added codegen'ed choosers

6 files changed, 171 insertions, 75 deletions

diff --git a/src/mesa/tnl/t_save_api.c b/src/mesa/tnl/t_save_api.c
index db338bbef6c..18fa46951d0 100644
--- a/src/mesa/tnl/t_save_api.c
+++ b/src/mesa/tnl/t_save_api.c
@@ -712,6 +712,15 @@ do {						\
 
 #define DISPATCH_ATTR1F( ATTR, S ) DISPATCH_ATTRFV( ATTR, 1, &(S) )
 
+#ifdef USE_X86_ASM
+/* Naughty cheat:
+ */
+#define DISPATCH_ATTR2F( ATTR, S,T ) DISPATCH_ATTRFV( ATTR, 2, &(S) )
+#define DISPATCH_ATTR3F( ATTR, S,T,R ) DISPATCH_ATTRFV( ATTR, 3, &(S) )
+#define DISPATCH_ATTR4F( ATTR, S,T,R,Q ) DISPATCH_ATTRFV( ATTR, 4, &(S) )
+#else
+/* Safe:
+ */
 #define DISPATCH_ATTR2F( ATTR, S,T ) 		\
 do { 						\
    GLfloat v[2]; 				\
@@ -730,6 +739,7 @@ do { 						\
    v[0] = S; v[1] = T; v[2] = R; v[3] = Q;	\
    DISPATCH_ATTR4FV( ATTR, v );			\
 } while (0)
+#endif
 
 
 static void enum_error( void )
diff --git a/src/mesa/tnl/t_vtx_api.c b/src/mesa/tnl/t_vtx_api.c
index 807d99952f0..f47114cf623 100644
--- a/src/mesa/tnl/t_vtx_api.c
+++ b/src/mesa/tnl/t_vtx_api.c
@@ -883,6 +883,8 @@ void _tnl_vtx_init( GLcontext *ctx )
       choose[ERROR_ATTRIB][2] = error_attrib;
       choose[ERROR_ATTRIB][3] = error_attrib;
 
+      _tnl_x86choosers(choose, do_choose); /* x86 INIT_CHOOSERS */
+
       _tnl_generic_attr_table_init( generic_attr_func );
    }
 
@@ -901,7 +903,7 @@ 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 */
+   _tnl_x86_exec_vtxfmt_init( ctx ); /* 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 46700fcd0ae..f58461332ec 100644
--- a/src/mesa/tnl/t_vtx_api.h
+++ b/src/mesa/tnl/t_vtx_api.h
@@ -80,6 +80,10 @@ extern void _tnl_InitX86Codegen( struct _tnl_dynfn_generators *gen );
 
 extern void _tnl_x86_exec_vtxfmt_init( GLcontext *ctx );
 
+extern void _tnl_x86choosers( attrfv_func (*choose)[4],
+			      attrfv_func (*do_choose)( GLuint attr,
+							GLuint sz ));
+
 
 
 
diff --git a/src/mesa/tnl/t_vtx_generic.c b/src/mesa/tnl/t_vtx_generic.c
index 00dd2e89075..ea03dff05bd 100644
--- a/src/mesa/tnl/t_vtx_generic.c
+++ b/src/mesa/tnl/t_vtx_generic.c
@@ -151,6 +151,15 @@ do {						\
 
 #define DISPATCH_ATTR1F( ATTR, S ) DISPATCH_ATTRFV( ATTR, 1, &(S) )
 
+#ifdef USE_X86_ASM
+/* Naughty cheat:
+ */
+#define DISPATCH_ATTR2F( ATTR, S,T ) DISPATCH_ATTRFV( ATTR, 2, &(S) )
+#define DISPATCH_ATTR3F( ATTR, S,T,R ) DISPATCH_ATTRFV( ATTR, 3, &(S) )
+#define DISPATCH_ATTR4F( ATTR, S,T,R,Q ) DISPATCH_ATTRFV( ATTR, 4, &(S) )
+#else
+/* Safe:
+ */
 #define DISPATCH_ATTR2F( ATTR, S,T ) 		\
 do { 						\
    GLfloat v[2]; 				\
@@ -169,6 +178,7 @@ do { 						\
    v[0] = S; v[1] = T; v[2] = R; v[3] = Q;	\
    DISPATCH_ATTR4FV( ATTR, v );			\
 } while (0)
+#endif
 
 
 static void GLAPIENTRY _tnl_Vertex2f( GLfloat x, GLfloat y )
@@ -408,8 +418,8 @@ static void GLAPIENTRY _tnl_VertexAttrib4fvNV( GLuint index,
 }
 
 
-/* Install the generic versions of the 2nd level dispatch functions.
- * [dBorca] Some of these have a codegen alternative.
+/* Install the generic versions of the 2nd level dispatch
+ * functions.  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 4f04a4efbb3..66950e70e19 100644
--- a/src/mesa/tnl/t_vtx_x86.c
+++ b/src/mesa/tnl/t_vtx_x86.c
@@ -28,6 +28,7 @@ USE OR OTHER DEALINGS IN THE SOFTWARE.
 /*
  * Authors:
  *   Keith Whitwell <[email protected]>
+ *   Daniel Borca <[email protected]>
  */
 
 
@@ -66,6 +67,8 @@ EXTERN( _tnl_x86_dispatch_multitexcoordfv );
 EXTERN( _tnl_x86_dispatch_vertexattribf );
 EXTERN( _tnl_x86_dispatch_vertexattribfv );
 
+EXTERN( _tnl_x86_choose_fv );
+
 
 static void notify( void )
 {
@@ -77,7 +80,7 @@ static void notify( void )
 
 
 #define DFN( FUNC, CACHE, KEY )				\
-   struct _tnl_dynfn *dfn = MALLOC_STRUCT( _tnl_dynfn );          \
+   struct _tnl_dynfn *dfn = MALLOC_STRUCT( _tnl_dynfn );\
    char *start = (char *)&FUNC;				\
    char *end = (char *)&FUNC##_end;			\
    int offset = 0;               			\
@@ -254,20 +257,23 @@ void _tnl_InitX86Codegen( struct _tnl_dynfn_generators *gen )
    gen->Attribute[3] = makeX86Attribute4fv;
 }
 
-void _do_choose( void )
+
+static attrfv_func
+_do_choose( GLuint attr, GLuint sz )
 {
+   return NULL;
 }
 
 
-/* [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...
+/* 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 *)&_tnl_x86_dispatch_attr##TYPE;			\
-   char *end = (char *)&_tnl_x86_dispatch_attr##TYPE##_end;			\
+   char *end = (char *)&_tnl_x86_dispatch_attr##TYPE##_end;		\
    int offset = 0;							\
    code = ALIGN_MALLOC( end - start, 16 );				\
    memcpy (code, start, end - start);					\
@@ -279,7 +285,7 @@ do {									\
 #define MAKE_DISPATCH_MULTITEXCOORD(FUNC, SIZE, TYPE, ATTR)		\
 do {									\
    char *code;								\
-   char *start = (char *)&_tnl_x86_dispatch_multitexcoord##TYPE;		\
+   char *start = (char *)&_tnl_x86_dispatch_multitexcoord##TYPE;	\
    char *end = (char *)&_tnl_x86_dispatch_multitexcoord##TYPE##_end;	\
    int offset = 0;							\
    code = ALIGN_MALLOC( end - start, 16 );				\
@@ -293,7 +299,7 @@ do {									\
 do {									\
    char *code;								\
    char *start = (char *)&_tnl_x86_dispatch_vertexattrib##TYPE;		\
-   char *end = (char *)&_tnl_x86_dispatch_vertexattrib##TYPE##_end;		\
+   char *end = (char *)&_tnl_x86_dispatch_vertexattrib##TYPE##_end;	\
    int offset = 0;							\
    code = ALIGN_MALLOC( end - start, 16 );				\
    memcpy (code, start, end - start);					\
@@ -301,7 +307,8 @@ do {									\
    vfmt->FUNC##SIZE##TYPE##NV = code;					\
 } while (0)
 
-/* [dBorca] Install the codegen'ed versions of the 2nd level dispatch
+
+/* 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 )
@@ -312,21 +319,70 @@ void _tnl_x86_exec_vtxfmt_init( GLcontext *ctx )
    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);
+/* vfmt->FogCoordfEXT = _tnl_FogCoordfEXT;
+   vfmt->FogCoordfvEXT = _tnl_FogCoordfvEXT;*/
    MAKE_DISPATCH_ATTR(Normal,3,f,    _TNL_ATTRIB_NORMAL);
    MAKE_DISPATCH_ATTR(Normal,3,fv,   _TNL_ATTRIB_NORMAL);
+/* vfmt->SecondaryColor3fEXT = _tnl_SecondaryColor3fEXT;
+   vfmt->SecondaryColor3fvEXT = _tnl_SecondaryColor3fvEXT; */
+   MAKE_DISPATCH_ATTR(TexCoord,1,f,  _TNL_ATTRIB_TEX0);
+   MAKE_DISPATCH_ATTR(TexCoord,1,fv, _TNL_ATTRIB_TEX0);
    MAKE_DISPATCH_ATTR(TexCoord,2,f,  _TNL_ATTRIB_TEX0);
    MAKE_DISPATCH_ATTR(TexCoord,2,fv, _TNL_ATTRIB_TEX0);
+   MAKE_DISPATCH_ATTR(TexCoord,3,f,  _TNL_ATTRIB_TEX0);
+   MAKE_DISPATCH_ATTR(TexCoord,3,fv, _TNL_ATTRIB_TEX0);
+   MAKE_DISPATCH_ATTR(TexCoord,4,f,  _TNL_ATTRIB_TEX0);
+   MAKE_DISPATCH_ATTR(TexCoord,4,fv, _TNL_ATTRIB_TEX0);
+   MAKE_DISPATCH_ATTR(Vertex,2,f,    _TNL_ATTRIB_POS);
+   MAKE_DISPATCH_ATTR(Vertex,2,fv,   _TNL_ATTRIB_POS);
    MAKE_DISPATCH_ATTR(Vertex,3,f,    _TNL_ATTRIB_POS);
    MAKE_DISPATCH_ATTR(Vertex,3,fv,   _TNL_ATTRIB_POS);
-   /* just add more */
+   MAKE_DISPATCH_ATTR(Vertex,4,f,    _TNL_ATTRIB_POS);
+   MAKE_DISPATCH_ATTR(Vertex,4,fv,   _TNL_ATTRIB_POS);
 
+   MAKE_DISPATCH_MULTITEXCOORD(MultiTexCoord,1,f,  0);
+   MAKE_DISPATCH_MULTITEXCOORD(MultiTexCoord,1,fv, 0);
    MAKE_DISPATCH_MULTITEXCOORD(MultiTexCoord,2,f,  0);
    MAKE_DISPATCH_MULTITEXCOORD(MultiTexCoord,2,fv, 0);
-   /* just add more */
+   MAKE_DISPATCH_MULTITEXCOORD(MultiTexCoord,3,f,  0);
+   MAKE_DISPATCH_MULTITEXCOORD(MultiTexCoord,3,fv, 0);
+   MAKE_DISPATCH_MULTITEXCOORD(MultiTexCoord,4,f,  0);
+   MAKE_DISPATCH_MULTITEXCOORD(MultiTexCoord,4,fv, 0);
 
+   MAKE_DISPATCH_VERTEXATTRIB(VertexAttrib,1,f,  0);
+   MAKE_DISPATCH_VERTEXATTRIB(VertexAttrib,1,fv, 0);
    MAKE_DISPATCH_VERTEXATTRIB(VertexAttrib,2,f,  0);
    MAKE_DISPATCH_VERTEXATTRIB(VertexAttrib,2,fv, 0);
-   /* just add more */
+   MAKE_DISPATCH_VERTEXATTRIB(VertexAttrib,3,f,  0);
+   MAKE_DISPATCH_VERTEXATTRIB(VertexAttrib,3,fv, 0);
+   MAKE_DISPATCH_VERTEXATTRIB(VertexAttrib,4,f,  0);
+   MAKE_DISPATCH_VERTEXATTRIB(VertexAttrib,4,fv, 0);
+}
+
+
+/* Install the codegen'ed choosers.
+ * We should keep a list and free them in the end...
+ */
+void _tnl_x86choosers( attrfv_func (*choose)[4],
+		       attrfv_func (*do_choose)( GLuint attr,
+						 GLuint sz ))
+{
+   int attr, size;
+
+   for (attr = 0; attr < _TNL_MAX_ATTR_CODEGEN; attr++) {
+      for (size = 0; size < 4; size++) {
+         char *code;
+         char *start = (char *)&_tnl_x86_choose_fv;
+         char *end = (char *)&_tnl_x86_choose_fv_end;
+         int offset = 0;
+         code = ALIGN_MALLOC( end - start, 16 );
+         memcpy (code, start, end - start);
+         FIXUP(code, 0, 0, attr);
+         FIXUP(code, 0, 1, size + 1);
+         FIXUPREL(code, 0, 2, do_choose);
+         choose[attr][size] = code;
+      }
+   }
 }
 
 #else 
@@ -336,9 +392,19 @@ void _tnl_InitX86Codegen( struct _tnl_dynfn_generators *gen )
    (void) gen;
 }
 
+
 void _tnl_x86_exec_vtxfmt_init( GLcontext *ctx )
 {
    (void) ctx;
 }
 
+
+void _tnl_x86choosers( attrfv_func (*choose)[4],
+		       attrfv_func (*do_choose)( GLuint attr,
+						 GLuint sz ))
+{
+   (void) choose;
+   (void) do_choose;
+}
+
 #endif
diff --git a/src/mesa/tnl/t_vtx_x86_gcc.S b/src/mesa/tnl/t_vtx_x86_gcc.S
index 2a2e933f971..5a1adc0f33d 100644
--- a/src/mesa/tnl/t_vtx_x86_gcc.S
+++ b/src/mesa/tnl/t_vtx_x86_gcc.S
@@ -36,16 +36,12 @@ USE OR OTHER DEALINGS IN THE SOFTWARE.
 .globl x;		\
 x:
 
-#define EXTRN( x )	x
-
 #else  /* defined(__DJGPP__) */
 
 #define GLOBL( x )	\
 .globl _##x;		\
 _##x:
 
-#define EXTRN( x )	_##x
-
 #endif /* defined(__DJGPP__) */
 
 .data
@@ -55,17 +51,22 @@ _##x:
 // macro to note current offsets, etc in a special region of the
 // object file & just make everything work out neat.  I don't know
 // enough to do that...
-	
-#define SUBST( x ) (0x10101010 + x)	
-	
+
+#define SUBST( x ) (0x10101010 + x)
+
 
 
 // [dBorca] TODO
 // Unfold functions for each vertex size?
 // Build super-specialized MMX/SSE versions?
+// STDCALL woes (HAVE_NONSTANDARD_GLAPIENTRY):
+//   need separate routine for the non "fv" case,
+//   to clean up the stack (I guess we could codegen
+//   'ret nn' insn)! Also we need to call notify, then
+//   return, instead of jump!
 
 GLOBL ( _tnl_x86_Vertex1fv )
-	movl    4(%esp), %ecx	
+	movl	4(%esp), %ecx
 	push	%edi
 	push	%esi
 	movl	SUBST(0), %edi	// 0x0 --> tnl->vtx.vbptr
@@ -75,7 +76,7 @@ GLOBL ( _tnl_x86_Vertex1fv )
 	movl	$SUBST(1), %ecx	// 0x1 --> (tnl->vtx.vertex_size - 1)
 	movl	$SUBST(2), %esi	// 0x2 --> (tnl->vtx.vertex + 1)
 	repz
-	movsl   %ds:(%esi), %es:(%edi)
+	movsl	%ds:(%esi), %es:(%edi)
 	movl	%edi, SUBST(0)	// 0x0 --> tnl->vtx.vbptr
 	movl	SUBST(3), %edx	// 0x3 --> counter
 	pop	%esi
@@ -90,7 +91,7 @@ GLOBL ( _tnl_x86_Vertex1fv_end )
 
 .align 4
 GLOBL ( _tnl_x86_Vertex2fv )
-	movl    4(%esp), %ecx	
+	movl	4(%esp), %ecx
 	push	%edi
 	push	%esi
 	movl	SUBST(0), %edi	// load tnl->vtx.vbptr
@@ -102,7 +103,7 @@ GLOBL ( _tnl_x86_Vertex2fv )
 	movl	$SUBST(1), %ecx	// vertex_size - 2
 	movl	$SUBST(2), %esi	// tnl->vtx.vertex + 2
 	repz
-	movsl %ds:(%esi), %es:(%edi)
+	movsl	%ds:(%esi), %es:(%edi)
 	movl	%edi, SUBST(0)	// save tnl->vtx.vbptr
 	movl	SUBST(3), %edx	// load counter
 	pop	%esi
@@ -116,7 +117,7 @@ GLOBL ( _tnl_x86_Vertex2fv_end )
 
 .align 4
 GLOBL ( _tnl_x86_Vertex3fv )
-	movl    4(%esp), %ecx	
+	movl	4(%esp), %ecx
 	push	%edi
 	push	%esi
 	movl	SUBST(0), %edi	// load tnl->vtx.vbptr
@@ -130,7 +131,7 @@ GLOBL ( _tnl_x86_Vertex3fv )
 	movl	$SUBST(1), %ecx	// vertex_size - 3
 	movl	$SUBST(2), %esi	// tnl->vtx.vertex + 3
 	repz
-	movsl %ds:(%esi), %es:(%edi)
+	movsl	%ds:(%esi), %es:(%edi)
 	movl	%edi, SUBST(0)	// save tnl->vtx.vbptr
 	movl	SUBST(3), %edx	// load counter
 	pop	%esi
@@ -142,10 +143,10 @@ GLOBL ( _tnl_x86_Vertex3fv )
 	ret			// return
 GLOBL ( _tnl_x86_Vertex3fv_end )
 
-			
+
 .align 4
 GLOBL ( _tnl_x86_Vertex4fv )
-	movl    4(%esp), %ecx	
+	movl	4(%esp), %ecx
 	push	%edi
 	push	%esi
 	movl	SUBST(0), %edi	// load tnl->vtx.vbptr
@@ -161,7 +162,7 @@ GLOBL ( _tnl_x86_Vertex4fv )
 	movl	$SUBST(1), %ecx	// vertex_size - 4
 	movl	$SUBST(2), %esi	// tnl->vtx.vertex + 3
 	repz
-	movsl   %ds:(%esi), %es:(%edi)
+	movsl	%ds:(%esi), %es:(%edi)
 	movl	%edi, SUBST(0)	// save tnl->vtx.vbptr
 	movl	SUBST(3), %edx	// load counter
 	pop	%esi
@@ -174,49 +175,49 @@ GLOBL ( _tnl_x86_Vertex4fv )
 GLOBL ( _tnl_x86_Vertex4fv_end )
 
 
-	
+
 /**
  * Generic handlers for vector format data. 
  */
 
 GLOBL( _tnl_x86_Attribute1fv)
-	movl 4(%esp), %ecx	
-	movl (%ecx), %eax       /* load v[0] */
-	movl %eax, SUBST(0)    	/* store v[0] to current vertex */
+	movl	4(%esp), %ecx
+	movl	(%ecx), %eax	/* load v[0] */
+	movl	%eax, SUBST(0)	/* store v[0] to current vertex */
 	ret
 GLOBL ( _tnl_x86_Attribute1fv_end )
 
 GLOBL( _tnl_x86_Attribute2fv)
-	movl 4(%esp), %ecx	
-	movl (%ecx), %eax       /* load v[0] */
-	movl 4(%ecx), %edx      /* load v[1] */
-	movl %eax, SUBST(0)    	/* store v[0] to current vertex */
-	movl %edx, SUBST(1)    	/* store v[1] to current vertex */
+	movl	4(%esp), %ecx
+	movl	(%ecx), %eax	/* load v[0] */
+	movl	4(%ecx), %edx	/* load v[1] */
+	movl	%eax, SUBST(0)	/* store v[0] to current vertex */
+	movl	%edx, SUBST(1)	/* store v[1] to current vertex */
 	ret
 GLOBL ( _tnl_x86_Attribute2fv_end )
 
 
 GLOBL( _tnl_x86_Attribute3fv)
-	movl 4(%esp), %ecx	
-	movl (%ecx), %eax       /* load v[0] */
-	movl 4(%ecx), %edx      /* load v[1] */
-	movl 8(%ecx), %ecx      /* load v[2] */
-	movl %eax, SUBST(0)    	/* store v[0] to current vertex */
-	movl %edx, SUBST(1)    	/* store v[1] to current vertex */
-	movl %ecx, SUBST(2)   	/* store v[2] to current vertex */
+	movl	4(%esp), %ecx
+	movl	(%ecx), %eax	/* load v[0] */
+	movl	4(%ecx), %edx	/* load v[1] */
+	movl	8(%ecx), %ecx	/* load v[2] */
+	movl	%eax, SUBST(0)	/* store v[0] to current vertex */
+	movl	%edx, SUBST(1)	/* store v[1] to current vertex */
+	movl	%ecx, SUBST(2)	/* store v[2] to current vertex */
 	ret
 GLOBL ( _tnl_x86_Attribute3fv_end )
 
 GLOBL( _tnl_x86_Attribute4fv)
-	movl 4(%esp), %ecx	
-	movl (%ecx), %eax       /* load v[0] */
-	movl 4(%ecx), %edx      /* load v[1] */
-	movl %eax, SUBST(0)    	/* store v[0] to current vertex */
-	movl %edx, SUBST(1)    	/* store v[1] to current vertex */
-	movl 8(%ecx), %eax      /* load v[2] */
-	movl 12(%ecx), %edx     /* load v[3] */
-	movl %eax, SUBST(2)    	/* store v[2] to current vertex */
-	movl %edx, SUBST(3)    	/* store v[3] to current vertex */
+	movl	4(%esp), %ecx
+	movl	(%ecx), %eax	/* load v[0] */
+	movl	4(%ecx), %edx	/* load v[1] */
+	movl	%eax, SUBST(0)	/* store v[0] to current vertex */
+	movl	%edx, SUBST(1)	/* store v[1] to current vertex */
+	movl	8(%ecx), %eax	/* load v[2] */
+	movl	12(%ecx), %edx	/* load v[3] */
+	movl	%eax, SUBST(2)	/* store v[2] to current vertex */
+	movl	%edx, SUBST(3)	/* store v[3] to current vertex */
 	ret
 GLOBL ( _tnl_x86_Attribute4fv_end )
 
@@ -225,29 +226,24 @@ GLOBL ( _tnl_x86_Attribute4fv_end )
 
 // Must generate all of these ahead of first usage.  Generate at
 // compile-time?  
-	
-// NOT CURRENTLY USED
 
 
 GLOBL( _tnl_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    EXTRN(_do_choose)	// new function returned in %eax
-	add     $12, %esp		// tear down stack frame
-	jmp     *%eax			// jump to new func
-GLOBL ( _tnl_x86_choosefv_end )
-	
-	
+	.byte	0xe8			// call ...
+	.long	SUBST(2)		// ... do_choose
+	add	$12, %esp		// tear down stack frame
+	jmp	*%eax			// jump to new func
+GLOBL ( _tnl_x86_choose_fv_end )
+
 
 
 // FIRST LEVEL FUNCTIONS -- these are plugged directly into GL dispatch.
-	
 
-// NOT CURRENTLY USED
 
-	
-		
+
 // In the 1st level dispatch functions, switch to a different
 // calling convention -- (const GLfloat *v) in %ecx.
 // 
@@ -256,7 +252,7 @@ GLOBL ( _tnl_x86_choosefv_end )
 // back to the original caller.
 
 
-	
+
 // Vertex/Normal/Color, etc: the address of the function pointer
 // is known at codegen time.
 
@@ -282,6 +278,13 @@ GLOBL( _tnl_x86_dispatch_attrfv_end )
 // MultiTexcoord: the address of the function pointer must be
 // calculated, but can use the index argument slot to hold 'v', and
 // avoid setting up a new stack frame.
+//
+// [dBorca]
+// right, this would be the preferred approach, but gcc does not
+// clean up the stack after each function call when optimizing (-fdefer-pop);
+// can it make assumptions about what's already on the stack?  I dunno,
+// but in this case, we can't mess with the caller's stack frame, and
+// we must use a model like `_x86_dispatch_attrfv' above.  Caveat emptor!
 
 // Also, will only need a maximum of four of each of these per context:
 // 
@@ -302,15 +305,16 @@ GLOBL( _tnl_x86_dispatch_multitexcoordfv )
 	sall	$4, %ecx
 	jmp	*SUBST(0)(%ecx)	// 0x0 - tabfv[tex0][n]
 GLOBL( _tnl_x86_dispatch_multitexcoordfv_end )
-				
+
 // VertexAttrib: the address of the function pointer must be
 // calculated.
 
 GLOBL( _tnl_x86_dispatch_vertexattribf )
-	movl	$16, %ecx
 	movl	4(%esp), %eax
 	cmpl	$16, %eax
-	cmovge	%ecx, %eax	// [dBorca] BADBAD! might not be supported
+	jb	.0		// "cmovge" is not supported on all CPUs
+	movl	$16, %eax
+.0:
 	leal	8(%esp), %ecx	// calculate 'v'
 	movl	%ecx, 4(%esp)	// save in 1st arg slot
 	sall	$4, %eax
@@ -318,13 +322,13 @@ GLOBL( _tnl_x86_dispatch_vertexattribf )
 GLOBL( _tnl_x86_dispatch_vertexattribf_end )
 
 GLOBL( _tnl_x86_dispatch_vertexattribfv )
-	movl	$16, %ecx
 	movl	4(%esp), %eax
 	cmpl	$16, %eax
-	cmovge	%ecx, %eax	// [dBorca] BADBAD! might not be supported
+	jb	.1		// "cmovge" is not supported on all CPUs
+	movl	$16, %eax
+.1:
 	movl	8(%esp), %ecx	// load 'v'
 	movl	%ecx, 4(%esp)	// save in 1st arg slot
 	sall	$4, %eax
 	jmp	*SUBST(0)(%eax)	// 0x0 - tabfv[0][n]
 GLOBL( _tnl_x86_dispatch_vertexattribfv_end )
-