For most purposes, the information given in the earlier chapters is sufficient to allow adequate and efficient programs to be written. Occasionally, however, it may be possible to increase the efficiency of program writing or execution; the following sections describe how this may be effected.
In the great majority of programs, the user will wish to take advantage of the one level store concept and will regard the core store and drums as a single, large main store. Programs are written as if the entire store were core store, and the Supervisor will automatically control the transfer of 512-word blocks between the drums and the core store as needed.
However, circumstances can arise in which it is useful to exercise some degree of control over these block transfers, both to ensure that blocks of information are already available in the core store when required, and to clear space in the core store by releasing blocks to the drums as soon as they are no longer needed; the extracodes provided for these purposes are all designed to assist towards greater economy of time by the avoidance of unnecessary Supervisor drum transfers. To understand just how the programmer may assist the Supervisor, it is necessary to consider the means provided for the regulation of automatic drum transfers.
In addition to any store location explicit in each instruction, there is implicit a store reference to the location containing the instruction word itself; in either case, the address is taken to specify both a block and a word within that block. The block address is invariably interpreted as a store request, and the Supervisor will initiate a drum transfer if the block is not already in the core store. Normally there will be only one copy of a particular block, occupying either a page in the core store or a sector on one of the drums. With each 512-word page of the core store there is associated a Page Address Register (P.A.R.) containing the number of the block occupying the page at any particular time; there is also a lock-out digit which is set whenever the page is involved in a drum or peripheral transfer and so is not available to the main program. At every store request, the block address is automatically compared with the content of each P.A.R.; if a coincidence is found, the store reference is completed by the extraction of the required word from the appropriate page. Otherwise a non-equivalence interrupt occurs and the Supervisor drum transfer program is entered; this is in two parts, one to carry out the actual block transfers and the other to decide which page of information should next be transferred to a drum to make space available in the core store.
All requests for information transfers between the core and the drum stores, whether originated by the Supervisor or called for directly by an object program, are placed in a drum queue holding up to 64 entries which are dealt with in the order of their occurrence. The drum transfer routine is re-entered repeatedly until the queue is cleared.
It is arranged that there shall always be at least one free page in the core store, so that, whenever the drum transfer routine is entered, the first read request in the drum queue can be implemented. Then, whilst this transfer is taking place, the drum transfer learning program decides which page may next be freed by writing its contents away to a drum. This decision is made on the basis of the frequency of past references to each block of information, and with the intention of choosing the core store page least likely to be referred to.
The drum learning program only attempts to predict future store needs in the light of past requests; it anticipates neither the termination of references to a particular block of information nor the imminent requirement for a new block. Hence, there arise in the main two ways of assisting the Supervisor by means of programmed drum transfers; firstly by releasing core store pages no longer required, and secondly by initiating the reading of a new block of information from a drum to the core store before it is actually referred to. These are both of marked advantage to the system as a whole: the first plainly helps towards efficient utilisation of the available facilities, and the second can often prove of even greater benefit and economy by reducing the time spent in waiting for drum transfers to be completed - this is especially significant in the inner loops of a program. It will be appreciated that the time during which a program is held up waiting for a drum transfer is still wasteful, notwithstanding time-sharing, since it may take from 1.3 to 2.7 milliseconds, depending on the core store size, to switch to another program and a similar length of time to switch back later.
Ideally, a read transfer should be timed to reach completion only just before the first reference is made to the block; otherwise the Supervisor may choose to write the block away to the drum again before the program comes to use it. The actual transfer of a block of 512 words takes 2 milliseconds, but there is an initial delay of up to 12 milliseconds, the revolution-time of the drum.
The core store of Atlas is arranged in 4096 word stacks, with 16 pages of information sharing each pair of stacks, and each stack having its own access equipment; to take advantage of this, and so to attain maximum speed, operands and instructions should be arranged, as far as is possible, in different pairs of stacks. At Manchester, the Supervisor endeavours to read down instructions to pages 0 to 15 and operands to the remaining pages of the machine; in the event of a non-equivalence interrupt the preference is automatic, being determined by the non-availability in the core store of an operand or an instruction as the case may be; in the case of a programmed drum transfer, the preference must be indicated by affixing a bit 1 before the address of a block of instructions, and 0 before the address of a block of operands. This preference bit will be the most significant bit of n (singly modified) or of ba where appropriate. At the other installations where there is more store, instructions and operands are placed in different pairs of stacks whenever possible. Pages 0-15 form one stack pair, as do pages 15-31, 32-47, etc.
Extracodes are available for the purposes we have discussed and these will now be described with the help of the following notation:
In all but two of the extracodes which follow, whenever information is transferred to a new block, the old block is made free. The exceptions are 1162 and 1163, where a block is to be duplicated leaving the original copy intact.
Further, when a block is quoted as the destination of an information transfer, either directly or as the result of renaming, any existing block of the same name is lost. This will apply even if the name quoted as that of the source is unallocated, and will in fact be the only action taken in such a case. Also, in those instructions referring to two block addresses, these addresses should not be the same.
b91' = c and c' = n if block number ≥ ba newly defined.
Henceforth, each time a block with a number ≥ ba is newly defined by a non-equivalence, store current control in B91 and jump to n. The block number in ba occupies bits 1 to 11 and the remaining bits of ba are ignored. The contents of Ba are undisturbed. The instruction causing the nonequivalence is not executed. n is singly modified.
ba' = smallest block label ≥ n defined.
Place in bits 1 to 11 of ba the smallest block number ≥ n which is defined for this program. The remaining bits of ba are left cleared. Only bits 1 to 11 of n are used. n is singly modified. If all the program's blocks are < n, then bit 0 of ba' is made 1 and the remaining bits are cleared.
Read block P.
If P is not already in the core store, the transfer request is inserted in the drum queue exactly as if a non-equivalence interrupt had occurred, but control is restored to the object program immediately the drum queue entry has been made. Should the queue be already full, the object program will be halted. until the entry can be inserted.
Release block P from the core store.
This extracode adjusts the parameters used by the drum learning program so as to cause it to choose block P next for writing away to the drum store, if this has not already occurred. No entry is made in the drum queue and the transfer will in general take place earlier than if extracode 1165 (below) had been used.
Duplicate block P1 as P2 in the core store.
Any existing block P2 is always lost, and, if P1 is allocated, a oopy of it will be formed as P2 in the core store. Unless the drum store is full, block P1 will finally be located there; otherwise P1 will be left in the core store. P1 ≠ P2.
Duplicate block P1 as P2 in the drum store.
Provided P1 is allocated, the effect of this extracode is to form a duplicate copy of it. It will be arranged that one copy shall always be left in the core store and named P1; the second copy, named P2, will be put in the drum store unless this is full, in whioh case it will be left in the core store. Any previously existing block P2 will be lost in all cases. P1 ≠ P2.
Rename block P1 as P2.
If P1 is allocated, the appropriate entry in the drum directory or the core store P.A.R. is altered to P2. Any P2 previously existing will be lost. There must be at least one more block allowed for in the job description than is defined at that moment. P1 ≠ P2.
Write block P.
Provided P is allocated and. is not already on a drum, it is transferred to the next empty sector. Should the drum store be full, block P is released, precisely as in 1161.
Read block P to absolute page d.
This extracode makes possible full control of the store by those exceptional. programs for which this may be worthwhile. Before using 1166, the program must set a trap in case page d is locked down and reserved by the Supervisor. d is in the integer position of ba and defines the absolute number of a page in the core store to which block P is to be transferred. Before this transfer takes place, any existing contents of d are copied to a free page.
Lose block P.
If P is allocated, the page or sector occupied by it is made free.
Clear new blocks / Do not clear new blocks.
When a program refers to a main store block for the first time, the Supervisor allocates a free page of the core store; floating-point zero will be written in all 512 words if the clear blocks switch is set. Initially, this switch is set to clear all new blocks, but it may subsequently be set or reset by means of extracode 1170 according to the sign bit of n:-
n ≥ 0 Clear new blocks.
n < 0 Do not clear new blocks.
Change store allocation to n blocks.
Each program has some number of main store blocks assigned to it. This number may be altered during the execution of the program by the use of extracode 1171. If there are less than n blocks available in the store, then the program will be faulted for ILLEGAL FUNCTION and EXCESS BLOCKS.
Set ba' = number of pages available.
This extracode provides an estimate of the number of core store pages available to the program at a particular moment. It cannot be assumed that this number of pages will continue to be available, since the core store allocations are always fluctuating.
Set ba' = number of blocks available.
At a particular moment, this extracode records the maximum number of main store blocks available, consisting of all unallocated blocks together with those already allocated to the program itself.
Reserve band D.
A complete band of the drum store is reserved for the program and may subsequently be referred to as band D.
Read K + 1 blocks from band d, starting at sector k.
d must already be defined by extracode 1174. The K + 1 successive sectors k, k + 1,..., k + K are read to store blocks P, P + 1,..., P + K. If K is 6 (or 7), sectors k (or k and k + 1) will be read twice. Thus if K = 6, blocks P and P + 6 will both contain sector k. If k is 6 or 7, it is taken as 0 or 1 respectively. All blocks involved are locked down until the entire transfer is complete.
Write K + 1 blocks to drum band d starting at sector k.
d must already be defined by extracode 1174. This extracode writes store blocks P, P + 1..., P + K to drum sectors k, k + 1 ,..., k + K. Sectors 6 and 7 are the same as sectors 0 and 1. If K exceeds 5 some of the earlier blocks are overwritten. Thus if K= 6, sector k will finally contain block P + K rather than block P.
Lose band D.
The band of the drum store previously reserved as logical band D is freed and made available for general use.
The following table gives the approximate times in microseconds achieved on Atlas for various instructions. The figures are averages for obeying long sequences of each instruction, with the instructions and operands in different stacks of the core store.
Type of Instruction | Number of Address Modifications |
Time µsecs |
---|---|---|
am' = am + s | 0 | 1.6 |
1 | 1.9 | |
2 | 2.4 | |
am' = am × s | 0, 1 or 2 | 5.9 |
am' = am / s | 0, 1 or 2 | 22.5 |
ba' = ba + s | 0 | 1.7 |
1 | 2.0 |
It is not possible to time a single instruction because, in general this is dependent on
This is illustrated when evaluating a polynomial, using a central loop involving a singly modified accumulator addition, an accumulator multiplication, and a test, count and jump instruction. The average time for this loop is 8.5 µsec, of which the accumulator operations would take 7.8 µsec if their individual times are simply added. This leaves only 0.5 µsec for the test, count and jump, although the average figure for a series of jump instructions on their own might be ten times as large.
We shall consider those factors which control the time taken to obey instructions, to show what advantage can be taken of them in optimizing a program loop which has to be executed many times.
The main core store consists of pairs of 4096-word stacks. Each stack can be regarded as a physically independent store, and sequential address positions occur in the two stacks of a pair alternately, the even addresses in one stack, the odd in the other. The cycle time of the core store is 2 µs., that is, the time after reading or writing a number before another number can be read from or written to the same stack is 2 µs.
To reduce the effective access time, instructions are always read in pairs and held in two buffer registers called Present Instruction Even (PIE) and Present Instruction Odd (PIO) whilst waiting to be obeyed. Instructions are executed from PIE, the odd instruction being copied from PIO into PIE as soon as the even instruction has been initiated.
Because of the 2 µs cycle time for each stack the programmer should separate instructions which refer to operands in the same stack.
For example, the instructions
121 1 0 0 324 1 0 A4 362 1 0 A4
would be executed more quickly if written as
324 1 0 A4 121 1 0 0 362 1 0 A4
The maximum overlap is obtained when alternate operands come from alternate stacks.
The Supervisor attempts to organise the store so that instructions and operands are placed in different pairs of stacks. On the Manchester University Atlas, wherever possible, instructions are kept in pages 0-15 and operands in pages 16-31. The programmer can assist the Supervisor to do this by using the extracodes described in 12.1. These are of most use for jobs with a large amount of data. It is then useful to request drum transfers in anticipation, and to release from the core store blocks which will not be wanted again for some time.
Instructions on Atlas are overlapped as far as possible. For example, in a sequence of singly-modified accumulator instructions, the computer is obeying four instructions for one quarter of the time, two instructions for one fifth of the time, and three instructions for the remainder of the time.
This overlapping is possible because the accumulator arithmetic, the B-register arithmetic, the function decoding, the B-store, and the main core store are independent of each other to a large extent. A number of rules which enable the programmer to gain as much advantage as possible from the overlapping are given below. It should be noted that these rules cannot always be guaranteed to establish the best way of arranging any particular loop, as in some cases this can only be done by actually running the program; nevertheless the application of these rules, as far as possible, will normally lead to a time reasonably close to the optimum being obtained.
Following a store-write instruction, no further instructions or operands can be extracted until the writing operation is completed. Many typical program loops, however, include such an instruction. It is usually possible to have this instruction at the beginning of the loop, and this enables the B-type instruction and return jump to be obeyed and overlap any accumulator arithmetic still going on. As mentioned in (a) above, the store-write instruction should. preferably be in an odd-numbered address. From these two rules, two possible ways of arranging a loop, depending on whether it has an odd or an even number of instructions, emerge.
Examples:
1. Odd number of instructions
0 Even 1 Odd Store-write 2 Even Accumulator Instruction 3 Odd Accumulator Instruction 4 Even Step B-line 5 Odd Return jump to 1
2. Even number of instructions
0 Even Step B-line 1 Odd Store-write 2 Even Accumulator Instruction 3 Odd Return jump to 0
A delay occurs if a B-register is operated on in the Ba position and then used as a modifier in the next instruction. This should therefore be avoided if possible e.g. by inserting some other instruction in between. Note, however, that
124 1 0 1 300 1 2 0
is preferable to
124 1 0 1 300 2 1 0
Where the above rules conflict, the order in which they are given should be taken as the order of importance.
Branching is a facility which enables different parts of the same program to operate in parallel, using the time-sharing process. Such parallel operation is of value if some parts of the program are liable to be held up waiting for peripheral transfers whilst other parts are still able to proceed. It is important to note that simple operation of peripheral devices in parallel with computing is available without recourse to branching; normally, the program itself is only held up if it attempts to refer to the locations involved in a transfer before the transfer has been completed. Branching is an additional facility which is intended to permit parallel operation of two or more different processes which are liable to be held up by peripheral transfers, where each process involves some computation or organization and does not consist merely of peripheral transfers.
When a block transfer to or from a drum or magnetic tape has been initiated, by means of a drum or tape block transfer extracode, the program is allowed to proceed as long as it does not refer to the main store block involved in the transfer. If it does refer to that block, it is held up until the transfer has been completed.
Variable length tape transfers operate by using part of the main store as a buffer. It is usually possible to keep sufficient information in the buffer to permit the actual transfer, between the buffer and the specified store address, to take place as soon as the transfer instruction is encountered. Otherwise the program will be held up until the transfer is complete.
Other peripheral devices, apart from the drums and magnetic tapes, are not normally controlled directly by the program. Instead, the input documents are read and stored on a system magnetic tape before the program is initiated, and output documents are stored on a system tape and printed after the program has been completed.
Permit Ba Branches (2 ≤ Ba ≤ 32)
Before any branching can take place, the program must obey an 1103 instruction, which enables the Supervisor to prepare for branching.
This instruction normally takes the form
1103 Ba 0 pD1
After obeying it, the program is permitted to have up to Ba live branches, including the main program, in progress at anyone time; the main program is defined as branch O. When the Supervisor switches from one branch to another it will preserve certain standard information and. also the contents of index registers Bp, B(p + 1), B(p + 2),....., B99. Note that if the N-address is zero all index registers are preserved, and if p = 91 only the extracode index-registers are preserved (these are usually essential).
Start Branch Ba at n (0 ≤ Ba ≤ 63)
The current branch of the program continues at the next instruction, but a new branch, with number and priority Ba, is started at address n. The highest priority is given to the highest-numbered branch: if other branches with the same number Ba have been defined previously, they will take higher priority than the new branch. The main program is initially defined as branch number O.
Kill Branch Ba or Current Branch
Kill all branches with the number Ba. If Ba = 64, kill the current branch. This prevents any further instructions being obeyed in the specified branches, but peripheral transfers already requested will be completed.
Wait until Branch Ba is Dead.
Halt the current branch of the program if any branch numbered Ba is still live. Proceed to the next instruction when all branches numbered Ba are dead.
Jump if Branch Ba Live.
Transfer control to address n if any branch numbered Ba is still live. Otherwise proceed to the next instruction.
A branch is usually started at some point in a program where it is required to carry out two different processes, at least one of which is liable to be held up by peripheral transfers. Usually, the more severely peripheral limited process is put in the new branch and this is given higher priority. When the program is obeyed, the higher priority branch is allowed to proceed until it is held up waiting for a peripheral transfer; control is then transferred to the other branch, which proceeds either until it is held up, or until the higher priority branch is ready to resume. Similarly, if there are several branches, the Supervisor ensures that control always passes to the highest-priority branch able to proceed. Each time control is switched from one branch to another, the Supervisor stores and restores the contents of the following registers and indicators:
Thus, each branch can use these registers as though it were one single program uninterrupted by other branches. It is, however, necessary to ensure that two branches which may operate simultaneously do not use the same main store locations, or index registers which are not preserved. It should be noted that the selected Input and Output are not preserved, and therefore input and output can each take place in only one branch at a time.
Once a branch has been started it can be regarded as a live branch, and it remains live, even when it is held up, until its task has been completed. When a branch has completed its task, it must die, and this it does by obeying an 1105 instruction, usually with Ba = 64.
When one branch of a program is ready to make use of the work done by another, it must first ensure that the work has been completed. This may be done by obeying an 1106 instruction, which causes the current branch of the program to be held up until the specified branch is dead, having completed its task.
A simple example of the need for branching arises when it is required to scan a magnetic tape in order to process a selected sample of the information on it. The processing routine and the tape-scanning routine can then be written as two separate branches, with the tape-scanning routine as the higher-priority branch.
Example:
It is required to scan sections 1 to 3000 of tape 4 and to apply a lengthy processing routine R3 to the information in about 25% of these sections. The sections to be processed are to be identified by having a number greater than 0.32 in the first word of the section. The program to do this could be written as follows:
Branch 0 1103 2 0 89D1 Prepare to use 2 branches, preserving b89-b99. 1001 4 0 1 Search for section 1, tape 4 121 89 0 2999 Set count for 3000 sections 121 16 0 0 Clear marker in B16 1104 1 0 A6 Start branch 1 at A6 5) 1106 1 0 0 Wait until branch 1 dead 215 127 16 A12 Exit if last section processed 121 10 0 4: Rename block 3 1164 10 0 3: as block 4 1104 1 0 A7 Start branch 1 at A7 121 90 0 A5 Set link for return 121 127 0 A1/3 Enter R3 to process block 4 Branch 1 6) 1002 4 0 3: Next section to block 3 324 0 0 3: First number in section 321 0 0 1A8 Subtract 0.32 236 127 0 A8 Exit if number ≥ 0.32 7) 203 127 89 A6 Count tape sections 121 16 0 7 Mark b16 non-zero 1105 64 0 0 Kill current branch +0.32 R3 R3 (Part of Branch 0) 1) ..... Routine to process ..... the information ..... in block 4
The chart below shows how control would pass from one branch to the other in a typical sequence of operations when the program is obeyed. Interruptions from the Supervisor and higher-priority programs have been excluded because they would complicate the chart without altering the sequence significantly. The sequence of operations starts at the top with the beginning of the program, runs through the first entry to branch 1, and then cycles round a loop in which branches 0 and 1 operate in parallel. The chart shows branch 0 completing its work before branch 1 has found the next section to be processed, but branch 1 might equally well be finished first, that is, a further required section may be found during processing. It should be remembered that each entry to branch 1 takes only a few microseconds, whereas 64 milliseconds must elapse between successive entries to branch 1 to read one more tape section.
Notes:
A. This loop will be repeated until a required section is found.
B. If a required section is found., then Branch 1 will be killed.
When an 1103 instruction is obeyed, the Supervisor assigns sufficient storage space for the specified number of branches. This storage space is taken out of the main store allocated to the program, either by the job description or by a subsequent use of the extracode 1171, and will be counted in the estimates made by extracodes 1172 and 1173; it is therefore necessary for the programmer to know how much store is required by the Supervisor for branching purposes. Many cases should be covered by the following table, showing the maximum number of branches that can be accommodated in 1, 2 or 3 blocks, depending on the number of index registers preserved.
Index Registers Preserved |
Storage Space Allocated | ||
---|---|---|---|
1 Block | 2 Blocks | 3 Blocks | |
Maximum Number of Branches Permitted | |||
B0 to B99 | 3 | 10 | 18 |
B30 to B99 | 4 | 14 | 24 |
B50 to B99 | 5 | 17 | 29 |
B70 to B99 | 6 | 22 | 32 |
B80 to B99 | 8 | 27 | 32 |
B90 to B99 | 10 | 32 | - |
If it is necessary to estimate the store required in some case not covered by the above table, it is probably easiest to do this by considering the way in which the Supervisor allocates this store. It takes 300 words at the beginning of the first block to store branching routines, and follows these by 5 words for each branch requested in the 1103 instruction. Each branch is then allocated a further (11 + ½m) words, where m is the number of index registers in the range 0 to 99 that are to be preserved. The (11 + ½m) words for one branch must all be in the same block, and if less than (11 + ½m) words are left at the end of a b lock the (11 + ½m) words for the next branch will start at the beginning of a new block.
As each basic instruction is obeyed, an instruction counter is stepped on, normally by one, but by two for multiplication orders, and by four for division. Each time the counter reaches 2048, an interrupt occurs and an instruction interrupt counter is stepped on by one. This latter counter is used by the Supervisor in monitoring the program, but may also be read by the program using extracode 1136.
Read instruction count.
Set am' to the number of instructions obeyed from the start of the program; this will be a fixed-point integer with exponent 16, and will be a multiple of 2048.
Besides this count, the program may also use a local instruction counter. A trappable fault will be recognized when this count expires, which may provide a convenient way to end an iterative loop, since the counter may be set as well as read by program.
Set local timer.
Set the local instruction counter to n × 2048 instructions. The Supervisor will override any attempt to set the counter to a figure in excess of the amount of allotted time remaining.
local timer' = 2048n
Read local timer.
Read local instruction counter into Ba in units of 2048 instructions.
ba' = local timer
Most programs are compiled completely before they are entered, and therefore it is not normally necessary to retain the compiler in store during the program's execution. The E-type of directive is the only enter directive which deletes the compiler from the store before transferring control to the object program, and so is the most commonly used.
In some circumstances, however, it is necessary to enter the program, and then compile more program later. The first entry may be to actually execute part of the object program, or it may be only to set certain parameters. The compiler must be retained in store for these purposes, and so either an ER or EX type of enter directive must be used. The compiler uses store locations J3 (3/4 × 220) and above which should not normally be altered by the program, although no check is made except when actually compiling.
The EX-directive is intended for obeying interludes during compiling; an interlude would normally consist of a few instructions only, or of none at all. For example, if it were required to have any ABL fault printing on some output stream other than Output 0, then a one instruction interlude to select output would suffice. If it is only required to set parameters, then the address specified in the EX-directive should cause immediate re-entry to the compiler; such a directive in fact occurs near the beginning of L100, the general input routine, to determine the various optional parameter settings. The EX-directive does not call down any library routines; if these are required in the interlude, they must be called by one of the L-directives before obeying the EX entry. No distinction is made by the Supervisor between compiling proper and obeying an interlude, i.e. the Compile/Execute switch is not changed.
The ER-directive is designed to allow part of a program to be compiled and executed before reading more program, and provides most of the facilities of an E-directive, including the compilation of any library routines mentioned but not called earlier in the program. The routine current when the ER-directive is obeyed will be terminated before more program is read. (The EX-directive does not do this.) The Supervisor recognises that an object program is being executed, and as with the E-type of directive, the Compile/Execute switch is set to Execute.
Two types of list within the compiler are used in connection with parameters in a program. The parameter lists contain all those parameters which are determinate, and if the program refers to a set parameter, these lists are used to replace the parameter by its value. If a program refers to a routine or global parameter before it has been set, then this is noted on a forward reference list, from which it is deleted when the parameter is determinate, and hence, so long as this list is not empty, there are some parameters still to be set. When the compiler is retained in store, these lists also remain, in the same state as when the enter directive was obeyed. If more program is to be read which uses parameters to refer back to the program compiled previously, then it is essential that the lists remain unaltered. If, however, the subsequent sections of program are to be compiled independently of the earlier part, or if the same parameters are to be used again with different values, then the lists must be cleared on re-entry to the compiler. Different re-entry points provide for both requirements, and are listed below. In every case, re-entry to the compiler does not alter the Compile/Execute switch. After compiling program, B1-B88 will be cleared, and B89 will contain the final transfer address. The other B-lines may be destroyed.
The compiler may also be used as a subroutine by a program, control returning to the main program when no more items are to be read. Again there are two modes of entry, depending on what is required of the compiler lists.
If the transfer address is written to location Y4P121, and the link to 129P121, then re-entry to the compiler at P120 will read more program, retaining the compiler lists. Return to the link address in the main program is effected by
EP129 ERP129 or EXP129
If the transfer address is written to B89, and the link to B90, when the compiler is re-entered at P120BY6, more program will be read as if the re-entry were to P120B, except that an attempt to compile into store at an address less than b89 will be faulted, and that the transfer address will be taken as *=b89 unless b89=0, when *=1:0. Return to the main program is again by EP129 etc.
P120, P121, and P129 are examples of special preset parameters, which are described in the next section.
Although for normal purposes only Preset Parameters 0 to 99 may be used, some above 100 do exist and these are used in special ways for special purposes. In some cases use of them causes special action by the compiler; in other cases they are used to convey information between the compiler and the program.
P100 to P109 are in many ways like ordinary Preset Parameters; they can be reset by the programmer and no special action is taken by the compiler on encountering them. However, they are initially set by the compiler at the start of compilation and they are referred to by the compiler during the course of compilation.
P110 to P119, if defined, may be reset by the programmer but will have initial values set for them by the compiler. However, whenever an Equation Directive for resetting them is encountered, special action is required by the compiler. An attempt to set one of these parameters not listed below is faulted.
P120 to P129 are preset by the compiler, but may not be reset by program. An attempt to do so is faulted. They are used to convey information from the compiler to the program.
P100 - Optional Printing
At the start of compiling ABL sets P100 to zero. Non-zero settings of P100 cause ABL to print various kinds of information during compiling.
P100 is treated by ABL as made up of 8 octal digits abcdefgh. Each octal digit controls the printing of one kind of information, as indicated.
If the least significant bit of an octal digit is 1 the information controlled by this digit will be printed - on a new line if the middle bit of the digit is 1 and on the same line if the middle bit is 0. If the least significant bit is 0, then the other two bits are ignored. If the most significant bit of the second digit (b) is 1, then printing on a new line will occur when a library routine is compiled. Otherwise the most significant bits of the digits are ignored.
P100 is preserved and set to zero before compilation of each library routine and restored afterwards, so that there will be no other optional printing, unless the library routine contains a P100 = directive.
The kind of printing controlled by each octal digit is as follows. All printing is preceded by a space, except R. L is printed on a new line.
Octal Digit |
ABL prints this | when it meets this |
---|---|---|
a | *=p | *=expression |
b | Ra *=q La.b N *=q |
Ra Library routine name N compiled |
c | Z *=q | Z |
d | ignored | unassigned |
e | p | P111=expression (see P111) |
f | E p | E expression |
g | E R p | ER expression |
h | E X p | EX expression |
where
Some examples of useful settings of P100 are
P101 - Permitted Number of Errors
At the start of compiling ABL sets P101 to 0.2. This is equivalent to infinity since for each error met ABL reduces P101 by 1, and when it reaches zero stops compiling and ends the run after printing
TOO MANY ERRORS
The program may set P101 = n where n is any expression. Compiling will stop when n + 1 errors have been met. P101 = 0 causes ABL to stop on the first error met, which may be useful for a developed program.
No matter how many labels remain unset when the E directive is met ABL lumps them all together as one error for the purpose of counting errors.
P102 - Entry Despite Faults
At the start of compiling ABL sets P102 = 0.2. If any errors have been found, an EX directive will be obeyed, but an E or ER will not, and ABL will print
ERRORS DO NOT ENTER
and end the run.
P102 = 0 allows all 3 E directives to be obeyed despite errors
P102 = 0.3 forbids all 3 E directives after errors
P102 = 0.1 allows E and ER but forbids EX after errors
P104 - Setting Private Monitor
P104 is the address of a private monitor routine, which is set up each time any Enter Directive is encountered. Thus if any monitors occur after the Enter Directive (including an immediate entry to an address holding an Illegal Function due to a wrong Enter Directive address), these will give rise to an entry to Private Monitor according to the rules of extracode 1112 (see section 11.3). If P104 is negative any current setting is terminated.
P110 - Change of Program Location
At the start of compiling, ABL sets P110 = O. A non-zero setting of, P110 specifies the difference between * as evaluated in expressions (say *1) and * indicating where items are to be stored (say *2), i.e. P110 = *1 - *2. Setting P110≠0 permits the compiling of program into one set of store addresses - *2 - for later execution in another set of store addresses - *1 (e.g. after Renaming or after storing on magnetic tape). Thus, for example, a program which is to be executed starting at address J3 but compiled initially into store starting at J1 would have the directives
P110 = J2 * = J3 at its start
P111 - Expression printing
When ABL meets the equation
P111 = expression
it evaluates the expression and sets P111 in the usual way. If the appropriate bit of P100 is set, the value of the expression is immediately printed out.
P112 - Unused
This parameter is used by ABL on Atlas 2. No fault will occur if the program sets P112, the value being assigned in the usual way, but the remainder of the line on which the setting occurs will be ignored. The compiler initially sets P112 to J3. P112 should not normally be used with ABL on Atlas 1.
P115 - Change of Input Stream
The equation P115 = n (n is any expression) causes ABL to start reading program from the programmer's input stream n. The rest of the line on which P115 occurs will be ignored.
P120 - Re-entry to the Compiler
This is described in the previous section.
P121 - Preset Parameter List
P121 is the address of the start of the compiler's Preset Parameter list. Half-word P121+n contains the value of parameter n if it has been set. This list can of course only be referred to by a program entered by an ER or an EX directive. This is a list of alternate half-words, the other half-words of which are used for other purposes by the compiler and should not be disturbed.
P122 - State of Preset Parameters
P122 is the address of the start of the compiler's list which indicates whether the Preset Parameters are set or not. Bit 1 of half-word P122+n is a 1 if parameter n is unset, 0 if set. This is a list of alternate half-words, and neither the other bits of the half-words, nor the other half-words should be disturbed if subsequent use of the compiler is intended.
Example:
An interlude to set P17 = 0 if' P35 = 0 and to leave P17 unaltered otherwise
5) 121 1 0 P35 215 127 1 P120 113 0 0 P121+17 121 1 0 J2' 117 1 0 P122+17 121 127 0 P120 EXA5
P123 - Characters Count with C directives
P123 indicates the number of characters read by means of the previous C directive, described in section 5.10.
P129 - Return from Compiler
P129 is the return address when the ABL compiler is used as a subroutine (see previous section).
Library routines are given numbers and names; the program refers to them by number, but the name may be useful to indicate their purpose. The standard input and output routines, described in Chapter 8, have been given names as follows:
L1 GENERAL OUTPUT L100 GENERAL INPUT L199 LINE RECONSTRUCTION
L199 is used by L100, and hence, if the library routine is being compiled implicitly, (whether by an L directive or by an Enter directive), L199 will automatically be compiled with L100 because the latter refers to it, but, if it is required to compile the input library explicitly into a part of the programmer's store area, then it is necessary to use both of the directives
L100 L199
All undefined library routines have the name NONEXISTENT and there is a special device to make the optional printing as described for P100 (section 12.6) compulsory for non-existent routines. For example, output such as
L10 NONEXISTENT* = 2:36.3
will result from either an attempt to call for L10 explicitly or when an attempt is made to compile it by an L or Enter Directive when it has been referred to implicitly. The latter would also result in monitor printing about unset labels.
Any defining of a Private library routine (see below) will cause suspension of the NONEXISTENT monitoring for that routine.
Private Library Routines may be incorporated in the normal program input stream and may be referred to in the program in the same way as public library routines.
The routine is headed by two lines:
RLc < Name>
where c is the number assigned to the library routine and < Name > is the name of the routine. If no name is required, this line must be left blank, and then a blank title will appear in any optional printing. The Name must not consist of the two-character record ZL.
The routine is terminated by the two-character record ZL. This is not line-reconstructed and may not contain any spaces, erases, backspaces, tabs etc.
The library routine may consist of a single routine (routine 0) or of one or more routines headed by routine directives and optionally terminated by Z directives in the usual way. It may contain any of the normal ABL forms except the directives
L, La, La.b, ER expression, E expression, RLc
All will be monitored. No T or C directive within the library routine may be followed by the two character record ZL.
When the R1 directive is encountered, the library routine following is simply copied character by character into the compiler's store area. The routine is not, at that time, compiled or placed in the programmer's store area. This is achieved in the normal way, by an E, ER or L - type of directive.
If a private library routine is given the same number as a public library routine, it replaces the public one for the remainder of that program. This is convenient for the development of new versions of existing public routines.
Private Library Routines must precede any calls for them in the body of the program. The best place for them is at the beginning of the program stream.
A private library of routines required by people working in some limited field (e.g. properties of steam) may be formed by putting the routines on to a titled paper tape (as pseudo-data) and terminating them by P115=O. The master programmer may then write, for example,
INPUT 15 STEAM LIBRARY
in his Job Description, and
P115 = 15
at the head of his program to incorporate these routines effectively as above.
Each private library routine incorporated in the way described in this section counts as one line from the point of view of line counting for error monitoring of the subsequent program.
The directive
P115 = expression
within a library routine, will not cause monitoring, although it should never normally be needed or used. Its effect is in fact to cause switching of the input stream after completing the compilation of the current library routine or routines at the point where these are compiled (i.e. at an L or Enter directive).
This is done by means of a special job, using a standard program of the system. Essentially, the routine to be incorporated is put at the head of the program stream in the normal way, as described above, and is followed by the standard program. This program uses no labels as the non-empty parameter list would otherwise become part of the compiler.
If the routine being incorporated is a new version of an already existing routine, then this latter is not destroyed or overwritten on the compiler tape. The reference to it in the compiler's library routine list is simply changed and so it becomes dead. A separate special program (or prelude to the above program) may be used from time to time to clear out all dead library routines, but this clears out all live ones as well, so that after this clearing out operation all public library routines must be reincorporated.
The following conventions are recommended for library routine writers and users:-
A routine parameter of the master program can be referred to within a library routine by treating the master program as if it were library routine 0, e.g. .A6/3L0 is A6/3 of the master program.
To correct a small program, it is usually simplest to re-punch the tape or cards, making alterations as necessary. This is impracticable for larger programs, but the facilities of ABL may be used to help make corrections.
Very often it is possible to make corrections by overwriting certain store locations, using a *-directive. The corrections, however, must be compiled after the faulty items, or the faults will overwrite the alterations. Hence, the corrections are normally placed just before the enter directive. An enter directive may cause library routines to be compiled from the current transfer address, and so to prevent the program being overwritten from the faulty item onwards it is necessary to insert the correction in one of the following ways.
a) 1) <Last item of program proper> *= 6A10/4 <Correction> * = 1A1 EA40 b) <Last item of program proper> L * = 6A10/4 <Correction> 1) EA40
In case b), however, if data is read to A1 onwards, this will also overwrite program when the correction is inserted.
It may be convenient to end a program with
R10 1) <Last item of program proper> P115 = 15 * = 1A1/10 EA40/3
so that corrections, if any, will be read from input stream 15, which ends with P115 = O.
When routine parameters are mentioned in a correction, without specifying any routine, i.e. /n is omitted, the routine to which they refer will be that current before the correction, rather than that at the location to be corrected.
Difficulties may be encountered when using *= to overwrite an item containing forward references. The result can be predicted from the following notes on ABL's handling of forward references.
Two lists are concerned, the forward reference list in which are partly evaluated expressions containing forward references, and the parameter list in which are all parameters found in the program with their values if set.
Condensation of the forward reference list
If, on reaching the end of the forward reference list, any parameters have been set during 1, the process is repeated from the beginning.
For example, suppose the program begins
*=0, HA3 A3=A4-1 *=0, HA3-1 A4=4
On reaching the 4th line the forward reference list will contain the expressions A3, A4-1, A3-1. When the 4th line is read the list becomes
A3, A4-1, A3-1, 4;
the 4 is evaluated immediately and A4 becomes set; the list is then completely re-condensed.
A3 is not yet set and so remains.
A4-1 is evaluated and A3 gets set.
A5-1 is evaluated and planted.
The list now consists of only A3 and since a parameter was set in the last condensation the list is re-condensed.
A3 gets evaluated and planted.
It is seen that in this case the half-word ends with the value of A3, i.e. 3.
If a routine or global parameter is optionally set more than once, but is not set otherwise then the first optional setting will be implemented. The subsequent optional settings will not be checked for faults.
A different mode of correction uses the library facilities of the ABL compiler. A copy of the compiler would be dumped initially onto a private magnetic tape, and subsequently used to compile routines as a library on to the tape. As these routines are corrected, the new versions are introduced, replacing the faulty library routines as described in section 12.7.4.
Floating-point numbers may be represented in the form a(b:c):d or Ka(b:c):d as described in section 5.11. Although the number may be within the range of the accumulator, compiling it may cause exponent overflow unless the following three limitations are observed.
When a parameter optionally set by a ?= directive is actually set elsewhere, the right hand side of the equation for the optional settings may not be checked.
After the final ABL fault printing, no new line is output.
ABL will read program incorrectly after 8191 = 213 - 1 printed lines without the implicit setting of a routine parameter by labelling.
In fault pointing, the line count will be taken modulo 212.
No check is made that function codes exist. One and two digit functions are right justified into the function bits.
When obeying program, the compiled value of * will be different from the current value of control, as b127 is stepped on by 1 before starting to obey an instruction. Thus
121 69 127 _*
would set b69' = 1.
121 127 127 -1
causes a loop stop.
When more than twelve digits in a number are printed by the general output routine L1, digits after the twelfth may be wrong. L1 also may give exponent overflow attempting to print the following numbers:-
Non-zero Mantissa Exponent -1 +127 ±x (x ≠ -1) -127 ±x -128
The extracodes given below are used mainly by system programmers. They complete the list of Atlas 1 extracodes.
v7' = n
and hoot if the least significant integer bit of n is 1. (bit 20)
ba' = v7 & n
Mask the digits of the engineer's handswitches with n, and read them to bits 16-23 of Ba.
Line 7 of the central computer V-store consists of 8-bits. Only bits 16-23 may be read, being set from the engineer's handswitches. Other bits are read as zero. Bit 20 controls the hooter and may be set by program; writing to the other bits is ignored.
Read 'parameter' Ba of program to store; starting at location S.
Ba | Parameter |
---|---|
0 | Job title (10 words) |
1 | Computing time estimate, in seconds, in digits 0-23 (One half-word) |
2 | Execution time estimate, (One half-word) |
3 | Number of store blocks required, in digits 1-11 ( One half-word) |
4 | 'Parameter' in Job Description (One-half-word) |
5 | Logical tape numbers defined (8 half-words). The jth digit (0 ≤ j ≤ 15) of the ith half-word is a 1 if tape number 16i + j is defined. |
6 | Inputs defined (One half-word). The ith digit (0 ≤ i :≤ 15) is 1 if input stream i is defined. |
7 | Outputs defined (One half-word) As 6. |
Define Compiler
Ba = Tape Number to which the compiler is to be written. If Ba = 127, and if the compiler name specified (see below) appears in the Supervisor Directory, the compiler will be written to the current Supervisor tape. In this case there will be two loop stops with J70707070 in B120 since this extracode will use 1143, 0, 0, 0.1 and 1143, 0, 0, 0.2 (see below).
The five half-words S to S+2 contain the following parameters:
Notes:
The compiler name should be right justified within each half-word, but the first four characters should be put in the first half-word, e.g.
ABL= J00414254, 0 HARTRAN= J50416264, J00624156
The following fault indications may be printed:
COMPILER NAME NOT LISTED COMPILER NOW U/S TAPEFAIL COMPILER TOO BIG WRITING TO SPECIFIED BLOCKS ON SYSTEM TAPE IS PROHIBITED
The first and third can only occur if Ba = 127. The fourth implies the facility described in Note (ii) has been used with Ba = 127.
End compiling
Reserve Supervisor Tape
Ba should be zero
n should be 0.1, 0.2 or 0, with the following meanings:
An operator request will be necessary before using this extracode.
Call Compiler
In both cases, if ba = 0, the standard entry point will be used.
With 1150 and 1151, ba, P and K are as defined in section 12.1.
Assign ba blocks, labels P to (P + ba ..1) to overflow K.
This extracode enables a program or compiler to temporarily hand blocks to the Supervisor, which may write them to the system dump tape. Subsequent use of these labels in the program causes new blocks to be assigned. The block labels are retained in the overflow region and additions to this region must bear distinct labels. If ba = 0, one block is transferred.
Set up ba blocks, labels P onwards, from overflow K.
This extracode recalls blocks previously written to the overflow region by use of 1150. Any existing blocks having these labels are overwritten. If ba = 0, one block is recalled. If these blocks do not exist in the overflow region, the program is monitored.
Enter extracode control at n if the 'In Supervisor' switch is set.
This extracode is used by various Supervisor Extracode Routines which are obeyed on main control. If the In Supervisor switch is not set, the program will be monitored.
Enter extracode control at n if the 'Process' switch is set.
This may only be used by Supervisor routines such as the monitor called in during the running of a main program.