i#58 MacOS: assembly support

The default assembler uses Intel mnemonics yet swaps the source and
destination order.  So for now I'm requiring a recent nasm (recent for
64-bit support) to be installed.

Several other wacky differences (including kind of shocking shortcomings in
cpp) required tweaking how assembly is done on Mac.

SVN-Revision: 2063

core/x86/asm_defines.asm

@@ -148,15 +148,20 @@ ASSUME fs:_DATA @N@\
 # define END_FUNC(symbol) /* nothing */
 # define DECLARE_GLOBAL(symbol) global symbol
 # define GLOBAL_LABEL(label) label
+# define ADDRTAKEN_LABEL(label) label
 # define BYTE byte
 # define WORD word
 # define DWORD dword
 # define QWORD qword
-# define SYMREF(sym) [sym]
+# ifdef X64
+#  define SYMREF(sym) [rel sym]
+# else
+#  define SYMREF(sym) [sym]
+# endif
 # define HEX(n) 0x##n
 # define SEGMEM(seg,mem) [seg:mem]
 # define DECL_EXTERN(symbol) EXTERN symbol
-# define RAW(n) error_not_implemented
+# define RAW(n) DB HEX(n) @N@
 # define DECLARE_FUNC_SEH(symbol) DECLARE_FUNC(symbol)
 # define PUSH_SEH(reg) push reg
 # define PUSH_NONCALLEE_SEH(reg) push reg

core/x86/x86.asm

@@ -439,7 +439,7 @@ GLOBAL_LABEL(dr_app_start:)
         sub     REG_XSP, FRAME_ALIGNMENT - ARG_SZ  /* Maintain alignment. */
 
         /* grab exec state and pass as param in a priv_mcontext_t struct */
-        PUSH_PRIV_MCXT([FRAME_ALIGNMENT - ARG_SZ + REG_XSP -
+        PUSH_PRIV_MCXT(PTRSZ [FRAME_ALIGNMENT - ARG_SZ + REG_XSP -
                         PUSH_PRIV_MCXT_PRE_PC_SHIFT]) /* return address as pc */
 
         /* do the rest in C */
@@ -472,7 +472,7 @@ GLOBAL_LABEL(dynamorio_app_take_over:)
         sub     REG_XSP, FRAME_ALIGNMENT - ARG_SZ  /* Maintain alignment. */
 
         /* grab exec state and pass as param in a priv_mcontext_t struct */
-        PUSH_PRIV_MCXT([FRAME_ALIGNMENT - ARG_SZ + REG_XSP -
+        PUSH_PRIV_MCXT(PTRSZ [FRAME_ALIGNMENT - ARG_SZ + REG_XSP -
                         PUSH_PRIV_MCXT_PRE_PC_SHIFT]) /* return address as pc */
 
         /* do the rest in C */
@@ -2208,6 +2208,8 @@ make_val_word_size:
         jmp     do_memset
         END_FUNC(memset)
 
+
+# ifndef MACOS /* XXX: attribute alias issue, plus using nasm */
 /* gcc emits calls to these *_chk variants in release builds when the size of
  * dst is known at compile time.  In C, the caller is responsible for cleaning
  * up arguments on the stack, so we alias these *_chk routines to the non-chk
@@ -2220,6 +2222,7 @@ make_val_word_size:
 .global __memset_chk
 .hidden __memset_chk
 .set __memset_chk,memset
+# endif
 
 
 /* Replacement for _dl_runtime_resolve() used for catching module transitions

make/CMake_asm.cmake

@@ -34,4 +34,11 @@ string(REGEX REPLACE
   "\@N\@"
   "\n"
   string "${string}")
+# MacOS "cpp -E" on macros with token pasting leaves ## in place!
+# And while "gcc -E" does not, it has the linker warning, so we just
+# remove them here.
+string(REGEX REPLACE
+  "##"
+  ""
+  string "${string}")
 file(WRITE ${file} "${string}")

make/configure.cmake.h

@@ -219,7 +219,9 @@
 #  define NOLIBC
 #endif
 
-#ifdef UNIX
+#ifdef MACOS
+#  define ASSEMBLE_WITH_NASM
+#elif defined(UNIX)
 #  define ASSEMBLE_WITH_GAS
 #else
 #  define ASSEMBLE_WITH_MASM

make/cpp2asm_support.cmake

@@ -162,7 +162,22 @@ if (NOT "${CMAKE_GENERATOR}" MATCHES "Visual Studio")
   enable_language(ASM)
 endif ()
 
-if (UNIX)
+if (APPLE)
+  # See above: we can't use CMAKE_ASM_COMPILER and require nasm.
+  find_program(NASM nasm DOC "path to nasm assembler")
+  if (NOT NASM)
+    message(FATAL_ERROR "nasm assembler not found: required to build")
+  endif (NOT NASM)
+  message(STATUS "Found nasm: ${NASM}")
+  if (X64)
+    set(ASM_FLAGS "${ASM_FLAGS} -fmacho64")
+  else (X64)
+    set(ASM_FLAGS "${ASM_FLAGS} -fmacho32")
+  endif (X64)
+  if (DEBUG)
+    set(ASM_FLAGS "${ASM_FLAGS} -g")
+  endif (DEBUG)
+elseif (UNIX)
   set(ASM_FLAGS "-mmnemonic=intel -msyntax=intel -mnaked-reg --noexecstack")
   if (X64)
     set(ASM_FLAGS "${ASM_FLAGS} --64")
@@ -173,7 +188,7 @@ if (UNIX)
   if (DEBUG)
     set(ASM_FLAGS "${ASM_FLAGS} -g")
   endif (DEBUG)
-else (UNIX)
+else ()
   if (X64)
     find_program(CMAKE_ASM_COMPILER ml64.exe HINTS "${cl_path}" DOC "path to assembler")
   else (X64)
@@ -189,11 +204,11 @@ else (UNIX)
     set(ASM_DBG "")
   endif ()
   set(ASM_FLAGS "/nologo ${ASM_DBG}")
-endif (UNIX)
+endif ()
 
 if (APPLE)
-  # XXX: xcode assembler doesn't seem to take any kind of --help or --version
-  # so for now we assume it supports Intel
+  # XXX: xcode assembler uses Intel mnemonics but opposite src,dst order!
+  # We rely on having nasm installed as a workaround.
   set(CMAKE_ASM_SUPPORTS_INTEL_SYNTAX ON)
 endif (APPLE)
 if (UNIX AND NOT APPLE)
@@ -230,7 +245,14 @@ endif (UNIX AND NOT APPLE)
 ##################################################
 # Assembler build rule for Makefile generators
 
-if (UNIX)
+if (APPLE)
+  # Despite the docs, -o does not work: cpp prints to stdout.
+  set(CMAKE_ASM_COMPILE_OBJECT
+    "${CMAKE_CPP} ${CMAKE_CPP_FLAGS} <FLAGS> <DEFINES> -E <SOURCE> > <OBJECT>.s"
+    "<CMAKE_COMMAND> -Dfile=<OBJECT>.s -P \"${cpp2asm_newline_script_path}\""
+    "<NASM> ${ASM_FLAGS} -o <OBJECT> <OBJECT>.s"
+    )
+elseif (UNIX)
   # we used to have ".ifdef FOO" and to not have it turn into ".ifdef 1" we'd say
   # "-DFOO=FOO", but we now use exclusively preprocessor defines, which is good
   # since our defines are mostly in configure.h where we can't as easily tweak them
@@ -253,7 +275,7 @@ if (UNIX)
     # comes up empty for me.
     "<CMAKE_ASM_COMPILER> ${ASM_FLAGS} -o <OBJECT> <OBJECT>.s"
     )
-else (UNIX)
+else ()
   # Even if we didn't preprocess we'd need our own rule since cmake doesn't
   # support ml.
   set(CMAKE_ASM_COMPILE_OBJECT
@@ -269,7 +291,7 @@ else (UNIX)
     "<CMAKE_COMMAND> -Dfile=<OBJECT>.s -P \"${cpp2asm_newline_script_path}\""
     "<CMAKE_ASM_COMPILER> ${ASM_FLAGS} /c /Fo<OBJECT> <OBJECT>.s"
     )
-endif (UNIX)
+endif ()
 
 ##################################################
 # Assembler build commands for Visual Studio generators and