Implement utils::memory_map_fd (partial)

Improve JIT profiling dump format (data + name, mmap)
Improve objdump interception util (better speed, fix bugs)
Rename spu_ubertrampoline to __ub+number

Utilities/JIT.cpp

@@ -24,26 +24,54 @@ void jit_announce(uptr func, usz size, std::string_view name)
 		return;
 	}
 
-	if (!name.empty())
-	{
 	// If directory ASMJIT doesn't exist, nothing will be written
-		static const fs::file s_asm = []()
+	static constexpr u64 c_dump_size = 0x1'0000'0000;
+	static constexpr u64 c_index_size = c_dump_size / 16;
+	static atomic_t<u64> g_index_off = 0;
+	static atomic_t<u64> g_data_off = c_index_size;
+
+	static void* g_asm = []() -> void*
 	{
 		fs::remove_all(fs::get_cache_dir() + "/ASMJIT/", false);
 
-			return fs::file(fmt::format("%s/ASMJIT/.objects", fs::get_cache_dir()), fs::rewrite + fs::append);
+		fs::file objs(fmt::format("%s/ASMJIT/.objects", fs::get_cache_dir()), fs::read + fs::rewrite);
-		}();
 
-		if (s_asm)
+		if (!objs || !objs.trunc(c_dump_size))
 		{
-			// Dump object: addr + size + bytes
+			return nullptr;
-			s_asm.write(fmt::format("%s%s%s",
-				std::string_view(reinterpret_cast<char*>(&func), 8),
-				std::string_view(reinterpret_cast<char*>(&size), 8),
-				std::string_view(reinterpret_cast<char*>(func), size)));
 		}
 
-		if (s_asm && name[0] != '_')
+		return utils::memory_map_fd(objs.get_handle(), c_dump_size, utils::protection::rw);
+	}();
+
+	if (g_asm && size < c_index_size)
+	{
+		struct entry
+		{
+			u64 addr; // RPCS3 process address
+			u32 size; // Function size
+			u32 off; // Function offset
+		};
+
+		// Write index entry at the beginning of file, and data + NTS name at fixed offset
+		const u64 index_off = g_index_off.fetch_add(1);
+		const u64 size_all = size + name.size() + 1;
+		const u64 data_off = g_data_off.fetch_add(size_all);
+
+		// If either index or data area is exhausted, nothing will be written
+		if (index_off < c_index_size / sizeof(entry) && data_off + size_all < c_dump_size)
+		{
+			entry& index = static_cast<entry*>(g_asm)[index_off];
+
+			std::memcpy(static_cast<char*>(g_asm) + data_off, reinterpret_cast<char*>(func), size);
+			std::memcpy(static_cast<char*>(g_asm) + data_off + size, name.data(), name.size());
+			index.size = static_cast<u32>(size);
+			index.off = static_cast<u32>(data_off);
+			atomic_storage<u64>::store(index.addr, func);
+		}
+	}
+
+	if (g_asm && !name.empty() && name[0] != '_')
 	{
 		// Save some objects separately
 		fs::file dump(fmt::format("%s/ASMJIT/%s", fs::get_cache_dir(), name), fs::rewrite);
@@ -53,7 +81,6 @@ void jit_announce(uptr func, usz size, std::string_view name)
 			dump.write(reinterpret_cast<uchar*>(func), size);
 		}
 	}
-	}
 
 #ifdef __linux__
 	static const fs::file s_map(fmt::format("/tmp/perf-%d.map", getpid()), fs::rewrite + fs::append);

objdump.cpp

@@ -21,10 +21,7 @@
 #include <cstdint>
 #include <unistd.h>
 #include <sys/file.h>
-#include <sys/wait.h>
+#include <sys/mman.h>
-#include <sys/sendfile.h>
-#include <spawn.h>
-#include <unordered_map>
 #include <string>
 #include <vector>
 #include <charconv>
@@ -53,58 +50,92 @@ int main(int argc, char* argv[])
 	// Get cache path
 	home += "/rpcs3/ASMJIT/";
 
-	// Get object names
+	// Get objects
 	int fd = open((home + ".objects").c_str(), O_RDONLY);
 
 	if (fd < 0)
 		return 1;
 
-	// Addr -> offset;size in .objects
+	// Map 4GiB (full size)
-	std::unordered_map<std::uint64_t, std::pair<std::uint64_t, std::uint64_t>> objects;
+	const auto data = mmap(nullptr, 0x10000'0000, PROT_READ, MAP_SHARED, fd, 0);
 
-	while (true)
+	struct entry
 	{
-		// Size is name size, not object size
+		std::uint64_t addr;
-		std::uint64_t ptr, size;
+		std::uint32_t size;
-		if (read(fd, &ptr, 8) != 8 || read(fd, &size, 8) != 8)
+		std::uint32_t off;
-			break;
+	};
-		std::uint64_t off = lseek(fd, 0, SEEK_CUR);
+
-		objects.emplace(ptr, std::make_pair(off, size));
+	// Index part (precedes actual data)
-		lseek(fd, size, SEEK_CUR);
+	const auto index = static_cast<const entry*>(data);
-	}
+
+	const entry* found = nullptr;
+
+	std::string out_file;
 
 	std::vector<std::string> args;
 
-	std::uint64_t addr = 0;
-
 	for (int i = 0; i < argc; i++)
 	{
 		// Replace args
 		std::string arg = argv[i];
 
-		if (arg.find("--start-address=0x") == 0)
+		if (std::uintptr_t(data) != -1 && arg.find("--start-address=0x") == 0)
 		{
+			// Decode address and try to find the object
+			std::uint64_t addr = -1;
+
 			std::from_chars(arg.data() + strlen("--start-address=0x"), arg.data() + arg.size(), addr, 16);
 
-			if (objects.count(addr))
+			for (int j = 0; j < 0x100'0000; j++)
 			{
-				// Extract object into a tmp file
+				if (index[j].addr == 0)
-				lseek(fd, objects[addr].first, SEEK_SET);
+				{
-				const int fd2 = open("/tmp/rpcs3.objdump.tmp", O_WRONLY | O_CREAT | O_TRUNC, S_IRUSR | S_IWUSR | S_IRGRP | S_IROTH);
+					break;
-				sendfile(fd2, fd, nullptr, objects[addr].second);
+				}
+
+				if (index[j].addr == addr)
+				{
+					found = index + j;
+					break;
+				}
+			}
+
+			if (found)
+			{
+				// Extract object into a new file (read file name from the mapped memory)
+				const char* name = static_cast<char*>(data) + found->off + found->size;
+
+				if (name[0])
+				{
+					out_file = home + name;
+				}
+				else
+				{
+					out_file = "/tmp/rpcs3.objdump." + std::to_string(getpid());
+					unlink(out_file.c_str());
+				}
+
+				const int fd2 = open(out_file.c_str(), O_WRONLY | O_CREAT | O_EXCL, S_IRUSR | S_IWUSR | S_IRGRP | S_IROTH);
+
+				if (fd2 > 0)
+				{
+					// Don't overwrite if exists
+					write(fd2, static_cast<char*>(data) + found->off, found->size);
 					close(fd2);
+				}
 
 				args.emplace_back("--adjust-vma=" + to_hex(addr));
 				continue;
 			}
 		}
 
-		if (objects.count(addr) && arg.find("--stop-address=0x") == 0)
+		if (found && arg.find("--stop-address=0x") == 0)
 		{
 			continue;
 		}
 
-		if (objects.count(addr) && arg == "-d")
+		if (found && arg == "-d")
 		{
 			arg = "-D";
 		}
@@ -117,14 +148,14 @@ int main(int argc, char* argv[])
 		args.emplace_back(std::move(arg));
 	}
 
-	if (objects.count(addr))
+	if (found)
 	{
 		args.pop_back();
 		args.emplace_back("-b");
 		args.emplace_back("binary");
 		args.emplace_back("-m");
-		args.emplace_back("i386");
+		args.emplace_back("i386:x86-64");
-		args.emplace_back("/tmp/rpcs3.objdump.tmp");
+		args.emplace_back(std::move(out_file));
 	}
 
 	args[0] = "/usr/bin/objdump";
@@ -138,12 +169,11 @@ int main(int argc, char* argv[])
 
 	new_argv.push_back(nullptr);
 
-	if (objects.count(addr))
+	if (found)
 	{
 		int fds[2];
 		pipe(fds);
 
-		// objdump is broken; fix address truncation
 		if (fork() > 0)
 		{
 			close(fds[1]);
@@ -158,20 +188,6 @@ int main(int argc, char* argv[])
 
 					if (c == '\n')
 					{
-						// Replace broken address
-						if ((buf[0] >= '0' && buf[0] <= '9') || (buf[0] >= 'a' && buf[0] <= 'f'))
-						{
-							std::uint64_t ptr = -1;
-							auto cvt = std::from_chars(buf.data(), buf.data() + buf.size(), ptr, 16);
-
-							if (cvt.ec == std::errc() && ptr < addr)
-							{
-								auto fix = to_hex((ptr - std::uint32_t(addr)) + addr, false);
-								write(STDOUT_FILENO, fix.data(), fix.size());
-								buf = std::string(cvt.ptr);
-							}
-						}
-
 						write(STDOUT_FILENO, buf.data(), buf.size());
 						buf.clear();
 					}

rpcs3/Emu/Cell/SPURecompiler.cpp

@@ -1074,7 +1074,9 @@ spu_function_t spu_runtime::rebuild_ubertrampoline(u32 id_inst)
 		workload.clear();
 		result = reinterpret_cast<spu_function_t>(reinterpret_cast<u64>(wxptr));
 
-		jit_announce(wxptr, raw - wxptr, "spu_ubertrampoline");
+		std::string fname;
+		fmt::append(fname, "__ub%u", m_flat_list.size());
+		jit_announce(wxptr, raw - wxptr, fname);
 	}
 
 	if (auto _old = stuff_it->trampoline.compare_and_swap(nullptr, result))

rpcs3/util/vm.hpp

@@ -5,6 +5,12 @@
 
 namespace utils
 {
+#ifdef _WIN32
+	using native_handle = void*;
+#else
+	using native_handle = int;
+#endif
+
 	// Memory protection type
 	enum class protection
 	{
@@ -43,6 +49,9 @@ namespace utils
 	// Lock pages in memory
 	bool memory_lock(void* pointer, usz size);
 
+	// Map file descriptor
+	void* memory_map_fd(native_handle fd, usz size, protection prot);
+
 	// Shared memory handle
 	class shm
 	{

rpcs3/util/vm_native.cpp

@@ -296,6 +296,23 @@ namespace utils
 #endif
 	}
 
+	void* memory_map_fd(native_handle fd, usz size, protection prot)
+	{
+#ifdef _WIN32
+		// TODO
+		return nullptr;
+#else
+		const auto result = ::mmap(nullptr, size, +prot, MAP_SHARED, fd, 0);
+
+		if (result == reinterpret_cast<void*>(uptr{umax}))
+		{
+			[[unlikely]] return nullptr;
+		}
+
+		return result;
+#endif
+	}
+
 	shm::shm(u32 size, u32 flags)
 		: m_flags(flags)
 		, m_size(utils::align(size, 0x10000))