Exverge/suyu
3
0
Fork 0
forked from suyu/suyu
Code Pull requests Activity

Merge pull request #4089 from MerryMage/macrojit-cleanup-1

Browse source 
macro_jit_x64: Cleanup
This commit is contained in:
bunnei 2020-06-16 23:44:48 -04:00 committed by GitHub
commit b660ef6c8a
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
2 changed files with 18 additions and 51 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

64
src/video_core/macro/macro_jit_x64.cpp
View file

@ -14,22 +14,16 @@ MICROPROFILE_DEFINE(MacroJitCompile, "GPU", "Compile macro JIT", MP_RGB(173, 255
MICROPROFILE_DEFINE(MacroJitExecute, "GPU", "Execute macro JIT", MP_RGB(255, 255, 0));
namespace Tegra {
static const Xbyak::Reg64 PARAMETERS = Xbyak::util::r9;
static const Xbyak::Reg64 REGISTERS = Xbyak::util::r10;
static const Xbyak::Reg64 STATE = Xbyak::util::r11;
static const Xbyak::Reg64 NEXT_PARAMETER = Xbyak::util::r12;
static const Xbyak::Reg32 RESULT = Xbyak::util::r13d;
static const Xbyak::Reg64 RESULT_64 = Xbyak::util::r13;
static const Xbyak::Reg64 STATE = Xbyak::util::rbx;
static const Xbyak::Reg32 RESULT = Xbyak::util::ebp;
static const Xbyak::Reg64 PARAMETERS = Xbyak::util::r12;
static const Xbyak::Reg32 METHOD_ADDRESS = Xbyak::util::r14d;
static const Xbyak::Reg64 METHOD_ADDRESS_64 = Xbyak::util::r14;
static const Xbyak::Reg64 BRANCH_HOLDER = Xbyak::util::r15;
static const std::bitset<32> PERSISTENT_REGISTERS = Common::X64::BuildRegSet({
PARAMETERS,
REGISTERS,
STATE,
NEXT_PARAMETER,
RESULT,
PARAMETERS,
METHOD_ADDRESS,
BRANCH_HOLDER,
});
@ -53,8 +47,7 @@ void MacroJITx64Impl::Execute(const std::vector<u32>& parameters, u32 method) {
JITState state{};
state.maxwell3d = &maxwell3d;
state.registers = {};
state.parameters = parameters.data();
program(&state);
program(&state, parameters.data());
}
void MacroJITx64Impl::Compile_ALU(Macro::Opcode opcode) {
@ -64,18 +57,18 @@ void MacroJITx64Impl::Compile_ALU(Macro::Opcode opcode) {
const bool is_move_operation = !is_a_zero && is_b_zero;
const bool has_zero_register = is_a_zero || is_b_zero;
Xbyak::Reg64 src_a;
Xbyak::Reg32 src_a;
Xbyak::Reg32 src_b;
if (!optimizer.zero_reg_skip) {
src_a = Compile_GetRegister(opcode.src_a, RESULT_64);
src_b = Compile_GetRegister(opcode.src_b, ebx);
src_a = Compile_GetRegister(opcode.src_a, RESULT);
src_b = Compile_GetRegister(opcode.src_b, eax);
} else {
if (!is_a_zero) {
src_a = Compile_GetRegister(opcode.src_a, RESULT_64);
src_a = Compile_GetRegister(opcode.src_a, RESULT);
}
if (!is_b_zero) {
src_b = Compile_GetRegister(opcode.src_b, ebx);
src_b = Compile_GetRegister(opcode.src_b, eax);
}
}
Xbyak::Label skip_carry{};
@ -329,7 +322,7 @@ void Tegra::MacroJITx64Impl::Compile_Send(Xbyak::Reg32 value) {
and_(METHOD_ADDRESS, 0xfff);
shr(ecx, 12);
and_(ecx, 0x3f);
lea(eax, ptr[rcx + METHOD_ADDRESS_64]);
lea(eax, ptr[rcx + METHOD_ADDRESS.cvt64()]);
sal(ecx, 12);
or_(eax, ecx);
@ -424,16 +417,12 @@ void MacroJITx64Impl::Compile() {
Common::X64::ABI_PushRegistersAndAdjustStackGPS(*this, Common::X64::ABI_ALL_CALLEE_SAVED, 8);
// JIT state
mov(STATE, Common::X64::ABI_PARAM1);
mov(PARAMETERS, qword[Common::X64::ABI_PARAM1 +
static_cast<Xbyak::uint32>(offsetof(JITState, parameters))]);
mov(REGISTERS, Common::X64::ABI_PARAM1);
add(REGISTERS, static_cast<Xbyak::uint32>(offsetof(JITState, registers)));
mov(PARAMETERS, Common::X64::ABI_PARAM2);
xor_(RESULT, RESULT);
xor_(METHOD_ADDRESS, METHOD_ADDRESS);
xor_(NEXT_PARAMETER, NEXT_PARAMETER);
xor_(BRANCH_HOLDER, BRANCH_HOLDER);
mov(dword[REGISTERS + 4], Compile_FetchParameter());
mov(dword[STATE + offsetof(JITState, registers) + 4], Compile_FetchParameter());
// Track get register for zero registers and mark it as no-op
optimizer.zero_reg_skip = true;
@ -537,8 +526,8 @@ bool MacroJITx64Impl::Compile_NextInstruction() {
}
Xbyak::Reg32 Tegra::MacroJITx64Impl::Compile_FetchParameter() {
mov(eax, dword[PARAMETERS + NEXT_PARAMETER * sizeof(u32)]);
inc(NEXT_PARAMETER);
mov(eax, dword[PARAMETERS]);
add(PARAMETERS, sizeof(u32));
return eax;
}
@ -547,31 +536,12 @@ Xbyak::Reg32 MacroJITx64Impl::Compile_GetRegister(u32 index, Xbyak::Reg32 dst) {
// Register 0 is always zero
xor_(dst, dst);
} else {
mov(dst, dword[REGISTERS + index * sizeof(u32)]);
mov(dst, dword[STATE + offsetof(JITState, registers) + index * sizeof(u32)]);
}
return dst;
}
Xbyak::Reg64 Tegra::MacroJITx64Impl::Compile_GetRegister(u32 index, Xbyak::Reg64 dst) {
if (index == 0) {
// Register 0 is always zero
xor_(dst, dst);
} else {
mov(dst, dword[REGISTERS + index * sizeof(u32)]);
}
return dst;
}
void Tegra::MacroJITx64Impl::Compile_WriteCarry(Xbyak::Reg64 dst) {
Xbyak::Label zero{}, end{};
xor_(ecx, ecx);
shr(dst, 32);
setne(cl);
mov(dword[STATE + offsetof(JITState, carry_flag)], ecx);
}
void MacroJITx64Impl::Compile_ProcessResult(Macro::ResultOperation operation, u32 reg) {
auto SetRegister = [=](u32 reg, Xbyak::Reg32 result) {
// Register 0 is supposed to always return 0. NOP is implemented as a store to the zero
@ -579,7 +549,7 @@ void MacroJITx64Impl::Compile_ProcessResult(Macro::ResultOperation operation, u3
if (reg == 0) {
return;
}
mov(dword[REGISTERS + reg * sizeof(u32)], result);
mov(dword[STATE + offsetof(JITState, registers) + reg * sizeof(u32)], result);
};
auto SetMethodAddress = [=](Xbyak::Reg32 reg) { mov(METHOD_ADDRESS, reg); };

5
src/video_core/macro/macro_jit_x64.h
View file

@ -55,8 +55,6 @@ private:
Xbyak::Reg32 Compile_FetchParameter();
Xbyak::Reg32 Compile_GetRegister(u32 index, Xbyak::Reg32 dst);
Xbyak::Reg64 Compile_GetRegister(u32 index, Xbyak::Reg64 dst);
void Compile_WriteCarry(Xbyak::Reg64 dst);
void Compile_ProcessResult(Macro::ResultOperation operation, u32 reg);
void Compile_Send(Xbyak::Reg32 value);
@ -67,11 +65,10 @@ private:
struct JITState {
Engines::Maxwell3D* maxwell3d{};
std::array<u32, Macro::NUM_MACRO_REGISTERS> registers{};
const u32* parameters{};
u32 carry_flag{};
};
static_assert(offsetof(JITState, maxwell3d) == 0, "Maxwell3D is not at 0x0");
using ProgramType = void (*)(JITState*);
using ProgramType = void (*)(JITState*, const u32*);
struct OptimizerState {
bool can_skip_carry{};