rust/compiler/rustc_llvm/llvm-wrapper/PassWrapper.cpp

1679 lines
56 KiB
C++
Raw Normal View History

2013-08-22 22:58:42 -05:00
#include <stdio.h>
#include <vector>
#include <set>
2020-07-26 12:11:30 -05:00
#include "LLVMWrapper.h"
#include "llvm/Analysis/TargetLibraryInfo.h"
#include "llvm/Analysis/TargetTransformInfo.h"
#include "llvm/CodeGen/TargetSubtargetInfo.h"
2019-12-31 07:02:22 -06:00
#include "llvm/InitializePasses.h"
#include "llvm/IR/AutoUpgrade.h"
#include "llvm/IR/AssemblyAnnotationWriter.h"
#include "llvm/IR/IntrinsicInst.h"
#include "llvm/IR/Verifier.h"
Store LLVM bitcode in object files, not compressed This commit is an attempted resurrection of #70458 where LLVM bitcode emitted by rustc into rlibs is stored into object file sections rather than in a separate file. The main rationale for doing this is that when rustc emits bitcode it will no longer use a custom compression scheme which makes it both easier to interoperate with existing tools and also cuts down on compile time since this compression isn't happening. The blocker for this in #70458 turned out to be that native linkers didn't handle the new sections well, causing the sections to either trigger bugs in the linker or actually end up in the final linked artifact. This commit attempts to address these issues by ensuring that native linkers ignore the new sections by inserting custom flags with module-level inline assembly. Note that this does not currently change the API of the compiler at all. The pre-existing `-C bitcode-in-rlib` flag is co-opted to indicate whether the bitcode should be present in the object file or not. Finally, note that an important consequence of this commit, which is also one of its primary purposes, is to enable rustc's `-Clto` bitcode loading to load rlibs produced with `-Clinker-plugin-lto`. The goal here is that when you're building with LTO Cargo will tell rustc to skip codegen of all intermediate crates and only generate LLVM IR. Today rustc will generate both object code and LLVM IR, but the object code is later simply thrown away, wastefully.
2020-04-23 13:45:55 -05:00
#include "llvm/Object/ObjectFile.h"
#include "llvm/Object/IRObjectFile.h"
#include "llvm/Passes/PassBuilder.h"
#include "llvm/Passes/StandardInstrumentations.h"
#include "llvm/Support/CBindingWrapping.h"
#include "llvm/Support/FileSystem.h"
#include "llvm/Support/Host.h"
#include "llvm/Target/TargetMachine.h"
2013-08-22 22:58:42 -05:00
#include "llvm/Transforms/IPO/PassManagerBuilder.h"
#include "llvm/Transforms/IPO/AlwaysInliner.h"
rustc: Implement ThinLTO This commit is an implementation of LLVM's ThinLTO for consumption in rustc itself. Currently today LTO works by merging all relevant LLVM modules into one and then running optimization passes. "Thin" LTO operates differently by having more sharded work and allowing parallelism opportunities between optimizing codegen units. Further down the road Thin LTO also allows *incremental* LTO which should enable even faster release builds without compromising on the performance we have today. This commit uses a `-Z thinlto` flag to gate whether ThinLTO is enabled. It then also implements two forms of ThinLTO: * In one mode we'll *only* perform ThinLTO over the codegen units produced in a single compilation. That is, we won't load upstream rlibs, but we'll instead just perform ThinLTO amongst all codegen units produced by the compiler for the local crate. This is intended to emulate a desired end point where we have codegen units turned on by default for all crates and ThinLTO allows us to do this without performance loss. * In anther mode, like full LTO today, we'll optimize all upstream dependencies in "thin" mode. Unlike today, however, this LTO step is fully parallelized so should finish much more quickly. There's a good bit of comments about what the implementation is doing and where it came from, but the tl;dr; is that currently most of the support here is copied from upstream LLVM. This code duplication is done for a number of reasons: * Controlling parallelism means we can use the existing jobserver support to avoid overloading machines. * We will likely want a slightly different form of incremental caching which integrates with our own incremental strategy, but this is yet to be determined. * This buys us some flexibility about when/where we run ThinLTO, as well as having it tailored to fit our needs for the time being. * Finally this allows us to reuse some artifacts such as our `TargetMachine` creation, where all our options we used today aren't necessarily supported by upstream LLVM yet. My hope is that we can get some experience with this copy/paste in tree and then eventually upstream some work to LLVM itself to avoid the duplication while still ensuring our needs are met. Otherwise I fear that maintaining these bindings may be quite costly over the years with LLVM updates!
2017-07-23 10:14:38 -05:00
#include "llvm/Transforms/IPO/FunctionImport.h"
#include "llvm/Transforms/Utils/FunctionImportUtils.h"
#include "llvm/LTO/LTO.h"
2013-08-22 22:58:42 -05:00
#include "llvm-c/Transforms/PassManagerBuilder.h"
#include "llvm/Transforms/Instrumentation.h"
#include "llvm/Transforms/Instrumentation/AddressSanitizer.h"
#include "llvm/Support/TimeProfiler.h"
#include "llvm/Transforms/Instrumentation/ThreadSanitizer.h"
#include "llvm/Transforms/Instrumentation/MemorySanitizer.h"
#include "llvm/Transforms/Utils/CanonicalizeAliases.h"
#include "llvm/Transforms/Utils/NameAnonGlobals.h"
2013-08-22 22:58:42 -05:00
using namespace llvm;
2013-08-22 22:58:42 -05:00
typedef struct LLVMOpaquePass *LLVMPassRef;
typedef struct LLVMOpaqueTargetMachine *LLVMTargetMachineRef;
DEFINE_STDCXX_CONVERSION_FUNCTIONS(Pass, LLVMPassRef)
DEFINE_STDCXX_CONVERSION_FUNCTIONS(TargetMachine, LLVMTargetMachineRef)
2020-06-25 20:52:41 -05:00
#if LLVM_VERSION_LT(11, 0)
DEFINE_STDCXX_CONVERSION_FUNCTIONS(PassManagerBuilder,
LLVMPassManagerBuilderRef)
2020-06-25 20:52:41 -05:00
#endif
extern "C" void LLVMInitializePasses() {
PassRegistry &Registry = *PassRegistry::getPassRegistry();
initializeCore(Registry);
initializeCodeGen(Registry);
initializeScalarOpts(Registry);
initializeVectorization(Registry);
initializeIPO(Registry);
initializeAnalysis(Registry);
initializeTransformUtils(Registry);
initializeInstCombine(Registry);
initializeInstrumentation(Registry);
initializeTarget(Registry);
}
extern "C" void LLVMTimeTraceProfilerInitialize() {
#if LLVM_VERSION_GE(10, 0)
timeTraceProfilerInitialize(
/* TimeTraceGranularity */ 0,
/* ProcName */ "rustc");
#else
timeTraceProfilerInitialize();
#endif
}
extern "C" void LLVMTimeTraceProfilerFinish(const char* FileName) {
StringRef FN(FileName);
std::error_code EC;
raw_fd_ostream OS(FN, EC, sys::fs::CD_CreateAlways);
timeTraceProfilerWrite(OS);
timeTraceProfilerCleanup();
}
enum class LLVMRustPassKind {
Other,
Function,
Module,
};
static LLVMRustPassKind toRust(PassKind Kind) {
switch (Kind) {
case PT_Function:
return LLVMRustPassKind::Function;
case PT_Module:
return LLVMRustPassKind::Module;
default:
return LLVMRustPassKind::Other;
}
}
extern "C" LLVMPassRef LLVMRustFindAndCreatePass(const char *PassName) {
StringRef SR(PassName);
PassRegistry *PR = PassRegistry::getPassRegistry();
const PassInfo *PI = PR->getPassInfo(SR);
if (PI) {
return wrap(PI->createPass());
}
return nullptr;
}
extern "C" LLVMPassRef LLVMRustCreateAddressSanitizerFunctionPass(bool Recover) {
const bool CompileKernel = false;
const bool UseAfterScope = true;
return wrap(createAddressSanitizerFunctionPass(CompileKernel, Recover, UseAfterScope));
}
extern "C" LLVMPassRef LLVMRustCreateModuleAddressSanitizerPass(bool Recover) {
const bool CompileKernel = false;
return wrap(createModuleAddressSanitizerLegacyPassPass(CompileKernel, Recover));
}
extern "C" LLVMPassRef LLVMRustCreateMemorySanitizerPass(int TrackOrigins, bool Recover) {
const bool CompileKernel = false;
return wrap(createMemorySanitizerLegacyPassPass(
MemorySanitizerOptions{TrackOrigins, Recover, CompileKernel}));
}
extern "C" LLVMPassRef LLVMRustCreateThreadSanitizerPass() {
return wrap(createThreadSanitizerLegacyPassPass());
}
extern "C" LLVMRustPassKind LLVMRustPassKind(LLVMPassRef RustPass) {
assert(RustPass);
Pass *Pass = unwrap(RustPass);
return toRust(Pass->getPassKind());
}
extern "C" void LLVMRustAddPass(LLVMPassManagerRef PMR, LLVMPassRef RustPass) {
assert(RustPass);
Pass *Pass = unwrap(RustPass);
PassManagerBase *PMB = unwrap(PMR);
PMB->add(Pass);
2013-08-22 22:58:42 -05:00
}
rustc: Implement ThinLTO This commit is an implementation of LLVM's ThinLTO for consumption in rustc itself. Currently today LTO works by merging all relevant LLVM modules into one and then running optimization passes. "Thin" LTO operates differently by having more sharded work and allowing parallelism opportunities between optimizing codegen units. Further down the road Thin LTO also allows *incremental* LTO which should enable even faster release builds without compromising on the performance we have today. This commit uses a `-Z thinlto` flag to gate whether ThinLTO is enabled. It then also implements two forms of ThinLTO: * In one mode we'll *only* perform ThinLTO over the codegen units produced in a single compilation. That is, we won't load upstream rlibs, but we'll instead just perform ThinLTO amongst all codegen units produced by the compiler for the local crate. This is intended to emulate a desired end point where we have codegen units turned on by default for all crates and ThinLTO allows us to do this without performance loss. * In anther mode, like full LTO today, we'll optimize all upstream dependencies in "thin" mode. Unlike today, however, this LTO step is fully parallelized so should finish much more quickly. There's a good bit of comments about what the implementation is doing and where it came from, but the tl;dr; is that currently most of the support here is copied from upstream LLVM. This code duplication is done for a number of reasons: * Controlling parallelism means we can use the existing jobserver support to avoid overloading machines. * We will likely want a slightly different form of incremental caching which integrates with our own incremental strategy, but this is yet to be determined. * This buys us some flexibility about when/where we run ThinLTO, as well as having it tailored to fit our needs for the time being. * Finally this allows us to reuse some artifacts such as our `TargetMachine` creation, where all our options we used today aren't necessarily supported by upstream LLVM yet. My hope is that we can get some experience with this copy/paste in tree and then eventually upstream some work to LLVM itself to avoid the duplication while still ensuring our needs are met. Otherwise I fear that maintaining these bindings may be quite costly over the years with LLVM updates!
2017-07-23 10:14:38 -05:00
extern "C"
void LLVMRustPassManagerBuilderPopulateThinLTOPassManager(
rustc: Implement ThinLTO This commit is an implementation of LLVM's ThinLTO for consumption in rustc itself. Currently today LTO works by merging all relevant LLVM modules into one and then running optimization passes. "Thin" LTO operates differently by having more sharded work and allowing parallelism opportunities between optimizing codegen units. Further down the road Thin LTO also allows *incremental* LTO which should enable even faster release builds without compromising on the performance we have today. This commit uses a `-Z thinlto` flag to gate whether ThinLTO is enabled. It then also implements two forms of ThinLTO: * In one mode we'll *only* perform ThinLTO over the codegen units produced in a single compilation. That is, we won't load upstream rlibs, but we'll instead just perform ThinLTO amongst all codegen units produced by the compiler for the local crate. This is intended to emulate a desired end point where we have codegen units turned on by default for all crates and ThinLTO allows us to do this without performance loss. * In anther mode, like full LTO today, we'll optimize all upstream dependencies in "thin" mode. Unlike today, however, this LTO step is fully parallelized so should finish much more quickly. There's a good bit of comments about what the implementation is doing and where it came from, but the tl;dr; is that currently most of the support here is copied from upstream LLVM. This code duplication is done for a number of reasons: * Controlling parallelism means we can use the existing jobserver support to avoid overloading machines. * We will likely want a slightly different form of incremental caching which integrates with our own incremental strategy, but this is yet to be determined. * This buys us some flexibility about when/where we run ThinLTO, as well as having it tailored to fit our needs for the time being. * Finally this allows us to reuse some artifacts such as our `TargetMachine` creation, where all our options we used today aren't necessarily supported by upstream LLVM yet. My hope is that we can get some experience with this copy/paste in tree and then eventually upstream some work to LLVM itself to avoid the duplication while still ensuring our needs are met. Otherwise I fear that maintaining these bindings may be quite costly over the years with LLVM updates!
2017-07-23 10:14:38 -05:00
LLVMPassManagerBuilderRef PMBR,
LLVMPassManagerRef PMR
) {
unwrap(PMBR)->populateThinLTOPassManager(*unwrap(PMR));
}
extern "C"
void LLVMRustAddLastExtensionPasses(
LLVMPassManagerBuilderRef PMBR, LLVMPassRef *Passes, size_t NumPasses) {
auto AddExtensionPasses = [Passes, NumPasses](
const PassManagerBuilder &Builder, PassManagerBase &PM) {
for (size_t I = 0; I < NumPasses; I++) {
PM.add(unwrap(Passes[I]));
}
};
// Add the passes to both of the pre-finalization extension points,
// so they are run for optimized and non-optimized builds.
unwrap(PMBR)->addExtension(PassManagerBuilder::EP_OptimizerLast,
AddExtensionPasses);
unwrap(PMBR)->addExtension(PassManagerBuilder::EP_EnabledOnOptLevel0,
AddExtensionPasses);
}
#ifdef LLVM_COMPONENT_X86
#define SUBTARGET_X86 SUBTARGET(X86)
#else
#define SUBTARGET_X86
#endif
#ifdef LLVM_COMPONENT_ARM
#define SUBTARGET_ARM SUBTARGET(ARM)
#else
#define SUBTARGET_ARM
#endif
#ifdef LLVM_COMPONENT_AARCH64
#define SUBTARGET_AARCH64 SUBTARGET(AArch64)
#else
#define SUBTARGET_AARCH64
#endif
2016-05-06 08:32:10 -05:00
#ifdef LLVM_COMPONENT_AVR
#define SUBTARGET_AVR SUBTARGET(AVR)
#else
#define SUBTARGET_AVR
#endif
#ifdef LLVM_COMPONENT_MIPS
#define SUBTARGET_MIPS SUBTARGET(Mips)
#else
#define SUBTARGET_MIPS
#endif
#ifdef LLVM_COMPONENT_POWERPC
#define SUBTARGET_PPC SUBTARGET(PPC)
#else
#define SUBTARGET_PPC
#endif
#ifdef LLVM_COMPONENT_SYSTEMZ
#define SUBTARGET_SYSTEMZ SUBTARGET(SystemZ)
#else
#define SUBTARGET_SYSTEMZ
#endif
#ifdef LLVM_COMPONENT_MSP430
#define SUBTARGET_MSP430 SUBTARGET(MSP430)
#else
#define SUBTARGET_MSP430
#endif
2018-07-24 05:03:28 -05:00
#ifdef LLVM_COMPONENT_RISCV
#define SUBTARGET_RISCV SUBTARGET(RISCV)
#else
#define SUBTARGET_RISCV
#endif
2016-12-03 10:53:31 -06:00
#ifdef LLVM_COMPONENT_SPARC
#define SUBTARGET_SPARC SUBTARGET(Sparc)
#else
#define SUBTARGET_SPARC
#endif
#ifdef LLVM_COMPONENT_HEXAGON
#define SUBTARGET_HEXAGON SUBTARGET(Hexagon)
#else
#define SUBTARGET_HEXAGON
#endif
#define GEN_SUBTARGETS \
SUBTARGET_X86 \
SUBTARGET_ARM \
SUBTARGET_AARCH64 \
2016-05-06 08:32:10 -05:00
SUBTARGET_AVR \
SUBTARGET_MIPS \
SUBTARGET_PPC \
SUBTARGET_SYSTEMZ \
2016-12-31 21:40:10 -06:00
SUBTARGET_MSP430 \
SUBTARGET_SPARC \
2018-07-24 05:03:28 -05:00
SUBTARGET_HEXAGON \
SUBTARGET_RISCV \
#define SUBTARGET(x) \
namespace llvm { \
extern const SubtargetFeatureKV x##FeatureKV[]; \
extern const SubtargetFeatureKV x##SubTypeKV[]; \
}
GEN_SUBTARGETS
#undef SUBTARGET
extern "C" bool LLVMRustHasFeature(LLVMTargetMachineRef TM,
const char *Feature) {
TargetMachine *Target = unwrap(TM);
const MCSubtargetInfo *MCInfo = Target->getMCSubtargetInfo();
return MCInfo->checkFeatures(std::string("+") + Feature);
}
enum class LLVMRustCodeModel {
Tiny,
Small,
Kernel,
Medium,
Large,
None,
};
static Optional<CodeModel::Model> fromRust(LLVMRustCodeModel Model) {
switch (Model) {
case LLVMRustCodeModel::Tiny:
return CodeModel::Tiny;
case LLVMRustCodeModel::Small:
return CodeModel::Small;
case LLVMRustCodeModel::Kernel:
return CodeModel::Kernel;
case LLVMRustCodeModel::Medium:
return CodeModel::Medium;
case LLVMRustCodeModel::Large:
return CodeModel::Large;
case LLVMRustCodeModel::None:
return None;
default:
report_fatal_error("Bad CodeModel.");
}
}
enum class LLVMRustCodeGenOptLevel {
None,
Less,
Default,
Aggressive,
};
static CodeGenOpt::Level fromRust(LLVMRustCodeGenOptLevel Level) {
switch (Level) {
case LLVMRustCodeGenOptLevel::None:
return CodeGenOpt::None;
case LLVMRustCodeGenOptLevel::Less:
return CodeGenOpt::Less;
case LLVMRustCodeGenOptLevel::Default:
return CodeGenOpt::Default;
case LLVMRustCodeGenOptLevel::Aggressive:
return CodeGenOpt::Aggressive;
default:
report_fatal_error("Bad CodeGenOptLevel.");
}
}
enum class LLVMRustPassBuilderOptLevel {
O0,
O1,
O2,
O3,
Os,
Oz,
};
static PassBuilder::OptimizationLevel fromRust(LLVMRustPassBuilderOptLevel Level) {
switch (Level) {
case LLVMRustPassBuilderOptLevel::O0:
2020-06-25 20:52:41 -05:00
return PassBuilder::OptimizationLevel::O0;
case LLVMRustPassBuilderOptLevel::O1:
2020-06-25 20:52:41 -05:00
return PassBuilder::OptimizationLevel::O1;
case LLVMRustPassBuilderOptLevel::O2:
2020-06-25 20:52:41 -05:00
return PassBuilder::OptimizationLevel::O2;
case LLVMRustPassBuilderOptLevel::O3:
2020-06-25 20:52:41 -05:00
return PassBuilder::OptimizationLevel::O3;
case LLVMRustPassBuilderOptLevel::Os:
2020-06-25 20:52:41 -05:00
return PassBuilder::OptimizationLevel::Os;
case LLVMRustPassBuilderOptLevel::Oz:
2020-06-25 20:52:41 -05:00
return PassBuilder::OptimizationLevel::Oz;
default:
report_fatal_error("Bad PassBuilderOptLevel.");
}
}
enum class LLVMRustRelocModel {
Static,
PIC,
DynamicNoPic,
ROPI,
RWPI,
ROPIRWPI,
};
static Reloc::Model fromRust(LLVMRustRelocModel RustReloc) {
switch (RustReloc) {
case LLVMRustRelocModel::Static:
return Reloc::Static;
case LLVMRustRelocModel::PIC:
return Reloc::PIC_;
case LLVMRustRelocModel::DynamicNoPic:
return Reloc::DynamicNoPIC;
case LLVMRustRelocModel::ROPI:
return Reloc::ROPI;
case LLVMRustRelocModel::RWPI:
return Reloc::RWPI;
case LLVMRustRelocModel::ROPIRWPI:
return Reloc::ROPI_RWPI;
}
report_fatal_error("Bad RelocModel.");
}
#ifdef LLVM_RUSTLLVM
/// getLongestEntryLength - Return the length of the longest entry in the table.
template<typename KV>
static size_t getLongestEntryLength(ArrayRef<KV> Table) {
size_t MaxLen = 0;
for (auto &I : Table)
MaxLen = std::max(MaxLen, std::strlen(I.Key));
return MaxLen;
}
extern "C" void LLVMRustPrintTargetCPUs(LLVMTargetMachineRef TM) {
const TargetMachine *Target = unwrap(TM);
const MCSubtargetInfo *MCInfo = Target->getMCSubtargetInfo();
2017-09-07 20:07:16 -05:00
const Triple::ArchType HostArch = Triple(sys::getProcessTriple()).getArch();
const Triple::ArchType TargetArch = Target->getTargetTriple().getArch();
const ArrayRef<SubtargetSubTypeKV> CPUTable = MCInfo->getCPUTable();
unsigned MaxCPULen = getLongestEntryLength(CPUTable);
printf("Available CPUs for this target:\n");
2017-09-07 20:07:16 -05:00
if (HostArch == TargetArch) {
const StringRef HostCPU = sys::getHostCPUName();
printf(" %-*s - Select the CPU of the current host (currently %.*s).\n",
MaxCPULen, "native", (int)HostCPU.size(), HostCPU.data());
}
for (auto &CPU : CPUTable)
printf(" %-*s\n", MaxCPULen, CPU.Key);
printf("\n");
}
extern "C" void LLVMRustPrintTargetFeatures(LLVMTargetMachineRef TM) {
const TargetMachine *Target = unwrap(TM);
const MCSubtargetInfo *MCInfo = Target->getMCSubtargetInfo();
const ArrayRef<SubtargetFeatureKV> FeatTable = MCInfo->getFeatureTable();
unsigned MaxFeatLen = getLongestEntryLength(FeatTable);
printf("Available features for this target:\n");
for (auto &Feature : FeatTable)
printf(" %-*s - %s.\n", MaxFeatLen, Feature.Key, Feature.Desc);
printf("\nRust-specific features:\n");
2020-06-16 13:53:30 -05:00
printf(" %-*s - %s.\n",
MaxFeatLen,
"crt-static",
2020-06-16 13:14:32 -05:00
"Enables libraries with C Run-time Libraries(CRT) to be statically linked"
);
printf("\n");
printf("Use +feature to enable a feature, or -feature to disable it.\n"
"For example, rustc -C -target-cpu=mycpu -C "
"target-feature=+feature1,-feature2\n\n");
}
#else
extern "C" void LLVMRustPrintTargetCPUs(LLVMTargetMachineRef) {
printf("Target CPU help is not supported by this LLVM version.\n\n");
}
extern "C" void LLVMRustPrintTargetFeatures(LLVMTargetMachineRef) {
printf("Target features help is not supported by this LLVM version.\n\n");
}
#endif
extern "C" const char* LLVMRustGetHostCPUName(size_t *len) {
StringRef Name = sys::getHostCPUName();
*len = Name.size();
return Name.data();
}
extern "C" LLVMTargetMachineRef LLVMRustCreateTargetMachine(
const char *TripleStr, const char *CPU, const char *Feature,
const char *ABIStr, LLVMRustCodeModel RustCM, LLVMRustRelocModel RustReloc,
LLVMRustCodeGenOptLevel RustOptLevel, bool UseSoftFloat,
bool FunctionSections,
std: Add a new wasm32-unknown-unknown target This commit adds a new target to the compiler: wasm32-unknown-unknown. This target is a reimagining of what it looks like to generate WebAssembly code from Rust. Instead of using Emscripten which can bring with it a weighty runtime this instead is a target which uses only the LLVM backend for WebAssembly and a "custom linker" for now which will hopefully one day be direct calls to lld. Notable features of this target include: * There is zero runtime footprint. The target assumes nothing exists other than the wasm32 instruction set. * There is zero toolchain footprint beyond adding the target. No custom linker is needed, rustc contains everything. * Very small wasm modules can be generated directly from Rust code using this target. * Most of the standard library is stubbed out to return an error, but anything related to allocation works (aka `HashMap`, `Vec`, etc). * Naturally, any `#[no_std]` crate should be 100% compatible with this new target. This target is currently somewhat janky due to how linking works. The "linking" is currently unconditional whole program LTO (aka LLVM is being used as a linker). Naturally that means compiling programs is pretty slow! Eventually though this target should have a linker. This target is also intended to be quite experimental. I'm hoping that this can act as a catalyst for further experimentation in Rust with WebAssembly. Breaking changes are very likely to land to this target, so it's not recommended to rely on it in any critical capacity yet. We'll let you know when it's "production ready". --- Currently testing-wise this target is looking pretty good but isn't complete. I've got almost the entire `run-pass` test suite working with this target (lots of tests ignored, but many passing as well). The `core` test suite is still getting LLVM bugs fixed to get that working and will take some time. Relatively simple programs all seem to work though! --- It's worth nothing that you may not immediately see the "smallest possible wasm module" for the input you feed to rustc. For various reasons it's very difficult to get rid of the final "bloat" in vanilla rustc (again, a real linker should fix all this). For now what you'll have to do is: cargo install --git https://github.com/alexcrichton/wasm-gc wasm-gc foo.wasm bar.wasm And then `bar.wasm` should be the smallest we can get it! --- In any case for now I'd love feedback on this, particularly on the various integration points if you've got better ideas of how to approach them!
2017-10-22 22:01:00 -05:00
bool DataSections,
bool TrapUnreachable,
bool Singlethread,
2018-09-13 12:43:15 -05:00
bool AsmComments,
bool EmitStackSizeSection,
bool RelaxELFRelocations,
bool UseInitArray,
const char *SplitDwarfFile) {
auto OptLevel = fromRust(RustOptLevel);
auto RM = fromRust(RustReloc);
auto CM = fromRust(RustCM);
std::string Error;
Triple Trip(Triple::normalize(TripleStr));
const llvm::Target *TheTarget =
TargetRegistry::lookupTarget(Trip.getTriple(), Error);
if (TheTarget == nullptr) {
LLVMRustSetLastError(Error.c_str());
return nullptr;
}
2013-08-22 22:58:42 -05:00
TargetOptions Options;
Options.FloatABIType = FloatABI::Default;
if (UseSoftFloat) {
Options.FloatABIType = FloatABI::Soft;
}
Options.DataSections = DataSections;
Options.FunctionSections = FunctionSections;
Options.MCOptions.AsmVerbose = AsmComments;
Options.MCOptions.PreserveAsmComments = AsmComments;
Options.MCOptions.ABIName = ABIStr;
if (SplitDwarfFile) {
Options.MCOptions.SplitDwarfFile = SplitDwarfFile;
}
Options.RelaxELFRelocations = RelaxELFRelocations;
Options.UseInitArray = UseInitArray;
if (TrapUnreachable) {
2018-05-08 08:10:16 -05:00
// Tell LLVM to codegen `unreachable` into an explicit trap instruction.
// This limits the extent of possible undefined behavior in some cases, as
// it prevents control flow from "falling through" into whatever code
// happens to be laid out next in memory.
Options.TrapUnreachable = true;
}
std: Add a new wasm32-unknown-unknown target This commit adds a new target to the compiler: wasm32-unknown-unknown. This target is a reimagining of what it looks like to generate WebAssembly code from Rust. Instead of using Emscripten which can bring with it a weighty runtime this instead is a target which uses only the LLVM backend for WebAssembly and a "custom linker" for now which will hopefully one day be direct calls to lld. Notable features of this target include: * There is zero runtime footprint. The target assumes nothing exists other than the wasm32 instruction set. * There is zero toolchain footprint beyond adding the target. No custom linker is needed, rustc contains everything. * Very small wasm modules can be generated directly from Rust code using this target. * Most of the standard library is stubbed out to return an error, but anything related to allocation works (aka `HashMap`, `Vec`, etc). * Naturally, any `#[no_std]` crate should be 100% compatible with this new target. This target is currently somewhat janky due to how linking works. The "linking" is currently unconditional whole program LTO (aka LLVM is being used as a linker). Naturally that means compiling programs is pretty slow! Eventually though this target should have a linker. This target is also intended to be quite experimental. I'm hoping that this can act as a catalyst for further experimentation in Rust with WebAssembly. Breaking changes are very likely to land to this target, so it's not recommended to rely on it in any critical capacity yet. We'll let you know when it's "production ready". --- Currently testing-wise this target is looking pretty good but isn't complete. I've got almost the entire `run-pass` test suite working with this target (lots of tests ignored, but many passing as well). The `core` test suite is still getting LLVM bugs fixed to get that working and will take some time. Relatively simple programs all seem to work though! --- It's worth nothing that you may not immediately see the "smallest possible wasm module" for the input you feed to rustc. For various reasons it's very difficult to get rid of the final "bloat" in vanilla rustc (again, a real linker should fix all this). For now what you'll have to do is: cargo install --git https://github.com/alexcrichton/wasm-gc wasm-gc foo.wasm bar.wasm And then `bar.wasm` should be the smallest we can get it! --- In any case for now I'd love feedback on this, particularly on the various integration points if you've got better ideas of how to approach them!
2017-10-22 22:01:00 -05:00
if (Singlethread) {
Options.ThreadModel = ThreadModel::Single;
}
2018-09-13 12:43:15 -05:00
Options.EmitStackSizeSection = EmitStackSizeSection;
TargetMachine *TM = TheTarget->createTargetMachine(
Trip.getTriple(), CPU, Feature, Options, RM, CM, OptLevel);
return wrap(TM);
2013-08-22 22:58:42 -05:00
}
extern "C" void LLVMRustDisposeTargetMachine(LLVMTargetMachineRef TM) {
delete unwrap(TM);
2013-08-22 22:58:42 -05:00
}
extern "C" void LLVMRustConfigurePassManagerBuilder(
LLVMPassManagerBuilderRef PMBR, LLVMRustCodeGenOptLevel OptLevel,
bool MergeFunctions, bool SLPVectorize, bool LoopVectorize, bool PrepareForThinLTO,
const char* PGOGenPath, const char* PGOUsePath) {
unwrap(PMBR)->MergeFunctions = MergeFunctions;
unwrap(PMBR)->SLPVectorize = SLPVectorize;
unwrap(PMBR)->OptLevel = fromRust(OptLevel);
unwrap(PMBR)->LoopVectorize = LoopVectorize;
unwrap(PMBR)->PrepareForThinLTO = PrepareForThinLTO;
if (PGOGenPath) {
assert(!PGOUsePath);
unwrap(PMBR)->EnablePGOInstrGen = true;
unwrap(PMBR)->PGOInstrGen = PGOGenPath;
}
if (PGOUsePath) {
assert(!PGOGenPath);
unwrap(PMBR)->PGOInstrUse = PGOUsePath;
}
2013-08-22 22:58:42 -05:00
}
// Unfortunately, the LLVM C API doesn't provide a way to set the `LibraryInfo`
// field of a PassManagerBuilder, we expose our own method of doing so.
extern "C" void LLVMRustAddBuilderLibraryInfo(LLVMPassManagerBuilderRef PMBR,
LLVMModuleRef M,
bool DisableSimplifyLibCalls) {
Triple TargetTriple(unwrap(M)->getTargetTriple());
TargetLibraryInfoImpl *TLI = new TargetLibraryInfoImpl(TargetTriple);
if (DisableSimplifyLibCalls)
TLI->disableAllFunctions();
unwrap(PMBR)->LibraryInfo = TLI;
2013-08-22 22:58:42 -05:00
}
// Unfortunately, the LLVM C API doesn't provide a way to create the
// TargetLibraryInfo pass, so we use this method to do so.
extern "C" void LLVMRustAddLibraryInfo(LLVMPassManagerRef PMR, LLVMModuleRef M,
bool DisableSimplifyLibCalls) {
Triple TargetTriple(unwrap(M)->getTargetTriple());
TargetLibraryInfoImpl TLII(TargetTriple);
if (DisableSimplifyLibCalls)
TLII.disableAllFunctions();
unwrap(PMR)->add(new TargetLibraryInfoWrapperPass(TLII));
2013-08-22 22:58:42 -05:00
}
// Unfortunately, the LLVM C API doesn't provide an easy way of iterating over
// all the functions in a module, so we do that manually here. You'll find
// similar code in clang's BackendUtil.cpp file.
extern "C" void LLVMRustRunFunctionPassManager(LLVMPassManagerRef PMR,
LLVMModuleRef M) {
llvm::legacy::FunctionPassManager *P =
unwrap<llvm::legacy::FunctionPassManager>(PMR);
P->doInitialization();
// Upgrade all calls to old intrinsics first.
for (Module::iterator I = unwrap(M)->begin(), E = unwrap(M)->end(); I != E;)
UpgradeCallsToIntrinsic(&*I++); // must be post-increment, as we remove
for (Module::iterator I = unwrap(M)->begin(), E = unwrap(M)->end(); I != E;
++I)
if (!I->isDeclaration())
P->run(*I);
P->doFinalization();
2013-08-22 22:58:42 -05:00
}
extern "C" void LLVMRustSetLLVMOptions(int Argc, char **Argv) {
// Initializing the command-line options more than once is not allowed. So,
// check if they've already been initialized. (This could happen if we're
// being called from rustpkg, for example). If the arguments change, then
// that's just kinda unfortunate.
static bool Initialized = false;
if (Initialized)
return;
Initialized = true;
cl::ParseCommandLineOptions(Argc, Argv);
2013-08-22 22:58:42 -05:00
}
enum class LLVMRustFileType {
AssemblyFile,
ObjectFile,
};
#if LLVM_VERSION_GE(10, 0)
static CodeGenFileType fromRust(LLVMRustFileType Type) {
switch (Type) {
case LLVMRustFileType::AssemblyFile:
return CGFT_AssemblyFile;
case LLVMRustFileType::ObjectFile:
return CGFT_ObjectFile;
default:
report_fatal_error("Bad FileType.");
}
}
#else
static TargetMachine::CodeGenFileType fromRust(LLVMRustFileType Type) {
switch (Type) {
case LLVMRustFileType::AssemblyFile:
return TargetMachine::CGFT_AssemblyFile;
case LLVMRustFileType::ObjectFile:
return TargetMachine::CGFT_ObjectFile;
default:
report_fatal_error("Bad FileType.");
}
}
#endif
extern "C" LLVMRustResult
LLVMRustWriteOutputFile(LLVMTargetMachineRef Target, LLVMPassManagerRef PMR,
LLVMModuleRef M, const char *Path, const char *DwoPath,
LLVMRustFileType RustFileType) {
llvm::legacy::PassManager *PM = unwrap<llvm::legacy::PassManager>(PMR);
auto FileType = fromRust(RustFileType);
2013-08-22 22:58:42 -05:00
std::string ErrorInfo;
2014-09-30 16:20:22 -05:00
std::error_code EC;
raw_fd_ostream OS(Path, EC, sys::fs::F_None);
2014-09-30 16:20:22 -05:00
if (EC)
ErrorInfo = EC.message();
2013-08-22 22:58:42 -05:00
if (ErrorInfo != "") {
LLVMRustSetLastError(ErrorInfo.c_str());
return LLVMRustResult::Failure;
2013-08-22 22:58:42 -05:00
}
buffer_ostream BOS(OS);
if (DwoPath) {
raw_fd_ostream DOS(DwoPath, EC, sys::fs::F_None);
EC.clear();
if (EC)
ErrorInfo = EC.message();
if (ErrorInfo != "") {
LLVMRustSetLastError(ErrorInfo.c_str());
return LLVMRustResult::Failure;
}
buffer_ostream DBOS(DOS);
unwrap(Target)->addPassesToEmitFile(*PM, BOS, &DBOS, FileType, false);
PM->run(*unwrap(M));
} else {
unwrap(Target)->addPassesToEmitFile(*PM, BOS, nullptr, FileType, false);
PM->run(*unwrap(M));
}
// Apparently `addPassesToEmitFile` adds a pointer to our on-the-stack output
// stream (OS), so the only real safe place to delete this is here? Don't we
// wish this was written in Rust?
LLVMDisposePassManager(PMR);
return LLVMRustResult::Success;
2013-08-22 22:58:42 -05:00
}
extern "C" typedef void (*LLVMRustSelfProfileBeforePassCallback)(void*, // LlvmSelfProfiler
const char*, // pass name
const char*); // IR name
extern "C" typedef void (*LLVMRustSelfProfileAfterPassCallback)(void*); // LlvmSelfProfiler
std::string LLVMRustwrappedIrGetName(const llvm::Any &WrappedIr) {
if (any_isa<const Module *>(WrappedIr))
return any_cast<const Module *>(WrappedIr)->getName().str();
if (any_isa<const Function *>(WrappedIr))
return any_cast<const Function *>(WrappedIr)->getName().str();
if (any_isa<const Loop *>(WrappedIr))
return any_cast<const Loop *>(WrappedIr)->getName().str();
if (any_isa<const LazyCallGraph::SCC *>(WrappedIr))
return any_cast<const LazyCallGraph::SCC *>(WrappedIr)->getName();
return "<UNKNOWN>";
}
void LLVMSelfProfileInitializeCallbacks(
PassInstrumentationCallbacks& PIC, void* LlvmSelfProfiler,
LLVMRustSelfProfileBeforePassCallback BeforePassCallback,
LLVMRustSelfProfileAfterPassCallback AfterPassCallback) {
PIC.registerBeforePassCallback([LlvmSelfProfiler, BeforePassCallback](
StringRef Pass, llvm::Any Ir) {
std::string PassName = Pass.str();
std::string IrName = LLVMRustwrappedIrGetName(Ir);
BeforePassCallback(LlvmSelfProfiler, PassName.c_str(), IrName.c_str());
return true;
});
PIC.registerAfterPassCallback(
[LlvmSelfProfiler, AfterPassCallback](StringRef Pass, llvm::Any Ir) {
AfterPassCallback(LlvmSelfProfiler);
});
PIC.registerAfterPassInvalidatedCallback(
[LlvmSelfProfiler, AfterPassCallback](StringRef Pass) {
AfterPassCallback(LlvmSelfProfiler);
});
PIC.registerBeforeAnalysisCallback([LlvmSelfProfiler, BeforePassCallback](
StringRef Pass, llvm::Any Ir) {
std::string PassName = Pass.str();
std::string IrName = LLVMRustwrappedIrGetName(Ir);
BeforePassCallback(LlvmSelfProfiler, PassName.c_str(), IrName.c_str());
});
PIC.registerAfterAnalysisCallback(
[LlvmSelfProfiler, AfterPassCallback](StringRef Pass, llvm::Any Ir) {
AfterPassCallback(LlvmSelfProfiler);
});
}
enum class LLVMRustOptStage {
PreLinkNoLTO,
PreLinkThinLTO,
PreLinkFatLTO,
ThinLTO,
FatLTO,
};
struct LLVMRustSanitizerOptions {
bool SanitizeAddress;
bool SanitizeAddressRecover;
bool SanitizeMemory;
bool SanitizeMemoryRecover;
int SanitizeMemoryTrackOrigins;
bool SanitizeThread;
};
extern "C" void
LLVMRustOptimizeWithNewPassManager(
LLVMModuleRef ModuleRef,
LLVMTargetMachineRef TMRef,
LLVMRustPassBuilderOptLevel OptLevelRust,
LLVMRustOptStage OptStage,
bool NoPrepopulatePasses, bool VerifyIR, bool UseThinLTOBuffers,
bool MergeFunctions, bool UnrollLoops, bool SLPVectorize, bool LoopVectorize,
bool DisableSimplifyLibCalls, bool EmitLifetimeMarkers,
LLVMRustSanitizerOptions *SanitizerOptions,
const char *PGOGenPath, const char *PGOUsePath,
void* LlvmSelfProfiler,
LLVMRustSelfProfileBeforePassCallback BeforePassCallback,
LLVMRustSelfProfileAfterPassCallback AfterPassCallback) {
Module *TheModule = unwrap(ModuleRef);
TargetMachine *TM = unwrap(TMRef);
PassBuilder::OptimizationLevel OptLevel = fromRust(OptLevelRust);
// FIXME: MergeFunctions is not supported by NewPM yet.
(void) MergeFunctions;
PipelineTuningOptions PTO;
PTO.LoopUnrolling = UnrollLoops;
PTO.LoopInterleaving = UnrollLoops;
PTO.LoopVectorization = LoopVectorize;
PTO.SLPVectorization = SLPVectorize;
PassInstrumentationCallbacks PIC;
StandardInstrumentations SI;
SI.registerCallbacks(PIC);
if (LlvmSelfProfiler){
LLVMSelfProfileInitializeCallbacks(PIC,LlvmSelfProfiler,BeforePassCallback,AfterPassCallback);
}
Optional<PGOOptions> PGOOpt;
if (PGOGenPath) {
assert(!PGOUsePath);
PGOOpt = PGOOptions(PGOGenPath, "", "", PGOOptions::IRInstr);
} else if (PGOUsePath) {
assert(!PGOGenPath);
PGOOpt = PGOOptions(PGOUsePath, "", "", PGOOptions::IRUse);
}
PassBuilder PB(TM, PTO, PGOOpt, &PIC);
// FIXME: We may want to expose this as an option.
bool DebugPassManager = false;
LoopAnalysisManager LAM(DebugPassManager);
FunctionAnalysisManager FAM(DebugPassManager);
CGSCCAnalysisManager CGAM(DebugPassManager);
ModuleAnalysisManager MAM(DebugPassManager);
FAM.registerPass([&] { return PB.buildDefaultAAPipeline(); });
Triple TargetTriple(TheModule->getTargetTriple());
std::unique_ptr<TargetLibraryInfoImpl> TLII(new TargetLibraryInfoImpl(TargetTriple));
if (DisableSimplifyLibCalls)
TLII->disableAllFunctions();
FAM.registerPass([&] { return TargetLibraryAnalysis(*TLII); });
PB.registerModuleAnalyses(MAM);
PB.registerCGSCCAnalyses(CGAM);
PB.registerFunctionAnalyses(FAM);
PB.registerLoopAnalyses(LAM);
PB.crossRegisterProxies(LAM, FAM, CGAM, MAM);
// We manually collect pipeline callbacks so we can apply them at O0, where the
// PassBuilder does not create a pipeline.
std::vector<std::function<void(ModulePassManager &)>> PipelineStartEPCallbacks;
2020-06-25 20:52:41 -05:00
#if LLVM_VERSION_GE(11, 0)
std::vector<std::function<void(ModulePassManager &, PassBuilder::OptimizationLevel)>>
OptimizerLastEPCallbacks;
#else
std::vector<std::function<void(FunctionPassManager &, PassBuilder::OptimizationLevel)>>
OptimizerLastEPCallbacks;
2020-06-25 20:52:41 -05:00
#endif
if (VerifyIR) {
PipelineStartEPCallbacks.push_back([VerifyIR](ModulePassManager &MPM) {
MPM.addPass(VerifierPass());
});
}
if (SanitizerOptions) {
if (SanitizerOptions->SanitizeMemory) {
MemorySanitizerOptions Options(
SanitizerOptions->SanitizeMemoryTrackOrigins,
SanitizerOptions->SanitizeMemoryRecover,
/*CompileKernel=*/false);
2020-06-25 20:52:41 -05:00
#if LLVM_VERSION_GE(11, 0)
OptimizerLastEPCallbacks.push_back(
[Options](ModulePassManager &MPM, PassBuilder::OptimizationLevel Level) {
MPM.addPass(MemorySanitizerPass(Options));
MPM.addPass(createModuleToFunctionPassAdaptor(MemorySanitizerPass(Options)));
}
);
#else
#if LLVM_VERSION_GE(10, 0)
PipelineStartEPCallbacks.push_back([Options](ModulePassManager &MPM) {
MPM.addPass(MemorySanitizerPass(Options));
});
#endif
OptimizerLastEPCallbacks.push_back(
[Options](FunctionPassManager &FPM, PassBuilder::OptimizationLevel Level) {
FPM.addPass(MemorySanitizerPass(Options));
}
);
2020-06-25 20:52:41 -05:00
#endif
}
if (SanitizerOptions->SanitizeThread) {
2020-06-25 20:52:41 -05:00
#if LLVM_VERSION_GE(11, 0)
OptimizerLastEPCallbacks.push_back(
[](ModulePassManager &MPM, PassBuilder::OptimizationLevel Level) {
MPM.addPass(ThreadSanitizerPass());
MPM.addPass(createModuleToFunctionPassAdaptor(ThreadSanitizerPass()));
}
);
#else
#if LLVM_VERSION_GE(10, 0)
PipelineStartEPCallbacks.push_back([](ModulePassManager &MPM) {
MPM.addPass(ThreadSanitizerPass());
});
#endif
OptimizerLastEPCallbacks.push_back(
[](FunctionPassManager &FPM, PassBuilder::OptimizationLevel Level) {
FPM.addPass(ThreadSanitizerPass());
}
);
2020-06-25 20:52:41 -05:00
#endif
}
if (SanitizerOptions->SanitizeAddress) {
2020-06-25 20:52:41 -05:00
#if LLVM_VERSION_GE(11, 0)
OptimizerLastEPCallbacks.push_back(
[SanitizerOptions](ModulePassManager &MPM, PassBuilder::OptimizationLevel Level) {
MPM.addPass(RequireAnalysisPass<ASanGlobalsMetadataAnalysis, Module>());
MPM.addPass(ModuleAddressSanitizerPass(
/*CompileKernel=*/false, SanitizerOptions->SanitizeAddressRecover));
MPM.addPass(createModuleToFunctionPassAdaptor(AddressSanitizerPass(
/*CompileKernel=*/false, SanitizerOptions->SanitizeAddressRecover,
/*UseAfterScope=*/true)));
}
);
#else
PipelineStartEPCallbacks.push_back([&](ModulePassManager &MPM) {
MPM.addPass(RequireAnalysisPass<ASanGlobalsMetadataAnalysis, Module>());
});
OptimizerLastEPCallbacks.push_back(
[SanitizerOptions](FunctionPassManager &FPM, PassBuilder::OptimizationLevel Level) {
FPM.addPass(AddressSanitizerPass(
/*CompileKernel=*/false, SanitizerOptions->SanitizeAddressRecover,
/*UseAfterScope=*/true));
}
);
PipelineStartEPCallbacks.push_back(
[SanitizerOptions](ModulePassManager &MPM) {
MPM.addPass(ModuleAddressSanitizerPass(
/*CompileKernel=*/false, SanitizerOptions->SanitizeAddressRecover));
}
);
2020-06-25 20:52:41 -05:00
#endif
}
}
ModulePassManager MPM(DebugPassManager);
if (!NoPrepopulatePasses) {
2020-06-25 20:52:41 -05:00
if (OptLevel == PassBuilder::OptimizationLevel::O0) {
for (const auto &C : PipelineStartEPCallbacks)
C(MPM);
2020-06-25 20:52:41 -05:00
#if LLVM_VERSION_GE(11, 0)
for (const auto &C : OptimizerLastEPCallbacks)
C(MPM, OptLevel);
#else
if (!OptimizerLastEPCallbacks.empty()) {
FunctionPassManager FPM(DebugPassManager);
for (const auto &C : OptimizerLastEPCallbacks)
C(FPM, OptLevel);
MPM.addPass(createModuleToFunctionPassAdaptor(std::move(FPM)));
}
2020-06-25 20:52:41 -05:00
#endif
MPM.addPass(AlwaysInlinerPass(EmitLifetimeMarkers));
#if LLVM_VERSION_GE(10, 0)
if (PGOOpt) {
PB.addPGOInstrPassesForO0(
MPM, DebugPassManager, PGOOpt->Action == PGOOptions::IRInstr,
/*IsCS=*/false, PGOOpt->ProfileFile, PGOOpt->ProfileRemappingFile);
}
#endif
} else {
for (const auto &C : PipelineStartEPCallbacks)
PB.registerPipelineStartEPCallback(C);
if (OptStage != LLVMRustOptStage::PreLinkThinLTO) {
for (const auto &C : OptimizerLastEPCallbacks)
PB.registerOptimizerLastEPCallback(C);
}
switch (OptStage) {
case LLVMRustOptStage::PreLinkNoLTO:
MPM = PB.buildPerModuleDefaultPipeline(OptLevel, DebugPassManager);
break;
case LLVMRustOptStage::PreLinkThinLTO:
MPM = PB.buildThinLTOPreLinkDefaultPipeline(OptLevel, DebugPassManager);
2020-06-25 20:52:41 -05:00
#if LLVM_VERSION_GE(11, 0)
for (const auto &C : OptimizerLastEPCallbacks)
C(MPM, OptLevel);
#else
if (!OptimizerLastEPCallbacks.empty()) {
FunctionPassManager FPM(DebugPassManager);
for (const auto &C : OptimizerLastEPCallbacks)
C(FPM, OptLevel);
MPM.addPass(createModuleToFunctionPassAdaptor(std::move(FPM)));
}
2020-06-25 20:52:41 -05:00
#endif
break;
case LLVMRustOptStage::PreLinkFatLTO:
MPM = PB.buildLTOPreLinkDefaultPipeline(OptLevel, DebugPassManager);
break;
case LLVMRustOptStage::ThinLTO:
// FIXME: Does it make sense to pass the ModuleSummaryIndex?
// It only seems to be needed for C++ specific optimizations.
MPM = PB.buildThinLTODefaultPipeline(OptLevel, DebugPassManager, nullptr);
break;
case LLVMRustOptStage::FatLTO:
MPM = PB.buildLTODefaultPipeline(OptLevel, DebugPassManager, nullptr);
break;
}
}
}
if (UseThinLTOBuffers) {
MPM.addPass(CanonicalizeAliasesPass());
MPM.addPass(NameAnonGlobalPass());
}
// Upgrade all calls to old intrinsics first.
for (Module::iterator I = TheModule->begin(), E = TheModule->end(); I != E;)
UpgradeCallsToIntrinsic(&*I++); // must be post-increment, as we remove
MPM.run(*TheModule, MAM);
}
// Callback to demangle function name
// Parameters:
// * name to be demangled
// * name len
// * output buffer
// * output buffer len
// Returns len of demangled string, or 0 if demangle failed.
typedef size_t (*DemangleFn)(const char*, size_t, char*, size_t);
namespace {
class RustAssemblyAnnotationWriter : public AssemblyAnnotationWriter {
DemangleFn Demangle;
std::vector<char> Buf;
public:
RustAssemblyAnnotationWriter(DemangleFn Demangle) : Demangle(Demangle) {}
// Return empty string if demangle failed
// or if name does not need to be demangled
StringRef CallDemangle(StringRef name) {
if (!Demangle) {
return StringRef();
}
if (Buf.size() < name.size() * 2) {
// Semangled name usually shorter than mangled,
// but allocate twice as much memory just in case
Buf.resize(name.size() * 2);
}
auto R = Demangle(name.data(), name.size(), Buf.data(), Buf.size());
if (!R) {
// Demangle failed.
return StringRef();
}
auto Demangled = StringRef(Buf.data(), R);
if (Demangled == name) {
// Do not print anything if demangled name is equal to mangled.
return StringRef();
}
return Demangled;
}
void emitFunctionAnnot(const Function *F,
formatted_raw_ostream &OS) override {
StringRef Demangled = CallDemangle(F->getName());
if (Demangled.empty()) {
return;
}
OS << "; " << Demangled << "\n";
}
void emitInstructionAnnot(const Instruction *I,
formatted_raw_ostream &OS) override {
const char *Name;
const Value *Value;
if (const CallInst *CI = dyn_cast<CallInst>(I)) {
Name = "call";
2020-06-25 20:52:41 -05:00
Value = CI->getCalledOperand();
} else if (const InvokeInst* II = dyn_cast<InvokeInst>(I)) {
Name = "invoke";
2020-06-25 20:52:41 -05:00
Value = II->getCalledOperand();
} else {
// Could demangle more operations, e. g.
// `store %place, @function`.
return;
}
if (!Value->hasName()) {
return;
}
StringRef Demangled = CallDemangle(Value->getName());
if (Demangled.empty()) {
return;
}
OS << "; " << Name << " " << Demangled << "\n";
}
};
} // namespace
extern "C" LLVMRustResult
LLVMRustPrintModule(LLVMModuleRef M, const char *Path, DemangleFn Demangle) {
2013-08-22 22:58:42 -05:00
std::string ErrorInfo;
2014-09-30 16:20:22 -05:00
std::error_code EC;
raw_fd_ostream OS(Path, EC, sys::fs::F_None);
2014-09-30 16:20:22 -05:00
if (EC)
ErrorInfo = EC.message();
if (ErrorInfo != "") {
LLVMRustSetLastError(ErrorInfo.c_str());
return LLVMRustResult::Failure;
}
RustAssemblyAnnotationWriter AAW(Demangle);
2013-08-22 22:58:42 -05:00
formatted_raw_ostream FOS(OS);
unwrap(M)->print(FOS, &AAW);
return LLVMRustResult::Success;
2013-08-22 22:58:42 -05:00
}
extern "C" void LLVMRustPrintPasses() {
LLVMInitializePasses();
struct MyListener : PassRegistrationListener {
void passEnumerate(const PassInfo *Info) {
StringRef PassArg = Info->getPassArgument();
StringRef PassName = Info->getPassName();
if (!PassArg.empty()) {
// These unsigned->signed casts could theoretically overflow, but
// realistically never will (and even if, the result is implementation
// defined rather plain UB).
printf("%15.*s - %.*s\n", (int)PassArg.size(), PassArg.data(),
(int)PassName.size(), PassName.data());
}
}
} Listener;
2013-08-22 22:58:42 -05:00
PassRegistry *PR = PassRegistry::getPassRegistry();
PR->enumerateWith(&Listener);
}
extern "C" void LLVMRustAddAlwaysInlinePass(LLVMPassManagerBuilderRef PMBR,
bool AddLifetimes) {
unwrap(PMBR)->Inliner = llvm::createAlwaysInlinerLegacyPass(AddLifetimes);
}
Implement LTO This commit implements LTO for rust leveraging LLVM's passes. What this means is: * When compiling an rlib, in addition to insdering foo.o into the archive, also insert foo.bc (the LLVM bytecode) of the optimized module. * When the compiler detects the -Z lto option, it will attempt to perform LTO on a staticlib or binary output. The compiler will emit an error if a dylib or rlib output is being generated. * The actual act of performing LTO is as follows: 1. Force all upstream libraries to have an rlib version available. 2. Load the bytecode of each upstream library from the rlib. 3. Link all this bytecode into the current LLVM module (just using llvm apis) 4. Run an internalization pass which internalizes all symbols except those found reachable for the local crate of compilation. 5. Run the LLVM LTO pass manager over this entire module 6a. If assembling an archive, then add all upstream rlibs into the output archive. This ignores all of the object/bitcode/metadata files rust generated and placed inside the rlibs. 6b. If linking a binary, create copies of all upstream rlibs, remove the rust-generated object-file, and then link everything as usual. As I have explained in #10741, this process is excruciatingly slow, so this is *not* turned on by default, and it is also why I have decided to hide it behind a -Z flag for now. The good news is that the binary sizes are about as small as they can be as a result of LTO, so it's definitely working. Closes #10741 Closes #10740
2013-12-03 01:19:29 -06:00
extern "C" void LLVMRustRunRestrictionPass(LLVMModuleRef M, char **Symbols,
size_t Len) {
llvm::legacy::PassManager passes;
auto PreserveFunctions = [=](const GlobalValue &GV) {
for (size_t I = 0; I < Len; I++) {
if (GV.getName() == Symbols[I]) {
return true;
}
}
return false;
};
passes.add(llvm::createInternalizePass(PreserveFunctions));
passes.run(*unwrap(M));
Implement LTO This commit implements LTO for rust leveraging LLVM's passes. What this means is: * When compiling an rlib, in addition to insdering foo.o into the archive, also insert foo.bc (the LLVM bytecode) of the optimized module. * When the compiler detects the -Z lto option, it will attempt to perform LTO on a staticlib or binary output. The compiler will emit an error if a dylib or rlib output is being generated. * The actual act of performing LTO is as follows: 1. Force all upstream libraries to have an rlib version available. 2. Load the bytecode of each upstream library from the rlib. 3. Link all this bytecode into the current LLVM module (just using llvm apis) 4. Run an internalization pass which internalizes all symbols except those found reachable for the local crate of compilation. 5. Run the LLVM LTO pass manager over this entire module 6a. If assembling an archive, then add all upstream rlibs into the output archive. This ignores all of the object/bitcode/metadata files rust generated and placed inside the rlibs. 6b. If linking a binary, create copies of all upstream rlibs, remove the rust-generated object-file, and then link everything as usual. As I have explained in #10741, this process is excruciatingly slow, so this is *not* turned on by default, and it is also why I have decided to hide it behind a -Z flag for now. The good news is that the binary sizes are about as small as they can be as a result of LTO, so it's definitely working. Closes #10741 Closes #10740
2013-12-03 01:19:29 -06:00
}
extern "C" void LLVMRustMarkAllFunctionsNounwind(LLVMModuleRef M) {
for (Module::iterator GV = unwrap(M)->begin(), E = unwrap(M)->end(); GV != E;
++GV) {
GV->setDoesNotThrow();
Function *F = dyn_cast<Function>(GV);
if (F == nullptr)
continue;
for (Function::iterator B = F->begin(), BE = F->end(); B != BE; ++B) {
for (BasicBlock::iterator I = B->begin(), IE = B->end(); I != IE; ++I) {
if (isa<InvokeInst>(I)) {
InvokeInst *CI = cast<InvokeInst>(I);
CI->setDoesNotThrow();
}
}
}
}
}
extern "C" void
LLVMRustSetDataLayoutFromTargetMachine(LLVMModuleRef Module,
LLVMTargetMachineRef TMR) {
TargetMachine *Target = unwrap(TMR);
unwrap(Module)->setDataLayout(Target->createDataLayout());
}
extern "C" void LLVMRustSetModulePICLevel(LLVMModuleRef M) {
unwrap(M)->setPICLevel(PICLevel::Level::BigPIC);
}
extern "C" void LLVMRustSetModulePIELevel(LLVMModuleRef M) {
unwrap(M)->setPIELevel(PIELevel::Level::Large);
}
rustc: Implement ThinLTO This commit is an implementation of LLVM's ThinLTO for consumption in rustc itself. Currently today LTO works by merging all relevant LLVM modules into one and then running optimization passes. "Thin" LTO operates differently by having more sharded work and allowing parallelism opportunities between optimizing codegen units. Further down the road Thin LTO also allows *incremental* LTO which should enable even faster release builds without compromising on the performance we have today. This commit uses a `-Z thinlto` flag to gate whether ThinLTO is enabled. It then also implements two forms of ThinLTO: * In one mode we'll *only* perform ThinLTO over the codegen units produced in a single compilation. That is, we won't load upstream rlibs, but we'll instead just perform ThinLTO amongst all codegen units produced by the compiler for the local crate. This is intended to emulate a desired end point where we have codegen units turned on by default for all crates and ThinLTO allows us to do this without performance loss. * In anther mode, like full LTO today, we'll optimize all upstream dependencies in "thin" mode. Unlike today, however, this LTO step is fully parallelized so should finish much more quickly. There's a good bit of comments about what the implementation is doing and where it came from, but the tl;dr; is that currently most of the support here is copied from upstream LLVM. This code duplication is done for a number of reasons: * Controlling parallelism means we can use the existing jobserver support to avoid overloading machines. * We will likely want a slightly different form of incremental caching which integrates with our own incremental strategy, but this is yet to be determined. * This buys us some flexibility about when/where we run ThinLTO, as well as having it tailored to fit our needs for the time being. * Finally this allows us to reuse some artifacts such as our `TargetMachine` creation, where all our options we used today aren't necessarily supported by upstream LLVM yet. My hope is that we can get some experience with this copy/paste in tree and then eventually upstream some work to LLVM itself to avoid the duplication while still ensuring our needs are met. Otherwise I fear that maintaining these bindings may be quite costly over the years with LLVM updates!
2017-07-23 10:14:38 -05:00
// Here you'll find an implementation of ThinLTO as used by the Rust compiler
// right now. This ThinLTO support is only enabled on "recent ish" versions of
// LLVM, and otherwise it's just blanket rejected from other compilers.
//
// Most of this implementation is straight copied from LLVM. At the time of
// this writing it wasn't *quite* suitable to reuse more code from upstream
// for our purposes, but we should strive to upstream this support once it's
// ready to go! I figure we may want a bit of testing locally first before
// sending this upstream to LLVM. I hear though they're quite eager to receive
// feedback like this!
//
// If you're reading this code and wondering "what in the world" or you're
// working "good lord by LLVM upgrade is *still* failing due to these bindings"
// then fear not! (ok maybe fear a little). All code here is mostly based
// on `lib/LTO/ThinLTOCodeGenerator.cpp` in LLVM.
//
// You'll find that the general layout here roughly corresponds to the `run`
// method in that file as well as `ProcessThinLTOModule`. Functions are
// specifically commented below as well, but if you're updating this code
// or otherwise trying to understand it, the LLVM source will be useful in
// interpreting the mysteries within.
//
// Otherwise I'll apologize in advance, it probably requires a relatively
// significant investment on your part to "truly understand" what's going on
// here. Not saying I do myself, but it took me awhile staring at LLVM's source
// and various online resources about ThinLTO to make heads or tails of all
// this.
// This is a shared data structure which *must* be threadsafe to share
// read-only amongst threads. This also corresponds basically to the arguments
// of the `ProcessThinLTOModule` function in the LLVM source.
struct LLVMRustThinLTOData {
// The combined index that is the global analysis over all modules we're
// performing ThinLTO for. This is mostly managed by LLVM.
ModuleSummaryIndex Index;
// All modules we may look at, stored as in-memory serialized versions. This
// is later used when inlining to ensure we can extract any module to inline
// from.
StringMap<MemoryBufferRef> ModuleMap;
// A set that we manage of everything we *don't* want internalized. Note that
// this includes all transitive references right now as well, but it may not
// always!
DenseSet<GlobalValue::GUID> GUIDPreservedSymbols;
// Not 100% sure what these are, but they impact what's internalized and
// what's inlined across modules, I believe.
StringMap<FunctionImporter::ImportMapTy> ImportLists;
StringMap<FunctionImporter::ExportSetTy> ExportLists;
StringMap<GVSummaryMapTy> ModuleToDefinedGVSummaries;
Use llvm::computeLTOCacheKey to determine post-ThinLTO CGU reuse During incremental ThinLTO compilation, we attempt to re-use the optimized (post-ThinLTO) bitcode file for a module if it is 'safe' to do so. Up until now, 'safe' has meant that the set of modules that our current modules imports from/exports to is unchanged from the previous compilation session. See PR #67020 and PR #71131 for more details. However, this turns out be insufficient to guarantee that it's safe to reuse the post-LTO module (i.e. that optimizing the pre-LTO module would produce the same result). When LLVM optimizes a module during ThinLTO, it may look at other information from the 'module index', such as whether a (non-imported!) global variable is used. If this information changes between compilation runs, we may end up re-using an optimized module that (for example) had dead-code elimination run on a function that is now used by another module. Fortunately, LLVM implements its own ThinLTO module cache, which is used when ThinLTO is performed by a linker plugin (e.g. when clang is used to compile a C proect). Using this cache directly would require extensive refactoring of our code - but fortunately for us, LLVM provides a function that does exactly what we need. The function `llvm::computeLTOCacheKey` is used to compute a SHA-1 hash from all data that might influence the result of ThinLTO on a module. In addition to the module imports/exports that we manually track, it also hashes information about global variables (e.g. their liveness) which might be used during optimization. By using this function, we shouldn't have to worry about new LLVM passes breaking our module re-use behavior. In LLVM, the output of this function forms part of the filename used to store the post-ThinLTO module. To keep our current filename structure intact, this PR just writes out the mapping 'CGU name -> Hash' to a file. To determine if a post-LTO module should be reused, we compare hashes from the previous session. This should unblock PR #75199 - by sheer chance, it seems to have hit this issue due to the particular CGU partitioning and optimization decisions that end up getting made.
2020-09-17 16:36:13 -05:00
StringMap<std::map<GlobalValue::GUID, GlobalValue::LinkageTypes>> ResolvedODR;
LLVMRustThinLTOData() : Index(/* HaveGVs = */ false) {}
rustc: Implement ThinLTO This commit is an implementation of LLVM's ThinLTO for consumption in rustc itself. Currently today LTO works by merging all relevant LLVM modules into one and then running optimization passes. "Thin" LTO operates differently by having more sharded work and allowing parallelism opportunities between optimizing codegen units. Further down the road Thin LTO also allows *incremental* LTO which should enable even faster release builds without compromising on the performance we have today. This commit uses a `-Z thinlto` flag to gate whether ThinLTO is enabled. It then also implements two forms of ThinLTO: * In one mode we'll *only* perform ThinLTO over the codegen units produced in a single compilation. That is, we won't load upstream rlibs, but we'll instead just perform ThinLTO amongst all codegen units produced by the compiler for the local crate. This is intended to emulate a desired end point where we have codegen units turned on by default for all crates and ThinLTO allows us to do this without performance loss. * In anther mode, like full LTO today, we'll optimize all upstream dependencies in "thin" mode. Unlike today, however, this LTO step is fully parallelized so should finish much more quickly. There's a good bit of comments about what the implementation is doing and where it came from, but the tl;dr; is that currently most of the support here is copied from upstream LLVM. This code duplication is done for a number of reasons: * Controlling parallelism means we can use the existing jobserver support to avoid overloading machines. * We will likely want a slightly different form of incremental caching which integrates with our own incremental strategy, but this is yet to be determined. * This buys us some flexibility about when/where we run ThinLTO, as well as having it tailored to fit our needs for the time being. * Finally this allows us to reuse some artifacts such as our `TargetMachine` creation, where all our options we used today aren't necessarily supported by upstream LLVM yet. My hope is that we can get some experience with this copy/paste in tree and then eventually upstream some work to LLVM itself to avoid the duplication while still ensuring our needs are met. Otherwise I fear that maintaining these bindings may be quite costly over the years with LLVM updates!
2017-07-23 10:14:38 -05:00
};
// Just an argument to the `LLVMRustCreateThinLTOData` function below.
struct LLVMRustThinLTOModule {
const char *identifier;
const char *data;
size_t len;
};
// This is copied from `lib/LTO/ThinLTOCodeGenerator.cpp`, not sure what it
// does.
static const GlobalValueSummary *
getFirstDefinitionForLinker(const GlobalValueSummaryList &GVSummaryList) {
auto StrongDefForLinker = llvm::find_if(
GVSummaryList, [](const std::unique_ptr<GlobalValueSummary> &Summary) {
auto Linkage = Summary->linkage();
return !GlobalValue::isAvailableExternallyLinkage(Linkage) &&
!GlobalValue::isWeakForLinker(Linkage);
});
if (StrongDefForLinker != GVSummaryList.end())
return StrongDefForLinker->get();
auto FirstDefForLinker = llvm::find_if(
GVSummaryList, [](const std::unique_ptr<GlobalValueSummary> &Summary) {
auto Linkage = Summary->linkage();
return !GlobalValue::isAvailableExternallyLinkage(Linkage);
});
if (FirstDefForLinker == GVSummaryList.end())
return nullptr;
return FirstDefForLinker->get();
}
// The main entry point for creating the global ThinLTO analysis. The structure
// here is basically the same as before threads are spawned in the `run`
// function of `lib/LTO/ThinLTOCodeGenerator.cpp`.
extern "C" LLVMRustThinLTOData*
LLVMRustCreateThinLTOData(LLVMRustThinLTOModule *modules,
int num_modules,
const char **preserved_symbols,
int num_symbols) {
2019-12-31 07:08:25 -06:00
#if LLVM_VERSION_GE(10, 0)
auto Ret = std::make_unique<LLVMRustThinLTOData>();
#else
rustc: Implement ThinLTO This commit is an implementation of LLVM's ThinLTO for consumption in rustc itself. Currently today LTO works by merging all relevant LLVM modules into one and then running optimization passes. "Thin" LTO operates differently by having more sharded work and allowing parallelism opportunities between optimizing codegen units. Further down the road Thin LTO also allows *incremental* LTO which should enable even faster release builds without compromising on the performance we have today. This commit uses a `-Z thinlto` flag to gate whether ThinLTO is enabled. It then also implements two forms of ThinLTO: * In one mode we'll *only* perform ThinLTO over the codegen units produced in a single compilation. That is, we won't load upstream rlibs, but we'll instead just perform ThinLTO amongst all codegen units produced by the compiler for the local crate. This is intended to emulate a desired end point where we have codegen units turned on by default for all crates and ThinLTO allows us to do this without performance loss. * In anther mode, like full LTO today, we'll optimize all upstream dependencies in "thin" mode. Unlike today, however, this LTO step is fully parallelized so should finish much more quickly. There's a good bit of comments about what the implementation is doing and where it came from, but the tl;dr; is that currently most of the support here is copied from upstream LLVM. This code duplication is done for a number of reasons: * Controlling parallelism means we can use the existing jobserver support to avoid overloading machines. * We will likely want a slightly different form of incremental caching which integrates with our own incremental strategy, but this is yet to be determined. * This buys us some flexibility about when/where we run ThinLTO, as well as having it tailored to fit our needs for the time being. * Finally this allows us to reuse some artifacts such as our `TargetMachine` creation, where all our options we used today aren't necessarily supported by upstream LLVM yet. My hope is that we can get some experience with this copy/paste in tree and then eventually upstream some work to LLVM itself to avoid the duplication while still ensuring our needs are met. Otherwise I fear that maintaining these bindings may be quite costly over the years with LLVM updates!
2017-07-23 10:14:38 -05:00
auto Ret = llvm::make_unique<LLVMRustThinLTOData>();
2019-12-31 07:08:25 -06:00
#endif
rustc: Implement ThinLTO This commit is an implementation of LLVM's ThinLTO for consumption in rustc itself. Currently today LTO works by merging all relevant LLVM modules into one and then running optimization passes. "Thin" LTO operates differently by having more sharded work and allowing parallelism opportunities between optimizing codegen units. Further down the road Thin LTO also allows *incremental* LTO which should enable even faster release builds without compromising on the performance we have today. This commit uses a `-Z thinlto` flag to gate whether ThinLTO is enabled. It then also implements two forms of ThinLTO: * In one mode we'll *only* perform ThinLTO over the codegen units produced in a single compilation. That is, we won't load upstream rlibs, but we'll instead just perform ThinLTO amongst all codegen units produced by the compiler for the local crate. This is intended to emulate a desired end point where we have codegen units turned on by default for all crates and ThinLTO allows us to do this without performance loss. * In anther mode, like full LTO today, we'll optimize all upstream dependencies in "thin" mode. Unlike today, however, this LTO step is fully parallelized so should finish much more quickly. There's a good bit of comments about what the implementation is doing and where it came from, but the tl;dr; is that currently most of the support here is copied from upstream LLVM. This code duplication is done for a number of reasons: * Controlling parallelism means we can use the existing jobserver support to avoid overloading machines. * We will likely want a slightly different form of incremental caching which integrates with our own incremental strategy, but this is yet to be determined. * This buys us some flexibility about when/where we run ThinLTO, as well as having it tailored to fit our needs for the time being. * Finally this allows us to reuse some artifacts such as our `TargetMachine` creation, where all our options we used today aren't necessarily supported by upstream LLVM yet. My hope is that we can get some experience with this copy/paste in tree and then eventually upstream some work to LLVM itself to avoid the duplication while still ensuring our needs are met. Otherwise I fear that maintaining these bindings may be quite costly over the years with LLVM updates!
2017-07-23 10:14:38 -05:00
// Load each module's summary and merge it into one combined index
for (int i = 0; i < num_modules; i++) {
auto module = &modules[i];
StringRef buffer(module->data, module->len);
MemoryBufferRef mem_buffer(buffer, module->identifier);
Ret->ModuleMap[module->identifier] = mem_buffer;
2017-10-10 14:29:14 -05:00
if (Error Err = readModuleSummaryIndex(mem_buffer, Ret->Index, i)) {
LLVMRustSetLastError(toString(std::move(Err)).c_str());
return nullptr;
}
rustc: Implement ThinLTO This commit is an implementation of LLVM's ThinLTO for consumption in rustc itself. Currently today LTO works by merging all relevant LLVM modules into one and then running optimization passes. "Thin" LTO operates differently by having more sharded work and allowing parallelism opportunities between optimizing codegen units. Further down the road Thin LTO also allows *incremental* LTO which should enable even faster release builds without compromising on the performance we have today. This commit uses a `-Z thinlto` flag to gate whether ThinLTO is enabled. It then also implements two forms of ThinLTO: * In one mode we'll *only* perform ThinLTO over the codegen units produced in a single compilation. That is, we won't load upstream rlibs, but we'll instead just perform ThinLTO amongst all codegen units produced by the compiler for the local crate. This is intended to emulate a desired end point where we have codegen units turned on by default for all crates and ThinLTO allows us to do this without performance loss. * In anther mode, like full LTO today, we'll optimize all upstream dependencies in "thin" mode. Unlike today, however, this LTO step is fully parallelized so should finish much more quickly. There's a good bit of comments about what the implementation is doing and where it came from, but the tl;dr; is that currently most of the support here is copied from upstream LLVM. This code duplication is done for a number of reasons: * Controlling parallelism means we can use the existing jobserver support to avoid overloading machines. * We will likely want a slightly different form of incremental caching which integrates with our own incremental strategy, but this is yet to be determined. * This buys us some flexibility about when/where we run ThinLTO, as well as having it tailored to fit our needs for the time being. * Finally this allows us to reuse some artifacts such as our `TargetMachine` creation, where all our options we used today aren't necessarily supported by upstream LLVM yet. My hope is that we can get some experience with this copy/paste in tree and then eventually upstream some work to LLVM itself to avoid the duplication while still ensuring our needs are met. Otherwise I fear that maintaining these bindings may be quite costly over the years with LLVM updates!
2017-07-23 10:14:38 -05:00
}
// Collect for each module the list of function it defines (GUID -> Summary)
Ret->Index.collectDefinedGVSummariesPerModule(Ret->ModuleToDefinedGVSummaries);
// Convert the preserved symbols set from string to GUID, this is then needed
// for internalization.
rustc: Implement ThinLTO This commit is an implementation of LLVM's ThinLTO for consumption in rustc itself. Currently today LTO works by merging all relevant LLVM modules into one and then running optimization passes. "Thin" LTO operates differently by having more sharded work and allowing parallelism opportunities between optimizing codegen units. Further down the road Thin LTO also allows *incremental* LTO which should enable even faster release builds without compromising on the performance we have today. This commit uses a `-Z thinlto` flag to gate whether ThinLTO is enabled. It then also implements two forms of ThinLTO: * In one mode we'll *only* perform ThinLTO over the codegen units produced in a single compilation. That is, we won't load upstream rlibs, but we'll instead just perform ThinLTO amongst all codegen units produced by the compiler for the local crate. This is intended to emulate a desired end point where we have codegen units turned on by default for all crates and ThinLTO allows us to do this without performance loss. * In anther mode, like full LTO today, we'll optimize all upstream dependencies in "thin" mode. Unlike today, however, this LTO step is fully parallelized so should finish much more quickly. There's a good bit of comments about what the implementation is doing and where it came from, but the tl;dr; is that currently most of the support here is copied from upstream LLVM. This code duplication is done for a number of reasons: * Controlling parallelism means we can use the existing jobserver support to avoid overloading machines. * We will likely want a slightly different form of incremental caching which integrates with our own incremental strategy, but this is yet to be determined. * This buys us some flexibility about when/where we run ThinLTO, as well as having it tailored to fit our needs for the time being. * Finally this allows us to reuse some artifacts such as our `TargetMachine` creation, where all our options we used today aren't necessarily supported by upstream LLVM yet. My hope is that we can get some experience with this copy/paste in tree and then eventually upstream some work to LLVM itself to avoid the duplication while still ensuring our needs are met. Otherwise I fear that maintaining these bindings may be quite costly over the years with LLVM updates!
2017-07-23 10:14:38 -05:00
for (int i = 0; i < num_symbols; i++) {
auto GUID = GlobalValue::getGUID(preserved_symbols[i]);
Ret->GUIDPreservedSymbols.insert(GUID);
rustc: Implement ThinLTO This commit is an implementation of LLVM's ThinLTO for consumption in rustc itself. Currently today LTO works by merging all relevant LLVM modules into one and then running optimization passes. "Thin" LTO operates differently by having more sharded work and allowing parallelism opportunities between optimizing codegen units. Further down the road Thin LTO also allows *incremental* LTO which should enable even faster release builds without compromising on the performance we have today. This commit uses a `-Z thinlto` flag to gate whether ThinLTO is enabled. It then also implements two forms of ThinLTO: * In one mode we'll *only* perform ThinLTO over the codegen units produced in a single compilation. That is, we won't load upstream rlibs, but we'll instead just perform ThinLTO amongst all codegen units produced by the compiler for the local crate. This is intended to emulate a desired end point where we have codegen units turned on by default for all crates and ThinLTO allows us to do this without performance loss. * In anther mode, like full LTO today, we'll optimize all upstream dependencies in "thin" mode. Unlike today, however, this LTO step is fully parallelized so should finish much more quickly. There's a good bit of comments about what the implementation is doing and where it came from, but the tl;dr; is that currently most of the support here is copied from upstream LLVM. This code duplication is done for a number of reasons: * Controlling parallelism means we can use the existing jobserver support to avoid overloading machines. * We will likely want a slightly different form of incremental caching which integrates with our own incremental strategy, but this is yet to be determined. * This buys us some flexibility about when/where we run ThinLTO, as well as having it tailored to fit our needs for the time being. * Finally this allows us to reuse some artifacts such as our `TargetMachine` creation, where all our options we used today aren't necessarily supported by upstream LLVM yet. My hope is that we can get some experience with this copy/paste in tree and then eventually upstream some work to LLVM itself to avoid the duplication while still ensuring our needs are met. Otherwise I fear that maintaining these bindings may be quite costly over the years with LLVM updates!
2017-07-23 10:14:38 -05:00
}
// Collect the import/export lists for all modules from the call-graph in the
// combined index
//
// This is copied from `lib/LTO/ThinLTOCodeGenerator.cpp`
auto deadIsPrevailing = [&](GlobalValue::GUID G) {
return PrevailingType::Unknown;
};
// We don't have a complete picture in our use of ThinLTO, just our immediate
// crate, so we need `ImportEnabled = false` to limit internalization.
// Otherwise, we sometimes lose `static` values -- see #60184.
computeDeadSymbolsWithConstProp(Ret->Index, Ret->GUIDPreservedSymbols,
deadIsPrevailing, /* ImportEnabled = */ false);
ComputeCrossModuleImport(
Ret->Index,
Ret->ModuleToDefinedGVSummaries,
Ret->ImportLists,
Ret->ExportLists
);
rustc: Implement ThinLTO This commit is an implementation of LLVM's ThinLTO for consumption in rustc itself. Currently today LTO works by merging all relevant LLVM modules into one and then running optimization passes. "Thin" LTO operates differently by having more sharded work and allowing parallelism opportunities between optimizing codegen units. Further down the road Thin LTO also allows *incremental* LTO which should enable even faster release builds without compromising on the performance we have today. This commit uses a `-Z thinlto` flag to gate whether ThinLTO is enabled. It then also implements two forms of ThinLTO: * In one mode we'll *only* perform ThinLTO over the codegen units produced in a single compilation. That is, we won't load upstream rlibs, but we'll instead just perform ThinLTO amongst all codegen units produced by the compiler for the local crate. This is intended to emulate a desired end point where we have codegen units turned on by default for all crates and ThinLTO allows us to do this without performance loss. * In anther mode, like full LTO today, we'll optimize all upstream dependencies in "thin" mode. Unlike today, however, this LTO step is fully parallelized so should finish much more quickly. There's a good bit of comments about what the implementation is doing and where it came from, but the tl;dr; is that currently most of the support here is copied from upstream LLVM. This code duplication is done for a number of reasons: * Controlling parallelism means we can use the existing jobserver support to avoid overloading machines. * We will likely want a slightly different form of incremental caching which integrates with our own incremental strategy, but this is yet to be determined. * This buys us some flexibility about when/where we run ThinLTO, as well as having it tailored to fit our needs for the time being. * Finally this allows us to reuse some artifacts such as our `TargetMachine` creation, where all our options we used today aren't necessarily supported by upstream LLVM yet. My hope is that we can get some experience with this copy/paste in tree and then eventually upstream some work to LLVM itself to avoid the duplication while still ensuring our needs are met. Otherwise I fear that maintaining these bindings may be quite costly over the years with LLVM updates!
2017-07-23 10:14:38 -05:00
// Resolve LinkOnce/Weak symbols, this has to be computed early be cause it
// impacts the caching.
//
// This is copied from `lib/LTO/ThinLTOCodeGenerator.cpp` with some of this
// being lifted from `lib/LTO/LTO.cpp` as well
rustc: Implement ThinLTO This commit is an implementation of LLVM's ThinLTO for consumption in rustc itself. Currently today LTO works by merging all relevant LLVM modules into one and then running optimization passes. "Thin" LTO operates differently by having more sharded work and allowing parallelism opportunities between optimizing codegen units. Further down the road Thin LTO also allows *incremental* LTO which should enable even faster release builds without compromising on the performance we have today. This commit uses a `-Z thinlto` flag to gate whether ThinLTO is enabled. It then also implements two forms of ThinLTO: * In one mode we'll *only* perform ThinLTO over the codegen units produced in a single compilation. That is, we won't load upstream rlibs, but we'll instead just perform ThinLTO amongst all codegen units produced by the compiler for the local crate. This is intended to emulate a desired end point where we have codegen units turned on by default for all crates and ThinLTO allows us to do this without performance loss. * In anther mode, like full LTO today, we'll optimize all upstream dependencies in "thin" mode. Unlike today, however, this LTO step is fully parallelized so should finish much more quickly. There's a good bit of comments about what the implementation is doing and where it came from, but the tl;dr; is that currently most of the support here is copied from upstream LLVM. This code duplication is done for a number of reasons: * Controlling parallelism means we can use the existing jobserver support to avoid overloading machines. * We will likely want a slightly different form of incremental caching which integrates with our own incremental strategy, but this is yet to be determined. * This buys us some flexibility about when/where we run ThinLTO, as well as having it tailored to fit our needs for the time being. * Finally this allows us to reuse some artifacts such as our `TargetMachine` creation, where all our options we used today aren't necessarily supported by upstream LLVM yet. My hope is that we can get some experience with this copy/paste in tree and then eventually upstream some work to LLVM itself to avoid the duplication while still ensuring our needs are met. Otherwise I fear that maintaining these bindings may be quite costly over the years with LLVM updates!
2017-07-23 10:14:38 -05:00
DenseMap<GlobalValue::GUID, const GlobalValueSummary *> PrevailingCopy;
for (auto &I : Ret->Index) {
2017-10-10 14:29:14 -05:00
if (I.second.SummaryList.size() > 1)
PrevailingCopy[I.first] = getFirstDefinitionForLinker(I.second.SummaryList);
rustc: Implement ThinLTO This commit is an implementation of LLVM's ThinLTO for consumption in rustc itself. Currently today LTO works by merging all relevant LLVM modules into one and then running optimization passes. "Thin" LTO operates differently by having more sharded work and allowing parallelism opportunities between optimizing codegen units. Further down the road Thin LTO also allows *incremental* LTO which should enable even faster release builds without compromising on the performance we have today. This commit uses a `-Z thinlto` flag to gate whether ThinLTO is enabled. It then also implements two forms of ThinLTO: * In one mode we'll *only* perform ThinLTO over the codegen units produced in a single compilation. That is, we won't load upstream rlibs, but we'll instead just perform ThinLTO amongst all codegen units produced by the compiler for the local crate. This is intended to emulate a desired end point where we have codegen units turned on by default for all crates and ThinLTO allows us to do this without performance loss. * In anther mode, like full LTO today, we'll optimize all upstream dependencies in "thin" mode. Unlike today, however, this LTO step is fully parallelized so should finish much more quickly. There's a good bit of comments about what the implementation is doing and where it came from, but the tl;dr; is that currently most of the support here is copied from upstream LLVM. This code duplication is done for a number of reasons: * Controlling parallelism means we can use the existing jobserver support to avoid overloading machines. * We will likely want a slightly different form of incremental caching which integrates with our own incremental strategy, but this is yet to be determined. * This buys us some flexibility about when/where we run ThinLTO, as well as having it tailored to fit our needs for the time being. * Finally this allows us to reuse some artifacts such as our `TargetMachine` creation, where all our options we used today aren't necessarily supported by upstream LLVM yet. My hope is that we can get some experience with this copy/paste in tree and then eventually upstream some work to LLVM itself to avoid the duplication while still ensuring our needs are met. Otherwise I fear that maintaining these bindings may be quite costly over the years with LLVM updates!
2017-07-23 10:14:38 -05:00
}
auto isPrevailing = [&](GlobalValue::GUID GUID, const GlobalValueSummary *S) {
const auto &Prevailing = PrevailingCopy.find(GUID);
if (Prevailing == PrevailingCopy.end())
return true;
return Prevailing->second == S;
};
auto recordNewLinkage = [&](StringRef ModuleIdentifier,
GlobalValue::GUID GUID,
GlobalValue::LinkageTypes NewLinkage) {
Use llvm::computeLTOCacheKey to determine post-ThinLTO CGU reuse During incremental ThinLTO compilation, we attempt to re-use the optimized (post-ThinLTO) bitcode file for a module if it is 'safe' to do so. Up until now, 'safe' has meant that the set of modules that our current modules imports from/exports to is unchanged from the previous compilation session. See PR #67020 and PR #71131 for more details. However, this turns out be insufficient to guarantee that it's safe to reuse the post-LTO module (i.e. that optimizing the pre-LTO module would produce the same result). When LLVM optimizes a module during ThinLTO, it may look at other information from the 'module index', such as whether a (non-imported!) global variable is used. If this information changes between compilation runs, we may end up re-using an optimized module that (for example) had dead-code elimination run on a function that is now used by another module. Fortunately, LLVM implements its own ThinLTO module cache, which is used when ThinLTO is performed by a linker plugin (e.g. when clang is used to compile a C proect). Using this cache directly would require extensive refactoring of our code - but fortunately for us, LLVM provides a function that does exactly what we need. The function `llvm::computeLTOCacheKey` is used to compute a SHA-1 hash from all data that might influence the result of ThinLTO on a module. In addition to the module imports/exports that we manually track, it also hashes information about global variables (e.g. their liveness) which might be used during optimization. By using this function, we shouldn't have to worry about new LLVM passes breaking our module re-use behavior. In LLVM, the output of this function forms part of the filename used to store the post-ThinLTO module. To keep our current filename structure intact, this PR just writes out the mapping 'CGU name -> Hash' to a file. To determine if a post-LTO module should be reused, we compare hashes from the previous session. This should unblock PR #75199 - by sheer chance, it seems to have hit this issue due to the particular CGU partitioning and optimization decisions that end up getting made.
2020-09-17 16:36:13 -05:00
Ret->ResolvedODR[ModuleIdentifier][GUID] = NewLinkage;
rustc: Implement ThinLTO This commit is an implementation of LLVM's ThinLTO for consumption in rustc itself. Currently today LTO works by merging all relevant LLVM modules into one and then running optimization passes. "Thin" LTO operates differently by having more sharded work and allowing parallelism opportunities between optimizing codegen units. Further down the road Thin LTO also allows *incremental* LTO which should enable even faster release builds without compromising on the performance we have today. This commit uses a `-Z thinlto` flag to gate whether ThinLTO is enabled. It then also implements two forms of ThinLTO: * In one mode we'll *only* perform ThinLTO over the codegen units produced in a single compilation. That is, we won't load upstream rlibs, but we'll instead just perform ThinLTO amongst all codegen units produced by the compiler for the local crate. This is intended to emulate a desired end point where we have codegen units turned on by default for all crates and ThinLTO allows us to do this without performance loss. * In anther mode, like full LTO today, we'll optimize all upstream dependencies in "thin" mode. Unlike today, however, this LTO step is fully parallelized so should finish much more quickly. There's a good bit of comments about what the implementation is doing and where it came from, but the tl;dr; is that currently most of the support here is copied from upstream LLVM. This code duplication is done for a number of reasons: * Controlling parallelism means we can use the existing jobserver support to avoid overloading machines. * We will likely want a slightly different form of incremental caching which integrates with our own incremental strategy, but this is yet to be determined. * This buys us some flexibility about when/where we run ThinLTO, as well as having it tailored to fit our needs for the time being. * Finally this allows us to reuse some artifacts such as our `TargetMachine` creation, where all our options we used today aren't necessarily supported by upstream LLVM yet. My hope is that we can get some experience with this copy/paste in tree and then eventually upstream some work to LLVM itself to avoid the duplication while still ensuring our needs are met. Otherwise I fear that maintaining these bindings may be quite costly over the years with LLVM updates!
2017-07-23 10:14:38 -05:00
};
thinLTOResolvePrevailingInIndex(Ret->Index, isPrevailing, recordNewLinkage,
Ret->GUIDPreservedSymbols);
// Here we calculate an `ExportedGUIDs` set for use in the `isExported`
// callback below. This callback below will dictate the linkage for all
// summaries in the index, and we basically just only want to ensure that dead
// symbols are internalized. Otherwise everything that's already external
// linkage will stay as external, and internal will stay as internal.
std::set<GlobalValue::GUID> ExportedGUIDs;
for (auto &List : Ret->Index) {
for (auto &GVS: List.second.SummaryList) {
if (GlobalValue::isLocalLinkage(GVS->linkage()))
continue;
auto GUID = GVS->getOriginalName();
if (GVS->flags().Live)
ExportedGUIDs.insert(GUID);
}
}
#if LLVM_VERSION_GE(10, 0)
auto isExported = [&](StringRef ModuleIdentifier, ValueInfo VI) {
const auto &ExportList = Ret->ExportLists.find(ModuleIdentifier);
return (ExportList != Ret->ExportLists.end() &&
ExportList->second.count(VI)) ||
ExportedGUIDs.count(VI.getGUID());
};
thinLTOInternalizeAndPromoteInIndex(Ret->Index, isExported, isPrevailing);
#else
rustc: Implement ThinLTO This commit is an implementation of LLVM's ThinLTO for consumption in rustc itself. Currently today LTO works by merging all relevant LLVM modules into one and then running optimization passes. "Thin" LTO operates differently by having more sharded work and allowing parallelism opportunities between optimizing codegen units. Further down the road Thin LTO also allows *incremental* LTO which should enable even faster release builds without compromising on the performance we have today. This commit uses a `-Z thinlto` flag to gate whether ThinLTO is enabled. It then also implements two forms of ThinLTO: * In one mode we'll *only* perform ThinLTO over the codegen units produced in a single compilation. That is, we won't load upstream rlibs, but we'll instead just perform ThinLTO amongst all codegen units produced by the compiler for the local crate. This is intended to emulate a desired end point where we have codegen units turned on by default for all crates and ThinLTO allows us to do this without performance loss. * In anther mode, like full LTO today, we'll optimize all upstream dependencies in "thin" mode. Unlike today, however, this LTO step is fully parallelized so should finish much more quickly. There's a good bit of comments about what the implementation is doing and where it came from, but the tl;dr; is that currently most of the support here is copied from upstream LLVM. This code duplication is done for a number of reasons: * Controlling parallelism means we can use the existing jobserver support to avoid overloading machines. * We will likely want a slightly different form of incremental caching which integrates with our own incremental strategy, but this is yet to be determined. * This buys us some flexibility about when/where we run ThinLTO, as well as having it tailored to fit our needs for the time being. * Finally this allows us to reuse some artifacts such as our `TargetMachine` creation, where all our options we used today aren't necessarily supported by upstream LLVM yet. My hope is that we can get some experience with this copy/paste in tree and then eventually upstream some work to LLVM itself to avoid the duplication while still ensuring our needs are met. Otherwise I fear that maintaining these bindings may be quite costly over the years with LLVM updates!
2017-07-23 10:14:38 -05:00
auto isExported = [&](StringRef ModuleIdentifier, GlobalValue::GUID GUID) {
const auto &ExportList = Ret->ExportLists.find(ModuleIdentifier);
return (ExportList != Ret->ExportLists.end() &&
ExportList->second.count(GUID)) ||
ExportedGUIDs.count(GUID);
rustc: Implement ThinLTO This commit is an implementation of LLVM's ThinLTO for consumption in rustc itself. Currently today LTO works by merging all relevant LLVM modules into one and then running optimization passes. "Thin" LTO operates differently by having more sharded work and allowing parallelism opportunities between optimizing codegen units. Further down the road Thin LTO also allows *incremental* LTO which should enable even faster release builds without compromising on the performance we have today. This commit uses a `-Z thinlto` flag to gate whether ThinLTO is enabled. It then also implements two forms of ThinLTO: * In one mode we'll *only* perform ThinLTO over the codegen units produced in a single compilation. That is, we won't load upstream rlibs, but we'll instead just perform ThinLTO amongst all codegen units produced by the compiler for the local crate. This is intended to emulate a desired end point where we have codegen units turned on by default for all crates and ThinLTO allows us to do this without performance loss. * In anther mode, like full LTO today, we'll optimize all upstream dependencies in "thin" mode. Unlike today, however, this LTO step is fully parallelized so should finish much more quickly. There's a good bit of comments about what the implementation is doing and where it came from, but the tl;dr; is that currently most of the support here is copied from upstream LLVM. This code duplication is done for a number of reasons: * Controlling parallelism means we can use the existing jobserver support to avoid overloading machines. * We will likely want a slightly different form of incremental caching which integrates with our own incremental strategy, but this is yet to be determined. * This buys us some flexibility about when/where we run ThinLTO, as well as having it tailored to fit our needs for the time being. * Finally this allows us to reuse some artifacts such as our `TargetMachine` creation, where all our options we used today aren't necessarily supported by upstream LLVM yet. My hope is that we can get some experience with this copy/paste in tree and then eventually upstream some work to LLVM itself to avoid the duplication while still ensuring our needs are met. Otherwise I fear that maintaining these bindings may be quite costly over the years with LLVM updates!
2017-07-23 10:14:38 -05:00
};
thinLTOInternalizeAndPromoteInIndex(Ret->Index, isExported);
#endif
rustc: Implement ThinLTO This commit is an implementation of LLVM's ThinLTO for consumption in rustc itself. Currently today LTO works by merging all relevant LLVM modules into one and then running optimization passes. "Thin" LTO operates differently by having more sharded work and allowing parallelism opportunities between optimizing codegen units. Further down the road Thin LTO also allows *incremental* LTO which should enable even faster release builds without compromising on the performance we have today. This commit uses a `-Z thinlto` flag to gate whether ThinLTO is enabled. It then also implements two forms of ThinLTO: * In one mode we'll *only* perform ThinLTO over the codegen units produced in a single compilation. That is, we won't load upstream rlibs, but we'll instead just perform ThinLTO amongst all codegen units produced by the compiler for the local crate. This is intended to emulate a desired end point where we have codegen units turned on by default for all crates and ThinLTO allows us to do this without performance loss. * In anther mode, like full LTO today, we'll optimize all upstream dependencies in "thin" mode. Unlike today, however, this LTO step is fully parallelized so should finish much more quickly. There's a good bit of comments about what the implementation is doing and where it came from, but the tl;dr; is that currently most of the support here is copied from upstream LLVM. This code duplication is done for a number of reasons: * Controlling parallelism means we can use the existing jobserver support to avoid overloading machines. * We will likely want a slightly different form of incremental caching which integrates with our own incremental strategy, but this is yet to be determined. * This buys us some flexibility about when/where we run ThinLTO, as well as having it tailored to fit our needs for the time being. * Finally this allows us to reuse some artifacts such as our `TargetMachine` creation, where all our options we used today aren't necessarily supported by upstream LLVM yet. My hope is that we can get some experience with this copy/paste in tree and then eventually upstream some work to LLVM itself to avoid the duplication while still ensuring our needs are met. Otherwise I fear that maintaining these bindings may be quite costly over the years with LLVM updates!
2017-07-23 10:14:38 -05:00
return Ret.release();
}
extern "C" void
LLVMRustFreeThinLTOData(LLVMRustThinLTOData *Data) {
delete Data;
}
// Below are the various passes that happen *per module* when doing ThinLTO.
//
// In other words, these are the functions that are all run concurrently
// with one another, one per module. The passes here correspond to the analysis
// passes in `lib/LTO/ThinLTOCodeGenerator.cpp`, currently found in the
// `ProcessThinLTOModule` function. Here they're split up into separate steps
// so rustc can save off the intermediate bytecode between each step.
2020-06-25 20:52:41 -05:00
#if LLVM_VERSION_GE(11, 0)
static bool
clearDSOLocalOnDeclarations(Module &Mod, TargetMachine &TM) {
// When linking an ELF shared object, dso_local should be dropped. We
// conservatively do this for -fpic.
bool ClearDSOLocalOnDeclarations =
TM.getTargetTriple().isOSBinFormatELF() &&
TM.getRelocationModel() != Reloc::Static &&
Mod.getPIELevel() == PIELevel::Default;
return ClearDSOLocalOnDeclarations;
}
#endif
rustc: Implement ThinLTO This commit is an implementation of LLVM's ThinLTO for consumption in rustc itself. Currently today LTO works by merging all relevant LLVM modules into one and then running optimization passes. "Thin" LTO operates differently by having more sharded work and allowing parallelism opportunities between optimizing codegen units. Further down the road Thin LTO also allows *incremental* LTO which should enable even faster release builds without compromising on the performance we have today. This commit uses a `-Z thinlto` flag to gate whether ThinLTO is enabled. It then also implements two forms of ThinLTO: * In one mode we'll *only* perform ThinLTO over the codegen units produced in a single compilation. That is, we won't load upstream rlibs, but we'll instead just perform ThinLTO amongst all codegen units produced by the compiler for the local crate. This is intended to emulate a desired end point where we have codegen units turned on by default for all crates and ThinLTO allows us to do this without performance loss. * In anther mode, like full LTO today, we'll optimize all upstream dependencies in "thin" mode. Unlike today, however, this LTO step is fully parallelized so should finish much more quickly. There's a good bit of comments about what the implementation is doing and where it came from, but the tl;dr; is that currently most of the support here is copied from upstream LLVM. This code duplication is done for a number of reasons: * Controlling parallelism means we can use the existing jobserver support to avoid overloading machines. * We will likely want a slightly different form of incremental caching which integrates with our own incremental strategy, but this is yet to be determined. * This buys us some flexibility about when/where we run ThinLTO, as well as having it tailored to fit our needs for the time being. * Finally this allows us to reuse some artifacts such as our `TargetMachine` creation, where all our options we used today aren't necessarily supported by upstream LLVM yet. My hope is that we can get some experience with this copy/paste in tree and then eventually upstream some work to LLVM itself to avoid the duplication while still ensuring our needs are met. Otherwise I fear that maintaining these bindings may be quite costly over the years with LLVM updates!
2017-07-23 10:14:38 -05:00
extern "C" bool
2020-06-25 20:52:41 -05:00
LLVMRustPrepareThinLTORename(const LLVMRustThinLTOData *Data, LLVMModuleRef M,
LLVMTargetMachineRef TM) {
rustc: Implement ThinLTO This commit is an implementation of LLVM's ThinLTO for consumption in rustc itself. Currently today LTO works by merging all relevant LLVM modules into one and then running optimization passes. "Thin" LTO operates differently by having more sharded work and allowing parallelism opportunities between optimizing codegen units. Further down the road Thin LTO also allows *incremental* LTO which should enable even faster release builds without compromising on the performance we have today. This commit uses a `-Z thinlto` flag to gate whether ThinLTO is enabled. It then also implements two forms of ThinLTO: * In one mode we'll *only* perform ThinLTO over the codegen units produced in a single compilation. That is, we won't load upstream rlibs, but we'll instead just perform ThinLTO amongst all codegen units produced by the compiler for the local crate. This is intended to emulate a desired end point where we have codegen units turned on by default for all crates and ThinLTO allows us to do this without performance loss. * In anther mode, like full LTO today, we'll optimize all upstream dependencies in "thin" mode. Unlike today, however, this LTO step is fully parallelized so should finish much more quickly. There's a good bit of comments about what the implementation is doing and where it came from, but the tl;dr; is that currently most of the support here is copied from upstream LLVM. This code duplication is done for a number of reasons: * Controlling parallelism means we can use the existing jobserver support to avoid overloading machines. * We will likely want a slightly different form of incremental caching which integrates with our own incremental strategy, but this is yet to be determined. * This buys us some flexibility about when/where we run ThinLTO, as well as having it tailored to fit our needs for the time being. * Finally this allows us to reuse some artifacts such as our `TargetMachine` creation, where all our options we used today aren't necessarily supported by upstream LLVM yet. My hope is that we can get some experience with this copy/paste in tree and then eventually upstream some work to LLVM itself to avoid the duplication while still ensuring our needs are met. Otherwise I fear that maintaining these bindings may be quite costly over the years with LLVM updates!
2017-07-23 10:14:38 -05:00
Module &Mod = *unwrap(M);
2020-06-25 20:52:41 -05:00
TargetMachine &Target = *unwrap(TM);
#if LLVM_VERSION_GE(11, 0)
bool ClearDSOLocal = clearDSOLocalOnDeclarations(Mod, Target);
bool error = renameModuleForThinLTO(Mod, Data->Index, ClearDSOLocal);
#else
bool error = renameModuleForThinLTO(Mod, Data->Index);
#endif
if (error) {
rustc: Implement ThinLTO This commit is an implementation of LLVM's ThinLTO for consumption in rustc itself. Currently today LTO works by merging all relevant LLVM modules into one and then running optimization passes. "Thin" LTO operates differently by having more sharded work and allowing parallelism opportunities between optimizing codegen units. Further down the road Thin LTO also allows *incremental* LTO which should enable even faster release builds without compromising on the performance we have today. This commit uses a `-Z thinlto` flag to gate whether ThinLTO is enabled. It then also implements two forms of ThinLTO: * In one mode we'll *only* perform ThinLTO over the codegen units produced in a single compilation. That is, we won't load upstream rlibs, but we'll instead just perform ThinLTO amongst all codegen units produced by the compiler for the local crate. This is intended to emulate a desired end point where we have codegen units turned on by default for all crates and ThinLTO allows us to do this without performance loss. * In anther mode, like full LTO today, we'll optimize all upstream dependencies in "thin" mode. Unlike today, however, this LTO step is fully parallelized so should finish much more quickly. There's a good bit of comments about what the implementation is doing and where it came from, but the tl;dr; is that currently most of the support here is copied from upstream LLVM. This code duplication is done for a number of reasons: * Controlling parallelism means we can use the existing jobserver support to avoid overloading machines. * We will likely want a slightly different form of incremental caching which integrates with our own incremental strategy, but this is yet to be determined. * This buys us some flexibility about when/where we run ThinLTO, as well as having it tailored to fit our needs for the time being. * Finally this allows us to reuse some artifacts such as our `TargetMachine` creation, where all our options we used today aren't necessarily supported by upstream LLVM yet. My hope is that we can get some experience with this copy/paste in tree and then eventually upstream some work to LLVM itself to avoid the duplication while still ensuring our needs are met. Otherwise I fear that maintaining these bindings may be quite costly over the years with LLVM updates!
2017-07-23 10:14:38 -05:00
LLVMRustSetLastError("renameModuleForThinLTO failed");
return false;
}
return true;
}
extern "C" bool
LLVMRustPrepareThinLTOResolveWeak(const LLVMRustThinLTOData *Data, LLVMModuleRef M) {
Module &Mod = *unwrap(M);
const auto &DefinedGlobals = Data->ModuleToDefinedGVSummaries.lookup(Mod.getModuleIdentifier());
2018-11-09 15:09:32 -06:00
thinLTOResolvePrevailingInModule(Mod, DefinedGlobals);
rustc: Implement ThinLTO This commit is an implementation of LLVM's ThinLTO for consumption in rustc itself. Currently today LTO works by merging all relevant LLVM modules into one and then running optimization passes. "Thin" LTO operates differently by having more sharded work and allowing parallelism opportunities between optimizing codegen units. Further down the road Thin LTO also allows *incremental* LTO which should enable even faster release builds without compromising on the performance we have today. This commit uses a `-Z thinlto` flag to gate whether ThinLTO is enabled. It then also implements two forms of ThinLTO: * In one mode we'll *only* perform ThinLTO over the codegen units produced in a single compilation. That is, we won't load upstream rlibs, but we'll instead just perform ThinLTO amongst all codegen units produced by the compiler for the local crate. This is intended to emulate a desired end point where we have codegen units turned on by default for all crates and ThinLTO allows us to do this without performance loss. * In anther mode, like full LTO today, we'll optimize all upstream dependencies in "thin" mode. Unlike today, however, this LTO step is fully parallelized so should finish much more quickly. There's a good bit of comments about what the implementation is doing and where it came from, but the tl;dr; is that currently most of the support here is copied from upstream LLVM. This code duplication is done for a number of reasons: * Controlling parallelism means we can use the existing jobserver support to avoid overloading machines. * We will likely want a slightly different form of incremental caching which integrates with our own incremental strategy, but this is yet to be determined. * This buys us some flexibility about when/where we run ThinLTO, as well as having it tailored to fit our needs for the time being. * Finally this allows us to reuse some artifacts such as our `TargetMachine` creation, where all our options we used today aren't necessarily supported by upstream LLVM yet. My hope is that we can get some experience with this copy/paste in tree and then eventually upstream some work to LLVM itself to avoid the duplication while still ensuring our needs are met. Otherwise I fear that maintaining these bindings may be quite costly over the years with LLVM updates!
2017-07-23 10:14:38 -05:00
return true;
}
extern "C" bool
LLVMRustPrepareThinLTOInternalize(const LLVMRustThinLTOData *Data, LLVMModuleRef M) {
Module &Mod = *unwrap(M);
const auto &DefinedGlobals = Data->ModuleToDefinedGVSummaries.lookup(Mod.getModuleIdentifier());
thinLTOInternalizeModule(Mod, DefinedGlobals);
return true;
}
extern "C" bool
2020-06-25 20:52:41 -05:00
LLVMRustPrepareThinLTOImport(const LLVMRustThinLTOData *Data, LLVMModuleRef M,
LLVMTargetMachineRef TM) {
rustc: Implement ThinLTO This commit is an implementation of LLVM's ThinLTO for consumption in rustc itself. Currently today LTO works by merging all relevant LLVM modules into one and then running optimization passes. "Thin" LTO operates differently by having more sharded work and allowing parallelism opportunities between optimizing codegen units. Further down the road Thin LTO also allows *incremental* LTO which should enable even faster release builds without compromising on the performance we have today. This commit uses a `-Z thinlto` flag to gate whether ThinLTO is enabled. It then also implements two forms of ThinLTO: * In one mode we'll *only* perform ThinLTO over the codegen units produced in a single compilation. That is, we won't load upstream rlibs, but we'll instead just perform ThinLTO amongst all codegen units produced by the compiler for the local crate. This is intended to emulate a desired end point where we have codegen units turned on by default for all crates and ThinLTO allows us to do this without performance loss. * In anther mode, like full LTO today, we'll optimize all upstream dependencies in "thin" mode. Unlike today, however, this LTO step is fully parallelized so should finish much more quickly. There's a good bit of comments about what the implementation is doing and where it came from, but the tl;dr; is that currently most of the support here is copied from upstream LLVM. This code duplication is done for a number of reasons: * Controlling parallelism means we can use the existing jobserver support to avoid overloading machines. * We will likely want a slightly different form of incremental caching which integrates with our own incremental strategy, but this is yet to be determined. * This buys us some flexibility about when/where we run ThinLTO, as well as having it tailored to fit our needs for the time being. * Finally this allows us to reuse some artifacts such as our `TargetMachine` creation, where all our options we used today aren't necessarily supported by upstream LLVM yet. My hope is that we can get some experience with this copy/paste in tree and then eventually upstream some work to LLVM itself to avoid the duplication while still ensuring our needs are met. Otherwise I fear that maintaining these bindings may be quite costly over the years with LLVM updates!
2017-07-23 10:14:38 -05:00
Module &Mod = *unwrap(M);
2020-06-25 20:52:41 -05:00
TargetMachine &Target = *unwrap(TM);
rustc: Implement ThinLTO This commit is an implementation of LLVM's ThinLTO for consumption in rustc itself. Currently today LTO works by merging all relevant LLVM modules into one and then running optimization passes. "Thin" LTO operates differently by having more sharded work and allowing parallelism opportunities between optimizing codegen units. Further down the road Thin LTO also allows *incremental* LTO which should enable even faster release builds without compromising on the performance we have today. This commit uses a `-Z thinlto` flag to gate whether ThinLTO is enabled. It then also implements two forms of ThinLTO: * In one mode we'll *only* perform ThinLTO over the codegen units produced in a single compilation. That is, we won't load upstream rlibs, but we'll instead just perform ThinLTO amongst all codegen units produced by the compiler for the local crate. This is intended to emulate a desired end point where we have codegen units turned on by default for all crates and ThinLTO allows us to do this without performance loss. * In anther mode, like full LTO today, we'll optimize all upstream dependencies in "thin" mode. Unlike today, however, this LTO step is fully parallelized so should finish much more quickly. There's a good bit of comments about what the implementation is doing and where it came from, but the tl;dr; is that currently most of the support here is copied from upstream LLVM. This code duplication is done for a number of reasons: * Controlling parallelism means we can use the existing jobserver support to avoid overloading machines. * We will likely want a slightly different form of incremental caching which integrates with our own incremental strategy, but this is yet to be determined. * This buys us some flexibility about when/where we run ThinLTO, as well as having it tailored to fit our needs for the time being. * Finally this allows us to reuse some artifacts such as our `TargetMachine` creation, where all our options we used today aren't necessarily supported by upstream LLVM yet. My hope is that we can get some experience with this copy/paste in tree and then eventually upstream some work to LLVM itself to avoid the duplication while still ensuring our needs are met. Otherwise I fear that maintaining these bindings may be quite costly over the years with LLVM updates!
2017-07-23 10:14:38 -05:00
const auto &ImportList = Data->ImportLists.lookup(Mod.getModuleIdentifier());
auto Loader = [&](StringRef Identifier) {
const auto &Memory = Data->ModuleMap.lookup(Identifier);
auto &Context = Mod.getContext();
auto MOrErr = getLazyBitcodeModule(Memory, Context, true, true);
if (!MOrErr)
return MOrErr;
// The rest of this closure is a workaround for
// https://bugs.llvm.org/show_bug.cgi?id=38184 where during ThinLTO imports
// we accidentally import wasm custom sections into different modules,
// duplicating them by in the final output artifact.
//
// The issue is worked around here by manually removing the
// `wasm.custom_sections` named metadata node from any imported module. This
// we know isn't used by any optimization pass so there's no need for it to
// be imported.
//
// Note that the metadata is currently lazily loaded, so we materialize it
// here before looking up if there's metadata inside. The `FunctionImporter`
// will immediately materialize metadata anyway after an import, so this
// shouldn't be a perf hit.
if (Error Err = (*MOrErr)->materializeMetadata()) {
Expected<std::unique_ptr<Module>> Ret(std::move(Err));
return Ret;
}
auto *WasmCustomSections = (*MOrErr)->getNamedMetadata("wasm.custom_sections");
if (WasmCustomSections)
WasmCustomSections->eraseFromParent();
return MOrErr;
rustc: Implement ThinLTO This commit is an implementation of LLVM's ThinLTO for consumption in rustc itself. Currently today LTO works by merging all relevant LLVM modules into one and then running optimization passes. "Thin" LTO operates differently by having more sharded work and allowing parallelism opportunities between optimizing codegen units. Further down the road Thin LTO also allows *incremental* LTO which should enable even faster release builds without compromising on the performance we have today. This commit uses a `-Z thinlto` flag to gate whether ThinLTO is enabled. It then also implements two forms of ThinLTO: * In one mode we'll *only* perform ThinLTO over the codegen units produced in a single compilation. That is, we won't load upstream rlibs, but we'll instead just perform ThinLTO amongst all codegen units produced by the compiler for the local crate. This is intended to emulate a desired end point where we have codegen units turned on by default for all crates and ThinLTO allows us to do this without performance loss. * In anther mode, like full LTO today, we'll optimize all upstream dependencies in "thin" mode. Unlike today, however, this LTO step is fully parallelized so should finish much more quickly. There's a good bit of comments about what the implementation is doing and where it came from, but the tl;dr; is that currently most of the support here is copied from upstream LLVM. This code duplication is done for a number of reasons: * Controlling parallelism means we can use the existing jobserver support to avoid overloading machines. * We will likely want a slightly different form of incremental caching which integrates with our own incremental strategy, but this is yet to be determined. * This buys us some flexibility about when/where we run ThinLTO, as well as having it tailored to fit our needs for the time being. * Finally this allows us to reuse some artifacts such as our `TargetMachine` creation, where all our options we used today aren't necessarily supported by upstream LLVM yet. My hope is that we can get some experience with this copy/paste in tree and then eventually upstream some work to LLVM itself to avoid the duplication while still ensuring our needs are met. Otherwise I fear that maintaining these bindings may be quite costly over the years with LLVM updates!
2017-07-23 10:14:38 -05:00
};
2020-06-25 20:52:41 -05:00
#if LLVM_VERSION_GE(11, 0)
bool ClearDSOLocal = clearDSOLocalOnDeclarations(Mod, Target);
FunctionImporter Importer(Data->Index, Loader, ClearDSOLocal);
#else
rustc: Implement ThinLTO This commit is an implementation of LLVM's ThinLTO for consumption in rustc itself. Currently today LTO works by merging all relevant LLVM modules into one and then running optimization passes. "Thin" LTO operates differently by having more sharded work and allowing parallelism opportunities between optimizing codegen units. Further down the road Thin LTO also allows *incremental* LTO which should enable even faster release builds without compromising on the performance we have today. This commit uses a `-Z thinlto` flag to gate whether ThinLTO is enabled. It then also implements two forms of ThinLTO: * In one mode we'll *only* perform ThinLTO over the codegen units produced in a single compilation. That is, we won't load upstream rlibs, but we'll instead just perform ThinLTO amongst all codegen units produced by the compiler for the local crate. This is intended to emulate a desired end point where we have codegen units turned on by default for all crates and ThinLTO allows us to do this without performance loss. * In anther mode, like full LTO today, we'll optimize all upstream dependencies in "thin" mode. Unlike today, however, this LTO step is fully parallelized so should finish much more quickly. There's a good bit of comments about what the implementation is doing and where it came from, but the tl;dr; is that currently most of the support here is copied from upstream LLVM. This code duplication is done for a number of reasons: * Controlling parallelism means we can use the existing jobserver support to avoid overloading machines. * We will likely want a slightly different form of incremental caching which integrates with our own incremental strategy, but this is yet to be determined. * This buys us some flexibility about when/where we run ThinLTO, as well as having it tailored to fit our needs for the time being. * Finally this allows us to reuse some artifacts such as our `TargetMachine` creation, where all our options we used today aren't necessarily supported by upstream LLVM yet. My hope is that we can get some experience with this copy/paste in tree and then eventually upstream some work to LLVM itself to avoid the duplication while still ensuring our needs are met. Otherwise I fear that maintaining these bindings may be quite costly over the years with LLVM updates!
2017-07-23 10:14:38 -05:00
FunctionImporter Importer(Data->Index, Loader);
2020-06-25 20:52:41 -05:00
#endif
rustc: Implement ThinLTO This commit is an implementation of LLVM's ThinLTO for consumption in rustc itself. Currently today LTO works by merging all relevant LLVM modules into one and then running optimization passes. "Thin" LTO operates differently by having more sharded work and allowing parallelism opportunities between optimizing codegen units. Further down the road Thin LTO also allows *incremental* LTO which should enable even faster release builds without compromising on the performance we have today. This commit uses a `-Z thinlto` flag to gate whether ThinLTO is enabled. It then also implements two forms of ThinLTO: * In one mode we'll *only* perform ThinLTO over the codegen units produced in a single compilation. That is, we won't load upstream rlibs, but we'll instead just perform ThinLTO amongst all codegen units produced by the compiler for the local crate. This is intended to emulate a desired end point where we have codegen units turned on by default for all crates and ThinLTO allows us to do this without performance loss. * In anther mode, like full LTO today, we'll optimize all upstream dependencies in "thin" mode. Unlike today, however, this LTO step is fully parallelized so should finish much more quickly. There's a good bit of comments about what the implementation is doing and where it came from, but the tl;dr; is that currently most of the support here is copied from upstream LLVM. This code duplication is done for a number of reasons: * Controlling parallelism means we can use the existing jobserver support to avoid overloading machines. * We will likely want a slightly different form of incremental caching which integrates with our own incremental strategy, but this is yet to be determined. * This buys us some flexibility about when/where we run ThinLTO, as well as having it tailored to fit our needs for the time being. * Finally this allows us to reuse some artifacts such as our `TargetMachine` creation, where all our options we used today aren't necessarily supported by upstream LLVM yet. My hope is that we can get some experience with this copy/paste in tree and then eventually upstream some work to LLVM itself to avoid the duplication while still ensuring our needs are met. Otherwise I fear that maintaining these bindings may be quite costly over the years with LLVM updates!
2017-07-23 10:14:38 -05:00
Expected<bool> Result = Importer.importFunctions(Mod, ImportList);
if (!Result) {
LLVMRustSetLastError(toString(Result.takeError()).c_str());
return false;
}
return true;
}
extern "C" typedef void (*LLVMRustModuleNameCallback)(void*, // payload
const char*, // importing module name
const char*); // imported module name
// Calls `module_name_callback` for each module import done by ThinLTO.
// The callback is provided with regular null-terminated C strings.
extern "C" void
Use llvm::computeLTOCacheKey to determine post-ThinLTO CGU reuse During incremental ThinLTO compilation, we attempt to re-use the optimized (post-ThinLTO) bitcode file for a module if it is 'safe' to do so. Up until now, 'safe' has meant that the set of modules that our current modules imports from/exports to is unchanged from the previous compilation session. See PR #67020 and PR #71131 for more details. However, this turns out be insufficient to guarantee that it's safe to reuse the post-LTO module (i.e. that optimizing the pre-LTO module would produce the same result). When LLVM optimizes a module during ThinLTO, it may look at other information from the 'module index', such as whether a (non-imported!) global variable is used. If this information changes between compilation runs, we may end up re-using an optimized module that (for example) had dead-code elimination run on a function that is now used by another module. Fortunately, LLVM implements its own ThinLTO module cache, which is used when ThinLTO is performed by a linker plugin (e.g. when clang is used to compile a C proect). Using this cache directly would require extensive refactoring of our code - but fortunately for us, LLVM provides a function that does exactly what we need. The function `llvm::computeLTOCacheKey` is used to compute a SHA-1 hash from all data that might influence the result of ThinLTO on a module. In addition to the module imports/exports that we manually track, it also hashes information about global variables (e.g. their liveness) which might be used during optimization. By using this function, we shouldn't have to worry about new LLVM passes breaking our module re-use behavior. In LLVM, the output of this function forms part of the filename used to store the post-ThinLTO module. To keep our current filename structure intact, this PR just writes out the mapping 'CGU name -> Hash' to a file. To determine if a post-LTO module should be reused, we compare hashes from the previous session. This should unblock PR #75199 - by sheer chance, it seems to have hit this issue due to the particular CGU partitioning and optimization decisions that end up getting made.
2020-09-17 16:36:13 -05:00
LLVMRustGetThinLTOModules(const LLVMRustThinLTOData *data,
LLVMRustModuleNameCallback module_name_callback,
void* callback_payload) {
for (const auto& importing_module : data->ImportLists) {
const std::string importing_module_id = importing_module.getKey().str();
const auto& imports = importing_module.getValue();
for (const auto& imported_module : imports) {
const std::string imported_module_id = imported_module.getKey().str();
module_name_callback(callback_payload,
importing_module_id.c_str(),
imported_module_id.c_str());
}
}
}
rustc: Implement ThinLTO This commit is an implementation of LLVM's ThinLTO for consumption in rustc itself. Currently today LTO works by merging all relevant LLVM modules into one and then running optimization passes. "Thin" LTO operates differently by having more sharded work and allowing parallelism opportunities between optimizing codegen units. Further down the road Thin LTO also allows *incremental* LTO which should enable even faster release builds without compromising on the performance we have today. This commit uses a `-Z thinlto` flag to gate whether ThinLTO is enabled. It then also implements two forms of ThinLTO: * In one mode we'll *only* perform ThinLTO over the codegen units produced in a single compilation. That is, we won't load upstream rlibs, but we'll instead just perform ThinLTO amongst all codegen units produced by the compiler for the local crate. This is intended to emulate a desired end point where we have codegen units turned on by default for all crates and ThinLTO allows us to do this without performance loss. * In anther mode, like full LTO today, we'll optimize all upstream dependencies in "thin" mode. Unlike today, however, this LTO step is fully parallelized so should finish much more quickly. There's a good bit of comments about what the implementation is doing and where it came from, but the tl;dr; is that currently most of the support here is copied from upstream LLVM. This code duplication is done for a number of reasons: * Controlling parallelism means we can use the existing jobserver support to avoid overloading machines. * We will likely want a slightly different form of incremental caching which integrates with our own incremental strategy, but this is yet to be determined. * This buys us some flexibility about when/where we run ThinLTO, as well as having it tailored to fit our needs for the time being. * Finally this allows us to reuse some artifacts such as our `TargetMachine` creation, where all our options we used today aren't necessarily supported by upstream LLVM yet. My hope is that we can get some experience with this copy/paste in tree and then eventually upstream some work to LLVM itself to avoid the duplication while still ensuring our needs are met. Otherwise I fear that maintaining these bindings may be quite costly over the years with LLVM updates!
2017-07-23 10:14:38 -05:00
// This struct and various functions are sort of a hack right now, but the
// problem is that we've got in-memory LLVM modules after we generate and
// optimize all codegen-units for one compilation in rustc. To be compatible
// with the LTO support above we need to serialize the modules plus their
// ThinLTO summary into memory.
//
// This structure is basically an owned version of a serialize module, with
// a ThinLTO summary attached.
struct LLVMRustThinLTOBuffer {
std::string data;
};
extern "C" LLVMRustThinLTOBuffer*
LLVMRustThinLTOBufferCreate(LLVMModuleRef M) {
2019-12-31 07:08:25 -06:00
#if LLVM_VERSION_GE(10, 0)
auto Ret = std::make_unique<LLVMRustThinLTOBuffer>();
#else
rustc: Implement ThinLTO This commit is an implementation of LLVM's ThinLTO for consumption in rustc itself. Currently today LTO works by merging all relevant LLVM modules into one and then running optimization passes. "Thin" LTO operates differently by having more sharded work and allowing parallelism opportunities between optimizing codegen units. Further down the road Thin LTO also allows *incremental* LTO which should enable even faster release builds without compromising on the performance we have today. This commit uses a `-Z thinlto` flag to gate whether ThinLTO is enabled. It then also implements two forms of ThinLTO: * In one mode we'll *only* perform ThinLTO over the codegen units produced in a single compilation. That is, we won't load upstream rlibs, but we'll instead just perform ThinLTO amongst all codegen units produced by the compiler for the local crate. This is intended to emulate a desired end point where we have codegen units turned on by default for all crates and ThinLTO allows us to do this without performance loss. * In anther mode, like full LTO today, we'll optimize all upstream dependencies in "thin" mode. Unlike today, however, this LTO step is fully parallelized so should finish much more quickly. There's a good bit of comments about what the implementation is doing and where it came from, but the tl;dr; is that currently most of the support here is copied from upstream LLVM. This code duplication is done for a number of reasons: * Controlling parallelism means we can use the existing jobserver support to avoid overloading machines. * We will likely want a slightly different form of incremental caching which integrates with our own incremental strategy, but this is yet to be determined. * This buys us some flexibility about when/where we run ThinLTO, as well as having it tailored to fit our needs for the time being. * Finally this allows us to reuse some artifacts such as our `TargetMachine` creation, where all our options we used today aren't necessarily supported by upstream LLVM yet. My hope is that we can get some experience with this copy/paste in tree and then eventually upstream some work to LLVM itself to avoid the duplication while still ensuring our needs are met. Otherwise I fear that maintaining these bindings may be quite costly over the years with LLVM updates!
2017-07-23 10:14:38 -05:00
auto Ret = llvm::make_unique<LLVMRustThinLTOBuffer>();
2019-12-31 07:08:25 -06:00
#endif
rustc: Implement ThinLTO This commit is an implementation of LLVM's ThinLTO for consumption in rustc itself. Currently today LTO works by merging all relevant LLVM modules into one and then running optimization passes. "Thin" LTO operates differently by having more sharded work and allowing parallelism opportunities between optimizing codegen units. Further down the road Thin LTO also allows *incremental* LTO which should enable even faster release builds without compromising on the performance we have today. This commit uses a `-Z thinlto` flag to gate whether ThinLTO is enabled. It then also implements two forms of ThinLTO: * In one mode we'll *only* perform ThinLTO over the codegen units produced in a single compilation. That is, we won't load upstream rlibs, but we'll instead just perform ThinLTO amongst all codegen units produced by the compiler for the local crate. This is intended to emulate a desired end point where we have codegen units turned on by default for all crates and ThinLTO allows us to do this without performance loss. * In anther mode, like full LTO today, we'll optimize all upstream dependencies in "thin" mode. Unlike today, however, this LTO step is fully parallelized so should finish much more quickly. There's a good bit of comments about what the implementation is doing and where it came from, but the tl;dr; is that currently most of the support here is copied from upstream LLVM. This code duplication is done for a number of reasons: * Controlling parallelism means we can use the existing jobserver support to avoid overloading machines. * We will likely want a slightly different form of incremental caching which integrates with our own incremental strategy, but this is yet to be determined. * This buys us some flexibility about when/where we run ThinLTO, as well as having it tailored to fit our needs for the time being. * Finally this allows us to reuse some artifacts such as our `TargetMachine` creation, where all our options we used today aren't necessarily supported by upstream LLVM yet. My hope is that we can get some experience with this copy/paste in tree and then eventually upstream some work to LLVM itself to avoid the duplication while still ensuring our needs are met. Otherwise I fear that maintaining these bindings may be quite costly over the years with LLVM updates!
2017-07-23 10:14:38 -05:00
{
raw_string_ostream OS(Ret->data);
{
legacy::PassManager PM;
PM.add(createWriteThinLTOBitcodePass(OS));
PM.run(*unwrap(M));
}
}
return Ret.release();
}
extern "C" void
LLVMRustThinLTOBufferFree(LLVMRustThinLTOBuffer *Buffer) {
delete Buffer;
}
extern "C" const void*
LLVMRustThinLTOBufferPtr(const LLVMRustThinLTOBuffer *Buffer) {
return Buffer->data.data();
}
extern "C" size_t
LLVMRustThinLTOBufferLen(const LLVMRustThinLTOBuffer *Buffer) {
return Buffer->data.length();
}
// This is what we used to parse upstream bitcode for actual ThinLTO
// processing. We'll call this once per module optimized through ThinLTO, and
// it'll be called concurrently on many threads.
extern "C" LLVMModuleRef
LLVMRustParseBitcodeForLTO(LLVMContextRef Context,
const char *data,
size_t len,
const char *identifier) {
rustc: Implement ThinLTO This commit is an implementation of LLVM's ThinLTO for consumption in rustc itself. Currently today LTO works by merging all relevant LLVM modules into one and then running optimization passes. "Thin" LTO operates differently by having more sharded work and allowing parallelism opportunities between optimizing codegen units. Further down the road Thin LTO also allows *incremental* LTO which should enable even faster release builds without compromising on the performance we have today. This commit uses a `-Z thinlto` flag to gate whether ThinLTO is enabled. It then also implements two forms of ThinLTO: * In one mode we'll *only* perform ThinLTO over the codegen units produced in a single compilation. That is, we won't load upstream rlibs, but we'll instead just perform ThinLTO amongst all codegen units produced by the compiler for the local crate. This is intended to emulate a desired end point where we have codegen units turned on by default for all crates and ThinLTO allows us to do this without performance loss. * In anther mode, like full LTO today, we'll optimize all upstream dependencies in "thin" mode. Unlike today, however, this LTO step is fully parallelized so should finish much more quickly. There's a good bit of comments about what the implementation is doing and where it came from, but the tl;dr; is that currently most of the support here is copied from upstream LLVM. This code duplication is done for a number of reasons: * Controlling parallelism means we can use the existing jobserver support to avoid overloading machines. * We will likely want a slightly different form of incremental caching which integrates with our own incremental strategy, but this is yet to be determined. * This buys us some flexibility about when/where we run ThinLTO, as well as having it tailored to fit our needs for the time being. * Finally this allows us to reuse some artifacts such as our `TargetMachine` creation, where all our options we used today aren't necessarily supported by upstream LLVM yet. My hope is that we can get some experience with this copy/paste in tree and then eventually upstream some work to LLVM itself to avoid the duplication while still ensuring our needs are met. Otherwise I fear that maintaining these bindings may be quite costly over the years with LLVM updates!
2017-07-23 10:14:38 -05:00
StringRef Data(data, len);
MemoryBufferRef Buffer(Data, identifier);
unwrap(Context)->enableDebugTypeODRUniquing();
Expected<std::unique_ptr<Module>> SrcOrError =
parseBitcodeFile(Buffer, *unwrap(Context));
if (!SrcOrError) {
LLVMRustSetLastError(toString(SrcOrError.takeError()).c_str());
return nullptr;
}
return wrap(std::move(*SrcOrError).release());
}
Store LLVM bitcode in object files, not compressed This commit is an attempted resurrection of #70458 where LLVM bitcode emitted by rustc into rlibs is stored into object file sections rather than in a separate file. The main rationale for doing this is that when rustc emits bitcode it will no longer use a custom compression scheme which makes it both easier to interoperate with existing tools and also cuts down on compile time since this compression isn't happening. The blocker for this in #70458 turned out to be that native linkers didn't handle the new sections well, causing the sections to either trigger bugs in the linker or actually end up in the final linked artifact. This commit attempts to address these issues by ensuring that native linkers ignore the new sections by inserting custom flags with module-level inline assembly. Note that this does not currently change the API of the compiler at all. The pre-existing `-C bitcode-in-rlib` flag is co-opted to indicate whether the bitcode should be present in the object file or not. Finally, note that an important consequence of this commit, which is also one of its primary purposes, is to enable rustc's `-Clto` bitcode loading to load rlibs produced with `-Clinker-plugin-lto`. The goal here is that when you're building with LTO Cargo will tell rustc to skip codegen of all intermediate crates and only generate LLVM IR. Today rustc will generate both object code and LLVM IR, but the object code is later simply thrown away, wastefully.
2020-04-23 13:45:55 -05:00
// Find the bitcode section in the object file data and return it as a slice.
// Fail if the bitcode section is present but empty.
//
// On success, the return value is the pointer to the start of the slice and
// `out_len` is filled with the (non-zero) length. On failure, the return value
// is `nullptr` and `out_len` is set to zero.
extern "C" const char*
LLVMRustGetBitcodeSliceFromObjectData(const char *data,
size_t len,
size_t *out_len) {
*out_len = 0;
StringRef Data(data, len);
MemoryBufferRef Buffer(Data, ""); // The id is unused.
Expected<MemoryBufferRef> BitcodeOrError =
object::IRObjectFile::findBitcodeInMemBuffer(Buffer);
if (!BitcodeOrError) {
LLVMRustSetLastError(toString(BitcodeOrError.takeError()).c_str());
return nullptr;
}
*out_len = BitcodeOrError->getBufferSize();
return BitcodeOrError->getBufferStart();
}
// Rewrite all `DICompileUnit` pointers to the `DICompileUnit` specified. See
// the comment in `back/lto.rs` for why this exists.
extern "C" void
LLVMRustThinLTOGetDICompileUnit(LLVMModuleRef Mod,
DICompileUnit **A,
DICompileUnit **B) {
Module *M = unwrap(Mod);
DICompileUnit **Cur = A;
DICompileUnit **Next = B;
for (DICompileUnit *CU : M->debug_compile_units()) {
*Cur = CU;
Cur = Next;
Next = nullptr;
if (Cur == nullptr)
break;
}
}
// Rewrite all `DICompileUnit` pointers to the `DICompileUnit` specified. See
// the comment in `back/lto.rs` for why this exists.
extern "C" void
LLVMRustThinLTOPatchDICompileUnit(LLVMModuleRef Mod, DICompileUnit *Unit) {
Module *M = unwrap(Mod);
// If the original source module didn't have a `DICompileUnit` then try to
// merge all the existing compile units. If there aren't actually any though
// then there's not much for us to do so return.
if (Unit == nullptr) {
for (DICompileUnit *CU : M->debug_compile_units()) {
Unit = CU;
break;
}
if (Unit == nullptr)
return;
}
// Use LLVM's built-in `DebugInfoFinder` to find a bunch of debuginfo and
2020-06-25 20:52:41 -05:00
// process it recursively. Note that we used to specifically iterate over
// instructions to ensure we feed everything into it, but `processModule`
// started doing this the same way in LLVM 7 (commit d769eb36ab2b8).
DebugInfoFinder Finder;
Finder.processModule(*M);
// After we've found all our debuginfo, rewrite all subprograms to point to
// the same `DICompileUnit`.
for (auto &F : Finder.subprograms()) {
F->replaceUnit(Unit);
}
// Erase any other references to other `DICompileUnit` instances, the verifier
// will later ensure that we don't actually have any other stale references to
// worry about.
auto *MD = M->getNamedMetadata("llvm.dbg.cu");
MD->clearOperands();
MD->addOperand(Unit);
}
Use llvm::computeLTOCacheKey to determine post-ThinLTO CGU reuse During incremental ThinLTO compilation, we attempt to re-use the optimized (post-ThinLTO) bitcode file for a module if it is 'safe' to do so. Up until now, 'safe' has meant that the set of modules that our current modules imports from/exports to is unchanged from the previous compilation session. See PR #67020 and PR #71131 for more details. However, this turns out be insufficient to guarantee that it's safe to reuse the post-LTO module (i.e. that optimizing the pre-LTO module would produce the same result). When LLVM optimizes a module during ThinLTO, it may look at other information from the 'module index', such as whether a (non-imported!) global variable is used. If this information changes between compilation runs, we may end up re-using an optimized module that (for example) had dead-code elimination run on a function that is now used by another module. Fortunately, LLVM implements its own ThinLTO module cache, which is used when ThinLTO is performed by a linker plugin (e.g. when clang is used to compile a C proect). Using this cache directly would require extensive refactoring of our code - but fortunately for us, LLVM provides a function that does exactly what we need. The function `llvm::computeLTOCacheKey` is used to compute a SHA-1 hash from all data that might influence the result of ThinLTO on a module. In addition to the module imports/exports that we manually track, it also hashes information about global variables (e.g. their liveness) which might be used during optimization. By using this function, we shouldn't have to worry about new LLVM passes breaking our module re-use behavior. In LLVM, the output of this function forms part of the filename used to store the post-ThinLTO module. To keep our current filename structure intact, this PR just writes out the mapping 'CGU name -> Hash' to a file. To determine if a post-LTO module should be reused, we compare hashes from the previous session. This should unblock PR #75199 - by sheer chance, it seems to have hit this issue due to the particular CGU partitioning and optimization decisions that end up getting made.
2020-09-17 16:36:13 -05:00
// Computes the LTO cache key for the provided 'ModId' in the given 'Data',
// storing the result in 'KeyOut'.
// Currently, this cache key is a SHA-1 hash of anything that could affect
// the result of optimizing this module (e.g. module imports, exports, liveness
// of access globals, etc).
// The precise details are determined by LLVM in `computeLTOCacheKey`, which is
// used during the normal linker-plugin incremental thin-LTO process.
extern "C" void
LLVMRustComputeLTOCacheKey(RustStringRef KeyOut, const char *ModId, LLVMRustThinLTOData *Data) {
SmallString<40> Key;
llvm::lto::Config conf;
const auto &ImportList = Data->ImportLists.lookup(ModId);
const auto &ExportList = Data->ExportLists.lookup(ModId);
const auto &ResolvedODR = Data->ResolvedODR.lookup(ModId);
const auto &DefinedGlobals = Data->ModuleToDefinedGVSummaries.lookup(ModId);
std::set<GlobalValue::GUID> CfiFunctionDefs;
std::set<GlobalValue::GUID> CfiFunctionDecls;
// Based on the 'InProcessThinBackend' constructor in LLVM
for (auto &Name : Data->Index.cfiFunctionDefs())
CfiFunctionDefs.insert(
GlobalValue::getGUID(GlobalValue::dropLLVMManglingEscape(Name)));
for (auto &Name : Data->Index.cfiFunctionDecls())
CfiFunctionDecls.insert(
GlobalValue::getGUID(GlobalValue::dropLLVMManglingEscape(Name)));
llvm::computeLTOCacheKey(Key, conf, Data->Index, ModId,
ImportList, ExportList, ResolvedODR, DefinedGlobals, CfiFunctionDefs, CfiFunctionDecls
);
LLVMRustStringWriteImpl(KeyOut, Key.c_str(), Key.size());
}