Index: lib/Target/AMDGPU/AMDGPUSubtarget.h
===================================================================
--- lib/Target/AMDGPU/AMDGPUSubtarget.h
+++ lib/Target/AMDGPU/AMDGPUSubtarget.h
@@ -360,6 +360,10 @@
     return FlatAddressSpace;
   }
 
+  bool hasSoftClauses() const {
+    return getGeneration() >= AMDGPUSubtarget::VOLCANIC_ISLANDS;
+  }
+
   bool isMesaKernel(const MachineFunction &MF) const {
     return isMesa3DOS() && !AMDGPU::isShader(MF.getFunction()->getCallingConv());
   }
Index: lib/Target/AMDGPU/SIInsertWaits.cpp
===================================================================
--- lib/Target/AMDGPU/SIInsertWaits.cpp
+++ lib/Target/AMDGPU/SIInsertWaits.cpp
@@ -99,6 +99,9 @@
   /// \brief Registers defined by async instructions.
   RegCounters DefinedRegs;
 
+  /// Registers used as inputs in the current soft clause.
+  SmallSet<unsigned, 8> ClausePtrRegUnits;
+
   /// \brief Different export instruction types seen since last wait.
   unsigned ExpInstrTypesSeen = 0;
 
@@ -126,6 +129,13 @@
   RegInterval getRegInterval(const TargetRegisterClass *RC,
                              const MachineOperand &Reg) const;
 
+  bool inSoftClause() const {
+    return LastOpcodeType == VMEM;
+  }
+
+  void leaveSoftClause();
+  bool needToBreakSoftClause(MachineInstr &MI);
+
   /// \brief Handle instructions async components
   void pushInstruction(MachineBasicBlock &MBB,
                        MachineBasicBlock::iterator I,
@@ -299,6 +309,54 @@
   return Result;
 }
 
+// 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 SIInsertWaits::needToBreakSoftClause(MachineInstr &MI) {
+  // TODO: Handle SMEM clauses.
+
+  // TODO: A clause is broken if any of the counters will overflow. Do we need
+  // to handle this here?
+
+  // Don't need to break if this begins the clause.
+  if (inSoftClause()) {
+    for (MachineOperand &Dst : MI.defs()) {
+      for (MCRegUnitIterator R(Dst.getReg(), TRI); R.isValid(); ++R) {
+        if (ClausePtrRegUnits.count(*R)) {
+          // Overwriting a pointer def'd within the clause. Break the clause
+          // with a nop.
+          leaveSoftClause();
+          return true;
+        }
+      }
+    }
+  } else {
+    // Clear out state from previous clause.
+    ClausePtrRegUnits.clear();
+  }
+
+  MachineOperand *VAddr = TII->getNamedOperand(MI, AMDGPU::OpName::vaddr);
+  if (VAddr) {
+    // XXX - What about scalar loads?
+    for (MCRegUnitIterator R(VAddr->getReg(), TRI); R.isValid(); ++R) {
+      // XXX - What about same register for dest and pointer?
+      ClausePtrRegUnits.insert(*R);
+    }
+  }
+
+  return false;
+}
+
+void SIInsertWaits::leaveSoftClause() {
+  ClausePtrRegUnits.clear();
+}
+
 void SIInsertWaits::pushInstruction(MachineBasicBlock &MBB,
                                     MachineBasicBlock::iterator I,
                                     const Counters &Increment) {
@@ -322,29 +380,6 @@
     return;
   }
 
-  if (ST->getGeneration() >= SISubtarget::VOLCANIC_ISLANDS) {
-    // 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
-    if (LastOpcodeType == VMEM && Increment.Named.VM) {
-      // Insert a NOP to break the clause.
-      BuildMI(MBB, I, DebugLoc(), TII->get(AMDGPU::S_NOP))
-          .addImm(0);
-      LastInstWritesM0 = false;
-    }
-
-    if (TII->isSMRD(*I))
-      LastOpcodeType = SMEM;
-    else if (Increment.Named.VM)
-      LastOpcodeType = VMEM;
-  }
-
   // Remember which export instructions we have seen
   if (Increment.Named.EXP) {
     ExpInstrTypesSeen |= TII->isEXP(*I) ? 1 : 2;
@@ -368,6 +403,23 @@
         UsedRegs[j] = Limit;
     }
   }
+
+  if (ST->hasSoftClauses()) {
+    if (TII->isSMRD(*I))
+      LastOpcodeType = SMEM;
+    else if (Increment.Named.VM) {
+      if (needToBreakSoftClause(*I)) {
+        // Insert a NOP to break the clause.
+        BuildMI(MBB, I, DebugLoc(), TII->get(AMDGPU::S_NOP))
+          .addImm(0);
+        LastInstWritesM0 = false;
+      }
+
+      LastOpcodeType = VMEM;
+    } else {
+      LastOpcodeType = OTHER;
+    }
+  }
 }
 
 bool SIInsertWaits::insertWait(MachineBasicBlock &MBB,
Index: test/CodeGen/AMDGPU/break-vmem-soft-clauses.mir
===================================================================
--- /dev/null
+++ test/CodeGen/AMDGPU/break-vmem-soft-clauses.mir
@@ -0,0 +1,623 @@
+# RUN: llc -march=amdgcn -mcpu=fiji -verify-machineinstrs -run-pass si-insert-waits %s -o - | FileCheck -check-prefix=GCN -check-prefix=VI %s
+
+--- |
+  define void @trivial_clause_load_flat4_x1() {
+    ret void
+  }
+
+  define void @trivial_clause_load_flat4_x2() {
+    ret void
+  }
+
+  define void @trivial_clause_load_flat4_x3() {
+    ret void
+  }
+
+  define void @trivial_clause_load_flat4_x4() {
+    ret void
+  }
+
+  define void @trivial_clause_load_flat4_x2_sameptr() {
+    ret void
+  }
+
+  define void @flat_load4_overwrite_ptr_lo() {
+    ret void
+  }
+
+  define void @flat_load4_overwrite_ptr_hi() {
+    ret void
+  }
+
+  define void @flat_load8_overwrite_ptr() {
+    ret void
+  }
+
+  define void @break_clause_at_max_clause_size_flat_load4() {
+    ret void
+  }
+
+  define void @break_clause_simple_load_flat4_lo_ptr() {
+    ret void
+  }
+
+  define void @break_clause_simple_load_flat4_hi_ptr() {
+    ret void
+  }
+
+  define void @break_clause_simple_load_flat8_ptr() {
+    ret void
+  }
+
+  define void @break_clause_simple_load_flat16_ptr() {
+    ret void
+  }
+
+  define void @break_clause_block_boundary_load_flat8_ptr() {
+    ret void
+  }
+
+  define void @break_clause_store_load_into_ptr_flat4() {
+    ret void
+  }
+
+  define void @break_clause_store_load_into_data_flat4() {
+    ret void
+  }
+
+  define void @valu_inst_breaks_clause() {
+    ret void
+  }
+
+  define void @salu_inst_breaks_clause() {
+    ret void
+  }
+
+  define void @ds_inst_breaks_clause() {
+    ret void
+  }
+
+  define void @smrd_inst_breaks_clause() {
+    ret void
+  }
+
+  define void @implicit_use_breaks_clause() {
+    ret void
+  }
+
+  define void @trivial_clause_load_mubuf4_x2() {
+    ret void
+  }
+
+  define void @break_clause_simple_load_mubuf_offen_ptr() {
+    ret void
+  }
+
+  define void @mubuf_load4_overwrite_ptr() {
+    ret void
+  }
+
+  define void @break_clause_flat_load_mubuf_load() {
+    ret void
+  }
+
+  define void @break_clause_mubuf_load_flat_load() {
+    ret void
+  }
+
+  define void @break_clause_atomic_rtn_into_ptr_flat4() {
+    ret void
+  }
+
+  define void @break_clause_atomic_nortn_ptr_load_flat4() {
+    ret void
+  }
+
+  define void @break_clause_atomic_rtn_into_ptr_mubuf4() {
+    ret void
+  }
+
+  define void @break_clause_atomic_nortn_ptr_load_mubuf4() {
+    ret void
+  }
+
+  define void @no_break_clause_mubuf_load_novaddr() {
+    ret void
+  }
+
+...
+---
+# Trivial clause at beginning of program
+# GCN-LABEL: name: trivial_clause_load_flat4_x1
+# GCN: bb.0:
+# GCN-NEXT: %vgpr0 = FLAT_LOAD_DWORD %vgpr2_vgpr3
+name: trivial_clause_load_flat4_x1
+
+body: |
+  bb.0:
+    %vgpr0 = FLAT_LOAD_DWORD %vgpr2_vgpr3, 0, 0, 0, implicit %exec, implicit %flat_scr
+    S_ENDPGM
+...
+---
+# Trivial clause at beginning of program
+# GCN-LABEL: name: trivial_clause_load_flat4_x2
+# GCN: bb.0:
+# GCN-NEXT: %vgpr0 = FLAT_LOAD_DWORD %vgpr2_vgpr3
+# GCN-NEXT: %vgpr1 = FLAT_LOAD_DWORD %vgpr4_vgpr5
+name: trivial_clause_load_flat4_x2
+
+body: |
+  bb.0:
+    %vgpr0 = FLAT_LOAD_DWORD %vgpr2_vgpr3, 0, 0, 0, implicit %exec, implicit %flat_scr
+    %vgpr1 = FLAT_LOAD_DWORD %vgpr4_vgpr5, 0, 0, 0, implicit %exec, implicit %flat_scr
+    S_ENDPGM
+...
+---
+# Trivial clause at beginning of program
+# GCN-LABEL: name: trivial_clause_load_flat4_x3
+# GCN: %vgpr0 = FLAT_LOAD_DWORD %vgpr3_vgpr4, 0, 0, 0, implicit %exec, implicit %flat_scr
+# GCN: %vgpr1 = FLAT_LOAD_DWORD %vgpr5_vgpr6, 0, 0, 0, implicit %exec, implicit %flat_scr
+# GCN: %vgpr2 = FLAT_LOAD_DWORD %vgpr7_vgpr8, 0, 0, 0, implicit %exec, implicit %flat_scr
+# GCN: S_ENDPGM
+name: trivial_clause_load_flat4_x3
+
+body: |
+  bb.0:
+    %vgpr0 = FLAT_LOAD_DWORD %vgpr3_vgpr4, 0, 0, 0, implicit %exec, implicit %flat_scr
+    %vgpr1 = FLAT_LOAD_DWORD %vgpr5_vgpr6, 0, 0, 0, implicit %exec, implicit %flat_scr
+    %vgpr2 = FLAT_LOAD_DWORD %vgpr7_vgpr8, 0, 0, 0, implicit %exec, implicit %flat_scr
+    S_ENDPGM
+...
+---
+# Trivial clause at beginning of program
+# GCN-LABEL: name: trivial_clause_load_flat4_x4
+# GCN: %vgpr0 = FLAT_LOAD_DWORD %vgpr4_vgpr5, 0, 0, 0, implicit %exec, implicit %flat_scr
+# GCN: %vgpr1 = FLAT_LOAD_DWORD %vgpr6_vgpr7, 0, 0, 0, implicit %exec, implicit %flat_scr
+# GCN: %vgpr2 = FLAT_LOAD_DWORD %vgpr8_vgpr9, 0, 0, 0, implicit %exec, implicit %flat_scr
+# GCN: %vgpr3 = FLAT_LOAD_DWORD %vgpr10_vgpr11, 0, 0, 0, implicit %exec, implicit %flat_scr
+# GCN: S_ENDPGM
+name: trivial_clause_load_flat4_x4
+
+body: |
+  bb.0:
+    %vgpr0 = FLAT_LOAD_DWORD %vgpr4_vgpr5, 0, 0, 0, implicit %exec, implicit %flat_scr
+    %vgpr1 = FLAT_LOAD_DWORD %vgpr6_vgpr7, 0, 0, 0, implicit %exec, implicit %flat_scr
+    %vgpr2 = FLAT_LOAD_DWORD %vgpr8_vgpr9, 0, 0, 0, implicit %exec, implicit %flat_scr
+    %vgpr3 = FLAT_LOAD_DWORD %vgpr10_vgpr11, 0, 0, 0, implicit %exec, implicit %flat_scr
+    S_ENDPGM
+...
+---
+# Reuse of same input pointer is OK
+# GCN-LABEL: name: trivial_clause_load_flat4_x2_sameptr
+# GCN: %vgpr0 = FLAT_LOAD_DWORD %vgpr2_vgpr3
+# GCN-NEXT: %vgpr1 = FLAT_LOAD_DWORD %vgpr2_vgpr3
+# GCN-NEXT: S_ENDPGM
+
+name: trivial_clause_load_flat4_x2_sameptr
+body: |
+  bb.0:
+    %vgpr0 = FLAT_LOAD_DWORD %vgpr2_vgpr3, 0, 0, 0, implicit %exec, implicit %flat_scr
+    %vgpr1 = FLAT_LOAD_DWORD %vgpr2_vgpr3, 0, 0, 0, implicit %exec, implicit %flat_scr
+    S_ENDPGM
+...
+---
+# 32-bit load partially clobbers its own ptr reg
+# GCN-LABEL: name: flat_load4_overwrite_ptr_lo
+# GCN: %vgpr0 = FLAT_LOAD_DWORD %vgpr0_vgpr1
+name: flat_load4_overwrite_ptr_lo
+
+body: |
+  bb.0:
+    %vgpr0 = FLAT_LOAD_DWORD %vgpr0_vgpr1, 0, 0, 0, implicit %exec, implicit %flat_scr
+    S_ENDPGM
+...
+---
+# 32-bit load partially clobbers its own ptr reg
+# GCN-LABEL: name: flat_load4_overwrite_ptr_hi
+# GCN: %vgpr1 = FLAT_LOAD_DWORD %vgpr0_vgpr1
+name: flat_load4_overwrite_ptr_hi
+
+body: |
+  bb.0:
+    %vgpr1 = FLAT_LOAD_DWORD %vgpr0_vgpr1, 0, 0, 0, implicit %exec, implicit %flat_scr
+    S_ENDPGM
+...
+---
+# 64-bit load clobbers its own ptr reg
+# GCN-LABEL: name: flat_load8_overwrite_ptr
+# GCN: %vgpr2_vgpr3 = FLAT_LOAD_DWORDX2 %vgpr2_vgpr3
+name: flat_load8_overwrite_ptr
+
+body: |
+  bb.0:
+    %vgpr2_vgpr3 = FLAT_LOAD_DWORDX2 %vgpr2_vgpr3, 0, 0, 0, implicit %exec, implicit %flat_scr
+    S_ENDPGM
+...
+---
+# vmcnt has 4 bits, so maximum 16 outstanding loads. The waitcnt
+# breaks the clause.
+
+# GCN-LABEL: name: break_clause_at_max_clause_size_flat_load4
+# GCN: %vgpr17 = FLAT_LOAD_DWORD %vgpr0_vgpr1
+# GCN-NEXT: S_WAITCNT 112
+# GCN-NEXT: %vgpr0 = FLAT_LOAD_DWORD %vgpr2_vgpr3
+
+name: break_clause_at_max_clause_size_flat_load4
+
+body: |
+  bb.0:
+    %vgpr2 = FLAT_LOAD_DWORD %vgpr0_vgpr1, 0, 0, 0, implicit %exec, implicit %flat_scr
+    %vgpr3 = FLAT_LOAD_DWORD %vgpr0_vgpr1, 0, 0, 0, implicit %exec, implicit %flat_scr
+    %vgpr4 = FLAT_LOAD_DWORD %vgpr0_vgpr1, 0, 0, 0, implicit %exec, implicit %flat_scr
+    %vgpr5 = FLAT_LOAD_DWORD %vgpr0_vgpr1, 0, 0, 0, implicit %exec, implicit %flat_scr
+
+    %vgpr6 = FLAT_LOAD_DWORD %vgpr0_vgpr1, 0, 0, 0, implicit %exec, implicit %flat_scr
+    %vgpr7 = FLAT_LOAD_DWORD %vgpr0_vgpr1, 0, 0, 0, implicit %exec, implicit %flat_scr
+    %vgpr8 = FLAT_LOAD_DWORD %vgpr0_vgpr1, 0, 0, 0, implicit %exec, implicit %flat_scr
+    %vgpr9 = FLAT_LOAD_DWORD %vgpr0_vgpr1, 0, 0, 0, implicit %exec, implicit %flat_scr
+
+    %vgpr10 = FLAT_LOAD_DWORD %vgpr0_vgpr1, 0, 0, 0, implicit %exec, implicit %flat_scr
+    %vgpr11 = FLAT_LOAD_DWORD %vgpr0_vgpr1, 0, 0, 0, implicit %exec, implicit %flat_scr
+    %vgpr12 = FLAT_LOAD_DWORD %vgpr0_vgpr1, 0, 0, 0, implicit %exec, implicit %flat_scr
+    %vgpr13 = FLAT_LOAD_DWORD %vgpr0_vgpr1, 0, 0, 0, implicit %exec, implicit %flat_scr
+
+    %vgpr14 = FLAT_LOAD_DWORD %vgpr0_vgpr1, 0, 0, 0, implicit %exec, implicit %flat_scr
+    %vgpr15 = FLAT_LOAD_DWORD %vgpr0_vgpr1, 0, 0, 0, implicit %exec, implicit %flat_scr
+    %vgpr16 = FLAT_LOAD_DWORD %vgpr0_vgpr1, 0, 0, 0, implicit %exec, implicit %flat_scr
+    %vgpr17 = FLAT_LOAD_DWORD %vgpr0_vgpr1, 0, 0, 0, implicit %exec, implicit %flat_scr
+
+    %vgpr0 = FLAT_LOAD_DWORD %vgpr2_vgpr3, 0, 0, 0, implicit %exec, implicit %flat_scr
+    %sgpr0 = S_MOV_B32 %sgpr0, implicit %vgpr2, implicit %vgpr3, implicit %vgpr4, implicit %vgpr5, implicit %vgpr6, implicit %vgpr7, implicit %vgpr8, implicit %vgpr9, implicit %vgpr10, implicit %vgpr11, implicit %vgpr12, implicit %vgpr13, implicit %vgpr14, implicit %vgpr15, implicit %vgpr16, implicit %vgpr17, implicit %vgpr18
+    S_ENDPGM
+...
+---
+
+# GCN-LABEL: name: break_clause_simple_load_flat4_lo_ptr
+# GCN: %vgpr0 = FLAT_LOAD_DWORD %vgpr2_vgpr3
+# GCN-NEXT: S_NOP 0
+# GCN-NEXT: %vgpr2 = FLAT_LOAD_DWORD %vgpr4_vgpr5
+name: break_clause_simple_load_flat4_lo_ptr
+
+body: |
+  bb.0:
+    %vgpr0 = FLAT_LOAD_DWORD %vgpr2_vgpr3, 0, 0, 0, implicit %exec, implicit %flat_scr
+    %vgpr2 = FLAT_LOAD_DWORD %vgpr4_vgpr5, 0, 0, 0, implicit %exec, implicit %flat_scr
+    S_ENDPGM
+...
+---
+
+# GCN-LABEL: name: break_clause_simple_load_flat4_hi_ptr
+# GCN: %vgpr0 = FLAT_LOAD_DWORD %vgpr2_vgpr3
+# GCN-NEXT: S_NOP 0
+# GCN-NEXT: %vgpr3 = FLAT_LOAD_DWORD %vgpr4_vgpr5
+name: break_clause_simple_load_flat4_hi_ptr
+
+body: |
+  bb.0:
+    %vgpr0 = FLAT_LOAD_DWORD %vgpr2_vgpr3, 0, 0, 0, implicit %exec, implicit %flat_scr
+    %vgpr3 = FLAT_LOAD_DWORD %vgpr4_vgpr5, 0, 0, 0, implicit %exec, implicit %flat_scr
+    S_ENDPGM
+...
+---
+
+# GCN-LABEL: name: break_clause_simple_load_flat8_ptr
+# GCN: %vgpr0_vgpr1 = FLAT_LOAD_DWORDX2 %vgpr2_vgpr3, 0, 0, 0, implicit %exec, implicit %flat_scr
+# GCN-NEXT: S_NOP 0
+# GCN-NEXT: %vgpr2_vgpr3 = FLAT_LOAD_DWORDX2 %vgpr4_vgpr5, 0, 0, 0, implicit %exec, implicit %flat_scr
+name: break_clause_simple_load_flat8_ptr
+
+body: |
+  bb.0:
+    %vgpr0_vgpr1 = FLAT_LOAD_DWORDX2 %vgpr2_vgpr3, 0, 0, 0, implicit %exec, implicit %flat_scr
+    %vgpr2_vgpr3 = FLAT_LOAD_DWORDX2 %vgpr4_vgpr5, 0, 0, 0, implicit %exec, implicit %flat_scr
+    S_ENDPGM
+...
+---
+
+# GCN-LABEL: name: break_clause_simple_load_flat16_ptr
+# GCN: %vgpr0_vgpr1 = FLAT_LOAD_DWORDX2 %vgpr2_vgpr3
+# GCN-NEXT: S_NOP 0
+# GCN-NEXT: %vgpr2_vgpr3_vgpr4_vgpr5 = FLAT_LOAD_DWORDX4 %vgpr6_vgpr7
+
+name: break_clause_simple_load_flat16_ptr
+
+body: |
+  bb.0:
+    %vgpr0_vgpr1 = FLAT_LOAD_DWORDX2 %vgpr2_vgpr3, 0, 0, 0, implicit %exec, implicit %flat_scr
+    %vgpr2_vgpr3_vgpr4_vgpr5 = FLAT_LOAD_DWORDX4 %vgpr6_vgpr7, 0, 0, 0, implicit %exec, implicit %flat_scr
+    S_ENDPGM
+...
+---
+
+# The clause is broken by the waitcnt inserted at the end of the
+# block, so no nop is needed.
+# GCN-LABEL: name: break_clause_block_boundary_load_flat8_ptr
+# GCN: bb.0:
+# GCN-NEXT: %vgpr0_vgpr1 = FLAT_LOAD_DWORDX2 %vgpr2_vgpr3
+# GCN-NEXT: S_WAITCNT 112
+
+# GCN: bb.1:
+# GCN-NEXT: %vgpr2_vgpr3 = FLAT_LOAD_DWORDX2 %vgpr4_vgpr5
+# GCN-NEXT: S_ENDPGM
+
+name: break_clause_block_boundary_load_flat8_ptr
+
+body: |
+  bb.0:
+    %vgpr0_vgpr1 = FLAT_LOAD_DWORDX2 %vgpr2_vgpr3, 0, 0, 0, implicit %exec, implicit %flat_scr
+
+  bb.1:
+    %vgpr2_vgpr3 = FLAT_LOAD_DWORDX2 %vgpr4_vgpr5, 0, 0, 0, implicit %exec, implicit %flat_scr
+    S_ENDPGM
+...
+---
+# The load clobbers the pointer of the store, so it needs to break.
+
+# GCN-LABEL: name: break_clause_store_load_into_ptr_flat4
+# GCN: FLAT_STORE_DWORD %vgpr2_vgpr3, %vgpr0
+# GCN-NEXT: S_WAITCNT 112
+# GCN-NEXT: %vgpr2 = FLAT_LOAD_DWORD %vgpr4_vgpr5
+# GCN-NEXT: S_ENDPGM
+name: break_clause_store_load_into_ptr_flat4
+
+body: |
+  bb.0:
+    FLAT_STORE_DWORD %vgpr2_vgpr3, %vgpr0, 0, 0, 0, implicit %exec, implicit %flat_scr
+    %vgpr2 = FLAT_LOAD_DWORD %vgpr4_vgpr5, 0, 0, 0, implicit %exec, implicit %flat_scr
+    S_ENDPGM
+...
+---
+# The load clobbers the data of the store, so it needs to break.
+# FIXME: Would it be better to s_nop and wait later?
+
+# GCN-LABEL: name: break_clause_store_load_into_data_flat4
+# GCN: FLAT_STORE_DWORD %vgpr2_vgpr3, %vgpr0
+# GCN-NEXT: S_WAITCNT 112
+# GCN-NEXT: %vgpr0 = FLAT_LOAD_DWORD %vgpr4_vgpr5
+# GCN-NEXT: S_ENDPGM
+name: break_clause_store_load_into_data_flat4
+
+body: |
+  bb.0:
+    FLAT_STORE_DWORD %vgpr2_vgpr3, %vgpr0, 0, 0, 0, implicit %exec, implicit %flat_scr
+    %vgpr0 = FLAT_LOAD_DWORD %vgpr4_vgpr5, 0, 0, 0, implicit %exec, implicit %flat_scr
+    S_ENDPGM
+...
+---
+# Regular VALU instruction breaks clause, no nop needed
+
+# GCN-LABEL: name: valu_inst_breaks_clause
+# GCN: %vgpr0 = FLAT_LOAD_DWORD %vgpr2_vgpr3, 0, 0, 0, implicit %exec, implicit %flat_scr
+# GCN-NEXT: %vgpr8 = V_MOV_B32_e32 0, implicit %exec
+# GCN-NEXT: %vgpr2 = FLAT_LOAD_DWORD %vgpr4_vgpr5, 0, 0, 0, implicit %exec, implicit %flat_scr
+name: valu_inst_breaks_clause
+
+body: |
+  bb.0:
+    %vgpr0 = FLAT_LOAD_DWORD %vgpr2_vgpr3, 0, 0, 0, implicit %exec, implicit %flat_scr
+    %vgpr8 = V_MOV_B32_e32 0, implicit %exec
+    %vgpr2 = FLAT_LOAD_DWORD %vgpr4_vgpr5, 0, 0, 0, implicit %exec, implicit %flat_scr
+    S_ENDPGM
+...
+---
+# Regular SALU instruction breaks clause, no nop needed
+
+# GCN-LABEL: name: salu_inst_breaks_clause
+# GCN: bb.0:
+# GCN-NEXT: %vgpr0 = FLAT_LOAD_DWORD %vgpr2_vgpr3, 0, 0, 0, implicit %exec, implicit %flat_scr
+# GCN-NEXT: %sgpr8 = S_MOV_B32 0
+# GCN-NEXT: %vgpr2 = FLAT_LOAD_DWORD %vgpr4_vgpr5, 0, 0, 0, implicit %exec, implicit %flat_scr
+name: salu_inst_breaks_clause
+
+body: |
+  bb.0:
+    %vgpr0 = FLAT_LOAD_DWORD %vgpr2_vgpr3, 0, 0, 0, implicit %exec, implicit %flat_scr
+    %sgpr8 = S_MOV_B32 0
+    %vgpr2 = FLAT_LOAD_DWORD %vgpr4_vgpr5, 0, 0, 0, implicit %exec, implicit %flat_scr
+    S_ENDPGM
+...
+---
+
+# GCN-LABEL: name: ds_inst_breaks_clause
+# GCN: bb.0:
+# GCN-NEXT: %vgpr0 = FLAT_LOAD_DWORD %vgpr2_vgpr3, 0, 0, 0, implicit %exec, implicit %flat_scr
+# GCN-NEXT: %vgpr8 = DS_READ_B32 %vgpr9
+# GCN-NEXT: %vgpr2 = FLAT_LOAD_DWORD %vgpr4_vgpr5, 0, 0, 0, implicit %exec, implicit %flat_scr
+name: ds_inst_breaks_clause
+
+body: |
+  bb.0:
+    %vgpr0 = FLAT_LOAD_DWORD %vgpr2_vgpr3, 0, 0, 0, implicit %exec, implicit %flat_scr
+    %vgpr8 = DS_READ_B32 %vgpr9, 0, 0, implicit %m0, implicit %exec
+    %vgpr2 = FLAT_LOAD_DWORD %vgpr4_vgpr5, 0, 0, 0, implicit %exec, implicit %flat_scr
+    S_ENDPGM
+...
+---
+
+# GCN-LABEL: name: smrd_inst_breaks_clause
+# GCN: bb.0:
+# GCN-NEXT: %vgpr0 = FLAT_LOAD_DWORD %vgpr2_vgpr3, 0, 0, 0, implicit %exec, implicit %flat_scr
+# GCN-NEXT: %sgpr8 = S_LOAD_DWORD_IMM %sgpr0_sgpr1, 0, 0
+# GCN-NEXT: %vgpr2 = FLAT_LOAD_DWORD %vgpr4_vgpr5, 0, 0, 0, implicit %exec, implicit %flat_scr
+name: smrd_inst_breaks_clause
+
+body: |
+  bb.0:
+    %vgpr0 = FLAT_LOAD_DWORD %vgpr2_vgpr3, 0, 0, 0, implicit %exec, implicit %flat_scr
+    %sgpr8 = S_LOAD_DWORD_IMM %sgpr0_sgpr1, 0, 0
+    %vgpr2 = FLAT_LOAD_DWORD %vgpr4_vgpr5, 0, 0, 0, implicit %exec, implicit %flat_scr
+    S_ENDPGM
+...
+---
+# FIXME: Should this be handled?
+# GCN-LABEL: name: implicit_use_breaks_clause
+# GCN: %vgpr0_vgpr1 = FLAT_LOAD_DWORDX2 %vgpr2_vgpr3, 0, 0, 0, implicit %exec, implicit %flat_scr, implicit %vgpr4_vgpr5
+# GCN-NEXT: %vgpr4_vgpr5 = FLAT_LOAD_DWORDX2 %vgpr6_vgpr7, 0, 0, 0, implicit %exec, implicit %flat_scr
+name: implicit_use_breaks_clause
+
+body: |
+  bb.0:
+    %vgpr0_vgpr1 = FLAT_LOAD_DWORDX2 %vgpr2_vgpr3, 0, 0, 0, implicit %exec, implicit %flat_scr, implicit %vgpr4_vgpr5
+    %vgpr4_vgpr5 = FLAT_LOAD_DWORDX2 %vgpr6_vgpr7, 0, 0, 0, implicit %exec, implicit %flat_scr
+    S_ENDPGM
+...
+---
+# GCN-LABEL: name: trivial_clause_load_mubuf4_x2
+# GCN: bb.0:
+# GCN-NEXT: %vgpr1 = BUFFER_LOAD_DWORD_OFFEN %vgpr2, %sgpr0_sgpr1_sgpr2_sgpr3, %sgpr4
+# GCN-NEXT: %vgpr3 = BUFFER_LOAD_DWORD_OFFEN %vgpr4, %sgpr0_sgpr1_sgpr2_sgpr3, %sgpr4
+name: trivial_clause_load_mubuf4_x2
+
+body: |
+  bb.0:
+    %vgpr1 = BUFFER_LOAD_DWORD_OFFEN %vgpr2, %sgpr0_sgpr1_sgpr2_sgpr3, %sgpr4, 0, 0, 0, 0, implicit %exec
+    %vgpr3 = BUFFER_LOAD_DWORD_OFFEN %vgpr4, %sgpr0_sgpr1_sgpr2_sgpr3, %sgpr4, 0, 0, 0, 0, implicit %exec
+    S_ENDPGM
+...
+---
+# GCN-LABEL: name: break_clause_simple_load_mubuf_offen_ptr
+# GCN: bb.0:
+# GCN-NEXT: %vgpr1 = BUFFER_LOAD_DWORD_OFFEN %vgpr2, %sgpr0_sgpr1_sgpr2_sgpr3, %sgpr4
+# GCN-NEXT: S_NOP 0
+# GCN-NEXT: %vgpr2 = BUFFER_LOAD_DWORD_OFFEN %vgpr3, %sgpr0_sgpr1_sgpr2_sgpr3, %sgpr4
+name: break_clause_simple_load_mubuf_offen_ptr
+
+body: |
+  bb.0:
+    %vgpr1 = BUFFER_LOAD_DWORD_OFFEN %vgpr2, %sgpr0_sgpr1_sgpr2_sgpr3, %sgpr4, 0, 0, 0, 0, implicit %exec
+    %vgpr2 = BUFFER_LOAD_DWORD_OFFEN %vgpr3, %sgpr0_sgpr1_sgpr2_sgpr3, %sgpr4, 0, 0, 0, 0, implicit %exec
+    S_ENDPGM
+...
+---
+# BUFFER instructions overwriting their own inputs is supposedly OK.
+
+# GCN-LABEL: name: mubuf_load4_overwrite_ptr
+# GCN: bb.0:
+# GCN-NEXT: %vgpr0 = BUFFER_LOAD_DWORD_OFFEN %vgpr0, %sgpr0_sgpr1_sgpr2_sgpr3, %sgpr4, 0, 0, 0, 0, implicit %exec
+# GCN-NEXT: %vgpr1 = V_MOV_B32_e32 0, implicit %exec
+# GCN-NEXT: S_WAITCNT 3952
+# GCN-NEXT: %vgpr2 = V_MOV_B32_e32 %vgpr0, implicit %exec
+# GCN-NEXT: S_ENDPGM
+name: mubuf_load4_overwrite_ptr
+
+body: |
+  bb.0:
+    %vgpr0 = BUFFER_LOAD_DWORD_OFFEN %vgpr0, %sgpr0_sgpr1_sgpr2_sgpr3, %sgpr4, 0, 0, 0, 0, implicit %exec
+    %vgpr1 = V_MOV_B32_e32 0, implicit %exec
+    %vgpr2 = V_MOV_B32_e32 %vgpr0, implicit %exec
+    S_ENDPGM
+...
+---
+# Break a clause from interference between mubuf and flat instructions
+
+# GCN-LABEL: name: break_clause_flat_load_mubuf_load
+# GCN: bb.0:
+# GCN-NEXT: %vgpr0 = FLAT_LOAD_DWORD %vgpr2_vgpr3, 0, 0, 0, implicit %exec, implicit %flat_scr
+# GCN-NEXT: S_NOP 0
+# GCN-NEXT: %vgpr2 = BUFFER_LOAD_DWORD_OFFEN %vgpr1, %sgpr0_sgpr1_sgpr2_sgpr3, %sgpr4, 0, 0, 0, 0, implicit %exec
+# GCN-NEXT: S_ENDPGM
+name: break_clause_flat_load_mubuf_load
+
+body: |
+  bb.0:
+    %vgpr0 = FLAT_LOAD_DWORD %vgpr2_vgpr3, 0, 0, 0, implicit %exec, implicit %flat_scr
+    %vgpr2 = BUFFER_LOAD_DWORD_OFFEN %vgpr1, %sgpr0_sgpr1_sgpr2_sgpr3, %sgpr4, 0, 0, 0, 0, implicit %exec
+    S_ENDPGM
+...
+# Break a clause from interference between mubuf and flat instructions
+
+# GCN-LABEL: name: break_clause_mubuf_load_flat_load
+# GCN: bb.0:
+# GCN-NEXT: %vgpr0 = BUFFER_LOAD_DWORD_OFFEN %vgpr1, %sgpr0_sgpr1_sgpr2_sgpr3, %sgpr4
+# GCN-NEXT: S_NOP 0
+# GCN-NEXT: %vgpr1 = FLAT_LOAD_DWORD %vgpr2_vgpr3
+# GCN-NEXT: S_ENDPGM
+name: break_clause_mubuf_load_flat_load
+
+body: |
+  bb.0:
+    %vgpr0 = BUFFER_LOAD_DWORD_OFFEN %vgpr1, %sgpr0_sgpr1_sgpr2_sgpr3, %sgpr4, 0, 0, 0, 0, implicit %exec
+    %vgpr1 = FLAT_LOAD_DWORD %vgpr2_vgpr3, 0, 0, 0, implicit %exec, implicit %flat_scr
+
+    S_ENDPGM
+...
+---
+
+# GCN-LABEL: name: break_clause_atomic_rtn_into_ptr_flat4
+# GCN: bb.0:
+# GCN-NEXT: %vgpr2 = FLAT_LOAD_DWORD %vgpr4_vgpr5
+# GCN-NEXT: S_NOP 0
+# GCN-NEXT: %vgpr4 = FLAT_ATOMIC_ADD_RTN %vgpr5_vgpr6, %vgpr7
+# GCN-NEXT: S_ENDPGM
+name: break_clause_atomic_rtn_into_ptr_flat4
+
+body: |
+  bb.0:
+    %vgpr2 = FLAT_LOAD_DWORD %vgpr4_vgpr5, 0, 0, 0, implicit %exec, implicit %flat_scr
+    %vgpr4 = FLAT_ATOMIC_ADD_RTN %vgpr5_vgpr6, %vgpr7, 0, 0, implicit %exec, implicit %flat_scr
+    S_ENDPGM
+...
+---
+# GCN-LABEL: name: break_clause_atomic_nortn_ptr_load_flat4
+# GCN: bb.0:
+# GCN-NEXT: FLAT_ATOMIC_ADD %vgpr0_vgpr1, %vgpr2, 0, 0, implicit %exec, implicit %flat_scr
+# GCN-NEXT: S_WAITCNT 112
+# GCN-NEXT: %vgpr2 = FLAT_LOAD_DWORD %vgpr3_vgpr4, 0, 0, 0, implicit %exec, implicit %flat_scr
+# GCN-NEXT: S_ENDPGM
+name: break_clause_atomic_nortn_ptr_load_flat4
+
+body: |
+  bb.0:
+    FLAT_ATOMIC_ADD %vgpr0_vgpr1, %vgpr2, 0, 0, implicit %exec, implicit %flat_scr
+    %vgpr2 = FLAT_LOAD_DWORD %vgpr3_vgpr4, 0, 0, 0, implicit %exec, implicit %flat_scr
+    S_ENDPGM
+...
+---
+
+# GCN-LABEL: name: break_clause_atomic_rtn_into_ptr_mubuf4
+# GCN: bb.0:
+# GCN-NEXT: %vgpr1 = BUFFER_LOAD_DWORD_OFFEN %vgpr2, %sgpr0_sgpr1_sgpr2_sgpr3, %sgpr4
+# GCN-NEXT: S_NOP 0
+# GCN-NEXT: %vgpr2 = BUFFER_ATOMIC_ADD_RTN_OFFEN %vgpr2, %vgpr5, %sgpr0_sgpr1_sgpr2_sgpr3, %sgpr4
+# GCN-NEXT: S_ENDPGM
+name: break_clause_atomic_rtn_into_ptr_mubuf4
+
+body: |
+  bb.0:
+    %vgpr1 = BUFFER_LOAD_DWORD_OFFEN %vgpr2, %sgpr0_sgpr1_sgpr2_sgpr3, %sgpr4, 0, 0, 0, 0, implicit %exec
+    %vgpr2 = BUFFER_ATOMIC_ADD_RTN_OFFEN %vgpr2, %vgpr5, %sgpr0_sgpr1_sgpr2_sgpr3, %sgpr4, 0, 0, implicit %exec
+    S_ENDPGM
+...
+---
+
+# GCN-LABEL: name: break_clause_atomic_nortn_ptr_load_mubuf4
+# GCN: bb.0:
+# GCN-NEXT: BUFFER_ATOMIC_ADD_OFFEN %vgpr0, %vgpr1, %sgpr0_sgpr1_sgpr2_sgpr3, %sgpr4
+# GCN-NEXT: S_NOP 0
+# GCN-NEXT: %vgpr1 = BUFFER_LOAD_DWORD_OFFEN %vgpr2, %sgpr0_sgpr1_sgpr2_sgpr3, %sgpr4
+# GCN-NEXT: S_ENDPGM
+name: break_clause_atomic_nortn_ptr_load_mubuf4
+
+body: |
+  bb.0:
+    BUFFER_ATOMIC_ADD_OFFEN %vgpr0, %vgpr1, %sgpr0_sgpr1_sgpr2_sgpr3, %sgpr4, 0, 0, implicit %exec
+    %vgpr1 = BUFFER_LOAD_DWORD_OFFEN %vgpr2, %sgpr0_sgpr1_sgpr2_sgpr3, %sgpr4, 0, 0, 0, 0, implicit %exec
+    S_ENDPGM
+...
+---
+# Make sure there is no assert on mubuf instructions which do not have
+# vaddr, and don't add register to track.
+# GCN-LABEL: name: no_break_clause_mubuf_load_novaddr
+# GCN: bb.0:
+# GCN-NEXT: BUFFER_ATOMIC_ADD_OFFEN %vgpr0, %vgpr1, %sgpr0_sgpr1_sgpr2_sgpr3, %sgpr4
+# GCN-NEXT: S_NOP 0
+# GCN-NEXT: %vgpr1 = BUFFER_LOAD_DWORD_OFFEN %vgpr2, %sgpr0_sgpr1_sgpr2_sgpr3, %sgpr4
+# GCN-NEXT: S_ENDPGM
+name: no_break_clause_mubuf_load_novaddr
+
+body: |
+  bb.0:
+    %vgpr1 = BUFFER_LOAD_DWORD_OFFSET %sgpr0_sgpr1_sgpr2_sgpr3, %sgpr4, 0, 0, 0, 0, implicit %exec
+    %vgpr3 = BUFFER_LOAD_DWORD_OFFSET %sgpr0_sgpr1_sgpr2_sgpr3, %sgpr4, 0, 0, 0, 0, implicit %exec
+    S_ENDPGM
+...