Monday, April 13, 2020

Clocking a 6502 to 15GHz (!)

6-5-0-who?

The 6502 is an iconic processor that dominated home computing in the late 70s and early to mid 80s. It was used in machines ranging from the Apple II to the Atari 2600 to the Commodore 64 to the Nintendo Entertainment System. In pop culture, it powered Bender, not to mention the Terminator!



It often clocked at a modest 1MHz, with faster variants available. In the UK, a 2MHz variant powered the legendary BBC Micro, with a 3MHz 6502 co-processor box available to bolt on the side if you so desired.



Introducing beebjit

It was not the intention of this side project to produce a working, accurate emulator, but one appears to have popped out. The core of beebjit -- and original side project -- is a just-in-time style translation of 6502 code blocks to quasi-equivalent Intel x64 code blocks. Surrounding that core is emulation for the various supporting chips of the BBC Micro, such as the Hitachi 6845 CRTC for display; the 6522 VIA for timing and I/O; the Intel 8271 floppy disc controller; the Texas Instruments SN76489 for sound; etc.

beebjit is an open-source project on GitHub.

In tests, the translated x64 code has been able to demonstrate up to 15GHz of effective 6502 speed, for purely computational tasks such as BBC BASIC programs. For games, things can be a lot slower due to heavy interactions with the graphics, keyboard and timing hardware. To run games accurately (or in some cases at all) requires cycle perfect synchronization of different virtual hardware chips. Despite these challenges, speeds in the GHz range can still be expected.

What does a mutli-GHz BBC Micro system look like? It’s a lot of fun, perhaps best illustrated with a video. Here, we see the classic space shooter Galaforce. You can play it online here, in a JavaScript BBC Micro emulator called jsbeeb. In the video, we see Galaforce loading up from a disc simulated at real-time -- don’t forget to watch out for the track-to-track floppy disc seeks, visible as stutters in the loading of the title screen image! Then, after we get a feel for the speed at which the demo screens are running and cycling, we switch up into high gear. Will you be able to tell at which moment that occurs? (Hint: yes.)



Translating 6502 to x64

When starting out, I had no idea what sort of speed to expect. A naive initial guess would be “fast” because the 6502 instruction set is simple. There are only 256 available combinations of opcode type and addressing mode and only 4 or so 8-bit registers. This is a tiny subset of what the x64 can do in its instruction set, so you’d expect a trivial, fast mapping.

As is often the case, things are rarely so simple. A bit more thinking reveals a ton of factors why it might be fast and a ton why it might be slow:

Things that would tend to make 6502 -> x64 faster

  • In the “captain obvious” category, the 6502 in the BBC runs at 2MHz where a modern Intel chip is clocked in the GHz range.
  • The simplest 6502 instructions take 2 clock cycles. One example would be LDA #10, which sets register A to 10. The Intel processor can dispatch such a triviality in 1 clock cycle.
  • Depending on the level of parallelism inherent in the code, modern Intel processors can execute multiple instructions per clock cycle.
  • If you want to get into it, trivial optimizations are possible. For example, the 6502 can only shift its A register by 1 bit at a time with its LSR and ASL instructions, taking 2 clock cycles apiece. It’s common to see a few of these in a row. By contrast, modern Intel processors can shift a register by an arbitrary number of bits in just one clock cycle!
  • The x64 architecture has an abundance of registers compared to 6502, perhaps offering further opportunities to speed things up.
  • With 32KB RAM on the base model BBC model B, you could start to hope for a workload that fits into L1 caches.


Things that would tend to make 6502 -> x64 slower

  • Some 6502 instructions actually don’t map cleanly to a single x64 equivalent. There are some less common ones such as PHP (push processor flags), or SED (set decimal flag, eww!). A couple of extremely common ones, especially in critical loops, are LDA (),Y and STA (),Y. These read / write memory indirectly. The final read / write address is calculated based on issuing two preceding 8-bit reads to make a 16-bit pointer and then adding the Y register. Long story short, the “1 cycle per instruction” nirvana alluded to above certainly does not apply here. Even servicing reads from L1 cache, modern Intel processors take ~4 cycles.
  • The 6502 is from an era where performance was quite deterministic. Memory ran at the same speed (or in some cases faster!) then the processor, so caches and pipelines and reordering and pain had not yet been inflicted upon us. So, the 5-cycle 6502 instruction LDA (),Y that reads 2 bytes of instructions and 3 other bytes always takes 5 cycles. You won’t get icache misses, dcache misses, DLTB misses, ITLB misses, stalls for a million other reasons, etc. You get the point.
  • Self-modifying code. This was not only supported, but actually very common, particularly in the most performance critical code. I have to suspect that coders were enjoying themselves. Anyway, this poses huge problems. As we execute any 6502 instruction, we have to consider if it was blown away from under us. If we try to optimize anything, we have to beware of crazy stuff like an optimized-away instruction getting self-modify-nuked.
  • Synchronization and interrupts. At every 6502 instruction, we have to consider whether an interrupt condition exists that would cause us to need to execute the IRQ sequence instead. Related, there may be other hardware chips requiring cycle perfect synchronization with the 6502.
  • Annoying 6502 variable cycle counts. Some 6502 instructions that would otherwise translate very simply to x64 have a quirk where they might take a cycle longer depending on the X or Y register value. LDA ,Y and LDA (),Y are both affected. In a cycle perfect emulation, this needs accounting for.
  • Paging and hardware register access. Certain memory regions may not be stable. A lot of 6502 systems circumvented the 64KB address space restriction by allowing parts of it to be paged. Additional parts of the address space are typically given over to hardware registers. Hitting these under emulation is guaranteed to be painfully slow as some hardware chip or another will need to be emulated. ROM memory is another headache. Games can and do write to ROM memory, which must be detected and ignored!


We’ve already revealed that it’s possible to attain GHz speeds. Re-reading the above lists, it would seem that the “slows” outweigh the “fasts” so it’s worth a look at some tricks pulled to do it!

Tricks for fast 6502 -> x64 translation

Trick #1: make good use of x64’s 16 registers

To get us going, we can take a simple 6502 sample and x64 it!

   0x0000000020100031: mov    $0xa,%eax LDA #10
   0x0000000020100036: test   %al,%al
   0x0000000020100038: mov    %al,%bl TAX
   0x000000002010003a: test   %bl,%bl
   0x000000002010003c: mov    %al,0x40(%rbp) STA $C0
   0x000000002010003f: jmpq   0x20400000 JMP $4000

In the above (unoptimized) x64, we see that we’ve (permanently!) assigned the al register to the 6502 A register and bl to the X register. Note how the 6502 LDA instruction sets zero and negative flags but the x64 mov instruction does not, so the x64 code needs an extra test instruction -- another potential source of slowdown but see below. Some 6502 flags are actually stored in the x64 flags register. The rbp register is used to point to the 6502 memory (actually at offset 0x80 so that every zero page address can be accessed with a 1-byte displacement for a much shorter x64 instruction; the JIT code is performance sensitive to code size). Finally, note how the jmpq is jumping to a specific code location. By using a simple fixed mapping of 6502 code addresses to x64 code addresses, we can keep the generated code small and fast.

Trick #2: use the host’s memory management for ROM and paging

As mentioned above, a lot of software seems to like to write to ROM. In the BBC, the region $C000 - $FFFF is always the OS ROM (with some hardware registers mapped towards the end). The region $8000 - $BFFF is usually ROM, with the paged-in ROM selectable at runtime. (For example, BASIC might usually be paged in but the Disc Filing System ROM would be paged in when doing I/O to disc.) However, a lot of BBCs were upgraded to have some extra RAM that could be paged into this region.
Obviously, writes to ROM must be ignored. But we really do not want to be adding a test and branch instructions to every JIT memory write! Memory management and paging in the host comes to the rescue. JIT reads and JIT writes are aimed at two different 64KB mapping regions, representing views of the emulated BBC address space. For the JIT reads, ROMs are present as expected. For JIT writes, ROM regions are replaced with “dummy” host pages that are writable but write to a scratch area of memory. (An earlier version of beebjit used fault + fixup for ROM writes instead but this was too slow for Galaforce, which just loves to write to ROM.)
This all works nicely because the ROM boundaries are at 4KB alignments, which joyously happens to match the host granularity! There’s a minor wart caused by Windows having worse alignment granularity than other operating systems, but creative straddling of boundaries saves the day.

Trick #3: do simple self-modify invalidation on write

One of the most important decisions is how to handle self-modifying code. We don’t want to add in code to perform a check before executing every 6502 opcode, so the alternative is to consider doing something on writes. This is what we do and here is some code:

   0x000000002010003c: mov    %al,0x3001040(%rbp) STA $10C0
   0x0000000020100042: mov    0x4370(%rdi),%r8d
   0x0000000020100049: movw   $0x17ff,(%r8)

As can be seen, after the actual write to $10C0 is dispatched via the rbp register, there is a follow-up read and write. rdi permanently points to a context object that contains various important things.
In this instance, the JIT code is looking up a compiler maintained map of 6502 addresses to host code. The final x64 instruction writes to host code for the invalidation. But what is this magic constant 0x17ff? It is the 2 byte sequence for the x64 instruction callq *(%rdi). This ensures that if a 6502 address is invalidated, the x64 code representing it will cleanly call out of JIT and into some fix up and recompile routine.
It’s tempting to try and avoid the extra read and write: what if we can look up whether a given write address contains code or not? Whereas these sorts of optimization may be possible, adding the check adds a branch (extra code; bloats branch caches; very slow if not a predictable branch) and will almost certainly be slower than just unconditionally doing the write. It’s worth noting that nothing depends on the extra read / write and that’s exactly the sort of case where the single core parallelism of modern Intel processors can give a boost.

Trick #4: monitor excessive recompiles

If you consider the above scheme for handling self-modification, you’ll immediately notice a serious performance issue: every self-modification appears to incur a huge cost! Firstly, although Intel processors are documented as supporting self-modifying code, it is not fast. You’ll invalidate all sorts of caches and pipelines. Secondly, calling out of JIT to some C routine involves all sorts of register and calling convention fixups. Thirdly, the act of compiling is slow relative to the act of executing.
Fortunately, this can all be cleared up with a compiler that generates code differently based on monitoring what is going on. If a given 6502 opcode is being repeatedly recompiled, we can maintain some metadata about the exact situation and deal with it. To give a concrete example, here’s a couple of 6502 instructions from the Galaforce sprite routine at $B00:

   0B00: STA $0B4F
   …
   0B4E: LDA $2FA0,X

This sort of pattern covers many performance critical self-modifying situations and the compiler can make it go very fast. Firstly, the hard-coded $2FA0 value is replaced by a fetch from $0B4F and $0B50 so that the latest in-memory contents are used as the operand. Secondly, the write invalidations address table is set up so that 6502 writes to $0B4F and $0B50 no longer invalidate any x64 code because it no longer needs invalidation. And then execution proceeds, almost at 100% original speed but supporting self-modified operands!
beebjit doesn’t yet handle all self-modification situations in a recompile-free manner, but there is no fundamental reason it can’t / won’t in the future.

Trick #5: compile to blocks and check timing once per block

Looking at the JIT code for the above Galaforce sprite routine at $0B00, it starts off like this:

   0x00000000200b0000: lea    -0x33(%r15),%r15
   0x00000000200b0004: bt     $0x3f,%r15
   0x00000000200b0009: jb     0x3100b000
   0x00000000200b000f: mov    %al,0x3000acf(%rbp) STA $0B4F
   0x00000000200b0015: mov    0x2dac(%rdi),%r8d
   0x00000000200b001c: movw   $0x17ff,(%r8)

What are those three instructions at the start? The register r15 permanently stores a variable called “countdown”. Here it is being decremented by 51 (0x33) and checked to see if it goes negative, with a branch if so.
The whole timing model revolves around this countdown. As the 6502 code interacts with hardware and sets up timer IRQs, vsync IRQs, timers tick down towards 0 when some event must occur (raise IRQ perhaps). There’s also a timer ticking that fires every now and again to check the status of the physical keyboard for input. At any time, r15 tracks how soon the soonest timer is due to fire. Since we’re running a 2MHz processor, chances are that it could be thousands of processor ticks between interesting events.
In the case of the above block, it contains 51 6502 cycles worth of 6502 instructions. In the very common case, nothing interesting is due to happen in these 51 cycles, so the r15 decrement will not go negative and the branch will not be taken. This is great news for the performance of the block! Since nothing interesting is happening, we know that nothing will be raising any IRQ and the whole block can proceed without having to check for an interrupt condition every instruction. We also gain freedom to implement many optimizations without fear of some interrupt landing in the middle of a couple of coalesced 6502 instructions, for example. It’s also worth noting that nothing else in the block depends upon r15 so the Intel processor can do the check in parallel with other work.

Trick #6: optimize zero page writes until you can’t

This one is simple but highly effective. Zero page is the memory region $0000 - $00FF on the 6502. Reads / writes here are faster than accessing addresses >= $0100. So, a lot of 6502 code puts fast path variables in the zero page. We should make zero page handling as fast as possible. By default, the assumption is that no 6502 code is executed in the zero page, enabling zero page writes to go faster by not worrying about if they might be modifying code. In the unusual case that code is ever compiled in the zero page, it’s simple enough to handle it by invalidating all JIT code and regenerating on-demand with appropriate invalidation for zero page writes. (For an example of a game that triggers this, there’s Wizadore. It could be making the clever observation that it’s slightly faster to modify code in the zero page because zero page writes are faster.)
Note that very similar approaches apply to the stack page, which is $0100 - $01FF on the 6502. These are not yet implemented in beebjit.

Trick #7: use slow path faults to turbocharge some fast paths

Let’s look at a couple of 6502 reads from the same address -- a hardware register at $FE00 -- but with different addressing modes.

   0x0000000020100031: [1] lea    0xc0(%rbx),%edx LDA ($C0,X)
   0x0000000020100037: [2] movzbl %dl,%edx
   0x000000002010003a: [3] mov    0x12017f01(%rdx),%r9b
   0x0000000020100041:     mov    %rdx,%r8
   0x0000000020100044: [4] mov    -0x7f(%rdx,%rbp,1),%dh
   0x0000000020100048:     mov    -0x80(%r8,%rbp,1),%dl
   0x000000002010004d: [5] movzbl -0x80(%rdx,%rbp,1),%eax
   0x0000000020100052: [6] test   %al,%al
   0x0000000020100054:     mov    $0x12009010,%r10d LDA $FE00
   0x000000002010005a:     jmpq   0x42d613

There is plenty going on here. Unfortunately, the LDA (,X) addressing mode is not a trivial translation to x64 and implementing it correctly, as beebjit tries to do, incurs cost to handle the corner cases. It is worth a brief digression to break down the different parts corresponding to the numbers in blue above:

  1. Calculate 6502 X + the constant $C0.
  2. Corner case alert! Truncate the calculation result to 8-bits, as a 6502 does.
  3. Corner case alert! Arrange for a host access violation / SIGSEGV if the truncated result is exactly $FF. This is because the upcoming 16-bit pointer read would cross a page boundary and should wrap.
  4. Load a 16-bit 6502 pointer. NOTE! This is done in two 8-bit x64 reads. If done instead as a single 16-bit read, which you assume could be faster, you likely assume wrong as I did initially. You run the risk of colliding with store-to-load forwarding because of mixing 16-bit reads with 8-byte writes at the same memory address.
  5. Load the actual result into 6502 A from the 16-bit 6502 pointer we constructed.
  6. 6502 LDA updates zero and negative flags, so do the same in x64.

By contrast, look at the compiled code for LDA $FE00. The JIT engine has essentially given up on compiling this and is throwing it over to the slower but very capable interpreter. For this 6502 addressing mode, the address of the read is statically known at compile time, and it is also known to be a hardware register. So we jump to the interpreter to handle this complicated case.
However, for the indirect read of the LDA (,X) addressing mode, we don’t know what memory address will be eventually read so we don’t know if a hardware register will be hit or not. But wait -- there isn’t a check in the compiled code on the final address before it is read. So how does the hardware register work at all?
The answer is fault + fixup! The observation here is that it’s actually very uncommon for an indirect read to hit a hardware register, so we don’t check for that in the common fast path. However, the host virtual address space for the read is a special “indirect mapping” version of the main 6502. This mapping will cause a fault if the hardware register space is touched. The fault is slow and needs fixing up but it almost never happens so it’s a win.
You can also see the fault + fixup concept used to trap the exceptionally unusual case where the LDA (,X) instruction can fetch a 16-bit pointer from $FF, which wraps back to $00. We don’t want to pollute every (,X) mode instruction with code to compare against $FF and possibly branch so we can replace that with a single read which will fault approximately never. (The game Camelot does manage to somehow hit this.)

Trick #8: optimize!

The goal here is not to write a modern optimizing compiler, but to look for and address a couple of common patterns:

  • Common 6502 sequences that compile to x64 sequences with opportunities for coalescing at the x64 level.
  • Common 6502 sequences that can be expressed differently to take advantage of known details in the sequence.

In both cases, the word “sequence” is key. Whereas a 6502 interpreter has to consider each instruction in isolation, a 6502 JIT compiler can look across sequences and make use of knowledge of concrete values and concrete nearby instructions. Optimizing across sequences is not as gnarly as it could be because we know interrupts won’t be coming in at arbitrary points, but self-modification is still a hazard, as is the fault + fixup usage described above. To briefly get a flavor of the sort of optimizations possible, let’s look at one last 6502 compilation:

  LDA #1
   0x0000000020100034: movb   $0x1,0x3000000(%rbp) STA $80
   0x000000002010003b: mov    $0x2,%eax LDA #2
   0x0000000020100040: or     0x1(%rbp),%al ORA $81
ASL
ASL
   0x0000000020100043: shl    $0x3,%al ASL
CLC
   0x0000000020100046: add    0x2(%rbp),%al ADC $82
   0x0000000020100049: setb   %r14b
   0x000000002010004d: seto   %r12b

As can be seen, entire 6502 instructions have been eliminated; some have been coalesced; and the annoying test instructions after the LDA instructions have been eliminated because those flags are never observable on account of subsequent flag overwrites.


While the above tricks are enough to turbocharge a wide range of BBC software, some corner cases remain where speed drops below the GHz range, sometimes precipitously. These corner cases typically relate to the interaction between the 6502 and BBC hardware, not the 6502 -> x64 JIT compilation. There are also a few significant optimizations left on the TODO list that may raise performance for broad swathes of games. These optimizations will be undertaken as time and interest dictate.


Other beebjit tricks

Aside from the boosted 6502 core, beebjit is a reasonably capable BBC Micro emulator. Currently in beta-ish status, it should be able to emulate any BBC Micro game including classics such as Exile, Elite, Revs, etc.
It is cycle accurate enough to run the horror show known as Kevin Edwards’ copy protection for the Nightshade tape.

There are some major things beebjit doesn’t do well or at all yet. To satisfy these needs, you may wish to try jsbeeb or one of the many other BBC emulators.

  • Portability. beebjit is Linux native. An early attempt at a Windows build does exist. Have a look in the beebjit builds directory.
  • Broad model support. beebjit only emulates a BBC model B with a few optional bells and whistles such as sideways RAM support. Notable, BBC Master support is not (yet?) implemented.
  • GUI. beebjit renders to a window but does not have a GUI. There’s a fair amount of configurability but you will need to apply yourself to the command line!
  • Peripherals. beebjit does not emulate peripherals unless they were very common back in the day (e.g. disc drives and tape players). Things like joysticks, mice, second processors etc. are not (yet?) emulated because a vast majority of software does not need them. Other emulators do fill this need well, though.


There are also some things beebjit does that are not so common and may interest you:

  • Speed. As covered in this post.
  • Protected disc support. beebjit has a more accurate emulation of the disc, disc drive and disc controller than you will typically find. This means it will load images of original, protected discs. (The archival status of original protected BBC discs is not satisfactory and is something I am working on with others in the community.)
  • Capture and replay. Like most emulators, beebjit executes deterministically. Only external input -- currently just the keyboard -- affects the execution path. Therefore, beebjit has a mode to record external input and replay it (deterministically). This feature combines well with speed. You can replay very quickly, effectively providing an alternative to save states but with very small files and the ability to exit a replay early for a many-in-one save state file.


15GHz

Courtesy of Matt Godbolt (@mattgodbolt), here’s beebjit hitting 15GHz on a couple of the “CLOCKSP” BBC Basic tests, and 12GHz+ overall. This is on an i9-9980XE, which has robust single-core performance: 9th gen Intel core, boost clock @ 4.5GHz. No-one has tried on a 5GHz beast yet, or any 10th gen parts.



Collage of the GHz



With thanks

Special thanks to the following communities for their help, support and encouragement:

  • bitshifters. Pushing the BBC Micro to perform feats that no-one worked out how to do “back in the day”. I particularly recommend Twisted Brain.
  • StarDot.


No comments: