[CIR][ABI] Implement basic struct CC lowering for x86_64 (#784)

This patch adds the necessary bits for unraveling struct arguments and
return values for the x86_64 calling convention.

Author: sitio-couto

clang/include/clang/CIR/Dialect/IR/CIRDataLayout.h

@@ -12,51 +12,52 @@
 #ifndef LLVM_CLANG_CIR_DIALECT_IR_CIRDATALAYOUT_H
 #define LLVM_CLANG_CIR_DIALECT_IR_CIRDATALAYOUT_H
 
-#include "mlir/Dialect/DLTI/DLTI.h"
 #include "mlir/IR/BuiltinOps.h"
 #include "clang/CIR/Dialect/IR/CIRTypes.h"
-#include "llvm/ADT/StringRef.h"
+#include "llvm/IR/DataLayout.h"
+#include "llvm/Support/Alignment.h"
+#include "llvm/Support/TypeSize.h"
 
 namespace cir {
 
+class StructLayout;
+
+// FIXME(cir): This might be replaced by a CIRDataLayout interface which can
+// provide the same functionalities.
 class CIRDataLayout {
   bool bigEndian = false;
 
+  /// Primitive type alignment data. This is sorted by type and bit
+  /// width during construction.
+  llvm::DataLayout::PrimitiveSpec StructAlignment;
+
+  // The StructType -> StructLayout map.
+  mutable void *LayoutMap = nullptr;
+
 public:
   mlir::DataLayout layout;
 
-  /// Constructs a DataLayout from a specification string. See reset().
-  explicit CIRDataLayout(llvm::StringRef dataLayout, mlir::ModuleOp module)
-      : layout(module) {
-    reset(dataLayout);
-  }
+  /// Constructs a DataLayout the module's data layout attribute.
+  CIRDataLayout(mlir::ModuleOp modOp);
 
   /// Parse a data layout string (with fallback to default values).
-  void reset(llvm::StringRef dataLayout);
+  void reset();
 
   // Free all internal data structures.
   void clear();
 
-  CIRDataLayout(mlir::ModuleOp modOp);
-
   bool isBigEndian() const { return bigEndian; }
 
-  // `useABI` is `true` if not using prefered alignment.
-  unsigned getAlignment(mlir::Type ty, bool useABI) const {
-    if (llvm::isa<mlir::cir::StructType>(ty)) {
-      auto sTy = mlir::cast<mlir::cir::StructType>(ty);
-      if (sTy.getPacked() && useABI)
-        return 1;
-    } else if (llvm::isa<mlir::cir::ArrayType>(ty)) {
-      return getAlignment(mlir::cast<mlir::cir::ArrayType>(ty).getEltType(),
-                          useABI);
-    }
-
-    return useABI ? layout.getTypeABIAlignment(ty)
-                  : layout.getTypePreferredAlignment(ty);
-  }
+  /// Returns a StructLayout object, indicating the alignment of the
+  /// struct, its size, and the offsets of its fields.
+  ///
+  /// Note that this information is lazily cached.
+  const StructLayout *getStructLayout(mlir::cir::StructType Ty) const;
+
+  /// Internal helper method that returns requested alignment for type.
+  llvm::Align getAlignment(mlir::Type Ty, bool abi_or_pref) const;
 
-  unsigned getABITypeAlign(mlir::Type ty) const {
+  llvm::Align getABITypeAlign(mlir::Type ty) const {
     return getAlignment(ty, true);
   }
 
@@ -67,10 +68,10 @@ class CIRDataLayout {
   /// the runtime size will be a positive integer multiple of the base size.
   ///
   /// For example, returns 5 for i36 and 10 for x86_fp80.
-  unsigned getTypeStoreSize(mlir::Type Ty) const {
-    // FIXME: this is a bit inaccurate, see DataLayout::getTypeStoreSize for
-    // more information.
-    return llvm::divideCeil(layout.getTypeSizeInBits(Ty), 8);
+  llvm::TypeSize getTypeStoreSize(mlir::Type Ty) const {
+    llvm::TypeSize BaseSize = getTypeSizeInBits(Ty);
+    return {llvm::divideCeil(BaseSize.getKnownMinValue(), 8),
+            BaseSize.isScalable()};
   }
 
   /// Returns the offset in bytes between successive objects of the
@@ -81,20 +82,20 @@ class CIRDataLayout {
   ///
   /// This is the amount that alloca reserves for this type. For example,
   /// returns 12 or 16 for x86_fp80, depending on alignment.
-  unsigned getTypeAllocSize(mlir::Type Ty) const {
+  llvm::TypeSize getTypeAllocSize(mlir::Type Ty) const {
     // Round up to the next alignment boundary.
-    return llvm::alignTo(getTypeStoreSize(Ty), getABITypeAlign(Ty));
+    return llvm::alignTo(getTypeStoreSize(Ty), getABITypeAlign(Ty).value());
   }
 
-  unsigned getPointerTypeSizeInBits(mlir::Type Ty) const {
+  llvm::TypeSize getPointerTypeSizeInBits(mlir::Type Ty) const {
     assert(mlir::isa<mlir::cir::PointerType>(Ty) &&
            "This should only be called with a pointer type");
     return layout.getTypeSizeInBits(Ty);
   }
 
-  unsigned getTypeSizeInBits(mlir::Type Ty) const {
-    return layout.getTypeSizeInBits(Ty);
-  }
+  // The implementation of this method is provided inline as it is particularly
+  // well suited to constant folding when called on a specific Type subclass.
+  llvm::TypeSize getTypeSizeInBits(mlir::Type Ty) const;
 
   mlir::Type getIntPtrType(mlir::Type Ty) const {
     assert(mlir::isa<mlir::cir::PointerType>(Ty) && "Expected pointer type");
@@ -104,6 +105,58 @@ class CIRDataLayout {
   }
 };
 
+/// Used to lazily calculate structure layout information for a target machine,
+/// based on the DataLayout structure.
+class StructLayout final
+    : public llvm::TrailingObjects<StructLayout, llvm::TypeSize> {
+  llvm::TypeSize StructSize;
+  llvm::Align StructAlignment;
+  unsigned IsPadded : 1;
+  unsigned NumElements : 31;
+
+public:
+  llvm::TypeSize getSizeInBytes() const { return StructSize; }
+
+  llvm::TypeSize getSizeInBits() const { return 8 * StructSize; }
+
+  llvm::Align getAlignment() const { return StructAlignment; }
+
+  /// Returns whether the struct has padding or not between its fields.
+  /// NB: Padding in nested element is not taken into account.
+  bool hasPadding() const { return IsPadded; }
+
+  /// Given a valid byte offset into the structure, returns the structure
+  /// index that contains it.
+  unsigned getElementContainingOffset(uint64_t FixedOffset) const;
+
+  llvm::MutableArrayRef<llvm::TypeSize> getMemberOffsets() {
+    return llvm::MutableArrayRef(getTrailingObjects<llvm::TypeSize>(),
+                                 NumElements);
+  }
+
+  llvm::ArrayRef<llvm::TypeSize> getMemberOffsets() const {
+    return llvm::ArrayRef(getTrailingObjects<llvm::TypeSize>(), NumElements);
+  }
+
+  llvm::TypeSize getElementOffset(unsigned Idx) const {
+    assert(Idx < NumElements && "Invalid element idx!");
+    return getMemberOffsets()[Idx];
+  }
+
+  llvm::TypeSize getElementOffsetInBits(unsigned Idx) const {
+    return getElementOffset(Idx) * 8;
+  }
+
+private:
+  friend class CIRDataLayout; // Only DataLayout can create this class
+
+  StructLayout(mlir::cir::StructType ST, const CIRDataLayout &DL);
+
+  size_t numTrailingObjects(OverloadToken<llvm::TypeSize>) const {
+    return NumElements;
+  }
+};
+
 } // namespace cir
 
 #endif

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

@@ -589,6 +589,12 @@ def StoreOp : CIR_Op<"store", [
     ```
   }];
 
+  let builders = [
+    OpBuilder<(ins "Value":$value, "Value":$addr), [{
+      $_state.addOperands({value, addr});
+    }]>
+  ];
+
   let arguments = (ins CIR_AnyType:$value,
                        Arg<CIR_PointerType, "the address to store the value",
                            [MemWrite]>:$addr,

clang/include/clang/CIR/MissingFeatures.h

@@ -70,6 +70,7 @@ struct MissingFeatures {
   // ObjC
   static bool setObjCGCLValueClass() { return false; }
   static bool objCLifetime() { return false; }
+  static bool objCIvarDecls() { return false; }
 
   // Debug info
   static bool generateDebugInfo() { return false; }
@@ -205,17 +206,42 @@ struct MissingFeatures {
 
   //-- Missing AST queries
 
-  static bool recordDeclCanPassInRegisters() { return false; }
+  static bool CXXRecordDeclIsEmptyCXX11() { return false; }
+  static bool CXXRecordDeclIsPOD() { return false; }
+  static bool CXXRecordIsDynamicClass() { return false; }
+  static bool astContextGetExternalSource() { return false; }
+  static bool declGetMaxAlignment() { return false; }
+  static bool declHasAlignMac68kAttr() { return false; }
+  static bool declHasAlignNaturalAttr() { return false; }
+  static bool declHasMaxFieldAlignmentAttr() { return false; }
+  static bool fieldDeclIsBitfield() { return false; }
+  static bool fieldDeclIsPotentiallyOverlapping() { return false; }
+  static bool fieldDeclGetMaxFieldAlignment() { return false; }
+  static bool fieldDeclisUnnamedBitField() { return false; }
   static bool funcDeclIsCXXConstructorDecl() { return false; }
   static bool funcDeclIsCXXDestructorDecl() { return false; }
   static bool funcDeclIsCXXMethodDecl() { return false; }
   static bool funcDeclIsInlineBuiltinDeclaration() { return false; }
   static bool funcDeclIsReplaceableGlobalAllocationFunction() { return false; }
+  static bool isCXXRecordDecl() { return false; }
   static bool qualTypeIsReferenceType() { return false; }
-  static bool typeGetAsEnumType() { return false; }
+  static bool recordDeclCanPassInRegisters() { return false; }
+  static bool recordDeclHasAlignmentAttr() { return false; }
+  static bool recordDeclHasFlexibleArrayMember() { return false; }
+  static bool recordDeclIsCXXDecl() { return false; }
+  static bool recordDeclIsMSStruct() { return false; }
+  static bool recordDeclIsPacked() { return false; }
+  static bool recordDeclMayInsertExtraPadding() { return false; }
   static bool typeGetAsBuiltinType() { return false; }
+  static bool typeGetAsEnumType() { return false; }
+  static bool typeIsCXXRecordDecl() { return false; }
+  static bool typeIsScalableType() { return false; }
+  static bool typeIsSized() { return false; }
   static bool varDeclIsKNRPromoted() { return false; }
 
+  // We need to track parent (base) classes to determine the layout of a class.
+  static bool getCXXRecordBases() { return false; }
+
   //-- Missing types
 
   static bool fixedWidthIntegers() { return false; }
@@ -232,6 +258,18 @@ struct MissingFeatures {
 
   //-- Other missing features
 
+  // We need to track the parent record types that represent a field
+  // declaration. This is necessary to determine the layout of a class.
+  static bool fieldDeclAbstraction() { return false; }
+
+  // There are some padding diagnostic features for Itanium ABI that we might
+  // wanna add later.
+  static bool bitFieldPaddingDiagnostics() { return false; }
+
+  // Clang considers both enums and records as tag types. We don't have a way to
+  // transparently handle both these types yet. Might need an interface here.
+  static bool tagTypeClassAbstraction() { return false; }
+
   // Empty values might be passed as arguments to serve as padding, ensuring
   // alignment and compliance (e.g. MIPS). We do not yet support this.
   static bool argumentPadding() { return false; }
@@ -268,7 +306,7 @@ struct MissingFeatures {
   // evaluating ABI-specific lowering.
   static bool qualifiedTypes() { return false; }
 
-  // We're ignoring several details regarding ABI-halding for Swift.
+  // We're ignoring several details regarding ABI-handling for Swift.
   static bool swift() { return false; }
 
   // The AppleARM64 is using ItaniumCXXABI, which is not quite right.
@@ -281,6 +319,7 @@ struct MissingFeatures {
   // If a store op is guaranteed to execute before the retun value load op, we
   // can optimize away the store and load ops. Seems like an early optimization.
   static bool returnValueDominatingStoreOptmiization() { return false; }
+
   // Globals (vars and functions) may have attributes that are target depedent.
   static bool setTargetAttributes() { return false; }
 

clang/lib/CIR/CodeGen/CIRGenBuilder.h

@@ -878,7 +878,7 @@ class CIRGenBuilderTy : public CIRBaseBuilderTy {
       unsigned Pos = 0;
       for (size_t I = 0; I < Elts.size(); ++I) {
         auto EltSize = Layout.getTypeAllocSize(Elts[I]);
-        unsigned AlignMask = Layout.getABITypeAlign(Elts[I]) - 1;
+        unsigned AlignMask = Layout.getABITypeAlign(Elts[I]).value() - 1;
         Pos = (Pos + AlignMask) & ~AlignMask;
         if (Offset < Pos + EltSize) {
           Indices.push_back(I);