[CIR] Upstream global initialization for ArrayType (#131657)

This change adds global initialization for ArrayType

Issue #130197

Author: AmrDeveloper

clang/include/clang/CIR/Dialect/IR/CIRAttrs.td

@@ -154,6 +154,48 @@ def FPAttr : CIR_Attr<"FP", "fp", [TypedAttrInterface]> {
   }];
 }
 
+
+//===----------------------------------------------------------------------===//
+// ConstArrayAttr
+//===----------------------------------------------------------------------===//
+
+def ConstArrayAttr : CIR_Attr<"ConstArray", "const_array", [TypedAttrInterface]> {
+  let summary = "A constant array from ArrayAttr or StringRefAttr";
+  let description = [{
+    An CIR array attribute is an array of literals of the specified attr types.
+  }];
+
+  let parameters = (ins AttributeSelfTypeParameter<"">:$type,
+                        "mlir::Attribute":$elts,
+                        "int":$trailingZerosNum);
+
+  // Define a custom builder for the type; that removes the need to pass
+  // in an MLIRContext instance, as it can be infered from the `type`.
+  let builders = [
+    AttrBuilderWithInferredContext<(ins "cir::ArrayType":$type,
+                                        "mlir::Attribute":$elts), [{
+      int zeros = 0;
+      auto typeSize = mlir::cast<cir::ArrayType>(type).getSize();
+      if (auto str = mlir::dyn_cast<mlir::StringAttr>(elts))
+        zeros = typeSize - str.size();
+      else
+        zeros = typeSize - mlir::cast<mlir::ArrayAttr>(elts).size();
+
+      return $_get(type.getContext(), type, elts, zeros);
+    }]>
+  ];
+
+  // Printing and parsing available in CIRDialect.cpp
+  let hasCustomAssemblyFormat = 1;
+
+  // Enable verifier.
+  let genVerifyDecl = 1;
+
+  let extraClassDeclaration = [{
+    bool hasTrailingZeros() const { return getTrailingZerosNum() != 0; };
+  }];
+}
+
 //===----------------------------------------------------------------------===//
 // ConstPtrAttr
 //===----------------------------------------------------------------------===//

clang/lib/CIR/CodeGen/CIRGenBuilder.h

@@ -45,6 +45,42 @@ class CIRGenBuilderTy : public cir::CIRBaseBuilderTy {
     return false;
   }
 
+  // Return true if the value is a null constant such as null pointer, (+0.0)
+  // for floating-point or zero initializer
+  bool isNullValue(mlir::Attribute attr) const {
+    if (mlir::isa<cir::ZeroAttr>(attr))
+      return true;
+
+    if (const auto ptrVal = mlir::dyn_cast<cir::ConstPtrAttr>(attr))
+      return ptrVal.isNullValue();
+
+    if (const auto intVal = mlir::dyn_cast<cir::IntAttr>(attr))
+      return intVal.isNullValue();
+
+    if (const auto boolVal = mlir::dyn_cast<cir::BoolAttr>(attr))
+      return !boolVal.getValue();
+
+    if (auto fpAttr = mlir::dyn_cast<cir::FPAttr>(attr)) {
+      auto fpVal = fpAttr.getValue();
+      bool ignored;
+      llvm::APFloat fv(+0.0);
+      fv.convert(fpVal.getSemantics(), llvm::APFloat::rmNearestTiesToEven,
+                 &ignored);
+      return fv.bitwiseIsEqual(fpVal);
+    }
+
+    if (const auto arrayVal = mlir::dyn_cast<cir::ConstArrayAttr>(attr)) {
+      if (mlir::isa<mlir::StringAttr>(arrayVal.getElts()))
+        return false;
+      for (const auto elt : mlir::cast<mlir::ArrayAttr>(arrayVal.getElts())) {
+        if (!isNullValue(elt))
+          return false;
+      }
+      return true;
+    }
+    return false;
+  }
+
   bool isInt(mlir::Type i) { return mlir::isa<cir::IntType>(i); }
 
   // Creates constant nullptr for pointer type ty.

clang/lib/CIR/CodeGen/CIRGenConstantEmitter.h

@@ -20,7 +20,6 @@
 
 #include "CIRGenFunction.h"
 #include "CIRGenModule.h"
-#include "llvm/ADT/SmallVector.h"
 
 namespace clang::CIRGen {
 
@@ -41,6 +40,9 @@ class ConstantEmitter {
   /// block addresses or PredefinedExprs.
   ConstantEmitter(CIRGenFunction &cgf) : cgm(cgf.cgm), cgf(&cgf) {}
 
+  ConstantEmitter(CIRGenModule &cgm, CIRGenFunction *cgf = nullptr)
+      : cgm(cgm), cgf(cgf) {}
+
   ConstantEmitter(const ConstantEmitter &other) = delete;
   ConstantEmitter &operator=(const ConstantEmitter &other) = delete;
 
@@ -66,7 +68,7 @@ class ConstantEmitter {
   mlir::Attribute emitAbstract(SourceLocation loc, const APValue &value,
                                QualType t);
 
-  mlir::Attribute tryEmitConstantExpr(const ConstantExpr *CE);
+  mlir::Attribute tryEmitConstantExpr(const ConstantExpr *ce);
 
   // These are private helper routines of the constant emitter that
   // can't actually be private because things are split out into helper

clang/lib/CIR/CodeGen/CIRGenDecl.cpp

@@ -225,7 +225,6 @@ void CIRGenFunction::emitScalarInit(const Expr *init, mlir::Location loc,
   }
   assert(!cir::MissingFeatures::emitNullabilityCheck());
   emitStoreThroughLValue(RValue::get(value), lvalue, true);
-  return;
 }
 
 void CIRGenFunction::emitExprAsInit(const Expr *init, const ValueDecl *d,

clang/lib/CIR/CodeGen/CIRGenExprConstant.cpp

@@ -158,13 +158,63 @@ class ConstExprEmitter
 
 // TODO(cir): this can be shared with LLVM's codegen
 static QualType getNonMemoryType(CIRGenModule &cgm, QualType type) {
-  if (auto at = type->getAs<AtomicType>()) {
+  if (const auto *at = type->getAs<AtomicType>()) {
     return cgm.getASTContext().getQualifiedType(at->getValueType(),
                                                 type.getQualifiers());
   }
   return type;
 }
 
+static mlir::Attribute
+emitArrayConstant(CIRGenModule &cgm, mlir::Type desiredType,
+                  mlir::Type commonElementType, unsigned arrayBound,
+                  SmallVectorImpl<mlir::TypedAttr> &elements,
+                  mlir::TypedAttr filler) {
+  const CIRGenBuilderTy &builder = cgm.getBuilder();
+
+  unsigned nonzeroLength = arrayBound;
+  if (elements.size() < nonzeroLength && builder.isNullValue(filler))
+    nonzeroLength = elements.size();
+
+  if (nonzeroLength == elements.size()) {
+    while (nonzeroLength > 0 &&
+           builder.isNullValue(elements[nonzeroLength - 1]))
+      --nonzeroLength;
+  }
+
+  if (nonzeroLength == 0)
+    return cir::ZeroAttr::get(builder.getContext(), desiredType);
+
+  const unsigned trailingZeroes = arrayBound - nonzeroLength;
+
+  // Add a zeroinitializer array filler if we have lots of trailing zeroes.
+  if (trailingZeroes >= 8) {
+    assert(elements.size() >= nonzeroLength &&
+           "missing initializer for non-zero element");
+  } else if (elements.size() != arrayBound) {
+    elements.resize(arrayBound, filler);
+
+    if (filler.getType() != commonElementType)
+      commonElementType = {};
+  }
+
+  if (commonElementType) {
+    SmallVector<mlir::Attribute, 4> eles;
+    eles.reserve(elements.size());
+
+    for (const auto &element : elements)
+      eles.push_back(element);
+
+    return cir::ConstArrayAttr::get(
+        cir::ArrayType::get(builder.getContext(), commonElementType,
+                            arrayBound),
+        mlir::ArrayAttr::get(builder.getContext(), eles));
+  }
+
+  cgm.errorNYI("array with different type elements");
+  return {};
+}
+
 //===----------------------------------------------------------------------===//
 //                             ConstantEmitter
 //===----------------------------------------------------------------------===//
@@ -271,16 +321,57 @@ mlir::Attribute ConstantEmitter::tryEmitPrivate(const APValue &value,
         cgm.getASTContext().getTargetInfo().useFP16ConversionIntrinsics()) {
       cgm.errorNYI("ConstExprEmitter::tryEmitPrivate half");
       return {};
-    } else {
-      mlir::Type ty = cgm.convertType(destType);
-      assert(mlir::isa<cir::CIRFPTypeInterface>(ty) &&
-             "expected floating-point type");
-      return cgm.getBuilder().getAttr<cir::FPAttr>(ty, init);
     }
+
+    mlir::Type ty = cgm.convertType(destType);
+    assert(mlir::isa<cir::CIRFPTypeInterface>(ty) &&
+           "expected floating-point type");
+    return cgm.getBuilder().getAttr<cir::FPAttr>(ty, init);
   }
   case APValue::Array: {
-    cgm.errorNYI("ConstExprEmitter::tryEmitPrivate array");
-    return {};
+    const ArrayType *arrayTy = cgm.getASTContext().getAsArrayType(destType);
+    const QualType arrayElementTy = arrayTy->getElementType();
+    const unsigned numElements = value.getArraySize();
+    const unsigned numInitElts = value.getArrayInitializedElts();
+
+    mlir::Attribute filler;
+    if (value.hasArrayFiller()) {
+      filler = tryEmitPrivate(value.getArrayFiller(), arrayElementTy);
+      if (!filler)
+        return {};
+    }
+
+    SmallVector<mlir::TypedAttr, 16> elements;
+    if (filler && builder.isNullValue(filler))
+      elements.reserve(numInitElts + 1);
+    else
+      elements.reserve(numInitElts);
+
+    mlir::Type commonElementType;
+    for (unsigned i = 0; i < numInitElts; ++i) {
+      const APValue &arrayElement = value.getArrayInitializedElt(i);
+      const mlir::Attribute element =
+          tryEmitPrivateForMemory(arrayElement, arrayElementTy);
+      if (!element)
+        return {};
+
+      const mlir::TypedAttr elementTyped = mlir::cast<mlir::TypedAttr>(element);
+      if (i == 0)
+        commonElementType = elementTyped.getType();
+      else if (elementTyped.getType() != commonElementType) {
+        commonElementType = {};
+      }
+
+      elements.push_back(elementTyped);
+    }
+
+    mlir::TypedAttr typedFiller = llvm::cast_or_null<mlir::TypedAttr>(filler);
+    if (filler && !typedFiller)
+      cgm.errorNYI("array filler should always be typed");
+
+    mlir::Type desiredType = cgm.convertType(destType);
+    return emitArrayConstant(cgm, desiredType, commonElementType, numElements,
+                             elements, typedFiller);
   }
   case APValue::Vector: {
     cgm.errorNYI("ConstExprEmitter::tryEmitPrivate vector");
@@ -290,9 +381,23 @@ mlir::Attribute ConstantEmitter::tryEmitPrivate(const APValue &value,
     cgm.errorNYI("ConstExprEmitter::tryEmitPrivate member pointer");
     return {};
   }
-  case APValue::LValue:
-    cgm.errorNYI("ConstExprEmitter::tryEmitPrivate lvalue");
+  case APValue::LValue: {
+
+    if (value.getLValueBase()) {
+      cgm.errorNYI("non-null pointer initialization");
+    } else {
+
+      mlir::Type desiredType = cgm.convertType(destType);
+      if (const cir::PointerType ptrType =
+              mlir::dyn_cast<cir::PointerType>(desiredType)) {
+        return builder.getConstPtrAttr(ptrType,
+                                       value.getLValueOffset().getQuantity());
+      } else {
+        llvm_unreachable("non-pointer variable initialized with a pointer");
+      }
+    }
     return {};
+  }
   case APValue::Struct:
   case APValue::Union:
     cgm.errorNYI("ConstExprEmitter::tryEmitPrivate struct or union");

clang/lib/CIR/CodeGen/CIRGenModule.cpp

@@ -11,6 +11,7 @@
 //===----------------------------------------------------------------------===//
 
 #include "CIRGenModule.h"
+#include "CIRGenConstantEmitter.h"
 #include "CIRGenFunction.h"
 
 #include "clang/AST/ASTContext.h"
@@ -128,7 +129,8 @@ void CIRGenModule::emitGlobalFunctionDefinition(clang::GlobalDecl gd,
 
 void CIRGenModule::emitGlobalVarDefinition(const clang::VarDecl *vd,
                                            bool isTentative) {
-  mlir::Type type = convertType(vd->getType());
+  const QualType astTy = vd->getType();
+  const mlir::Type type = convertType(vd->getType());
   if (clang::IdentifierInfo *identifier = vd->getIdentifier()) {
     auto varOp = builder.create<cir::GlobalOp>(getLoc(vd->getSourceRange()),
                                                identifier->getName(), type);
@@ -141,38 +143,8 @@ void CIRGenModule::emitGlobalVarDefinition(const clang::VarDecl *vd,
     if (initExpr) {
       mlir::Attribute initializer;
       if (APValue *value = initDecl->evaluateValue()) {
-        switch (value->getKind()) {
-        case APValue::Int: {
-          if (mlir::isa<cir::BoolType>(type))
-            initializer =
-                builder.getCIRBoolAttr(value->getInt().getZExtValue());
-          else
-            initializer = builder.getAttr<cir::IntAttr>(type, value->getInt());
-          break;
-        }
-        case APValue::Float: {
-          initializer = builder.getAttr<cir::FPAttr>(type, value->getFloat());
-          break;
-        }
-        case APValue::LValue: {
-          if (value->getLValueBase()) {
-            errorNYI(initExpr->getSourceRange(),
-                     "non-null pointer initialization");
-          } else {
-            if (auto ptrType = mlir::dyn_cast<cir::PointerType>(type)) {
-              initializer = builder.getConstPtrAttr(
-                  ptrType, value->getLValueOffset().getQuantity());
-            } else {
-              llvm_unreachable(
-                  "non-pointer variable initialized with a pointer");
-            }
-          }
-          break;
-        }
-        default:
-          errorNYI(initExpr->getSourceRange(), "unsupported initializer kind");
-          break;
-        }
+        ConstantEmitter emitter(*this);
+        initializer = emitter.tryEmitPrivateForMemory(*value, astTy);
       } else {
         errorNYI(initExpr->getSourceRange(), "non-constant initializer");
       }