[PATCH] Make dataflow analyzer mark defs for call_fixed_regs

[Daniel Berlin]

Richard Henderson <rth@redhat.com> writes:

> On Sat, Jul 14, 2001 at 01:13:39AM -0400, Daniel Berlin wrote:
>> -             if (call_used_regs[i]
>> +             if ((call_used_regs[i] 
>>                   && ! global_regs[i]
>>                   && ! fixed_regs[i]
>>                   && ! df_insn_regno_def_p (df, bb, insn, i))
>> +                 || call_fixed_regs[i])
> 
> I don't think checking call_fixed_regs is necessary.  Seems
> to me that the df_insn_regno_def_p is probably incorrect.
> The only call_used registers that can be live after a call
> are FUNCTION_VALUE_REGNO_P.
Okeydokey.
> 
> Also, I think excluding global_regs is incorrect here, since
> the call can assign these registers a new value.
Okey.

> 
> In fact, excluding fixed_regs is also generally wrong -- 
> we don't let the target provide this information.  There are
> only a couple that we know are safe.  The test should be
> 
> 	if ((fixed_regs[i]
> 	     && i != STACK_POINTER_REGNUM
> 	     && i != FRAME_POINTER_REGNUM
> 	     && i != HARD_FRAME_POINTER_REGNUM
> #ifndef PIC_OFFSET_TABLE_REG_CALL_CLOBBERED
> 	     && i != PIC_OFFSET_TABLE_REGNUM
> #endif
> 	     && i != ARG_POINTER_REGNUM)
>	    || (! fixed_regs[i]
> 		&& (call_used_regs[i] || global_regs[i])))
> 	  {
> 	    // clobber the register
>	  }
> 
> Curiously, I see flow.c has not got this quite right either.
> I guess global registers and odd fixed registers don't get
> tested as much as they ought...

Probably.
I spent hours starting at dumps before I realized something simplfy hadn't
been marked clobbered by the analyzer. I thought we were just
computing one of the  store vectors wrong still.


> 
> 
> r~

-- 
"I saw a bank that said "24 Hour Banking", but I don't have that
much time.
"-Steven Wright