[clang] [llvm] [transforms] Inline simple variadic functions (PR #81058)
Matt Arsenault via cfe-commits
cfe-commits at lists.llvm.org
Wed Feb 7 22:52:29 PST 2024
================
@@ -0,0 +1,698 @@
+//===-- ExpandVariadicsPass.cpp --------------------------------*- C++ -*-=//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+//
+// This is an optimisation pass for variadic functions. If called from codegen,
+// it can serve as the implementation of variadic functions for a given target.
+//
+// The target-dependent parts are in namespace VariadicABIInfo. Enabling a new
+// target means adding a case to VariadicABIInfo::create() along with tests.
+//
+// The module pass using that information is class ExpandVariadics.
+//
+// The strategy is:
+// 1. Test whether a variadic function is sufficiently simple
+// 2. If it was, calls to it can be replaced with calls to a different function
+// 3. If it wasn't, try to split it into a simple function and a remainder
+// 4. Optionally rewrite the varadic function calling convention as well
+//
+// This pass considers "sufficiently simple" to mean a variadic function that
+// calls into a different function taking a va_list to do the real work. For
+// example, libc might implement fprintf as a single basic block calling into
+// vfprintf. This pass can then rewrite call to the variadic into some code
+// to construct a target-specific value to use for the va_list and a call
+// into the non-variadic implementation function. There's a test for that.
+//
+// Most other variadic functions whose definition is known can be converted into
+// that form. Create a new internal function taking a va_list where the original
+// took a ... parameter. Move the blocks across. Create a new block containing a
+// va_start that calls into the new function. This is nearly target independent.
+//
+// Where this transform is consistent with the ABI, e.g. AMDGPU or NVPTX, or
+// where the ABI can be chosen to align with this transform, the function
+// interface can be rewritten along with calls to unknown variadic functions.
+//
+// The aggregate effect is to unblock other transforms, most critically the
+// general purpose inliner. Known calls to variadic functions become zero cost.
+//
+// This pass does define some target specific information which is partially
+// redundant with other parts of the compiler. In particular, the call frame
+// it builds must be the exact complement of the va_arg lowering performed
+// by clang. The va_list construction is similar to work done by the backend
+// for targets that lower variadics there, though distinct in that this pass
+// constructs the pieces using alloca instead of relative to stack pointers.
+//
+// Consistency with clang is primarily tested by emitting va_arg using clang
+// then expanding the variadic functions using this pass, followed by trying
+// to constant fold the functions to no-ops.
+//
+// Target specific behaviour is tested in IR - mainly checking that values are
+// put into positions in call frames that make sense for that particular target.
+//
+//===----------------------------------------------------------------------===//
+
+#include "llvm/Transforms/IPO/ExpandVariadics.h"
+#include "llvm/ADT/SmallVector.h"
+#include "llvm/CodeGen/Passes.h"
+#include "llvm/IR/Constants.h"
+#include "llvm/IR/IRBuilder.h"
+#include "llvm/IR/IntrinsicInst.h"
+#include "llvm/IR/Module.h"
+#include "llvm/IR/PassManager.h"
+#include "llvm/InitializePasses.h"
+#include "llvm/Pass.h"
+#include "llvm/TargetParser/Triple.h"
+
+#define DEBUG_TYPE "expand-variadics"
+
+using namespace llvm;
+
+namespace {
+namespace VariadicABIInfo {
+
+// calling convention for passing as valist object, same as it would be in C
+// aarch64 uses byval
+enum class valistCC { value, pointer, /*byval*/ };
+
+struct Interface {
+protected:
+ Interface(uint32_t MinAlign, uint32_t MaxAlign)
+ : MinAlign(MinAlign), MaxAlign(MaxAlign) {}
+
+public:
+ virtual ~Interface() {}
+ const uint32_t MinAlign;
+ const uint32_t MaxAlign;
+
+ // Most ABIs use a void* or char* for va_list, others can specialise
+ virtual Type *vaListType(LLVMContext &Ctx) {
+ return PointerType::getUnqual(Ctx);
+ }
+
+ // How the vaListType is passed
+ virtual valistCC vaListCC() { return valistCC::value; }
+
+ // The valist might need to be stack allocated.
+ virtual bool valistOnStack() { return false; }
+
+ virtual void initializeVAList(LLVMContext &Ctx, IRBuilder<> &Builder,
+ AllocaInst * /*va_list*/, Value * /*buffer*/) {
+ // Function needs to be implemented if valist is on the stack
+ assert(!valistOnStack());
+ __builtin_unreachable();
+ }
+
+ // All targets currently implemented use a ptr for the valist parameter
+ Type *vaListParameterType(LLVMContext &Ctx) {
+ return PointerType::getUnqual(Ctx);
+ }
+
+ bool VAEndIsNop() { return true; }
+
+ bool VACopyIsMemcpy() { return true; }
+};
+
+struct X64SystemV final : public Interface {
+ // X64 documented behaviour:
+ // Slots are at least eight byte aligned and at most 16 byte aligned.
+ // If the type needs more than sixteen byte alignment, it still only gets
+ // that much alignment on the stack.
+ // X64 behaviour in clang:
+ // Slots are at least eight byte aligned and at most naturally aligned
+ // This matches clang, not the ABI docs.
+ X64SystemV() : Interface(8, 0) {}
+
+ Type *vaListType(LLVMContext &Ctx) override {
+ auto I32 = Type::getInt32Ty(Ctx);
+ auto Ptr = PointerType::getUnqual(Ctx);
+ return ArrayType::get(StructType::get(Ctx, {I32, I32, Ptr, Ptr}), 1);
+ }
+ valistCC vaListCC() override { return valistCC::pointer; }
+
+ bool valistOnStack() override { return true; }
+
+ void initializeVAList(LLVMContext &Ctx, IRBuilder<> &Builder,
+ AllocaInst *va_list, Value *voidBuffer) override {
+ assert(valistOnStack());
+ assert(va_list != nullptr);
+ assert(va_list->getAllocatedType() == vaListType(Ctx));
+
+ Type *va_list_ty = vaListType(Ctx);
+
+ Type *I32 = Type::getInt32Ty(Ctx);
+ Type *I64 = Type::getInt64Ty(Ctx);
+
+ Value *Idxs[3] = {
+ ConstantInt::get(I64, 0),
+ ConstantInt::get(I32, 0),
+ nullptr,
+ };
+
+ Idxs[2] = ConstantInt::get(I32, 0);
+ Builder.CreateStore(
+ ConstantInt::get(I32, 48),
+ Builder.CreateInBoundsGEP(va_list_ty, va_list, Idxs, "gp_offset"));
+
+ Idxs[2] = ConstantInt::get(I32, 1);
+ Builder.CreateStore(
+ ConstantInt::get(I32, 6 * 8 + 8 * 16),
+ Builder.CreateInBoundsGEP(va_list_ty, va_list, Idxs, "fp_offset"));
+
+ Idxs[2] = ConstantInt::get(I32, 2);
+ Builder.CreateStore(voidBuffer,
+ Builder.CreateInBoundsGEP(va_list_ty, va_list, Idxs,
+ "overfow_arg_area"));
+
+ Idxs[2] = ConstantInt::get(I32, 3);
+ Builder.CreateStore(
+ ConstantPointerNull::get(PointerType::getUnqual(Ctx)),
+ Builder.CreateInBoundsGEP(va_list_ty, va_list, Idxs, "reg_save_area"));
+ }
+};
+
+std::unique_ptr<Interface> create(Module &M) {
+ llvm::Triple Triple(M.getTargetTriple());
+ const bool IsLinuxABI = Triple.isOSLinux() || Triple.isOSCygMing();
+
+ switch (Triple.getArch()) {
+
+ case Triple::x86: {
+ // These seem to all fall out the same, despite getTypeStackAlign
+ // implying otherwise.
+ if (Triple.isOSDarwin()) {
+ struct X86Darwin final : public Interface {
+ // X86_32ABIInfo::getTypeStackAlignInBytes is misleading for this.
+ // The slotSize(4) implies a minimum alignment
+ // The AllowHigherAlign = true means there is no maximum alignment.
+ X86Darwin() : Interface(4, 0) {}
+ };
+
+ return std::make_unique<X86Darwin>();
+ }
+ if (Triple.getOS() == llvm::Triple::Win32) {
+ struct X86Windows final : public Interface {
+ X86Windows() : Interface(4, 0) {}
+ };
+ return std::make_unique<X86Windows>();
+ }
+
+ if (IsLinuxABI) {
+ struct X86Linux final : public Interface {
+ X86Linux() : Interface(4, 0) {}
+ };
+ return std::make_unique<X86Linux>();
+ }
+ break;
+ }
+
+ case Triple::x86_64: {
+ if (Triple.isWindowsMSVCEnvironment() || Triple.isOSWindows()) {
+ struct X64Windows final : public Interface {
+ X64Windows() : Interface(8, 8) {}
+ };
+ // x64 msvc emit vaarg passes > 8 byte values by pointer
+ // however the variadic call instruction created does not, e.g.
+ // a <4 x f32> will be passed as itself, not as a pointer or byval.
+ // Postponing resolution of that for now.
+ return nullptr;
+ }
+
+ if (Triple.isOSDarwin()) {
+ return std::make_unique<VariadicABIInfo::X64SystemV>();
+ }
+
+ if (IsLinuxABI) {
+ return std::make_unique<VariadicABIInfo::X64SystemV>();
+ }
+
+ break;
+ }
+
+ default:
+ return nullptr;
+ }
+
+ return nullptr;
+}
+
+} // namespace VariadicABIInfo
+
+class ExpandVariadics : public ModulePass {
+public:
+ static char ID;
+ std::unique_ptr<VariadicABIInfo::Interface> ABI;
+
+ ExpandVariadics() : ModulePass(ID) {}
+ StringRef getPassName() const override { return "Expand variadic functions"; }
+
+ // A predicate in that return nullptr means false
+ // Returns the function target to use when inlining on success
+ Function *isFunctionInlinable(Module &M, Function *F);
+
+ // Rewrite a call site.
+ void ExpandCall(Module &M, CallInst *CB, Function *VarargF, Function *NF);
+
+ // this could be partially target specific
+ bool expansionApplicableToFunction(Module &M, Function *F) {
+ if (F->isIntrinsic() || !F->isVarArg() ||
+ F->hasFnAttribute(Attribute::Naked))
+ return false;
+
+ if (F->getCallingConv() != CallingConv::C)
+ return false;
+
+ if (GlobalValue::isInterposableLinkage(F->getLinkage()))
+ return false;
----------------
arsenm wrote:
Simpler as F->isDefinitionExact?
Don't we want to do this unconditionally for AMDGPU?
https://github.com/llvm/llvm-project/pull/81058
More information about the cfe-commits
mailing list