From: Matt Arsenault <Matthew.Arsenault@amd.com>
Date: Mon, 30 Nov 2015 21:15:53 +0000 (+0000)
Subject: AMDGPU: Remove SIPrepareScratchRegs
X-Git-Url: https://granicus.if.org/sourcecode?a=commitdiff_plain;h=956f59ab56fcbaf4256b6dc92dc6d8a9ac592365;p=llvm

AMDGPU: Remove SIPrepareScratchRegs

It does not work because of emergency stack slots.
This pass was supposed to eliminate dummy registers for the
spill instructions, but the register scavenger can introduce
more during PrologEpilogInserter, so some would end up
left behind if they were needed.

The potential for spilling the scratch resource descriptor
and offset register makes doing something like this
overly complicated. Reserve registers to use for the resource
descriptor and use them directly in eliminateFrameIndex.

Also removes creating another scratch resource descriptor
when directly selecting scratch MUBUF instructions.

The choice of which registers are reserved is temporary.
For now it attempts to pick the next available registers
after the user and system SGPRs.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@254329 91177308-0d34-0410-b5e6-96231b3b80d8
---

diff --git a/lib/Target/AMDGPU/AMDGPU.h b/lib/Target/AMDGPU/AMDGPU.h
index 80766086e15..a620e85101e 100644
--- a/lib/Target/AMDGPU/AMDGPU.h
+++ b/lib/Target/AMDGPU/AMDGPU.h
@@ -48,7 +48,6 @@ FunctionPass *createSIFixSGPRCopiesPass();
 FunctionPass *createSIFixSGPRLiveRangesPass();
 FunctionPass *createSICodeEmitterPass(formatted_raw_ostream &OS);
 FunctionPass *createSIInsertWaits(TargetMachine &tm);
-FunctionPass *createSIPrepareScratchRegs();
 
 ModulePass *createAMDGPUAnnotateKernelFeaturesPass();
 void initializeAMDGPUAnnotateKernelFeaturesPass(PassRegistry &);
diff --git a/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp b/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp
index 04a0c1d06af..85a06882ffe 100644
--- a/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp
+++ b/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp
@@ -1064,34 +1064,12 @@ bool AMDGPUDAGToDAGISel::SelectMUBUFScratch(SDValue Addr, SDValue &Rsrc,
   MachineFunction &MF = CurDAG->getMachineFunction();
   const SIRegisterInfo *TRI =
       static_cast<const SIRegisterInfo *>(Subtarget->getRegisterInfo());
-  MachineRegisterInfo &MRI = MF.getRegInfo();
-  const SITargetLowering& Lowering =
-    *static_cast<const SITargetLowering*>(getTargetLowering());
-
-  unsigned ScratchOffsetReg =
-      TRI->getPreloadedValue(MF, SIRegisterInfo::SCRATCH_WAVE_OFFSET);
-  Lowering.CreateLiveInRegister(*CurDAG, &AMDGPU::SReg_32RegClass,
-                                ScratchOffsetReg, MVT::i32);
-  SDValue Sym0 = CurDAG->getExternalSymbol("SCRATCH_RSRC_DWORD0", MVT::i32);
-  SDValue ScratchRsrcDword0 =
-      SDValue(CurDAG->getMachineNode(AMDGPU::S_MOV_B32, DL, MVT::i32, Sym0), 0);
-
-  SDValue Sym1 = CurDAG->getExternalSymbol("SCRATCH_RSRC_DWORD1", MVT::i32);
-  SDValue ScratchRsrcDword1 =
-      SDValue(CurDAG->getMachineNode(AMDGPU::S_MOV_B32, DL, MVT::i32, Sym1), 0);
+  const SIMachineFunctionInfo *Info = MF.getInfo<SIMachineFunctionInfo>();
 
-  const SDValue RsrcOps[] = {
-      CurDAG->getTargetConstant(AMDGPU::SReg_64RegClassID, DL, MVT::i32),
-      ScratchRsrcDword0,
-      CurDAG->getTargetConstant(AMDGPU::sub0, DL, MVT::i32),
-      ScratchRsrcDword1,
-      CurDAG->getTargetConstant(AMDGPU::sub1, DL, MVT::i32),
-  };
-  SDValue ScratchPtr = SDValue(CurDAG->getMachineNode(AMDGPU::REG_SEQUENCE, DL,
-                                              MVT::v2i32, RsrcOps), 0);
-  Rsrc = SDValue(Lowering.buildScratchRSRC(*CurDAG, DL, ScratchPtr), 0);
-  SOffset = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), DL,
-      MRI.getLiveInVirtReg(ScratchOffsetReg), MVT::i32);
+  unsigned ScratchOffsetReg
+    = TRI->getPreloadedValue(MF, SIRegisterInfo::SCRATCH_WAVE_OFFSET);
+  Rsrc = CurDAG->getRegister(Info->getScratchRSrcReg(), MVT::v4i32);
+  SOffset = CurDAG->getRegister(ScratchOffsetReg, MVT::i32);
 
   // (add n0, c1)
   if (CurDAG->isBaseWithConstantOffset(Addr)) {
diff --git a/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp b/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp
index 4e31c7ab4d4..7b0445db4df 100644
--- a/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp
+++ b/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp
@@ -327,7 +327,6 @@ void GCNPassConfig::addOptimizedRegAlloc(FunctionPass *RegAllocPass) {
 }
 
 void GCNPassConfig::addPostRegAlloc() {
-  addPass(createSIPrepareScratchRegs(), false);
   addPass(createSIShrinkInstructionsPass(), false);
 }
 
diff --git a/lib/Target/AMDGPU/CMakeLists.txt b/lib/Target/AMDGPU/CMakeLists.txt
index 7a4b5bb6d35..64c9e1882e4 100644
--- a/lib/Target/AMDGPU/CMakeLists.txt
+++ b/lib/Target/AMDGPU/CMakeLists.txt
@@ -57,7 +57,6 @@ add_llvm_target(AMDGPUCodeGen
   SILowerControlFlow.cpp
   SILowerI1Copies.cpp
   SIMachineFunctionInfo.cpp
-  SIPrepareScratchRegs.cpp
   SIRegisterInfo.cpp
   SIShrinkInstructions.cpp
   SITypeRewriter.cpp
diff --git a/lib/Target/AMDGPU/SIFrameLowering.cpp b/lib/Target/AMDGPU/SIFrameLowering.cpp
index a2d8fa1b0a1..6aff4b5700d 100644
--- a/lib/Target/AMDGPU/SIFrameLowering.cpp
+++ b/lib/Target/AMDGPU/SIFrameLowering.cpp
@@ -8,17 +8,89 @@
 //==-----------------------------------------------------------------------===//
 
 #include "SIFrameLowering.h"
+#include "SIInstrInfo.h"
+#include "SIMachineFunctionInfo.h"
 #include "SIRegisterInfo.h"
 #include "llvm/CodeGen/MachineFrameInfo.h"
 #include "llvm/CodeGen/MachineFunction.h"
+#include "llvm/CodeGen/MachineInstrBuilder.h"
 #include "llvm/CodeGen/RegisterScavenging.h"
 
 using namespace llvm;
 
+
+static bool hasOnlySGPRSpills(const SIMachineFunctionInfo *FuncInfo,
+                              const MachineFrameInfo *FrameInfo) {
+  if (!FuncInfo->hasSpilledSGPRs())
+    return false;
+
+  if (FuncInfo->hasSpilledVGPRs())
+    return false;
+
+  for (int I = FrameInfo->getObjectIndexBegin(),
+         E = FrameInfo->getObjectIndexEnd(); I != E; ++I) {
+    if (!FrameInfo->isSpillSlotObjectIndex(I))
+      return false;
+  }
+
+  return true;
+}
+
+void SIFrameLowering::emitPrologue(MachineFunction &MF,
+                                   MachineBasicBlock &MBB) const {
+  if (!MF.getFrameInfo()->hasStackObjects())
+    return;
+
+  assert(&MF.front() == &MBB && "Shrink-wrapping not yet supported");
+
+  const SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
+
+  // If we only have SGPR spills, we won't actually be using scratch memory
+  // since these spill to VGPRs.
+  //
+  // FIXME: We should be cleaning up these unused SGPR spill frame indices
+  // somewhere.
+  if (hasOnlySGPRSpills(MFI, MF.getFrameInfo()))
+    return;
+
+  const SIInstrInfo *TII =
+      static_cast<const SIInstrInfo *>(MF.getSubtarget().getInstrInfo());
+  const SIRegisterInfo *TRI = &TII->getRegisterInfo();
+
+  // We need to insert initialization of the scratch resource descriptor.
+  unsigned ScratchRsrcReg = MFI->getScratchRSrcReg();
+  assert(ScratchRsrcReg != AMDGPU::NoRegister);
+
+  uint64_t Rsrc23 = TII->getScratchRsrcWords23();
+  MachineBasicBlock::iterator I = MBB.begin();
+  DebugLoc DL;
+
+  unsigned Rsrc0 = TRI->getSubReg(ScratchRsrcReg, AMDGPU::sub0);
+  unsigned Rsrc1 = TRI->getSubReg(ScratchRsrcReg, AMDGPU::sub1);
+  unsigned Rsrc2 = TRI->getSubReg(ScratchRsrcReg, AMDGPU::sub2);
+  unsigned Rsrc3 = TRI->getSubReg(ScratchRsrcReg, AMDGPU::sub3);
+
+  BuildMI(MBB, I, DL, TII->get(AMDGPU::S_MOV_B32), Rsrc0)
+    .addExternalSymbol("SCRATCH_RSRC_DWORD0");
+
+  BuildMI(MBB, I, DL, TII->get(AMDGPU::S_MOV_B32), Rsrc1)
+    .addExternalSymbol("SCRATCH_RSRC_DWORD1");
+
+  BuildMI(MBB, I, DL, TII->get(AMDGPU::S_MOV_B32), Rsrc2)
+    .addImm(Rsrc23 & 0xffffffff);
+
+  BuildMI(MBB, I, DL, TII->get(AMDGPU::S_MOV_B32), Rsrc3)
+    .addImm(Rsrc23 >> 32);
+}
+
 void SIFrameLowering::processFunctionBeforeFrameFinalized(
   MachineFunction &MF,
   RegScavenger *RS) const {
   MachineFrameInfo *MFI = MF.getFrameInfo();
+
+  if (!MFI->hasStackObjects())
+    return;
+
   bool MayNeedScavengingEmergencySlot = MFI->hasStackObjects();
 
   assert((RS || !MayNeedScavengingEmergencySlot) &&
diff --git a/lib/Target/AMDGPU/SIFrameLowering.h b/lib/Target/AMDGPU/SIFrameLowering.h
index 677128d6ce0..a9152fd8b2a 100644
--- a/lib/Target/AMDGPU/SIFrameLowering.h
+++ b/lib/Target/AMDGPU/SIFrameLowering.h
@@ -21,6 +21,9 @@ public:
     AMDGPUFrameLowering(D, StackAl, LAO, TransAl) {}
   ~SIFrameLowering() override {}
 
+  void emitPrologue(MachineFunction &MF,
+                    MachineBasicBlock &MBB) const override;
+
   void processFunctionBeforeFrameFinalized(
     MachineFunction &MF,
     RegScavenger *RS = nullptr) const override;
diff --git a/lib/Target/AMDGPU/SIISelLowering.cpp b/lib/Target/AMDGPU/SIISelLowering.cpp
index 94fad32c007..51cbc95bc07 100644
--- a/lib/Target/AMDGPU/SIISelLowering.cpp
+++ b/lib/Target/AMDGPU/SIISelLowering.cpp
@@ -552,6 +552,7 @@ SDValue SITargetLowering::LowerFormalArguments(
   MachineFunction &MF = DAG.getMachineFunction();
   FunctionType *FType = MF.getFunction()->getFunctionType();
   SIMachineFunctionInfo *Info = MF.getInfo<SIMachineFunctionInfo>();
+  const AMDGPUSubtarget &ST = MF.getSubtarget<AMDGPUSubtarget>();
 
   if (Subtarget->isAmdHsaOS() && Info->getShaderType() != ShaderType::COMPUTE) {
     const Function *Fn = MF.getFunction();
@@ -622,9 +623,9 @@ SDValue SITargetLowering::LowerFormalArguments(
   // The pointer to the scratch buffer is stored in SGPR2, SGPR3
   if (Info->getShaderType() == ShaderType::COMPUTE) {
     if (Subtarget->isAmdHsaOS())
-      Info->NumUserSGPRs = 2;  // FIXME: Need to support scratch buffers.
+      Info->NumUserSGPRs += 4;  // FIXME: Need to support scratch buffers.
     else
-      Info->NumUserSGPRs = 4;
+      Info->NumUserSGPRs += 4;
 
     unsigned InputPtrReg =
         TRI->getPreloadedValue(MF, SIRegisterInfo::INPUT_PTR);
@@ -750,6 +751,9 @@ SDValue SITargetLowering::LowerFormalArguments(
     Info->ScratchOffsetReg = AMDGPU::SGPR_32RegClass.getRegister(ScratchIdx);
   }
 
+  if (MF.getFrameInfo()->hasStackObjects() || ST.isVGPRSpillingEnabled(Info))
+    Info->setScratchRSrcReg(TRI);
+
   if (Chains.empty())
     return Chain;
 
@@ -2335,15 +2339,6 @@ MachineSDNode *SITargetLowering::buildRSRC(SelectionDAG &DAG,
   return DAG.getMachineNode(AMDGPU::REG_SEQUENCE, DL, MVT::v4i32, Ops);
 }
 
-MachineSDNode *SITargetLowering::buildScratchRSRC(SelectionDAG &DAG,
-                                                  SDLoc DL,
-                                                  SDValue Ptr) const {
-  const SIInstrInfo *TII =
-      static_cast<const SIInstrInfo *>(Subtarget->getInstrInfo());
-
-  return buildRSRC(DAG, DL, Ptr, 0, TII->getScratchRsrcWords23());
-}
-
 SDValue SITargetLowering::CreateLiveInRegister(SelectionDAG &DAG,
                                                const TargetRegisterClass *RC,
                                                unsigned Reg, EVT VT) const {
diff --git a/lib/Target/AMDGPU/SIISelLowering.h b/lib/Target/AMDGPU/SIISelLowering.h
index a358e3fc3c0..0659dd7d5d0 100644
--- a/lib/Target/AMDGPU/SIISelLowering.h
+++ b/lib/Target/AMDGPU/SIISelLowering.h
@@ -116,10 +116,6 @@ public:
                            SDValue Ptr,
                            uint32_t RsrcDword1,
                            uint64_t RsrcDword2And3) const;
-  MachineSDNode *buildScratchRSRC(SelectionDAG &DAG,
-                                  SDLoc DL,
-                                  SDValue Ptr) const;
-
   std::pair<unsigned, const TargetRegisterClass *>
   getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI,
                                StringRef Constraint, MVT VT) const override;
diff --git a/lib/Target/AMDGPU/SIInstrInfo.cpp b/lib/Target/AMDGPU/SIInstrInfo.cpp
index 9a85a1d515f..b7d2a471275 100644
--- a/lib/Target/AMDGPU/SIInstrInfo.cpp
+++ b/lib/Target/AMDGPU/SIInstrInfo.cpp
@@ -551,15 +551,16 @@ void SIInstrInfo::storeRegToStackSlot(MachineBasicBlock &MBB,
 
   assert(RI.hasVGPRs(RC) && "Only VGPR spilling expected");
 
+  unsigned ScratchOffsetPreloadReg
+    = RI.getPreloadedValue(*MF, SIRegisterInfo::SCRATCH_WAVE_OFFSET);
+
   unsigned Opcode = getVGPRSpillSaveOpcode(RC->getSize());
   MFI->setHasSpilledVGPRs();
   BuildMI(MBB, MI, DL, get(Opcode))
     .addReg(SrcReg)                   // src
     .addFrameIndex(FrameIndex)        // frame_idx
-    // Place-holder registers, these will be filled in by
-    // SIPrepareScratchRegs.
-    .addReg(AMDGPU::SGPR0_SGPR1_SGPR2_SGPR3, RegState::Undef)
-    .addReg(AMDGPU::SGPR0, RegState::Undef)
+    .addReg(MFI->getScratchRSrcReg()) // scratch_rsrc
+    .addReg(ScratchOffsetPreloadReg)  // scratch_offset
     .addMemOperand(MMO);
 }
 
@@ -637,13 +638,14 @@ void SIInstrInfo::loadRegFromStackSlot(MachineBasicBlock &MBB,
 
   assert(RI.hasVGPRs(RC) && "Only VGPR spilling expected");
 
+  unsigned ScratchOffsetPreloadReg
+    = RI.getPreloadedValue(*MF, SIRegisterInfo::SCRATCH_WAVE_OFFSET);
+
   unsigned Opcode = getVGPRSpillRestoreOpcode(RC->getSize());
   BuildMI(MBB, MI, DL, get(Opcode), DestReg)
     .addFrameIndex(FrameIndex)        // frame_idx
-    // Place-holder registers, these will be filled in by
-    // SIPrepareScratchRegs.
-    .addReg(AMDGPU::SGPR0_SGPR1_SGPR2_SGPR3, RegState::Undef)
-    .addReg(AMDGPU::SGPR0, RegState::Undef)
+    .addReg(MFI->getScratchRSrcReg()) // scratch_rsrc
+    .addReg(ScratchOffsetPreloadReg)  // scratch_offset
     .addMemOperand(MMO);
 }
 
diff --git a/lib/Target/AMDGPU/SIMachineFunctionInfo.cpp b/lib/Target/AMDGPU/SIMachineFunctionInfo.cpp
index 6269dce553f..d042844aa13 100644
--- a/lib/Target/AMDGPU/SIMachineFunctionInfo.cpp
+++ b/lib/Target/AMDGPU/SIMachineFunctionInfo.cpp
@@ -68,6 +68,14 @@ SIMachineFunctionInfo::SIMachineFunctionInfo(const MachineFunction &MF)
     WorkItemIDZ = true;
 }
 
+void SIMachineFunctionInfo::setScratchRSrcReg(const SIRegisterInfo *TRI) {
+  // We need to round up to next multiple of 4.
+  unsigned NextSReg128 = RoundUpToAlignment(NumUserSGPRs + 5, 4);
+  unsigned RegSub0 = AMDGPU::SReg_32RegClass.getRegister(NextSReg128);
+  ScratchRSrcReg = TRI->getMatchingSuperReg(RegSub0, AMDGPU::sub0,
+                                            &AMDGPU::SReg_128RegClass);
+}
+
 SIMachineFunctionInfo::SpilledReg SIMachineFunctionInfo::getSpilledReg(
                                                        MachineFunction *MF,
                                                        unsigned FrameIndex,
diff --git a/lib/Target/AMDGPU/SIPrepareScratchRegs.cpp b/lib/Target/AMDGPU/SIPrepareScratchRegs.cpp
deleted file mode 100644
index a6c22775e09..00000000000
--- a/lib/Target/AMDGPU/SIPrepareScratchRegs.cpp
+++ /dev/null
@@ -1,196 +0,0 @@
-//===-- SIPrepareScratchRegs.cpp - Use predicates for control flow --------===//
-//
-//                     The LLVM Compiler Infrastructure
-//
-// This file is distributed under the University of Illinois Open Source
-// License. See LICENSE.TXT for details.
-//
-//===----------------------------------------------------------------------===//
-//
-/// \file
-///
-/// This pass loads scratch pointer and scratch offset into a register or a
-/// frame index which can be used anywhere in the program.  These values will
-/// be used for spilling VGPRs.
-///
-//===----------------------------------------------------------------------===//
-
-#include "AMDGPU.h"
-#include "AMDGPUSubtarget.h"
-#include "SIDefines.h"
-#include "SIInstrInfo.h"
-#include "SIMachineFunctionInfo.h"
-#include "llvm/CodeGen/MachineFrameInfo.h"
-#include "llvm/CodeGen/MachineFunction.h"
-#include "llvm/CodeGen/MachineFunctionPass.h"
-#include "llvm/CodeGen/MachineInstrBuilder.h"
-#include "llvm/CodeGen/MachineRegisterInfo.h"
-#include "llvm/CodeGen/RegisterScavenging.h"
-#include "llvm/IR/Function.h"
-#include "llvm/IR/LLVMContext.h"
-
-using namespace llvm;
-
-namespace {
-
-class SIPrepareScratchRegs : public MachineFunctionPass {
-
-private:
-  static char ID;
-
-public:
-  SIPrepareScratchRegs() : MachineFunctionPass(ID) { }
-
-  bool runOnMachineFunction(MachineFunction &MF) override;
-
-  const char *getPassName() const override {
-    return "SI prepare scratch registers";
-  }
-
-  void getAnalysisUsage(AnalysisUsage &AU) const override {
-    AU.setPreservesCFG();
-    MachineFunctionPass::getAnalysisUsage(AU);
-  }
-};
-
-} // End anonymous namespace
-
-char SIPrepareScratchRegs::ID = 0;
-
-FunctionPass *llvm::createSIPrepareScratchRegs() {
-  return new SIPrepareScratchRegs();
-}
-
-bool SIPrepareScratchRegs::runOnMachineFunction(MachineFunction &MF) {
-  SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
-  const SIInstrInfo *TII =
-      static_cast<const SIInstrInfo *>(MF.getSubtarget().getInstrInfo());
-  const SIRegisterInfo *TRI = &TII->getRegisterInfo();
-  MachineRegisterInfo &MRI = MF.getRegInfo();
-  MachineFrameInfo *FrameInfo = MF.getFrameInfo();
-  MachineBasicBlock *Entry = &MF.front();
-  MachineBasicBlock::iterator I = Entry->begin();
-  DebugLoc DL = I->getDebugLoc();
-
-  // FIXME: If we don't have enough VGPRs for SGPR spilling we will need to
-  // run this pass.
-  if (!MFI->hasSpilledVGPRs())
-    return false;
-
-  unsigned ScratchPtrPreloadReg =
-      TRI->getPreloadedValue(MF, SIRegisterInfo::SCRATCH_PTR);
-  unsigned ScratchOffsetPreloadReg =
-      TRI->getPreloadedValue(MF, SIRegisterInfo::SCRATCH_WAVE_OFFSET);
-
-  if (!Entry->isLiveIn(ScratchPtrPreloadReg))
-    Entry->addLiveIn(ScratchPtrPreloadReg);
-
-  if (!Entry->isLiveIn(ScratchOffsetPreloadReg))
-    Entry->addLiveIn(ScratchOffsetPreloadReg);
-
-  // Load the scratch offset.
-  unsigned ScratchOffsetReg =
-      TRI->findUnusedRegister(MRI, &AMDGPU::SGPR_32RegClass);
-  int ScratchOffsetFI = -1;
-
-  if (ScratchOffsetReg != AMDGPU::NoRegister) {
-    // Found an SGPR to use
-    BuildMI(*Entry, I, DL, TII->get(AMDGPU::S_MOV_B32), ScratchOffsetReg)
-            .addReg(ScratchOffsetPreloadReg);
-  } else {
-    // No SGPR is available, we must spill.
-    ScratchOffsetFI = FrameInfo->CreateSpillStackObject(4,4);
-    BuildMI(*Entry, I, DL, TII->get(AMDGPU::SI_SPILL_S32_SAVE))
-            .addReg(ScratchOffsetPreloadReg)
-            .addFrameIndex(ScratchOffsetFI)
-            .addReg(AMDGPU::SGPR0_SGPR1_SGPR2_SGPR3, RegState::Undef)
-            .addReg(AMDGPU::SGPR0, RegState::Undef);
-  }
-
-
-  // Now that we have the scratch pointer and offset values, we need to
-  // add them to all the SI_SPILL_V* instructions.
-
-  RegScavenger RS;
-  unsigned ScratchRsrcFI = FrameInfo->CreateSpillStackObject(16, 4);
-  RS.addScavengingFrameIndex(ScratchRsrcFI);
-
-  for (MachineFunction::iterator BI = MF.begin(), BE = MF.end();
-       BI != BE; ++BI) {
-
-    MachineBasicBlock &MBB = *BI;
-    // Add the scratch offset reg as a live-in so that the register scavenger
-    // doesn't re-use it.
-    if (!MBB.isLiveIn(ScratchOffsetReg) &&
-        ScratchOffsetReg != AMDGPU::NoRegister)
-      MBB.addLiveIn(ScratchOffsetReg);
-    RS.enterBasicBlock(&MBB);
-
-    for (MachineBasicBlock::iterator I = MBB.begin(), E = MBB.end();
-         I != E; ++I) {
-      MachineInstr &MI = *I;
-      RS.forward(I);
-      DebugLoc DL = MI.getDebugLoc();
-      if (!TII->isVGPRSpill(MI))
-        continue;
-
-      // Scratch resource
-      unsigned ScratchRsrcReg =
-          RS.scavengeRegister(&AMDGPU::SReg_128RegClass, 0);
-
-      uint64_t Rsrc23 = TII->getScratchRsrcWords23();
-
-      unsigned Rsrc0 = TRI->getSubReg(ScratchRsrcReg, AMDGPU::sub0);
-      unsigned Rsrc1 = TRI->getSubReg(ScratchRsrcReg, AMDGPU::sub1);
-      unsigned Rsrc2 = TRI->getSubReg(ScratchRsrcReg, AMDGPU::sub2);
-      unsigned Rsrc3 = TRI->getSubReg(ScratchRsrcReg, AMDGPU::sub3);
-
-      BuildMI(MBB, I, DL, TII->get(AMDGPU::S_MOV_B32), Rsrc0)
-              .addExternalSymbol("SCRATCH_RSRC_DWORD0")
-              .addReg(ScratchRsrcReg, RegState::ImplicitDefine);
-
-      BuildMI(MBB, I, DL, TII->get(AMDGPU::S_MOV_B32), Rsrc1)
-              .addExternalSymbol("SCRATCH_RSRC_DWORD1")
-              .addReg(ScratchRsrcReg, RegState::ImplicitDefine);
-
-      BuildMI(MBB, I, DL, TII->get(AMDGPU::S_MOV_B32), Rsrc2)
-              .addImm(Rsrc23 & 0xffffffff)
-              .addReg(ScratchRsrcReg, RegState::ImplicitDefine);
-
-      BuildMI(MBB, I, DL, TII->get(AMDGPU::S_MOV_B32), Rsrc3)
-              .addImm(Rsrc23 >> 32)
-              .addReg(ScratchRsrcReg, RegState::ImplicitDefine);
-
-      // Scratch Offset
-      if (ScratchOffsetReg == AMDGPU::NoRegister) {
-        ScratchOffsetReg = RS.scavengeRegister(&AMDGPU::SGPR_32RegClass, 0);
-        BuildMI(MBB, I, DL, TII->get(AMDGPU::SI_SPILL_S32_RESTORE),
-                ScratchOffsetReg)
-                .addFrameIndex(ScratchOffsetFI)
-                .addReg(AMDGPU::SGPR0_SGPR1_SGPR2_SGPR3, RegState::Undef)
-                .addReg(AMDGPU::SGPR0, RegState::Undef);
-      } else if (!MBB.isLiveIn(ScratchOffsetReg)) {
-        MBB.addLiveIn(ScratchOffsetReg);
-      }
-
-      if (ScratchRsrcReg == AMDGPU::NoRegister ||
-          ScratchOffsetReg == AMDGPU::NoRegister) {
-        LLVMContext &Ctx = MF.getFunction()->getContext();
-        Ctx.emitError("ran out of SGPRs for spilling VGPRs");
-        ScratchRsrcReg = AMDGPU::SGPR0;
-        ScratchOffsetReg = AMDGPU::SGPR0;
-      }
-      MI.getOperand(2).setReg(ScratchRsrcReg);
-      MI.getOperand(2).setIsKill(true);
-      MI.getOperand(2).setIsUndef(false);
-      MI.getOperand(3).setReg(ScratchOffsetReg);
-      MI.getOperand(3).setIsUndef(false);
-      MI.getOperand(3).setIsKill(false);
-      MI.addOperand(MachineOperand::CreateReg(Rsrc0, false, true, true));
-      MI.addOperand(MachineOperand::CreateReg(Rsrc1, false, true, true));
-      MI.addOperand(MachineOperand::CreateReg(Rsrc2, false, true, true));
-      MI.addOperand(MachineOperand::CreateReg(Rsrc3, false, true, true));
-    }
-  }
-  return true;
-}
diff --git a/lib/Target/AMDGPU/SIRegisterInfo.cpp b/lib/Target/AMDGPU/SIRegisterInfo.cpp
index ab7539b6fb3..b392c86fa2e 100644
--- a/lib/Target/AMDGPU/SIRegisterInfo.cpp
+++ b/lib/Target/AMDGPU/SIRegisterInfo.cpp
@@ -68,6 +68,22 @@ BitVector SIRegisterInfo::getReservedRegs(const MachineFunction &MF) const {
     }
   }
 
+  const SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
+  unsigned ScratchRSrcReg = MFI->getScratchRSrcReg();
+  if (ScratchRSrcReg != AMDGPU::NoRegister) {
+    unsigned ScratchOffsetPreloadReg
+      = getPreloadedValue(MF, SIRegisterInfo::SCRATCH_WAVE_OFFSET);
+    // We will need to use this user SGPR argument for spilling, and thus never
+    // want it to be spilled.
+    reserveRegisterTuples(Reserved, ScratchOffsetPreloadReg);
+
+    // Reserve 4 SGPRs for the scratch buffer resource descriptor in case we need
+    // to spill.
+    // TODO: May need to reserve a VGPR if doing LDS spilling.
+    reserveRegisterTuples(Reserved, ScratchRSrcReg);
+    assert(!isSubRegister(ScratchRSrcReg, ScratchOffsetPreloadReg));
+  }
+
   return Reserved;
 }
 
@@ -243,6 +259,9 @@ void SIRegisterInfo::eliminateFrameIndex(MachineBasicBlock::iterator MI,
                 .addReg(SubReg)
                 .addImm(Spill.Lane);
 
+        // FIXME: Since this spills to another register instead of an actual
+        // frame index, we should delete the frame index when all references to
+        // it are fixed.
       }
       MI->eraseFromParent();
       break;
diff --git a/test/CodeGen/AMDGPU/kernel-args.ll b/test/CodeGen/AMDGPU/kernel-args.ll
index 803b2ecced0..e9d98ac89e7 100644
--- a/test/CodeGen/AMDGPU/kernel-args.ll
+++ b/test/CodeGen/AMDGPU/kernel-args.ll
@@ -294,8 +294,8 @@ entry:
 ; EG-DAG: T{{[0-9]\.[XYZW]}}, KC0[5].Z
 ; EG-DAG: T{{[0-9]\.[XYZW]}}, KC0[5].W
 ; EG-DAG: T{{[0-9]\.[XYZW]}}, KC0[6].X
-; SI: s_load_dwordx8 s{{\[[0-9]:[0-9]+\]}}, s[0:1], 0x11
-; VI: s_load_dwordx8 s{{\[[0-9]:[0-9]+\]}}, s[0:1], 0x44
+; SI: s_load_dwordx8 s{{\[[0-9]+:[0-9]+\]}}, s[0:1], 0x11
+; VI: s_load_dwordx8 s{{\[[0-9]+:[0-9]+\]}}, s[0:1], 0x44
 define void @v8i32_arg(<8 x i32> addrspace(1)* nocapture %out, <8 x i32> %in) nounwind {
 entry:
   store <8 x i32> %in, <8 x i32> addrspace(1)* %out, align 4
@@ -311,7 +311,7 @@ entry:
 ; EG-DAG: T{{[0-9]\.[XYZW]}}, KC0[5].Z
 ; EG-DAG: T{{[0-9]\.[XYZW]}}, KC0[5].W
 ; EG-DAG: T{{[0-9]\.[XYZW]}}, KC0[6].X
-; SI: s_load_dwordx8 s{{\[[0-9]:[0-9]+\]}}, s[0:1], 0x11
+; SI: s_load_dwordx8 s{{\[[0-9]+:[0-9]+\]}}, s[0:1], 0x11
 define void @v8f32_arg(<8 x float> addrspace(1)* nocapture %out, <8 x float> %in) nounwind {
 entry:
   store <8 x float> %in, <8 x float> addrspace(1)* %out, align 4
@@ -413,8 +413,8 @@ entry:
 ; EG-DAG: T{{[0-9]\.[XYZW]}}, KC0[9].Z
 ; EG-DAG: T{{[0-9]\.[XYZW]}}, KC0[9].W
 ; EG-DAG: T{{[0-9]\.[XYZW]}}, KC0[10].X
-; SI: s_load_dwordx16 s{{\[[0-9]:[0-9]+\]}}, s[0:1], 0x19
-; VI: s_load_dwordx16 s{{\[[0-9]:[0-9]+\]}}, s[0:1], 0x64
+; SI: s_load_dwordx16 s{{\[[0-9]+:[0-9]+\]}}, s[0:1], 0x19
+; VI: s_load_dwordx16 s{{\[[0-9]+:[0-9]+\]}}, s[0:1], 0x64
 define void @v16i32_arg(<16 x i32> addrspace(1)* nocapture %out, <16 x i32> %in) nounwind {
 entry:
   store <16 x i32> %in, <16 x i32> addrspace(1)* %out, align 4
@@ -438,8 +438,8 @@ entry:
 ; EG-DAG: T{{[0-9]\.[XYZW]}}, KC0[9].Z
 ; EG-DAG: T{{[0-9]\.[XYZW]}}, KC0[9].W
 ; EG-DAG: T{{[0-9]\.[XYZW]}}, KC0[10].X
-; SI: s_load_dwordx16 s{{\[[0-9]:[0-9]+\]}}, s[0:1], 0x19
-; VI: s_load_dwordx16 s{{\[[0-9]:[0-9]+\]}}, s[0:1], 0x64
+; SI: s_load_dwordx16 s{{\[[0-9]+:[0-9]+\]}}, s[0:1], 0x19
+; VI: s_load_dwordx16 s{{\[[0-9]+:[0-9]+\]}}, s[0:1], 0x64
 define void @v16f32_arg(<16 x float> addrspace(1)* nocapture %out, <16 x float> %in) nounwind {
 entry:
   store <16 x float> %in, <16 x float> addrspace(1)* %out, align 4
diff --git a/test/CodeGen/AMDGPU/large-alloca-compute.ll b/test/CodeGen/AMDGPU/large-alloca-compute.ll
new file mode 100644
index 00000000000..5e8cf5bb3d2
--- /dev/null
+++ b/test/CodeGen/AMDGPU/large-alloca-compute.ll
@@ -0,0 +1,42 @@
+; RUN: llc -march=amdgcn -mcpu=bonaire < %s | FileCheck -check-prefix=GCN -check-prefix=CI -check-prefix=ALL %s
+; RUN: llc -march=amdgcn -mcpu=tonga < %s | FileCheck -check-prefix=GCN -check-prefix=VI -check-prefix=ALL %s
+; XUN: llc -march=amdgcn -mcpu=bonaire -mtriple=amdgcn-unknown-amdhsa < %s | FileCheck -check-prefix=GCNHSA -check-prefix=CIHSA -check-prefix=ALL %s
+; XUN: llc -march=amdgcn -mcpu=tonga -mtriple=amdgcn-unknown-amdhsa < %s | FileCheck -check-prefix=GCNHSA -check-prefix=VIHSA -check-prefix=ALL %s
+
+; FIXME: align on alloca seems to be ignored for private_segment_alignment
+
+; ALL-LABEL: {{^}}large_alloca_compute_shader:
+
+; GCN: s_mov_b32 s12, SCRATCH_RSRC_DWORD0
+; GCN: s_mov_b32 s13, SCRATCH_RSRC_DWORD1
+; GCN: s_mov_b32 s14, -1
+; CI: s_mov_b32 s15, 0x80f000
+; VI: s_mov_b32 s15, 0x800000
+
+
+; GCNHSA: .amd_kernel_code_t
+; GCNHSA: private_segment_alignment = 4
+; GCNHSA: .end_amd_kernel_code_t
+
+; GCNHSA: s_mov_b32 s8, SCRATCH_RSRC_DWORD0
+; GCNHSA: s_mov_b32 s9, SCRATCH_RSRC_DWORD1
+; GCNHSA: s_mov_b32 s10, -1
+; CIHSA: s_mov_b32 s11, 0x180f000
+; VIHSA: s_mov_b32 s11, 0x11800000
+
+; GCNHSA: buffer_store_dword {{v[0-9]+}}, {{v[0-9]+}}, s[8:11], s6 offen
+; GCNHSA: buffer_load_dword {{v[0-9]+}}, {{v[0-9]+}}, s[8:11], s6 offen
+
+; Scratch size = alloca size + emergency stack slot
+; ALL: ; ScratchSize: 32772
+define void @large_alloca_compute_shader(i32 %x, i32 %y) #0 {
+  %large = alloca [8192 x i32], align 4
+  %gep = getelementptr [8192 x i32], [8192 x i32]* %large, i32 0, i32 8191
+  store volatile i32 %x, i32* %gep
+  %gep1 = getelementptr [8192 x i32], [8192 x i32]* %large, i32 0, i32 %y
+  %val = load volatile i32, i32* %gep1
+  store volatile i32 %val, i32 addrspace(1)* undef
+  ret void
+}
+
+attributes #0 = { nounwind  }
diff --git a/test/CodeGen/AMDGPU/large-alloca-graphics.ll b/test/CodeGen/AMDGPU/large-alloca-graphics.ll
new file mode 100644
index 00000000000..208b9a10050
--- /dev/null
+++ b/test/CodeGen/AMDGPU/large-alloca-graphics.ll
@@ -0,0 +1,47 @@
+; RUN: llc -march=amdgcn -mcpu=bonaire < %s | FileCheck -check-prefix=GCN -check-prefix=CI -check-prefix=ALL %s
+; RUN: llc -march=amdgcn -mcpu=tonga < %s | FileCheck -check-prefix=GCN -check-prefix=VI -check-prefix=ALL %s
+
+; ALL-LABEL: {{^}}large_alloca_pixel_shader:
+; GCN: s_mov_b32 s8, SCRATCH_RSRC_DWORD0
+; GCN: s_mov_b32 s9, SCRATCH_RSRC_DWORD1
+; GCN: s_mov_b32 s10, -1
+; CI: s_mov_b32 s11, 0x80f000
+; VI: s_mov_b32 s11, 0x800000
+
+; GCN: buffer_store_dword {{v[0-9]+}}, {{v[0-9]+}}, s[8:11], s0 offen
+; GCN: buffer_load_dword {{v[0-9]+}}, {{v[0-9]+}}, s[8:11], s0 offen
+
+; ALL: ; ScratchSize: 32772
+define void @large_alloca_pixel_shader(i32 %x, i32 %y) #1 {
+  %large = alloca [8192 x i32], align 4
+  %gep = getelementptr [8192 x i32], [8192 x i32]* %large, i32 0, i32 8191
+  store volatile i32 %x, i32* %gep
+  %gep1 = getelementptr [8192 x i32], [8192 x i32]* %large, i32 0, i32 %y
+  %val = load volatile i32, i32* %gep1
+  store volatile i32 %val, i32 addrspace(1)* undef
+  ret void
+}
+
+; ALL-LABEL: {{^}}large_alloca_pixel_shader_inreg:
+; GCN: s_mov_b32 s8, SCRATCH_RSRC_DWORD0
+; GCN: s_mov_b32 s9, SCRATCH_RSRC_DWORD1
+; GCN: s_mov_b32 s10, -1
+; CI: s_mov_b32 s11, 0x80f000
+; VI: s_mov_b32 s11, 0x800000
+
+; GCN: buffer_store_dword {{v[0-9]+}}, {{v[0-9]+}}, s[8:11], s2 offen
+; GCN: buffer_load_dword {{v[0-9]+}}, {{v[0-9]+}}, s[8:11], s2 offen
+
+; ALL: ; ScratchSize: 32772
+define void @large_alloca_pixel_shader_inreg(i32 inreg %x, i32 inreg %y) #1 {
+  %large = alloca [8192 x i32], align 4
+  %gep = getelementptr [8192 x i32], [8192 x i32]* %large, i32 0, i32 8191
+  store volatile i32 %x, i32* %gep
+  %gep1 = getelementptr [8192 x i32], [8192 x i32]* %large, i32 0, i32 %y
+  %val = load volatile i32, i32* %gep1
+  store volatile i32 %val, i32 addrspace(1)* undef
+  ret void
+}
+
+attributes #0 = { nounwind  }
+attributes #1 = { nounwind "ShaderType"="0" }
diff --git a/test/CodeGen/AMDGPU/large-alloca.ll b/test/CodeGen/AMDGPU/large-alloca.ll
deleted file mode 100644
index e1122da78ef..00000000000
--- a/test/CodeGen/AMDGPU/large-alloca.ll
+++ /dev/null
@@ -1,18 +0,0 @@
-; RUN: llc -march=amdgcn -mcpu=tahiti -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=SI %s
-; RUN: llc -march=amdgcn -mcpu=tonga -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=VI %s
-
-; GCN-LABEL: {{^}}large_alloca:
-; GCN: buffer_store_dword v{{[0-9]+}}, v{{[0-9]+}}, s{{\[[0-9]+:[0-9]+\]}}, s{{[0-9]+}} offen
-; GCN: buffer_load_dword v{{[0-9]+}}, v{{[0-9]+}}, s{{\[[0-9]+:[0-9]+\]}}, s{{[0-9]+}} offen
-; GCN: ScratchSize: 32776
-define void @large_alloca(i32 addrspace(1)* %out, i32 %x, i32 %y) #0 {
-  %large = alloca [8192 x i32], align 4
-  %gep = getelementptr [8192 x i32], [8192 x i32]* %large, i32 0, i32 8191
-  store i32 %x, i32* %gep
-  %gep1 = getelementptr [8192 x i32], [8192 x i32]* %large, i32 0, i32 %y
-  %load = load i32, i32* %gep1
-  store i32 %load, i32 addrspace(1)* %out
-  ret void
-}
-
-attributes #0 = { nounwind }
diff --git a/test/CodeGen/AMDGPU/si-sgpr-spill.ll b/test/CodeGen/AMDGPU/si-sgpr-spill.ll
index 84652701f77..d7b35fc631e 100644
--- a/test/CodeGen/AMDGPU/si-sgpr-spill.ll
+++ b/test/CodeGen/AMDGPU/si-sgpr-spill.ll
@@ -6,6 +6,16 @@
 
 ; CHECK-LABEL: {{^}}main:
 ; CHECK: s_wqm
+
+; Make sure not emitting unused scratch resource descriptor setup
+; CHECK-NOT: s_mov_b32
+; CHECK-NOT: s_mov_b32
+; CHECK-NOT: s_mov_b32
+; CHECK-NOT: s_mov_b32
+
+; CHECK: s_mov_b32 m0
+
+
 ; Writing to M0 from an SMRD instruction will hang the GPU.
 ; CHECK-NOT: s_buffer_load_dword m0
 ; CHECK: s_endpgm
diff --git a/test/CodeGen/AMDGPU/vgpr-spill-emergency-stack-slot-compute.ll b/test/CodeGen/AMDGPU/vgpr-spill-emergency-stack-slot-compute.ll
new file mode 100644
index 00000000000..2cbb67d85fb
--- /dev/null
+++ b/test/CodeGen/AMDGPU/vgpr-spill-emergency-stack-slot-compute.ll
@@ -0,0 +1,583 @@
+; RUN: llc -march=amdgcn -mcpu=tahiti -mattr=+vgpr-spilling -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=SI %s
+; RUN: llc -march=amdgcn -mcpu=fiji -mattr=+vgpr-spilling -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=VI %s
+; XUN: llc -march=amdgcn -mcpu=hawaii -mtriple=amdgcn-unknown-amdhsa -mattr=+vgpr-spilling -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=CIHSA %s
+; XUN: llc -march=amdgcn -mcpu=fiji -mtriple=amdgcn-unknown-amdhsa -mattr=+vgpr-spilling -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=VIHSA %s
+
+; This ends up using all 256 registers and requires register
+; scavenging which will fail to find an unsued register.
+
+; Check the ScratchSize to avoid regressions from spilling
+; intermediate register class copies.
+
+; FIXME: The same register is initialized to 0 for every spill.
+
+declare i32 @llvm.r600.read.tgid.x() #1
+declare i32 @llvm.r600.read.tgid.y() #1
+declare i32 @llvm.r600.read.tgid.z() #1
+
+; GCN-LABEL: {{^}}spill_vgpr_compute:
+
+; GCN: s_mov_b32 s12, SCRATCH_RSRC_DWORD0
+; GCN-NEXT: s_mov_b32 s13, SCRATCH_RSRC_DWORD1
+; GCN-NEXT: s_mov_b32 s14, -1
+; SI-NEXT: s_mov_b32 s15, 0x80f000
+; VI-NEXT: s_mov_b32 s15, 0x800000
+
+; GCN: buffer_store_dword {{v[0-9]+}}, s[12:15], s8 offset:{{[0-9]+}} ; 4-byte Folded Spill
+
+; GCN: buffer_store_dword {{v[0-9]+}}, {{v[0-9]+}}, s[12:15], s8 offen offset:{{[0-9]+}}
+; GCN: buffer_load_dword {{v[0-9]+}}, {{v[0-9]+}}, s[12:15], s8 offen offset:{{[0-9]+}}
+
+; GCN: NumVgprs: 256
+; GCN: ScratchSize: 1024
+
+; s[0:3] input user SGPRs. s4,s5,s6 = workgroup IDs. s8 scratch offset.
+define void @spill_vgpr_compute(<4 x float> %arg6, float addrspace(1)* %arg, i32 %arg1, i32 %arg2, float %arg3, float %arg4, float %arg5) #0 {
+bb:
+  %tmp = add i32 %arg1, %arg2
+  %tmp7 = extractelement <4 x float> %arg6, i32 0
+  %tmp8 = extractelement <4 x float> %arg6, i32 1
+  %tmp9 = extractelement <4 x float> %arg6, i32 2
+  %tmp10 = extractelement <4 x float> %arg6, i32 3
+  %tmp11 = bitcast float %arg5 to i32
+  br label %bb12
+
+bb12:                                             ; preds = %bb145, %bb
+  %tmp13 = phi float [ 0.000000e+00, %bb ], [ %tmp338, %bb145 ]
+  %tmp14 = phi float [ 0.000000e+00, %bb ], [ %tmp337, %bb145 ]
+  %tmp15 = phi float [ 0.000000e+00, %bb ], [ %tmp336, %bb145 ]
+  %tmp16 = phi float [ 0.000000e+00, %bb ], [ %tmp339, %bb145 ]
+  %tmp17 = phi float [ 0.000000e+00, %bb ], [ %tmp335, %bb145 ]
+  %tmp18 = phi float [ 0.000000e+00, %bb ], [ %tmp334, %bb145 ]
+  %tmp19 = phi float [ 0.000000e+00, %bb ], [ %tmp333, %bb145 ]
+  %tmp20 = phi float [ 0.000000e+00, %bb ], [ %tmp340, %bb145 ]
+  %tmp21 = phi float [ 0.000000e+00, %bb ], [ %tmp332, %bb145 ]
+  %tmp22 = phi float [ 0.000000e+00, %bb ], [ %tmp331, %bb145 ]
+  %tmp23 = phi float [ 0.000000e+00, %bb ], [ %tmp330, %bb145 ]
+  %tmp24 = phi float [ 0.000000e+00, %bb ], [ %tmp341, %bb145 ]
+  %tmp25 = phi float [ 0.000000e+00, %bb ], [ %tmp329, %bb145 ]
+  %tmp26 = phi float [ 0.000000e+00, %bb ], [ %tmp328, %bb145 ]
+  %tmp27 = phi float [ 0.000000e+00, %bb ], [ %tmp327, %bb145 ]
+  %tmp28 = phi float [ 0.000000e+00, %bb ], [ %tmp342, %bb145 ]
+  %tmp29 = phi float [ 0.000000e+00, %bb ], [ %tmp326, %bb145 ]
+  %tmp30 = phi float [ 0.000000e+00, %bb ], [ %tmp325, %bb145 ]
+  %tmp31 = phi float [ 0.000000e+00, %bb ], [ %tmp324, %bb145 ]
+  %tmp32 = phi float [ 0.000000e+00, %bb ], [ %tmp343, %bb145 ]
+  %tmp33 = phi float [ 0.000000e+00, %bb ], [ %tmp323, %bb145 ]
+  %tmp34 = phi float [ 0.000000e+00, %bb ], [ %tmp322, %bb145 ]
+  %tmp35 = phi float [ 0.000000e+00, %bb ], [ %tmp321, %bb145 ]
+  %tmp36 = phi float [ 0.000000e+00, %bb ], [ %tmp344, %bb145 ]
+  %tmp37 = phi float [ 0.000000e+00, %bb ], [ %tmp320, %bb145 ]
+  %tmp38 = phi float [ 0.000000e+00, %bb ], [ %tmp319, %bb145 ]
+  %tmp39 = phi float [ 0.000000e+00, %bb ], [ %tmp318, %bb145 ]
+  %tmp40 = phi float [ 0.000000e+00, %bb ], [ %tmp345, %bb145 ]
+  %tmp41 = phi float [ 0.000000e+00, %bb ], [ %tmp317, %bb145 ]
+  %tmp42 = phi float [ 0.000000e+00, %bb ], [ %tmp316, %bb145 ]
+  %tmp43 = phi float [ 0.000000e+00, %bb ], [ %tmp315, %bb145 ]
+  %tmp44 = phi float [ 0.000000e+00, %bb ], [ %tmp346, %bb145 ]
+  %tmp45 = phi float [ 0.000000e+00, %bb ], [ %tmp314, %bb145 ]
+  %tmp46 = phi float [ 0.000000e+00, %bb ], [ %tmp313, %bb145 ]
+  %tmp47 = phi float [ 0.000000e+00, %bb ], [ %tmp312, %bb145 ]
+  %tmp48 = phi float [ 0.000000e+00, %bb ], [ %tmp347, %bb145 ]
+  %tmp49 = phi float [ 0.000000e+00, %bb ], [ %tmp311, %bb145 ]
+  %tmp50 = phi float [ 0.000000e+00, %bb ], [ %tmp310, %bb145 ]
+  %tmp51 = phi float [ 0.000000e+00, %bb ], [ %tmp309, %bb145 ]
+  %tmp52 = phi float [ 0.000000e+00, %bb ], [ %tmp348, %bb145 ]
+  %tmp53 = phi float [ 0.000000e+00, %bb ], [ %tmp308, %bb145 ]
+  %tmp54 = phi float [ 0.000000e+00, %bb ], [ %tmp307, %bb145 ]
+  %tmp55 = phi float [ 0.000000e+00, %bb ], [ %tmp306, %bb145 ]
+  %tmp56 = phi float [ 0.000000e+00, %bb ], [ %tmp349, %bb145 ]
+  %tmp57 = phi float [ 0.000000e+00, %bb ], [ %tmp305, %bb145 ]
+  %tmp58 = phi float [ 0.000000e+00, %bb ], [ %tmp304, %bb145 ]
+  %tmp59 = phi float [ 0.000000e+00, %bb ], [ %tmp303, %bb145 ]
+  %tmp60 = phi float [ 0.000000e+00, %bb ], [ %tmp350, %bb145 ]
+  %tmp61 = phi float [ 0.000000e+00, %bb ], [ %tmp302, %bb145 ]
+  %tmp62 = phi float [ 0.000000e+00, %bb ], [ %tmp301, %bb145 ]
+  %tmp63 = phi float [ 0.000000e+00, %bb ], [ %tmp300, %bb145 ]
+  %tmp64 = phi float [ 0.000000e+00, %bb ], [ %tmp351, %bb145 ]
+  %tmp65 = phi float [ 0.000000e+00, %bb ], [ %tmp299, %bb145 ]
+  %tmp66 = phi float [ 0.000000e+00, %bb ], [ %tmp298, %bb145 ]
+  %tmp67 = phi float [ 0.000000e+00, %bb ], [ %tmp297, %bb145 ]
+  %tmp68 = phi float [ 0.000000e+00, %bb ], [ %tmp352, %bb145 ]
+  %tmp69 = phi float [ 0.000000e+00, %bb ], [ %tmp296, %bb145 ]
+  %tmp70 = phi float [ 0.000000e+00, %bb ], [ %tmp295, %bb145 ]
+  %tmp71 = phi float [ 0.000000e+00, %bb ], [ %tmp294, %bb145 ]
+  %tmp72 = phi float [ 0.000000e+00, %bb ], [ %tmp353, %bb145 ]
+  %tmp73 = phi float [ 0.000000e+00, %bb ], [ %tmp293, %bb145 ]
+  %tmp74 = phi float [ 0.000000e+00, %bb ], [ %tmp292, %bb145 ]
+  %tmp75 = phi float [ 0.000000e+00, %bb ], [ %tmp291, %bb145 ]
+  %tmp76 = phi float [ 0.000000e+00, %bb ], [ %tmp354, %bb145 ]
+  %tmp77 = phi float [ 0.000000e+00, %bb ], [ %tmp290, %bb145 ]
+  %tmp78 = phi float [ 0.000000e+00, %bb ], [ %tmp289, %bb145 ]
+  %tmp79 = phi float [ 0.000000e+00, %bb ], [ %tmp288, %bb145 ]
+  %tmp80 = phi float [ 0.000000e+00, %bb ], [ %tmp355, %bb145 ]
+  %tmp81 = phi float [ 0.000000e+00, %bb ], [ %tmp287, %bb145 ]
+  %tmp82 = phi float [ 0.000000e+00, %bb ], [ %tmp286, %bb145 ]
+  %tmp83 = phi float [ 0.000000e+00, %bb ], [ %tmp285, %bb145 ]
+  %tmp84 = phi float [ 0.000000e+00, %bb ], [ %tmp356, %bb145 ]
+  %tmp85 = phi float [ 0.000000e+00, %bb ], [ %tmp284, %bb145 ]
+  %tmp86 = phi float [ 0.000000e+00, %bb ], [ %tmp283, %bb145 ]
+  %tmp87 = phi float [ 0.000000e+00, %bb ], [ %tmp282, %bb145 ]
+  %tmp88 = phi float [ 0.000000e+00, %bb ], [ %tmp357, %bb145 ]
+  %tmp89 = phi float [ 0.000000e+00, %bb ], [ %tmp281, %bb145 ]
+  %tmp90 = phi float [ 0.000000e+00, %bb ], [ %tmp280, %bb145 ]
+  %tmp91 = phi float [ 0.000000e+00, %bb ], [ %tmp279, %bb145 ]
+  %tmp92 = phi float [ 0.000000e+00, %bb ], [ %tmp358, %bb145 ]
+  %tmp93 = phi float [ 0.000000e+00, %bb ], [ %tmp359, %bb145 ]
+  %tmp94 = phi float [ 0.000000e+00, %bb ], [ %tmp360, %bb145 ]
+  %tmp95 = phi float [ 0.000000e+00, %bb ], [ %tmp409, %bb145 ]
+  %tmp96 = phi float [ 0.000000e+00, %bb ], [ %tmp361, %bb145 ]
+  %tmp97 = phi float [ 0.000000e+00, %bb ], [ %tmp362, %bb145 ]
+  %tmp98 = phi float [ 0.000000e+00, %bb ], [ %tmp363, %bb145 ]
+  %tmp99 = phi float [ 0.000000e+00, %bb ], [ %tmp364, %bb145 ]
+  %tmp100 = phi float [ 0.000000e+00, %bb ], [ %tmp365, %bb145 ]
+  %tmp101 = phi float [ 0.000000e+00, %bb ], [ %tmp366, %bb145 ]
+  %tmp102 = phi float [ 0.000000e+00, %bb ], [ %tmp367, %bb145 ]
+  %tmp103 = phi float [ 0.000000e+00, %bb ], [ %tmp368, %bb145 ]
+  %tmp104 = phi float [ 0.000000e+00, %bb ], [ %tmp369, %bb145 ]
+  %tmp105 = phi float [ 0.000000e+00, %bb ], [ %tmp370, %bb145 ]
+  %tmp106 = phi float [ 0.000000e+00, %bb ], [ %tmp371, %bb145 ]
+  %tmp107 = phi float [ 0.000000e+00, %bb ], [ %tmp372, %bb145 ]
+  %tmp108 = phi float [ 0.000000e+00, %bb ], [ %tmp373, %bb145 ]
+  %tmp109 = phi float [ 0.000000e+00, %bb ], [ %tmp374, %bb145 ]
+  %tmp110 = phi float [ 0.000000e+00, %bb ], [ %tmp375, %bb145 ]
+  %tmp111 = phi float [ 0.000000e+00, %bb ], [ %tmp376, %bb145 ]
+  %tmp112 = phi float [ 0.000000e+00, %bb ], [ %tmp377, %bb145 ]
+  %tmp113 = phi float [ 0.000000e+00, %bb ], [ %tmp378, %bb145 ]
+  %tmp114 = phi float [ 0.000000e+00, %bb ], [ %tmp379, %bb145 ]
+  %tmp115 = phi float [ 0.000000e+00, %bb ], [ %tmp380, %bb145 ]
+  %tmp116 = phi float [ 0.000000e+00, %bb ], [ %tmp381, %bb145 ]
+  %tmp117 = phi float [ 0.000000e+00, %bb ], [ %tmp382, %bb145 ]
+  %tmp118 = phi float [ 0.000000e+00, %bb ], [ %tmp383, %bb145 ]
+  %tmp119 = phi float [ 0.000000e+00, %bb ], [ %tmp384, %bb145 ]
+  %tmp120 = phi float [ 0.000000e+00, %bb ], [ %tmp385, %bb145 ]
+  %tmp121 = phi float [ 0.000000e+00, %bb ], [ %tmp386, %bb145 ]
+  %tmp122 = phi float [ 0.000000e+00, %bb ], [ %tmp387, %bb145 ]
+  %tmp123 = phi float [ 0.000000e+00, %bb ], [ %tmp388, %bb145 ]
+  %tmp124 = phi float [ 0.000000e+00, %bb ], [ %tmp389, %bb145 ]
+  %tmp125 = phi float [ 0.000000e+00, %bb ], [ %tmp390, %bb145 ]
+  %tmp126 = phi float [ 0.000000e+00, %bb ], [ %tmp391, %bb145 ]
+  %tmp127 = phi float [ 0.000000e+00, %bb ], [ %tmp392, %bb145 ]
+  %tmp128 = phi float [ 0.000000e+00, %bb ], [ %tmp393, %bb145 ]
+  %tmp129 = phi float [ 0.000000e+00, %bb ], [ %tmp394, %bb145 ]
+  %tmp130 = phi float [ 0.000000e+00, %bb ], [ %tmp395, %bb145 ]
+  %tmp131 = phi float [ 0.000000e+00, %bb ], [ %tmp396, %bb145 ]
+  %tmp132 = phi float [ 0.000000e+00, %bb ], [ %tmp397, %bb145 ]
+  %tmp133 = phi float [ 0.000000e+00, %bb ], [ %tmp398, %bb145 ]
+  %tmp134 = phi float [ 0.000000e+00, %bb ], [ %tmp399, %bb145 ]
+  %tmp135 = phi float [ 0.000000e+00, %bb ], [ %tmp400, %bb145 ]
+  %tmp136 = phi float [ 0.000000e+00, %bb ], [ %tmp401, %bb145 ]
+  %tmp137 = phi float [ 0.000000e+00, %bb ], [ %tmp402, %bb145 ]
+  %tmp138 = phi float [ 0.000000e+00, %bb ], [ %tmp403, %bb145 ]
+  %tmp139 = phi float [ 0.000000e+00, %bb ], [ %tmp404, %bb145 ]
+  %tmp140 = phi float [ 0.000000e+00, %bb ], [ %tmp405, %bb145 ]
+  %tmp141 = phi float [ 0.000000e+00, %bb ], [ %tmp406, %bb145 ]
+  %tmp142 = bitcast float %tmp95 to i32
+  %tmp143 = icmp sgt i32 %tmp142, 125
+  br i1 %tmp143, label %bb144, label %bb145
+
+bb144:                                            ; preds = %bb12
+  store volatile float %arg3, float addrspace(1)* %arg
+  store volatile float %tmp91, float addrspace(1)* %arg
+  store volatile float %tmp90, float addrspace(1)* %arg
+  store volatile float %tmp89, float addrspace(1)* %arg
+  store volatile float %tmp87, float addrspace(1)* %arg
+  store volatile float %tmp86, float addrspace(1)* %arg
+  store volatile float %tmp85, float addrspace(1)* %arg
+  store volatile float %tmp83, float addrspace(1)* %arg
+  store volatile float %tmp82, float addrspace(1)* %arg
+  store volatile float %tmp81, float addrspace(1)* %arg
+  store volatile float %tmp79, float addrspace(1)* %arg
+  store volatile float %tmp78, float addrspace(1)* %arg
+  store volatile float %tmp77, float addrspace(1)* %arg
+  store volatile float %tmp75, float addrspace(1)* %arg
+  store volatile float %tmp74, float addrspace(1)* %arg
+  store volatile float %tmp73, float addrspace(1)* %arg
+  store volatile float %tmp71, float addrspace(1)* %arg
+  store volatile float %tmp70, float addrspace(1)* %arg
+  store volatile float %tmp69, float addrspace(1)* %arg
+  store volatile float %tmp67, float addrspace(1)* %arg
+  store volatile float %tmp66, float addrspace(1)* %arg
+  store volatile float %tmp65, float addrspace(1)* %arg
+  store volatile float %tmp63, float addrspace(1)* %arg
+  store volatile float %tmp62, float addrspace(1)* %arg
+  store volatile float %tmp61, float addrspace(1)* %arg
+  store volatile float %tmp59, float addrspace(1)* %arg
+  store volatile float %tmp58, float addrspace(1)* %arg
+  store volatile float %tmp57, float addrspace(1)* %arg
+  store volatile float %tmp55, float addrspace(1)* %arg
+  store volatile float %tmp54, float addrspace(1)* %arg
+  store volatile float %tmp53, float addrspace(1)* %arg
+  store volatile float %tmp51, float addrspace(1)* %arg
+  store volatile float %tmp50, float addrspace(1)* %arg
+  store volatile float %tmp49, float addrspace(1)* %arg
+  store volatile float %tmp47, float addrspace(1)* %arg
+  store volatile float %tmp46, float addrspace(1)* %arg
+  store volatile float %tmp45, float addrspace(1)* %arg
+  store volatile float %tmp43, float addrspace(1)* %arg
+  store volatile float %tmp42, float addrspace(1)* %arg
+  store volatile float %tmp41, float addrspace(1)* %arg
+  store volatile float %tmp39, float addrspace(1)* %arg
+  store volatile float %tmp38, float addrspace(1)* %arg
+  store volatile float %tmp37, float addrspace(1)* %arg
+  store volatile float %tmp35, float addrspace(1)* %arg
+  store volatile float %tmp34, float addrspace(1)* %arg
+  store volatile float %tmp33, float addrspace(1)* %arg
+  store volatile float %tmp31, float addrspace(1)* %arg
+  store volatile float %tmp30, float addrspace(1)* %arg
+  store volatile float %tmp29, float addrspace(1)* %arg
+  store volatile float %tmp27, float addrspace(1)* %arg
+  store volatile float %tmp26, float addrspace(1)* %arg
+  store volatile float %tmp25, float addrspace(1)* %arg
+  store volatile float %tmp23, float addrspace(1)* %arg
+  store volatile float %tmp22, float addrspace(1)* %arg
+  store volatile float %tmp21, float addrspace(1)* %arg
+  store volatile float %tmp19, float addrspace(1)* %arg
+  store volatile float %tmp18, float addrspace(1)* %arg
+  store volatile float %tmp17, float addrspace(1)* %arg
+  store volatile float %tmp15, float addrspace(1)* %arg
+  store volatile float %tmp14, float addrspace(1)* %arg
+  store volatile float %tmp13, float addrspace(1)* %arg
+  store volatile float %tmp16, float addrspace(1)* %arg
+  store volatile float %tmp20, float addrspace(1)* %arg
+  store volatile float %tmp24, float addrspace(1)* %arg
+  store volatile float %tmp28, float addrspace(1)* %arg
+  store volatile float %tmp32, float addrspace(1)* %arg
+  store volatile float %tmp36, float addrspace(1)* %arg
+  store volatile float %tmp40, float addrspace(1)* %arg
+  store volatile float %tmp44, float addrspace(1)* %arg
+  store volatile float %tmp48, float addrspace(1)* %arg
+  store volatile float %tmp52, float addrspace(1)* %arg
+  store volatile float %tmp56, float addrspace(1)* %arg
+  store volatile float %tmp60, float addrspace(1)* %arg
+  store volatile float %tmp64, float addrspace(1)* %arg
+  store volatile float %tmp68, float addrspace(1)* %arg
+  store volatile float %tmp72, float addrspace(1)* %arg
+  store volatile float %tmp76, float addrspace(1)* %arg
+  store volatile float %tmp80, float addrspace(1)* %arg
+  store volatile float %tmp84, float addrspace(1)* %arg
+  store volatile float %tmp88, float addrspace(1)* %arg
+  store volatile float %tmp92, float addrspace(1)* %arg
+  store volatile float %tmp93, float addrspace(1)* %arg
+  store volatile float %tmp94, float addrspace(1)* %arg
+  store volatile float %tmp96, float addrspace(1)* %arg
+  store volatile float %tmp97, float addrspace(1)* %arg
+  store volatile float %tmp98, float addrspace(1)* %arg
+  store volatile float %tmp99, float addrspace(1)* %arg
+  store volatile float %tmp100, float addrspace(1)* %arg
+  store volatile float %tmp101, float addrspace(1)* %arg
+  store volatile float %tmp102, float addrspace(1)* %arg
+  store volatile float %tmp103, float addrspace(1)* %arg
+  store volatile float %tmp104, float addrspace(1)* %arg
+  store volatile float %tmp105, float addrspace(1)* %arg
+  store volatile float %tmp106, float addrspace(1)* %arg
+  store volatile float %tmp107, float addrspace(1)* %arg
+  store volatile float %tmp108, float addrspace(1)* %arg
+  store volatile float %tmp109, float addrspace(1)* %arg
+  store volatile float %tmp110, float addrspace(1)* %arg
+  store volatile float %tmp111, float addrspace(1)* %arg
+  store volatile float %tmp112, float addrspace(1)* %arg
+  store volatile float %tmp113, float addrspace(1)* %arg
+  store volatile float %tmp114, float addrspace(1)* %arg
+  store volatile float %tmp115, float addrspace(1)* %arg
+  store volatile float %tmp116, float addrspace(1)* %arg
+  store volatile float %tmp117, float addrspace(1)* %arg
+  store volatile float %tmp118, float addrspace(1)* %arg
+  store volatile float %tmp119, float addrspace(1)* %arg
+  store volatile float %tmp120, float addrspace(1)* %arg
+  store volatile float %tmp121, float addrspace(1)* %arg
+  store volatile float %tmp122, float addrspace(1)* %arg
+  store volatile float %tmp123, float addrspace(1)* %arg
+  store volatile float %tmp124, float addrspace(1)* %arg
+  store volatile float %tmp125, float addrspace(1)* %arg
+  store volatile float %tmp126, float addrspace(1)* %arg
+  store volatile float %tmp127, float addrspace(1)* %arg
+  store volatile float %tmp128, float addrspace(1)* %arg
+  store volatile float %tmp129, float addrspace(1)* %arg
+  store volatile float %tmp130, float addrspace(1)* %arg
+  store volatile float %tmp131, float addrspace(1)* %arg
+  store volatile float %tmp132, float addrspace(1)* %arg
+  store volatile float %tmp133, float addrspace(1)* %arg
+  store volatile float %tmp134, float addrspace(1)* %arg
+  store volatile float %tmp135, float addrspace(1)* %arg
+  store volatile float %tmp136, float addrspace(1)* %arg
+  store volatile float %tmp137, float addrspace(1)* %arg
+  store volatile float %tmp138, float addrspace(1)* %arg
+  store volatile float %tmp139, float addrspace(1)* %arg
+  store volatile float %arg4, float addrspace(1)* %arg
+  store volatile float %tmp7, float addrspace(1)* %arg
+  store volatile float %tmp8, float addrspace(1)* %arg
+  store volatile float %tmp9, float addrspace(1)* %arg
+  store volatile float %tmp10, float addrspace(1)* %arg
+  ret void
+
+bb145:                                            ; preds = %bb12
+  %tmp146 = bitcast float %tmp95 to i32
+  %tmp147 = bitcast float %tmp95 to i32
+  %tmp148 = add i32 %tmp11, %tmp147
+  %tmp149 = bitcast i32 %tmp148 to float
+  %tmp150 = insertelement <128 x float> undef, float %tmp91, i32 0
+  %tmp151 = insertelement <128 x float> %tmp150, float %tmp90, i32 1
+  %tmp152 = insertelement <128 x float> %tmp151, float %tmp89, i32 2
+  %tmp153 = insertelement <128 x float> %tmp152, float %tmp87, i32 3
+  %tmp154 = insertelement <128 x float> %tmp153, float %tmp86, i32 4
+  %tmp155 = insertelement <128 x float> %tmp154, float %tmp85, i32 5
+  %tmp156 = insertelement <128 x float> %tmp155, float %tmp83, i32 6
+  %tmp157 = insertelement <128 x float> %tmp156, float %tmp82, i32 7
+  %tmp158 = insertelement <128 x float> %tmp157, float %tmp81, i32 8
+  %tmp159 = insertelement <128 x float> %tmp158, float %tmp79, i32 9
+  %tmp160 = insertelement <128 x float> %tmp159, float %tmp78, i32 10
+  %tmp161 = insertelement <128 x float> %tmp160, float %tmp77, i32 11
+  %tmp162 = insertelement <128 x float> %tmp161, float %tmp75, i32 12
+  %tmp163 = insertelement <128 x float> %tmp162, float %tmp74, i32 13
+  %tmp164 = insertelement <128 x float> %tmp163, float %tmp73, i32 14
+  %tmp165 = insertelement <128 x float> %tmp164, float %tmp71, i32 15
+  %tmp166 = insertelement <128 x float> %tmp165, float %tmp70, i32 16
+  %tmp167 = insertelement <128 x float> %tmp166, float %tmp69, i32 17
+  %tmp168 = insertelement <128 x float> %tmp167, float %tmp67, i32 18
+  %tmp169 = insertelement <128 x float> %tmp168, float %tmp66, i32 19
+  %tmp170 = insertelement <128 x float> %tmp169, float %tmp65, i32 20
+  %tmp171 = insertelement <128 x float> %tmp170, float %tmp63, i32 21
+  %tmp172 = insertelement <128 x float> %tmp171, float %tmp62, i32 22
+  %tmp173 = insertelement <128 x float> %tmp172, float %tmp61, i32 23
+  %tmp174 = insertelement <128 x float> %tmp173, float %tmp59, i32 24
+  %tmp175 = insertelement <128 x float> %tmp174, float %tmp58, i32 25
+  %tmp176 = insertelement <128 x float> %tmp175, float %tmp57, i32 26
+  %tmp177 = insertelement <128 x float> %tmp176, float %tmp55, i32 27
+  %tmp178 = insertelement <128 x float> %tmp177, float %tmp54, i32 28
+  %tmp179 = insertelement <128 x float> %tmp178, float %tmp53, i32 29
+  %tmp180 = insertelement <128 x float> %tmp179, float %tmp51, i32 30
+  %tmp181 = insertelement <128 x float> %tmp180, float %tmp50, i32 31
+  %tmp182 = insertelement <128 x float> %tmp181, float %tmp49, i32 32
+  %tmp183 = insertelement <128 x float> %tmp182, float %tmp47, i32 33
+  %tmp184 = insertelement <128 x float> %tmp183, float %tmp46, i32 34
+  %tmp185 = insertelement <128 x float> %tmp184, float %tmp45, i32 35
+  %tmp186 = insertelement <128 x float> %tmp185, float %tmp43, i32 36
+  %tmp187 = insertelement <128 x float> %tmp186, float %tmp42, i32 37
+  %tmp188 = insertelement <128 x float> %tmp187, float %tmp41, i32 38
+  %tmp189 = insertelement <128 x float> %tmp188, float %tmp39, i32 39
+  %tmp190 = insertelement <128 x float> %tmp189, float %tmp38, i32 40
+  %tmp191 = insertelement <128 x float> %tmp190, float %tmp37, i32 41
+  %tmp192 = insertelement <128 x float> %tmp191, float %tmp35, i32 42
+  %tmp193 = insertelement <128 x float> %tmp192, float %tmp34, i32 43
+  %tmp194 = insertelement <128 x float> %tmp193, float %tmp33, i32 44
+  %tmp195 = insertelement <128 x float> %tmp194, float %tmp31, i32 45
+  %tmp196 = insertelement <128 x float> %tmp195, float %tmp30, i32 46
+  %tmp197 = insertelement <128 x float> %tmp196, float %tmp29, i32 47
+  %tmp198 = insertelement <128 x float> %tmp197, float %tmp27, i32 48
+  %tmp199 = insertelement <128 x float> %tmp198, float %tmp26, i32 49
+  %tmp200 = insertelement <128 x float> %tmp199, float %tmp25, i32 50
+  %tmp201 = insertelement <128 x float> %tmp200, float %tmp23, i32 51
+  %tmp202 = insertelement <128 x float> %tmp201, float %tmp22, i32 52
+  %tmp203 = insertelement <128 x float> %tmp202, float %tmp21, i32 53
+  %tmp204 = insertelement <128 x float> %tmp203, float %tmp19, i32 54
+  %tmp205 = insertelement <128 x float> %tmp204, float %tmp18, i32 55
+  %tmp206 = insertelement <128 x float> %tmp205, float %tmp17, i32 56
+  %tmp207 = insertelement <128 x float> %tmp206, float %tmp15, i32 57
+  %tmp208 = insertelement <128 x float> %tmp207, float %tmp14, i32 58
+  %tmp209 = insertelement <128 x float> %tmp208, float %tmp13, i32 59
+  %tmp210 = insertelement <128 x float> %tmp209, float %tmp16, i32 60
+  %tmp211 = insertelement <128 x float> %tmp210, float %tmp20, i32 61
+  %tmp212 = insertelement <128 x float> %tmp211, float %tmp24, i32 62
+  %tmp213 = insertelement <128 x float> %tmp212, float %tmp28, i32 63
+  %tmp214 = insertelement <128 x float> %tmp213, float %tmp32, i32 64
+  %tmp215 = insertelement <128 x float> %tmp214, float %tmp36, i32 65
+  %tmp216 = insertelement <128 x float> %tmp215, float %tmp40, i32 66
+  %tmp217 = insertelement <128 x float> %tmp216, float %tmp44, i32 67
+  %tmp218 = insertelement <128 x float> %tmp217, float %tmp48, i32 68
+  %tmp219 = insertelement <128 x float> %tmp218, float %tmp52, i32 69
+  %tmp220 = insertelement <128 x float> %tmp219, float %tmp56, i32 70
+  %tmp221 = insertelement <128 x float> %tmp220, float %tmp60, i32 71
+  %tmp222 = insertelement <128 x float> %tmp221, float %tmp64, i32 72
+  %tmp223 = insertelement <128 x float> %tmp222, float %tmp68, i32 73
+  %tmp224 = insertelement <128 x float> %tmp223, float %tmp72, i32 74
+  %tmp225 = insertelement <128 x float> %tmp224, float %tmp76, i32 75
+  %tmp226 = insertelement <128 x float> %tmp225, float %tmp80, i32 76
+  %tmp227 = insertelement <128 x float> %tmp226, float %tmp84, i32 77
+  %tmp228 = insertelement <128 x float> %tmp227, float %tmp88, i32 78
+  %tmp229 = insertelement <128 x float> %tmp228, float %tmp92, i32 79
+  %tmp230 = insertelement <128 x float> %tmp229, float %tmp93, i32 80
+  %tmp231 = insertelement <128 x float> %tmp230, float %tmp94, i32 81
+  %tmp232 = insertelement <128 x float> %tmp231, float %tmp96, i32 82
+  %tmp233 = insertelement <128 x float> %tmp232, float %tmp97, i32 83
+  %tmp234 = insertelement <128 x float> %tmp233, float %tmp98, i32 84
+  %tmp235 = insertelement <128 x float> %tmp234, float %tmp99, i32 85
+  %tmp236 = insertelement <128 x float> %tmp235, float %tmp100, i32 86
+  %tmp237 = insertelement <128 x float> %tmp236, float %tmp101, i32 87
+  %tmp238 = insertelement <128 x float> %tmp237, float %tmp102, i32 88
+  %tmp239 = insertelement <128 x float> %tmp238, float %tmp103, i32 89
+  %tmp240 = insertelement <128 x float> %tmp239, float %tmp104, i32 90
+  %tmp241 = insertelement <128 x float> %tmp240, float %tmp105, i32 91
+  %tmp242 = insertelement <128 x float> %tmp241, float %tmp106, i32 92
+  %tmp243 = insertelement <128 x float> %tmp242, float %tmp107, i32 93
+  %tmp244 = insertelement <128 x float> %tmp243, float %tmp108, i32 94
+  %tmp245 = insertelement <128 x float> %tmp244, float %tmp109, i32 95
+  %tmp246 = insertelement <128 x float> %tmp245, float %tmp110, i32 96
+  %tmp247 = insertelement <128 x float> %tmp246, float %tmp111, i32 97
+  %tmp248 = insertelement <128 x float> %tmp247, float %tmp112, i32 98
+  %tmp249 = insertelement <128 x float> %tmp248, float %tmp113, i32 99
+  %tmp250 = insertelement <128 x float> %tmp249, float %tmp114, i32 100
+  %tmp251 = insertelement <128 x float> %tmp250, float %tmp115, i32 101
+  %tmp252 = insertelement <128 x float> %tmp251, float %tmp116, i32 102
+  %tmp253 = insertelement <128 x float> %tmp252, float %tmp117, i32 103
+  %tmp254 = insertelement <128 x float> %tmp253, float %tmp118, i32 104
+  %tmp255 = insertelement <128 x float> %tmp254, float %tmp119, i32 105
+  %tmp256 = insertelement <128 x float> %tmp255, float %tmp120, i32 106
+  %tmp257 = insertelement <128 x float> %tmp256, float %tmp121, i32 107
+  %tmp258 = insertelement <128 x float> %tmp257, float %tmp122, i32 108
+  %tmp259 = insertelement <128 x float> %tmp258, float %tmp123, i32 109
+  %tmp260 = insertelement <128 x float> %tmp259, float %tmp124, i32 110
+  %tmp261 = insertelement <128 x float> %tmp260, float %tmp125, i32 111
+  %tmp262 = insertelement <128 x float> %tmp261, float %tmp126, i32 112
+  %tmp263 = insertelement <128 x float> %tmp262, float %tmp127, i32 113
+  %tmp264 = insertelement <128 x float> %tmp263, float %tmp128, i32 114
+  %tmp265 = insertelement <128 x float> %tmp264, float %tmp129, i32 115
+  %tmp266 = insertelement <128 x float> %tmp265, float %tmp130, i32 116
+  %tmp267 = insertelement <128 x float> %tmp266, float %tmp131, i32 117
+  %tmp268 = insertelement <128 x float> %tmp267, float %tmp132, i32 118
+  %tmp269 = insertelement <128 x float> %tmp268, float %tmp133, i32 119
+  %tmp270 = insertelement <128 x float> %tmp269, float %tmp134, i32 120
+  %tmp271 = insertelement <128 x float> %tmp270, float %tmp135, i32 121
+  %tmp272 = insertelement <128 x float> %tmp271, float %tmp136, i32 122
+  %tmp273 = insertelement <128 x float> %tmp272, float %tmp137, i32 123
+  %tmp274 = insertelement <128 x float> %tmp273, float %tmp138, i32 124
+  %tmp275 = insertelement <128 x float> %tmp274, float %tmp139, i32 125
+  %tmp276 = insertelement <128 x float> %tmp275, float %tmp140, i32 126
+  %tmp277 = insertelement <128 x float> %tmp276, float %tmp141, i32 127
+  %tmp278 = insertelement <128 x float> %tmp277, float %tmp149, i32 %tmp146
+  %tmp279 = extractelement <128 x float> %tmp278, i32 0
+  %tmp280 = extractelement <128 x float> %tmp278, i32 1
+  %tmp281 = extractelement <128 x float> %tmp278, i32 2
+  %tmp282 = extractelement <128 x float> %tmp278, i32 3
+  %tmp283 = extractelement <128 x float> %tmp278, i32 4
+  %tmp284 = extractelement <128 x float> %tmp278, i32 5
+  %tmp285 = extractelement <128 x float> %tmp278, i32 6
+  %tmp286 = extractelement <128 x float> %tmp278, i32 7
+  %tmp287 = extractelement <128 x float> %tmp278, i32 8
+  %tmp288 = extractelement <128 x float> %tmp278, i32 9
+  %tmp289 = extractelement <128 x float> %tmp278, i32 10
+  %tmp290 = extractelement <128 x float> %tmp278, i32 11
+  %tmp291 = extractelement <128 x float> %tmp278, i32 12
+  %tmp292 = extractelement <128 x float> %tmp278, i32 13
+  %tmp293 = extractelement <128 x float> %tmp278, i32 14
+  %tmp294 = extractelement <128 x float> %tmp278, i32 15
+  %tmp295 = extractelement <128 x float> %tmp278, i32 16
+  %tmp296 = extractelement <128 x float> %tmp278, i32 17
+  %tmp297 = extractelement <128 x float> %tmp278, i32 18
+  %tmp298 = extractelement <128 x float> %tmp278, i32 19
+  %tmp299 = extractelement <128 x float> %tmp278, i32 20
+  %tmp300 = extractelement <128 x float> %tmp278, i32 21
+  %tmp301 = extractelement <128 x float> %tmp278, i32 22
+  %tmp302 = extractelement <128 x float> %tmp278, i32 23
+  %tmp303 = extractelement <128 x float> %tmp278, i32 24
+  %tmp304 = extractelement <128 x float> %tmp278, i32 25
+  %tmp305 = extractelement <128 x float> %tmp278, i32 26
+  %tmp306 = extractelement <128 x float> %tmp278, i32 27
+  %tmp307 = extractelement <128 x float> %tmp278, i32 28
+  %tmp308 = extractelement <128 x float> %tmp278, i32 29
+  %tmp309 = extractelement <128 x float> %tmp278, i32 30
+  %tmp310 = extractelement <128 x float> %tmp278, i32 31
+  %tmp311 = extractelement <128 x float> %tmp278, i32 32
+  %tmp312 = extractelement <128 x float> %tmp278, i32 33
+  %tmp313 = extractelement <128 x float> %tmp278, i32 34
+  %tmp314 = extractelement <128 x float> %tmp278, i32 35
+  %tmp315 = extractelement <128 x float> %tmp278, i32 36
+  %tmp316 = extractelement <128 x float> %tmp278, i32 37
+  %tmp317 = extractelement <128 x float> %tmp278, i32 38
+  %tmp318 = extractelement <128 x float> %tmp278, i32 39
+  %tmp319 = extractelement <128 x float> %tmp278, i32 40
+  %tmp320 = extractelement <128 x float> %tmp278, i32 41
+  %tmp321 = extractelement <128 x float> %tmp278, i32 42
+  %tmp322 = extractelement <128 x float> %tmp278, i32 43
+  %tmp323 = extractelement <128 x float> %tmp278, i32 44
+  %tmp324 = extractelement <128 x float> %tmp278, i32 45
+  %tmp325 = extractelement <128 x float> %tmp278, i32 46
+  %tmp326 = extractelement <128 x float> %tmp278, i32 47
+  %tmp327 = extractelement <128 x float> %tmp278, i32 48
+  %tmp328 = extractelement <128 x float> %tmp278, i32 49
+  %tmp329 = extractelement <128 x float> %tmp278, i32 50
+  %tmp330 = extractelement <128 x float> %tmp278, i32 51
+  %tmp331 = extractelement <128 x float> %tmp278, i32 52
+  %tmp332 = extractelement <128 x float> %tmp278, i32 53
+  %tmp333 = extractelement <128 x float> %tmp278, i32 54
+  %tmp334 = extractelement <128 x float> %tmp278, i32 55
+  %tmp335 = extractelement <128 x float> %tmp278, i32 56
+  %tmp336 = extractelement <128 x float> %tmp278, i32 57
+  %tmp337 = extractelement <128 x float> %tmp278, i32 58
+  %tmp338 = extractelement <128 x float> %tmp278, i32 59
+  %tmp339 = extractelement <128 x float> %tmp278, i32 60
+  %tmp340 = extractelement <128 x float> %tmp278, i32 61
+  %tmp341 = extractelement <128 x float> %tmp278, i32 62
+  %tmp342 = extractelement <128 x float> %tmp278, i32 63
+  %tmp343 = extractelement <128 x float> %tmp278, i32 64
+  %tmp344 = extractelement <128 x float> %tmp278, i32 65
+  %tmp345 = extractelement <128 x float> %tmp278, i32 66
+  %tmp346 = extractelement <128 x float> %tmp278, i32 67
+  %tmp347 = extractelement <128 x float> %tmp278, i32 68
+  %tmp348 = extractelement <128 x float> %tmp278, i32 69
+  %tmp349 = extractelement <128 x float> %tmp278, i32 70
+  %tmp350 = extractelement <128 x float> %tmp278, i32 71
+  %tmp351 = extractelement <128 x float> %tmp278, i32 72
+  %tmp352 = extractelement <128 x float> %tmp278, i32 73
+  %tmp353 = extractelement <128 x float> %tmp278, i32 74
+  %tmp354 = extractelement <128 x float> %tmp278, i32 75
+  %tmp355 = extractelement <128 x float> %tmp278, i32 76
+  %tmp356 = extractelement <128 x float> %tmp278, i32 77
+  %tmp357 = extractelement <128 x float> %tmp278, i32 78
+  %tmp358 = extractelement <128 x float> %tmp278, i32 79
+  %tmp359 = extractelement <128 x float> %tmp278, i32 80
+  %tmp360 = extractelement <128 x float> %tmp278, i32 81
+  %tmp361 = extractelement <128 x float> %tmp278, i32 82
+  %tmp362 = extractelement <128 x float> %tmp278, i32 83
+  %tmp363 = extractelement <128 x float> %tmp278, i32 84
+  %tmp364 = extractelement <128 x float> %tmp278, i32 85
+  %tmp365 = extractelement <128 x float> %tmp278, i32 86
+  %tmp366 = extractelement <128 x float> %tmp278, i32 87
+  %tmp367 = extractelement <128 x float> %tmp278, i32 88
+  %tmp368 = extractelement <128 x float> %tmp278, i32 89
+  %tmp369 = extractelement <128 x float> %tmp278, i32 90
+  %tmp370 = extractelement <128 x float> %tmp278, i32 91
+  %tmp371 = extractelement <128 x float> %tmp278, i32 92
+  %tmp372 = extractelement <128 x float> %tmp278, i32 93
+  %tmp373 = extractelement <128 x float> %tmp278, i32 94
+  %tmp374 = extractelement <128 x float> %tmp278, i32 95
+  %tmp375 = extractelement <128 x float> %tmp278, i32 96
+  %tmp376 = extractelement <128 x float> %tmp278, i32 97
+  %tmp377 = extractelement <128 x float> %tmp278, i32 98
+  %tmp378 = extractelement <128 x float> %tmp278, i32 99
+  %tmp379 = extractelement <128 x float> %tmp278, i32 100
+  %tmp380 = extractelement <128 x float> %tmp278, i32 101
+  %tmp381 = extractelement <128 x float> %tmp278, i32 102
+  %tmp382 = extractelement <128 x float> %tmp278, i32 103
+  %tmp383 = extractelement <128 x float> %tmp278, i32 104
+  %tmp384 = extractelement <128 x float> %tmp278, i32 105
+  %tmp385 = extractelement <128 x float> %tmp278, i32 106
+  %tmp386 = extractelement <128 x float> %tmp278, i32 107
+  %tmp387 = extractelement <128 x float> %tmp278, i32 108
+  %tmp388 = extractelement <128 x float> %tmp278, i32 109
+  %tmp389 = extractelement <128 x float> %tmp278, i32 110
+  %tmp390 = extractelement <128 x float> %tmp278, i32 111
+  %tmp391 = extractelement <128 x float> %tmp278, i32 112
+  %tmp392 = extractelement <128 x float> %tmp278, i32 113
+  %tmp393 = extractelement <128 x float> %tmp278, i32 114
+  %tmp394 = extractelement <128 x float> %tmp278, i32 115
+  %tmp395 = extractelement <128 x float> %tmp278, i32 116
+  %tmp396 = extractelement <128 x float> %tmp278, i32 117
+  %tmp397 = extractelement <128 x float> %tmp278, i32 118
+  %tmp398 = extractelement <128 x float> %tmp278, i32 119
+  %tmp399 = extractelement <128 x float> %tmp278, i32 120
+  %tmp400 = extractelement <128 x float> %tmp278, i32 121
+  %tmp401 = extractelement <128 x float> %tmp278, i32 122
+  %tmp402 = extractelement <128 x float> %tmp278, i32 123
+  %tmp403 = extractelement <128 x float> %tmp278, i32 124
+  %tmp404 = extractelement <128 x float> %tmp278, i32 125
+  %tmp405 = extractelement <128 x float> %tmp278, i32 126
+  %tmp406 = extractelement <128 x float> %tmp278, i32 127
+  %tmp407 = bitcast float %tmp95 to i32
+  %tmp408 = add i32 %tmp407, 1
+  %tmp409 = bitcast i32 %tmp408 to float
+  br label %bb12
+}
+
+attributes #0 = { nounwind }
+attributes #1 = { nounwind readnone }
diff --git a/test/CodeGen/AMDGPU/vgpr-spill-emergency-stack-slot.ll b/test/CodeGen/AMDGPU/vgpr-spill-emergency-stack-slot.ll
index 5ce65371b01..ef492cef576 100644
--- a/test/CodeGen/AMDGPU/vgpr-spill-emergency-stack-slot.ll
+++ b/test/CodeGen/AMDGPU/vgpr-spill-emergency-stack-slot.ll
@@ -1,7 +1,5 @@
-; RUN: llc -march=amdgcn -mcpu=tahiti -mattr=+vgpr-spilling < %s | FileCheck %s
-; RUN: llc -march=amdgcn -mcpu=fiji -mattr=+vgpr-spilling < %s | FileCheck %s
-
-; FIXME: Enable -verify-instructions
+; RUN: llc -march=amdgcn -mcpu=tahiti -mattr=+vgpr-spilling -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=SI %s
+; RUN: llc -march=amdgcn -mcpu=fiji -mattr=+vgpr-spilling -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=VI %s
 
 ; This ends up using all 255 registers and requires register
 ; scavenging which will fail to find an unsued register.
@@ -11,9 +9,19 @@
 
 ; FIXME: The same register is initialized to 0 for every spill.
 
-; CHECK-LABEL: {{^}}main:
-; CHECK: NumVgprs: 256
-; CHECK: ScratchSize: 1024
+; GCN-LABEL: {{^}}main:
+
+; GCN: s_mov_b32 s8, SCRATCH_RSRC_DWORD0
+; GCN-NEXT: s_mov_b32 s9, SCRATCH_RSRC_DWORD1
+; GCN-NEXT: s_mov_b32 s10, -1
+; SI-NEXT: s_mov_b32 s11, 0x80f000
+; VI-NEXT: s_mov_b32 s11, 0x800000
+
+; s12 is offset user SGPR
+; GCN: buffer_store_dword {{v[0-9]+}}, s[8:11], s12 offset:{{[0-9]+}} ; 4-byte Folded Spill
+
+; GCN: NumVgprs: 256
+; GCN: ScratchSize: 1024
 
 define void @main([9 x <16 x i8>] addrspace(2)* byval %arg, [17 x <16 x i8>] addrspace(2)* byval %arg1, [17 x <4 x i32>] addrspace(2)* byval %arg2, [34 x <8 x i32>] addrspace(2)* byval %arg3, [16 x <16 x i8>] addrspace(2)* byval %arg4, i32 inreg %arg5, i32 inreg %arg6, i32 %arg7, i32 %arg8, i32 %arg9, i32 %arg10) #0 {
 bb: