register variables vs asm?

register variables vs asm?

[Richard Henderson]

EGCS is not respecting the register assigned to the variable in the
appended example on Alpha.  Unlike the previous nonsense with longjmp,
I should be able to expect this to work, yes?

I'll look at the problem soon, as this also completely screws the kernel.


r~



typedef struct _pthread_descr_struct * pthread_descr;

struct _pthread_descr_struct {
  int p_cancelstate;            
  int p_canceltype;             
  int p_canceled;               
};

static inline pthread_descr thread_self (void) __attribute__ ((const));
static inline pthread_descr thread_self (void)
{
  register pthread_descr __self __asm__("$0");
  __asm__ ("call_pal %1 # %0" : "=r"(__self) : "i"(158 ));
  return __self;
}


int pthread_setcancelstate(int state, int * oldstate)
{
  pthread_descr self = thread_self();
  if (state < 0 || state > 5)
    return 22 ;
  if (oldstate != ((void *)0) ) *oldstate = self->p_cancelstate;
  self->p_cancelstate = state;
  if (self->p_canceled &&
      self->p_cancelstate == 1 &&
      self->p_canceltype == 2)
    pthread_exit(((void *) -1) );
  return 0;
}

Re: register variables vs asm?

[Richard Henderson]

On Thu, Oct 09, 1997 at 12:08:12PM -0700, Richard Henderson wrote:
> EGCS is not respecting the register assigned to the variable in the
> appended example on Alpha.  Unlike the previous nonsense with longjmp,
> I should be able to expect this to work, yes?

The following patch fixes the problem. 

I wonder if there is something similar with input operands, since
I don't see anything to prevent them from being substituted away,
but I can't put together something that triggers it either.


r~



Thu Oct  9 22:38:06 1997  Richard Henderson  <rth@cygnus.com>

	* combine.c (can_combine_p): Don't combine with an asm whose
	output is a hard register.

Index: combine.c
===================================================================
RCS file: /cvs/cvsfiles/egcs/gcc/combine.c,v
retrieving revision 1.3
diff -u -p -d -r1.3 combine.c
--- combine.c	1997/09/22 17:41:11	1.3
+++ combine.c	1997/10/10 05:37:42
@@ -1003,6 +1003,13 @@ can_combine_p (insn, i3, pred, succ, pde
 	  && p != succ && volatile_refs_p (PATTERN (p)))
 	return 0;
 
+  /* If INSN is an asm and DEST is a hard register, reject, since it has
+     to be an explicit register variable and was chosen for a reason.  */
+
+  if (GET_CODE (src) == ASM_OPERANDS
+      && GET_CODE (dest) == REG && REGNO (dest) < FIRST_PSEUDO_REGISTER)
+    return 0;
+
   /* If there are any volatile insns between INSN and I3, reject, because
      they might affect machine state.  */