qemu-devel.nongnu.org archive mirror
 help / color / mirror / Atom feed
* [Qemu-devel] [Bug 617528] [NEW] Incorrect translation of unary PPC/SPE instructions (efdneg etc.)
@ 2010-08-13 18:28 Mike Pall
  2010-08-13 18:28 ` [Qemu-devel] [Bug 617528] " Mike Pall
                   ` (5 more replies)
  0 siblings, 6 replies; 7+ messages in thread
From: Mike Pall @ 2010-08-13 18:28 UTC (permalink / raw)
  To: qemu-devel

Public bug reported:

The translation for the following PPC/SPE (e500) instructions is wrong
in QEMU git 6cbf4c8c:

    evfsabs, evfsnabs, evfsneg
    efdabs, efdnabs, efdneg
    efsabs, efsnabs, efsneg

As you can see from the provided patch, these ought to write their
result to the destination register (rD) and not modify the source
register (rA) in-place.

It's rather hard to generate a test-case for this with GCC, since it
likes to put the input and output of a unary operation into the same
registers (that's probably also the reason why this went unnoticed).
There is however a broken code path in the EGLIBC function for sin()
when compiled for e500v2. It returns nonsense results for e.g. -1.0.
Trivial test code follows:

    #include <stdio.h>
    #include <stdlib.h>
    #include <math.h>

    int main(int argc, char **argv)
    {
      double x = strtod(argv[1], NULL);
      printf("%.14g\n", sin(x));
      return 0;
    }

Result before the patch (WRONG):

$ qemu-ppc -cpu e500v2 sintest -1.0
-1

Result after the patch (OK):

$ qemu-ppc -cpu e500v2 sintest -1.0
-0.84147071838379

A self-contained test-case using inline assembler can be provided upon
request.

** Affects: qemu
     Importance: Undecided
         Status: New

-- 
Incorrect translation of unary PPC/SPE instructions (efdneg etc.)
https://bugs.launchpad.net/bugs/617528
You received this bug notification because you are a member of qemu-
devel-ml, which is subscribed to QEMU.

Status in QEMU: New

Bug description:
The translation for the following PPC/SPE (e500) instructions is wrong in QEMU git 6cbf4c8c:

    evfsabs, evfsnabs, evfsneg
    efdabs, efdnabs, efdneg
    efsabs, efsnabs, efsneg

As you can see from the provided patch, these ought to write their result to the destination register (rD) and not modify the source register (rA) in-place.

It's rather hard to generate a test-case for this with GCC, since it likes to put the input and output of a unary operation into the same registers (that's probably also the reason why this went unnoticed). There is however a broken code path in the EGLIBC function for sin() when compiled for e500v2. It returns nonsense results for e.g. -1.0. Trivial test code follows:

    #include <stdio.h>
    #include <stdlib.h>
    #include <math.h>

    int main(int argc, char **argv)
    {
      double x = strtod(argv[1], NULL);
      printf("%.14g\n", sin(x));
      return 0;
    }

Result before the patch (WRONG):

$ qemu-ppc -cpu e500v2 sintest -1.0
-1

Result after the patch (OK):

$ qemu-ppc -cpu e500v2 sintest -1.0
-0.84147071838379

A self-contained test-case using inline assembler can be provided upon request.

^ permalink raw reply	[flat|nested] 7+ messages in thread
end of thread, other threads:[~2011-02-20 17:21 UTC | newest]

Thread overview: 7+ messages (download: mbox.gz follow: Atom feed
-- links below jump to the message on this page --
2010-08-13 18:28 [Qemu-devel] [Bug 617528] [NEW] Incorrect translation of unary PPC/SPE instructions (efdneg etc.) Mike Pall
2010-08-13 18:28 ` [Qemu-devel] [Bug 617528] " Mike Pall
2010-08-13 18:52 ` Mike Pall
2010-08-19 14:04 ` Nathan Froyd
2010-08-19 14:30 ` Mike Pall
2011-01-01 14:29 ` Aurelien Jarno
2011-02-20 17:14 ` Aurelien Jarno

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).