fine-tuning x86 gcc codegen

2 files changed, 35 insertions, 20 deletions

diff --git a/src/mesa/tnl/t_vtx_x86.c b/src/mesa/tnl/t_vtx_x86.c
index d082ef82bb5..6ff1a52743d 100644
--- a/src/mesa/tnl/t_vtx_x86.c
+++ b/src/mesa/tnl/t_vtx_x86.c
@@ -109,6 +109,25 @@ do {							\
 
 
 
+#define FIXUPREL( CODE, KNOWN_OFFSET, CHECKVAL, NEWVAL )\
+do {							\
+   GLuint subst = 0x10101010 + CHECKVAL;		\
+							\
+   if (DONT_KNOW_OFFSETS) {				\
+      while (*(int *)(CODE+offset) != subst) offset++;	\
+      *(int *)(CODE+offset) = (int)(NEWVAL) - ((int)(CODE)+offset) - 4; \
+      if (0) fprintf(stderr, "%s/%d: offset %d, new value: 0x%x\n", __FILE__, __LINE__, offset, (int)(NEWVAL) - ((int)(CODE)+offset) - 4); \
+      offset += 4;					\
+   }							\
+   else {						\
+      int *icode = (int *)(CODE+KNOWN_OFFSET);		\
+      assert (*icode == subst);				\
+      *icode = (int)(NEWVAL) - (int)(icode) - 4;	\
+   }							\
+} while (0)
+
+
+
 
 /* Build specialized versions of the immediate calls on the fly for
  * the current state.  Generic x86 versions.
@@ -125,7 +144,7 @@ static struct _tnl_dynfn *makeX86Vertex1fv( GLcontext *ctx, int vertex_size )
    FIXUP(dfn->code, 0, 0, (int)&tnl->vtx.vbptr);
    FIXUP(dfn->code, 0, 3, (int)&tnl->vtx.counter);
    FIXUP(dfn->code, 0, 3, (int)&tnl->vtx.counter);
-   FIXUP(dfn->code, 0, 4, (int)&notify);
+   FIXUPREL(dfn->code, 0, 4, (int)&notify);
 
    return dfn;
 }
@@ -141,7 +160,7 @@ static struct _tnl_dynfn *makeX86Vertex2fv( GLcontext *ctx, int vertex_size )
    FIXUP(dfn->code, 0, 0, (int)&tnl->vtx.vbptr);
    FIXUP(dfn->code, 0, 3, (int)&tnl->vtx.counter);
    FIXUP(dfn->code, 0, 3, (int)&tnl->vtx.counter);
-   FIXUP(dfn->code, 0, 4, (int)&notify);
+   FIXUPREL(dfn->code, 0, 4, (int)&notify);
 
    return dfn;
 }
@@ -157,7 +176,7 @@ static struct _tnl_dynfn *makeX86Vertex3fv( GLcontext *ctx, int vertex_size )
    FIXUP(dfn->code, 0, 0, (int)&tnl->vtx.vbptr);
    FIXUP(dfn->code, 0, 3, (int)&tnl->vtx.counter);
    FIXUP(dfn->code, 0, 3, (int)&tnl->vtx.counter);
-   FIXUP(dfn->code, 0, 4, (int)&notify);
+   FIXUPREL(dfn->code, 0, 4, (int)&notify);
    return dfn;
 }
 
@@ -172,7 +191,7 @@ static struct _tnl_dynfn *makeX86Vertex4fv( GLcontext *ctx, int vertex_size )
    FIXUP(dfn->code, 0, 0, (int)&tnl->vtx.vbptr);
    FIXUP(dfn->code, 0, 3, (int)&tnl->vtx.counter);
    FIXUP(dfn->code, 0, 3, (int)&tnl->vtx.counter);
-   FIXUP(dfn->code, 0, 4, (int)&notify);
+   FIXUPREL(dfn->code, 0, 4, (int)&notify);
 
    return dfn;
 }
diff --git a/src/mesa/tnl/t_vtx_x86_gcc.S b/src/mesa/tnl/t_vtx_x86_gcc.S
index c06fef6c47d..937b53bfd3f 100644
--- a/src/mesa/tnl/t_vtx_x86_gcc.S
+++ b/src/mesa/tnl/t_vtx_x86_gcc.S
@@ -66,10 +66,9 @@ GLOBL ( _x86_Vertex1fv )
 	pop	%edi
 	dec	%edx		// counter--
 	movl	%edx, SUBST(3)	// 0x3 --> counter 
-	je	.5		// if (counter != 0)
-	ret			//    return
-.5:	mov     $SUBST(4), %eax	// else notify()
-        jmp     *%eax           // jmp $0x10101014 doesn't seem to work
+	.short	0x840f		// if (counter == 0)
+	.long	SUBST(4)	//    notify()
+	ret			// return
 GLOBL ( _x86_Vertex1fv_end )
 
 
@@ -94,10 +93,9 @@ GLOBL ( _x86_Vertex2fv )
 	pop	%edi
 	dec	%edx		// counter--
 	movl	%edx, SUBST(3)	// save counter
-	je	.6		// if (counter != 0)
-	ret			//    return
-.6:	mov     $SUBST(4), %eax	// else notify()
-        jmp     *%eax           // jmp $0x10101014 doesn't seem to work
+	.short	0x840f		// if (counter == 0)
+	.long	SUBST(4)	//    notify()
+	ret			// return
 GLOBL ( _x86_Vertex2fv_end )
 
 .align 4
@@ -123,10 +121,9 @@ GLOBL ( _x86_Vertex3fv )
 	pop	%edi
 	dec	%edx		// counter--
 	movl	%edx, SUBST(3)	// save counter
-	je	.7		// if (counter != 0)
-	ret			//    return
-.7:	mov     $SUBST(4), %eax	// else notify()
-        jmp     *%eax           // jmp $0x10101014 doesn't seem to work
+	.short	0x840f		// if (counter == 0)
+	.long	SUBST(4)	//    notify()
+	ret			// return
 GLOBL ( _x86_Vertex3fv_end )
 
 			
@@ -155,10 +152,9 @@ GLOBL ( _x86_Vertex4fv )
 	pop	%edi
 	dec	%edx		// counter--
 	movl	%edx, SUBST(3)	// save counter
-	je	.6		// if (counter != 0)
-	ret			//    return
-.8:	mov     $SUBST(4), %eax	// else notify()
-        jmp     *%eax           // jmp $0x10101014 doesn't seem to work
+	.short	0x840f		// if (counter == 0)
+	.long	SUBST(4)	//    notify()
+	ret			// return
 GLOBL ( _x86_Vertex4fv_end )