Fixes against AmigaOS 3.2.3.
Our rule for every open replacement is to behave exactly like AmigaOS 3.2.3, down to the registers a call leaves behind. To keep that promise we run each of our libraries and commands beside the real one and compare every answer. Along the way the comparison found places where 3.2.3 hangs, crashes or gives a wrong answer. Those are the ones we did not copy.
This page lists them: what goes wrong in 3.2.3, what ours does instead, and how we know. Everything marked measured was seen on a Kickstart 3.2.3 test machine in Cradle. Two items are inferred from the system's own documentation and say so.
Compatible first, then better.
We are writing the parts of AmigaOS 3.2.3 again, in the open, and until every part is ours and passes its side-by-side test we copy the original's observable behaviour: its answers, its version numbers, its boot order, even the values it happens to leave in registers. A program that runs on 3.2.3 must run on ours without noticing.
That rule has one exception, and it is a deliberate one. When the original hangs the machine, writes past its memory or gives an answer that is plainly wrong, copying it would mean building the bug on purpose. So each library's design notes carry a short list headed "differs on purpose", every entry has a reason, and the comparison test knows to step around it. This page is those lists, gathered in one place.
Every call compared
Each replacement has a test that calls the original and ours with the same arguments and prints both answers, on a copy of an AmigaOS 3.2.3 machine in Cradle. The counts run from 109 checks for lowlevel.library to 111,528 for the maths libraries.
Then run for real
After the comparison, ours goes in the original's place and the same test runs again, followed by the programs that use the library. Help in IconEdit and Exchange, for instance, runs on our amigaguide.library.
Measured or inferred
Where we watched 3.2.3 fail, the table says measured. Where we only read the warning in its documentation, or ours has run only on a PC so far, it says so.
Hangs and crashes.
Calls and commands that stop the machine, or take it down, on 3.2.3. Most came to light because our comparison test asked a question nobody had asked before, and 3.2.3 never came back.
| Where | AmigaOS 3.2.3 | Ours | How we know |
|---|---|---|---|
| rexxsyslib.library, DeleteRexxHostPort | Frees the port but leaves it in exec's port list. The next FindPort walks into freed memory; in our labs that hung the machine every time. | Takes the port out of the list before freeing it. | Measured in the lab |
| lowlevel.library, ReadJoyPort | Asked about port 4 or higher, it hangs. | Answers that the port is not available. | Measured in the lab |
| lowlevel.library, SystemControl | One SCON_RemCreateKeys more than were added hangs. | Ignores the extra one. | Measured in the lab |
| lowlevel.library, StartTimerInt | An interval of 0 sometimes floods the machine with interrupts. | Uses one E-clock tick. | Measured in the lab |
| amigaguide.library, GetAmigaGuideString | Returns junk for most ids and hangs past id 28. | Returns NULL. | Measured in the lab |
| expansion.library, AllocExpansionMem | Asked for 0 bytes at alignment 0, it never returns. | Answers -1 at once. | Measured in the lab |
| expansion.library, AddConfigDev | A board with no size takes about 11 seconds. | Returns at once. | Measured in the lab |
| dos.library, FGets | A length of 0 never comes back. | Returns NULL. | 3.2.3 measured; ours hosted |
| dos.library, LoadSeg | A CODE hunk longer than its header is read past its memory. In the lab it took the whole machine down. An overlay file never comes back. | Refuses both with ERROR_BAD_HUNK. | 3.2.3 measured; ours hosted |
| graphics.library, AttachPalExtra | A colour map smaller than the screen's depth makes it write past the map and jump to an odd address. | Caps at the map. | Measured in the lab |
| DefIcons | A file type named . in its prefs hangs it. | Such a type is never the answer. | Measured in the lab |
| Eval | An LFORMAT ending in %X or %O prints ? and runs on through memory. It never stops. | Prints ? and stops. | Measured in the lab |
| List | FLAT together with ALL never stops. | Lists each directory once. | Measured in the lab |
Wrong answers.
Where 3.2.3 returns, but with a result we could show is wrong. The maths library is the biggest: 1,138 of 111,528 checks differ, and in every one we could verify, ours is the correctly rounded answer.
| Where | AmigaOS 3.2.3 | Ours | How we know |
|---|---|---|---|
| mathieeedoubtrans.library | Pow is off by up to three ulps. Sin, Cos and Tan reduce with a short pi, so a double next to a multiple of pi keeps only a few good bits, and a huge argument gives an unrelated answer. | Correctly rounded. Of 48 differing trig, Atan and Log10 results, 46 are ours right; all 292 differing Pow results checked are ours right. Errors and NaNs stay as 3.2.3's. | 111,528 checks side by side |
| locale.library, the built-in time zone | With no Locale prefs the offset is -300, which FormatDate prints as +0500: a fresh machine runs five hours east of UTC. | Uses the host's zone on AmiCradle, or Europe/London with summer time, and carries real zone rules for 419 places. | Measured in the lab |
| locale.library, FormatString | A positional argument past a handful of positions reads memory it never set. | Reads up to the 64th argument and gives 0 after that. | Measured in the lab |
| lowlevel.library, ElapsedTime | A context from the future answers whatever its divide leaves. | Answers 0. | Measured in the lab |
| timer.device | A wait is counted to a moment on the system clock. Set the clock back an hour and every Delay, C:Wait and VBLANK wait already running holds for an hour. | Relative waits count on the E-clock, so setting the clock never moves them. | 3.2.3 measured; ours by design |
| Date | Date 15-Oct-26 Monday, given on a Tuesday, lands on Saturday the 17th. | Takes the day name: Monday the 19th. | Measured in the lab |
| dos.library, StrToLong | Stops one digit early past 2^32. | Reads the whole number. | 3.2.3 measured; ours hosted |
| dos.library, ReadItem | At the end of a CSource with no newline, answers the last character again and again. | Reports the end. | 3.2.3 measured; ours hosted |
Small glitches.
Nothing here would lose you work. They are the kind of thing you notice once and then stop seeing, and we would rather not carry them over.
| Where | AmigaOS 3.2.3 | Ours | How we know |
|---|---|---|---|
| Eval | An error after TO has opened its file leaves the file open: object is in use afterwards. | Closes it. | Measured in the lab |
| Version | A version string that ends the file with no newline is followed by a stray ) on a line of its own, left in its buffer from the date. | Prints the string and nothing else. | Measured in the lab |
| Version | After Could not find version information, Result2 is 0 or 212 from run to run. | Always 0. | Measured in the lab |
| Execute | A key that starts with $ gives odd results: <$> leaves its >, <$x$y> gives y. | Reads the key as empty. | Measured in the lab |
| dos.library, VFPrintf | Formatting several WORD fields stops after the first one. | Formats them all. | 3.2.3 measured; ours hosted |
| dos.library, Fault | Answers a length one or three more than it wrote. | Answers what it wrote. | 3.2.3 measured; ours hosted |
| nonvolatile.library | The autodoc warns that SetNVProtection with any bit but NVEF_DELETE will crash, and a call from a plain task crashes in dos.library. | Both fail cleanly. A volume over 4 GB reports its real size, capped at the largest ULONG. | From the autodoc, not run |
| printer.device, DumpRPort | No hook and no RastPort would crash it, as far as we can tell. | Returns PDERR_BADDIMENSION. | Inferred, not measured |
Where the fixes are today.
None of this changes a stock AmigaOS 3.2.3 install. The fixes live in our replacements, and you get them when you use those.
Swap one in
rexxsyslib, lowlevel, amigaguide, locale, nonvolatile, the maths libraries and the rest of the disk libraries are files you can put in LIBS: in place of the original. The fix is on whenever ours is the one that loads; there is no switch. The maths fixes need an FPU, because without one the original library stays in use.
dos, expansion, graphics and timer
These are ROM modules, so their fixes arrive with our open ROM. The dos.library fixes have so far run on a PC under our test harness, not yet side by side on an Amiga, and the table marks them as such.
The commands
Eval, List, Date, Version, Execute and DefIcons are GFOpenBench commands. Each one is checked case by case against 3.2.3's command in the lab, and each difference above is written into the test as a known one.
The last few dos differences
StrToLong, ReadItem, VFPrintf and Fault are listed in our dos notes as differences still open under the compatibility rule, with 3.2.3's answer recorded first. If a program turns out to depend on the old answer, we will copy it after all. The hangs and the LoadSeg checks stay.
Not on this list.
Some things look like fixes against 3.2.3 and are not, and we would rather say so than let them pad the page.
Our own bugs
Most of the differences the comparison finds are ours, not theirs: a LoadSeg that forgot to clear the caches, a Version that printed the wrong file date, a ConClip that quit when 3.2.3's keeps going. Those get fixed to match, and they are the bulk of the work.
Limits we keep
The battery clock's two-digit year, FFS's file name length and the one-byte quirks of the keymap library are the same in ours. A program may depend on them, so until the platform is feature complete, they stay.
Formats we keep off
Our icon.library can read NewIcons and PNG icons, which 3.2.3 cannot. It does not, by default: it reads them the way 3.2.3 does, and the extra readers are a build option, so a disk made for ours still looks right on a stock machine.