CALL saves A, B and Data Pointers 0 to 2 and nothing else, which is exactly
why Q and DP3 are how a subroutine hands something back. An interrupt saves
all of it, so a service with an answer had to reach into its own frame and
un-save two fields by hand:
MVSD.2
DPUP.2 0d02 ; the saved Q, by an offset it had to know
STA.2
RETI
Thirty places in CosmOS did that. Every one knew the frame's layout by
heart, and all thirty would have gone quietly wrong the day the frame
gained a field - the same duplicated fact this project keeps being bitten
by, except duplicated into thirty places AND into the CPU.
SRET is 0x76, in the seat the block split left for it. It is RETI's frame
with RET's rule applied: A, B and DP0 to DP2 come back, the saved Q and DP3
are dropped, and the Interrupt Flag is restored from the frame - only that
bit, so carry survives a service the way it survives a call, and there is
one rule rather than two. RETI stays exactly as it was: a hardware handler
has nothing to say and must leave no trace.
CosmOS is 10,969 bytes against 11,122, and no handler knows a frame offset.
TWO MISTAKES WORTH RECORDING, both mine, both caught by tests.
The first conversion matched STA.2 with a regular expression that did not
allow a trailing comment, so it ran past the end of one handler and into
the next. The second understood the pattern and still got it wrong: the old
frame write carried the answer from A into the saved Q slot, so simply
deleting the write left Q holding whatever it happened to hold. Services
that answer by calling something were fine - Q already had it - and
services that set A directly silently reported success for every failure.
cosmosCwd is what noticed, by saying "cannot go there" about a directory
that was there. Sixteen handlers move the answer into Q now.
Seven MVQA went with it. They copied Q into A so the frame write could
carry it; SRET puts A back, so they moved a value nobody would ever read.
Three blocks move and nothing else changes. Branches take 0x60, subroutines
take 0x70, and the ALU moves up into the 0x10 block the two of them used to
share. Order within each block is preserved exactly - this relocates them,
it does not rethink them.
WHAT IT BUYS IS AN EMPTY 0x00 TO 0x0F. Program Memory that was never
written, or a load that stopped part way and left zeroes in its tail, used
to read as a long run of ADDs: the machine carried on through them, arrived
somewhere unpredictable, and whatever broke there was a long way from the
byte that caused it. Now it faults where it is met:
Fault: 0x00 at Program Address 0x0004 is not an instruction.
That is the address of the byte after the last real instruction, which is
the difference between a diagnosis and a search. Reserving the whole nibble
rather than just 0x00 means a run into blank memory faults wherever it
starts rather than only when it lands on the right byte. runOffTest records
it, and the block is left empty for whatever turns out to want it.
The other half is room: branches and subroutines had filled 0x10 to 0x1F
between them, so a service return that keeps Q and DP3 had nowhere to sit
next to its family. It has 0x76 waiting now.
Five places wrote an opcode down that the scripted remap did not reach, and
four of them were found by tests rather than by looking:
- secondPass.c lists which opcodes take an address, and firstPass.c knows
SWI by number. Missing those made XOR read as a branch.
- Asm.asm knows SWI by number too, being the other assembler. Missing it
made the native and host assemblers disagree byte for byte, which is
exactly the check that exists to catch a thing known in two places.
- loaderTest.asm carries a hand written payload, and its RETI was 0x19. To
the assembler those are numbers and to the program they are data, so
nothing but running it could notice. It says so in a comment now.
- The Assembler Manual prints the bytes hello.asm assembles to, and two of
them were branches.
The monitor's recorded disassembly moved by exactly the bytes it should:
18 became 72 wherever SWI appears, with SETD and INIB untouched and every
disassembled line still reading the same.
./SplitBit --disk system.img
stage two
CosmOS
>
No boot image named. The emulator shadows its built in stage one into
Program Memory - boot vector included - and the CPU then does exactly what
it has always done: reads the boot vector and starts where it points.
NOTHING ABOUT THE CPU CHANGED to make a machine that starts itself, which
is what picking shadowing over a mapped ROM bought.
The ROM is generated from Programs/Boot/stage1.asm by the makefile rather
than committed beside it, because a copy of a program kept next to the
program is a copy that goes stale. That makes the assembler a real
dependency of the emulator, which it always sort of was and now says so.
od and awk rather than xxd, which is not everywhere, or python, which the
README does not ask anybody to install in order to build this.
loadROM is loadFile given bytes instead of a path: both go through one
reader over an fmemopen stream, because a ROM is a boot image and there is
no reason for the machine to have two ways of understanding one.
Naming an image still works and is what every other test here does. That
path is not a shortcut to apologise for - placing memory from outside is a
real thing real machines allow, and it is a debugger. The help says so now.
No image and no disk is the one case with nothing to run, and it says that
rather than printing a usage message about a missing file.
run.sh gained a "rom" mode which hands the emulator a disk and nothing
else. The source column still names stage1.asm, because that is what is IN
the ROM: assembling it there says the thing the emulator carries is a thing
that still assembles.
HALT is terminal - stepCPU returns at once when the Halt Flag is up, so a
halted machine does not execute, service devices, or take an interrupt -
and that has to stay true, because every test ends with a halt and "halted"
is how a program says it has finished. The consequence was that SplitBit
had no way to wait at all. Every wait was a spin, and a spin is bus
traffic: 11.5% of Type over a 14K file on a disk of ten thousand cycles,
after read-ahead had already hidden three quarters of the latency.
WAIT is 0xFE, one byte, no operands, sitting under HALT where the
instruction that almost stops the machine belongs. Three decisions in it:
- A line already standing means there is nothing to wait for, so WAIT does
nothing. That is what makes test-then-wait race-free.
- Any line ends the wait, masked or not, so a program can sleep on a device
it has no handler for and read its status afterwards. Masking says who
answers a request, not whether it happened.
- A line that wakes the CPU without being dispatched is taken down by the
WAIT. Left standing it would be found by the next WAIT, which would
return at once - the program would spin exactly as before while looking
as though it slept.
Waiting is NOT a Status bit, and that is the trap avoided rather than a
gap: Status rides into the interrupt frame and comes back out, so a machine
interrupted mid-wait would return from its handler still waiting, and wait
again for what it had already been given. An internal field instead.
Idle cycles are counted apart from bus cycles and the halt line says so
when there are any, which is what makes the difference observable at all -
with the line-clearing removed the total moves by ONE cycle, 20,100 against
20,099, and only the idle half changes, halving to 9,976. A test on
totals could never have seen it. Tests/terminal.sh asks that question,
being the file for things a recorded output cannot see, and fails with the
clear removed while "both reads finished" still passes.
Three collisions, all found by building it:
- 0xFE was the assembler's "not an instruction" sentinel. getOpcode now
answers a negative NOT_AN_OPCODE, which is outside the range of every
possible answer instead of inside the unused part of it.
- 0xFE was also what faultTest and faultResumeTest executed to provoke a
fault. They now use 0xFD and say why, because they did not fail when it
became an instruction - they HUNG, having started sleeping instead.
- Keys.asm has had a label called "wait" for a year, and mnemonics are
matched uppercased. What that reported was "Branch without label" at the
BRQ thirty lines away. The assembler now refuses a label that is already
an instruction, at the label, by name; every instruction added takes a
word out of the space of label names, so this will happen again.
The disk's status has always had a bit meaning "still going", and the header
beside it has always said to honour it. Nothing did, because nothing could: the
host finished the transfer inside the instruction that asked for it, so the bit
could never be seen up and asking about it was asking about something that
cannot happen.
--disk-cycles gives it a latency. The command is still checked at once, because
a refusal is not work - a block that is not there fails before any head moves -
but the transfer is remembered and done when the machine has run that far. Until
then the buffer holds the block BEFORE this one.
That last part is the point. A program that does not wait gets the wrong bytes
rather than an error, which is the failure the bit exists to prevent and the one
that would never have shown up. With a latency of two thousand, CosmOS could not
even mount: sbfsMount reads block zero and looks straight at the buffer.
deviceTick is the general shape rather than a disk feature. Called once per
instruction with the machine's clock, it lets anything whose moment has come
finish - which is what a display that refreshes, or a port that waits on the
host, would want in exactly the same way.
The filesystem watches the bit now, in one small routine reached with RCAL. That
is not decoration: what it hands back is the settled status in A, and CALL puts A
back the way it found it, so an ordinary call cannot carry the one thing this
exists to carry. Two bytes of Stack rather than ten, in a routine that runs on
every block the machine ever touches - the first place in the system where the
new call is the right one rather than merely a cheaper one.
The manifest takes a @N after a disk, the way it already takes :ro, so a test can
ask for a slow one. cosmosSlowDisk lists a directory at two thousand cycles a
block and gets the same listing as everything else, which is the whole assertion:
a filesystem that did not wait would print nonsense rather than fail.
Zero is the default and every other test runs at it. What waiting costs, on a
directory heavy run: 229k cycles at zero, 275k at five hundred, 415k at two
thousand, 1.16M at ten thousand.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01E2JrLzFvuFX9fgi1LDRjrW
A blit cost ten cycles, which were the five port writes that set it up. The
quarter of a kilobyte that moved cost nothing, and no hardware moves a quarter
of a kilobyte for nothing.
BANKS ARE SEPARATE MEMORIES, AND THAT IS WHAT SETS THE RATE. A move between two
of them can overlap its read and its write - fetch the next byte while the last
one is stored - so it settles at a byte a cycle. A move within one bank cannot,
and costs two. A fill has nothing to read and costs one whatever the banks are.
The odd cycle on each is the pipeline filling.
That is not a modelling choice so much as a reading of the structure the machine
already has: a Program to Data blit is inherently twice the rate of a Data to
Data one, and it is legible why.
Measured: 256 bytes is 297 cycles across banks and 518 within one, both
including the instructions that ask for it.
WHAT IT TAUGHT, which was not what I expected. Charging for movement costs the
native assembler 0.4 per cent and costs directory work 13.4. The assembler reads
a block and then thinks about it for a long time, so the move is amortised into
nothing; the filesystem reads a block in order to look at it and does nothing
else in between.
So the case for a blitter that runs alongside the CPU is weaker than it sounds.
Concurrency pays when there is other work to do during the transfer, and the
place that spends its time moving memory is exactly the place with nothing else
to do - it blits a block precisely so that it can read it. What that workload
wants is a FASTER controller, not a concurrent one: a wider data path halves the
wait, and the machine is waiting either way.
Video is the case that would still want concurrency, since a frame can be moved
while the next one is worked out. That is an argument about software nobody has
written yet, and it is now an argument with numbers on the other side of it.
The byte at a time port is charged too, for the byte it moves beyond reaching
the port. Nothing polls CTRL_STATUS, so the transfer stalls whoever asked for
it, which is the conservative reading and the one the software already assumes.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01E2JrLzFvuFX9fgi1LDRjrW
cycleCount used to tick once per instruction, so RSTA cost what SETD cost and a
CALL moving ten bytes of Stack cost what a branch cost. No machine anybody could
build works that way, and the emulator's job is to be the thing the hardware is
designed against.
Every touch of memory now goes through one of four accessors that charge for it:
fetching an opcode, fetching the bytes after it, reading or writing Data Memory,
and reaching a device port. One access, one cycle, nothing overlapped. The
accessors exist so the cost is counted where the access happens rather than in a
table of per instruction costs kept somewhere else - a table like that is a
second copy of what the code does, and the two drift.
The run loop spends a budget of cycles instead of running a count of
instructions, so the emulated rate means something: an instruction costs what it
touches, and a batch ends when the cycles are gone.
What the numbers say now: RSTA 1 and SETD 4, being one byte and four. LDA 3,
DPUA 2. CALL and RET together 24, RCAL and RRET together 8, because the first
pair moves twenty bytes of Stack and the second moves four. The average SplitBit
instruction costs 3.72 of these, measured over the native assembler assembling a
program.
And the measurement that prompted all of this: converting the filesystem's
hottest leaf routine to RCAL is 3.1 per cent cheaper on a directory heavy
workload. The old model said 0.0, which is what a model that cannot see memory
traffic must say about a change that is nothing else.
Three tests moved. settle() strips the cycle count from recorded output, so
nothing should have churned - but it was anchored to the start of a line and
replCalculator's last output has no newline on it, which leaves the halt message
mid line where the pattern never reached. Not anchored any more.
The two Life programs are bounded by a cycle count because they never end, and
that number was rescaled from 3,000,000 to 11,200,000 - the same amount of work
at 3.72 cycles to the instruction. Nothing about either program changed. No limit
reproduces the old output exactly, because the cut now lands elsewhere in a
frame, so they are recorded again rather than tuned to match.
Whether hardware overlaps a fetch with the end of the previous instruction is
left open on purpose. This is the conservative model; pipelining is a decision to
make while drawing the hardware, not one to inherit from an emulator.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01E2JrLzFvuFX9fgi1LDRjrW
The six settled back on the twenty fourth, built now.
RCAL and RRET are a call that puts nothing back. CALL restores A, B and Data
Pointers 0 through 2, which costs ten bytes of Stack and is why a subroutine
here can only hand anything back through Q, DP3 or memory. RCAL costs two and
restores nothing, which is what a short leaf routine wants and is unsafe in
exactly the way the name says.
They are a pair because the frames are different sizes: returning from one
through the other walks the Stack to somewhere that was never a return address.
That was the user's correction to the original proposal, which had a raw call
and no raw return.
DPUA and DPDA offset a Data Pointer by A; DPUW and DPDW by A and B together,
most significant first. DPUP and DPDN take a byte written into the program, so
moving a pointer by something just worked out meant storing it and loading it
back. Down as well as up on symmetry grounds, which was also the user's call -
the argument against it came from counting uses in a corpus written under the
constraint.
The opcodes sit where they belong: 0x16 and 0x1E immediately below CALL and RET,
and 0x4E through 0x51 at the end of the Data Pointer family. All six fit shapes
that already existed, so instructiontable.py needed only set membership and both
machine side copies of the table regenerated from it unchanged.
Checked at every level it exists at: the emulator runs them, the host assembler
encodes them, the monitor disassembles all six with the right lengths, and the
assembler that runs on the machine builds a program using them byte for byte
identically to the host - and that program runs.
The recorded test measures what the two calls COST as well as what they put
back, because an RCAL that quietly did what CALL does would still return to the
right place. It does not survive that: returned through RRET, it hangs.
docs.sh can read a two word number now. The count of instructions taking a Data
Pointer went past twenty, and the pattern only allowed one word, so the check
would have reported that the manual had stopped saying it rather than that the
number was wrong.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01E2JrLzFvuFX9fgi1LDRjrW
"Binary" was doing three jobs. It meant an SPBT file that the machine starts
from; it meant whatever the assembler happened to produce, which is now
either that or a loadable program; and it meant a compiled host tool. A word
that means three things means none of them, and the first of the three has a
name already - this project has been calling them boot images for a while
and the manuals had not caught up.
Where it means an SPBT file -> boot image
Where it means either output -> output
Where it means a host executable -> left alone
Where it means base two -> left alone
The user facing messages move with it:
Error: No boot image specified.
Usage: ./SplitBit [OPTIONS] <boot image>
Error: This is not a SplitBit boot image.
Error: This boot image is in format version 2, and this emulator reads 1.
Successfully wrote SplitBit boot image to "hello.bin".
The assembler's own help was the interesting case. Its -o writes either
format, so "the binary" there was never right - it is "the output" now, and
the message that names the format is the one that says which it wrote.
No recorded output contained the word, so nothing needed re-blessing.
Checked before starting rather than after.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01E2JrLzFvuFX9fgi1LDRjrW
Overflowing/Underflowing A or B with increment or decrement instructions now sets the Carry Flag.
An increment or decrement of A or B that doesn't result in an overflow/underflow now clears the Carry Flag.
Q is no longer saved and restored in subroutine calls, allowing it to be used to return a one byte result.
Documentation updated with more detail.
Added DPUP and DPDN, which take a one byte operand specifying how far up or down to offset the Data Pointer.
Three Fibonacci generators using the print.asm library.
- 8 bit values printing in decimal representation.
- 16 bit values printing in hexadecimal representation.
- 32 bit values printing in hexadecimal representation.
Anachronaut
Anachronaut