mindbox77 has written a complete graphical desktop for the 48K ZX Spectrum, in Z80 assembly, for a machine from 1982. Overlapping windows with a z-order and focus, a permanent menu bar with pull-downs, a sixteen-slot event ring, a heap with an owner byte, a two-pane file manager, a notepad, a clock, a calendar, and a settings panel that actually changes things. All of it fits in 48K, and it can drag a window to completion inside a single 69,888 T-state frame. It runs on the real machine, not an emulator.

His reason is plainly put: in the eighties he wanted an Atari ST and couldn't afford one. What he actually wanted was GEM — the desktop, the menu bar that was always there, the feeling that the machine was a place rather than a prompt. He had a Spectrum instead. He started writing bits of it and never finished. This is that, finished, forty years later. It isn't a port of GEM and doesn't pretend to be; it is what the idea turns into when you push it up against a 3.5 MHz Z80, 48K of RAM, a one-bit display with attribute clash, and a video chip that steals cycles from the CPU while it paints.

But the repository's real content is not the desktop. It is the measurement log.

The fastest way to fill a rectangle on this machine is to use SP as a write pointer — PUSH writes two bytes at a time, and a chain of them moves pixels faster than anything else available. The cost is that SP stops being a stack. The textbook answer is one instruction: DI. Disable interrupts, protect the region. He did that, and then found the Spectrum does not defer interrupts at all. The ULA asserts INT for exactly 32 T states and then withdraws it. A DI window covering those 32 T does not postpone the interrupt; it destroys it. He measured it by injecting an interrupt after every single instruction of a fill: 458 of 500 instruction boundaries refused it in one routine, 710 of 752 in the other.

The critical section — disable interrupts around code that must not be interrupted — is one of the most thoroughly already-construct objects in all of computing. It was a live cut once, in the mutual-exclusion problems of the sixties. It got named. It became construct, and after that nobody saw it again. And it carries a premise it never states: that the thing being kept out will wait. DI means "I am not available right now," which only holds if the world is willing to hold. The Spectrum's ULA is not willing. It speaks once, for 32 T states, and is done.

So the remainder is the temporal shape of the interruption itself — the difference between a message and a window. A message in a queue can be collected late. A window that has closed cannot. The critical section never made room for that difference, because on the machines where it took shape the difference did not exist.

The sharper point is which way the repair runs. Re-enabling interrupts between rows sounds right and fails, because the interrupt is not pending, it is gone — an EI window a few T states wide catches it about one row in thirty. His sentence is the hardest thing in the document: making the DI region short is not the same as making it absent, and only absent is a fix.

So the fills now run with interrupts enabled throughout. What DI was protecting was SP, so the real fix is to guarantee that the two bytes of return address an interrupt pushes always land somewhere that is about to be overwritten anyway. The push chain is made one push shorter than the row, and the leftmost two bytes are owed — paid one row later, once SP has moved on and can no longer reach them. The interrupt is no longer excluded. It is given somewhere to land. Severing becomes cultivation. It costs 48 T states a row, and he says it is worth it.

The repository holds a run of other things with the same shape. He built a frame watchdog to catch dropped frames, and the watchdog was caught by the thing it was built to catch: INC (IX+0) sets the flags, and a fill holds a live carry across the test that decides whether a row has crossed a cell boundary. An interrupt landing in that gap stole the carry and the row stepped to the wrong address. The roadmap had predicted the watchdog would catch the interrupt bug earlier. What happened is that the interrupt test caught the watchdog. The instrument is not outside the system.

Then the heap bug that healed itself. The free routine had one DEC HL too many and should have ruined the heap on the first release. It didn't — because the coalescer then read a zero-filled payload as an empty free block and absorbed it four bytes at a time until it arrived back at exactly the true total. Every number agreed. It only surfaced when a freed block held something other than zeros, which was the first block big enough to have held window pixels. The lesson is not about the allocator but about the test: a heap test that frees blocks it never wrote to is testing a zero fill. It writes $A5 over the payload before freeing now, and asserts on the header directly rather than through a total that can heal.

And one plainer one. Timings are taken by counting a sixteen-T loop until the next interrupt, so a slower routine leaves fewer turns. He read that backwards twice in one session, once reporting a 1,216 T cost as a 76 T saving — so a real regression went unnoticed for being displayed as an improvement. Thirty years of folklore say screen writes pay a 50% contention penalty; he swept a 1,024-byte fill across the whole frame and measured a worst case of 13,473 T against 11,744 T uncontended, which is 14.7%, so every budget he had built was a third too conservative. He put that in the document too, including the part where his own budgets were wrong.

These are the moments before naming. "Owing the last push" has no name in any operating-systems book. It still has the shape of the particular machine it was found on; it has not been generalised, has not become a word people reach for without looking. The day it gets a name it will be useful and repeatable and dead as a remainder — one more layer of sediment, like the critical section, that nobody sees any more. It is alive right now precisely because you still have to read the whole paragraph to understand it.

Han Qin (秦汉) · @hqinjarsy · SAE Judgment & Aesthetics