LI78438: INEFFICIENT CODE FOR VEC_ANY_EQ FOR 64-BIT VECTOR

XL C/C++ for Linux Fix Pack 2 (February 2015 Update) for 13.1

APAR status

• Closed as program error.

Error description

• The compiler does not generate efficient code for the following
  test case:
  
  ======COMPILE COMMAND:
  xlC -q64 -O2 -qarch=pwr7 -qlist test.cpp -qaltivec
  
  ======TESTCASE:
  $ cat test.cpp
  extern "C" int vanyeq(vector unsigned long long a, vector
  unsigned long
  long b)
  {
            return vec_any_eq(a,b);
  }
  $
  
  
  =======ACTUAL OUTPUT:
  Listing output:
       | 000000                           PDEF     vanyeq
      7|                                  PROC      a,b,vs34,vs35
      9| 000040 vcmpequw 10221886   1     VCMPEQUW  vs33=vs34,vs35
      9| 000044 xxlxor   F00004D7   1     VXOR      vs32=vs32,vs32
      9| 000048 addi     38600001   1     LI        gr3=1
      9| 00004C xxsldwi  F0010916   1     VSLDWI
  vs0=vs33,vs33,1
      9| 000050 xxland   F0000C12   1     VAND      vs0=vs0,vs33
      9| 000054 xxspltw  F0200290   1     VSPLTW    vs1=vs0,0
      9| 000058 xxspltw  F0020290   1     VSPLTW    vs0=vs0,2
      9| 00005C xxpermdi F0210051   1     VMRGHD    vs33=vs1,vs0
      9| 000060 vcmpgtuw 10010686   1     VCMPGTUW_
  vs32,cr6=vs33,vs32
      9| 000064 bclr     4C9A0020   1     BF
  CL.4,cr6,0x4/eq,taken=50%(0,0)
      9| 000068 addi     38600000   1     LI        gr3=0
     10|                              CL.4:
     10| 00006C bclr     4E800020   1     BA        lr
  
  
  =======EXPECTED OUTPUT:
  Optimal output:
       | 000000                           PDEF     vanyeq_opt
     12|                                  PROC      a,b,vs34,vs35
     14| 000080 vcmpequw 10021886   1     VCMPEQUW  vs32=vs34,vs35
     16| 000084 vspltisw 103F038C   1     VSPLTISW  vs33=-1
     16| 000088 addi     38600001   1     LI        gr3=1
     15| 00008C vpkuwum  1000004E   1     VPKUWUM   vs32=vs32,vs32
     16| 000090 vcmpequw 10000C86   1     VCMPEQUW_
  vs32,cr6=vs32,vs33
     16| 000094 bclr     4C9A0020   1     BF
  CL.5,cr6,0x4/eq,taken=50%(0,0)
     16| 000098 addi     38600000   1     LI        gr3=0
     17|                              CL.5:
     17| 00009C bclr     4E800020   1     BA        lr
  

Local fix

• The provided test case can be coded as follows instead:
  
  extern "C" int vanyeq_opt(vector unsigned long long a, vector
  unsigned
  long long b)
  {
     vector unsigned int cmp = (vector unsigned
  int)vec_cmpeq((vector
  unsigned int)a, (vector unsigned int)b);
     cmp = (vector unsigned int)vec_pack(cmp,cmp);
     return vec_any_eq(cmp,vec_splats((unsigned int)(-1)));
  }
  

Problem summary

• USERS AFFECTED:
  Users using vec built-ins vec_any_eq or similar routines maybe
  affected by this issue.
  
  PROBLEM DESCRIPTION:
  Compiler generates inefficient codes for vec_any_eq built-ins.
  

Problem conclusion

• Compiler has been fixed to generate more efficient codes for
  vector built-ins like vec_any_eq.
  

Temporary fix

Comments

APAR Information

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

  IV62060

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

Fix information

• Fixed component name

  XL C/C++ FOR LI

• Fixed component ID

  5725C7300

Applicable component levels

• RD10 PSN IV62060

     UP06/09/13