From f01cafc0ead695e8c5280e5e4a942b24e1922e13 Mon Sep 17 00:00:00 2001 From: _ov4 Date: Mon, 19 Oct 2026 13:21:24 +0000 Subject: refactor(driver): isolate MinGW JIT compatibility support Move COFF normalization, jump stubs and SEH registration behind two internal entry points. Keep runtime bindings in the embedded runner, remove redundant conditions and document the upstream requirements. Co-Authored-By: Claude Opus 5.5 --- tools/driver/EmbeddedRunner.cc | 335 +---------------------------------------- 1 file changed, 4 insertions(+), 331 deletions(-) (limited to 'tools/driver/EmbeddedRunner.cc') diff --git a/tools/driver/EmbeddedRunner.cc b/tools/driver/EmbeddedRunner.cc index 49375e1..587d664 100644 --- a/tools/driver/EmbeddedRunner.cc +++ b/tools/driver/EmbeddedRunner.cc @@ -1,11 +1,11 @@ #include "EmbeddedRunner.h" +#include "MinGWJIT.h" #include "NativeCompiler.h" #include "llvm/Object/ObjectFile.h" #include "cps/CodeGen.h" #include "llvm/ExecutionEngine/Orc/AbsoluteSymbols.h" #include "llvm/ExecutionEngine/Orc/ExecutionUtils.h" #include "llvm/ExecutionEngine/Orc/LLJIT.h" -#include "llvm/ExecutionEngine/Orc/ObjectFileInterface.h" #include "llvm/ExecutionEngine/Orc/ObjectLinkingLayer.h" #include "llvm/ExecutionEngine/Orc/ThreadSafeModule.h" #include "llvm/Support/Error.h" @@ -17,323 +17,12 @@ #include #include #include -#if defined(_WIN32) && defined(__x86_64__) -#ifndef NOMINMAX -#define NOMINMAX -#endif -#include -#include -#include -#include "llvm/ExecutionEngine/JITLink/JITLink.h" -#include "llvm/ExecutionEngine/JITLink/x86_64.h" -#include "llvm/ExecutionEngine/Orc/LinkGraphLinkingLayer.h" -#include "llvm/ExecutionEngine/Orc/Shared/WrapperFunctionUtils.h" -#include "llvm/Support/Endian.h" -#include "MinGWCOFF.h" -// Windows IMAGE_* macros conflict with LLVM COFF enums and inline helpers. -// Parse all LLVM COFF declarations and helpers before introducing those macros. -#include -#endif #ifdef __MINGW32__ #include #endif namespace { -#if defined(_WIN32) && defined(__x86_64__) -// In-process COFF/SEH support for the MinGW x86_64 JIT. -// -// The full upstream COFFPlatform starts the VC runtime loading machinery, -// which does not apply to the current MinGW build. This plugin provides only -// the linking and registration pieces needed for in-process execution: -// a per-graph __ImageBase shared by relocations and SEH registration, -// JITLink stubs for external direct branches and GNU personalities in .xdata, -// and RtlAddFunctionTable/RtlDeleteFunctionTable via JITLink allocation -// actions. Exception propagation still uses the existing C++ ABI, LLVM -// exception instructions and platform unwinder. - -static constexpr llvm::StringLiteral COFFSEHTableSection = "$__cpsc_seh"; - -static llvm::Error addCOFFJumpStubs(llvm::jitlink::LinkGraph &G) { - using namespace llvm; - using namespace llvm::jitlink; - using namespace llvm::jitlink::x86_64; - GOTTableManager GOT(G); - PLTTableManager PLT(G, GOT); - // Creating a stub may add sections/blocks. Visit only the original blocks. - std::vector Blocks(G.blocks().begin(), G.blocks().end()); - for (auto *B : Blocks) { - bool Executable = (B->getSection().getMemProt() & orc::MemProt::Exec) != - orc::MemProt::None; - for (auto &E : B->edges()) { - Symbol &Target = E.getTarget(); - if (!Target.isExternal()) - continue; - bool Personality = false; - if (isCOFFUnwindSection(B->getSection().getName(), ".xdata") && - Target.hasName()) { - StringRef Name = *Target.getName(); - Personality = Name == "__gxx_personality_seh0" || - Name == "__gxx_personality_sj0" || - Name == "__gxx_personality_v0"; - } - // COFF's private PCRel32 kind is lowered only in PreFixup. - // Use its name rather than confusing it with x86_64::PCRel32. - // Only E8/E9 direct branches with a plain symbol target need a - // jump stub; RIP-relative data and indirect calls must stay intact. - bool DirectBranch = false; - if (Executable && !B->isZeroFill() && E.getAddend() == 0 && - StringRef(G.getEdgeKindName(E.getKind())) == "PCRel32") { - auto Content = B->getContent(); - size_t Offset = E.getOffset(); - if (Offset > 0 && Offset <= Content.size() && - Content.size() - Offset >= 4) { - unsigned char Opcode = Content[Offset - 1]; - DirectBranch = Opcode == 0xe8 || Opcode == 0xe9; - } - } - if (Personality || DirectBranch) - E.setTarget(PLT.getEntryForTarget(G, Target)); - } - } - return Error::success(); -} - -static llvm::Error registerSEHTable(llvm::orc::ExecutorAddrRange TableRange, - llvm::orc::ExecutorAddr ImageBase) { - using llvm::make_error; - using llvm::StringError; - using llvm::inconvertibleErrorCode; - size_t Size = TableRange.size(); - if (Size == 0 || ImageBase.isNull()) - return make_error("invalid embedded SEH function table", - inconvertibleErrorCode()); - if (Size % sizeof(RUNTIME_FUNCTION)) - return make_error("invalid embedded SEH function table", - inconvertibleErrorCode()); - uint64_t Count64 = Size / sizeof(RUNTIME_FUNCTION); - if (Count64 == 0 || Count64 > static_cast(std::numeric_limits::max())) - return make_error("invalid embedded SEH function table", - inconvertibleErrorCode()); - auto *Table = TableRange.Start.toPtr(); - DWORD PrevBegin = 0; - for (uint64_t I = 0; I < Count64; ++I) { - DWORD Begin = llvm::support::endian::read32le( - reinterpret_cast(Table + I) + 0); - DWORD End = llvm::support::endian::read32le( - reinterpret_cast(Table + I) + 4); - if (!(Begin < End)) - return make_error("invalid embedded SEH function table", - inconvertibleErrorCode()); - if (I > 0 && !(PrevBegin < Begin)) - return make_error("invalid embedded SEH function table", - inconvertibleErrorCode()); - PrevBegin = Begin; - } - auto *MutableTable = const_cast(Table); - DWORD Count = static_cast(Count64); - DWORD64 Base = static_cast(ImageBase.getValue()); - if (!RtlAddFunctionTable(MutableTable, Count, Base)) - return make_error("cannot register embedded SEH function table", - inconvertibleErrorCode()); - return llvm::Error::success(); -} - -static llvm::Error deregisterSEHTable(llvm::orc::ExecutorAddrRange TableRange, - llvm::orc::ExecutorAddr ImageBase) { - (void)ImageBase; - auto *Table = TableRange.Start.toPtr(); - if (TableRange.size() == 0 || !Table) - return llvm::Error::success(); - RtlDeleteFunctionTable(Table); - return llvm::Error::success(); -} - -extern "C" llvm::orc::shared::CWrapperFunctionBuffer -cps_register_seh_table_alloc_action(const char *ArgData, size_t ArgSize) { - using namespace llvm::orc::shared; - return WrapperFunction::handle( - ArgData, ArgSize, registerSEHTable) - .release(); -} - -extern "C" llvm::orc::shared::CWrapperFunctionBuffer -cps_deregister_seh_table_alloc_action(const char *ArgData, size_t ArgSize) { - using namespace llvm::orc::shared; - return WrapperFunction::handle( - ArgData, ArgSize, deregisterSEHTable) - .release(); -} - -class MinGWSEHRegistrationPlugin final - : public llvm::orc::LinkGraphLinkingLayer::Plugin { -public: - void modifyPassConfig(llvm::orc::MaterializationResponsibility &MR, - llvm::jitlink::LinkGraph &G, - llvm::jitlink::PassConfiguration &Config) override { - using namespace llvm; - using namespace llvm::orc; - using namespace llvm::jitlink; - const Triple &TT = G.getTargetTriple(); - if (!TT.isOSBinFormatCOFF() || TT.getArch() != Triple::x86_64) - return; - - // SEH table liveness relies on the COFF backend's default - // SEHFrameKeepAlivePass(".pdata"), which keeps .pdata alive from its - // referenced function/unwind blocks. Normalized .pdata$/.xdata$ - // sections are kept alive by their associative COMDAT code edges. - // COFF unwind tables address the personality by 32-bit RVA, so it - // must live near JIT code. Retarget .xdata personality references - // to this graph's anonymous pointer jump stubs; the stub pointer - // itself uses a 64-bit relocation to reach the far host runtime. - // External direct calls/jumps from user objects also need nearby - // stubs: their PCRel32 displacement cannot reach a distant CRT DLL. - // The same target reuses the same stub, including personality uses. - Config.PostPrunePasses.push_back(addCOFFJumpStubs); - - // Tables sharing an image base have overlapping lookup ranges in the - // Windows unwinder. Reserve one table for all retained .pdata sections - // in this graph, then fill and sort its entries after relocations. - Config.PostPrunePasses.push_back([](LinkGraph &G) -> Error { - size_t Size = 0; - for (auto &Sec : G.sections()) { - if (!isCOFFUnwindSection(Sec.getName(), ".pdata")) - continue; - for (auto *B : Sec.blocks()) { - if (B->isZeroFill() || B->getSize() % sizeof(RUNTIME_FUNCTION)) - return make_error( - "invalid embedded SEH function table"); - Size += B->getSize(); - } - } - if (Size) { - auto &Sec = G.createSection(COFFSEHTableSection, MemProt::Read); - G.createMutableContentBlock(Sec, Size, ExecutorAddr(), - alignof(RUNTIME_FUNCTION), 0); - } - return Error::success(); - }); - - // COFF unwind tables use 32-bit image-relative addresses, so the - // image base must sit at or below the lowest JIT address. Alias the - // external __ImageBase to the lowest defined block once addresses - // are final; a trailing zero block would push RVAs negative. - Config.PostAllocationPasses.push_back([](LinkGraph &G) -> Error { - // A graph-local alias never collides at materialization time, - // no matter how many graphs define the same base name. - if (G.findDefinedSymbolByName(G.intern("__ImageBase"))) - return Error::success(); - auto *ImageBaseExt = - G.findExternalSymbolByName(G.intern("__ImageBase")); - if (!ImageBaseExt) - return Error::success(); - jitlink::Block *Low = nullptr; - for (auto &Sec : G.sections()) { - for (auto *B : Sec.blocks()) { - if (!Low || B->getAddress() < Low->getAddress()) - Low = B; - } - } - if (!Low) - return make_error( - "missing embedded __ImageBase content"); - G.makeDefined(*ImageBaseExt, *Low, 0, 8, Linkage::Strong, - Scope::Local, true); - return Error::success(); - }); - - // Copy fixed-up entries into one sorted, allocation-owned table. - // Register it once per graph, with deregistration on link failure or - // teardown before the table and its code are released. - Config.PostFixupPasses.push_back( - [](LinkGraph &G) -> Error { - using namespace llvm::orc::shared; - auto *TableSec = G.findSectionByName(COFFSEHTableSection); - if (!TableSec) - return Error::success(); - auto *Table = *TableSec->blocks().begin(); - std::vector Entries; - Entries.reserve(Table->getSize() / sizeof(RUNTIME_FUNCTION)); - for (auto &Sec : G.sections()) { - if (!isCOFFUnwindSection(Sec.getName(), ".pdata")) - continue; - for (auto *B : Sec.blocks()) { - size_t Start = Entries.size(); - Entries.resize(Start + B->getSize() / sizeof(RUNTIME_FUNCTION)); - if (B->getSize()) - std::memcpy(Entries.data() + Start, B->getContent().data(), - B->getSize()); - } - } - std::sort(Entries.begin(), Entries.end(), - [](const RUNTIME_FUNCTION &L, const RUNTIME_FUNCTION &R) { - return L.BeginAddress < R.BeginAddress; - }); - std::memcpy(Table->getAlreadyMutableContent().data(), Entries.data(), - Table->getSize()); - ExecutorAddrRange TableRange(Table->getAddress(), Table->getSize()); - auto *ImageBaseSym = - G.findDefinedSymbolByName(G.intern("__ImageBase")); - if (!ImageBaseSym) - return make_error( - "missing embedded __ImageBase for SEH registration"); - ExecutorAddr ImageBase = ImageBaseSym->getAddress(); - if (ImageBase.isNull()) - return make_error("invalid embedded SEH image base"); - auto RegisterAction = - WrapperFunctionCall::Create>( - ExecutorAddr::fromPtr(&cps_register_seh_table_alloc_action), - TableRange, ImageBase); - if (!RegisterAction) - return RegisterAction.takeError(); - auto DeregisterAction = - WrapperFunctionCall::Create>( - ExecutorAddr::fromPtr(&cps_deregister_seh_table_alloc_action), - TableRange, ImageBase); - if (!DeregisterAction) - return DeregisterAction.takeError(); - G.allocActions().push_back( - {std::move(*RegisterAction), std::move(*DeregisterAction)}); - return Error::success(); - }); - (void)MR; - } - - llvm::Error notifyFailed(llvm::orc::MaterializationResponsibility &) override { - return llvm::Error::success(); - } - - llvm::Error notifyRemovingResources(llvm::orc::JITDylib &, - llvm::orc::ResourceKey) override { - return llvm::Error::success(); - } - - void notifyTransferringResources(llvm::orc::JITDylib &, - llvm::orc::ResourceKey, - llvm::orc::ResourceKey) override {} -}; - -// Both user objects and lazily extracted runtime archive members pass through -// this layer, before JITLink builds their graphs or runs plugin passes. -class MinGWObjectLinkingLayer final : public llvm::orc::ObjectLinkingLayer { -public: - using ObjectLinkingLayer::ObjectLinkingLayer; - - void emit(std::unique_ptr R, - std::unique_ptr O) override { - auto Normalized = normalizeMinGWObject(std::move(O)); - if (!Normalized) { - getExecutionSession().reportError(Normalized.takeError()); - R->failMaterialization(); - return; - } - ObjectLinkingLayer::emit(std::move(R), std::move(*Normalized)); - } -}; -#endif - // Embedded JITLink object layer. Created before LLJIT takes ownership so // runEmbedded keeps only the execution flow. COFF flags stay conditional on // the object format; other formats use LLVM defaults. @@ -341,21 +30,15 @@ static std::unique_ptr createEmbeddedObjectLayer(llvm::orc::ExecutionSession &ES, llvm::jitlink::JITLinkMemoryManager &MM, const llvm::Triple &TT) { - std::unique_ptr Layer; #if defined(_WIN32) && defined(__x86_64__) if (TT.isOSBinFormatCOFF() && TT.getArch() == llvm::Triple::x86_64) - Layer = std::make_unique(ES, MM); - else + return createMinGWObjectLinkingLayer(ES, MM); #endif - Layer = std::make_unique(ES, MM); + auto Layer = std::make_unique(ES, MM); if (TT.isOSBinFormatCOFF()) { Layer->setOverrideObjectFlagsWithResponsibilityFlags(true); Layer->setAutoClaimResponsibilityForObjectSymbols(true); } -#if defined(_WIN32) && defined(__x86_64__) - if (TT.isOSBinFormatCOFF() && TT.getArch() == llvm::Triple::x86_64) - Layer->addPlugin(std::make_shared()); -#endif return Layer; } @@ -459,8 +142,6 @@ static llvm::Error defineRuntimeSymbols(llvm::orc::LLJIT &J) { // msvcrt DLL only exports __iob_func, so bind the host's shim directly. Symbols[Intern("__acrt_iob_func")] = ExecutorSymbolDef::fromPtr(&__acrt_iob_func); -#endif -#ifdef _WIN32 // The JIT-generated main references __main on MinGW; provide a no-op so // the host CRT is not initialized a second time. Symbols[Intern("__main")] = @@ -481,15 +162,7 @@ static llvm::Error loadCxxRuntime(llvm::orc::LLJIT &J) { for (auto &Archive : Archives) { llvm::orc::StaticLibraryDefinitionGenerator::GetObjectFileInterface GetInterface; #if defined(_WIN32) && defined(__x86_64__) - // Match the normalized object that MinGWObjectLinkingLayer::emit links. - GetInterface = [](llvm::orc::ExecutionSession &ES, llvm::MemoryBufferRef Ref) - -> llvm::Expected { - auto Normalized = normalizeMinGWObject( - llvm::MemoryBuffer::getMemBuffer(Ref, false)); - if (!Normalized) - return Normalized.takeError(); - return llvm::orc::getObjectFileInterface(ES, (*Normalized)->getMemBufferRef()); - }; + GetInterface = getMinGWObjectFileInterface; #endif auto Generator = llvm::orc::StaticLibraryDefinitionGenerator::Create( J.getObjLinkingLayer(), std::move(Archive), -- cgit v1.3.1-4-g156e