[Bug c/48510] New: Does not vectorize loops involving casts from floating point to integer types

[Bug c/48510] New: Does not vectorize loops involving casts from floating point to integer types

[jeremysalwen at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=48510

           Summary: Does not vectorize loops involving casts from floating
                    point to integer types
           Product: gcc
           Version: 4.5.2
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: c
        AssignedTo: unassigned@gcc.gnu.org
        ReportedBy: jeremysalwen@gmail.com


The following code vectorizes with the command line options:

-march=native -mtune=native -ftree-vectorizer-verbose=12 -O3 -std=c99
-ffast-math -funsafe-math-optimizations -lm  main.c

#include <stdio.h>
#include <math.h>
int main() {
  double g[1000];
  for(int i=0; i<1000; i++) {
    g[i]=2*(g[i]);
  }
  for(int i=0; i<1000; i++) {
   printf("%f\n",g[i]);
  }
}

but the following code does not with the same options:


#include <stdio.h>
#include <math.h>
int main() {
  double g[1000];
  for(int i=0; i<1000; i++) {
    g[i]=2*((unsigned long)g[i]);
  }
  for(int i=0; i<1000; i++) {
   printf("%f\n",g[i]);
  }
}

If I understand correctly, there are SSE instructions for casting doubles to
long integers on the platform I'm on (Intel Atom) which GCC could use.  (or
perhaps there could be a benefit to vectorizing other parts of the loop, even
if the cast does not utilize SIMD instructions.)

[Bug target/48510] Does not vectorize loops involving casts from floating point to unsigned integer types

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=48510

Richard Guenther <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Target|                            |x86_64-*-*, i?86-*-*
             Status|UNCONFIRMED                 |NEW
           Keywords|                            |missed-optimization
   Last reconfirmed|                            |2011.04.08 10:02:15
          Component|c                           |target
     Ever Confirmed|0                           |1
            Summary|Does not vectorize loops    |Does not vectorize loops
                   |involving casts from        |involving casts from
                   |floating point to integer   |floating point to unsigned
                   |types                       |integer types
           Severity|normal                      |enhancement

--- Comment #1 from Richard Guenther <rguenth at gcc dot gnu.org> 2011-04-08 10:02:15 UTC ---
There are no conversions to unsigned.  For scalar code we work around this
by biasing the input/output, but there is no vectorized version for this
available (yet).

[Bug c/48510] Does not vectorize loops involving casts from floating point to unsigned integer types

[jeremysalwen at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=48510

jeremysalwen at gmail dot com changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
          Component|target                      |c
           Severity|enhancement                 |normal

--- Comment #2 from jeremysalwen at gmail dot com 2011-04-08 17:43:25 UTC ---
As you mentioned, unsigned conversions to int and long int both fail to
vectorize.  However, signed conversions to long int still fail, while they
succeed for signed conversions to int.

I should note that the computer I ran these last tests on is a Phenom II.

Apologies if I'm mistaken, but I believe the intrinsic

__builtin_ia32_cvttsd2si64

should do the conversion to a signed long on my platform.

[Bug target/48510] Does not vectorize loops involving casts from floating point to unsigned integer types

[pinskia at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=48510

Andrew Pinski <pinskia at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
          Component|c                           |target
           Severity|normal                      |enhancement

[Bug target/48510] Does not vectorize loops involving casts from floating point to unsigned integer types

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=48510

Richard Guenther <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Blocks|                            |53947

--- Comment #3 from Richard Guenther <rguenth at gcc dot gnu.org> 2012-07-13 08:42:23 UTC ---
Link to vectorizer missed-optimization meta-bug.

[Bug target/48510] Does not vectorize loops involving casts from floating point to unsigned integer types

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=48510

--- Comment #4 from Richard Guenther <rguenth at gcc dot gnu.org> 2012-07-19 10:57:07 UTC ---
We can now vectorize the conversion to unsigned int, it is not possible to
directly convert from double to unsigned long as far as I can see - only
cvttpd2dq exists which is a signed conversion.  Biasing the input value
from [0, ULONG_MAX] to [-LONG_MIN, LONG_MAX] is not possible because of
the different truncation behavior for signed/unsigned values.

As a side-note, C specifies that float -> unsigned integer truncation
only has defined behavior for inputs in the range (-1, Utype_MAX+1).

ICC seems to use some clever range-dependent operations, finally mixing
three cases together with bitwise operations.