Re: Let fold_rtx's unsigned_fix be consistent with other cases

[Geoff Keating <geoffk at geoffk dot org>]

Alexandre Oliva <aoliva@redhat.com> writes:

> --=-=-=
> 
> The value of `(unsigned char)((unsigned char)3 - (float)5)' is usually
> expected to be `(unsigned char)0-(unsigned char)2', even though
> technically it invokes undefined behavior.  However, it is quite
> surprising that we get the expected result when the expression is
> folded upfront, in the tree level.  However, if the expression is
> convoluted enough that it makes it to cse, when cse applies fold_rtx()
> to it, it gets 0 instead.
> 
> Here's a patch that arranges for real.c to compute the expected value.
> Ok to install?  (Will bootstrap and test on i686-pc-linux-gnu, in
> addition to the already-tested mn10300-elf, if I get encouraging
> feedback)

I don't think this is very helpful.  It will ensure that on machines
where the hardware computes 0, that gcc's constant folding is
consistently inconsistent with the hardware.

A quick survey indicates that this program:

static int t(double f);
#include <stdio.h>
int main(void)
{
  printf ("%d\t%d\n", t(-9.), t(4294967287.));
  return 0;
}
static int t(double f)
{
  return (unsigned char) f;
}


prints (`soft' means with -msoft-float):

Target		CPU	Result
powerpc-linux	750	247	255
powerpc-aix	603e	0	247
powerpc-aix	power3	0	247
powerpc-linux	soft	247	247
powerpc-eabi	sim	247	255
mips-tx39-elf	sim	0	247
mips-tx39-elf	soft	0	247
sparc-solaris	usparc	247	255
i386-linux	p3	247	0	with gcc from RH 7
i386-linux	k6	247	247	with gcc from RH 6.2
hpux11.0	?	247	255

Note that these results very much depend on whether the chip has a fp
-> unsigned conversion instruction; in some cases, gcc will compile
this as '(unsigned char) (int) f' and in some as 
'(unsigned char)(unsigned)f'.  They may depend on the gcc version, too.

I believe the right thing to do is that when this case is detected in
the constant folder, it should produce a warning, just like in the case
of 0/0.

-- 
- Geoffrey Keating <geoffk@geoffk.org>