5d9b39514b0daf3c16b1178b82e18ba6fce3e8e3
333
Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
  AnachronautandClaude Opus 5
|
83623a3df3 |
Give the Voyager a screen
A tile engine on ports 0x30 to 0x3F, bringing one bank of video memory registered the way the disk's buffer is. The CPU writes cell indices and the device turns them into pixels, which is the whole reason a screen is affordable at a megahertz: a frame is 16,667 cycles, a full 320 by 200 picture is 64,000 bytes, and a 40 by 25 map is 2,000. A program that changes two cells writes four bytes. The cost of a screen becomes the number of cells that changed rather than the number of pixels on it. Which makes colour depth free, so the tiles are eight bits: an 8 by 8 cell is 64 pixels and each picks independently out of 256 colours, with no per-cell limit of the kind that made a Spectrum two and C64 multicolour four. The low nibble of a cell's attribute is ADDED to every index in its tile, sixteen at a time, so a tile drawn in 0 to 15 appears in any of sixteen schemes without a second copy in tile memory - and a tile wanting all 256 leaves the nibble at zero and gets them. Neither use costs the other anything. Two decisions are arithmetic rather than taste, and both come from the machine having no multiply. A map row is a page whether the mode fills it or not, so a cell address is the row number as the high byte and the doubled column as the low byte with no arithmetic at all; otherwise every cursor move on a 40 column screen would cost a row-times-40 in software. And a palette entry is four bytes rather than three, so entry n is at n times four, a shift. THE MAP IS A RING and the Scroll register says which of its 128 rows is on top. Scrolling moves a register and no memory: blitting a 40 by 25 screen up one line is 1,920 bytes inside one bank, which is twelve percent of a frame even with the controller widened, and a program printing one page would spend six frames shuffling memory. It is now one port write - and the rows that scrolled off are still there, which is where a terminal gets scrollback it never had. The device is part of the machine rather than part of the window. It renders into a buffer that is a pure function of video memory, so the same program draws the same picture with nobody watching; Voyager puts that buffer on the glass and decides nothing. Both binaries take --screen, which saves a PPM when the machine stops, and that is what makes a screen checkable on a host with no display at all. Tests/video.sh checks fourteen named behaviours rather than comparing a recorded image, because a recorded image would say "something changed" and leave which of the palette, the tile, the attribute, the map or the scroll register broke to be found by hand. Verified by breaking three things in turn: the additive nibble failed exactly one check, the scroll origin exactly two, and moving every cell one pixel sideways exactly the four about placement. Tests/docs.sh could not count past nine, which is how a suite of ten scripts reported itself as wrong for the wrong reason. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01E2JrLzFvuFX9fgi1LDRjrW |
||
|
AnachronautandClaude Opus 5
|
e3ef25e3b3 |
Split the machine from its front end, and add Voyager
The Segan Voyager is the same SplitBit with a screen and a speaker instead of a terminal, and this is the rung that makes there be two of them at all. Everything that is actually the machine - the CPU, the controller, the devices, the run loop, the reporting - moves to machine.c, and each front end brings one file of its own. emulator.c is now sixty lines of argument handling and a three line loop. The machine runs in SLICES rather than to completion, because that is the cut a window needs: run a slice, present a frame, run another. A terminal runs slices until the machine stops. Both loops are three lines, which is why the cut is there rather than anywhere else. At this stage Voyager's window is empty. There is no video device yet and inventing a temporary way to draw would mean building something to throw away. PLAIN MAKE STILL WORKS WITH NO GRAPHICS LIBRARY. Raylib is probed by compiling and linking against it rather than by looking for a file, because a header with no library behind it passes a file check and then fails at link time. Where it is missing, make says so once and builds everything else - the machine, the assembler, the disk tool, the linter and the whole suite. A project about a small understandable CPU should not need OpenGL to run its tests. That nearly broke here: make strict globs Source/Emulator/*.c, so it would have tried to compile voyager.c and failed on precisely the machines the split exists to support, and this machine has Raylib so nothing would have caught it. Tests/voyager.sh runs the WHOLE MANIFEST through Voyager and holds it to the recorded results SplitBit is held to. Not that the two look alike: that one satisfies every recording the other does, byte for byte, exit status included. It reuses run.sh, which now takes the machine from SPLITBIT_EMULATOR, rather than keeping a second copy of the runner that would drift. Voyager not being built is not a failure - it says so and passes. Verified both ways. Made Voyager print one extra line, and 114 of 165 failed: exactly the tests that run the emulator, with the 51 assemble-only and xfail cases correctly untouched. Removed the binary, and the script skipped. Built with HAVE_RAYLIB=no, and everything else still built and checked clean. --headless is taken out of the arguments in voyager.c rather than in the shared parser, which should not learn about a window only one binary has. It exists so the suite can run this binary at all: a front end that could only be exercised by a person looking at it would be a front end nothing checks. loadFile takes a const char * now, which it always should have. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01E2JrLzFvuFX9fgi1LDRjrW |
||
|
AnachronautandClaude Opus 5
|
4c3eac8d9c |
Widen the memory controller's path to sixteen bits
The controller now reaches bank memory two bytes at a time, so a transfer whose source, destination and length are all even moves two bytes a cycle between banks and one within a bank - twice what each was. A 256 byte block between banks falls from 257 cycles to 129. Alignment is required all three ways because a word is read at an even address and written at an even address; an odd anything would mean shifting bytes across word boundaries to line them up, which is a different design. A misaligned transfer falls back to the byte a cycle it cost before, so nothing already written got slower. THE CPU DOES NOT CHANGE. It still sees eight bits, a Data Pointer still addresses a byte, and no instruction means anything different. This is a peripheral getting faster, which is why it is worth doing now rather than after more is built on top of it. The rule is deliberately visible rather than smoothed over: aligning a buffer costs nothing and halves what moving it costs, and a cost a program cannot see is a cost it cannot avoid. Tests/cycles.sh is new, and is the test the Test Manual has always said this kind of change would need - run.sh strips the cycle count from every recorded result, so nothing else in the suite can see any of this. It pins the RATE rather than a total: each case runs twice from programs whose instructions are identical but for the byte written to the Command port, once asking for the transfer and once for GuardOff, which costs nothing beyond the port write. The difference is the transfer and nothing else. Verified by disabling the widening, which failed exactly the three aligned cases and left the five misaligned ones passing. The Programming Manual gains a section saying what a transfer costs, which it never said at all - it only promised a transfer does not wait, which is a different claim and could be read as promising it is free. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01E2JrLzFvuFX9fgi1LDRjrW |
||
|
AnachronautandClaude Opus 5
|
c3c2451afe |
Say what the strict build and the sanitizers actually establish
Three claims in the Test Manual were true enough to pass and loose enough to mislead. The headline said the two assemblers produce identical bytes and the two filesystems produce identical disks, with no qualifier. That reads as universal and is not: it is evidence about the corpus assembled and the operations performed. The detailed sections already said so; the headline now says so too, because a document arguing about what evidence is worth cannot overstate its own. "Compiles under strict ISO C" omitted -D_XOPEN_SOURCE=700. The check is strict C11 with the POSIX interfaces the code uses explicitly selected - realpath, strdup, dirname and getopt - not freestanding ISO C, and the flag is part of the check rather than a hole in it. And the sanitizer section leaned on AddressSanitizer's junk fill, which is a toolchain default this build does not configure and which buys almost nothing here anyway: there are six heap allocations in the whole repository, all in the assembler, the largest a deliberate calloc, and the machine's own memories are static arrays the sanitizers neither fill nor bound-check. That last part is the overrun blind spot seen from the other side, so it now points at it. The same claim is corrected in the makefile, where it originated. Found by ChatGPT reviewing the manual. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01E2JrLzFvuFX9fgi1LDRjrW |
||
|
AnachronautandClaude Opus 5
|
8fbbeb6ec9 |
Make xfail compare the diagnostic, not just the failure
The Test Manual said an xfail test records the assembler's refusal message and so catches both an error that stops being detected and a message that changes without anybody meaning it to. It did not. run.sh checked only that the assembler exited non zero, printed the first line for a person to read, and compared nothing; --bless recorded nothing for these sixteen tests at all. So an xfail passed four different ways that look identical from outside: the intended error fired, an unrelated error fired, the message changed, or the assembler fell over on its way to the point. That is the documentation describing behaviour the code does not have, which is the exact failure Tests/docs.sh exists to prevent, in the manual that argues for knowing what your evidence is worth. The diagnostic is now stripped of colour, given the same [exit N] line every other recorded result carries, and compared through check() like anything else. Sixteen results recorded; every existing one is byte for byte unchanged. Verified the way the manual asks: one diagnostic was broken on purpose, its test failed with the changed line in the diff, and its neighbour passed. Found by ChatGPT reviewing the manual. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01E2JrLzFvuFX9fgi1LDRjrW |
||
|
AnachronautandClaude Opus 5
|
999451e935 |
Ignore External, which is other people's repositories
soundThing and Hatchet-GPU are kept in the working tree to be read, not built: one is where the synthesizer comes from, the other is where the 8x8 font and the scroll register idea come from. Neither belongs to this repository. They carry their own .git directories, their own build outputs and three compiled binaries between them, and a plain 'git add -A' would have taken all of it. What this repository takes from them will be vendored deliberately, with a note in the file saying where it came from and what changed. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01E2JrLzFvuFX9fgi1LDRjrW |
||
|
AnachronautandClaude Opus 5
|
d6bc416698 |
Write the Test Manual, and make the suite check it
The test system had grown to seven scripts making five genuinely different kinds of claim, and nothing said which was which. A recorded transcript and a byte-for-byte comparison against a second implementation both print [ok ] and are worth wildly different amounts, so the fourth manual exists to say so: what each script can and cannot answer, why every determinism rule is there, how to add a test, and - the part written nowhere else - where the suite is blind. That last section is the reason for the document. Three buffer overruns into adjacent variables were all found by a person using the machine and none by the suite, the sanitizers cannot see them because emulated Data Memory is one legitimate host array, and there is no second opinion about the CPU at all. A document listing only strengths teaches the wrong lesson. The bullets describing each script move out of the README, so docs.sh now reads the manual for them, and five more numbers in it are settled from the source rather than trusted: the shape of the manifest, the xfail count, how many fixture disks makedisks.sh builds, how large the lint baseline is, and the tool count in either document. Each of the new checks was broken on purpose and watched to report before being kept, which is the discipline the manual itself argues for. Also drops the stale "70 instructions" from instructiontable.py's docstring. There are 72, and a number that carries no meaning is better removed than corrected. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01E2JrLzFvuFX9fgi1LDRjrW |
||
|
Anachronaut
|
79727044b7 |
Reboot, and the machine device that makes it possible
Until now the only way to restart was to stop the emulator and run it again, which meant the one thing the machine could not do was the thing Once was written for. The loop now closes without leaving it: > Once /System/Boot/bare.bin next start: /System/Boot/bare.bin, once > Reboot starting again stage two just this once: /System/Boot/bare.bin bare metal: no system, just this Writing 1 to port 0x13 asks the machine to start over. A PORT RATHER THAN A SERVICE, because a reset has to work when the system does not: something only askable through SWI would be unavailable in exactly the case that wants it most, and a program that owns the whole machine has no system to ask. It is device class 0x04, in the range kept for the machine rather than among the peripherals, because it is not one - it is not attached to anything and cannot be unplugged. WHAT A RESET REPEATS IS HOW THE MACHINE STARTED. Named an image, the emulator places it again; named none, the ROM is shadowed again and reads the disk. Anything else would mean a reset changed what the machine IS, which is the one thing a reset must not do. Both are tested. Taken between instructions, because a device cannot restart the machine from inside the instruction that asked: the CPU is part way through a step and its state is not yet anything a reset could leave behind consistently. The disk stays attached and keeps everything written to it - that is what warm means. The vector table is cleared, which is the one deliberate departure from leaving memory alone: a vector points into whatever installed it, and after a reset that program is not running, so a handler left behind would aim an interrupt at an address belonging to something gone. It is the argument CosmOS already makes at exit, applied to the machine. Reboot is 45 bytes, most of them the word it prints. |
||
|
Anachronaut
|
7b28f48f52 |
Once: start something else on the next start, and only that one
A program that owns the whole machine had nowhere to run. It cannot be started from the shell, because starting it means there is no shell, and pointing boot.cfg at it means a machine that keeps starting it - which is a poor place to find a mistake in something written five minutes ago. Once writes /System/Boot/once.cfg, in the same format as boot.cfg and read with the same routines, because a second format for one setting would be a second format. The loader reads it before boot.cfg and DELETES IT BEFORE IT JUMPS, which is the only moment there is: after the jump the loader does not exist. Consumed by being read rather than by working, so a one shot that hangs cannot hang twice - the request is gone before the image ran, and the next start reads boot.cfg like any other. THE BOOT STATE IS NOT TOUCHED, and the first version got that wrong. It marked the start the way any other start is marked, and then every successful bare metal boot reported that it had never arrived - because a program with the whole machine has no filesystem to clear a mark with and is doing nothing wrong by not having one. Found by running it: the image printed its line and the next start still said the last one did not. Three disks, each a start further along, so none of the tests depends on another having run. The loop is closed on the machine now: write it in Edit, assemble it with Asm, ask for it with Once, restart, watch it own the machine, and the system comes back without being asked. |
||
|
Anachronaut
|
89c667848b |
Edit read a file into a buffer it never checked the size of
Opening hello.asm showed a thirty one line file as three, one of them cut short. Opening it again hung the machine: the emulator kept running and nothing ever answered. Entry is the buffer a line is read into, and it is followed in memory by TextHead and ArenaFree - the head of the document, and the pointer its line allocator hands out. The loop that splits a file into lines copied characters in WITH NO BOUND AT ALL, so a 94 character line wrote thirteen bytes over both of them. The list head then pointed into the middle of the text and the allocator handed out an address inside the file, which is why the second open walked a list that led back into itself for ever. Typing was always safe. osReadLine is told how much room there is, so a new document behaved perfectly and a source file did not - which is exactly how the user found it, and why it looked like a mystery rather than a bug. The bound is there now, and the buffer is 128 characters: what a line is everywhere else on this machine, the same number configuration files use, rather than a second answer to a question already answered. hello.asm fits. A file with a longer line is REFUSED rather than shortened. This is an editor - a line cut on the way in would be written back cut, and the file damaged by having been looked at. It says so and exits with a status of one, which it can do since this afternoon; the file is byte identical afterwards, and the test checks that. Opened twice in the test, because once is not enough to see it: the first open does the damage and the second is what never returns. This is the third time this shape has turned up: a buffer written past its end into the variables that happened to follow it. The prompt walked off CwdText into the shell's own command names; the assembler's output ran into its label table. Every one was found by a person using the machine. |
||
|
Anachronaut
|
87d819847e |
A program can say how it went
SWI osExit takes a status in A, and the shell keeps it. Fifty eight exits across twenty three programs now say deliberately whether they worked: 25 did what they were asked, 24 did not, 9 were asked wrongly. Compare is the exception and says so - one there means the files differ, which is a result rather than a failure, the way diff has always had it. IN A RATHER THAN Q, which is not a departure from the rule that a service answers in Q. This one takes an ARGUMENT, the way osPrintNumber takes A and B, and it never returns to answer anything. A is free precisely because a return would have put it back - and Q is the ALU's output, so a small number costs four instructions there against one in A. The shell does not print it. A program that failed has already said so in words and a number beside that is noise, so osLastStatus hands it back and Status is the program that shows it. That indirection is the point: the number exists for the thing that cannot read words. MARKING THE EXITS FOUND A DEFECT ON THE FIRST RUN. Type and More printed why they had failed and then fell through into the success exit, reporting that all was well. Nobody had noticed, because while the only reader was a person, the person could see both the complaint and the claim. Two smaller things. Snake sets the console to line mode and then exits with zero, and the linter flagged the second RSTA as redundant - an exit status and a console mode, equal by accident, which is the class that must never be collapsed. And the README still taught answering by writing into the frame, three months of habit that SRET replaced yesterday; that section is gone and the one describing SRET stands in its place. |
||
|
Anachronaut
|
b4206673a6 |
The clean install disk is one the machine can start itself from
make run-cosmos handed the emulator a boot image AND a disk, so it took the direct path every time: memory placed from outside, nothing on the disk consulted about it. The whole boot chain was built and tested and then not used by the one command that runs the system. The disk is formatted with a boot area now, stage two goes into slot zero as raw bytes, and the system is an ordinary file at /System/Boot/cosmos.bin - which is what boot.cfg would choose between if there were one. There is not: stage two falls back to that name when the file is missing, and a clean install with nothing to configure is the right default. run-cosmos names no image, so the machine starts itself. The old behaviour is run-cosmos-direct, which is worth keeping and worth naming honestly: it is what a debugger does, and it is what to use when the thing being debugged is the boot chain, since it skips the boot chain. /Source gains stage1.asm and stage2.asm, so the machine can rebuild what starts it. Everything stage two includes is already in /Lib, so Asm stage2.asm on the machine produces the bytes a boot slot takes. Stage one is the exception and always will be: it is the ROM, and the one part of this that a disk cannot replace. Second slot left empty, which is what makes replacing the first survivable. |
||
|
Anachronaut
|
f1cc2e56b2 |
The last handler that knew a frame offset
handleReadLine used DP1 for the frame and the conversion only looked for DP2, so it kept the whole dance - and the dance had become a no-op that looked like work: MVQA copied Q into A, A was written over the saved Q, and RETI restored it, which is where it started. readLine leaves the length in Q and SRET keeps Q, so the answer simply stands. No handler in CosmOS knows what an interrupt frame looks like now. The 0d02 and 0d05 offsets still in the file are structures - an SBEX header, an entry in the instruction table - and not frames. |
||
|
Anachronaut
|
61a80ae13d |
Put the manual's sections back in opcode order
The user's edit. Moving branches to 0x60 and subroutines to 0x70 left their sections sitting where they used to be numerically, between the ALU and the registers, so the manual read in an order the opcode map no longer did. The list has always been arranged by opcode sequence rather than by importance or by how often a thing is used - arbitrary, and now canon. |
||
|
Anachronaut
|
c8c9f0b363 |
SRET: a handler answers the way a subroutine does
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. |
||
|
Anachronaut
|
cd5f548736 |
Move the opcode map: nothing in 0x0X, and room for a return variant
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. |
||
|
Anachronaut
|
ce2a2cd7e6 |
Settle is a program, and a machine with no fallback still starts
The boot state opened a loop that could not be closed from inside: the machine said "settle it to try again" and gave you no way to do so. Settle closes it, in 349 bytes. A PROGRAM RATHER THAN A SHELL WORD. The shell is for the things that cannot be done without it, and this is not one - it reaches the system through SWI like anything else, which means it can be replaced, left off a disk, or called by whatever comes to call programs in turn. That last one is the point: a shell word is not callable by anything. Two services for it. osBootState answers in Q, and a machine with no disk answers settled, because there is nothing there to be unsettled about. osBootSettle puts it back. SETTLING IS THE ONLY WRITE A PROGRAM GETS - marking a start as trying or fallen back is the loader's business, and a service that let a program claim either would let it lie about something the loader has no way to check. And a hole the tests walked into, which was mine rather than theirs. With no fallback configured, a failed start left the machine unable to start at all: the mark said do not use the system, and there was nothing else to use. That turns "the last start failed" into "no start is permitted", which is worse than the problem the mark was added to solve. With nothing to fall back to it now tries the configuration again and says so - a failure that was passing recovers, and one that is not leaves the machine exactly where it would have been without any of this, which is the most that can be promised when there is only one thing to start. docs.sh caught both new services having no row in the services table before anything else did. |
||
|
Anachronaut
|
dc74149321 |
B4: the disk remembers whether the last start arrived
The loader marks the superblock before it hands over and the system clears the mark when it reaches its prompt, so a system that crashes on the way there leaves it set. The loader finding it still set next time is how a machine that will not start says so to the only thing in a position to do anything about it. Without that, pointing boot.cfg at something that dies before the shell is a machine that can never be told anything again - the shell is the only way to change the file, and the file is what stops the shell from starting. Three states rather than two, and the third is the one worth having: 0 settled the last start arrived; use the configuration 1 trying handed over, and nothing came back to say it got there 2 fell back a try failed and the fallback was used, until settled With only 0 and 1 the machine alternates for ever: fall back, reach a prompt, clear the mark, retry the broken system, crash, fall back. State 2 stops that. A system known not to start is not tried again until somebody says the situation has changed. REACHING THE PROMPT IS A DELIBERATE THRESHOLD. It is not a claim that the system works - a shell can be reached by something broken in every other way. It is the point where a person can type, which is exactly what the fallback exists to give back: anything wrong past there is fixable from the prompt and nothing wrong before it is fixable at all. The routines live in sbfs.asm because both the loader and the system read and write this byte, and two pieces of code with their own idea of where a byte lives is what this format has two implementations and a byte for byte comparison to avoid. And the trap this system documents in its own manual caught me anyway: the first version handed the state back in A, which CALL restores, so every read got whatever the caller happened to be holding. It comes back in memory now, and the comment says why. Three disks differing only in the state on them, so the tests read as three consecutive starts of one machine while none depends on another running. |
||
|
Anachronaut
|
546f336823 |
Configuration files, and boot.cfg as the first of them
One setting to a line: a key, a space, the rest of the line is the value. A semicolon starts a comment. The format was noticed rather than designed - textSplit already cuts the first word off a line and leaves the rest, and textSame already insists two strings end together, so reading a setting is those two routines and a loop. It is also what the shell reads, which makes a configuration line a command line the machine reads instead of a person typing one. The format was chosen by asking what the BOOT LOADER could manage, because it is the worst case in every direction: a few kilobytes, no operating system to report to, and if it fails the machine does not start. Two formats would be worse than one and the loader cannot have the richer one. CONFIGURATION IS ADVICE. A missing file, a missing key, an unusable value, a line too long to read: all of them mean use the default and none is a failure. BUT QUIET IS NOT SILENT - a setting somebody meant, which did not take effect, says so. That was the user's addition and it is the better rule: the default alone leaves the only symptom being that the machine did not do what somebody asked. So two routines. cfgGet reads and says nothing, because reading three settings should not report one bad line three times. cfgCheck reads the file once and reports, and is handed the caller's list of keys - whether a key means anything is the only part of this a shared reader cannot judge. /System/Boot/ holds the boot files, and stage two reads boot.cfg for what to start, with a fallback to try if it does not work and a name compiled in for when the file says nothing. THE TEST FOUND A REAL BUG, and it is the interaction I would not have thought to look for. First-match-wins met an empty value: a file with system system /System/Boot/bare.bin matched the first line, handed back nothing, and the machine tried to start a file with no name while a good setting sat underneath. An unusable value is an absent one - which is what "configuration is advice" says, and this is where it earns its keep. cfgBare starts an image with no operating system in it at all, which is what loading an ordinary boot image buys: a program wanting the whole machine is a file like any other, chosen the same way the system is. Three disks differing ONLY in boot.cfg, so each is a test of the file rather than of the machinery under it. |
||
|
Anachronaut
|
54ff7196c9 |
Stage 0: the emulator carries the ROM, so a disk is enough
./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. |
||
|
Anachronaut
|
c312853f8e |
The machine starts itself
stage two CosmOS > saved it read it back, 22 bytes: Stage one hands over to stage two out of a boot slot; stage two mounts the filesystem, finds /System/cosmos.bin, takes the image apart and places its code, its data and its vector table, and jumps to the entry point the vectors named. Nothing placed memory for it. What it loads is an ORDINARY BOOT IMAGE, the same SPBT file the emulator has always been handed. That was the user's call and it is the whole trick: a second stage that loads the machine's normal image format is not a boot-specific mechanism, so bare metal SplitBit stops being a special case. A program wanting no operating system under it is just an image, written under CosmOS like any other, and startable because it is a file. Three things in it worth knowing: - THE ENTRY POINT IS CAUGHT ON ITS WAY PAST. Program Memory cannot be read back, so the boot vector cannot be looked up after being installed; the vector loop notices the one addressed at 0xFC00 and keeps it. - A missing "VEC" is not a fault. An image written before vectors existed simply ends after its data, and then the entry point is zero, which is what every such image has always relied on. - Feature flags that are set mean an image asking for a machine this may not be, and the honest answer to a request that cannot be understood is to refuse rather than to run it anyway. The test records that the system WORKS afterwards rather than that it started. A loaded program running is what says the vector table arrived, because a program reaches the system through SWI and nothing else; the file written and the directory entered say the filesystem and the console came up with it. A second disk has a boot slot and nothing to start, and says so rather than jumping somewhere. |
||
|
Anachronaut
|
82adeeb193 |
A boot payload can arrange its own Data Segment
Stage one places Program Memory and nothing else, because knowing where a payload's data ends and its code begins would mean knowing a format, and knowing formats is what ROM must do as little of as possible. But the real second stage needs a Data Segment: sbfs.asm has variables and a string it compares against. The answer needs nothing new. A loadable image is written into the slot as code followed by data, so the data image is already in Program Memory just past the code - and the payload's first instructions blit it down to where it was assembled for. Proved by slotData.asm, which prints from a string it placed itself. The padding is the part worth recording. The blit needs a length and the assembler will not work out the difference between two labels, so the segment is padded to a round number and that number is what gets copied. The first draft padded to 257 and copied 256, and the byte that did not arrive was padding, so it worked by luck. It is exact now and says why. This is the shape the user asked for and it goes further than the mechanism: the second stage becomes a loader for the machine's ORDINARY image format rather than for anything boot-specific, so bare metal SplitBit stops being a special case. A program that wants no operating system is just an image, developed under CosmOS like any other, and selectable at boot because it is a file. |
||
|
Anachronaut
|
d07b23f90b |
Rung 2: the machine starts itself off a disk
Stage one exists and works. It is 330 bytes of program and everything it knows is a thing that will be true forever: which port the disk is on, that a SplitBit disk begins with its own name, and where two numbers sit in that first block. Not what a file is, not what a directory is, not that SBFS has versions. It reads the live boot slot into Program Memory, jumps to the first byte, and prints one character and halts if there is nothing there. It is an ordinary boot image for now, so the whole chain runs on machinery that already exists and the emulator has not been touched. Nothing about it changes when it moves into ROM except who puts it in memory. SplitDisk gained "boot" to write a slot and "bootslot" to choose one, kept apart on purpose: writing a slot and starting from it are different decisions, and joining them would make every write a commitment. A slot is always written WHOLE, because one still holding the tail of what was there before is one whose contents depend on its history, and stage one reads all of it without knowing where the file stopped. Three recorded tests, and the pair is the point: two disks differing only in which slot the superblock names, with payloads that say different things. One prints "booted" and the other does not, so this is a test of CHOOSING a slot rather than a test that some bytes were read. The third boots a disk with no boot area and gets the one character a ROM has room for. Eight more host checks, including that a slot is padded whole. Two things worth recording. The first draft used #Align to put the scratch buffer at 0x8000 and produced a 33K file - thirty two kilobytes of zeroes in something meant to be a ROM. It is an address, not storage, which is exactly what the assembler's own scratch map exists to say. And SplitLint caught the second in code written an hour after the baseline that catches it. In the blit set-up, RSTA writes a source address of zero and then RSTA writes a bank number of zero - two unrelated quantities that are equal by accident, in the most safety critical file in the repository. It is marked with a reason rather than removed. |
||
|
Anachronaut
|
612bd1b97c |
B1: a boot area on the disk, reserved by arithmetic that was already there
The first rung of booting from disk. A boot area is blocks between the superblock and the directory that the filesystem never allocates and never sees, and NOTHING WAS ADDED TO RESERVE THEM: both implementations work out the first usable block as directoryStart + directoryBlocks, and directoryStart has always been a field rather than a constant. Formatting with the directory moved up reserves everything below it. Neither allocator changed, on either side. Two new superblock fields in bytes that were reserved: bootBlocks at 14, per slot, and bootSlot at 16. A disk made before this has zero in both, which reads as "no boot area" - true, and the same shape as the version two parent field, where the value an older disk already held was the right answer without conversion. TWO SLOTS, ALWAYS. A boot slot is raw blocks with no entry to rename, so the write-a-temporary-and-rename ordering that protects every file cannot protect it, and a machine interrupted while updating its only slot would not boot at all - the one failure on this disk with no way back. Writing the slot that is not live and then moving one byte makes that a machine that boots what it had before. bootBlocks and directoryStart say the same thing from two sides, so a disk where they disagree is refused rather than guessed at, as is one naming a slot that does not exist. Checked where it matters: the HOST formats a disk with a boot area and the MACHINE fills it, then the reserved blocks are compared against zero. The machine's allocator is the one that had no idea any of this was happening, which is what makes that the check worth having. Six host checks besides, including both halves of the superblock disagreeing. |
||
|
Anachronaut
|
0a2965bc63 |
Name the rules, say when there is nothing to say, and hold a baseline
Four things SplitLint wanted, and they build on each other. EVERY WARNING NAMES ITS RULE, in brackets at the end the way a compiler names the flag that produced it. Twelve rules, listed by --help. That makes the other three possible: suppressions can name one rule and leave the line honest about the others, the harness can assert on a rule's identity rather than on the wording of its message, and --machine can print one tab separated line per warning - file, line, rule, message, help - so nothing downstream reads prose. This file's own output was parsed with regular expressions three times in one day before it had a shape to rely on. A CLEAN RUN SAYS SO: No style warnings: 121 files checked against 12 rules. It used to exit in silence, which does not tell you it found nothing - it tells you nothing at all, and from outside the two are identical. A MARKER THAT SILENCES NOTHING IS ITSELF REPORTED, as dead-suppression. An exception that outlived whatever made it necessary is the thing the required reason exists to prevent, and naming the wrong rule now gets you both the warning you meant to silence and a note that your suppression is doing nothing. AND THE CORPUS IS HELD TO A BASELINE. Sixty one warnings are left in it deliberately and nothing stopped a sixty second. Tests/lint-baseline.txt records how many of each rule each file should produce, so a new one fails make test while the sixty one stay quiet; confirmed by adding an INIA 0d0 to Say.asm and watching it name the file, the rule and the count. It counts per file and rule rather than recording line numbers, because line numbers would churn the whole baseline whenever anything was inserted above a warning - the same reason cycle counts are stripped from recorded output here. ./Tests/lint.sh --bless records it again. One thing to know for next time: the rule name was inserted before the line number at all twenty one call sites, and the signature was changed to match rather than the twenty one call sites being fixed. (path, rule, line) reads no worse than (path, line, rule) and one edit has fewer ways to go wrong than twenty one. |
||
|
Anachronaut
|
9c144469b4 |
Take the SplitLint findings that are one operation, leave the rest
Twenty four more sites, and the interesting part is which ones were left alone. A rule emerged while reading them and it held all the way through: apply where the repetition is INSIDE one operation, skip where the author's own structure says it is a new thought, and never where two equal values mean different things. Taken: - Five registers reassigned to a value they already held, where both are the same quantity: two masks in one expression in Snake, two spaces printed by the monitor, both halves of block zero in waitTest, and a RSTA in Pour that the very next instruction overwrote. - Eighteen SETDs that reload a pointer inside one operation - a store back into the variable just read, or an INCD stepping to the second byte of a two byte value. Those read correctly without the reload. - sbfsNext, which branched to the label on the line below it. Left, with reasons that are the useful part of this: - Eight registers where the same number means two different things. CosmOS and the loader set A to 1 for a blit command and then to 1 again for a bank number; Asm compares a type against 3 and then a status against 3. Removing those couples one quantity to another that is equal by accident and would part company silently. - Ten RSTAs that open the RSTA/RSTB/CCF/ADD "return zero" block. The redundancy is what makes that idiom self contained; taking it out makes the return value depend on the line above. - Eleven SETDs that begin an arm of a comparison chain. Each arm loads, compares and branches, and they get reordered - the repetition is the reason a new arm can be dropped in anywhere. - Twenty five SETDs separated from their pointer by a blank line or a comment, which is the author saying a new thought starts here. - Two CCFs before arithmetic, which this codebase writes unconditionally. - Three redundant branches in test programs whose recorded output includes addresses, where three fewer bytes moves what the test demonstrates. Nine recorded outputs moved and every one is a size in a listing or, for Life, five more generations inside the same cycle budget. Behaviour is unchanged everywhere: cosmosSnake and cosmosEdit pass byte for byte while Snake loses eight bytes and Edit twelve. CosmOS is 10,902 bytes of program against 10,937, and the native assembler 12,173 against 12,183. The CosmOS README's size for Edit moved twice in one sitting, and this morning's check caught it both times - which it could not have done before that claim was reworded to name what it was about. |
||
|
Anachronaut
|
e1273337c4 |
Two mechanical fixes SplitLint found: MVQA, and RSTA for zero
Twenty four places moved Q into A or B by pushing it and popping it back. That is four bus cycles and two bytes to do what MVQA does in one of each, and several of them are inside loops - Life, the calculator, int8. Nineteen more loaded zero with INIA 0d0 where RSTA says the same thing in one byte. Both are equivalent at the CPU rather than by assertion: RSTA and INIA both leave Status alone, and PSHQ followed by POPA nets to A = Q with the Stack Pointer where it started. The one difference is that the pair leaves a copy of Q in memory just below the Stack Pointer and MVQA does not, which nothing here reads. Five recorded outputs moved and every one of them says the change worked: - 16x16Life fits five more generations into the same cycle budget, the first 457 lines identical, because the loop got cheaper. - Life.sbx is 1409 bytes rather than 1411, in three tests that list it. - Edit.sbx is 1995 rather than 1996. That last one broke a check I added this morning, and the hole is worth recording: the CosmOS README's claim about Edit's size did not have the word "Edit" on the same line as the number, because the subject was in the sentence before, so the check that measures quoted sizes skipped it silently. The sentence now names what it is talking about, which makes it both checkable and clearer, and the check fails on a wrong number there. Comments on either half of a replaced pair are carried onto the instruction that replaces them, so nothing anybody wrote was lost. |
||
|
Anachronaut
|
c146d98588 |
Stop the linter recommending a change that a faster helper would break
SplitLint knew that CALL restores A, B and Data Pointers 0 to 2, so a pointer set before a call is still set after it. That is true, and it made the tool give advice that was correct today and unsafe to take. Of the 178 redundant SETDs it found across the corpus, 122 were redundant ONLY because of that restore - the shape is everywhere, because it is how a helper is given its arguments: SETD.0 SbfsBlock SETD.2 SbfsFileStart CALL sbfsSetWord SETD.0 SbfsBlock <- flagged Removing that last line is right until sbfsSetWord is reached with RCAL, which restores nothing - and that is not hypothetical, it is what RCAL was added to this machine for, measured at close to halving the assembler's memory traffic. The failure would also be silent from the linter's side: it forgets everything across an RCAL, so it would stop reporting while the removals stayed removed. So a claim now ends at any call, for pointers and for registers, the way a claim about carry already did. 257 warnings become 127, and the redundant SETDs 178 become 54 - which is exactly the number an independent count of "no CALL in between" had arrived at separately. The fixture gained a SETD and an INIA repeated across a CALL, which must stay quiet, and the harness fails with the old behaviour put back. Two mistakes worth recording: the new expectations first pointed at the LABEL above the repeats rather than the repeats, which passes for free because nothing ever warns about a label; and the block landed in the middle of another check's comment, leaving that comment describing the code below it instead of its own. |
||
|
Anachronaut
|
8f4cc5878d |
Repair a table cut in half, and measure the numbers that had drifted
The README's emulator options table was split by forty lines of prose: two rows, then the whole discussion of the cost model, then five more rows with no header above them. Markdown renders that second half as something other than a table, so three of the seven options were not being shown as options. The rows are back together and the prose follows them. Four numbers had gone stale, in three different ways, and none was noticed: - "Five more scripts run alongside it" - there are six, and lint.sh had no bullet saying what it was for. - "rebuild all three tools" - there are four. - Files.asm quoted at 645 bytes in two places; it is 665. - Edit quoted at 1,983; it is 1,996. The last two are the most quotable sentences in the CosmOS README and the least likely to be rechecked by hand: the programs kept being made better and the sentences about how small they are stayed where they were. So docs.sh measures all four now. It counts the scripts in Tests/ that are not the driver or the disk builder and checks the README says that many and explains each one; it counts what the makefile's all target builds and checks the tool count in both phrasings, which took two attempts because one sentence says "the four tools" and the other "all four tools"; and it assembles every app the CosmOS README quotes a size for and compares. Each check was confirmed by making the fact wrong and watching it fail. WAIT also added a second kind of cycle this morning and the cost model section still described only one. It now says what an idle cycle is, why the two are counted apart, and what that distinction is FOR - a machine that slept through a slow disk and one that spun on it take the same elapsed time and print the same characters, and only the split tells them apart. The duplicated sentence about pipelining is gone; it was said twice, eleven lines apart, in nearly the same words. |
||
|
Anachronaut
|
2b079324ae |
Add SplitLint, and let it be told when something is deliberate
SplitLint reports valid assembly that has a shorter direct expression: zero loads that could be RSTA or RSTB, Q moved through the stack where MVQA would do, self-cancelling push and pop pairs, assignments overwritten before use, unreachable fallthrough, one-byte pointer moves that could be INCD or DECD, a branch to the label directly below it, a SETD reloading an address the pointer already holds, and branches whose carry is known. Its model is deliberately local and conservative: every label and every directive forgets all known state, so a claim only ever lives inside a straight-line region. It knows the calling convention - CALL forgets DP3 and keeps the rest, RCAL and SWI forget everything - and it shares assembly.o with the assembler, so an added opcode cannot leave it holding a private copy of the instruction table. 260 warnings across the corpus, of which three were wrong in the way that matters: branchTest.asm and interruptFlagTest.asm exist to check that a branch whose carry is known behaves correctly, so a diagnostic saying the outcome is known is exactly right and exactly unwanted. A line whose comment says "splitlint: <reason>" is now not reported on. THE REASON IS REQUIRED and a bare marker is refused, because a suppression nobody explained outlives whatever made it necessary. Suppressed warnings are not counted, so --fatal-warnings does not fail on one, and the number of them is printed at the end so the claim is visible rather than silent. Tests/lint.sh checked a TOTAL: twenty three warnings expected, twenty three found. That number stays right while the thing behind it goes wrong - a rule that stopped firing while another fired twice would pass, and so would a rule reporting at the wrong line. It now checks which warning came out and at which line, that nothing else came out, and that the four lines meant to stay quiet did. Confirmed by breaking one rule's message and watching it name that rule: the old assertion passed the same sabotage, because the warning still fired and the count never moved. Written with the user while I was away; my part is the suppression mechanism, the harness rewrite, and the three marks in the test programs. |
||
|
Anachronaut
|
c216c83e12 |
The filesystem waits for the disk instead of asking it repeatedly
sbfsWaitDisk spun on the status port. Read ahead had already hidden about
three quarters of the latency, and what remained was still 11.5% of Type
over a 14K file on a ten thousand cycle disk - all of it memory traffic
spent finding out that nothing had happened yet.
It tests the port first and only waits if the disk is still busy, and that
order is the whole of what makes it safe: the disk raises its line when it
finishes, so a disk that finished in the gap between the test and the WAIT
has its line standing already and the WAIT does nothing rather than
sleeping through the answer. No handler and no vector - the shell keeps the
Interrupt Flag down, and a WAIT wakes on a line whether or not anybody
means to answer it.
Printing a 14K file, by where the cycles go:
cycles a block total bus waiting
0 922,570 922,570 0
2,000 946,474 922,702 23,772
10,000 1,042,474 922,702 119,772
The bus column stops moving. What the program costs in memory is now the
same whatever the disk does. On this emulator nothing observable changes;
on hardware it is a CPU standing out of the way of the memory controller
rather than competing with it for every one of those 119,772 cycles.
The first version cost 660 cycles more at latency zero because it read the
status port again on the way out. AND writes to Q and leaves A alone, so
the status was already there - which is what the original said in its own
comment, and what I stopped believing while rewriting around it.
|
||
|
Anachronaut
|
c3188ed657 |
Seventy becomes seventy one: a machine that can wait
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. |
||
|
Anachronaut
|
6b41354f8f |
Delete a comment describing a design that was removed, and check the rest
Two comments in the native assembler survived the changes that made them false, and both are the kind that misleads rather than merely ages. Asm.asm still explained an eighteen kilobyte buffer that the whole output had to fit in "because a file is written in one call and there is nowhere to put half of one" - which stopped being true when the assembler learned to stream, and the variable it described, ImgRoom, does not exist any more. It was sitting in front of the comment that replaced it, so the paragraph a reader met first described the design that had been taken out. Replaced by what is actually there: no limit but the disk, one block at a time through a window in scratch. scratch.asm said the system keeps below 0x1000 twelve lines above the paragraph explaining that the system's half of Data Memory was doubled to 0x1FFF. A stale number next to its own correction is worse than a stale number alone, because which one a reader believes is a coin toss. docs.sh now checks both of the facts scratch.asm asserts about the machine around it: the floor it claims the system keeps below, against the CosmOS README, and the address its buffers start at, against where the assembler's own data actually ends. Neither is enforced by a line of code anywhere - the map is a comment, deliberately, because reserving the buffers would put 22K of zeroes in the file and the assembler could not load itself - so a check is the only thing that can hold them. Both fail when broken; the second reports the exact overlap. |
||
|
Anachronaut
|
c74075dc51 |
The review notes are all done, so the file that said to delete it goes
Six items, all fixed and all tested: the streamed commit bounded against its reservation, the memory map checked against itself, the save guarantee stated precisely, temporaries marked in the entry rather than by name, the path limits measured and left alone, and the directory bounded to what the parent field can name. The list of design strengths at the end went with it. It was a reader describing what the format gets right, which the README says in its own words and at more length; keeping a second copy in a file headed 'delete me' would only be a second copy to keep current. It is in the history if anybody wants to read it back. |
||
|
Anachronaut
|
ce0f18f4ef |
Refuse a directory whose last entries cannot be named as a parent
A parent is an entry index PLUS ONE in two bytes, so entry 65535 has no parent number: adding one wraps to zero, and zero is the root. Eight entries to a block, so 8192 directory blocks reaches it and SplitDisk formatted that happily. It does not fail by refusing, which is why it was worth chasing rather than reasoning about. Reproduced on a disk built for it: mkdir /deep/child, with /deep at entry 65535, printed 'Made "/deep/child" as entry 0' and put child in the ROOT. Listing /deep then showed nothing, because the search is for a parent of 65536 and the entry carries zero - so the same mkdir succeeded again, and again, and five entries called /child piled up in the root. Duplicate names in one directory are the one thing rename refuses outright, on the grounds that a search answers with whichever it meets first and the rest can never be reached; this manufactured them one per attempt. 8191 blocks is the most, giving 65528 entries. Refused when formatting and again when reading, in both implementations, because a disk claiming more was made by something that never checked. On the machine only the high byte of the count has to be looked at: anything from 0x20 up is too many. Three checks, all of which fail with their guard removed. The machine's disk claims the size rather than having it, so the test image is 64 blocks that lie rather than sixteen megabytes that do not - mounting is refused at the geometry, which is read out of block 0. |
||
|
Anachronaut
|
634650cab9 |
The path limits are settled: leave them, and say what they are
Measured rather than adjusted. The longest path on a full install is 21 characters against a native cap of about 117, so neither that nor SplitDisk's 512 binds anything, and both can be raised if that ever changes. Looking for the evidence turned up the limit that did bind - the prompt's 127 bytes, fixed in the commit before this one. |
||
|
Anachronaut
|
2b5506ee70 |
Stop the prompt writing off the front of its own buffer
The prompt is the working directory's path, worked out each time by walking the chain of parents up to the root. The names arrive deepest first, so they are written backwards from the end of a 127 byte buffer - and nothing bounded that walk. Nothing bounds the depth either. A path given to one operation is capped at 95 characters and a 22 character name, but "mkdir a" and "cd a" are each far inside that and can be repeated forever. Six directories of 22 characters is 132 characters of path, and at that point the walk wrote down past the front of CwdText and into what the assembler had laid out below it: the shell's own command names. ExitName sits five bytes under, so the word "exit" went first and the shell stopped recognising the command for leaving. Measured, not deduced: fine at five levels, gone at six. The walk now counts the room it has left, byte by byte, and stops. What is already written is the DEEP end of the path, which is the end worth showing, so it is cut at the front and three dots say so - out of three bytes held back from the count, so there is always somewhere to put them. Twenty levels deep the prompt shows the last five and every command still works. cosmosDeep records that, and records it by running help, cd and exit from down there rather than by looking at the prompt: a wrong prompt is cosmetic, and this was writing into other variables. It fails with the bound removed. The tree is built by SplitDisk because a path that long cannot be given to mkdir in one piece - which is the same fact that makes the depth unbounded. The three path limits are written down in the README now, including which one actually binds. The other two do not: the longest path on a full install is 21 characters. |
||
|
Anachronaut
|
0c240f7ad3 |
The temporary namespace is no longer a note
Two left: one path length both implementations agree on, and a bound on directory entries that keeps every one of them nameable as a parent. |
||
|
Anachronaut
|
ce8fb721fe |
Say a temporary is one in the entry, not in its name
Saving something that already exists writes a temporary, deletes the original and gives the temporary its name, so that nothing is lost if the writing fails. The temporary was told apart from a real file by being called sbfs.part or sbfs.out - and those are legal names. Starting a save deleted whatever answered to one as stale scratch, so saving anything at all in a directory destroyed your own file of that name there, silently. Flag bit 0x04 now says it. The property is not in the contents - the same bytes become the finished file the instant the rename lands - so it belongs in the entry, which is the thing the commit changes. sbfsCreateTempAt is the door temporaries come in by, the commit writes the flags flat along with the name, and cleanup wipes what it finds only if the entry says it is ours. Anything else stops the save instead. The bit is also the recovery. Both listings show an unfinished write rather than sizing it, because the size in the entry is the room that was asked for and not what was written: "<unfinished>" from dir, and a line from SplitDisk saying the blocks are held and a rename brings the data back. That was the gap in what the last commit documented - the data survived a crash and nothing would show you where it was. Four new agreement checks, three of which fail with the guards removed. The fourth needed rebuilding first: both tests started on one disk, and the first save ate the sbfs.part that was the second test's SOURCE, so the copy failed for want of a file, never opened a stream, and passed while reporting on nothing. A disk each. The fifth check forges the wreckage by setting the flag on a finished file, since nothing here can crash a save half way through. No version bump: a committed file never carries the bit, so a disk this writes is byte for byte the disk the old code wrote, which the whole-image comparisons confirm. Only the wreckage differs, and older code reads that as an ordinary file - which is what it did before. |
||
|
Anachronaut
|
d6cbbb5034 |
Strike the three review items that are now fixed
The streamed commit is bounded, the memory map is checked against itself, and the save guarantee says what it is. Three left: the temporary-file namespace, one path length both implementations agree on, and a bound on directory entries that keeps every one of them nameable as a parent. |
||
|
Anachronaut
|
9e2aa0122e |
Say what a safe save actually promises, and where it stops
Both sbfsSaveFile and sbfsStreamDone write a temporary, delete the original and rename the temporary onto its name, and the README explained why that order and not the obvious one. It never said what the resulting guarantee is, which invited the stronger reading: "safe save" sounds like it survives anything. It survives everything that goes wrong while it is running - no run long enough, no free blocks, a refused write, a name that turns out to be a directory, a writer that gives up - and in all of those the original is untouched. It does not survive the machine stopping, because the commit is two block writes with a gap between them: stop there and the old file is gone and the new one is still called sbfs.part. The bytes are all present and one rename by hand recovers them, but nothing does that on its own. Written down in the README beside the ordering it qualifies, and at both commit points in sbfs.asm, where a reader is when the question occurs to them. Closing the gap wants a journal or a second directory, which is a lot of disk for two writes; a boot-time consistency check is the cheaper answer if it ever matters, and it would want temporaries flagged in the entry rather than recognised by name. |
||
|
Anachronaut
|
aa7bdc6acd |
Check the memory map against itself, not only against the code
The CosmOS README's Data row gave the system 0x0000-0x3FFF and a loaded application 0x2000 and above: two columns of one row that cannot both be true. Program was doubled to 0x3FFF when CosmOS outgrew its first map and that number was copied into the Data row as well, where the answer is 0x1FFF. docs.sh measured both segments against the CosmOS column and passed the table anyway, because it never read the column beside it. A number checked against the code and not against the number next to it is still unchecked, so it now reads both and compares them - and compares two further copies of the same fact that had gone stale on their own: the minimal application in the README, still based where applications lived before the doubling, and the map cosmos.asm opens with, which somebody reading the system reads before they read the README. Each of the three checks was confirmed by breaking the fact and watching it fail; the first reproduces exactly the text this commit removes. While in that header, the command list said five commands and CosmOS has eleven and a search path besides, and "dump is next" outlived the monitor. |
||
|
AnachronautandClaude Opus 5
|
db3d349da8 |
Read the next block while the program is busy with this one
A file is read front to back, so when something asks for a block, the one after it is almost certainly wanted next. sbfsReadOne asks the disk for it straight away and hands back the block that was wanted - so the transfer happens while the caller is busy, and the waiting is mostly over by the time it comes back. Printing a fourteen kilobyte file: 1,064,498 cycles at two thousand a block becomes 976,882, and 1,576,562 at ten thousand becomes 1,032,889. The second figure barely moves between those - seven per cent from an instant disk to a slow one, where without it the same change costs sixty eight. A machine that reads ahead stops caring very much how fast its disk is. NOT FOR DIRECTORY SEARCHES, and that was tried first and thrown away. A scan stops the moment it matches, so the block it would read ahead is one nobody will ever look at - a transfer to fetch and another wait to throw away. It was nineteen per cent SLOWER on a lookup at ten thousand cycles a block. Reading ahead is a bet that the next block is wanted, and a search is exactly the case that hopes it is not. The scan loop is untouched. Three per cent is what it costs when the disk is instant, which is the bookkeeping with nothing to hide behind it, and the default. The disk has one buffer, so the only way to know what is in it is to remember what was last asked for. Every read records that; a write clears it, because a write fills the buffer from memory and no read asked for what is in it. Getting that wrong would blit whatever happened to be there and call it the block somebody wanted. The read ahead is bounded by the file's own length, so reading the last block does not fetch whatever follows the file on the disk - which belongs to somebody else and would be paid for twice, once to fetch and once to discard. It cost one bug, and an instructive one: the next index was worked out by stepping SbfsIndex, which is the CALLER'S and is still wanted after the return - handleFileBlock compares it against the file's length to see whether this was the short last block. Every block reported the wrong number of bytes, and the output of Type skipped five lines in the middle. It has a place of its own now. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01E2JrLzFvuFX9fgi1LDRjrW |
||
|
AnachronautandClaude Opus 5
|
d4cba36c5e |
Devices that take time, and a filesystem that waits for one
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 |
||
|
AnachronautandClaude Opus 5
|
e0cf0a9a25 |
Charge the memory controller for the memory it moves
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 |
||
|
AnachronautandClaude Opus 5
|
f1e5cc46f6 |
A cycle is an access to memory, not an instruction
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 |
||
|
AnachronautandClaude Opus 5
|
54f5cfe8a4 |
Add Copy and Compare, which stream in both directions at once
Copy joins the read and write streaming services: source and destination are both larger than Data Memory while the program owns one block. Compare reads two files through separate blocks and ignores the bytes past a short final block, which belong to neither file. Between them they exercise empty, exact-block, part-block and 84,000 byte files, and the host extracts the copy afterwards so that two native programs agreeing with each other is not the only oracle. Written by ChatGPT, as their headers record, along with the agree.sh section and the manifest entry that drive them. THIS SHOULD HAVE COME FIRST. The commit before it staged whole files rather than the hunks it meant, so its manifest already names these two programs while their source was still untracked - that commit will not build on its own. Left in place rather than rewritten, since the pair is right and only their order is wrong. NOTES.md is their review of the streaming work. The first item in it is fixed by the commit before this one; the rest are still open. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01E2JrLzFvuFX9fgi1LDRjrW |
||
|
AnachronautandClaude Opus 5
|
a7d3e09d94 |
Refuse a streamed file that commits more than it reserved
osFileStart sets an extent aside and osFileWrite refuses a block index outside it, so writing off the end was already barred. Committing a larger size was not, and reaches the same neighbour by simply claiming it: a directory entry is the only record of what a file owns, so an entry claiming a block it was never given owns it, and so does whatever owned it before. Both files then look perfectly well formed. The free count went backwards past zero on the same path. Found by ChatGPT's review of the streaming work, in NOTES.md. I had bounded the index because writing off the end was the obvious way to reach a neighbour, and had not noticed that the other end of the same reservation was open. THE SIZE IS COMPARED, NOT THE ROOM IT TAKES UP. One block and a tail occupies exactly what two whole blocks occupy, so bounding the blocks alone would let a file reserve the first, commit the second, claim no block it was not given, and still report two hundred and forty six bytes that were never written to it - whatever the disk had there before. Checked before anything is touched, which is why the temporary is found twice. The old file is deleted a few lines down and a refusal after that point would have destroyed the thing it was protecting. AND IT CAUGHT A REAL ONE IMMEDIATELY. The assembler reserves the file plus room for its vectors, and asked for four bytes per vector DECLARED - which looks like a safe bound and is not, because a device is declared during the SECOND pass, in the line that implements it. A program with a device installs a vector that was not counted when the room was measured. CosmOS reserved 14,163 bytes and committed 14,167, writing four bytes past what it had been given on every build since S2. It landed inside the last block it owned, and would not have if the boundary had fallen four bytes earlier. It reserves against the vector table's LIMIT now, which cannot go stale whenever things are counted. Claim.asm is the program that tries it: reserve one block and a tail of ten, write them, then tell osFileDone the file came to two whole blocks. The refusal and the honest commit that follows are both recorded. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01E2JrLzFvuFX9fgi1LDRjrW |
||
|
AnachronautandClaude Opus 5
|
19ab36a201 |
The assembler can say where everything ended up
-S writes every label and the address it was given, in address order. Nothing else knows that. A program on the disk is bytes; the monitor can disassemble it but has no idea what any of it is called. So counting which addresses a program calls says a great deal and names nothing - the answer arrives as a column of numbers and somebody works out by hand which routine each one is inside. It was deferred when the native assembler was planned, as a listing and symbol dump nobody needed yet. Finding out where the assembler spends its time is what needed it: the top six call targets were addresses until this existed and are numStep, numCompare, tokGet, srcNext, numAddByte and clsSameName with it. Sorted by address rather than by name, because the question asked of it is always "what is at this address". Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01E2JrLzFvuFX9fgi1LDRjrW |
||
|
AnachronautandClaude Opus 5
|
af0360128b |
Sixty four instructions becomes seventy
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 |