meter_signal crashes when nfaults > 1.

[anonymous]

meter_signal has a command line option -nfaults to specify the number of times the test should be repeated to gather statistical data. The default value is one; specifying a larger value results in program failure.

meter_signal -nfaults 1
The following environment will be established:

1 stack frames will be laid down.
0 dummy interrupt handlers will be established in each frame.
1 zerodivide faults will be signalled.


Following are the times in microseconds for each fault:


       419

Minimum value = 419   Mean = 419

meter_signal -nfaults 2
The following environment will be established:

1 stack frames will be laid down.
0 dummy interrupt handlers will be established in each frame.
2 zerodivide faults will be signalled.


Following are the times in microseconds for each fault:


       436
Error:  illegal_procedure condition by meter_signal$|1461
(>system_library_tools>bound_system_test_)
referencing stack_4|4010 (in process dir)

The illegal procedure condition is occuring in meter_signal at the DTB instruction shown below:

               q = divide (1, 0, 17, 0);

001455  aa  100 004 227 404   dv3d      (ic),(ic),(pr)
001456  aa   000322 01 0002   desc9ls   210,2,0             001777 = 053060000000
001457  aa   000323 01 0002   desc9ls   211,2,0             002000 = 053061000000
001460  aa  6 00156 01 0022   desc9ls   pr6|110,18,0
001461  aa  000 100 305 500   dtb       (pr),(pr)
001462  aa  6 00156 01 0022   desc9ls   pr6|110,18,0
001463  aa  6 00123 00 0004   desc9a    pr6|83,4            q

The program intentionally divides by zero to generate the signal to be measured. The program's condition handler collects the timing data and if more faults are to be generated, returns from the condition handler. The execution is resumed at the DTB instruction. DV3D, by specification does not change the value of the result operand for the divide-by-zero case, so pr6|110 contains random stack values, and the DTB instruction faults on an illegal digit value.

This sequence seems predictable, so it is unclear why this issue was never reported.

In discussions with Olin Sibert, he raised the point that a DV3D/DTB sequence was almost certainly not what the program authors intended, and suggested re-coding the divide as:

                q = divide (1b, 0b, 17, 0);                  /* divide by zero */

This causes the pl1 compiler to generate the desired sequence:

               q = divide (1b, 0b, 17, 0);

001455  aa   000001 2360 07   ldq       1,dl
001456  aa   000000 5060 07   div       0,dl
001457  aa  6 00123 7561 00   stq       pr6|83     

This, in turn, corrects the issue:

meter_signal -nfaults 2
The following environment will be established:

1 stack frames will be laid down.
0 dummy interrupt handlers will be established in each frame.
2 zerodivide faults will be signalled.


Following are the times in microseconds for each fault:


       424       403

Minimum value = 403   Mean = 413

Examining the history of the source code, we see:

/* Originally coded by Paul Karger August 17, 1971 */
/* Updated by Alan Bier - March l974. */
/* Fixed to eliminate ERROR 295, 05/01/81, W. Olin Sibert */

This implies that meter_signal was rebuilt in 1981, ten years after the original implementation. I hypothesize that a change in the compiler in that period caused the unwanted DV3D/DTB code generation, and the issue can be blamed on compiler regression.

The following change fixes the issue:

cpa meter_signal.pl1.orig meter_signal.pl1.new

A166                     q = divide (1, 0, 17, 0);                    /* divide
\c by zero */
Changed by B to:
B166                     q = divide (1b, 0b, 17, 0);                  /* divide
\c by zero */

Comparison finished: 1 difference, 2 lines.

[Eric Swenson]

​https://s3.amazonaws.com/eswenson-multics/public/mcrs/MCR10123.pdf

[Eric Swenson]

MCR replaced with this one:

​https://s3.amazonaws.com/eswenson-multics/public/mcrs/MCR10123_v1.1.pdf

[Eric Swenson]

Note: the fix mentioned in this ticket is incorrect. See the MCR for the correct fix.

[Eric Swenson]

Fixed with install MR12.8-1040.