block_of_code: Remove far code machinery

This commit is contained in:
Merry 2022-07-13 13:26:29 +01:00 committed by merry
parent dd60f4b7d8
commit 840982be95
10 changed files with 16 additions and 69 deletions

View file

@ -114,9 +114,6 @@ A32EmitX64::BlockDescriptor A32EmitX64::Emit(IR::Block& block) {
// Start emitting. // Start emitting.
code.align(); code.align();
const u8* const entrypoint = code.getCurr(); const u8* const entrypoint = code.getCurr();
code.SwitchToFarCode();
const u8* const entrypoint_far = code.getCurr();
code.SwitchToNearCode();
EmitCondPrelude(ctx); EmitCondPrelude(ctx);
@ -168,7 +165,7 @@ A32EmitX64::BlockDescriptor A32EmitX64::Emit(IR::Block& block) {
const auto range = boost::icl::discrete_interval<u32>::closed(descriptor.PC(), end_location.PC() - 1); const auto range = boost::icl::discrete_interval<u32>::closed(descriptor.PC(), end_location.PC() - 1);
block_ranges.AddRange(range, descriptor); block_ranges.AddRange(range, descriptor);
return RegisterBlock(descriptor, entrypoint, entrypoint_far, size); return RegisterBlock(descriptor, entrypoint, size);
} }
void A32EmitX64::ClearCache() { void A32EmitX64::ClearCache() {

View file

@ -60,7 +60,7 @@ static Optimization::PolyfillOptions GenPolyfillOptions(const BlockOfCode& code)
struct Jit::Impl { struct Jit::Impl {
Impl(Jit* jit, A32::UserConfig conf) Impl(Jit* jit, A32::UserConfig conf)
: block_of_code(GenRunCodeCallbacks(conf.callbacks, &GetCurrentBlockThunk, this, conf), JitStateInfo{jit_state}, conf.code_cache_size, conf.far_code_offset, GenRCP(conf)) : block_of_code(GenRunCodeCallbacks(conf.callbacks, &GetCurrentBlockThunk, this, conf), JitStateInfo{jit_state}, conf.code_cache_size, GenRCP(conf))
, emitter(block_of_code, conf, jit) , emitter(block_of_code, conf, jit)
, polyfill_options(GenPolyfillOptions(block_of_code)) , polyfill_options(GenPolyfillOptions(block_of_code))
, conf(std::move(conf)) , conf(std::move(conf))

View file

@ -85,9 +85,6 @@ A64EmitX64::BlockDescriptor A64EmitX64::Emit(IR::Block& block) {
// Start emitting. // Start emitting.
code.align(); code.align();
const u8* const entrypoint = code.getCurr(); const u8* const entrypoint = code.getCurr();
code.SwitchToFarCode();
const u8* const entrypoint_far = code.getCurr();
code.SwitchToNearCode();
ASSERT(block.GetCondition() == IR::Cond::AL); ASSERT(block.GetCondition() == IR::Cond::AL);
@ -139,7 +136,7 @@ A64EmitX64::BlockDescriptor A64EmitX64::Emit(IR::Block& block) {
const auto range = boost::icl::discrete_interval<u64>::closed(descriptor.PC(), end_location.PC() - 1); const auto range = boost::icl::discrete_interval<u64>::closed(descriptor.PC(), end_location.PC() - 1);
block_ranges.AddRange(range, descriptor); block_ranges.AddRange(range, descriptor);
return RegisterBlock(descriptor, entrypoint, entrypoint_far, size); return RegisterBlock(descriptor, entrypoint, size);
} }
void A64EmitX64::ClearCache() { void A64EmitX64::ClearCache() {

View file

@ -58,7 +58,7 @@ struct Jit::Impl final {
public: public:
Impl(Jit* jit, UserConfig conf) Impl(Jit* jit, UserConfig conf)
: conf(conf) : conf(conf)
, block_of_code(GenRunCodeCallbacks(conf.callbacks, &GetCurrentBlockThunk, this, conf), JitStateInfo{jit_state}, conf.code_cache_size, conf.far_code_offset, GenRCP(conf)) , block_of_code(GenRunCodeCallbacks(conf.callbacks, &GetCurrentBlockThunk, this, conf), JitStateInfo{jit_state}, conf.code_cache_size, GenRCP(conf))
, emitter(block_of_code, conf, jit) , emitter(block_of_code, conf, jit)
, polyfill_options(GenPolyfillOptions(block_of_code)) { , polyfill_options(GenPolyfillOptions(block_of_code)) {
ASSERT(conf.page_table_address_space_bits >= 12 && conf.page_table_address_space_bits <= 64); ASSERT(conf.page_table_address_space_bits >= 12 && conf.page_table_address_space_bits <= 64);

View file

@ -185,22 +185,19 @@ HostFeature GetHostFeatures() {
} // anonymous namespace } // anonymous namespace
BlockOfCode::BlockOfCode(RunCodeCallbacks cb, JitStateInfo jsi, size_t total_code_size, size_t far_code_offset, std::function<void(BlockOfCode&)> rcp) BlockOfCode::BlockOfCode(RunCodeCallbacks cb, JitStateInfo jsi, size_t total_code_size, std::function<void(BlockOfCode&)> rcp)
: Xbyak::CodeGenerator(total_code_size, nullptr, &s_allocator) : Xbyak::CodeGenerator(total_code_size, nullptr, &s_allocator)
, cb(std::move(cb)) , cb(std::move(cb))
, jsi(jsi) , jsi(jsi)
, far_code_offset(far_code_offset)
, constant_pool(*this, CONSTANT_POOL_SIZE) , constant_pool(*this, CONSTANT_POOL_SIZE)
, host_features(GetHostFeatures()) { , host_features(GetHostFeatures()) {
ASSERT(total_code_size > far_code_offset);
EnableWriting(); EnableWriting();
GenRunCode(rcp); GenRunCode(rcp);
} }
void BlockOfCode::PreludeComplete() { void BlockOfCode::PreludeComplete() {
prelude_complete = true; prelude_complete = true;
near_code_begin = getCurr(); code_begin = getCurr();
far_code_begin = getCurr() + far_code_offset;
ClearCache(); ClearCache();
DisableWriting(); DisableWriting();
} }
@ -219,21 +216,15 @@ void BlockOfCode::DisableWriting() {
void BlockOfCode::ClearCache() { void BlockOfCode::ClearCache() {
ASSERT(prelude_complete); ASSERT(prelude_complete);
in_far_code = false; SetCodePtr(code_begin);
near_code_ptr = near_code_begin;
far_code_ptr = far_code_begin;
SetCodePtr(near_code_begin);
} }
size_t BlockOfCode::SpaceRemaining() const { size_t BlockOfCode::SpaceRemaining() const {
ASSERT(prelude_complete); ASSERT(prelude_complete);
const u8* current_near_ptr = in_far_code ? reinterpret_cast<const u8*>(near_code_ptr) : getCurr<const u8*>(); const u8* current_ptr = getCurr<const u8*>();
const u8* current_far_ptr = in_far_code ? getCurr<const u8*>() : reinterpret_cast<const u8*>(far_code_ptr); if (current_ptr >= &top_[maxSize_])
if (current_near_ptr >= far_code_begin)
return 0; return 0;
if (current_far_ptr >= &top_[maxSize_]) return &top_[maxSize_] - current_ptr;
return 0;
return std::min(reinterpret_cast<const u8*>(far_code_begin) - current_near_ptr, &top_[maxSize_] - current_far_ptr);
} }
HaltReason BlockOfCode::RunCode(void* jit_state, CodePtr code_ptr) const { HaltReason BlockOfCode::RunCode(void* jit_state, CodePtr code_ptr) const {
@ -406,26 +397,8 @@ Xbyak::Address BlockOfCode::XmmConst(const Xbyak::AddressFrame& frame, u64 lower
return constant_pool.GetConstant(frame, lower, upper); return constant_pool.GetConstant(frame, lower, upper);
} }
void BlockOfCode::SwitchToFarCode() {
ASSERT(prelude_complete);
ASSERT(!in_far_code);
in_far_code = true;
near_code_ptr = getCurr();
SetCodePtr(far_code_ptr);
ASSERT_MSG(near_code_ptr < far_code_begin, "Near code has overwritten far code!");
}
void BlockOfCode::SwitchToNearCode() {
ASSERT(prelude_complete);
ASSERT(in_far_code);
in_far_code = false;
far_code_ptr = getCurr();
SetCodePtr(near_code_ptr);
}
CodePtr BlockOfCode::GetCodeBegin() const { CodePtr BlockOfCode::GetCodeBegin() const {
return near_code_begin; return code_begin;
} }
size_t BlockOfCode::GetTotalCodeSize() const { size_t BlockOfCode::GetTotalCodeSize() const {

View file

@ -36,7 +36,7 @@ struct RunCodeCallbacks {
class BlockOfCode final : public Xbyak::CodeGenerator { class BlockOfCode final : public Xbyak::CodeGenerator {
public: public:
BlockOfCode(RunCodeCallbacks cb, JitStateInfo jsi, size_t total_code_size, size_t far_code_offset, std::function<void(BlockOfCode&)> rcp); BlockOfCode(RunCodeCallbacks cb, JitStateInfo jsi, size_t total_code_size, std::function<void(BlockOfCode&)> rcp);
BlockOfCode(const BlockOfCode&) = delete; BlockOfCode(const BlockOfCode&) = delete;
/// Call when external emitters have finished emitting their preludes. /// Call when external emitters have finished emitting their preludes.
@ -49,7 +49,7 @@ public:
/// Clears this block of code and resets code pointer to beginning. /// Clears this block of code and resets code pointer to beginning.
void ClearCache(); void ClearCache();
/// Calculates how much space is remaining to use. This is the minimum of near code and far code. /// Calculates how much space is remaining to use.
size_t SpaceRemaining() const; size_t SpaceRemaining() const;
/// Runs emulated code from code_ptr. /// Runs emulated code from code_ptr.
@ -125,11 +125,6 @@ public:
mcl::bit::replicate_element<u64>(esize, value)); mcl::bit::replicate_element<u64>(esize, value));
} }
/// Far code sits far away from the near code. Execution remains primarily in near code.
/// "Cold" / Rarely executed instructions sit in far code, so the CPU doesn't fetch them unless necessary.
void SwitchToFarCode();
void SwitchToNearCode();
CodePtr GetCodeBegin() const; CodePtr GetCodeBegin() const;
size_t GetTotalCodeSize() const; size_t GetTotalCodeSize() const;
@ -180,18 +175,12 @@ public:
private: private:
RunCodeCallbacks cb; RunCodeCallbacks cb;
JitStateInfo jsi; JitStateInfo jsi;
size_t far_code_offset;
bool prelude_complete = false; bool prelude_complete = false;
CodePtr near_code_begin = nullptr; CodePtr code_begin = nullptr;
CodePtr far_code_begin = nullptr;
ConstantPool constant_pool; ConstantPool constant_pool;
bool in_far_code = false;
CodePtr near_code_ptr;
CodePtr far_code_ptr;
using RunCodeFuncType = HaltReason (*)(void*, CodePtr); using RunCodeFuncType = HaltReason (*)(void*, CodePtr);
RunCodeFuncType run_code = nullptr; RunCodeFuncType run_code = nullptr;
RunCodeFuncType step_code = nullptr; RunCodeFuncType step_code = nullptr;

View file

@ -276,11 +276,8 @@ Xbyak::Label EmitX64::EmitCond(IR::Cond cond) {
return pass; return pass;
} }
EmitX64::BlockDescriptor EmitX64::RegisterBlock(const IR::LocationDescriptor& descriptor, CodePtr entrypoint, CodePtr entrypoint_far, size_t size) { EmitX64::BlockDescriptor EmitX64::RegisterBlock(const IR::LocationDescriptor& descriptor, CodePtr entrypoint, size_t size) {
PerfMapRegister(entrypoint, code.getCurr(), LocationDescriptorToFriendlyName(descriptor)); PerfMapRegister(entrypoint, code.getCurr(), LocationDescriptorToFriendlyName(descriptor));
code.SwitchToFarCode();
PerfMapRegister(entrypoint_far, code.getCurr(), LocationDescriptorToFriendlyName(descriptor) + "_far");
code.SwitchToNearCode();
Patch(descriptor, entrypoint); Patch(descriptor, entrypoint);
BlockDescriptor block_desc{entrypoint, size}; BlockDescriptor block_desc{entrypoint, size};

View file

@ -105,7 +105,7 @@ protected:
virtual std::string LocationDescriptorToFriendlyName(const IR::LocationDescriptor&) const = 0; virtual std::string LocationDescriptorToFriendlyName(const IR::LocationDescriptor&) const = 0;
void EmitAddCycles(size_t cycles); void EmitAddCycles(size_t cycles);
Xbyak::Label EmitCond(IR::Cond cond); Xbyak::Label EmitCond(IR::Cond cond);
BlockDescriptor RegisterBlock(const IR::LocationDescriptor& location_descriptor, CodePtr entrypoint, CodePtr entrypoint_far, size_t size); BlockDescriptor RegisterBlock(const IR::LocationDescriptor& location_descriptor, CodePtr entrypoint, size_t size);
void PushRSBHelper(Xbyak::Reg64 loc_desc_reg, Xbyak::Reg64 index_reg, IR::LocationDescriptor target); void PushRSBHelper(Xbyak::Reg64 loc_desc_reg, Xbyak::Reg64 index_reg, IR::LocationDescriptor target);
// Terminal instruction emitters // Terminal instruction emitters

View file

@ -229,9 +229,6 @@ struct UserConfig {
// Minimum size is about 8MiB. Maximum size is about 2GiB. Maximum size is limited by // Minimum size is about 8MiB. Maximum size is about 2GiB. Maximum size is limited by
// the maximum length of a x64 jump. // the maximum length of a x64 jump.
size_t code_cache_size = 256 * 1024 * 1024; // bytes size_t code_cache_size = 256 * 1024 * 1024; // bytes
// Determines the relative size of the near and far code caches. Must be smaller than
// code_cache_size.
size_t far_code_offset = 200 * 1024 * 1024; // bytes
}; };
} // namespace A32 } // namespace A32

View file

@ -288,9 +288,6 @@ struct UserConfig {
// Minimum size is about 8MiB. Maximum size is about 2GiB. Maximum size is limited by // Minimum size is about 8MiB. Maximum size is about 2GiB. Maximum size is limited by
// the maximum length of a x64 jump. // the maximum length of a x64 jump.
size_t code_cache_size = 256 * 1024 * 1024; // bytes size_t code_cache_size = 256 * 1024 * 1024; // bytes
// Determines the relative size of the near and far code caches. Must be smaller than
// code_cache_size.
size_t far_code_offset = 200 * 1024 * 1024; // bytes
}; };
} // namespace A64 } // namespace A64