[Bug tree-optimization/55600] New: excessive size of vectorized code

[Bug tree-optimization/55600] New: excessive size of vectorized code

[neleai at seznam dot cz]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=55600

             Bug #: 55600
           Summary: excessive size of vectorized code
    Classification: Unclassified
           Product: gcc
           Version: 4.8.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: tree-optimization
        AssignedTo: unassigned@gcc.gnu.org
        ReportedBy: neleai@seznam.cz


Consider following code:

int sum(int *s){long i;
  int su=0;
  for(i=0;i<128;i+=2) su+=s[i]*s[i+1];
  return su;
}

When compiled by latest gcc with -O3 then generated assembly has 1099 bytes.

[Bug tree-optimization/55600] excessive size of vectorized code

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=55600

--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> 2012-12-05 10:33:06 UTC ---
GCC fully unrolls the vectorized looo.  ICC does not.

The loop rolls 16 times:

  <bb 3>:
  # vect_p.5_30 = PHI <vect_p.5_45(4), vect_p.8_31(2)>
  # vect_su.12_52 = PHI <vect_su.12_53(4), { 0, 0, 0, 0 }(2)>
  # ivtmp_61 = PHI <ivtmp_62(4), 0(2)>
  vect_var_.9_46 = MEM[(int *)vect_p.5_30];
  vect_p.5_47 = vect_p.5_30 + 16;
  vect_var_.10_48 = MEM[(int *)vect_p.5_47];
  vect_perm_even_49 = VEC_PERM_EXPR <vect_var_.9_46, vect_var_.10_48, { 0, 2,
4, 6 }>;
  vect_perm_odd_50 = VEC_PERM_EXPR <vect_var_.9_46, vect_var_.10_48, { 1, 3, 5,
7 }>;
  vect_var_.11_51 = vect_perm_even_49 * vect_perm_odd_50;
  vect_su.12_53 = vect_var_.11_51 + vect_su.12_52;
  vect_p.5_45 = vect_p.5_47 + 16;
  ivtmp_62 = ivtmp_61 + 1;
  if (ivtmp_62 < 16)
    goto <bb 4>;
  else
    goto <bb 5>;

  <bb 4>:
  goto <bb 3>;

but at -O3 we don't care too much about code size in this case.  So I'm not
sure you can call this a "bug".  Does it run slower?

[Bug tree-optimization/55600] excessive size of vectorized code

[neleai at seznam dot cz]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=55600

--- Comment #2 from Ondrej Bilka <neleai at seznam dot cz> 2012-12-26 22:03:59 UTC ---
Yes when 128 is replaced by smaller constant. Attached patch gives on my i5
following:
size 32
vector

real    0m0.224s
user    0m0.220s
sys    0m0.000s
unroll

real    0m0.155s
user    0m0.148s
sys    0m0.004s
size 64
vector

real    0m0.398s
user    0m0.396s
sys    0m0.000s
unroll

real    0m0.380s
user    0m0.376s
sys    0m0.000s
size 128
vector

real    0m0.703s
user    0m0.700s
sys    0m0.000s
unroll

real    0m0.817s
user    0m0.812s
sys    0m0.000s

[Bug tree-optimization/55600] excessive size of vectorized code

[neleai at seznam dot cz]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=55600

--- Comment #3 from Ondrej Bilka <neleai at seznam dot cz> 2012-12-26 22:05:37 UTC ---
Created attachment 29052
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=29052
benchmark

[Bug target/55600] excessive size of vectorized code

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=55600

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Target|                            |x86_64-linux-gnu
           Keywords|                            |missed-optimization
          Component|tree-optimization           |target

--- Comment #4 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
So here is what the current state for this.
GCC vectorizers it and unroll it fully at 128 and 64
clang keeps it as **scalars** but unrolls the loop 4 times at 128 and fully at
64
ICC vectorizers it and unrolls it half way (that is 32 times) at 128 and fully
at 64
MSVC keeps it as **scalars** but unrolls it half way (that is 32 times) at 128

So it looks all compilers do stuff hugely different here really.