После полной миграции на merge флаги g_cur_merged/g_next_merged всегда
равны 1 (все типы блоков строит форвард-сканер). Убраны сами флаги и ~30
гвардов `if (g_cur_merged)` в горячем посимвольном цикле — это и небольшое
ускорение (ветка на символ), и чистка.
Проверено замером: инлайн gc_put (макрос/inline) в этот регистро-нагруженный
цикл, наоборот, ЗАМЕДЛЯЕТ (~11→13с) из-за роста спиллов — оставлен функцией.
Вывод: per-char микрооптимизации здесь исчерпаны (asm-fb не помог, инлайн
навредил, снятие гвардов дало ~1с).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Раньше препроцессинг сканировал файл ДВАЖДЫ: index_lines считал точки
переноса (forward-сканеры scan_join_stream/inline_scan), а отдельный
render_line_to_cache заново сканировал каждый сегмент для сборки
(char,attr)-ячеек. Профилирование показало, что всё время — в этих двух
посимвольных проходах (bank I/O и cont-walk ≈ 0).
Теперь forward-сканер собирает ячейки в g_cells ПО ХОДУ единственного
прохода; emit_seg флашит ячейки предыдущего сегмента в кэш (lag-1),
последний — после цикла. Перенос строки усекает буфер до снимка на
последнем пробеле (g_ncells_at_space), continuation-сегменты получают
префикс (отступ списка / маркер цитаты 0xB3 / title только в 1-й строке
заголовка). Прямые типы (код verbatim, HR, таблица через nowrap-inline,
blank/fence) строят ячейки на месте.
Миграция шла по типам блоков с dual-verify (старый render строил эталон,
merge сверял ячейки) — найдены и согласованы расхождения forward-сканера
со старым рендером: backtick внутри эмфазиса = литерал; одиночный маркер
закрывает ЛЮБОЙ активный эмфазис; soft-join пропускает ведущие пробелы
строки-продолжения; хвостовые пробелы нерелевантны. Все типы дали 0
расхождений, после чего render_line_to_cache / cc_put / cc_fill /
handle_inline_marker и verify-каркас удалены (−185 строк).
Итог на 52КБ README: 22→11с (2×). Совокупно с прошлым коммитом 24→11с.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Профилирование (52КБ README, раздельный замер): препроцессинг 24с, из них
bank I/O ~0с, cont-walk ~1с — всё время в посимвольных скан-циклах
(индексация 8с + рендер 13с, два прохода по файлу).
- fb() переписана на ассемблере (была закомментированная заготовка): убран
вызов функции и 32-битная арифметика на каждый байт. Раскладка __sdcccall(1)
для uint32 аргумента (p=HLDE) и возврат char в A сверены через sdcc -S;
координация маппинга с bank_read держится на том, что _io_page_w3 — порт
(__sfr 0xE2), пишем OUT — bank_read читает IN.
- render_line_to_cache: горячий путь обычного текста больше не вызывает
handle_inline_marker (6 аргументов) на каждый символ — только на маркерах
` * _ ~; cc_put заинлайнен. Поведение идентично.
Итого 24→18с. Индекс-скан (8с) лёгкого инлайна не имеет (нет per-char
вызовов). Следующий шаг — merge двух проходов в один.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Горизонтальный скролл (Фаза 5, финал):
- Скроллим по ТИПУ, а не по длине: новый флаг IF_HSCROLL ставится только
на код и таблицы; HR и границы fence (IF_NOWRAP без IF_HSCROLL) не
двигаются. Блок едет целиком, включая строки короче 80.
- Обход бага кодогенерации SDCC z80: `if (n!=g) g=n;` пишет (n-g) вместо n
(SUB сравнения затирает A, store переиспользует испорченный A). Лечится
записью viewport_x ДО сравнения. Минимальный репродьюсер и оба описания
для трекера — в docs/bugs/sdcc-z80-cmp-store-a/ (воспроизводится на чистом
sdcc 4.5, в т.ч. с --no-peep → это кодогенератор, не peephole).
Рендеринг:
- Отступленный fence (```c внутри списка) теперь распознаётся: is_fence_raw
пропускает ведущие пробелы/табы; то же в рендере прячет строку-границу.
- Строки-разделители (HR, ровно 80) больше не участвуют в скролле.
Чистка: удалён мёртвый код (is_fence_delim, get_init_style[_raw], is_cont,
seg_flags). Makefile (mdview/mdview2): iconv UTF-8→CP866 завершается ненулевым
кодом при отбрасывании символов — игнорируем (|| true).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
New examples/mdview2 — render-cache version of mdview: each logical
line is rendered into an EMM (char,attr) buffer once during file load
(interleaved with index_lines()), then scrolling draws straight from
the cache via ESTEX WINREST, with no re-parsing or fb() access in the
steady state. Horizontal scroll still uses the old live-render path
(Phase 5, not yet migrated).
Format and budget were derisked empirically first (tests/winrest):
confirmed ESTEX WINCOPY/WINREST buffer layout and measured EMM free
space, both folded into the implementation.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
hello2 — stdio vs conio output API comparison (fast/no-attr vs
slower/attributed), also exercises textcolor/textbackground.
simple — packed 2-bit style array bit-twiddling smoke test.
banktest — __banked function calls across two explicit banks in huge
memory mode.
hello2.c updated to call the renamed gettextmode/settextmode (see the
conio rename in 5e5e70d).
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
mem_alloc_pages/mem_free_block/mem_get_page/mem_info, implemented twice:
mem_estex.c via ESTEX $3C-$3E (INFOMEM/GETMEM/FREEMEM), mem_bios.c via
the BIOS $C0/$C2-$C4/$C6 EMM_FN* calls. lib/Makefile already listed both
under LIBC_C.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
palette.h/video/palette.c wrap BIOS $A4 PIC_SET_PAL/PIC_GET_PAL and $A6
SET_PAL_INIT for the 8 palette pages (0-3 graphics, 4-7 text planes).
gfx_palette.c and conio/text_palette.c are thin per-domain wrappers;
text_palette.c maps the text "plane" 0..3 (paper/ink/blink-paper/
blink-ink) onto BIOS pages 4..7. Covered by tests/text_palette.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
_get_cursor/_set_cursor now use BIOS GetCursor/SetCursor (RST 8, 08Eh/084h)
instead of ESTEX CURSOR/LOCATE; cursor position packed into a two_bytes
union. get_videotextmode/set_videotextmode renamed to gettextmode/settextmode.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Replace per-character wrchar() loops in render_line/fill_row/help screens
with batched BIOS calls (bios_writeattr/bios_fillcharattr), and cache the
per-line index record (idx_get) instead of refetching it ~13 times per
rendered line — each refetch cost two W3 bank switches via bank_read().
Also fixes a pre-existing indexer bug: when word-wrap pushes a token that
starts with an emphasis/code marker (e.g. `_text_`) onto a new line, the
marker got re-scanned a second time during the wrap continuation, flipping
line_style back off and corrupting the rendered attribute of the next
word. Fixed by snapshotting line_style at the last seen space and rolling
back to that snapshot (both the saved continuation style and the live
scan state) when a wrap is taken.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Direct __naked z80 wrappers over BIOS LP_PRINT_*/LP_SET_PLACE/LP_GET_PLACE
(RST 8, opcodes 081h-08Eh) for fast text rendering — a single BIOS call
replaces a wrchar() loop for constant-attribute/constant-character runs.
Covered by tests/bios_text.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- исправлен rewind до первого non-cont сегмента для continuation
- для quote-потока newline обрабатывается как soft join с пропуском сырого ' > ' маркера
- восстановлен quote-префикс на continuation строках
Co-Authored-By: Oz <oz-agent@warp.dev>
- Quote blocks now join consecutive quote lines into one paragraph with
soft breaks and wrap continuation under quote prefix.
- Empty quote lines ('>' / '> ') act as quote-paragraph separators.
- Nested quote starts ('> > ...') are not merged into the parent text and
stay separate rows.
- Keeps existing inline emphasis handling inside quote content.
Co-Authored-By: Oz <oz-agent@warp.dev>
Replace status-line printf formatting with compact manual decimal rendering
(u16/u8 right-aligned helpers + wrchar/put_str_attr). This removes runtime
printf usage from mdview.c and keeps fixed status columns without stale digits.
Co-Authored-By: Oz <oz-agent@warp.dev>
- List items (UL/OL) can now span multiple source lines: non-marker
lines are joined into the current item as lazy continuation.
- Continuation-line leading indentation is trimmed before joining so
wrapped item text is separated by a single space.
- A single blank line between adjacent list markers is suppressed
(same visual list), while 2+ blank lines still produce a separator.
Co-Authored-By: Oz <oz-agent@warp.dev>
- Paragraph scanning no longer swallows the blank line separating it from
the next block; runs of blank lines collapse to one row. Restores blank
separation between paragraphs, headers and horizontal rules.
- Detect table rows (first non-space char '|') as nowrap segments; they are
no longer merged into surrounding text or each other.
- Continuation (wrapped) rows render content as plain text and are no longer
re-classified, so a wrapped word starting with '-'/'#'/'>' is not mis-drawn
as a list/heading/quote marker.
Co-Authored-By: Oz <oz-agent@warp.dev>
- A trailing backslash before a newline now forces an in-paragraph line
break (like two trailing spaces); render consumes the marker (non-code).
- Add ~~strikethrough~~ inline style (INIT_STYLE_STRIKE / EM_STRIKE),
parsed in inline_scan, the paragraph merger and render, mirroring **.
- Horizontal pan is bounded by the widest nowrap segment on screen, and a
'<' indicator marks hidden content off the left edge.
Co-Authored-By: Oz <oz-agent@warp.dev>
- Index is now an 8-byte record per visible segment in a dedicated EMM
block (idx_get/idx_put), freeing ~11 KB of near RAM and lifting the
old 2048-line cap (dynamic max_lines = index_pages * 2048).
- The per-byte scan keeps the previous segment offset in a near var
(cur_seg_off) and mirrors the last record (cur_rec), so it never reads
the index back from the bank.
- fb()/map_page() are inlined now that there is code headroom, removing
per-byte call + 32-bit argument marshalling overhead.
Co-Authored-By: Oz <oz-agent@warp.dev>
Decode page/offset from the 32-bit offset's bytes to avoid SDCC z80
32-bit shift/mask helpers on the hot path; map_page() uses a cached
file_phys[] table to skip mem_get_page() on every W3 swap.
Behaviour-preserving.
Co-Authored-By: Oz <oz-agent@warp.dev>
- Add fenced code block tracking to the wrap-pass in index_lines().
line_style is reset to PLAIN at every ``` delimiter, and all segs
inside a fenced block get init_style=PLAIN. This prevents emphasis
markers (e.g. _ in __var) inside code blocks from leaking into later
normal text.
- Also carry init_style across wrap continuation segs so that a long
bold/italic line that is wrapped continues with the correct style on
the next segment.
- The fence bitmap pass now only updates in_code[], since init_style is
already set correctly by the wrap pass.
- Split tests/ (libc feature tests) and examples/ (real apps); shared
app.mk in repo root, was examples/example.mk
- libc/io/* split into libc/{conio,env,errno,file,mouse,string,sys,
time,video}/ — clearer module boundaries
- New examples/mdview/: markdown viewer (Phases 1-5 + light nested
lists). Headers (H1-H4), HR, ulist/olist/quote with nesting via
leading spaces, fenced code blocks, inline emphasis (bold/italic/
underscore/code), wrap/unwrap mode with soft wrap (F2), horizontal
pan (← →) with '>' truncation indicator
- libc additions: scroll() in conio (ESTEX SCROLL), strlwr/strupr,
gets() test
- Makefile updates across tests/ for the new shared app.mk path
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
print_hex(uint8_t) was an early bare-metal helper doing exactly what
hex8() in the freshly-ported hex_print.c does (two-digit hex via
putchar()). hex8() is smaller (asm cp/sbc/daa nibble trick, no LUT)
and consistent with the dec8/hex16/dec16/hex32/dec32 family.
• Replaced print_hex() calls with hex8() in examples/banked and
examples/bankedbg.
• Removed libc/stdio/print_hex.c, dropped its prototype from
sprinter.h and its entry from lib/Makefile.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
libc/mem/:
• Split bank_io.c into bank_io_w3.c (existing W3 helpers, base 0xC000,
port 0xE2) and bank_io_w1.c (new mirror through W1, base 0x4000,
port 0xA2). Two .rel files so DCE picks only the needed group:
a W3-only user pulls ~70 bytes instead of all 134. W1 variants are
`--memory tiny`-only (any other mode runs code from W1 or uses W1
for the banked-code segment, swapping it crashes).
• mem_alloc.c: add CF=err checks for mem_free_block and mem_get_page
(were silently ignored), per-function docstrings on alloc/free/
get_page/info, drop the confused "wait wrong order" comment in
mem_info. Header sprinter_mem.h gets matching per-function doc.
libc/stdio/:
• Add hex_print.c (hex8/hex16/hex32, ~26 bytes) and dec_print.c
(dec8/dec16/dec32, ~170 bytes) ported from solid-c STDLIB.ASM.
Replaces the printf("%u"/"%X") wrappers in solid_helpers.c that
dragged in the 3-5 KB printf machinery.
- hex* use the classic cp 10 / sbc 0x69 / daa nibble→ASCII trick;
hex8 self-calls for the high nibble, hex16/hex32 tail-call hex8.
- dec32 is the master routine; dec8/dec16 jump into shared entry
points (__dec_entry3 / __dec_entry5). 32-bit subtract-power-of-10
keeps the high 16 bits in HL alt (shadow set).
- DISCOVERY: ESTEX PUTCHAR ($5B) on our Sprinter build preserves
the main register set + IX but CLOBBERS the shadow set
(BC'/DE'/HL'). solid-c's original code assumed otherwise and
garbled output for values ≥ 6 digits. Fix: save/restore HL alt
around the RST 10 in _dec_emit_or_skip. Documented in
memory/estex_putchar_abi.md.
• file.c: drop stdaux/stdprn (no Sprinter printer API), change
stdin/stdout/stderr fd markers to 0/-1/-2 (positive fds clash with
ESTEX OPEN return values), add TODO header pointing at v2 buffered
FILE rewrite (see docs/TODO.md for the Solid-C reference struct).
bin/sprinter-cc:
• --memory big and --memory huge now always use crt0_banked.s (was:
only with --bank flags), matching docs/memory_modes_implemented.md.
When the user has no --bank flags, generate a tiny stub with
`const uint8_t n_banks = 0;` and assemble bank.s for _bank_pages.
Without this fix, openenv with --memory big could not see the
estex_file_handle symbol exported by crt0_banked.
examples/openenv:
• Add usage of estex_file_handle to confirm the crt0_banked startup-
info is reachable. Local extern decl — keeps the symbol out of
sprinter.h since it only exists in big/huge builds.
examples/dec_test:
• New regression test covering hex8/16/32 and dec8/16/32 across the
interesting boundary values.
.gitignore: add .kilo/ (editor session cache).
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
snark13