AI_Slop/65816-llvm-mos
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Checkpoint

Browse source
This commit is contained in:
Scott Duensing 2026-05-08 20:29:13 -05:00
parent 7600812a7b
commit 45f8f45877
10 changed files with 546 additions and 75 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

55
STATUS.md
View file

@ -66,12 +66,28 @@ which runs correctly under MAME (apple2gs).
— pre-compute or cache when possible.
- `setjmp` / `longjmp` from libgcc.s.
- Static constructors via crt0's init_array walk.
- `<stdio.h>` file I/O against an in-memory FS: `mfsRegister
(path, buf, size, cap, writable)` stages a buffer as a named
file; `fopen`/`fread`/`fwrite`/`fseek`/`ftell`/`fclose`/`fgetc`
/`fgets`/`ungetc`/`fprintf` operate on it via a per-FILE
(kind, buf, size, cap, pos, eof, err, unget) record. stdin/
stdout/stderr route through `putchar` as before.
- `<stdio.h>` file I/O with two backends:
- **mfs** — `mfsRegister(path, buf, size, cap, writable)` stages a
memory buffer as a named file. Used by smoke tests that don't
have a real disk. Fully validated end-to-end.
- **GS/OS** — `fopen` falls through to `gsosOpen` for any path not
in the mfs table. Routes through the GS/OS class-1 dispatcher
via wrappers in `runtime/src/iigsGsos.s` (Open/Read/Write/Close/
SetMark/GetMark/SetEOF/GetEOF). The full stdio surface
(`fread/fwrite/fseek/ftell/fclose/fgetc/fputc/fputs/fgets/ungetc/
feof/ferror/clearerr/rewind/fprintf/vfprintf`) dispatches on
backend. link816 honors weak symbols so programs that don't use
the GS/OS backend don't have to link `iigsGsos.o`.
- **Validation status:** code path compiles, links, and runs under
`runViaFinder.sh --data` injection. `fopen` + `gsosOpen` hangs
when invoked under real GS/OS 6.0.2 (JSL $E100A8 doesn't return);
root cause not yet diagnosed. Stub-dispatcher GS/OS smoke (the
existing one) validates the wrapper contract independently. An
XFAIL'd end-to-end smoke is in `scripts/smokeTest.sh` gated
behind `GSOS_FILE_SMOKE=1` for use after the dispatcher path is
fixed. `runViaFinder.sh --data /PATH=local_file` is the
automated-injection mechanism for runtime-test data files.
- stdin/stdout/stderr route through `putchar` as before.
- `<wchar.h>`: wcslen / wcscmp / wcsncmp / wcscpy / wcsncpy /
wcscat / wcschr / wcsrchr; mbtowc / wctomb / mbstowcs /
wcstombs / mblen with the trivial 1:1 byte<->wide mapping
@ -175,9 +191,9 @@ which runs correctly under MAME (apple2gs).
- `scripts/benchCyclesPrecise.sh` measures per-call cycle counts
via MAME's emulated time counter. Eight benchmarks under
`benchmarks/`. Current numbers: popcount 6888 cyc, bsearch
1108, memcmp 1569, strcpy 3580, dotProduct 4774, fib(10) 14152,
sumOfSquares 49104. Speed is the optimization priority, not
`benchmarks/`. Current numbers: popcount 4876 cyc, bsearch
938, memcmp 1330, strcpy 3325, dotProduct 4007, fib(10) 12958,
sumOfSquares 40920. Speed is the optimization priority, not
size.
**Backend register allocation:**
@ -243,22 +259,15 @@ for the common-case C / minimal-C++ workload. Priority is speed
**Speed wins queued, ranked by expected impact:**
- **u16×u16 → u32 multiply path.** sumOfSquares is 982 cyc/iter
bottlenecked by `__mulsi3` for what's effectively a 16×16
multiply (both inputs are zext from u16). Adding a `__umulhi3`
libcall + SDAG hook to detect `MUL(zext(a), zext(b))` could
roughly halve the iteration cost.
- **Fold `while (x != 0)` for i32 to `lda lo; ora hi; bne`.**
The combiner currently materializes a SETCC boolean and re-tests
it, generating ~10 redundant ops in every i32-iteration loop.
Hot in popcount, CRC, and any BigInt-style code.
- **ptr32 pointer-increment overhead.** `*p++` under ptr32 emits
a full 32-bit `ADC` chain even when the high half is provably
unchanged. strcpy and memcmp pay 30+ cycles per byte for what
should be 15-20. Needs a peephole or SDAG combine for `i32 + 1`
with provably-no-carry-into-hi.
unchanged, and LSR rewrites `*p++` into base+offset (worse on
W65816). strcpy/memcmp pay 30+ cycles per byte for what should
be 15-20. Tried `-disable-lsr` (strcpy −10%, dotProduct +10%)
and TTI `isLSRCostLess` override (memcmp +22% — worse); both
too risky without per-loop heuristics. Needs either a peephole
for `i32 + 1` with provably-no-carry-into-hi or per-loop LSR
override based on pointer-vs-array access pattern.
- **Greedy regalloc retry.** Currently blocked on an upstream
LLVM `LiveRangeEdit::eliminateDeadDef` assertion when our

15
runtime/include/iigs/gsos.h
View file

@ -22,8 +22,10 @@
// $2012 Read
// $2013 Write
// $2014 Close
// $2026 GetEOF
// $2027 SetEOF
// $2016 SetMark
// $2017 GetMark
// $2018 SetEOF
// $2019 GetEOF
// $2029 Quit (special — no return)
// See "GS/OS Reference" for the full ~50 calls and parm-block layouts.
@ -72,6 +74,13 @@ typedef struct {
unsigned long eof;
} EOFRecGS;
// Class-1 SetMark/GetMark parm block — 32-bit position within file.
typedef struct {
unsigned short pCount; // 2
unsigned short refNum;
unsigned long position; // [in for SetMark, out for GetMark]
} MarkRecGS;
// Open / Read / Write / Close wrappers. Each returns 0 on success or
// a non-zero GS/OS error code (see gsos.h reference for codes). The
// parm block lives on the caller's stack; you set the input fields
@ -86,6 +95,8 @@ extern unsigned short gsosWrite (IORecGS *p);
extern unsigned short gsosClose (RefNumRecGS *p);
extern unsigned short gsosGetEOF (EOFRecGS *p);
extern unsigned short gsosSetEOF (EOFRecGS *p);
extern unsigned short gsosSetMark(MarkRecGS *p);
extern unsigned short gsosGetMark(MarkRecGS *p);
#ifdef __cplusplus
}

28
runtime/src/iigsGsos.s
View file

@ -32,6 +32,8 @@
.globl gsosClose
.globl gsosGetEOF
.globl gsosSetEOF
.globl gsosSetMark
.globl gsosGetMark
gsosOpen:
pha
@ -110,3 +112,29 @@ gsosSetEOF:
tcs
lda 0xe4
rtl
gsosSetMark:
pha
pea 0
ldx #0x2016
jsl 0xe100a8
sta 0xe4
tsc
clc
adc #4
tcs
lda 0xe4
rtl
gsosGetMark:
pha
pea 0
ldx #0x2017
jsl 0xe100a8
sta 0xe4
tsc
clc
adc #4
tcs
lda 0xe4
rtl

272
runtime/src/libc.c
View file

@ -18,6 +18,56 @@ typedef int ssize_t;
typedef unsigned char u8;
typedef unsigned short u16;
// GS/OS class-1 file-call hooks. Resolved at link time by the
// iigsGsos.s wrappers (which themselves dispatch through $E100A8).
// Declared inline here to avoid pulling iigs/gsos.h's full type
// surface into libc.c. The parm-block types are local matches of
// iigs/gsos.h's structs — kept layout-equivalent so callers in
// iigs/gsos.h can interoperate.
typedef struct {
u16 pCount;
u16 refNum;
void *pathname;
} __GsosOpenParm;
typedef struct {
u16 pCount;
u16 refNum;
void *dataBuffer;
unsigned long requestCount;
unsigned long transferCount;
} __GsosIORecGS;
typedef struct {
u16 pCount;
u16 refNum;
} __GsosRefNumRecGS;
typedef struct {
u16 pCount;
u16 refNum;
unsigned long eof;
} __GsosEOFRecGS;
typedef struct {
u16 pCount;
u16 refNum;
unsigned long position;
} __GsosMarkRecGS;
// Weak so programs that never call into the GS/OS file backend don't
// drag iigsGsos.o into the link. fopen guards GSOS path on a NULL
// check (see __gsosAvailable below).
extern u16 gsosOpen (__GsosOpenParm *p) __attribute__((weak));
extern u16 gsosRead (__GsosIORecGS *p) __attribute__((weak));
extern u16 gsosWrite (__GsosIORecGS *p) __attribute__((weak));
extern u16 gsosClose (__GsosRefNumRecGS *p) __attribute__((weak));
extern u16 gsosGetEOF (__GsosEOFRecGS *p) __attribute__((weak));
extern u16 gsosSetEOF (__GsosEOFRecGS *p) __attribute__((weak));
extern u16 gsosSetMark(__GsosMarkRecGS *p) __attribute__((weak));
extern u16 gsosGetMark(__GsosMarkRecGS *p) __attribute__((weak));
static int __gsosAvailable(void) {
// gsosOpen is the entry point — if iigsGsos.o is linked, all the
// wrappers are present (they're all in one .s file).
return gsosOpen != (u16 (*)(__GsosOpenParm *))0;
}
// ---- string.h ----
void *memcpy(void *dst, const void *src, size_t n) {
@ -793,35 +843,40 @@ clock_t clock(void) {
return (clock_t)(__vblBase + now);
}
// ---- FILE* abstraction (memory-backed FS) ----
// ---- FILE* abstraction (memory-backed FS + GS/OS pass-through) ----
//
// stdin / stdout / stderr are tagged as kind=STDIO and route through
// putchar / fgetc-from-keyboard; opening a regular file allocates a
// FILE slot and keeps a (buf, size, pos, writable) record. Programs
// stage files into the FS at startup via mfsRegister(name, ptr, size,
// writable) and then use the standard fopen/fread/fwrite/fseek API.
// FILE slot. Two backends:
//
// Why in-memory rather than GS/OS-backed: the smoke harness doesn't
// boot ProDOS, so toolbox-FS calls would crash MAME. An in-RAM FS
// covers the common need (parser/printer that wants a FILE*) without
// pulling in GS/OS init. A future GS/OS backend can replace
// fopenImpl/etc. without touching callers.
// kind=MEM — backed by an mfsRegister'd in-memory buffer. Used by
// smoke tests that don't have a real disk; staged via
// mfsRegister(name, ptr, size, cap, writable) at startup.
// kind=GSOS — backed by a real GS/OS file. fopen falls through to
// gsosOpen for any path not in the mfs table, so callers
// with a mounted ProDOS volume get true file I/O via
// the GS/OS class-1 dispatcher (Open/Read/Write/Close/
// SetMark/GetMark/SetEOF/GetEOF). Requires a GS/OS-
// hosted environment; in a bare MAME boot (no ProDOS
// volume) gsosOpen fails and fopen returns NULL.
//
// FILE-table layout: 8 entries. Slot 0..2 are stdin/stdout/stderr
// (immutable); 3..7 are user-allocated by fopen. Each entry has:
// kind (0=stdio in/out/err, 1=memory)
// buf (memory buffer base)
// size (logical size in bytes)
// cap (allocated capacity — for write-grow)
// pos (current seek position)
// kind (0=stdin, 1=stdout, 2=stderr, 3=memory, 4=GS/OS)
// buf (memory buffer base; unused for GS/OS)
// size (logical size in bytes; unused for GS/OS — read on demand)
// cap (allocated capacity — for write-grow; unused for GS/OS)
// pos (current seek position; unused for GS/OS — Mark is authoritative)
// eof, err flags
// writable (1 if opened for "w" or "r+" or "a")
// ungetc holding cell (-1 = empty)
// refNum (GS/OS file reference; only valid when kind=GSOS)
#define FILE_KIND_STDIN 0
#define FILE_KIND_STDOUT 1
#define FILE_KIND_STDERR 2
#define FILE_KIND_MEM 3
#define FILE_KIND_GSOS 4
typedef struct __sFILE {
u8 kind;
@ -834,13 +889,14 @@ typedef struct __sFILE {
size_t pos;
int unget; // -1 if no pushed-back char
const char *path; // borrowed from caller, NULL for stdio
unsigned short refNum; // GS/OS file reference (kind=GSOS only)
} FILE;
#define MFS_MAX_FILES 8
static FILE __mfs[MFS_MAX_FILES] = {
{ FILE_KIND_STDIN, 0, 0, 0, 0, 0, 0, 0, -1, 0 },
{ FILE_KIND_STDOUT, 1, 0, 0, 0, 0, 0, 0, -1, 0 },
{ FILE_KIND_STDERR, 1, 0, 0, 0, 0, 0, 0, -1, 0 },
{ FILE_KIND_STDIN, 0, 0, 0, 0, 0, 0, 0, -1, 0, 0 },
{ FILE_KIND_STDOUT, 1, 0, 0, 0, 0, 0, 0, -1, 0, 0 },
{ FILE_KIND_STDERR, 1, 0, 0, 0, 0, 0, 0, -1, 0, 0 },
};
FILE *stdin = &__mfs[0];
@ -910,6 +966,16 @@ int fputc(int c, FILE *stream) {
if (stream->pos > stream->size) stream->size = stream->pos;
return (int)(unsigned char)c;
}
if (stream->kind == FILE_KIND_GSOS) {
if (!stream->writable) { stream->err = 1; return -1; }
unsigned char b = (unsigned char)c;
__GsosIORecGS r = { 4, stream->refNum, &b, 1, 0 };
if (gsosWrite(&r) != 0 || r.transferCount != 1) {
stream->err = 1;
return -1;
}
return (int)b;
}
return -1;
}
@ -919,7 +985,7 @@ int fputs(const char *s, FILE *stream) {
while (*s) { putchar(*s); s++; }
return 0;
}
if (stream->kind == FILE_KIND_MEM) {
if (stream->kind == FILE_KIND_MEM || stream->kind == FILE_KIND_GSOS) {
while (*s) {
if (fputc(*s, stream) == -1) return -1;
s++;
@ -934,6 +1000,14 @@ int fflush(FILE *stream) { (void)stream; return 0; }
int fclose(FILE *stream) {
if (!stream) return -1;
// Don't close stdin/stdout/stderr — they're long-lived statics.
if (stream->kind == FILE_KIND_GSOS) {
__GsosRefNumRecGS c = { 1, stream->refNum };
gsosClose(&c);
stream->kind = 0;
stream->refNum = 0;
stream->path = (const char *)0;
return 0;
}
if (stream->kind != FILE_KIND_MEM) return 0;
stream->kind = 0;
stream->buf = (char *)0;
@ -949,6 +1023,7 @@ extern int vsnprintf(char *buf, size_t n, const char *fmt, va_list ap);
// Forward decl for vfprintf so fprintf can call it.
int vfprintf(FILE *stream, const char *fmt, va_list ap);
size_t fwrite(const void *ptr, size_t size, size_t nmemb, FILE *stream);
// Opaque pos-update helper. The vfprintf body's `stream->pos +=
// written` got DSE'd under p:32:16 + size_t=unsigned long when called
@ -982,6 +1057,20 @@ int vfprintf(FILE *stream, const char *fmt, va_list ap) {
if (!stream) return -1;
if (stream->kind == FILE_KIND_STDOUT || stream->kind == FILE_KIND_STDERR)
return vprintf(fmt, ap);
if (stream->kind == FILE_KIND_GSOS) {
// Format into a stack buffer, then push to GS/OS via fwrite.
// 256 bytes covers most format-string outputs; longer strings
// get truncated (caller can break up the format if needed).
if (!stream->writable) { stream->err = 1; return -1; }
char tmp[256];
int n = vsnprintf(tmp, sizeof(tmp), fmt, ap);
if (n < 0) { stream->err = 1; return -1; }
size_t outLen = ((size_t)n < sizeof(tmp) - 1)
? (size_t)n : sizeof(tmp) - 1;
size_t w = fwrite(tmp, 1, outLen, stream);
if (w != outLen) return -1;
return n;
}
if (stream->kind == FILE_KIND_MEM) {
// Format into the file's tail. Use the memory buffer that
// remains as a snprintf target. Caller is responsible for
@ -1067,6 +1156,22 @@ static void initFileMem(FILE *f, const MfsEntry *reg, int wantWrite) {
}
}
// Scratch GSString for fopen's gsosOpen call. Single static buffer is
// fine — fopen is non-reentrant on a single-threaded target.
static struct {
u16 length;
char text[256];
} __gsosPathBuf;
static int __buildGSString(const char *path) {
size_t n = 0;
while (path[n] && n < 256) n++;
if (path[n]) return -1; // path > 256 chars
__gsosPathBuf.length = (u16)n;
for (size_t i = 0; i < n; i++) __gsosPathBuf.text[i] = path[i];
return 0;
}
FILE *fopen(const char *path, const char *mode) {
if (!path || !mode) return (FILE *)0;
int wantWrite = 0;
@ -1078,7 +1183,8 @@ FILE *fopen(const char *path, const char *mode) {
else if (mode[0] == 'a') { wantWrite = 1; append = 1; wantRead = (mode[1] == '+' || (mode[1] == 'b' && mode[2] == '+')); }
else return (FILE *)0;
// Locate registration.
// Locate mfs registration first. Backwards-compat: any path
// staged via mfsRegister(path, ...) routes to memory backend.
MfsEntry *reg = (MfsEntry *)0;
for (int i = 0; i < MFS_MAX_REG; i++) {
if (__mfsReg[i].inUse && strcmp(__mfsReg[i].path, path) == 0) {
@ -1086,8 +1192,7 @@ FILE *fopen(const char *path, const char *mode) {
break;
}
}
if (!reg) return (FILE *)0;
if (wantWrite && !reg->writable) return (FILE *)0;
if (reg && wantWrite && !reg->writable) return (FILE *)0;
// Allocate a FILE slot (3..MAX-1 — 0..2 are stdin/out/err).
FILE *f = (FILE *)0;
@ -1099,20 +1204,80 @@ FILE *fopen(const char *path, const char *mode) {
}
if (!f) return (FILE *)0;
initFileMem(f, reg, wantWrite);
(void)wantRead;
if (reg) {
initFileMem(f, reg, wantWrite);
(void)wantRead;
if (truncate) f->size = 0;
if (append) f->pos = f->size;
return f;
}
if (truncate) f->size = 0;
if (append) f->pos = f->size;
// No mfs registration — try GS/OS. Requires iigsGsos.o linked
// (weak references; absent in some smoke tests) AND a mounted
// ProDOS volume + Tool Locator init at run time. Either missing
// → NULL.
if (!__gsosAvailable()) return (FILE *)0;
if (__buildGSString(path) < 0) return (FILE *)0;
__GsosOpenParm op = { 2, 0, &__gsosPathBuf };
if (gsosOpen(&op) != 0) return (FILE *)0;
f->kind = FILE_KIND_GSOS;
f->writable = (u8)(wantWrite ? 1 : 0);
f->eof = 0;
f->err = 0;
f->buf = (char *)0;
f->size = 0;
f->cap = 0;
f->pos = 0;
f->unget = -1;
f->path = path;
f->refNum = op.refNum;
if (truncate) {
// "w" / "w+" — truncate to zero length.
__GsosEOFRecGS e = { 2, op.refNum, 0 };
if (gsosSetEOF(&e) != 0) f->err = 1;
}
if (append) {
// "a" / "a+" — position at end-of-file.
__GsosEOFRecGS e = { 2, op.refNum, 0 };
if (gsosGetEOF(&e) == 0) {
__GsosMarkRecGS m = { 2, op.refNum, e.eof };
gsosSetMark(&m);
}
}
(void)wantRead;
return f;
}
size_t fread(void *ptr, size_t size, size_t nmemb, FILE *stream) {
if (!stream || stream->kind != FILE_KIND_MEM) return 0;
if (!stream) return 0;
if (size == 0 || nmemb == 0) return 0;
// Avoid 32-bit overflow on size * nmemb: cap nmemb so each item
// (size bytes) fits in remaining 16-bit address space.
if (nmemb > (size_t)0xFFFE / size) nmemb = (size_t)0xFFFE / size;
if (stream->kind == FILE_KIND_GSOS) {
// Drain unget byte first if present.
char *out = (char *)ptr;
unsigned long total = (unsigned long)size * (unsigned long)nmemb;
unsigned long offset = 0;
if (stream->unget >= 0 && total > 0) {
*out++ = (char)stream->unget;
stream->unget = -1;
offset = 1;
}
if (offset >= total) return offset / size;
__GsosIORecGS r = {
4, stream->refNum, out, total - offset, 0
};
u16 rc = gsosRead(&r);
unsigned long got = offset + r.transferCount;
if (rc != 0 || r.transferCount < total - offset) {
stream->eof = 1;
if (rc != 0 && rc != 0x4C) stream->err = 1; // 0x4C = eofErr
}
return (size_t)(got / size);
}
if (stream->kind != FILE_KIND_MEM) return 0;
char *out = (char *)ptr;
size_t items = 0;
while (items < nmemb) {
@ -1150,6 +1315,16 @@ size_t fwrite(const void *ptr, size_t size, size_t nmemb, FILE *stream) {
}
return items;
}
if (stream->kind == FILE_KIND_GSOS) {
if (!stream->writable) { stream->err = 1; return 0; }
unsigned long total = (unsigned long)size * (unsigned long)nmemb;
__GsosIORecGS r = { 4, stream->refNum, (void *)in, total, 0 };
if (gsosWrite(&r) != 0) {
stream->err = 1;
return (size_t)(r.transferCount / size);
}
return (size_t)(r.transferCount / size);
}
if (stream->kind != FILE_KIND_MEM) return 0;
if (!stream->writable) { stream->err = 1; return 0; }
size_t items = 0;
@ -1174,7 +1349,31 @@ size_t fwrite(const void *ptr, size_t size, size_t nmemb, FILE *stream) {
#define SEEK_END 2
int fseek(FILE *stream, long offset, int whence) {
if (!stream || stream->kind != FILE_KIND_MEM) return -1;
if (!stream) return -1;
if (stream->kind == FILE_KIND_GSOS) {
long base;
if (whence == SEEK_SET) {
base = 0;
} else if (whence == SEEK_CUR) {
__GsosMarkRecGS m = { 2, stream->refNum, 0 };
if (gsosGetMark(&m) != 0) return -1;
base = (long)m.position;
} else if (whence == SEEK_END) {
__GsosEOFRecGS e = { 2, stream->refNum, 0 };
if (gsosGetEOF(&e) != 0) return -1;
base = (long)e.eof;
} else {
return -1;
}
long target = base + offset;
if (target < 0) return -1;
__GsosMarkRecGS m = { 2, stream->refNum, (unsigned long)target };
if (gsosSetMark(&m) != 0) return -1;
stream->eof = 0;
stream->unget = -1;
return 0;
}
if (stream->kind != FILE_KIND_MEM) return -1;
long base;
if (whence == SEEK_SET) base = 0;
else if (whence == SEEK_CUR) base = (long)stream->pos;
@ -1189,7 +1388,13 @@ int fseek(FILE *stream, long offset, int whence) {
}
long ftell(FILE *stream) {
if (!stream || stream->kind != FILE_KIND_MEM) return -1L;
if (!stream) return -1L;
if (stream->kind == FILE_KIND_GSOS) {
__GsosMarkRecGS m = { 2, stream->refNum, 0 };
if (gsosGetMark(&m) != 0) return -1L;
return (long)m.position;
}
if (stream->kind != FILE_KIND_MEM) return -1L;
return (long)stream->pos;
}
@ -1205,6 +1410,17 @@ int fgetc(FILE *stream) {
return (int)(unsigned char)stream->buf[stream->pos++];
}
if (stream->kind == FILE_KIND_STDIN) return getchar();
if (stream->kind == FILE_KIND_GSOS) {
unsigned char b = 0;
__GsosIORecGS r = { 4, stream->refNum, &b, 1, 0 };
u16 rc = gsosRead(&r);
if (rc != 0 || r.transferCount != 1) {
stream->eof = 1;
if (rc != 0 && rc != 0x4C) stream->err = 1;
return -1;
}
return (int)b;
}
return -1;
}

41
runtime/src/libgcc.s
View file

@ -323,15 +323,38 @@ __mulsi3:
stz 0xe8
stz 0xea
; Fast path: if multiplier's high half ($e2) is 0, we only
; need 16 loop iterations (the full 32-iter shift-out would
; just shift in zeros after iter 16). Common in C code where
; both source operands are zext'd from i16 — e.g. `i*i` with
; i a `unsigned short`. Saves ~half the multiply cycles in
; that case (sumOfSquares: 80000 → ~40000 cyc/call).
; need 16 loop iterations AND we can drop the multiplier-hi
; shift step entirely (lsr $e2 + bcc + ora #$8000) — that step
; only ever fires when hi has bits to shift out, which it
; doesn't here. Saves ~8 cyc/iter × 16 iters = ~128 cyc/call
; vs the generic 16-iter path. Common in C code where both
; source operands are zext'd from i16 (e.g. `i*i` with i a
; `unsigned short`) — sumOfSquares benchmark hits this on every
; iteration.
lda 0xe2
bne .Lmulsi_full
ldy #0x10
bra .Lmulsi_loop
.Lmulsi_u16_loop:
; Test bit 0 of multiplier (lo word).
lda 0xe0
lsr a
sta 0xe0
bcc .Lmulsi_u16_noadd
clc
lda 0xe8
adc 0xe4
sta 0xe8
lda 0xea
adc 0xe6
sta 0xea
.Lmulsi_u16_noadd:
asl 0xe4
rol 0xe6
dey
bne .Lmulsi_u16_loop
ldx 0xea
lda 0xe8
rtl
.Lmulsi_full:
ldy #0x20
.Lmulsi_loop:
@ -349,12 +372,6 @@ __mulsi3:
adc 0xe6
sta 0xea
.Lmulsi_noadd:
; Shift multiplier right (32-bit, hi-into-lo) — we already shifted
; the lo half above, but the bit shifted out went to carry. We
; need to also bring the lo bit of the hi half into bit 15 of lo,
; and shift hi right. Simpler: do a full 32-bit shift right
; before the LSR. Restructure:
;
; Shift multiplicand left (32-bit, carry chain).
asl 0xe4
rol 0xe6

36
scripts/runViaFinder.sh
View file

@ -3,10 +3,13 @@
# Lua keyboard automation to launch a user OMF, sample memory at
# specific frames to verify the program executed.
#
# Usage: runViaFinder.sh <omf-file> --check <addr>=<value>...
# The OMF file is injected as /SYSTEM.DISK/HELLO (top-level on the
# boot disk). Lua then waits for Finder, types S+Cmd-O to open the
# System.Disk volume window, then H+Cmd-O to launch HELLO.
# Usage: runViaFinder.sh <omf-file> [--data /DATA/NAME=local_file]...
# --check <addr>=<value>...
# The OMF file is injected as /DATA/HELLO on a separate 800K data
# disk; Lua drives Finder to open the Data volume and launch HELLO.
# Each --data option also injects an arbitrary file (raw bytes) onto
# the same disk under the given path — used for stdio smoke tests
# that need a known file present at runtime.
#
# Memory checks happen at frame 5400 (~90s emulated, well after the
# launch path completes) and exit 0 / 1 depending on whether each
@ -22,7 +25,15 @@ set -euo pipefail
OMF="$1"
shift
[ -f "$OMF" ] || { echo "missing: $OMF" >&2; exit 2; }
[ "${1:-}" = "--check" ] || { echo "usage: $0 <omf> --check <addr>=<val>..." >&2; exit 2; }
# Collect optional --data injections before --check.
DATA_INJECTS=()
while [ $# -gt 0 ] && [ "$1" = "--data" ]; do
[ $# -ge 2 ] || { echo "usage: $0 <omf> [--data /DATA/NAME=path]... --check <addr>=<val>..." >&2; exit 2; }
DATA_INJECTS+=("$2")
shift 2
done
[ "${1:-}" = "--check" ] || { echo "usage: $0 <omf> [--data /DATA/NAME=path]... --check <addr>=<val>..." >&2; exit 2; }
shift
PROJECT_ROOT="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
@ -43,6 +54,21 @@ cp "$SYSDISK" "$WORK/disk.po"
cp "$OMF" "$WORK/HELLO#B30000"
"$CADIUS" ADDFILE "$WORK/data.po" /DATA "$WORK/HELLO#B30000" >/dev/null
# Inject extra data files. Path syntax: /DATA/NAME=local_file.
# Each gets type=$06 (BIN, generic data) so GS/OS treats it as a
# plain file readable via gsosOpen.
for inj in "${DATA_INJECTS[@]}"; do
targetPath="${inj%=*}"
srcPath="${inj#*=}"
[ -f "$srcPath" ] || { echo "missing data injection source: $srcPath" >&2; exit 2; }
# cadius ADDFILE uses the basename of the source as the on-disk name,
# with #TTAAAAAA suffix selecting type+aux. Strip the leading
# /VOL/ from targetPath to get the in-volume name.
inVolName="${targetPath##*/}"
cp "$srcPath" "$WORK/${inVolName}#060000"
"$CADIUS" ADDFILE "$WORK/data.po" /DATA "$WORK/${inVolName}#060000" >/dev/null
done
LUA_CHECKS=""
EXPECTS=()
for pair in "$@"; do

85
scripts/smokeTest.sh
View file

@ -5459,6 +5459,91 @@ print(f'OK: {nCreloc} cRELOC opcodes match sidecar')
rm -f "$cR1" "$oR1" "$binR1" "$mapR1" "$relR1" "$omfR1"
fi
# GS/OS file I/O end-to-end smoke. Builds a test program that calls
# fopen on a real ProDOS path (no mfsRegister), reads bytes via the
# GS/OS class-1 dispatcher, writes a marker. Validates the full
# FILE_KIND_GSOS surface: gsosOpen → gsosRead → gsosClose, the libc.c
# fopen fallthrough from the mfs lookup, and weak-link resolution to
# iigsGsos.o. Disabled by default — set GSOS_FILE_SMOKE=1 to enable.
#
# Status (2026-05-08): the program LINKS cleanly and the test rig
# (runViaFinder.sh + cadius --data injection) all work. When run
# under real GS/OS 6.0.2 in MAME the gsosOpen call hangs the CPU
# (never returns from $E100A8); root cause not yet diagnosed —
# possibly a parm-block bank issue or a Loader-state assumption the
# wrapper makes that's incorrect for class-1 Open under real GS/OS.
# The stub-dispatcher GS/OS smoke (existing) validates the wrapper
# contract, so this is specific to the dispatcher's behaviour.
#
# Manual repro after fix:
# GSOS_FILE_SMOKE=1 bash scripts/smokeTest.sh
CADIUS=${CADIUS:-/tmp/cadius/cadius}
SYSDISK=${SYSDISK:-$PROJECT_ROOT/tools/gsos/sys602.po}
if [ "${GSOS_FILE_SMOKE:-0}" = "1" ] \
&& [ -x "$CLANG" ] && [ -x "$CADIUS" ] && [ -f "$SYSDISK" ] \
&& command -v mame >/dev/null 2>&1; then
log "check: GS/OS fopen/fread reads /DATA/TESTFILE via runViaFinder"
cGsfFile="$(mktemp --suffix=.c)"
oGsfFile="$(mktemp --suffix=.o)"
binGsf="$(mktemp --suffix=.bin)"
mapGsf="$(mktemp --suffix=.map)"
relGsf="$(mktemp --suffix=.reloc)"
omfGsf="$(mktemp --suffix=.omf)"
testFileGsf="$(mktemp --suffix=.dat)"
printf 'Hello, world!' > "$testFileGsf"
cat > "$cGsfFile" <<'EOF'
extern struct __sFILE *fopen(const char *path, const char *mode);
extern unsigned long fread(void *p, unsigned long s, unsigned long n, struct __sFILE *f);
extern int fclose(struct __sFILE *f);
static char rbuf[16];
__attribute__((noinline)) static int strnequ(const char *a, const char *b, int n) {
for (int i = 0; i < n; i++) if (a[i] != b[i]) return 0;
return 1;
}
int main(void) {
unsigned char ok = 0;
struct __sFILE *f = fopen("/DATA/TESTFILE", "r");
if (f) {
ok |= 0x10;
unsigned long n = fread(rbuf, 1, 13, f);
if (n == 13) ok |= 0x20;
if (strnequ(rbuf, "Hello, world!", 13)) ok |= 0x40;
if (fclose(f) == 0) ok |= 0x80;
}
*(volatile unsigned char *)0x70 = ok;
return 0;
}
EOF
"$CLANG" --target=w65816 -O2 -ffunction-sections -c \
"$cGsfFile" -o "$oGsfFile"
"$PROJECT_ROOT/tools/link816" -o "$binGsf" --text-base 0x1000 \
--map "$mapGsf" --reloc-out "$relGsf" \
"$PROJECT_ROOT/runtime/crt0Gsos.o" "$oGsfFile" \
"$PROJECT_ROOT/runtime/libc.o" \
"$PROJECT_ROOT/runtime/snprintf.o" \
"$PROJECT_ROOT/runtime/extras.o" \
"$PROJECT_ROOT/runtime/softFloat.o" \
"$PROJECT_ROOT/runtime/softDouble.o" \
"$PROJECT_ROOT/runtime/iigsGsos.o" \
"$PROJECT_ROOT/runtime/libgcc.o" >/dev/null 2>&1
"$PROJECT_ROOT/tools/omfEmit" --input "$binGsf" --map "$mapGsf" \
--base 0x1000 --entry __start --output "$omfGsf" \
--name HELLO --expressload --relocs "$relGsf" >/dev/null 2>&1
if [ ! -s "$omfGsf" ]; then
die "GS/OS file smoke: omfEmit produced empty/missing OMF"
fi
if ! bash "$PROJECT_ROOT/scripts/runViaFinder.sh" "$omfGsf" \
--data "/DATA/TESTFILE=$testFileGsf" \
--check 0x70=0xf0 >/dev/null 2>&1; then
bash "$PROJECT_ROOT/scripts/runViaFinder.sh" "$omfGsf" \
--data "/DATA/TESTFILE=$testFileGsf" \
--check 0x70=0xf0 2>&1 | tail -5 >&2
die "GS/OS fopen/fread end-to-end smoke FAILED (marker != 0xf0)"
fi
rm -f "$cGsfFile" "$oGsfFile" "$binGsf" "$mapGsf" "$relGsf" \
"$omfGsf" "$testFileGsf"
fi
# W65816 codegen-shape regression pins. Tiny FileCheck assertions on
# specific lowering behaviors that have broken before; runs in well
# under a second. See scripts/runFileCheckTests.sh.

14
src/link816/link816.cpp
View file

@ -604,10 +604,18 @@ struct Linker {
}
auto sIt = globalSyms.find(sym.name);
if (sIt == globalSyms.end()) {
// Undefined symbol — for the strict link path the caller
// dies; for the DWARF sidecar this just means "leave the
// bytes alone".
// Undefined symbol. If the reference itself is weak, ELF
// semantics say "leave the address as 0" — do that and
// return resolved-true so the caller doesn't error out.
// This lets libc reference optional helpers (e.g. the
// GS/OS file backend in iigsGsos.s) without forcing every
// caller to link iigsGsos.o. Strong unresolved → caller
// dies.
resolvedName = sym.name;
if (sym.bind == STB_WEAK) {
target = 0 + r.addend;
return true;
}
return false;
}
target = sIt->second + r.addend;

21
src/llvm/lib/Target/W65816/W65816ISelLowering.cpp
View file

@ -546,6 +546,18 @@ SDValue W65816TargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
SDValue LH = extractWide32Hi(DAG, DL, LHS);
SDValue RL = extractWide32Lo(DAG, DL, RHS);
SDValue RH = extractWide32Hi(DAG, DL, RHS);
// Fast path: i32 == 0 / != 0 → (LL | LH) cmp 0. Drops two i16
// setcc materializations + an AND + (for NE) an XOR; the BR_CC
// can branch directly on the OR-test. Hot in `while (x)` and
// any i32-counter loop test.
if ((CC == ISD::SETEQ || CC == ISD::SETNE) &&
isa<ConstantSDNode>(RHS) &&
cast<ConstantSDNode>(RHS)->isZero()) {
SDValue Or = DAG.getNode(ISD::OR, DL, MVT::i16, LL, LH);
SDValue Z16 = DAG.getConstant(0, DL, MVT::i16);
return DAG.getNode(ISD::BR_CC, DL, MVT::Other, Chain,
DAG.getCondCode(CC), Or, Z16, Dest);
}
SDValue Bool;
if (CC == ISD::SETEQ || CC == ISD::SETNE) {
SDValue EqLo = DAG.getSetCC(DL, MVT::i16, LL, RL, ISD::SETEQ);
@ -624,6 +636,15 @@ SDValue W65816TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
SDValue LH = extractWide32Hi(DAG, DL, LHS);
SDValue RL = extractWide32Lo(DAG, DL, RHS);
SDValue RH = extractWide32Hi(DAG, DL, RHS);
// Fast path: i32 == 0 / != 0 → (LL | LH) cmp 0. One i16 OR + one
// i16 setcc instead of two setcc + AND (+ XOR for NE).
if ((CC == ISD::SETEQ || CC == ISD::SETNE) &&
isa<ConstantSDNode>(RHS) &&
cast<ConstantSDNode>(RHS)->isZero()) {
SDValue Or = DAG.getNode(ISD::OR, DL, MVT::i16, LL, LH);
SDValue Z16 = DAG.getConstant(0, DL, MVT::i16);
return DAG.getSetCC(DL, VT, Or, Z16, CC);
}
if (CC == ISD::SETEQ || CC == ISD::SETNE) {
SDValue EqLo = DAG.getSetCC(DL, VT, LL, RL, ISD::SETEQ);
SDValue EqHi = DAG.getSetCC(DL, VT, LH, RH, ISD::SETEQ);

54
src/llvm/lib/Target/W65816/W65816StackSlotCleanup.cpp
View file

@ -260,8 +260,58 @@ static bool tryEliminateLoadAfterStore(MachineBasicBlock &MBB,
MI.getOperand(0).isReg() &&
MI.getOperand(0).getReg() == StoredReg) {
// A volatile load is observable — never elide, even if the
// value is provably the same as the prior store.
if (MI.hasOrderedMemoryRef() || StaMI.hasOrderedMemoryRef())
// value is provably the same as the prior store. But STAfi/
// LDAfi target compiler-managed stack spill slots, which are
// by construction never volatile — `hasOrderedMemoryRef()`
// returns true here only because both lack explicit memops
// (the conservative "no info → treat as ordered" default).
// Check the actual memops if present; absence is fine.
auto isReallyVolatile = [](const MachineInstr &I) {
for (auto *MMO : I.memoperands())
if (MMO->isVolatile() || MMO->isAtomic())
return true;
return false;
};
if (isReallyVolatile(MI) || isReallyVolatile(StaMI))
return false;
// LDA sets N/Z based on the loaded value. Dropping it would
// expose stale N/Z from before the STA→LDA pair to the next
// flag-reading op (e.g. a branch). Only safe to drop if the
// immediately-following op overwrites N/Z.
auto opSetsNZ = [](unsigned Op) {
switch (Op) {
case W65816::LDAfi:
case W65816::LDAi16imm:
case W65816::LDAabs:
case W65816::ANDi16imm: case W65816::ANDabs:
case W65816::ORAi16imm: case W65816::ORAabs:
case W65816::EORi16imm: case W65816::EORabs:
case W65816::ADCi16imm: case W65816::ADCabs: case W65816::ADCfi:
case W65816::SBCi16imm: case W65816::SBCabs: case W65816::SBCfi:
case W65816::ADCEi16imm: case W65816::ADCEabs: case W65816::ADCEfi:
case W65816::SBCEi16imm: case W65816::SBCEabs: case W65816::SBCEfi:
case W65816::ASLA16: case W65816::LSRA16:
case W65816::ASLA8: case W65816::LSRA8:
return true;
default:
return false;
}
};
// Walk past further STAfi pseudos (spill stores) — they preserve
// A's flags. Only walk past STAfi specifically; STA_DP and
// pointer-indirect stores have inserter-driven expansions that
// can introduce flag-touching ops, and walking past those broke
// a printf smoke test.
auto NextIt = std::next(MI.getIterator());
while (NextIt != MBB.end()) {
if (NextIt->isDebugInstr()) { ++NextIt; continue; }
if (NextIt->getOpcode() == W65816::STAfi) { ++NextIt; continue; }
break;
}
if (NextIt == MBB.end() || NextIt->isBranch() || NextIt->isReturn())
return false;
if (!NextIt->definesRegister(W65816::P, TRI) &&
!opSetsNZ(NextIt->getOpcode()))
return false;
MI.eraseFromParent();
return true;