I know BCD is like more intuitive datatype if you don't know binary. But I don't know why to use this encoding, its like don't makes a lot of sense since its waste representation in 4bits (when representation is bigger than 9).
Also I think x86 only supports adds and subs directly (you can convert them via FPU).
Its possible that this comes from old machines, or other architectures?
BCD arithmetic is useful for exact decimal calculations, which is often a requirement for financial applications, accountancy, etc. It also makes things like multiplying/dividing by powers of 10 easier. These days there are better alternatives.
There's a good Wikipedia article which discusses the pro and cons.
BCD is useful at the very low end of the electronics spectrum, when the value in a register is displayed by some output device. For example, say you have a calculator with a number of seven-segment displays that show a number. It is convenient if each display is controlled by separate bits.
It may seem implausible that a modern x86 processor would be used in a device with these kinds of displays, but x86 goes back a long way, and the ISA maintains a great deal of backward compatibility.
BCD exists in modern x86 CPU's since it was in the original 8086 processor, and all x86 CPU's are 8086 compatible. BCD operations in x86 were used to support business applications way back then. BCD support in the processor itself isn't really used anymore.
Note that BCD is an exact representation of decimal numbers, which floating point is not, and that implementing BCD in hardware is far simpler than implementing floating point. These sort of things mattered more back when processors had less than a million transistors that ran at a few megahertz.
BCD is space-wise wasteful, that's true, but it has the advantage of being a "fixed pitch" format, making it easy to find the nth digit in a particular number.
Another advantage is that is allows for exact arithmetic calculations on arbitrary size numbers. Also, using the "fixed pitch" characteristics mentioned, such arithmetic operations can easily be chunked into multiple threads (parallel processing).
I think BCD is useful for many things, the reasons given above. One thing that is sort of obvious that seems to have been overlooked is providing an instruction to go from binary to BCD and the opposite. This could be very useful in converting an ASCII number to binary for arithmatic.
One of the posters was wrong about numbers being stored often in ASCII, actually a lot of binary number storage is done because its more efficient. And converting ASCII to binary is a little complicated. BCD is sort of go between ASCII and binary, if there were a bsdtoint and inttobcd instructions it would make conversions as such really easy. All ASCII values must be converted to binary for arithmatic. So, BCD is actually useful in that ASCII to binary conversion.
Nowadays, it's common to store numbers in binary format, and convert them to decimal format for display purposes, but the conversion does take some time. If the primary purpose of a number is to be displayed, or to be added to a number which will be displayed, it may be more practical to perform computations in a decimal format than to perform computations in binary and convert to decimal. Many devices with numerical readouts, and many video games, stored numbers in packed BCD format, which stores two digits per byte. This is why many score counters overflow at 1,000,000 points rather than some power-of-two value. If hardware did not facilitate packed-BCD arithmetic, the alternative would not be to use binary, but to use unpacked decimal. Converting packed BCD to unpacked decimal at the moment it's displayed can easily be done a digit at a time. Converting binary to decimal, by contrast, is much slower, and requires operating on the entire quantity.
Incidentally, the 8086 instruction set is the only one I've seen with instructions for "ASCII Adjust for Division" and "ASCII Adjust for Multiplication", one of which multiplies a byte by ten and the other of which divides by ten. Curiously, the value "0A" is part of the machine instructions, and substituting a different number will cause those instructions to multiply or divide by other quantities, but the instructions are not documented as being general-purpose multiply/divide-by-constant instructions. I wonder why that feature wasn't documented, given that it could have been useful?
It's also interesting to note the variety of approaches processors used for adding or subtracting packed BCD. Many perform a binary addition but use a flag to keep track of whether a carry occurred from bit 3 to bit 4 during an addition; they may then expect code to clean up the result (e.g. PIC), supply an opcode to cleanup addition but not subtraction, supply one opcode to clean up addition and another for subtraction (e.g. x86), or use a flag to track whether the last operation was addition or subtraction and use the same opcode to clean up both (e.g. Z80). Some use separate opcodes for BCD arithmetic (e.g. 68000), and some use a flag to indicate whether add/subtract operations should use binary or BCD (e.g. 6502 derivatives). Interestingly, the original 6502 performs BCD math at the same speed as binary math, but CMOS derivatives of it require an extra cycle for BCD operations.
I'm sure the Wiki article linked to earlier goes into more detail, but I used BCD on IBM mainframe programming (in PL/I). BCD not only guaranteed that you could look at particular areas of a byte to find an individual digit - which is useful sometimes - but also allowed the hardware to apply simple rules to calculate the required precision and scale for e.g. adding or multiplying two numbers together.
As I recall, I was told that on mainframes, support for BCD was implemented in hardware and at that time, was our only option for representing floating point numbers. (We're talking 18 years go here!)
When I was in college over 30 years ago, I was told the reasons why BCD (COMP-3 in COBOL) was a good format.
None of those reasons are still relevant with modern hardware. We have fast, binary fixed point arithmetic. We no longer need to be able to convert BCD to a displayable format by adding an offset to each BCD digit. We rarely store numbers as eight bits per digit, so the fact that BCD only takes four bits per digit isn't very interesting.
BCD is a relic, and should be left in the past, where it belongs.
Modern computing has emphasized coding that captures the design logic rather than optimizing a few cpu cycles here or there. The value of the time and/or memory saved often isn't worth writing special bit-level routines.
That being said, BCD is still occasionally useful.
The one example I can think of is when you have a huge database flatfiles or other such big data that's in an ASCII format like CSV. BCD is awesome if all you're doing is looking for value between some limits. To convert all of the values as you scan all that data would greatly increase processing time.
Very few humans can size amounts expressed in hexa, so it is useful to show or at least allow to view intermediary result in decimal. Specially in the financial or accounting world.
