AMDGPU: Expand flat atomics that may access private memory

If the runtime flat address resolves to a scratch address,
64-bit atomics do not work correctly. Insert a runtime address
space check (which is quite likely to be uniform) and select between
the non-atomic and real atomic cases.

Consider noalias.addrspace metadata and avoid this expansion when
possible (we also need to consider it to avoid infinitely expanding
after adding the predication code).

Author: arsenm

llvm/lib/Target/AMDGPU/SIISelLowering.cpp

@@ -39,6 +39,7 @@
 #include "llvm/IR/IntrinsicInst.h"
 #include "llvm/IR/IntrinsicsAMDGPU.h"
 #include "llvm/IR/IntrinsicsR600.h"
+#include "llvm/IR/MDBuilder.h"
 #include "llvm/Support/CommandLine.h"
 #include "llvm/Support/KnownBits.h"
 #include "llvm/Support/ModRef.h"
@@ -16243,12 +16244,39 @@ atomicSupportedIfLegalIntType(const AtomicRMWInst *RMW) {
              : TargetLowering::AtomicExpansionKind::CmpXChg;
 }
 
+/// Return if a flat address space atomicrmw can access private memory.
+static bool flatInstrMayAccessPrivate(const Instruction *I) {
+  const MDNode *NoaliasAddrSpaceMD =
+      I->getMetadata(LLVMContext::MD_noalias_addrspace);
+  if (!NoaliasAddrSpaceMD)
+    return true;
+
+  // FIXME: Can this actually fail? Why is this optional?
+  if (std::optional<ConstantRange> CR =
+          getConstantRangeFromMetadata(*NoaliasAddrSpaceMD)) {
+    return !CR->contains(APInt(32, AMDGPUAS::PRIVATE_ADDRESS));
+  }
+
+  llvm_unreachable("Why is getConstantRangeFromMetadata optional");
+}
+
 TargetLowering::AtomicExpansionKind
 SITargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *RMW) const {
   unsigned AS = RMW->getPointerAddressSpace();
   if (AS == AMDGPUAS::PRIVATE_ADDRESS)
     return AtomicExpansionKind::NotAtomic;
 
+  // 64-bit flat atomics that dynamically reside in private memory will silently
+  // be dropped.
+  //
+  // Note that we will emit a new copy of the original atomic in the expansion,
+  // which will be incrementally relegalized.
+  const DataLayout &DL = RMW->getFunction()->getDataLayout();
+  if (AS == AMDGPUAS::FLAT_ADDRESS &&
+      DL.getTypeSizeInBits(RMW->getType()) == 64 &&
+      flatInstrMayAccessPrivate(RMW))
+    return AtomicExpansionKind::Expand;
+
   auto ReportUnsafeHWInst = [=](TargetLowering::AtomicExpansionKind Kind) {
     OptimizationRemarkEmitter ORE(RMW->getFunction());
     ORE.emit([=]() {
@@ -16647,20 +16675,34 @@ void SITargetLowering::emitExpandAtomicRMW(AtomicRMWInst *AI) const {
 
   if (Op == AtomicRMWInst::Sub || Op == AtomicRMWInst::Or ||
       Op == AtomicRMWInst::Xor) {
-    // atomicrmw or %ptr, 0 -> atomicrmw add %ptr, 0
-    assert(cast<Constant>(AI->getValOperand())->isNullValue() &&
-           "this cannot be replaced with add");
-    AI->setOperation(AtomicRMWInst::Add);
-    return;
+    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.
+    }
   }
 
-  assert(Subtarget->hasAtomicFaddInsts() &&
-         "target should have atomic fadd instructions");
-  assert(AI->getType()->isFloatTy() &&
-         AI->getPointerAddressSpace() == AMDGPUAS::FLAT_ADDRESS &&
-         "generic atomicrmw expansion only supports FP32 operand in flat "
-         "address space");
-  assert(Op == AtomicRMWInst::FAdd && "only fadd is supported for now");
+  // 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();
 
   // Given: atomicrmw fadd ptr %addr, float %val ordering
   //
@@ -16700,6 +16742,10 @@ void SITargetLowering::emitExpandAtomicRMW(AtomicRMWInst *AI) const {
   //
   // atomicrmw.end:
   //    [...]
+  //
+  //
+  // For 64-bit atomics which may reside in private memory, we perform a simpler
+  // version that only inserts the private check, and uses the flat operation.
 
   IRBuilder<> Builder(AI);
   LLVMContext &Ctx = Builder.getContext();
@@ -16711,9 +16757,15 @@ void SITargetLowering::emitExpandAtomicRMW(AtomicRMWInst *AI) const {
   Function *F = BB->getParent();
   BasicBlock *ExitBB =
       BB->splitBasicBlock(Builder.GetInsertPoint(), "atomicrmw.end");
-  BasicBlock *SharedBB = BasicBlock::Create(Ctx, "atomicrmw.shared", F, ExitBB);
-  BasicBlock *CheckPrivateBB =
-      BasicBlock::Create(Ctx, "atomicrmw.check.private", F, ExitBB);
+  BasicBlock *SharedBB = nullptr;
+
+  BasicBlock *CheckPrivateBB = BB;
+  if (FullFlatEmulation) {
+    SharedBB = BasicBlock::Create(Ctx, "atomicrmw.shared", F, ExitBB);
+    CheckPrivateBB =
+        BasicBlock::Create(Ctx, "atomicrmw.check.private", F, ExitBB);
+  }
+
   BasicBlock *PrivateBB =
       BasicBlock::Create(Ctx, "atomicrmw.private", F, ExitBB);
   BasicBlock *GlobalBB = BasicBlock::Create(Ctx, "atomicrmw.global", F, ExitBB);
@@ -16726,23 +16778,26 @@ void SITargetLowering::emitExpandAtomicRMW(AtomicRMWInst *AI) const {
 
   std::prev(BB->end())->eraseFromParent();
   Builder.SetInsertPoint(BB);
-  CallInst *IsShared = Builder.CreateIntrinsic(Intrinsic::amdgcn_is_shared, {},
-                                               {Addr}, nullptr, "is.shared");
-  Builder.CreateCondBr(IsShared, SharedBB, CheckPrivateBB);
 
-  Builder.SetInsertPoint(SharedBB);
-  Value *CastToLocal = Builder.CreateAddrSpaceCast(
-      Addr, PointerType::get(Ctx, AMDGPUAS::LOCAL_ADDRESS));
+  Value *LoadedShared = nullptr;
+  if (FullFlatEmulation) {
+    CallInst *IsShared = Builder.CreateIntrinsic(
+        Intrinsic::amdgcn_is_shared, {}, {Addr}, nullptr, "is.shared");
+    Builder.CreateCondBr(IsShared, SharedBB, CheckPrivateBB);
+    Builder.SetInsertPoint(SharedBB);
+    Value *CastToLocal = Builder.CreateAddrSpaceCast(
+        Addr, PointerType::get(Ctx, AMDGPUAS::LOCAL_ADDRESS));
 
-  Instruction *Clone = AI->clone();
-  Clone->insertInto(SharedBB, SharedBB->end());
-  Clone->getOperandUse(AtomicRMWInst::getPointerOperandIndex())
-      .set(CastToLocal);
-  Instruction *LoadedShared = Clone;
+    Instruction *Clone = AI->clone();
+    Clone->insertInto(SharedBB, SharedBB->end());
+    Clone->getOperandUse(AtomicRMWInst::getPointerOperandIndex())
+        .set(CastToLocal);
+    LoadedShared = Clone;
 
-  Builder.CreateBr(PhiBB);
+    Builder.CreateBr(PhiBB);
+    Builder.SetInsertPoint(CheckPrivateBB);
+  }
 
-  Builder.SetInsertPoint(CheckPrivateBB);
   CallInst *IsPrivate = Builder.CreateIntrinsic(
       Intrinsic::amdgcn_is_private, {}, {Addr}, nullptr, "is.private");
   Builder.CreateCondBr(IsPrivate, PrivateBB, GlobalBB);
@@ -16759,23 +16814,41 @@ void SITargetLowering::emitExpandAtomicRMW(AtomicRMWInst *AI) const {
   Builder.CreateBr(PhiBB);
 
   Builder.SetInsertPoint(GlobalBB);
-  Value *CastToGlobal = Builder.CreateAddrSpaceCast(
-      Addr, PointerType::get(Ctx, AMDGPUAS::GLOBAL_ADDRESS));
-  Value *LoadedGlobal = AI;
 
-  AI->getOperandUse(AtomicRMWInst::getPointerOperandIndex()).set(CastToGlobal);
+  // Continue using a flat instruction if we only emitted the check for private.
+  Instruction *LoadedGlobal = AI;
+  if (FullFlatEmulation) {
+    Value *CastToGlobal = Builder.CreateAddrSpaceCast(
+        Addr, PointerType::get(Ctx, AMDGPUAS::GLOBAL_ADDRESS));
+    AI->getOperandUse(AtomicRMWInst::getPointerOperandIndex())
+        .set(CastToGlobal);
+  }
 
   AI->removeFromParent();
   AI->insertInto(GlobalBB, GlobalBB->end());
 
+  // The new atomicrmw may go through another round of legalization later.
+  if (!FullFlatEmulation) {
+    // We inserted the runtime check already, make sure we do not try to
+    // re-expand this.
+    // TODO: Should union with any existing metadata.
+    MDBuilder MDB(F->getContext());
+    MDNode *RangeNotPrivate =
+        MDB.createRange(APInt(32, AMDGPUAS::PRIVATE_ADDRESS),
+                        APInt(32, AMDGPUAS::PRIVATE_ADDRESS + 1));
+    LoadedGlobal->setMetadata(LLVMContext::MD_noalias_addrspace,
+                              RangeNotPrivate);
+  }
+
   Builder.CreateBr(PhiBB);
 
   Builder.SetInsertPoint(PhiBB);
 
   if (ReturnValueIsUsed) {
     PHINode *Loaded = Builder.CreatePHI(ValTy, 3);
     AI->replaceAllUsesWith(Loaded);
-    Loaded->addIncoming(LoadedShared, SharedBB);
+    if (FullFlatEmulation)
+      Loaded->addIncoming(LoadedShared, SharedBB);
     Loaded->addIncoming(LoadedPrivate, PrivateBB);
     Loaded->addIncoming(LoadedGlobal, GlobalBB);
     Loaded->takeName(AI);

llvm/test/CodeGen/AMDGPU/GlobalISel/atomicrmw_fmax.ll

@@ -1343,7 +1343,7 @@ define double @flat_agent_atomic_fmax_ret_f64__amdgpu_no_fine_grained_memory(ptr
 ; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX7-NEXT:    buffer_wbinvl1
 ; GFX7-NEXT:    s_setpc_b64 s[30:31]
-  %result = atomicrmw fmax ptr %ptr, double %val syncscope("agent") seq_cst, !amdgpu.no.fine.grained.memory !0
+  %result = atomicrmw fmax ptr %ptr, double %val syncscope("agent") seq_cst, !noalias.addrspace !1, !amdgpu.no.fine.grained.memory !0
   ret double %result
 }
 
@@ -1494,7 +1494,7 @@ define void @flat_agent_atomic_fmax_noret_f64__amdgpu_no_fine_grained_memory(ptr
 ; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX7-NEXT:    buffer_wbinvl1
 ; GFX7-NEXT:    s_setpc_b64 s[30:31]
-  %unused = atomicrmw fmax ptr %ptr, double %val syncscope("agent") seq_cst, !amdgpu.no.fine.grained.memory !0
+  %unused = atomicrmw fmax ptr %ptr, double %val syncscope("agent") seq_cst, !noalias.addrspace !1, !amdgpu.no.fine.grained.memory !0
   ret void
 }
 
@@ -2230,3 +2230,4 @@ define void @buffer_fat_ptr_agent_atomic_fmax_noret_f64__amdgpu_no_fine_grained_
 }
 
 !0 = !{}
+!1 = !{i32 5, i32 6}

llvm/test/CodeGen/AMDGPU/GlobalISel/atomicrmw_fmin.ll

@@ -1343,7 +1343,7 @@ define double @flat_agent_atomic_fmin_ret_f64__amdgpu_no_fine_grained_memory(ptr
 ; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX7-NEXT:    buffer_wbinvl1
 ; GFX7-NEXT:    s_setpc_b64 s[30:31]
-  %result = atomicrmw fmin ptr %ptr, double %val syncscope("agent") seq_cst, !amdgpu.no.fine.grained.memory !0
+  %result = atomicrmw fmin ptr %ptr, double %val syncscope("agent") seq_cst, !noalias.addrspace !1, !amdgpu.no.fine.grained.memory !0
   ret double %result
 }
 
@@ -1494,7 +1494,7 @@ define void @flat_agent_atomic_fmin_noret_f64__amdgpu_no_fine_grained_memory(ptr
 ; GFX7-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX7-NEXT:    buffer_wbinvl1
 ; GFX7-NEXT:    s_setpc_b64 s[30:31]
-  %unused = atomicrmw fmin ptr %ptr, double %val syncscope("agent") seq_cst, !amdgpu.no.fine.grained.memory !0
+  %unused = atomicrmw fmin ptr %ptr, double %val syncscope("agent") seq_cst, !noalias.addrspace !1, !amdgpu.no.fine.grained.memory !0
   ret void
 }
 
@@ -2230,3 +2230,4 @@ define void @buffer_fat_ptr_agent_atomic_fmin_noret_f64__amdgpu_no_fine_grained_
 }
 
 !0 = !{}
+!1 = !{i32 5, i32 6}

llvm/test/CodeGen/AMDGPU/GlobalISel/atomicrmw_udec_wrap.ll

@@ -1657,7 +1657,7 @@ define amdgpu_kernel void @flat_atomic_dec_ret_i64(ptr %out, ptr %ptr) #1 {
 ; GFX11-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-NEXT:    flat_store_b64 v[2:3], v[0:1]
 ; GFX11-NEXT:    s_endpgm
-  %result = atomicrmw udec_wrap ptr %ptr, i64 42 syncscope("agent") seq_cst, align 8
+  %result = atomicrmw udec_wrap ptr %ptr, i64 42 syncscope("agent") seq_cst, align 8, !noalias.addrspace !0
   store i64 %result, ptr %out, align 4
   ret void
 }
@@ -1759,7 +1759,7 @@ define amdgpu_kernel void @flat_atomic_dec_ret_i64_offset(ptr %out, ptr %ptr) #1
 ; GFX11-NEXT:    flat_store_b64 v[2:3], v[0:1]
 ; GFX11-NEXT:    s_endpgm
   %gep = getelementptr i64, ptr %ptr, i32 4
-  %result = atomicrmw udec_wrap ptr %gep, i64 42 syncscope("agent") seq_cst, align 8
+  %result = atomicrmw udec_wrap ptr %gep, i64 42 syncscope("agent") seq_cst, align 8, !noalias.addrspace !0
   store i64 %result, ptr %out, align 4
   ret void
 }
@@ -1832,7 +1832,7 @@ define amdgpu_kernel void @flat_atomic_dec_noret_i64(ptr %ptr) #1 {
 ; GFX11-NEXT:    buffer_gl1_inv
 ; GFX11-NEXT:    buffer_gl0_inv
 ; GFX11-NEXT:    s_endpgm
-  %result = atomicrmw udec_wrap ptr %ptr, i64 42 syncscope("agent") seq_cst, align 8
+  %result = atomicrmw udec_wrap ptr %ptr, i64 42 syncscope("agent") seq_cst, align 8, !noalias.addrspace !0
   ret void
 }
 
@@ -1911,7 +1911,7 @@ define amdgpu_kernel void @flat_atomic_dec_noret_i64_offset(ptr %ptr) #1 {
 ; GFX11-NEXT:    buffer_gl0_inv
 ; GFX11-NEXT:    s_endpgm
   %gep = getelementptr i64, ptr %ptr, i32 4
-  %result = atomicrmw udec_wrap ptr %gep, i64 42 syncscope("agent") seq_cst, align 8
+  %result = atomicrmw udec_wrap ptr %gep, i64 42 syncscope("agent") seq_cst, align 8, !noalias.addrspace !0
   ret void
 }
 
@@ -1990,7 +1990,7 @@ define amdgpu_kernel void @flat_atomic_dec_noret_i64_offset_system(ptr %ptr) #1
 ; GFX11-NEXT:    buffer_gl0_inv
 ; GFX11-NEXT:    s_endpgm
   %gep = getelementptr i64, ptr %ptr, i32 4
-  %result = atomicrmw udec_wrap ptr %gep, i64 42 seq_cst, align 8
+  %result = atomicrmw udec_wrap ptr %gep, i64 42 seq_cst, align 8, !noalias.addrspace !0
   ret void
 }
 
@@ -2118,7 +2118,7 @@ define amdgpu_kernel void @flat_atomic_dec_ret_i64_offset_addr64(ptr %out, ptr %
   %gep.tid = getelementptr i64, ptr %ptr, i32 %id
   %out.gep = getelementptr i64, ptr %out, i32 %id
   %gep = getelementptr i64, ptr %gep.tid, i32 5
-  %result = atomicrmw udec_wrap ptr %gep, i64 42 syncscope("agent") seq_cst, align 8
+  %result = atomicrmw udec_wrap ptr %gep, i64 42 syncscope("agent") seq_cst, align 8, !noalias.addrspace !0
   store i64 %result, ptr %out.gep, align 4
   ret void
 }
@@ -2217,7 +2217,7 @@ define amdgpu_kernel void @flat_atomic_dec_noret_i64_offset_addr64(ptr %ptr) #1
   %id = call i32 @llvm.amdgcn.workitem.id.x()
   %gep.tid = getelementptr i64, ptr %ptr, i32 %id
   %gep = getelementptr i64, ptr %gep.tid, i32 5
-  %result = atomicrmw udec_wrap ptr %gep, i64 42 syncscope("agent") seq_cst, align 8
+  %result = atomicrmw udec_wrap ptr %gep, i64 42 syncscope("agent") seq_cst, align 8, !noalias.addrspace !0
   ret void
 }
 
@@ -3340,7 +3340,7 @@ define amdgpu_kernel void @atomic_dec_shl_base_lds_0_i64(ptr addrspace(1) %out,
   %tid.x = tail call i32 @llvm.amdgcn.workitem.id.x() #2
   %idx.0 = add nsw i32 %tid.x, 2
   %arrayidx0 = getelementptr inbounds [512 x i64], ptr addrspace(3) @lds1, i32 0, i32 %idx.0
-  %result = atomicrmw udec_wrap ptr addrspace(3) %arrayidx0, i64 9 syncscope("agent") seq_cst, align 8
+  %result = atomicrmw udec_wrap ptr addrspace(3) %arrayidx0, i64 9 syncscope("agent") seq_cst, align 8, !noalias.addrspace !0
   store i32 %idx.0, ptr addrspace(1) %add_use, align 4
   store i64 %result, ptr addrspace(1) %out, align 4
   ret void
@@ -3349,5 +3349,8 @@ define amdgpu_kernel void @atomic_dec_shl_base_lds_0_i64(ptr addrspace(1) %out,
 attributes #0 = { nounwind speculatable willreturn memory(none) }
 attributes #1 = { nounwind }
 attributes #2 = { nounwind memory(none) }
+
+!0 = !{i32 5, i32 6}
+
 ;; NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line:
 ; GCN: {{.*}}