[CIR] Add limited support for array new (#1286)

This change adds initial support for array new expressions where the
array size is constant and the element does not require a cookie.

Author: andykaylor

clang/lib/CIR/CodeGen/CIRGenCXXABI.cpp

@@ -85,3 +85,18 @@ bool CIRGenCXXABI::isZeroInitializable(const MemberPointerType *MPT) {
   // Fake answer.
   return true;
 }
+
+CharUnits CIRGenCXXABI::getArrayCookieSize(const CXXNewExpr *E) {
+  if (!requiresArrayCookie(E))
+    return CharUnits::Zero();
+  llvm_unreachable("NYI");
+}
+
+bool CIRGenCXXABI::requiresArrayCookie(const CXXNewExpr *E) {
+  // If the class's usual deallocation function takes two arguments,
+  // it needs a cookie.
+  if (E->doesUsualArrayDeleteWantSize())
+    return true;
+
+  return E->getAllocatedType().isDestructedType();
+}

clang/lib/CIR/CodeGen/CIRGenCXXABI.h

@@ -38,6 +38,8 @@ class CIRGenCXXABI {
 
   clang::ASTContext &getContext() const { return CGM.getASTContext(); }
 
+  virtual bool requiresArrayCookie(const CXXNewExpr *E);
+
 public:
   /// Similar to AddedStructorArgs, but only notes the number of additional
   /// arguments.
@@ -347,6 +349,19 @@ class CIRGenCXXABI {
 
   virtual cir::MethodAttr buildVirtualMethodAttr(cir::MethodType MethodTy,
                                                  const CXXMethodDecl *MD) = 0;
+
+  /**************************** Array cookies ******************************/
+
+  /// Returns the extra size required in order to store the array
+  /// cookie for the given new-expression.  May return 0 to indicate that no
+  /// array cookie is required.
+  ///
+  /// Several cases are filtered out before this method is called:
+  ///   - non-array allocations never need a cookie
+  ///   - calls to \::operator new(size_t, void*) never need a cookie
+  ///
+  /// \param E - the new-expression being allocated.
+  virtual CharUnits getArrayCookieSize(const CXXNewExpr *E);
 };
 
 /// Creates and Itanium-family ABI

clang/lib/CIR/CodeGen/CIRGenExprCXX.cpp

@@ -13,6 +13,7 @@
 #include "clang/CIR/Dialect/IR/CIRAttrs.h"
 #include "clang/CIR/MissingFeatures.h"
 #include <CIRGenCXXABI.h>
+#include <CIRGenCstEmitter.h>
 #include <CIRGenFunction.h>
 #include <CIRGenModule.h>
 #include <CIRGenValue.h>
@@ -549,11 +550,25 @@ static UsualDeleteParams getUsualDeleteParams(const FunctionDecl *FD) {
   return Params;
 }
 
+static CharUnits CalculateCookiePadding(CIRGenFunction &CGF,
+                                        const CXXNewExpr *E) {
+  if (!E->isArray())
+    return CharUnits::Zero();
+
+  // No cookie is required if the operator new[] being used is the
+  // reserved placement operator new[].
+  if (E->getOperatorNew()->isReservedGlobalPlacementOperator())
+    return CharUnits::Zero();
+
+  return CGF.CGM.getCXXABI().getArrayCookieSize(E);
+}
+
 static mlir::Value emitCXXNewAllocSize(CIRGenFunction &CGF, const CXXNewExpr *e,
                                        unsigned minElements,
                                        mlir::Value &numElements,
                                        mlir::Value &sizeWithoutCookie) {
   QualType type = e->getAllocatedType();
+  mlir::Location Loc = CGF.getLoc(e->getSourceRange());
 
   if (!e->isArray()) {
     CharUnits typeSize = CGF.getContext().getTypeSizeInChars(type);
@@ -563,7 +578,96 @@ static mlir::Value emitCXXNewAllocSize(CIRGenFunction &CGF, const CXXNewExpr *e,
     return sizeWithoutCookie;
   }
 
-  llvm_unreachable("NYI");
+  // The width of size_t.
+  unsigned sizeWidth = CGF.CGM.getDataLayout().getTypeSizeInBits(CGF.SizeTy);
+
+  // The number of elements can be have an arbitrary integer type;
+  // essentially, we need to multiply it by a constant factor, add a
+  // cookie size, and verify that the result is representable as a
+  // size_t.  That's just a gloss, though, and it's wrong in one
+  // important way: if the count is negative, it's an error even if
+  // the cookie size would bring the total size >= 0.
+  //
+  // If the array size is constant, Sema will have prevented negative
+  // values and size overflow.
+
+  // Compute the constant factor.
+  llvm::APInt arraySizeMultiplier(sizeWidth, 1);
+  while (const ConstantArrayType *CAT =
+             CGF.getContext().getAsConstantArrayType(type)) {
+    type = CAT->getElementType();
+    arraySizeMultiplier *= CAT->getSize();
+  }
+
+  CharUnits typeSize = CGF.getContext().getTypeSizeInChars(type);
+  llvm::APInt typeSizeMultiplier(sizeWidth, typeSize.getQuantity());
+  typeSizeMultiplier *= arraySizeMultiplier;
+
+  // Figure out the cookie size.
+  llvm::APInt cookieSize(sizeWidth,
+                         CalculateCookiePadding(CGF, e).getQuantity());
+
+  // This will be a size_t.
+  mlir::Value size;
+
+  // Emit the array size expression.
+  // We multiply the size of all dimensions for NumElements.
+  // e.g for 'int[2][3]', ElemType is 'int' and NumElements is 6.
+  const Expr *arraySize = *e->getArraySize();
+  mlir::Attribute constNumElements =
+      ConstantEmitter(CGF.CGM, &CGF)
+          .tryEmitAbstract(arraySize, arraySize->getType());
+  if (constNumElements) {
+    // Get an APInt from the constant
+    const llvm::APInt &count =
+        mlir::cast<cir::IntAttr>(constNumElements).getValue();
+
+    unsigned numElementsWidth = count.getBitWidth();
+
+    // The equivalent code in CodeGen/CGExprCXX.cpp handles these cases as
+    // overflow, but they should never happen. The size argument is implicitly
+    // cast to a size_t, so it can never be negative and numElementsWidth will
+    // always equal sizeWidth.
+    assert(!count.isNegative() && "Expected non-negative array size");
+    assert(numElementsWidth == sizeWidth &&
+           "Expected a size_t array size constant");
+
+    // Okay, compute a count at the right width.
+    llvm::APInt adjustedCount = count.zextOrTrunc(sizeWidth);
+
+    // Scale numElements by that.  This might overflow, but we don't
+    // care because it only overflows if allocationSize does, too, and
+    // if that overflows then we shouldn't use this.
+    // This emits a constant that may not be used, but we can't tell here
+    // whether it will be needed or not.
+    numElements =
+        CGF.getBuilder().getConstInt(Loc, adjustedCount * arraySizeMultiplier);
+
+    // Compute the size before cookie, and track whether it overflowed.
+    bool overflow;
+    llvm::APInt allocationSize =
+        adjustedCount.umul_ov(typeSizeMultiplier, overflow);
+
+    // Sema prevents us from hitting this case
+    assert(!overflow && "Overflow in array allocation size");
+
+    // Add in the cookie, and check whether it's overflowed.
+    if (cookieSize != 0) {
+      llvm_unreachable("NYI");
+    }
+
+    size = CGF.getBuilder().getConstInt(Loc, allocationSize);
+  } else {
+    // TODO: Handle the variable size case
+    llvm_unreachable("NYI");
+  }
+
+  if (cookieSize == 0)
+    sizeWithoutCookie = size;
+  else
+    assert(sizeWithoutCookie && "didn't set sizeWithoutCookie?");
+
+  return size;
 }
 
 namespace {
@@ -745,33 +849,32 @@ static void StoreAnyExprIntoOneUnit(CIRGenFunction &CGF, const Expr *Init,
   llvm_unreachable("bad evaluation kind");
 }
 
+void CIRGenFunction::emitNewArrayInitializer(
+    const CXXNewExpr *E, QualType ElementType, mlir::Type ElementTy,
+    Address BeginPtr, mlir::Value NumElements,
+    mlir::Value AllocSizeWithoutCookie) {
+  // If we have a type with trivial initialization and no initializer,
+  // there's nothing to do.
+  if (!E->hasInitializer())
+    return;
+
+  llvm_unreachable("NYI");
+}
+
 static void emitNewInitializer(CIRGenFunction &CGF, const CXXNewExpr *E,
                                QualType ElementType, mlir::Type ElementTy,
                                Address NewPtr, mlir::Value NumElements,
                                mlir::Value AllocSizeWithoutCookie) {
   assert(!cir::MissingFeatures::generateDebugInfo());
   if (E->isArray()) {
-    llvm_unreachable("NYI");
+    CGF.emitNewArrayInitializer(E, ElementType, ElementTy, NewPtr, NumElements,
+                                AllocSizeWithoutCookie);
   } else if (const Expr *Init = E->getInitializer()) {
     StoreAnyExprIntoOneUnit(CGF, Init, E->getAllocatedType(), NewPtr,
                             AggValueSlot::DoesNotOverlap);
   }
 }
 
-static CharUnits CalculateCookiePadding(CIRGenFunction &CGF,
-                                        const CXXNewExpr *E) {
-  if (!E->isArray())
-    return CharUnits::Zero();
-
-  // No cookie is required if the operator new[] being used is the
-  // reserved placement operator new[].
-  if (E->getOperatorNew()->isReservedGlobalPlacementOperator())
-    return CharUnits::Zero();
-
-  llvm_unreachable("NYI");
-  // return CGF.CGM.getCXXABI().GetArrayCookieSize(E);
-}
-
 namespace {
 /// Calls the given 'operator delete' on a single object.
 struct CallObjectDelete final : EHScopeStack::Cleanup {
@@ -1129,9 +1232,6 @@ mlir::Value CIRGenFunction::emitCXXNewExpr(const CXXNewExpr *E) {
   emitNewInitializer(*this, E, allocType, elementTy, result, numElements,
                      allocSizeWithoutCookie);
   auto resultPtr = result.getPointer();
-  if (E->isArray()) {
-    llvm_unreachable("NYI");
-  }
 
   // Deactivate the 'operator delete' cleanup if we finished
   // initialization.

clang/lib/CIR/CodeGen/CIRGenFunction.h

@@ -726,6 +726,11 @@ class CIRGenFunction : public CIRGenTypeCache {
   mlir::Value emitCXXNewExpr(const CXXNewExpr *E);
   void emitCXXDeleteExpr(const CXXDeleteExpr *E);
 
+  void emitNewArrayInitializer(const CXXNewExpr *E, QualType ElementType,
+                               mlir::Type ElementTy, Address BeginPtr,
+                               mlir::Value NumElements,
+                               mlir::Value AllocSizeWithoutCookie);
+
   void emitCXXAggrConstructorCall(const CXXConstructorDecl *D,
                                   const clang::ArrayType *ArrayTy,
                                   Address ArrayPtr, const CXXConstructExpr *E,

clang/test/CIR/CodeGen/new.cpp

@@ -56,3 +56,37 @@ void t() {
   B b;
   b.construct(&b);
 }
+
+
+void t_new_constant_size() {
+  auto p = new double[16];
+}
+
+// In this test, NUM_ELEMENTS isn't used because no cookie is needed and there
+//   are no constructor calls needed.
+
+// CHECK:   cir.func @_Z19t_new_constant_sizev()
+// CHECK:    %0 = cir.alloca !cir.ptr<!cir.double>, !cir.ptr<!cir.ptr<!cir.double>>, ["p", init] {alignment = 8 : i64}
+// CHECK:    %[[#NUM_ELEMENTS:]] = cir.const #cir.int<16> : !u64i
+// CHECK:    %[[#ALLOCATION_SIZE:]] = cir.const #cir.int<128> : !u64i
+// CHECK:    %3 = cir.call @_Znam(%[[#ALLOCATION_SIZE]]) : (!u64i) -> !cir.ptr<!void>
+// CHECK:    %4 = cir.cast(bitcast, %3 : !cir.ptr<!void>), !cir.ptr<!cir.double>
+// CHECK:    cir.store %4, %0 : !cir.ptr<!cir.double>, !cir.ptr<!cir.ptr<!cir.double>>
+// CHECK:    cir.return
+// CHECK:  }
+
+void t_new_multidim_constant_size() {
+  auto p = new double[2][3][4];
+}
+
+// As above, NUM_ELEMENTS isn't used.
+
+// CHECK:   cir.func @_Z28t_new_multidim_constant_sizev()
+// CHECK:    %0 = cir.alloca !cir.ptr<!cir.array<!cir.array<!cir.double x 4> x 3>>, !cir.ptr<!cir.ptr<!cir.array<!cir.array<!cir.double x 4> x 3>>>, ["p", init] {alignment = 8 : i64}
+// CHECK:    %[[#NUM_ELEMENTS:]] = cir.const #cir.int<24> : !u64i
+// CHECK:    %[[#ALLOCATION_SIZE:]] = cir.const #cir.int<192> : !u64i
+// CHECK:    %3 = cir.call @_Znam(%[[#ALLOCATION_SIZE]]) : (!u64i) -> !cir.ptr<!void>
+// CHECK:    %4 = cir.cast(bitcast, %3 : !cir.ptr<!void>), !cir.ptr<!cir.double>
+// CHECK:    %5 = cir.cast(bitcast, %0 : !cir.ptr<!cir.ptr<!cir.array<!cir.array<!cir.double x 4> x 3>>>), !cir.ptr<!cir.ptr<!cir.double>>
+// CHECK:    cir.store %4, %5 : !cir.ptr<!cir.double>, !cir.ptr<!cir.ptr<!cir.double>>
+// CHECK:  }

clang/test/CIR/Lowering/new.cpp

@@ -0,0 +1,20 @@
+// RUN: %clang_cc1 -triple x86_64-unknown-linux-gnu -fclangir -emit-llvm %s -o %t.ll
+// RUN: FileCheck --input-file=%t.ll %s -check-prefix=LLVM
+
+void t_new_constant_size() {
+  auto p = new double[16];
+}
+
+// LLVM: @_Z19t_new_constant_sizev()
+// LLVM:   %[[ALLOCA:.*]] = alloca ptr, i64 1, align 8
+// LLVM:   %[[ADDR:.*]] = call ptr @_Znam(i64 128)
+// LLVM:   store ptr %[[ADDR]], ptr %[[ALLOCA]], align 8
+
+void t_new_multidim_constant_size() {
+  auto p = new double[2][3][4];
+}
+
+// LLVM: @_Z28t_new_multidim_constant_sizev()
+// LLVM:   %[[ALLOCA:.*]] = alloca ptr, i64 1, align 8
+// LLVM:   %[[ADDR:.*]] = call ptr @_Znam(i64 192)
+// LLVM:   store ptr %[[ADDR]], ptr %[[ALLOCA]], align 8