feat(codegen): arrange EXTERN signatures with clang's C ABI

Every EXTERN "C" signature is now arranged by clang's CodeGen
(arrangeFreeFunctionCall) for the host triple and applied by one generic
lowering routine, replacing the hand-written signext/zeroext rule.
clangCodeGen is linked statically into CPSCodeGen. Records still cross
only through pointer types, so generated code is unchanged.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

7 files changed, 480 insertions, 26 deletions

diff --git a/CMakeLists.txt b/CMakeLists.txt
index e1cf41b..8207da4 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -83,6 +83,8 @@ add_library(CPSCodeGen STATIC
     lib/CodeGen/MemoryHandler.cc
     lib/CodeGen/FileHandler.cc
     lib/CodeGen/FileRuntimeIO.cc
+    lib/CodeGen/CABI/ClangABI.cc
+    lib/CodeGen/CABI/Lowering.cc
 )
 
 llvm_map_components_to_libnames(llvm_libs core support native passes object)
diff --git a/cmake/NativeCompiler.cmake b/cmake/NativeCompiler.cmake
index e0b4b3f..5ff49cb 100644
--- a/cmake/NativeCompiler.cmake
+++ b/cmake/NativeCompiler.cmake
@@ -19,7 +19,7 @@ else()
     set(cpsc_linker_libraries lldELF)
 endif()
 
-foreach(cpsc_native_library clangDriver ${cpsc_linker_libraries})
+foreach(cpsc_native_library clangDriver clangCodeGen ${cpsc_linker_libraries})
     get_target_property(cpsc_native_type ${cpsc_native_library} TYPE)
     if(NOT cpsc_native_type STREQUAL "STATIC_LIBRARY")
         message(FATAL_ERROR "Native output requires static ${cpsc_native_library}; found ${cpsc_native_type}. See doc/build.md for static dependency setup.")
@@ -75,6 +75,9 @@ target_sources(cpsc PRIVATE tools/driver/NativeCompiler.cc)
 target_include_directories(cpsc PRIVATE ${CLANG_INCLUDE_DIRS} ${LLD_INCLUDE_DIRS}
     "${CMAKE_BINARY_DIR}/native-runtime")
 target_link_libraries(cpsc PRIVATE clangDriver ${cpsc_linker_libraries})
+# EXTERN signatures use clang's C ABI code (lib/CodeGen/CABI/ClangABI.cc).
+target_include_directories(CPSCodeGen PRIVATE ${CLANG_INCLUDE_DIRS})
+target_link_libraries(CPSCodeGen PRIVATE clangCodeGen clangDriver)
 if(MINGW AND TARGET LibXml2::LibXml2)
     # LLD's COFF manifest support reaches libxml2 through LLVMWindowsManifest.
     # An explicit .dll.a in that interface bypasses the driver's -static flag.
diff --git a/include/cps/CABI.h b/include/cps/CABI.h
new file mode 100644
index 0000000..ba776e9
--- /dev/null
+++ b/include/cps/CABI.h
@@ -0,0 +1,76 @@
+#pragma once
+#include "llvm/ADT/ArrayRef.h"
+#include "llvm/IR/Attributes.h"
+#include "llvm/IR/DerivedTypes.h"
+#include "llvm/IR/IRBuilder.h"
+#include <cstdint>
+#include <memory>
+#include <string>
+#include <vector>
+
+namespace cps {
+
+struct TypeInfo;
+class TypeSystem;
+
+// One EXTERN "C" parameter or result as clang's ABIArgInfo arranges it. Only
+// the kinds our targets use exist here; ClangABI reports any other kind as an
+// internal error.
+struct CABIArg {
+    enum class Kind { Direct, Extend, Indirect, Ignore };
+    Kind K = Kind::Direct;
+    llvm::Type *ValueType = nullptr;  // the pseudocode value; void for VOID
+    llvm::Type *CoerceType = nullptr; // Direct, Extend: the IR type passed
+    unsigned Offset = 0;              // Direct, Extend: byte offset of the coerced part
+    bool Flatten = false;             // Direct: a struct CoerceType passed element by element
+    unsigned DirectAlign = 0;         // Direct, Extend: stack alignment, if any
+    llvm::Attribute::AttrKind Extension = llvm::Attribute::None; // Extend: SExt, ZExt or NoExt
+    bool ByVal = false;               // Indirect: byval, otherwise a pointer to a copy
+    uint64_t Align = 0;               // Indirect: alignment of the memory in bytes
+    bool InReg = false;
+};
+
+// An arranged EXTERN signature. Args holds the Fixed prototype parameters and,
+// for one variadic call, the promoted variadic arguments after them.
+struct CABISignature {
+    CABIArg Ret;
+    std::vector<CABIArg> Args;
+    size_t Fixed = 0;
+    bool Variadic = false;
+    unsigned CallingConv = 0;
+};
+
+// Arranges signatures for the host C ABI with clang's CodeGen. The clang
+// context is created on the first call; ClangABI.cc is the only file that
+// includes clang headers.
+class ClangABI {
+public:
+    ClangABI(llvm::LLVMContext &Context, const TypeSystem &Types);
+    ~ClangABI();
+    // Ret is nullptr for VOID; the first Fixed of Args are the prototype's
+    // parameters. On failure Err holds the reason for an internal error.
+    bool arrange(const TypeInfo *Ret, const std::vector<const TypeInfo *> &Args, size_t Fixed,
+                 bool Variadic, CABISignature &Sig, std::string &Err);
+
+private:
+    struct State;
+    llvm::LLVMContext &Context;
+    const TypeSystem &Types;
+    std::unique_ptr<State> S;
+};
+
+// Generic lowering of an arranged signature (Lowering.cc), without
+// per-architecture code.
+// The declared function type, from the fixed parameters.
+llvm::FunctionType *getCABIFunctionType(const CABISignature &Sig);
+// The call attributes, for every argument in Sig.
+llvm::AttributeList getCABIAttributes(llvm::LLVMContext &C, const CABISignature &Sig);
+// The IR arguments for one pseudocode value per Sig.Args entry. SRet receives
+// the result memory when the result is returned indirectly.
+std::vector<llvm::Value *> lowerCABIArguments(llvm::IRBuilder<> &B, const CABISignature &Sig,
+                                              llvm::ArrayRef<llvm::Value *> Args, llvm::Value *&SRet);
+// The pseudocode result of Call; nullptr for VOID.
+llvm::Value *lowerCABIResult(llvm::IRBuilder<> &B, const CABISignature &Sig, llvm::Value *Call,
+                             llvm::Value *SRet);
+
+} // namespace cps
diff --git a/include/cps/FunctionGen.h b/include/cps/FunctionGen.h
index 1b4a07a..8b63d67 100644
--- a/include/cps/FunctionGen.h
+++ b/include/cps/FunctionGen.h
@@ -1,4 +1,5 @@
 #pragma once
+#include "cps/CABI.h"
 #include "cps/Identifier.h"
 #include "llvm/IR/IRBuilder.h"
 #include "llvm/IR/Module.h"
@@ -45,9 +46,11 @@ struct FuncSig {
     std::vector<ParamSig> Params;
     size_t OwnerUnit = 0;
     // EXTERN "C": called through its declared C type, which may differ from a
-    // runtime declaration of the same libc symbol.
+    // runtime declaration of the same libc symbol. ExternABI is clang's
+    // arrangement of the fixed parameters; ExternType and ExternAttrs lower it.
     bool IsExternal = false;
     bool IsVariadic = false;
+    CABISignature ExternABI;
     llvm::FunctionType *ExternType = nullptr;
     llvm::AttributeList ExternAttrs;
 };
@@ -83,6 +86,7 @@ private:
     std::function<void(const FuncSig &)> BeforeCall;
     ArrayParamBinder BindArrayParam;
     std::function<void(llvm::Value *, const std::string &, int)> RegisterCopy;
+    ClangABI ABI{Context, Types}; // clang context created on the first EXTERN
 
     void createArgumentAllocas(llvm::Function *F, const std::vector<ParamDecl> &Params, bool IsMethod);
     llvm::Type *getLLVMType(const std::string &TypeName);
diff --git a/lib/CodeGen/CABI/ClangABI.cc b/lib/CodeGen/CABI/ClangABI.cc
new file mode 100644
index 0000000..3757a94
--- /dev/null
+++ b/lib/CodeGen/CABI/ClangABI.cc
@@ -0,0 +1,227 @@
+// The only file that talks to clang. Keep it that way: clang::CodeGen clashes
+// with cps::CodeGen, clang is built without RTTI (so never subclass its
+// classes here), and the CodeGenerator must be destroyed before the ASTContext.
+#include "cps/CABI.h"
+#include "cps/TypeSystem.h"
+#include "clang/AST/ASTContext.h"
+#include "clang/AST/Decl.h"
+#include "clang/Basic/Builtins.h"
+#include "clang/Basic/Diagnostic.h"
+#include "clang/Basic/DiagnosticOptions.h"
+#include "clang/Basic/FileManager.h"
+#include "clang/Basic/IdentifierTable.h"
+#include "clang/Basic/SourceManager.h"
+#include "clang/Basic/TargetInfo.h"
+#include "clang/CodeGen/CGFunctionInfo.h"
+#include "clang/CodeGen/CodeGenABITypes.h"
+#include "clang/CodeGen/ModuleBuilder.h"
+#include "clang/Driver/CreateInvocationFromArgs.h"
+#include "clang/Frontend/CompilerInvocation.h"
+#include "llvm/IR/Module.h"
+#include "llvm/Support/VirtualFileSystem.h"
+#include "llvm/TargetParser/Host.h"
+#include <map>
+
+using namespace cps;
+using clang::CodeGen::ABIArgInfo;
+
+// Members are destroyed in reverse order, so Gen goes before AST.
+struct ClangABI::State {
+    const TypeSystem *Types = nullptr;
+    clang::DiagnosticOptions DiagOpts;
+    llvm::IntrusiveRefCntPtr<clang::DiagnosticsEngine> Diags;
+    std::shared_ptr<clang::CompilerInvocation> Invocation;
+    llvm::IntrusiveRefCntPtr<clang::FileManager> Files;
+    std::unique_ptr<clang::SourceManager> Sources;
+    llvm::IntrusiveRefCntPtr<clang::TargetInfo> Target;
+    std::unique_ptr<clang::IdentifierTable> Idents;
+    clang::SelectorTable Selectors;
+    clang::Builtin::Context Builtins;
+    std::unique_ptr<clang::ASTContext> AST;
+    std::unique_ptr<clang::CodeGenerator> Gen;
+    std::map<const TypeInfo *, clang::CanQualType> Records;
+
+    // Target and language options come from a driver invocation for the host
+    // triple, as the C compiler would use them (e.g. Apple's darwinpcs and the
+    // signedness of char).
+    bool init(llvm::LLVMContext &Context, std::string &Err) {
+        Diags = llvm::makeIntrusiveRefCnt<clang::DiagnosticsEngine>(
+            clang::DiagnosticIDs::create(), DiagOpts, new clang::IgnoringDiagConsumer());
+        std::string Triple = llvm::sys::getProcessTriple();
+        std::string TargetArg = "--target=" + Triple;
+        const char *Args[] = {"clang", TargetArg.c_str(), "-fsyntax-only", "-x", "c", "cps-abi.c"};
+        clang::CreateInvocationOptions Opts;
+        Opts.Diags = Diags;
+        Invocation = clang::createInvocation(Args, Opts);
+        if (Invocation) Target = clang::TargetInfo::CreateTargetInfo(*Diags, Invocation->getTargetOpts());
+        if (!Target) {
+            Err = "cannot create a clang C ABI context for " + Triple;
+            return false;
+        }
+        clang::LangOptions &Lang = Invocation->getLangOpts();
+        Target->adjust(*Diags, Lang, nullptr);
+        auto VFS = llvm::vfs::getRealFileSystem();
+        Files = llvm::makeIntrusiveRefCnt<clang::FileManager>(clang::FileSystemOptions(), VFS);
+        Sources = std::make_unique<clang::SourceManager>(*Diags, *Files);
+        Idents = std::make_unique<clang::IdentifierTable>(Lang);
+        AST = std::make_unique<clang::ASTContext>(Lang, *Sources, *Idents, Selectors, Builtins,
+                                                  clang::TU_Complete);
+        AST->InitBuiltinTypes(*Target);
+        Gen = clang::CreateLLVMCodeGen(*Diags, "cps-abi", VFS, Invocation->getHeaderSearchOpts(),
+                                       Invocation->getPreprocessorOpts(),
+                                       Invocation->getCodeGenOpts(), Context);
+        Gen->Initialize(*AST);
+        return true;
+    }
+
+    // C scalars map by name and every pointer to void *; records are built
+    // once and checked against cpsc's layout.
+    bool map(const TypeInfo *Info, clang::CanQualType &Out, std::string &Err) {
+        using Member = clang::CanQualType clang::ASTContext::*;
+        static const std::map<std::string, Member> Scalars = {
+            {"_Bool", &clang::ASTContext::BoolTy},
+            {"char", &clang::ASTContext::CharTy},
+            {"signed char", &clang::ASTContext::SignedCharTy},
+            {"unsigned char", &clang::ASTContext::UnsignedCharTy},
+            {"short", &clang::ASTContext::ShortTy},
+            {"unsigned short", &clang::ASTContext::UnsignedShortTy},
+            {"int", &clang::ASTContext::IntTy},
+            {"unsigned int", &clang::ASTContext::UnsignedIntTy},
+            {"long", &clang::ASTContext::LongTy},
+            {"unsigned long", &clang::ASTContext::UnsignedLongTy},
+            {"long long", &clang::ASTContext::LongLongTy},
+            {"unsigned long long", &clang::ASTContext::UnsignedLongLongTy},
+            {"float", &clang::ASTContext::FloatTy},
+            {"double", &clang::ASTContext::DoubleTy},
+            {"long double", &clang::ASTContext::LongDoubleTy},
+        };
+        if (Info && (Info->Kind == TypeKind::CPointer || Info->Kind == TypeKind::Pointer)) {
+            Out = AST->VoidPtrTy;
+            return true;
+        }
+        if (Info && Info->Kind == TypeKind::CScalar) {
+            auto It = Scalars.find(Info->CIdentity);
+            if (It != Scalars.end()) {
+                Out = (*AST).*(It->second);
+                return true;
+            }
+        }
+        if (!Info || !Info->isRecord()) {
+            Err = "no C type for " + (Info ? Info->Name : std::string("an unknown type"));
+            return false;
+        }
+        auto Cached = Records.find(Info);
+        if (Cached != Records.end()) {
+            Out = Cached->second;
+            return true;
+        }
+        clang::TranslationUnitDecl *Unit = AST->getTranslationUnitDecl();
+        auto *Record = clang::RecordDecl::Create(*AST, clang::TagTypeKind::Struct, Unit, {}, {},
+                                                 &AST->Idents.get(Info->Name));
+        Record->startDefinition();
+        for (const auto &[Name, Field] : Info->Record.Fields) {
+            clang::CanQualType FieldType;
+            if (!map(Types->resolve(Field.TypeName), FieldType, Err)) return false;
+            auto *Decl = clang::FieldDecl::Create(*AST, Record, {}, {}, &AST->Idents.get(Name),
+                                                  FieldType, nullptr, nullptr, false,
+                                                  clang::ICIS_NoInit);
+            Decl->setAccess(clang::AS_public);
+            Record->addDecl(Decl);
+        }
+        Record->completeDefinition();
+        Unit->addDecl(Record);
+        Out = AST->getCanonicalTagType(Record);
+        uint64_t CSize = AST->getTypeSizeInChars(Out).getQuantity();
+        uint64_t Size = Gen->GetModule()->getDataLayout().getTypeAllocSize(Info->LLVMType);
+        if (CSize != Size) {
+            Err = "record " + Info->Name + " has " + std::to_string(Size) + " bytes in cpsc but " +
+                  std::to_string(CSize) + " in C";
+            return false;
+        }
+        Records.emplace(Info, Out);
+        return true;
+    }
+};
+
+namespace {
+
+bool translate(const ABIArgInfo &I, llvm::Type *ValueType, const std::string &What, CABIArg &Out,
+               std::string &Err) {
+    Out.ValueType = ValueType;
+    const char *Unsupported = nullptr;
+    switch (I.getKind()) {
+    case ABIArgInfo::Direct:
+    case ABIArgInfo::Extend:
+        Out.K = I.isDirect() ? CABIArg::Kind::Direct : CABIArg::Kind::Extend;
+        Out.CoerceType = I.getCoerceToType();
+        Out.Offset = I.getDirectOffset();
+        Out.DirectAlign = I.getDirectAlign();
+        Out.Flatten = I.isDirect() && I.getCanBeFlattened() && llvm::isa<llvm::StructType>(Out.CoerceType);
+        if (I.isExtend())
+            Out.Extension = I.isSignExt() ? llvm::Attribute::SExt
+                : I.isZeroExt() ? llvm::Attribute::ZExt : llvm::Attribute::NoExt;
+        Out.InReg = I.getInReg();
+        break;
+    case ABIArgInfo::Indirect:
+        Out.K = CABIArg::Kind::Indirect;
+        Out.ByVal = I.getIndirectByVal();
+        Out.Align = I.getIndirectAlign().getQuantity();
+        Out.InReg = I.getInReg();
+        break;
+    case ABIArgInfo::Ignore:
+        Out.K = CABIArg::Kind::Ignore;
+        break;
+    case ABIArgInfo::IndirectAliased: Unsupported = "IndirectAliased"; break;
+    case ABIArgInfo::Expand: Unsupported = "Expand"; break;
+    case ABIArgInfo::CoerceAndExpand: Unsupported = "CoerceAndExpand"; break;
+    case ABIArgInfo::InAlloca: Unsupported = "InAlloca"; break;
+    case ABIArgInfo::TargetSpecific: Unsupported = "TargetSpecific"; break;
+    }
+    if (!Unsupported && I.getPaddingType()) Unsupported = "with a padding type";
+    if (Unsupported) {
+        Err = "unsupported C ABI kind " + std::string(Unsupported) + " for " + What;
+        return false;
+    }
+    return true;
+}
+
+} // namespace
+
+ClangABI::ClangABI(llvm::LLVMContext &Context, const TypeSystem &Types)
+    : Context(Context), Types(Types) {}
+
+ClangABI::~ClangABI() = default;
+
+bool ClangABI::arrange(const TypeInfo *Ret, const std::vector<const TypeInfo *> &Args, size_t Fixed,
+                       bool Variadic, CABISignature &Sig, std::string &Err) {
+    if (!S) {
+        auto Fresh = std::make_unique<State>();
+        Fresh->Types = &Types;
+        if (!Fresh->init(Context, Err)) return false;
+        S = std::move(Fresh);
+    }
+    clang::CanQualType RetType = S->AST->VoidTy;
+    if (Ret && !S->map(Ret, RetType, Err)) return false;
+    std::vector<clang::CanQualType> ArgTypes(Args.size());
+    for (size_t I = 0; I < Args.size(); ++I)
+        if (!S->map(Args[I], ArgTypes[I], Err)) return false;
+    clang::CodeGen::RequiredArgs Required = Variadic
+        ? clang::CodeGen::RequiredArgs(Fixed) : clang::CodeGen::RequiredArgs::All;
+    const clang::CodeGen::CGFunctionInfo &Info = clang::CodeGen::arrangeFreeFunctionCall(
+        S->Gen->CGM(), RetType, ArgTypes, clang::FunctionType::ExtInfo(), Required);
+
+    Sig = CABISignature();
+    Sig.Fixed = Fixed;
+    Sig.Variadic = Variadic;
+    Sig.CallingConv = Info.getEffectiveCallingConvention();
+    if (!translate(Info.getReturnInfo(), Ret ? Ret->LLVMType : llvm::Type::getVoidTy(Context),
+                   "the result", Sig.Ret, Err))
+        return false;
+    Sig.Args.resize(Args.size());
+    for (size_t I = 0; I < Args.size(); ++I)
+        if (!translate(Info.arguments()[I].info, Args[I]->LLVMType,
+                       (I < Fixed ? "parameter " : "variadic argument ") + std::to_string(I + 1),
+                       Sig.Args[I], Err))
+            return false;
+    return true;
+}
diff --git a/lib/CodeGen/CABI/Lowering.cc b/lib/CodeGen/CABI/Lowering.cc
new file mode 100644
index 0000000..32305ef
--- /dev/null
+++ b/lib/CodeGen/CABI/Lowering.cc
@@ -0,0 +1,140 @@
+// Applies a clang-arranged CABISignature with no per-architecture code. Sizes
+// are never computed here: the module has no DataLayout until emission, so a
+// coerced value goes through memory that holds both the value and its coerced
+// form.
+#include "cps/CABI.h"
+#include "llvm/IR/Function.h"
+#include "llvm/IR/Instructions.h"
+
+using namespace llvm;
+using namespace cps;
+
+namespace {
+
+bool isDirect(const CABIArg &A) {
+    return A.K == CABIArg::Kind::Direct || A.K == CABIArg::Kind::Extend;
+}
+
+// The IR values one argument occupies: none, one, or a flattened struct's elements.
+unsigned countIRArgs(const CABIArg &A) {
+    if (A.K == CABIArg::Kind::Ignore) return 0;
+    return A.Flatten ? cast<StructType>(A.CoerceType)->getNumElements() : 1;
+}
+
+void appendIRTypes(const CABIArg &A, std::vector<Type *> &Out) {
+    if (A.K == CABIArg::Kind::Indirect) Out.push_back(PointerType::getUnqual(A.ValueType->getContext()));
+    else if (A.Flatten) for (Type *Element : cast<StructType>(A.CoerceType)->elements()) Out.push_back(Element);
+    else if (isDirect(A)) Out.push_back(A.CoerceType);
+}
+
+AllocaInst *entryAlloca(IRBuilder<> &B, Type *Ty, uint64_t MinAlign, const Twine &Name) {
+    Function *F = B.GetInsertBlock()->getParent();
+    IRBuilder<> Entry(&F->getEntryBlock(), F->getEntryBlock().begin());
+    AllocaInst *Memory = Entry.CreateAlloca(Ty, nullptr, Name);
+    if (MinAlign > Memory->getAlign().value()) Memory->setAlignment(Align(MinAlign));
+    return Memory;
+}
+
+// Memory for the value at offset 0 and its coerced form at A.Offset; the
+// struct is at least as large and as aligned as either.
+struct CoercionSlot {
+    AllocaInst *Value;
+    llvm::Value *Coerced;
+    Align CoercedAlign;
+};
+
+CoercionSlot coercionSlot(IRBuilder<> &B, const CABIArg &A) {
+    LLVMContext &C = A.ValueType->getContext();
+    auto *Ty = StructType::get(C, {A.ValueType, ArrayType::get(Type::getInt8Ty(C), A.Offset), A.CoerceType});
+    AllocaInst *Memory = entryAlloca(B, Ty, 0, "c.coerce");
+    Value *Coerced = B.CreateConstInBoundsGEP1_32(Type::getInt8Ty(C), Memory, A.Offset);
+    return {Memory, Coerced, commonAlignment(Memory->getAlign(), A.Offset)};
+}
+
+AttributeSet argumentAttributes(LLVMContext &C, const CABIArg &A) {
+    AttrBuilder Attrs(C);
+    if (A.K == CABIArg::Kind::Extend) Attrs.addAttribute(A.Extension);
+    if (A.InReg) Attrs.addAttribute(Attribute::InReg);
+    if (isDirect(A) && A.DirectAlign) Attrs.addStackAlignmentAttr(A.DirectAlign);
+    if (A.K == CABIArg::Kind::Indirect && A.ByVal) {
+        Attrs.addByValAttr(A.ValueType);
+        Attrs.addAlignmentAttr(A.Align);
+    }
+    return AttributeSet::get(C, Attrs);
+}
+
+} // namespace
+
+FunctionType *cps::getCABIFunctionType(const CABISignature &Sig) {
+    LLVMContext &C = Sig.Ret.ValueType->getContext();
+    std::vector<Type *> Params;
+    if (Sig.Ret.K == CABIArg::Kind::Indirect) Params.push_back(PointerType::getUnqual(C));
+    for (size_t I = 0; I < Sig.Fixed; ++I) appendIRTypes(Sig.Args[I], Params);
+    Type *Ret = isDirect(Sig.Ret) ? Sig.Ret.CoerceType : Type::getVoidTy(C);
+    return FunctionType::get(Ret, Params, Sig.Variadic);
+}
+
+AttributeList cps::getCABIAttributes(LLVMContext &C, const CABISignature &Sig) {
+    AttrBuilder RetAttrs(C);
+    std::vector<AttributeSet> Params;
+    if (Sig.Ret.K == CABIArg::Kind::Indirect) {
+        AttrBuilder SRet(C);
+        SRet.addStructRetAttr(Sig.Ret.ValueType);
+        SRet.addAlignmentAttr(Sig.Ret.Align);
+        if (Sig.Ret.InReg) SRet.addAttribute(Attribute::InReg);
+        Params.push_back(AttributeSet::get(C, SRet));
+    } else if (isDirect(Sig.Ret)) {
+        if (Sig.Ret.K == CABIArg::Kind::Extend) RetAttrs.addAttribute(Sig.Ret.Extension);
+        if (Sig.Ret.InReg) RetAttrs.addAttribute(Attribute::InReg);
+    }
+    for (const CABIArg &A : Sig.Args)
+        Params.insert(Params.end(), countIRArgs(A), argumentAttributes(C, A));
+    return AttributeList::get(C, AttributeSet(), AttributeSet::get(C, RetAttrs), Params);
+}
+
+std::vector<Value *> cps::lowerCABIArguments(IRBuilder<> &B, const CABISignature &Sig,
+                                             ArrayRef<Value *> Args, Value *&SRet) {
+    std::vector<Value *> Out;
+    SRet = nullptr;
+    if (Sig.Ret.K == CABIArg::Kind::Indirect) {
+        SRet = entryAlloca(B, Sig.Ret.ValueType, Sig.Ret.Align, "c.result");
+        Out.push_back(SRet);
+    }
+    for (size_t I = 0; I < Args.size(); ++I) {
+        const CABIArg &A = Sig.Args[I];
+        Value *V = Args[I];
+        if (A.K == CABIArg::Kind::Ignore) continue;
+        if (A.K == CABIArg::Kind::Indirect) {
+            // byval or not, the callee receives memory of its own.
+            AllocaInst *Copy = entryAlloca(B, A.ValueType, A.Align, "c.arg");
+            B.CreateStore(V, Copy);
+            Out.push_back(Copy);
+            continue;
+        }
+        if (!A.Flatten && !A.Offset && A.CoerceType == V->getType()) {
+            Out.push_back(V);
+            continue;
+        }
+        CoercionSlot Slot = coercionSlot(B, A);
+        B.CreateStore(V, Slot.Value);
+        if (!A.Flatten) {
+            Out.push_back(B.CreateAlignedLoad(A.CoerceType, Slot.Coerced, Slot.CoercedAlign));
+            continue;
+        }
+        auto *Struct = cast<StructType>(A.CoerceType);
+        for (unsigned E = 0; E < Struct->getNumElements(); ++E)
+            Out.push_back(B.CreateLoad(Struct->getElementType(E), B.CreateStructGEP(Struct, Slot.Coerced, E)));
+    }
+    return Out;
+}
+
+Value *cps::lowerCABIResult(IRBuilder<> &B, const CABISignature &Sig, Value *Call, Value *SRet) {
+    const CABIArg &R = Sig.Ret;
+    if (R.ValueType->isVoidTy()) return nullptr;
+    if (R.K == CABIArg::Kind::Ignore) return Constant::getNullValue(R.ValueType);
+    if (R.K == CABIArg::Kind::Indirect) return B.CreateLoad(R.ValueType, SRet);
+    if (!R.Offset && R.CoerceType == R.ValueType) return Call;
+    CoercionSlot Slot = coercionSlot(B, R);
+    B.CreateAlignedStore(Call, Slot.Coerced, Slot.CoercedAlign);
+    return B.CreateLoad(R.ValueType, Slot.Value);
+}
diff --git a/lib/CodeGen/FunctionGen.cc b/lib/CodeGen/FunctionGen.cc
index 1cbf19f..2cf6da3 100644
--- a/lib/CodeGen/FunctionGen.cc
+++ b/lib/CodeGen/FunctionGen.cc
@@ -274,7 +274,7 @@ Function *FunctionGen::emitPrototype(PrototypeAST *Proto) {
 }
 
 // The prototype names the exact C symbol; parameters and the result must be C
-// types. Narrow integers carry the C ABI's sign/zero extension attributes.
+// types. clang arranges the signature for the host C ABI.
 bool FunctionGen::emitExternPrototype(PrototypeAST *Proto) {
     if (PrototypeLinks.count(Proto)) return true;
     const std::string &Name = Proto->getName();
@@ -291,13 +291,6 @@ bool FunctionGen::emitExternPrototype(PrototypeAST *Proto) {
         return false;
     }
 
-    auto Extension = [](const TypeInfo *Info) {
-        if (Info->Kind != TypeKind::CScalar || !Info->LLVMType->isIntegerTy() ||
-            Info->LLVMType->getIntegerBitWidth() >= 32)
-            return Attribute::None;
-        return Info->CSigned ? Attribute::SExt : Attribute::ZExt;
-    };
-
     // Records cross the C boundary only as C structs, for now through a
     // pointer type. What names the parameter or return type in diagnostics.
     auto CheckRecord = [&](const TypeInfo *Info, const std::string &What) {
@@ -327,9 +320,7 @@ bool FunctionGen::emitExternPrototype(PrototypeAST *Proto) {
     Sig.IsExternal = true;
     Sig.IsVariadic = Proto->isVariadic();
 
-    std::vector<Type *> ArgTypes;
-    AttrBuilder RetAttrs(Context);
-    std::vector<AttributeSet> ParamAttrs;
+    std::vector<const TypeInfo *> ArgTypes;
     for (size_t I = 0; I < Proto->getParams().size(); ++I) {
         const ParamDecl &P = Proto->getParams()[I];
         const TypeInfo *Info = Types.resolve(P.TypeName);
@@ -344,15 +335,12 @@ bool FunctionGen::emitExternPrototype(PrototypeAST *Proto) {
                         I + 1, Name.c_str(), P.TypeName.c_str());
             return false;
         }
-        ArgTypes.push_back(Info->LLVMType);
-        AttrBuilder B(Context);
-        if (Attribute::AttrKind K = Extension(Info); K != Attribute::None) B.addAttribute(K);
-        ParamAttrs.push_back(AttributeSet::get(Context, B));
+        ArgTypes.push_back(Info);
         ParamSig PS;
         PS.TypeName = P.TypeName;
         Sig.Params.push_back(std::move(PS));
     }
-    Type *RetType = Type::getVoidTy(Context);
+    const TypeInfo *RetType = nullptr;
     if (!sameIdentifier(Proto->getReturnType(), "VOID")) {
         const TypeInfo *Info = Types.resolve(Proto->getReturnType());
         if (!Info) {
@@ -365,8 +353,7 @@ bool FunctionGen::emitExternPrototype(PrototypeAST *Proto) {
                         Name.c_str(), Proto->getReturnType().c_str());
             return false;
         }
-        RetType = Info->LLVMType;
-        if (Attribute::AttrKind K = Extension(Info); K != Attribute::None) RetAttrs.addAttribute(K);
+        RetType = Info;
     }
 
     // Repeated identical EXTERN declarations, as in C, share one symbol; each
@@ -388,6 +375,7 @@ bool FunctionGen::emitExternPrototype(PrototypeAST *Proto) {
             reportError("Conflicting EXTERN declarations of %s", Name.c_str());
             return false;
         }
+        Sig.ExternABI = First.ExternABI;
         Sig.ExternType = First.ExternType;
         Sig.ExternAttrs = First.ExternAttrs;
         PrototypeLinks.emplace(Proto, Name);
@@ -401,10 +389,16 @@ bool FunctionGen::emitExternPrototype(PrototypeAST *Proto) {
         else reportError("EXTERN %s conflicts with FUNCTION/PROCEDURE %s", Name.c_str(), Name.c_str());
         return false;
     }
-    Sig.ExternType = FunctionType::get(RetType, ArgTypes, Proto->isVariadic());
-    Sig.ExternAttrs = AttributeList::get(Context, AttributeSet(),
-                                         AttributeSet::get(Context, RetAttrs), ParamAttrs);
-    Module.getOrInsertFunction(Name, Sig.ExternType);
+    std::string Err;
+    if (!ABI.arrange(RetType, ArgTypes, ArgTypes.size(), Sig.IsVariadic, Sig.ExternABI, Err)) {
+        reportError("Internal: cannot arrange EXTERN %s: %s", Name.c_str(), Err.c_str());
+        return false;
+    }
+    Sig.ExternType = getCABIFunctionType(Sig.ExternABI);
+    Sig.ExternAttrs = getCABIAttributes(Context, Sig.ExternABI);
+    FunctionCallee Declared = Module.getOrInsertFunction(Name, Sig.ExternType);
+    if (auto *F = dyn_cast<Function>(Declared.getCallee()))
+        F->setCallingConv(Sig.ExternABI.CallingConv);
     PrototypeLinks.emplace(Proto, Name);
     Overloads.push_back(std::move(Sig));
     return true;
@@ -485,10 +479,18 @@ llvm::Value *FunctionGen::emitCall(const FuncSig &Sig, const std::vector<llvm::V
     if (BeforeCall) BeforeCall(Sig);
     if (!Sig.IsExternal) return emitCall(Sig.LinkName, Args);
     FunctionCallee Callee(Sig.ExternType, Module.getNamedValue(Sig.LinkName));
+    // Variadic arguments arrive promoted and pass as they are.
+    size_t Fixed = Sig.ExternABI.Fixed;
+    Value *SRet = nullptr;
+    std::vector<Value *> IRArgs = lowerCABIArguments(
+        Builder, Sig.ExternABI, ArrayRef<Value *>(Args).take_front(Fixed), SRet);
+    IRArgs.insert(IRArgs.end(), Args.begin() + Fixed, Args.end());
     Value *Call = Sig.ExternType->getReturnType()->isVoidTy()
-        ? Exceptions.emitCall(Callee, Args) : Exceptions.emitCall(Callee, Args, "calltmp");
+        ? Exceptions.emitCall(Callee, IRArgs) : Exceptions.emitCall(Callee, IRArgs, "calltmp");
     cast<CallBase>(Call)->setAttributes(Sig.ExternAttrs);
-    return Call;
+    cast<CallBase>(Call)->setCallingConv(Sig.ExternABI.CallingConv);
+    Value *Result = lowerCABIResult(Builder, Sig.ExternABI, Call, SRet);
+    return Result ? Result : Call;
 }
 
 void FunctionGen::emitReturn(llvm::Value *RetVal) {