Index: lib/Target/AMDGPU/AMDGPURegisterBankInfo.cpp
===================================================================
--- lib/Target/AMDGPU/AMDGPURegisterBankInfo.cpp
+++ lib/Target/AMDGPU/AMDGPURegisterBankInfo.cpp
@@ -323,6 +323,8 @@
   }
 }
 
+// FIXME: Returns uniform if there's no source value information. This is
+// probably wrong.
 static bool isInstrUniformNonExtLoadAlign4(const MachineInstr &MI) {
   if (!MI.hasOneMemOperand())
     return false;
@@ -1047,8 +1049,13 @@
   SmallVector<unsigned, 1> SrcRegs(OpdMapper.getVRegs(1));
 
   // If the pointer is an SGPR, we have nothing to do.
-  if (SrcRegs.empty())
-    return false;
+  if (SrcRegs.empty()) {
+    Register PtrReg = MI.getOperand(1).getReg();
+    const RegisterBank *PtrBank = getRegBank(PtrReg, MRI, *TRI);
+    if (PtrBank == &AMDGPU::SGPRRegBank)
+      return false;
+    SrcRegs.push_back(PtrReg);
+  }
 
   assert(LoadSize % MaxNonSmrdLoadSize == 0);
 
@@ -2025,7 +2032,7 @@
 
   const MachineFunction &MF = *MI.getParent()->getParent();
   const MachineRegisterInfo &MRI = MF.getRegInfo();
-  SmallVector<const ValueMapping*, 8> OpdsMapping(MI.getNumOperands());
+  SmallVector<const ValueMapping*, 2> OpdsMapping(2);
   unsigned Size = getSizeInBits(MI.getOperand(0).getReg(), MRI, *TRI);
   LLT LoadTy = MRI.getType(MI.getOperand(0).getReg());
   Register PtrReg = MI.getOperand(1).getReg();
@@ -2036,7 +2043,10 @@
   const ValueMapping *ValMapping;
   const ValueMapping *PtrMapping;
 
-  if ((AS != AMDGPUAS::LOCAL_ADDRESS && AS != AMDGPUAS::REGION_ADDRESS &&
+  const RegisterBank *PtrBank = getRegBank(PtrReg, MRI, *TRI);
+
+  if (PtrBank == &AMDGPU::SGPRRegBank &&
+      (AS != AMDGPUAS::LOCAL_ADDRESS && AS != AMDGPUAS::REGION_ADDRESS &&
        AS != AMDGPUAS::PRIVATE_ADDRESS) &&
       isInstrUniformNonExtLoadAlign4(MI)) {
     // We have a uniform instruction so we want to use an SMRD load
Index: test/CodeGen/AMDGPU/GlobalISel/legalize-extract.mir
===================================================================
--- test/CodeGen/AMDGPU/GlobalISel/legalize-extract.mir
+++ test/CodeGen/AMDGPU/GlobalISel/legalize-extract.mir
@@ -937,6 +937,7 @@
 ...
 
 ---
+
 name: extract_s16_v2s16_offset0
 body: |
   bb.0:
@@ -1113,3 +1114,18 @@
     %2:_(s32) = G_ANYEXT %1
     $vgpr0 = COPY %2
 ...
+
+name: extract_s16_s64_offset18
+
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+    ; CHECK-LABEL: name: extract_s16_s64_offset18
+    ; CHECK: [[COPY:%[0-9]+]]:_(s64) = COPY $vgpr0_vgpr1
+    ; CHECK: [[EXTRACT:%[0-9]+]]:_(s16) = G_EXTRACT [[COPY]](s64), 18
+    ; CHECK: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[EXTRACT]](s16)
+    ; CHECK: $vgpr0 = COPY [[ANYEXT]](s32)
+  %0:_(s64) = COPY $vgpr0_vgpr1
+  %1:_(s16) = G_EXTRACT %0, 18
+  %2:_(s32) = G_ANYEXT %1
+  $vgpr0 = COPY %2
Index: test/CodeGen/AMDGPU/GlobalISel/regbankselect-load.mir
===================================================================
--- test/CodeGen/AMDGPU/GlobalISel/regbankselect-load.mir
+++ test/CodeGen/AMDGPU/GlobalISel/regbankselect-load.mir
@@ -70,6 +70,7 @@
   define amdgpu_kernel void @load_constant_i32_uniform_align1() {ret void}
   define amdgpu_kernel void @load_private_uniform_sgpr_i32() {ret void}
   define amdgpu_kernel void @load_constant_v8i32_vgpr_crash() { ret void }
+  define amdgpu_kernel void @load_constant_v8i32_vgpr_crash_loop_phi() { ret void }
 
   declare i32 @llvm.amdgcn.workitem.id.x() #0
   attributes #0 = { nounwind readnone }
@@ -657,12 +658,43 @@
 ---
 name: load_constant_v8i32_vgpr_crash
 legalized: true
+tracksRegLiveness: true
 
 body: |
   bb.0:
     liveins: $vgpr0_vgpr1
 
-  %0:_(p4) = COPY $vgpr0_vgpr1
-  %1:_(<8 x s32>) = G_LOAD %0 :: (load 32, addrspace 4)
+    ; CHECK-LABEL: name: load_constant_v8i32_vgpr_crash
+    ; CHECK: %0:vgpr(p4) = COPY $vgpr0_vgpr1
+    ; CHECK: vgpr(<4 x s32>) = G_LOAD %0(p4)
+    ; CHECK: vgpr(<4 x s32>) = G_LOAD
+    ; CHECK: G_CONCAT_VECTORS
+    %0:_(p4) = COPY $vgpr0_vgpr1
+    %1:_(<8 x s32>) = G_LOAD %0 :: (load 32, addrspace 4)
+...
+
+---
+name: load_constant_v8i32_vgpr_crash_loop_phi
+legalized: true
+tracksRegLiveness: true
 
+body: |
+  bb.0:
+    liveins: $sgpr0_sgpr1, $sgpr2_sgpr3
+
+    ; CHECK-LABEL: name: load_constant_v8i32_vgpr_crash_loop_phi
+    ; CHECK: G_PHI
+    ; CHECK: vgpr(<4 x s32>) = G_LOAD
+    ; CHECK: vgpr(<4 x s32>) = G_LOAD
+    ; CHECK: G_CONCAT_VECTORS
+
+    %0:_(p4) = COPY $sgpr0_sgpr1
+    %1:_(p4) = COPY $sgpr2_sgpr3
+    G_BR %bb.1
+
+  bb.1:
+    %2:_(p4) = G_PHI %0, %bb.0, %4, %bb.1
+    %3:_(<8 x s32>) = G_LOAD %2 :: (load 32, addrspace 4)
+    %4:_(p4) = COPY %1
+    G_BR %bb.1
 ...