[PATCH] global register problem

[PATCH] global register problem

[Richard Zidlicky]

Hi,

I have a problem with this testcase:

  void bar(void);
  extern int x1;
  register int g1 asm("%d7");

  void foo(void)
   {
     g1=x1;
     bar();
   }

gcc 3.1.1 will ICE as

# m68k-linux-gcc -O2 -da -S -fverbose-asm x4.i
x4.i: In function `foo':
x4.i:12: Internal compiler error in verify_wide_reg, at flow.c:557

x4.i.15.life:

;; Function foo

30 registers.

Register 29 used 2 times across 2 insns in block 0; set 1 time; pointer.

1 basic blocks, 2 edges.

Basic block 0: first insn 17, last 13, loop_depth 0, count 0, freq 10000.
Predecessors:  ENTRY [100.0%]  (fallthru)
Successors:  EXIT [100.0%]  (fallthru)
Registers live at start: 14 [%a6] 15 [%sp]
Registers live at end: 7 [%d7] 14 [%a6] 15 [%sp]

(note 2 0 3 NOTE_INSN_DELETED)

(note 3 2 17 NOTE_INSN_FUNCTION_BEG)

;; Start of basic block 0, registers live: 14 [%a6] 15 [%sp]
(note 17 3 9 [bb 0] NOTE_INSN_BASIC_BLOCK)

(insn 9 17 12 (set (reg/v:SI 7 %d7)
        (mem/f:SI (symbol_ref:SI ("x1")) [2 x1+0 S4 A16])) 29 {*m68k.md:976} (n
il)
    (nil))   

(insn 12 9 13 (set (reg/f:SI 29)
        (symbol_ref:SI ("bar"))) 29 {*m68k.md:976} (nil)
    (expr_list:REG_EQUAL (symbol_ref:SI ("bar"))
        (nil)))

(call_insn 13 12 14 (call (mem:QI (reg/f:SI 29) [0 S1 A8])
        (const_int 0 [0x0])) 370 {*m68k.md:7234} (insn_list 9 (insn_list 12 (ni
l)))
    (expr_list:REG_DEAD (reg:SI 7 %d7)
        (expr_list:REG_DEAD (reg/f:SI 29)
            (nil)))
    (nil))
;; End of basic block 0, registers live:
 7 [%d7] 14 [%a6] 15 [%sp]

(note 14 13 0 NOTE_INSN_FUNCTION_END)



Because %d7 is in regs_invalidated_by_call %d7 it is mark_set_1(..)
for the call instruction. This removes it from pbi->reg_live
and the subsequent attempt to make it live again by calling 
mark_used_reg(..) in flow.c:1773 results in adding an REG_DEAD note 
for %d7 to the call instruction instead of making it live again. 
Later the store to g1 gets eliminated somewhere.

This is the backtrace at the point where the REG_DEAD note for %d7
is added:

#0  mark_used_reg (pbi=0x82d4d30, reg=<incomplete type>, 
    cond=<incomplete type>, insn=<incomplete type>)
    at ../../gcc-3.1-new/gcc/flow.c:3569
#1  0x80f58a6 in propagate_one_insn (pbi=0x82d4d30, insn=<incomplete type>)
    at ../../gcc-3.1-new/gcc/flow.c:1773
#2  0x80f5c47 in propagate_block (bb=<error type>, live=<error type>, 
    local_set=<error type>, cond_local_set=<error type>, flags=103)
    at ../../gcc-3.1-new/gcc/flow.c:2026
#3  0x80f4033 in update_life_info (blocks=<error type>, extent=<error type>, 
    prop_flags=127) at ../../gcc-3.1-new/gcc/flow.c:726
#4  0x80f3ac7 in life_analysis (f=<incomplete type>, file=0x82d23d8, flags=127)
    at ../../gcc-3.1-new/gcc/flow.c:477
#5  0x818dbed in rest_of_compilation (decl=<incomplete type>)
    at ../../gcc-3.1-new/gcc/toplev.c:3028

  

This patch seems to cure all regression in my program that heavilly
relies on global register variables, any opinions?


Richard

*** flow.c.rz	Mon Jul 29 12:01:06 2002
--- flow.c	Thu Aug  1 17:24:06 2002
***************
*** 1770,1777 ****
  	     so they are made live.  */
  	  for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
  	    if (global_regs[i])
! 	      mark_used_reg (pbi, gen_rtx_REG (reg_raw_mode[i], i),
! 			     cond, insn);
  	}
      }
  
--- 1770,1780 ----
  	     so they are made live.  */
  	  for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
  	    if (global_regs[i])
! 	      {
! 		SET_REGNO_REG_SET (pbi->reg_live, i);
! 		mark_used_reg (pbi, gen_rtx_REG (reg_raw_mode[i], i),
! 			       cond, insn);
! 	      }
  	}
      }
  

Re: [PATCH] global register problem

[Richard Henderson]

On Thu, Aug 01, 2002 at 06:01:23PM +0200, Richard Zidlicky wrote:
> Because %d7 is in regs_invalidated_by_call %d7 it is mark_set_1(..)
> for the call instruction....

Should be fixed by

2002-06-16  Richard Henderson  <rth@redhat.com>

        PR opt/6722
        * regclass.c (globalize_reg): Update regs_invalidated_by_call.



r~

Re: [PATCH] global register problem

[Richard Zidlicky]

On Thu, Aug 01, 2002 at 07:09:43PM -0700, Richard Henderson wrote:
> On Thu, Aug 01, 2002 at 06:01:23PM +0200, Richard Zidlicky wrote:
> > Because %d7 is in regs_invalidated_by_call %d7 it is mark_set_1(..)
> > for the call instruction....
> 
> Should be fixed by
> 
> 2002-06-16  Richard Henderson  <rth@redhat.com>
> 
>         PR opt/6722
>         * regclass.c (globalize_reg): Update regs_invalidated_by_call.

it is in the Changelog of my CVS snapshot and certainly doesn't fix the
problem - perhaps even aggravates it. For example 3.1 release was almost 
correct wrt global asm register handling, 3.1.1 is hopelessly broken 
for my purposes.

What happens in my case, the loop 1720-1727 in propagate_one_insn
removes the register from the pbi->reg_live set. The loop in lines
1771 tries to add it again, instead it 
 - fails to add it to reg_live set
 - adds a REG_DEAD note for %d7

Perhaps mark_used_reg needs to take a flags arg like mark_set_1?
Beeing a bloody beginner to gcc hacking I would find it more intuitive
to have mark_reg_used_{r,w,rw} btw.

Richard