Re: fix strlensi in i386.md

[grahams <grahams at rcp dot co dot uk>]

Jan

Here's the contents of the message "BUG x86 define_expand "strlensi"
which I originally posted to gcc@gcc.gnu.org on 27/2/00 and later 
forwared on 17/2/00 to gcc-bugs@gcc.gnu.org.

=====================================================================
I have been investigating problems with the define_expand "strlensi"
pattern on the x86 and have come to the conclusion that it cannot
use FAIL (i.e. it must expand or abort).

There are two reasons for coming to this conclusion

1.) The expand_builtin_strlen () in builtins.c calls the MD 
    backend expander (i.e. gen_strlensi) using the following

        emit_insn (GEN_FCN (icode)(result, ...);

    If the MD gen_strlensi expander uses FAIL (can't expand) then 
    a NULL_RTX is returned and it make no sense to call emit_insn 
    with a NULL_RTX

2.) Assuming the gen_strlensi () expander is correct in using
    FAIL then the above code would need to be written

    expanded_rtx = GEN_FCN (icode)(result, ...);
    if (!expanded_rtx)
       return NULL_RTX;
    emit_insn (expanded_rtx);


    But this in itself can cause aborts in expand_expr ()
    if the rtl being expanded has a RTL_EXPR with a non-zero 
    RTL_EXPR_SEQUENCE (). 

    The abort occurs because the when expand_expr processes a RTL_EXPR
    with a non-zero RTL_EXPR_SEQUENCE() it finds it's value is const0_rtx.

    Note that when expand_expr processing a RTL_EXPR with a non zero 
    RTL_EXPR_SEQUENCE() has the side effect of setting the RTL_EXPR_SEQUENCE () 
    to const0_rtx.

    I found that RTL_EXPR node was being expanded twice the 1st time via a
    call to expand_expr() in expand_builtin_strlen() and subsequently if the
    MD expander bailed out using FAIL a call to expand_expr() in expand_call().

This is an extract from i386.md file which shows its use of FAIL.

(define_expand "strlensi"  [(set (match_operand:SI 0 "register_operand" "")
       (unspec:SI [(match_operand:BLK 1 "general_operand" "")
                   (match_operand:QI 2 "immediate_operand" "")
                   (match_operand:SI 3 "immediate_operand" "")] 0))]  "" 
 "
{

  rtx out, addr, eoschar, align, scratch1, scratch2, scratch3;

  /* The generic case of strlen expander is long.  Avoid it's
     expanding unless TARGET_INLINE_ALL_STRINGOPS.  */

  if (TARGET_UNROLL_STRLEN && eoschar == const0_rtx && optimize > 1
      && !optimize_size
      && (GET_CODE (align) != CONST_INT || INTVAL (align) < 4))    
    FAIL;

This code is using eoschar and align before they are initialized (
reported this in a separate mailing to gcc-bugs a couple of days ago)

I don't understand how the comment matches the actual test performed?

I think this test should be deleted so that the expander always
returns using DONE.

Comments anyone?
=============================================================================

Graham