Mark Smotherman
Last updated: July 2017
The Model 91 performs all the interruption functions defined for the IBM System/360. ... The supervisor call, external, machine check, and I/O interruptions are logically handled as defined. However, the performance objectives of the Model 91 require some deviations in the handling of program exceptions. The program-exception deviations are basically those resulting from an operation that has been sent by the instruction processor to another element for execution, and the current PSW no longer references the operation. Consequently, the interruption-causing instruction cannot be directly identified, and such a program interruption is described as being imprecise.[Almost two pages of details follow, including the description of a no-operation instruction (a particular encoding of the BCR instruction) that acts as a barrier instruction and causes all previously-decoded instructions to complete before it can execute. There is a note that stores can execute out of order and that channel programs may require the use of the nop-barrier instruction.]
Interrupts, as architecturally constrained, are a major bottleneck to performance in the assembly line organization. Strict adherence to a specification which states that an interrupt on instruction n should logically precede and inhibit any action from being taken on instruction n + 1 leaves two alternatives. The first would be to force sequentialism between instructions which may lead to an interrupt. In view of the variety of interrupt possibilities defined, this course would totally thwart high performance and is necessarily discarded. The second is to set aside sufficient information to permit recovery from any interrupt which might arise. In view of the pipeline and execution concurrency which allows the Model 91 to advance many instructions beyond n prior to its execution, and to execute independent instructions out of sequence (n + m before n), the recovery problem becomes extremely complex and costly. Taking this approach would entail hardware additions to the extent that it would severely degrade the performance one is seeking to enhance. The impracticality of both alternatives by which the interrupt specifications could be met made it mandatory that the specifications themselves be altered. The architecture was compromised by removing the above-mentioned "precedence" and "inhibit" requirements. The specification change led to what is termed the "imprecise interrupt" philosophy of the Model 91 and reduced the interrupt bottleneck to an instruction supply discontinuity.
Both the 6600 and the Model 91 (and, in certain circumstances, the Model 75) suffer from the "imprecise interrupt" problem. This means that when an interrupt -- or its equivalent, in the case of the 6600 -- is caused by execution of an instruction in the program, the location of the particular instruction causing the interrupt is not indicated precisely. If a jump instruction intervened, then the location of the offending instruction is not determined at all. This impreciseness is a consequence of the parallelism of the arithmetic units of both machines.* Imprecise interrupts will unquestionably cause difficulty in debugging programs, particularly system programs.
The Model 91 designers have provided a switch that puts the machine into non-overlapped mode, in which it runs at about Model 75 speed but with precise interrupts. This switch will be important for program debugging, and it can be set and reset by programming.[footnote as given in report] * Note: However, it is not a necessary consequence and could have been eliminated with some extra hardware in the CPU.
[Question from David Gifford] Have architectural deviations been allowed, to improve performance in a particular model?
[Andris] Padegs Performance normally is not the motivation for a deviation from architecture. A deviation normally is allowed after the fact. Furthermore, we do not grant deviations for the basic part of the architecture that would affect the normal operation of a program. For example, in the past we have allowed a deviation when the test light on the panel doesn't come on at the right time. A deviation normally is requested to avoid the cost of correcting the design.
[Question from David Gifford] Was the imprecise interrupt in the Model 91 an architectural deviation?
Padegs I guess you would have to call it a deviation, yes. I look at deviations as minor things that don't affect the basic functioning of a machine. Normally there are very few deviations in a machine, perhaps a couple or none at all. Imprecise interrupts in the Model 61 were a design decision. It was really almost like adding a new function to the machine. We told the users of that machine that in exchange for increased performance they could not recover from floating-point errors.
[Richard] Case Imprecise interrupts did not survive in the successor to the Model 91, by the way.
Corrections are welcome!
[History page] [Mark's homepage] [CPSC homepage] [Clemson Univ. homepage]
mark@cs.clemson.edu