Wow, blast from the past! I was on the architecture / micro-architecture team for the 29K and 29050, and your story of writing and debugging the emulator brought up similar memories of working on the performance simulator for these chips.
Wow! Thank you! The floating-point implementation in the Am29050 is really great as it ran in parallel with the integer unit. Sorry, I need to ask, maybe you don't remember. What happens if you write into gr2-gr63? Not really needed to know, but I've been curious the last 30 years.
The FPU in the 29050 was based on the Am29325 bit-slice family FPU; an extra write port was added to the register file to allow for asynchronous FPU result writeback, which allowed the parallel operation.
gr2 and gr3 were used in the 29050 for FP condition code registers, but otherwise I'm pretty sure accessing the rest caused a trap.
Great article! I remember reading about the Toledo family's computer systems ( http://www.biyubi.com/ ) years ago, but never saw any specifics so they were always shrouded in mystery to me. It's interesting reading about how you wrote the OS and seeing it run in an emulator. It's really impressive that you were able to write such a complex system in machine code.
> The subtraction instruction has a set with reversed operands, so we have SUB, SUBU, SUBS, SUBC, SUBCU, SUBCS, but also SUBR, SUBRU, SUBRS, SUBRC, SUBRCU, SUBRCS.
Pretty much the only use of "reversed subtraction" instruction on a 3-operand machine is to implement NEG (subtract from immediate zero) and NOT (subtract from immediate -1, which Am29000 can't actually do since it zero-extends the immediates)... but NOT is usually not a very useful operation on its own anyway, especially when you have an actual ANDN instruction (most of the uses of ~ in C are of "var &= ~mask" or "expr & ~mask" variety) and a whole slew of NAND/NOR/XNOR instructions as well. Oh, and also, if you know for sure that your value is either 1 or 0, you can logically negate it by subtracting it from immediate 1.
On the other hand, on a 2-operand machine "SUBR reg, reg" is a surprisingly useful instruction.
> I made a phantom ROM that reads the disk using a special instruction coded in the emulator, and returns on unhandled services.
Really puts into perspective just how useful BIOS/UEFI utilities are, doesn't it? Just give me the device tree info (or the equivalent) and load the bloody OS image, thank you, that's all I need.
> I find fun how they implement the multiplication instruction using a trap and 32 continuous MUL instructions (bit-shifting all the way).
They did division the same way; I believe it was a done as a clever to have an interruptible MULTIPLY/DIVIDE instructions: otherwise, those 32-35 cycle beasts would need some special handling in the pipeline.
The BIOS originates from an era where most or all of the OS software was burned to a ROM on the computer. Then disks arrived and we wanted more, but we didn't want the OS to implement drivers for every hardware configuration. There wasn't enough room. Then there was enough room and it turned out BIOS functions mostly sucked.
Somewhere in my collection I have an AMD 29K-based tablet that was used in São Paulo's stock exchange trading floor. It had a huge IR LED array that, I assume, was used to transmit orders to receivers installed near the monitors. I managed to boot it from ROM, but it seemed the machine was waiting for the transmitter to send it its boot image.
Would have been an interesting project. Considering the size of the array, I'd be able to control all TVs from the building one block down.
"The AMD 29000 improved the design by allowing the windows to be of variable size, which helps utilization in the common case where fewer than eight registers are needed for a call. It also separated the registers into a global set of 64, and an additional 128 for the windows."
There was so much stuff to cover but there's a slight mention saying " had to add the LOADM and STOREM instructions as these handle the spilling of local registers to memory when the register stack is full" I can add an extra explanation there.
It was more an expression of surprise than anything else. The defining feature of the 29K, when I was introduced to it at least, was what appeared to be an enormous register set and the way of dealing with the stack better than the SPARC equivalents.
The fact you got this stuff working at all is completely mind bending.
The registers could also be partitioned into fixed sizes and protected, to allow for a banked register file that allowed for amazingly fast context switches in RTOS systems.
> The AMD Am29000 processor came out in 1987, also announced as the world's fastest processor (do you see a meme here?) and it could process one instruction every cycle, so indeed it was fast.
The AMD 29k - the fastest CPU you never heard of? No, seriously, I am quite interested in computer hardware, and Computer Architecture by Patterson/Hennessy was required reading during my studies, but I can't say I ever heard of this CPU architecture, which is "based on the seminal Berkeley RISC" (according to https://en.wikipedia.org/wiki/AMD_Am29000). And it seems to have been quite widely used, but not in an actual computer (except for this one)...
> Well that was basically an am29000 but with an x86 instruction decoder front end.
The K5 was based upon a 29K-family design, but it was actually an un-released superscalar project code-named "Jaguar", rather than just a 29K with an x86 decoder.
AMD bought the company NexGen and K6, K6-2 and K6-III were based on the NexGen design. I had been the happy owner of a K6-2, which did not have the integrated L2 cache of K6-III, but I had a friend who had bought earlier an AMD K5.
The K6 family also did not have successors, as Athlon had a very different microarchitecture, with a lot of inspiration from DEC Alpha.
The AMD K6 family of CPUs has the distinction of being one of the very few modern CPUs that are described in an entire book: "The anatomy of a high-performance microprocessor : a systems perspective" by Bruce Shriver.
I am not aware about any comparable description for any more recent CPU, because nowadays the companies have become much more secretive.
It is likely that the publication of many details about K6 was also approved because it was an architecture abandoned by AMD, as they focused on Athlon for their future (this was not yet known publicly when the book was launched, in 1998).
Yep, it wound up as the main CPU in a bunch of laser-printers, as well as Apple's 8*24GC QuickDraw accelerator card for the Macintosh, and as part of the flight-control computer on Boeing 777s, among many other embedded system designs.
Wow, blast from the past! I was on the architecture / micro-architecture team for the 29K and 29050, and your story of writing and debugging the emulator brought up similar memories of working on the performance simulator for these chips.
Very impressive work -- congratulations!
Wow! Thank you! The floating-point implementation in the Am29050 is really great as it ran in parallel with the integer unit. Sorry, I need to ask, maybe you don't remember. What happens if you write into gr2-gr63? Not really needed to know, but I've been curious the last 30 years.
The FPU in the 29050 was based on the Am29325 bit-slice family FPU; an extra write port was added to the register file to allow for asynchronous FPU result writeback, which allowed the parallel operation.
gr2 and gr3 were used in the 29050 for FP condition code registers, but otherwise I'm pretty sure accessing the rest caused a trap.
Amazing! I didn't knew about the gr2 and gr3 register. I just searched for these in the user's manual, very sparse documentation. Thank you!
Great article! I remember reading about the Toledo family's computer systems ( http://www.biyubi.com/ ) years ago, but never saw any specifics so they were always shrouded in mystery to me. It's interesting reading about how you wrote the OS and seeing it run in an emulator. It's really impressive that you were able to write such a complex system in machine code.
> The subtraction instruction has a set with reversed operands, so we have SUB, SUBU, SUBS, SUBC, SUBCU, SUBCS, but also SUBR, SUBRU, SUBRS, SUBRC, SUBRCU, SUBRCS.
Pretty much the only use of "reversed subtraction" instruction on a 3-operand machine is to implement NEG (subtract from immediate zero) and NOT (subtract from immediate -1, which Am29000 can't actually do since it zero-extends the immediates)... but NOT is usually not a very useful operation on its own anyway, especially when you have an actual ANDN instruction (most of the uses of ~ in C are of "var &= ~mask" or "expr & ~mask" variety) and a whole slew of NAND/NOR/XNOR instructions as well. Oh, and also, if you know for sure that your value is either 1 or 0, you can logically negate it by subtracting it from immediate 1.
On the other hand, on a 2-operand machine "SUBR reg, reg" is a surprisingly useful instruction.
> I made a phantom ROM that reads the disk using a special instruction coded in the emulator, and returns on unhandled services.
Really puts into perspective just how useful BIOS/UEFI utilities are, doesn't it? Just give me the device tree info (or the equivalent) and load the bloody OS image, thank you, that's all I need.
> I find fun how they implement the multiplication instruction using a trap and 32 continuous MUL instructions (bit-shifting all the way).
They did division the same way; I believe it was a done as a clever to have an interruptible MULTIPLY/DIVIDE instructions: otherwise, those 32-35 cycle beasts would need some special handling in the pipeline.
The BIOS originates from an era where most or all of the OS software was burned to a ROM on the computer. Then disks arrived and we wanted more, but we didn't want the OS to implement drivers for every hardware configuration. There wasn't enough room. Then there was enough room and it turned out BIOS functions mostly sucked.
Somewhere in my collection I have an AMD 29K-based tablet that was used in São Paulo's stock exchange trading floor. It had a huge IR LED array that, I assume, was used to transmit orders to receivers installed near the monitors. I managed to boot it from ROM, but it seemed the machine was waiting for the transmitter to send it its boot image.
Would have been an interesting project. Considering the size of the array, I'd be able to control all TVs from the building one block down.
Writing a windowing OS straight in machine code for the Am29000 is wild. The emulator debugging story is the part I would love expanded.
The article ended so big for my own measures, but I'm glad to know you want more of my wall of text. Thank you!
How is the phrase "register window" not in this article? https://en.wikipedia.org/wiki/Register_window
"The AMD 29000 improved the design by allowing the windows to be of variable size, which helps utilization in the common case where fewer than eight registers are needed for a call. It also separated the registers into a global set of 64, and an additional 128 for the windows."
Amazing stuff. (Both the 29K and the article).
There was so much stuff to cover but there's a slight mention saying " had to add the LOADM and STOREM instructions as these handle the spilling of local registers to memory when the register stack is full" I can add an extra explanation there.
It was more an expression of surprise than anything else. The defining feature of the 29K, when I was introduced to it at least, was what appeared to be an enormous register set and the way of dealing with the stack better than the SPARC equivalents.
The fact you got this stuff working at all is completely mind bending.
The registers could also be partitioned into fixed sizes and protected, to allow for a banked register file that allowed for amazingly fast context switches in RTOS systems.
Yes, I had an architecture professor that was a fan of it (and so your work, presumably!) It's the only way I'd encountered it, sadly.
That was back around 2000 at Loughborough in the UK.
Thanks Oscar - as always, a very very nice writing style - a good read!
I'm glad you enjoyed it. It is my longest article so far and it took a lot of effort. Thank you!
> The AMD Am29000 processor came out in 1987, also announced as the world's fastest processor (do you see a meme here?) and it could process one instruction every cycle, so indeed it was fast.
The AMD 29k - the fastest CPU you never heard of? No, seriously, I am quite interested in computer hardware, and Computer Architecture by Patterson/Hennessy was required reading during my studies, but I can't say I ever heard of this CPU architecture, which is "based on the seminal Berkeley RISC" (according to https://en.wikipedia.org/wiki/AMD_Am29000). And it seems to have been quite widely used, but not in an actual computer (except for this one)...
You might have heard of the x86 cpu the AMD K5 that competed against the Pentium?
Well that was basically an am29000 but with an x86 instruction decoder front end. The architecture was that good it was competitive 10years later.
Think of the am29000 as an incredibly influential research project that flowed into all of AMDs future CPUs in some way or another.
> Well that was basically an am29000 but with an x86 instruction decoder front end.
The K5 was based upon a 29K-family design, but it was actually an un-released superscalar project code-named "Jaguar", rather than just a 29K with an x86 decoder.
> Well that was basically an am29000 but with an x86 instruction decoder front end.
Sad. A 29K saddled with an ugly and kludgy ISA.
Wow, I wasn't aware of that either! I didn't have an AMD K5 PC, but (IIRC) a K6-III - did that still have the same underlying architecture?
No. AMD K5 did not have successors.
AMD bought the company NexGen and K6, K6-2 and K6-III were based on the NexGen design. I had been the happy owner of a K6-2, which did not have the integrated L2 cache of K6-III, but I had a friend who had bought earlier an AMD K5.
The K6 family also did not have successors, as Athlon had a very different microarchitecture, with a lot of inspiration from DEC Alpha.
The AMD K6 family of CPUs has the distinction of being one of the very few modern CPUs that are described in an entire book: "The anatomy of a high-performance microprocessor : a systems perspective" by Bruce Shriver.
I am not aware about any comparable description for any more recent CPU, because nowadays the companies have become much more secretive.
It is likely that the publication of many details about K6 was also approved because it was an architecture abandoned by AMD, as they focused on Athlon for their future (this was not yet known publicly when the book was launched, in 1998).
Yep, it wound up as the main CPU in a bunch of laser-printers, as well as Apple's 8*24GC QuickDraw accelerator card for the Macintosh, and as part of the flight-control computer on Boeing 777s, among many other embedded system designs.
It's the 'other' major processor that shipped with register windows along with SPARC and i960. Very popular in the embedded world.
This is because history gets written by the victors.
F 15 Strike Eagle, Infiltrator and 688 Attack Sub were some of my first games.
And then came Leisure Suit Larry and the tweenage years.
Don't leave out F4U Secret Sortie: https://www.oldgame.cz/en/f4u-secret-sortie
The breadboard computer from 1997 deserves it's own post.
Very interesting article, I wasn't aware of the Am29000.
You sir, are a rockstar!
OP just casually dropped awesomeness on a Thursday.