From cff862344f02df4679e057673a477d0b9deec23d Mon Sep 17 00:00:00 2001
From: Pavel Krajcevski <pavel@cs.unc.edu>
Date: Tue, 28 Aug 2012 15:36:26 -0400
Subject: [PATCH] Add compiler specific syntax for inline assembly.

---
 BPTCEncoder/CMakeLists.txt        | 78 +++++++++++++++++++++++++++++--
 BPTCEncoder/config/BC7Config.h.in |  1 +
 BPTCEncoder/src/RGBAEndpoints.cpp | 21 ++++++---
 3 files changed, 89 insertions(+), 11 deletions(-)

diff --git a/BPTCEncoder/CMakeLists.txt b/BPTCEncoder/CMakeLists.txt
index 604258c..06fc889 100644
--- a/BPTCEncoder/CMakeLists.txt
+++ b/BPTCEncoder/CMakeLists.txt
@@ -5,7 +5,67 @@ INCLUDE(CheckCXXSourceRuns)
 
 SET(OLD_CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS})
 IF(CMAKE_COMPILER_IS_GNUCC)
-	SET(CMAKE_REQUIRED_FLAGS -msse4.1 -E)
+
+        # Test whether or not the compiler allows inline
+	# assmbly...
+	CHECK_CXX_SOURCE_RUNS("
+            \#ifdef _MSC_VER
+             int main() {
+             int x = 1, y = 0;
+               __asm {
+                mov eax, x
+                mov y, eax
+               }
+             return !y;
+             }
+            \#else
+             int main() {
+             int x = 1, y = 0;
+             __asm__ (\"movl %1, %%eax;\"
+                      \"movl %%eax, %0;\"
+                       :\"=r\"(y)        /* output */
+                       :\"r\"(x)         /* input */
+                       :\"%eax\"         /* clobbered register */
+             ); 
+
+             return !y;
+             }
+            \#endif"
+            HAS_INLINE_ASSEMBLY
+        )
+
+        # If the compiler doesn't allow it, then try with a
+        # compiler flag...
+	IF( NOT HAS_INLINE_ASSEMBLY )
+          SET(CMAKE_REQUIRED_FLAGS -fasm-blocks)
+          CHECK_CXX_SOURCE_RUNS("
+            \#ifdef _MSC_VER
+             int main() {
+             int x = 1, y = 0;
+               __asm {
+                mov eax, x
+                mov y, eax
+               }
+             return !y;
+             }
+            \#else
+             int main() {
+             int x = 1, y = 0;
+             __asm__ (\"movl %1, %%eax;\"
+                      \"movl %%eax, %0;\"
+                       :\"=r\"(y)        /* output */
+                       :\"r\"(x)         /* input */
+                       :\"%eax\"         /* clobbered register */
+             ); 
+
+             return !y;
+             }
+            \#endif"
+            HAS_INLINE_ASSEMBLY_WITH_FLAGS
+          )
+        ENDIF()
+
+	SET(CMAKE_REQUIRED_FLAGS -msse4.1)
 	CHECK_CXX_SOURCE_RUNS("
 		#include <smmintrin.h>
 		int main() {
@@ -36,15 +96,15 @@ ELSEIF(MSVC)
 ENDIF()
 SET(CMAKE_REQUIRED_FLAGS ${OLD_CMAKE_REQUIRED_FLAGS})
 
+IF( NOT HAS_INLINE_ASSEMBLY AND NOT HAS_INLINE_ASSEMBLY_WITH_FLAGS )
+  SET( NO_INLINE_ASSEMBLY true )
+ENDIF()
+
 CONFIGURE_FILE(
 	"config/BC7Config.h.in"
 	"include/BC7Config.h"
 )
 
-IF(CMAKE_COMPILER_IS_GNUCC)
-	ADD_DEFINITIONS(-fasm-blocks)
-ENDIF(CMAKE_COMPILER_IS_GNUCC)
-
 SET( HEADERS
 	src/BC7CompressionMode.h
 	src/BC7IntTypes.h
@@ -86,6 +146,14 @@ IF( HAS_SSE_41 )
 	)
 ENDIF( HAS_SSE_41 )
 
+IF( HAS_INLINE_ASSEMBLY_WITH_FLAGS )
+  IF( MSVC )
+    # !FIXME!
+  ELSE()
+    ADD_DEFINITIONS( -fasm-blocks )
+  ENDIF()
+ENDIF()
+
 ADD_LIBRARY( BPTCEncoder
 	${HEADERS}
 	${SOURCES}
diff --git a/BPTCEncoder/config/BC7Config.h.in b/BPTCEncoder/config/BC7Config.h.in
index 599835a..174a962 100644
--- a/BPTCEncoder/config/BC7Config.h.in
+++ b/BPTCEncoder/config/BC7Config.h.in
@@ -5,4 +5,5 @@
 // explicitly by the CMake build process.
 
 // Do we have the proper popcnt instruction defined?
+#cmakedefine NO_INLINE_ASSEMBLY
 #cmakedefine HAS_SSE_POPCNT
diff --git a/BPTCEncoder/src/RGBAEndpoints.cpp b/BPTCEncoder/src/RGBAEndpoints.cpp
index 0b7c541..fd64498 100755
--- a/BPTCEncoder/src/RGBAEndpoints.cpp
+++ b/BPTCEncoder/src/RGBAEndpoints.cpp
@@ -15,6 +15,7 @@
 //
 //--------------------------------------------------------------------------------------
 
+#include "BC7Config.h"
 #include "BC7IntTypes.h"
 #include "RGBAEndpoints.h"
 #include "BC7Compressor.h"
@@ -66,7 +67,7 @@ static const float kFloatConversion[256] = {
 ///////////////////////////////////////////////////////////////////////////////
 static inline uint32 CountBitsInMask(uint8 n) {
 
-#if _WIN64
+#if defined(_WIN64) || defined(NO_INLINE_ASSEMBLY)
 	if(!n) return 0; // no bits set
 	if(!(n & (n-1))) return 1; // power of two
 
@@ -76,15 +77,23 @@ static inline uint32 CountBitsInMask(uint8 n) {
 	}
 	return c;
 #else
-
-	__asm
-	{
+#ifdef _MSC_VER
+	__asm {
 		mov eax, 8
 		movzx ecx, n
 		bsf ecx, ecx
 		sub eax, ecx
-	}
-
+        }
+#else
+	__asm__("mov %%eax, 8;"
+		"movzbl %0, %%ecx;"
+		"bsf %%ecx, %%ecx;"
+		"sub %%eax, %%ecx;"
+		: // No output registers
+		: "r"(n)
+		: "%eax", "%ecx"
+	);
+#endif	
 #endif
 }