IV37038: INEFFICIENT CODE FOR VEC_ANY_NAN

March 2013 PTF for XL C for AIX, V12.1
March 2013 PTF for XL C/C++ for AIX, V12.1

APAR status

• Closed as program error.

Error description

• The following test case shows how the compiler produces
  non-optimal code:
  
  ====Compile Line:
  xlC -q64 -O2 -qarch=pwr7 -c -qlist -qaltivec test.C
  
  ====Testcase:
  $ cat test.C
  #include <builtins.h>
  
  extern "C" bool vec_nan(vector double x)
  {
    return vec_any_nan(x);
  }
  $
  
  Generates:
       | 000000                           PDEF     vec_nan
      4|                                  PROC      x,vs34
      6| 000000 xvcmpeqd F002171E   1     VCDCEQ_R
  vs0,cr6=vs34,vs34
      6| 000004 mfcr     7C102026   1     LFCR1     gr0=cr6,6
      6| 000008 rlwinm   5400CFFE   1     RN4       gr0=gr0,25,0x1
      6| 00000C subfic   20000001   1     SUBC      gr0,ca=1,gr0
      6| 000010 cntlzw   7C030034   1     CNTLZ4    gr3=gr0
      6| 000014 addi     3803FFE0   1     AI        gr0=gr3,-32
      6| 000018 rlwinm   54030FFE   1     SRL4      gr3=gr0,31
      7| 00001C bclr     4E800020   1     BA        lr
  
  Optimal code:
       | 000000                           PDEF     vec_nan_opt
      9|                                  PROC      x,vs34
     11| 000030 xvcmpeqd F002171E   1     VCDCEQ_R
  vs0,cr6=vs34,vs34
     11| 000034 addi     38600001   1     LI        gr3=1
     11| 000038 bclr     4C980020   1     BF
  CL.3,cr6,0x1/lt,taken=50%(0,0)
     11| 00003C addi     38600000   1     LI        gr3=0
     12|                              CL.3:
     12| 000040 bclr     4E800020   1     BA        lr
  

Local fix

• The optimal code can be created using:
  extern "C" bool vec_nan_opt(vector double x)
  {
    return vec_any_ne(x,x);
  }
  

Problem summary

• USERS AFFECTED:
  Users that want peak performance using the vec_any_nan builtin
  may be affected by this issue.
  
  PROBLEM DESCRIPTION:
  There have been minor performance issues for the vec_any_nan
  builtin function that may not be noticable, but can potentially
  add up if numerous calls are made.
  

Problem conclusion

• The minor performance issues for vec_any_nan builtin have been
  improved.
  

Temporary fix

Comments

APAR Information

• APAR is sysrouted FROM one or more of the following:

• APAR is sysrouted TO one or more of the following:

  LI77366

Fix information

• Fixed component name

  XL C FOR AIX

• Fixed component ID

  5725C7100

Applicable component levels

• RC10 PSY U858497

     UP C