4.0KB

Исходник Blame История

story: E01-S05 owner: e01_memory_map status: done started: 2026-11-07 depends-on: [E01-S02, E01-S03, E01-S04]

E01-S05 Claim

Intended change

Freeze the compiler memory map from E01 evidence: common-RAM placement, a private compiler zero page via P0, bank-1 table space, and a fixed test mailbox. Provide cfg/compiler.cfg, src/common/c128.inc, src/common/mailbox.inc, and the production src/compiler/mmu.s (bank switching, KERNAL wrapper, NMI guard), then prove them with a linked skeleton in x128.

File area

cfg/compiler.cfg
src/common/c128.inc, src/common/mailbox.inc
src/compiler/mmu.s
tests/platform/map-skeleton.s
scripts/build-map-skeleton.ps1, scripts/verify-map-skeleton.ps1
docs/architecture.md section 4, docs/decisions.md
planning/claims/E01-S05.md
generated files under build/

Acceptance evidence

docs/architecture.md lists verified regions and reservations (section 4.4; full table and evidence in docs/memory-map.md).
cfg/compiler.cfg can link a capacity-checking skeleton (normal link passes; ZP, COMMON, MAIN, SWSTACK, and TABLES overflow builds are each rejected by ld65 in the expected memory area).
src/common/c128.inc defines observed hardware/KERNAL constants, each annotated with the document holding its evidence.
The test mailbox address and protocol are documented and exercised ($1BF0, src/common/mailbox.inc; x128 run reads 4E 38 01 A5 00 00 ...).

Coordination notes

Absorbs the E01-S03 open NMI risk: the skeleton takes a CIA #2 timer NMI while bank 1 and the private zero page are active.

Work log

Read from x128 ROM: $FFFA -> $FF05 (SEI, push A/X/Y/CR, CR=0, JMP ($0318)); $FF33 pops CR/Y/X/A and RTIs; $FF17 IRQ path jumps via $0314 into BASIC, which uses system zero page. $1300-$1BFF holds only the power-on fill pattern after boot.
First x128 run never touched the mailbox: mmu.s linked first, so mmu_init sat at the SYS 7181 target. Added a STARTUP segment placed first and a build check that the entry label is at $1C0D.
Memory-area and segment define symbols collided (__DISKBUF_SIZE__); only the memory areas define symbols now.
A CIA #2 timer NMI taken in the bank-1 state with the private zero page returns cleanly through ROM $FF05 -> mmu_nmi_guard -> $FF33, which closes the E01-S03 NMI risk.
Negative control: forcing P0L to $00 in mmu_bank0/1 gives mailbox FAIL code $000F (ZP, KCALL, BANK1, NMI). Source restored afterwards.
Found that the earlier branch switch/merge had rewritten all committed newLang/ text files with CRLF (core.autocrlf=true), breaking the grammar validator. Added newLang/** text=auto eol=lf to .gitattributes and converted the 79 files back to LF; they match the committed blobs again.

Completion

Tests run: scripts/verify-map-skeleton.ps1 (includes five overflow builds), mutation run, tests/grammar/validate.ps1, tests/disk/prepare-disk.tests.ps1, scripts/verify-minimal.ps1, scripts/verify-display-clock.ps1, scripts/verify-mmu-probe.ps1, scripts/verify-kernal-probe.ps1.
Results: skeleton PASS with code $0000; one NMI under CR $4E / P0 $1A with sentinel $C5 intact; screenshot shows both KCALL lines and NOVA128 MAP SKELETON OK; all earlier regressions pass, and the disk builder still produces SHA-256 208EF297....
Files changed: cfg/compiler.cfg, src/common/c128.inc, src/common/mailbox.inc, src/compiler/mmu.s, tests/platform/map-skeleton.s, scripts/build-map-skeleton.ps1, scripts/verify-map-skeleton.ps1, docs/memory-map.md, docs/architecture.md, docs/decisions.md (D117), this claim, and the repository .gitattributes.
Remaining risks: generated-program map still open (E04); bank-1 block-copy routines not yet written (only byte peek/poke); RESTORE is ignored during compilation (documented in D117); IRQs stay masked, so long compiler phases freeze the jiffy clock and keyboard except during KERNAL calls; VICE-only evidence.

4.0KB Исходник Blame История