archive/rpcs3
1
0
Fork 0
mirror of https://github.com/RPCS3/rpcs3.git synced 2025-07-09 00:11:24 +12:00
Code Issues Projects Releases Packages Wiki Activity Actions

Transactions: move loops inside

Browse source 
Rewrite loops in assembler (minor optimization)
This commit is contained in:
Nekotekina 2018-05-18 18:51:48 +03:00
parent f6f45b8699
commit 33a1c743a4
4 changed files with 81 additions and 106 deletions
Download patch file Download diff file Expand all files Collapse all files

93
rpcs3/Emu/Cell/PPUThread.cpp
View file

@ -944,8 +944,23 @@ static T ppu_load_acquire_reservation(ppu_thread& ppu, u32 addr)
ppu.raddr = addr;
while (g_use_rtm)
{
ppu.rtime = vm::reservation_acquire(addr, sizeof(T));
ppu.rdata = data;
if (LIKELY(vm::reservation_acquire(addr, sizeof(T)) == ppu.rtime))
{
return static_cast<T>(ppu.rdata);
}
else
{
_mm_pause();
}
}
// Do several attemps
for (uint i = 0; g_use_rtm || i < 5; i++)
for (uint i = 0; i < 5; i++)
{
ppu.rtime = vm::reservation_acquire(addr, sizeof(T));
_mm_lfence();
@ -981,7 +996,7 @@ extern u64 ppu_ldarx(ppu_thread& ppu, u32 addr)
return ppu_load_acquire_reservation<u64>(ppu, addr);
}
const auto ppu_stwcx_tx = build_function_asm<int(*)(u32 raddr, u64 rtime, u64 rdata, u32 value)>([](asmjit::X86Assembler& c, auto& args)
const auto ppu_stwcx_tx = build_function_asm<bool(*)(u32 raddr, u64 rtime, u64 rdata, u32 value)>([](asmjit::X86Assembler& c, auto& args)
{
using namespace asmjit;
@ -998,9 +1013,10 @@ const auto ppu_stwcx_tx = build_function_asm<int(*)(u32 raddr, u64 rtime, u64 rd
c.lea(x86::r10, x86::qword_ptr(x86::r10, args[0], 3));
c.bswap(args[2].r32());
c.bswap(args[3].r32());
c.mov(args[0].r32(), 5);
// Begin transaction
build_transaction_enter(c, fall);
Label begin = build_transaction_enter(c, fall);
c.cmp(x86::qword_ptr(x86::r10), args[1]);
c.jne(fail);
c.cmp(x86::dword_ptr(x86::r11), args[2].r32());
@ -1017,14 +1033,18 @@ const auto ppu_stwcx_tx = build_function_asm<int(*)(u32 raddr, u64 rtime, u64 rd
// Touch memory after transaction failure
c.bind(fall);
c.sub(args[0].r32(), 1);
c.jz(fail);
c.sar(x86::eax, 24);
c.js(fail);
c.lock().add(x86::dword_ptr(x86::r11), 0);
c.lock().add(x86::qword_ptr(x86::r10), 0);
c.sar(x86::eax, 24);
c.ret();
c.jmp(begin);
c.bind(fail);
build_transaction_abort(c, 0xff);
c.int3();
c.xor_(x86::eax, x86::eax);
c.ret();
});
extern bool ppu_stwcx(ppu_thread& ppu, u32 addr, u32 reg_value)
@ -1039,27 +1059,14 @@ extern bool ppu_stwcx(ppu_thread& ppu, u32 addr, u32 reg_value)
if (g_use_rtm)
{
// Do several attempts (TODO)
for (u32 i = 0; i < 5; i++)
if (ppu_stwcx_tx(addr, ppu.rtime, ppu.rdata, reg_value))
{
const int r = ppu_stwcx_tx(addr, ppu.rtime, ppu.rdata, reg_value);
if (r > 0)
{
vm::reservation_notifier(addr, sizeof(u32)).notify_all();
ppu.raddr = 0;
return true;
}
if (r < 0)
{
// Reservation lost
ppu.raddr = 0;
return false;
}
vm::reservation_notifier(addr, sizeof(u32)).notify_all();
ppu.raddr = 0;
return true;
}
// Give up
// Reservation lost
ppu.raddr = 0;
return false;
}
@ -1078,7 +1085,7 @@ extern bool ppu_stwcx(ppu_thread& ppu, u32 addr, u32 reg_value)
return result;
}
const auto ppu_stdcx_tx = build_function_asm<int(*)(u32 raddr, u64 rtime, u64 rdata, u64 value)>([](asmjit::X86Assembler& c, auto& args)
const auto ppu_stdcx_tx = build_function_asm<bool(*)(u32 raddr, u64 rtime, u64 rdata, u64 value)>([](asmjit::X86Assembler& c, auto& args)
{
using namespace asmjit;
@ -1095,9 +1102,10 @@ const auto ppu_stdcx_tx = build_function_asm<int(*)(u32 raddr, u64 rtime, u64 rd
c.lea(x86::r10, x86::qword_ptr(x86::r10, args[0], 3));
c.bswap(args[2]);
c.bswap(args[3]);
c.mov(args[0].r32(), 5);
// Begin transaction
build_transaction_enter(c, fall);
Label begin = build_transaction_enter(c, fall);
c.cmp(x86::qword_ptr(x86::r10), args[1]);
c.jne(fail);
c.cmp(x86::qword_ptr(x86::r11), args[2]);
@ -1114,14 +1122,18 @@ const auto ppu_stdcx_tx = build_function_asm<int(*)(u32 raddr, u64 rtime, u64 rd
// Touch memory after transaction failure
c.bind(fall);
c.sub(args[0].r32(), 1);
c.jz(fail);
c.sar(x86::eax, 24);
c.js(fail);
c.lock().add(x86::qword_ptr(x86::r11), 0);
c.lock().add(x86::qword_ptr(x86::r10), 0);
c.sar(x86::eax, 24);
c.ret();
c.jmp(begin);
c.bind(fail);
build_transaction_abort(c, 0xff);
c.int3();
c.xor_(x86::eax, x86::eax);
c.ret();
});
extern bool ppu_stdcx(ppu_thread& ppu, u32 addr, u64 reg_value)
@ -1136,27 +1148,14 @@ extern bool ppu_stdcx(ppu_thread& ppu, u32 addr, u64 reg_value)
if (g_use_rtm)
{
// Do several attempts (TODO)
for (u32 i = 0; i < 5; i++)
if (ppu_stdcx_tx(addr, ppu.rtime, ppu.rdata, reg_value))
{
const int r = ppu_stdcx_tx(addr, ppu.rtime, ppu.rdata, reg_value);
if (r > 0)
{
vm::reservation_notifier(addr, sizeof(u64)).notify_all();
ppu.raddr = 0;
return true;
}
if (r < 0)
{
// Reservation lost
ppu.raddr = 0;
return false;
}
vm::reservation_notifier(addr, sizeof(u64)).notify_all();
ppu.raddr = 0;
return true;
}
// Give up
// Reservation lost
ppu.raddr = 0;
return false;
}