Re: [PATCH] Segfault in rtx_varies_p

[Mark Mitchell <mark at codesourcery dot com>]

Paul Brook wrote:

> When complied with -O2 on arm-none-elf, the following testcase causes a 
> segfault.
>
> static const char digs[] = "0123456789ABCDEF";
> int __attribute__((pure)) bar();
>
> int foo (int i)
> {
>  int len;
>  if (i)
>    return 0;
>  len = bar();
>  return digs[len];
> }
>
> For the call to bar() we generate a libcall sequence. This includes a 
> REG_EQUAL note containing a EXPR_LIST of the parameters of the libcall (and 
> symbol for bar).
>
> (insn (set (reg 101) (...))
> 	(insn_list:REG_RETVAL ...
> 	(expr_list:REG_EQUAL (expr_list (use (mem:blk scratch))
> 					(expr_list (symbol_ref "bar")
> 					(nil)))))
> 	(nil))
> (insn (set (reg 102) (symbol_ref "digs)))
> (insn (set (reg 103) (mem (plus (reg 101) (reg 102))))
> 	(expr_list:REG_EQUAL (plus (reg 101) (symbol_ref "bar")))
> 	(nil))
>
>  
This last REG_EQUAL note doesn't look right to me.

It references "bar", rather than "digs".  Why is that?

> During the gcse pass we then propagate the equality for reg101 from the first 
> REG_EQUAL note into the second REG_EQUAL note (the code for for digs[len]). 
> In a later pass rtx_varies_p falls over on the NULL terminator of the 
> EXPR_LIST.
Jeff, Richard: do y'all have thoughts about this?

--
Mark Mitchell
CodeSourcery, LLC
(916) 791-8304
mark@codesourcery.com