Reading the 1962 book about Stretch. There are a lot of strange ideas and terminology in here.
“Edit” seems to be used in a sense I don’t understand.
IBM, especially at that time, had a habit of plagiarizing other people’s ideas, so it’s hard to tell what’s really new here; but Stretch was the first IBM computer, at least, to be pipelined, though this was called “virtual memory,” predating the use of that term to mean swapping to disk.
The idea of a “program” solving a “problem” for a “sponsor” is pretty far from our current understanding. So, too, the justification of “multiprogramming” (also “time-sharing”): “it becomes economically practical for a person seated at a console to observe his program during execution and interrupt it while considering the next step”, and also to “avoid delaying the calculator [!] for input-output”.
The disk storage (intended for 250,000 words per second, though only half of that was achieved: 8 megabits per second, seek time 150 ms; the terms “DASD” and “fixed disk” do not appear) was considered to avoid the need for “very-high-speed magnetic tapes.”
“On line” and “off line” also appear, in quotes, to describe different ways of entering input data: on directly connected devices, or using physically separate devices that produced magnetic tapes, “the fastest possible medium”.
“To an increasing extent, bits in even a scientific computer represent things other than numerical quantities: elements of a program metalanguage, alphabetic material, representations of graphs...”
A rather striking feature is that Stretch’s memory, in 1961, had a 2.1-μs cycle time, which they considered rather slow (thus its 64-bit word size). Nowadays the main memory of my laptop (whose CPU uses a 64-bit word size) can produce bursts substantially faster than this, at close to 1 ns, but random accesses still cost it nearly 100 ns, which is only 21× faster.
Stretch could address 2¹⁸ words, but the first 32 words were its registers; I imagine this wasn’t useful to user programs because of the memory-protection scheme explained below. Register 0 was the zero register, “a bottomless pit.”
Memory protection was provided as follows: “The interpretation and execution of instructions is monitored to make sure that the effective addresses are within boundaries defined by two [address-]boundary registers.”
Only floating-point math was bit-parallel; logical operations and fixed-point arithmetic were done by “the serial section” of “the arithmetic unit”, which was not yet called the ALU (though “Instructions that combine bits by logical and, or, and exclusive or functions have been available in earlier computers.”). This had the advantage that the operands didn’t have to be word-aligned, or even byte-aligned, but could instead be at arbitrary bit offsets. “Byte-aligned” wouldn’t have made any sense anyway — “The number of bits used to encode individual characters may be varied. Thus a decimal digit may be compactly represented by a binary code of 4 bits, or it may be expanded to 6 or more bits when intermixed with alphabetic information.”
It still had only a single accumulator (and 15 index registers) but the accumulator was two words, which I think means 128 bits! This despite extensive handwringing about how hard they worked to wring the most out of their slow memory (this was the reason for the pipelining, which prefetched).
The index registers were 64 bits wide even though memory addresses were only 24 bits (“A complete word-and-bit address forms a 24-bit number.”). So they used the extra bits to count loop iterations and specify a “refill address” for reloading the register, and there’s a “progressive indexing” mode “in which the index quantities may be advanced each time they are used”.
Its indirect addressing mode was wild: “The term indirect address refers to an address that gives the location of another address. An indirect address may select an immediate address, a direct address, or yet another indirect address. Indirect addresses are obtained in the 7030 by the instruction LOAD VALUE EFFECTIVE, which places the effective address found at the specified memory location into an index register for indexing a subsequent instruction. Multiple-level indirect addressing is obtained when LOAD VALUE EFFECTIVE finds at the selected location another instruction LOAD VALUE EFFECTIVE which causes the indirect addressing process to be repeated.”
The interrupt vector table was called “a table of fix-up instructions”.
When talking about data formats for address arithmetic, the data formats shown are actually instruction word formats, suggesting that despite the ample index registers on the machine, self-modifying code was common.
One of the strange uses of “edit”: “For example, a floating-point datum may be developed as a unit in a computation, its components then used in radix-conversion arithmetic, and the characters of the result finally used as units in editing for printing.” I think maybe “edit” means “format” (although the Fortran statement is FORMAT, not EDIT.)
Aha, there’s a definition: “A final class, editing operations, includes all operations in which data are transformed from one format to another, checked for consistency with a source format, or tested for controlling the course of the program.” So “editing” includes comparisons for control flow!