[AMDGPU] New Waitcnt Insertion Pass

[kanarayan]

Summary

This pass implements the algorithm deployed by our internal compiler for inserting waitcnt instructions. The pass performs cross basic-block analysis and tracks individual registers, and provides predicted performance improvements over the current implementation.

There are further improvements forthcoming, including relaxing overtly conservative assumptions about LDS access, integration of memory model pass, and more targeted tests for the corners.

[kzhuravl]

@rampitec wrote:
Please add test with s_barrier.

lib/Target/AMDGPU/SIInsertWaitcnts.cpp

75

@rampitec wrote:
Should not we get these numbers from TD files for target which we already have?

1018

@rampitec wrote:
For a barrier it will always insert strongest:

s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)

regardless of what was an argument of the barrier.

This also seems to completely ignore atomic fences inserted around the barrier from the library, which shall be a real source of wait argument. Note, that semantics of needWaitcntBeforeBarrier() is not that we always need to insert wait with barrier, but that we may need to insert it.

Also note that existing pass does not seem to do it for a barrier.

1614

@rampitec wrote:
Need to check for XNACK support.

[rampitec]

Please add test with s_barrier.

lib/Target/AMDGPU/SIInsertWaitcnts.cpp

75

Should not we get these numbers from TD files for target which we already have?

1018

For a barrier it will always insert strongest:

s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)

regardless of what was an argument of the barrier.

This also seems to completely ignore atomic fences inserted around the barrier from the library, which shall be a real source of wait argument. Note, that semantics of needWaitcntBeforeBarrier() is not that we always need to insert wait with barrier, but that we may need to insert it.

Also note that existing pass does not seem to do it for a barrier.

1614

Need to check for XNACK support.

[kzhuravl]

lib/Target/AMDGPU/SIInsertWaitcnts.cpp

1018

Atomics are currently handled in a separate pass, which we determined to be conservatively correct. We plan to integrate atomics into waitcnt insertion but later and in a separate patch.

The *needWaitcntBeforeBarrier* tells you whether you need a waitcnt before the barrier or not. For >=GFX9, waitcnt is automatically inserted before the barrier, so we do not need to generate it, and *needWaitcntBeforeBarrier* returns false for >=GFX9.

Also note that existing pass does not seem to do it for a barrier.

https://github.com/llvm-mirror/llvm/blob/master/lib/Target/AMDGPU/SIInsertWaits.cpp#L633

[kzhuravl]

In D31161#709066, @rampitec wrote:

Please add test with s_barrier.

https://github.com/llvm-mirror/llvm/blob/master/test/CodeGen/AMDGPU/llvm.amdgcn.s.barrier.ll

[rampitec]

In D31161#709081, @kzhuravl wrote:

In D31161#709066, @rampitec wrote:

Please add test with s_barrier.

https://github.com/llvm-mirror/llvm/blob/master/test/CodeGen/AMDGPU/llvm.amdgcn.s.barrier.ll

How about waitcnt operands in that test?

lib/Target/AMDGPU/SIInsertWaitcnts.cpp

1018

Barrier itself does not need fence. OpenCL barrier needs fence, but these are generated in the library.

[kanarayan]

This addresses two issues:

1. SQ_MAX_PGM_VGPRS and other constants are used to map the llvm register map to a data structure internal to this algorithm. For the register ampping used by the algorithm see the comments before the enum. They are maximum values across all targets. It is ideal to use dynamically sized arrays that fit the particular architecture target. As an interim step, I have changed these constants to enum values, asserted that no target has a larger register file in the main entry to this pass. (In response to comments from Tony and Konstantin) I have also updated the getRegInterval call. Please notice that the previous version was identical to the pass used by the old pass except I do the necessary adjustments for the mapping used by this algorithm. In a next step, I will remove the assert and the associated code.
2. Barriers no longer force a zero waitcnt. For GFX9 and above, barrier needs no additional waitcnt. For lesser targets, waitcnts are added only if needed. The following tests already test barrier: LLVM :: CodeGen/AMDGPU/addrspacecast.ll LLVM :: CodeGen/AMDGPU/array-ptr-calc-i32.ll LLVM :: CodeGen/AMDGPU/ds-negative-offset-addressing-mode-loop.ll LLVM :: CodeGen/AMDGPU/ds_read2.ll LLVM :: CodeGen/AMDGPU/indirect-private-64.ll LLVM :: CodeGen/AMDGPU/llvm.amdgcn.s.barrier.ll LLVM :: CodeGen/AMDGPU/local-memory.amdgcn.ll LLVM :: CodeGen/AMDGPU/merge-stores.ll LLVM :: CodeGen/AMDGPU/schedule-vs-if-nested-loop-failure.ll LLVM :: CodeGen/AMDGPU/si-triv-disjoint-mem-access.ll LLVM :: CodeGen/AMDGPU/store-barrier.ll LLVM :: CodeGen/AMDGPU/wait.ll

There is also a CodeGen/AMDGPU/llvm.amdgcn.s.barrier.ll

This patch does not yet address the XNACK changes in https://reviews.llvm.org/D30302

[rampitec]

Test with barrier surrounded with fences is needed. All relevant combinations of fences needs to be checked and pattern shall check that only needed counters are used, wait is produced, and that is only one wait. Please refer to OpenCL barrier() implementation for the code to check.

lib/Target/AMDGPU/SIInsertWaitcnts.cpp

1006

Not needed anymore.

[rampitec]

lib/Target/AMDGPU/SIInsertWaitcnts.cpp

819

I believe there is no SI_RETURN anymore, it is renamed.

[kanarayan]

1. Rename SI_RETURN to SI_RETURN_TO_EPILOG to reflect recent change.
2. Condition the nop insertion to break soft clauses on isXNACKEnabled(). Add a note to also condition this code on hasSoftClauses when that code is put back.
3. Fix tests added since last patch to reflect the new pass.
4. Remove experimental code.

[kanarayan]

Fix the last patch that inadvertently deleted "amdgpu_kernel" added recently.

1. Rename SI_RETURN to SI_RETURN_TO_EPILOG to reflect recent change.
2. Condition the nop insertion to break soft clauses on isXNACKEnabled(). Add a note to also condition this code on hasSoftClauses when that code is put back.
3. Fix tests added since last patch to reflect the new pass.
4. Remove experimental code.

[kzhuravl]

• Your pass does not take into account encoding changes for gfx9.
• New pass name is confusing.
• Make a note somewhere that the new pass is on by default, and you can switch it off by such and such option.

lib/Target/AMDGPU/SIInsertWaitcnts.cpp

1070–1072

Use helper functions from AMDGPUBaseInfo.h (they also have logic for gfx9 in it)

1099–1100

Use helper functions from AMDGPUBaseInfo.h (they also have logic for gfx9 in it)

[kanarayan]

Use the architecture neutral interfaces to decode wait counts.

In response to the other comments, please note:

1. The old pass will be removed shortly. I agree the pass names are close enough to cause some confusion in the interim, but the names are apt.
2. Please refer to the comment here below (from AMDGPUTargetMachine.cpp) that indicates the new pass is the default and with -enable-si-insert-waitcnts=0 one can revert to the old pass.

// Option to enable new waitcnt insertion pass.
static cl::opt<bool> EnableSIInsertWaitcntsPass(

"enable-si-insert-waitcnts",
cl::desc("Use new waitcnt insertion pass"),
cl::init(true));

[arsenm]

lib/Target/AMDGPU/SIInsertWaitcnts.cpp

3

Extra blank line

89–92

This should be replaced with an iterator range or simply removed

511–516

Can't you just get the ::data operand and check if it is a use or def? Same for the other places checking getAtomicNoRetOp

761–763

C++ style comments

780

nullptr

786

Linemapping->LineMapping

787

MI.isDebugValue()

930

I'm concerned by relying on the memoperands since it's possible they were dropped. The uses LDS memory check could at least be factored into a predicate function

969–972

Ditto

1141–1147

This should check the gds operand, not the mem operands. This also needs MIR tests since we don't emit GDS operations now

1194

getNamedOperand

1609–1610

Braces around the block and addImm on next line

1688

Use BuildMI rather than the low level instruction creation APIs

[kzhuravl]

In D31161#709509, @rampitec wrote:

Test with barrier surrounded with fences is needed. All relevant combinations of fences needs to be checked and pattern shall check that only needed counters are used, wait is produced, and that is only one wait. Please refer to OpenCL barrier() implementation for the code to check.

these tests must go downstream however (separate change). we do not have fence implementation upstreamed yet.

[rampitec]

In D31161#713546, @kzhuravl wrote:

In D31161#709509, @rampitec wrote:

Test with barrier surrounded with fences is needed. All relevant combinations of fences needs to be checked and pattern shall check that only needed counters are used, wait is produced, and that is only one wait. Please refer to OpenCL barrier() implementation for the code to check.

these tests must go downstream however (separate change). we do not have fence implementation upstreamed yet.

It can be in the internal repo, but we need these tests to be sure we are producing right fences for OpenCL barriers.

[kzhuravl]

In D31161#713550, @rampitec wrote:

In D31161#713546, @kzhuravl wrote:

In D31161#709509, @rampitec wrote:

Test with barrier surrounded with fences is needed. All relevant combinations of fences needs to be checked and pattern shall check that only needed counters are used, wait is produced, and that is only one wait. Please refer to OpenCL barrier() implementation for the code to check.

these tests must go downstream however (separate change). we do not have fence implementation upstreamed yet.

It can be in the internal repo, but we need these tests to be sure we are producing right fences for OpenCL barriers.

that is exactly what i meant by my previous comment (tests should go to amd-common branch).

[kanarayan]

In response to review comments from arsenm, the following updates/responses:

The following updates to the code:

1. Removed Extra blank line.
2. Removed macro defining iterator on enum and replace it with inline code.
3. Replace with C++ style comments.
4. Use nullptr.
5. Replace Linemapping with LineMapping.
6. Use MI.isDebugValue()
7. Use getNamedOperand
8. Braces around the block (but clang-format likes addImm on same line).

The following will be addressed in a subsequent put-back.

1. LDS access check & GDS tests: The code can be simplified further as suggested. The generated code is overly conservative but correct.
2. Uses of getAtomicNoRetOp: The code is looking for atomic operations that do not return. The atomic operations that do return are covered by the code that look for stores. Again, this is conservatively correct as is written, but needs to be fixed.

CreateMachineInstr - BuildMI was unusable in this context. Besides, X86 and some other code written for our target also use CreateMachineInstr.

[kanarayan]

The update includes the following after rebasing:

Re-organize the getRegInterval as follows:

1. Support the registers covered under isSGPRReg that are allocatable.
2. Assert that the register numbers are within bounds.

Check the GDS bit and avoid the loop on the memory operands. For DS operations, additionally check they are either load or a store.

Fix the LGKM bit for ds_bpermute, ds_permute operations that do not access DS.

Add comments on TODO items to the code. The following are some of the other items:

1. We currently list all !VGPR registers under isSGPRReg that causes confusion.
2. There are waintcnt flags/bits which do not appear to be consistently set (example: ds_swizzle). I need to verify the flag bits are set property and then could use the bits in the new pass.

[kanarayan]

clang-format

[t-tye]

lib/Target/AMDGPU/SIInsertWaitcnts.cpp

424

Could M0, EXEC, etc. also potentially be a source or dest for SMEM and VMEM load/store?

Perhaps add an assert to ensure they never happen.

[arsenm]

lib/Target/AMDGPU/SIInsertWaitcnts.cpp

424

Those can't even be encoded for VMEM. M0 isn't in an allocatable class, and exec is reserved so neither one will appear there either. They are also disallowed by the operand constraints so will be a verifier error

[rampitec]

According to the showed ISA for barrier with global and local memory fence this pass produces incorrect code. Please fix.

[nhaehnle]

When applied to current trunk, this code can get into an infinite loop. Please see the sample LLVM IR at https://paste.debian.net/926533/

When compiled with llc -march=amdgcn -mcpu=bonaire, the new pass gets into an infinite loop. When compiled for Tonga, no infinite loop is encountered.

[kanarayan]

Fix the issue related to barrier.

[kanarayan]

Fix the looping issue found with the GL input. The bug only triggers for some targets under some conditions. This is due to an earlier fix for a VCCZ bug; that code needs to go away once the scheduler is fixed to handle this scenario. I need to add a lit test.

[kanarayan]

This update includes all the fixes so far, but keeps the old pass the default. We would like to get broader test coverage under the option first, and then add the tests and turn the new pass on by default.

[t-tye]

LGTM

[arsenm]

lib/Target/AMDGPU/SIInsertWaitcnts.cpp

569–580

This can be deleted

lib/Target/AMDGPU/AMDGPU.h

 FunctionPass *createSILowerI1CopiesPass();
 FunctionPass *createSIShrinkInstructionsPass();
 FunctionPass *createSILoadStoreOptimizerPass(TargetMachine &tm);
 FunctionPass *createSIWholeQuadModePass();
 FunctionPass *createSIFixControlFlowLiveIntervalsPass();
 FunctionPass *createSIFixSGPRCopiesPass();
 FunctionPass *createSIDebuggerInsertNopsPass();
 FunctionPass *createSIInsertWaitsPass();
+FunctionPass *createSIInsertWaitcntsPass();
 FunctionPass *createAMDGPUCodeGenPreparePass(const GCNTargetMachine *TM = nullptr);
 
 ModulePass *createAMDGPUAnnotateKernelFeaturesPass(const TargetMachine *TM = nullptr);
 void initializeAMDGPUAnnotateKernelFeaturesPass(PassRegistry &);
 extern char &AMDGPUAnnotateKernelFeaturesID;
 
 ModulePass *createAMDGPULowerIntrinsicsPass();
 void initializeAMDGPULowerIntrinsicsPass(PassRegistry &);

 extern char &SIAnnotateControlFlowPassID;
 
 void initializeSIDebuggerInsertNopsPass(PassRegistry&);
 extern char &SIDebuggerInsertNopsID;
 
 void initializeSIInsertWaitsPass(PassRegistry&);
 extern char &SIInsertWaitsID;
 
+void initializeSIInsertWaitcntsPass(PassRegistry&);
+extern char &SIInsertWaitcntsID;
+
 void initializeAMDGPUUnifyDivergentExitNodesPass(PassRegistry&);
 extern char &AMDGPUUnifyDivergentExitNodesID;
 
 ImmutablePass *createAMDGPUAAWrapperPass();
 void initializeAMDGPUAAWrapperPassPass(PassRegistry&);
 
 Target &getTheAMDGPUTarget();
 Target &getTheGCNTarget();

lib/Target/AMDGPU/AMDGPUTargetMachine.cpp

static cl::opt<bool> EnableSDWAPeephole(
   cl::desc("Enable SDWA peepholer"),
   cl::init(true));
 
 // Enable address space based alias analysis
 static cl::opt<bool> EnableAMDGPUAliasAnalysis("enable-amdgpu-aa", cl::Hidden,
   cl::desc("Enable AMDGPU Alias Analysis"),
   cl::init(true));
 
+// Option to enable new waitcnt insertion pass.
+static cl::opt<bool> EnableSIInsertWaitcntsPass(
+  "enable-si-insert-waitcnts",
+  cl::desc("Use new waitcnt insertion pass"),
+  cl::init(true));
+
 extern "C" void LLVMInitializeAMDGPUTarget() {
   // Register the target
   RegisterTargetMachine<R600TargetMachine> X(getTheAMDGPUTarget());
   RegisterTargetMachine<GCNTargetMachine> Y(getTheGCNTarget());
 
   PassRegistry *PR = PassRegistry::getPassRegistry();
   initializeSILowerI1CopiesPass(*PR);
   initializeSIFixSGPRCopiesPass(*PR);
   initializeSIFixVGPRCopiesPass(*PR);
   initializeSIFoldOperandsPass(*PR);
   initializeSIPeepholeSDWAPass(*PR);
   initializeSIShrinkInstructionsPass(*PR);
   initializeSIFixControlFlowLiveIntervalsPass(*PR);
   initializeSILoadStoreOptimizerPass(*PR);
   initializeAMDGPUAnnotateKernelFeaturesPass(*PR);
   initializeAMDGPUAnnotateUniformValuesPass(*PR);
   initializeAMDGPULowerIntrinsicsPass(*PR);
   initializeAMDGPUPromoteAllocaPass(*PR);
   initializeAMDGPUCodeGenPreparePass(*PR);
   initializeAMDGPUUnifyMetadataPass(*PR);
   initializeSIAnnotateControlFlowPass(*PR);
   initializeSIInsertWaitsPass(*PR);
+  initializeSIInsertWaitcntsPass(*PR);
   initializeSIWholeQuadModePass(*PR);
   initializeSILowerControlFlowPass(*PR);
   initializeSIInsertSkipsPass(*PR);
   initializeSIDebuggerInsertNopsPass(*PR);
   initializeSIOptimizeExecMaskingPass(*PR);
   initializeAMDGPUUnifyDivergentExitNodesPass(*PR);
   initializeAMDGPUAAWrapperPassPass(*PR);
 }
void GCNPassConfig::addPreEmitPass() {
   // are multiple scheduling regions in a basic block, the regions are scheduled
   // bottom up, so when we begin to schedule a region we don't know what
   // instructions were emitted directly before it.
   //
   // Here we add a stand-alone hazard recognizer pass which can handle all
   // cases.
   addPass(&PostRAHazardRecognizerID);
 
+  if (EnableSIInsertWaitcntsPass)
+    addPass(createSIInsertWaitcntsPass());
+  else
     addPass(createSIInsertWaitsPass());
   addPass(createSIShrinkInstructionsPass());
   addPass(&SIInsertSkipsPassID);
   addPass(createSIDebuggerInsertNopsPass());
   addPass(&BranchRelaxationPassID);
 }
 
 TargetPassConfig *GCNTargetMachine::createPassConfig(PassManagerBase &PM) {
   return new GCNPassConfig(this, PM);
 }
 

lib/Target/AMDGPU/CMakeLists.txt

add_llvm_target(AMDGPUCodeGen
   SIDebuggerInsertNops.cpp
   SIFixControlFlowLiveIntervals.cpp
   SIFixSGPRCopies.cpp
   SIFixVGPRCopies.cpp
   SIFoldOperands.cpp
   SIFrameLowering.cpp
   SIInsertSkips.cpp
   SIInsertWaits.cpp
+  SIInsertWaitcnts.cpp
   SIInstrInfo.cpp
   SIISelLowering.cpp
   SILoadStoreOptimizer.cpp
   SILowerControlFlow.cpp
   SILowerI1Copies.cpp
   SIMachineFunctionInfo.cpp
   SIMachineScheduler.cpp
   SIOptimizeExecMasking.cpp

lib/Target/AMDGPU/DSInstructions.td

 class DS_1A1D_PERMUTE <string opName, SDPatternOperator node = null_frag>
 : DS_Pseudo<opName,
   (outs VGPR_32:$vdst),
   (ins VGPR_32:$addr, VGPR_32:$data0, offset:$offset),
   "$vdst, $addr, $data0$offset",
   [(set i32:$vdst,
    (node (DS1Addr1Offset i32:$addr, i16:$offset), i32:$data0))] > {
 
+  let LGKM_CNT = 0;
+
   let mayLoad = 0;
   let mayStore = 0;
   let isConvergent = 1;
 
   let has_data1 = 0;
   let has_gds = 0;
 }
 

lib/Target/AMDGPU/SIInsertWaitcnts.cpp

+//===-- SIInsertWaitcnts.cpp - Insert Wait Instructions --------------------===/
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+/// \file
+/// \brief Insert wait instructions for memory reads and writes.
+///
+/// Memory reads and writes are issued asynchronously, so we need to insert
+/// S_WAITCNT instructions when we want to access any of their results or
+/// overwrite any register that's used asynchronously.
+//
+//===----------------------------------------------------------------------===//
+
+#include "AMDGPU.h"
+#include "AMDGPUSubtarget.h"
+#include "SIDefines.h"
+#include "SIInstrInfo.h"
+#include "SIMachineFunctionInfo.h"
+#include "Utils/AMDGPUBaseInfo.h"
+#include "llvm/ADT/PostOrderIterator.h"
+#include "llvm/CodeGen/MachineFunction.h"
+#include "llvm/CodeGen/MachineFunctionPass.h"
+#include "llvm/CodeGen/MachineInstrBuilder.h"
+#include "llvm/CodeGen/MachineRegisterInfo.h"
+
+#define DEBUG_TYPE "si-insert-waitcnts"
+
+using namespace llvm;
+
+namespace {
+
+// Class of object that encapsulates latest instruction counter score
+// associated with the operand.  Used for determining whether
+// s_waitcnt instruction needs to be emited.
+
+#define CNT_MASK(t) (1u << (t))
+
+enum InstCounterType { VM_CNT = 0, LGKM_CNT, EXP_CNT, NUM_INST_CNTS };
+
+typedef std::pair<signed, signed> RegInterval;
+
+struct {
+  int32_t VmcntMax;
+  int32_t ExpcntMax;
+  int32_t LgkmcntMax;
+  int32_t NumVGPRsMax;
+  int32_t NumSGPRsMax;
+} HardwareLimits;
+
+struct {
+  unsigned VGPR0;
+  unsigned VGPRL;
+  unsigned SGPR0;
+  unsigned SGPRL;
+} RegisterEncoding;
+
+enum WaitEventType {
+  VMEM_ACCESS,      // vector-memory read & write
+  LDS_ACCESS,       // lds read & write
+  GDS_ACCESS,       // gds read & write
+  SQ_MESSAGE,       // send message
+  SMEM_ACCESS,      // scalar-memory read & write
+  EXP_GPR_LOCK,     // export holding on its data src
+  GDS_GPR_LOCK,     // GDS holding on its data and addr src
+  EXP_POS_ACCESS,   // write to export position
+  EXP_PARAM_ACCESS, // write to export parameter
+  VMW_GPR_LOCK,     // vector-memory write holding on its data src
+  NUM_WAIT_EVENTS,
+};
+
+// The mapping is:
+//  0                .. SQ_MAX_PGM_VGPRS-1               real VGPRs
+//  SQ_MAX_PGM_VGPRS .. NUM_ALL_VGPRS-1                  extra VGPR-like slots
+//  NUM_ALL_VGPRS    .. NUM_ALL_VGPRS+SQ_MAX_PGM_SGPRS-1 real SGPRs
+// We reserve a fixed number of VGPR slots in the scoring tables for
+// special tokens like SCMEM_LDS (needed for buffer load to LDS).
+enum RegisterMapping {
+  SQ_MAX_PGM_VGPRS = 256, // Maximum programmable VGPRs across all targets.
+  SQ_MAX_PGM_SGPRS = 256, // Maximum programmable SGPRs across all targets.
+  NUM_EXTRA_VGPRS = 1,    // A reserved slot for DS.
+  EXTRA_VGPR_LDS = 0,     // This is a placeholder the Shader algorithm uses.
+  NUM_ALL_VGPRS = SQ_MAX_PGM_VGPRS + NUM_EXTRA_VGPRS, // Where SGPR starts.
+};
+
+#define ForAllWaitEventType(w)                                                 \
+  for (enum WaitEventType w = (enum WaitEventType)0;                           \
+       (w) < (enum WaitEventType)NUM_WAIT_EVENTS;                              \
+       (w) = (enum WaitEventType)((w) + 1))
+
+// This is a per-basic-block object that maintains current score brackets
+// of each wait-counter, and a per-register scoreboard for each wait-couner.
+// We also maintain the latest score for every event type that can change the
+// waitcnt in order to know if there are multiple types of events within
+// the brackets. When multiple types of event happen in the bracket,
+// wait-count may get decreased out of order, therefore we need to put in
+// "s_waitcnt 0" before use.
+class BlockWaitcntBrackets {
+public:
+  static int32_t getWaitCountMax(InstCounterType T) {
+    switch (T) {
+    case VM_CNT:
+      return HardwareLimits.VmcntMax;
+    case LGKM_CNT:
+      return HardwareLimits.LgkmcntMax;
+    case EXP_CNT:
+      return HardwareLimits.ExpcntMax;
+    default:
+      break;
+    }
+    return 0;
+  };
+
+  void setScoreLB(InstCounterType T, int32_t Val) {
+    assert(T < NUM_INST_CNTS);
+    if (T >= NUM_INST_CNTS)
+      return;
+    ScoreLBs[T] = Val;
+  };
+
+  void setScoreUB(InstCounterType T, int32_t Val) {
+    assert(T < NUM_INST_CNTS);
+    if (T >= NUM_INST_CNTS)
+      return;
+    ScoreUBs[T] = Val;
+    if (T == EXP_CNT) {
+      int32_t UB = (int)(ScoreUBs[T] - getWaitCountMax(EXP_CNT));
+      if (ScoreLBs[T] < UB)
+        ScoreLBs[T] = UB;
+    }
+  };
+
+  int32_t getScoreLB(InstCounterType T) {
+    assert(T < NUM_INST_CNTS);
+    if (T >= NUM_INST_CNTS)
+      return 0;
+    return ScoreLBs[T];
+  };
+
+  int32_t getScoreUB(InstCounterType T) {
+    assert(T < NUM_INST_CNTS);
+    if (T >= NUM_INST_CNTS)
+      return 0;
+    return ScoreUBs[T];
+  };
+
+  // Mapping from event to counter.
+  InstCounterType eventCounter(WaitEventType E) {
+    switch (E) {
+    case VMEM_ACCESS:
+      return VM_CNT;
+    case LDS_ACCESS:
+    case GDS_ACCESS:
+    case SQ_MESSAGE:
+    case SMEM_ACCESS:
+      return LGKM_CNT;
+    case EXP_GPR_LOCK:
+    case GDS_GPR_LOCK:
+    case VMW_GPR_LOCK:
+    case EXP_POS_ACCESS:
+    case EXP_PARAM_ACCESS:
+      return EXP_CNT;
+    default:
+      llvm_unreachable("unhandled event type");
+    }
+    return NUM_INST_CNTS;
+  }
+
+  void setRegScore(int GprNo, InstCounterType T, int32_t Val) {
+    if (GprNo < NUM_ALL_VGPRS) {
+      if (GprNo > VgprUB) {
+        VgprUB = GprNo;
+      }
+      VgprScores[T][GprNo] = Val;
+    } else {
+      assert(T == LGKM_CNT);
+      if (GprNo - NUM_ALL_VGPRS > SgprUB) {
+        SgprUB = GprNo - NUM_ALL_VGPRS;
+      }
+      SgprScores[GprNo - NUM_ALL_VGPRS] = Val;
+    }
+  }
+
+  int32_t getRegScore(int GprNo, InstCounterType T) {
+    if (GprNo < NUM_ALL_VGPRS) {
+      return VgprScores[T][GprNo];
+    }
+    return SgprScores[GprNo - NUM_ALL_VGPRS];
+  }
+
+  void clear() {
+    memset(ScoreLBs, 0, sizeof(ScoreLBs));
+    memset(ScoreUBs, 0, sizeof(ScoreUBs));
+    memset(EventUBs, 0, sizeof(EventUBs));
+    for (enum InstCounterType T = VM_CNT; T < NUM_INST_CNTS;
+         T = (enum InstCounterType)(T + 1)) {
+      memset(VgprScores[T], 0, sizeof(VgprScores[T]));
+    }
+    memset(SgprScores, 0, sizeof(SgprScores));
+  }
+
+  RegInterval getRegInterval(const MachineInstr *MI, const SIInstrInfo *TII,
+                             const MachineRegisterInfo *MRI,
+                             const SIRegisterInfo *TRI, unsigned OpNo,
+                             bool Def) const;
+
+  void setExpScore(const MachineInstr *MI, const SIInstrInfo *TII,
+                   const SIRegisterInfo *TRI, const MachineRegisterInfo *MRI,
+                   unsigned OpNo, int32_t Val);
+
+  void setWaitAtBeginning() { WaitAtBeginning = true; }
+  void clearWaitAtBeginning() { WaitAtBeginning = false; }
+  bool getWaitAtBeginning() const { return WaitAtBeginning; }
+  void setEventUB(enum WaitEventType W, int32_t Val) { EventUBs[W] = Val; }
+  int32_t getMaxVGPR() const { return VgprUB; }
+  int32_t getMaxSGPR() const { return SgprUB; }
+  int32_t getEventUB(enum WaitEventType W) const {
+    assert(W < NUM_WAIT_EVENTS);
+    return EventUBs[W];
+  }
+  bool counterOutOfOrder(InstCounterType T);
+  unsigned int updateByWait(InstCounterType T, int ScoreToWait);
+  void updateByEvent(const SIInstrInfo *TII, const SIRegisterInfo *TRI,
+                     const MachineRegisterInfo *MRI, WaitEventType E,
+                     MachineInstr &MI);
+
+  BlockWaitcntBrackets()
+      : WaitAtBeginning(false), ValidLoop(false), MixedExpTypes(false),
+        LoopRegion(NULL), PostOrder(0), Waitcnt(NULL), VgprUB(0), SgprUB(0) {
+    for (enum InstCounterType T = VM_CNT; T < NUM_INST_CNTS;
+         T = (enum InstCounterType)(T + 1)) {
+      memset(VgprScores[T], 0, sizeof(VgprScores[T]));
+    }
+  }
+  ~BlockWaitcntBrackets(){};
+
+  bool hasPendingSMEM() const {
+    return (EventUBs[SMEM_ACCESS] > ScoreLBs[LGKM_CNT] &&
+            EventUBs[SMEM_ACCESS] <= ScoreUBs[LGKM_CNT]);
+  }
+
+  bool hasPendingFlat() const {
+    return ((LastFlat[LGKM_CNT] > ScoreLBs[LGKM_CNT] &&
+             LastFlat[LGKM_CNT] <= ScoreUBs[LGKM_CNT]) ||
+            (LastFlat[VM_CNT] > ScoreLBs[VM_CNT] &&
+             LastFlat[VM_CNT] <= ScoreUBs[VM_CNT]));
+  }
+
+  void setPendingFlat() {
+    LastFlat[VM_CNT] = ScoreUBs[VM_CNT];
+    LastFlat[LGKM_CNT] = ScoreUBs[LGKM_CNT];
+  }
+
+  int pendingFlat(InstCounterType Ct) const { return LastFlat[Ct]; }
+
+  void setLastFlat(InstCounterType Ct, int Val) { LastFlat[Ct] = Val; }
+
+  bool getRevisitLoop() const { return RevisitLoop; }
+  void setRevisitLoop(bool RevisitLoopIn) { RevisitLoop = RevisitLoopIn; }
+
+  void setPostOrder(int32_t PostOrderIn) { PostOrder = PostOrderIn; }
+  int32_t getPostOrder() const { return PostOrder; }
+
+  void setWaitcnt(MachineInstr *WaitcntIn) { Waitcnt = WaitcntIn; }
+  void clearWaitcnt() { Waitcnt = NULL; }
+  MachineInstr *getWaitcnt() const { return Waitcnt; }
+
+  bool mixedExpTypes() const { return MixedExpTypes; }
+  void setMixedExpTypes(bool MixedExpTypesIn) {
+    MixedExpTypes = MixedExpTypesIn;
+  }
+
+  void print(raw_ostream &);
+  void dump() { print(dbgs()); }
+
+private:
+  bool WaitAtBeginning;
+  bool RevisitLoop;
+  bool ValidLoop;
+  bool MixedExpTypes;
+  MachineLoop *LoopRegion;
+  int32_t PostOrder;
+  MachineInstr *Waitcnt;
+  int32_t ScoreLBs[NUM_INST_CNTS] = {0};
+  int32_t ScoreUBs[NUM_INST_CNTS] = {0};
+  int32_t EventUBs[NUM_WAIT_EVENTS] = {0};
+  // Remember the last flat memory operation.
+  int32_t LastFlat[NUM_INST_CNTS] = {0};
+  // wait_cnt scores for every vgpr.
+  // Keep track of the VgprUB and SgprUB to make merge at join efficient.
+  int32_t VgprUB;
+  int32_t SgprUB;
+  int32_t VgprScores[NUM_INST_CNTS][NUM_ALL_VGPRS];
+  // Wait cnt scores for every sgpr, only lgkmcnt is relevant.
+  int32_t SgprScores[SQ_MAX_PGM_SGPRS] = {0};
+};
+
+// This is a per-loop-region object that records waitcnt status at the end of
+// loop footer from the previous iteration. We also maintain an iteration
+// count to track the number of times the loop has been visited. When it
+// doesn't converge naturally, we force convergence by inserting s_waitcnt 0
+// at the end of the loop footer.
+class LoopWaitcntData {
+public:
+  void incIterCnt() { IterCnt++; }
+  void resetIterCnt() { IterCnt = 0; }
+  int32_t getIterCnt() { return IterCnt; }
+
+  LoopWaitcntData() : LfWaitcnt(NULL), IterCnt(0) {}
+  ~LoopWaitcntData(){};
+
+  void setWaitcnt(MachineInstr *WaitcntIn) { LfWaitcnt = WaitcntIn; }
+  MachineInstr *getWaitcnt() const { return LfWaitcnt; }
+
+  void print() {
+    DEBUG(dbgs() << "  iteration " << IterCnt << '\n';);
+    return;
+  }
+
+private:
+  // s_waitcnt added at the end of loop footer to stablize wait scores
+  // at the end of the loop footer.
+  MachineInstr *LfWaitcnt;
+  // Number of iterations the loop has been visited, not including the initial
+  // walk over.
+  int32_t IterCnt;
+};
+
+class SIInsertWaitcnts : public MachineFunctionPass {
+
+private:
+  const SISubtarget *ST;
+  const SIInstrInfo *TII;
+  const SIRegisterInfo *TRI;
+  const MachineRegisterInfo *MRI;
+  const MachineLoopInfo *MLI;
+  AMDGPU::IsaInfo::IsaVersion IV;
+  AMDGPUAS AMDGPUASI;
+
+  DenseSet<MachineBasicBlock *> BlockVisitedSet;
+  DenseSet<MachineInstr *> CompilerGeneratedWaitcntSet;
+
+  DenseMap<MachineBasicBlock *, std::unique_ptr<BlockWaitcntBrackets>>
+      BlockWaitcntBracketsMap;
+
+  DenseSet<MachineBasicBlock *> BlockWaitcntProcessedSet;
+
+  DenseMap<MachineLoop *, std::unique_ptr<LoopWaitcntData>> LoopWaitcntDataMap;
+
+  std::vector<std::unique_ptr<BlockWaitcntBrackets>> KillWaitBrackets;
+
+public:
+  static char ID;
+
+  SIInsertWaitcnts()
+      : MachineFunctionPass(ID), ST(nullptr), TII(nullptr), TRI(nullptr),
+        MRI(nullptr), MLI(nullptr) {}
+
+  bool runOnMachineFunction(MachineFunction &MF) override;
+
+  StringRef getPassName() const override {
+    return "SI insert wait instructions";
+  }
+
+  void getAnalysisUsage(AnalysisUsage &AU) const override {
+    AU.setPreservesCFG();
+    AU.addRequired<MachineLoopInfo>();
+    MachineFunctionPass::getAnalysisUsage(AU);
+  }
+
+  void addKillWaitBracket(BlockWaitcntBrackets *Bracket) {
+    // The waitcnt information is copied because it changes as the block is
+    // traversed.
+    KillWaitBrackets.push_back(make_unique<BlockWaitcntBrackets>(*Bracket));
+  }
+
+  MachineInstr *generateSWaitCntInstBefore(MachineInstr &MI,
+                                           BlockWaitcntBrackets *ScoreBrackets);
+  void updateEventWaitCntAfter(MachineInstr &Inst,
+                               BlockWaitcntBrackets *ScoreBrackets);
+  void mergeInputScoreBrackets(MachineBasicBlock &Block);
+  MachineBasicBlock *loopBottom(const MachineLoop *Loop);
+  void insertWaitcntInBlock(MachineFunction &MF, MachineBasicBlock &Block);
+  void insertWaitcntBeforeCF(MachineBasicBlock &Block, MachineInstr *Inst);
+};
+
+} // End anonymous namespace.
+
+RegInterval BlockWaitcntBrackets::getRegInterval(const MachineInstr *MI,
+                                                 const SIInstrInfo *TII,
+                                                 const MachineRegisterInfo *MRI,
+                                                 const SIRegisterInfo *TRI,
+                                                 unsigned OpNo,
+                                                 bool Def) const {
+  const MachineOperand &Op = MI->getOperand(OpNo);
+  if (!Op.isReg() || !TRI->isInAllocatableClass(Op.getReg()) ||
+      (Def && !Op.isDef()))
+    return {-1, -1};
+
+  // A use via a PW operand does not need a waitcnt.
+  // A partial write is not a WAW.
+  assert(!Op.getSubReg() || !Op.isUndef());
+
+  RegInterval Result;
+  const MachineRegisterInfo &MRIA = *MRI;
+
+  unsigned Reg = TRI->getEncodingValue(Op.getReg());
+
+  if (TRI->isVGPR(MRIA, Op.getReg())) {
+    assert(Reg >= RegisterEncoding.VGPR0 && Reg <= RegisterEncoding.VGPRL);
+    Result.first = Reg - RegisterEncoding.VGPR0;
+    assert(Result.first >= 0 && Result.first < SQ_MAX_PGM_VGPRS);
+  } else if (TRI->isSGPRReg(MRIA, Op.getReg())) {
+    assert(Reg >= RegisterEncoding.SGPR0 && Reg < SQ_MAX_PGM_SGPRS);
+    Result.first = Reg - RegisterEncoding.SGPR0 + NUM_ALL_VGPRS;
+    assert(Result.first >= NUM_ALL_VGPRS &&
+           Result.first < SQ_MAX_PGM_SGPRS + NUM_ALL_VGPRS);
+  }
+  // TODO: Handle TTMP
+  // else if (TRI->isTTMP(MRIA, Reg.getReg())) ...
+  else
+    return {-1, -1};
+
+  const MachineInstr &MIA = *MI;
+  const TargetRegisterClass *RC = TII->getOpRegClass(MIA, OpNo);
+  unsigned Size = RC->getSize();
+  Result.second = Result.first + (Size / 4);
+
+  return Result;
+}
+
+void BlockWaitcntBrackets::setExpScore(const MachineInstr *MI,
+                                       const SIInstrInfo *TII,
+                                       const SIRegisterInfo *TRI,
+                                       const MachineRegisterInfo *MRI,
+                                       unsigned OpNo, int32_t Val) {
+  RegInterval Interval = getRegInterval(MI, TII, MRI, TRI, OpNo, false);
+  DEBUG({
+    const MachineOperand &Opnd = MI->getOperand(OpNo);
+    assert(TRI->isVGPR(*MRI, Opnd.getReg()));
+  });
+  for (signed RegNo = Interval.first; RegNo < Interval.second; ++RegNo) {
+    setRegScore(RegNo, EXP_CNT, Val);
+  }
+}
+
+void BlockWaitcntBrackets::updateByEvent(const SIInstrInfo *TII,
+                                         const SIRegisterInfo *TRI,
+                                         const MachineRegisterInfo *MRI,
+                                         WaitEventType E, MachineInstr &Inst) {
+  const MachineRegisterInfo &MRIA = *MRI;
+  InstCounterType T = eventCounter(E);
+  int32_t CurrScore = getScoreUB(T) + 1;
+  // EventUB and ScoreUB need to be update regardless if this event changes
+  // the score of a register or not.
+  // Examples including vm_cnt when buffer-store or lgkm_cnt when send-message.
+  EventUBs[E] = CurrScore;
+  setScoreUB(T, CurrScore);
+
+  if (T == EXP_CNT) {
+    // Check for mixed export types. If they are mixed, then a waitcnt exp(0)
+    // is required.
+    if (!MixedExpTypes) {
+      MixedExpTypes = counterOutOfOrder(EXP_CNT);
+    }
+
+    // Put score on the source vgprs. If this is a store, just use those
+    // specific register(s).
+    if (TII->isDS(Inst) && (Inst.mayStore() || Inst.mayLoad())) {
+      // All GDS operations must protect their address register (same as
+      // export.)
+      if (Inst.getOpcode() != AMDGPU::DS_APPEND &&
+          Inst.getOpcode() != AMDGPU::DS_CONSUME) {
+        setExpScore(
+            &Inst, TII, TRI, MRI,
+            AMDGPU::getNamedOperandIdx(Inst.getOpcode(), AMDGPU::OpName::addr),
+            CurrScore);
+      }
+      if (Inst.mayStore()) {
+        setExpScore(
+            &Inst, TII, TRI, MRI,
+            AMDGPU::getNamedOperandIdx(Inst.getOpcode(), AMDGPU::OpName::data0),
+            CurrScore);
+        if (AMDGPU::getNamedOperandIdx(Inst.getOpcode(),
+                                       AMDGPU::OpName::data1) != -1) {
+          setExpScore(&Inst, TII, TRI, MRI,
+                      AMDGPU::getNamedOperandIdx(Inst.getOpcode(),
+                                                 AMDGPU::OpName::data1),
+                      CurrScore);
+        }
+      } else if (AMDGPU::getAtomicNoRetOp(Inst.getOpcode()) != -1 &&
+                 Inst.getOpcode() != AMDGPU::DS_GWS_INIT &&
+                 Inst.getOpcode() != AMDGPU::DS_GWS_SEMA_V &&
+                 Inst.getOpcode() != AMDGPU::DS_GWS_SEMA_BR &&
+                 Inst.getOpcode() != AMDGPU::DS_GWS_SEMA_P &&
+                 Inst.getOpcode() != AMDGPU::DS_GWS_BARRIER &&
+                 Inst.getOpcode() != AMDGPU::DS_APPEND &&
+                 Inst.getOpcode() != AMDGPU::DS_CONSUME &&
+                 Inst.getOpcode() != AMDGPU::DS_ORDERED_COUNT) {
+        for (unsigned I = 0, E = Inst.getNumOperands(); I != E; ++I) {
+          const MachineOperand &Op = Inst.getOperand(I);
+          if (Op.isReg() && !Op.isDef() && TRI->isVGPR(MRIA, Op.getReg())) {
+            setExpScore(&Inst, TII, TRI, MRI, I, CurrScore);
+          }
+        }
+      }
+    } else if (TII->isFLAT(Inst)) {
+      if (Inst.mayStore()) {
+        setExpScore(
+            &Inst, TII, TRI, MRI,
+            AMDGPU::getNamedOperandIdx(Inst.getOpcode(), AMDGPU::OpName::data),
+            CurrScore);
+      } else if (AMDGPU::getAtomicNoRetOp(Inst.getOpcode()) != -1) {
+        setExpScore(
+            &Inst, TII, TRI, MRI,
+            AMDGPU::getNamedOperandIdx(Inst.getOpcode(), AMDGPU::OpName::data),
+            CurrScore);
+      }
+    } else if (TII->isMIMG(Inst)) {
+      if (Inst.mayStore()) {
+        setExpScore(&Inst, TII, TRI, MRI, 0, CurrScore);
+      } else if (AMDGPU::getAtomicNoRetOp(Inst.getOpcode()) != -1) {
+        setExpScore(
+            &Inst, TII, TRI, MRI,
+            AMDGPU::getNamedOperandIdx(Inst.getOpcode(), AMDGPU::OpName::data),
+            CurrScore);
+      }
+    } else if (TII->isMTBUF(Inst)) {
+      if (Inst.mayStore()) {
+        setExpScore(&Inst, TII, TRI, MRI, 0, CurrScore);
+      }
+    } else if (TII->isMUBUF(Inst)) {
+      if (Inst.mayStore()) {
+        setExpScore(&Inst, TII, TRI, MRI, 0, CurrScore);
+      } else if (AMDGPU::getAtomicNoRetOp(Inst.getOpcode()) != -1) {
+        setExpScore(
+            &Inst, TII, TRI, MRI,
+            AMDGPU::getNamedOperandIdx(Inst.getOpcode(), AMDGPU::OpName::data),
+            CurrScore);
+      }
+    } else {
+      if (TII->isEXP(Inst)) {
+        // For export the destination registers are really temps that
+        // can be used as the actual source after export patching, so
+        // we need to treat them like sources and set the EXP_CNT
+        // score.
+        for (unsigned I = 0, E = Inst.getNumOperands(); I != E; ++I) {
+          MachineOperand &DefMO = Inst.getOperand(I);
+          if (DefMO.isReg() && DefMO.isDef() &&
+              TRI->isVGPR(MRIA, DefMO.getReg())) {
+            setRegScore(TRI->getEncodingValue(DefMO.getReg()), EXP_CNT,
+                        CurrScore);
+          }
+        }
+      }
+      for (unsigned I = 0, E = Inst.getNumOperands(); I != E; ++I) {
+        MachineOperand &MO = Inst.getOperand(I);
+        if (MO.isReg() && !MO.isDef() && TRI->isVGPR(MRIA, MO.getReg())) {
+          setExpScore(&Inst, TII, TRI, MRI, I, CurrScore);
+        }
+      }
+    }
+#if 0 // TODO: check if this is handled by MUBUF code above.
+  } else if (Inst.getOpcode() == AMDGPU::BUFFER_STORE_DWORD ||
+       Inst.getOpcode() == AMDGPU::BUFFER_STORE_DWORDX2 ||
+       Inst.getOpcode() == AMDGPU::BUFFER_STORE_DWORDX4) {
+    MachineOperand *MO = TII->getNamedOperand(Inst, AMDGPU::OpName::data);
+    unsigned OpNo;//TODO: find the OpNo for this operand;
+    RegInterval Interval = getRegInterval(&Inst, TII, MRI, TRI, OpNo, false);
+    for (signed RegNo = Interval.first; RegNo < Interval.second;
+   ++RegNo) {
+      setRegScore(RegNo + NUM_ALL_VGPRS, t, CurrScore);
+    }
+#endif
+  } else {
+    // Match the score to the destination registers.
+    for (unsigned I = 0, E = Inst.getNumOperands(); I != E; ++I) {
+      RegInterval Interval = getRegInterval(&Inst, TII, MRI, TRI, I, true);
+      if (T == VM_CNT && Interval.first >= NUM_ALL_VGPRS)
+        continue;
+      for (signed RegNo = Interval.first; RegNo < Interval.second; ++RegNo) {
+        setRegScore(RegNo, T, CurrScore);
+      }
+    }
+    if (TII->isDS(Inst) && Inst.mayStore()) {
+      setRegScore(SQ_MAX_PGM_VGPRS + EXTRA_VGPR_LDS, T, CurrScore);
+    }
+  }
+}
+
+void BlockWaitcntBrackets::print(raw_ostream &OS) {
+  OS << '\n';
+  for (enum InstCounterType T = VM_CNT; T < NUM_INST_CNTS;
+       T = (enum InstCounterType)(T + 1)) {
+    int LB = getScoreLB(T);
+    int UB = getScoreUB(T);
+
+    switch (T) {
+    case VM_CNT:
+      OS << "    VM_CNT(" << UB - LB << "): ";
+      break;
+    case LGKM_CNT:
+      OS << "    LGKM_CNT(" << UB - LB << "): ";
+      break;
+    case EXP_CNT:
+      OS << "    EXP_CNT(" << UB - LB << "): ";
+      break;
+    default:
+      OS << "    UNKNOWN(" << UB - LB << "): ";
+      break;
+    }
+
+    if (LB < UB) {
+      // Print vgpr scores.
+      for (int J = 0; J <= getMaxVGPR(); J++) {
+        int RegScore = getRegScore(J, T);
+        if (RegScore <= LB)
+          continue;
+        int RelScore = RegScore - LB - 1;
+        if (J < SQ_MAX_PGM_VGPRS + EXTRA_VGPR_LDS) {
+          OS << RelScore << ":v" << J << " ";
+        } else {
+          OS << RelScore << ":ds ";
+        }
+      }
+      // Also need to print sgpr scores for lgkm_cnt.
+      if (T == LGKM_CNT) {
+        for (int J = 0; J <= getMaxSGPR(); J++) {
+          int RegScore = getRegScore(J + NUM_ALL_VGPRS, LGKM_CNT);
+          if (RegScore <= LB)
+            continue;
+          int RelScore = RegScore - LB - 1;
+          OS << RelScore << ":s" << J << " ";
+        }
+      }
+    }
+    OS << '\n';
+  }
+  OS << '\n';
+  return;
+}
+
+unsigned int BlockWaitcntBrackets::updateByWait(InstCounterType T,
+                                                int ScoreToWait) {
+  unsigned int NeedWait = 0;
+  if (ScoreToWait == -1) {
+    // The score to wait is unknown. This implies that it was not encountered
+    // during the path of the CFG walk done during the current traversal but
+    // may be seen on a different path. Emit an s_wait counter with a
+    // conservative value of 0 for the counter.
+    NeedWait = CNT_MASK(T);
+    setScoreLB(T, getScoreUB(T));
+    return NeedWait;
+  }
+
+  // If the score of src_operand falls within the bracket, we need an
+  // s_waitcnt instruction.
+  const int32_t LB = getScoreLB(T);
+  const int32_t UB = getScoreUB(T);
+  if ((UB >= ScoreToWait) && (ScoreToWait > LB)) {
+    if (T == VM_CNT && hasPendingFlat()) {
+      // If there is a pending FLAT operation, and this is a VM waitcnt,
+      // then we need to force a waitcnt 0 for VM.
+      NeedWait = CNT_MASK(T);
+      setScoreLB(T, getScoreUB(T));
+    } else if (counterOutOfOrder(T)) {
+      // Counter can get decremented out-of-order when there
+      // are multiple types event in the brack. Also emit an s_wait counter
+      // with a conservative value of 0 for the counter.
+      NeedWait = CNT_MASK(T);
+      setScoreLB(T, getScoreUB(T));
+    } else {
+      NeedWait = CNT_MASK(T);
+      setScoreLB(T, ScoreToWait);
+    }
+  }
+
+  return NeedWait;
+}
+
+// Where there are multiple types of event in the bracket of a counter,
+// the decrement may go out of order.
+bool BlockWaitcntBrackets::counterOutOfOrder(InstCounterType T) {
+  switch (T) {
+  case VM_CNT:
+    return false;
+  case LGKM_CNT: {
+    if (EventUBs[SMEM_ACCESS] > ScoreLBs[LGKM_CNT] &&
+        EventUBs[SMEM_ACCESS] <= ScoreUBs[LGKM_CNT]) {
+      // Scalar memory read always can go out of order.
+      return true;
+    }
+    int NumEventTypes = 0;
+    if (EventUBs[LDS_ACCESS] > ScoreLBs[LGKM_CNT] &&
+        EventUBs[LDS_ACCESS] <= ScoreUBs[LGKM_CNT]) {
+      NumEventTypes++;
+    }
+    if (EventUBs[GDS_ACCESS] > ScoreLBs[LGKM_CNT] &&
+        EventUBs[GDS_ACCESS] <= ScoreUBs[LGKM_CNT]) {
+      NumEventTypes++;
+    }
+    if (EventUBs[SQ_MESSAGE] > ScoreLBs[LGKM_CNT] &&
+        EventUBs[SQ_MESSAGE] <= ScoreUBs[LGKM_CNT]) {
+      NumEventTypes++;
+    }
+    if (NumEventTypes <= 1) {
+      return false;
+    }
+    break;
+  }
+  case EXP_CNT: {
+    // If there has been a mixture of export types, then a waitcnt exp(0) is
+    // required.
+    if (MixedExpTypes)
+      return true;
+    int NumEventTypes = 0;
+    if (EventUBs[EXP_GPR_LOCK] > ScoreLBs[EXP_CNT] &&
+        EventUBs[EXP_GPR_LOCK] <= ScoreUBs[EXP_CNT]) {
+      NumEventTypes++;
+    }
+    if (EventUBs[GDS_GPR_LOCK] > ScoreLBs[EXP_CNT] &&
+        EventUBs[GDS_GPR_LOCK] <= ScoreUBs[EXP_CNT]) {
+      NumEventTypes++;
+    }
+    if (EventUBs[VMW_GPR_LOCK] > ScoreLBs[EXP_CNT] &&
+        EventUBs[VMW_GPR_LOCK] <= ScoreUBs[EXP_CNT]) {
+      NumEventTypes++;
+    }
+    if (EventUBs[EXP_PARAM_ACCESS] > ScoreLBs[EXP_CNT] &&
+        EventUBs[EXP_PARAM_ACCESS] <= ScoreUBs[EXP_CNT]) {
+      NumEventTypes++;
+    }
+
+    if (EventUBs[EXP_POS_ACCESS] > ScoreLBs[EXP_CNT] &&
+        EventUBs[EXP_POS_ACCESS] <= ScoreUBs[EXP_CNT]) {
+      NumEventTypes++;
+    }
+
+    if (NumEventTypes <= 1) {
+      return false;
+    }
+    break;
+  }
+  default:
+    break;
+  }
+  return true;
+}
+
+INITIALIZE_PASS_BEGIN(SIInsertWaitcnts, DEBUG_TYPE, "SI Insert Waitcnts", false,
+                      false)
+INITIALIZE_PASS_END(SIInsertWaitcnts, DEBUG_TYPE, "SI Insert Waitcnts", false,
+                    false)
+
+char SIInsertWaitcnts::ID = 0;
+
+char &llvm::SIInsertWaitcntsID = SIInsertWaitcnts::ID;
+
+FunctionPass *llvm::createSIInsertWaitcntsPass() {
+  return new SIInsertWaitcnts();
+}
+
+static bool readsVCCZ(const MachineInstr &MI) {
+  unsigned Opc = MI.getOpcode();
+  return (Opc == AMDGPU::S_CBRANCH_VCCNZ || Opc == AMDGPU::S_CBRANCH_VCCZ) &&
+         !MI.getOperand(1).isUndef();
+}
+
+///  \brief Generate s_waitcnt instruction to be placed before cur_Inst.
+///  Instructions of a given type are returned in order,
+///  but instructions of different types can complete out of order.
+///  We rely on this in-order completion
+///  and simply assign a score to the memory access instructions.
+///  We keep track of the active "score bracket" to determine
+///  if an access of a memory read requires an s_waitcnt
+///  and if so what the value of each counter is.
+///  The "score bracket" is bound by the lower bound and upper bound
+///  scores (*_score_LB and *_score_ub respectively).
+MachineInstr *SIInsertWaitcnts::generateSWaitCntInstBefore(
+    MachineInstr &MI, BlockWaitcntBrackets *ScoreBrackets) {
+  // To emit, or not to emit - that's the question!
+  // Start with an assumption that there is no need to emit.
+  unsigned int EmitSwaitcnt = 0;
+  // s_waitcnt instruction to return; default is NULL.
+  MachineInstr *SWaitInst = nullptr;
+  // No need to wait before phi. If a phi-move exists, then the wait should
+  // has been inserted before the move. If a phi-move does not exist, then
+  // wait should be inserted before the real use. The same is true for
+  // sc-merge. It is not a coincident that all these cases correspond to the
+  // instructions that are skipped in the assembling loop.
+  bool NeedLineMapping = false; // TODO: Check on this.
+  if (MI.isDebugValue() &&
+      // TODO: any other opcode?
+      !NeedLineMapping) {
+    return SWaitInst;
+  }
+
+  // See if an s_waitcnt is forced at block entry, or is needed at
+  // program end.
+  if (ScoreBrackets->getWaitAtBeginning()) {
+    // Note that we have already cleared the state, so we don't need to update
+    // it.
+    ScoreBrackets->clearWaitAtBeginning();
+    for (enum InstCounterType T = VM_CNT; T < NUM_INST_CNTS;
+         T = (enum InstCounterType)(T + 1)) {
+      EmitSwaitcnt |= CNT_MASK(T);
+      ScoreBrackets->setScoreLB(T, ScoreBrackets->getScoreUB(T));
+    }
+  }
+
+  // See if this instruction has a forced S_WAITCNT VM.
+  // TODO: Handle other cases of NeedsWaitcntVmBefore()
+  else if (MI.getOpcode() == AMDGPU::BUFFER_WBINVL1 ||
+           MI.getOpcode() == AMDGPU::BUFFER_WBINVL1_SC ||
+           MI.getOpcode() == AMDGPU::BUFFER_WBINVL1_VOL) {
+    EmitSwaitcnt |=
+        ScoreBrackets->updateByWait(VM_CNT, ScoreBrackets->getScoreUB(VM_CNT));
+  }
+
+  // All waits must be resolved at call return.
+  // NOTE: this could be improved with knowledge of all call sites or
+  //   with knowledge of the called routines.
+  if (MI.getOpcode() == AMDGPU::RETURN ||
+      MI.getOpcode() == AMDGPU::SI_RETURN_TO_EPILOG) {
+    for (enum InstCounterType T = VM_CNT; T < NUM_INST_CNTS;
+         T = (enum InstCounterType)(T + 1)) {
+      if (ScoreBrackets->getScoreUB(T) > ScoreBrackets->getScoreLB(T)) {
+        ScoreBrackets->setScoreLB(T, ScoreBrackets->getScoreUB(T));
+        EmitSwaitcnt |= CNT_MASK(T);
+      }
+    }
+  }
+  // Resolve vm waits before gs-done.
+  else if ((MI.getOpcode() == AMDGPU::S_SENDMSG ||
+            MI.getOpcode() == AMDGPU::S_SENDMSGHALT) &&
+           ((MI.getOperand(0).getImm() & AMDGPU::SendMsg::ID_MASK_) ==
+            AMDGPU::SendMsg::ID_GS_DONE)) {
+    if (ScoreBrackets->getScoreUB(VM_CNT) > ScoreBrackets->getScoreLB(VM_CNT)) {
+      ScoreBrackets->setScoreLB(VM_CNT, ScoreBrackets->getScoreUB(VM_CNT));
+      EmitSwaitcnt |= CNT_MASK(VM_CNT);
+    }
+  }
+#if 0 // TODO: the following blocks of logic when we have fence.
+  else if (MI.getOpcode() == SC_FENCE) {
+    const unsigned int group_size =
+      context->shader_info->GetMaxThreadGroupSize();
+    // group_size == 0 means thread group size is unknown at compile time
+    const bool group_is_multi_wave =
+      (group_size == 0 || group_size > target_info->GetWaveFrontSize());
+    const bool fence_is_global = !((SCInstInternalMisc*)Inst)->IsGroupFence();
+
+    for (unsigned int i = 0; i < Inst->NumSrcOperands(); i++) {
+      SCRegType src_type = Inst->GetSrcType(i);
+      switch (src_type) {
+        case SCMEM_LDS:
+          if (group_is_multi_wave ||
+        context->OptFlagIsOn(OPT_R1100_LDSMEM_FENCE_CHICKEN_BIT)) {
+            EmitSwaitcnt |= ScoreBrackets->updateByWait(LGKM_CNT,
+                               ScoreBrackets->getScoreUB(LGKM_CNT));
+            // LDS may have to wait for VM_CNT after buffer load to LDS
+            if (target_info->HasBufferLoadToLDS()) {
+              EmitSwaitcnt |= ScoreBrackets->updateByWait(VM_CNT,
+                                 ScoreBrackets->getScoreUB(VM_CNT));
+            }
+          }
+          break;
+
+        case SCMEM_GDS:
+          if (group_is_multi_wave || fence_is_global) {
+            EmitSwaitcnt |= ScoreBrackets->updateByWait(EXP_CNT,
+             ScoreBrackets->getScoreUB(EXP_CNT));
+            EmitSwaitcnt |= ScoreBrackets->updateByWait(LGKM_CNT,
+             ScoreBrackets->getScoreUB(LGKM_CNT));
+          }
+          break;
+
+        case SCMEM_UAV:
+        case SCMEM_TFBUF:
+        case SCMEM_RING:
+        case SCMEM_SCATTER:
+          if (group_is_multi_wave || fence_is_global) {
+            EmitSwaitcnt |= ScoreBrackets->updateByWait(EXP_CNT,
+             ScoreBrackets->getScoreUB(EXP_CNT));
+            EmitSwaitcnt |= ScoreBrackets->updateByWait(VM_CNT,
+             ScoreBrackets->getScoreUB(VM_CNT));
+          }
+          break;
+
+        case SCMEM_SCRATCH:
+        default:
+          break;
+      }
+    }
+  }
+#endif
+
+  // Export & GDS instructions do not read the EXEC mask until after the export
+  // is granted (which can occur well after the instruction is issued).
+  // The shader program must flush all EXP operations on the export-count
+  // before overwriting the EXEC mask.
+  else {
+    if (MI.modifiesRegister(AMDGPU::EXEC, TRI)) {
+      // Export and GDS are tracked individually, either may trigger a waitcnt
+      // for EXEC.
+      EmitSwaitcnt |= ScoreBrackets->updateByWait(
+          EXP_CNT, ScoreBrackets->getEventUB(EXP_GPR_LOCK));
+      EmitSwaitcnt |= ScoreBrackets->updateByWait(
+          EXP_CNT, ScoreBrackets->getEventUB(EXP_PARAM_ACCESS));
+      EmitSwaitcnt |= ScoreBrackets->updateByWait(
+          EXP_CNT, ScoreBrackets->getEventUB(EXP_POS_ACCESS));
+      EmitSwaitcnt |= ScoreBrackets->updateByWait(
+          EXP_CNT, ScoreBrackets->getEventUB(GDS_GPR_LOCK));
+    }
+
+#if 0 // TODO: the following code to handle CALL.
+    // The argument passing for CALLs should suffice for VM_CNT and LGKM_CNT.
+    // However, there is a problem with EXP_CNT, because the call cannot
+    // easily tell if a register is used in the function, and if it did, then
+    // the referring instruction would have to have an S_WAITCNT, which is
+    // dependent on all call sites. So Instead, force S_WAITCNT for EXP_CNTs
+    // before the call.
+    if (MI.getOpcode() == SC_CALL) {
+      if (ScoreBrackets->getScoreUB(EXP_CNT) >
+    ScoreBrackets->getScoreLB(EXP_CNT)) {
+        ScoreBrackets->setScoreLB(EXP_CNT, ScoreBrackets->getScoreUB(EXP_CNT));
+        EmitSwaitcnt |= CNT_MASK(EXP_CNT);
+      }
+    }
+#endif
+
+    // Look at the source operands of every instruction to see if
+    // any of them results from a previous memory operation that affects
+    // its current usage. If so, an s_waitcnt instruction needs to be
+    // emitted.
+    // If the source operand was defined by a load, add the s_waitcnt
+    // instruction.
+    for (const MachineMemOperand *Memop : MI.memoperands()) {
+      unsigned AS = Memop->getAddrSpace();
+      if (AS != AMDGPUASI.LOCAL_ADDRESS)
+        continue;
+      unsigned RegNo = SQ_MAX_PGM_VGPRS + EXTRA_VGPR_LDS;
+      // VM_CNT is only relevant to vgpr or LDS.
+      EmitSwaitcnt |= ScoreBrackets->updateByWait(
+          VM_CNT, ScoreBrackets->getRegScore(RegNo, VM_CNT));
+    }
+    for (unsigned I = 0, E = MI.getNumOperands(); I != E; ++I) {
+      const MachineOperand &Op = MI.getOperand(I);
+      const MachineRegisterInfo &MRIA = *MRI;
+      RegInterval Interval =
+          ScoreBrackets->getRegInterval(&MI, TII, MRI, TRI, I, false);
+      for (signed RegNo = Interval.first; RegNo < Interval.second; ++RegNo) {
+        if (TRI->isVGPR(MRIA, Op.getReg())) {
+          // VM_CNT is only relevant to vgpr or LDS.
+          EmitSwaitcnt |= ScoreBrackets->updateByWait(
+              VM_CNT, ScoreBrackets->getRegScore(RegNo, VM_CNT));
+        }
+        EmitSwaitcnt |= ScoreBrackets->updateByWait(
+            LGKM_CNT, ScoreBrackets->getRegScore(RegNo, LGKM_CNT));
+      }
+    }
+    // End of for loop that looks at all source operands to decide vm_wait_cnt
+    // and lgk_wait_cnt.
+
+    // Two cases are handled for destination operands:
+    // 1) If the destination operand was defined by a load, add the s_waitcnt
+    // instruction to guarantee the right WAW order.
+    // 2) If a destination operand that was used by a recent export/store ins,
+    // add s_waitcnt on exp_cnt to guarantee the WAR order.
+    if (MI.mayStore()) {
+      for (const MachineMemOperand *Memop : MI.memoperands()) {
+        unsigned AS = Memop->getAddrSpace();
+        if (AS != AMDGPUASI.LOCAL_ADDRESS)
+          continue;
+        unsigned RegNo = SQ_MAX_PGM_VGPRS + EXTRA_VGPR_LDS;
+        EmitSwaitcnt |= ScoreBrackets->updateByWait(
+            VM_CNT, ScoreBrackets->getRegScore(RegNo, VM_CNT));
+        EmitSwaitcnt |= ScoreBrackets->updateByWait(
+            EXP_CNT, ScoreBrackets->getRegScore(RegNo, EXP_CNT));
+      }
+    }
+    for (unsigned I = 0, E = MI.getNumOperands(); I != E; ++I) {
+      MachineOperand &Def = MI.getOperand(I);
+      const MachineRegisterInfo &MRIA = *MRI;
+      RegInterval Interval =
+          ScoreBrackets->getRegInterval(&MI, TII, MRI, TRI, I, true);
+      for (signed RegNo = Interval.first; RegNo < Interval.second; ++RegNo) {
+        if (TRI->isVGPR(MRIA, Def.getReg())) {
+          EmitSwaitcnt |= ScoreBrackets->updateByWait(
+              VM_CNT, ScoreBrackets->getRegScore(RegNo, VM_CNT));
+          EmitSwaitcnt |= ScoreBrackets->updateByWait(
+              EXP_CNT, ScoreBrackets->getRegScore(RegNo, EXP_CNT));
+        }
+        EmitSwaitcnt |= ScoreBrackets->updateByWait(
+            LGKM_CNT, ScoreBrackets->getRegScore(RegNo, LGKM_CNT));
+      }
+    } // End of for loop that looks at all dest operands.
+  }
+
+  // TODO: Tie force zero to a compiler triage option.
+  bool ForceZero = false;
+
+  // Check to see if this is an S_BARRIER, and if an implicit S_WAITCNT 0
+  // occurs before the instruction. Doing it here prevents any additional
+  // S_WAITCNTs from being emitted if the instruction was marked as
+  // requiring a WAITCNT beforehand.
+  if (MI.getOpcode() == AMDGPU::S_BARRIER && ST->needWaitcntBeforeBarrier()) {
+    EmitSwaitcnt |=
+        ScoreBrackets->updateByWait(VM_CNT, ScoreBrackets->getScoreUB(VM_CNT));
+    EmitSwaitcnt |= ScoreBrackets->updateByWait(
+        EXP_CNT, ScoreBrackets->getScoreUB(EXP_CNT));
+    EmitSwaitcnt |= ScoreBrackets->updateByWait(
+        LGKM_CNT, ScoreBrackets->getScoreUB(LGKM_CNT));
+  }
+
+  // TODO: Remove this work-around, enable the assert for Bug 457939
+  //       after fixing the scheduler. Also, the Shader Compiler code is
+  //       independent of target.
+  if (readsVCCZ(MI) && ST->getGeneration() <= SISubtarget::SEA_ISLANDS) {
+    if (ScoreBrackets->getScoreLB(LGKM_CNT) <
+            ScoreBrackets->getScoreUB(LGKM_CNT) &&
+        ScoreBrackets->hasPendingSMEM()) {
+      // Wait on everything, not just LGKM.  vccz reads usually come from
+      // terminators, and we always wait on everything at the end of the
+      // block, so if we only wait on LGKM here, we might end up with
+      // another s_waitcnt inserted right after this if there are non-LGKM
+      // instructions still outstanding.
+      ForceZero = true;
+      EmitSwaitcnt = true;
+    }
+  }
+
+  // Does this operand processing indicate s_wait counter update?
+  if (EmitSwaitcnt) {
+    int CntVal[NUM_INST_CNTS];
+
+    bool UseDefaultWaitcntStrategy = true;
+    if (ForceZero) {
+      // Force all waitcnts to 0.
+      for (enum InstCounterType T = VM_CNT; T < NUM_INST_CNTS;
+           T = (enum InstCounterType)(T + 1)) {
+        ScoreBrackets->setScoreLB(T, ScoreBrackets->getScoreUB(T));
+      }
+      CntVal[VM_CNT] = 0;
+      CntVal[EXP_CNT] = 0;
+      CntVal[LGKM_CNT] = 0;
+      UseDefaultWaitcntStrategy = false;
+    }
+
+    if (UseDefaultWaitcntStrategy) {
+      for (enum InstCounterType T = VM_CNT; T < NUM_INST_CNTS;
+           T = (enum InstCounterType)(T + 1)) {
+        if (EmitSwaitcnt & CNT_MASK(T)) {
+          int Delta =
+              ScoreBrackets->getScoreUB(T) - ScoreBrackets->getScoreLB(T);
+          int MaxDelta = ScoreBrackets->getWaitCountMax(T);
+          if (Delta >= MaxDelta) {
+            Delta = -1;
+            if (T != EXP_CNT) {
+              ScoreBrackets->setScoreLB(
+                  T, ScoreBrackets->getScoreUB(T) - MaxDelta);
+            }
+            EmitSwaitcnt &= ~CNT_MASK(T);
+          }
+          CntVal[T] = Delta;
+        } else {
+          // If we are not waiting for a particular counter then encode
+          // it as -1 which means "don't care."
+          CntVal[T] = -1;
+        }
+      }
+    }
+
+    // If we are not waiting on any counter we can skip the wait altogether.
+    if (EmitSwaitcnt != 0) {
+      MachineInstr *OldWaitcnt = ScoreBrackets->getWaitcnt();
+      int Imm = (!OldWaitcnt) ? 0 : OldWaitcnt->getOperand(0).getImm();
+      if (!OldWaitcnt || (AMDGPU::decodeVmcnt(IV, Imm) !=
+                          (CntVal[VM_CNT] & AMDGPU::getVmcntBitMask(IV))) ||
+          (AMDGPU::decodeExpcnt(IV, Imm) !=
+           (CntVal[EXP_CNT] & AMDGPU::getExpcntBitMask(IV))) ||
+          (AMDGPU::decodeLgkmcnt(IV, Imm) !=
+           (CntVal[LGKM_CNT] & AMDGPU::getLgkmcntBitMask(IV)))) {
+        MachineLoop *ContainingLoop = MLI->getLoopFor(MI.getParent());
+        if (ContainingLoop) {
+          MachineBasicBlock *TBB = ContainingLoop->getTopBlock();
+          BlockWaitcntBrackets *ScoreBracket =
+              BlockWaitcntBracketsMap[TBB].get();
+          if (!ScoreBracket) {
+            assert(BlockVisitedSet.find(TBB) == BlockVisitedSet.end());
+            BlockWaitcntBracketsMap[TBB] = make_unique<BlockWaitcntBrackets>();
+            ScoreBracket = BlockWaitcntBracketsMap[TBB].get();
+          }
+          ScoreBracket->setRevisitLoop(true);
+          DEBUG(dbgs() << "set-revisit: block"
+                       << ContainingLoop->getTopBlock()->getNumber() << '\n';);
+        }
+      }
+
+      // Update an existing waitcount, or make a new one.
+      MachineFunction &MF = *MI.getParent()->getParent();
+      if (OldWaitcnt && OldWaitcnt->getOpcode() != AMDGPU::S_WAITCNT) {
+        SWaitInst = OldWaitcnt;
+      } else {
+        SWaitInst = MF.CreateMachineInstr(TII->get(AMDGPU::S_WAITCNT),
+                                          MI.getDebugLoc());
+        CompilerGeneratedWaitcntSet.insert(SWaitInst);
+      }
+
+      const MachineOperand &Op =
+          MachineOperand::CreateImm(AMDGPU::encodeWaitcnt(
+              IV, CntVal[VM_CNT], CntVal[EXP_CNT], CntVal[LGKM_CNT]));
+      SWaitInst->addOperand(MF, Op);
+
+      if (CntVal[EXP_CNT] == 0) {
+        ScoreBrackets->setMixedExpTypes(false);
+      }
+    }
+  }
+
+  return SWaitInst;
+}
+
+void SIInsertWaitcnts::insertWaitcntBeforeCF(MachineBasicBlock &MBB,
+                                             MachineInstr *Waitcnt) {
+  if (MBB.empty()) {
+    MBB.push_back(Waitcnt);
+    return;
+  }
+
+  MachineBasicBlock::iterator It = MBB.end();
+  MachineInstr *MI = &*(--It);
+  if (MI->isBranch()) {
+    MBB.insert(It, Waitcnt);
+  } else {
+    MBB.push_back(Waitcnt);
+  }
+
+  return;
+}
+
+void SIInsertWaitcnts::updateEventWaitCntAfter(
+    MachineInstr &Inst, BlockWaitcntBrackets *ScoreBrackets) {
+  // Now look at the instruction opcode. If it is a memory access
+  // instruction, update the upper-bound of the appropriate counter's
+  // bracket and the destination operand scores.
+  // TODO: Use the (TSFlags & SIInstrFlags::LGKM_CNT) property everywhere.
+  if (TII->isDS(Inst) && (Inst.mayLoad() || Inst.mayStore())) {
+    if (TII->getNamedOperand(Inst, AMDGPU::OpName::gds)->getImm() != 0) {
+      ScoreBrackets->updateByEvent(TII, TRI, MRI, GDS_ACCESS, Inst);
+      ScoreBrackets->updateByEvent(TII, TRI, MRI, GDS_GPR_LOCK, Inst);
+    } else {
+      ScoreBrackets->updateByEvent(TII, TRI, MRI, LDS_ACCESS, Inst);
+    }
+  } else if (TII->isFLAT(Inst)) {
+    assert(Inst.mayLoad() || Inst.mayStore());
+    ScoreBrackets->updateByEvent(TII, TRI, MRI, VMEM_ACCESS, Inst);
+    ScoreBrackets->updateByEvent(TII, TRI, MRI, LDS_ACCESS, Inst);
+
+    // This is a flat memory operation. Check to see if it has memory
+    // tokens for both LDS and Memory, and if so mark it as a flat.
+    bool FoundLDSMem = false;
+    for (const MachineMemOperand *Memop : Inst.memoperands()) {
+      unsigned AS = Memop->getAddrSpace();
+      if (AS == AMDGPUASI.LOCAL_ADDRESS || AS == AMDGPUASI.FLAT_ADDRESS)
+        FoundLDSMem = true;
+    }
+
+    // This is a flat memory operation, so note it - it will require
+    // that both the VM and LGKM be flushed to zero if it is pending when
+    // a VM or LGKM dependency occurs.
+    if (FoundLDSMem) {
+      ScoreBrackets->setPendingFlat();
+    }
+  } else if (SIInstrInfo::isVMEM(Inst) &&
+             // TODO: get a better carve out.
+             Inst.getOpcode() != AMDGPU::BUFFER_WBINVL1 &&
+             Inst.getOpcode() != AMDGPU::BUFFER_WBINVL1_SC &&
+             Inst.getOpcode() != AMDGPU::BUFFER_WBINVL1_VOL) {
+    ScoreBrackets->updateByEvent(TII, TRI, MRI, VMEM_ACCESS, Inst);
+    if ( // TODO: assumed yes -- target_info->MemWriteNeedsExpWait() &&
+        (Inst.mayStore() || AMDGPU::getAtomicNoRetOp(Inst.getOpcode()))) {
+      ScoreBrackets->updateByEvent(TII, TRI, MRI, VMW_GPR_LOCK, Inst);
+    }
+  } else if (TII->isSMRD(Inst)) {
+    ScoreBrackets->updateByEvent(TII, TRI, MRI, SMEM_ACCESS, Inst);
+  } else {
+    switch (Inst.getOpcode()) {
+    case AMDGPU::S_SENDMSG:
+    case AMDGPU::S_SENDMSGHALT:
+      ScoreBrackets->updateByEvent(TII, TRI, MRI, SQ_MESSAGE, Inst);
+      break;
+    case AMDGPU::EXP:
+    case AMDGPU::EXP_DONE: {
+      int Imm = TII->getNamedOperand(Inst, AMDGPU::OpName::tgt)->getImm();
+      if (Imm >= 32 && Imm <= 63)
+        ScoreBrackets->updateByEvent(TII, TRI, MRI, EXP_PARAM_ACCESS, Inst);
+      else if (Imm >= 12 && Imm <= 15)
+        ScoreBrackets->updateByEvent(TII, TRI, MRI, EXP_POS_ACCESS, Inst);
+      else
+        ScoreBrackets->updateByEvent(TII, TRI, MRI, EXP_GPR_LOCK, Inst);
+      break;
+    }
+    case AMDGPU::S_MEMTIME:
+    case AMDGPU::S_MEMREALTIME:
+      ScoreBrackets->updateByEvent(TII, TRI, MRI, SMEM_ACCESS, Inst);
+      break;
+    default:
+      break;
+    }
+  }
+}
+
+void SIInsertWaitcnts::mergeInputScoreBrackets(MachineBasicBlock &Block) {
+  BlockWaitcntBrackets *ScoreBrackets = BlockWaitcntBracketsMap[&Block].get();
+  int32_t MaxPending[NUM_INST_CNTS] = {0};
+  int32_t MaxFlat[NUM_INST_CNTS] = {0};
+  bool MixedExpTypes = false;
+
+  // Clear the score bracket state.
+  ScoreBrackets->clear();
+
+  // Compute the number of pending elements on block entry.
+
+  // IMPORTANT NOTE: If iterative handling of loops is added, the code will
+  // need to handle single BBs with backedges to themselves. This means that
+  // they will need to retain and not clear their initial state.
+
+  // See if there are any uninitialized predecessors. If so, emit an
+  // s_waitcnt 0 at the beginning of the block.
+  for (MachineBasicBlock *pred : Block.predecessors()) {
+    BlockWaitcntBrackets *PredScoreBrackets =
+        BlockWaitcntBracketsMap[pred].get();
+    bool Visited = BlockVisitedSet.find(pred) != BlockVisitedSet.end();
+    if (!Visited || PredScoreBrackets->getWaitAtBeginning()) {
+      break;
+    }
+    for (enum InstCounterType T = VM_CNT; T < NUM_INST_CNTS;
+         T = (enum InstCounterType)(T + 1)) {
+      int span =
+          PredScoreBrackets->getScoreUB(T) - PredScoreBrackets->getScoreLB(T);
+      MaxPending[T] = std::max(MaxPending[T], span);
+      span =
+          PredScoreBrackets->pendingFlat(T) - PredScoreBrackets->getScoreLB(T);
+      MaxFlat[T] = std::max(MaxFlat[T], span);
+    }
+
+    MixedExpTypes |= PredScoreBrackets->mixedExpTypes();
+  }
+
+  // TODO: Is SC Block->IsMainExit() same as Block.succ_empty()?
+  // Also handle kills for exit block.
+  if (Block.succ_empty() && !KillWaitBrackets.empty()) {
+    for (unsigned int I = 0; I < KillWaitBrackets.size(); I++) {
+      for (enum InstCounterType T = VM_CNT; T < NUM_INST_CNTS;
+           T = (enum InstCounterType)(T + 1)) {
+        int Span = KillWaitBrackets[I]->getScoreUB(T) -
+                   KillWaitBrackets[I]->getScoreLB(T);
+        MaxPending[T] = std::max(MaxPending[T], Span);
+        Span = KillWaitBrackets[I]->pendingFlat(T) -
+               KillWaitBrackets[I]->getScoreLB(T);
+        MaxFlat[T] = std::max(MaxFlat[T], Span);
+      }
+
+      MixedExpTypes |= KillWaitBrackets[I]->mixedExpTypes();
+    }
+  }
+
+  // Special handling for GDS_GPR_LOCK and EXP_GPR_LOCK.
+  for (MachineBasicBlock *Pred : Block.predecessors()) {
+    BlockWaitcntBrackets *PredScoreBrackets =
+        BlockWaitcntBracketsMap[Pred].get();
+    bool Visited = BlockVisitedSet.find(Pred) != BlockVisitedSet.end();
+    if (!Visited || PredScoreBrackets->getWaitAtBeginning()) {
+      break;
+    }
+
+    int GDSSpan = PredScoreBrackets->getEventUB(GDS_GPR_LOCK) -
+                  PredScoreBrackets->getScoreLB(EXP_CNT);
+    MaxPending[EXP_CNT] = std::max(MaxPending[EXP_CNT], GDSSpan);
+    int EXPSpan = PredScoreBrackets->getEventUB(EXP_GPR_LOCK) -
+                  PredScoreBrackets->getScoreLB(EXP_CNT);
+    MaxPending[EXP_CNT] = std::max(MaxPending[EXP_CNT], EXPSpan);
+  }
+
+  // TODO: Is SC Block->IsMainExit() same as Block.succ_empty()?
+  if (Block.succ_empty() && !KillWaitBrackets.empty()) {
+    for (unsigned int I = 0; I < KillWaitBrackets.size(); I++) {
+      int GDSSpan = KillWaitBrackets[I]->getEventUB(GDS_GPR_LOCK) -
+                    KillWaitBrackets[I]->getScoreLB(EXP_CNT);
+      MaxPending[EXP_CNT] = std::max(MaxPending[EXP_CNT], GDSSpan);
+      int EXPSpan = KillWaitBrackets[I]->getEventUB(EXP_GPR_LOCK) -
+                    KillWaitBrackets[I]->getScoreLB(EXP_CNT);
+      MaxPending[EXP_CNT] = std::max(MaxPending[EXP_CNT], EXPSpan);
+    }
+  }
+
+#if 0
+  // LC does not (unlike) add a waitcnt at beginning. Leaving it as marker.
+  // TODO: how does LC distinguish between function entry and main entry?
+  // If this is the entry to a function, force a wait.
+  MachineBasicBlock &Entry = Block.getParent()->front();
+  if (Entry.getNumber() == Block.getNumber()) {
+    ScoreBrackets->setWaitAtBeginning();
+    return;
+  }
+#endif
+
+  // Now set the current Block's brackets to the largest ending bracket.
+  for (enum InstCounterType T = VM_CNT; T < NUM_INST_CNTS;
+       T = (enum InstCounterType)(T + 1)) {
+    ScoreBrackets->setScoreUB(T, MaxPending[T]);
+    ScoreBrackets->setScoreLB(T, 0);
+    ScoreBrackets->setLastFlat(T, MaxFlat[T]);
+  }
+
+  ScoreBrackets->setMixedExpTypes(MixedExpTypes);
+
+  // Set the register scoreboard.
+  for (MachineBasicBlock *Pred : Block.predecessors()) {
+    if (BlockVisitedSet.find(Pred) == BlockVisitedSet.end()) {
+      break;
+    }
+
+    BlockWaitcntBrackets *PredScoreBrackets =
+        BlockWaitcntBracketsMap[Pred].get();
+
+    // Now merge the gpr_reg_score information
+    for (enum InstCounterType T = VM_CNT; T < NUM_INST_CNTS;
+         T = (enum InstCounterType)(T + 1)) {
+      int PredLB = PredScoreBrackets->getScoreLB(T);
+      int PredUB = PredScoreBrackets->getScoreUB(T);
+      if (PredLB < PredUB) {
+        int PredScale = MaxPending[T] - PredUB;
+        // Merge vgpr scores.
+        for (int J = 0; J <= PredScoreBrackets->getMaxVGPR(); J++) {
+          int PredRegScore = PredScoreBrackets->getRegScore(J, T);
+          if (PredRegScore <= PredLB)
+            continue;
+          int NewRegScore = PredScale + PredRegScore;
+          ScoreBrackets->setRegScore(
+              J, T, std::max(ScoreBrackets->getRegScore(J, T), NewRegScore));
+        }
+        // Also need to merge sgpr scores for lgkm_cnt.
+        if (T == LGKM_CNT) {
+          for (int J = 0; J <= PredScoreBrackets->getMaxSGPR(); J++) {
+            int PredRegScore =
+                PredScoreBrackets->getRegScore(J + NUM_ALL_VGPRS, LGKM_CNT);
+            if (PredRegScore <= PredLB)
+              continue;
+            int NewRegScore = PredScale + PredRegScore;
+            ScoreBrackets->setRegScore(
+                J + NUM_ALL_VGPRS, LGKM_CNT,
+                std::max(
+                    ScoreBrackets->getRegScore(J + NUM_ALL_VGPRS, LGKM_CNT),
+                    NewRegScore));
+          }
+        }
+      }
+    }
+
+    // Also merge the WaitEvent information.
+    ForAllWaitEventType(W) {
+      enum InstCounterType T = PredScoreBrackets->eventCounter(W);
+      int PredEventUB = PredScoreBrackets->getEventUB(W);
+      if (PredEventUB > PredScoreBrackets->getScoreLB(T)) {
+        int NewEventUB =
+            MaxPending[T] + PredEventUB - PredScoreBrackets->getScoreUB(T);
+        if (NewEventUB > 0) {
+          ScoreBrackets->setEventUB(
+              W, std::max(ScoreBrackets->getEventUB(W), NewEventUB));
+        }
+      }
+    }
+  }
+
+  // TODO: Is SC Block->IsMainExit() same as Block.succ_empty()?
+  // Set the register scoreboard.
+  if (Block.succ_empty() && !KillWaitBrackets.empty()) {
+    for (unsigned int I = 0; I < KillWaitBrackets.size(); I++) {
+      // Now merge the gpr_reg_score information.
+      for (enum InstCounterType T = VM_CNT; T < NUM_INST_CNTS;
+           T = (enum InstCounterType)(T + 1)) {
+        int PredLB = KillWaitBrackets[I]->getScoreLB(T);
+        int PredUB = KillWaitBrackets[I]->getScoreUB(T);
+        if (PredLB < PredUB) {
+          int PredScale = MaxPending[T] - PredUB;
+          // Merge vgpr scores.
+          for (int J = 0; J <= KillWaitBrackets[I]->getMaxVGPR(); J++) {
+            int PredRegScore = KillWaitBrackets[I]->getRegScore(J, T);
+            if (PredRegScore <= PredLB)
+              continue;
+            int NewRegScore = PredScale + PredRegScore;
+            ScoreBrackets->setRegScore(
+                J, T, std::max(ScoreBrackets->getRegScore(J, T), NewRegScore));
+          }
+          // Also need to merge sgpr scores for lgkm_cnt.
+          if (T == LGKM_CNT) {
+            for (int J = 0; J <= KillWaitBrackets[I]->getMaxSGPR(); J++) {
+              int PredRegScore =
+                  KillWaitBrackets[I]->getRegScore(J + NUM_ALL_VGPRS, LGKM_CNT);
+              if (PredRegScore <= PredLB)
+                continue;
+              int NewRegScore = PredScale + PredRegScore;
+              ScoreBrackets->setRegScore(
+                  J + NUM_ALL_VGPRS, LGKM_CNT,
+                  std::max(
+                      ScoreBrackets->getRegScore(J + NUM_ALL_VGPRS, LGKM_CNT),
+                      NewRegScore));
+            }
+          }
+        }
+      }
+
+      // Also merge the WaitEvent information.
+      ForAllWaitEventType(W) {
+        enum InstCounterType T = KillWaitBrackets[I]->eventCounter(W);
+        int PredEventUB = KillWaitBrackets[I]->getEventUB(W);
+        if (PredEventUB > KillWaitBrackets[I]->getScoreLB(T)) {
+          int NewEventUB =
+              MaxPending[T] + PredEventUB - KillWaitBrackets[I]->getScoreUB(T);
+          if (NewEventUB > 0) {
+            ScoreBrackets->setEventUB(
+                W, std::max(ScoreBrackets->getEventUB(W), NewEventUB));
+          }
+        }
+      }
+    }
+  }
+
+  // Special case handling of GDS_GPR_LOCK and EXP_GPR_LOCK. Merge this for the
+  // sequencing predecessors, because changes to EXEC require waitcnts due to
+  // the delayed nature of these operations.
+  for (MachineBasicBlock *Pred : Block.predecessors()) {
+    if (BlockVisitedSet.find(Pred) == BlockVisitedSet.end()) {
+      break;
+    }
+
+    BlockWaitcntBrackets *PredScoreBrackets =
+        BlockWaitcntBracketsMap[Pred].get();
+
+    int pred_gds_ub = PredScoreBrackets->getEventUB(GDS_GPR_LOCK);
+    if (pred_gds_ub > PredScoreBrackets->getScoreLB(EXP_CNT)) {
+      int new_gds_ub = MaxPending[EXP_CNT] + pred_gds_ub -
+                       PredScoreBrackets->getScoreUB(EXP_CNT);
+      if (new_gds_ub > 0) {
+        ScoreBrackets->setEventUB(
+            GDS_GPR_LOCK,
+            std::max(ScoreBrackets->getEventUB(GDS_GPR_LOCK), new_gds_ub));
+      }
+    }
+    int pred_exp_ub = PredScoreBrackets->getEventUB(EXP_GPR_LOCK);
+    if (pred_exp_ub > PredScoreBrackets->getScoreLB(EXP_CNT)) {
+      int new_exp_ub = MaxPending[EXP_CNT] + pred_exp_ub -
+                       PredScoreBrackets->getScoreUB(EXP_CNT);
+      if (new_exp_ub > 0) {
+        ScoreBrackets->setEventUB(
+            EXP_GPR_LOCK,
+            std::max(ScoreBrackets->getEventUB(EXP_GPR_LOCK), new_exp_ub));
+      }
+    }
+  }
+}
+
+/// Return the "bottom" block of a loop. This differs from
+/// MachineLoop::getBottomBlock in that it works even if the loop is
+/// discontiguous.
+MachineBasicBlock *SIInsertWaitcnts::loopBottom(const MachineLoop *Loop) {
+  MachineBasicBlock *Bottom = Loop->getHeader();
+  for (MachineBasicBlock *MBB : Loop->blocks())
+    if (MBB->getNumber() > Bottom->getNumber())
+      Bottom = MBB;
+  return Bottom;
+}
+
+// Generate s_waitcnt instructions where needed.
+void SIInsertWaitcnts::insertWaitcntInBlock(MachineFunction &MF,
+                                            MachineBasicBlock &Block) {
+  // Initialize the state information.
+  mergeInputScoreBrackets(Block);
+
+  BlockWaitcntBrackets *ScoreBrackets = BlockWaitcntBracketsMap[&Block].get();
+
+  DEBUG({
+    dbgs() << "Block" << Block.getNumber();
+    ScoreBrackets->dump();
+  });
+
+  bool InsertNOP = false;
+
+  // Walk over the instructions.
+  for (MachineBasicBlock::iterator Iter = Block.begin(), E = Block.end();
+       Iter != E;) {
+    MachineInstr &Inst = *Iter;
+    // Remove any previously existing waitcnts.
+    if (Inst.getOpcode() == AMDGPU::S_WAITCNT) {
+      // TODO: Register the old waitcnt and optimize the following waitcnts.
+      // Leaving the previously existing waitcnts is conservatively correct.
+      if (CompilerGeneratedWaitcntSet.find(&Inst) ==
+          CompilerGeneratedWaitcntSet.end())
+        ++Iter;
+      else {
+        ScoreBrackets->setWaitcnt(&Inst);
+        ++Iter;
+        Inst.removeFromParent();
+      }
+      continue;
+    }
+
+    // Kill instructions generate a conditional branch to the endmain block.
+    // Merge the current waitcnt state into the endmain block information.
+    // TODO: Are there other flavors of KILL instruction?
+    if (Inst.getOpcode() == AMDGPU::KILL) {
+      addKillWaitBracket(ScoreBrackets);
+    }
+
+    bool VCCZBugWorkAround = false;
+    if (readsVCCZ(Inst)) {
+      if (ScoreBrackets->getScoreLB(LGKM_CNT) <
+              ScoreBrackets->getScoreUB(LGKM_CNT) &&
+          ScoreBrackets->hasPendingSMEM()) {
+        if (ST->getGeneration() <= SISubtarget::SEA_ISLANDS)
+          VCCZBugWorkAround = true;
+      }
+    }
+
+    // Generate an s_waitcnt instruction to be placed before
+    // cur_Inst, if needed.
+    MachineInstr *SWaitInst = generateSWaitCntInstBefore(Inst, ScoreBrackets);
+
+    if (SWaitInst) {
+      Block.insert(Inst, SWaitInst);
+      if (ScoreBrackets->getWaitcnt() != SWaitInst) {
+        DEBUG(dbgs() << "insertWaitcntInBlock\n"
+                     << "Old Instr: " << Inst << '\n'
+                     << "New Instr: " << *SWaitInst << '\n';);
+      }
+    }
+
+    updateEventWaitCntAfter(Inst, ScoreBrackets);
+
+#if 0 // TODO: implement resource type check controlled by options with ub = LB.
+    // If this instruction generates a S_SETVSKIP because it is an
+    // indexed resource, and we are on Tahiti, then it will also force
+    // an S_WAITCNT vmcnt(0)
+    if (RequireCheckResourceType(Inst, context)) {
+      // Force the score to as if an S_WAITCNT vmcnt(0) is emitted.
+      ScoreBrackets->setScoreLB(VM_CNT,
+           ScoreBrackets->getScoreUB(VM_CNT));
+    }
+#endif
+
+    ScoreBrackets->clearWaitcnt();
+
+    if (SWaitInst) {
+      DEBUG({ SWaitInst->print(dbgs() << '\n'); });
+    }
+    DEBUG({
+      Inst.print(dbgs());
+      ScoreBrackets->dump();
+    });
+
+    // Check to see if this is a GWS instruction. If so, and if this is CI or
+    // VI, then the generated code sequence will include an S_WAITCNT 0.
+    // TODO: Are these the only GWS instructions?
+    if (Inst.getOpcode() == AMDGPU::DS_GWS_INIT ||
+        Inst.getOpcode() == AMDGPU::DS_GWS_SEMA_V ||
+        Inst.getOpcode() == AMDGPU::DS_GWS_SEMA_BR ||
+        Inst.getOpcode() == AMDGPU::DS_GWS_SEMA_P ||
+        Inst.getOpcode() == AMDGPU::DS_GWS_BARRIER) {
+      // TODO: && context->target_info->GwsRequiresMemViolTest() ) {
+      ScoreBrackets->updateByWait(VM_CNT, ScoreBrackets->getScoreUB(VM_CNT));
+      ScoreBrackets->updateByWait(EXP_CNT, ScoreBrackets->getScoreUB(EXP_CNT));
+      ScoreBrackets->updateByWait(LGKM_CNT,
+                                  ScoreBrackets->getScoreUB(LGKM_CNT));
+    }
+
+    // TODO: Remove this work-around after fixing the scheduler and enable the
+    // assert above.
+    if (VCCZBugWorkAround) {
+      // Restore the vccz bit.  Any time a value is written to vcc, the vcc
+      // bit is updated, so we can restore the bit by reading the value of
+      // vcc and then writing it back to the register.
+      BuildMI(Block, Inst, Inst.getDebugLoc(), TII->get(AMDGPU::S_MOV_B64),
+              AMDGPU::VCC)
+          .addReg(AMDGPU::VCC);
+    }
+
+    if (ST->getGeneration() >= SISubtarget::VOLCANIC_ISLANDS) {
+
+      // This avoids a s_nop after a waitcnt has just been inserted.
+      if (!SWaitInst && InsertNOP) {
+        BuildMI(Block, Inst, DebugLoc(), TII->get(AMDGPU::S_NOP)).addImm(0);
+      }
+      InsertNOP = false;
+
+      // Any occurrence of consecutive VMEM or SMEM instructions forms a VMEM
+      // or SMEM clause, respectively.
+      //
+      // The temporary workaround is to break the clauses with S_NOP.
+      //
+      // The proper solution would be to allocate registers such that all source
+      // and destination registers don't overlap, e.g. this is illegal:
+      //   r0 = load r2
+      //   r2 = load r0
+      bool IsSMEM = false;
+      bool IsVMEM = false;
+      if (TII->isSMRD(Inst))
+        IsSMEM = true;
+      else if (TII->usesVM_CNT(Inst))
+        IsVMEM = true;
+
+      ++Iter;
+      if (Iter == E)
+        break;
+
+      MachineInstr &Next = *Iter;
+
+      // TODO: How about consecutive SMEM instructions?
+      //       The comments above says break the clause but the code does not.
+      // if ((TII->isSMRD(next) && isSMEM) ||
+      if (!IsSMEM && TII->usesVM_CNT(Next) && IsVMEM &&
+          // TODO: Enable this check when hasSoftClause is upstreamed.
+          // ST->hasSoftClauses() &&
+          ST->isXNACKEnabled()) {
+        // Insert a NOP to break the clause.
+        InsertNOP = true;
+        continue;
+      }
+
+      // There must be "S_NOP 0" between an instruction writing M0 and
+      // S_SENDMSG.
+      if ((Next.getOpcode() == AMDGPU::S_SENDMSG ||
+           Next.getOpcode() == AMDGPU::S_SENDMSGHALT) &&
+          Inst.definesRegister(AMDGPU::M0))
+        InsertNOP = true;
+
+      continue;
+    }
+
+    ++Iter;
+  }
+
+  // Check if we need to force convergence at loop footer.
+  MachineLoop *ContainingLoop = MLI->getLoopFor(&Block);
+  if (ContainingLoop && loopBottom(ContainingLoop) == &Block) {
+    LoopWaitcntData *WaitcntData = LoopWaitcntDataMap[ContainingLoop].get();
+    WaitcntData->print();
+    DEBUG(dbgs() << '\n';);
+
+    // The iterative waitcnt insertion algorithm aims for optimal waitcnt
+    // placement and doesn't always guarantee convergence for a loop. Each
+    // loop should take at most 2 iterations for it to converge naturally.
+    // When this max is reached and result doesn't converge, we force
+    // convergence by inserting a s_waitcnt at the end of loop footer.
+    if (WaitcntData->getIterCnt() > 2) {
+      // To ensure convergence, need to make wait events at loop footer be no
+      // more than those from the previous iteration.
+      // As a simplification, Instead of tracking individual scores and
+      // generate the precise wait count, just wait on 0.
+      bool HasPending = false;
+      MachineInstr *SWaitInst = WaitcntData->getWaitcnt();
+      for (enum InstCounterType T = VM_CNT; T < NUM_INST_CNTS;
+           T = (enum InstCounterType)(T + 1)) {
+        if (ScoreBrackets->getScoreUB(T) > ScoreBrackets->getScoreLB(T)) {
+          ScoreBrackets->setScoreLB(T, ScoreBrackets->getScoreUB(T));
+          HasPending = true;
+        }
+      }
+
+      if (HasPending) {
+        if (!SWaitInst) {
+          SWaitInst = Block.getParent()->CreateMachineInstr(
+              TII->get(AMDGPU::S_WAITCNT), DebugLoc());
+          CompilerGeneratedWaitcntSet.insert(SWaitInst);
+          const MachineOperand &Op = MachineOperand::CreateImm(0);
+          SWaitInst->addOperand(MF, Op);
+#if 0 // TODO: Format the debug output
+          OutputTransformBanner("insertWaitcntInBlock",0,"Create:",context);
+          OutputTransformAdd(SWaitInst, context);
+#endif
+        }
+#if 0 // TODO: ??
+        _DEV( REPORTED_STATS->force_waitcnt_converge = 1; )
+#endif
+      }
+
+      if (SWaitInst) {
+        DEBUG({
+          SWaitInst->print(dbgs());
+          dbgs() << "\nAdjusted score board:";
+          ScoreBrackets->dump();
+        });
+
+        // Add this waitcnt to the block. It is either newly created or
+        // created in previous iterations and added back since block traversal
+        // always remove waitcnt.
+        insertWaitcntBeforeCF(Block, SWaitInst);
+        WaitcntData->setWaitcnt(SWaitInst);
+      }
+    }
+  }
+}
+
+bool SIInsertWaitcnts::runOnMachineFunction(MachineFunction &MF) {
+  ST = &MF.getSubtarget<SISubtarget>();
+  TII = ST->getInstrInfo();
+  TRI = &TII->getRegisterInfo();
+  MRI = &MF.getRegInfo();
+  MLI = &getAnalysis<MachineLoopInfo>();
+  IV = AMDGPU::IsaInfo::getIsaVersion(ST->getFeatureBits());
+  AMDGPUASI = ST->getAMDGPUAS();
+
+  HardwareLimits.VmcntMax = AMDGPU::getVmcntBitMask(IV);
+  HardwareLimits.ExpcntMax = AMDGPU::getExpcntBitMask(IV);
+  HardwareLimits.LgkmcntMax = AMDGPU::getLgkmcntBitMask(IV);
+
+  HardwareLimits.NumVGPRsMax = ST->getAddressableNumVGPRs();
+  HardwareLimits.NumSGPRsMax = ST->getAddressableNumSGPRs();
+  assert(HardwareLimits.NumVGPRsMax <= SQ_MAX_PGM_VGPRS);
+  assert(HardwareLimits.NumSGPRsMax <= SQ_MAX_PGM_SGPRS);
+
+  RegisterEncoding.VGPR0 = TRI->getEncodingValue(AMDGPU::VGPR0);
+  RegisterEncoding.VGPRL =
+      RegisterEncoding.VGPR0 + HardwareLimits.NumVGPRsMax - 1;
+  RegisterEncoding.SGPR0 = TRI->getEncodingValue(AMDGPU::SGPR0);
+  RegisterEncoding.SGPRL =
+      RegisterEncoding.SGPR0 + HardwareLimits.NumSGPRsMax - 1;
+
+  // Walk over the blocks in reverse post-dominator order, inserting
+  // s_waitcnt where needed.
+  ReversePostOrderTraversal<MachineFunction *> RPOT(&MF);
+  bool Modified = false;
+  for (ReversePostOrderTraversal<MachineFunction *>::rpo_iterator
+           I = RPOT.begin(),
+           E = RPOT.end(), J = RPOT.begin();
+       I != E;) {
+    MachineBasicBlock &MBB = **I;
+
+    BlockVisitedSet.insert(&MBB);
+
+    BlockWaitcntBrackets *ScoreBrackets = BlockWaitcntBracketsMap[&MBB].get();
+    if (!ScoreBrackets) {
+      BlockWaitcntBracketsMap[&MBB] = make_unique<BlockWaitcntBrackets>();
+      ScoreBrackets = BlockWaitcntBracketsMap[&MBB].get();
+    }
+    ScoreBrackets->setPostOrder(MBB.getNumber());
+    MachineLoop *ContainingLoop = MLI->getLoopFor(&MBB);
+    if (ContainingLoop && LoopWaitcntDataMap[ContainingLoop] == nullptr)
+      LoopWaitcntDataMap[ContainingLoop] = make_unique<LoopWaitcntData>();
+
+    // If we are walking into the block from before the loop, then guarantee
+    // at least 1 re-walk over the loop to propagate the information, even if
+    // no S_WAITCNT instructions were generated.
+    if (ContainingLoop && ContainingLoop->getTopBlock() == &MBB && J < I &&
+        (BlockWaitcntProcessedSet.find(&MBB) ==
+         BlockWaitcntProcessedSet.end())) {
+      BlockWaitcntBracketsMap[&MBB]->setRevisitLoop(true);
+      DEBUG(dbgs() << "set-revisit: block"
+                   << ContainingLoop->getTopBlock()->getNumber() << '\n';);
+    }
+
+    // Walk over the instructions.
+    insertWaitcntInBlock(MF, MBB);
+
+    // Flag that waitcnts have been processed at least once.
+    BlockWaitcntProcessedSet.insert(&MBB);
+
+    // See if we want to revisit the loop.
+    if (ContainingLoop && loopBottom(ContainingLoop) == &MBB) {
+      MachineBasicBlock *EntryBB = ContainingLoop->getTopBlock();
+      BlockWaitcntBrackets *EntrySB = BlockWaitcntBracketsMap[EntryBB].get();
+      if (EntrySB && EntrySB->getRevisitLoop()) {
+        EntrySB->setRevisitLoop(false);
+        J = I;
+        int32_t PostOrder = EntrySB->getPostOrder();
+        // TODO: Avoid this loop. Find another way to set I.
+        for (ReversePostOrderTraversal<MachineFunction *>::rpo_iterator
+                 X = RPOT.begin(),
+                 Y = RPOT.end();
+             X != Y; ++X) {
+          MachineBasicBlock &MBBX = **X;
+          if (MBBX.getNumber() == PostOrder) {
+            I = X;
+            break;
+          }
+        }
+        LoopWaitcntData *WaitcntData = LoopWaitcntDataMap[ContainingLoop].get();
+        WaitcntData->incIterCnt();
+        DEBUG(dbgs() << "revisit: block" << EntryBB->getNumber() << '\n';);
+        continue;
+      } else {
+        LoopWaitcntData *WaitcntData = LoopWaitcntDataMap[ContainingLoop].get();
+        // Loop converged, reset iteration count. If this loop gets revisited,
+        // it must be from an outer loop, the counter will restart, this will
+        // ensure we don't force convergence on such revisits.
+        WaitcntData->resetIterCnt();
+      }
+    }
+
+    J = I;
+    ++I;
+  }
+
+  SmallVector<MachineBasicBlock *, 4> EndPgmBlocks;
+
+  bool HaveScalarStores = false;
+
+  for (MachineFunction::iterator BI = MF.begin(), BE = MF.end(); BI != BE;
+       ++BI) {
+
+    MachineBasicBlock &MBB = *BI;
+
+    for (MachineBasicBlock::iterator I = MBB.begin(), E = MBB.end(); I != E;
+         ++I) {
+
+      if (!HaveScalarStores && TII->isScalarStore(*I))
+        HaveScalarStores = true;
+
+      if (I->getOpcode() == AMDGPU::S_ENDPGM ||
+          I->getOpcode() == AMDGPU::SI_RETURN_TO_EPILOG)
+        EndPgmBlocks.push_back(&MBB);
+    }
+  }
+
+  if (HaveScalarStores) {
+    // If scalar writes are used, the cache must be flushed or else the next
+    // wave to reuse the same scratch memory can be clobbered.
+    //
+    // Insert s_dcache_wb at wave termination points if there were any scalar
+    // stores, and only if the cache hasn't already been flushed. This could be
+    // improved by looking across blocks for flushes in postdominating blocks
+    // from the stores but an explicitly requested flush is probably very rare.
+    for (MachineBasicBlock *MBB : EndPgmBlocks) {
+      bool SeenDCacheWB = false;
+
+      for (MachineBasicBlock::iterator I = MBB->begin(), E = MBB->end(); I != E;
+           ++I) {
+
+        if (I->getOpcode() == AMDGPU::S_DCACHE_WB)
+          SeenDCacheWB = true;
+        else if (TII->isScalarStore(*I))
+          SeenDCacheWB = false;
+
+        // FIXME: It would be better to insert this before a waitcnt if any.
+        if ((I->getOpcode() == AMDGPU::S_ENDPGM ||
+             I->getOpcode() == AMDGPU::SI_RETURN_TO_EPILOG) &&
+            !SeenDCacheWB) {
+          Modified = true;
+          BuildMI(*MBB, I, I->getDebugLoc(), TII->get(AMDGPU::S_DCACHE_WB));
+        }
+      }
+    }
+  }
+
+  return Modified;
+}

test/CodeGen/AMDGPU/basic-branch.ll

 
 ; GCN-LABEL: {{^}}test_brcc_i1:
 ; GCN: buffer_load_ubyte
 ; GCN: v_and_b32_e32 v{{[0-9]+}}, 1,
 ; GCN: v_cmp_eq_u32_e32 vcc,
 ; GCN: s_cbranch_vccnz [[END:BB[0-9]+_[0-9]+]]
 
 ; GCN: buffer_store_dword
-; GCNOPT-NEXT: s_waitcnt vmcnt(0) expcnt(0)
-; TODO: This waitcnt can be eliminated
 
 ; GCN: {{^}}[[END]]:
 ; GCN: s_endpgm
 define amdgpu_kernel void @test_brcc_i1(i32 addrspace(1)* noalias %out, i32 addrspace(1)* noalias %in, i1 %val) #0 {
   %cmp0 = icmp ne i1 %val, 0
   br i1 %cmp0, label %store, label %end
 
 store:
   store i32 222, i32 addrspace(1)* %out
   ret void
 
 end:
   ret void
 }
 
 attributes #0 = { nounwind }

test/CodeGen/AMDGPU/branch-condition-and.ll

 ; GCN-DAG: v_cmp_lt_f32_e32 vcc,
 ; GCN: s_and_b64 [[AND:s\[[0-9]+:[0-9]+\]]], vcc, [[OTHERCC]]
 ; GCN: s_and_saveexec_b64 [[SAVED:s\[[0-9]+:[0-9]+\]]], [[AND]]
 ; GCN: s_xor_b64 {{s\[[0-9]+:[0-9]+\]}}, exec, [[SAVED]]
 ; GCN: ; mask branch [[BB5:BB[0-9]+_[0-9]+]]
 
 ; GCN-NEXT: BB{{[0-9]+_[0-9]+}}: ; %bb4
 ; GCN: ds_write_b32
-; GCN: s_waitcnt
 
-; GCN-NEXT: [[BB5]]
+; GCN: [[BB5]]
 ; GCN: s_or_b64 exec, exec
 ; GCN-NEXT: s_endpgm
 ; GCN-NEXT: .Lfunc_end
 define amdgpu_ps void @ham(float %arg, float %arg1) #0 {
 bb:
   %tmp = fcmp ogt float %arg, 0.000000e+00
   %tmp2 = fcmp ogt float %arg1, 0.000000e+00
   %tmp3 = and i1 %tmp, %tmp2

test/CodeGen/AMDGPU/branch-relaxation.ll

 ; GCN-NEXT: s_getpc_b64 vcc
 ; GCN-NEXT: s_add_u32 vcc_lo, vcc_lo, [[BB3:BB[0-9]_[0-9]+]]-([[LONG_JUMP0]]+4)
 ; GCN-NEXT: s_addc_u32 vcc_hi, vcc_hi, 0{{$}}
 ; GCN-NEXT: s_setpc_b64 vcc
 
 ; GCN-NEXT: [[BB2]]: ; %bb2
 ; GCN: v_mov_b32_e32 [[BB2_K:v[0-9]+]], 17
 ; GCN: buffer_store_dword [[BB2_K]]
-; GCN: s_waitcnt vmcnt(0)
 
 ; GCN-NEXT: [[LONG_JUMP1:BB[0-9]+_[0-9]+]]: ; %bb2
 ; GCN-NEXT: s_getpc_b64 vcc
 ; GCN-NEXT: s_add_u32 vcc_lo, vcc_lo, [[BB4:BB[0-9]_[0-9]+]]-([[LONG_JUMP1]]+4)
 ; GCN-NEXT: s_addc_u32 vcc_hi, vcc_hi, 0{{$}}
 ; GCN-NEXT: s_setpc_b64 vcc
 
 ; GCN: [[BB3]]: ; %bb3

 
 ; GCN-NEXT: [[IF]]: ; %if
 ; GCN: buffer_store_dword
 ; GCN: s_cmp_lg_u32
 ; GCN: s_cbranch_scc1 [[ENDIF]]
 
 ; GCN-NEXT: ; BB#2: ; %if_uniform
 ; GCN: buffer_store_dword
-; GCN: s_waitcnt vmcnt(0)
 
 ; GCN-NEXT: [[ENDIF]]: ; %endif
 ; GCN-NEXT: s_or_b64 exec, exec, [[MASK]]
 ; GCN-NEXT: s_sleep 5
 ; GCN-NEXT: s_endpgm
 define amdgpu_kernel void @uniform_inside_divergent(i32 addrspace(1)* %out, i32 %cond) #0 {
 entry:
   %tid = call i32 @llvm.amdgcn.workitem.id.x()

test/CodeGen/AMDGPU/control-flow-fastregalloc.ll

 ; VMEM: v_mov_b32_e32 v[[V_SAVEEXEC_HI:[0-9]+]], s[[SAVEEXEC_HI]]
 ; VMEM: buffer_store_dword v[[V_SAVEEXEC_HI]], off, s[0:3], s7 offset:8 ; 4-byte Folded Spill
 
 ; Spill load
 ; GCN: buffer_store_dword [[LOAD0]], off, s[0:3], s7 offset:[[LOAD0_OFFSET:[0-9]+]] ; 4-byte Folded Spill
 
 ; GCN: s_mov_b64 exec, s{{\[}}[[ANDEXEC_LO]]:[[ANDEXEC_HI]]{{\]}}
 
-; GCN: s_waitcnt vmcnt(0) expcnt(0)
 ; GCN: mask branch [[ENDIF:BB[0-9]+_[0-9]+]]
 
 ; GCN: {{^}}BB{{[0-9]+}}_1: ; %if
 ; GCN: s_mov_b32 m0, -1
 ; GCN: ds_read_b32 [[LOAD1:v[0-9]+]]
+; GCN: s_waitcnt lgkmcnt(0)
 ; GCN: buffer_load_dword [[RELOAD_LOAD0:v[0-9]+]], off, s[0:3], s7 offset:[[LOAD0_OFFSET]] ; 4-byte Folded Reload
-; GCN: s_waitcnt vmcnt(0)
 
 ; Spill val register
 ; GCN: v_add_i32_e32 [[VAL:v[0-9]+]], vcc, [[LOAD1]], [[RELOAD_LOAD0]]
 ; GCN: buffer_store_dword [[VAL]], off, s[0:3], s7 offset:[[VAL_OFFSET:[0-9]+]] ; 4-byte Folded Spill
-; GCN: s_waitcnt vmcnt(0)
 
 ; VMEM: [[ENDIF]]:
 ; Reload and restore exec mask
+; VGPR: s_waitcnt lgkmcnt(0)
 ; VGPR: v_readlane_b32 s[[S_RELOAD_SAVEEXEC_LO:[0-9]+]], [[SPILL_VGPR]], [[SAVEEXEC_LO_LANE]]
 ; VGPR: v_readlane_b32 s[[S_RELOAD_SAVEEXEC_HI:[0-9]+]], [[SPILL_VGPR]], [[SAVEEXEC_HI_LANE]]
 
 
 
 ; VMEM: buffer_load_dword v[[V_RELOAD_SAVEEXEC_LO:[0-9]+]], off, s[0:3], s7 offset:4 ; 4-byte Folded Reload
 ; VMEM: s_waitcnt vmcnt(0)
 ; VMEM: v_readfirstlane_b32 s[[S_RELOAD_SAVEEXEC_LO:[0-9]+]], v[[V_RELOAD_SAVEEXEC_LO]]

 
 ; VMEM: v_mov_b32_e32 v[[V_SAVEEXEC_LO:[0-9]+]], s[[SAVEEXEC_LO]]
 ; VMEM: buffer_store_dword v[[V_SAVEEXEC_LO]], off, s[0:3], s7 offset:20 ; 4-byte Folded Spill
 ; VMEM: v_mov_b32_e32 v[[V_SAVEEXEC_HI:[0-9]+]], s[[SAVEEXEC_HI]]
 ; VMEM: buffer_store_dword v[[V_SAVEEXEC_HI]], off, s[0:3], s7 offset:24 ; 4-byte Folded Spill
 
 ; GCN: s_mov_b64 exec, s{{\[}}[[ANDEXEC_LO]]:[[ANDEXEC_HI]]{{\]}}
 
-; GCN: s_waitcnt vmcnt(0) expcnt(0)
 ; GCN-NEXT: ; mask branch [[END:BB[0-9]+_[0-9]+]]
 ; GCN-NEXT: s_cbranch_execz [[END]]
 
 
 ; GCN: [[LOOP:BB[0-9]+_[0-9]+]]:
 ; GCN: buffer_load_dword v[[VAL_LOOP_RELOAD:[0-9]+]], off, s[0:3], s7 offset:4 ; 4-byte Folded Reload
 ; GCN: v_subrev_i32_e32 [[VAL_LOOP:v[0-9]+]], vcc, v{{[0-9]+}}, v[[VAL_LOOP_RELOAD]]
 ; GCN: v_cmp_ne_u32_e32 vcc,
 ; GCN: s_and_b64 vcc, exec, vcc
 ; GCN: buffer_store_dword [[VAL_LOOP]], off, s[0:3], s7 offset:[[VAL_SUB_OFFSET:[0-9]+]] ; 4-byte Folded Spill
-; GCN: s_waitcnt vmcnt(0) expcnt(0)
 ; GCN-NEXT: s_cbranch_vccnz [[LOOP]]
 
 
 ; GCN: [[END]]:
 ; VGPR: v_readlane_b32 s[[S_RELOAD_SAVEEXEC_LO:[0-9]+]], [[SPILL_VGPR]], [[SAVEEXEC_LO_LANE]]
 ; VGPR: v_readlane_b32 s[[S_RELOAD_SAVEEXEC_HI:[0-9]+]], [[SPILL_VGPR]], [[SAVEEXEC_HI_LANE]]
 
 ; VMEM: buffer_load_dword v[[V_RELOAD_SAVEEXEC_LO:[0-9]+]], off, s[0:3], s7 offset:20 ; 4-byte Folded Reload

 ; VGPR: v_writelane_b32 [[SPILL_VGPR]], s[[SAVEEXEC_HI]], [[SAVEEXEC_HI_LANE:[0-9]+]]
 
 ; VMEM: v_mov_b32_e32 v[[V_SAVEEXEC_LO:[0-9]+]], s[[SAVEEXEC_LO]]
 ; VMEM: buffer_store_dword v[[V_SAVEEXEC_LO]], off, s[0:3], s7 offset:[[SAVEEXEC_LO_OFFSET:[0-9]+]] ; 4-byte Folded Spill
 ; VMEM: v_mov_b32_e32 v[[V_SAVEEXEC_HI:[0-9]+]], s[[SAVEEXEC_HI]]
 ; VMEM: buffer_store_dword v[[V_SAVEEXEC_HI]], off, s[0:3], s7 offset:[[SAVEEXEC_HI_OFFSET:[0-9]+]] ; 4-byte Folded Spill
 
 ; GCN: s_mov_b64 exec, [[CMP0]]
-; GCN: s_waitcnt vmcnt(0) expcnt(0)
 
 ; FIXME: It makes no sense to put this skip here
 ; GCN-NEXT: ; mask branch [[FLOW:BB[0-9]+_[0-9]+]]
 ; GCN: s_cbranch_execz [[FLOW]]
 ; GCN-NEXT: s_branch [[ELSE:BB[0-9]+_[0-9]+]]
 
 ; GCN: [[FLOW]]: ; %Flow
 ; VGPR: v_readlane_b32 s[[FLOW_S_RELOAD_SAVEEXEC_LO:[0-9]+]], [[SPILL_VGPR]], [[SAVEEXEC_LO_LANE]]

 
 ; VMEM: v_mov_b32_e32 v[[FLOW_V_SAVEEXEC_LO:[0-9]+]], s[[FLOW_S_RELOAD_SAVEEXEC_LO]]
 ; VMEM: buffer_store_dword v[[FLOW_V_SAVEEXEC_LO]], off, s[0:3], s7 offset:[[FLOW_SAVEEXEC_LO_OFFSET:[0-9]+]] ; 4-byte Folded Spill
 ; VMEM: v_mov_b32_e32 v[[FLOW_V_SAVEEXEC_HI:[0-9]+]], s[[FLOW_S_RELOAD_SAVEEXEC_HI]]
 ; VMEM: buffer_store_dword v[[FLOW_V_SAVEEXEC_HI]], off, s[0:3], s7 offset:[[FLOW_SAVEEXEC_HI_OFFSET:[0-9]+]] ; 4-byte Folded Spill
 
 ; GCN: buffer_store_dword [[FLOW_VAL]], off, s[0:3], s7 offset:[[RESULT_OFFSET:[0-9]+]] ; 4-byte Folded Spill
 ; GCN: s_xor_b64 exec, exec, s{{\[}}[[FLOW_S_RELOAD_SAVEEXEC_LO]]:[[FLOW_S_RELOAD_SAVEEXEC_HI]]{{\]}}
-; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0)
 ; GCN-NEXT: ; mask branch [[ENDIF:BB[0-9]+_[0-9]+]]
 ; GCN-NEXT: s_cbranch_execz [[ENDIF]]
 
 
 ; GCN: BB{{[0-9]+}}_2: ; %if
 ; GCN: ds_read_b32
 ; GCN: buffer_load_dword v[[LOAD0_RELOAD:[0-9]+]], off, s[0:3], s7 offset:4 ; 4-byte Folded Reload
 ; GCN: v_add_i32_e32 [[ADD:v[0-9]+]], vcc, v{{[0-9]+}}, v[[LOAD0_RELOAD]]
 ; GCN: buffer_store_dword [[ADD]], off, s[0:3], s7 offset:[[RESULT_OFFSET]] ; 4-byte Folded Spill
-; GCN: s_waitcnt vmcnt(0) expcnt(0)
 ; GCN-NEXT: s_branch [[ENDIF:BB[0-9]+_[0-9]+]]
 
 ; GCN: [[ELSE]]: ; %else
 ; GCN: buffer_load_dword v[[LOAD0_RELOAD:[0-9]+]], off, s[0:3], s7 offset:4 ; 4-byte Folded Reload
 ; GCN: v_subrev_i32_e32 [[SUB:v[0-9]+]], vcc, v{{[0-9]+}}, v[[LOAD0_RELOAD]]
 ; GCN: buffer_store_dword [[ADD]], off, s[0:3], s7 offset:[[FLOW_RESULT_OFFSET:[0-9]+]] ; 4-byte Folded Spill
-; GCN: s_waitcnt vmcnt(0) expcnt(0)
 ; GCN-NEXT: s_branch [[FLOW]]
 
 ; GCN: [[ENDIF]]:
 ; VGPR: v_readlane_b32 s[[S_RELOAD_SAVEEXEC_LO:[0-9]+]], [[SPILL_VGPR]], [[FLOW_SAVEEXEC_LO_LANE]]
 ; VGPR: v_readlane_b32 s[[S_RELOAD_SAVEEXEC_HI:[0-9]+]], [[SPILL_VGPR]], [[FLOW_SAVEEXEC_HI_LANE]]
 
 
 ; VMEM: buffer_load_dword v[[V_RELOAD_SAVEEXEC_LO:[0-9]+]], off, s[0:3], s7 offset:[[FLOW_SAVEEXEC_LO_OFFSET]] ; 4-byte Folded Reload

test/CodeGen/AMDGPU/indirect-addressing-si.ll

 }
 
 ; GCN-LABEL: {{^}}extract_neg_offset_vgpr:
 ; The offset depends on the register that holds the first element of the vector.
 
 ; FIXME: The waitcnt for the argument load can go after the loop
 ; IDXMODE: s_set_gpr_idx_on 0, src0
 ; GCN: s_mov_b64 s{{\[[0-9]+:[0-9]+\]}}, exec
-; GCN: s_waitcnt lgkmcnt(0)
-
+; GCN: [[LOOPBB:BB[0-9]+_[0-9]+]]:
 ; GCN: v_readfirstlane_b32 [[READLANE:s[0-9]+]], v{{[0-9]+}}
 
 ; MOVREL: s_add_i32 m0, [[READLANE]], 0xfffffe0
 ; MOVREL: s_and_saveexec_b64 vcc, vcc
 ; MOVREL: v_movrels_b32_e32 [[RESULT:v[0-9]+]], v1
 
 ; IDXMODE: s_addk_i32 [[ADD_IDX:s[0-9]+]], 0xfe00
 ; IDXMODE: s_set_gpr_idx_idx [[ADD_IDX]]

 ; The offset depends on the register that holds the first element of the vector.
 
 ; GCN-DAG: v_mov_b32_e32 [[VEC_ELT0:v[0-9]+]], 1{{$}}
 ; GCN-DAG: v_mov_b32_e32 [[VEC_ELT1:v[0-9]+]], 2{{$}}
 ; GCN-DAG: v_mov_b32_e32 [[VEC_ELT2:v[0-9]+]], 3{{$}}
 ; GCN-DAG: v_mov_b32_e32 [[VEC_ELT3:v[0-9]+]], 4{{$}}
 
 ; GCN: s_mov_b64 [[SAVEEXEC:s\[[0-9]+:[0-9]+\]]], exec
-; GCN: s_waitcnt lgkmcnt(0)
-
 ; GCN: [[LOOPBB:BB[0-9]+_[0-9]+]]:
 ; GCN: v_readfirstlane_b32 [[READLANE:s[0-9]+]]
 
 ; MOVREL: s_add_i32 m0, [[READLANE]], 0xfffffe00
 ; MOVREL: s_and_saveexec_b64 vcc, vcc
 ; MOVREL: v_movreld_b32_e32 [[VEC_ELT0]], 5
 
 ; IDXMODE: s_addk_i32 [[ADD_IDX:s[0-9]+]], 0xfe00{{$}}

 ; GCN-DAG: v_mov_b32_e32 [[VEC_ELT1:v[0-9]+]], 2{{$}}
 ; GCN-DAG: v_mov_b32_e32 [[VEC_ELT2:v[0-9]+]], 3{{$}}
 ; GCN-DAG: v_mov_b32_e32 [[VEC_ELT3:v[0-9]+]], 4{{$}}
 ; GCN-DAG: v_mov_b32_e32 [[VAL:v[0-9]+]], 0x1f4{{$}}
 
 ; IDXMODE: s_set_gpr_idx_on 0, dst
 
 ; GCN: s_mov_b64 [[SAVEEXEC:s\[[0-9]+:[0-9]+\]]], exec
-; GCN: s_waitcnt lgkmcnt(0)
 
 ; The offset depends on the register that holds the first element of the vector.
 ; GCN: v_readfirstlane_b32 [[READLANE:s[0-9]+]]
 
 ; MOVREL: s_add_i32 m0, [[READLANE]], -16
 ; MOVREL: v_movreld_b32_e32 [[VEC_ELT0]], [[VAL]]
 
 ; IDXMODE: s_add_i32 [[ADD_IDX:s[0-9]+]], [[READLANE]], -16

 ; GCN-DAG: s_mov_b32 [[S_ELT1:s[0-9]+]], 9
 ; GCN-DAG: s_mov_b32 [[S_ELT0:s[0-9]+]], 7
 ; GCN-DAG: v_mov_b32_e32 [[VEC_ELT0:v[0-9]+]], [[S_ELT0]]
 ; GCN-DAG: v_mov_b32_e32 [[VEC_ELT1:v[0-9]+]], [[S_ELT1]]
 
 ; IDXMODE: s_set_gpr_idx_on 0, src0
 
 ; GCN: s_mov_b64 [[MASK:s\[[0-9]+:[0-9]+\]]], exec
-; GCN: s_waitcnt vmcnt(0)
 
 ; GCN: [[LOOP0:BB[0-9]+_[0-9]+]]:
+; GCN-NEXT: s_waitcnt vmcnt(0)
 ; GCN-NEXT: v_readfirstlane_b32 [[READLANE:s[0-9]+]], [[IDX0]]
 ; GCN: v_cmp_eq_u32_e32 vcc, [[READLANE]], [[IDX0]]
 
 ; MOVREL: s_mov_b32 m0, [[READLANE]]
 ; MOVREL: s_and_saveexec_b64 vcc, vcc
 ; MOVREL: v_movrels_b32_e32 [[MOVREL0:v[0-9]+]], [[VEC_ELT0]]
 
 ; IDXMODE: s_set_gpr_idx_idx [[READLANE]]

 ; GCN-DAG: v_mov_b32_e32 v[[VEC_ELT3:[0-9]+]], s[[S_ELT3]]
 ; GCN: v_mov_b32_e32 v[[VEC_ELT2:[0-9]+]], s{{[0-9]+}}
 ; GCN: v_mov_b32_e32 v[[VEC_ELT1:[0-9]+]], s{{[0-9]+}}
 ; GCN: v_mov_b32_e32 v[[VEC_ELT0:[0-9]+]], s[[S_ELT0]]
 
 ; IDXMODE: s_set_gpr_idx_on 0, dst
 
 ; GCN: [[LOOP0:BB[0-9]+_[0-9]+]]:
+; GCN-NEXT: s_waitcnt vmcnt(0)
 ; GCN-NEXT: v_readfirstlane_b32 [[READLANE:s[0-9]+]], [[IDX0]]
 ; GCN: v_cmp_eq_u32_e32 vcc, [[READLANE]], [[IDX0]]
 
 ; MOVREL: s_mov_b32 m0, [[READLANE]]
 ; MOVREL: s_and_saveexec_b64 vcc, vcc
 ; MOVREL-NEXT: v_movreld_b32_e32 v[[VEC_ELT0]], [[INS0]]
 
 ; IDXMODE: s_set_gpr_idx_idx [[READLANE]]

test/CodeGen/AMDGPU/infinite-loop.ll

 ; RUN: llc -march=amdgcn -verify-machineinstrs < %s | FileCheck -check-prefix=SI %s
 ; RUN: llc -march=amdgcn -mcpu=tonga -mattr=-flat-for-global -verify-machineinstrs < %s | FileCheck -check-prefix=SI %s
 
 ; SI-LABEL: {{^}}infinite_loop:
 ; SI: v_mov_b32_e32 [[REG:v[0-9]+]], 0x3e7
 ; SI: BB0_1:
+; SI: s_waitcnt lgkmcnt(0)
 ; SI: buffer_store_dword [[REG]]
-; SI: s_waitcnt vmcnt(0) expcnt(0)
 ; SI: s_branch BB0_1
 define amdgpu_kernel void @infinite_loop(i32 addrspace(1)* %out) {
 entry:
   br label %for.body
 
 for.body:                                         ; preds = %entry, %for.body
   store i32 999, i32 addrspace(1)* %out, align 4
   br label %for.body
 }
 

test/CodeGen/AMDGPU/llvm.amdgcn.buffer.store.format.ll

main_body:
   call void @llvm.amdgcn.buffer.store.format.v4f32(<4 x float> %1, <4 x i32> %0, i32 %3, i32 %2, i1 0, i1 0)
   ret void
 }
 
 ; Ideally, the register allocator would avoid the wait here
 ;
 ;CHECK-LABEL: {{^}}buffer_store_wait:
 ;CHECK: buffer_store_format_xyzw v[0:3], v4, s[0:3], 0 idxen
-;CHECK: s_waitcnt vmcnt(0) expcnt(0)
+;CHECK: s_waitcnt expcnt(0)
 ;CHECK: buffer_load_format_xyzw v[0:3], v5, s[0:3], 0 idxen
 ;CHECK: s_waitcnt vmcnt(0)
 ;CHECK: buffer_store_format_xyzw v[0:3], v6, s[0:3], 0 idxen
 define amdgpu_ps void @buffer_store_wait(<4 x i32> inreg, <4 x float>, i32, i32, i32) {
 main_body:
   call void @llvm.amdgcn.buffer.store.format.v4f32(<4 x float> %1, <4 x i32> %0, i32 %2, i32 0, i1 0, i1 0)
   %data = call <4 x float> @llvm.amdgcn.buffer.load.format.v4f32(<4 x i32> %0, i32 %3, i32 0, i1 0, i1 0)
   call void @llvm.amdgcn.buffer.store.format.v4f32(<4 x float> %data, <4 x i32> %0, i32 %4, i32 0, i1 0, i1 0)

test/CodeGen/AMDGPU/llvm.amdgcn.buffer.store.ll

main_body:
   call void @llvm.amdgcn.buffer.store.v4f32(<4 x float> %1, <4 x i32> %0, i32 %3, i32 %2, i1 0, i1 0)
   ret void
 }
 
 ; Ideally, the register allocator would avoid the wait here
 ;
 ;CHECK-LABEL: {{^}}buffer_store_wait:
 ;CHECK: buffer_store_dwordx4 v[0:3], v4, s[0:3], 0 idxen
-;CHECK: s_waitcnt vmcnt(0) expcnt(0)
+;CHECK: s_waitcnt expcnt(0)
 ;CHECK: buffer_load_dwordx4 v[0:3], v5, s[0:3], 0 idxen
 ;CHECK: s_waitcnt vmcnt(0)
 ;CHECK: buffer_store_dwordx4 v[0:3], v6, s[0:3], 0 idxen
 define amdgpu_ps void @buffer_store_wait(<4 x i32> inreg, <4 x float>, i32, i32, i32) {
 main_body:
   call void @llvm.amdgcn.buffer.store.v4f32(<4 x float> %1, <4 x i32> %0, i32 %2, i32 0, i1 0, i1 0)
   %data = call <4 x float> @llvm.amdgcn.buffer.load.v4f32(<4 x i32> %0, i32 %3, i32 0, i1 0, i1 0)
   call void @llvm.amdgcn.buffer.store.v4f32(<4 x float> %data, <4 x i32> %0, i32 %4, i32 0, i1 0, i1 0)

test/CodeGen/AMDGPU/llvm.amdgcn.ds.bpermute.ll

 ; RUN: llc -mtriple=amdgcn--amdhsa -mcpu=fiji -verify-machineinstrs < %s | FileCheck %s
 
 declare i32 @llvm.amdgcn.ds.bpermute(i32, i32) #0
 
 ; FUNC-LABEL: {{^}}ds_bpermute:
 ; CHECK: ds_bpermute_b32 v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+}}
-; CHECK: s_waitcnt lgkmcnt
 define amdgpu_kernel void @ds_bpermute(i32 addrspace(1)* %out, i32 %index, i32 %src) nounwind {
   %bpermute = call i32 @llvm.amdgcn.ds.bpermute(i32 %index, i32 %src) #0
   store i32 %bpermute, i32 addrspace(1)* %out, align 4
   ret void
 }
 
 ; CHECK-LABEL: {{^}}ds_bpermute_imm_offset:
 ; CHECK: ds_bpermute_b32 v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+}} offset:4
-; CHECK: s_waitcnt lgkmcnt
 define amdgpu_kernel void @ds_bpermute_imm_offset(i32 addrspace(1)* %out, i32 %base_index, i32 %src) nounwind {
   %index = add i32 %base_index, 4
   %bpermute = call i32 @llvm.amdgcn.ds.bpermute(i32 %index, i32 %src) #0
   store i32 %bpermute, i32 addrspace(1)* %out, align 4
   ret void
 }
 
 ; CHECK-LABEL: {{^}}ds_bpermute_imm_index:
 ; CHECK: ds_bpermute_b32 v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+}} offset:64
-; CHECK: s_waitcnt lgkmcnt
 define amdgpu_kernel void @ds_bpermute_imm_index(i32 addrspace(1)* %out, i32 %base_index, i32 %src) nounwind {
   %bpermute = call i32 @llvm.amdgcn.ds.bpermute(i32 64, i32 %src) #0
   store i32 %bpermute, i32 addrspace(1)* %out, align 4
   ret void
 }
 
 attributes #0 = { nounwind readnone convergent }

test/CodeGen/AMDGPU/llvm.amdgcn.ds.permute.ll

 ; RUN: llc -mtriple=amdgcn--amdhsa -mcpu=fiji -verify-machineinstrs < %s | FileCheck %s
 
 declare i32 @llvm.amdgcn.ds.permute(i32, i32) #0
 
 ; CHECK-LABEL: {{^}}ds_permute:
 ; CHECK: ds_permute_b32 v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+}}
-; CHECK: s_waitcnt lgkmcnt
 define amdgpu_kernel void @ds_permute(i32 addrspace(1)* %out, i32 %index, i32 %src) nounwind {
   %bpermute = call i32 @llvm.amdgcn.ds.permute(i32 %index, i32 %src) #0
   store i32 %bpermute, i32 addrspace(1)* %out, align 4
   ret void
 }
 
 ; CHECK-LABEL: {{^}}ds_permute_imm_offset:
 ; CHECK: ds_permute_b32 v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+}} offset:4
-; CHECK: s_waitcnt lgkmcnt
 define amdgpu_kernel void @ds_permute_imm_offset(i32 addrspace(1)* %out, i32 %base_index, i32 %src) nounwind {
   %index = add i32 %base_index, 4
   %bpermute = call i32 @llvm.amdgcn.ds.permute(i32 %index, i32 %src) #0
   store i32 %bpermute, i32 addrspace(1)* %out, align 4
   ret void
 }
 
 attributes #0 = { nounwind readnone convergent }

test/CodeGen/AMDGPU/llvm.amdgcn.ds.swizzle.ll

 ; RUN: llc -mtriple=amdgcn--amdhsa -mcpu=hawaii -verify-machineinstrs < %s | FileCheck %s
 ; RUN: llc -mtriple=amdgcn--amdhsa -mcpu=fiji -verify-machineinstrs < %s | FileCheck %s
 
 declare i32 @llvm.amdgcn.ds.swizzle(i32, i32) #0
 
 ; FUNC-LABEL: {{^}}ds_swizzle:
 ; CHECK: ds_swizzle_b32 v{{[0-9]+}}, v{{[0-9]+}} offset:100
-; CHECK: s_waitcnt lgkmcnt
 define amdgpu_kernel void @ds_swizzle(i32 addrspace(1)* %out, i32 %src) nounwind {
   %swizzle = call i32 @llvm.amdgcn.ds.swizzle(i32 %src, i32 100) #0
   store i32 %swizzle, i32 addrspace(1)* %out, align 4
   ret void
 }
 
 attributes #0 = { nounwind readnone convergent }

test/CodeGen/AMDGPU/llvm.amdgcn.image.ll

main_body:
   call void @llvm.amdgcn.exp.f32(i32 0, i32 15, float %r0, float %r1, float %r2, float %r3, i1 true, i1 true) #0
   ret void
 }
 
 ; Ideally, the register allocator would avoid the wait here
 ;
 ; GCN-LABEL: {{^}}image_store_wait:
 ; GCN: image_store v[0:3], v4, s[0:7] dmask:0xf unorm
-; GCN: s_waitcnt vmcnt(0) expcnt(0)
+; GCN: s_waitcnt expcnt(0)
 ; GCN: image_load v[0:3], v4, s[8:15] dmask:0xf unorm
 ; GCN: s_waitcnt vmcnt(0)
 ; GCN: image_store v[0:3], v4, s[16:23] dmask:0xf unorm
 define amdgpu_ps void @image_store_wait(<8 x i32> inreg %arg, <8 x i32> inreg %arg1, <8 x i32> inreg %arg2, <4 x float> %arg3, i32 %arg4) #0 {
 main_body:
   call void @llvm.amdgcn.image.store.v4f32.i32.v8i32(<4 x float> %arg3, i32 %arg4, <8 x i32> %arg, i32 15, i1 false, i1 false, i1 false, i1 false)
   %data = call <4 x float> @llvm.amdgcn.image.load.v4f32.i32.v8i32(i32 %arg4, <8 x i32> %arg1, i32 15, i1 false, i1 false, i1 false, i1 false)
   call void @llvm.amdgcn.image.store.v4f32.i32.v8i32(<4 x float> %data, i32 %arg4, <8 x i32> %arg2, i32 15, i1 false, i1 false, i1 false, i1 false)

test/CodeGen/AMDGPU/llvm.amdgcn.s.dcache.inv.ll

 }
 
 ; GCN-LABEL: {{^}}test_s_dcache_inv_insert_wait:
 ; GCN-NEXT: ; BB#0:
 ; GCN: s_dcache_inv
 ; GCN: s_waitcnt lgkmcnt(0) ; encoding
 define amdgpu_kernel void @test_s_dcache_inv_insert_wait() #0 {
   call void @llvm.amdgcn.s.dcache.inv()
-  call void @llvm.amdgcn.s.waitcnt(i32 0)
+  call void @llvm.amdgcn.s.waitcnt(i32 127)
   br label %end
 
 end:
   store volatile i32 3, i32 addrspace(1)* undef
   ret void
 }
 
 attributes #0 = { nounwind }

test/CodeGen/AMDGPU/llvm.amdgcn.s.dcache.inv.vol.ll

 }
 
 ; GCN-LABEL: {{^}}test_s_dcache_inv_vol_insert_wait:
 ; GCN-NEXT: ; BB#0:
 ; GCN-NEXT: s_dcache_inv_vol
 ; GCN: s_waitcnt lgkmcnt(0) ; encoding
 define amdgpu_kernel void @test_s_dcache_inv_vol_insert_wait() #0 {
   call void @llvm.amdgcn.s.dcache.inv.vol()
-  call void @llvm.amdgcn.s.waitcnt(i32 0)
+  call void @llvm.amdgcn.s.waitcnt(i32 127)
   br label %end
 
 end:
   store volatile i32 3, i32 addrspace(1)* undef
   ret void
 }
 
 attributes #0 = { nounwind }

test/CodeGen/AMDGPU/llvm.amdgcn.s.dcache.wb.ll

 }
 
 ; VI-LABEL: {{^}}test_s_dcache_wb_insert_wait:
 ; VI-NEXT: ; BB#0:
 ; VI-NEXT: s_dcache_wb
 ; VI: s_waitcnt lgkmcnt(0) ; encoding
 define amdgpu_kernel void @test_s_dcache_wb_insert_wait() #0 {
   call void @llvm.amdgcn.s.dcache.wb()
-  call void @llvm.amdgcn.s.waitcnt(i32 0)
+  call void @llvm.amdgcn.s.waitcnt(i32 127)
   br label %end
 
 end:
   store volatile i32 3, i32 addrspace(1)* undef
   ret void
 }
 
 attributes #0 = { nounwind }

test/CodeGen/AMDGPU/llvm.amdgcn.s.dcache.wb.vol.ll

 }
 
 ; VI-LABEL: {{^}}test_s_dcache_wb_vol_insert_wait:
 ; VI-NEXT: ; BB#0:
 ; VI-NEXT: s_dcache_wb_vol
 ; VI: s_waitcnt lgkmcnt(0) ; encoding
 define amdgpu_kernel void @test_s_dcache_wb_vol_insert_wait() #0 {
   call void @llvm.amdgcn.s.dcache.wb.vol()
-  call void @llvm.amdgcn.s.waitcnt(i32 0)
+  call void @llvm.amdgcn.s.waitcnt(i32 127)
   br label %end
 
 end:
   store volatile i32 3, i32 addrspace(1)* undef
   ret void
 }
 
 attributes #0 = { nounwind }

test/CodeGen/AMDGPU/llvm.amdgcn.s.waitcnt.ll

define amdgpu_ps void @test1(<8 x i32> inreg %rsrc, <4 x float> %d0, <4 x float> %d1, i32 %c0, i32 %c1) {
   ret void
 }
 
 ; Test that the intrinsic is merged with automatically generated waits and
 ; emitted as late as possible.
 ;
 ; CHECK-LABEL: {{^}}test2:
 ; CHECK: image_load
-; CHECK-NOT: s_waitcnt vmcnt(0){{$}}
-; CHECK: s_waitcnt
+; CHECK-NEXT: s_waitcnt
+; CHECK: s_waitcnt vmcnt(0){{$}}
 ; CHECK-NEXT: image_store
 define amdgpu_ps void @test2(<8 x i32> inreg %rsrc, i32 %c) {
   %t = call <4 x float> @llvm.amdgcn.image.load.v4f32.i32.v8i32(i32 %c, <8 x i32> %rsrc, i32 15, i1 0, i1 0, i1 0, i1 0)
   call void @llvm.amdgcn.s.waitcnt(i32 3840) ; 0xf00
   %c.1 = mul i32 %c, 2
   call void @llvm.amdgcn.image.store.v4f32.i32.v8i32(<4 x float> %t, i32 %c.1, <8 x i32> %rsrc, i32 15, i1 0, i1 0, i1 0, i1 0)
   ret void
 }
 
 declare void @llvm.amdgcn.s.waitcnt(i32) #0
 
 declare <4 x float> @llvm.amdgcn.image.load.v4f32.i32.v8i32(i32, <8 x i32>, i32, i1, i1, i1, i1) #1
 declare void @llvm.amdgcn.image.store.v4f32.i32.v8i32(<4 x float>, i32, <8 x i32>, i32, i1, i1, i1, i1) #0
 
 attributes #0 = { nounwind }
 attributes #1 = { nounwind readonly }

test/CodeGen/AMDGPU/multi-divergent-exit-region.ll

 
 ; GCN: v_mov_b32_e32 v0, 2.0
 ; GCN: s_or_b64 exec, exec
 ; GCN: s_and_b64 exec, exec
 ; GCN: v_mov_b32_e32 v0, 1.0
 
 ; GCN: {{^BB[0-9]+_[0-9]+}}: ; %UnifiedReturnBlock
 ; GCN-NEXT: s_or_b64 exec, exec
+; GCN-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN-NEXT: ; return
 
 define amdgpu_ps float @uniform_branch_to_multi_divergent_region_exit_ret_ret_return_value(i32 inreg %sgpr, i32 %vgpr) #0 {
 entry:
   %uniform.cond = icmp slt i32 %sgpr, 2
   br i1 %uniform.cond, label %LeafBlock, label %LeafBlock1
 
 LeafBlock:                                        ; preds = %entry

test/CodeGen/AMDGPU/ret_jump.ll

 
 ; GCN: ; BB#{{[0-9]+}}: ; %else
 ; GCN: s_and_saveexec_b64 [[SAVE_EXEC:s\[[0-9]+:[0-9]+\]]], vcc
 ; GCN-NEXT: s_xor_b64 [[XOR_EXEC:s\[[0-9]+:[0-9]+\]]], exec, [[SAVE_EXEC]]
 ; GCN-NEXT: ; mask branch [[FLOW1:BB[0-9]+_[0-9]+]]
 
 ; GCN-NEXT:  ; %unreachable.bb
 ; GCN: ds_write_b32
-; GCN: s_waitcnt
 ; GCN: ; divergent unreachable
 
 ; GCN: ; %ret.bb
 ; GCN: store_dword
 
 ; GCN: ; %UnifiedReturnBlock
 ; GCN-NEXT: s_or_b64 exec, exec
+; GCN-NEXT: s_waitcnt 
 ; GCN-NEXT: ; return
 ; GCN-NEXT: .Lfunc_end
 define amdgpu_ps <{ i32, i32, i32, i32, i32, i32, i32, i32, i32, float, float, float, float, float, float, float, float, float, float, float, float, float, float }> @uniform_br_nontrivial_ret_divergent_br_nontrivial_unreachable([9 x <16 x i8>] addrspace(2)* byval %arg, [17 x <16 x i8>] addrspace(2)* byval %arg1, [17 x <8 x i32>] addrspace(2)* byval %arg2, i32 addrspace(2)* byval %arg3, float inreg %arg4, i32 inreg %arg5, <2 x i32> %arg6, <2 x i32> %arg7, <2 x i32> %arg8, <3 x i32> %arg9, <2 x i32> %arg10, <2 x i32> %arg11, <2 x i32> %arg12, float %arg13, float %arg14, float %arg15, float %arg16, float %arg17, i32 inreg %arg18, i32 %arg19, float %arg20, i32 %arg21) #0 {
 main_body:
   %i.i = extractelement <2 x i32> %arg7, i32 0
   %j.i = extractelement <2 x i32> %arg7, i32 1
   %i.f.i = bitcast i32 %i.i to float
   %j.f.i = bitcast i32 %j.i to float

test/CodeGen/AMDGPU/si-lower-control-flow-unreachable-block.ll

 ; RUN: llc -march=amdgcn -verify-machineinstrs < %s | FileCheck -check-prefix=GCN %s
 
 ; GCN-LABEL: {{^}}lower_control_flow_unreachable_terminator:
 ; GCN: v_cmp_eq_u32
 ; GCN: s_and_saveexec_b64
 ; GCN: s_xor_b64
 ; GCN: ; mask branch [[RET:BB[0-9]+_[0-9]+]]
 
 ; GCN-NEXT: BB{{[0-9]+_[0-9]+}}: ; %unreachable
 ; GCN: ds_write_b32
 ; GCN: ; divergent unreachable
-; GCN: s_waitcnt
 
 ; GCN-NEXT: [[RET]]: ; %UnifiedReturnBlock
 ; GCN-NEXT: s_or_b64 exec, exec
 ; GCN: s_endpgm
 
 define amdgpu_kernel void @lower_control_flow_unreachable_terminator() #0 {
 bb:
   %tmp15 = tail call i32 @llvm.amdgcn.workitem.id.y()

 ; GCN: v_cmp_ne_u32
 ; GCN: s_and_saveexec_b64
 ; GCN: s_xor_b64
 ; GCN: ; mask branch [[RETURN:BB[0-9]+_[0-9]+]]
 
 ; GCN-NEXT: {{^BB[0-9]+_[0-9]+}}: ; %unreachable
 ; GCN: ds_write_b32
 ; GCN: ; divergent unreachable
-; GCN: s_waitcnt
 
 ; GCN: [[RETURN]]:
 ; GCN-NEXT: s_or_b64 exec, exec
 ; GCN-NEXT: s_endpgm
 define amdgpu_kernel void @lower_control_flow_unreachable_terminator_swap_block_order() #0 {
 bb:
   %tmp15 = tail call i32 @llvm.amdgcn.workitem.id.y()
   %tmp63 = icmp eq i32 %tmp15, 32

 ; GCN: s_cmp_lg_u32
 ; GCN: s_cbranch_scc0 [[UNREACHABLE:BB[0-9]+_[0-9]+]]
 
 ; GCN-NEXT: BB#{{[0-9]+}}: ; %ret
 ; GCN-NEXT: s_endpgm
 
 ; GCN: [[UNREACHABLE]]:
 ; GCN: ds_write_b32
-; GCN: s_waitcnt
 define amdgpu_kernel void @uniform_lower_control_flow_unreachable_terminator(i32 %arg0) #0 {
 bb:
   %tmp63 = icmp eq i32 %arg0, 32
   br i1 %tmp63, label %unreachable, label %ret
 
 unreachable:
   store volatile i32 0, i32 addrspace(3)* undef, align 4
   unreachable

test/CodeGen/AMDGPU/smrd-vccz-bug.ll

 ; RUN: llc -march=amdgcn -mcpu=verde -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=VCCZ-BUG %s
 ; RUN: llc -march=amdgcn -mcpu=bonaire -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=VCCZ-BUG %s
 ; RUN: llc -march=amdgcn -mcpu=tonga -mattr=-flat-for-global -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=NOVCCZ-BUG %s
 
 ; GCN-FUNC: {{^}}vccz_workaround:
 ; GCN: s_load_dword s{{[0-9]+}}, s[{{[0-9]+:[0-9]+}}], 0x0
 ; GCN: v_cmp_neq_f32_e64 vcc, s{{[0-9]+}}, 0{{$}}
-; GCN: s_waitcnt lgkmcnt(0)
+; VCCZ-BUG: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; VCCZ-BUG: s_mov_b64 vcc, vcc
 ; NOVCCZ-BUG-NOT: s_mov_b64 vcc, vcc
 ; GCN: s_cbranch_vccnz [[EXIT:[0-9A-Za-z_]+]]
 ; GCN: buffer_store_dword
 ; GCN: [[EXIT]]:
 ; GCN: s_endpgm
 define amdgpu_kernel void @vccz_workaround(i32 addrspace(2)* %in, i32 addrspace(1)* %out, float %cond) {
 entry:

test/CodeGen/AMDGPU/spill-m0.ll

 ; GCN-DAG: s_cmp_lg_u32
 
 ; TOVGPR-DAG: s_mov_b32 [[M0_COPY:s[0-9]+]], m0
 ; TOVGPR: v_writelane_b32 [[SPILL_VREG:v[0-9]+]], [[M0_COPY]], 0
 
 ; TOVMEM-DAG: s_mov_b32 [[M0_COPY:s[0-9]+]], m0
 ; TOVMEM-DAG: v_mov_b32_e32 [[SPILL_VREG:v[0-9]+]], [[M0_COPY]]
 ; TOVMEM: buffer_store_dword [[SPILL_VREG]], off, s{{\[[0-9]+:[0-9]+\]}}, s{{[0-9]+}} offset:4 ; 4-byte Folded Spill
-; TOVMEM: s_waitcnt vmcnt(0)
 
 ; TOSMEM-DAG: s_mov_b32 [[M0_COPY:s[0-9]+]], m0
 ; TOSMEM: s_add_u32 m0, s3, 0x100{{$}}
 ; TOSMEM-NOT: [[M0_COPY]]
 ; TOSMEM: s_buffer_store_dword [[M0_COPY]], s{{\[}}[[LO]]:[[HI]]], m0 ; 4-byte Folded Spill
-; TOSMEM: s_waitcnt lgkmcnt(0)
 
 ; GCN: s_cbranch_scc1 [[ENDIF:BB[0-9]+_[0-9]+]]
 
 ; GCN: [[ENDIF]]:
 ; TOVGPR: v_readlane_b32 [[M0_RESTORE:s[0-9]+]], [[SPILL_VREG]], 0
 ; TOVGPR: s_mov_b32 m0, [[M0_RESTORE]]
 
 ; TOVMEM: buffer_load_dword [[RELOAD_VREG:v[0-9]+]], off, s{{\[[0-9]+:[0-9]+\]}}, s{{[0-9]+}} offset:4 ; 4-byte Folded Reload

test/CodeGen/AMDGPU/valu-i1.ll

 ; RUN: llc -march=amdgcn -verify-machineinstrs -enable-misched -asm-verbose < %s | FileCheck -check-prefix=SI %s
 
 declare i32 @llvm.amdgcn.workitem.id.x() nounwind readnone
 
 ; SI-LABEL: {{^}}test_if:
 ; Make sure the i1 values created by the cfg structurizer pass are
 ; moved using VALU instructions
 
 
 ; waitcnt should be inserted after exec modification
 ; SI: v_cmp_lt_i32_e32 vcc, 0,
 ; SI-NEXT: s_and_saveexec_b64 [[SAVE1:s\[[0-9]+:[0-9]+\]]], vcc
 ; SI-NEXT: s_xor_b64 [[SAVE2:s\[[0-9]+:[0-9]+\]]], exec, [[SAVE1]]
-; SI-NEXT: s_waitcnt lgkmcnt(0)
 ; SI-NEXT: ; mask branch [[FLOW_BB:BB[0-9]+_[0-9]+]]
 ; SI-NEXT: s_cbranch_execz [[FLOW_BB]]
 
 ; SI-NEXT: BB{{[0-9]+}}_1: ; %LeafBlock3
 ; SI-NOT: s_mov_b64 s[{{[0-9]:[0-9]}}], -1
 ; SI: v_mov_b32_e32 v{{[0-9]}}, -1
 ; SI: s_and_saveexec_b64
 ; SI-NEXT: s_xor_b64

 ; SI-LABEL: {{^}}simple_test_v_if:
 ; SI: v_cmp_ne_u32_e32 vcc, 0, v{{[0-9]+}}
 ; SI: s_and_saveexec_b64 [[BR_SREG:s\[[0-9]+:[0-9]+\]]], vcc
 ; SI: s_xor_b64 [[BR_SREG]], exec, [[BR_SREG]]
 ; SI: ; mask branch [[EXIT:BB[0-9]+_[0-9]+]]
 
 ; SI-NEXT: BB{{[0-9]+_[0-9]+}}:
 ; SI: buffer_store_dword
-; SI-NEXT: s_waitcnt
 
 ; SI-NEXT: {{^}}[[EXIT]]:
 ; SI: s_or_b64 exec, exec, [[BR_SREG]]
 ; SI: s_endpgm
 define amdgpu_kernel void @simple_test_v_if(i32 addrspace(1)* %dst, i32 addrspace(1)* %src) #1 {
   %tid = call i32 @llvm.amdgcn.workitem.id.x() nounwind readnone
   %is.0 = icmp ne i32 %tid, 0
   br i1 %is.0, label %then, label %exit

 ; SI-LABEL: {{^}}simple_test_v_if_ret_else_ret:
 ; SI: v_cmp_ne_u32_e32 vcc, 0, v{{[0-9]+}}
 ; SI: s_and_saveexec_b64 [[BR_SREG:s\[[0-9]+:[0-9]+\]]], vcc
 ; SI: s_xor_b64 [[BR_SREG]], exec, [[BR_SREG]]
 ; SI: ; mask branch [[EXIT:BB[0-9]+_[0-9]+]]
 
 ; SI-NEXT: BB{{[0-9]+_[0-9]+}}:
 ; SI: buffer_store_dword
-; SI-NEXT: s_waitcnt
 
 ; SI-NEXT: {{^}}[[EXIT]]:
 ; SI: s_or_b64 exec, exec, [[BR_SREG]]
 ; SI: s_endpgm
 define amdgpu_kernel void @simple_test_v_if_ret_else_ret(i32 addrspace(1)* %dst, i32 addrspace(1)* %src) #1 {
   %tid = call i32 @llvm.amdgcn.workitem.id.x()
   %is.0 = icmp ne i32 %tid, 0
   br i1 %is.0, label %then, label %exit

 ; SI-LABEL: {{^}}simple_test_v_if_ret_else_code_ret:
 ; SI: v_cmp_eq_u32_e32 vcc, 0, v{{[0-9]+}}
 ; SI: s_and_saveexec_b64 [[BR_SREG:s\[[0-9]+:[0-9]+\]]], vcc
 ; SI: s_xor_b64 [[BR_SREG]], exec, [[BR_SREG]]
 ; SI: ; mask branch [[FLOW:BB[0-9]+_[0-9]+]]
 
 ; SI-NEXT: {{^BB[0-9]+_[0-9]+}}: ; %exit
 ; SI: ds_write_b32
-; SI: s_waitcnt
 
 ; SI-NEXT: {{^}}[[FLOW]]:
 ; SI-NEXT: s_or_saveexec_b64
 ; SI-NEXT: s_xor_b64 exec, exec
 ; SI-NEXT: ; mask branch [[UNIFIED_RETURN:BB[0-9]+_[0-9]+]]
 
 ; SI-NEXT: {{^BB[0-9]+_[0-9]+}}: ; %then
-; SI: buffer_store_dword
-; SI-NEXT: s_waitcnt
+; SI: s_waitcnt
+; SI-NEXT: buffer_store_dword
 
 ; SI-NEXT: {{^}}[[UNIFIED_RETURN]]: ; %UnifiedReturnBlock
 ; SI: s_or_b64 exec, exec
 ; SI: s_endpgm
 define amdgpu_kernel void @simple_test_v_if_ret_else_code_ret(i32 addrspace(1)* %dst, i32 addrspace(1)* %src) #1 {
   %tid = call i32 @llvm.amdgcn.workitem.id.x()
   %is.0 = icmp ne i32 %tid, 0
   br i1 %is.0, label %then, label %exit