[HLSL] Adding Flatten and Branch if attributes with test fixes (llvm#122157)

- Adding the changes from PRs: 
  - llvm#116331 
  - llvm#121852 
- Fixes test `tools/dxil-dis/debug-info.ll`
- Address some missed comments in the previous PR

---------

Co-authored-by: joaosaffran <[email protected]>

Author: joaosaffran

clang/include/clang/Basic/Attr.td

@@ -4353,6 +4353,16 @@ def HLSLLoopHint: StmtAttr {
   let Documentation = [HLSLLoopHintDocs, HLSLUnrollHintDocs];
 }
 
+def HLSLControlFlowHint: StmtAttr {
+  /// [branch]
+  /// [flatten]
+  let Spellings = [Microsoft<"branch">, Microsoft<"flatten">];
+  let Subjects = SubjectList<[IfStmt],
+                              ErrorDiag, "'if' statements">;
+  let LangOpts = [HLSL];
+  let Documentation = [InternalOnly];
+}
+
 def CapturedRecord : InheritableAttr {
   // This attribute has no spellings as it is only ever created implicitly.
   let Spellings = [];

clang/lib/CodeGen/CGStmt.cpp

@@ -760,6 +760,8 @@ void CodeGenFunction::EmitAttributedStmt(const AttributedStmt &S) {
   bool noinline = false;
   bool alwaysinline = false;
   bool noconvergent = false;
+  HLSLControlFlowHintAttr::Spelling flattenOrBranch =
+      HLSLControlFlowHintAttr::SpellingNotCalculated;
   const CallExpr *musttail = nullptr;
 
   for (const auto *A : S.getAttrs()) {
@@ -791,13 +793,17 @@ void CodeGenFunction::EmitAttributedStmt(const AttributedStmt &S) {
         Builder.CreateAssumption(AssumptionVal);
       }
     } break;
+    case attr::HLSLControlFlowHint: {
+      flattenOrBranch = cast<HLSLControlFlowHintAttr>(A)->getSemanticSpelling();
+    } break;
     }
   }
   SaveAndRestore save_nomerge(InNoMergeAttributedStmt, nomerge);
   SaveAndRestore save_noinline(InNoInlineAttributedStmt, noinline);
   SaveAndRestore save_alwaysinline(InAlwaysInlineAttributedStmt, alwaysinline);
   SaveAndRestore save_noconvergent(InNoConvergentAttributedStmt, noconvergent);
   SaveAndRestore save_musttail(MustTailCall, musttail);
+  SaveAndRestore save_flattenOrBranch(HLSLControlFlowAttr, flattenOrBranch);
   EmitStmt(S.getSubStmt(), S.getAttrs());
 }
 

clang/lib/CodeGen/CodeGenFunction.cpp

@@ -40,6 +40,7 @@
 #include "llvm/IR/DataLayout.h"
 #include "llvm/IR/Dominators.h"
 #include "llvm/IR/FPEnv.h"
+#include "llvm/IR/Instruction.h"
 #include "llvm/IR/IntrinsicInst.h"
 #include "llvm/IR/Intrinsics.h"
 #include "llvm/IR/MDBuilder.h"
@@ -2086,7 +2087,30 @@ void CodeGenFunction::EmitBranchOnBoolExpr(
     Weights = createProfileWeights(TrueCount, CurrentCount - TrueCount);
   }
 
-  Builder.CreateCondBr(CondV, TrueBlock, FalseBlock, Weights, Unpredictable);
+  llvm::Instruction *BrInst = Builder.CreateCondBr(CondV, TrueBlock, FalseBlock,
+                                                   Weights, Unpredictable);
+  switch (HLSLControlFlowAttr) {
+  case HLSLControlFlowHintAttr::Microsoft_branch:
+  case HLSLControlFlowHintAttr::Microsoft_flatten: {
+    llvm::MDBuilder MDHelper(CGM.getLLVMContext());
+
+    llvm::ConstantInt *BranchHintConstant =
+        HLSLControlFlowAttr ==
+                HLSLControlFlowHintAttr::Spelling::Microsoft_branch
+            ? llvm::ConstantInt::get(CGM.Int32Ty, 1)
+            : llvm::ConstantInt::get(CGM.Int32Ty, 2);
+
+    SmallVector<llvm::Metadata *, 2> Vals(
+        {MDHelper.createString("hlsl.controlflow.hint"),
+         MDHelper.createConstant(BranchHintConstant)});
+    BrInst->setMetadata("hlsl.controlflow.hint",
+                        llvm::MDNode::get(CGM.getLLVMContext(), Vals));
+    break;
+  }
+  // This is required to avoid warnings during compilation
+  case HLSLControlFlowHintAttr::SpellingNotCalculated:
+    break;
+  }
 }
 
 /// ErrorUnsupported - Print out an error that codegen doesn't support the

clang/lib/CodeGen/CodeGenFunction.h

@@ -615,6 +615,10 @@ class CodeGenFunction : public CodeGenTypeCache {
   /// True if the current statement has noconvergent attribute.
   bool InNoConvergentAttributedStmt = false;
 
+  /// HLSL Branch attribute.
+  HLSLControlFlowHintAttr::Spelling HLSLControlFlowAttr =
+      HLSLControlFlowHintAttr::SpellingNotCalculated;
+
   // The CallExpr within the current statement that the musttail attribute
   // applies to.  nullptr if there is no 'musttail' on the current statement.
   const CallExpr *MustTailCall = nullptr;

clang/lib/Sema/SemaStmtAttr.cpp

@@ -619,6 +619,12 @@ static Attr *handleHLSLLoopHintAttr(Sema &S, Stmt *St, const ParsedAttr &A,
   return ::new (S.Context) HLSLLoopHintAttr(S.Context, A, UnrollFactor);
 }
 
+static Attr *handleHLSLControlFlowHint(Sema &S, Stmt *St, const ParsedAttr &A,
+                                       SourceRange Range) {
+
+  return ::new (S.Context) HLSLControlFlowHintAttr(S.Context, A);
+}
+
 static Attr *ProcessStmtAttribute(Sema &S, Stmt *St, const ParsedAttr &A,
                                   SourceRange Range) {
   if (A.isInvalid() || A.getKind() == ParsedAttr::IgnoredAttribute)
@@ -655,6 +661,8 @@ static Attr *ProcessStmtAttribute(Sema &S, Stmt *St, const ParsedAttr &A,
     return handleLoopHintAttr(S, St, A, Range);
   case ParsedAttr::AT_HLSLLoopHint:
     return handleHLSLLoopHintAttr(S, St, A, Range);
+  case ParsedAttr::AT_HLSLControlFlowHint:
+    return handleHLSLControlFlowHint(S, St, A, Range);
   case ParsedAttr::AT_OpenCLUnrollHint:
     return handleOpenCLUnrollHint(S, St, A, Range);
   case ParsedAttr::AT_Suppress:

clang/test/AST/HLSL/HLSLControlFlowHint.hlsl

@@ -0,0 +1,43 @@
+// RUN: %clang_cc1 -triple dxil-pc-shadermodel6.3-compute -ast-dump %s | FileCheck %s
+
+// CHECK: FunctionDecl 0x{{[0-9A-Fa-f]+}} <{{.*}}> {{.*}} used branch 'int (int)'
+// CHECK: AttributedStmt 0x{{[0-9A-Fa-f]+}} <<invalid sloc>
+// CHECK-NEXT: -HLSLControlFlowHintAttr 0x{{[0-9A-Fa-f]+}} <{{.*}}> branch
+export int branch(int X){
+    int resp;
+    [branch] if (X > 0) {
+        resp = -X;
+    } else {
+        resp = X * 2;
+    }
+
+    return resp;
+}
+
+// CHECK: FunctionDecl 0x{{[0-9A-Fa-f]+}} <{{.*}}> {{.*}} used flatten 'int (int)'
+// CHECK: AttributedStmt 0x{{[0-9A-Fa-f]+}} <<invalid sloc>
+// CHECK-NEXT: -HLSLControlFlowHintAttr 0x{{[0-9A-Fa-f]+}} <{{.*}}> flatten
+export int flatten(int X){
+    int resp;
+    [flatten] if (X > 0) {
+        resp = -X;
+    } else {
+        resp = X * 2;
+    }
+
+    return resp;
+}
+
+// CHECK: FunctionDecl 0x{{[0-9A-Fa-f]+}} <{{.*}}> {{.*}} used no_attr 'int (int)'
+// CHECK-NOT: AttributedStmt 0x{{[0-9A-Fa-f]+}} <<invalid sloc>
+// CHECK-NOT: -HLSLControlFlowHintAttr
+export int no_attr(int X){
+    int resp;
+    if (X > 0) {
+        resp = -X;
+    } else {
+        resp = X * 2;
+    }
+
+    return resp;
+}

clang/test/CodeGenHLSL/HLSLControlFlowHint.hlsl

@@ -0,0 +1,48 @@
+// RUN: %clang_cc1 -finclude-default-header -x hlsl -triple dxil-pc-shadermodel6.3-library %s -fnative-half-type -emit-llvm -o - | FileCheck %s
+// RUN: %clang_cc1 -finclude-default-header -x hlsl -triple spirv-vulkan-library %s -fnative-half-type -emit-llvm -o - | FileCheck %s
+
+// CHECK: define {{.*}} i32 {{.*}}test_branch{{.*}}(i32 {{.*}} [[VALD:%.*]])
+// CHECK: [[PARAM:%.*]] = load i32, ptr [[VALD]].addr, align 4
+// CHECK: [[CMP:%.*]] = icmp sgt i32 [[PARAM]], 0
+// CHECK: br i1 [[CMP]], label %if.then, label %if.else, !hlsl.controlflow.hint [[HINT_BRANCH:![0-9]+]]
+export int test_branch(int X){
+    int resp;
+    [branch] if (X > 0) {
+        resp = -X;
+    } else {
+        resp = X * 2;
+    }
+
+    return resp;
+}
+
+// CHECK: define {{.*}} i32 {{.*}}test_flatten{{.*}}(i32 {{.*}} [[VALD:%.*]])
+// CHECK: [[PARAM:%.*]] = load i32, ptr [[VALD]].addr, align 4
+// CHECK: [[CMP:%.*]] = icmp sgt i32 [[PARAM]], 0
+// CHECK: br i1 [[CMP]], label %if.then, label %if.else, !hlsl.controlflow.hint [[HINT_FLATTEN:![0-9]+]]
+export int test_flatten(int X){
+    int resp;
+    [flatten] if (X > 0) {
+        resp = -X;
+    } else {
+        resp = X * 2;
+    }
+
+    return resp;
+}
+
+// CHECK: define {{.*}} i32 {{.*}}test_no_attr{{.*}}(i32 {{.*}} [[VALD:%.*]])
+// CHECK-NOT: !hlsl.controlflow.hint
+export int test_no_attr(int X){
+    int resp;
+    if (X > 0) {
+        resp = -X;
+    } else {
+        resp = X * 2;
+    }
+
+    return resp;
+}
+
+//CHECK: [[HINT_BRANCH]] = !{!"hlsl.controlflow.hint", i32 1}
+//CHECK: [[HINT_FLATTEN]] = !{!"hlsl.controlflow.hint", i32 2}

llvm/include/llvm/IR/IntrinsicsSPIRV.td

@@ -33,7 +33,7 @@ let TargetPrefix = "spv" in {
   def int_spv_ptrcast : Intrinsic<[llvm_any_ty], [llvm_any_ty, llvm_metadata_ty, llvm_i32_ty], [ImmArg<ArgIndex<2>>]>;
   def int_spv_switch : Intrinsic<[], [llvm_any_ty, llvm_vararg_ty]>;
   def int_spv_loop_merge : Intrinsic<[], [llvm_vararg_ty]>;
-  def int_spv_selection_merge : Intrinsic<[], [llvm_vararg_ty]>;
+  def int_spv_selection_merge : Intrinsic<[], [llvm_any_ty, llvm_i32_ty], [ImmArg<ArgIndex<1>>]>;
   def int_spv_cmpxchg : Intrinsic<[llvm_i32_ty], [llvm_any_ty, llvm_vararg_ty]>;
   def int_spv_unreachable : Intrinsic<[], []>;
   def int_spv_alloca : Intrinsic<[llvm_any_ty], [llvm_i8_ty], [ImmArg<ArgIndex<0>>]>;

llvm/lib/Target/DirectX/DXILTranslateMetadata.cpp

@@ -15,12 +15,14 @@
 #include "llvm/ADT/Twine.h"
 #include "llvm/Analysis/DXILMetadataAnalysis.h"
 #include "llvm/Analysis/DXILResource.h"
+#include "llvm/IR/BasicBlock.h"
 #include "llvm/IR/Constants.h"
 #include "llvm/IR/DiagnosticInfo.h"
 #include "llvm/IR/DiagnosticPrinter.h"
 #include "llvm/IR/Function.h"
 #include "llvm/IR/IRBuilder.h"
 #include "llvm/IR/LLVMContext.h"
+#include "llvm/IR/MDBuilder.h"
 #include "llvm/IR/Metadata.h"
 #include "llvm/IR/Module.h"
 #include "llvm/InitializePasses.h"
@@ -300,6 +302,38 @@ static MDTuple *emitTopLevelLibraryNode(Module &M, MDNode *RMD,
   return constructEntryMetadata(nullptr, nullptr, RMD, Properties, Ctx);
 }
 
+// TODO: We might need to refactor this to be more generic,
+// in case we need more metadata to be replaced.
+static void translateBranchMetadata(Module &M) {
+  for (Function &F : M) {
+    for (BasicBlock &BB : F) {
+      Instruction *BBTerminatorInst = BB.getTerminator();
+
+      MDNode *HlslControlFlowMD =
+          BBTerminatorInst->getMetadata("hlsl.controlflow.hint");
+
+      if (!HlslControlFlowMD)
+        continue;
+
+      assert(HlslControlFlowMD->getNumOperands() == 2 &&
+             "invalid operands for hlsl.controlflow.hint");
+
+      MDBuilder MDHelper(M.getContext());
+      ConstantInt *Op1 =
+          mdconst::extract<ConstantInt>(HlslControlFlowMD->getOperand(1));
+
+      SmallVector<llvm::Metadata *, 2> Vals(
+          ArrayRef<Metadata *>{MDHelper.createString("dx.controlflow.hints"),
+                               MDHelper.createConstant(Op1)});
+
+      MDNode *MDNode = llvm::MDNode::get(M.getContext(), Vals);
+
+      BBTerminatorInst->setMetadata("dx.controlflow.hints", MDNode);
+      BBTerminatorInst->setMetadata("hlsl.controlflow.hint", nullptr);
+    }
+  }
+}
+
 static void translateMetadata(Module &M, DXILBindingMap &DBM,
                               DXILResourceTypeMap &DRTM,
                               const Resources &MDResources,
@@ -372,6 +406,7 @@ PreservedAnalyses DXILTranslateMetadata::run(Module &M,
   const dxil::ModuleMetadataInfo MMDI = MAM.getResult<DXILMetadataAnalysis>(M);
 
   translateMetadata(M, DBM, DRTM, MDResources, ShaderFlags, MMDI);
+  translateBranchMetadata(M);
 
   return PreservedAnalyses::all();
 }
@@ -409,6 +444,7 @@ class DXILTranslateMetadataLegacy : public ModulePass {
         getAnalysis<DXILMetadataAnalysisWrapperPass>().getModuleMetadata();
 
     translateMetadata(M, DBM, DRTM, MDResources, ShaderFlags, MMDI);
+    translateBranchMetadata(M);
     return true;
   }
 };

llvm/lib/Target/SPIRV/SPIRVInstructionSelector.cpp

@@ -33,6 +33,7 @@
 #include "llvm/CodeGen/TargetOpcodes.h"
 #include "llvm/IR/IntrinsicsSPIRV.h"
 #include "llvm/Support/Debug.h"
+#include "llvm/Support/ErrorHandling.h"
 
 #define DEBUG_TYPE "spirv-isel"
 
@@ -45,6 +46,17 @@ using ExtInstList =
 
 namespace {
 
+llvm::SPIRV::SelectionControl::SelectionControl
+getSelectionOperandForImm(int Imm) {
+  if (Imm == 2)
+    return SPIRV::SelectionControl::Flatten;
+  if (Imm == 1)
+    return SPIRV::SelectionControl::DontFlatten;
+  if (Imm == 0)
+    return SPIRV::SelectionControl::None;
+  llvm_unreachable("Invalid immediate");
+}
+
 #define GET_GLOBALISEL_PREDICATE_BITSET
 #include "SPIRVGenGlobalISel.inc"
 #undef GET_GLOBALISEL_PREDICATE_BITSET
@@ -2818,19 +2830,24 @@ bool SPIRVInstructionSelector::selectIntrinsic(Register ResVReg,
     }
     return MIB.constrainAllUses(TII, TRI, RBI);
   }
-  case Intrinsic::spv_loop_merge:
-  case Intrinsic::spv_selection_merge: {
-    const auto Opcode = IID == Intrinsic::spv_selection_merge
-                            ? SPIRV::OpSelectionMerge
-                            : SPIRV::OpLoopMerge;
-    auto MIB = BuildMI(BB, I, I.getDebugLoc(), TII.get(Opcode));
+  case Intrinsic::spv_loop_merge: {
+    auto MIB = BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpLoopMerge));
     for (unsigned i = 1; i < I.getNumExplicitOperands(); ++i) {
       assert(I.getOperand(i).isMBB());
       MIB.addMBB(I.getOperand(i).getMBB());
     }
     MIB.addImm(SPIRV::SelectionControl::None);
     return MIB.constrainAllUses(TII, TRI, RBI);
   }
+  case Intrinsic::spv_selection_merge: {
+    auto MIB =
+        BuildMI(BB, I, I.getDebugLoc(), TII.get(SPIRV::OpSelectionMerge));
+    assert(I.getOperand(1).isMBB() &&
+           "operand 1 to spv_selection_merge must be a basic block");
+    MIB.addMBB(I.getOperand(1).getMBB());
+    MIB.addImm(getSelectionOperandForImm(I.getOperand(2).getImm()));
+    return MIB.constrainAllUses(TII, TRI, RBI);
+  }
   case Intrinsic::spv_cmpxchg:
     return selectAtomicCmpXchg(ResVReg, ResType, I);
   case Intrinsic::spv_unreachable: