carbon-language/carbon-lang
1
0
Fork 0
mirror of https://github.com/carbon-language/carbon-lang.git synced 2026-09-24 20:50:13 +01:00
Code Issues Packages Projects Releases 10 Wiki Activity
Files
709776ad1c0b4392b0b702cbdac4a0b86c2e1e34
carbon-lang/toolchain/lower/file_context.cpp
T
Richard Smith 709776ad1c Support for locations in transitively imported C++ code (#7093) 
Instead of treating all C++ code as coming from a single synthetic
`CheckIRId`, track the `SemIR::File` associated with each C++ location.
This is necessary since each `SemIR::File` has a distinct `CppFile` and
therefore distinct `SourceLocation`s and `ClangSourceLocId`s.

Assisted-by: Gemini via Antigravity
2026-04-23 20:11:04 +00:00

1635 lines
67 KiB
C++
Raw Blame History

// Part of the Carbon Language project, under the Apache License v2.0 with LLVM
// Exceptions. See /LICENSE for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
#include "toolchain/lower/file_context.h"
#include <memory>
#include <optional>
#include <string>
#include <utility>
#include "clang/CodeGen/ModuleBuilder.h"
#include "common/check.h"
#include "common/pretty_stack_trace_function.h"
#include "common/vlog.h"
#include "llvm/ADT/STLExtras.h"
#include "llvm/ADT/Sequence.h"
#include "llvm/Linker/Linker.h"
#include "llvm/Transforms/Utils/BasicBlockUtils.h"
#include "llvm/Transforms/Utils/ModuleUtils.h"
#include "toolchain/base/kind_switch.h"
#include "toolchain/lower/clang_global_decl.h"
#include "toolchain/lower/constant.h"
#include "toolchain/lower/function_context.h"
#include "toolchain/lower/options.h"
#include "toolchain/lower/specific_coalescer.h"
#include "toolchain/sem_ir/absolute_node_ref.h"
#include "toolchain/sem_ir/diagnostic_loc_converter.h"
#include "toolchain/sem_ir/entry_point.h"
#include "toolchain/sem_ir/expr_info.h"
#include "toolchain/sem_ir/file.h"
#include "toolchain/sem_ir/function.h"
#include "toolchain/sem_ir/generic.h"
#include "toolchain/sem_ir/ids.h"
#include "toolchain/sem_ir/inst.h"
#include "toolchain/sem_ir/inst_categories.h"
#include "toolchain/sem_ir/inst_kind.h"
#include "toolchain/sem_ir/mangler.h"
#include "toolchain/sem_ir/pattern.h"
#include "toolchain/sem_ir/stringify.h"
#include "toolchain/sem_ir/type_info.h"
#include "toolchain/sem_ir/typed_insts.h"
namespace Carbon::Lower {
FileContext::FileContext(Context& context, const SemIR::File& sem_ir,
const SemIR::InstNamer* inst_namer,
llvm::raw_ostream* vlog_stream)
: context_(&context),
sem_ir_(&sem_ir),
inst_namer_(inst_namer),
vlog_stream_(vlog_stream),
functions_(LoweredFunctionStore::MakeForOverwrite(sem_ir.functions())),
specific_functions_(sem_ir.specifics(), std::nullopt),
types_(LoweredTypeStore::MakeWithExplicitSize(
sem_ir.constant_values().ConcreteStoreSize(),
sem_ir.constant_values().GetTypeIdTag(), {nullptr, nullptr})),
constants_(LoweredConstantStore::MakeWithExplicitSize(
sem_ir.insts().size(), sem_ir.insts().GetIdTag(), nullptr)),
lowered_specifics_(sem_ir.generics(),
llvm::SmallVector<SemIR::SpecificId>()),
coalescer_(vlog_stream_, sem_ir.specifics()),
vtables_(decltype(vtables_)::MakeForOverwrite(sem_ir.vtables())),
specific_vtables_(sem_ir.specifics(), nullptr) {
// Initialization that relies on invariants of the class.
cpp_code_generator_ = cpp_file() ? cpp_file()->GetCodeGenerator() : nullptr;
CARBON_CHECK(
!cpp_code_generator_ ||
(&cpp_code_generator_->GetModule()->getContext() == &llvm_context()));
CARBON_CHECK(!sem_ir.has_errors(),
"Generating LLVM IR from invalid SemIR::File is unsupported.");
}
// TODO: Move this to lower.cpp.
auto FileContext::PrepareToLower() -> void {
// Lower all types that were required to be complete.
for (auto type_id : sem_ir_->types().complete_types()) {
if (type_id.index >= 0) {
types_.Set(type_id, BuildType(sem_ir_->types().GetTypeInstId(type_id)));
}
}
// Lower function declarations.
for (auto [id, function] : sem_ir_->functions().enumerate()) {
if (id == sem_ir().global_ctor_id()) {
// The global constructor is only lowered when we generate its definition.
// LLVM doesn't allow an internal linkage function to be undefined.
continue;
}
if (function.evaluation_mode == SemIR::Function::EvaluationMode::MustEval) {
// musteval functions are never lowered.
continue;
}
functions_.Set(id, BuildFunctionDecl(id));
}
// TODO: Split vtable declaration creation from definition creation to avoid
// redundant vtable definitions for imported vtables.
for (const auto& [id, vtable] : sem_ir_->vtables().enumerate()) {
const auto& class_info = sem_ir().classes().Get(vtable.class_id);
// Vtables can't be generated for generics, only for their specifics - and
// must be done lazily based on the use of those specifics.
if (!class_info.generic_id.has_value()) {
vtables_.Set(id, BuildVtable(vtable, SemIR::SpecificId::None));
}
}
// Lower constants.
LowerConstants(*this, constants_);
}
// TODO: Move this to lower.cpp.
auto FileContext::LowerDefinitions() -> void {
// Lower global variable definitions.
// TODO: Storing both a `constants_` array and a separate `global_variables_`
// map is redundant.
for (auto inst_id :
sem_ir().inst_blocks().Get(sem_ir().top_inst_block_id())) {
// Only `VarStorage` indicates a global variable declaration in the
// top instruction block.
if (auto var = sem_ir().insts().TryGetAs<SemIR::VarStorage>(inst_id)) {
// Get the global variable declaration. We created this when lowering the
// constant unless the variable is unnamed, in which case we need to
// create it now.
llvm::GlobalVariable* llvm_var = nullptr;
if (auto const_id = sem_ir().constant_values().Get(inst_id);
const_id.is_constant()) {
llvm_var = cast<llvm::GlobalVariable>(GetConstant(const_id, inst_id));
} else {
// We should never be emitting a definition for a C++ global variable.
llvm_var = BuildNonCppGlobalVariableDecl(*var);
}
// Convert the declaration of this variable into a definition by adding an
// initializer.
global_variables_.Insert(inst_id, llvm_var);
llvm_var->setInitializer(
llvm::Constant::getNullValue(llvm_var->getValueType()));
}
}
// Lower function definitions.
for (auto [id, fn_info] : sem_ir_->functions().enumerate()) {
// If we created a declaration and the function definition is not imported,
// build a definition.
if (functions_.Get(id) && fn_info.definition_id.has_value() &&
!sem_ir().insts().GetImportSource(fn_info.definition_id).has_value()) {
BuildFunctionDefinition(id);
}
}
// Append `__global_init` to `llvm::global_ctors` to initialize global
// variables.
if (auto global_ctor_id = sem_ir().global_ctor_id();
global_ctor_id.has_value()) {
auto llvm_function = BuildFunctionDecl(global_ctor_id);
functions_.Set(global_ctor_id, llvm_function);
const auto& global_ctor = sem_ir().functions().Get(global_ctor_id);
BuildFunctionBody(global_ctor_id, SemIR::SpecificId::None, global_ctor,
*this, global_ctor);
llvm::appendToGlobalCtors(llvm_module(), llvm_function->llvm_function,
/*Priority=*/0);
}
}
auto FileContext::Finalize() -> void {
if (cpp_code_generator_) {
// Clang code generation should not actually modify the AST, but isn't
// const-correct.
cpp_code_generator_->HandleTranslationUnit(
const_cast<clang::ASTContext&>(cpp_file()->ast_context()));
}
// Find equivalent specifics (from the same generic), replace all uses and
// remove duplicately lowered function definitions.
coalescer_.CoalesceEquivalentSpecifics(lowered_specifics_,
specific_functions_);
}
auto FileContext::GetConstant(SemIR::ConstantId const_id,
SemIR::InstId use_inst_id) -> llvm::Value* {
auto const_inst_id = sem_ir().constant_values().GetInstId(const_id);
auto* const_value = constants_.Get(const_inst_id);
// For value expressions and initializing expressions, the value produced by
// a constant instruction is a value representation of the constant. For
// initializing expressions, `FinishInit` will perform a copy if needed.
switch (auto cat = SemIR::GetExprCategory(sem_ir(), const_inst_id)) {
case SemIR::ExprCategory::Value:
case SemIR::ExprCategory::ReprInitializing:
case SemIR::ExprCategory::InPlaceInitializing:
break;
case SemIR::ExprCategory::DurableRef:
case SemIR::ExprCategory::EphemeralRef:
// Constant reference expressions lower to an address.
return const_value;
case SemIR::ExprCategory::NotExpr:
case SemIR::ExprCategory::Error:
case SemIR::ExprCategory::Pattern:
case SemIR::ExprCategory::Mixed:
case SemIR::ExprCategory::RefTagged:
case SemIR::ExprCategory::Dependent:
CARBON_FATAL("Unexpected category {0} for lowered constant {1}", cat,
sem_ir().insts().Get(const_inst_id));
};
auto value_rep = SemIR::ValueRepr::ForType(
sem_ir(), sem_ir().insts().Get(const_inst_id).type_id());
if (value_rep.kind != SemIR::ValueRepr::Pointer) {
return const_value;
}
// The value representation is a pointer. Generate a variable to hold the
// value, or find and reuse an existing one.
if (auto result = global_variables().Lookup(const_inst_id)) {
return result.value();
}
// Include both the name of the constant, if any, and the point of use in
// the name of the variable.
llvm::StringRef const_name;
llvm::StringRef use_name;
if (inst_namer_) {
const_name = inst_namer_->GetUnscopedNameFor(const_inst_id);
if (use_inst_id.has_value()) {
use_name = inst_namer_->GetUnscopedNameFor(use_inst_id);
}
}
// We always need to give the global a name even if the instruction namer
// doesn't have one to use.
if (const_name.empty()) {
const_name = "const";
}
if (use_name.empty()) {
use_name = "anon";
}
llvm::StringRef sep = (use_name[0] == '.') ? "" : ".";
auto* global_variable = new llvm::GlobalVariable(
llvm_module(), GetType(sem_ir().GetPointeeType(value_rep.type_id)),
/*isConstant=*/true, llvm::GlobalVariable::InternalLinkage, const_value,
const_name + sep + use_name);
global_variables_.Insert(const_inst_id, global_variable);
return global_variable;
}
auto FileContext::GetOrCreateFunctionInfo(
SemIR::FunctionId function_id, SemIR::SpecificId specific_id,
FileContext* fallback_file, SemIR::FunctionId fallback_function_id,
SemIR::SpecificId fallback_specific_id) -> std::optional<FunctionInfo>& {
// If we have already lowered a declaration of this function, just return it.
// TODO: If the existing declaration is inexact, and we now have a fallback,
// we should try again.
auto& result = GetFunctionInfo(function_id, specific_id);
if (!result) {
result = BuildFunctionDecl(function_id, specific_id, fallback_file,
fallback_function_id, fallback_specific_id);
}
return result;
}
// State machine for building a FunctionTypeInfo from SemIR.
//
// The main difficulty this class encapsulates is that each abstraction level
// has different expectations about how the return is reflected in the parameter
// list.
// - In SemIR, if the function has an initializing return form, it has a
// corresponding output parameter at the end of the parameter list.
// - In LLVM IR, if the SemIR has an output parameter _and_ that parameter's
// type has an in-place initializing representation, we emit a corresponding
// `sret` output parameter (and the function's return type is void). By
// convention the output parameter goes at the start of the parameter list.
// - In LLVM debug info, the list of parameter types always starts with the
// return type (which doubles as the type of the return parameter, if there
// is one).
//
// Furthermore, SemIR is designed to eventually support compound return forms,
// in which case there can be multiple output parameters for different pieces of
// the return form, but it's not yet clear how we will lower such functions.
//
// We also deal with the case where the function signature involves incomplete
// types. This can happen if the function is declared but never defined nor
// called in this file. Declarations of such functions can still need to be
// emitted; currently this happens if they are part of a class's vtable. Such
// uses do not need an exact signature, so we emit them with the LLVM type
// `void()` and set `inexact` on the result to indicate the type is not known.
// LLVM can handle merging inexact and exact signatures, and this matches how
// Clang handles the corresponding situation in C++.
//
// One additional complexity is that we may need to fetch information about the
// same function from multiple different files. For a call to a generic
// function, there may be no single file in which all the relevant types are
// complete, so we will look at both the specific function definition that is
// the resolved callee, as well as the partially-specific function from the call
// site.
//
// In general, we support being given a list of variants of the function, in
// which the first function in the list is the primary declaration and should be
// the most specific function, and the others are used as fallbacks if an
// incomplete type is encountered.
class FileContext::FunctionTypeInfoBuilder {
public:
struct FunctionInContext {
FileContext* context;
SemIR::FunctionId function_id;
SemIR::SpecificId specific_id;
};
// Creates a FunctionTypeInfoBuilder that uses the given functions.
explicit FunctionTypeInfoBuilder(llvm::ArrayRef<FunctionInContext> functions)
: context_(&functions.front().context->context()), functions_(functions) {
CARBON_CHECK(!functions_.empty());
}
// Retrieves various features of the function's type useful for constructing
// the `llvm::Type` and `llvm::DISubroutineType` for the `llvm::Function`. If
// any part of the type can't be manifest (eg: incomplete return or parameter
// types), then the result is as if the type was `void()`. Should only be
// called once on a given builder.
auto Build() && -> FunctionTypeInfo;
private:
// By convention, state transition methods return false (without changing the
// accumulated information about the function) to indicate that we could not
// manifest the complete function type successfully in this context.
// Information about how a function is called in SemIR.
struct SemIRIndexInfo {
// The number of parameters in the SemIR call signature.
int num_params;
// The index of the first return parameter in the SemIR call signature.
int return_param_index;
friend auto operator==(const SemIRIndexInfo& lhs, const SemIRIndexInfo& rhs)
-> bool = default;
};
// Get information about the SemIR function signature.
auto GetSemIRIndexInfo(const FunctionInContext& fn_in_context)
-> SemIRIndexInfo {
const auto& sem_ir = fn_in_context.context->sem_ir();
const auto& function = sem_ir.functions().Get(fn_in_context.function_id);
int num_params =
sem_ir.inst_blocks().Get(function.call_param_patterns_id).size();
int return_param_index = -1;
if (function.call_param_ranges.return_size() > 0) {
CARBON_CHECK(function.call_param_ranges.return_size() == 1,
"TODO: support multiple return forms");
return_param_index = function.call_param_ranges.return_begin().index;
}
return {.num_params = num_params, .return_param_index = return_param_index};
}
// Handles the function's return form.
//
// This should be called before `HandleParameter`. It handles the return form
// by trying each `FunctionInContext` until one succeeds, and returns false if
// all attempts failed.
auto HandleReturnForm() -> bool;
// Tries to handle the return form using the given context. Delegates to
// exactly one of `SetReturnByCopy`, `SetReturnByReference`, or
// `SetReturnInPlace`, or returns false if the return type is incomplete.
auto TryHandleReturnForm(const FunctionInContext& func_ctx) -> bool;
// Records that the LLVM function returns by copy, with type `return_type_id`.
// `return_type_id` can be `None`, which is treated as equivalent to the
// default return type `()`.
auto SetReturnByCopy(const FunctionInContext& func_ctx,
SemIR::TypeId return_type_id) -> bool {
CARBON_CHECK(return_type_ == nullptr);
CARBON_CHECK(param_di_types_.empty());
auto lowered_return_types = GetLoweredTypes(func_ctx, return_type_id);
return_type_ = lowered_return_types.llvm_ir_type;
param_di_types_.push_back(lowered_return_types.llvm_di_type);
return true;
}
// Records that the LLVM function returns by reference, with type
// `return_type_id`.
auto SetReturnByReference(const FunctionInContext& func_ctx,
SemIR::TypeId /*return_type_id*/) -> bool {
return_type_ = llvm::PointerType::get(func_ctx.context->llvm_context(),
/*AddressSpace=*/0);
// TODO: replace this with a reference type.
param_di_types_.push_back(GetPointerDIType(nullptr));
return true;
}
// Records that the LLVM function returns in place, with type
// `return_type_id`.
auto SetReturnInPlace(const FunctionInContext& func_ctx,
SemIR::TypeId return_type_id) -> bool {
return_type_ = llvm::Type::getVoidTy(func_ctx.context->llvm_context());
sret_type_ = func_ctx.context->GetType(return_type_id);
// We don't add to param_di_types_ because that will be handled by the
// loop over the SemIR parameters.
return true;
}
// Handles `Call` parameter pattern at the given index. This should be called
// on parameter patterns in the order that they should appear in the LLVM IR
// parameter list, so in particular it should be called on the
// `OutParamPattern` (if any) first. It should be called on all `Call`
// parameters; it will determine which parameters belong in the LLVM IR
// parameter list.
//
// This tries each `FunctionInContext` until one succeeds, and returns false
// if all attempts failed.
auto HandleParameter(SemIR::CallParamIndex index) -> bool;
// Tries to handle the parameter pattern at the given index using the given
// context. Delegates to either `AddLoweredParam` or `IgnoreParam`, or returns
// false if the parameter type is incomplete.
auto TryHandleParameter(const FunctionInContext& func_ctx,
SemIR::CallParamIndex index) -> bool;
// Records that the parameter pattern at the given index has the given ID, and
// lowers to the given IR and DI types.
auto AddLoweredParam(const FunctionInContext& func_ctx,
SemIR::CallParamIndex index,
SemIR::InstId param_pattern_id, LoweredTypes param_types)
-> bool {
lowered_param_indices_.push_back(index);
param_name_ids_.push_back(SemIR::GetPrettyNameFromPatternId(
func_ctx.context->sem_ir(), param_pattern_id));
param_types_.push_back(param_types.llvm_ir_type);
param_di_types_.push_back(param_types.llvm_di_type);
return true;
}
// Records that the `Call` parameter pattern at the given index is not lowered
// to an LLVM parameter.
auto IgnoreParam(SemIR::CallParamIndex index) -> bool {
unused_param_indices_.push_back(index);
return true;
}
// Builds and returns a FunctionTypeInfo from the accumulated information.
auto Finalize() -> FunctionTypeInfo;
// Clears out accumulated state and returns a FunctionTypeInfo with the
// fallback state `void()`.
auto Abort() -> FunctionTypeInfo;
// Returns LLVM IR and DI types for the given SemIR type. This is not a state
// transition. It mostly delegates to context_.GetTypeAndDIType, but treats
// TypeId::None as equivalent to the unit type, and uses an untyped pointer as
// a placeholder DI type if context_ doesn't provide one.
auto GetLoweredTypes(const FunctionInContext& func_ctx, SemIR::TypeId type_id)
-> LoweredTypes;
// Returns a DI type for a pointer to the given pointee. The pointee type may
// be null.
auto GetPointerDIType(llvm::DIType* pointee_type, unsigned address_space = 0)
-> llvm::DIDerivedType* {
const auto& data_layout = context_->llvm_module().getDataLayout();
return context_->di_builder().createPointerType(
pointee_type, data_layout.getPointerSizeInBits(address_space));
}
Context* context_;
llvm::ArrayRef<FunctionInContext> functions_;
// The number of input `Call` parameter patterns.
int num_params_ = 0;
// The types of the parameters in the LLVM IR function. Each one corresponds
// to a SemIR `Call` parameter, but some `Call` parameters may be omitted
// (e.g. if they are stateless) or reordered (e.g. the return parameter, if
// any, always goes first).
llvm::SmallVector<llvm::Type*> param_types_;
// The LLLVM DI representation of the parameter list. As required by LLVM DI
// convention, this starts with the function's return type, and ends with the
// DI representations of param_types_ (in the same order). Note that those
// two ranges may overlap: if the first element of param_types_ represents
// a return parameter, the first element of param_di_types_ corresponds to it
// while also representing the return type.
llvm::SmallVector<llvm::Metadata*> param_di_types_;
// The indices of the `Call` parameters that correspond to `param_types_`, in
// the same order.
llvm::SmallVector<SemIR::CallParamIndex> lowered_param_indices_;
// The names of the `Call` parameters that correspond to `param_types_`, in
// the same order.
llvm::SmallVector<SemIR::NameId> param_name_ids_;
// The indices of any `Call` param patterns that aren't present in
// lowered_param_indices_.
llvm::SmallVector<SemIR::CallParamIndex> unused_param_indices_;
// The LLVM function's return type.
llvm::Type* return_type_ = nullptr;
// If not null, the LLVM function's first parameter should have a `sret`
// attribute with this type.
llvm::Type* sret_type_ = nullptr;
// Whether we failed to form an exact description of the function type. This
// can happen if a parameter or return type is incomplete. In this case, we
// can still sometimes need to emit a declaration of the function, for example
// because it appears in a vtable, but we cannot emit a definition or a call.
bool inexact_ = false;
};
auto FileContext::FunctionTypeInfoBuilder::Build() && -> FunctionTypeInfo {
// TODO: For the `Run` entry point, remap return type to i32 if it doesn't
// return a value.
// Determine how the parameters are numbered in SemIR, and make sure it's the
// same for all versions of the function.
auto semir_info = GetSemIRIndexInfo(functions_.front());
CARBON_CHECK(
llvm::all_of(functions_.drop_front(), [&](const auto& fn_in_context) {
return GetSemIRIndexInfo(fn_in_context) == semir_info;
}));
num_params_ = semir_info.num_params;
lowered_param_indices_.reserve(num_params_);
param_name_ids_.reserve(num_params_);
param_types_.reserve(num_params_);
param_di_types_.reserve(num_params_);
if (!HandleReturnForm()) {
return Abort();
}
int params_end = num_params_;
if (semir_info.return_param_index >= 0) {
CARBON_CHECK(semir_info.return_param_index == semir_info.num_params - 1,
"Unexpected parameter order");
params_end = semir_info.return_param_index;
// Handle the return parameter first, because it goes first in the LLVM
// convention.
if (!HandleParameter(
SemIR::CallParamIndex(semir_info.return_param_index))) {
return Abort();
}
}
for (int i : llvm::seq(params_end)) {
if (!HandleParameter(SemIR::CallParamIndex(i))) {
return Abort();
}
}
return Finalize();
}
auto FileContext::FunctionTypeInfoBuilder::HandleReturnForm() -> bool {
for (const auto& func_ctx : functions_) {
if (TryHandleReturnForm(func_ctx)) {
return true;
}
}
return false;
}
auto FileContext::FunctionTypeInfoBuilder::TryHandleReturnForm(
const FunctionInContext& func_ctx) -> bool {
const auto& function =
func_ctx.context->sem_ir().functions().Get(func_ctx.function_id);
auto return_form_inst_id = function.return_form_inst_id;
if (!return_form_inst_id.has_value()) {
return SetReturnByCopy(func_ctx, SemIR::TypeId::None);
}
auto return_form_const_id = SemIR::GetConstantValueInSpecific(
func_ctx.context->sem_ir(), func_ctx.specific_id, return_form_inst_id);
auto return_form_inst = func_ctx.context->sem_ir().insts().Get(
func_ctx.context->sem_ir().constant_values().GetInstId(
return_form_const_id));
CARBON_KIND_SWITCH(return_form_inst) {
case CARBON_KIND(SemIR::InitForm init_form): {
auto return_type_id =
func_ctx.context->sem_ir().types().GetTypeIdForTypeConstantId(
SemIR::GetConstantValueInSpecific(
func_ctx.context->sem_ir(), func_ctx.specific_id,
init_form.type_component_inst_id));
switch (
SemIR::InitRepr::ForType(func_ctx.context->sem_ir(), return_type_id)
.kind) {
case SemIR::InitRepr::InPlace: {
return SetReturnInPlace(func_ctx, return_type_id);
}
case SemIR::InitRepr::ByCopy: {
return SetReturnByCopy(func_ctx, return_type_id);
}
case SemIR::InitRepr::None:
return SetReturnByCopy(func_ctx, SemIR::TypeId::None);
case SemIR::InitRepr::Dependent:
CARBON_FATAL("Lowering function return with dependent type: {0}",
return_form_inst);
case SemIR::InitRepr::Incomplete:
case SemIR::InitRepr::Abstract:
return false;
}
}
case CARBON_KIND(SemIR::RefForm ref_form): {
auto return_type_id =
func_ctx.context->sem_ir().types().GetTypeIdForTypeConstantId(
SemIR::GetConstantValueInSpecific(
func_ctx.context->sem_ir(), func_ctx.specific_id,
ref_form.type_component_inst_id));
return SetReturnByReference(func_ctx, return_type_id);
}
case CARBON_KIND(SemIR::ValueForm val_form): {
auto return_type_id =
func_ctx.context->sem_ir().types().GetTypeIdForTypeConstantId(
SemIR::GetConstantValueInSpecific(
func_ctx.context->sem_ir(), func_ctx.specific_id,
val_form.type_component_inst_id));
switch (
SemIR::ValueRepr::ForType(func_ctx.context->sem_ir(), return_type_id)
.kind) {
case SemIR::ValueRepr::Unknown:
return false;
case SemIR::ValueRepr::Dependent:
CARBON_FATAL("Lowering function return with dependent type: {0}",
return_form_inst);
case SemIR::ValueRepr::None:
return SetReturnByCopy(func_ctx, SemIR::TypeId::None);
case SemIR::ValueRepr::Copy:
return SetReturnByCopy(func_ctx, return_type_id);
case SemIR::ValueRepr::Pointer:
case SemIR::ValueRepr::Custom:
return SetReturnByReference(func_ctx, return_type_id);
}
}
default:
CARBON_FATAL("Unexpected inst kind: {0}", return_form_inst);
}
}
auto FileContext::FunctionTypeInfoBuilder::HandleParameter(
SemIR::CallParamIndex index) -> bool {
for (const auto& func_ctx : functions_) {
if (TryHandleParameter(func_ctx, index)) {
return true;
}
}
return false;
}
auto FileContext::FunctionTypeInfoBuilder::TryHandleParameter(
const FunctionInContext& func_ctx, SemIR::CallParamIndex index) -> bool {
const auto& sem_ir = func_ctx.context->sem_ir();
auto param_pattern_id =
sem_ir.inst_blocks().Get(sem_ir.functions()
.Get(func_ctx.function_id)
.call_param_patterns_id)[index.index];
auto param_pattern = sem_ir.insts().Get(param_pattern_id);
auto param_type_id = ExtractScrutineeType(
sem_ir, SemIR::GetTypeOfInstInSpecific(sem_ir, func_ctx.specific_id,
param_pattern_id));
// Returns the appropriate LoweredTypes for reference-like parameters.
auto ref_lowered_types = [&]() -> LoweredTypes {
return {
.llvm_ir_type = llvm::PointerType::get(func_ctx.context->llvm_context(),
/*AddressSpace=*/0),
// TODO: replace this with a reference type.
.llvm_di_type = GetLoweredTypes(func_ctx, param_type_id).llvm_di_type};
};
CARBON_CHECK(
!param_type_id.AsConstantId().is_symbolic(),
"Found symbolic type id after resolution when lowering type {0}.",
param_pattern.type_id());
auto param_kind = param_pattern.kind();
// Treat a form parameter pattern like the kind of param pattern that
// corresponds to its form.
if (auto form_param_pattern =
param_pattern.TryAs<SemIR::FormParamPattern>()) {
auto form_kind = sem_ir.insts().Get(form_param_pattern->form_id).kind();
switch (form_kind) {
case SemIR::InitForm::Kind:
param_kind = SemIR::VarParamPattern::Kind;
break;
case SemIR::RefForm::Kind:
param_kind = SemIR::RefParamPattern::Kind;
break;
case SemIR::ValueForm::Kind:
param_kind = SemIR::ValueParamPattern::Kind;
break;
default:
CARBON_FATAL("Unexpected kind {0} for form inst", form_kind);
}
}
switch (param_kind) {
case SemIR::RefParamPattern::Kind:
case SemIR::VarParamPattern::Kind: {
return AddLoweredParam(func_ctx, index, param_pattern_id,
ref_lowered_types());
}
case SemIR::OutParamPattern::Kind: {
switch (SemIR::InitRepr::ForType(sem_ir, param_type_id).kind) {
case SemIR::InitRepr::InPlace:
return AddLoweredParam(func_ctx, index, param_pattern_id,
ref_lowered_types());
case SemIR::InitRepr::ByCopy:
case SemIR::InitRepr::None:
return IgnoreParam(index);
case SemIR::InitRepr::Dependent:
CARBON_FATAL("Lowering function parameter with dependent type: {0}",
param_pattern);
case SemIR::InitRepr::Incomplete:
case SemIR::InitRepr::Abstract:
return false;
}
}
case SemIR::ValueParamPattern::Kind: {
switch (auto value_rep = SemIR::ValueRepr::ForType(sem_ir, param_type_id);
value_rep.kind) {
case SemIR::ValueRepr::Unknown:
return false;
case SemIR::ValueRepr::Dependent:
CARBON_FATAL("Lowering function parameter with dependent type: {0}",
param_pattern);
case SemIR::ValueRepr::None:
return IgnoreParam(index);
case SemIR::ValueRepr::Copy:
case SemIR::ValueRepr::Custom:
case SemIR::ValueRepr::Pointer: {
if (value_rep.type_id.has_value()) {
return AddLoweredParam(
func_ctx, index, param_pattern_id,
GetLoweredTypes(func_ctx, value_rep.type_id));
} else {
return IgnoreParam(index);
}
}
}
}
default:
CARBON_FATAL("Unexpected inst kind: {0}", param_pattern);
}
}
auto FileContext::FunctionTypeInfoBuilder::Finalize() -> FunctionTypeInfo {
CARBON_CHECK(lowered_param_indices_.size() + unused_param_indices_.size() ==
static_cast<size_t>(num_params_));
CARBON_CHECK(!param_di_types_.empty());
auto& di_builder = context_->di_builder();
return {.type = llvm::FunctionType::get(return_type_, param_types_,
/*isVarArg=*/false),
.di_type = di_builder.createSubroutineType(
di_builder.getOrCreateTypeArray(param_di_types_),
llvm::DINode::FlagZero),
.lowered_param_indices = std::move(lowered_param_indices_),
.unused_param_indices = std::move(unused_param_indices_),
.param_name_ids = std::move(param_name_ids_),
.sret_type = sret_type_,
.inexact = inexact_};
}
auto FileContext::FunctionTypeInfoBuilder::Abort() -> FunctionTypeInfo {
num_params_ = 0;
lowered_param_indices_.clear();
unused_param_indices_.clear();
param_name_ids_.clear();
param_types_.clear();
param_di_types_.clear();
return_type_ = llvm::Type::getVoidTy(context_->llvm_context());
param_di_types_.push_back(nullptr);
inexact_ = true;
return Finalize();
}
auto FileContext::FunctionTypeInfoBuilder::GetLoweredTypes(
const FunctionInContext& func_ctx, SemIR::TypeId type_id) -> LoweredTypes {
if (!type_id.has_value()) {
return {
.llvm_ir_type = llvm::Type::getVoidTy(func_ctx.context->llvm_context()),
.llvm_di_type = nullptr};
}
auto result = func_ctx.context->GetTypeAndDIType(type_id);
if (result.llvm_di_type == nullptr) {
// TODO: figure out what type should go here, or ensure this doesn't
// happen.
result.llvm_di_type = GetPointerDIType(nullptr);
}
return result;
}
auto FileContext::HandleReferencedCppFunction(clang::FunctionDecl* cpp_decl)
-> llvm::Function* {
// Create the LLVM function (`CodeGenModule::GetOrCreateLLVMFunction()`)
// so that code generation (`CodeGenModule::EmitGlobal()`) would see this
// function name (`CodeGenModule::getMangledName()`), and will generate
// its definition.
auto* function_address = dyn_cast<llvm::Function>(
cpp_code_generator_->GetAddrOfGlobal(CreateGlobalDecl(cpp_decl),
/*isForDefinition=*/false));
CARBON_CHECK(function_address);
return function_address;
}
auto FileContext::HandleReferencedSpecificFunction(
SemIR::FunctionId function_id, SemIR::SpecificId specific_id,
llvm::Type* llvm_type) -> void {
CARBON_CHECK(specific_id.has_value());
// Add this specific function to a list of specific functions whose
// definitions we need to emit.
// TODO: Don't do this if we know this function is emitted as a
// non-discardable symbol in the IR for some other file.
context().AddPendingSpecificFunctionDefinition({.context = this,
.function_id = function_id,
.specific_id = specific_id});
// Create a unique fingerprint for the function type.
// For now, we compute the function type fingerprint only for specifics,
// though we might need it for all functions in order to create a canonical
// fingerprint across translation units.
coalescer_.CreateTypeFingerprint(specific_id, llvm_type);
}
auto FileContext::GetOrCreateLLVMFunction(
const FunctionTypeInfo& function_type_info, SemIR::FunctionId function_id,
SemIR::SpecificId specific_id) -> llvm::Function* {
// If this is a C++ function, tell Clang that we referenced it.
if (auto clang_decl_id = sem_ir().functions().Get(function_id).clang_decl_id;
clang_decl_id.has_value()) {
CARBON_CHECK(!specific_id.has_value(),
"Specific functions cannot have C++ definitions");
return HandleReferencedCppFunction(
sem_ir().clang_decls().Get(clang_decl_id).key.decl->getAsFunction());
}
SemIR::Mangler m(sem_ir(), context().total_ir_count());
std::string mangled_name = m.Mangle(function_id, specific_id);
if (auto* existing = llvm_module().getFunction(mangled_name)) {
// We might have already lowered this function while lowering a different
// file. That's OK.
// TODO: If the prior function was inexact and the new one is not, we should
// lower this new one and replace the existing function with it.
// TODO: Check-fail or maybe diagnose if the two LLVM functions are not
// produced by declarations of the same Carbon function. Name collisions
// between non-private members of the same library should have been
// diagnosed by check if detected, but it's not clear that check will
// always be able to see this problem. In theory, name collisions could
// also occur due to fingerprint collision.
return existing;
}
// If this is a specific function, we may need to do additional work to
// emit its definition.
if (specific_id.has_value()) {
HandleReferencedSpecificFunction(function_id, specific_id,
function_type_info.type);
}
// TODO: For an imported inline function, consider generating an
// `available_externally` definition.
auto linkage = llvm::Function::ExternalLinkage;
if (function_id == sem_ir().global_ctor_id()) {
// The global constructor name would collide with global constructors for
// other files in the same package, so use an internal linkage symbol.
linkage = llvm::Function::InternalLinkage;
} else if (specific_id.has_value()) {
// Specific functions are allowed to be duplicated across files.
// TODO: CoreWitness should have the same behavior; see its use of
// WeakODRLinkage in BuildFunctionDefinition.
linkage = llvm::Function::LinkOnceODRLinkage;
}
auto* llvm_function = llvm::Function::Create(function_type_info.type, linkage,
mangled_name, llvm_module());
CARBON_CHECK(llvm_function->getName() == mangled_name,
"Mangled name collision: {0}", mangled_name);
// Set up parameters and the return slot.
for (auto [name_id, arg] : llvm::zip_equal(function_type_info.param_name_ids,
llvm_function->args())) {
arg.setName(sem_ir().names().GetIRBaseName(name_id));
}
if (function_type_info.sret_type != nullptr) {
auto& return_arg = *llvm_function->args().begin();
return_arg.addAttr(llvm::Attribute::getWithStructRetType(
llvm_context(), function_type_info.sret_type));
}
return llvm_function;
}
auto FileContext::BuildFunctionDecl(SemIR::FunctionId function_id,
SemIR::SpecificId specific_id,
FileContext* fallback_file,
SemIR::FunctionId fallback_function_id,
SemIR::SpecificId fallback_specific_id)
-> std::optional<FunctionInfo> {
const auto& function = sem_ir().functions().Get(function_id);
// Don't lower generic functions. Note that associated functions in interfaces
// have `Self` in scope, so are implicitly generic functions.
if (function.generic_id.has_value() && !specific_id.has_value()) {
return std::nullopt;
}
// Don't lower builtins.
if (function.builtin_function_kind() != SemIR::BuiltinFunctionKind::None) {
return std::nullopt;
}
// Don't lower C++ functions that use a thunk. We will never reference them
// directly, and their signatures would not be expected to match the
// corresponding C++ function anyway.
if (function.special_function_kind ==
SemIR::Function::SpecialFunctionKind::HasCppThunk) {
return std::nullopt;
}
// TODO: Consider tracking whether the function has been used, and only
// lowering it if it's needed.
FunctionTypeInfoBuilder::FunctionInContext func_infos[] = {
{this, function_id, specific_id},
{fallback_file, fallback_function_id, fallback_specific_id}};
auto function_type_info =
FunctionTypeInfoBuilder(llvm::ArrayRef(func_infos, fallback_file ? 2 : 1))
.Build();
auto* llvm_function =
GetOrCreateLLVMFunction(function_type_info, function_id, specific_id);
return {{.type = function_type_info.type,
.di_type = function_type_info.di_type,
.lowered_param_indices =
std::move(function_type_info.lowered_param_indices),
.unused_param_indices =
std::move(function_type_info.unused_param_indices),
.llvm_function = llvm_function,
.inexact = function_type_info.inexact}};
}
// Find the file and function ID describing the definition of a function.
static auto GetFunctionDefinition(const SemIR::File* decl_ir,
SemIR::FunctionId function_id)
-> std::pair<const SemIR::File*, SemIR::FunctionId> {
// Find the file containing the definition.
auto decl_id = decl_ir->functions().Get(function_id).definition_id;
if (!decl_id.has_value()) {
// Function is not defined.
return {nullptr, SemIR::FunctionId::None};
}
// Find the function declaration this function was originally imported from.
while (true) {
auto import_inst_id = decl_ir->insts().GetImportSource(decl_id);
if (!import_inst_id.has_value()) {
break;
}
auto import_inst = decl_ir->import_ir_insts().Get(import_inst_id);
decl_ir = decl_ir->import_irs().Get(import_inst.ir_id()).sem_ir;
decl_id = import_inst.inst_id();
}
auto decl_ir_function_id =
decl_ir->insts().GetAs<SemIR::FunctionDecl>(decl_id).function_id;
return {decl_ir, decl_ir_function_id};
}
auto FileContext::BuildFunctionDefinition(SemIR::FunctionId function_id,
SemIR::SpecificId specific_id)
-> void {
auto [definition_ir, definition_ir_function_id] =
GetFunctionDefinition(&sem_ir(), function_id);
if (!definition_ir) {
// Function is probably defined in another file; not an error.
return;
}
const auto& definition_function =
definition_ir->functions().Get(definition_ir_function_id);
BuildFunctionBody(
function_id, specific_id, sem_ir().functions().Get(function_id),
context().GetFileContext(definition_ir), definition_function);
}
auto FileContext::BuildFunctionBody(SemIR::FunctionId function_id,
SemIR::SpecificId specific_id,
const SemIR::Function& declaration_function,
FileContext& definition_context,
const SemIR::Function& definition_function)
-> void {
// On crash, report the function we were lowering.
PrettyStackTraceFunction stack_trace_entry([&](llvm::raw_ostream& output) {
SemIR::DiagnosticLocConverter converter(
&context().tree_and_subtrees_getters(), &sem_ir());
auto converted =
converter.Convert(SemIR::LocId(declaration_function.definition_id),
/*token_only=*/false);
converted.loc.FormatLocation(output);
output << "Lowering function ";
if (specific_id.has_value()) {
output << SemIR::StringifySpecific(sem_ir(), specific_id);
} else {
output << SemIR::StringifyConstantInst(
sem_ir(), declaration_function.definition_id);
}
output << "\n";
// Crash output has a tab indent; try to indent slightly past that.
converted.loc.FormatSnippet(output, /*indent=*/10);
});
// Note that `definition_function` is potentially from a different SemIR::File
// than the one that this file context represents. Any lowering done for
// values derived from `definition_function` should use `definition_context`
// instead of our context.
const auto& definition_ir = definition_context.sem_ir();
auto function_info = GetFunctionInfo(function_id, specific_id);
CARBON_CHECK(function_info && function_info->llvm_function,
"Attempting to define function that was not declared");
CARBON_CHECK(!function_info->inexact,
"Attempting to emit definition of inexact function: {0}",
*function_info->llvm_function);
// TODO: Build CoreWitness functions when they're called instead of when
// they're defined. That should allow LinkOnceODRLinkage.
if (declaration_function.special_function_kind ==
SemIR::Function::SpecialFunctionKind::CoreWitness) {
function_info->llvm_function->setLinkage(llvm::Function::WeakODRLinkage);
}
const auto& body_block_ids = definition_function.body_block_ids;
CARBON_DCHECK(!body_block_ids.empty(),
"No function body blocks found during lowering.");
// Store which specifics were already lowered (with definitions) for each
// generic.
if (declaration_function.generic_id.has_value() && specific_id.has_value()) {
// TODO: We should track this in the definition context instead so that we
// can deduplicate specifics from different files.
AddLoweredSpecificForGeneric(declaration_function.generic_id, specific_id);
}
// Set attributes on the function definition.
{
llvm::AttrBuilder attr_builder(llvm_context());
attr_builder.addAttribute(llvm::Attribute::NoUnwind);
// TODO: We should take the opt level from the SemIR file; it might not be
// the same for all files in a compilation.
if (context().opt_level() == Lower::OptimizationLevel::None) {
// --optimize=none disables all optimizations for this function.
attr_builder.addAttribute(llvm::Attribute::OptimizeNone);
attr_builder.addAttribute(llvm::Attribute::NoInline);
} else {
// Otherwise, always inline thunks.
if (definition_function.special_function_kind ==
SemIR::Function::SpecialFunctionKind::Thunk) {
attr_builder.addAttribute(llvm::Attribute::AlwaysInline);
}
// Convert --optimize=size into optsize and minsize.
if (context().opt_level() == Lower::OptimizationLevel::Size) {
attr_builder.addAttribute(llvm::Attribute::OptimizeForSize);
attr_builder.addAttribute(llvm::Attribute::MinSize);
}
// TODO: Should we generate an InlineHint for some functions? Perhaps for
// those defined in the API file?
}
function_info->llvm_function->addFnAttrs(attr_builder);
}
auto* subprogram = BuildDISubprogram(declaration_function, *function_info);
FunctionContext function_lowering(
definition_context, function_info->llvm_function, *this, specific_id,
coalescer_.InitializeFingerprintForSpecific(specific_id), subprogram,
vlog_stream_);
auto call_param_ids = definition_ir.inst_blocks().GetOrEmpty(
definition_function.call_params_id);
// Add local variables for the parameters.
for (auto [llvm_index, index] :
llvm::enumerate(function_info->lowered_param_indices)) {
function_lowering.SetLocal(
call_param_ids[index.index],
function_info->llvm_function->getArg(llvm_index));
}
// Add local variables for the SemIR parameters that aren't LLVM parameters.
// These shouldn't actually be used, so they're set to poison values.
for (auto [llvm_index, index] :
llvm::enumerate(function_info->unused_param_indices)) {
auto param_id = call_param_ids[index.index];
function_lowering.SetLocal(
param_id,
llvm::PoisonValue::get(function_lowering.GetTypeOfInst(param_id)));
}
auto decl_block_id = SemIR::InstBlockId::None;
if (function_id == sem_ir().global_ctor_id()) {
decl_block_id = SemIR::InstBlockId::Empty;
} else {
decl_block_id =
definition_ir.insts()
.GetAs<SemIR::FunctionDecl>(definition_function.latest_decl_id())
.decl_block_id;
}
// Lowers the contents of decl_block_id into the corresponding LLVM block,
// creating it if it doesn't already exist.
auto lower_block = [&](SemIR::InstBlockId block_id) {
CARBON_VLOG("Lowering {0}\n", block_id);
auto* llvm_block = function_lowering.GetBlock(block_id);
// Keep the LLVM blocks in lexical order.
llvm_block->moveBefore(function_info->llvm_function->end());
function_lowering.builder().SetInsertPoint(llvm_block);
function_lowering.LowerBlockContents(block_id);
};
lower_block(decl_block_id);
// If the decl block is empty, reuse it as the first body block. We don't do
// this when the decl block is non-empty so that any branches back to the
// first body block don't also re-execute the decl.
llvm::BasicBlock* block = function_lowering.builder().GetInsertBlock();
if (block->empty() &&
function_lowering.TryToReuseBlock(body_block_ids.front(), block)) {
// Reuse this block as the first block of the function body.
} else {
function_lowering.builder().CreateBr(
function_lowering.GetBlock(body_block_ids.front()));
}
// Lower all blocks.
for (auto block_id : body_block_ids) {
lower_block(block_id);
}
// LLVM requires that the entry block has no predecessors.
auto* entry_block = &function_info->llvm_function->getEntryBlock();
if (entry_block->hasNPredecessorsOrMore(1)) {
auto* new_entry_block = llvm::BasicBlock::Create(
llvm_context(), "entry", function_info->llvm_function, entry_block);
llvm::UncondBrInst::Create(entry_block, new_entry_block);
}
// Emit fingerprint accumulated inside the function context.
function_lowering.EmitFinalFingerprint();
context().di_builder().finalizeSubprogram(subprogram);
}
auto FileContext::BuildDISubprogram(const SemIR::Function& function,
const FunctionInfo& function_info)
-> llvm::DISubprogram* {
if (!context().di_compile_unit()) {
return nullptr;
}
auto name = sem_ir().names().GetAsStringIfIdentifier(function.name_id);
CARBON_CHECK(name, "Unexpected special name for function: {0}",
function.name_id);
auto loc = GetLocForDI(function.definition_id);
llvm::DISubroutineType* subroutine_type = function_info.di_type;
auto* subprogram = context().di_builder().createFunction(
context().di_compile_unit(), *name,
function_info.llvm_function->getName(),
/*File=*/context().di_builder().createFile(loc.filename, ""),
/*LineNo=*/loc.line_number, subroutine_type,
/*ScopeLine=*/0, llvm::DINode::FlagZero,
llvm::DISubprogram::SPFlagDefinition);
// Add a variable for each parameter, as that is where DWARF debug information
// comes from.
// TODO: this doesn't declare a variable for the output parameter. Is that
// what we want?
for (auto [argument_number, type] :
llvm::enumerate(llvm::drop_begin(subroutine_type->getTypeArray()))) {
context().di_builder().createParameterVariable(
subprogram, "", argument_number + 1, nullptr, 0, type,
/*AlwaysPreserve=*/true);
}
return subprogram;
}
// Given an LLVM type, build a corresponding type with `padding_bytes` bytes of
// explicit tail padding.
static auto BuildTailPaddedType(llvm::Type* subtype, int64_t padding_bytes)
-> llvm::Type* {
if (padding_bytes == 0) {
return subtype;
}
// Build the type `<{subtype, [i8 x padding_bytes]}>`.
llvm::Type* type_with_padding[2] = {
subtype,
llvm::ArrayType::get(llvm::Type::getInt8Ty(subtype->getContext()),
padding_bytes)};
return llvm::StructType::get(subtype->getContext(), type_with_padding,
/*isPacked=*/true);
}
// BuildTypeForInst is used to construct types for FileContext::BuildType below.
// Implementations return the LLVM type for the instruction. This first overload
// is the fallback handler for non-type instructions.
template <typename InstT>
requires(InstT::Kind.is_type() == SemIR::InstIsType::Never)
static auto BuildTypeForInst(FileContext& /*context*/, InstT inst)
-> FileContext::LoweredTypes {
CARBON_FATAL("Cannot use inst as type: {0}", inst);
}
template <typename InstT>
requires(InstT::Kind.is_symbolic_when_type())
static auto BuildTypeForInst(FileContext& context, InstT /*inst*/)
-> FileContext::LoweredTypes {
// Treat non-monomorphized symbolic types as opaque.
return {llvm::StructType::get(context.llvm_context()), nullptr};
}
static auto BuildTypeForInst(FileContext& context, SemIR::ArrayType inst)
-> FileContext::LoweredTypes {
auto elem_type_id = context.sem_ir().types().GetTypeIdForTypeInstId(
inst.element_type_inst_id);
auto stride = context.sem_ir()
.types()
.GetCompleteTypeInfo(elem_type_id)
.object_layout.ArrayStride();
auto* elem_type = context.GetType(elem_type_id);
auto elem_size = SemIR::ObjectSize::Bytes(
context.llvm_module().getDataLayout().getTypeAllocSize(elem_type));
if (elem_size != stride) {
CARBON_CHECK(elem_size < stride, "Array element type too large");
elem_type = BuildTailPaddedType(context.GetType(elem_type_id),
stride.bytes() - elem_size.bytes());
}
return {llvm::ArrayType::get(
elem_type, *context.sem_ir().GetZExtIntValue(inst.bound_id)),
nullptr};
}
static auto BuildTypeForInst(FileContext& context, SemIR::BoolType /*inst*/)
-> FileContext::LoweredTypes {
// TODO: We may want to have different representations for `bool` storage
// (`i8`) versus for `bool` values (`i1`).
return {llvm::Type::getInt1Ty(context.llvm_context()), nullptr};
}
static auto BuildTypeForInst(FileContext& context, SemIR::ClassType inst)
-> FileContext::LoweredTypes {
auto object_repr_id = context.sem_ir()
.classes()
.Get(inst.class_id)
.GetObjectRepr(context.sem_ir(), inst.specific_id);
return context.GetTypeAndDIType(object_repr_id);
}
template <typename InstT>
requires(SemIR::Internal::HasInstCategory<SemIR::AnyQualifiedType, InstT>)
static auto BuildTypeForInst(FileContext& context, InstT inst)
-> FileContext::LoweredTypes {
return {context.GetType(
context.sem_ir().types().GetTypeIdForTypeInstId(inst.inner_id)),
nullptr};
}
static auto BuildTypeForInst(FileContext& context, SemIR::CustomLayoutType inst)
-> FileContext::LoweredTypes {
auto layout = context.sem_ir().custom_layouts().Get(inst.layout_id);
return {
llvm::ArrayType::get(llvm::Type::getInt8Ty(context.llvm_context()),
layout[SemIR::CustomLayoutId::SizeIndex].bytes()),
nullptr};
}
static auto BuildTypeForInst(FileContext& context,
SemIR::ImplWitnessAssociatedConstant inst)
-> FileContext::LoweredTypes {
return {context.GetType(inst.type_id), nullptr};
}
static auto BuildTypeForInst(FileContext& /*context*/,
SemIR::ErrorInst /*inst*/)
-> FileContext::LoweredTypes {
// This is a complete type but uses of it should never be lowered.
return {nullptr, nullptr};
}
static auto BuildTypeForInst(FileContext& context, SemIR::FloatType inst)
-> FileContext::LoweredTypes {
return {llvm::Type::getFloatingPointTy(context.llvm_context(),
inst.float_kind.Semantics()),
nullptr};
}
static auto BuildTypeForInst(FileContext& context, SemIR::IntType inst)
-> FileContext::LoweredTypes {
auto width_inst =
context.sem_ir().insts().TryGetAs<SemIR::IntValue>(inst.bit_width_id);
CARBON_CHECK(width_inst, "Can't lower int type with symbolic width");
auto width = context.sem_ir().ints().Get(width_inst->int_id).getZExtValue();
return {llvm::IntegerType::get(context.llvm_context(), width),
context.context().di_builder().createBasicType(
"int", width,
inst.int_kind.is_signed() ? llvm::dwarf::DW_ATE_signed
: llvm::dwarf::DW_ATE_unsigned)};
}
static auto BuildTypeForInst(FileContext& /*context*/,
SemIR::ImplWitnessAccess /*inst*/)
-> FileContext::LoweredTypes {
CARBON_FATAL("Unexpected ImplWitnessAccess in lowering");
}
static auto BuildTypeForInst(FileContext& context, SemIR::PointerType /*inst*/)
-> FileContext::LoweredTypes {
return {llvm::PointerType::get(context.llvm_context(), /*AddressSpace=*/0),
nullptr};
}
static auto BuildTypeForInst(FileContext& /*context*/,
SemIR::PatternType /*inst*/)
-> FileContext::LoweredTypes {
CARBON_FATAL("Unexpected pattern type in lowering");
}
// Builds an LLVM packed struct type whose layout matches the Carbon layout for
// an aggregate with the given field types and field layouts.
static auto BuildPackedStructType(FileContext& context,
llvm::MutableArrayRef<llvm::Type*> subtypes,
llvm::ArrayRef<SemIR::ObjectLayout> layouts)
-> llvm::StructType* {
const auto& data_layout = context.llvm_module().getDataLayout();
auto struct_layout = SemIR::ObjectLayout::Empty();
auto size_so_far = SemIR::ObjectSize::Zero();
llvm::Type** previous_type = nullptr;
for (auto [type, layout] : llvm::zip_equal(subtypes, layouts)) {
auto offset = struct_layout.FieldOffset(layout);
// If this field has padding before it, represent that padding explicitly as
// part of the previous field. This allows us to always use GEP indexes that
// match the field indexes.
if (offset != size_so_far) {
CARBON_CHECK(previous_type, "Padding before first field?");
CARBON_CHECK(offset > size_so_far, "Extraneous padding after field {0}",
**previous_type);
int64_t padding_bytes = offset.bytes() - struct_layout.size.bytes();
*previous_type = BuildTailPaddedType(*previous_type, padding_bytes);
size_so_far += SemIR::ObjectSize::Bytes(padding_bytes);
CARBON_CHECK(offset == size_so_far, "Field at non-byte offset");
}
size_so_far += SemIR::ObjectSize::Bytes(data_layout.getTypeAllocSize(type));
struct_layout.AppendField(layout);
previous_type = &type;
}
return llvm::StructType::get(context.llvm_context(), subtypes,
/*isPacked=*/true);
}
// Returns whether the given LLVM layout matches the expected Carbon layout for
// an aggregate with the given field layouts.
static auto StructLayoutMatches(llvm::ArrayRef<SemIR::ObjectLayout> layouts,
const llvm::StructLayout& llvm_layout) -> bool {
auto struct_layout = SemIR::ObjectLayout::Empty();
// Check each field is at the right offset.
for (auto [i, layout] : llvm::enumerate(layouts)) {
if (static_cast<int64_t>(llvm_layout.getElementOffsetInBits(i)) !=
struct_layout.FieldOffset(layout).bits()) {
return false;
}
struct_layout.AppendField(layout);
}
// Treat the LLVM layout as being acceptable if it's the right byte size and
// does not require more alignment than the Carbon type. We could ignore the
// alignment, but an overaligned LLVM type will prevent the type from being
// used in non-packed structs in more situations.
return static_cast<int64_t>(llvm_layout.getSizeInBytes()) ==
struct_layout.size.bytes() &&
llvm_layout.getAlignment() <=
llvm::Align(struct_layout.alignment.bytes());
}
// Builds an LLVM struct type whose layout matches the Carbon layout for an
// aggregate with the given field types and field layouts.
static auto BuildStructType(FileContext& context,
llvm::MutableArrayRef<llvm::Type*> subtypes,
llvm::ArrayRef<SemIR::ObjectLayout> layouts)
-> FileContext::LoweredTypes {
// Opportunistically try building an llvm StructType from the subtypes. If it
// has the right layout, we're done. We prefer to use a non-packed struct type
// where possible to produce a smaller LLVM IR representation for the type and
// for constant values of the type, and to improve the readability of the IR.
auto* struct_type = llvm::StructType::get(context.llvm_context(), subtypes);
if (!StructLayoutMatches(
layouts, *context.llvm_module().getDataLayout().getStructLayout(
struct_type))) {
struct_type = BuildPackedStructType(context, subtypes, layouts);
}
return {struct_type, nullptr};
}
static auto BuildTypeForInst(FileContext& context, SemIR::StructType inst)
-> FileContext::LoweredTypes {
auto fields = context.sem_ir().struct_type_fields().Get(inst.fields_id);
llvm::SmallVector<llvm::Type*> subtypes;
llvm::SmallVector<SemIR::ObjectLayout> layouts;
subtypes.reserve(fields.size());
layouts.reserve(fields.size());
for (auto field : fields) {
auto type_id =
context.sem_ir().types().GetTypeIdForTypeInstId(field.type_inst_id);
subtypes.push_back(context.GetType(type_id));
layouts.push_back(
context.sem_ir().types().GetCompleteTypeInfo(type_id).object_layout);
}
return BuildStructType(context, subtypes, layouts);
}
static auto BuildTypeForInst(FileContext& context, SemIR::TupleType inst)
-> FileContext::LoweredTypes {
// TODO: Investigate special-casing handling of empty tuples so that they
// can be collectively replaced with LLVM's void, particularly around
// function returns. LLVM doesn't allow declaring variables with a void
// type, so that may require significant special casing.
auto elements = context.sem_ir().inst_blocks().Get(inst.type_elements_id);
llvm::SmallVector<llvm::Type*> subtypes;
llvm::SmallVector<SemIR::ObjectLayout> layouts;
subtypes.reserve(elements.size());
layouts.reserve(elements.size());
for (auto type_id : context.sem_ir().types().GetBlockAsTypeIds(elements)) {
subtypes.push_back(context.GetType(type_id));
layouts.push_back(
context.sem_ir().types().GetCompleteTypeInfo(type_id).object_layout);
}
return BuildStructType(context, subtypes, layouts);
}
static auto BuildTypeForInst(FileContext& context, SemIR::TypeType /*inst*/)
-> FileContext::LoweredTypes {
return {context.GetTypeType(), nullptr};
}
static auto BuildTypeForInst(FileContext& context, SemIR::FormType /*inst*/)
-> FileContext::LoweredTypes {
return {context.GetFormType(), nullptr};
}
static auto BuildTypeForInst(FileContext& context, SemIR::VtableType /*inst*/)
-> FileContext::LoweredTypes {
return {llvm::Type::getVoidTy(context.llvm_context()), nullptr};
}
template <typename InstT>
requires(InstT::Kind.template IsAnyOf<
SemIR::AssociatedEntityType, SemIR::AutoType, SemIR::BoundMethodType,
SemIR::CharLiteralType, SemIR::CppOverloadSetType,
SemIR::CppTemplateNameType, SemIR::FacetType,
SemIR::FloatLiteralType, SemIR::FunctionType,
SemIR::FunctionTypeWithSelfType, SemIR::GenericClassType,
SemIR::GenericInterfaceType, SemIR::GenericNamedConstraintType,
SemIR::InstType, SemIR::IntLiteralType, SemIR::NamespaceType,
SemIR::RequireSpecificDefinitionType, SemIR::SpecificFunctionType,
SemIR::UnboundElementType, SemIR::WhereExpr, SemIR::WitnessType>())
static auto BuildTypeForInst(FileContext& context, InstT /*inst*/)
-> FileContext::LoweredTypes {
// Return an empty struct as a placeholder.
// TODO: Should we model an interface as a witness table, or an associated
// entity as an index?
return {llvm::StructType::get(context.llvm_context()), nullptr};
}
auto FileContext::BuildType(SemIR::InstId inst_id) -> LoweredTypes {
// Use overload resolution to select the implementation, producing compile
// errors when BuildTypeForInst isn't defined for a given instruction.
LoweredTypes result;
CARBON_KIND_SWITCH(sem_ir_->insts().Get(inst_id)) {
#define CARBON_SEM_IR_INST_KIND(Name) \
case CARBON_KIND(SemIR::Name inst): { \
result = BuildTypeForInst(*this, inst); \
break; \
}
#include "toolchain/sem_ir/inst_kind.def"
}
// In debug builds, check that the type we built has the expected size.
CARBON_DCHECK([&] {
if (!result.llvm_ir_type) {
return true;
}
const auto& layout = llvm_module().getDataLayout();
auto expected_layout =
sem_ir()
.types()
.GetCompleteTypeInfo(
sem_ir().types().GetTypeIdForTypeInstId(inst_id))
.object_layout;
CARBON_CHECK(expected_layout.has_value());
auto size =
SemIR::ObjectSize::Bits(layout.getTypeSizeInBits(result.llvm_ir_type));
// Round up to byte granularity for this check, since LLVM doesn't support
// non-byte-sized packed structs.
CARBON_CHECK(
size.bytes() == expected_layout.size.bytes(),
"Lowered type {0} for {1} has unexpected size {2}, expected {3}",
*result.llvm_ir_type, sem_ir().insts().Get(inst_id), size,
expected_layout.size);
return true;
}());
return result;
}
auto FileContext::BuildGlobalVariableDecl(SemIR::VarStorage var_storage)
-> llvm::Constant* {
auto var_name_id =
SemIR::GetFirstBindingNameFromPatternId(sem_ir(), var_storage.pattern_id);
if (auto cpp_global_var_id =
sem_ir().cpp_global_vars().Lookup({.entity_name_id = var_name_id});
cpp_global_var_id.has_value()) {
SemIR::ClangDeclId clang_decl_id =
sem_ir().cpp_global_vars().Get(cpp_global_var_id).clang_decl_id;
CARBON_CHECK(clang_decl_id.has_value(),
"CppGlobalVar should have a clang_decl_id");
return cpp_code_generator_->GetAddrOfGlobal(
cast<clang::VarDecl>(
sem_ir().clang_decls().Get(clang_decl_id).key.decl),
/*isForDefinition=*/false);
}
return BuildNonCppGlobalVariableDecl(var_storage);
}
auto FileContext::BuildNonCppGlobalVariableDecl(SemIR::VarStorage var_storage)
-> llvm::GlobalVariable* {
SemIR::Mangler m(sem_ir(), context().total_ir_count());
auto mangled_name = m.MangleGlobalVariable(var_storage.pattern_id);
auto linkage = llvm::GlobalVariable::ExternalLinkage;
// If the variable doesn't have an externally-visible name, demote it to
// internal linkage and invent a plausible name that shouldn't collide with
// any of our real manglings.
if (mangled_name.empty()) {
linkage = llvm::GlobalVariable::InternalLinkage;
if (inst_namer_) {
mangled_name =
("var.anon" + inst_namer_->GetUnscopedNameFor(var_storage.pattern_id))
.str();
}
}
auto* type = GetType(var_storage.type_id);
return new llvm::GlobalVariable(llvm_module(), type,
/*isConstant=*/false, linkage,
/*Initializer=*/nullptr, mangled_name);
}
auto FileContext::GetLocForDI(SemIR::InstId inst_id) -> Context::LocForDI {
auto abs_node_ref = GetAbsoluteNodeRef(sem_ir_, SemIR::LocId(inst_id)).back();
return context().GetLocForDI(abs_node_ref);
}
auto FileContext::BuildVtable(const SemIR::Vtable& vtable,
SemIR::SpecificId specific_id)
-> llvm::GlobalVariable* {
const auto& class_info = sem_ir().classes().Get(vtable.class_id);
SemIR::Mangler m(sem_ir(), context().total_ir_count());
std::string mangled_name = m.MangleVTable(class_info, specific_id);
if (sem_ir()
.insts()
.GetImportSource(class_info.first_owning_decl_id)
.has_value()) {
// Emit a declaration of an imported vtable using a(n opaque) pointer type.
// This doesn't have to match the definition that appears elsewhere, it'll
// still get merged correctly.
auto* gv = new llvm::GlobalVariable(
llvm_module(),
llvm::PointerType::get(llvm_context(), /*AddressSpace=*/0),
/*isConstant=*/true, llvm::GlobalValue::ExternalLinkage, nullptr,
mangled_name);
gv->setUnnamedAddr(llvm::GlobalValue::UnnamedAddr::Global);
return gv;
}
auto vtable_inst_block =
sem_ir().inst_blocks().Get(vtable.virtual_functions_id);
auto* entry_type = llvm::IntegerType::getInt32Ty(llvm_context());
auto* table_type = llvm::ArrayType::get(entry_type, vtable_inst_block.size());
auto* llvm_vtable = new llvm::GlobalVariable(
llvm_module(), table_type, /*isConstant=*/true,
llvm::GlobalValue::ExternalLinkage, nullptr, mangled_name);
auto* i32_type = llvm::IntegerType::getInt32Ty(llvm_context());
auto* i64_type = llvm::IntegerType::getInt64Ty(llvm_context());
auto* vtable_const_int =
llvm::ConstantExpr::getPtrToInt(llvm_vtable, i64_type);
llvm::SmallVector<llvm::Constant*> vfuncs;
vfuncs.reserve(vtable_inst_block.size());
for (auto fn_decl_id : vtable_inst_block) {
auto [_1, _2, fn_id, fn_specific_id] =
DecomposeVirtualFunction(sem_ir(), fn_decl_id, specific_id);
vfuncs.push_back(llvm::ConstantExpr::getTrunc(
llvm::ConstantExpr::getSub(
llvm::ConstantExpr::getPtrToInt(
GetOrCreateFunctionInfo(fn_id, fn_specific_id)->llvm_function,
i64_type),
vtable_const_int),
i32_type));
}
llvm_vtable->setInitializer(llvm::ConstantArray::get(table_type, vfuncs));
llvm_vtable->setUnnamedAddr(llvm::GlobalValue::UnnamedAddr::Global);
return llvm_vtable;
}
} // namespace Carbon::Lower
Reference in New Issue View Git Blame Copy Permalink