sh specific bits needed for nptl

 * unified atomic.h compare and exchange macros
 * clone.S with RESET_PID support
 * sh specific versions of pread/pwrite with cancellation support
 * check SHARED instead of PIC

Signed-off-by: Austin Foxley <austinf@cetoncorp.com>

1 files changed, 195 insertions, 321 deletions

diff --git a/libc/sysdeps/linux/sh/bits/atomic.h b/libc/sysdeps/linux/sh/bits/atomic.h
index 6bb7255c5..dd6e5f97d 100644
--- a/libc/sysdeps/linux/sh/bits/atomic.h
+++ b/libc/sysdeps/linux/sh/bits/atomic.h
@@ -54,6 +54,10 @@ typedef uintmax_t uatomic_max_t;
     Japan. http://lc.linux.or.jp/lc2002/papers/niibe0919h.pdf (in
     Japanese).
 
+    Niibe Yutaka, "gUSA: User Space Atomicity with Little Kernel
+    Modification", LinuxTag 2003, Rome.
+    http://www.semmel.ch/Linuxtag-DVD/talks/170/paper.html (in English).
+
     B.N. Bershad, D. Redell, and J. Ellis, "Fast Mutual Exclusion for
     Uniprocessors",  Proceedings of the Fifth Architectural Support for
     Programming Languages and Operating Systems (ASPLOS), pp. 223-233,
@@ -65,56 +69,44 @@ typedef uintmax_t uatomic_max_t;
       r1:     saved stack pointer
 */
 
-#define __arch_compare_and_exchange_val_8_acq(mem, newval, oldval) \
-  ({ __typeof (*(mem)) __result; \
-     __asm__ __volatile__ ("\
+/* Avoid having lots of different versions of compare and exchange,
+   by having this one complicated version. Parameters:
+      bwl:     b, w or l for 8, 16 and 32 bit versions.
+      version: val or bool, depending on whether the result is the
+               previous value or a bool indicating whether the transfer
+               did happen (note this needs inverting before being
+               returned in atomic_compare_and_exchange_bool).
+*/
+
+#define __arch_compare_and_exchange_n(mem, newval, oldval, bwl, version) \
+  ({ signed long __result; \
+     __asm __volatile ("\
 	.align 2\n\
 	mova 1f,r0\n\
 	nop\n\
 	mov r15,r1\n\
 	mov #-8,r15\n\
-     0: mov.b @%1,%0\n\
+     0: mov." #bwl " @%1,%0\n\
 	cmp/eq %0,%3\n\
 	bf 1f\n\
-	mov.b %2,@%1\n\
-     1: mov r1,r15"\
-	: "=&r" (__result) : "r" (mem), "r" (newval), "r" (oldval) \
-	: "r0", "r1", "t", "memory"); \
+	mov." #bwl " %2,@%1\n\
+     1: mov r1,r15\n\
+     .ifeqs \"bool\",\"" #version "\"\n\
+        movt %0\n\
+     .endif\n"					\
+	: "=&r" (__result)			\
+	: "r" (mem), "r" (newval), "r" (oldval)	\
+	: "r0", "r1", "t", "memory");		\
      __result; })
 
+#define __arch_compare_and_exchange_val_8_acq(mem, newval, oldval) \
+  __arch_compare_and_exchange_n(mem, newval, (int8_t)(oldval), b, val)
+
 #define __arch_compare_and_exchange_val_16_acq(mem, newval, oldval) \
-  ({ __typeof (*(mem)) __result; \
-     __asm__ __volatile__ ("\
-	.align 2\n\
-	mova 1f,r0\n\
-	nop\n\
-	mov r15,r1\n\
-	mov #-8,r15\n\
-     0: mov.w @%1,%0\n\
-	cmp/eq %0,%3\n\
-	bf 1f\n\
-	mov.w %2,@%1\n\
-     1: mov r1,r15"\
-	: "=&r" (__result) : "r" (mem), "r" (newval), "r" (oldval) \
-	: "r0", "r1", "t", "memory"); \
-     __result; })
+  __arch_compare_and_exchange_n(mem, newval, (int16_t)(oldval), w, val)
 
 #define __arch_compare_and_exchange_val_32_acq(mem, newval, oldval) \
-  ({ __typeof (*(mem)) __result; \
-     __asm__ __volatile__ ("\
-	.align 2\n\
-	mova 1f,r0\n\
-	nop\n\
-	mov r15,r1\n\
-	mov #-8,r15\n\
-     0: mov.l @%1,%0\n\
-	cmp/eq %0,%3\n\
-	bf 1f\n\
-	mov.l %2,@%1\n\
-     1: mov r1,r15"\
-	: "=&r" (__result) : "r" (mem), "r" (newval), "r" (oldval) \
-	: "r0", "r1", "t", "memory"); \
-     __result; })
+  __arch_compare_and_exchange_n(mem, newval, (int32_t)(oldval), l, val)
 
 /* XXX We do not really need 64-bit compare-and-exchange.  At least
    not in the moment.  Using it would mean causing portability
@@ -122,298 +114,180 @@ typedef uintmax_t uatomic_max_t;
    such an operation.  So don't define any code for now.  */
 
 # define __arch_compare_and_exchange_val_64_acq(mem, newval, oldval) \
-  (abort (), (__typeof (*mem)) 0)
+  (abort (), 0)
+
+/* For "bool" routines, return if the exchange did NOT occur */
+
+#define __arch_compare_and_exchange_bool_8_acq(mem, newval, oldval) \
+  (! __arch_compare_and_exchange_n(mem, newval, (int8_t)(oldval), b, bool))
+
+#define __arch_compare_and_exchange_bool_16_acq(mem, newval, oldval) \
+  (! __arch_compare_and_exchange_n(mem, newval, (int16_t)(oldval), w, bool))
+
+#define __arch_compare_and_exchange_bool_32_acq(mem, newval, oldval) \
+  (! __arch_compare_and_exchange_n(mem, newval, (int32_t)(oldval), l, bool))
+
+# define __arch_compare_and_exchange_bool_64_acq(mem, newval, oldval) \
+  (abort (), 0)
+
+/* Similar to the above, have one template which can be used in a
+   number of places. This version returns both the old and the new
+   values of the location. Parameters:
+      bwl:     b, w or l for 8, 16 and 32 bit versions.
+      oper:    The instruction to perform on the old value.
+   Note old is not sign extended, so should be an unsigned long.
+*/
+
+#define __arch_operate_old_new_n(mem, value, old, new, bwl, oper)	\
+  (void) ({ __asm __volatile ("\
+	.align 2\n\
+	mova 1f,r0\n\
+	mov r15,r1\n\
+	nop\n\
+	mov #-8,r15\n\
+     0: mov." #bwl " @%2,%0\n\
+	mov %0,%1\n\
+	" #oper " %3,%1\n\
+	mov." #bwl " %1,@%2\n\
+     1: mov r1,r15"			\
+	: "=&r" (old), "=&r"(new)	\
+	: "r" (mem), "r" (value)	\
+	: "r0", "r1", "memory");	\
+    })
+
+#define __arch_exchange_and_add_8_int(mem, value)			\
+  ({ int32_t __value = (value), __new, __old;				\
+    __arch_operate_old_new_n((mem), __value, __old, __new, b, add);	\
+    __old; })
+
+#define __arch_exchange_and_add_16_int(mem, value)			\
+  ({ int32_t __value = (value), __new, __old;				\
+    __arch_operate_old_new_n((mem), __value, __old, __new, w, add);	\
+    __old; })
+
+#define __arch_exchange_and_add_32_int(mem, value)			\
+  ({ int32_t __value = (value), __new, __old;				\
+    __arch_operate_old_new_n((mem), __value, __old, __new, l, add);	\
+    __old; })
+
+#define __arch_exchange_and_add_64_int(mem, value)			\
+  (abort (), 0)
 
 #define atomic_exchange_and_add(mem, value) \
-  ({ __typeof (*(mem)) __result, __tmp, __value = (value); \
-     if (sizeof (*(mem)) == 1) \
-       __asm__ __volatile__ ("\
-	  .align 2\n\
-	  mova 1f,r0\n\
-	  mov r15,r1\n\
-	  mov #-6,r15\n\
-       0: mov.b @%2,%0\n\
-	  add %0,%1\n\
-	  mov.b %1,@%2\n\
-       1: mov r1,r15"\
-	: "=&r" (__result), "=&r" (__tmp) : "r" (mem), "1" (__value) \
-	: "r0", "r1", "memory"); \
-     else if (sizeof (*(mem)) == 2) \
-       __asm__ __volatile__ ("\
-	  .align 2\n\
-	  mova 1f,r0\n\
-	  mov r15,r1\n\
-	  mov #-6,r15\n\
-       0: mov.w @%2,%0\n\
-	  add %0,%1\n\
-	  mov.w %1,@%2\n\
-       1: mov r1,r15"\
-	: "=&r" (__result), "=&r" (__tmp) : "r" (mem), "1" (__value) \
-	: "r0", "r1", "memory"); \
-     else if (sizeof (*(mem)) == 4) \
-       __asm__ __volatile__ ("\
-	  .align 2\n\
-	  mova 1f,r0\n\
-	  mov r15,r1\n\
-	  mov #-6,r15\n\
-       0: mov.l @%2,%0\n\
-	  add %0,%1\n\
-	  mov.l %1,@%2\n\
-       1: mov r1,r15"\
-	: "=&r" (__result), "=&r" (__tmp) : "r" (mem), "1" (__value) \
-	: "r0", "r1", "memory"); \
-     else \
-       { \
-	 __typeof (mem) memp = (mem); \
-	 do \
-	   __result = *memp; \
-	 while (__arch_compare_and_exchange_val_64_acq \
-		 (memp,	__result + __value, __result) == __result); \
-	 (void) __value; \
-       } \
-     __result; })
+  __atomic_val_bysize (__arch_exchange_and_add, int, mem, value)
+
+
+/* Again, another template. We get a slight optimisation when the old value
+   does not need to be returned. Parameters:
+      bwl:     b, w or l for 8, 16 and 32 bit versions.
+      oper:    The instruction to perform on the old value.
+*/
+
+#define __arch_operate_new_n(mem, value, bwl, oper)	 \
+  ({ int32_t __value = (value), __new; \
+     __asm __volatile ("\
+	.align 2\n\
+	mova 1f,r0\n\
+	mov r15,r1\n\
+	mov #-6,r15\n\
+     0: mov." #bwl " @%1,%0\n\
+	" #oper " %2,%0\n\
+	mov." #bwl " %0,@%1\n\
+     1: mov r1,r15"			\
+	: "=&r" (__new)			\
+	: "r" (mem), "r" (__value)	\
+	: "r0", "r1", "memory");	\
+     __new;				\
+  })
+
+#define __arch_add_8_int(mem, value)		\
+  __arch_operate_new_n(mem, value, b, add)
+
+#define __arch_add_16_int(mem, value)		\
+  __arch_operate_new_n(mem, value, w, add)
+
+#define __arch_add_32_int(mem, value)		\
+  __arch_operate_new_n(mem, value, l, add)
+
+#define __arch_add_64_int(mem, value)		\
+  (abort (), 0)
 
 #define atomic_add(mem, value) \
-  (void) ({ __typeof (*(mem)) __tmp, __value = (value); \
-	    if (sizeof (*(mem)) == 1) \
-	      __asm__ __volatile__ ("\
-		.align 2\n\
-		mova 1f,r0\n\
-		mov r15,r1\n\
-		mov #-6,r15\n\
-	     0: mov.b @%1,r2\n\
-		add r2,%0\n\
-		mov.b %0,@%1\n\
-	     1: mov r1,r15"\
-		: "=&r" (__tmp) : "r" (mem), "0" (__value) \
-		: "r0", "r1", "r2", "memory"); \
-	    else if (sizeof (*(mem)) == 2) \
-	      __asm__ __volatile__ ("\
-		.align 2\n\
-		mova 1f,r0\n\
-		mov r15,r1\n\
-		mov #-6,r15\n\
-	     0: mov.w @%1,r2\n\
-		add r2,%0\n\
-		mov.w %0,@%1\n\
-	     1: mov r1,r15"\
-		: "=&r" (__tmp) : "r" (mem), "0" (__value) \
-		: "r0", "r1", "r2", "memory"); \
-	    else if (sizeof (*(mem)) == 4) \
-	      __asm__ __volatile__ ("\
-		.align 2\n\
-		mova 1f,r0\n\
-		mov r15,r1\n\
-		mov #-6,r15\n\
-	     0: mov.l @%1,r2\n\
-		add r2,%0\n\
-		mov.l %0,@%1\n\
-	     1: mov r1,r15"\
-		: "=&r" (__tmp) : "r" (mem), "0" (__value) \
-		: "r0", "r1", "r2", "memory"); \
-	    else \
-	      { \
-		__typeof (*(mem)) oldval; \
-		__typeof (mem) memp = (mem); \
-		do \
-		  oldval = *memp; \
-		while (__arch_compare_and_exchange_val_64_acq \
-			(memp, oldval + __value, oldval) == oldval); \
-		(void) __value; \
-	      } \
-	    })
+  ((void) __atomic_val_bysize (__arch_add, int, mem, value))
+
+
+#define __arch_add_negative_8_int(mem, value)		\
+  (__arch_operate_new_n(mem, value, b, add) < 0)
+
+#define __arch_add_negative_16_int(mem, value)		\
+  (__arch_operate_new_n(mem, value, w, add) < 0)
+
+#define __arch_add_negative_32_int(mem, value)		\
+  (__arch_operate_new_n(mem, value, l, add) < 0)
+
+#define __arch_add_negative_64_int(mem, value)		\
+  (abort (), 0)
 
 #define atomic_add_negative(mem, value) \
-  ({ unsigned char __result; \
-     __typeof (*(mem)) __tmp, __value = (value); \
-     if (sizeof (*(mem)) == 1) \
-       __asm__ __volatile__ ("\
-	  .align 2\n\
-	  mova 1f,r0\n\
-	  mov r15,r1\n\
-	  mov #-6,r15\n\
-       0: mov.b @%2,r2\n\
-	  add r2,%1\n\
-	  mov.b %1,@%2\n\
-       1: mov r1,r15\n\
-	  shal %1\n\
-	  movt %0"\
-	: "=r" (__result), "=&r" (__tmp) : "r" (mem), "1" (__value) \
-	: "r0", "r1", "r2", "t", "memory"); \
-     else if (sizeof (*(mem)) == 2) \
-       __asm__ __volatile__ ("\
-	  .align 2\n\
-	  mova 1f,r0\n\
-	  mov r15,r1\n\
-	  mov #-6,r15\n\
-       0: mov.w @%2,r2\n\
-	  add r2,%1\n\
-	  mov.w %1,@%2\n\
-       1: mov r1,r15\n\
-	  shal %1\n\
-	  movt %0"\
-	: "=r" (__result), "=&r" (__tmp) : "r" (mem), "1" (__value) \
-	: "r0", "r1", "r2", "t", "memory"); \
-     else if (sizeof (*(mem)) == 4) \
-       __asm__ __volatile__ ("\
-	  .align 2\n\
-	  mova 1f,r0\n\
-	  mov r15,r1\n\
-	  mov #-6,r15\n\
-       0: mov.l @%2,r2\n\
-	  add r2,%1\n\
-	  mov.l %1,@%2\n\
-       1: mov r1,r15\n\
-	  shal %1\n\
-	  movt %0"\
-	: "=r" (__result), "=&r" (__tmp) : "r" (mem), "1" (__value) \
-	: "r0", "r1", "r2", "t", "memory"); \
-     else \
-       abort (); \
-     __result; })
+  __atomic_bool_bysize (__arch_add_negative, int, mem, value)
+
+
+#define __arch_add_zero_8_int(mem, value)		\
+  (__arch_operate_new_n(mem, value, b, add) == 0)
+
+#define __arch_add_zero_16_int(mem, value)		\
+  (__arch_operate_new_n(mem, value, w, add) == 0)
+
+#define __arch_add_zero_32_int(mem, value)		\
+  (__arch_operate_new_n(mem, value, l, add) == 0)
+
+#define __arch_add_zero_64_int(mem, value)		\
+  (abort (), 0)
 
 #define atomic_add_zero(mem, value) \
-  ({ unsigned char __result; \
-     __typeof (*(mem)) __tmp, __value = (value); \
-     if (sizeof (*(mem)) == 1) \
-       __asm__ __volatile__ ("\
-	  .align 2\n\
-	  mova 1f,r0\n\
-	  mov r15,r1\n\
-	  mov #-6,r15\n\
-       0: mov.b @%2,r2\n\
-	  add r2,%1\n\
-	  mov.b %1,@%2\n\
-       1: mov r1,r15\n\
-	  tst %1,%1\n\
-	  movt %0"\
-	: "=r" (__result), "=&r" (__tmp) : "r" (mem), "1" (__value) \
-	: "r0", "r1", "r2", "t", "memory"); \
-     else if (sizeof (*(mem)) == 2) \
-       __asm__ __volatile__ ("\
-	  .align 2\n\
-	  mova 1f,r0\n\
-	  mov r15,r1\n\
-	  mov #-6,r15\n\
-       0: mov.w @%2,r2\n\
-	  add r2,%1\n\
-	  mov.w %1,@%2\n\
-       1: mov r1,r15\n\
-	  tst %1,%1\n\
-	  movt %0"\
-	: "=r" (__result), "=&r" (__tmp) : "r" (mem), "1" (__value) \
-	: "r0", "r1", "r2", "t", "memory"); \
-     else if (sizeof (*(mem)) == 4) \
-       __asm__ __volatile__ ("\
-	  .align 2\n\
-	  mova 1f,r0\n\
-	  mov r15,r1\n\
-	  mov #-6,r15\n\
-       0: mov.l @%2,r2\n\
-	  add r2,%1\n\
-	  mov.l %1,@%2\n\
-       1: mov r1,r15\n\
-	  tst %1,%1\n\
-	  movt %0"\
-	: "=r" (__result), "=&r" (__tmp) : "r" (mem), "1" (__value) \
-	: "r0", "r1", "r2", "t", "memory"); \
-     else \
-       abort (); \
-     __result; })
+  __atomic_bool_bysize (__arch_add_zero, int, mem, value)
+
 
 #define atomic_increment_and_test(mem) atomic_add_zero((mem), 1)
 #define atomic_decrement_and_test(mem) atomic_add_zero((mem), -1)
 
-#define atomic_bit_set(mem, bit) \
-  (void) ({ unsigned int __mask = 1 << (bit); \
-	    if (sizeof (*(mem)) == 1) \
-	      __asm__ __volatile__ ("\
-		.align 2\n\
-		mova 1f,r0\n\
-		mov r15,r1\n\
-		mov #-6,r15\n\
-	     0: mov.b @%0,r2\n\
-		or %1,r2\n\
-		mov.b r2,@%0\n\
-	     1: mov r1,r15"\
-		: : "r" (mem), "r" (__mask) \
-		: "r0", "r1", "r2", "memory"); \
-	    else if (sizeof (*(mem)) == 2) \
-	      __asm__ __volatile__ ("\
-		.align 2\n\
-		mova 1f,r0\n\
-		mov r15,r1\n\
-		mov #-6,r15\n\
-	     0: mov.w @%0,r2\n\
-		or %1,r2\n\
-		mov.w r2,@%0\n\
-	     1: mov r1,r15"\
-		: : "r" (mem), "r" (__mask) \
-		: "r0", "r1", "r2", "memory"); \
-	    else if (sizeof (*(mem)) == 4) \
-	      __asm__ __volatile__ ("\
-		.align 2\n\
-		mova 1f,r0\n\
-		mov r15,r1\n\
-		mov #-6,r15\n\
-	     0: mov.l @%0,r2\n\
-		or %1,r2\n\
-		mov.l r2,@%0\n\
-	     1: mov r1,r15"\
-		: : "r" (mem), "r" (__mask) \
-		: "r0", "r1", "r2", "memory"); \
-	    else \
-	      abort (); \
-	    })
-
-#define atomic_bit_test_set(mem, bit) \
-  ({ unsigned int __mask = 1 << (bit); \
-     unsigned int __result = __mask; \
-     if (sizeof (*(mem)) == 1) \
-       __asm__ __volatile__ ("\
-	  .align 2\n\
-	  mova 1f,r0\n\
-	  nop\n\
-	  mov r15,r1\n\
-	  mov #-8,r15\n\
-       0: mov.b @%2,r2\n\
-	  or r2,%1\n\
-	  and r2,%0\n\
-	  mov.b %1,@%2\n\
-       1: mov r1,r15"\
-	: "=&r" (__result), "=&r" (__mask) \
-	: "r" (mem), "0" (__result), "1" (__mask) \
-	: "r0", "r1", "r2", "memory"); \
-     else if (sizeof (*(mem)) == 2) \
-       __asm__ __volatile__ ("\
-	  .align 2\n\
-	  mova 1f,r0\n\
-	  nop\n\
-	  mov r15,r1\n\
-	  mov #-8,r15\n\
-       0: mov.w @%2,r2\n\
-	  or r2,%1\n\
-	  and r2,%0\n\
-	  mov.w %1,@%2\n\
-       1: mov r1,r15"\
-	: "=&r" (__result), "=&r" (__mask) \
-	: "r" (mem), "0" (__result), "1" (__mask) \
-	: "r0", "r1", "r2", "memory"); \
-     else if (sizeof (*(mem)) == 4) \
-       __asm__ __volatile__ ("\
-	  .align 2\n\
-	  mova 1f,r0\n\
-	  nop\n\
-	  mov r15,r1\n\
-	  mov #-8,r15\n\
-       0: mov.l @%2,r2\n\
-	  or r2,%1\n\
-	  and r2,%0\n\
-	  mov.l %1,@%2\n\
-       1: mov r1,r15"\
-	: "=&r" (__result), "=&r" (__mask) \
-	: "r" (mem), "0" (__result), "1" (__mask) \
-	: "r0", "r1", "r2", "memory"); \
-     else \
-       abort (); \
-     __result; })
+
+#define __arch_bit_set_8_int(mem, value)		\
+  __arch_operate_new_n(mem, 1<<(value), b, or)
+
+#define __arch_bit_set_16_int(mem, value)		\
+  __arch_operate_new_n(mem, 1<<(value), w, or)
+
+#define __arch_bit_set_32_int(mem, value)		\
+  __arch_operate_new_n(mem, 1<<(value), l, or)
+  
+#define __arch_bit_set_64_int(mem, value)		\
+  (abort (), 0)
+
+#define __arch_add_64_int(mem, value)			\
+  (abort (), 0)
+
+#define atomic_bit_set(mem, value) \
+  ((void) __atomic_val_bysize (__arch_bit_set, int, mem, value))
+
+
+#define __arch_bit_test_set_8_int(mem, value)				\
+  ({ int32_t __value = 1<<(value), __new, __old;			\
+    __arch_operate_old_new_n((mem), __value, __old, __new, b, or);	\
+    __old & __value; })
+
+#define __arch_bit_test_set_16_int(mem, value)				\
+  ({ int32_t __value = 1<<(value), __new, __old;			\
+    __arch_operate_old_new_n((mem), __value, __old, __new, w, or);	\
+    __old & __value; })
+
+#define __arch_bit_test_set_32_int(mem, value)				\
+  ({ int32_t __value = 1<<(value), __new, __old;			\
+    __arch_operate_old_new_n((mem), __value, __old, __new, l, or);	\
+    __old & __value; })
+
+#define __arch_bit_test_set_64_int(mem, value)	\
+  (abort (), 0)
+
+#define atomic_bit_test_set(mem, value) \
+  __atomic_val_bysize (__arch_bit_test_set, int, mem, value)