[DirectX] Lower llvm.lifetime.* intrinsics to stores when DXIL version is lower than 1.6 (#147432)

Fixes #147394

References DXC for the implementation logic:

https://github.com/microsoft/DirectXShaderCompiler/blob/d751c827ed3b61e87fdf57d0f424cb2d7af30cd0/lib/HLSL/DxilPreparePasses.cpp#L693-L699

If DXIL Version < 1.6 then replace lifetime intrinsics with stores

- For validator version >= 1.6, store an undef
- For validator version < 1.6, store zeros

else keep the lifetime intrinsics in the DXIL.

After this PR, the number of DML shaders failing validation due to
#146974 is reduced from 157 to 50.

Author: Icohedron

llvm/lib/Target/DirectX/DXILOpLowering.cpp

@@ -39,11 +39,13 @@ class OpLowerer {
   DXILOpBuilder OpBuilder;
   DXILResourceMap &DRM;
   DXILResourceTypeMap &DRTM;
+  const ModuleMetadataInfo &MMDI;
   SmallVector<CallInst *> CleanupCasts;
 
 public:
-  OpLowerer(Module &M, DXILResourceMap &DRM, DXILResourceTypeMap &DRTM)
-      : M(M), OpBuilder(M), DRM(DRM), DRTM(DRTM) {}
+  OpLowerer(Module &M, DXILResourceMap &DRM, DXILResourceTypeMap &DRTM,
+            const ModuleMetadataInfo &MMDI)
+      : M(M), OpBuilder(M), DRM(DRM), DRTM(DRTM), MMDI(MMDI) {}
 
   /// Replace every call to \c F using \c ReplaceCall, and then erase \c F. If
   /// there is an error replacing a call, we emit a diagnostic and return true.
@@ -316,8 +318,7 @@ class OpLowerer {
   /// model and taking into account binding information from
   /// DXILResourceAnalysis.
   bool lowerHandleFromBinding(Function &F) {
-    const Triple &TT = M.getTargetTriple();
-    if (TT.getDXILVersion() < VersionTuple(1, 6))
+    if (MMDI.DXILVersion < VersionTuple(1, 6))
       return lowerToCreateHandle(F);
     return lowerToBindAndAnnotateHandle(F);
   }
@@ -486,8 +487,6 @@ class OpLowerer {
   }
 
   [[nodiscard]] bool lowerRawBufferLoad(Function &F) {
-    const Triple &TT = M.getTargetTriple();
-    VersionTuple DXILVersion = TT.getDXILVersion();
     const DataLayout &DL = F.getDataLayout();
     IRBuilder<> &IRB = OpBuilder.getIRB();
     Type *Int8Ty = IRB.getInt8Ty();
@@ -511,7 +510,7 @@ class OpLowerer {
           ConstantInt::get(Int32Ty, DL.getPrefTypeAlign(ScalarTy).value());
 
       Expected<CallInst *> OpCall =
-          DXILVersion >= VersionTuple(1, 2)
+          MMDI.DXILVersion >= VersionTuple(1, 2)
               ? OpBuilder.tryCreateOp(OpCode::RawBufferLoad,
                                       {Handle, Index0, Index1, Mask, Align},
                                       CI->getName(), NewRetTy)
@@ -586,8 +585,6 @@ class OpLowerer {
   }
 
   [[nodiscard]] bool lowerBufferStore(Function &F, bool IsRaw) {
-    const Triple &TT = M.getTargetTriple();
-    VersionTuple DXILVersion = TT.getDXILVersion();
     const DataLayout &DL = F.getDataLayout();
     IRBuilder<> &IRB = OpBuilder.getIRB();
     Type *Int8Ty = IRB.getInt8Ty();
@@ -654,7 +651,7 @@ class OpLowerer {
       SmallVector<Value *, 9> Args{
           Handle,          Index0,          Index1,          DataElements[0],
           DataElements[1], DataElements[2], DataElements[3], Mask};
-      if (IsRaw && DXILVersion >= VersionTuple(1, 2)) {
+      if (IsRaw && MMDI.DXILVersion >= VersionTuple(1, 2)) {
         Op = OpCode::RawBufferStore;
         // RawBufferStore requires the alignment
         Args.push_back(
@@ -745,6 +742,37 @@ class OpLowerer {
     });
   }
 
+  [[nodiscard]] bool lowerLifetimeIntrinsic(Function &F) {
+    IRBuilder<> &IRB = OpBuilder.getIRB();
+    return replaceFunction(F, [&](CallInst *CI) -> Error {
+      IRB.SetInsertPoint(CI);
+      Value *Ptr = CI->getArgOperand(1);
+      assert(Ptr->getType()->isPointerTy() &&
+             "Expected operand of lifetime intrinsic to be a pointer");
+
+      auto ZeroOrUndef = [&](Type *Ty) {
+        return MMDI.ValidatorVersion < VersionTuple(1, 6)
+                   ? Constant::getNullValue(Ty)
+                   : UndefValue::get(Ty);
+      };
+
+      Value *Val = nullptr;
+      if (auto *GV = dyn_cast<GlobalVariable>(Ptr)) {
+        if (GV->hasInitializer() || GV->isExternallyInitialized())
+          return Error::success();
+        Val = ZeroOrUndef(GV->getValueType());
+      } else if (auto *AI = dyn_cast<AllocaInst>(Ptr))
+        Val = ZeroOrUndef(AI->getAllocatedType());
+
+      assert(Val && "Expected operand of lifetime intrinsic to be a global "
+                    "variable or alloca instruction");
+      IRB.CreateStore(Val, Ptr, false);
+
+      CI->eraseFromParent();
+      return Error::success();
+    });
+  }
+
   [[nodiscard]] bool lowerIsFPClass(Function &F) {
     IRBuilder<> &IRB = OpBuilder.getIRB();
     Type *RetTy = IRB.getInt1Ty();
@@ -803,8 +831,6 @@ class OpLowerer {
       case Intrinsic::dx_resource_casthandle:
       // NOTE: llvm.dbg.value is supported as is in DXIL.
       case Intrinsic::dbg_value:
-      case Intrinsic::lifetime_start:
-      case Intrinsic::lifetime_end:
       case Intrinsic::not_intrinsic:
         if (F.use_empty())
           F.eraseFromParent();
@@ -855,6 +881,17 @@ class OpLowerer {
       case Intrinsic::ctpop:
         HasErrors |= lowerCtpopToCountBits(F);
         break;
+      case Intrinsic::lifetime_start:
+      case Intrinsic::lifetime_end:
+        if (F.use_empty())
+          F.eraseFromParent();
+        else {
+          if (MMDI.DXILVersion < VersionTuple(1, 6))
+            HasErrors |= lowerLifetimeIntrinsic(F);
+          else
+            continue;
+        }
+        break;
       case Intrinsic::is_fpclass:
         HasErrors |= lowerIsFPClass(F);
         break;
@@ -872,8 +909,9 @@ class OpLowerer {
 PreservedAnalyses DXILOpLowering::run(Module &M, ModuleAnalysisManager &MAM) {
   DXILResourceMap &DRM = MAM.getResult<DXILResourceAnalysis>(M);
   DXILResourceTypeMap &DRTM = MAM.getResult<DXILResourceTypeAnalysis>(M);
+  const ModuleMetadataInfo MMDI = MAM.getResult<DXILMetadataAnalysis>(M);
 
-  bool MadeChanges = OpLowerer(M, DRM, DRTM).lowerIntrinsics();
+  const bool MadeChanges = OpLowerer(M, DRM, DRTM, MMDI).lowerIntrinsics();
   if (!MadeChanges)
     return PreservedAnalyses::all();
   PreservedAnalyses PA;
@@ -891,8 +929,10 @@ class DXILOpLoweringLegacy : public ModulePass {
         getAnalysis<DXILResourceWrapperPass>().getResourceMap();
     DXILResourceTypeMap &DRTM =
         getAnalysis<DXILResourceTypeWrapperPass>().getResourceTypeMap();
+    const ModuleMetadataInfo MMDI =
+        getAnalysis<DXILMetadataAnalysisWrapperPass>().getModuleMetadata();
 
-    return OpLowerer(M, DRM, DRTM).lowerIntrinsics();
+    return OpLowerer(M, DRM, DRTM, MMDI).lowerIntrinsics();
   }
   StringRef getPassName() const override { return "DXIL Op Lowering"; }
   DXILOpLoweringLegacy() : ModulePass(ID) {}
@@ -901,6 +941,7 @@ class DXILOpLoweringLegacy : public ModulePass {
   void getAnalysisUsage(llvm::AnalysisUsage &AU) const override {
     AU.addRequired<DXILResourceTypeWrapperPass>();
     AU.addRequired<DXILResourceWrapperPass>();
+    AU.addRequired<DXILMetadataAnalysisWrapperPass>();
     AU.addPreserved<DXILResourceWrapperPass>();
     AU.addPreserved<DXILMetadataAnalysisWrapperPass>();
     AU.addPreserved<ShaderFlagsAnalysisWrapper>();

llvm/test/CodeGen/DirectX/legalize-lifetimes.ll renamed to llvm/test/CodeGen/DirectX/legalize-lifetimes-valver-1.5.ll

@@ -3,9 +3,9 @@
 ; CHECK-LABEL: define void @test_legal_lifetime() {
 ; CHECK-NEXT:    [[ACCUM_I_FLAT:%.*]] = alloca [1 x i32], align 4
 ; CHECK-NEXT:    [[GEP:%.*]] = getelementptr i32, ptr [[ACCUM_I_FLAT]], i32 0
-; CHECK-NEXT:    call void @llvm.lifetime.start.p0(i64 4, ptr nonnull [[ACCUM_I_FLAT]])
+; CHECK-NEXT:    store [1 x i32] zeroinitializer, ptr [[ACCUM_I_FLAT]], align 4
 ; CHECK-NEXT:    store i32 0, ptr [[GEP]], align 4
-; CHECK-NEXT:    call void @llvm.lifetime.end.p0(i64 4, ptr nonnull [[ACCUM_I_FLAT]])
+; CHECK-NEXT:    store [1 x i32] zeroinitializer, ptr [[ACCUM_I_FLAT]], align 4
 ; CHECK-NEXT:    ret void
 ;
 define void @test_legal_lifetime()  {
@@ -16,3 +16,7 @@ define void @test_legal_lifetime()  {
   call void @llvm.lifetime.end.p0(i64 4, ptr nonnull %accum.i.flat)
   ret void
 }
+
+; Set the validator version to 1.5
+!dx.valver = !{!0}
+!0 = !{i32 1, i32 5}

llvm/test/CodeGen/DirectX/legalize-lifetimes-valver-1.6.ll

@@ -0,0 +1,31 @@
+; RUN: opt -S -passes='dxil-op-lower' -mtriple=dxil-pc-shadermodel6.3-library %s | FileCheck %s --check-prefixes=CHECK,CHECK-SM63
+; RUN: opt -S -passes='dxil-op-lower' -mtriple=dxil-pc-shadermodel6.6-library %s | FileCheck %s --check-prefixes=CHECK,CHECK-SM66
+
+; CHECK-LABEL: define void @test_legal_lifetime() {
+; 
+; CHECK-SM63-NEXT:    [[ACCUM_I_FLAT:%.*]] = alloca [1 x i32], align 4
+; CHECK-SM63-NEXT:    [[GEP:%.*]] = getelementptr i32, ptr [[ACCUM_I_FLAT]], i32 0
+; CHECK-SM63-NEXT:    store [1 x i32] undef, ptr [[ACCUM_I_FLAT]], align 4
+; CHECK-SM63-NEXT:    store i32 0, ptr [[GEP]], align 4
+; CHECK-SM63-NEXT:    store [1 x i32] undef, ptr [[ACCUM_I_FLAT]], align 4
+; 
+; CHECK-SM66-NEXT:    [[ACCUM_I_FLAT:%.*]] = alloca [1 x i32], align 4
+; CHECK-SM66-NEXT:    [[GEP:%.*]] = getelementptr i32, ptr [[ACCUM_I_FLAT]], i32 0
+; CHECK-SM66-NEXT:    call void @llvm.lifetime.start.p0(i64 4, ptr nonnull [[ACCUM_I_FLAT]])
+; CHECK-SM66-NEXT:    store i32 0, ptr [[GEP]], align 4
+; CHECK-SM66-NEXT:    call void @llvm.lifetime.end.p0(i64 4, ptr nonnull [[ACCUM_I_FLAT]])
+; 
+; CHECK-NEXT:    ret void
+;
+define void @test_legal_lifetime()  {
+  %accum.i.flat = alloca [1 x i32], align 4
+  %gep = getelementptr i32, ptr %accum.i.flat, i32 0
+  call void @llvm.lifetime.start.p0(i64 4, ptr nonnull %accum.i.flat)
+  store i32 0, ptr %gep, align 4
+  call void @llvm.lifetime.end.p0(i64 4, ptr nonnull %accum.i.flat)
+  ret void
+}
+
+; Set the validator version to 1.6
+!dx.valver = !{!0}
+!0 = !{i32 1, i32 6}