problem with call clobbered regs

[Daniel Bayer]

On Thu, Feb 12, 2004 at 03:31:19PM -0800, Jim Wilson wrote:
> Daniel Bayer wrote:
> >I already did that. The save and restore instructions are generated by
> >the caller-save code as insn 187 and 188 (24.greg). The code that
> >calculates the address for the saving is also generated in this pass
> >(some time later as insns 230 and 231). But there is no code to
> >calculate the address for the restore.
> 
> This isn't a complete description of the problem.  What does the RTL 
> look like?

In save_call_clobbered_regs() this RTL is generated:
|  (set (mem:QI (plus:QI (reg/f:QI 8 fp)
|              (const_int -4 [0xfffffffc])) [0 S1 A16])
| 	         (reg:QI 2 p2))
| 
|  (set (reg:QI 2 p2)
|      (mem:QI (plus:QI (reg/f:QI 8 fp)
|                  (const_int -4 [0xfffffffc])) [0 S1 A16]))

> I would assume the address was calculated into a pseudo, in which case 
> using it across the call is fine.  It should be reg allocated to a 
> call-saved register.  Are there any call-saved registers?  If not, that 
> could be a problem.

By "call-saved register" you mean registers that are not clobbered?
There are 5 of them. But I disabled 2 and one is used as framepointer.
So there are 2 available.

> Is there are REG_EQUAL note the insn that calculates the address? 

There is a REG_EQUIV note in the code generated by
emit_reload_insns(save_insn):
| (insn 230 0 231 (nil) (set (reg:QI 3 p3)
|         (const_int -4 [0xfffffffc])) -1 (nil)
| 	    (nil))
| (insn 231 230 0 (nil) (set (reg:QI 3 p3)
|         (plus:QI (reg:QI 3 p3)
|             (reg/f:QI 8 fp))) 11 {addqi3} (nil)
|         (expr_list:REG_EQUIV (plus:QI (reg/f:QI 8 fp)
|             (const_int -4 [0xfffffffc]))
|             (nil)))

After emit_reload_insns() subst_reloads() is called with the save-insn
as argument. It is changed to
| (insn 187 231 79 5 (nil) (set (mem:QI (reg:QI 3 p3) [0 S1 A16])
|         (reg:QI 2 p2)) 4 {*movqi_register_speicher_insn} (nil)
|     (nil))

But in this very moment it also changes the restore insn to
| (insn 188 180 64 5 (nil) (set (reg:QI 2 p2)
|         (mem:QI (reg:QI 3 p3) [0 S1 A16])) 2 {*movqi_speicher_register_insn} (nil)
|     (nil))

I think this shouldn't happen because the comment to subst_reloads()
says that it only changes the argument and insns "just made for
reloading".

> If so, then reload could recalculate the address after the call.  If
> the address is in a hard reg instead of a pseudo, then why is it in
> a hard reg?  Who put it in a hard reg, and why was it put in a
> call-clobbered reg?
> 
> You will probably have to step through the caller-save code and
> answer these questions.  The caller-save code is not very
> complicated, so this shouldn't be very hard.

The caller-save code just puts the 2 insns I have listed above into the
reload_insn_chain. I think, that is right.


Daniel