistent extractParts result");
+
+  for (unsigned I = 0, E = Src1Regs.size(); I != E; ++I) {
+    auto Select = MIRBuilder.buildSelect(NarrowTy,
+                                         CondReg, Src1Regs[I], Src2Regs[I]);
+    DstRegs.push_back(Select->getOperand(0).getReg());
+  }
+
+  for (unsigned I = 0, E = Src1LeftoverRegs.size(); I != E; ++I) {
+    auto Select = MIRBuilder.buildSelect(
+      LeftoverTy, CondReg, Src1LeftoverRegs[I], Src2LeftoverRegs[I]);
+    DstLeftoverRegs.push_back(Select->getOperand(0).getReg());
+  }
+
+  insertParts(DstReg, DstTy, NarrowTy, DstRegs,
+              LeftoverTy, DstLeftoverRegs);
+
+  MI.eraseFromParent();
+  return Legalized;
+}
+
 LegalizerHelper::LegalizeResult
 LegalizerHelper::lowerBitCount(MachineInstr &MI, unsigned TypeIdx, LLT Ty) {
   unsigned Opc = MI.getOpcode();
Index: test/CodeGen/AMDGPU/GlobalISel/legalize-select.mir
===================================================================
--- test/CodeGen/AMDGPU/GlobalISel/legalize-select.mir
+++ test/CodeGen/AMDGPU/GlobalISel/legalize-select.mir
@@ -1,5 +1,5 @@
 # NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
-# RUN: llc -O0 -march=amdgcn -mcpu=tahiti -run-pass=legalizer %s -o - | FileCheck %s
+# RUN: llc -O0 -march=amdgcn -mcpu=tahiti -run-pass=legalizer -global-isel-abort=0  %s -o - | FileCheck %s
 
 ---
 name: test_select_s32
@@ -132,6 +132,67 @@
     $vgpr0 = COPY %6
 
 ...
+
+---
+name: test_select_s96
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1_vgpr2, $vgpr3_vgpr4_vgpr5 , $vgpr6
+    ; CHECK-LABEL: name: test_select_s96
+    ; CHECK: [[COPY:%[0-9]+]]:_(s96) = COPY $vgpr0_vgpr1_vgpr2
+    ; CHECK: [[COPY1:%[0-9]+]]:_(s96) = COPY $vgpr3_vgpr4_vgpr5
+    ; CHECK: [[COPY2:%[0-9]+]]:_(s32) = COPY $vgpr6
+    ; CHECK: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 0
+    ; CHECK: [[ICMP:%[0-9]+]]:_(s1) = G_ICMP intpred(ne), [[COPY2]](s32), [[C]]
+    ; CHECK: [[EXTRACT:%[0-9]+]]:_(s64) = G_EXTRACT [[COPY]](s96), 0
+    ; CHECK: [[EXTRACT1:%[0-9]+]]:_(s32) = G_EXTRACT [[COPY]](s96), 64
+    ; CHECK: [[EXTRACT2:%[0-9]+]]:_(s64) = G_EXTRACT [[COPY1]](s96), 0
+    ; CHECK: [[EXTRACT3:%[0-9]+]]:_(s32) = G_EXTRACT [[COPY1]](s96), 64
+    ; CHECK: [[SELECT:%[0-9]+]]:_(s64) = G_SELECT [[ICMP]](s1), [[EXTRACT]], [[EXTRACT2]]
+    ; CHECK: [[SELECT1:%[0-9]+]]:_(s32) = G_SELECT [[ICMP]](s1), [[EXTRACT1]], [[EXTRACT3]]
+    ; CHECK: [[DEF:%[0-9]+]]:_(s96) = G_IMPLICIT_DEF
+    ; CHECK: [[INSERT:%[0-9]+]]:_(s96) = G_INSERT [[DEF]], [[SELECT]](s64), 0
+    ; CHECK: [[INSERT1:%[0-9]+]]:_(s96) = G_INSERT [[INSERT]], [[SELECT1]](s32), 64
+    ; CHECK: $vgpr0_vgpr1_vgpr2 = COPY [[INSERT1]](s96)
+    %0:_(s96) = COPY $vgpr0_vgpr1_vgpr2
+    %1:_(s96) = COPY $vgpr3_vgpr4_vgpr5
+    %2:_(s32) = COPY $vgpr6
+    %3:_(s32) = G_CONSTANT i32 0
+
+    %4:_(s1) = G_ICMP intpred(ne), %2, %3
+    %5:_(s96) = G_SELECT %4, %0, %1
+    $vgpr0_vgpr1_vgpr2 = COPY %5
+
+...
+
+---
+name: test_select_s128
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1_vgpr2_vgpr3, $vgpr4_vgpr5_vgpr6_vgpr7 , $vgpr8
+
+    ; CHECK-LABEL: name: test_select_s128
+    ; CHECK: [[COPY:%[0-9]+]]:_(s128) = COPY $vgpr0_vgpr1_vgpr2_vgpr3
+    ; CHECK: [[COPY1:%[0-9]+]]:_(s128) = COPY $vgpr4_vgpr5_vgpr6_vgpr7
+    ; CHECK: [[COPY2:%[0-9]+]]:_(s32) = COPY $vgpr8
+    ; CHECK: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 0
+    ; CHECK: [[ICMP:%[0-9]+]]:_(s1) = G_ICMP intpred(ne), [[COPY2]](s32), [[C]]
+    ; CHECK: [[UV:%[0-9]+]]:_(s64), [[UV1:%[0-9]+]]:_(s64) = G_UNMERGE_VALUES [[COPY]](s128)
+    ; CHECK: [[UV2:%[0-9]+]]:_(s64), [[UV3:%[0-9]+]]:_(s64) = G_UNMERGE_VALUES [[COPY1]](s128)
+    ; CHECK: [[SELECT:%[0-9]+]]:_(s64) = G_SELECT [[ICMP]](s1), [[UV]], [[UV2]]
+    ; CHECK: [[SELECT1:%[0-9]+]]:_(s64) = G_SELECT [[ICMP]](s1), [[UV1]], [[UV3]]
+    ; CHECK: [[MV:%[0-9]+]]:_(s128) = G_MERGE_VALUES [[SELECT]](s64), [[SELECT1]](s64)
+    ; CHECK: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[MV]](s128)
+    %0:_(s128) = COPY $vgpr0_vgpr1_vgpr2_vgpr3
+    %1:_(s128) = COPY $vgpr4_vgpr5_vgpr6_vgpr7
+    %2:_(s32) = COPY $vgpr8
+    %3:_(s32) = G_CONSTANT i32 0
+    %4:_(s1) = G_ICMP intpred(ne), %2, %3
+    %5:_(s128) = G_SELECT %4, %0, %1
+    $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %5
+
+...
+
 ---
 name: test_select_v2s16
 body: |
@@ -390,6 +451,7 @@
 body: |
   bb.0:
     liveins: $vgpr0, $vgpr1, $vgpr2
+
     ; CHECK-LABEL: name: test_select_p5
     ; CHECK: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 0
     ; CHECK: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0
@@ -402,7 +464,6 @@
     %1:_(s32) = COPY $vgpr0
     %2:_(p5) = COPY $vgpr1
     %3:_(p5) = COPY $vgpr2
-
     %4:_(s1) = G_ICMP intpred(ne), %0, %1
     %5:_(p5) = G_SELECT %4, %2, %3
     $vgpr0 = COPY %5
@@ -414,6 +475,7 @@
 body: |
   bb.0:
     liveins: $vgpr0, $vgpr1_vgpr2, $vgpr3_vgpr4
+
     ; CHECK-LABEL: name: test_select_p999
     ; CHECK: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 0
     ; CHECK: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0
@@ -426,7 +488,6 @@
     %1:_(s32) = COPY $vgpr0
     %2:_(p999) = COPY $vgpr1_vgpr2
     %3:_(p999) = COPY $vgpr3_vgpr4
-
     %4:_(s1) = G_ICMP intpred(ne), %0, %1
     %5:_(p999) = G_SELECT %4, %2, %3
     $vgpr0_vgpr1 = COPY %5
@@ -434,6 +495,7 @@
 ...
 
 ---
+
 name: test_select_v2p3
 body: |
   bb.0:
@@ -606,6 +668,51 @@
 ...
 
 ---
+name: test_select_v2s96
+body: |
+  bb.0:
+    liveins: $vgpr0
+
+    ; CHECK-LABEL: name: test_select_v2s96
+    ; CHECK: [[DEF:%[0-9]+]]:_(<2 x s96>) = G_IMPLICIT_DEF
+    ; CHECK: [[DEF1:%[0-9]+]]:_(<2 x s96>) = G_IMPLICIT_DEF
+    ; CHECK: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0
+    ; CHECK: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 0
+    ; CHECK: [[ICMP:%[0-9]+]]:_(s1) = G_ICMP intpred(ne), [[COPY]](s32), [[C]]
+    ; CHECK: [[UV:%[0-9]+]]:_(s96), [[UV1:%[0-9]+]]:_(s96) = G_UNMERGE_VALUES [[DEF]](<2 x s96>)
+    ; CHECK: [[UV2:%[0-9]+]]:_(s96), [[UV3:%[0-9]+]]:_(s96) = G_UNMERGE_VALUES [[DEF1]](<2 x s96>)
+    ; CHECK: [[EXTRACT:%[0-9]+]]:_(s64) = G_EXTRACT [[UV]](s96), 0
+    ; CHECK: [[EXTRACT1:%[0-9]+]]:_(s32) = G_EXTRACT [[UV]](s96), 64
+    ; CHECK: [[EXTRACT2:%[0-9]+]]:_(s64) = G_EXTRACT [[UV2]](s96), 0
+    ; CHECK: [[EXTRACT3:%[0-9]+]]:_(s32) = G_EXTRACT [[UV2]](s96), 64
+    ; CHECK: [[SELECT:%[0-9]+]]:_(s64) = G_SELECT [[ICMP]](s1), [[EXTRACT]], [[EXTRACT2]]
+    ; CHECK: [[SELECT1:%[0-9]+]]:_(s32) = G_SELECT [[ICMP]](s1), [[EXTRACT1]], [[EXTRACT3]]
+    ; CHECK: [[DEF2:%[0-9]+]]:_(s96) = G_IMPLICIT_DEF
+    ; CHECK: [[INSERT:%[0-9]+]]:_(s96) = G_INSERT [[DEF2]], [[SELECT]](s64), 0
+    ; CHECK: [[INSERT1:%[0-9]+]]:_(s96) = G_INSERT [[INSERT]], [[SELECT1]](s32), 64
+    ; CHECK: [[EXTRACT4:%[0-9]+]]:_(s64) = G_EXTRACT [[UV1]](s96), 0
+    ; CHECK: [[EXTRACT5:%[0-9]+]]:_(s32) = G_EXTRACT [[UV1]](s96), 64
+    ; CHECK: [[EXTRACT6:%[0-9]+]]:_(s64) = G_EXTRACT [[UV3]](s96), 0
+    ; CHECK: [[EXTRACT7:%[0-9]+]]:_(s32) = G_EXTRACT [[UV3]](s96), 64
+    ; CHECK: [[SELECT2:%[0-9]+]]:_(s64) = G_SELECT [[ICMP]](s1), [[EXTRACT4]], [[EXTRACT6]]
+    ; CHECK: [[SELECT3:%[0-9]+]]:_(s32) = G_SELECT [[ICMP]](s1), [[EXTRACT5]], [[EXTRACT7]]
+    ; CHECK: [[DEF3:%[0-9]+]]:_(s96) = G_IMPLICIT_DEF
+    ; CHECK: [[INSERT2:%[0-9]+]]:_(s96) = G_INSERT [[DEF3]], [[SELECT2]](s64), 0
+    ; CHECK: [[INSERT3:%[0-9]+]]:_(s96) = G_INSERT [[INSERT2]], [[SELECT3]](s32), 64
+    ; CHECK: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s96>) = G_BUILD_VECTOR [[INSERT1]](s96), [[INSERT3]](s96)
+    ; CHECK: S_NOP 0, implicit [[BUILD_VECTOR]](<2 x s96>)
+    %0:_(<2 x s96>) = G_IMPLICIT_DEF
+    %1:_(<2 x s96>) = G_IMPLICIT_DEF
+    %2:_(s32) = COPY $vgpr0
+    %3:_(s32) = G_CONSTANT i32 0
+    %4:_(s1) = G_ICMP intpred(ne), %2, %3
+    %5:_(<2 x s96>) = G_SELECT %4, %0, %1
+    S_NOP 0, implicit %5
+
+...
+
+---
+
 name: test_select_v8p0
 body: |
   bb.0:
@@ -637,6 +744,40 @@
     %4:_(s1) = G_ICMP intpred(ne), %0, %1
     %5:_(<8 x p0>) = G_SELECT %4, %2, %3
     S_NOP 0, implicit %5
+...
+---
+name: test_select_v2s128
+body: |
+  bb.0:
+    liveins:  $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7, $vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15, $vgpr16
+
+    ; CHECK-LABEL: name: test_select_v2s128
+    ; CHECK: [[COPY:%[0-9]+]]:_(<2 x s128>) = COPY $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7
+    ; CHECK: [[COPY1:%[0-9]+]]:_(<2 x s128>) = COPY $vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
+    ; CHECK: [[COPY2:%[0-9]+]]:_(s32) = COPY $vgpr16
+    ; CHECK: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 0
+    ; CHECK: [[ICMP:%[0-9]+]]:_(s1) = G_ICMP intpred(ne), [[COPY2]](s32), [[C]]
+    ; CHECK: [[UV:%[0-9]+]]:_(s128), [[UV1:%[0-9]+]]:_(s128) = G_UNMERGE_VALUES [[COPY]](<2 x s128>)
+    ; CHECK: [[UV2:%[0-9]+]]:_(s128), [[UV3:%[0-9]+]]:_(s128) = G_UNMERGE_VALUES [[COPY1]](<2 x s128>)
+    ; CHECK: [[UV4:%[0-9]+]]:_(s64), [[UV5:%[0-9]+]]:_(s64) = G_UNMERGE_VALUES [[UV]](s128)
+    ; CHECK: [[UV6:%[0-9]+]]:_(s64), [[UV7:%[0-9]+]]:_(s64) = G_UNMERGE_VALUES [[UV2]](s128)
+    ; CHECK: [[SELECT:%[0-9]+]]:_(s64) = G_SELECT [[ICMP]](s1), [[UV4]], [[UV6]]
+    ; CHECK: [[SELECT1:%[0-9]+]]:_(s64) = G_SELECT [[ICMP]](s1), [[UV5]], [[UV7]]
+    ; CHECK: [[MV:%[0-9]+]]:_(s128) = G_MERGE_VALUES [[SELECT]](s64), [[SELECT1]](s64)
+    ; CHECK: [[UV8:%[0-9]+]]:_(s64), [[UV9:%[0-9]+]]:_(s64) = G_UNMERGE_VALUES [[UV1]](s128)
+    ; CHECK: [[UV10:%[0-9]+]]:_(s64), [[UV11:%[0-9]+]]:_(s64) = G_UNMERGE_VALUES [[UV3]](s128)
+    ; CHECK: [[SELECT2:%[0-9]+]]:_(s64) = G_SELECT [[ICMP]](s1), [[UV8]], [[UV10]]
+    ; CHECK: [[SELECT3:%[0-9]+]]:_(s64) = G_SELECT [[ICMP]](s1), [[UV9]], [[UV11]]
+    ; CHECK: [[MV1:%[0-9]+]]:_(s128) = G_MERGE_VALUES [[SELECT2]](s64), [[SELECT3]](s64)
+    ; CHECK: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s128>) = G_BUILD_VECTOR [[MV]](s128), [[MV1]](s128)
+    ; CHECK: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 = COPY [[BUILD_VECTOR]](<2 x s128>)
+    %0:_(<2 x s128>) = COPY $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 ,
+    %1:_(<2 x s128>) = COPY $vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
+    %2:_(s32) = COPY $vgpr16
+    %3:_(s32) = G_CONSTANT i32 0
+    %4:_(s1) = G_ICMP intpred(ne), %2, %3
+    %5:_(<2 x s128>) = G_SELECT %4, %0, %1
+    $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7   = COPY %5
 
 ...