fix profiling of EVALUATE

[lefessan]

Try to fix issue reported in #110 (comment) and meyfa/CobolCraft#214 (comment).

[GitMensch]

please add the test case from the original example to the testsuite, then that's good to upstream

[lefessan]

I added the test, and also fixed the same code for break, though I don't have a test case for that one...

[GitMensch]

Please include also the expected profiling output here.

[lefessan]

Is it still ok for SVN with the change to the break function ?

[GitMensch]

I'm not sure yet to get the complete picture,

What are the elements of the chain(s) with and what without -fperf?
How does this change if we actually have a GO TO at the last position of the imperative statements of a WHEN?

[I'll check fir the break part in the meantime]

[lefessan]

For example, in the test, the chain would be [ CALL TestRecurse ] in the non-profiling case, and [ C_CALL "prof_begin_call" ; CALL TestRecurse ; C_CALL "prof_end_call" ] in the profiling case. The C_CALL are not "CB_STATEMENT_P", so the previous test was failing and behaving the same as if it was a GOTO, because the test was not correctly written.

I think that, in the case of a GOTO, profiling does not add any information before and after (it adds information to the GOTO itself), so it should be the same as without profiling.

[GitMensch]

Updated testcase:

       IDENTIFICATION DIVISION.
       PROGRAM-ID. Main.
       
       DATA DIVISION.
       WORKING-STORAGE SECTION.
           01 STR PIC X(6) VALUE "..-..-".
           01 OFFSET BINARY-LONG UNSIGNED.
       
       PROCEDURE DIVISION.
           MOVE 1 TO OFFSET
           PERFORM UNTIL OFFSET > LENGTH OF STR
               CALL "TestRecurse" USING STR OFFSET
           END-PERFORM.
       
       END PROGRAM Main.
       
       
       IDENTIFICATION DIVISION.
       PROGRAM-ID. TestRecurse IS RECURSIVE.
       
       DATA DIVISION.
       LINKAGE SECTION.
           01 LK-STR.
              03 LK-CHR PIC X OCCURS 6
        >> IF BREAK IS DEFINED
                 INDEXED BY LK-IDX
        >> END-IF

                 .
           01 LK-OFFSET BINARY-LONG UNSIGNED.
       
       PROCEDURE DIVISION USING LK-STR LK-OFFSET.
           DISPLAY "LK-OFFSET: " LK-OFFSET
        >> IF BREAK IS DEFINED
           SET LK-IDX TO LK-OFFSET
           SEARCH LK-CHR
               AT END
                   IF LK-OFFSET > 20
                      GOBACK
                   END-IF
                   ADD 1 TO LK-OFFSET
                   CALL "TestRecurse" USING LK-STR LK-OFFSET
               WHEN LK-CHR (LK-IDX) = "."
                   ADD 1 TO LK-OFFSET
           END-SEARCH.
        >> ELSE
           EVALUATE LK-STR(LK-OFFSET:1)
               WHEN "."
                   ADD 1 TO LK-OFFSET
                   CALL "TestRecurse" USING LK-STR LK-OFFSET
               WHEN OTHER
                   ADD 1 TO LK-OFFSET
           END-EVALUATE.
        >> END-IF
       
       END PROGRAM TestRecurse.

compile with -DBREAK for the second testcase (which actually breaks at runtime because of memory issues with current codegen).

maybe rename the testcase as "profiling and codegen" or something similar, with keywords EVALUATE and SEARCH.

Please add both expected profiling results to the testcase as well.

[lefessan]

In your example, where is the break supposed to be inserted ? In the WHEN of the SEARCH ?

Note that on my computer, this program with -DBREAK fails to compile without profiling, but works ok with profiling...

[GitMensch]

The break is not inserted with the current svn version:

        /* Line: 42        : CALL               : prof.cob */
#line 42 "prof.cob"
        cob_nop ();
#line 448 "prof.c"
        cob_procedure_params[0] = COB_SET_DATA (f_36, b_36);
        cob_procedure_params[1] = COB_SET_DATA (f_39, b_39);
        cob_glob_ptr->cob_call_params = 2;
        cob_glob_ptr->cob_stmt_exception = 0;
        if (unlikely((cob_glob_ptr->cob_exception_code & 0x0b00) == 0x0b00)) cob_glob_ptr->cob_exception_code = 0;
                if (unlikely(call_TestRecurse.funcvoid == NULL || cob_glob_ptr->cob_physical_cancel == 1))
        {
          call_TestRecurse.funcvoid = cob_resolve_cobol ("TestRecurse", 0, 1);
        }
        b_21 = ((int (*)(void *, void *))call_TestRecurse.funcint) (b_36, b_39);
        cob_prof_exit_procedure (prof_info, 3);
      }

      /* Line: 43        : WHEN               : prof.cob */

without:

        /* Line: 42        : CALL               : prof.cob */
#line 42 "prof.cob"
        cob_nop ();
#line 425 "prof.c"
        cob_procedure_params[0] = COB_SET_DATA (f_36, b_36);
        cob_procedure_params[1] = COB_SET_DATA (f_39, b_39);
        cob_glob_ptr->cob_call_params = 2;
        cob_glob_ptr->cob_stmt_exception = 0;
        if (unlikely((cob_glob_ptr->cob_exception_code & 0x0b00) == 0x0b00)) cob_glob_ptr->cob_exception_code = 0;
                if (unlikely(call_TestRecurse.funcvoid == NULL || cob_glob_ptr->cob_physical_cancel == 1))
        {
          call_TestRecurse.funcvoid = cob_resolve_cobol ("TestRecurse", 0, 1);
        }
        b_21 = ((int (*)(void *, void *))call_TestRecurse.funcint) (b_36, b_39);
        break;
      }

      /* Line: 43        : WHEN               : prof.cob */

[GitMensch]

If the program fails to compile, then please recode it to just do some addition in the sub-program no recursive calls (at least that would be my guess for the error). This also helps to keep that test to be more focused.

rechecked: The compile error got in with GnuCOBOL 3.2 (3.1.2 compiles it clean, 3.3 dev doesn't) - that should get a new upstream bug report (and be fixed until 3.3 final, rc1 would be nice, but not necessary as the error was already released).

[lefessan]

I think I found the bug with the program: the INDEXED BY appearing in LINKAGE SECTION was not correctly handled by gnucobol, I had to modify parser.y to set its index_type to CB_INT_INDEX, as if it were defined in the LOCAL-STORAGE section.