LI77235: INEFFICIENT COMPARE WITH 32-BIT IMMEDIATES

February 2013 Update for XL C/C++ for Blue Gene/Q, V12.1

APAR status

• Closed as program error.

Error description

• ===== DESCRIPTION:
  xlC compile does not issue optimal code for compare with 32-bit
  immediates.
  
  extern "C" unsigned int eqFEFEEFEF(unsigned int *in)
  {
     unsigned int i = *in;
     if (i == 0xFEFEEFEF) i++;
     return i;
  }
  
  Output:
       | 000000                           PDEF     eqFEFEEFEF
     17|                                  PROC      in,gr3
     19| 000080 lwz      80830000   1     L4Z       gr4=(unsigned
  int)(gr3,0)
     20| 000084 ld       E8620018   1     L8
  gr3=.+CONSTANT_AREA (gr2,0)          <---- here
     20| 000088 ld       E8030000   1     L8
  gr0=+CONSTANT_AREA (gr3,0)           <---- here
     20| 00008C cmplw    7C040040   1     CL4       cr0=gr4,gr0
     20| 000090 bc       40820010   1     BF
  CL.23,cr0,0x4/eq,taken=50%(0,0)
     20| 000094 addi     38040001   1     AI        gr0=gr4,1
     20| 000098 rlwinm   5403003E   1     RN4
  gr3=gr0,0,0xFFFFFFFF
      0| 00009C bclr     4E800020   1     BA        lr
      0|                              CL.23:
     20| 0000A0 rlwinm   5483003E   1     RN4
  gr3=gr4,0,0xFFFFFFFF
     22| 0000A4 bclr     4E800020   1     BA        lr
  
  Optimal C input:
  extern "C" unsigned int eqFEFEEFEF_opt(unsigned int *in)
  {
     unsigned int i = *in;
     if ((i ^ 0xFEFE0000) == 0xEFEF) i++;
     return i;
  }
  
  Optimal output:
     26| 0000C0 lwz      80630000   1     L4Z       gr3=(unsigned
  int)(gr3,0)
     27| 0000C4 xoris    6C60FEFE   1     XIU       gr0=gr3,0xFEFE
     27| 0000C8 cmplwi   2800EFEF   1     CL4       cr0=gr0,61423
     27| 0000CC bc       40820010   1     BF
  CL.20,cr0,0x4/eq,taken=50%(0,0)
     27| 0000D0 addi     38830001   1     AI        gr4=gr3,1
     27| 0000D4 rlwinm   5483003E   1     RN4
  gr3=gr4,0,0xFFFFFFFF
      0| 0000D8 bclr     4E800020   1     BA        lr
      0|                              CL.20:
     27| 0000DC rlwinm   5463003E   1     RN4
  gr3=gr3,0,0xFFFFFFFF
     29| 0000E0 bclr     4E800020   1     BA        lr
  
  
  ===== TESTCASE:
  $ cat test.C
  extern "C" unsigned int eqFEFEEFEF(unsigned int *in)
  {
     unsigned int i = *in;
     if (i == 0xFEFEEFEF) i++;
     return i;
  }
  $
  
  ===== COMPILE COMMAND:
  xlC -q64 -O2 test.C -qlist -c
  

Local fix

• n/a
  

Problem summary

• USERS AFFECTED:
  Potentially any user of the compiler is affected by this issue.
  
  PROBLEM DESCRIPTION:
  When doing comparisons with large 32-bit constant values, we
  are generating 3 instructions; ie 2 to generate the constant,
  and 1 to do the comparison.
  When the comparison is a check for equality, we could
  accomplish it with 2 instructions.
  

Problem conclusion

• Added code in the compiler to generate the 2 instruction
  sequence when appropriate.
  

Temporary fix

Comments

APAR Information

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

  IV33245

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

Fix information

• Fixed component name

  XL C/C++ FOR BG

• Fixed component ID

  5799AG100

Applicable component levels

• RC10 PSY

     UP