PPU LLVM: remove unnecessary code

rpcs3/Emu/Cell/PPUThread.cpp

@@ -875,141 +875,6 @@ static bool adde_carry(u64 a, u64 b, bool c)
 #endif
 }
 
-static std::string ppu_context_prologue()
-{
-	std::string c;
-	//c += "\xCC";
-#ifndef _WIN32
-	c += "\x48\x89\xF9"; // mov rcx, rdi
-#endif
-	c += "\x48\xB8"; // mov rax, imm64
-	uptr ptr = (uptr)&vm::g_base_addr;
-	c.append((const char*)&ptr, 8);
-	c += "\x48\x8B"; // mov rax, [rax]
-	c += '\0';
-	c += "\x48\x03\x41"; // add rax, [ppu+r3]
-	c += char(offset32(&ppu_thread::gpr, 3));
-	c += "\x48\x83\xC0\x0F"; // add rax, 15
-	c += "\x48\x83\xE0\xF0"; // and rax, -16
-	return c;
-}
-
-const auto ppu_get_context = []() -> std::string
-{
-	std::string c = ppu_context_prologue();
-	c += "\x48\x8B\x51"; // mov rdx, [rcx+r1]
-	c += char(offset32(&ppu_thread::gpr, 1));
-	c += "\x48\x89\x10"; // mov [rax], rdx
-	c += "\x48\x8B\x51"; // mov rdx, [rcx+r2]
-	c += char(offset32(&ppu_thread::gpr, 2));
-	c += "\x48\x89\x50\x08"; // mov [rax+8], rdx
-	c += "\x48\x8B\x54\x24\xF8"; // mov rdx, [rsp-8]
-	c += "\x48\x89\x50\x10"; // mov [rax+0x10], rdx
-
-	c += "\x48\x89\x60\x18"; // mov [rax+0x18], rsp
-	c += "\x48\x89\x58\x20"; // mov [rax+0x20], rbx
-	c += "\x48\x89\x68\x28"; // mov [rax+0x28], rbp
-#ifdef _WIN32
-	c += "\x48\x89\x70\x30"; // mov [rax+0x30], rsi
-	c += "\x48\x89\x78\x38"; // mov [rax+0x38], rdi
-#endif
-	c += "\x4C\x89\x60\x40"; // mov [rax+0x40], r12
-	c += "\x4C\x89\x68\x48"; // mov [rax+0x48], r13
-	c += "\x4C\x89\x70\x50"; // mov [rax+0x50], r14
-	c += "\x4C\x89\x78\x58"; // mov [rax+0x58], r15
-
-#ifdef _WIN32
-	c += "\x66\x0F\x7F\x70\x60"; // movdqa [rax+0x60], xmm6
-	c += "\x66\x0F\x7F\x78\x70"; // movdqa [rax+0x70], xmm7
-	c += "\x66\x44\x0F\x7F\x80\x80\x00\x00\x00"s; // ...
-	c += "\x66\x44\x0F\x7F\x88\x90\x00\x00\x00"s;
-	c += "\x66\x44\x0F\x7F\x90\xA0\x00\x00\x00"s;
-	c += "\x66\x44\x0F\x7F\x98\xB0\x00\x00\x00"s;
-	c += "\x66\x44\x0F\x7F\xA0\xC0\x00\x00\x00"s;
-	c += "\x66\x44\x0F\x7F\xA8\xD0\x00\x00\x00"s;
-	c += "\x66\x44\x0F\x7F\xB0\xE0\x00\x00\x00"s;
-	c += "\x66\x44\x0F\x7F\xB8\xF0\x00\x00\x00"s;
-#endif
-
-	c += "\x48\xC7\x41"; // mov [rcx+r3], 0
-	c += char(offset32(&ppu_thread::gpr, 3));
-	c.append(4, '\0');
-	//c += "\xCC";
-	c += "\xC3"; // ret
-	return c;
-}();
-
-const auto ppu_set_context = []() -> std::string
-{
-	std::string c = ppu_context_prologue();
-	c += "\xCC";
-	c += "\x48\x8B\x10"; // mov rdx, [rax]
-	c += "\x48\x89\x51"; // mov [rcx+r1], rdx
-	c += char(offset32(&ppu_thread::gpr, 1));
-
-	c += "\x48\x8B\x50\x08"; // mov rdx, [rax+8]
-	c += "\x48\x89\x51"; // mov [rcx+r2], rdx
-	c += char(offset32(&ppu_thread::gpr, 2));
-
-	c += "\x48\x8B\x60\x18"; // mov rsp, [rax+0x18]
-	c += "\x48\x8B\x58\x20"; // mov rbx, [rax+0x20]
-	c += "\x48\x8B\x68\x28"; // mov rbp, [rax+0x28]
-#ifdef _WIN32
-	c += "\x48\x8B\x70\x30"; // mov rsi, [rax+0x30]
-	c += "\x48\x8B\x78\x38"; // mov rdi, [rax+0x38]
-#endif
-	c += "\x4C\x8B\x60\x40"; // mov r12, [rax+0x40]
-	c += "\x4C\x8B\x68\x48"; // mov r13, [rax+0x48]
-	c += "\x4C\x8B\x70\x50"; // mov r14, [rax+0x50]
-	c += "\x4C\x8B\x78\x58"; // mov r15, [rax+0x58]
-
-#ifdef _WIN32
-	c += "\x66\x0F\x6F\x70\x60"; // movdqa xmm6, [rax+0x60]
-	c += "\x66\x0F\x6F\x78\x70"; // movdqa xmm7, [rax+0x70]
-	c += "\x66\x44\x0F\x6F\x80\x80\x00\x00\x00"s; // ...
-	c += "\x66\x44\x0F\x6F\x88\x90\x00\x00\x00"s;
-	c += "\x66\x44\x0F\x6F\x90\xA0\x00\x00\x00"s;
-	c += "\x66\x44\x0F\x6F\x98\xB0\x00\x00\x00"s;
-	c += "\x66\x44\x0F\x6F\xA0\xC0\x00\x00\x00"s;
-	c += "\x66\x44\x0F\x6F\xA8\xD0\x00\x00\x00"s;
-	c += "\x66\x44\x0F\x6F\xB0\xE0\x00\x00\x00"s;
-	c += "\x66\x44\x0F\x6F\xB8\xF0\x00\x00\x00"s;
-#endif
-
-	c += "\x48\x8B\x50\x10"; // mov rdx, [rax+0x10]
-	c += "\x48\x89\x54\x24\xF8"; // mov [rsp-8], rdx
-	c += "\x48\x8B\x51"; // mov rdx, [rcx+r4]
-	c += char(offset32(&ppu_thread::gpr, 4));
-	c += "\x48\x85\xD2"; // test rdx, rdx
-	c += "\x0F\x94\xC2"; // setz dl
-	c += "\x48\x0F\xB6\xD2"; // movzx rdx, dl
-	c += "\x48\x89\x51"; // mov [rcx+r3], rdx
-	c += char(offset32(&ppu_thread::gpr, 3));
-	c += "\xC3"; // ret
-	return c;
-}();
-
-const auto ppu_use_context = []() -> std::string
-{
-	std::string c;
-	c += "\x48\xB8"; // mov rax, imm64
-	uptr ptr = (uptr)&vm::g_exec_addr;
-	c.append((const char*)&ptr, 8);
-	c += "\x48\x8B\x20"; // mov rsp, [rax]
-#ifdef _WIN32
-	c += "\x48\x01\xD4"; // add rsp,rdx
-#else
-	c += "\x48\x01\xFC"; // add rsp,rsi
-#endif
-	//c += "\x48\x83\xE4\xE0"; // and rsp, -0x20
-#ifdef _WIN32
-	c += "\x41\xFF\xD0"; // call r8
-#else
-	c += "\xFF\xD2"; // call rdx
-#endif
-	return c;
-}();
-
 extern void ppu_initialize()
 {
 	const auto _funcs = fxm::withdraw<std::vector<ppu_function>>();
@@ -1380,60 +1245,6 @@ static void ppu_initialize2(jit_compiler& jit, const ppu_module& module_part, co
 
 				// Run optimization passes
 				pm.run(*func);
-
-				const auto _syscall = module->getFunction("__syscall");
-
-				for (auto i = inst_begin(*func), end = inst_end(*func); i != end;)
-				{
-					const auto inst = &*i++;
-
-					if (const auto ci = dyn_cast<CallInst>(inst))
-					{
-						const auto cif = ci->getCalledFunction();
-						const auto op1 = ci->getNumArgOperands() > 1 ? ci->getArgOperand(1) : nullptr;
-
-						if (cif == _syscall && op1 && isa<ConstantInt>(op1))
-						{
-							// Try to determine syscall using the value from r11 (requires constant propagation)
-							const u64 index = cast<ConstantInt>(op1)->getZExtValue();
-
-							if (const auto ptr = ppu_get_syscall(index))
-							{
-								const auto n = ppu_get_syscall_name(index);
-								const auto f = cast<Function>(module->getOrInsertFunction(n, _func));
-
-								// Call the syscall directly
-								ReplaceInstWithInst(ci, CallInst::Create(f, {ci->getArgOperand(0)}));
-							}
-						}
-
-						continue;
-					}
-
-					if (const auto li = dyn_cast<LoadInst>(inst))
-					{
-						// TODO: more careful check
-						if (li->getNumUses() == 0)
-						{
-							// Remove unreferenced volatile loads
-							li->eraseFromParent();
-						}
-
-						continue;
-					}
-
-					if (const auto si = dyn_cast<StoreInst>(inst))
-					{
-						// TODO: more careful check
-						if (isa<UndefValue>(si->getOperand(0)) && si->getParent() == &func->getEntryBlock())
-						{
-							// Remove undef volatile stores
-							si->eraseFromParent();
-						}
-
-						continue;
-					}
-				}
 			}
 		}
 

rpcs3/Emu/Cell/PPUTranslator.cpp

@@ -7,6 +7,8 @@
 #include "../Utilities/Log.h"
 #include <algorithm>
 
+extern std::string ppu_get_syscall_name(u64 code);
+
 using namespace llvm;
 
 const ppu_decoder<PPUTranslator> s_ppu_decoder;
@@ -1636,6 +1638,21 @@ void PPUTranslator::SC(ppu_opcode_t op)
 	const auto num = GetGpr(11);
 	RegStore(m_ir->getInt32(m_current_addr), m_cia);
 	FlushRegisters();
+
+	if (!op.lev && isa<ConstantInt>(num))
+	{
+		// Try to determine syscall using the constant value from r11
+		const u64 index = cast<ConstantInt>(num)->getZExtValue();
+
+		if (index < 1024)
+		{
+			// Call the syscall directly
+			Call(GetType<void>(), ppu_get_syscall_name(index), m_thread);
+			m_ir->CreateRetVoid();
+			return;
+		}
+	}
+
 	Call(GetType<void>(), op.lev ? "__lv1call" : "__syscall", m_thread, num);
 	m_ir->CreateRetVoid();
 }

rpcs3/Emu/Cell/lv2/lv2.cpp

@@ -33,7 +33,13 @@
 
 extern std::string ppu_get_syscall_name(u64 code);
 
-static constexpr ppu_function_t null_func = nullptr;
+static bool null_func(ppu_thread& ppu)
+{
+	LOG_TODO(HLE, "Unimplemented syscall %s -> CELL_OK", ppu_get_syscall_name(ppu.gpr[11]));
+	ppu.gpr[3] = 0;
+	ppu.cia += 4;
+	return false;
+}
 
 std::array<ppu_function_t, 1024> g_ppu_syscall_table{};
 
@@ -974,15 +980,8 @@ extern void ppu_execute_syscall(ppu_thread& ppu, u64 code)
 		{
 			func(ppu);
 			LOG_TRACE(PPU, "Syscall '%s' (%llu) finished, r3=0x%llx", ppu_get_syscall_name(code), code, ppu.gpr[3]);
+			return;
 		}
-		else
-		{
-			LOG_TODO(HLE, "Unimplemented syscall %s -> CELL_OK", ppu_get_syscall_name(code));
-			ppu.gpr[3] = 0;
-			ppu.cia += 4;
-		}
-
-		return;
 	}
 
 	fmt::throw_exception("Invalid syscall number (%llu)", code);