IV62255: MISSED RLDIMI OPTIMIZATION FOR __BPERMD AND __POPCNT8

XL C/C++ for AIX Fix Pack 2 (December 2014 PTF) for 13.1
XL C/C++ for AIX Fix Pack 3 (April 2015 PTF) for 13.1
XL C/C++ for AIX Fix Pack 4 (July 2015 PTF) for 13.1
XL C for AIX Fix Pack 5 (November 2015 Update) for 13.1
XL C/C++ for AIX Fix Pack 5 (November 2015 PTF) for 13.1
XL C/C++ for AIX Fix Pack 6 (February 2016 PTF) for 13.1
XL C for AIX Fix Pack 6 (February 2016 Update) for 13.1
XL C for AIX Fix Pack 7 (May 2016 Update) for 13.1
XL C/C++ for AIX Fix Pack 7 (May 2016 PTF) for 13.1
XL C for AIX Fix Pack 8 (August 2016 Update) for 13.1
XL C/C++ for AIX Fix Pack 8 (August 2016 PTF) for 13.1
XL C for AIX Fix Pack 9 (June 2017 Update) for 13.1
XL C/C++ for AIX Fix Pack 9 (June 2017 PTF) for 13.1
XL C for AIX Fix Pack 10 (February 2018 Update) for 13.1
XL C/C++ for AIX Fix Pack 10 (February 2018 PTF) for 13.1
XL C for AIX Fix Pack 2 (December 2014 PTF) for 13.1
XL C for AIX Fix Pack 3 (April 2015 PTF) for 13.1
XL C for AIX Fix Pack 4 (July 2015 PTF) for 13.1
XL C/C++ for AIX Fix Pack 11 (November 2018 PTF) for 13.1
XL C for AIX Fix Pack 11 (November 2018) for 13.1

APAR status

• Closed as program error.

Error description

• When doing bitwise operations after the __bpermd and __popcnt8
  built-ins, the compiler is not recognizing a code pattern and is
  not replacing a rldicr/or with a rldimi instruction:
  
  
     === TEST CASE ===
  
  unsigned long long mm(unsigned long long *p)
  {
     unsigned long long a = __bpermd(0x0001, p[0]);
     unsigned long long b = __bpermd(0x0001, p[1]);
     return (a << 8) | b;
  }
  
  
  The assembly generated is using a rldicr and or instead of a
  combined rldimi:
  
    12| 00002C bpermd   7C8001F8   1     BPERMD   gr0=gr0,gr4
    13| 000030 bpermd   7C8319F8   1     BPERMD   gr3=gr3,gr4
    14| 000034 rldicr   780045E4   1     SLL8     gr0=gr0,8
    14| 000038 or       7C031B78   1     O        gr3=gr0,gr3
    15| 00003C bclr     4E800020   1     BA        lr
  
  Other built-ins have the rldimi generated, ex:
  
  unsigned long long mm3(unsigned long long *p)
  {
     unsigned long long a = __cntlz8(p[0]);
     unsigned long long b = __cntlz8(p[1]);
     return (a << 8) | b;
  }
  
    26| 000088 cntlzd 7C000074 1 CNTLZ8  gr0=gr0
    27| 00008C cntlzd 7C630074 1 CNTLZ8  gr3=gr3
    28| 000090 rldimi 7803402C 1 RI8     gr3=gr0,8,gr3,0xFFFFFF00
    29| 000094 bclr   4E800020 1 BA      lr
  

Local fix

Problem summary

• PROBLEM DESCRIPTION:
  When doing bitwise operations after the __bpermd and __popcnt8
  built-ins, the compiler is not generating efficient code.
  
  USERS AFFECTED:
  Users who use bitwise operations after the __bpermd and
  __popcnt8 built-ins.
  

Problem conclusion

• The fix will help the compiler generate efficient code when
  mixing bitwise operations after the __bpermd and __popcnt8
  built-ins.
  

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:

  LI78452

Fix information

• Fixed component name

  XL C FOR AIX

• Fixed component ID

  5725C7100

Applicable component levels