[clang] [CIR] Support for static variables (PR #143980)

[Andy Kaylor via cfe-commits]

================
@@ -219,6 +236,252 @@ void CIRGenFunction::emitVarDecl(const VarDecl &d) {
   return emitAutoVarDecl(d);
 }
 
+static std::string getStaticDeclName(CIRGenModule &cgm, const VarDecl &d) {
+  if (cgm.getLangOpts().CPlusPlus)
+    return cgm.getMangledName(&d).str();
+
+  // If this isn't C++, we don't need a mangled name, just a pretty one.
+  assert(!d.isExternallyVisible() && "name shouldn't matter");
+  std::string contextName;
+  const DeclContext *dc = d.getDeclContext();
+  if (auto *cd = dyn_cast<CapturedDecl>(dc))
+    dc = cast<DeclContext>(cd->getNonClosureContext());
+  if (const auto *fd = dyn_cast<FunctionDecl>(dc))
+    contextName = std::string(cgm.getMangledName(fd));
+  else if (isa<BlockDecl>(dc))
+    cgm.errorNYI(d.getSourceRange(), "block decl context for static var");
+  else if (isa<ObjCMethodDecl>(dc))
+    cgm.errorNYI(d.getSourceRange(), "ObjC decl context for static var");
+  else
+    cgm.errorNYI(d.getSourceRange(), "Unknown context for static var decl");
+
+  contextName += "." + d.getNameAsString();
+  return contextName;
+}
+
+// TODO(cir): LLVM uses a Constant base class. Maybe CIR could leverage an
+// interface for all constants?
+cir::GlobalOp
+CIRGenModule::getOrCreateStaticVarDecl(const VarDecl &d,
+                                       cir::GlobalLinkageKind linkage) {
+  // In general, we don't always emit static var decls once before we reference
+  // them. It is possible to reference them before emitting the function that
+  // contains them, and it is possible to emit the containing function multiple
+  // times.
+  if (cir::GlobalOp existingGV = getStaticLocalDeclAddress(&d))
+    return existingGV;
+
+  QualType ty = d.getType();
+  assert(ty->isConstantSizeType() && "VLAs can't be static");
+
+  // Use the label if the variable is renamed with the asm-label extension.
+  std::string name;
+  if (d.hasAttr<AsmLabelAttr>())
+    errorNYI(d.getSourceRange(), "getOrCreateStaticVarDecl: asm label");
+  else
+    name = getStaticDeclName(*this, d);
+
+  mlir::Type lty = getTypes().convertTypeForMem(ty);
+  assert(!cir::MissingFeatures::addressSpace());
+
+  // OpenCL variables in local address space and CUDA shared
+  // variables cannot have an initializer.
+  mlir::Attribute init = nullptr;
+  if (d.hasAttr<LoaderUninitializedAttr>())
+    errorNYI(d.getSourceRange(),
+             "getOrCreateStaticVarDecl: LoaderUninitializedAttr");
+
+  init = builder.getZeroInitAttr(convertType(ty));
+
+  cir::GlobalOp gv = builder.createVersionedGlobal(
+      getModule(), getLoc(d.getLocation()), name, lty, linkage);
+  // TODO(cir): infer visibility from linkage in global op builder.
+  gv.setVisibility(getMLIRVisibilityFromCIRLinkage(linkage));
+  gv.setInitialValueAttr(init);
+  gv.setAlignment(getASTContext().getDeclAlign(&d).getAsAlign().value());
+
+  if (supportsCOMDAT() && gv.isWeakForLinker())
+    gv.setComdat(true);
+
+  assert(!cir::MissingFeatures::opGlobalThreadLocal());
+
+  setGVProperties(gv, &d);
+
+  // OG checks if the expected address space, denoted by the type, is the
+  // same as the actual address space indicated by attributes. If they aren't
+  // the same, an addrspacecast is emitted when this variable is accessed.
+  // In CIR however, cir.get_global already carries that information in
+  // !cir.ptr type - if this global is in OpenCL local address space, then its
+  // type would be !cir.ptr<..., addrspace(offload_local)>. Therefore we don't
+  // need an explicit address space cast in CIR: they will get emitted when
+  // lowering to LLVM IR.
+
+  // Ensure that the static local gets initialized by making sure the parent
+  // function gets emitted eventually.
+  const Decl *dc = cast<Decl>(d.getDeclContext());
+
+  // We can't name blocks or captured statements directly, so try to emit their
+  // parents.
+  if (isa<BlockDecl>(dc) || isa<CapturedDecl>(dc)) {
+    dc = dc->getNonClosureContext();
+    // FIXME: Ensure that global blocks get emitted.
+    if (!dc)
+      errorNYI(d.getSourceRange(), "non-closure context");
+  }
+
+  GlobalDecl gd;
+  if (isa<CXXConstructorDecl>(dc))
+    errorNYI(d.getSourceRange(), "C++ constructors static var context");
+  else if (isa<CXXDestructorDecl>(dc))
+    errorNYI(d.getSourceRange(), "C++ destructors static var context");
+  else if (const auto *fd = dyn_cast<FunctionDecl>(dc))
+    gd = GlobalDecl(fd);
+  else {
+    // Don't do anything for Obj-C method decls or global closures. We should
+    // never defer them.
+    assert(isa<ObjCMethodDecl>(dc) && "unexpected parent code decl");
+  }
+  if (gd.getDecl() && cir::MissingFeatures::openMP()) {
+    // Disable emission of the parent function for the OpenMP device codegen.
+    errorNYI(d.getSourceRange(), "OpenMP");
+  }
+
+  return gv;
+}
+
+/// Add the initializer for 'd' to the global variable that has already been
+/// created for it. If the initializer has a different type than gv does, this
+/// may free gv and return a different one. Otherwise it just returns gv.
+cir::GlobalOp CIRGenFunction::addInitializerToStaticVarDecl(
+    const VarDecl &d, cir::GlobalOp gv, cir::GetGlobalOp gvAddr) {
+  ConstantEmitter emitter(*this);
+  mlir::TypedAttr init =
+      mlir::cast<mlir::TypedAttr>(emitter.tryEmitForInitializer(d));
+
+  // If constant emission failed, then this should be a C++ static
+  // initializer.
+  if (!init) {
+    cgm.errorNYI(d.getSourceRange(), "static var without initializer");
+    return gv;
+  }
+
+  // TODO(cir): There should be debug code here to assert that the decl size
+  // matches the CIR data layout alloca size, but the code for calculating the
+  // alloca size is not implemented yet.
+  assert(!cir::MissingFeatures::dataLayoutAllocaSize());
+
+  // The initializer may differ in type from the global. Rewrite
+  // the global to match the initializer.  (We have to do this
+  // because some types, like unions, can't be completely represented
+  // in the LLVM type system.)
+  if (gv.getSymType() != init.getType()) {
+    gv.setSymType(init.getType());
+
+    // Normally this should be done with a call to cgm.replaceGlobal(oldGV, gv),
+    // but since at this point the current block hasn't been really attached,
+    // there's no visibility into the GetGlobalOp corresponding to this Global.
+    // Given those constraints, thread in the GetGlobalOp and update it
+    // directly.
+    assert(!cir::MissingFeatures::addressSpace());
+    gvAddr.getAddr().setType(builder.getPointerTo(init.getType()));
+  }
+
+  bool needsDtor =
+      d.needsDestruction(getContext()) == QualType::DK_cxx_destructor;
+
+  assert(!cir::MissingFeatures::opGlobalConstant());
+  gv.setInitialValueAttr(init);
+
+  emitter.finalize(gv);
+
+  if (needsDtor) {
+    // We have a constant initializer, but a nontrivial destructor. We still
+    // need to perform a guarded "initialization" in order to register the
+    // destructor.
+    cgm.errorNYI(d.getSourceRange(), "C++ guarded init");
+  }
+
+  return gv;
+}
+
+void CIRGenFunction::emitStaticVarDecl(const VarDecl &d,
+                                       cir::GlobalLinkageKind linkage) {
+  // Check to see if we already have a global variable for this
+  // declaration.  This can happen when double-emitting function
+  // bodies, e.g. with complete and base constructors.
+  cir::GlobalOp globalOp = cgm.getOrCreateStaticVarDecl(d, linkage);
+  // TODO(cir): we should have a way to represent global ops as values without
+  // having to emit a get global op. Sometimes these emissions are not used.
+  mlir::Value addr = builder.createGetGlobal(globalOp);
+  auto getAddrOp = mlir::cast<cir::GetGlobalOp>(addr.getDefiningOp());
+
+  CharUnits alignment = getContext().getDeclAlign(&d);
+
+  // Store into LocalDeclMap before generating initializer to handle
+  // circular references.
+  mlir::Type elemTy = convertTypeForMem(d.getType());
+  setAddrOfLocalVar(&d, Address(addr, elemTy, alignment));
+
+  // We can't have a VLA here, but we can have a pointer to a VLA,
+  // even though that doesn't really make any sense.
+  // Make sure to evaluate VLA bounds now so that we have them for later.
+  if (d.getType()->isVariablyModifiedType()) {
+    cgm.errorNYI(d.getSourceRange(),
+                 "emitStaticVarDecl: variably modified type");
+  }
+
+  // Save the type in case adding the initializer forces a type change.
+  mlir::Type expectedType = addr.getType();
+
+  cir::GlobalOp var = globalOp;
+
+  assert(!cir::MissingFeatures::cudaSupport());
+
+  // If this value has an initializer, emit it.
+  if (d.getInit())
+    var = addInitializerToStaticVarDecl(d, var, getAddrOp);
+
+  var.setAlignment(alignment.getAsAlign().value());
+
+  if (d.hasAttr<AnnotateAttr>())
+    cgm.errorNYI(d.getSourceRange(), "static var annotation");
+
+  if (d.getAttr<PragmaClangBSSSectionAttr>())
+    cgm.errorNYI(d.getSourceRange(), "static var BSS section attribute");
+  if (d.getAttr<PragmaClangDataSectionAttr>())
+    cgm.errorNYI(d.getSourceRange(), "static var Data section attribute");
+  if (d.getAttr<PragmaClangRodataSectionAttr>())
+    cgm.errorNYI(d.getSourceRange(), "static var Rodata section attribute");
+  if (d.getAttr<PragmaClangRelroSectionAttr>())
+    cgm.errorNYI(d.getSourceRange(), "static var Relro section attribute");
+
+  if (d.getAttr<SectionAttr>())
+    cgm.errorNYI(d.getSourceRange(),
+                 "static var object file section attribute");
+
+  if (d.hasAttr<RetainAttr>())
+    cgm.errorNYI(d.getSourceRange(), "static var retain attribute");
+  else if (d.hasAttr<UsedAttr>())
----------------
andykaylor wrote:

I suppose I should just have one check for any of a list of attributes. This is also not implemented in the incubator, FWIW. It seemed wrong to just put a pile of MissingFeatures here, so I did this. Classic codegen does this with `retain` and `used`:

```
  if (D.hasAttr<RetainAttr>())
    CGM.addUsedGlobal(var);
  else if (D.hasAttr<UsedAttr>())
    CGM.addUsedOrCompilerUsedGlobal(var);
```
I guess if it had `retain`, also handling `used` is redundant.

https://github.com/llvm/llvm-project/pull/143980