jit: Add custom memory manager (still segfaulting)

This commit is contained in:
Zack Corr 2012-08-26 22:40:21 +10:00 committed by Brian Anderson
parent d7aa9918ef
commit 7993f48209
5 changed files with 275 additions and 184 deletions

View file

@ -1,167 +0,0 @@
# This file is generated by gyp; do not edit.
TOOLSET := target
TARGET := uv
DEFS_Debug := '-DWIN32' \
'-D_CRT_SECURE_NO_DEPRECATE' \
'-D_CRT_NONSTDC_NO_DEPRECATE' \
'-DHAVE_CONFIG_H' \
'-D_WIN32_WINNT=0x0600' \
'-DEIO_STACKSIZE=262144' \
'-D_GNU_SOURCE' \
'-DDEBUG' \
'-D_DEBUG'
# Flags passed to all source files.
CFLAGS_Debug := -g \
-O0
# Flags passed to only C files.
CFLAGS_C_Debug :=
# Flags passed to only C++ files.
CFLAGS_CC_Debug :=
INCS_Debug := -I$(srcdir)\src\libuv\include \
-I$(srcdir)\src\libuv\include\uv-private \
-I$(srcdir)\src\libuv\src \
-I$(srcdir)\src\libuv\src\ares\config_win32
DEFS_Release := '-DWIN32' \
'-D_CRT_SECURE_NO_DEPRECATE' \
'-D_CRT_NONSTDC_NO_DEPRECATE' \
'-DHAVE_CONFIG_H' \
'-D_WIN32_WINNT=0x0600' \
'-DEIO_STACKSIZE=262144' \
'-D_GNU_SOURCE' \
'-DNDEBUG'
# Flags passed to all source files.
CFLAGS_Release := -O3 \
-fomit-frame-pointer \
-fdata-sections \
-ffunction-sections
# Flags passed to only C files.
CFLAGS_C_Release :=
# Flags passed to only C++ files.
CFLAGS_CC_Release :=
INCS_Release := -I$(srcdir)\src\libuv\include \
-I$(srcdir)\src\libuv\include\uv-private \
-I$(srcdir)\src\libuv\src \
-I$(srcdir)\src\libuv\src\ares\config_win32
OBJS := $(obj).target\$(TARGET)\src\libuv\src\uv-common.o \
$(obj).target\$(TARGET)\src\libuv\src\ares\ares_cancel.o \
$(obj).target\$(TARGET)\src\libuv\src\ares\ares__close_sockets.o \
$(obj).target\$(TARGET)\src\libuv\src\ares\ares_data.o \
$(obj).target\$(TARGET)\src\libuv\src\ares\ares_destroy.o \
$(obj).target\$(TARGET)\src\libuv\src\ares\ares_expand_name.o \
$(obj).target\$(TARGET)\src\libuv\src\ares\ares_expand_string.o \
$(obj).target\$(TARGET)\src\libuv\src\ares\ares_fds.o \
$(obj).target\$(TARGET)\src\libuv\src\ares\ares_free_hostent.o \
$(obj).target\$(TARGET)\src\libuv\src\ares\ares_free_string.o \
$(obj).target\$(TARGET)\src\libuv\src\ares\ares_gethostbyaddr.o \
$(obj).target\$(TARGET)\src\libuv\src\ares\ares_gethostbyname.o \
$(obj).target\$(TARGET)\src\libuv\src\ares\ares__get_hostent.o \
$(obj).target\$(TARGET)\src\libuv\src\ares\ares_getnameinfo.o \
$(obj).target\$(TARGET)\src\libuv\src\ares\ares_getopt.o \
$(obj).target\$(TARGET)\src\libuv\src\ares\ares_getsock.o \
$(obj).target\$(TARGET)\src\libuv\src\ares\ares_init.o \
$(obj).target\$(TARGET)\src\libuv\src\ares\ares_library_init.o \
$(obj).target\$(TARGET)\src\libuv\src\ares\ares_llist.o \
$(obj).target\$(TARGET)\src\libuv\src\ares\ares_mkquery.o \
$(obj).target\$(TARGET)\src\libuv\src\ares\ares_nowarn.o \
$(obj).target\$(TARGET)\src\libuv\src\ares\ares_options.o \
$(obj).target\$(TARGET)\src\libuv\src\ares\ares_parse_aaaa_reply.o \
$(obj).target\$(TARGET)\src\libuv\src\ares\ares_parse_a_reply.o \
$(obj).target\$(TARGET)\src\libuv\src\ares\ares_parse_mx_reply.o \
$(obj).target\$(TARGET)\src\libuv\src\ares\ares_parse_ns_reply.o \
$(obj).target\$(TARGET)\src\libuv\src\ares\ares_parse_ptr_reply.o \
$(obj).target\$(TARGET)\src\libuv\src\ares\ares_parse_srv_reply.o \
$(obj).target\$(TARGET)\src\libuv\src\ares\ares_parse_txt_reply.o \
$(obj).target\$(TARGET)\src\libuv\src\ares\ares_process.o \
$(obj).target\$(TARGET)\src\libuv\src\ares\ares_query.o \
$(obj).target\$(TARGET)\src\libuv\src\ares\ares__read_line.o \
$(obj).target\$(TARGET)\src\libuv\src\ares\ares_search.o \
$(obj).target\$(TARGET)\src\libuv\src\ares\ares_send.o \
$(obj).target\$(TARGET)\src\libuv\src\ares\ares_strcasecmp.o \
$(obj).target\$(TARGET)\src\libuv\src\ares\ares_strdup.o \
$(obj).target\$(TARGET)\src\libuv\src\ares\ares_strerror.o \
$(obj).target\$(TARGET)\src\libuv\src\ares\ares_timeout.o \
$(obj).target\$(TARGET)\src\libuv\src\ares\ares__timeval.o \
$(obj).target\$(TARGET)\src\libuv\src\ares\ares_version.o \
$(obj).target\$(TARGET)\src\libuv\src\ares\ares_writev.o \
$(obj).target\$(TARGET)\src\libuv\src\ares\bitncmp.o \
$(obj).target\$(TARGET)\src\libuv\src\ares\inet_net_pton.o \
$(obj).target\$(TARGET)\src\libuv\src\ares\inet_ntop.o \
$(obj).target\$(TARGET)\src\libuv\src\ares\windows_port.o \
$(obj).target\$(TARGET)\src\libuv\src\ares\ares_getenv.o \
$(obj).target\$(TARGET)\src\libuv\src\ares\ares_platform.o \
$(obj).target\$(TARGET)\src\libuv\src\win\async.o \
$(obj).target\$(TARGET)\src\libuv\src\win\cares.o \
$(obj).target\$(TARGET)\src\libuv\src\win\core.o \
$(obj).target\$(TARGET)\src\libuv\src\win\dl.o \
$(obj).target\$(TARGET)\src\libuv\src\win\error.o \
$(obj).target\$(TARGET)\src\libuv\src\win\fs.o \
$(obj).target\$(TARGET)\src\libuv\src\win\fs-event.o \
$(obj).target\$(TARGET)\src\libuv\src\win\getaddrinfo.o \
$(obj).target\$(TARGET)\src\libuv\src\win\handle.o \
$(obj).target\$(TARGET)\src\libuv\src\win\loop-watcher.o \
$(obj).target\$(TARGET)\src\libuv\src\win\pipe.o \
$(obj).target\$(TARGET)\src\libuv\src\win\thread.o \
$(obj).target\$(TARGET)\src\libuv\src\win\process.o \
$(obj).target\$(TARGET)\src\libuv\src\win\req.o \
$(obj).target\$(TARGET)\src\libuv\src\win\stream.o \
$(obj).target\$(TARGET)\src\libuv\src\win\tcp.o \
$(obj).target\$(TARGET)\src\libuv\src\win\tty.o \
$(obj).target\$(TARGET)\src\libuv\src\win\threadpool.o \
$(obj).target\$(TARGET)\src\libuv\src\win\timer.o \
$(obj).target\$(TARGET)\src\libuv\src\win\udp.o \
$(obj).target\$(TARGET)\src\libuv\src\win\util.o \
$(obj).target\$(TARGET)\src\libuv\src\win\winapi.o \
$(obj).target\$(TARGET)\src\libuv\src\win\winsock.o
# Add to the list of files we specially track dependencies for.
all_deps += $(OBJS)
# CFLAGS et al overrides must be target-local.
# See "Target-specific Variable Values" in the GNU Make manual.
$(OBJS): TOOLSET := $(TOOLSET)
$(OBJS): GYP_CFLAGS := $(DEFS_$(BUILDTYPE)) $(INCS_$(BUILDTYPE)) $(CFLAGS_$(BUILDTYPE)) $(CFLAGS_C_$(BUILDTYPE))
$(OBJS): GYP_CXXFLAGS := $(DEFS_$(BUILDTYPE)) $(INCS_$(BUILDTYPE)) $(CFLAGS_$(BUILDTYPE)) $(CFLAGS_CC_$(BUILDTYPE))
# Suffix rules, putting all outputs into $(obj).
$(obj).$(TOOLSET)\$(TARGET)\%.o: $(srcdir)\%.c FORCE_DO_CMD
@$(call do_cmd,cc,1)
# Try building from generated source, too.
$(obj).$(TOOLSET)\$(TARGET)\%.o: $(obj).$(TOOLSET)\%.c FORCE_DO_CMD
@$(call do_cmd,cc,1)
$(obj).$(TOOLSET)\$(TARGET)\%.o: $(obj)\%.c FORCE_DO_CMD
@$(call do_cmd,cc,1)
# End of this set of suffix rules
### Rules for final target.
LDFLAGS_Debug :=
LDFLAGS_Release :=
LIBS :=
$(obj).target\src\libuv\libuv.a: GYP_LDFLAGS := $(LDFLAGS_$(BUILDTYPE))
$(obj).target\src\libuv\libuv.a: LIBS := $(LIBS)
$(obj).target\src\libuv\libuv.a: TOOLSET := $(TOOLSET)
$(obj).target\src\libuv\libuv.a: $(OBJS) FORCE_DO_CMD
$(call do_cmd,alink)
all_deps += $(obj).target\src\libuv\libuv.a
# Add target alias
.PHONY: uv
uv: $(obj).target\src\libuv\libuv.a

View file

@ -160,14 +160,37 @@ mod write {
llvm::LLVMAddCFGSimplificationPass(pm.llpm);*/
// JIT execution takes ownership of the module,
// so don't dispose and return.
// so don't dispose and return. Due to a weird bug
// with dynamic libraries, we need to separate jitting
// into two functions and load crates inbetween.
if !llvm::LLVMRustJIT(pm.llpm,
llmod,
CodeGenOptLevel,
true) {
if !llvm::LLVMRustPrepareJIT(pm.llpm,
llmod,
CodeGenOptLevel,
true) {
llvm_err(sess, ~"Could not JIT");
}
// We need to tell LLVM where to resolve all linked
// symbols from. The equivalent of -lstd, -lcore, etc.
/*let cstore = sess.cstore;
for cstore::get_used_crate_files(cstore).each |cratepath| {
debug!{"linking: %s", cratepath};
let _: () = str::as_c_str(
cratepath,
|buf_t| {
if !llvm::LLVMRustLoadLibrary(buf_t) {
llvm_err(sess, ~"Could not link");
}
debug!{"linked: %s", cratepath};
});
}*/
if !llvm::LLVMRustExecuteJIT() {
llvm_err(sess, ~"Could not JIT");
}
if sess.time_llvm_passes() { llvm::LLVMRustPrintPassTimings(); }
return;
}

View file

@ -986,12 +986,18 @@ extern mod llvm {
call. */
fn LLVMRustGetLastError() -> *c_char;
/** JIT the module. **/
fn LLVMRustJIT(PM: PassManagerRef,
/** Load a shared library to resolve symbols against. */
fn LLVMRustLoadLibrary(Filename: *c_char) -> bool;
/** Create the JIT engine. */
fn LLVMRustPrepareJIT(PM: PassManagerRef,
M: ModuleRef,
OptLevel: c_int,
EnableSegmentedStacks: bool) -> bool;
/** Execute the JIT engine. */
fn LLVMRustExecuteJIT() -> bool;
/** Parses the bitcode in the given memory buffer. */
fn LLVMRustParseBitcode(MemBuf: MemoryBufferRef) -> ModuleRef;

View file

@ -29,8 +29,11 @@
#include "llvm/Support/Host.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/DynamicLibrary.h"
#include "llvm/Support/Memory.h"
#include "llvm/ExecutionEngine/ExecutionEngine.h"
#include "llvm/ExecutionEngine/JIT.h"
#include "llvm/ExecutionEngine/JITMemoryManager.h"
#include "llvm/ExecutionEngine/MCJIT.h"
#include "llvm/ExecutionEngine/Interpreter.h"
#include "llvm/ExecutionEngine/GenericValue.h"
#include "llvm-c/Core.h"
@ -38,6 +41,14 @@
#include "llvm-c/Object.h"
#include <cstdlib>
// Used by RustMCJITMemoryManager::getPointerToNamedFunction()
// to get around glibc issues. See the function for more information.
#ifdef __linux__
#include <sys/stat.h>
#include <fcntl.h>
#include <unistd.h>
#endif
using namespace llvm;
static const char *LLVMRustError;
@ -85,7 +96,208 @@ void LLVMInitializeX86AsmParser();
LLVMInitializeX86AsmParser();
extern "C" bool
LLVMRustJIT(LLVMPassManagerRef PMR,
LLVMRustLoadLibrary(const char* file) {
std::string err;
if(llvm::sys::DynamicLibrary::LoadLibraryPermanently(file, &err)) {
LLVMRustError = err.c_str();
return false;
}
return true;
}
ExecutionEngine* EE;
// Custom memory manager for MCJITting. It needs special features
// that the generic JIT memory manager doesn't entail. Based on
// code from LLI, change where needed for Rust.
class RustMCJITMemoryManager : public JITMemoryManager {
public:
SmallVector<sys::MemoryBlock, 16> AllocatedDataMem;
SmallVector<sys::MemoryBlock, 16> AllocatedCodeMem;
SmallVector<sys::MemoryBlock, 16> FreeCodeMem;
RustMCJITMemoryManager() { }
~RustMCJITMemoryManager();
virtual uint8_t *allocateCodeSection(uintptr_t Size, unsigned Alignment,
unsigned SectionID);
virtual uint8_t *allocateDataSection(uintptr_t Size, unsigned Alignment,
unsigned SectionID);
virtual void *getPointerToNamedFunction(const std::string &Name,
bool AbortOnFailure = true);
// Invalidate instruction cache for code sections. Some platforms with
// separate data cache and instruction cache require explicit cache flush,
// otherwise JIT code manipulations (like resolved relocations) will get to
// the data cache but not to the instruction cache.
virtual void invalidateInstructionCache();
// The MCJITMemoryManager doesn't use the following functions, so we don't
// need implement them.
virtual void setMemoryWritable() {
llvm_unreachable("Unimplemented call");
}
virtual void setMemoryExecutable() {
llvm_unreachable("Unimplemented call");
}
virtual void setPoisonMemory(bool poison) {
llvm_unreachable("Unimplemented call");
}
virtual void AllocateGOT() {
llvm_unreachable("Unimplemented call");
}
virtual uint8_t *getGOTBase() const {
llvm_unreachable("Unimplemented call");
return 0;
}
virtual uint8_t *startFunctionBody(const Function *F,
uintptr_t &ActualSize){
llvm_unreachable("Unimplemented call");
return 0;
}
virtual uint8_t *allocateStub(const GlobalValue* F, unsigned StubSize,
unsigned Alignment) {
llvm_unreachable("Unimplemented call");
return 0;
}
virtual void endFunctionBody(const Function *F, uint8_t *FunctionStart,
uint8_t *FunctionEnd) {
llvm_unreachable("Unimplemented call");
}
virtual uint8_t *allocateSpace(intptr_t Size, unsigned Alignment) {
llvm_unreachable("Unimplemented call");
return 0;
}
virtual uint8_t *allocateGlobal(uintptr_t Size, unsigned Alignment) {
llvm_unreachable("Unimplemented call");
return 0;
}
virtual void deallocateFunctionBody(void *Body) {
llvm_unreachable("Unimplemented call");
}
virtual uint8_t* startExceptionTable(const Function* F,
uintptr_t &ActualSize) {
llvm_unreachable("Unimplemented call");
return 0;
}
virtual void endExceptionTable(const Function *F, uint8_t *TableStart,
uint8_t *TableEnd, uint8_t* FrameRegister) {
llvm_unreachable("Unimplemented call");
}
virtual void deallocateExceptionTable(void *ET) {
llvm_unreachable("Unimplemented call");
}
};
uint8_t *RustMCJITMemoryManager::allocateDataSection(uintptr_t Size,
unsigned Alignment,
unsigned SectionID) {
if (!Alignment)
Alignment = 16;
uint8_t *Addr = (uint8_t*)calloc((Size + Alignment - 1)/Alignment, Alignment);
AllocatedDataMem.push_back(sys::MemoryBlock(Addr, Size));
return Addr;
}
uint8_t *RustMCJITMemoryManager::allocateCodeSection(uintptr_t Size,
unsigned Alignment,
unsigned SectionID) {
if (!Alignment)
Alignment = 16;
unsigned NeedAllocate = Alignment * ((Size + Alignment - 1)/Alignment + 1);
uintptr_t Addr = 0;
// Look in the list of free code memory regions and use a block there if one
// is available.
for (int i = 0, e = FreeCodeMem.size(); i != e; ++i) {
sys::MemoryBlock &MB = FreeCodeMem[i];
if (MB.size() >= NeedAllocate) {
Addr = (uintptr_t)MB.base();
uintptr_t EndOfBlock = Addr + MB.size();
// Align the address.
Addr = (Addr + Alignment - 1) & ~(uintptr_t)(Alignment - 1);
// Store cutted free memory block.
FreeCodeMem[i] = sys::MemoryBlock((void*)(Addr + Size),
EndOfBlock - Addr - Size);
return (uint8_t*)Addr;
}
}
// No pre-allocated free block was large enough. Allocate a new memory region.
sys::MemoryBlock MB = sys::Memory::AllocateRWX(NeedAllocate, 0, 0);
AllocatedCodeMem.push_back(MB);
Addr = (uintptr_t)MB.base();
uintptr_t EndOfBlock = Addr + MB.size();
// Align the address.
Addr = (Addr + Alignment - 1) & ~(uintptr_t)(Alignment - 1);
// The AllocateRWX may allocate much more memory than we need. In this case,
// we store the unused memory as a free memory block.
unsigned FreeSize = EndOfBlock-Addr-Size;
if (FreeSize > 16)
FreeCodeMem.push_back(sys::MemoryBlock((void*)(Addr + Size), FreeSize));
// Return aligned address
return (uint8_t*)Addr;
}
void RustMCJITMemoryManager::invalidateInstructionCache() {
for (int i = 0, e = AllocatedCodeMem.size(); i != e; ++i)
sys::Memory::InvalidateInstructionCache(AllocatedCodeMem[i].base(),
AllocatedCodeMem[i].size());
}
void *RustMCJITMemoryManager::getPointerToNamedFunction(const std::string &Name,
bool AbortOnFailure) {
#ifdef __linux__
// Force the following functions to be linked in to anything that uses the
// JIT. This is a hack designed to work around the all-too-clever Glibc
// strategy of making these functions work differently when inlined vs. when
// not inlined, and hiding their real definitions in a separate archive file
// that the dynamic linker can't see. For more info, search for
// 'libc_nonshared.a' on Google, or read http://llvm.org/PR274.
if (Name == "stat") return (void*)(intptr_t)&stat;
if (Name == "fstat") return (void*)(intptr_t)&fstat;
if (Name == "lstat") return (void*)(intptr_t)&lstat;
if (Name == "stat64") return (void*)(intptr_t)&stat64;
if (Name == "fstat64") return (void*)(intptr_t)&fstat64;
if (Name == "lstat64") return (void*)(intptr_t)&lstat64;
if (Name == "atexit") return (void*)(intptr_t)&atexit;
if (Name == "mknod") return (void*)(intptr_t)&mknod;
#endif
const char *NameStr = Name.c_str();
void *Ptr = sys::DynamicLibrary::SearchForAddressOfSymbol(NameStr);
if (Ptr) return Ptr;
// If it wasn't found and if it starts with an underscore ('_') character,
// try again without the underscore.
if (NameStr[0] == '_') {
Ptr = sys::DynamicLibrary::SearchForAddressOfSymbol(NameStr+1);
if (Ptr) return Ptr;
}
if (AbortOnFailure)
report_fatal_error("Program used external function '" + Name +
"' which could not be resolved!");
return 0;
}
RustMCJITMemoryManager::~RustMCJITMemoryManager() {
for (unsigned i = 0, e = AllocatedCodeMem.size(); i != e; ++i)
sys::Memory::ReleaseRWX(AllocatedCodeMem[i]);
for (unsigned i = 0, e = AllocatedDataMem.size(); i != e; ++i)
free(AllocatedDataMem[i].base());
}
// Separated functions because loading libraries before creating
// an execution engine seems to break stuff.
extern "C" bool
LLVMRustPrepareJIT(LLVMPassManagerRef PMR,
LLVMModuleRef M,
CodeGenOpt::Level OptLevel,
bool EnableSegmentedStacks) {
@ -96,15 +308,16 @@ LLVMRustJIT(LLVMPassManagerRef PMR,
std::string Err;
TargetOptions Options;
Options.NoFramePointerElim = true;
Options.EnableSegmentedStacks = EnableSegmentedStacks;
Options.JITEmitDebugInfo = true;
//Options.NoFramePointerElim = true;
//Options.EnableSegmentedStacks = EnableSegmentedStacks;
PassManager *PM = unwrap<PassManager>(PMR);
unwrap<PassManager>(PMR)->run(*unwrap(M));
PM->run(*unwrap(M));
ExecutionEngine* EE = EngineBuilder(unwrap(M))
RustMCJITMemoryManager* MM = new RustMCJITMemoryManager();
EE = EngineBuilder(unwrap(M))
.setTargetOptions(Options)
.setJITMemoryManager(MM)
.setOptLevel(OptLevel)
.setUseMCJIT(true)
.create();
@ -114,6 +327,16 @@ LLVMRustJIT(LLVMPassManagerRef PMR,
return false;
}
MM->invalidateInstructionCache();
return true;
}
extern "C" bool
LLVMRustExecuteJIT() {
assert(EE);
std::string Err;
Function* func = EE->FindFunctionNamed("main");
if(!func || Err != "") {
@ -121,9 +344,13 @@ LLVMRustJIT(LLVMPassManagerRef PMR,
return false;
}
std::vector<GenericValue> args;
//std::vector<GenericValue> args;
typedef int (*entry_t)(int, int);
entry_t entry = (entry_t) EE->getPointerToFunction(func);
EE->runFunction(func, args);
assert(entry);
entry(0, 0);
//EE->runFunction(func, args);
return true;
}

View file

@ -4,7 +4,9 @@ LLVMRustWriteOutputFile
LLVMRustGetLastError
LLVMRustConstSmallInt
LLVMRustConstInt
LLVMRustJIT
LLVMRustLoadLibrary
LLVMRustPrepareJIT
LLVMRustExecuteJIT
LLVMRustParseBitcode
LLVMRustParseAssemblyFile
LLVMRustPrintPassTimings