Re: 19991027 ICE compiling x11perf

[Richard Henderson <rth at cygnus dot com>]

On Thu, Oct 28, 1999 at 08:31:08AM +0930, Alan Modra wrote:
> x11perf.c: In function `usage':
> x11perf.c:472: Internal compiler error in `verify_local_live_at_start', at
> flow.c:2533

This was a different problem.

We split 

(insn:QI 89 88 90 (set (reg:CCNO 17 flags)
        (compare:CCNO (mem/s:SI (plus:SI (mult:SI (reg/s:SI 0 eax)
                        (const_int 4 [0x4]))
                    (reg/s:SI 4 esi)) 11)
            (const_int 0 [0x0]))) 4 {cmpsi_0} (insn_list 83 (insn_list 86 (insn_
list 86 (insn_list 83 (nil)))))
    (expr_list:REG_DEAD (reg/s:SI 0 eax)
        (nil)))

into 

(insn 371 88 372 (set (reg:SI 6 ebp)
        (mem/s:SI (plus:SI (mult:SI (reg/s:SI 0 eax)
                    (const_int 4 [0x4]))
                (reg/s:SI 4 esi)) 11)) -1 (nil)
    (expr_list:REG_DEAD (reg/s:SI 0 eax)
        (nil)))

(insn 372 371 90 (set (reg:CCNO 17 flags)
        (compare:CCNO (reg:SI 6 ebp)
            (const_int 0 [0x0]))) -1 (nil)
    (nil))

Note the use of ebp, without having compiled with -fomit-frame-pointer.

This ordinarily would have been fine, since ebp was in fact dead,
however mark_used_regs short-circuits the stack pointer and the
frame pointer, and doesn't perform complete life analysis for them. 
Note the lack of a REG_DEAD note on insn 372. 

This should probably be changed.

However, I decided to go ahead and prevent find_free_register from
getting us into this situation, because as-is we'd be scrogging the
backtrace from eg abort.


r~


        * resource.c (find_free_register): Don't use the frame pointer
        if frame_pointer_needed.

Index: resource.c
===================================================================
RCS file: /cvs/gcc/egcs/gcc/resource.c,v
retrieving revision 1.18
diff -c -p -d -r1.18 resource.c
*** resource.c	1999/10/27 19:27:41	1.18
--- resource.c	1999/10/28 19:28:36
*************** find_free_register (current_insn, last_i
*** 1261,1266 ****
--- 1261,1270 ----
        /* And that we don't create an extra save/restore.  */
        if (! call_used_regs[regno] && ! regs_ever_live[regno])
  	continue;
+       /* And we don't clobber traceback for noreturn functions.  */
+       if ((regno == FRAME_POINTER_REGNUM || regno == HARD_FRAME_POINTER_REGNUM)
+ 	  && (! reload_completed || frame_pointer_needed))
+         continue;
  
        success = 1;
        for (j = HARD_REGNO_NREGS (regno, mode) - 1; j >= 0; j--)