diff --git a/llvm/lib/Target/AMDGPU/SIInsertWaitcnts.cpp b/llvm/lib/Target/AMDGPU/SIInsertWaitcnts.cpp
--- a/llvm/lib/Target/AMDGPU/SIInsertWaitcnts.cpp
+++ b/llvm/lib/Target/AMDGPU/SIInsertWaitcnts.cpp
@@ -292,6 +292,11 @@
     VgprVmemTypes[GprNo] = 0;
   }
 
+  void setNonKernelFunctionInitialState() {
+    for (InstCounterType Counter : inst_counter_types())
+      setScoreUB(Counter, getWaitCountMax(Counter));
+  }
+
   void print(raw_ostream &);
   void dump() { print(dbgs()); }
 
@@ -1857,6 +1862,12 @@
       ;
     BuildMI(EntryBB, I, DebugLoc(), TII->get(AMDGPU::S_WAITCNT)).addImm(0);
 
+    // initialize the incoming as if all counters dirty
+    auto NonKernelInitialState =
+        std::make_unique<WaitcntBrackets>(ST, Limits, Encoding);
+    NonKernelInitialState->setNonKernelFunctionInitialState();
+    BlockInfos[&EntryBB].Incoming = std::move(NonKernelInitialState);
+
     Modified = true;
   }
 
diff --git a/llvm/test/CodeGen/AMDGPU/amd.endpgm.ll b/llvm/test/CodeGen/AMDGPU/amd.endpgm.ll
--- a/llvm/test/CodeGen/AMDGPU/amd.endpgm.ll
+++ b/llvm/test/CodeGen/AMDGPU/amd.endpgm.ll
@@ -33,6 +33,8 @@
 ; GFX11-LABEL: test1:
 ; GFX11:       ; %bb.0:
 ; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NEXT:    s_nop 0
+; GFX11-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-NEXT:    s_endpgm
   tail call void @llvm.amdgcn.endpgm()
   unreachable
diff --git a/llvm/test/CodeGen/AMDGPU/back-off-barrier-subtarget-feature.ll b/llvm/test/CodeGen/AMDGPU/back-off-barrier-subtarget-feature.ll
--- a/llvm/test/CodeGen/AMDGPU/back-off-barrier-subtarget-feature.ll
+++ b/llvm/test/CodeGen/AMDGPU/back-off-barrier-subtarget-feature.ll
@@ -45,6 +45,7 @@
 ; GFX11-BACKOFF-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX11-BACKOFF-NEXT:    flat_load_b32 v0, v[0:1]
 ; GFX11-BACKOFF-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-BACKOFF-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-BACKOFF-NEXT:    s_barrier
 ; GFX11-BACKOFF-NEXT:    flat_store_b32 v[2:3], v0
 ; GFX11-BACKOFF-NEXT:    s_waitcnt lgkmcnt(0)
diff --git a/llvm/test/CodeGen/AMDGPU/waitcnt-overflow.mir b/llvm/test/CodeGen/AMDGPU/waitcnt-overflow.mir
--- a/llvm/test/CodeGen/AMDGPU/waitcnt-overflow.mir
+++ b/llvm/test/CodeGen/AMDGPU/waitcnt-overflow.mir
@@ -48,6 +48,7 @@
     ; GFX9-NEXT: $vgpr4 = V_MAC_F32_e32 0, $vgpr5, $vgpr4, implicit $mode, implicit $exec
     ; GFX9-NEXT: $vgpr6 = V_MAC_F32_e32 0, $vgpr7, $vgpr6, implicit $mode, implicit $exec
     ; GFX9-NEXT: S_ENDPGM 0
+    ;
     ; GFX10-LABEL: name: max-counter-lgkmcnt
     ; GFX10: liveins: $vgpr99
     ; GFX10-NEXT: {{  $}}
@@ -79,6 +80,7 @@
     ; GFX10-NEXT: S_WAITCNT 52863
     ; GFX10-NEXT: $vgpr6 = V_MAC_F32_e32 0, $vgpr7, $vgpr6, implicit $mode, implicit $exec
     ; GFX10-NEXT: S_ENDPGM 0
+    ;
     ; GFX11-LABEL: name: max-counter-lgkmcnt
     ; GFX11: liveins: $vgpr99
     ; GFX11-NEXT: {{  $}}
@@ -109,6 +111,8 @@
     ; GFX11-NEXT: $vgpr4 = V_MAC_F32_e32 0, $vgpr5, $vgpr4, implicit $mode, implicit $exec
     ; GFX11-NEXT: S_WAITCNT 64743
     ; GFX11-NEXT: $vgpr6 = V_MAC_F32_e32 0, $vgpr7, $vgpr6, implicit $mode, implicit $exec
+    ; GFX11-NEXT: S_NOP 0
+    ; GFX11-NEXT: S_SENDMSG 3, implicit $exec, implicit $m0
     ; GFX11-NEXT: S_ENDPGM 0
     $vgpr0_vgpr1 = DS_READ2_B32_gfx9 renamable $vgpr99, 0, 1, 0, implicit $exec
     $vgpr2_vgpr3 = DS_READ2_B32_gfx9 renamable $vgpr99, 2, 3, 0, implicit $exec
@@ -219,6 +223,7 @@
     ; GFX9-NEXT: $vgpr2 = V_MAC_F32_e32 0, $vgpr3, $vgpr2, implicit $mode, implicit $exec
     ; GFX9-NEXT: $vgpr3 = V_MAC_F32_e32 0, $vgpr4, $vgpr3, implicit $mode, implicit $exec
     ; GFX9-NEXT: S_ENDPGM 0
+    ;
     ; GFX10-LABEL: name: max-counter-vmcnt
     ; GFX10: liveins: $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4
     ; GFX10-NEXT: {{  $}}
@@ -296,6 +301,7 @@
     ; GFX10-NEXT: $vgpr2 = V_MAC_F32_e32 0, $vgpr3, $vgpr2, implicit $mode, implicit $exec
     ; GFX10-NEXT: $vgpr3 = V_MAC_F32_e32 0, $vgpr4, $vgpr3, implicit $mode, implicit $exec
     ; GFX10-NEXT: S_ENDPGM 0
+    ;
     ; GFX11-LABEL: name: max-counter-vmcnt
     ; GFX11: liveins: $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4
     ; GFX11-NEXT: {{  $}}
@@ -372,6 +378,8 @@
     ; GFX11-NEXT: $vgpr1 = V_MAC_F32_e32 0, $vgpr2, $vgpr1, implicit $mode, implicit $exec
     ; GFX11-NEXT: $vgpr2 = V_MAC_F32_e32 0, $vgpr3, $vgpr2, implicit $mode, implicit $exec
     ; GFX11-NEXT: $vgpr3 = V_MAC_F32_e32 0, $vgpr4, $vgpr3, implicit $mode, implicit $exec
+    ; GFX11-NEXT: S_NOP 0
+    ; GFX11-NEXT: S_SENDMSG 3, implicit $exec, implicit $m0
     ; GFX11-NEXT: S_ENDPGM 0
     $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 0, 0, 0, implicit $exec
     $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 4, 0, 0, implicit $exec
@@ -468,6 +476,7 @@
     ; GFX9-NEXT: EXP 0, $vgpr1, $vgpr1, $vgpr1, $vgpr1, -1, -1, 15, implicit $exec
     ; GFX9-NEXT: $vgpr0 = V_MAC_F32_e32 0, $vgpr1, $vgpr0, implicit $mode, implicit $exec
     ; GFX9-NEXT: S_ENDPGM 0
+    ;
     ; GFX10-LABEL: name: max-counter-expcnt
     ; GFX10: liveins: $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, $vgpr0, $vgpr1
     ; GFX10-NEXT: {{  $}}
@@ -482,6 +491,7 @@
     ; GFX10-NEXT: EXP 0, $vgpr1, $vgpr1, $vgpr1, $vgpr1, -1, -1, 15, implicit $exec
     ; GFX10-NEXT: $vgpr0 = V_MAC_F32_e32 0, $vgpr1, $vgpr0, implicit $mode, implicit $exec
     ; GFX10-NEXT: S_ENDPGM 0
+    ;
     ; GFX11-LABEL: name: max-counter-expcnt
     ; GFX11: liveins: $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, $vgpr0, $vgpr1
     ; GFX11-NEXT: {{  $}}
@@ -495,6 +505,8 @@
     ; GFX11-NEXT: EXP 0, $vgpr1, $vgpr1, $vgpr1, $vgpr1, -1, -1, 15, implicit $exec
     ; GFX11-NEXT: EXP 0, $vgpr1, $vgpr1, $vgpr1, $vgpr1, -1, -1, 15, implicit $exec
     ; GFX11-NEXT: $vgpr0 = V_MAC_F32_e32 0, $vgpr1, $vgpr0, implicit $mode, implicit $exec
+    ; GFX11-NEXT: S_NOP 0
+    ; GFX11-NEXT: S_SENDMSG 3, implicit $exec, implicit $m0
     ; GFX11-NEXT: S_ENDPGM 0
     EXP 0, $vgpr0, $vgpr0, $vgpr0, $vgpr0, -1, -1, 15, implicit $exec
     EXP 0, $vgpr1, $vgpr1, $vgpr1, $vgpr1, -1, -1, 15, implicit $exec