Merge pull request #47 from notaz/asm_rework

Asm rework
icculus · May 9, 2016 · a8c6c77 · a8c6c77
2 parents 056d77f + ce46bd1
commit a8c6c77
Show file tree

Hide file tree

Showing 24 changed files with 1,080 additions and 1,296 deletions.
diff --git a/Sources/CMakeLists.txt b/Sources/CMakeLists.txt
@@ -188,13 +188,20 @@ else()
     set(DEBUGSUFFIX "")
 endif()
 
-# This should not be needed anymore, but might be faster on 32bit x86
-option(USE_I386_ASM "Use X86 ASM" FALSE)
+option(USE_ASM "Use ASM code" TRUE)
+if (USE_ASM)
+    MESSAGE(STATUS "Using assembler code (when available)")
+else()
+    add_definitions(-DUSE_PORTABLE_C=1)
+    MESSAGE(STATUS "Using portable C instead of all ASM")
+endif()
 
-if (USE_I386_ASM)
+option(USE_I386_NASM_ASM "Use i386 nasm ASM code" FALSE)
+
+if (USE_ASM AND USE_I386_NASM_ASM)
     # You need the Netwide Assembler (NASM) to build this on Intel systems.
     #   http://nasm.sf.net/
-    add_definitions(-DUSE_I386_ASM=1)
+    add_definitions(-DUSE_I386_NASM_ASM=1)
     if (MACOSX)
         set(ASMOBJFMT "macho")
         list(APPEND ASMFLAGS --prefix _)
@@ -203,10 +210,9 @@ if (USE_I386_ASM)
     else()
         set(ASMOBJFMT "elf")
     endif()
-    MESSAGE(STATUS "Using i386 assembler")
+    MESSAGE(STATUS "Using i386 nasm ASM")
 else()
-    add_definitions(-DUSE_PORTABLE_C=1)
-    MESSAGE(STATUS "Using portable C instead of ASM")
+    MESSAGE(STATUS "Not using i386 nasm ASM")
 endif()
 
 option(PANDORA "Compile for Pandora" FALSE)
@@ -655,7 +661,7 @@ add_dependencies(${SHADERSLIB} ParseEntities)
 add_parser_and_scanner("Engine/Base/Parser" "Engine/Base/Scanner")
 add_parser_and_scanner("Engine/Ska/smcPars" "Engine/Ska/smcScan")
 
-if (USE_I386_ASM)
+if (USE_I386_NASM_ASM)
     add_custom_command(
         OUTPUT "SoundMixer386.o"
         MAIN_DEPENDENCY "${CMAKE_CURRENT_SOURCE_DIR}/Engine/Sound/SoundMixer386.asm"

diff --git a/Sources/Engine/Base/Base.h b/Sources/Engine/Base/Base.h
@@ -65,9 +65,7 @@ with this program; if not, write to the Free Software Foundation, Inc.,
 #else
   #warning "UNKNOWN PLATFORM IDENTIFIED!!!!"
   #define PLATFORM_UNKNOWN 1
-  #warning "USING PORTABLE C!!!"
-  #define USE_PORTABLE_C
-#endif 
+#endif
 
 #if PLATFORM_LINUX || PLATFORM_MACOSX
   #ifndef PLATFORM_UNIX

diff --git a/Sources/Engine/Base/Profiling.cpp b/Sources/Engine/Base/Profiling.cpp
@@ -21,24 +21,13 @@ with this program; if not, write to the Free Software Foundation, Inc.,
 template class CStaticArray<CProfileCounter>;
 template class CStaticArray<CProfileTimer>;
 
-#if (defined USE_PORTABLE_C)
+#if (defined PLATFORM_UNIX) && !defined(__GNU_INLINE_X86_32__)
 #include <sys/time.h>
 #endif
 
 static inline __int64 ReadTSC_profile(void)
 {
-#if (defined USE_PORTABLE_C)
-  #ifdef __arm__
-  struct timespec tv;
-  clock_gettime(CLOCK_MONOTONIC, &tv);
-  return( (((__int64) tv.tv_sec) * 1000) + (((__int64) tv.tv_nsec) / 1000000) );
-  #else
-  struct timeval tv;
-  gettimeofday(&tv, NULL);
-  return( (((__int64) tv.tv_sec) * 1000) + (((__int64) tv.tv_usec) / 1000) );
-  #endif
-
-#elif (defined __MSVC_INLINE__)
+#if (defined __MSVC_INLINE__)
   __int64 mmRet;
   __asm {
     rdtsc
@@ -47,7 +36,7 @@ static inline __int64 ReadTSC_profile(void)
   }
   return mmRet;
 
-#elif (defined __GNU_INLINE__)
+#elif (defined __GNU_INLINE_X86_32__)
   __int64 mmRet;
   __asm__ __volatile__ (
     "rdtsc                    \n\t"
@@ -60,7 +49,16 @@ static inline __int64 ReadTSC_profile(void)
   return(mmRet);
 
 #else
-  #error Please implement for your platform/compiler.
+  #ifdef __arm__
+  struct timespec tv;
+  clock_gettime(CLOCK_MONOTONIC, &tv);
+  return( (((__int64) tv.tv_sec) * 1000) + (((__int64) tv.tv_nsec) / 1000000) );
+  #else
+  struct timeval tv;
+  gettimeofday(&tv, NULL);
+  return( (((__int64) tv.tv_sec) * 1000) + (((__int64) tv.tv_usec) / 1000) );
+  #endif
+
 #endif
 }
 

diff --git a/Sources/Engine/Base/Timer.cpp b/Sources/Engine/Base/Timer.cpp
@@ -29,7 +29,7 @@ with this program; if not, write to the Free Software Foundation, Inc.,
 #include <Engine/Base/Priority.inl>
 
 // !!! FIXME: use SDL timer code instead and rdtsc never?
-#if (USE_PORTABLE_C) 
+#if (defined PLATFORM_UNIX) && !defined(__GNU_INLINE_X86_32__)
 #define USE_GETTIMEOFDAY 1
 #endif
 
@@ -64,7 +64,7 @@ static inline __int64 ReadTSC(void)
   }
   return mmRet;
 
-#elif (defined __GNU_INLINE__)
+#elif (defined __GNU_INLINE_X86_32__)
   __int64 mmRet;
   __asm__ __volatile__ (
     "rdtsc                    \n\t"

diff --git a/Sources/Engine/Base/Types.h b/Sources/Engine/Base/Types.h
@@ -109,6 +109,30 @@ MY_STATIC_ASSERT(size_tSize, sizeof(size_t) == sizeof(void*));
   #define ASMSYM(x) #x
 #endif
 
+/* should we enable inline asm? */
+#ifndef USE_PORTABLE_C
+  #if defined(__MSVC_INLINE__)
+    /* the build system selected __MSVC_INLINE__ */
+  #elif defined(__GNU_INLINE_X86_32__)
+    /* the build system selected __GNU_INLINE_X86_32__ */
+  #elif defined(_MSC_VER) && defined(_M_IX86)
+    #define __MSVC_INLINE__
+  #elif defined (__GNUC__) && defined(__i386)
+    #define __GNU_INLINE_X86_32__
+  #elif defined (__GNUC__) && defined(__x86_64__)
+    #define __GNU_INLINE_X86_64__
+  #endif
+
+  #if defined(__GNU_INLINE_X86_32__) || defined(__GNU_INLINE_X86_64__)
+    #define __GNU_INLINE_X86__
+  #endif
+
+  #if defined(__GNU_INLINE_X86__)
+    #define FPU_REGS "st", "st(1)", "st(2)", "st(3)", "st(4)", "st(5)", "st(6)", "st(7)"
+    #define MMX_REGS "mm0", "mm1", "mm2", "mm3", "mm4", "mm5", "mm6", "mm7"
+  #endif
+#endif
+
 #ifdef PLATFORM_UNIX  /* rcg10042001 */
     #include <stdio.h>
     #include <string.h>
@@ -134,25 +158,6 @@ MY_STATIC_ASSERT(size_tSize, sizeof(size_t) == sizeof(void*));
       #endif
     #endif
 
-    #if ((defined __GNUC__) && (!defined __GNU_INLINE__))
-      #define __GNU_INLINE__
-    #endif
-
-    #if (defined __INTEL_COMPILER)
-      #if ((!defined __GNU_INLINE__) && (!defined __MSVC_INLINE__))
-        #error Please define __GNU_INLINE__ or __MSVC_INLINE__ with Intel C++.
-      #endif
-
-      #if ((defined __GNU_INLINE__) && (defined __MSVC_INLINE__))
-        #error Define either __GNU_INLINE__ or __MSVC_INLINE__ with Intel C++.
-      #endif
-    #endif
-
-    #if defined(__GNU_INLINE__) && defined(__i386__)
-      #define FPU_REGS "st", "st(1)", "st(2)", "st(3)", "st(4)", "st(5)", "st(6)", "st(7)"
-      #define MMX_REGS "mm0", "mm1", "mm2", "mm3", "mm4", "mm5", "mm6", "mm7"
-    #endif
-
     #ifndef PAGESIZE
       #define PAGESIZE 4096
     #endif
@@ -230,10 +235,7 @@ MY_STATIC_ASSERT(size_tSize, sizeof(size_t) == sizeof(void*));
 
     inline ULONG _rotl(ULONG ul, int bits)
     {
-        #if (defined USE_PORTABLE_C)
-            // DG: according to http://blog.regehr.org/archives/1063 this is fast
-            return (ul<<bits) | (ul>>(-bits&31));
-        #elif (defined __GNU_INLINE__)
+        #if (defined __GNU_INLINE_X86_32__)
             // This, on the other hand, is wicked fast.  :)
             __asm__ __volatile__ (
                 "roll %%cl, %%eax    \n\t"
@@ -255,7 +257,8 @@ MY_STATIC_ASSERT(size_tSize, sizeof(size_t) == sizeof(void*));
             return(ul);
 
         #else
-            #error need inline asm for your platform.
+            // DG: according to http://blog.regehr.org/archives/1063 this is fast
+            return (ul<<bits) | (ul>>(-bits&31));
         #endif
     }
 

diff --git a/Sources/Engine/Engine.cpp b/Sources/Engine/Engine.cpp
@@ -125,14 +125,10 @@ BOOL APIENTRY DllMain( HANDLE hModule, DWORD  ul_reason_for_call, LPVOID lpReser
 
 static void DetectCPU(void)
 {
-#if (defined USE_PORTABLE_C)  // rcg10072001
-  CPrintF(TRANSV("  (No CPU detection in this binary.)\n"));
-
-#else
-  char strVendor[12+1];
+  char strVendor[12+1] = { 0 };
   strVendor[12] = 0;
-  ULONG ulTFMS;
-  ULONG ulFeatures;
+  ULONG ulTFMS = 0;
+  ULONG ulFeatures = 0;
 
   #if (defined __MSVC_INLINE__)
   // test MMX presence and update flag
@@ -148,43 +144,47 @@ static void DetectCPU(void)
     mov     dword ptr [ulFeatures], edx
   }
 
-  #elif (defined __GNU_INLINE__)
+  #elif (defined __GNU_INLINE_X86__)
+    ULONG eax, ebx, ecx, edx;
     // test MMX presence and update flag
     __asm__ __volatile__ (
-        "pushl   %%ebx            \n\t"
-        "xorl    %%eax,%%eax      \n\t"  // request for basic id
+    #if (defined __GNU_INLINE_X86_64__)
+        "cpuid                    \n\t"
+            : "=a" (eax), "=b" (ebx), "=c" (ecx), "=d" (edx)
+    #else
+        "movl    %%ebx, %%esi     \n\t"
         "cpuid                    \n\t"
-        "movl    %%ebx,  (%%esi)  \n\t"
-        "movl    %%edx, 4(%%esi)  \n\t"
-        "movl    %%ecx, 8(%%esi)  \n\t"
-        "popl    %%ebx            \n\t"
-            : // no specific outputs.
-            : "S" (strVendor)
-            : "eax", "ecx", "edx", "memory"
+        "xchgl   %%ebx, %%esi     \n\t"
+            : "=a" (eax), "=S" (ebx), "=c" (ecx), "=d" (edx)
+    #endif
+            : "a" (0) // request for basic id
     );
-
-        // need to break this into a separate asm block, since I'm clobbering
-        //  too many registers. There's something to be said for letting MSVC
-        //  figure out where on the stack your locals are resting, but yeah,
-        //  I know, that's x86-specific anyhow...
-        // !!! FIXME: can probably do this right with modern GCC.
+    memcpy(strVendor + 0, &ebx, 4);
+    memcpy(strVendor + 4, &edx, 4);
+    memcpy(strVendor + 8, &ecx, 4);
 
     __asm__ __volatile__ (
-        "pushl   %%ebx                  \n\t"
-        "movl    $1, %%eax              \n\t"  // request for TFMS feature flags
-        "cpuid                          \n\t"
-        "mov     %%eax, (%%esi)         \n\t"  // remember type, family, model and stepping
-        "mov     %%edx, (%%edi)         \n\t"
-        "popl    %%ebx                  \n\t"
-            : // no specific outputs.
-            : "S" (&ulTFMS), "D" (&ulFeatures)
-            : "eax", "ecx", "edx", "memory"
+    #if (defined __GNU_INLINE_X86_64__)
+        "cpuid                    \n\t"
+            : "=a" (eax), "=b" (ebx), "=c" (ecx), "=d" (edx)
+    #else
+        "movl    %%ebx, %%esi     \n\t"
+        "cpuid                    \n\t"
+        "xchgl   %%ebx, %%esi     \n\t"
+            : "=a" (eax), "=S" (ebx), "=c" (ecx), "=d" (edx)
+    #endif
+            : "a" (1) // request for TFMS feature flags
     );
+    ulTFMS = eax;
+    ulFeatures = edx;
 
-  #else
-    #error Please implement for your platform or define USE_PORTABLE_C.
   #endif
 
+  if (ulTFMS == 0) {
+    CPrintF(TRANSV("  (No CPU detection in this binary.)\n"));
+    return;
+  }
+
   INDEX iType     = (ulTFMS>>12)&0x3;
   INDEX iFamily   = (ulTFMS>> 8)&0xF;
   INDEX iModel    = (ulTFMS>> 4)&0xF;
@@ -215,8 +215,6 @@ static void DetectCPU(void)
   sys_iCPUMHz = INDEX(_pTimer->tm_llCPUSpeedHZ/1E6);
 
   if( !bMMX) FatalError( TRANS("MMX support required but not present!"));
-
-#endif  // defined USE_PORTABLE_C
 }
 
 static void DetectCPUWrapper(void)