[Bug driver/50740] CPUID leaf 7 for BMI/BMI2/AVX2 feature detection not qualified with max_level and doesn't use subleaf

[warnerme at ptd dot net]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=50740

Mark Warner  changed:

   What|Removed |Added

 CC||warnerme at ptd dot net

--- Comment #4 from Mark Warner  ---
In the file cpuid.h, there is a problem with the definition:

#define __cpuid(level, a, b, c, d)  \
  __asm__ ("cpuid\n\t"  \
   : "=a" (a), "=b" (b), "=c" (c), "=d" (d) \
   : "0" (level))

With the ecx undefined, the leaf 7 (Extended Features %eax == 7) can not be
accessed.  Leaf 7 (%eax) can only be accessed when sub-leaf (%ecx) is 0

I suggest instead:
#define __cpuid(level, a, b, c, d)  \
  __asm__ ("cpuid\n\t"  \
   : "=a" (a), "=b" (b), "=c" (c), "=d" (d) \
   : "0" (level), "2" (0))

[Bug target/83401] New: cpuid function for leaf 7 may not be accessed when sub-leaf (%ecx) is undefined

[warnerme at ptd dot net]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=83401

Bug ID: 83401
   Summary: cpuid function for leaf 7 may not be accessed when
sub-leaf (%ecx) is undefined
   Product: gcc
   Version: 5.4.0
Status: UNCONFIRMED
  Severity: normal
  Priority: P3
 Component: target
  Assignee: unassigned at gcc dot gnu.org
  Reporter: warnerme at ptd dot net
  Target Milestone: ---

In the file cpuid.h, there is a problem with the definition:

#define __cpuid(level, a, b, c, d)  \
  __asm__ ("cpuid\n\t"  \
   : "=a" (a), "=b" (b), "=c" (c), "=d" (d) \
   : "0" (level))

With the ecx undefined, the leaf 7 (Extended Features %eax == 7) can not be
accessed.  Leaf 7 (%eax) can only be accessed when sub-leaf (%ecx) is 0

I suggest instead:
#define __cpuid(level, a, b, c, d)  \
  __asm__ ("cpuid\n\t"  \
   : "=a" (a), "=b" (b), "=c" (c), "=d" (d) \
   : "0" (level), "2" (0))

[Bug c/59933] New: for loop goes wild with assert() enabled

[warnerme at ptd dot net]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=59933

Bug ID: 59933
   Summary: for loop goes wild with assert() enabled
   Product: gcc
   Version: 4.8.2
Status: UNCONFIRMED
  Severity: major
  Priority: P3
 Component: c
  Assignee: unassigned at gcc dot gnu.org
  Reporter: warnerme at ptd dot net

I have this odd case where a for loop goes wild.
But it only fails when assert() is enabled.

gcc  -Iinc -g -O2 -DDEBUG -fstack-protector-all -W -Wstrict-prototypes -Wall
-Wextra -Wcast-align -Wnested-externs -Wshadow -c -o build/NSQ_del_dec.o
c/NSQ_del_dec.c

If -DNDEBUG is used, the code works fine, although there is a slight difference
between CYGWIN32 (4.8.2-2) and CYGWIN64 (4.8.2-1).  This was not a problem with
gcc 4.7.

if (RDmin_Q10 < RDmax_Q10)
{
#if 1
/* THIS IS THE CODE THAT FAILS */
for (k = i; k < (int)(sizeof(NSQ_del_dec_struct) /
sizeof(opus_int32)); ++k)
{
psDelDec[RDmax_ind].sLPC_Q14[k] =
psDelDec[RDmin_ind].sLPC_Q14[k];
}
#else
/* THIS IS THE WORK-AROUND */
int n = (sizeof(NSQ_del_dec_struct) / sizeof(opus_int32)) - i;
opus_int32 *src = &psDelDec[RDmin_ind].sLPC_Q14[i];
opus_int32 *dst = &psDelDec[RDmax_ind].sLPC_Q14[i];
while (n-- > 0)
*dst++ = *src++;
#endif
psSampleState[RDmax_ind][0] = psSampleState[RDmin_ind][1];
}

I've tried lots of combinations of code to get the work around, and this even
fails when I insert printf-s, but the most common with this exact code is that
it does sizeof(NSQ_del_dec_struct) number of loops leaving out the /
sizeof(opus_int32).  I have had some test test where the loop didn't stop till
it destroyed enough stack to crash it.

Sorry, that the file is a bit big but trying to whittle it down usually made
the problem go away.

[Bug c/59933] for loop goes wild with assert() enabled

[warnerme at ptd dot net]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=59933

--- Comment #1 from Mark Warner  ---
Created attachment 31945
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=31945&action=edit
C source of subroutines which contain problem for loop

This is a file from OPUS.  As sent it can't be run, but the problem was bad
enough that I thought looking at the generated code might be enough.
All, except system, include files are inlined.

[Bug c/59933] for loop goes wild with assert() enabled

[warnerme at ptd dot net]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=59933

--- Comment #5 from Mark Warner  ---
sizeof(NSQ_del_dec_struct) / sizeof(opus_int32) is guaranteed to produced a
even number with a remainder of 0.
Note the __attribute__ ((__aligned__ (8))) to make it a multiple of 8 in size.

[Bug c/59933] for loop goes wild with assert() enabled

[warnerme at ptd dot net]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=59933

--- Comment #6 from Mark Warner  ---
If it is invalid, why does -Wall not trigger anything ?

[Bug c/59933] for loop goes wild with assert() enabled

[warnerme at ptd dot net]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=59933

--- Comment #11 from Mark Warner  ---
I'm confused .. what about..
for (k = i; k < (int)(sizeof(NSQ_del_dec_struct) / sizeof(opus_int32)); ++k)
... is illegal or invalid ?
Why does it only fail if -DDEBUG is defined ?
I mean, this code worked fine for months .. and now