Stefan Allius writes:

    I attached a patch, which revise the clone.S and vfork.S:
        - Use PIC code.
        - include new file syscall.S, so we can simply make a branch to
          __syscall_error instead of a PLT/GOT call
        - call errno_location to store the syscall error (for pthreads)
        - avoid to use the 'shad' statement on SH2 targets
        - call fork if vfork isn't available
        - some cleanups and optimization

libc/sysdeps/linux/sh/clone.S

@@ -35,7 +35,6 @@
 /* int clone(int (*fn)(void *arg), void *child_stack, int flags, void *arg); */
 
         .text
-.extern __syscall_error
 
 .text
 .align 4
@@ -44,71 +43,50 @@
 __clone:
 	/* sanity check arguments.  */
 	tst	r4, r4
+	bt	0f
+	tst	r5, r5
 	bf/s	1f
-	 tst	r5, r5
-	bf/s	1f
-	 mov.l	.L1, r1
-#ifdef __HAVE_SHARED__
-	mov.l	r12, @-r15
-	sts.l	pr, @-r15
-        mov.l	.LG, r12
-	mova	.LG, r0
-	add	r0, r12
-	mova	.L1, r0
-	add	r0, r1
-	jsr	@r1
+	 mov	#+__NR_clone, r3
+0:		
+	bra __syscall_error
 	 mov	#-EINVAL, r4
-	lds.l	@r15+, pr
-	rts
-	 mov.l	@r15+, r12
-#else
-	jmp	@r1
-	 mov	#-EINVAL, r4
-#endif
-	.align	2
-.L1:
-	.long	PLTJMP(__syscall_error)
+
 1:
 	/* insert the args onto the new stack */
 	mov.l	r7, @-r5
 	/* save the function pointer as the 0th element */
 	mov.l	r4, @-r5
-
+	
 	/* do the system call */
 	mov	r6, r4
-	mov	#+__NR_clone, r3
 	trapa	#0x12
 	mov     r0, r1
+#ifdef __sh2__
+// 12 arithmetic shifts for the crappy sh2, because shad doesn't exist!	
+	shar	r1
+	shar	r1
+	shar	r1
+	shar	r1
+	shar	r1
+	shar	r1
+	shar	r1
+	shar	r1
+	shar	r1
+	shar	r1
+	shar	r1
+	shar	r1
+#else		
 	mov	#-12, r2
 	shad	r2, r1
+#endif
 	not	r1, r1			// r1=0 means r0 = -1 to -4095
 	tst	r1, r1			// i.e. error in linux
-	bf	2f
-	mov.l	.L2, r1
-#ifdef __HAVE_SHARED__
-	mov	r0, r4
-	mov.l	r12, @-r15
-	sts.l	pr, @-r15
-        mov.l	.LG, r12
-	mova	.LG, r0
-	add	r0, r12
-	mova	.L2, r0
-	add	r0, r1
-	jsr	@r1
-	 nop
-	lds.l	@r15+, pr
-	rts
-	 mov.l	@r15+, r12
-#else
-	jmp	@r1
+	bf/s	2f
+	 tst	r0, r0
+        bra __syscall_error
 	 mov	r0, r4
-#endif
-	.align	2
-.L2:
-	.long	PLTJMP(__syscall_error)
 
 2:
-	tst	r0, r0
 	bt	3f
 	rts
 	 nop
@@ -119,15 +97,15 @@ __clone:
 	 mov.l	@(4,r15), r4
 
 	/* we are done, passing the return value through r0  */
-	mov.l	.L3, r1
-#ifdef __HAVE_SHARED__
+	mov.l	.L1, r1
+#if defined __HAVE_ELF__ && defined __HAVE_SHARED__ 
 	mov.l	r12, @-r15
 	sts.l	pr, @-r15
 	mov	r0, r4
-	mova	.LG, r0
+	mova	.LG, r0  /* .LG from syscall_error.S */
 	mov.l	.LG, r12
 	add	r0, r12
-	mova	.L3, r0
+	mova	.L1, r0
 	add	r0, r1
 	jsr	@r1
 	 nop
@@ -139,11 +117,11 @@ __clone:
 	 mov	r0, r4
 #endif
 	.align	2
-.LG:
-	.long	_GLOBAL_OFFSET_TABLE_
-.L3:
+.L1:
 	.long	PLTJMP(_exit)
 .size __clone,.-__clone;
 
 .globl  clone;
     clone = __clone
+
+#include "syscall_error.S"

libc/sysdeps/linux/sh/syscall_error.S

@@ -0,0 +1,37 @@
+	.align 4
+__syscall_error:
+	/* Call errno_location, store '-r4' in errno and return -1 */
+	mov.l	r12, @-r15
+	sts.l	pr, @-r15
+#if defined __HAVE_ELF__ && defined __HAVE_SHARED__ 
+	mova	.LG, r0
+	mov.l	.LG, r12
+	add	r0, r12
+	mov.l	1f, r0
+	mov.l	@(r0,r12),r0
+ 	jsr	@r0
+	 neg	r4, r12
+#else	
+	mov.l	1f, r0
+	bsrf	r0
+	 neg	r4, r12
+.jmp_loc:		
+#endif
+	mov.l	r12, @r0
+	lds.l	@r15+, pr
+	mov.l	@r15+,r12
+
+	/* And just kick back a -1.  */
+	rts
+	 mov	#-1, r0
+
+	.align	4
+
+#if defined __HAVE_ELF__ && defined __HAVE_SHARED__ 
+1:	.long   __errno_location@GOT
+.LG:	.long	_GLOBAL_OFFSET_TABLE_
+#else
+1:	.long   __errno_location - .jmp_loc
+#endif		
+
+

libc/sysdeps/linux/sh/vfork.S

@@ -20,6 +20,9 @@
  respective copyright holders.
 */
 
+#include <features.h>
+#define _SYSCALL_H
+#include <bits/sysnum.h>
 #define _ERRNO_H	1
 #include <bits/errno.h>
 
@@ -28,26 +31,15 @@
    replaced by a call to `execve'.  Return -1 for errors, 0 to the new process,
    and the process ID of the new process to the old process.  */
 
-.global errno
-
 .text
 .align 4
 .type	__vfork,@function
 .globl	__vfork;
 __vfork:
-	mov.l	@r15+,r3		// pop value from the stack
-	mov.l	.L5,r1
-	mov.l	r3,@r1			// save it in .sav_stack
-
-	mov.w	.L3, r3
-
-#ifdef HIOS
-	trapa	#0x28
-#else
+	mov.w	.L2, r3
 	trapa	#0x10
-#endif
-
 	mov     r0, r1
+#ifdef __sh2__
 // 12 arithmetic shifts for the crappy sh2, because shad doesn't exist!	
 	shar	r1
 	shar	r1
@@ -61,66 +53,60 @@ __vfork:
 	shar	r1
 	shar	r1
 	shar	r1
+#else		
+	mov	#-12, r2
+	shad	r2, r1
+#endif
 
-//	mov	#-12, r2
-//	shad	r2, r1
 	not	r1, r1			// r1=0 means r0 = -1 to -4095
 	tst	r1, r1			// i.e. error in linux
-	bf	1f
+	bf	2f
 	mov.w	.L1, r1
 	cmp/eq	r1, r0
-	bt	2f
-	mov.l	.L2, r1
-	jmp	@r1
+	bf/s	__syscall_error
 	 mov	r0, r4
 
-	.align	4
-
-1:
-	mov.l	.L5,r1
-	mov.l	@r1,r3			// get it from .sav_stack
-	mov.l	r3,@-r15		// restore value to the stack
-
-	rts
-	 nop
-
-	.align	4
+	/* If we don't have vfork, use fork.  */
+	mov.w	.L3, r3
+	trapa	#0x10
+	mov     r0, r1
+#ifdef __sh2__
+// 12 arithmetic shifts for the crappy sh2, because shad doesn't exist!	
+	shar	r1
+	shar	r1
+	shar	r1
+	shar	r1
+	shar	r1
+	shar	r1
+	shar	r1
+	shar	r1
+	shar	r1
+	shar	r1
+	shar	r1
+	shar	r1
+#else		
+	mov	#-12, r2
+	shad	r2, r1
+#endif
 
+	not	r1, r1			// r1=0 means r0 = -1 to -4095
+	tst	r1, r1			// i.e. error in linux
+	bt/s	__syscall_error
+	 mov	r0, r4
 2:
-.global __syscall_error
-__syscall_error:
-	/* Store it in errno... */
-	mov.l	.L4, r1
-	mov.l	r0, @r1
-
-	mov.l	.L5,r1
-	mov.l	@r1,r3			// get it from .sav_stack
-	mov.l	r3,@-r15		// restore value to the stack
-
-	/* And just kick back a -1.  */
 	rts
-	 mov	#-1, r0
+	 nop
 
-	.align	4
+	.align	2
 .L1:
 	.word	-ENOSYS
-.L3:
-	.word	190			//__NR_vfork
-
-	.align	4			// Shouldn't be necessary as previously with have two words
 .L2:
-	.long	__syscall_error
-
-.L4:	.long	errno
-
-.L5:	.long	.sav_stack
-
-	.data
-
-	.align 4
+	.word	__NR_vfork
+.L3:
+	.word	__NR_fork
+        .size   __vfork, .-__vfork
+.weak	vfork
+	vfork =  __vfork
 
-.sav_stack:					//area to temporary save the stach
-	.long	0
+#include "syscall_error.S"
 
-.weak	vfork
-vfork = __vfork