llvm dev - Jul 2009 - [LLVMdev] GCC DejaGNU regressions

The GCC DejaGNU testsuite has discovered some regressions.  Here's  
one; this was reduced from testsuite/gcc.apple/4656532.c:

typedef long long __m64 __attribute__ ((__vector_size__ (8),  
__may_alias__));
static __inline __m64 __attribute__((__always_inline__, __nodebug__))  
_mm_slli_si64 (__m64 __m, int __count) {
}
__m64 x, y;
void t1(int n) {
   y = _mm_slli_si64(x, n);
}

Compiled with LLVM-GCC (v76963) on Darwin/x86, this generates an ICE  
in the GCC/LLVM conversion layer.

Here's another test that want's to provoke lots of null pointer  
diagnostics; several such diagnostics are no longer happening, on  
lines 69, 71, 72, 75, 77, and 78:

-------------- next part --------------
A non-text attachment was scrubbed...
Name: nonnull-3.c
Type: application/octet-stream
Size: 5404 bytes
Desc: not available
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20090724/6b1e5042/attachment.obj>
-------------- next part --------------



Here's another test that aborts at runtime (it's gcc.dg/pr26719.c):

/* { dg-do compile } */
/* { dg-do run } */
/* { dg-options "-O2" } */

void abort (void);

int table[32][256];

int main(void)
{
   int i, j;

   for (i = 0; i < 32; i++)
     for (j = 0; j < 256; j++)
       table[i][j] = ((signed char)j) * i;

   if (table[9][132] != -1116)
     abort ();

   return 0;
}

I've only scratched the surface here; my low-priority testing rig  
thinks it has found 400+ regressions.  (It's wrong; it didn't locate  
libstdc++ this time, so most of the C++ tests failed.)

Do any of these "ring a bell" with anyone?

The morbidly curious can see the awfulness here.  Warning: there's a  
lot of noise here; the C++ results are mostly useless.

-------------- next part --------------
An embedded and charset-unspecified text was scrubbed...
Name: regressions.txt
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20090724/6b1e5042/attachment.txt>
-------------- next part --------------



Note that this test rig does not record "failures."   Instead, when a
test fails, it records that some previous revision of LLVM-GCC passed  
that given test (i.e. it records regressions).  Ergo, a line like this:

	74228.prev/gcc.sum:PASS: gcc.dg/i386-bitfield4.c scan-assembler-not  
movb

says "The r74228 checkout of LLVM-GCC passed the gcc.dg/i386- 
bitfield4.c scan-assembler-not movb test, but the most recent LLVM-GCC  
(r76963) failed."

stuart

On Jul 24, 2009, at 5:01 PM, Stuart Hastings wrote:
> The GCC DejaGNU testsuite has discovered some regressions.  Here's  
> one; this was reduced from testsuite/gcc.apple/4656532.c:
>
> typedef long long __m64 __attribute__ ((__vector_size__ (8),  
> __may_alias__));
> static __inline __m64 __attribute__((__always_inline__,  
> __nodebug__)) _mm_slli_si64 (__m64 __m, int __count) {
> }
> __m64 x, y;
> void t1(int n) {
>  y = _mm_slli_si64(x, n);
> }
>
> Compiled with LLVM-GCC (v76963) on Darwin/x86, this generates an ICE  
> in the GCC/LLVM conversion layer.
>
> Here's another test that want's to provoke lots of null pointer  
> diagnostics; several such diagnostics are no longer happening, on  
> lines 69, 71, 72, 75, 77, and 78:
>
> <nonnull-3.c>
>
>
> Here's another test that aborts at runtime (it's gcc.dg/pr26719.c):
>
> /* { dg-do compile } */
> /* { dg-do run } */
> /* { dg-options "-O2" } */
>
> void abort (void);
>
> int table[32][256];
>
> int main(void)
> {
>  int i, j;
>
>  for (i = 0; i < 32; i++)
>    for (j = 0; j < 256; j++)
>      table[i][j] = ((signed char)j) * i;
>
>  if (table[9][132] != -1116)
>    abort ();
>
>  return 0;
> }
>
> I've only scratched the surface here; my low-priority testing rig  
> thinks it has found 400+ regressions.  (It's wrong; it didn't
locate
> libstdc++ this time, so most of the C++ tests failed.)
>
> Do any of these "ring a bell" with anyone?
>
It doesn't sound familiar. Go ahead and file PRs for them. Thanks!

-bw
> The morbidly curious can see the awfulness here.  Warning: there's a  
> lot of noise here; the C++ results are mostly useless.
>
> <regressions.txt>
>
>
> Note that this test rig does not record "failures."   Instead,
when
> a test fails, it records that some previous revision of LLVM-GCC  
> passed that given test (i.e. it records regressions).  Ergo, a line  
> like this:
>
> 	74228.prev/gcc.sum:PASS: gcc.dg/i386-bitfield4.c scan-assembler-not  
> movb
>
> says "The r74228 checkout of LLVM-GCC passed the gcc.dg/i386- 
> bitfield4.c scan-assembler-not movb test, but the most recent LLVM- 
> GCC (r76963) failed."
>
> stuart_______________________________________________
> LLVM Developers mailing list
> LLVMdev at cs.uiuc.edu         http://llvm.cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev

On Jul 24, 2009, at 5:01 PM, Stuart Hastings wrote:
> Here's another test that aborts at runtime (it's gcc.dg/pr26719.c):
>
> /* { dg-do compile } */
> /* { dg-do run } */
> /* { dg-options "-O2" } */
>
> void abort (void);
>
> int table[32][256];
>
> int main(void)
> {
>  int i, j;
>
>  for (i = 0; i < 32; i++)
>    for (j = 0; j < 256; j++)
>      table[i][j] = ((signed char)j) * i;
>
>  if (table[9][132] != -1116)
>    abort ();
>
>  return 0;
> }
This one is fixed in r77078.

Dan