assembler error in XFree86 snapshot

[Thomas Moestl]

On Mon, 2003/04/07 at 17:33:32 -0700, Kris Kennaway wrote:
> On Thu, Jan 16, 2003 at 09:17:28PM +0100, Thomas Moestl wrote:
> > This is a arguably a gcc bug. All (13-bit) immediate operands are
> > sign-extended, even those to instructions which operate on unsigned
> > values, so umul can handle a range of very small and a range of very
> > large operands. gcc correctly recognizes that it can use an immediate
> > here; however, it chooses to output it as an unsigned number and does
> > not sign-extended it from 32 to 64 bit.
> > 
> > All sign extensions for instructions are made to the full 64 bit
> > however (even if umul only happens to use 32 of those), so when the
> > assembler checks whether a value is representable as an immediate, it
> > will check that the 64-bit sign extension of the immediate creates
> > the desired value (in sparc64 mode), i.e. it doesn't ignore the upper
> > 32 bits even if a particular instruction does not use them.
> > 
> > One solution is to generate negative literals for immediates if we
> > mean them to be sign-extended (which gcc does already for some other
> > instructions). The attached patch implements this, I'm not sure it
> > uses the best possible way to do this though, and it also needs a bit
> > more testing.
> 
> *Ping*
> 
> Someone needs to take this up with the gcc developers so it can get fixed.

Sorry, I didn't have time to get this done for 5.2. The attached patch
should work around the bug however; Eric, could you please add it to
XFree86-4-libraries, until the problem is resolved in gcc or gas, so
that there can be a package for the release?

Thanks,
	- Thomas

-- 
Thomas Moestl <t.moestl at tu-bs.de>	http://www.tu-bs.de/~y0015675/
              <tmm at FreeBSD.org>		http://people.FreeBSD.org/~tmm/
PGP fingerprint: 1C97 A604 2BD0 E492 51D0  9C0F 1FE6 4F1D 419C 776C
-------------- next part --------------
--- lib/XRes/XRes.c.orig	Wed Oct 16 02:37:26 2002
+++ lib/XRes/XRes.c	Wed May  7 23:44:38 2003
@@ -218,7 +218,13 @@
     }
 
 #ifdef LONG64
+#ifdef __sparc64__
+    /* The first assignment is to work around a bug in gcc/gas on sparc64. */
+    *bytes = rep.bytes_overflow;
+    *bytes = (*bytes * 4294967295) + rep.bytes;
+#else
     *bytes = (rep.bytes_overflow * 4294967295) + rep.bytes;
+#endif
 #else
     *bytes = rep.bytes_overflow ? 0xffffffff : rep.bytes;
 #endif