RPCS3/rpcs3
1
0
Fork 0
mirror of https://github.com/RPCS3/rpcs3.git synced 2025-12-06 07:12:28 +01:00
Code Issues Actions 1 Packages Projects Releases Wiki Activity

SPU: Tame PUTLLC16

Browse source
This commit is contained in:
Elad 2025-11-30 15:11:16 +02:00
parent 7001940483
commit d822d85ea1
1 changed files with 39 additions and 6 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

43
rpcs3/Emu/Cell/SPUCommonRecompiler.cpp
View file

@ -4973,6 +4973,7 @@ spu_program spu_recompiler_base::analyse(const be_t<u32>* ls, u32 entry_point, s
u32 lsa_last_pc = SPU_LS_SIZE; // PC of first LSA write u32 lsa_last_pc = SPU_LS_SIZE; // PC of first LSA write
u32 get_pc = SPU_LS_SIZE; // PC of GETLLAR u32 get_pc = SPU_LS_SIZE; // PC of GETLLAR
u32 put_pc = SPU_LS_SIZE; // PC of PUTLLC u32 put_pc = SPU_LS_SIZE; // PC of PUTLLC
u32 rdatomic_pc = SPU_LS_SIZE; // PC of last RdAtomcStat read
reg_state_t ls{}; // state of LS load/store address register reg_state_t ls{}; // state of LS load/store address register
reg_state_t ls_offs = reg_state_t::from_value(0); // Added value to ls reg_state_t ls_offs = reg_state_t::from_value(0); // Added value to ls
reg_state_t lsa{}; // state of LSA register on GETLLAR reg_state_t lsa{}; // state of LSA register on GETLLAR
@ -6333,6 +6334,8 @@ spu_program spu_recompiler_base::analyse(const be_t<u32>* ls, u32 entry_point, s
break; break;
} }
atomic16->rdatomic_pc = pos;
const auto it = atomic16_all.find(pos); const auto it = atomic16_all.find(pos);
if (it == atomic16_all.end()) if (it == atomic16_all.end())
@ -7273,7 +7276,7 @@ spu_program spu_recompiler_base::analyse(const be_t<u32>* ls, u32 entry_point, s
for (const auto& [pc_commited, pattern] : atomic16_all) for (const auto& [pc_commited, pattern] : atomic16_all)
{ {
if (!pattern.active) if (!pattern.active || pattern.lsa_pc >= pattern.rdatomic_pc)
{ {
continue; continue;
} }
@ -7283,6 +7286,17 @@ spu_program spu_recompiler_base::analyse(const be_t<u32>* ls, u32 entry_point, s
continue; continue;
} }
std::string pattern_hash;
{
sha1_context ctx;
u8 output[20]{};
sha1_starts(&ctx);
sha1_update(&ctx, reinterpret_cast<const u8*>(result.data.data()) + (pattern.lsa_pc - result.lower_bound), pattern.rdatomic_pc - pattern.lsa_pc);
sha1_finish(&ctx, output);
fmt::append(pattern_hash, "%s", fmt::base57(output));
}
union putllc16_or_0_info union putllc16_or_0_info
{ {
u64 data; u64 data;
@ -7373,16 +7387,35 @@ spu_program spu_recompiler_base::analyse(const be_t<u32>* ls, u32 entry_point, s
value.reg2 = pattern.reg2; value.reg2 = pattern.reg2;
} }
bool allow_pattern = true;
if (g_cfg.core.spu_accurate_reservations) if (g_cfg.core.spu_accurate_reservations)
{ {
// Because enabling it is a hack, as it turns out // The problem with PUTLLC16 optimization, that it is in theory correct at the bounds of the spu function.
// continue; // But if the SPU code reuses the cache line data observed, it is not truly atomic.
// So we may enable it only for known cases where SPU atomic data is not used after the function leaves.
// So the two options are:
// 1. Atomic compare exchange 16 bytes operation. (rest of data is not read) -> good for RPCS3 to optimize.
// 2. Fetch 128 bytes (read them later), modify only 16 bytes. -> Bad for RPCS3 to optimize.
// This difference cannot be known at analyzer time but from observing callers.
static constexpr std::initializer_list<std::string_view> allowed_patterns =
{
"620oYSe8uQqq9eTkhWfMqoEXX0us"sv, // CellSpurs JobChain acquire pattern
};
allow_pattern = std::any_of(allowed_patterns.begin(), allowed_patterns.end(), FN(pattern_hash == x));
} }
if (allow_pattern)
{
add_pattern(false, inst_attr::putllc16, pattern.put_pc - result.entry_point, value.data); add_pattern(false, inst_attr::putllc16, pattern.put_pc - result.entry_point, value.data);
}
spu_log.success("PUTLLC16 Pattern Detected! (mem_count=%d, put_pc=0x%x, pc_rel=%d, offset=0x%x, const=%u, two_regs=%d, reg=%u, runtime=%d, 0x%x-%s) (putllc0=%d, putllc16+0=%d, all=%d)" spu_log.success("PUTLLC16 Pattern Detected! (mem_count=%d, put_pc=0x%x, pc_rel=%d, offset=0x%x, const=%u, two_regs=%d, reg=%u, runtime=%d, 0x%x-%s, pattern-hash=%s) (putllc0=%d, putllc16+0=%d, all=%d)"
, pattern.mem_count, pattern.put_pc, value.type == v_relative, value.off18, value.type == v_const, value.type == v_reg2, value.reg, value.runtime16_select, entry_point, func_hash, +stats.nowrite, ++stats.single, +stats.all); , pattern.mem_count, pattern.put_pc, value.type == v_relative, value.off18, value.type == v_const, value.type == v_reg2, value.reg, value.runtime16_select, entry_point, func_hash, pattern_hash, +stats.nowrite, ++stats.single, +stats.all);
} }
for (const auto& [read_pc, pattern] : rchcnt_loop_all) for (const auto& [read_pc, pattern] : rchcnt_loop_all)