codegen'ed versions of the 2nd level dispatch

5 files changed, 101 insertions, 12 deletions

diff --git a/src/mesa/tnl/t_vtx_api.c b/src/mesa/tnl/t_vtx_api.c
index 6dcd8b43a0b..807d99952f0 100644
--- a/src/mesa/tnl/t_vtx_api.c
+++ b/src/mesa/tnl/t_vtx_api.c
@@ -901,6 +901,7 @@ void _tnl_vtx_init( GLcontext *ctx )
    _tnl_current_init( ctx );
    _tnl_exec_vtxfmt_init( ctx );
    _tnl_generic_exec_vtxfmt_init( ctx );
+   _tnl_x86_exec_vtxfmt_init( ctx ); /* [dBorca] x86 DISPATCH_ATTRFV */
 
    _mesa_install_exec_vtxfmt( ctx, &tnl->exec_vtxfmt );
 
diff --git a/src/mesa/tnl/t_vtx_api.h b/src/mesa/tnl/t_vtx_api.h
index 85cc33f0718..46700fcd0ae 100644
--- a/src/mesa/tnl/t_vtx_api.h
+++ b/src/mesa/tnl/t_vtx_api.h
@@ -78,6 +78,8 @@ extern void _tnl_generic_attr_table_init( attrfv_func (*tab)[4] );
  */
 extern void _tnl_InitX86Codegen( struct _tnl_dynfn_generators *gen );
 
+extern void _tnl_x86_exec_vtxfmt_init( GLcontext *ctx );
+
 
 
 
diff --git a/src/mesa/tnl/t_vtx_generic.c b/src/mesa/tnl/t_vtx_generic.c
index 25dd07a52c7..00dd2e89075 100644
--- a/src/mesa/tnl/t_vtx_generic.c
+++ b/src/mesa/tnl/t_vtx_generic.c
@@ -409,8 +409,7 @@ static void GLAPIENTRY _tnl_VertexAttrib4fvNV( GLuint index,
 
 
 /* Install the generic versions of the 2nd level dispatch functions.
- * There's currently no codegen alternative to these, though one is in
- * the works.
+ * [dBorca] Some of these have a codegen alternative.
  */
 void _tnl_generic_exec_vtxfmt_init( GLcontext *ctx )
 {
diff --git a/src/mesa/tnl/t_vtx_x86.c b/src/mesa/tnl/t_vtx_x86.c
index 6ff1a52743d..3f3a198a243 100644
--- a/src/mesa/tnl/t_vtx_x86.c
+++ b/src/mesa/tnl/t_vtx_x86.c
@@ -59,8 +59,6 @@ EXTERN( _x86_Vertex2fv );
 EXTERN( _x86_Vertex3fv );
 EXTERN( _x86_Vertex4fv );
 
-/* None of these used yet:
- */
 EXTERN( _x86_dispatch_attrf );
 EXTERN( _x86_dispatch_attrfv );
 EXTERN( _x86_dispatch_multitexcoordf );
@@ -260,6 +258,77 @@ void _do_choose( void )
 {
 }
 
+
+/* [dBorca] I purposely avoided one single macro, since they might need to
+ * be handled in different ways. Ohwell, once things get much clearer, they
+ * could collapse...
+ */
+#define MAKE_DISPATCH_ATTR(FUNC, SIZE, TYPE, ATTR)			\
+do {									\
+   char *code;								\
+   char *start = (char *)&_x86_dispatch_attr##TYPE;			\
+   char *end = (char *)&_x86_dispatch_attr##TYPE##_end;			\
+   int offset = 0;							\
+   code = ALIGN_MALLOC( end - start, 16 );				\
+   memcpy (code, start, end - start);					\
+   FIXUP(code, 0, 0, (int)&(TNL_CONTEXT(ctx)->vtx.tabfv[ATTR][SIZE-1]));\
+   vfmt->FUNC##SIZE##TYPE = code;					\
+} while (0)
+
+
+#define MAKE_DISPATCH_MULTITEXCOORD(FUNC, SIZE, TYPE, ATTR)		\
+do {									\
+   char *code;								\
+   char *start = (char *)&_x86_dispatch_multitexcoord##TYPE;		\
+   char *end = (char *)&_x86_dispatch_multitexcoord##TYPE##_end;	\
+   int offset = 0;							\
+   code = ALIGN_MALLOC( end - start, 16 );				\
+   memcpy (code, start, end - start);					\
+   FIXUP(code, 0, 0, (int)&(TNL_CONTEXT(ctx)->vtx.tabfv[_TNL_ATTRIB_TEX0][SIZE-1]));\
+   vfmt->FUNC##SIZE##TYPE##ARB = code;					\
+} while (0)
+
+
+#define MAKE_DISPATCH_VERTEXATTRIB(FUNC, SIZE, TYPE, ATTR)		\
+do {									\
+   char *code;								\
+   char *start = (char *)&_x86_dispatch_vertexattrib##TYPE;		\
+   char *end = (char *)&_x86_dispatch_vertexattrib##TYPE##_end;		\
+   int offset = 0;							\
+   code = ALIGN_MALLOC( end - start, 16 );				\
+   memcpy (code, start, end - start);					\
+   FIXUP(code, 0, 0, (int)&(TNL_CONTEXT(ctx)->vtx.tabfv[0][SIZE-1]));	\
+   vfmt->FUNC##SIZE##TYPE##NV = code;					\
+} while (0)
+
+/* [dBorca] Install the codegen'ed versions of the 2nd level dispatch
+ * functions.  We should keep a list and free them in the end...
+ */
+void _tnl_x86_exec_vtxfmt_init( GLcontext *ctx )
+{
+   GLvertexformat *vfmt = &(TNL_CONTEXT(ctx)->exec_vtxfmt);
+
+   MAKE_DISPATCH_ATTR(Color,3,f,     _TNL_ATTRIB_COLOR0);
+   MAKE_DISPATCH_ATTR(Color,3,fv,    _TNL_ATTRIB_COLOR0);
+   MAKE_DISPATCH_ATTR(Color,4,f,     _TNL_ATTRIB_COLOR0);
+   MAKE_DISPATCH_ATTR(Color,4,fv,    _TNL_ATTRIB_COLOR0);
+   MAKE_DISPATCH_ATTR(Normal,3,f,    _TNL_ATTRIB_NORMAL);
+   MAKE_DISPATCH_ATTR(Normal,3,fv,   _TNL_ATTRIB_NORMAL);
+   MAKE_DISPATCH_ATTR(TexCoord,2,f,  _TNL_ATTRIB_TEX0);
+   MAKE_DISPATCH_ATTR(TexCoord,2,fv, _TNL_ATTRIB_TEX0);
+   MAKE_DISPATCH_ATTR(Vertex,3,f,    _TNL_ATTRIB_POS);
+   MAKE_DISPATCH_ATTR(Vertex,3,fv,   _TNL_ATTRIB_POS);
+   /* just add more */
+
+   MAKE_DISPATCH_MULTITEXCOORD(MultiTexCoord,2,f,  0);
+   MAKE_DISPATCH_MULTITEXCOORD(MultiTexCoord,2,fv, 0);
+   /* just add more */
+
+   MAKE_DISPATCH_VERTEXATTRIB(VertexAttrib,2,f,  0);
+   MAKE_DISPATCH_VERTEXATTRIB(VertexAttrib,2,fv, 0);
+   /* just add more */
+}
+
 #else 
 
 void _tnl_InitX86Codegen( struct _tnl_dynfn_generators *gen )
diff --git a/src/mesa/tnl/t_vtx_x86_gcc.S b/src/mesa/tnl/t_vtx_x86_gcc.S
index 937b53bfd3f..e932faff757 100644
--- a/src/mesa/tnl/t_vtx_x86_gcc.S
+++ b/src/mesa/tnl/t_vtx_x86_gcc.S
@@ -31,11 +31,25 @@ USE OR OTHER DEALINGS IN THE SOFTWARE.
  */
 
 #if defined(USE_X86_ASM) && !defined(HAVE_NONSTANDARD_GLAPIENTRY)
-			
+
+#if !defined (__DJGPP__)
+
 #define GLOBL( x )	\
 .globl x;		\
 x:
 
+#define EXTRN( x )	x
+
+#else  /* defined(__DJGPP__) */
+
+#define GLOBL( x )	\
+.globl _##x;		\
+_##x:
+
+#define EXTRN( x )	_##x
+
+#endif /* defined(__DJGPP__) */
+
 .data
 .align 4
 
@@ -48,6 +62,10 @@ x:
 	
 
 
+// [dBorca] TODO
+// Unfold functions for each vertex size?
+// Build super-specialized MMX/SSE versions?
+
 GLOBL ( _x86_Vertex1fv )
 	movl    4(%esp), %ecx	
 	push	%edi
@@ -217,9 +235,9 @@ GLOBL( _x86_choose_fv)
 	subl	$12, %esp	// gcc does 16 byte alignment of stack frames?
 	movl	$SUBST(0), (%esp)	// arg 0 - attrib
 	movl	$SUBST(1), 4(%esp)	// arg 1 - N
-	call    _do_choose	// new function returned in %eax
-	add     $12, %esp	// tear down stack frame
-	jmp     *%eax		// jump to new func
+	call    EXTRN(_do_choose)	// new function returned in %eax
+	add     $12, %esp		// tear down stack frame
+	jmp     *%eax			// jump to new func
 GLOBL ( _x86_choosefv_end )
 	
 	
@@ -251,7 +269,7 @@ GLOBL( _x86_dispatch_attrf )
 	subl	$12, %esp	// gcc does 16 byte alignment of stack frames?
 	leal	16(%esp), %edx	// address of first float on stack
 	movl	%edx, (%esp)	// save as 'v'
-	call	SUBST(0)	// 0x0 --> tabfv[attr][n]
+	call	*SUBST(0)	// 0x0 --> tabfv[attr][n]
 	addl	$12, %esp	// tear down frame
 	ret			// return
 GLOBL( _x86_dispatch_attrf_end )
@@ -259,7 +277,7 @@ GLOBL( _x86_dispatch_attrf_end )
 // The fv case is simpler:
 // 
 GLOBL( _x86_dispatch_attrfv )
-	jmp	SUBST(0)	// 0x0 --> tabfv[attr][n]
+	jmp	*SUBST(0)	// 0x0 --> tabfv[attr][n]
 GLOBL( _x86_dispatch_attrfv_end )
 
 
@@ -294,7 +312,7 @@ GLOBL( _x86_dispatch_vertexattribf )
 	movl	$16, %ecx
 	movl	4(%esp), %eax
 	cmpl	$16, %eax
-	cmovge	%ecx, %eax
+	cmovge	%ecx, %eax	// [dBorca] BADBAD! might not be supported
 	leal	8(%esp), %ecx	// calculate 'v'
 	movl	%ecx, 4(%esp)	// save in 1st arg slot
 	sall	$4, %eax
@@ -305,7 +323,7 @@ GLOBL( _x86_dispatch_vertexattribfv )
 	movl	$16, %ecx
 	movl	4(%esp), %eax
 	cmpl	$16, %eax
-	cmovge	%ecx, %eax
+	cmovge	%ecx, %eax	// [dBorca] BADBAD! might not be supported
 	movl	8(%esp), %ecx	// load 'v'
 	movl	%ecx, 4(%esp)	// save in 1st arg slot
 	sall	$4, %eax