Index: llvm/lib/Target/AMDGPU/AMDGPURegisterBankInfo.cpp
===================================================================
--- llvm/lib/Target/AMDGPU/AMDGPURegisterBankInfo.cpp
+++ llvm/lib/Target/AMDGPU/AMDGPURegisterBankInfo.cpp
@@ -847,7 +847,18 @@
         continue;
       }
 
-      LLT OpTy = MRI.getType(Op.getReg());
+      Register OpReg = Op.getReg();
+      LLT OpTy = MRI.getType(OpReg);
+
+      const RegisterBank *OpBank = getRegBank(OpReg, MRI, *TRI);
+      if (OpBank != &AMDGPU::VGPRRegBank) {
+        // Insert copy from AGPR to VGPR before the loop.
+        B.setMBB(MBB);
+        OpReg = B.buildCopy(OpTy, OpReg).getReg(0);
+        MRI.setRegBank(OpReg, AMDGPU::VGPRRegBank);
+        B.setInstr(*I);
+      }
+
       unsigned OpSize = OpTy.getSizeInBits();
 
       // Can only do a readlane of 32-bit pieces.
@@ -857,11 +868,11 @@
           = MRI.createVirtualRegister(&AMDGPU::SReg_32_XM0RegClass);
         MRI.setType(CurrentLaneOpReg, OpTy);
 
-        constrainGenericRegister(Op.getReg(), AMDGPU::VGPR_32RegClass, MRI);
+        constrainGenericRegister(OpReg, AMDGPU::VGPR_32RegClass, MRI);
         // Read the next variant <- also loop target.
         BuildMI(*LoopBB, I, DL, TII->get(AMDGPU::V_READFIRSTLANE_B32),
                 CurrentLaneOpReg)
-          .addReg(Op.getReg());
+          .addReg(OpReg);
 
         Register NewCondReg = MRI.createVirtualRegister(WaveRC);
         bool First = CondReg == AMDGPU::NoRegister;
@@ -872,7 +883,7 @@
         B.buildInstr(AMDGPU::V_CMP_EQ_U32_e64)
           .addDef(NewCondReg)
           .addReg(CurrentLaneOpReg)
-          .addReg(Op.getReg());
+          .addReg(OpReg);
         Op.setReg(CurrentLaneOpReg);
 
         if (!First) {
@@ -904,7 +915,7 @@
         // Insert the unmerge before the loop.
 
         B.setMBB(MBB);
-        auto Unmerge = B.buildUnmerge(UnmergeTy, Op.getReg());
+        auto Unmerge = B.buildUnmerge(UnmergeTy, OpReg);
         B.setInstr(*I);
 
         unsigned NumPieces = Unmerge->getNumOperands() - 1;
@@ -1048,7 +1059,7 @@
     assert(MI.getOperand(Op).isUse());
     Register Reg = MI.getOperand(Op).getReg();
     const RegisterBank *OpBank = getRegBank(Reg, MRI, *TRI);
-    if (OpBank->getID() == AMDGPU::VGPRRegBankID)
+    if (OpBank->getID() != AMDGPU::SGPRRegBankID)
       SGPROperandRegs.insert(Reg);
   }
 
@@ -1083,16 +1094,24 @@
     MachineInstr &MI, MachineRegisterInfo &MRI, unsigned OpIdx) const {
   Register Reg = MI.getOperand(OpIdx).getReg();
   const RegisterBank *Bank = getRegBank(Reg, MRI, *TRI);
-  if (Bank != &AMDGPU::VGPRRegBank)
+  if (Bank == &AMDGPU::SGPRRegBank)
     return;
 
+  LLT Ty = MRI.getType(Reg);
   MachineIRBuilder B(MI);
+
+  if (Bank != &AMDGPU::VGPRRegBank) {
+    // We need to copy from AGPR to VGPR
+    Reg = B.buildCopy(Ty, Reg).getReg(0);
+    MRI.setRegBank(Reg, AMDGPU::VGPRRegBank);
+  }
+
   Register SGPR = MRI.createVirtualRegister(&AMDGPU::SReg_32RegClass);
   B.buildInstr(AMDGPU::V_READFIRSTLANE_B32)
     .addDef(SGPR)
     .addReg(Reg);
 
-  MRI.setType(SGPR, MRI.getType(Reg));
+  MRI.setType(SGPR, Ty);
 
   const TargetRegisterClass *Constrained =
       constrainGenericRegister(Reg, AMDGPU::VGPR_32RegClass, MRI);
@@ -1922,7 +1941,7 @@
   const RegisterBank &IdxBank =
     *OpdMapper.getInstrMapping().getOperandMapping(2).BreakDown[0].RegBank;
 
-  bool IsDivergentIdx = IdxBank == AMDGPU::VGPRRegBank;
+  bool IsDivergentIdx = IdxBank != AMDGPU::SGPRRegBank;
 
   LLT VecTy = MRI.getType(VecReg);
   unsigned EltSize = VecTy.getScalarSizeInBits();
@@ -2004,7 +2023,7 @@
   const RegisterBank &IdxBank =
     *OpdMapper.getInstrMapping().getOperandMapping(3).BreakDown[0].RegBank;
 
-  bool IsDivergentIdx = IdxBank == AMDGPU::VGPRRegBank;
+  bool IsDivergentIdx = IdxBank != AMDGPU::SGPRRegBank;
 
   LLT VecTy = MRI.getType(VecReg);
   unsigned EltSize = VecTy.getScalarSizeInBits();
Index: llvm/test/CodeGen/AMDGPU/GlobalISel/regbankselect-amdgcn.readlane.mir
===================================================================
--- llvm/test/CodeGen/AMDGPU/GlobalISel/regbankselect-amdgcn.readlane.mir
+++ llvm/test/CodeGen/AMDGPU/GlobalISel/regbankselect-amdgcn.readlane.mir
@@ -1,6 +1,6 @@
 # NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
-# RUN: llc -march=amdgcn -mcpu=fiji -run-pass=regbankselect %s -verify-machineinstrs -o - -regbankselect-fast | FileCheck %s
-# RUN: llc -march=amdgcn -mcpu=fiji -run-pass=regbankselect %s -verify-machineinstrs -o - -regbankselect-greedy | FileCheck %s
+# RUN: llc -march=amdgcn -mcpu=gfx908 -run-pass=regbankselect %s -verify-machineinstrs -o - -regbankselect-fast | FileCheck %s
+# RUN: llc -march=amdgcn -mcpu=gfx908 -run-pass=regbankselect %s -verify-machineinstrs -o - -regbankselect-greedy | FileCheck %s
 
 ---
 name: readlane_ss
@@ -69,3 +69,78 @@
     %1:_(s32) = COPY $vgpr0
     %2:_(s32) = G_INTRINSIC intrinsic(@llvm.amdgcn.readlane), %0, %1
 ...
+
+---
+name: readlane_aa
+legalized: true
+
+body: |
+  bb.0:
+    liveins: $agpr0, $agpr1
+    ; CHECK-LABEL: name: readlane_aa
+    ; CHECK: [[COPY:%[0-9]+]]:agpr(s32) = COPY $agpr0
+    ; CHECK: [[COPY1:%[0-9]+]]:agpr(s32) = COPY $agpr1
+    ; CHECK: [[COPY2:%[0-9]+]]:vgpr(s32) = COPY [[COPY]](s32)
+    ; CHECK: [[COPY3:%[0-9]+]]:vgpr_32(s32) = COPY [[COPY1]](s32)
+    ; CHECK: [[V_READFIRSTLANE_B32_:%[0-9]+]]:sreg_32(s32) = V_READFIRSTLANE_B32 [[COPY3]](s32), implicit $exec
+    ; CHECK: [[INT:%[0-9]+]]:sgpr(s32) = G_INTRINSIC intrinsic(@llvm.amdgcn.readlane), [[COPY2]](s32), [[V_READFIRSTLANE_B32_]](s32)
+    ; CHECK: S_ENDPGM 0, implicit [[INT]](s32)
+    %0:_(s32) = COPY $agpr0
+    %1:_(s32) = COPY $agpr1
+    %2:_(s32) = G_INTRINSIC intrinsic(@llvm.amdgcn.readlane), %0, %1
+    S_ENDPGM 0, implicit %2
+...
+
+---
+name: readlane_as
+legalized: true
+
+body: |
+  bb.0:
+    liveins: $agpr0, $sgpr0
+    ; CHECK-LABEL: name: readlane_as
+    ; CHECK: [[COPY:%[0-9]+]]:agpr(s32) = COPY $agpr0
+    ; CHECK: [[COPY1:%[0-9]+]]:sgpr(s32) = COPY $sgpr0
+    ; CHECK: [[COPY2:%[0-9]+]]:vgpr(s32) = COPY [[COPY]](s32)
+    ; CHECK: [[INT:%[0-9]+]]:sgpr(s32) = G_INTRINSIC intrinsic(@llvm.amdgcn.readlane), [[COPY2]](s32), [[COPY1]](s32)
+    %0:_(s32) = COPY $agpr0
+    %1:_(s32) = COPY $sgpr0
+    %2:_(s32) = G_INTRINSIC intrinsic(@llvm.amdgcn.readlane), %0, %1
+...
+
+---
+name: readlane_sa
+legalized: true
+
+body: |
+  bb.0:
+    liveins: $agpr0, $sgpr0
+    ; CHECK-LABEL: name: readlane_sa
+    ; CHECK: [[COPY:%[0-9]+]]:sgpr(s32) = COPY $sgpr0
+    ; CHECK: [[COPY1:%[0-9]+]]:agpr(s32) = COPY $agpr0
+    ; CHECK: [[COPY2:%[0-9]+]]:vgpr(s32) = COPY [[COPY]](s32)
+    ; CHECK: [[COPY3:%[0-9]+]]:vgpr_32(s32) = COPY [[COPY1]](s32)
+    ; CHECK: [[V_READFIRSTLANE_B32_:%[0-9]+]]:sreg_32(s32) = V_READFIRSTLANE_B32 [[COPY3]](s32), implicit $exec
+    ; CHECK: [[INT:%[0-9]+]]:sgpr(s32) = G_INTRINSIC intrinsic(@llvm.amdgcn.readlane), [[COPY2]](s32), [[V_READFIRSTLANE_B32_]](s32)
+    %0:_(s32) = COPY $sgpr0
+    %1:_(s32) = COPY $agpr0
+    %2:_(s32) = G_INTRINSIC intrinsic(@llvm.amdgcn.readlane), %0, %1
+...
+
+---
+name: readlane_va
+legalized: true
+
+body: |
+  bb.0:
+    liveins: $vgpr0, $agpr0
+    ; CHECK-LABEL: name: readlane_va
+    ; CHECK: [[COPY:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
+    ; CHECK: [[COPY1:%[0-9]+]]:agpr(s32) = COPY $agpr0
+    ; CHECK: [[COPY2:%[0-9]+]]:vgpr_32(s32) = COPY [[COPY1]](s32)
+    ; CHECK: [[V_READFIRSTLANE_B32_:%[0-9]+]]:sreg_32(s32) = V_READFIRSTLANE_B32 [[COPY2]](s32), implicit $exec
+    ; CHECK: [[INT:%[0-9]+]]:sgpr(s32) = G_INTRINSIC intrinsic(@llvm.amdgcn.readlane), [[COPY]](s32), [[V_READFIRSTLANE_B32_]](s32)
+    %0:_(s32) = COPY $vgpr0
+    %1:_(s32) = COPY $agpr0
+    %2:_(s32) = G_INTRINSIC intrinsic(@llvm.amdgcn.readlane), %0, %1
+...
Index: llvm/test/CodeGen/AMDGPU/GlobalISel/regbankselect-waterfall-agpr.mir
===================================================================
--- /dev/null
+++ llvm/test/CodeGen/AMDGPU/GlobalISel/regbankselect-waterfall-agpr.mir
@@ -0,0 +1,107 @@
+# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
+# RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx908 -verify-machineinstrs -run-pass=regbankselect -o - %s | FileCheck %s
+
+# Make sure that an arbitrary AGPR is treated as a divergent value
+# that needs to be copied to VGPR, and then waterfalled
+
+# 32-bit case
+---
+name:            raw_buffer_store__sgpr_rsrc__vgpr_val__vgpr_voffset__agpr_soffset
+legalized:       true
+tracksRegLiveness: true
+body:             |
+  bb.0:
+    liveins: $sgpr0_sgpr1_sgpr2_sgpr3, $vgpr0, $vgpr1, $agpr0
+
+    ; CHECK-LABEL: name: raw_buffer_store__sgpr_rsrc__vgpr_val__vgpr_voffset__agpr_soffset
+    ; CHECK: successors: %bb.1(0x80000000)
+    ; CHECK: liveins: $sgpr0_sgpr1_sgpr2_sgpr3, $vgpr0, $vgpr1, $agpr0
+    ; CHECK: %val:vgpr(s32) = COPY $vgpr0
+    ; CHECK: %rsrc:sgpr(<4 x s32>) = COPY $sgpr0_sgpr1_sgpr2_sgpr3
+    ; CHECK: %agpr:agpr(s32) = COPY $agpr0
+    ; CHECK: %voffset:vgpr(s32) = COPY $vgpr1
+    ; CHECK: %zero:sgpr(s32) = G_CONSTANT i32 0
+    ; CHECK: [[COPY:%[0-9]+]]:vgpr(s32) = COPY %zero(s32)
+    ; CHECK: [[DEF:%[0-9]+]]:sreg_64_xexec = IMPLICIT_DEF
+    ; CHECK: [[COPY1:%[0-9]+]]:vgpr_32(s32) = COPY %agpr(s32)
+    ; CHECK: [[S_MOV_B64_term:%[0-9]+]]:sreg_64_xexec = S_MOV_B64_term $exec
+    ; CHECK: .1:
+    ; CHECK: successors: %bb.2(0x40000000), %bb.1(0x40000000)
+    ; CHECK: [[PHI:%[0-9]+]]:sreg_64_xexec = PHI [[DEF]], %bb.0, %9, %bb.1
+    ; CHECK: [[V_READFIRSTLANE_B32_:%[0-9]+]]:sreg_32_xm0(s32) = V_READFIRSTLANE_B32 [[COPY1]](s32), implicit $exec
+    ; CHECK: [[V_CMP_EQ_U32_e64_:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U32_e64 [[V_READFIRSTLANE_B32_]](s32), [[COPY1]](s32), implicit $exec
+    ; CHECK: G_AMDGPU_BUFFER_STORE %val(s32), %rsrc(<4 x s32>), [[COPY]](s32), %voffset, [[V_READFIRSTLANE_B32_]], 0, 0, 0 :: (dereferenceable store 4, addrspace 4)
+    ; CHECK: [[S_AND_SAVEEXEC_B64_:%[0-9]+]]:sreg_64_xexec = S_AND_SAVEEXEC_B64 killed [[V_CMP_EQ_U32_e64_]], implicit-def $exec, implicit-def $scc, implicit $exec
+    ; CHECK: $exec = S_XOR_B64_term $exec, [[S_AND_SAVEEXEC_B64_]], implicit-def $scc
+    ; CHECK: S_CBRANCH_EXECNZ %bb.1, implicit $exec
+    ; CHECK: .2:
+    ; CHECK: successors: %bb.3(0x80000000)
+    ; CHECK: $exec = S_MOV_B64_term [[S_MOV_B64_term]]
+    ; CHECK: .3:
+    ; CHECK: S_ENDPGM 0
+    %val:_(s32) = COPY $vgpr0
+    %rsrc:_(<4 x s32>) = COPY $sgpr0_sgpr1_sgpr2_sgpr3
+    %agpr:_(s32) = COPY $agpr0
+    %voffset:_(s32) = COPY $vgpr1
+    %zero:_(s32) = G_CONSTANT i32 0
+    G_AMDGPU_BUFFER_STORE %val, %rsrc, %zero, %voffset, %agpr, 0, 0, 0 :: (dereferenceable store 4, addrspace 4)
+    S_ENDPGM 0
+
+...
+
+# Register tuple case
+---
+name:            load_1d_vgpr_vaddr__agpr_srsrc
+legalized:       true
+tracksRegLiveness: true
+body:             |
+  bb.0:
+    liveins: $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7, $vgpr0
+    ; CHECK-LABEL: name: load_1d_vgpr_vaddr__agpr_srsrc
+    ; CHECK: successors: %bb.1(0x80000000)
+    ; CHECK: liveins: $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7, $vgpr0
+    ; CHECK: [[COPY:%[0-9]+]]:agpr(<8 x s32>) = COPY $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7
+    ; CHECK: [[COPY1:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
+    ; CHECK: [[DEF:%[0-9]+]]:vgpr(<4 x s32>) = G_IMPLICIT_DEF
+    ; CHECK: [[DEF1:%[0-9]+]]:sreg_64_xexec = IMPLICIT_DEF
+    ; CHECK: [[COPY2:%[0-9]+]]:vgpr(<8 x s32>) = COPY [[COPY]](<8 x s32>)
+    ; CHECK: [[UV:%[0-9]+]]:vreg_64(s64), [[UV1:%[0-9]+]]:vreg_64(s64), [[UV2:%[0-9]+]]:vreg_64(s64), [[UV3:%[0-9]+]]:vreg_64(s64) = G_UNMERGE_VALUES [[COPY2]](<8 x s32>)
+    ; CHECK: [[S_MOV_B64_term:%[0-9]+]]:sreg_64_xexec = S_MOV_B64_term $exec
+    ; CHECK: .1:
+    ; CHECK: successors: %bb.2(0x40000000), %bb.1(0x40000000)
+    ; CHECK: [[PHI:%[0-9]+]]:sreg_64_xexec = PHI [[DEF1]], %bb.0, %8, %bb.1
+    ; CHECK: [[PHI1:%[0-9]+]]:vgpr(<4 x s32>) = G_PHI [[DEF]](<4 x s32>), %bb.0, %2(<4 x s32>), %bb.1
+    ; CHECK: [[V_READFIRSTLANE_B32_:%[0-9]+]]:sreg_32_xm0(s32) = V_READFIRSTLANE_B32 [[UV]].sub0(s64), implicit $exec
+    ; CHECK: [[V_READFIRSTLANE_B32_1:%[0-9]+]]:sreg_32_xm0(s32) = V_READFIRSTLANE_B32 [[UV]].sub1(s64), implicit $exec
+    ; CHECK: [[MV:%[0-9]+]]:sreg_64_xexec(s64) = G_MERGE_VALUES [[V_READFIRSTLANE_B32_]](s32), [[V_READFIRSTLANE_B32_1]](s32)
+    ; CHECK: [[V_CMP_EQ_U64_e64_:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U64_e64 [[MV]](s64), [[UV]](s64), implicit $exec
+    ; CHECK: [[V_READFIRSTLANE_B32_2:%[0-9]+]]:sreg_32_xm0(s32) = V_READFIRSTLANE_B32 [[UV1]].sub0(s64), implicit $exec
+    ; CHECK: [[V_READFIRSTLANE_B32_3:%[0-9]+]]:sreg_32_xm0(s32) = V_READFIRSTLANE_B32 [[UV1]].sub1(s64), implicit $exec
+    ; CHECK: [[MV1:%[0-9]+]]:sreg_64_xexec(s64) = G_MERGE_VALUES [[V_READFIRSTLANE_B32_2]](s32), [[V_READFIRSTLANE_B32_3]](s32)
+    ; CHECK: [[V_CMP_EQ_U64_e64_1:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U64_e64 [[MV1]](s64), [[UV1]](s64), implicit $exec
+    ; CHECK: [[S_AND_B64_:%[0-9]+]]:sreg_64_xexec = S_AND_B64 [[V_CMP_EQ_U64_e64_1]], [[V_CMP_EQ_U64_e64_]], implicit-def $scc
+    ; CHECK: [[V_READFIRSTLANE_B32_4:%[0-9]+]]:sreg_32_xm0(s32) = V_READFIRSTLANE_B32 [[UV2]].sub0(s64), implicit $exec
+    ; CHECK: [[V_READFIRSTLANE_B32_5:%[0-9]+]]:sreg_32_xm0(s32) = V_READFIRSTLANE_B32 [[UV2]].sub1(s64), implicit $exec
+    ; CHECK: [[MV2:%[0-9]+]]:sreg_64_xexec(s64) = G_MERGE_VALUES [[V_READFIRSTLANE_B32_4]](s32), [[V_READFIRSTLANE_B32_5]](s32)
+    ; CHECK: [[V_CMP_EQ_U64_e64_2:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U64_e64 [[MV2]](s64), [[UV2]](s64), implicit $exec
+    ; CHECK: [[S_AND_B64_1:%[0-9]+]]:sreg_64_xexec = S_AND_B64 [[V_CMP_EQ_U64_e64_2]], [[S_AND_B64_]], implicit-def $scc
+    ; CHECK: [[V_READFIRSTLANE_B32_6:%[0-9]+]]:sreg_32_xm0(s32) = V_READFIRSTLANE_B32 [[UV3]].sub0(s64), implicit $exec
+    ; CHECK: [[V_READFIRSTLANE_B32_7:%[0-9]+]]:sreg_32_xm0(s32) = V_READFIRSTLANE_B32 [[UV3]].sub1(s64), implicit $exec
+    ; CHECK: [[MV3:%[0-9]+]]:sreg_64_xexec(s64) = G_MERGE_VALUES [[V_READFIRSTLANE_B32_6]](s32), [[V_READFIRSTLANE_B32_7]](s32)
+    ; CHECK: [[V_CMP_EQ_U64_e64_3:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U64_e64 [[MV3]](s64), [[UV3]](s64), implicit $exec
+    ; CHECK: [[S_AND_B64_2:%[0-9]+]]:sreg_64_xexec = S_AND_B64 [[V_CMP_EQ_U64_e64_3]], [[S_AND_B64_1]], implicit-def $scc
+    ; CHECK: [[BUILD_VECTOR:%[0-9]+]]:sgpr(<8 x s32>) = G_BUILD_VECTOR [[V_READFIRSTLANE_B32_]](s32), [[V_READFIRSTLANE_B32_1]](s32), [[V_READFIRSTLANE_B32_2]](s32), [[V_READFIRSTLANE_B32_3]](s32), [[V_READFIRSTLANE_B32_4]](s32), [[V_READFIRSTLANE_B32_5]](s32), [[V_READFIRSTLANE_B32_6]](s32), [[V_READFIRSTLANE_B32_7]](s32)
+    ; CHECK: [[AMDGPU_INTRIN_IMAGE_LOAD:%[0-9]+]]:vgpr(<4 x s32>) = G_AMDGPU_INTRIN_IMAGE_LOAD intrinsic(@llvm.amdgcn.image.load.1d), 15, [[COPY1]](s32), [[BUILD_VECTOR]](<8 x s32>), 0, 0, 0 :: (dereferenceable load 16)
+    ; CHECK: [[S_AND_SAVEEXEC_B64_:%[0-9]+]]:sreg_64_xexec = S_AND_SAVEEXEC_B64 killed [[S_AND_B64_2]], implicit-def $exec, implicit-def $scc, implicit $exec
+    ; CHECK: $exec = S_XOR_B64_term $exec, [[S_AND_SAVEEXEC_B64_]], implicit-def $scc
+    ; CHECK: S_CBRANCH_EXECNZ %bb.1, implicit $exec
+    ; CHECK: .2:
+    ; CHECK: successors: %bb.3(0x80000000)
+    ; CHECK: $exec = S_MOV_B64_term [[S_MOV_B64_term]]
+    ; CHECK: .3:
+    ; CHECK: S_ENDPGM 0, implicit [[AMDGPU_INTRIN_IMAGE_LOAD]](<4 x s32>)
+    %0:_(<8 x s32>) = COPY $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7
+    %1:_(s32) = COPY $vgpr0
+    %2:_(<4 x s32>) = G_AMDGPU_INTRIN_IMAGE_LOAD intrinsic(@llvm.amdgcn.image.load.1d), 15, %1(s32), %0(<8 x s32>), 0, 0, 0 :: (dereferenceable load 16)
+    S_ENDPGM 0, implicit %2
+...