A Complete History Of Mainframe Computing

On 2 Jul 2009 at 16:36, Warren Wolfe wrote: Uh, you mean 12 bits, not 18. The mantissa was 48 bits. A and B registers were 18 bits (as was the A register in the PPUs). The high-order bit of an FP number was the sign of the mantissa; the exponent was biased by 2000 octal (i.e. range from 0000 to 3777). That was convenient because the 6000 had no condition codes-- conditional branches were performed based on the value of a specific register, so a sign test on the high-order bit worked for both integers and floating point numbers. 0000...000 was zero for both. The 6000 series had only add and subtract for 60 bit integers; multiply and divide on integers were performed by using the floating- point instructions. (There was later an "Integer Multiply" special feature that performed an integer multiply with the mantissas=0000 instead of 2000 (octal). It wasn't any faster, but you didn't have to pack and unpack (i.e. add and remove and exponent) the 2000 bias to the numbers). One oddity is that a double-precision operation involved issuing separate instructions to obtain the upper and lower parts of a result. So, to multiply two integers, you used the double-precision multiply and obtained the lower 48 bits of a 96 bit product. Normalization of the result mantissa was not guaranteed unless both operands were normalized. But in COBOL, a lot of fancy bit-twiddling (courtesy mostly of Rick James) allowed for direct addition of decimal numbers (10 digits at a pass), zero or space filling and quick conversion from decimal to binary with no branching. It was possible to write a memory-to- memory move that kept the both store units busy. Coding for that machine was fun--the scheduling rules for instructions were very well documented. --Chuck