Update SPU cache (v2)

Improve SPU analyser: filter unreachable fragments
More strict NOP/LNOP analysis
Fill block predecessors info
ASMJIT: fix assertion and improve indirect branch

rpcs3/Emu/Cell/SPUASMJITRecompiler.cpp

@@ -1156,11 +1156,11 @@ void spu_recompiler::branch_fixed(u32 target)
 	c->jmp(x86::rax);
 }
 
-void spu_recompiler::branch_indirect(spu_opcode_t op, bool jt)
+void spu_recompiler::branch_indirect(spu_opcode_t op, bool local)
 {
 	using namespace asmjit;
 
-	if (g_cfg.core.spu_block_size == spu_block_size_type::safe && !jt)
+	if (g_cfg.core.spu_block_size == spu_block_size_type::safe && !local)
 	{
 		// Simply external call (return or indirect call)
 		c->mov(x86::r10, x86::qword_ptr(*cpu, addr->r64(), 1, offset32(&SPUThread::jit_dispatcher)));
@@ -2660,13 +2660,13 @@ void spu_recompiler::BIZ(spu_opcode_t op)
 	c->cmp(SPU_OFF_32(gpr, op.rt, &v128::_u32, 3), 0);
 	c->je(branch_label);
 
-	after.emplace_back([=]
+	after.emplace_back([=, jt = m_targets[m_pos].size() > 1]
 	{
 		c->align(asmjit::kAlignCode, 16);
 		c->bind(branch_label);
 		c->mov(*addr, SPU_OFF_32(gpr, op.ra, &v128::_u32, 3));
 		c->and_(*addr, 0x3fffc);
-		branch_indirect(op);
+		branch_indirect(op, jt);
 	});
 }
 
@@ -2676,13 +2676,13 @@ void spu_recompiler::BINZ(spu_opcode_t op)
 	c->cmp(SPU_OFF_32(gpr, op.rt, &v128::_u32, 3), 0);
 	c->jne(branch_label);
 
-	after.emplace_back([=]
+	after.emplace_back([=, jt = m_targets[m_pos].size() > 1]
 	{
 		c->align(asmjit::kAlignCode, 16);
 		c->bind(branch_label);
 		c->mov(*addr, SPU_OFF_32(gpr, op.ra, &v128::_u32, 3));
 		c->and_(*addr, 0x3fffc);
-		branch_indirect(op);
+		branch_indirect(op, jt);
 	});
 }
 
@@ -2692,13 +2692,13 @@ void spu_recompiler::BIHZ(spu_opcode_t op)
 	c->cmp(SPU_OFF_16(gpr, op.rt, &v128::_u16, 6), 0);
 	c->je(branch_label);
 
-	after.emplace_back([=]
+	after.emplace_back([=, jt = m_targets[m_pos].size() > 1]
 	{
 		c->align(asmjit::kAlignCode, 16);
 		c->bind(branch_label);
 		c->mov(*addr, SPU_OFF_32(gpr, op.ra, &v128::_u32, 3));
 		c->and_(*addr, 0x3fffc);
-		branch_indirect(op);
+		branch_indirect(op, jt);
 	});
 }
 
@@ -2708,13 +2708,13 @@ void spu_recompiler::BIHNZ(spu_opcode_t op)
 	c->cmp(SPU_OFF_16(gpr, op.rt, &v128::_u16, 6), 0);
 	c->jne(branch_label);
 
-	after.emplace_back([=]
+	after.emplace_back([=, jt = m_targets[m_pos].size() > 1]
 	{
 		c->align(asmjit::kAlignCode, 16);
 		c->bind(branch_label);
 		c->mov(*addr, SPU_OFF_32(gpr, op.ra, &v128::_u32, 3));
 		c->and_(*addr, 0x3fffc);
-		branch_indirect(op);
+		branch_indirect(op, jt);
 	});
 }
 
@@ -2750,7 +2750,7 @@ void spu_recompiler::BI(spu_opcode_t op)
 {
 	c->mov(*addr, SPU_OFF_32(gpr, op.ra, &v128::_u32, 3));
 	c->and_(*addr, 0x3fffc);
-	branch_indirect(op, verify(HERE, m_targets[m_pos].size()) > 2);
+	branch_indirect(op, m_targets.find(m_pos) != m_targets.end());
 	m_pos = -1;
 }
 
@@ -2761,7 +2761,7 @@ void spu_recompiler::BISL(spu_opcode_t op)
 	const XmmLink& vr = XmmAlloc();
 	c->movdqa(vr, XmmConst(_mm_set_epi32(spu_branch_target(m_pos + 4), 0, 0, 0)));
 	c->movdqa(SPU_OFF_128(gpr, op.rt), vr);
-	branch_indirect(op);
+	branch_indirect(op, m_targets.find(m_pos) != m_targets.end());
 	m_pos = -1;
 }
 

rpcs3/Emu/Cell/SPUASMJITRecompiler.h

@@ -107,7 +107,7 @@ private:
 	asmjit::X86Mem XmmConst(__m128i data);
 
 	void branch_fixed(u32 target);
-	void branch_indirect(spu_opcode_t op, bool jt = false);
+	void branch_indirect(spu_opcode_t op, bool local = false);
 	void fall(spu_opcode_t op);
 	void save_rcx();
 	void load_rcx();

rpcs3/Emu/Cell/SPURecompiler.cpp

@@ -89,7 +89,7 @@ void spu_cache::initialize()
 	}
 
 	// SPU cache file (version + block size type)
-	const std::string loc = _main->cache + u8"spu-§" + fmt::to_lower(g_cfg.core.spu_block_size.to_string()) + "-v1.dat";
+	const std::string loc = _main->cache + u8"spu-§" + fmt::to_lower(g_cfg.core.spu_block_size.to_string()) + "-v2.dat";
 
 	auto cache = std::make_shared<spu_cache>(loc);
 
@@ -140,6 +140,11 @@ void spu_cache::initialize()
 			// Call analyser
 			std::vector<u32> func2 = compiler->block(ls.data(), func[0]);
 
+			if (func2.size() != func.size())
+			{
+				LOG_ERROR(SPU, "[0x%05x] SPU Analyser failed, %u vs %u", func2[0], func2.size() - 1, func.size() - 1);
+			}
+
 			compiler->compile(std::move(func));
 
 			// Clear fake LS
@@ -281,6 +286,7 @@ std::vector<u32> spu_recompiler_base::block(const be_t<u32>* ls, u32 lsa)
 
 	m_regmod.fill(0xff);
 	m_targets.clear();
+	m_preds.clear();
 
 	// Value flags (TODO)
 	enum class vf : u32
@@ -306,6 +312,8 @@ std::vector<u32> spu_recompiler_base::block(const be_t<u32>* ls, u32 lsa)
 			wi++;
 		};
 
+		const u32 pos = wl[wi];
+
 		const auto add_block = [&](u32 target)
 		{
 			// Verify validity of the new target (TODO)
@@ -316,12 +324,21 @@ std::vector<u32> spu_recompiler_base::block(const be_t<u32>* ls, u32 lsa)
 				{
 					m_block_info[target / 4] = true;
 					wl.push_back(target);
-					return;
 				}
+
+				// Add predecessor (check if already exists)
+				for (u32 pred : m_preds[target])
+				{
+					if (pred == pos)
+					{
+						return;
+					}
+				}
+
+				m_preds[target].push_back(pos);
 			}
 		};
 
-		const u32 pos = wl[wi];
 		const u32 data = ls[pos / 4];
 		const auto op = spu_opcode_t{data};
 
@@ -779,8 +796,33 @@ std::vector<u32> spu_recompiler_base::block(const be_t<u32>* ls, u32 lsa)
 		}
 	}
 
-	if (g_cfg.core.spu_block_size == spu_block_size_type::safe)
+	while (g_cfg.core.spu_block_size == spu_block_size_type::safe)
 	{
+		const u32 initial_size = result.size();
+
+		// Check unreachable blocks in safe mode (TODO)
+		u32 limit = lsa + result.size() * 4 - 4;
+
+		for (auto& pair : m_preds)
+		{
+			bool reachable = false;
+
+			for (u32 pred : pair.second)
+			{
+				if (pred >= lsa && pred < limit)
+				{
+					reachable = true;
+				}
+			}
+
+			if (!reachable && pair.first < limit)
+			{
+				limit = pair.first;
+			}
+		}
+
+		result.resize((limit - lsa) / 4 + 1);
+
 		// Check holes in safe mode (TODO)
 		u32 valid_size = 0;
 
@@ -790,13 +832,13 @@ std::vector<u32> spu_recompiler_base::block(const be_t<u32>* ls, u32 lsa)
 			{
 				const u32 pos = lsa + (i - 1) * 4;
 				const u32 data = ls[pos / 4];
-				const auto type = s_spu_itype.decode(data);
 
 				// Allow only NOP or LNOP instructions in holes
-				if (type == spu_itype::NOP || type == spu_itype::LNOP)
+				if (data == 0x200000 || (data & 0xffffff80) == 0x40200000)
 				{
 					if (i + 1 < result.size())
 					{
+						result[i] = se_storage<u32>::swap(data);
 						continue;
 					}
 				}
@@ -809,6 +851,12 @@ std::vector<u32> spu_recompiler_base::block(const be_t<u32>* ls, u32 lsa)
 				valid_size = i;
 			}
 		}
+
+		// Repeat if blocks were removed
+		if (result.size() == initial_size)
+		{
+			break;
+		}
 	}
 
 	if (result.size() == 1)

rpcs3/Emu/Cell/SPURecompiler.h

@@ -45,6 +45,9 @@ protected:
 	// List of possible targets for the instruction ({} = next instruction, {-1} = no targets)
 	std::unordered_map<u32, std::basic_string<u32>, value_hash<u32, 2>> m_targets;
 
+	// List of block predecessors
+	std::unordered_map<u32, std::basic_string<u32>, value_hash<u32, 2>> m_preds;
+
 	std::shared_ptr<spu_cache> m_cache;
 
 public: