Skip to content

Commit 660a37c

Browse files
elad335elad335
committed
g
1 parent b38ce94 commit 660a37c

File tree

3 files changed

+164
-6
lines changed

3 files changed

+164
-6
lines changed

‎rpcs3/Emu/Cell/PPUAnalyser.cpp‎

Lines changed: 132 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -2078,6 +2078,138 @@ bool ppu_module::analyse(u32 lib_toc, u32 entry, const u32 sec_end, const std::b
20782078
}
20792079

20802080
ppu_log.notice("Block analysis: %zu blocks (%zu enqueued)", funcs.size(), block_queue.size());
2081+
2082+
std::unordered_map<std::string_view, std::pair<u32, u32>> duplicate_data_map;
2083+
duplicate_map.clear();
2084+
2085+
for (auto& func : funcs)
2086+
{
2087+
if (func.size == 0 || func.size > 10000u)
2088+
{
2089+
continue;
2090+
}
2091+
2092+
auto& data = duplicate_data_map[std::string_view{get_ptr<char>(func.addr), func.size}];
2093+
2094+
const usz count = data.first;
2095+
2096+
if (!count)
2097+
{
2098+
data.first++;
2099+
data.second = func.addr;
2100+
continue;
2101+
}
2102+
2103+
if (!data.second)
2104+
{
2105+
continue;
2106+
}
2107+
2108+
if (count == 1)
2109+
{
2110+
const u32 faddr = func.addr;
2111+
const u32 fend = func.addr + func.size;
2112+
2113+
bool fail = false;
2114+
2115+
//for (const auto [addr, size] : func.blocks)
2116+
const u32 addr = func.addr;
2117+
const u32 size = func.size;
2118+
{
2119+
if (size == 0)
2120+
{
2121+
continue;
2122+
}
2123+
2124+
auto i_ptr = ensure(get_ptr<u32>(addr));
2125+
2126+
for (u32 i = addr; i < fend; i += 4, i_ptr++)
2127+
{
2128+
/ const ppu_opcode_t op{*i_ptr};
2129+
const auto itype = s_ppu_itype.decode(op.opcode);
2130+
2131+
if (itype != ppu_itype::BC && itype != ppu_itype::B)
2132+
{
2133+
if (i == fend - 4)
2134+
{
2135+
if (!(itype & ppu_itype::branch))
2136+
{
2137+
// Inserts a branch to following code
2138+
fail = true;
2139+
break;
2140+
}
2141+
}
2142+
2143+
continue;
2144+
}
2145+
2146+
if (!op.aa)
2147+
{
2148+
fail = true;
2149+
break;
2150+
}
2151+
2152+
if (itype == ppu_itype::BC && (op.bo & 0x14) != 0x14)
2153+
{
2154+
if (i == fend - 4)
2155+
{
2156+
// Can branch to next
2157+
fail = true;
2158+
break;
2159+
}
2160+
}
2161+
}
2162+
}
2163+
2164+
if (fail)
2165+
{
2166+
data.first = 1;
2167+
data.second = 0;
2168+
continue;
2169+
}
2170+
}
2171+
2172+
data.first++;
2173+
2174+
// Choose the lowest function as the source
2175+
data.second = std::min<u32>(data.second, func.addr);
2176+
}
2177+
2178+
usz dups_count = 0;
2179+
2180+
for (auto& func : funcs)
2181+
{
2182+
if (func.size == 0 || func.size > 10000u)
2183+
{
2184+
continue;
2185+
}
2186+
2187+
const auto data = ::at32(duplicate_data_map, std::string_view{get_ptr<char>(func.addr), func.size});
2188+
2189+
if (data.first > 1)
2190+
{
2191+
duplicate_map[func.addr] = data.second;
2192+
2193+
for (const auto [addr, size] : func.blocks)
2194+
{
2195+
if (size == 0 || addr >= func.addr + func.size)
2196+
{
2197+
continue;
2198+
}
2199+
2200+
duplicate_map[addr] = data.second + (addr - func.addr);
2201+
}
2202+
2203+
if (func.addr != data.second)
2204+
{
2205+
dups_count++;
2206+
}
2207+
2208+
ppu_log.trace("Found PPU function duplicate: func 0x%x vs 0x%x (%d times) (size=%d)", func.addr, data.second, data.first, func.size);
2209+
}
2210+
}
2211+
2212+
ppu_log.success("Function duplication count: %d/%d (%g%)", dups_count, duplicate_data_map.size(), dups_count * 100.0 / duplicate_data_map.size());
20812213
return true;
20822214
}
20832215

‎rpcs3/Emu/Cell/PPUAnalyser.h‎

Lines changed: 2 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -95,6 +95,7 @@ struct ppu_module
9595
std::vector<ppu_function> funcs{};
9696
std::deque<std::shared_ptr<void>> allocations;
9797
std::map<u32, u32> addr_to_seg_index;
98+
std::unordered_map<u32, u32> duplicate_map;
9899

99100
// Copy info without functions
100101
void copy_part(const ppu_module& info)
@@ -107,6 +108,7 @@ struct ppu_module
107108
secs = info.secs;
108109
allocations = info.allocations;
109110
addr_to_seg_index = info.addr_to_seg_index;
111+
duplicate_map = info.duplicate_map;
110112
}
111113

112114
bool analyse(u32 lib_toc, u32 entry, u32 end, const std::basic_string<u32>& applied, std::function<bool()> check_aborted = {});

‎rpcs3/Emu/Cell/PPUThread.cpp‎

Lines changed: 30 additions & 6 deletions
Original file line numberDiff line numberDiff line change
@@ -3918,13 +3918,13 @@ extern void ppu_precompile(std::vector<std::string>& dir_queue, std::vector<ppu_
39183918
mself_header hdr{};
39193919

39203920
if (mself.read(hdr) && hdr.get_count(mself.size()))
3921-
{
3921+
{
3922+
std::set<u64> offs;
3923+
39223924
for (u32 j = 0; j < hdr.count; j++)
39233925
{
39243926
mself_record rec{};
39253927

3926-
std::set<u64> offs;
3927-
39283928
if (mself.read(rec) && rec.get_pos(mself.size()))
39293929
{
39303930
if (rec.size <= 0x20)
@@ -4165,7 +4165,7 @@ extern void ppu_precompile(std::vector<std::string>& dir_queue, std::vector<ppu_
41654165
}
41664166

41674167
ppu_log.notice("Failed to precompile '%s' (prx: %s, ovl: %s): Attempting compilation as executable file", path, prx_err, ovl_err);
4168-
possible_exec_file_paths.push(path, offset, file_size);
4168+
possible_exec_file_paths.push(file_queue[func_i]);
41694169
inc_fdone = 0;
41704170
}
41714171

@@ -4680,8 +4680,15 @@ bool ppu_initialize(const ppu_module& info, bool check_only, u64 file_size)
46804680
// Copy block or function entry
46814681
ppu_function& entry = part.funcs.emplace_back(func);
46824682

4683+
u32 og_func = entry.addr;
4684+
4685+
if (auto it = info.duplicate_map.find(entry.addr); it != info.duplicate_map.end())
4686+
{
4687+
og_func = it->second;
4688+
}
4689+
46834690
// Fixup some information
4684-
entry.name = fmt::format("__0x%x", entry.addr - reloc);
4691+
entry.name = fmt::format("__0x%x", og_func - reloc);
46854692

46864693
if (has_mfvscr && g_cfg.core.ppu_set_sat_bit)
46874694
{
@@ -4848,7 +4855,7 @@ bool ppu_initialize(const ppu_module& info, bool check_only, u64 file_size)
48484855
settings += ppu_settings::contains_symbol_resolver; // Avoid invalidating all modules for this purpose
48494856

48504857
// Write version, hash, CPU, settings
4851-
fmt::append(obj_name, "v6-kusa-%s-%s-%s.obj", fmt::base57(output, 16), fmt::base57(settings), jit_compiler::cpu(g_cfg.core.llvm_cpu));
4858+
fmt::append(obj_name, "v7-kusa-%s-%s-%s.obj", fmt::base57(output, 16), fmt::base57(settings), jit_compiler::cpu(g_cfg.core.llvm_cpu));
48524859
}
48534860

48544861
if (cpu ? cpu->state.all_of(cpu_flag::exit) : Emu.IsStopped())
@@ -5086,6 +5093,9 @@ bool ppu_initialize(const ppu_module& info, bool check_only, u64 file_size)
50865093
jit_mod.symbol_resolver(vm::g_exec_addr, info.segs[0].addr);
50875094

50885095
// Find a BLR-only function in order to copy it to all BLRs (some games need it)
5096+
bool early_exit = false;
5097+
5098+
// Get and install function addresses
50895099
for (const auto& func : info.funcs)
50905100
{
50915101
if (func.size == 4 && *info.get_ptr<u32>(func.addr) == ppu_instructions::BLR())
@@ -5156,6 +5166,11 @@ static void ppu_initialize2(jit_compiler& jit, const ppu_module& module_part, co
51565166
{
51575167
if (func.size)
51585168
{
5169+
if (auto it = module_part.duplicate_map.find(func.addr); it != module_part.duplicate_map.end() && it->second != it->first)
5170+
{
5171+
continue;
5172+
}
5173+
51595174
const auto f = cast<Function>(_module->getOrInsertFunction(func.name, _func).getCallee());
51605175
f->setCallingConv(CallingConv::GHC);
51615176
f->addParamAttr(1, llvm::Attribute::NoAlias);
@@ -5229,6 +5244,15 @@ static void ppu_initialize2(jit_compiler& jit, const ppu_module& module_part, co
52295244

52305245
if (module_part.funcs[fi].size)
52315246
{
5247+
const u32 faddr = module_part.funcs[fi].addr;
5248+
auto it = module_part.duplicate_map.find(faddr);
5249+
5250+
if (it != module_part.duplicate_map.end() && it->second != faddr)
5251+
{
5252+
ppu_log.trace("LLVM: Function 0x%x was skipped (duplicate)", faddr);
5253+
continue;
5254+
}
5255+
52325256
// Translate
52335257
if (const auto func = translator.Translate(module_part.funcs[fi]))
52345258
{

0 commit comments

Comments
 (0)