Clean up asm for format (#164)

* clean up asm for format

* fix windows

* remove rpc-call hack

* revert cmake version change

* clang format

* asm fix for windows

CMakeLists.txt

@@ -111,4 +111,4 @@ add_subdirectory("third-party/zydis")
 # windows memory management lib
 IF (WIN32)
     add_subdirectory(third-party/mman)
-ENDIF ()
+ENDIF ()

common/util/CMakeLists.txt

@@ -3,3 +3,7 @@ add_library(common_util
         FileUtil.cpp
         DgoWriter.cpp
         Timer.cpp)
+
+IF(UNIX)
+target_link_libraries(common_util stdc++fs)
+ENDIF()

game/kernel/asm_funcs.asm

@@ -4,20 +4,47 @@
 
 ;; GOAL Runtime assembly functions. These exist only in the x86 version of GOAL.
 
-;; declaration of the extern "C" function format_impl
-extern format_impl
-
 SECTION .TEXT
 
-;; This _format function which will be exported to the GOAL symbol table at runtime start as "_format"
+;; this function pushes all 8 OpenGOAL registers into a stack array.
-;; This function accepts 8 GOAL arguments and puts them on the stack, then calls format_impl and passes
+;; then it calls the function pointed to by rax with a pointer to this array.
-;; a pointer to this array of GOAL arguments as the argument. The reason for this is that GOAL and
+;; it returns the return value of the called function.
-;; the standard System V ABI used in Linux are different for 8 argument function calls.
+global _stack_call_linux
-
+_stack_call_linux:
-global _format_win32
+  pop rax
-_format_win32:
+  ; align stack
-  ; GOAL will call with regs  RDI, RSI, RDX, RCX, R8, R9, R10, R11
+  sub rsp, 8
+  ; create stack array of arguments
+  push r11
+  push r10
+  push r9
+  push r8
+  push rcx
+  push rdx
+  push rsi
+  push rdi
+  ; set first argument
+  mov rdi, rsp
+  ; call function
+  call rax
+  ; restore arguments (probably don't need to really do this...)
+  pop rdi
+  pop rsi
+  pop rdx
+  pop rcx
+  pop r8
+  pop r9
+  pop r10
+  pop r11
+  ; restore stack
+  add rsp, 8
+  ; return!
+  ret
 
+;; windows implementation of stack_call
+global _stack_call_win32
+_stack_call_win32:
+  pop rax
   ; to make sure the stack frame is aligned
   sub rsp, 8
 
@@ -36,7 +63,7 @@ _format_win32:
   sub rsp, 32
 
   ; call C function to do format, result will go in RAX
-  call format_impl
+  call rax
   add rsp, 32
 
   ; restore
@@ -52,47 +79,6 @@ _format_win32:
   add rsp, 8
   ret
 
-global _format_linux
-_format_linux:
-  ; GOAL will call with regs  RDI, RSI, RDX, RCX, R8, R9, R10, R11
-
-  ; to make sure the stack frame is aligned
-  sub rsp, 8
-
-  ; push all registers and create the register array on the stack
-  push r11
-  push r10
-  push r9
-  push r8
-  push rcx
-  push rdx
-  push rsi
-  push rdi
-
-  ; set the first argument register to the stack argument array
-  mov rdi, rsp
-
-  ; call C function to do format, result will go in RAX
-  call format_impl
-
-  ; restore
-  ; (note - this could probably just be add rsp 72, we don't care about the value of these register)
-  pop rdi
-  pop rsi
-  pop rdx
-  pop rcx
-  pop r8
-  pop r9
-  pop r10
-  pop r11
-  add rsp, 8
-  ret
-;; NOTE: calling format has a _lot_ of indirection...
-;; symbol table lookup to find the GOAL "format" symbol value
-;; run the GOAL-to-C trampoline (on GOAL heap) to jump to this _format
-;; run this wrapper to call the real format_impl
-
-
 ;; The _call_goal_asm function is used to call a GOAL function from C.
 ;; It supports up to 3 arguments and a return value.
 ;; This should be called with the arguments:

game/kernel/kdgo.cpp

@@ -50,38 +50,31 @@ s32 RpcCall(s32 rpcChannel,
                        nullptr);
 }
 
-// Terrible hack! Remove soon!
-
 namespace {
-struct RpcCallArgCache {
+struct GoalStackArgs {
-  s32 rpcChannel;
+  u64 args[8];
-  u32 fno;
+  template <typename T>
-  u32 async;
+  T get_as(int i) {
-} rpc_arg_cache;
+    static_assert(sizeof(T) <= 8, "arg size");
+    T result;
+    memcpy(&result, args + i, sizeof(T));
+    return result;
+  }
+};
 }  // namespace
 
-void RpcCall_wrapper_part1(s32 rpcChannel, u32 fno, u32 async) {
-  rpc_arg_cache.rpcChannel = rpcChannel;
-  rpc_arg_cache.fno = fno;
-  rpc_arg_cache.async = async;
-}
-
-u64 RpcCall_wrapper_part2(u64 send_buff, s32 send_size, u64 recv_buff, s32 recv_size) {
-  return RpcCall_wrapper(rpc_arg_cache.rpcChannel, rpc_arg_cache.fno, rpc_arg_cache.async,
-                         send_buff, send_size, recv_buff, recv_size);
-}
-
 /*!
  * GOAL Wrapper for RpcCall.
  */
-u64 RpcCall_wrapper(s32 rpcChannel,
+u64 RpcCall_wrapper(void* _args) {
-                    u32 fno,
+  GoalStackArgs* args = (GoalStackArgs*)_args;
-                    u32 async,
+  auto rpcChannel = args->get_as<s32>(0);
-                    u64 send_buff,
+  auto fno = args->get_as<u32>(1);
-                    s32 send_size,
+  auto async = args->get_as<u32>(2);
-                    u64 recv_buff,
+  auto send_buff = args->get_as<u64>(3);
-                    s32 recv_size) {
+  auto send_size = args->get_as<s32>(4);
-  fprintf(stderr, "size in c is %d\n", recv_size);
+  auto recv_buff = args->get_as<u64>(5);
+  auto recv_size = args->get_as<s32>(6);
   return sceSifCallRpc(&cd[rpcChannel], fno, async, Ptr<u8>(send_buff).c(), send_size,
                        Ptr<u8>(recv_buff).c(), recv_size, nullptr, nullptr);
 }

game/kernel/kdgo.h

@@ -19,16 +19,7 @@ void load_and_link_dgo_from_c(const char* name, Ptr<kheapinfo> heap, u32 linkFla
 void load_and_link_dgo(u64 name_gstr, u64 heap_info, u64 flag, u64 buffer_size);
 void StopIOP();
 
-u64 RpcCall_wrapper(s32 rpcChannel,
+u64 RpcCall_wrapper(void* args);
-                    u32 fno,
-                    u32 async,
-                    u64 send_buff,
-                    s32 send_size,
-                    u64 recv_buff,
-                    s32 recv_size);
 u32 RpcBusy(s32 channel);
 void LoadDGOTest();
-
-void RpcCall_wrapper_part1(s32 rpcChannel, u32 fno, u32 async);
-u64 RpcCall_wrapper_part2(u64 send_buff, s32 send_size, u64 recv_buff, s32 recv_size);
 #endif  // JAK_V2_KDGO_H

game/kernel/kscheme.cpp

@@ -315,6 +315,10 @@ u64 make_string_from_c(const char* c_str) {
   return mem;
 }
 
+/*!
+ * This creates an OpenGOAL function from a C++ function. Only 6 arguments can be accepted.
+ * But calling this function is very fast and doesn't use the stack.
+ */
 Ptr<Function> make_function_from_c_linux(void* func) {
   // allocate a function object on the global heap
   auto mem = Ptr<u8>(
@@ -393,16 +397,59 @@ ret
   return mem.cast<Function>();
 }
 
+extern "C" {
+void _stack_call_linux();
+void _stack_call_win32();
+}
+
+Ptr<Function> make_stack_arg_function_from_c_linux(void* func) {
+  // allocate a function object on the global heap
+  auto mem = Ptr<u8>(
+      alloc_heap_object(s7.offset + FIX_SYM_GLOBAL_HEAP, *(s7 + FIX_SYM_FUNCTION_TYPE), 0x40));
+  auto f = (uint64_t)func;
+  auto target_function = (u8*)&f;
+  auto trampoline_function_addr = _stack_call_linux;
+  auto trampoline = (u8*)&trampoline_function_addr;
+
+  // movabs rax, target_function
+  int offset = 0;
+  mem.c()[offset++] = 0x48;
+  mem.c()[offset++] = 0xb8;
+  for (int i = 0; i < 8; i++) {
+    mem.c()[offset++] = target_function[i];
+  }
+
+  // push rax
+  mem.c()[offset++] = 0x50;
+
+  // movabs rax, trampoline
+  mem.c()[offset++] = 0x48;
+  mem.c()[offset++] = 0xb8;
+  for (int i = 0; i < 8; i++) {
+    mem.c()[offset++] = trampoline[i];
+  }
+
+  // jmp rax
+  mem.c()[offset++] = 0xff;
+  mem.c()[offset++] = 0xe0;
+
+  // CacheFlush(mem, 0x34);
+
+  return mem.cast<Function>();
+}
+
 /*!
  * Create a GOAL function from a C function.  This calls a windows function, but doesn't scramble
  * the argument order.  It's supposed to be used with _format_win32 which assumes GOAL order.
  */
-Ptr<Function> make_function_for_format_from_c_win32(void* func) {
+Ptr<Function> make_stack_arg_function_from_c_win32(void* func) {
   // allocate a function object on the global heap
   auto mem = Ptr<u8>(
       alloc_heap_object(s7.offset + FIX_SYM_GLOBAL_HEAP, *(s7 + FIX_SYM_FUNCTION_TYPE), 0x80));
   auto f = (uint64_t)func;
   auto fp = (u8*)&f;
+  auto trampoline_function_addr = _stack_call_win32;
+  auto trampoline = (u8*)&trampoline_function_addr;
 
   int i = 0;
   // we will put the function address in RAX with a movabs rax, imm8
@@ -412,13 +459,20 @@ Ptr<Function> make_function_for_format_from_c_win32(void* func) {
     mem.c()[i++] = fp[j];
   }
 
+  // push rax
+  mem.c()[i++] = 0x50;
+
+  // we will put the function address in RAX with a movabs rax, imm8
+  mem.c()[i++] = 0x48;
+  mem.c()[i++] = 0xb8;
+  for (int j = 0; j < 8; j++) {
+    mem.c()[i++] = trampoline[j];
+  }
+
   /*
-   * sub rsp, 40
+   * jmp rax
-   * call rax
-   * add rsp, 40
-   * ret
    */
-  for (auto x : {0x48, 0x83, 0xEC, 0x28, 0xFF, 0xD0, 0x48, 0x83, 0xC4, 0x28, 0xC3}) {
+  for (auto x : {0xFF, 0xE0}) {
     mem.c()[i++] = x;
   }
 
@@ -439,6 +493,14 @@ Ptr<Function> make_function_from_c(void* func) {
 #endif
 }
 
+Ptr<Function> make_stack_arg_function_from_c(void* func) {
+#ifdef __linux__
+  return make_stack_arg_function_from_c_linux(func);
+#elif _WIN32
+  return make_stack_arg_function_from_c_win32(func);
+#endif
+}
+
 /*!
  * Create a GOAL function which does nothing and immediately returns.
  */
@@ -471,6 +533,8 @@ Ptr<Function> make_zero_func() {
  * Given a C function and a name, create a GOAL function and store it in the symbol with the given
  * name. This effectively creates a global GOAL function with the given name which calls the given C
  * function.
+ *
+ * This work on both Linux and Windows, but only supports up to 6 arguments.
  */
 Ptr<Function> make_function_symbol_from_c(const char* name, void* f) {
   auto sym = intern_from_c(name);
@@ -480,12 +544,12 @@ Ptr<Function> make_function_symbol_from_c(const char* name, void* f) {
 }
 
 /*!
- * Given a C function and a name, create a GOAL function and store it in the symbol with the given
+ * Like make_function_symbol_from_c, but all 8 GOAL arguments are put into an array on the stack.
- * name. This is designed for _format_win32, which is special because it takes 8 arguments.
+ * The address of this array is passed as the first and only argument to f.
  */
-Ptr<Function> make_format_function_symbol_from_c_win32(const char* name, void* f) {
+Ptr<Function> make_stack_arg_function_symbol_from_c(const char* name, void* f) {
   auto sym = intern_from_c(name);
-  auto func = make_function_for_format_from_c_win32(f);
+  auto func = make_stack_arg_function_from_c(f);
   sym->value = func.offset;
   return func;
 }
@@ -1587,15 +1651,6 @@ s32 test_function(s32 arg0, s32 arg1, s32 arg2, s32 arg3) {
   return arg0 + 2 * arg1 + 3 * arg2 + 4 * arg3;
 }
 
-extern "C" {
-// defined in asm_funcs. It calls format_impl and sets up arguments correctly.
-#ifdef __linux__
-void _format_linux();
-#elif _WIN32
-void _format_win32();
-#endif
-}
-
 /*!
  * Initializes the GOAL Heap, GOAL Symbol Table, GOAL Funcdamental Types, loads the GOAL kernel,
  * exports Machine functions, loads the game engine, and calls "play" to initialize the engine.
@@ -1845,11 +1900,7 @@ s32 InitHeapAndSymbol() {
   make_function_symbol_from_c("load", (void*)load);
   make_function_symbol_from_c("loado", (void*)loado);
   make_function_symbol_from_c("unload", (void*)unload);
-#ifdef __linux__
+  make_stack_arg_function_symbol_from_c("_format", (void*)format_impl);
-  make_function_symbol_from_c("_format", (void*)_format_linux);
-#elif _WIN32
-  make_format_function_symbol_from_c_win32("_format", (void*)_format_win32);
-#endif
 
   // allocations
   make_function_symbol_from_c("malloc", (void*)alloc_heap_memory);

game/kernel/kscheme.h

@@ -101,6 +101,7 @@ u64 load(u32 file_name_in, u32 heap_in);
 u64 loado(u32 file_name_in, u32 heap_in);
 u64 unload(u32 name);
 Ptr<Function> make_function_symbol_from_c(const char* name, void* f);
+Ptr<Function> make_stack_arg_function_symbol_from_c(const char* name, void* f);
 u64 call_goal_function_by_name(const char* name);
 u64 call_method_of_type_arg2(u32 arg, Ptr<Type> type, u32 method_id, u32 a1, u32 a2);
 Ptr<Type> alloc_and_init_type(Ptr<Symbol> sym, u32 method_count);

game/kernel/ksound.cpp

@@ -25,8 +25,5 @@ void InitSoundScheme() {
   make_function_symbol_from_c("rpc-call", (void*)RpcCall_wrapper);
   make_function_symbol_from_c("rpc-busy?", (void*)RpcBusy);
   make_function_symbol_from_c("test-load-dgo-c", (void*)LoadDGOTest);
-
+  make_stack_arg_function_symbol_from_c("rpc-call", (void*)RpcCall_wrapper);
-  // terrible hack!
-  make_function_symbol_from_c("rpc-call-p1", (void*)RpcCall_wrapper_part1);
-  make_function_symbol_from_c("rpc-call-p2", (void*)RpcCall_wrapper_part2);
 }

goal_src/engine/ps2/rpc-h.gc

@@ -130,12 +130,6 @@
   '#f
   )
 
-(defmacro hack-rpc-call (a0 a1 a2 a3 a4 a5 a6)
-  `(begin
-     (rpc-call-p1 ,a0 ,a1 ,a2)
-  	 (rpc-call-p2 ,a3 ,a4 ,a5 ,a6)
-    )
-  )
 
 ;; method 9
 (defmethod call rpc-buffer-pair ((obj rpc-buffer-pair) (fno uint) (recv-buff pointer) (recv-size uint))
@@ -170,7 +164,7 @@
       (let ((current-buffer (-> obj current)))
         ;; rpc_channel, fno, async, send_buff, send_size, recv_buff, recv_size
         (format 0 "recv-size is ~D~%" recv-size)
-        (hack-rpc-call (-> obj rpc-port)
+        (rpc-call (-> obj rpc-port)
                   fno
                   (the uint 1)
                   (the uint (-> current-buffer base))

goal_src/kernel-defs.gc

@@ -154,6 +154,3 @@
 
 (define-extern rpc-call (function int uint uint uint int uint int uint))
 (define-extern rpc-busy? (function int uint))
-;; hack! remove!
-(define-extern rpc-call-p1 (function int uint uint none))
-(define-extern rpc-call-p2 (function uint int uint int uint))