RuntimeIC::RuntimeIC(void* func_addr, int num_slots, int slot_size) : eh_frame(RUNTIMEICS_OMIT_FRAME_PTR) {
static StatCounter sc("runtime_ics_num");
sc.log();
if (ENABLE_RUNTIME_ICS) {
assert(SCRATCH_BYTES >= 0);
assert(SCRATCH_BYTES < 0x80); // This would break both the instruction encoding and the dwarf encoding
assert(SCRATCH_BYTES % 8 == 0);
#if RUNTIMEICS_OMIT_FRAME_PTR
/*
* prologue:
* sub $0x28, %rsp # 48 83 ec 28
*
* epilogue:
* add $0x28, %rsp # 48 83 c4 28
* retq # c3
*
*/
static const int PROLOGUE_SIZE = 4;
static const int EPILOGUE_SIZE = 5;
assert(SCRATCH_BYTES % 16 == 8);
#else
/*
* The prologue looks like:
* push %rbp # 55
* mov %rsp, %rbp # 48 89 e5
* sub $0x30, %rsp # 48 83 ec 30
*
* The epilogue is:
* add $0x30, %rsp # 48 83 c4 30
* pop %rbp # 5d
* retq # c3
*/
static const int PROLOGUE_SIZE = 8;
static const int EPILOGUE_SIZE = 6;
assert(SCRATCH_BYTES % 16 == 0);
#endif
static const int CALL_SIZE = 13;
int patchable_size = num_slots * slot_size;
#ifdef NVALGRIND
int total_size = PROLOGUE_SIZE + patchable_size + CALL_SIZE + EPILOGUE_SIZE;
addr = malloc(total_size);
#else
total_size = PROLOGUE_SIZE + patchable_size + CALL_SIZE + EPILOGUE_SIZE;
addr = mmap(NULL, (total_size + (PAGE_SIZE - 1)) & ~(PAGE_SIZE - 1), PROT_READ | PROT_WRITE | PROT_EXEC,
MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
RELEASE_ASSERT(addr != MAP_FAILED, "");
#endif
// printf("Allocated runtime IC at %p\n", addr);
std::unique_ptr<ICSetupInfo> setup_info(
ICSetupInfo::initialize(true, num_slots, slot_size, ICSetupInfo::Generic, NULL));
uint8_t* pp_start = (uint8_t*)addr + PROLOGUE_SIZE;
uint8_t* pp_end = pp_start + patchable_size + CALL_SIZE;
SpillMap _spill_map;
PatchpointInitializationInfo initialization_info
= initializePatchpoint3(func_addr, pp_start, pp_end, 0 /* scratch_offset */, 0 /* scratch_size */,
std::unordered_set<int>(), _spill_map);
assert(_spill_map.size() == 0);
assert(initialization_info.slowpath_start == pp_start + patchable_size);
assert(initialization_info.slowpath_rtn_addr == pp_end);
assert(initialization_info.continue_addr == pp_end);
StackInfo stack_info(SCRATCH_BYTES, 0);
icinfo = registerCompiledPatchpoint(pp_start, pp_start + patchable_size, pp_end, pp_end, setup_info.get(),
stack_info, std::unordered_set<int>());
assembler::Assembler prologue_assem((uint8_t*)addr, PROLOGUE_SIZE);
#if RUNTIMEICS_OMIT_FRAME_PTR
// If SCRATCH_BYTES is 8 or less, we could use more compact instruction encodings
// (push instead of sub), but it doesn't seem worth it for now.
prologue_assem.sub(assembler::Immediate(SCRATCH_BYTES), assembler::RSP);
#else
prologue_assem.push(assembler::RBP);
prologue_assem.mov(assembler::RSP, assembler::RBP);
prologue_assem.sub(assembler::Immediate(SCRATCH_BYTES), assembler::RSP);
#endif
assert(!prologue_assem.hasFailed());
assert(prologue_assem.isExactlyFull());
assembler::Assembler epilogue_assem(pp_end, EPILOGUE_SIZE);
#if RUNTIMEICS_OMIT_FRAME_PTR
epilogue_assem.add(assembler::Immediate(SCRATCH_BYTES), assembler::RSP);
#else
epilogue_assem.add(assembler::Immediate(SCRATCH_BYTES), assembler::RSP);
epilogue_assem.pop(assembler::RBP);
#endif
epilogue_assem.retq();
assert(!epilogue_assem.hasFailed());
assert(epilogue_assem.isExactlyFull());
// TODO: ideally would be more intelligent about allocation strategies.
// The code sections should be together and the eh sections together
eh_frame.writeAndRegister(addr, total_size);
} else {
addr = func_addr;
}
}
RuntimeIC::RuntimeIC(void* func_addr, int total_size) {
static StatCounter sc("num_runtime_ics");
sc.log();
if (ENABLE_RUNTIME_ICS) {
assert(SCRATCH_BYTES >= 0);
assert(SCRATCH_BYTES < 0x80); // This would break both the instruction encoding and the dwarf encoding
assert(SCRATCH_BYTES % 8 == 0);
#if RUNTIMEICS_OMIT_FRAME_PTR
/*
* prologue:
* sub $0x28, %rsp # 48 83 ec 28
*
* epilogue:
* add $0x28, %rsp # 48 83 c4 28
* retq # c3
*
*/
static const int PROLOGUE_SIZE = 4;
static const int EPILOGUE_SIZE = 5;
assert(SCRATCH_BYTES % 16 == 8);
#else
/*
* The prologue looks like:
* push %rbp # 55
* mov %rsp, %rbp # 48 89 e5
* sub $0x30, %rsp # 48 83 ec 30
*
* The epilogue is:
* add $0x30, %rsp # 48 83 c4 30
* pop %rbp # 5d
* retq # c3
*/
static const int PROLOGUE_SIZE = 8;
static const int EPILOGUE_SIZE = 6;
assert(SCRATCH_BYTES % 16 == 0);
#endif
static const int CALL_SIZE = 13;
int total_code_size = total_size - EH_FRAME_SIZE;
int patchable_size = total_code_size - (PROLOGUE_SIZE + CALL_SIZE + EPILOGUE_SIZE);
int total_size = total_code_size + EH_FRAME_SIZE;
assert(total_size == 512 && "we currently only have a 512 byte block memory manager");
addr = memory_manager_512b.alloc();
// the memory block contains the EH frame directly followed by the generated machine code.
void* eh_frame_addr = addr;
addr = (char*)addr + EH_FRAME_SIZE;
// printf("Allocated runtime IC at %p\n", addr);
std::unique_ptr<ICSetupInfo> setup_info(ICSetupInfo::initialize(true, patchable_size, ICSetupInfo::Generic));
uint8_t* pp_start = (uint8_t*)addr + PROLOGUE_SIZE;
uint8_t* pp_end = pp_start + patchable_size + CALL_SIZE;
SpillMap _spill_map;
PatchpointInitializationInfo initialization_info = initializePatchpoint3(
func_addr, pp_start, pp_end, 0 /* scratch_offset */, 0 /* scratch_size */, LiveOutSet(), _spill_map);
assert(_spill_map.size() == 0);
assert(initialization_info.slowpath_start == pp_start + patchable_size);
assert(initialization_info.slowpath_rtn_addr == pp_end);
assert(initialization_info.continue_addr == pp_end);
StackInfo stack_info(SCRATCH_BYTES, 0);
icinfo = registerCompiledPatchpoint(pp_start, pp_start + patchable_size, pp_end, pp_end, setup_info.get(),
stack_info, LiveOutSet());
assembler::Assembler prologue_assem((uint8_t*)addr, PROLOGUE_SIZE);
#if RUNTIMEICS_OMIT_FRAME_PTR
// If SCRATCH_BYTES is 8 or less, we could use more compact instruction encodings
// (push instead of sub), but it doesn't seem worth it for now.
prologue_assem.sub(assembler::Immediate(SCRATCH_BYTES), assembler::RSP);
#else
prologue_assem.push(assembler::RBP);
prologue_assem.mov(assembler::RSP, assembler::RBP);
prologue_assem.sub(assembler::Immediate(SCRATCH_BYTES), assembler::RSP);
#endif
assert(!prologue_assem.hasFailed());
assert(prologue_assem.isExactlyFull());
assembler::Assembler epilogue_assem(pp_end, EPILOGUE_SIZE);
#if RUNTIMEICS_OMIT_FRAME_PTR
epilogue_assem.add(assembler::Immediate(SCRATCH_BYTES), assembler::RSP);
#else
epilogue_assem.add(assembler::Immediate(SCRATCH_BYTES), assembler::RSP);
epilogue_assem.pop(assembler::RBP);
#endif
epilogue_assem.retq();
assert(!epilogue_assem.hasFailed());
assert(epilogue_assem.isExactlyFull());
if (RUNTIMEICS_OMIT_FRAME_PTR)
memcpy(eh_frame_addr, _eh_frame_template_ofp, _eh_frame_template_ofp_size);
else
memcpy(eh_frame_addr, _eh_frame_template_fp, _eh_frame_template_fp_size);
register_eh_frame.updateAndRegisterFrameFromTemplate((uint64_t)addr, total_code_size, (uint64_t)eh_frame_addr,
EH_FRAME_SIZE);
} else {
addr = func_addr;
}
}
