your mail

[Jan Hubicka]

> 
> Boy, this was loads of fun to track down...
> 
> This code has a thinko.
> 
> 
> revision 1.173
> date: 2001/01/07 13:37:32;  author: hubicka;  state: Exp;  lines: +27 -9
> 
>         * combine.c (combine_simplify_rtx): Recognize the unordered compares.
>         (nonzero_bits): Likewise.
>         (simplify_comparison): Likewise.
>         (num_sign_bit_copies): Likewise; return more sane value depending
>         on STORE_FLAG_VALUE.
>         (known_cond): Do not assume EQ to be always true for equivalent
>         operands.
> 
> 
> It introduced the !FLOAT_MODE_P part of the conditional below in 
> combine.c::known_cond.
> 
>   if (cond == EQ && rtx_equal_p (x, reg) && !FLOAT_MODE_P (cond))
>     return val;
> 
> Applying FLOAT_MODE_P to an RTX_CODE doesn't produce a particularly useful
> or predictable value :(  Depending on precisely where things land in memory
> the condition above can give different results for the same values of COND,
> X & REG.  Not good.
> 
> Presumably you wanted to check that neither operand of COND is a floating
> point type.  Right?
Yes, thats precisely what I was shooting for.  It is amazing that such a bug
survived so long time in the compiler.

Should I prepare patch to fix the condtiional?

Thanks for letting me know!
Honza
> 
> The net result was that I was getting different code for the following
> function between a stage1 and stage2 compiler on the PA...
> 
> int optind = 0;
> static enum
> {
>   REQUIRE_ORDER, PERMUTE, RETURN_IN_ORDER
> } ordering;
> static int first_nonopt;
> static int last_nonopt;
> int
> _getopt_internal (argc, argv, optstring, longopts, longind, long_only)
>      int argc;
>      char *const *argv;
>      const char *optstring;
>      const struct option *longopts;
>      int *longind;
>      int long_only;
> {
>           if (first_nonopt != last_nonopt && last_nonopt != optind)
>             exchange ((char **) argv);
>           else if (last_nonopt != optind)
>             first_nonopt = optind;
> }
> 
> [ Note this code is from getopt.c which is not bootstrapped when we
>   do a make bootstrap which is why it went undiscovered for so long. ]
> 
> jeff