LI77662: SUBOPTIMAL CODE FOR VECTOR LONG LONG SUBTRACT

XL C/C++ for Linux Fix Pack 8 (October 2015 Update) for 12.1
XL C/C++ for Linux Fix Pack 4 (October 2013 Update) for 12.1
XL C/C++ for Linux Fix Pack 5 (December 2013 Update) for 12.1
XL C/C++ for Linux Fix Pack 6 (February 2014 Update) for 12.1
XL C/C++ for Linux Fix Pack 7 (May 2014 Update) for 12.1

APAR status

• Closed as program error.

Error description

• The code generated by the compiler could be further improved
  for the below test case.
  
  $cat vecadd64.C
  extern "C" vector unsigned long long sub64(vector unsigned long
  long a,
  vector unsigned long long b)
  {
    return a - b;
  }
  
  Command line:
  xlC -q64 vecadd64.C -O2 -qarch=pwr7 -qaltivec -qlist
  
  Actual compiler output:
       | 000000                           PDEF     sub64
     33|                                  PROC      a,b,vs34,vs35
     35| 0013D0 xxlnor   F0031D17   1     VNOR      vs32=vs35,vs35
     35| 0013D4 ld       E8620010   1     L8
  gr3=.+CONSTANT_AREA
  (gr2,0)
     35| 0013D8 addi     38000080   1     LI        gr0=128
     35| 0013DC addi     38800088   1     LI        gr4=136
     35| 0013E0 lxvdsx   7C230299   1     VLDS
  vs33=+CONSTANT_AREA
  (gr3,gr0,0)
     35| 0013E4 lxvdsx   7C032298   1     VLDS
  vs0=+CONSTANT_AREA
  (gr3,gr4,0)
     35| 0013E8 vadduwm  10600880   1     VADDUWM   vs35=vs32,vs33
     35| 0013EC vaddcuw  10000980   1     VADDCUW   vs32=vs32,vs33
     35| 0013F0 xxsldwi  F0000117   1     VSLDWI
  vs32=vs32,vs32,1
     35| 0013F4 vadduwm  10001880   1     VADDUWM   vs32=vs32,vs35
     35| 0013F8 vadduwm  10201080   1     VADDUWM   vs33=vs32,vs34
     35| 0013FC vaddcuw  10001180   1     VADDCUW   vs32=vs32,vs34
     35| 001400 xxsldwi  F0200116   1     VSLDWI
  vs1=vs32,vs32,1
     35| 001404 xxland   F0000C11   1     VAND      vs32=vs0,vs1
     35| 001408 vadduwm  10400880   1     VADDUWM   vs34=vs32,vs33
     36| 00140C bclr     4E800020   1     BA        lr
       |               Tag Table
       | 001410        00000000 00092200 00000000 00000040
       |               Instruction count           16
       |               Straight-line exec time     16
  
  
  The compiler could consider generating the following to save 7
  instructions:
  
       | 000000                           PDEF     sub64_opt
     38|                                  PROC      a,b,vs34,vs35
     42| 001420 vcmpgtuw 10231286   1     VCMPGTUW  vs33=vs35,vs34
     46| 001424 addi     38000090   1     LI        gr0=144
     43| 001428 vsubuwm  10621C80   1     VSUBUWM   vs35=vs34,vs35
     46| 00142C ld       E8620010   1     L8
  gr3=.+CONSTANT_AREA
  (gr2,0)
     46| 001430 xxlxor   F00004D7   1     VXOR      vs32=vs32,vs32
     46| 001434 lxvd2x   7C430699   1     VLQD
  vs34=+CONSTANT_AREA
  (gr3,gr0,0)
     46| 001438 vperm    100100AB   1     VPERM
  vs32=vs33,vs32,vs34
     47| 00143C vadduwm  10401880   1     VADDUWM   vs34=vs32,vs35
     49| 001440 bclr     4E800020   1     BA        lr
       |               Tag Table
       | 001444        00000000 00092000 00000000 00000024
       |               Instruction count            9
       |               Straight-line exec time      9
  

Local fix

• The following code would generate optimal binary code:
  
  extern "C" vector unsigned long long sub64_opt(vector unsigned
  long
  long a, vector unsigned long long b)
  {
    vector unsigned int ai = (vector unsigned int)a;
    vector unsigned int bi = (vector unsigned int)b;
    vector unsigned int ov = (vector unsigned
  int)vec_cmpgt(bi,ai);
    vector unsigned int diff = ai - bi;
    vector unsigned int vn = { 0, 0, 0, 0 };
    vector unsigned char vp =
  {0x07,0x07,0x07,0x7,0x1F,0x1F,0x1F,0x1F,0xF,
  0xF,0xF,0xF,0x1F,0x1F,0x1F,0x1F};
    ov = vec_perm(ov,vn,vp);
    diff = diff + ov;
    return (vector unsigned long long)diff;
  }
  

Problem summary

• PROBLEM DESCRIPTION: Inefficient code generated for vector
  unsigned 64-bit subtract
  
  USERS AFFECTED: Users of V12/14 with -qarch=pwr7 and up with
  -qaltivec with code using vec_sub for vector unsigned long long
  types
  

Problem conclusion

• Code generation for vector subtraction is improved by reducing
  number of generated instructions. Apply provided service.
  

Temporary fix

Comments

APAR Information

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

  IV37235

• 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

• RC10 PSN IV37235

     UP06/09/13