Totally bogus string optimizations

[Bernd Schmidt]

On Sat, 28 Oct 2000, Joseph S. Myers wrote:

> If this gets improved in GCC, then the strcpy macro should best be
> disabled for GCC 2.97 and later: the code generated by the macro in GCC
> CVS is terrible:
> 
> .LC0:
>         .string "Foobar"
>         .text
>         .align 16
> .globl foo
>         .type   foo,@function
> foo:
>         movl    $.LC0+1, %eax
>         subl    $12, %esp
>         cmpl    %eax, %eax
>         je      .L158

Note that there's definitely a bug in gcc's constant folding here.  This is the
part of the macro that tries to determine whether the string is made of 1 unit
wide characters by doing
  (((char *)&"Foobar"[1]) - (char *)&"Foobar"[0]) == 1
and gcc ought to optimize this away.

> while GCC 2.95.2 generates with the macro the code we want, apart from
> leaving the string constant in the code (so there *is* a regression in
> GCC):

Yes.  All this, however, is with the unaligned version of the strcpy macro.
The aligned version generates completely bogus code with all gcc versions I've
tried (2.7.2.3, 2.95.2, current).


Bernd