[llvm-branch-commits] [llvm] AMDGPU: Custom expand flat cmpxchg which may access private (PR #109410)

Matt Arsenault via llvm-branch-commits llvm-branch-commits at lists.llvm.org
Thu Oct 31 08:12:26 PDT 2024


https://github.com/arsenm updated https://github.com/llvm/llvm-project/pull/109410

>From 834ff3b40bd82cb54bb33532a54ad36870ea2b24 Mon Sep 17 00:00:00 2001
From: Matt Arsenault <Matthew.Arsenault at amd.com>
Date: Wed, 14 Aug 2024 13:57:14 +0400
Subject: [PATCH 1/2] AMDGPU: Custom expand flat cmpxchg which may access
 private

64-bit flat cmpxchg instructions do not work correctly for scratch
addresses, and need to be expanded as non-atomic.

Allow custom expansion of cmpxchg in AtomicExpand, as is
already the case for atomicrmw.
---
 llvm/include/llvm/CodeGen/TargetLowering.h    |    5 +
 .../llvm/Transforms/Utils/LowerAtomic.h       |    7 +
 llvm/lib/CodeGen/AtomicExpandPass.cpp         |    4 +
 llvm/lib/Target/AMDGPU/SIISelLowering.cpp     |  146 ++-
 llvm/lib/Target/AMDGPU/SIISelLowering.h       |    3 +
 llvm/lib/Transforms/Utils/LowerAtomic.cpp     |   21 +-
 llvm/test/CodeGen/AMDGPU/flat_atomics_i64.ll  | 1019 +++++++++++++++--
 ...expand-atomicrmw-flat-noalias-addrspace.ll |    6 +-
 ...expand-atomicrmw-integer-ops-0-to-add-0.ll |    6 +-
 .../expand-cmpxchg-flat-maybe-private.ll      |  104 +-
 10 files changed, 1157 insertions(+), 164 deletions(-)

diff --git a/llvm/include/llvm/CodeGen/TargetLowering.h b/llvm/include/llvm/CodeGen/TargetLowering.h
index 8e0cdc6f1a5e77..e0b638201a0474 100644
--- a/llvm/include/llvm/CodeGen/TargetLowering.h
+++ b/llvm/include/llvm/CodeGen/TargetLowering.h
@@ -2204,6 +2204,11 @@ class TargetLoweringBase {
         "Generic atomicrmw expansion unimplemented on this target");
   }
 
+  /// Perform a cmpxchg expansion using a target-specific method.
+  virtual void emitExpandAtomicCmpXchg(AtomicCmpXchgInst *CI) const {
+    llvm_unreachable("Generic cmpxchg expansion unimplemented on this target");
+  }
+
   /// Perform a bit test atomicrmw using a target-specific intrinsic. This
   /// represents the combined bit test intrinsic which will be lowered at a late
   /// stage by the backend.
diff --git a/llvm/include/llvm/Transforms/Utils/LowerAtomic.h b/llvm/include/llvm/Transforms/Utils/LowerAtomic.h
index b25b281667f9cb..295c2bd2b4b47e 100644
--- a/llvm/include/llvm/Transforms/Utils/LowerAtomic.h
+++ b/llvm/include/llvm/Transforms/Utils/LowerAtomic.h
@@ -23,6 +23,13 @@ class IRBuilderBase;
 /// Convert the given Cmpxchg into primitive load and compare.
 bool lowerAtomicCmpXchgInst(AtomicCmpXchgInst *CXI);
 
+/// Emit IR to implement the given cmpxchg operation on values in registers,
+/// returning the new value.
+std::pair<Value *, Value *> buildAtomicCmpXchgValue(IRBuilderBase &Builder,
+                                                    Value *Ptr, Value *Cmp,
+                                                    Value *Val,
+                                                    Align Alignment);
+
 /// Convert the given RMWI into primitive load and stores,
 /// assuming that doing so is legal. Return true if the lowering
 /// succeeds.
diff --git a/llvm/lib/CodeGen/AtomicExpandPass.cpp b/llvm/lib/CodeGen/AtomicExpandPass.cpp
index 0aff4f1f5cf1cb..1471e3d7cbc29d 100644
--- a/llvm/lib/CodeGen/AtomicExpandPass.cpp
+++ b/llvm/lib/CodeGen/AtomicExpandPass.cpp
@@ -1674,6 +1674,10 @@ bool AtomicExpandImpl::tryExpandAtomicCmpXchg(AtomicCmpXchgInst *CI) {
     return true;
   case TargetLoweringBase::AtomicExpansionKind::NotAtomic:
     return lowerAtomicCmpXchgInst(CI);
+  case TargetLoweringBase::AtomicExpansionKind::Expand: {
+    TLI->emitExpandAtomicCmpXchg(CI);
+    return true;
+  }
   }
 }
 
diff --git a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
index d66610ae0a160d..c8a46875bda408 100644
--- a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
@@ -16577,9 +16577,21 @@ SITargetLowering::shouldExpandAtomicStoreInIR(StoreInst *SI) const {
 
 TargetLowering::AtomicExpansionKind
 SITargetLowering::shouldExpandAtomicCmpXchgInIR(AtomicCmpXchgInst *CmpX) const {
-  return CmpX->getPointerAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS
-             ? AtomicExpansionKind::NotAtomic
-             : AtomicExpansionKind::None;
+  unsigned AddrSpace = CmpX->getPointerAddressSpace();
+  if (AddrSpace == AMDGPUAS::PRIVATE_ADDRESS)
+    return AtomicExpansionKind::NotAtomic;
+
+  if (AddrSpace != AMDGPUAS::FLAT_ADDRESS || !flatInstrMayAccessPrivate(CmpX))
+    return AtomicExpansionKind::None;
+
+  const DataLayout &DL = CmpX->getDataLayout();
+
+  Type *ValTy = CmpX->getNewValOperand()->getType();
+
+  // If a 64-bit flat atomic may alias private, we need to avoid using the
+  // atomic in the private case.
+  return DL.getTypeSizeInBits(ValTy) == 64 ? AtomicExpansionKind::Expand
+                                           : AtomicExpansionKind::None;
 }
 
 const TargetRegisterClass *
@@ -16745,40 +16757,8 @@ bool SITargetLowering::checkForPhysRegDependency(
   return false;
 }
 
-void SITargetLowering::emitExpandAtomicRMW(AtomicRMWInst *AI) const {
-  AtomicRMWInst::BinOp Op = AI->getOperation();
-
-  if (Op == AtomicRMWInst::Sub || Op == AtomicRMWInst::Or ||
-      Op == AtomicRMWInst::Xor) {
-    if (auto *ConstVal = dyn_cast<Constant>(AI->getValOperand());
-        ConstVal && ConstVal->isNullValue()) {
-      // atomicrmw or %ptr, 0 -> atomicrmw add %ptr, 0
-      AI->setOperation(AtomicRMWInst::Add);
-
-      // TODO: Turn the below private handling into a no-op for idempotent
-      // cases.
-    }
-  }
-
-  // The non-flat expansions should only perform the de-canonicalization of
-  // identity values.
-  if (AI->getPointerAddressSpace() != AMDGPUAS::FLAT_ADDRESS)
-    return;
-
-  // FullFlatEmulation is true if we need to issue the private, shared, and
-  // global cases.
-  //
-  // If this is false, we are only dealing with the flat-targeting-private case,
-  // where we only insert a check for private and still use the flat instruction
-  // for global and shared.
-
-  // TODO: Avoid the private check for the fadd case depending on
-  // noalias.addrspace.
-
-  bool FullFlatEmulation = Op == AtomicRMWInst::FAdd &&
-                           Subtarget->hasAtomicFaddInsts() &&
-                           AI->getType()->isFloatTy();
-
+void SITargetLowering::emitExpandAtomicAddrSpacePredicate(
+    Instruction *AI) const {
   // Given: atomicrmw fadd ptr %addr, float %val ordering
   //
   // With this expansion we produce the following code:
@@ -16825,6 +16805,34 @@ void SITargetLowering::emitExpandAtomicRMW(AtomicRMWInst *AI) const {
   IRBuilder<> Builder(AI);
   LLVMContext &Ctx = Builder.getContext();
 
+  auto *RMW = dyn_cast<AtomicRMWInst>(AI);
+  const unsigned PtrOpIdx = RMW ? AtomicRMWInst::getPointerOperandIndex()
+                                : AtomicCmpXchgInst::getPointerOperandIndex();
+  Value *Addr = AI->getOperand(PtrOpIdx);
+
+  /// TODO: Only need to check private, then emit flat-known-not private (no
+  /// need for shared block, or cast to global).
+  AtomicCmpXchgInst *CX = dyn_cast<AtomicCmpXchgInst>(AI);
+
+  Align Alignment;
+  if (RMW)
+    Alignment = RMW->getAlign();
+  else if (CX)
+    Alignment = CX->getAlign();
+  else
+    llvm_unreachable("unhandled atomic operation");
+
+  // FullFlatEmulation is true if we need to issue the private, shared, and
+  // global cases.
+  //
+  // If this is false, we are only dealing with the flat-targeting-private case,
+  // where we only insert a check for private and still use the flat instruction
+  // for global and shared.
+
+  bool FullFlatEmulation = RMW && RMW->getOperation() == AtomicRMWInst::FAdd &&
+                           Subtarget->hasAtomicFaddInsts() &&
+                           RMW->getType()->isFloatTy();
+
   // If the return value isn't used, do not introduce a false use in the phi.
   bool ReturnValueIsUsed = !AI->use_empty();
 
@@ -16846,11 +16854,6 @@ void SITargetLowering::emitExpandAtomicRMW(AtomicRMWInst *AI) const {
   BasicBlock *GlobalBB = BasicBlock::Create(Ctx, "atomicrmw.global", F, ExitBB);
   BasicBlock *PhiBB = BasicBlock::Create(Ctx, "atomicrmw.phi", F, ExitBB);
 
-  Value *Val = AI->getValOperand();
-  Type *ValTy = Val->getType();
-  Value *Addr = AI->getPointerOperand();
-  Align Alignment = AI->getAlign();
-
   std::prev(BB->end())->eraseFromParent();
   Builder.SetInsertPoint(BB);
 
@@ -16865,8 +16868,7 @@ void SITargetLowering::emitExpandAtomicRMW(AtomicRMWInst *AI) const {
 
     Instruction *Clone = AI->clone();
     Clone->insertInto(SharedBB, SharedBB->end());
-    Clone->getOperandUse(AtomicRMWInst::getPointerOperandIndex())
-        .set(CastToLocal);
+    Clone->getOperandUse(PtrOpIdx).set(CastToLocal);
     LoadedShared = Clone;
 
     Builder.CreateBr(PhiBB);
@@ -16878,14 +16880,29 @@ void SITargetLowering::emitExpandAtomicRMW(AtomicRMWInst *AI) const {
   Builder.CreateCondBr(IsPrivate, PrivateBB, GlobalBB);
 
   Builder.SetInsertPoint(PrivateBB);
+
   Value *CastToPrivate = Builder.CreateAddrSpaceCast(
       Addr, PointerType::get(Ctx, AMDGPUAS::PRIVATE_ADDRESS));
-  Value *LoadedPrivate = Builder.CreateAlignedLoad(ValTy, CastToPrivate,
-                                                   Alignment, "loaded.private");
 
-  Value *NewVal = buildAtomicRMWValue(Op, Builder, LoadedPrivate, Val);
+  Value *LoadedPrivate;
+  if (RMW) {
+    LoadedPrivate = Builder.CreateAlignedLoad(
+        RMW->getType(), CastToPrivate, RMW->getAlign(), "loaded.private");
+
+    Value *NewVal = buildAtomicRMWValue(RMW->getOperation(), Builder,
+                                        LoadedPrivate, RMW->getValOperand());
+
+    Builder.CreateAlignedStore(NewVal, CastToPrivate, RMW->getAlign());
+  } else {
+    auto [ResultLoad, Equal] =
+        buildAtomicCmpXchgValue(Builder, CastToPrivate, CX->getCompareOperand(),
+                                CX->getNewValOperand(), CX->getAlign());
+
+    Value *Insert = Builder.CreateInsertValue(PoisonValue::get(CX->getType()),
+                                              ResultLoad, 0);
+    LoadedPrivate = Builder.CreateInsertValue(Insert, Equal, 1);
+  }
 
-  Builder.CreateAlignedStore(NewVal, CastToPrivate, Alignment);
   Builder.CreateBr(PhiBB);
 
   Builder.SetInsertPoint(GlobalBB);
@@ -16895,8 +16912,7 @@ void SITargetLowering::emitExpandAtomicRMW(AtomicRMWInst *AI) const {
   if (FullFlatEmulation) {
     Value *CastToGlobal = Builder.CreateAddrSpaceCast(
         Addr, PointerType::get(Ctx, AMDGPUAS::GLOBAL_ADDRESS));
-    AI->getOperandUse(AtomicRMWInst::getPointerOperandIndex())
-        .set(CastToGlobal);
+    AI->getOperandUse(PtrOpIdx).set(CastToGlobal);
   }
 
   AI->removeFromParent();
@@ -16920,7 +16936,7 @@ void SITargetLowering::emitExpandAtomicRMW(AtomicRMWInst *AI) const {
   Builder.SetInsertPoint(PhiBB);
 
   if (ReturnValueIsUsed) {
-    PHINode *Loaded = Builder.CreatePHI(ValTy, 3);
+    PHINode *Loaded = Builder.CreatePHI(AI->getType(), 3);
     AI->replaceAllUsesWith(Loaded);
     if (FullFlatEmulation)
       Loaded->addIncoming(LoadedShared, SharedBB);
@@ -16932,6 +16948,34 @@ void SITargetLowering::emitExpandAtomicRMW(AtomicRMWInst *AI) const {
   Builder.CreateBr(ExitBB);
 }
 
+void SITargetLowering::emitExpandAtomicRMW(AtomicRMWInst *AI) const {
+  AtomicRMWInst::BinOp Op = AI->getOperation();
+
+  if (Op == AtomicRMWInst::Sub || Op == AtomicRMWInst::Or ||
+      Op == AtomicRMWInst::Xor) {
+    if (const auto *ConstVal = dyn_cast<Constant>(AI->getValOperand());
+        ConstVal && ConstVal->isNullValue()) {
+      // atomicrmw or %ptr, 0 -> atomicrmw add %ptr, 0
+      AI->setOperation(AtomicRMWInst::Add);
+
+      // We may still need the private-alias-flat handling below.
+
+      // TODO: Skip this for cases where we cannot access remote memory.
+    }
+  }
+
+  // The non-flat expansions should only perform the de-canonicalization of
+  // identity values.
+  if (AI->getPointerAddressSpace() != AMDGPUAS::FLAT_ADDRESS)
+    return;
+
+  emitExpandAtomicAddrSpacePredicate(AI);
+}
+
+void SITargetLowering::emitExpandAtomicCmpXchg(AtomicCmpXchgInst *CI) const {
+  emitExpandAtomicAddrSpacePredicate(CI);
+}
+
 LoadInst *
 SITargetLowering::lowerIdempotentRMWIntoFencedLoad(AtomicRMWInst *AI) const {
   IRBuilder<> Builder(AI);
diff --git a/llvm/lib/Target/AMDGPU/SIISelLowering.h b/llvm/lib/Target/AMDGPU/SIISelLowering.h
index 6c3edf37945e24..32e110fdfa84d4 100644
--- a/llvm/lib/Target/AMDGPU/SIISelLowering.h
+++ b/llvm/lib/Target/AMDGPU/SIISelLowering.h
@@ -544,7 +544,10 @@ class SITargetLowering final : public AMDGPUTargetLowering {
   AtomicExpansionKind shouldExpandAtomicStoreInIR(StoreInst *SI) const override;
   AtomicExpansionKind
   shouldExpandAtomicCmpXchgInIR(AtomicCmpXchgInst *AI) const override;
+
+  void emitExpandAtomicAddrSpacePredicate(Instruction *AI) const;
   void emitExpandAtomicRMW(AtomicRMWInst *AI) const override;
+  void emitExpandAtomicCmpXchg(AtomicCmpXchgInst *CI) const override;
 
   LoadInst *
   lowerIdempotentRMWIntoFencedLoad(AtomicRMWInst *AI) const override;
diff --git a/llvm/lib/Transforms/Utils/LowerAtomic.cpp b/llvm/lib/Transforms/Utils/LowerAtomic.cpp
index 8b3a0ce338e577..89c49d4a0732f6 100644
--- a/llvm/lib/Transforms/Utils/LowerAtomic.cpp
+++ b/llvm/lib/Transforms/Utils/LowerAtomic.cpp
@@ -25,13 +25,11 @@ bool llvm::lowerAtomicCmpXchgInst(AtomicCmpXchgInst *CXI) {
   Value *Cmp = CXI->getCompareOperand();
   Value *Val = CXI->getNewValOperand();
 
-  LoadInst *Orig =
-      Builder.CreateAlignedLoad(Val->getType(), Ptr, CXI->getAlign());
-  Value *Equal = Builder.CreateICmpEQ(Orig, Cmp);
-  Value *Res = Builder.CreateSelect(Equal, Val, Orig);
-  Builder.CreateAlignedStore(Res, Ptr, CXI->getAlign());
+  auto [Orig, Equal] =
+      buildAtomicCmpXchgValue(Builder, Ptr, Cmp, Val, CXI->getAlign());
 
-  Res = Builder.CreateInsertValue(PoisonValue::get(CXI->getType()), Orig, 0);
+  Value *Res =
+      Builder.CreateInsertValue(PoisonValue::get(CXI->getType()), Orig, 0);
   Res = Builder.CreateInsertValue(Res, Equal, 1);
 
   CXI->replaceAllUsesWith(Res);
@@ -39,6 +37,17 @@ bool llvm::lowerAtomicCmpXchgInst(AtomicCmpXchgInst *CXI) {
   return true;
 }
 
+std::pair<Value *, Value *>
+llvm::buildAtomicCmpXchgValue(IRBuilderBase &Builder, Value *Ptr, Value *Cmp,
+                              Value *Val, Align Alignment) {
+  LoadInst *Orig = Builder.CreateAlignedLoad(Val->getType(), Ptr, Alignment);
+  Value *Equal = Builder.CreateICmpEQ(Orig, Cmp);
+  Value *Res = Builder.CreateSelect(Equal, Val, Orig);
+  Builder.CreateAlignedStore(Res, Ptr, Alignment);
+
+  return {Orig, Equal};
+}
+
 Value *llvm::buildAtomicRMWValue(AtomicRMWInst::BinOp Op,
                                  IRBuilderBase &Builder, Value *Loaded,
                                  Value *Val) {
diff --git a/llvm/test/CodeGen/AMDGPU/flat_atomics_i64.ll b/llvm/test/CodeGen/AMDGPU/flat_atomics_i64.ll
index 839f4a18508e5b..f2959dc19ba4c0 100644
--- a/llvm/test/CodeGen/AMDGPU/flat_atomics_i64.ll
+++ b/llvm/test/CodeGen/AMDGPU/flat_atomics_i64.ll
@@ -12232,11 +12232,29 @@ entry:
 define amdgpu_kernel void @atomic_cmpxchg_i64_offset(ptr %out, i64 %in, i64 %old) {
 ; GCN1-LABEL: atomic_cmpxchg_i64_offset:
 ; GCN1:       ; %bb.0: ; %entry
+; GCN1-NEXT:    s_mov_b32 s12, SCRATCH_RSRC_DWORD0
+; GCN1-NEXT:    s_mov_b32 s13, SCRATCH_RSRC_DWORD1
+; GCN1-NEXT:    s_mov_b32 s14, -1
+; GCN1-NEXT:    s_mov_b32 s15, 0xe8f000
 ; GCN1-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x9
+; GCN1-NEXT:    s_load_dword s8, s[2:3], 0x3f
 ; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0xd
+; GCN1-NEXT:    s_add_u32 s12, s12, s9
+; GCN1-NEXT:    s_addc_u32 s13, s13, 0
 ; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    s_add_u32 s2, s4, 32
 ; GCN1-NEXT:    s_addc_u32 s3, s5, 0
+; GCN1-NEXT:    s_cmp_eq_u32 s3, s8
+; GCN1-NEXT:    s_cselect_b64 s[4:5], -1, 0
+; GCN1-NEXT:    s_andn2_b64 vcc, exec, s[4:5]
+; GCN1-NEXT:    s_mov_b64 s[4:5], -1
+; GCN1-NEXT:    s_cbranch_vccnz .LBB90_3
+; GCN1-NEXT:  ; %bb.1: ; %Flow
+; GCN1-NEXT:    s_andn2_b64 vcc, exec, s[4:5]
+; GCN1-NEXT:    s_cbranch_vccz .LBB90_4
+; GCN1-NEXT:  .LBB90_2: ; %atomicrmw.phi
+; GCN1-NEXT:    s_endpgm
+; GCN1-NEXT:  .LBB90_3: ; %atomicrmw.global
 ; GCN1-NEXT:    v_mov_b32_e32 v5, s3
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s6
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s7
@@ -12246,15 +12264,51 @@ define amdgpu_kernel void @atomic_cmpxchg_i64_offset(ptr %out, i64 %in, i64 %old
 ; GCN1-NEXT:    flat_atomic_cmpswap_x2 v[4:5], v[0:3]
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    s_cbranch_execnz .LBB90_2
+; GCN1-NEXT:  .LBB90_4: ; %atomicrmw.private
+; GCN1-NEXT:    v_cmp_ne_u64_e64 s[4:5], s[2:3], 0
+; GCN1-NEXT:    v_mov_b32_e32 v5, s6
+; GCN1-NEXT:    s_and_b64 s[4:5], s[4:5], exec
+; GCN1-NEXT:    s_cselect_b32 s2, s2, -1
+; GCN1-NEXT:    v_mov_b32_e32 v2, s2
+; GCN1-NEXT:    s_add_i32 s2, s2, 4
+; GCN1-NEXT:    v_mov_b32_e32 v3, s2
+; GCN1-NEXT:    buffer_load_dword v0, v2, s[12:15], 0 offen
+; GCN1-NEXT:    buffer_load_dword v1, v3, s[12:15], 0 offen
+; GCN1-NEXT:    v_mov_b32_e32 v4, s7
+; GCN1-NEXT:    s_waitcnt vmcnt(0)
+; GCN1-NEXT:    v_cmp_eq_u64_e32 vcc, s[0:1], v[0:1]
+; GCN1-NEXT:    v_cndmask_b32_e32 v0, v0, v5, vcc
+; GCN1-NEXT:    v_cndmask_b32_e32 v1, v1, v4, vcc
+; GCN1-NEXT:    buffer_store_dword v0, v2, s[12:15], 0 offen
+; GCN1-NEXT:    buffer_store_dword v1, v3, s[12:15], 0 offen
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_cmpxchg_i64_offset:
 ; GCN2:       ; %bb.0: ; %entry
+; GCN2-NEXT:    s_mov_b32 s88, SCRATCH_RSRC_DWORD0
+; GCN2-NEXT:    s_mov_b32 s89, SCRATCH_RSRC_DWORD1
+; GCN2-NEXT:    s_mov_b32 s90, -1
+; GCN2-NEXT:    s_mov_b32 s91, 0xe80000
 ; GCN2-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
+; GCN2-NEXT:    s_load_dword s8, s[2:3], 0xfc
 ; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x34
+; GCN2-NEXT:    s_add_u32 s88, s88, s9
+; GCN2-NEXT:    s_addc_u32 s89, s89, 0
 ; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    s_add_u32 s2, s4, 32
 ; GCN2-NEXT:    s_addc_u32 s3, s5, 0
+; GCN2-NEXT:    s_cmp_eq_u32 s3, s8
+; GCN2-NEXT:    s_cselect_b64 s[4:5], -1, 0
+; GCN2-NEXT:    s_andn2_b64 vcc, exec, s[4:5]
+; GCN2-NEXT:    s_mov_b64 s[4:5], -1
+; GCN2-NEXT:    s_cbranch_vccnz .LBB90_3
+; GCN2-NEXT:  ; %bb.1: ; %Flow
+; GCN2-NEXT:    s_andn2_b64 vcc, exec, s[4:5]
+; GCN2-NEXT:    s_cbranch_vccz .LBB90_4
+; GCN2-NEXT:  .LBB90_2: ; %atomicrmw.phi
+; GCN2-NEXT:    s_endpgm
+; GCN2-NEXT:  .LBB90_3: ; %atomicrmw.global
 ; GCN2-NEXT:    v_mov_b32_e32 v5, s3
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s6
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s7
@@ -12264,6 +12318,23 @@ define amdgpu_kernel void @atomic_cmpxchg_i64_offset(ptr %out, i64 %in, i64 %old
 ; GCN2-NEXT:    flat_atomic_cmpswap_x2 v[4:5], v[0:3]
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    s_cbranch_execnz .LBB90_2
+; GCN2-NEXT:  .LBB90_4: ; %atomicrmw.private
+; GCN2-NEXT:    s_cmp_lg_u64 s[2:3], 0
+; GCN2-NEXT:    s_cselect_b32 s2, s2, -1
+; GCN2-NEXT:    v_mov_b32_e32 v2, s2
+; GCN2-NEXT:    s_add_i32 s2, s2, 4
+; GCN2-NEXT:    v_mov_b32_e32 v3, s2
+; GCN2-NEXT:    buffer_load_dword v0, v2, s[88:91], 0 offen
+; GCN2-NEXT:    buffer_load_dword v1, v3, s[88:91], 0 offen
+; GCN2-NEXT:    v_mov_b32_e32 v5, s6
+; GCN2-NEXT:    v_mov_b32_e32 v4, s7
+; GCN2-NEXT:    s_waitcnt vmcnt(0)
+; GCN2-NEXT:    v_cmp_eq_u64_e32 vcc, s[0:1], v[0:1]
+; GCN2-NEXT:    v_cndmask_b32_e32 v0, v0, v5, vcc
+; GCN2-NEXT:    v_cndmask_b32_e32 v1, v1, v4, vcc
+; GCN2-NEXT:    buffer_store_dword v0, v2, s[88:91], 0 offen
+; GCN2-NEXT:    buffer_store_dword v1, v3, s[88:91], 0 offen
 ; GCN2-NEXT:    s_endpgm
 ;
 ; GFX12-LABEL: atomic_cmpxchg_i64_offset:
@@ -12271,13 +12342,38 @@ define amdgpu_kernel void @atomic_cmpxchg_i64_offset(ptr %out, i64 %in, i64 %old
 ; GFX12-NEXT:    s_clause 0x1
 ; GFX12-NEXT:    s_load_b128 s[4:7], s[2:3], 0x24
 ; GFX12-NEXT:    s_load_b64 s[0:1], s[2:3], 0x34
+; GFX12-NEXT:    s_mov_b64 s[8:9], src_private_base
 ; GFX12-NEXT:    s_wait_kmcnt 0x0
-; GFX12-NEXT:    v_dual_mov_b32 v4, s4 :: v_dual_mov_b32 v5, s5
+; GFX12-NEXT:    s_add_nc_u64 s[2:3], s[4:5], 32
+; GFX12-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
+; GFX12-NEXT:    s_cmp_eq_u32 s3, s9
+; GFX12-NEXT:    s_cselect_b32 s4, -1, 0
+; GFX12-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
+; GFX12-NEXT:    s_mov_b32 s4, -1
+; GFX12-NEXT:    s_cbranch_vccnz .LBB90_3
+; GFX12-NEXT:  ; %bb.1: ; %Flow
+; GFX12-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
+; GFX12-NEXT:    s_cbranch_vccz .LBB90_4
+; GFX12-NEXT:  .LBB90_2: ; %atomicrmw.phi
+; GFX12-NEXT:    s_endpgm
+; GFX12-NEXT:  .LBB90_3: ; %atomicrmw.global
 ; GFX12-NEXT:    v_dual_mov_b32 v0, s6 :: v_dual_mov_b32 v1, s7
 ; GFX12-NEXT:    v_dual_mov_b32 v2, s0 :: v_dual_mov_b32 v3, s1
-; GFX12-NEXT:    flat_atomic_cmpswap_b64 v[4:5], v[0:3] offset:32 scope:SCOPE_DEV
+; GFX12-NEXT:    v_dual_mov_b32 v5, s3 :: v_dual_mov_b32 v4, s2
+; GFX12-NEXT:    flat_atomic_cmpswap_b64 v[4:5], v[0:3] scope:SCOPE_DEV
 ; GFX12-NEXT:    s_wait_storecnt_dscnt 0x0
 ; GFX12-NEXT:    global_inv scope:SCOPE_DEV
+; GFX12-NEXT:    s_cbranch_execnz .LBB90_2
+; GFX12-NEXT:  .LBB90_4: ; %atomicrmw.private
+; GFX12-NEXT:    s_wait_alu 0xfffe
+; GFX12-NEXT:    s_cmp_lg_u64 s[2:3], 0
+; GFX12-NEXT:    s_cselect_b32 s2, s2, -1
+; GFX12-NEXT:    scratch_load_b64 v[0:1], off, s2
+; GFX12-NEXT:    s_wait_loadcnt 0x0
+; GFX12-NEXT:    v_cmp_eq_u64_e32 vcc_lo, s[0:1], v[0:1]
+; GFX12-NEXT:    v_cndmask_b32_e64 v1, v1, s7, vcc_lo
+; GFX12-NEXT:    v_cndmask_b32_e64 v0, v0, s6, vcc_lo
+; GFX12-NEXT:    scratch_store_b64 off, v[0:1], s2
 ; GFX12-NEXT:    s_endpgm
 entry:
   %gep = getelementptr i64, ptr %out, i64 4
@@ -12288,11 +12384,29 @@ entry:
 define amdgpu_kernel void @atomic_cmpxchg_i64_soffset(ptr %out, i64 %in, i64 %old) {
 ; GCN1-LABEL: atomic_cmpxchg_i64_soffset:
 ; GCN1:       ; %bb.0: ; %entry
+; GCN1-NEXT:    s_mov_b32 s12, SCRATCH_RSRC_DWORD0
+; GCN1-NEXT:    s_mov_b32 s13, SCRATCH_RSRC_DWORD1
+; GCN1-NEXT:    s_mov_b32 s14, -1
+; GCN1-NEXT:    s_mov_b32 s15, 0xe8f000
 ; GCN1-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x9
+; GCN1-NEXT:    s_load_dword s8, s[2:3], 0x3f
 ; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0xd
+; GCN1-NEXT:    s_add_u32 s12, s12, s9
+; GCN1-NEXT:    s_addc_u32 s13, s13, 0
 ; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    s_add_u32 s2, s4, 0x11940
 ; GCN1-NEXT:    s_addc_u32 s3, s5, 0
+; GCN1-NEXT:    s_cmp_eq_u32 s3, s8
+; GCN1-NEXT:    s_cselect_b64 s[4:5], -1, 0
+; GCN1-NEXT:    s_andn2_b64 vcc, exec, s[4:5]
+; GCN1-NEXT:    s_mov_b64 s[4:5], -1
+; GCN1-NEXT:    s_cbranch_vccnz .LBB91_3
+; GCN1-NEXT:  ; %bb.1: ; %Flow
+; GCN1-NEXT:    s_andn2_b64 vcc, exec, s[4:5]
+; GCN1-NEXT:    s_cbranch_vccz .LBB91_4
+; GCN1-NEXT:  .LBB91_2: ; %atomicrmw.phi
+; GCN1-NEXT:    s_endpgm
+; GCN1-NEXT:  .LBB91_3: ; %atomicrmw.global
 ; GCN1-NEXT:    v_mov_b32_e32 v5, s3
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s6
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s7
@@ -12302,15 +12416,51 @@ define amdgpu_kernel void @atomic_cmpxchg_i64_soffset(ptr %out, i64 %in, i64 %ol
 ; GCN1-NEXT:    flat_atomic_cmpswap_x2 v[4:5], v[0:3]
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    s_cbranch_execnz .LBB91_2
+; GCN1-NEXT:  .LBB91_4: ; %atomicrmw.private
+; GCN1-NEXT:    v_cmp_ne_u64_e64 s[4:5], s[2:3], 0
+; GCN1-NEXT:    v_mov_b32_e32 v5, s6
+; GCN1-NEXT:    s_and_b64 s[4:5], s[4:5], exec
+; GCN1-NEXT:    s_cselect_b32 s2, s2, -1
+; GCN1-NEXT:    v_mov_b32_e32 v2, s2
+; GCN1-NEXT:    s_add_i32 s2, s2, 4
+; GCN1-NEXT:    v_mov_b32_e32 v3, s2
+; GCN1-NEXT:    buffer_load_dword v0, v2, s[12:15], 0 offen
+; GCN1-NEXT:    buffer_load_dword v1, v3, s[12:15], 0 offen
+; GCN1-NEXT:    v_mov_b32_e32 v4, s7
+; GCN1-NEXT:    s_waitcnt vmcnt(0)
+; GCN1-NEXT:    v_cmp_eq_u64_e32 vcc, s[0:1], v[0:1]
+; GCN1-NEXT:    v_cndmask_b32_e32 v0, v0, v5, vcc
+; GCN1-NEXT:    v_cndmask_b32_e32 v1, v1, v4, vcc
+; GCN1-NEXT:    buffer_store_dword v0, v2, s[12:15], 0 offen
+; GCN1-NEXT:    buffer_store_dword v1, v3, s[12:15], 0 offen
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_cmpxchg_i64_soffset:
 ; GCN2:       ; %bb.0: ; %entry
+; GCN2-NEXT:    s_mov_b32 s88, SCRATCH_RSRC_DWORD0
+; GCN2-NEXT:    s_mov_b32 s89, SCRATCH_RSRC_DWORD1
+; GCN2-NEXT:    s_mov_b32 s90, -1
+; GCN2-NEXT:    s_mov_b32 s91, 0xe80000
 ; GCN2-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
+; GCN2-NEXT:    s_load_dword s8, s[2:3], 0xfc
 ; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x34
+; GCN2-NEXT:    s_add_u32 s88, s88, s9
+; GCN2-NEXT:    s_addc_u32 s89, s89, 0
 ; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    s_add_u32 s2, s4, 0x11940
 ; GCN2-NEXT:    s_addc_u32 s3, s5, 0
+; GCN2-NEXT:    s_cmp_eq_u32 s3, s8
+; GCN2-NEXT:    s_cselect_b64 s[4:5], -1, 0
+; GCN2-NEXT:    s_andn2_b64 vcc, exec, s[4:5]
+; GCN2-NEXT:    s_mov_b64 s[4:5], -1
+; GCN2-NEXT:    s_cbranch_vccnz .LBB91_3
+; GCN2-NEXT:  ; %bb.1: ; %Flow
+; GCN2-NEXT:    s_andn2_b64 vcc, exec, s[4:5]
+; GCN2-NEXT:    s_cbranch_vccz .LBB91_4
+; GCN2-NEXT:  .LBB91_2: ; %atomicrmw.phi
+; GCN2-NEXT:    s_endpgm
+; GCN2-NEXT:  .LBB91_3: ; %atomicrmw.global
 ; GCN2-NEXT:    v_mov_b32_e32 v5, s3
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s6
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s7
@@ -12320,6 +12470,23 @@ define amdgpu_kernel void @atomic_cmpxchg_i64_soffset(ptr %out, i64 %in, i64 %ol
 ; GCN2-NEXT:    flat_atomic_cmpswap_x2 v[4:5], v[0:3]
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    s_cbranch_execnz .LBB91_2
+; GCN2-NEXT:  .LBB91_4: ; %atomicrmw.private
+; GCN2-NEXT:    s_cmp_lg_u64 s[2:3], 0
+; GCN2-NEXT:    s_cselect_b32 s2, s2, -1
+; GCN2-NEXT:    v_mov_b32_e32 v2, s2
+; GCN2-NEXT:    s_add_i32 s2, s2, 4
+; GCN2-NEXT:    v_mov_b32_e32 v3, s2
+; GCN2-NEXT:    buffer_load_dword v0, v2, s[88:91], 0 offen
+; GCN2-NEXT:    buffer_load_dword v1, v3, s[88:91], 0 offen
+; GCN2-NEXT:    v_mov_b32_e32 v5, s6
+; GCN2-NEXT:    v_mov_b32_e32 v4, s7
+; GCN2-NEXT:    s_waitcnt vmcnt(0)
+; GCN2-NEXT:    v_cmp_eq_u64_e32 vcc, s[0:1], v[0:1]
+; GCN2-NEXT:    v_cndmask_b32_e32 v0, v0, v5, vcc
+; GCN2-NEXT:    v_cndmask_b32_e32 v1, v1, v4, vcc
+; GCN2-NEXT:    buffer_store_dword v0, v2, s[88:91], 0 offen
+; GCN2-NEXT:    buffer_store_dword v1, v3, s[88:91], 0 offen
 ; GCN2-NEXT:    s_endpgm
 ;
 ; GFX12-LABEL: atomic_cmpxchg_i64_soffset:
@@ -12327,13 +12494,38 @@ define amdgpu_kernel void @atomic_cmpxchg_i64_soffset(ptr %out, i64 %in, i64 %ol
 ; GFX12-NEXT:    s_clause 0x1
 ; GFX12-NEXT:    s_load_b128 s[4:7], s[2:3], 0x24
 ; GFX12-NEXT:    s_load_b64 s[0:1], s[2:3], 0x34
+; GFX12-NEXT:    s_mov_b64 s[8:9], src_private_base
 ; GFX12-NEXT:    s_wait_kmcnt 0x0
-; GFX12-NEXT:    v_dual_mov_b32 v4, s4 :: v_dual_mov_b32 v5, s5
+; GFX12-NEXT:    s_add_nc_u64 s[2:3], s[4:5], 0x11940
+; GFX12-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
+; GFX12-NEXT:    s_cmp_eq_u32 s3, s9
+; GFX12-NEXT:    s_cselect_b32 s4, -1, 0
+; GFX12-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
+; GFX12-NEXT:    s_mov_b32 s4, -1
+; GFX12-NEXT:    s_cbranch_vccnz .LBB91_3
+; GFX12-NEXT:  ; %bb.1: ; %Flow
+; GFX12-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
+; GFX12-NEXT:    s_cbranch_vccz .LBB91_4
+; GFX12-NEXT:  .LBB91_2: ; %atomicrmw.phi
+; GFX12-NEXT:    s_endpgm
+; GFX12-NEXT:  .LBB91_3: ; %atomicrmw.global
 ; GFX12-NEXT:    v_dual_mov_b32 v0, s6 :: v_dual_mov_b32 v1, s7
 ; GFX12-NEXT:    v_dual_mov_b32 v2, s0 :: v_dual_mov_b32 v3, s1
-; GFX12-NEXT:    flat_atomic_cmpswap_b64 v[4:5], v[0:3] offset:72000 scope:SCOPE_DEV
+; GFX12-NEXT:    v_dual_mov_b32 v5, s3 :: v_dual_mov_b32 v4, s2
+; GFX12-NEXT:    flat_atomic_cmpswap_b64 v[4:5], v[0:3] scope:SCOPE_DEV
 ; GFX12-NEXT:    s_wait_storecnt_dscnt 0x0
 ; GFX12-NEXT:    global_inv scope:SCOPE_DEV
+; GFX12-NEXT:    s_cbranch_execnz .LBB91_2
+; GFX12-NEXT:  .LBB91_4: ; %atomicrmw.private
+; GFX12-NEXT:    s_wait_alu 0xfffe
+; GFX12-NEXT:    s_cmp_lg_u64 s[2:3], 0
+; GFX12-NEXT:    s_cselect_b32 s2, s2, -1
+; GFX12-NEXT:    scratch_load_b64 v[0:1], off, s2
+; GFX12-NEXT:    s_wait_loadcnt 0x0
+; GFX12-NEXT:    v_cmp_eq_u64_e32 vcc_lo, s[0:1], v[0:1]
+; GFX12-NEXT:    v_cndmask_b32_e64 v1, v1, s7, vcc_lo
+; GFX12-NEXT:    v_cndmask_b32_e64 v0, v0, s6, vcc_lo
+; GFX12-NEXT:    scratch_store_b64 off, v[0:1], s2
 ; GFX12-NEXT:    s_endpgm
 entry:
   %gep = getelementptr i64, ptr %out, i64 9000
@@ -12344,54 +12536,144 @@ entry:
 define amdgpu_kernel void @atomic_cmpxchg_i64_ret_offset(ptr %out, ptr %out2, i64 %in, i64 %old) {
 ; GCN1-LABEL: atomic_cmpxchg_i64_ret_offset:
 ; GCN1:       ; %bb.0: ; %entry
-; GCN1-NEXT:    s_load_dwordx8 s[0:7], s[2:3], 0x9
+; GCN1-NEXT:    s_mov_b32 s12, SCRATCH_RSRC_DWORD0
+; GCN1-NEXT:    s_mov_b32 s13, SCRATCH_RSRC_DWORD1
+; GCN1-NEXT:    s_mov_b32 s14, -1
+; GCN1-NEXT:    s_mov_b32 s15, 0xe8f000
+; GCN1-NEXT:    s_add_u32 s12, s12, s9
+; GCN1-NEXT:    s_load_dwordx8 s[4:11], s[2:3], 0x9
+; GCN1-NEXT:    s_load_dword s2, s[2:3], 0x41
+; GCN1-NEXT:    s_addc_u32 s13, s13, 0
 ; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN1-NEXT:    s_add_u32 s0, s0, 32
-; GCN1-NEXT:    s_addc_u32 s1, s1, 0
+; GCN1-NEXT:    s_add_u32 s0, s4, 32
+; GCN1-NEXT:    s_addc_u32 s1, s5, 0
+; GCN1-NEXT:    s_cmp_eq_u32 s1, s2
+; GCN1-NEXT:    s_cselect_b64 s[2:3], -1, 0
+; GCN1-NEXT:    s_andn2_b64 vcc, exec, s[2:3]
+; GCN1-NEXT:    s_cbranch_vccz .LBB92_2
+; GCN1-NEXT:  ; %bb.1: ; %atomicrmw.global
 ; GCN1-NEXT:    v_mov_b32_e32 v5, s1
-; GCN1-NEXT:    v_mov_b32_e32 v0, s4
-; GCN1-NEXT:    v_mov_b32_e32 v1, s5
-; GCN1-NEXT:    v_mov_b32_e32 v2, s6
-; GCN1-NEXT:    v_mov_b32_e32 v3, s7
+; GCN1-NEXT:    v_mov_b32_e32 v0, s8
+; GCN1-NEXT:    v_mov_b32_e32 v1, s9
+; GCN1-NEXT:    v_mov_b32_e32 v2, s10
+; GCN1-NEXT:    v_mov_b32_e32 v3, s11
 ; GCN1-NEXT:    v_mov_b32_e32 v4, s0
 ; GCN1-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN1-NEXT:    buffer_wbinvl1_vol
-; GCN1-NEXT:    v_mov_b32_e32 v2, s2
-; GCN1-NEXT:    v_mov_b32_e32 v3, s3
+; GCN1-NEXT:    s_cbranch_execz .LBB92_3
+; GCN1-NEXT:    s_branch .LBB92_4
+; GCN1-NEXT:  .LBB92_2:
+; GCN1-NEXT:    ; implicit-def: $vgpr0_vgpr1
+; GCN1-NEXT:  .LBB92_3: ; %atomicrmw.private
+; GCN1-NEXT:    v_cmp_ne_u64_e64 s[2:3], s[0:1], 0
+; GCN1-NEXT:    v_mov_b32_e32 v5, s8
+; GCN1-NEXT:    s_and_b64 s[2:3], s[2:3], exec
+; GCN1-NEXT:    s_cselect_b32 s0, s0, -1
+; GCN1-NEXT:    v_mov_b32_e32 v2, s0
+; GCN1-NEXT:    s_add_i32 s0, s0, 4
+; GCN1-NEXT:    v_mov_b32_e32 v3, s0
+; GCN1-NEXT:    buffer_load_dword v0, v2, s[12:15], 0 offen
+; GCN1-NEXT:    buffer_load_dword v1, v3, s[12:15], 0 offen
+; GCN1-NEXT:    v_mov_b32_e32 v4, s9
+; GCN1-NEXT:    s_waitcnt vmcnt(0)
+; GCN1-NEXT:    v_cmp_eq_u64_e32 vcc, s[10:11], v[0:1]
+; GCN1-NEXT:    v_cndmask_b32_e32 v5, v0, v5, vcc
+; GCN1-NEXT:    v_cndmask_b32_e32 v4, v1, v4, vcc
+; GCN1-NEXT:    buffer_store_dword v5, v2, s[12:15], 0 offen
+; GCN1-NEXT:    buffer_store_dword v4, v3, s[12:15], 0 offen
+; GCN1-NEXT:  .LBB92_4: ; %atomicrmw.end
+; GCN1-NEXT:    v_mov_b32_e32 v2, s6
+; GCN1-NEXT:    v_mov_b32_e32 v3, s7
 ; GCN1-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_cmpxchg_i64_ret_offset:
 ; GCN2:       ; %bb.0: ; %entry
-; GCN2-NEXT:    s_load_dwordx8 s[0:7], s[2:3], 0x24
+; GCN2-NEXT:    s_mov_b32 s88, SCRATCH_RSRC_DWORD0
+; GCN2-NEXT:    s_mov_b32 s89, SCRATCH_RSRC_DWORD1
+; GCN2-NEXT:    s_mov_b32 s90, -1
+; GCN2-NEXT:    s_mov_b32 s91, 0xe80000
+; GCN2-NEXT:    s_add_u32 s88, s88, s9
+; GCN2-NEXT:    s_load_dwordx8 s[4:11], s[2:3], 0x24
+; GCN2-NEXT:    s_load_dword s2, s[2:3], 0x104
+; GCN2-NEXT:    s_addc_u32 s89, s89, 0
 ; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN2-NEXT:    s_add_u32 s0, s0, 32
-; GCN2-NEXT:    s_addc_u32 s1, s1, 0
+; GCN2-NEXT:    s_add_u32 s0, s4, 32
+; GCN2-NEXT:    s_addc_u32 s1, s5, 0
+; GCN2-NEXT:    s_cmp_eq_u32 s1, s2
+; GCN2-NEXT:    s_cselect_b64 s[2:3], -1, 0
+; GCN2-NEXT:    s_andn2_b64 vcc, exec, s[2:3]
+; GCN2-NEXT:    s_cbranch_vccz .LBB92_2
+; GCN2-NEXT:  ; %bb.1: ; %atomicrmw.global
 ; GCN2-NEXT:    v_mov_b32_e32 v5, s1
-; GCN2-NEXT:    v_mov_b32_e32 v0, s4
-; GCN2-NEXT:    v_mov_b32_e32 v1, s5
-; GCN2-NEXT:    v_mov_b32_e32 v2, s6
-; GCN2-NEXT:    v_mov_b32_e32 v3, s7
+; GCN2-NEXT:    v_mov_b32_e32 v0, s8
+; GCN2-NEXT:    v_mov_b32_e32 v1, s9
+; GCN2-NEXT:    v_mov_b32_e32 v2, s10
+; GCN2-NEXT:    v_mov_b32_e32 v3, s11
 ; GCN2-NEXT:    v_mov_b32_e32 v4, s0
 ; GCN2-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN2-NEXT:    buffer_wbinvl1_vol
-; GCN2-NEXT:    v_mov_b32_e32 v2, s2
-; GCN2-NEXT:    v_mov_b32_e32 v3, s3
+; GCN2-NEXT:    s_cbranch_execz .LBB92_3
+; GCN2-NEXT:    s_branch .LBB92_4
+; GCN2-NEXT:  .LBB92_2:
+; GCN2-NEXT:    ; implicit-def: $vgpr0_vgpr1
+; GCN2-NEXT:  .LBB92_3: ; %atomicrmw.private
+; GCN2-NEXT:    s_cmp_lg_u64 s[0:1], 0
+; GCN2-NEXT:    s_cselect_b32 s0, s0, -1
+; GCN2-NEXT:    v_mov_b32_e32 v2, s0
+; GCN2-NEXT:    s_add_i32 s0, s0, 4
+; GCN2-NEXT:    v_mov_b32_e32 v3, s0
+; GCN2-NEXT:    buffer_load_dword v0, v2, s[88:91], 0 offen
+; GCN2-NEXT:    buffer_load_dword v1, v3, s[88:91], 0 offen
+; GCN2-NEXT:    v_mov_b32_e32 v5, s8
+; GCN2-NEXT:    v_mov_b32_e32 v4, s9
+; GCN2-NEXT:    s_waitcnt vmcnt(0)
+; GCN2-NEXT:    v_cmp_eq_u64_e32 vcc, s[10:11], v[0:1]
+; GCN2-NEXT:    v_cndmask_b32_e32 v5, v0, v5, vcc
+; GCN2-NEXT:    v_cndmask_b32_e32 v4, v1, v4, vcc
+; GCN2-NEXT:    buffer_store_dword v5, v2, s[88:91], 0 offen
+; GCN2-NEXT:    buffer_store_dword v4, v3, s[88:91], 0 offen
+; GCN2-NEXT:  .LBB92_4: ; %atomicrmw.end
+; GCN2-NEXT:    v_mov_b32_e32 v2, s6
+; GCN2-NEXT:    v_mov_b32_e32 v3, s7
 ; GCN2-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
 ; GCN2-NEXT:    s_endpgm
 ;
 ; GFX12-LABEL: atomic_cmpxchg_i64_ret_offset:
 ; GFX12:       ; %bb.0: ; %entry
 ; GFX12-NEXT:    s_load_b256 s[0:7], s[2:3], 0x24
+; GFX12-NEXT:    s_mov_b64 s[8:9], src_private_base
 ; GFX12-NEXT:    s_wait_kmcnt 0x0
-; GFX12-NEXT:    v_dual_mov_b32 v4, s0 :: v_dual_mov_b32 v5, s1
+; GFX12-NEXT:    s_add_nc_u64 s[0:1], s[0:1], 32
+; GFX12-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
+; GFX12-NEXT:    s_cmp_eq_u32 s1, s9
+; GFX12-NEXT:    s_cselect_b32 s8, -1, 0
+; GFX12-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s8
+; GFX12-NEXT:    s_cbranch_vccz .LBB92_2
+; GFX12-NEXT:  ; %bb.1: ; %atomicrmw.global
 ; GFX12-NEXT:    v_dual_mov_b32 v0, s4 :: v_dual_mov_b32 v1, s5
 ; GFX12-NEXT:    v_dual_mov_b32 v2, s6 :: v_dual_mov_b32 v3, s7
-; GFX12-NEXT:    flat_atomic_cmpswap_b64 v[0:1], v[4:5], v[0:3] offset:32 th:TH_ATOMIC_RETURN scope:SCOPE_DEV
+; GFX12-NEXT:    v_dual_mov_b32 v5, s1 :: v_dual_mov_b32 v4, s0
+; GFX12-NEXT:    flat_atomic_cmpswap_b64 v[0:1], v[4:5], v[0:3] th:TH_ATOMIC_RETURN scope:SCOPE_DEV
 ; GFX12-NEXT:    s_wait_loadcnt_dscnt 0x0
 ; GFX12-NEXT:    global_inv scope:SCOPE_DEV
+; GFX12-NEXT:    s_cbranch_execz .LBB92_3
+; GFX12-NEXT:    s_branch .LBB92_4
+; GFX12-NEXT:  .LBB92_2:
+; GFX12-NEXT:    ; implicit-def: $vgpr0_vgpr1
+; GFX12-NEXT:  .LBB92_3: ; %atomicrmw.private
+; GFX12-NEXT:    s_wait_alu 0xfffe
+; GFX12-NEXT:    s_cmp_lg_u64 s[0:1], 0
+; GFX12-NEXT:    s_cselect_b32 s0, s0, -1
+; GFX12-NEXT:    scratch_load_b64 v[0:1], off, s0
+; GFX12-NEXT:    s_wait_loadcnt 0x0
+; GFX12-NEXT:    v_cmp_eq_u64_e32 vcc_lo, s[6:7], v[0:1]
+; GFX12-NEXT:    v_cndmask_b32_e64 v3, v1, s5, vcc_lo
+; GFX12-NEXT:    v_cndmask_b32_e64 v2, v0, s4, vcc_lo
+; GFX12-NEXT:    scratch_store_b64 off, v[2:3], s0
+; GFX12-NEXT:  .LBB92_4: ; %atomicrmw.end
 ; GFX12-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
 ; GFX12-NEXT:    flat_store_b64 v[2:3], v[0:1]
 ; GFX12-NEXT:    s_endpgm
@@ -12406,42 +12688,113 @@ entry:
 define amdgpu_kernel void @atomic_cmpxchg_i64_addr64_offset(ptr %out, i64 %in, i64 %index, i64 %old) {
 ; GCN1-LABEL: atomic_cmpxchg_i64_addr64_offset:
 ; GCN1:       ; %bb.0: ; %entry
-; GCN1-NEXT:    s_load_dwordx8 s[0:7], s[2:3], 0x9
+; GCN1-NEXT:    s_mov_b32 s12, SCRATCH_RSRC_DWORD0
+; GCN1-NEXT:    s_mov_b32 s13, SCRATCH_RSRC_DWORD1
+; GCN1-NEXT:    s_mov_b32 s14, -1
+; GCN1-NEXT:    s_mov_b32 s15, 0xe8f000
+; GCN1-NEXT:    s_add_u32 s12, s12, s9
+; GCN1-NEXT:    s_load_dwordx8 s[4:11], s[2:3], 0x9
+; GCN1-NEXT:    s_load_dword s2, s[2:3], 0x41
+; GCN1-NEXT:    s_addc_u32 s13, s13, 0
 ; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN1-NEXT:    s_lshl_b64 s[4:5], s[4:5], 3
-; GCN1-NEXT:    s_add_u32 s0, s0, s4
-; GCN1-NEXT:    s_addc_u32 s1, s1, s5
+; GCN1-NEXT:    s_lshl_b64 s[0:1], s[8:9], 3
+; GCN1-NEXT:    s_add_u32 s0, s4, s0
+; GCN1-NEXT:    s_addc_u32 s1, s5, s1
 ; GCN1-NEXT:    s_add_u32 s0, s0, 32
 ; GCN1-NEXT:    s_addc_u32 s1, s1, 0
+; GCN1-NEXT:    s_cmp_eq_u32 s1, s2
+; GCN1-NEXT:    s_cselect_b64 s[2:3], -1, 0
+; GCN1-NEXT:    s_andn2_b64 vcc, exec, s[2:3]
+; GCN1-NEXT:    s_mov_b64 s[2:3], -1
+; GCN1-NEXT:    s_cbranch_vccnz .LBB93_3
+; GCN1-NEXT:  ; %bb.1: ; %Flow
+; GCN1-NEXT:    s_andn2_b64 vcc, exec, s[2:3]
+; GCN1-NEXT:    s_cbranch_vccz .LBB93_4
+; GCN1-NEXT:  .LBB93_2: ; %atomicrmw.phi
+; GCN1-NEXT:    s_endpgm
+; GCN1-NEXT:  .LBB93_3: ; %atomicrmw.global
 ; GCN1-NEXT:    v_mov_b32_e32 v5, s1
-; GCN1-NEXT:    v_mov_b32_e32 v0, s2
-; GCN1-NEXT:    v_mov_b32_e32 v1, s3
-; GCN1-NEXT:    v_mov_b32_e32 v2, s6
-; GCN1-NEXT:    v_mov_b32_e32 v3, s7
+; GCN1-NEXT:    v_mov_b32_e32 v0, s6
+; GCN1-NEXT:    v_mov_b32_e32 v1, s7
+; GCN1-NEXT:    v_mov_b32_e32 v2, s10
+; GCN1-NEXT:    v_mov_b32_e32 v3, s11
 ; GCN1-NEXT:    v_mov_b32_e32 v4, s0
 ; GCN1-NEXT:    flat_atomic_cmpswap_x2 v[4:5], v[0:3]
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    s_cbranch_execnz .LBB93_2
+; GCN1-NEXT:  .LBB93_4: ; %atomicrmw.private
+; GCN1-NEXT:    v_cmp_ne_u64_e64 s[2:3], s[0:1], 0
+; GCN1-NEXT:    v_mov_b32_e32 v5, s6
+; GCN1-NEXT:    s_and_b64 s[2:3], s[2:3], exec
+; GCN1-NEXT:    s_cselect_b32 s0, s0, -1
+; GCN1-NEXT:    v_mov_b32_e32 v2, s0
+; GCN1-NEXT:    s_add_i32 s0, s0, 4
+; GCN1-NEXT:    v_mov_b32_e32 v3, s0
+; GCN1-NEXT:    buffer_load_dword v0, v2, s[12:15], 0 offen
+; GCN1-NEXT:    buffer_load_dword v1, v3, s[12:15], 0 offen
+; GCN1-NEXT:    v_mov_b32_e32 v4, s7
+; GCN1-NEXT:    s_waitcnt vmcnt(0)
+; GCN1-NEXT:    v_cmp_eq_u64_e32 vcc, s[10:11], v[0:1]
+; GCN1-NEXT:    v_cndmask_b32_e32 v0, v0, v5, vcc
+; GCN1-NEXT:    v_cndmask_b32_e32 v1, v1, v4, vcc
+; GCN1-NEXT:    buffer_store_dword v0, v2, s[12:15], 0 offen
+; GCN1-NEXT:    buffer_store_dword v1, v3, s[12:15], 0 offen
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_cmpxchg_i64_addr64_offset:
 ; GCN2:       ; %bb.0: ; %entry
-; GCN2-NEXT:    s_load_dwordx8 s[0:7], s[2:3], 0x24
+; GCN2-NEXT:    s_mov_b32 s88, SCRATCH_RSRC_DWORD0
+; GCN2-NEXT:    s_mov_b32 s89, SCRATCH_RSRC_DWORD1
+; GCN2-NEXT:    s_mov_b32 s90, -1
+; GCN2-NEXT:    s_mov_b32 s91, 0xe80000
+; GCN2-NEXT:    s_add_u32 s88, s88, s9
+; GCN2-NEXT:    s_load_dwordx8 s[4:11], s[2:3], 0x24
+; GCN2-NEXT:    s_load_dword s2, s[2:3], 0x104
+; GCN2-NEXT:    s_addc_u32 s89, s89, 0
 ; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN2-NEXT:    s_lshl_b64 s[4:5], s[4:5], 3
-; GCN2-NEXT:    s_add_u32 s0, s0, s4
-; GCN2-NEXT:    s_addc_u32 s1, s1, s5
+; GCN2-NEXT:    s_lshl_b64 s[0:1], s[8:9], 3
+; GCN2-NEXT:    s_add_u32 s0, s4, s0
+; GCN2-NEXT:    s_addc_u32 s1, s5, s1
 ; GCN2-NEXT:    s_add_u32 s0, s0, 32
 ; GCN2-NEXT:    s_addc_u32 s1, s1, 0
+; GCN2-NEXT:    s_cmp_eq_u32 s1, s2
+; GCN2-NEXT:    s_cselect_b64 s[2:3], -1, 0
+; GCN2-NEXT:    s_andn2_b64 vcc, exec, s[2:3]
+; GCN2-NEXT:    s_mov_b64 s[2:3], -1
+; GCN2-NEXT:    s_cbranch_vccnz .LBB93_3
+; GCN2-NEXT:  ; %bb.1: ; %Flow
+; GCN2-NEXT:    s_andn2_b64 vcc, exec, s[2:3]
+; GCN2-NEXT:    s_cbranch_vccz .LBB93_4
+; GCN2-NEXT:  .LBB93_2: ; %atomicrmw.phi
+; GCN2-NEXT:    s_endpgm
+; GCN2-NEXT:  .LBB93_3: ; %atomicrmw.global
 ; GCN2-NEXT:    v_mov_b32_e32 v5, s1
-; GCN2-NEXT:    v_mov_b32_e32 v0, s2
-; GCN2-NEXT:    v_mov_b32_e32 v1, s3
-; GCN2-NEXT:    v_mov_b32_e32 v2, s6
-; GCN2-NEXT:    v_mov_b32_e32 v3, s7
+; GCN2-NEXT:    v_mov_b32_e32 v0, s6
+; GCN2-NEXT:    v_mov_b32_e32 v1, s7
+; GCN2-NEXT:    v_mov_b32_e32 v2, s10
+; GCN2-NEXT:    v_mov_b32_e32 v3, s11
 ; GCN2-NEXT:    v_mov_b32_e32 v4, s0
 ; GCN2-NEXT:    flat_atomic_cmpswap_x2 v[4:5], v[0:3]
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    s_cbranch_execnz .LBB93_2
+; GCN2-NEXT:  .LBB93_4: ; %atomicrmw.private
+; GCN2-NEXT:    s_cmp_lg_u64 s[0:1], 0
+; GCN2-NEXT:    s_cselect_b32 s0, s0, -1
+; GCN2-NEXT:    v_mov_b32_e32 v2, s0
+; GCN2-NEXT:    s_add_i32 s0, s0, 4
+; GCN2-NEXT:    v_mov_b32_e32 v3, s0
+; GCN2-NEXT:    buffer_load_dword v0, v2, s[88:91], 0 offen
+; GCN2-NEXT:    buffer_load_dword v1, v3, s[88:91], 0 offen
+; GCN2-NEXT:    v_mov_b32_e32 v5, s6
+; GCN2-NEXT:    v_mov_b32_e32 v4, s7
+; GCN2-NEXT:    s_waitcnt vmcnt(0)
+; GCN2-NEXT:    v_cmp_eq_u64_e32 vcc, s[10:11], v[0:1]
+; GCN2-NEXT:    v_cndmask_b32_e32 v0, v0, v5, vcc
+; GCN2-NEXT:    v_cndmask_b32_e32 v1, v1, v4, vcc
+; GCN2-NEXT:    buffer_store_dword v0, v2, s[88:91], 0 offen
+; GCN2-NEXT:    buffer_store_dword v1, v3, s[88:91], 0 offen
 ; GCN2-NEXT:    s_endpgm
 ;
 ; GFX12-LABEL: atomic_cmpxchg_i64_addr64_offset:
@@ -12449,13 +12802,39 @@ define amdgpu_kernel void @atomic_cmpxchg_i64_addr64_offset(ptr %out, i64 %in, i
 ; GFX12-NEXT:    s_load_b256 s[0:7], s[2:3], 0x24
 ; GFX12-NEXT:    s_wait_kmcnt 0x0
 ; GFX12-NEXT:    s_lshl_b64 s[4:5], s[4:5], 3
-; GFX12-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX12-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
 ; GFX12-NEXT:    s_add_nc_u64 s[0:1], s[0:1], s[4:5]
+; GFX12-NEXT:    s_mov_b64 s[4:5], src_private_base
+; GFX12-NEXT:    s_add_nc_u64 s[0:1], s[0:1], 32
+; GFX12-NEXT:    s_cmp_eq_u32 s1, s5
+; GFX12-NEXT:    s_cselect_b32 s4, -1, 0
+; GFX12-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
+; GFX12-NEXT:    s_mov_b32 s4, -1
+; GFX12-NEXT:    s_cbranch_vccnz .LBB93_3
+; GFX12-NEXT:  ; %bb.1: ; %Flow
+; GFX12-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
+; GFX12-NEXT:    s_cbranch_vccz .LBB93_4
+; GFX12-NEXT:  .LBB93_2: ; %atomicrmw.phi
+; GFX12-NEXT:    s_endpgm
+; GFX12-NEXT:  .LBB93_3: ; %atomicrmw.global
+; GFX12-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX12-NEXT:    v_dual_mov_b32 v2, s6 :: v_dual_mov_b32 v3, s7
 ; GFX12-NEXT:    v_dual_mov_b32 v5, s1 :: v_dual_mov_b32 v4, s0
-; GFX12-NEXT:    flat_atomic_cmpswap_b64 v[4:5], v[0:3] offset:32 scope:SCOPE_DEV
+; GFX12-NEXT:    flat_atomic_cmpswap_b64 v[4:5], v[0:3] scope:SCOPE_DEV
 ; GFX12-NEXT:    s_wait_storecnt_dscnt 0x0
 ; GFX12-NEXT:    global_inv scope:SCOPE_DEV
+; GFX12-NEXT:    s_cbranch_execnz .LBB93_2
+; GFX12-NEXT:  .LBB93_4: ; %atomicrmw.private
+; GFX12-NEXT:    s_wait_alu 0xfffe
+; GFX12-NEXT:    s_cmp_lg_u64 s[0:1], 0
+; GFX12-NEXT:    s_cselect_b32 s0, s0, -1
+; GFX12-NEXT:    scratch_load_b64 v[0:1], off, s0
+; GFX12-NEXT:    s_wait_loadcnt 0x0
+; GFX12-NEXT:    v_cmp_eq_u64_e32 vcc_lo, s[6:7], v[0:1]
+; GFX12-NEXT:    v_cndmask_b32_e64 v1, v1, s3, vcc_lo
+; GFX12-NEXT:    v_cndmask_b32_e64 v0, v0, s2, vcc_lo
+; GFX12-NEXT:    scratch_store_b64 off, v[0:1], s0
 ; GFX12-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i64, ptr %out, i64 %index
@@ -12467,23 +12846,57 @@ entry:
 define amdgpu_kernel void @atomic_cmpxchg_i64_ret_addr64_offset(ptr %out, ptr %out2, i64 %in, i64 %index, i64 %old) {
 ; GCN1-LABEL: atomic_cmpxchg_i64_ret_addr64_offset:
 ; GCN1:       ; %bb.0: ; %entry
+; GCN1-NEXT:    s_mov_b32 s16, SCRATCH_RSRC_DWORD0
+; GCN1-NEXT:    s_mov_b32 s17, SCRATCH_RSRC_DWORD1
+; GCN1-NEXT:    s_mov_b32 s18, -1
+; GCN1-NEXT:    s_mov_b32 s19, 0xe8f000
+; GCN1-NEXT:    s_add_u32 s16, s16, s9
 ; GCN1-NEXT:    s_load_dwordx8 s[4:11], s[2:3], 0x9
 ; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x11
+; GCN1-NEXT:    s_load_dword s12, s[2:3], 0x43
+; GCN1-NEXT:    s_addc_u32 s17, s17, 0
 ; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    s_lshl_b64 s[2:3], s[10:11], 3
-; GCN1-NEXT:    v_mov_b32_e32 v2, s0
-; GCN1-NEXT:    s_add_u32 s0, s4, s2
+; GCN1-NEXT:    s_add_u32 s2, s4, s2
 ; GCN1-NEXT:    s_addc_u32 s3, s5, s3
-; GCN1-NEXT:    s_add_u32 s2, s0, 32
+; GCN1-NEXT:    s_add_u32 s2, s2, 32
 ; GCN1-NEXT:    s_addc_u32 s3, s3, 0
+; GCN1-NEXT:    s_cmp_eq_u32 s3, s12
+; GCN1-NEXT:    s_cselect_b64 s[4:5], -1, 0
+; GCN1-NEXT:    s_andn2_b64 vcc, exec, s[4:5]
+; GCN1-NEXT:    s_cbranch_vccz .LBB94_2
+; GCN1-NEXT:  ; %bb.1: ; %atomicrmw.global
 ; GCN1-NEXT:    v_mov_b32_e32 v5, s3
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s8
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s9
+; GCN1-NEXT:    v_mov_b32_e32 v2, s0
 ; GCN1-NEXT:    v_mov_b32_e32 v3, s1
 ; GCN1-NEXT:    v_mov_b32_e32 v4, s2
 ; GCN1-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    s_cbranch_execz .LBB94_3
+; GCN1-NEXT:    s_branch .LBB94_4
+; GCN1-NEXT:  .LBB94_2:
+; GCN1-NEXT:    ; implicit-def: $vgpr0_vgpr1
+; GCN1-NEXT:  .LBB94_3: ; %atomicrmw.private
+; GCN1-NEXT:    v_cmp_ne_u64_e64 s[4:5], s[2:3], 0
+; GCN1-NEXT:    v_mov_b32_e32 v5, s8
+; GCN1-NEXT:    s_and_b64 s[4:5], s[4:5], exec
+; GCN1-NEXT:    s_cselect_b32 s2, s2, -1
+; GCN1-NEXT:    v_mov_b32_e32 v2, s2
+; GCN1-NEXT:    s_add_i32 s2, s2, 4
+; GCN1-NEXT:    v_mov_b32_e32 v3, s2
+; GCN1-NEXT:    buffer_load_dword v0, v2, s[16:19], 0 offen
+; GCN1-NEXT:    buffer_load_dword v1, v3, s[16:19], 0 offen
+; GCN1-NEXT:    v_mov_b32_e32 v4, s9
+; GCN1-NEXT:    s_waitcnt vmcnt(0)
+; GCN1-NEXT:    v_cmp_eq_u64_e32 vcc, s[0:1], v[0:1]
+; GCN1-NEXT:    v_cndmask_b32_e32 v5, v0, v5, vcc
+; GCN1-NEXT:    v_cndmask_b32_e32 v4, v1, v4, vcc
+; GCN1-NEXT:    buffer_store_dword v5, v2, s[16:19], 0 offen
+; GCN1-NEXT:    buffer_store_dword v4, v3, s[16:19], 0 offen
+; GCN1-NEXT:  .LBB94_4: ; %atomicrmw.end
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s6
 ; GCN1-NEXT:    v_mov_b32_e32 v3, s7
 ; GCN1-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
@@ -12491,23 +12904,56 @@ define amdgpu_kernel void @atomic_cmpxchg_i64_ret_addr64_offset(ptr %out, ptr %o
 ;
 ; GCN2-LABEL: atomic_cmpxchg_i64_ret_addr64_offset:
 ; GCN2:       ; %bb.0: ; %entry
+; GCN2-NEXT:    s_mov_b32 s88, SCRATCH_RSRC_DWORD0
+; GCN2-NEXT:    s_mov_b32 s89, SCRATCH_RSRC_DWORD1
+; GCN2-NEXT:    s_mov_b32 s90, -1
+; GCN2-NEXT:    s_mov_b32 s91, 0xe80000
+; GCN2-NEXT:    s_add_u32 s88, s88, s9
 ; GCN2-NEXT:    s_load_dwordx8 s[4:11], s[2:3], 0x24
 ; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x44
+; GCN2-NEXT:    s_load_dword s12, s[2:3], 0x10c
+; GCN2-NEXT:    s_addc_u32 s89, s89, 0
 ; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    s_lshl_b64 s[2:3], s[10:11], 3
-; GCN2-NEXT:    v_mov_b32_e32 v2, s0
-; GCN2-NEXT:    s_add_u32 s0, s4, s2
+; GCN2-NEXT:    s_add_u32 s2, s4, s2
 ; GCN2-NEXT:    s_addc_u32 s3, s5, s3
-; GCN2-NEXT:    s_add_u32 s2, s0, 32
+; GCN2-NEXT:    s_add_u32 s2, s2, 32
 ; GCN2-NEXT:    s_addc_u32 s3, s3, 0
+; GCN2-NEXT:    s_cmp_eq_u32 s3, s12
+; GCN2-NEXT:    s_cselect_b64 s[4:5], -1, 0
+; GCN2-NEXT:    s_andn2_b64 vcc, exec, s[4:5]
+; GCN2-NEXT:    s_cbranch_vccz .LBB94_2
+; GCN2-NEXT:  ; %bb.1: ; %atomicrmw.global
 ; GCN2-NEXT:    v_mov_b32_e32 v5, s3
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s8
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s9
+; GCN2-NEXT:    v_mov_b32_e32 v2, s0
 ; GCN2-NEXT:    v_mov_b32_e32 v3, s1
 ; GCN2-NEXT:    v_mov_b32_e32 v4, s2
 ; GCN2-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    s_cbranch_execz .LBB94_3
+; GCN2-NEXT:    s_branch .LBB94_4
+; GCN2-NEXT:  .LBB94_2:
+; GCN2-NEXT:    ; implicit-def: $vgpr0_vgpr1
+; GCN2-NEXT:  .LBB94_3: ; %atomicrmw.private
+; GCN2-NEXT:    s_cmp_lg_u64 s[2:3], 0
+; GCN2-NEXT:    s_cselect_b32 s2, s2, -1
+; GCN2-NEXT:    v_mov_b32_e32 v2, s2
+; GCN2-NEXT:    s_add_i32 s2, s2, 4
+; GCN2-NEXT:    v_mov_b32_e32 v3, s2
+; GCN2-NEXT:    buffer_load_dword v0, v2, s[88:91], 0 offen
+; GCN2-NEXT:    buffer_load_dword v1, v3, s[88:91], 0 offen
+; GCN2-NEXT:    v_mov_b32_e32 v5, s8
+; GCN2-NEXT:    v_mov_b32_e32 v4, s9
+; GCN2-NEXT:    s_waitcnt vmcnt(0)
+; GCN2-NEXT:    v_cmp_eq_u64_e32 vcc, s[0:1], v[0:1]
+; GCN2-NEXT:    v_cndmask_b32_e32 v5, v0, v5, vcc
+; GCN2-NEXT:    v_cndmask_b32_e32 v4, v1, v4, vcc
+; GCN2-NEXT:    buffer_store_dword v5, v2, s[88:91], 0 offen
+; GCN2-NEXT:    buffer_store_dword v4, v3, s[88:91], 0 offen
+; GCN2-NEXT:  .LBB94_4: ; %atomicrmw.end
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s6
 ; GCN2-NEXT:    v_mov_b32_e32 v3, s7
 ; GCN2-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
@@ -12519,14 +12965,38 @@ define amdgpu_kernel void @atomic_cmpxchg_i64_ret_addr64_offset(ptr %out, ptr %o
 ; GFX12-NEXT:    s_load_b256 s[4:11], s[2:3], 0x24
 ; GFX12-NEXT:    s_load_b64 s[0:1], s[2:3], 0x44
 ; GFX12-NEXT:    s_wait_kmcnt 0x0
-; GFX12-NEXT:    s_lshl_b64 s[2:3], s[10:11], 3
+; GFX12-NEXT:    s_lshl_b64 s[10:11], s[10:11], 3
+; GFX12-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
+; GFX12-NEXT:    s_add_nc_u64 s[2:3], s[4:5], s[10:11]
+; GFX12-NEXT:    s_mov_b64 s[4:5], src_private_base
+; GFX12-NEXT:    s_add_nc_u64 s[2:3], s[2:3], 32
+; GFX12-NEXT:    s_cmp_eq_u32 s3, s5
+; GFX12-NEXT:    s_cselect_b32 s4, -1, 0
+; GFX12-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
+; GFX12-NEXT:    s_cbranch_vccz .LBB94_2
+; GFX12-NEXT:  ; %bb.1: ; %atomicrmw.global
 ; GFX12-NEXT:    v_dual_mov_b32 v0, s8 :: v_dual_mov_b32 v1, s9
-; GFX12-NEXT:    s_add_nc_u64 s[2:3], s[4:5], s[2:3]
 ; GFX12-NEXT:    v_dual_mov_b32 v2, s0 :: v_dual_mov_b32 v3, s1
 ; GFX12-NEXT:    v_dual_mov_b32 v5, s3 :: v_dual_mov_b32 v4, s2
-; GFX12-NEXT:    flat_atomic_cmpswap_b64 v[0:1], v[4:5], v[0:3] offset:32 th:TH_ATOMIC_RETURN scope:SCOPE_DEV
+; GFX12-NEXT:    flat_atomic_cmpswap_b64 v[0:1], v[4:5], v[0:3] th:TH_ATOMIC_RETURN scope:SCOPE_DEV
 ; GFX12-NEXT:    s_wait_loadcnt_dscnt 0x0
 ; GFX12-NEXT:    global_inv scope:SCOPE_DEV
+; GFX12-NEXT:    s_cbranch_execz .LBB94_3
+; GFX12-NEXT:    s_branch .LBB94_4
+; GFX12-NEXT:  .LBB94_2:
+; GFX12-NEXT:    ; implicit-def: $vgpr0_vgpr1
+; GFX12-NEXT:  .LBB94_3: ; %atomicrmw.private
+; GFX12-NEXT:    s_wait_alu 0xfffe
+; GFX12-NEXT:    s_cmp_lg_u64 s[2:3], 0
+; GFX12-NEXT:    s_cselect_b32 s2, s2, -1
+; GFX12-NEXT:    scratch_load_b64 v[0:1], off, s2
+; GFX12-NEXT:    s_wait_loadcnt 0x0
+; GFX12-NEXT:    v_cmp_eq_u64_e32 vcc_lo, s[0:1], v[0:1]
+; GFX12-NEXT:    v_cndmask_b32_e64 v3, v1, s9, vcc_lo
+; GFX12-NEXT:    v_cndmask_b32_e64 v2, v0, s8, vcc_lo
+; GFX12-NEXT:    scratch_store_b64 off, v[2:3], s2
+; GFX12-NEXT:  .LBB94_4: ; %atomicrmw.end
 ; GFX12-NEXT:    v_dual_mov_b32 v2, s6 :: v_dual_mov_b32 v3, s7
 ; GFX12-NEXT:    flat_store_b64 v[2:3], v[0:1]
 ; GFX12-NEXT:    s_endpgm
@@ -12542,34 +13012,105 @@ entry:
 define amdgpu_kernel void @atomic_cmpxchg_i64(ptr %out, i64 %in, i64 %old) {
 ; GCN1-LABEL: atomic_cmpxchg_i64:
 ; GCN1:       ; %bb.0: ; %entry
+; GCN1-NEXT:    s_mov_b32 s12, SCRATCH_RSRC_DWORD0
+; GCN1-NEXT:    s_mov_b32 s13, SCRATCH_RSRC_DWORD1
+; GCN1-NEXT:    s_mov_b32 s14, -1
+; GCN1-NEXT:    s_load_dword s8, s[2:3], 0x3f
 ; GCN1-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x9
 ; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0xd
+; GCN1-NEXT:    s_mov_b32 s15, 0xe8f000
+; GCN1-NEXT:    s_add_u32 s12, s12, s9
+; GCN1-NEXT:    s_addc_u32 s13, s13, 0
 ; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN1-NEXT:    s_cmp_eq_u32 s5, s8
+; GCN1-NEXT:    s_cselect_b64 s[2:3], -1, 0
+; GCN1-NEXT:    s_andn2_b64 vcc, exec, s[2:3]
+; GCN1-NEXT:    s_mov_b64 s[2:3], -1
+; GCN1-NEXT:    s_cbranch_vccnz .LBB95_3
+; GCN1-NEXT:  ; %bb.1: ; %Flow
+; GCN1-NEXT:    s_andn2_b64 vcc, exec, s[2:3]
+; GCN1-NEXT:    s_cbranch_vccz .LBB95_4
+; GCN1-NEXT:  .LBB95_2: ; %atomicrmw.phi
+; GCN1-NEXT:    s_endpgm
+; GCN1-NEXT:  .LBB95_3: ; %atomicrmw.global
 ; GCN1-NEXT:    v_mov_b32_e32 v4, s4
-; GCN1-NEXT:    v_mov_b32_e32 v5, s5
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s6
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s7
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s0
 ; GCN1-NEXT:    v_mov_b32_e32 v3, s1
+; GCN1-NEXT:    v_mov_b32_e32 v5, s5
 ; GCN1-NEXT:    flat_atomic_cmpswap_x2 v[4:5], v[0:3]
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    s_cbranch_execnz .LBB95_2
+; GCN1-NEXT:  .LBB95_4: ; %atomicrmw.private
+; GCN1-NEXT:    v_cmp_ne_u64_e64 s[2:3], s[4:5], 0
+; GCN1-NEXT:    v_mov_b32_e32 v5, s6
+; GCN1-NEXT:    s_and_b64 s[2:3], s[2:3], exec
+; GCN1-NEXT:    s_cselect_b32 s2, s4, -1
+; GCN1-NEXT:    v_mov_b32_e32 v2, s2
+; GCN1-NEXT:    s_add_i32 s2, s2, 4
+; GCN1-NEXT:    v_mov_b32_e32 v3, s2
+; GCN1-NEXT:    buffer_load_dword v0, v2, s[12:15], 0 offen
+; GCN1-NEXT:    buffer_load_dword v1, v3, s[12:15], 0 offen
+; GCN1-NEXT:    v_mov_b32_e32 v4, s7
+; GCN1-NEXT:    s_waitcnt vmcnt(0)
+; GCN1-NEXT:    v_cmp_eq_u64_e32 vcc, s[0:1], v[0:1]
+; GCN1-NEXT:    v_cndmask_b32_e32 v0, v0, v5, vcc
+; GCN1-NEXT:    v_cndmask_b32_e32 v1, v1, v4, vcc
+; GCN1-NEXT:    buffer_store_dword v0, v2, s[12:15], 0 offen
+; GCN1-NEXT:    buffer_store_dword v1, v3, s[12:15], 0 offen
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_cmpxchg_i64:
 ; GCN2:       ; %bb.0: ; %entry
+; GCN2-NEXT:    s_mov_b32 s88, SCRATCH_RSRC_DWORD0
+; GCN2-NEXT:    s_mov_b32 s89, SCRATCH_RSRC_DWORD1
+; GCN2-NEXT:    s_mov_b32 s90, -1
+; GCN2-NEXT:    s_load_dword s8, s[2:3], 0xfc
 ; GCN2-NEXT:    s_load_dwordx4 s[4:7], s[2:3], 0x24
 ; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x34
+; GCN2-NEXT:    s_mov_b32 s91, 0xe80000
+; GCN2-NEXT:    s_add_u32 s88, s88, s9
+; GCN2-NEXT:    s_addc_u32 s89, s89, 0
 ; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
+; GCN2-NEXT:    s_cmp_eq_u32 s5, s8
+; GCN2-NEXT:    s_cselect_b64 s[2:3], -1, 0
+; GCN2-NEXT:    s_andn2_b64 vcc, exec, s[2:3]
+; GCN2-NEXT:    s_mov_b64 s[2:3], -1
+; GCN2-NEXT:    s_cbranch_vccnz .LBB95_3
+; GCN2-NEXT:  ; %bb.1: ; %Flow
+; GCN2-NEXT:    s_andn2_b64 vcc, exec, s[2:3]
+; GCN2-NEXT:    s_cbranch_vccz .LBB95_4
+; GCN2-NEXT:  .LBB95_2: ; %atomicrmw.phi
+; GCN2-NEXT:    s_endpgm
+; GCN2-NEXT:  .LBB95_3: ; %atomicrmw.global
 ; GCN2-NEXT:    v_mov_b32_e32 v4, s4
-; GCN2-NEXT:    v_mov_b32_e32 v5, s5
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s6
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s7
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s0
 ; GCN2-NEXT:    v_mov_b32_e32 v3, s1
+; GCN2-NEXT:    v_mov_b32_e32 v5, s5
 ; GCN2-NEXT:    flat_atomic_cmpswap_x2 v[4:5], v[0:3]
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    s_cbranch_execnz .LBB95_2
+; GCN2-NEXT:  .LBB95_4: ; %atomicrmw.private
+; GCN2-NEXT:    s_cmp_lg_u64 s[4:5], 0
+; GCN2-NEXT:    s_cselect_b32 s2, s4, -1
+; GCN2-NEXT:    v_mov_b32_e32 v2, s2
+; GCN2-NEXT:    s_add_i32 s2, s2, 4
+; GCN2-NEXT:    v_mov_b32_e32 v3, s2
+; GCN2-NEXT:    buffer_load_dword v0, v2, s[88:91], 0 offen
+; GCN2-NEXT:    buffer_load_dword v1, v3, s[88:91], 0 offen
+; GCN2-NEXT:    v_mov_b32_e32 v5, s6
+; GCN2-NEXT:    v_mov_b32_e32 v4, s7
+; GCN2-NEXT:    s_waitcnt vmcnt(0)
+; GCN2-NEXT:    v_cmp_eq_u64_e32 vcc, s[0:1], v[0:1]
+; GCN2-NEXT:    v_cndmask_b32_e32 v0, v0, v5, vcc
+; GCN2-NEXT:    v_cndmask_b32_e32 v1, v1, v4, vcc
+; GCN2-NEXT:    buffer_store_dword v0, v2, s[88:91], 0 offen
+; GCN2-NEXT:    buffer_store_dword v1, v3, s[88:91], 0 offen
 ; GCN2-NEXT:    s_endpgm
 ;
 ; GFX12-LABEL: atomic_cmpxchg_i64:
@@ -12577,13 +13118,36 @@ define amdgpu_kernel void @atomic_cmpxchg_i64(ptr %out, i64 %in, i64 %old) {
 ; GFX12-NEXT:    s_clause 0x1
 ; GFX12-NEXT:    s_load_b128 s[4:7], s[2:3], 0x24
 ; GFX12-NEXT:    s_load_b64 s[0:1], s[2:3], 0x34
+; GFX12-NEXT:    s_mov_b64 s[2:3], src_private_base
 ; GFX12-NEXT:    s_wait_kmcnt 0x0
-; GFX12-NEXT:    v_dual_mov_b32 v4, s4 :: v_dual_mov_b32 v5, s5
+; GFX12-NEXT:    s_cmp_eq_u32 s5, s3
+; GFX12-NEXT:    s_cselect_b32 s2, -1, 0
+; GFX12-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s2
+; GFX12-NEXT:    s_mov_b32 s2, -1
+; GFX12-NEXT:    s_cbranch_vccnz .LBB95_3
+; GFX12-NEXT:  ; %bb.1: ; %Flow
+; GFX12-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s2
+; GFX12-NEXT:    s_cbranch_vccz .LBB95_4
+; GFX12-NEXT:  .LBB95_2: ; %atomicrmw.phi
+; GFX12-NEXT:    s_endpgm
+; GFX12-NEXT:  .LBB95_3: ; %atomicrmw.global
 ; GFX12-NEXT:    v_dual_mov_b32 v0, s6 :: v_dual_mov_b32 v1, s7
 ; GFX12-NEXT:    v_dual_mov_b32 v2, s0 :: v_dual_mov_b32 v3, s1
+; GFX12-NEXT:    v_dual_mov_b32 v4, s4 :: v_dual_mov_b32 v5, s5
 ; GFX12-NEXT:    flat_atomic_cmpswap_b64 v[4:5], v[0:3] scope:SCOPE_DEV
 ; GFX12-NEXT:    s_wait_storecnt_dscnt 0x0
 ; GFX12-NEXT:    global_inv scope:SCOPE_DEV
+; GFX12-NEXT:    s_cbranch_execnz .LBB95_2
+; GFX12-NEXT:  .LBB95_4: ; %atomicrmw.private
+; GFX12-NEXT:    s_cmp_lg_u64 s[4:5], 0
+; GFX12-NEXT:    s_cselect_b32 s2, s4, -1
+; GFX12-NEXT:    scratch_load_b64 v[0:1], off, s2
+; GFX12-NEXT:    s_wait_loadcnt 0x0
+; GFX12-NEXT:    v_cmp_eq_u64_e32 vcc_lo, s[0:1], v[0:1]
+; GFX12-NEXT:    v_cndmask_b32_e64 v1, v1, s7, vcc_lo
+; GFX12-NEXT:    v_cndmask_b32_e64 v0, v0, s6, vcc_lo
+; GFX12-NEXT:    scratch_store_b64 off, v[0:1], s2
 ; GFX12-NEXT:    s_endpgm
 entry:
   %val = cmpxchg volatile ptr %out, i64 %old, i64 %in syncscope("agent") seq_cst seq_cst
@@ -12593,50 +13157,138 @@ entry:
 define amdgpu_kernel void @atomic_cmpxchg_i64_ret(ptr %out, ptr %out2, i64 %in, i64 %old) {
 ; GCN1-LABEL: atomic_cmpxchg_i64_ret:
 ; GCN1:       ; %bb.0: ; %entry
-; GCN1-NEXT:    s_load_dwordx8 s[0:7], s[2:3], 0x9
+; GCN1-NEXT:    s_mov_b32 s12, SCRATCH_RSRC_DWORD0
+; GCN1-NEXT:    s_mov_b32 s13, SCRATCH_RSRC_DWORD1
+; GCN1-NEXT:    s_mov_b32 s14, -1
+; GCN1-NEXT:    s_mov_b32 s15, 0xe8f000
+; GCN1-NEXT:    s_add_u32 s12, s12, s9
+; GCN1-NEXT:    s_load_dwordx8 s[4:11], s[2:3], 0x9
+; GCN1-NEXT:    s_load_dword s0, s[2:3], 0x41
+; GCN1-NEXT:    s_addc_u32 s13, s13, 0
 ; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN1-NEXT:    v_mov_b32_e32 v4, s0
-; GCN1-NEXT:    v_mov_b32_e32 v5, s1
-; GCN1-NEXT:    v_mov_b32_e32 v0, s4
-; GCN1-NEXT:    v_mov_b32_e32 v1, s5
-; GCN1-NEXT:    v_mov_b32_e32 v2, s6
-; GCN1-NEXT:    v_mov_b32_e32 v3, s7
+; GCN1-NEXT:    s_cmp_eq_u32 s5, s0
+; GCN1-NEXT:    s_cselect_b64 s[0:1], -1, 0
+; GCN1-NEXT:    s_andn2_b64 vcc, exec, s[0:1]
+; GCN1-NEXT:    s_cbranch_vccz .LBB96_2
+; GCN1-NEXT:  ; %bb.1: ; %atomicrmw.global
+; GCN1-NEXT:    v_mov_b32_e32 v4, s4
+; GCN1-NEXT:    v_mov_b32_e32 v0, s8
+; GCN1-NEXT:    v_mov_b32_e32 v1, s9
+; GCN1-NEXT:    v_mov_b32_e32 v2, s10
+; GCN1-NEXT:    v_mov_b32_e32 v3, s11
+; GCN1-NEXT:    v_mov_b32_e32 v5, s5
 ; GCN1-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN1-NEXT:    buffer_wbinvl1_vol
-; GCN1-NEXT:    v_mov_b32_e32 v2, s2
-; GCN1-NEXT:    v_mov_b32_e32 v3, s3
+; GCN1-NEXT:    s_cbranch_execz .LBB96_3
+; GCN1-NEXT:    s_branch .LBB96_4
+; GCN1-NEXT:  .LBB96_2:
+; GCN1-NEXT:    ; implicit-def: $vgpr0_vgpr1
+; GCN1-NEXT:  .LBB96_3: ; %atomicrmw.private
+; GCN1-NEXT:    v_cmp_ne_u64_e64 s[0:1], s[4:5], 0
+; GCN1-NEXT:    v_mov_b32_e32 v5, s8
+; GCN1-NEXT:    s_and_b64 s[0:1], s[0:1], exec
+; GCN1-NEXT:    s_cselect_b32 s0, s4, -1
+; GCN1-NEXT:    v_mov_b32_e32 v2, s0
+; GCN1-NEXT:    s_add_i32 s0, s0, 4
+; GCN1-NEXT:    v_mov_b32_e32 v3, s0
+; GCN1-NEXT:    buffer_load_dword v0, v2, s[12:15], 0 offen
+; GCN1-NEXT:    buffer_load_dword v1, v3, s[12:15], 0 offen
+; GCN1-NEXT:    v_mov_b32_e32 v4, s9
+; GCN1-NEXT:    s_waitcnt vmcnt(0)
+; GCN1-NEXT:    v_cmp_eq_u64_e32 vcc, s[10:11], v[0:1]
+; GCN1-NEXT:    v_cndmask_b32_e32 v5, v0, v5, vcc
+; GCN1-NEXT:    v_cndmask_b32_e32 v4, v1, v4, vcc
+; GCN1-NEXT:    buffer_store_dword v5, v2, s[12:15], 0 offen
+; GCN1-NEXT:    buffer_store_dword v4, v3, s[12:15], 0 offen
+; GCN1-NEXT:  .LBB96_4: ; %atomicrmw.end
+; GCN1-NEXT:    v_mov_b32_e32 v2, s6
+; GCN1-NEXT:    v_mov_b32_e32 v3, s7
 ; GCN1-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_cmpxchg_i64_ret:
 ; GCN2:       ; %bb.0: ; %entry
-; GCN2-NEXT:    s_load_dwordx8 s[0:7], s[2:3], 0x24
+; GCN2-NEXT:    s_mov_b32 s88, SCRATCH_RSRC_DWORD0
+; GCN2-NEXT:    s_mov_b32 s89, SCRATCH_RSRC_DWORD1
+; GCN2-NEXT:    s_mov_b32 s90, -1
+; GCN2-NEXT:    s_mov_b32 s91, 0xe80000
+; GCN2-NEXT:    s_add_u32 s88, s88, s9
+; GCN2-NEXT:    s_load_dwordx8 s[4:11], s[2:3], 0x24
+; GCN2-NEXT:    s_load_dword s0, s[2:3], 0x104
+; GCN2-NEXT:    s_addc_u32 s89, s89, 0
 ; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN2-NEXT:    v_mov_b32_e32 v4, s0
-; GCN2-NEXT:    v_mov_b32_e32 v5, s1
-; GCN2-NEXT:    v_mov_b32_e32 v0, s4
-; GCN2-NEXT:    v_mov_b32_e32 v1, s5
-; GCN2-NEXT:    v_mov_b32_e32 v2, s6
-; GCN2-NEXT:    v_mov_b32_e32 v3, s7
+; GCN2-NEXT:    s_cmp_eq_u32 s5, s0
+; GCN2-NEXT:    s_cselect_b64 s[0:1], -1, 0
+; GCN2-NEXT:    s_andn2_b64 vcc, exec, s[0:1]
+; GCN2-NEXT:    s_cbranch_vccz .LBB96_2
+; GCN2-NEXT:  ; %bb.1: ; %atomicrmw.global
+; GCN2-NEXT:    v_mov_b32_e32 v4, s4
+; GCN2-NEXT:    v_mov_b32_e32 v0, s8
+; GCN2-NEXT:    v_mov_b32_e32 v1, s9
+; GCN2-NEXT:    v_mov_b32_e32 v2, s10
+; GCN2-NEXT:    v_mov_b32_e32 v3, s11
+; GCN2-NEXT:    v_mov_b32_e32 v5, s5
 ; GCN2-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN2-NEXT:    buffer_wbinvl1_vol
-; GCN2-NEXT:    v_mov_b32_e32 v2, s2
-; GCN2-NEXT:    v_mov_b32_e32 v3, s3
+; GCN2-NEXT:    s_cbranch_execz .LBB96_3
+; GCN2-NEXT:    s_branch .LBB96_4
+; GCN2-NEXT:  .LBB96_2:
+; GCN2-NEXT:    ; implicit-def: $vgpr0_vgpr1
+; GCN2-NEXT:  .LBB96_3: ; %atomicrmw.private
+; GCN2-NEXT:    s_cmp_lg_u64 s[4:5], 0
+; GCN2-NEXT:    s_cselect_b32 s0, s4, -1
+; GCN2-NEXT:    v_mov_b32_e32 v2, s0
+; GCN2-NEXT:    s_add_i32 s0, s0, 4
+; GCN2-NEXT:    v_mov_b32_e32 v3, s0
+; GCN2-NEXT:    buffer_load_dword v0, v2, s[88:91], 0 offen
+; GCN2-NEXT:    buffer_load_dword v1, v3, s[88:91], 0 offen
+; GCN2-NEXT:    v_mov_b32_e32 v5, s8
+; GCN2-NEXT:    v_mov_b32_e32 v4, s9
+; GCN2-NEXT:    s_waitcnt vmcnt(0)
+; GCN2-NEXT:    v_cmp_eq_u64_e32 vcc, s[10:11], v[0:1]
+; GCN2-NEXT:    v_cndmask_b32_e32 v5, v0, v5, vcc
+; GCN2-NEXT:    v_cndmask_b32_e32 v4, v1, v4, vcc
+; GCN2-NEXT:    buffer_store_dword v5, v2, s[88:91], 0 offen
+; GCN2-NEXT:    buffer_store_dword v4, v3, s[88:91], 0 offen
+; GCN2-NEXT:  .LBB96_4: ; %atomicrmw.end
+; GCN2-NEXT:    v_mov_b32_e32 v2, s6
+; GCN2-NEXT:    v_mov_b32_e32 v3, s7
 ; GCN2-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
 ; GCN2-NEXT:    s_endpgm
 ;
 ; GFX12-LABEL: atomic_cmpxchg_i64_ret:
 ; GFX12:       ; %bb.0: ; %entry
 ; GFX12-NEXT:    s_load_b256 s[0:7], s[2:3], 0x24
+; GFX12-NEXT:    s_mov_b64 s[8:9], src_private_base
 ; GFX12-NEXT:    s_wait_kmcnt 0x0
-; GFX12-NEXT:    v_dual_mov_b32 v4, s0 :: v_dual_mov_b32 v5, s1
+; GFX12-NEXT:    s_cmp_eq_u32 s1, s9
+; GFX12-NEXT:    s_cselect_b32 s8, -1, 0
+; GFX12-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s8
+; GFX12-NEXT:    s_cbranch_vccz .LBB96_2
+; GFX12-NEXT:  ; %bb.1: ; %atomicrmw.global
 ; GFX12-NEXT:    v_dual_mov_b32 v0, s4 :: v_dual_mov_b32 v1, s5
 ; GFX12-NEXT:    v_dual_mov_b32 v2, s6 :: v_dual_mov_b32 v3, s7
+; GFX12-NEXT:    v_dual_mov_b32 v5, s1 :: v_dual_mov_b32 v4, s0
 ; GFX12-NEXT:    flat_atomic_cmpswap_b64 v[0:1], v[4:5], v[0:3] th:TH_ATOMIC_RETURN scope:SCOPE_DEV
 ; GFX12-NEXT:    s_wait_loadcnt_dscnt 0x0
 ; GFX12-NEXT:    global_inv scope:SCOPE_DEV
+; GFX12-NEXT:    s_cbranch_execz .LBB96_3
+; GFX12-NEXT:    s_branch .LBB96_4
+; GFX12-NEXT:  .LBB96_2:
+; GFX12-NEXT:    ; implicit-def: $vgpr0_vgpr1
+; GFX12-NEXT:  .LBB96_3: ; %atomicrmw.private
+; GFX12-NEXT:    s_cmp_lg_u64 s[0:1], 0
+; GFX12-NEXT:    s_cselect_b32 s0, s0, -1
+; GFX12-NEXT:    scratch_load_b64 v[0:1], off, s0
+; GFX12-NEXT:    s_wait_loadcnt 0x0
+; GFX12-NEXT:    v_cmp_eq_u64_e32 vcc_lo, s[6:7], v[0:1]
+; GFX12-NEXT:    v_cndmask_b32_e64 v3, v1, s5, vcc_lo
+; GFX12-NEXT:    v_cndmask_b32_e64 v2, v0, s4, vcc_lo
+; GFX12-NEXT:    scratch_store_b64 off, v[2:3], s0
+; GFX12-NEXT:  .LBB96_4: ; %atomicrmw.end
 ; GFX12-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
 ; GFX12-NEXT:    flat_store_b64 v[2:3], v[0:1]
 ; GFX12-NEXT:    s_endpgm
@@ -12650,52 +13302,148 @@ entry:
 define amdgpu_kernel void @atomic_cmpxchg_i64_addr64(ptr %out, i64 %in, i64 %index, i64 %old) {
 ; GCN1-LABEL: atomic_cmpxchg_i64_addr64:
 ; GCN1:       ; %bb.0: ; %entry
-; GCN1-NEXT:    s_load_dwordx8 s[0:7], s[2:3], 0x9
+; GCN1-NEXT:    s_mov_b32 s12, SCRATCH_RSRC_DWORD0
+; GCN1-NEXT:    s_mov_b32 s13, SCRATCH_RSRC_DWORD1
+; GCN1-NEXT:    s_mov_b32 s14, -1
+; GCN1-NEXT:    s_mov_b32 s15, 0xe8f000
+; GCN1-NEXT:    s_add_u32 s12, s12, s9
+; GCN1-NEXT:    s_load_dwordx8 s[4:11], s[2:3], 0x9
+; GCN1-NEXT:    s_load_dword s2, s[2:3], 0x41
+; GCN1-NEXT:    s_addc_u32 s13, s13, 0
 ; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN1-NEXT:    s_lshl_b64 s[4:5], s[4:5], 3
-; GCN1-NEXT:    s_add_u32 s0, s0, s4
-; GCN1-NEXT:    s_addc_u32 s1, s1, s5
+; GCN1-NEXT:    s_lshl_b64 s[0:1], s[8:9], 3
+; GCN1-NEXT:    s_add_u32 s0, s4, s0
+; GCN1-NEXT:    s_addc_u32 s1, s5, s1
+; GCN1-NEXT:    s_cmp_eq_u32 s1, s2
+; GCN1-NEXT:    s_cselect_b64 s[2:3], -1, 0
+; GCN1-NEXT:    s_andn2_b64 vcc, exec, s[2:3]
+; GCN1-NEXT:    s_mov_b64 s[2:3], -1
+; GCN1-NEXT:    s_cbranch_vccnz .LBB97_3
+; GCN1-NEXT:  ; %bb.1: ; %Flow
+; GCN1-NEXT:    s_andn2_b64 vcc, exec, s[2:3]
+; GCN1-NEXT:    s_cbranch_vccz .LBB97_4
+; GCN1-NEXT:  .LBB97_2: ; %atomicrmw.phi
+; GCN1-NEXT:    s_endpgm
+; GCN1-NEXT:  .LBB97_3: ; %atomicrmw.global
 ; GCN1-NEXT:    v_mov_b32_e32 v5, s1
-; GCN1-NEXT:    v_mov_b32_e32 v0, s2
-; GCN1-NEXT:    v_mov_b32_e32 v1, s3
-; GCN1-NEXT:    v_mov_b32_e32 v2, s6
-; GCN1-NEXT:    v_mov_b32_e32 v3, s7
+; GCN1-NEXT:    v_mov_b32_e32 v0, s6
+; GCN1-NEXT:    v_mov_b32_e32 v1, s7
+; GCN1-NEXT:    v_mov_b32_e32 v2, s10
+; GCN1-NEXT:    v_mov_b32_e32 v3, s11
 ; GCN1-NEXT:    v_mov_b32_e32 v4, s0
 ; GCN1-NEXT:    flat_atomic_cmpswap_x2 v[4:5], v[0:3]
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    s_cbranch_execnz .LBB97_2
+; GCN1-NEXT:  .LBB97_4: ; %atomicrmw.private
+; GCN1-NEXT:    v_cmp_ne_u64_e64 s[2:3], s[0:1], 0
+; GCN1-NEXT:    v_mov_b32_e32 v5, s6
+; GCN1-NEXT:    s_and_b64 s[2:3], s[2:3], exec
+; GCN1-NEXT:    s_cselect_b32 s0, s0, -1
+; GCN1-NEXT:    v_mov_b32_e32 v2, s0
+; GCN1-NEXT:    s_add_i32 s0, s0, 4
+; GCN1-NEXT:    v_mov_b32_e32 v3, s0
+; GCN1-NEXT:    buffer_load_dword v0, v2, s[12:15], 0 offen
+; GCN1-NEXT:    buffer_load_dword v1, v3, s[12:15], 0 offen
+; GCN1-NEXT:    v_mov_b32_e32 v4, s7
+; GCN1-NEXT:    s_waitcnt vmcnt(0)
+; GCN1-NEXT:    v_cmp_eq_u64_e32 vcc, s[10:11], v[0:1]
+; GCN1-NEXT:    v_cndmask_b32_e32 v0, v0, v5, vcc
+; GCN1-NEXT:    v_cndmask_b32_e32 v1, v1, v4, vcc
+; GCN1-NEXT:    buffer_store_dword v0, v2, s[12:15], 0 offen
+; GCN1-NEXT:    buffer_store_dword v1, v3, s[12:15], 0 offen
 ; GCN1-NEXT:    s_endpgm
 ;
 ; GCN2-LABEL: atomic_cmpxchg_i64_addr64:
 ; GCN2:       ; %bb.0: ; %entry
-; GCN2-NEXT:    s_load_dwordx8 s[0:7], s[2:3], 0x24
+; GCN2-NEXT:    s_mov_b32 s88, SCRATCH_RSRC_DWORD0
+; GCN2-NEXT:    s_mov_b32 s89, SCRATCH_RSRC_DWORD1
+; GCN2-NEXT:    s_mov_b32 s90, -1
+; GCN2-NEXT:    s_mov_b32 s91, 0xe80000
+; GCN2-NEXT:    s_add_u32 s88, s88, s9
+; GCN2-NEXT:    s_load_dwordx8 s[4:11], s[2:3], 0x24
+; GCN2-NEXT:    s_load_dword s2, s[2:3], 0x104
+; GCN2-NEXT:    s_addc_u32 s89, s89, 0
 ; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
-; GCN2-NEXT:    s_lshl_b64 s[4:5], s[4:5], 3
-; GCN2-NEXT:    s_add_u32 s0, s0, s4
-; GCN2-NEXT:    s_addc_u32 s1, s1, s5
+; GCN2-NEXT:    s_lshl_b64 s[0:1], s[8:9], 3
+; GCN2-NEXT:    s_add_u32 s0, s4, s0
+; GCN2-NEXT:    s_addc_u32 s1, s5, s1
+; GCN2-NEXT:    s_cmp_eq_u32 s1, s2
+; GCN2-NEXT:    s_cselect_b64 s[2:3], -1, 0
+; GCN2-NEXT:    s_andn2_b64 vcc, exec, s[2:3]
+; GCN2-NEXT:    s_mov_b64 s[2:3], -1
+; GCN2-NEXT:    s_cbranch_vccnz .LBB97_3
+; GCN2-NEXT:  ; %bb.1: ; %Flow
+; GCN2-NEXT:    s_andn2_b64 vcc, exec, s[2:3]
+; GCN2-NEXT:    s_cbranch_vccz .LBB97_4
+; GCN2-NEXT:  .LBB97_2: ; %atomicrmw.phi
+; GCN2-NEXT:    s_endpgm
+; GCN2-NEXT:  .LBB97_3: ; %atomicrmw.global
 ; GCN2-NEXT:    v_mov_b32_e32 v5, s1
-; GCN2-NEXT:    v_mov_b32_e32 v0, s2
-; GCN2-NEXT:    v_mov_b32_e32 v1, s3
-; GCN2-NEXT:    v_mov_b32_e32 v2, s6
-; GCN2-NEXT:    v_mov_b32_e32 v3, s7
+; GCN2-NEXT:    v_mov_b32_e32 v0, s6
+; GCN2-NEXT:    v_mov_b32_e32 v1, s7
+; GCN2-NEXT:    v_mov_b32_e32 v2, s10
+; GCN2-NEXT:    v_mov_b32_e32 v3, s11
 ; GCN2-NEXT:    v_mov_b32_e32 v4, s0
 ; GCN2-NEXT:    flat_atomic_cmpswap_x2 v[4:5], v[0:3]
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    s_cbranch_execnz .LBB97_2
+; GCN2-NEXT:  .LBB97_4: ; %atomicrmw.private
+; GCN2-NEXT:    s_cmp_lg_u64 s[0:1], 0
+; GCN2-NEXT:    s_cselect_b32 s0, s0, -1
+; GCN2-NEXT:    v_mov_b32_e32 v2, s0
+; GCN2-NEXT:    s_add_i32 s0, s0, 4
+; GCN2-NEXT:    v_mov_b32_e32 v3, s0
+; GCN2-NEXT:    buffer_load_dword v0, v2, s[88:91], 0 offen
+; GCN2-NEXT:    buffer_load_dword v1, v3, s[88:91], 0 offen
+; GCN2-NEXT:    v_mov_b32_e32 v5, s6
+; GCN2-NEXT:    v_mov_b32_e32 v4, s7
+; GCN2-NEXT:    s_waitcnt vmcnt(0)
+; GCN2-NEXT:    v_cmp_eq_u64_e32 vcc, s[10:11], v[0:1]
+; GCN2-NEXT:    v_cndmask_b32_e32 v0, v0, v5, vcc
+; GCN2-NEXT:    v_cndmask_b32_e32 v1, v1, v4, vcc
+; GCN2-NEXT:    buffer_store_dword v0, v2, s[88:91], 0 offen
+; GCN2-NEXT:    buffer_store_dword v1, v3, s[88:91], 0 offen
 ; GCN2-NEXT:    s_endpgm
 ;
 ; GFX12-LABEL: atomic_cmpxchg_i64_addr64:
 ; GFX12:       ; %bb.0: ; %entry
 ; GFX12-NEXT:    s_load_b256 s[0:7], s[2:3], 0x24
+; GFX12-NEXT:    s_mov_b64 s[8:9], src_private_base
 ; GFX12-NEXT:    s_wait_kmcnt 0x0
 ; GFX12-NEXT:    s_lshl_b64 s[4:5], s[4:5], 3
-; GFX12-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX12-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
 ; GFX12-NEXT:    s_add_nc_u64 s[0:1], s[0:1], s[4:5]
+; GFX12-NEXT:    s_cmp_eq_u32 s1, s9
+; GFX12-NEXT:    s_cselect_b32 s4, -1, 0
+; GFX12-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX12-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
+; GFX12-NEXT:    s_mov_b32 s4, -1
+; GFX12-NEXT:    s_cbranch_vccnz .LBB97_3
+; GFX12-NEXT:  ; %bb.1: ; %Flow
+; GFX12-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
+; GFX12-NEXT:    s_cbranch_vccz .LBB97_4
+; GFX12-NEXT:  .LBB97_2: ; %atomicrmw.phi
+; GFX12-NEXT:    s_endpgm
+; GFX12-NEXT:  .LBB97_3: ; %atomicrmw.global
+; GFX12-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX12-NEXT:    v_dual_mov_b32 v2, s6 :: v_dual_mov_b32 v3, s7
 ; GFX12-NEXT:    v_dual_mov_b32 v5, s1 :: v_dual_mov_b32 v4, s0
 ; GFX12-NEXT:    flat_atomic_cmpswap_b64 v[4:5], v[0:3] scope:SCOPE_DEV
 ; GFX12-NEXT:    s_wait_storecnt_dscnt 0x0
 ; GFX12-NEXT:    global_inv scope:SCOPE_DEV
+; GFX12-NEXT:    s_cbranch_execnz .LBB97_2
+; GFX12-NEXT:  .LBB97_4: ; %atomicrmw.private
+; GFX12-NEXT:    s_wait_alu 0xfffe
+; GFX12-NEXT:    s_cmp_lg_u64 s[0:1], 0
+; GFX12-NEXT:    s_cselect_b32 s0, s0, -1
+; GFX12-NEXT:    scratch_load_b64 v[0:1], off, s0
+; GFX12-NEXT:    s_wait_loadcnt 0x0
+; GFX12-NEXT:    v_cmp_eq_u64_e32 vcc_lo, s[6:7], v[0:1]
+; GFX12-NEXT:    v_cndmask_b32_e64 v1, v1, s3, vcc_lo
+; GFX12-NEXT:    v_cndmask_b32_e64 v0, v0, s2, vcc_lo
+; GFX12-NEXT:    scratch_store_b64 off, v[0:1], s0
 ; GFX12-NEXT:    s_endpgm
 entry:
   %ptr = getelementptr i64, ptr %out, i64 %index
@@ -12706,12 +13454,24 @@ entry:
 define amdgpu_kernel void @atomic_cmpxchg_i64_ret_addr64(ptr %out, ptr %out2, i64 %in, i64 %index, i64 %old) {
 ; GCN1-LABEL: atomic_cmpxchg_i64_ret_addr64:
 ; GCN1:       ; %bb.0: ; %entry
+; GCN1-NEXT:    s_mov_b32 s16, SCRATCH_RSRC_DWORD0
+; GCN1-NEXT:    s_mov_b32 s17, SCRATCH_RSRC_DWORD1
+; GCN1-NEXT:    s_mov_b32 s18, -1
+; GCN1-NEXT:    s_mov_b32 s19, 0xe8f000
+; GCN1-NEXT:    s_add_u32 s16, s16, s9
 ; GCN1-NEXT:    s_load_dwordx8 s[4:11], s[2:3], 0x9
+; GCN1-NEXT:    s_load_dword s12, s[2:3], 0x43
 ; GCN1-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x11
+; GCN1-NEXT:    s_addc_u32 s17, s17, 0
 ; GCN1-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN1-NEXT:    s_lshl_b64 s[2:3], s[10:11], 3
 ; GCN1-NEXT:    s_add_u32 s2, s4, s2
 ; GCN1-NEXT:    s_addc_u32 s3, s5, s3
+; GCN1-NEXT:    s_cmp_eq_u32 s3, s12
+; GCN1-NEXT:    s_cselect_b64 s[4:5], -1, 0
+; GCN1-NEXT:    s_andn2_b64 vcc, exec, s[4:5]
+; GCN1-NEXT:    s_cbranch_vccz .LBB98_2
+; GCN1-NEXT:  ; %bb.1: ; %atomicrmw.global
 ; GCN1-NEXT:    v_mov_b32_e32 v5, s3
 ; GCN1-NEXT:    v_mov_b32_e32 v0, s8
 ; GCN1-NEXT:    v_mov_b32_e32 v1, s9
@@ -12721,6 +13481,28 @@ define amdgpu_kernel void @atomic_cmpxchg_i64_ret_addr64(ptr %out, ptr %out2, i6
 ; GCN1-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
 ; GCN1-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN1-NEXT:    buffer_wbinvl1_vol
+; GCN1-NEXT:    s_cbranch_execz .LBB98_3
+; GCN1-NEXT:    s_branch .LBB98_4
+; GCN1-NEXT:  .LBB98_2:
+; GCN1-NEXT:    ; implicit-def: $vgpr0_vgpr1
+; GCN1-NEXT:  .LBB98_3: ; %atomicrmw.private
+; GCN1-NEXT:    v_cmp_ne_u64_e64 s[4:5], s[2:3], 0
+; GCN1-NEXT:    v_mov_b32_e32 v5, s8
+; GCN1-NEXT:    s_and_b64 s[4:5], s[4:5], exec
+; GCN1-NEXT:    s_cselect_b32 s2, s2, -1
+; GCN1-NEXT:    v_mov_b32_e32 v2, s2
+; GCN1-NEXT:    s_add_i32 s2, s2, 4
+; GCN1-NEXT:    v_mov_b32_e32 v3, s2
+; GCN1-NEXT:    buffer_load_dword v0, v2, s[16:19], 0 offen
+; GCN1-NEXT:    buffer_load_dword v1, v3, s[16:19], 0 offen
+; GCN1-NEXT:    v_mov_b32_e32 v4, s9
+; GCN1-NEXT:    s_waitcnt vmcnt(0)
+; GCN1-NEXT:    v_cmp_eq_u64_e32 vcc, s[0:1], v[0:1]
+; GCN1-NEXT:    v_cndmask_b32_e32 v5, v0, v5, vcc
+; GCN1-NEXT:    v_cndmask_b32_e32 v4, v1, v4, vcc
+; GCN1-NEXT:    buffer_store_dword v5, v2, s[16:19], 0 offen
+; GCN1-NEXT:    buffer_store_dword v4, v3, s[16:19], 0 offen
+; GCN1-NEXT:  .LBB98_4: ; %atomicrmw.end
 ; GCN1-NEXT:    v_mov_b32_e32 v2, s6
 ; GCN1-NEXT:    v_mov_b32_e32 v3, s7
 ; GCN1-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
@@ -12728,12 +13510,24 @@ define amdgpu_kernel void @atomic_cmpxchg_i64_ret_addr64(ptr %out, ptr %out2, i6
 ;
 ; GCN2-LABEL: atomic_cmpxchg_i64_ret_addr64:
 ; GCN2:       ; %bb.0: ; %entry
+; GCN2-NEXT:    s_mov_b32 s88, SCRATCH_RSRC_DWORD0
+; GCN2-NEXT:    s_mov_b32 s89, SCRATCH_RSRC_DWORD1
+; GCN2-NEXT:    s_mov_b32 s90, -1
+; GCN2-NEXT:    s_mov_b32 s91, 0xe80000
+; GCN2-NEXT:    s_add_u32 s88, s88, s9
 ; GCN2-NEXT:    s_load_dwordx8 s[4:11], s[2:3], 0x24
+; GCN2-NEXT:    s_load_dword s12, s[2:3], 0x10c
 ; GCN2-NEXT:    s_load_dwordx2 s[0:1], s[2:3], 0x44
+; GCN2-NEXT:    s_addc_u32 s89, s89, 0
 ; GCN2-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN2-NEXT:    s_lshl_b64 s[2:3], s[10:11], 3
 ; GCN2-NEXT:    s_add_u32 s2, s4, s2
 ; GCN2-NEXT:    s_addc_u32 s3, s5, s3
+; GCN2-NEXT:    s_cmp_eq_u32 s3, s12
+; GCN2-NEXT:    s_cselect_b64 s[4:5], -1, 0
+; GCN2-NEXT:    s_andn2_b64 vcc, exec, s[4:5]
+; GCN2-NEXT:    s_cbranch_vccz .LBB98_2
+; GCN2-NEXT:  ; %bb.1: ; %atomicrmw.global
 ; GCN2-NEXT:    v_mov_b32_e32 v5, s3
 ; GCN2-NEXT:    v_mov_b32_e32 v0, s8
 ; GCN2-NEXT:    v_mov_b32_e32 v1, s9
@@ -12743,6 +13537,27 @@ define amdgpu_kernel void @atomic_cmpxchg_i64_ret_addr64(ptr %out, ptr %out2, i6
 ; GCN2-NEXT:    flat_atomic_cmpswap_x2 v[0:1], v[4:5], v[0:3] glc
 ; GCN2-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GCN2-NEXT:    buffer_wbinvl1_vol
+; GCN2-NEXT:    s_cbranch_execz .LBB98_3
+; GCN2-NEXT:    s_branch .LBB98_4
+; GCN2-NEXT:  .LBB98_2:
+; GCN2-NEXT:    ; implicit-def: $vgpr0_vgpr1
+; GCN2-NEXT:  .LBB98_3: ; %atomicrmw.private
+; GCN2-NEXT:    s_cmp_lg_u64 s[2:3], 0
+; GCN2-NEXT:    s_cselect_b32 s2, s2, -1
+; GCN2-NEXT:    v_mov_b32_e32 v2, s2
+; GCN2-NEXT:    s_add_i32 s2, s2, 4
+; GCN2-NEXT:    v_mov_b32_e32 v3, s2
+; GCN2-NEXT:    buffer_load_dword v0, v2, s[88:91], 0 offen
+; GCN2-NEXT:    buffer_load_dword v1, v3, s[88:91], 0 offen
+; GCN2-NEXT:    v_mov_b32_e32 v5, s8
+; GCN2-NEXT:    v_mov_b32_e32 v4, s9
+; GCN2-NEXT:    s_waitcnt vmcnt(0)
+; GCN2-NEXT:    v_cmp_eq_u64_e32 vcc, s[0:1], v[0:1]
+; GCN2-NEXT:    v_cndmask_b32_e32 v5, v0, v5, vcc
+; GCN2-NEXT:    v_cndmask_b32_e32 v4, v1, v4, vcc
+; GCN2-NEXT:    buffer_store_dword v5, v2, s[88:91], 0 offen
+; GCN2-NEXT:    buffer_store_dword v4, v3, s[88:91], 0 offen
+; GCN2-NEXT:  .LBB98_4: ; %atomicrmw.end
 ; GCN2-NEXT:    v_mov_b32_e32 v2, s6
 ; GCN2-NEXT:    v_mov_b32_e32 v3, s7
 ; GCN2-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
@@ -12755,13 +13570,35 @@ define amdgpu_kernel void @atomic_cmpxchg_i64_ret_addr64(ptr %out, ptr %out2, i6
 ; GFX12-NEXT:    s_load_b64 s[0:1], s[2:3], 0x44
 ; GFX12-NEXT:    s_wait_kmcnt 0x0
 ; GFX12-NEXT:    s_lshl_b64 s[2:3], s[10:11], 3
-; GFX12-NEXT:    v_dual_mov_b32 v0, s8 :: v_dual_mov_b32 v1, s9
+; GFX12-NEXT:    s_mov_b64 s[10:11], src_private_base
 ; GFX12-NEXT:    s_add_nc_u64 s[2:3], s[4:5], s[2:3]
+; GFX12-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
+; GFX12-NEXT:    s_cmp_eq_u32 s3, s11
+; GFX12-NEXT:    s_cselect_b32 s4, -1, 0
+; GFX12-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s4
+; GFX12-NEXT:    s_cbranch_vccz .LBB98_2
+; GFX12-NEXT:  ; %bb.1: ; %atomicrmw.global
+; GFX12-NEXT:    v_dual_mov_b32 v0, s8 :: v_dual_mov_b32 v1, s9
 ; GFX12-NEXT:    v_dual_mov_b32 v2, s0 :: v_dual_mov_b32 v3, s1
 ; GFX12-NEXT:    v_dual_mov_b32 v5, s3 :: v_dual_mov_b32 v4, s2
 ; GFX12-NEXT:    flat_atomic_cmpswap_b64 v[0:1], v[4:5], v[0:3] th:TH_ATOMIC_RETURN scope:SCOPE_DEV
 ; GFX12-NEXT:    s_wait_loadcnt_dscnt 0x0
 ; GFX12-NEXT:    global_inv scope:SCOPE_DEV
+; GFX12-NEXT:    s_cbranch_execz .LBB98_3
+; GFX12-NEXT:    s_branch .LBB98_4
+; GFX12-NEXT:  .LBB98_2:
+; GFX12-NEXT:    ; implicit-def: $vgpr0_vgpr1
+; GFX12-NEXT:  .LBB98_3: ; %atomicrmw.private
+; GFX12-NEXT:    s_wait_alu 0xfffe
+; GFX12-NEXT:    s_cmp_lg_u64 s[2:3], 0
+; GFX12-NEXT:    s_cselect_b32 s2, s2, -1
+; GFX12-NEXT:    scratch_load_b64 v[0:1], off, s2
+; GFX12-NEXT:    s_wait_loadcnt 0x0
+; GFX12-NEXT:    v_cmp_eq_u64_e32 vcc_lo, s[0:1], v[0:1]
+; GFX12-NEXT:    v_cndmask_b32_e64 v3, v1, s9, vcc_lo
+; GFX12-NEXT:    v_cndmask_b32_e64 v2, v0, s8, vcc_lo
+; GFX12-NEXT:    scratch_store_b64 off, v[2:3], s2
+; GFX12-NEXT:  .LBB98_4: ; %atomicrmw.end
 ; GFX12-NEXT:    v_dual_mov_b32 v2, s6 :: v_dual_mov_b32 v3, s7
 ; GFX12-NEXT:    flat_store_b64 v[2:3], v[0:1]
 ; GFX12-NEXT:    s_endpgm
diff --git a/llvm/test/Transforms/AtomicExpand/AMDGPU/expand-atomicrmw-flat-noalias-addrspace.ll b/llvm/test/Transforms/AtomicExpand/AMDGPU/expand-atomicrmw-flat-noalias-addrspace.ll
index 9e9503dfbd3819..380f376ce9c801 100644
--- a/llvm/test/Transforms/AtomicExpand/AMDGPU/expand-atomicrmw-flat-noalias-addrspace.ll
+++ b/llvm/test/Transforms/AtomicExpand/AMDGPU/expand-atomicrmw-flat-noalias-addrspace.ll
@@ -43,7 +43,7 @@ define i64 @test_flat_atomicrmw_sub_0_i64_agent(ptr %ptr) {
 ; ALL:       [[ATOMICRMW_PRIVATE]]:
 ; ALL-NEXT:    [[TMP1:%.*]] = addrspacecast ptr [[PTR]] to ptr addrspace(5)
 ; ALL-NEXT:    [[LOADED_PRIVATE:%.*]] = load i64, ptr addrspace(5) [[TMP1]], align 8
-; ALL-NEXT:    [[NEW:%.*]] = sub i64 [[LOADED_PRIVATE]], 0
+; ALL-NEXT:    [[NEW:%.*]] = add i64 [[LOADED_PRIVATE]], 0
 ; ALL-NEXT:    store i64 [[NEW]], ptr addrspace(5) [[TMP1]], align 8
 ; ALL-NEXT:    br label %[[ATOMICRMW_PHI:.*]]
 ; ALL:       [[ATOMICRMW_GLOBAL]]:
@@ -67,7 +67,7 @@ define i64 @test_flat_atomicrmw_or_0_i64_agent(ptr %ptr) {
 ; ALL:       [[ATOMICRMW_PRIVATE]]:
 ; ALL-NEXT:    [[TMP1:%.*]] = addrspacecast ptr [[PTR]] to ptr addrspace(5)
 ; ALL-NEXT:    [[LOADED_PRIVATE:%.*]] = load i64, ptr addrspace(5) [[TMP1]], align 8
-; ALL-NEXT:    [[NEW:%.*]] = or i64 [[LOADED_PRIVATE]], 0
+; ALL-NEXT:    [[NEW:%.*]] = add i64 [[LOADED_PRIVATE]], 0
 ; ALL-NEXT:    store i64 [[NEW]], ptr addrspace(5) [[TMP1]], align 8
 ; ALL-NEXT:    br label %[[ATOMICRMW_PHI:.*]]
 ; ALL:       [[ATOMICRMW_GLOBAL]]:
@@ -91,7 +91,7 @@ define i64 @test_flat_atomicrmw_xor_0_i64_agent(ptr %ptr) {
 ; ALL:       [[ATOMICRMW_PRIVATE]]:
 ; ALL-NEXT:    [[TMP1:%.*]] = addrspacecast ptr [[PTR]] to ptr addrspace(5)
 ; ALL-NEXT:    [[LOADED_PRIVATE:%.*]] = load i64, ptr addrspace(5) [[TMP1]], align 8
-; ALL-NEXT:    [[NEW:%.*]] = xor i64 [[LOADED_PRIVATE]], 0
+; ALL-NEXT:    [[NEW:%.*]] = add i64 [[LOADED_PRIVATE]], 0
 ; ALL-NEXT:    store i64 [[NEW]], ptr addrspace(5) [[TMP1]], align 8
 ; ALL-NEXT:    br label %[[ATOMICRMW_PHI:.*]]
 ; ALL:       [[ATOMICRMW_GLOBAL]]:
diff --git a/llvm/test/Transforms/AtomicExpand/AMDGPU/expand-atomicrmw-integer-ops-0-to-add-0.ll b/llvm/test/Transforms/AtomicExpand/AMDGPU/expand-atomicrmw-integer-ops-0-to-add-0.ll
index 80058b3cef4ea1..2bfcc5897c3824 100644
--- a/llvm/test/Transforms/AtomicExpand/AMDGPU/expand-atomicrmw-integer-ops-0-to-add-0.ll
+++ b/llvm/test/Transforms/AtomicExpand/AMDGPU/expand-atomicrmw-integer-ops-0-to-add-0.ll
@@ -40,14 +40,14 @@ define i32 @test_atomicrmw_or_0_flat_system(ptr %ptr) {
 ; CHECK:       atomicrmw.private:
 ; CHECK-NEXT:    [[TMP1:%.*]] = addrspacecast ptr [[PTR]] to ptr addrspace(5)
 ; CHECK-NEXT:    [[LOADED_PRIVATE:%.*]] = load i32, ptr addrspace(5) [[TMP1]], align 4
-; CHECK-NEXT:    [[NEW:%.*]] = or i32 [[LOADED_PRIVATE]], 0
+; CHECK-NEXT:    [[NEW:%.*]] = add i32 [[LOADED_PRIVATE]], 0
 ; CHECK-NEXT:    store i32 [[NEW]], ptr addrspace(5) [[TMP1]], align 4
 ; CHECK-NEXT:    br label [[ATOMICRMW_PHI:%.*]]
 ; CHECK:       atomicrmw.global:
-; CHECK-NEXT:    [[RES:%.*]] = atomicrmw add ptr [[PTR]], i32 0 seq_cst, align 4, !noalias.addrspace [[META1:![0-9]+]]
+; CHECK-NEXT:    [[TMP2:%.*]] = atomicrmw add ptr [[PTR]], i32 0 seq_cst, align 4, !noalias.addrspace [[META1:![0-9]+]]
 ; CHECK-NEXT:    br label [[ATOMICRMW_PHI]]
 ; CHECK:       atomicrmw.phi:
-; CHECK-NEXT:    [[RES1:%.*]] = phi i32 [ [[LOADED_PRIVATE]], [[ATOMICRMW_PRIVATE]] ], [ [[RES]], [[ATOMICRMW_GLOBAL]] ]
+; CHECK-NEXT:    [[RES1:%.*]] = phi i32 [ [[LOADED_PRIVATE]], [[ATOMICRMW_PRIVATE]] ], [ [[TMP2]], [[ATOMICRMW_GLOBAL]] ]
 ; CHECK-NEXT:    br label [[ATOMICRMW_END:%.*]]
 ; CHECK:       atomicrmw.end:
 ; CHECK-NEXT:    ret i32 [[RES1]]
diff --git a/llvm/test/Transforms/AtomicExpand/AMDGPU/expand-cmpxchg-flat-maybe-private.ll b/llvm/test/Transforms/AtomicExpand/AMDGPU/expand-cmpxchg-flat-maybe-private.ll
index 6b3c27be8688c2..ff550cbfd75d4f 100644
--- a/llvm/test/Transforms/AtomicExpand/AMDGPU/expand-cmpxchg-flat-maybe-private.ll
+++ b/llvm/test/Transforms/AtomicExpand/AMDGPU/expand-cmpxchg-flat-maybe-private.ll
@@ -84,7 +84,24 @@ define { i32, i1 } @cmpxchg_flat_agent_i32(ptr %ptr, i32 %val, i32 %swap) {
 define { i64, i1 } @cmpxchg_flat_agent_i64(ptr %ptr, i64 %val, i64 %swap) {
 ; CHECK-LABEL: define { i64, i1 } @cmpxchg_flat_agent_i64(
 ; CHECK-SAME: ptr [[PTR:%.*]], i64 [[VAL:%.*]], i64 [[SWAP:%.*]]) {
-; CHECK-NEXT:    [[RESULT:%.*]] = cmpxchg ptr [[PTR]], i64 [[VAL]], i64 [[SWAP]] syncscope("agent") monotonic seq_cst, align 8
+; CHECK-NEXT:    [[IS_PRIVATE:%.*]] = call i1 @llvm.amdgcn.is.private(ptr [[PTR]])
+; CHECK-NEXT:    br i1 [[IS_PRIVATE]], label %[[ATOMICRMW_PRIVATE:.*]], label %[[ATOMICRMW_GLOBAL:.*]]
+; CHECK:       [[ATOMICRMW_PRIVATE]]:
+; CHECK-NEXT:    [[TMP3:%.*]] = addrspacecast ptr [[PTR]] to ptr addrspace(5)
+; CHECK-NEXT:    [[TMP4:%.*]] = load i64, ptr addrspace(5) [[TMP3]], align 8
+; CHECK-NEXT:    [[TMP5:%.*]] = icmp eq i64 [[TMP4]], [[VAL]]
+; CHECK-NEXT:    [[TMP6:%.*]] = select i1 [[TMP5]], i64 [[SWAP]], i64 [[TMP4]]
+; CHECK-NEXT:    store i64 [[TMP6]], ptr addrspace(5) [[TMP3]], align 8
+; CHECK-NEXT:    [[TMP7:%.*]] = insertvalue { i64, i1 } poison, i64 [[TMP4]], 0
+; CHECK-NEXT:    [[TMP8:%.*]] = insertvalue { i64, i1 } [[TMP7]], i1 [[TMP5]], 1
+; CHECK-NEXT:    br label %[[ATOMICRMW_PHI:.*]]
+; CHECK:       [[ATOMICRMW_GLOBAL]]:
+; CHECK-NEXT:    [[TMP9:%.*]] = cmpxchg ptr [[PTR]], i64 [[VAL]], i64 [[SWAP]] syncscope("agent") monotonic seq_cst, align 8, !noalias.addrspace [[META0:![0-9]+]]
+; CHECK-NEXT:    br label %[[ATOMICRMW_PHI]]
+; CHECK:       [[ATOMICRMW_PHI]]:
+; CHECK-NEXT:    [[RESULT:%.*]] = phi { i64, i1 } [ [[TMP8]], %[[ATOMICRMW_PRIVATE]] ], [ [[TMP9]], %[[ATOMICRMW_GLOBAL]] ]
+; CHECK-NEXT:    br label %[[ATOMICRMW_END:.*]]
+; CHECK:       [[ATOMICRMW_END]]:
 ; CHECK-NEXT:    ret { i64, i1 } [[RESULT]]
 ;
   %result = cmpxchg ptr %ptr, i64 %val, i64 %swap syncscope("agent") monotonic seq_cst
@@ -94,7 +111,24 @@ define { i64, i1 } @cmpxchg_flat_agent_i64(ptr %ptr, i64 %val, i64 %swap) {
 define { i64, i1 } @cmpxchg_flat_agent_i64_volatile(ptr %ptr, i64 %val, i64 %swap) {
 ; CHECK-LABEL: define { i64, i1 } @cmpxchg_flat_agent_i64_volatile(
 ; CHECK-SAME: ptr [[PTR:%.*]], i64 [[VAL:%.*]], i64 [[SWAP:%.*]]) {
-; CHECK-NEXT:    [[RESULT:%.*]] = cmpxchg volatile ptr [[PTR]], i64 [[VAL]], i64 [[SWAP]] syncscope("agent") monotonic seq_cst, align 8
+; CHECK-NEXT:    [[IS_PRIVATE:%.*]] = call i1 @llvm.amdgcn.is.private(ptr [[PTR]])
+; CHECK-NEXT:    br i1 [[IS_PRIVATE]], label %[[ATOMICRMW_PRIVATE:.*]], label %[[ATOMICRMW_GLOBAL:.*]]
+; CHECK:       [[ATOMICRMW_PRIVATE]]:
+; CHECK-NEXT:    [[TMP3:%.*]] = addrspacecast ptr [[PTR]] to ptr addrspace(5)
+; CHECK-NEXT:    [[TMP4:%.*]] = load i64, ptr addrspace(5) [[TMP3]], align 8
+; CHECK-NEXT:    [[TMP5:%.*]] = icmp eq i64 [[TMP4]], [[VAL]]
+; CHECK-NEXT:    [[TMP6:%.*]] = select i1 [[TMP5]], i64 [[SWAP]], i64 [[TMP4]]
+; CHECK-NEXT:    store i64 [[TMP6]], ptr addrspace(5) [[TMP3]], align 8
+; CHECK-NEXT:    [[TMP7:%.*]] = insertvalue { i64, i1 } poison, i64 [[TMP4]], 0
+; CHECK-NEXT:    [[TMP8:%.*]] = insertvalue { i64, i1 } [[TMP7]], i1 [[TMP5]], 1
+; CHECK-NEXT:    br label %[[ATOMICRMW_PHI:.*]]
+; CHECK:       [[ATOMICRMW_GLOBAL]]:
+; CHECK-NEXT:    [[TMP9:%.*]] = cmpxchg volatile ptr [[PTR]], i64 [[VAL]], i64 [[SWAP]] syncscope("agent") monotonic seq_cst, align 8, !noalias.addrspace [[META0]]
+; CHECK-NEXT:    br label %[[ATOMICRMW_PHI]]
+; CHECK:       [[ATOMICRMW_PHI]]:
+; CHECK-NEXT:    [[RESULT:%.*]] = phi { i64, i1 } [ [[TMP8]], %[[ATOMICRMW_PRIVATE]] ], [ [[TMP9]], %[[ATOMICRMW_GLOBAL]] ]
+; CHECK-NEXT:    br label %[[ATOMICRMW_END:.*]]
+; CHECK:       [[ATOMICRMW_END]]:
 ; CHECK-NEXT:    ret { i64, i1 } [[RESULT]]
 ;
   %result = cmpxchg volatile ptr %ptr, i64 %val, i64 %swap syncscope("agent") monotonic seq_cst
@@ -144,7 +178,7 @@ define { i16, i1 } @cmpxchg_flat_agent_i16__noprivate(ptr %ptr, i16 %val, i16 %s
 define { i32, i1 } @cmpxchg_flat_agent_i32__noprivate(ptr %ptr, i32 %val, i32 %swap) {
 ; CHECK-LABEL: define { i32, i1 } @cmpxchg_flat_agent_i32__noprivate(
 ; CHECK-SAME: ptr [[PTR:%.*]], i32 [[VAL:%.*]], i32 [[SWAP:%.*]]) {
-; CHECK-NEXT:    [[RESULT:%.*]] = cmpxchg ptr [[PTR]], i32 [[VAL]], i32 [[SWAP]] syncscope("agent") monotonic seq_cst, align 4, !noalias.addrspace [[META0:![0-9]+]]
+; CHECK-NEXT:    [[RESULT:%.*]] = cmpxchg ptr [[PTR]], i32 [[VAL]], i32 [[SWAP]] syncscope("agent") monotonic seq_cst, align 4, !noalias.addrspace [[META0]]
 ; CHECK-NEXT:    ret { i32, i1 } [[RESULT]]
 ;
   %result = cmpxchg ptr %ptr, i32 %val, i32 %swap syncscope("agent") monotonic seq_cst, !noalias.addrspace !0
@@ -164,7 +198,24 @@ define { i64, i1 } @cmpxchg_flat_agent_i64__noprivate(ptr %ptr, i64 %val, i64 %s
 define { i64, i1 } @cmpxchg_flat_agent_i64__nolocal(ptr %ptr, i64 %val, i64 %swap) {
 ; CHECK-LABEL: define { i64, i1 } @cmpxchg_flat_agent_i64__nolocal(
 ; CHECK-SAME: ptr [[PTR:%.*]], i64 [[VAL:%.*]], i64 [[SWAP:%.*]]) {
-; CHECK-NEXT:    [[RESULT:%.*]] = cmpxchg ptr [[PTR]], i64 [[VAL]], i64 [[SWAP]] syncscope("agent") monotonic seq_cst, align 8, !noalias.addrspace [[META1:![0-9]+]]
+; CHECK-NEXT:    [[IS_PRIVATE:%.*]] = call i1 @llvm.amdgcn.is.private(ptr [[PTR]])
+; CHECK-NEXT:    br i1 [[IS_PRIVATE]], label %[[ATOMICRMW_PRIVATE:.*]], label %[[ATOMICRMW_GLOBAL:.*]]
+; CHECK:       [[ATOMICRMW_PRIVATE]]:
+; CHECK-NEXT:    [[TMP3:%.*]] = addrspacecast ptr [[PTR]] to ptr addrspace(5)
+; CHECK-NEXT:    [[TMP4:%.*]] = load i64, ptr addrspace(5) [[TMP3]], align 8
+; CHECK-NEXT:    [[TMP5:%.*]] = icmp eq i64 [[TMP4]], [[VAL]]
+; CHECK-NEXT:    [[TMP6:%.*]] = select i1 [[TMP5]], i64 [[SWAP]], i64 [[TMP4]]
+; CHECK-NEXT:    store i64 [[TMP6]], ptr addrspace(5) [[TMP3]], align 8
+; CHECK-NEXT:    [[TMP7:%.*]] = insertvalue { i64, i1 } poison, i64 [[TMP4]], 0
+; CHECK-NEXT:    [[TMP8:%.*]] = insertvalue { i64, i1 } [[TMP7]], i1 [[TMP5]], 1
+; CHECK-NEXT:    br label %[[ATOMICRMW_PHI:.*]]
+; CHECK:       [[ATOMICRMW_GLOBAL]]:
+; CHECK-NEXT:    [[TMP9:%.*]] = cmpxchg ptr [[PTR]], i64 [[VAL]], i64 [[SWAP]] syncscope("agent") monotonic seq_cst, align 8, !noalias.addrspace [[META0]]
+; CHECK-NEXT:    br label %[[ATOMICRMW_PHI]]
+; CHECK:       [[ATOMICRMW_PHI]]:
+; CHECK-NEXT:    [[RESULT:%.*]] = phi { i64, i1 } [ [[TMP8]], %[[ATOMICRMW_PRIVATE]] ], [ [[TMP9]], %[[ATOMICRMW_GLOBAL]] ]
+; CHECK-NEXT:    br label %[[ATOMICRMW_END:.*]]
+; CHECK:       [[ATOMICRMW_END]]:
 ; CHECK-NEXT:    ret { i64, i1 } [[RESULT]]
 ;
   %result = cmpxchg ptr %ptr, i64 %val, i64 %swap syncscope("agent") monotonic seq_cst, !noalias.addrspace !1
@@ -174,7 +225,24 @@ define { i64, i1 } @cmpxchg_flat_agent_i64__nolocal(ptr %ptr, i64 %val, i64 %swa
 define { i64, i1 } @cmpxchg_flat_agent_i64_mmra(ptr %ptr, i64 %val, i64 %swap) {
 ; CHECK-LABEL: define { i64, i1 } @cmpxchg_flat_agent_i64_mmra(
 ; CHECK-SAME: ptr [[PTR:%.*]], i64 [[VAL:%.*]], i64 [[SWAP:%.*]]) {
-; CHECK-NEXT:    [[RESULT:%.*]] = cmpxchg ptr [[PTR]], i64 [[VAL]], i64 [[SWAP]] syncscope("agent") monotonic seq_cst, align 8, !mmra [[META2:![0-9]+]]
+; CHECK-NEXT:    [[IS_PRIVATE:%.*]] = call i1 @llvm.amdgcn.is.private(ptr [[PTR]])
+; CHECK-NEXT:    br i1 [[IS_PRIVATE]], label %[[ATOMICRMW_PRIVATE:.*]], label %[[ATOMICRMW_GLOBAL:.*]]
+; CHECK:       [[ATOMICRMW_PRIVATE]]:
+; CHECK-NEXT:    [[TMP1:%.*]] = addrspacecast ptr [[PTR]] to ptr addrspace(5)
+; CHECK-NEXT:    [[TMP2:%.*]] = load i64, ptr addrspace(5) [[TMP1]], align 8
+; CHECK-NEXT:    [[TMP3:%.*]] = icmp eq i64 [[TMP2]], [[VAL]]
+; CHECK-NEXT:    [[TMP4:%.*]] = select i1 [[TMP3]], i64 [[SWAP]], i64 [[TMP2]]
+; CHECK-NEXT:    store i64 [[TMP4]], ptr addrspace(5) [[TMP1]], align 8
+; CHECK-NEXT:    [[TMP5:%.*]] = insertvalue { i64, i1 } poison, i64 [[TMP2]], 0
+; CHECK-NEXT:    [[TMP6:%.*]] = insertvalue { i64, i1 } [[TMP5]], i1 [[TMP3]], 1
+; CHECK-NEXT:    br label %[[ATOMICRMW_PHI:.*]]
+; CHECK:       [[ATOMICRMW_GLOBAL]]:
+; CHECK-NEXT:    [[TMP7:%.*]] = cmpxchg ptr [[PTR]], i64 [[VAL]], i64 [[SWAP]] syncscope("agent") monotonic seq_cst, align 8, !mmra [[META1:![0-9]+]], !noalias.addrspace [[META0]]
+; CHECK-NEXT:    br label %[[ATOMICRMW_PHI]]
+; CHECK:       [[ATOMICRMW_PHI]]:
+; CHECK-NEXT:    [[RESULT:%.*]] = phi { i64, i1 } [ [[TMP6]], %[[ATOMICRMW_PRIVATE]] ], [ [[TMP7]], %[[ATOMICRMW_GLOBAL]] ]
+; CHECK-NEXT:    br label %[[ATOMICRMW_END:.*]]
+; CHECK:       [[ATOMICRMW_END]]:
 ; CHECK-NEXT:    ret { i64, i1 } [[RESULT]]
 ;
   %result = cmpxchg ptr %ptr, i64 %val, i64 %swap syncscope("agent") monotonic seq_cst, !mmra !4
@@ -184,7 +252,24 @@ define { i64, i1 } @cmpxchg_flat_agent_i64_mmra(ptr %ptr, i64 %val, i64 %swap) {
 define { i64, i1 } @cmpxchg_flat_agent_i64_mmra_noprivate(ptr %ptr, i64 %val, i64 %swap) {
 ; CHECK-LABEL: define { i64, i1 } @cmpxchg_flat_agent_i64_mmra_noprivate(
 ; CHECK-SAME: ptr [[PTR:%.*]], i64 [[VAL:%.*]], i64 [[SWAP:%.*]]) {
-; CHECK-NEXT:    [[RESULT:%.*]] = cmpxchg ptr [[PTR]], i64 [[VAL]], i64 [[SWAP]] syncscope("agent") monotonic seq_cst, align 8, !mmra [[META2]], !noalias.addrspace [[META1]]
+; CHECK-NEXT:    [[IS_PRIVATE:%.*]] = call i1 @llvm.amdgcn.is.private(ptr [[PTR]])
+; CHECK-NEXT:    br i1 [[IS_PRIVATE]], label %[[ATOMICRMW_PRIVATE:.*]], label %[[ATOMICRMW_GLOBAL:.*]]
+; CHECK:       [[ATOMICRMW_PRIVATE]]:
+; CHECK-NEXT:    [[TMP1:%.*]] = addrspacecast ptr [[PTR]] to ptr addrspace(5)
+; CHECK-NEXT:    [[TMP2:%.*]] = load i64, ptr addrspace(5) [[TMP1]], align 8
+; CHECK-NEXT:    [[TMP3:%.*]] = icmp eq i64 [[TMP2]], [[VAL]]
+; CHECK-NEXT:    [[TMP4:%.*]] = select i1 [[TMP3]], i64 [[SWAP]], i64 [[TMP2]]
+; CHECK-NEXT:    store i64 [[TMP4]], ptr addrspace(5) [[TMP1]], align 8
+; CHECK-NEXT:    [[TMP5:%.*]] = insertvalue { i64, i1 } poison, i64 [[TMP2]], 0
+; CHECK-NEXT:    [[TMP6:%.*]] = insertvalue { i64, i1 } [[TMP5]], i1 [[TMP3]], 1
+; CHECK-NEXT:    br label %[[ATOMICRMW_PHI:.*]]
+; CHECK:       [[ATOMICRMW_GLOBAL]]:
+; CHECK-NEXT:    [[TMP7:%.*]] = cmpxchg ptr [[PTR]], i64 [[VAL]], i64 [[SWAP]] syncscope("agent") monotonic seq_cst, align 8, !mmra [[META1]], !noalias.addrspace [[META0]]
+; CHECK-NEXT:    br label %[[ATOMICRMW_PHI]]
+; CHECK:       [[ATOMICRMW_PHI]]:
+; CHECK-NEXT:    [[RESULT:%.*]] = phi { i64, i1 } [ [[TMP6]], %[[ATOMICRMW_PRIVATE]] ], [ [[TMP7]], %[[ATOMICRMW_GLOBAL]] ]
+; CHECK-NEXT:    br label %[[ATOMICRMW_END:.*]]
+; CHECK:       [[ATOMICRMW_END]]:
 ; CHECK-NEXT:    ret { i64, i1 } [[RESULT]]
 ;
   %result = cmpxchg ptr %ptr, i64 %val, i64 %swap syncscope("agent") monotonic seq_cst, !noalias.addrspace !1, !mmra !4
@@ -201,8 +286,7 @@ define { i64, i1 } @cmpxchg_flat_agent_i64_mmra_noprivate(ptr %ptr, i64 %val, i6
 
 ;.
 ; CHECK: [[META0]] = !{i32 5, i32 6}
-; CHECK: [[META1]] = !{i32 3, i32 4}
-; CHECK: [[META2]] = !{[[META3:![0-9]+]], [[META4:![0-9]+]]}
-; CHECK: [[META3]] = !{!"foo", !"bar"}
-; CHECK: [[META4]] = !{!"bux", !"baz"}
+; CHECK: [[META1]] = !{[[META2:![0-9]+]], [[META3:![0-9]+]]}
+; CHECK: [[META2]] = !{!"foo", !"bar"}
+; CHECK: [[META3]] = !{!"bux", !"baz"}
 ;.

>From 2793c97b42bd3183a2422b35c0603bba9ea83d63 Mon Sep 17 00:00:00 2001
From: Matt Arsenault <Matthew.Arsenault at amd.com>
Date: Mon, 30 Sep 2024 11:08:44 +0400
Subject: [PATCH 2/2] Rename helper function

---
 llvm/include/llvm/Transforms/Utils/LowerAtomic.h | 7 +++----
 llvm/lib/Target/AMDGPU/SIISelLowering.cpp        | 4 ++--
 llvm/lib/Transforms/Utils/LowerAtomic.cpp        | 9 +++++----
 3 files changed, 10 insertions(+), 10 deletions(-)

diff --git a/llvm/include/llvm/Transforms/Utils/LowerAtomic.h b/llvm/include/llvm/Transforms/Utils/LowerAtomic.h
index 295c2bd2b4b47e..d34f8c7fd88822 100644
--- a/llvm/include/llvm/Transforms/Utils/LowerAtomic.h
+++ b/llvm/include/llvm/Transforms/Utils/LowerAtomic.h
@@ -25,10 +25,9 @@ bool lowerAtomicCmpXchgInst(AtomicCmpXchgInst *CXI);
 
 /// Emit IR to implement the given cmpxchg operation on values in registers,
 /// returning the new value.
-std::pair<Value *, Value *> buildAtomicCmpXchgValue(IRBuilderBase &Builder,
-                                                    Value *Ptr, Value *Cmp,
-                                                    Value *Val,
-                                                    Align Alignment);
+std::pair<Value *, Value *> buildCmpXchgValue(IRBuilderBase &Builder,
+                                              Value *Ptr, Value *Cmp,
+                                              Value *Val, Align Alignment);
 
 /// Convert the given RMWI into primitive load and stores,
 /// assuming that doing so is legal. Return true if the lowering
diff --git a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
index c8a46875bda408..204ffa2920eefe 100644
--- a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
@@ -16895,8 +16895,8 @@ void SITargetLowering::emitExpandAtomicAddrSpacePredicate(
     Builder.CreateAlignedStore(NewVal, CastToPrivate, RMW->getAlign());
   } else {
     auto [ResultLoad, Equal] =
-        buildAtomicCmpXchgValue(Builder, CastToPrivate, CX->getCompareOperand(),
-                                CX->getNewValOperand(), CX->getAlign());
+        buildCmpXchgValue(Builder, CastToPrivate, CX->getCompareOperand(),
+                          CX->getNewValOperand(), CX->getAlign());
 
     Value *Insert = Builder.CreateInsertValue(PoisonValue::get(CX->getType()),
                                               ResultLoad, 0);
diff --git a/llvm/lib/Transforms/Utils/LowerAtomic.cpp b/llvm/lib/Transforms/Utils/LowerAtomic.cpp
index 89c49d4a0732f6..b51c32485411da 100644
--- a/llvm/lib/Transforms/Utils/LowerAtomic.cpp
+++ b/llvm/lib/Transforms/Utils/LowerAtomic.cpp
@@ -26,7 +26,7 @@ bool llvm::lowerAtomicCmpXchgInst(AtomicCmpXchgInst *CXI) {
   Value *Val = CXI->getNewValOperand();
 
   auto [Orig, Equal] =
-      buildAtomicCmpXchgValue(Builder, Ptr, Cmp, Val, CXI->getAlign());
+      buildCmpXchgValue(Builder, Ptr, Cmp, Val, CXI->getAlign());
 
   Value *Res =
       Builder.CreateInsertValue(PoisonValue::get(CXI->getType()), Orig, 0);
@@ -37,9 +37,10 @@ bool llvm::lowerAtomicCmpXchgInst(AtomicCmpXchgInst *CXI) {
   return true;
 }
 
-std::pair<Value *, Value *>
-llvm::buildAtomicCmpXchgValue(IRBuilderBase &Builder, Value *Ptr, Value *Cmp,
-                              Value *Val, Align Alignment) {
+std::pair<Value *, Value *> llvm::buildCmpXchgValue(IRBuilderBase &Builder,
+                                                    Value *Ptr, Value *Cmp,
+                                                    Value *Val,
+                                                    Align Alignment) {
   LoadInst *Orig = Builder.CreateAlignedLoad(Val->getType(), Ptr, Alignment);
   Value *Equal = Builder.CreateICmpEQ(Orig, Cmp);
   Value *Res = Builder.CreateSelect(Equal, Val, Orig);



More information about the llvm-branch-commits mailing list