merge work/actors: wasm seed bootstrap (seed/MANIFEST), shared copy-on-write actor memory, top-level hold blocks on all backends and lsp

3 files changed, 344 insertions, 43 deletions

diff --git a/src/core/core.c b/src/core/core.c
index ce9e5ca..7d1b91e 100644
--- a/src/core/core.c
+++ b/src/core/core.c
@@ -74,6 +74,10 @@
 #include "memmap.h"
 #include "fops.h"
 #include "bops.h"
+#if !defined(_WIN32) && !defined(__wasm__)
+#define HOLD_THREADS
+#endif
+#include "hold.h"
 
 #ifndef JTYPE
 #define JTYPE char
@@ -256,6 +260,7 @@ static atomic_int context_records_live, context_code_live, context_memory_live;
 size_t io_context_size(void);
 void io_context_snapshot(void *destination);
 void io_context_restore(const void *source);
+void io_context_release(void);
 #else
 static VM_TLS int context_default_stop;
 #endif
@@ -294,6 +299,12 @@ static Cell hostcall(Cell op, Cell a, Cell b, Cell c)
 #else
   return -1;
 #endif
+ case 57: {
+  Cell r = hold_enter(m, (uint64_t)MW, a, b);
+  if (r == -2) { fflush(stdout); fprintf(stderr, "zy: hold block %d was entered while block %d was held; hold blocks are entered in the order they are declared\n", (int)a, hold_top); exit(2); }
+  return r;
+ }
+ case 58: return hold_leave(m, (uint64_t)MW, a, b);
  case 56: return cb_r;	/* the return stack pointer as of the gate store: the word after the top return address */
 	case 1:
 		if (zcstr(a, path, 1024) < 0) return -1;
@@ -622,11 +633,44 @@ typedef struct {
 	Uop *u;
 	signed char *reg;	/* physical register per value, for the jit; 0 if it didn't fit */
 } Enh;
-static VM_TLS Enh enh[NENH];
+/* large per-thread tables live behind a pointer and are mapped on first use, so a
+   worker thread that never compiles anything pays nothing for them (as static tls
+   they were zeroed into every thread's stack mapping: about 13 MB per actor) */
+static VM_TLS void **tls_big_slots[48]; static VM_TLS size_t tls_big_sizes[48]; static VM_TLS int tls_big_n;
+static void *tls_big(size_t bytes, void **slot)
+{
+	void *p;
+#if !defined(_WIN32) && !defined(__wasm__)
+	p = mmap(0, bytes, PROT_READ | PROT_WRITE, MAP_PRIVATE | MAP_ANON, -1, 0); if (p == MAP_FAILED) p = 0;
+#else
+	p = calloc(1, bytes);
+#endif
+	if (!p) { fprintf(stderr, "zy: out of memory for compiler tables\n"); exit(2); }
+	if (tls_big_n < 48) { tls_big_slots[tls_big_n] = slot; tls_big_sizes[tls_big_n++] = bytes; }
+	*slot = p; return p;
+}
+static void tls_big_free(void)
+{
+	while (tls_big_n > 0) {
+		tls_big_n--;
+#if !defined(_WIN32) && !defined(__wasm__)
+		munmap(*tls_big_slots[tls_big_n], tls_big_sizes[tls_big_n]);
+#else
+		free(*tls_big_slots[tls_big_n]);
+#endif
+		*tls_big_slots[tls_big_n] = 0;
+	}
+}
+#define TLS_BIG(p) (*((p) ? (p) : (__typeof__(p))tls_big(sizeof *(p), (void **)&(p))))
+static VM_TLS Enh (*enh_tls)[NENH];
+#define enh TLS_BIG(enh_tls)
+
 static VM_TLS int need_flush;
 
 #define MAXUOP 4096
-static VM_TLS Uop ub[MAXUOP]; static VM_TLS int un, unv, uerr;	/* the program being built */
+static VM_TLS Uop (*ub_tls)[MAXUOP];
+#define ub TLS_BIG(ub_tls)
+static VM_TLS int un, unv, uerr;	/* the program being built */
 static int unew(int op, int a, int b, int c, Cell imm)
 {
 	Uop *u;
@@ -838,7 +882,9 @@ static void umagic(UCell d, UCell *mul, int *add, int *sh)	/* hacker's delight,
 	*mul = q2 + 1; *sh = p - 64;
 }
 
-static VM_TLS Uop ob[MAXUOP]; static VM_TLS int on, onv;
+static VM_TLS Uop (*ob_tls)[MAXUOP];
+#define ob TLS_BIG(ob_tls)
+static VM_TLS int on, onv;
 static VM_TLS int *omap, *isk; static VM_TLS Cell *kv;	/* old value -> new value; constness of new values */
 static int pure(int op) { return !is_memop(op); }
 static int ispow2(UCell x) { return x && !(x & (x - 1)); }
@@ -939,8 +985,8 @@ static int oe(int op, int a, int b, int c, Cell imm)
 	return raw(op, a, b, c, 0);
 }
 static void optimizer_storage(void) {
- static VM_TLS int omap_[MAXUOP + 16], isk_[MAXUOP * 4]; static VM_TLS Cell kv_[MAXUOP * 4];
- omap = omap_; isk = isk_; kv = kv_;
+ static VM_TLS int (*omap_)[MAXUOP + 16], (*isk_)[MAXUOP * 4]; static VM_TLS Cell (*kv_)[MAXUOP * 4];
+ omap = TLS_BIG(omap_); isk = TLS_BIG(isk_); kv = TLS_BIG(kv_);
 }
 /* runs the optimizer over ub[0..un) with nin inputs; leaves the result in e */
 static int optimize(Enh *e, int nin, int *out, int nout)
@@ -1121,11 +1167,19 @@ typedef struct {
 	unsigned gen;			/* bumped every time the region's code is dropped: a copy made from an older one is stale */
 } Rg;
 #define MAXRG 4096
-static VM_TLS Rg *rgs[MAXRG];
+static VM_TLS Rg * (*rgs_tls)[MAXRG];
+#define rgs TLS_BIG(rgs_tls)
+
 static VM_TLS int nrg;
-static VM_TLS int fn_pc[MAXRG];
-static VM_TLS Rg *fn_rg_[MAXRG];
-static VM_TLS Rg *fn_cache[1024];
+static VM_TLS int (*fn_pc_tls)[MAXRG];
+#define fn_pc TLS_BIG(fn_pc_tls)
+
+static VM_TLS Rg * (*fn_rg__tls)[MAXRG];
+#define fn_rg_ TLS_BIG(fn_rg__tls)
+
+static VM_TLS Rg * (*fn_cache_tls)[1024];
+#define fn_cache TLS_BIG(fn_cache_tls)
+
 static VM_TLS int nfn;
 
 /* ---- bookkeeping for translated code, shared by the backends ----
@@ -1142,12 +1196,18 @@ static VM_TLS int nfn;
 #define NPG ((code_words >> PGSH) + 1)
 #define MAXPGN (1 << 19)
 #define MAXDEP (1 << 18)
-static VM_TLS struct { int a, b; } ix_blk[MAXB];
+static VM_TLS struct { int a, b; } (*ix_blk_tls)[MAXB];
+#define ix_blk TLS_BIG(ix_blk_tls)
+
 static VM_TLS int ix_n;
 static VM_TLS int *pg_head, pg_free, pg_top, pg_live;		/* node numbers are stored plus one: 0 is none */
-static VM_TLS struct { int blk, next; } pg_node[MAXPGN];
+static VM_TLS struct { int blk, next; } (*pg_node_tls)[MAXPGN];
+#define pg_node TLS_BIG(pg_node_tls)
+
 static VM_TLS int dep_head[NENH + 1], dep_n;				/* slot NENH: depends on every install */
-static VM_TLS struct { int blk, next; } dep_node[MAXDEP];
+static VM_TLS struct { int blk, next; } (*dep_node_tls)[MAXDEP];
+#define dep_node TLS_BIG(dep_node_tls)
+
 static VM_TLS int drop_epoch;						/* changes whenever any translation dies */
 static void be_drop(int j);					/* the backend forgets block j's code and coverage */
 static void runmark(Rg *R);					/* the backend stops watching the region's words */
@@ -1770,7 +1830,9 @@ static Cell ffi_cbmake(Cell entry, Cell sig)
 /* a native call the o3 lifter has resolved: the micro-ops narg and ncall carry
    an index into this table. imm of narg is the argument's position */
 typedef struct { void *f; Cell sig; int n; char name[48]; } Ncd;
-static VM_TLS Ncd ncd[256];
+static VM_TLS Ncd (*ncd_tls)[256];
+#define ncd TLS_BIG(ncd_tls)
+
 static VM_TLS int nncd;
 static int ncd_get(Cell blk)
 {
@@ -1878,8 +1940,12 @@ static VM_TLS int lv_[RMAXV], lw_[RMAXV];
 #define SB 1100
 static VM_TLS int skv[2 * SB], stp, slo, shi;
 static VM_TLS unsigned char skk[2 * SB];
-static VM_TLS int afp[MAXUOP * 4 + 16];	/* value -> 1 + its offset from fp, if it is fp + k */
-static VM_TLS int emap[MAXUOP * 4 + 16];
+static VM_TLS int (*afp_tls)[MAXUOP * 4 + 16];
+#define afp TLS_BIG(afp_tls)
+/* value -> 1 + its offset from fp, if it is fp + k */
+static VM_TLS int (*emap_tls)[MAXUOP * 4 + 16];
+#define emap TLS_BIG(emap_tls)
+
 
 static int rvar(int v)
 {
@@ -2360,7 +2426,13 @@ enum { C_UNSEEN, C_CONST, C_VAR, C_COPY };
 /* C_COPY: the variable holds the same value as variable v (both at the same point) */
 static int copy_ok(const Rg *R, int j, int k) { return j > 0 && j < RMAXV && k > 0 && k < R->nv && !R->fv[j] && !R->fv[k]; }
 typedef struct { unsigned char st[RMAXV]; Cell v[RMAXV]; } CIn;
-static VM_TLS int g_isk[MAXUOP * 4]; static VM_TLS Cell g_kv[MAXUOP * 4]; static VM_TLS int g_omap[MAXUOP + 16];
+static VM_TLS int (*g_isk_tls)[MAXUOP * 4];
+#define g_isk TLS_BIG(g_isk_tls)
+static VM_TLS Cell (*g_kv_tls)[MAXUOP * 4];
+#define g_kv TLS_BIG(g_kv_tls)
+static VM_TLS int (*g_omap_tls)[MAXUOP + 16];
+#define g_omap TLS_BIG(g_omap_tls)
+
 static int cjoin(CIn *d, const CIn *s)
 {
 	int j, ch = 0;
@@ -3313,7 +3385,9 @@ static VM_TLS void **tab;		/* vm address -> translation */
 typedef struct { void *stub; int fslots; int argc; } FnNat;
 static VM_TLS FnNat *fn_nat;
 #define LINKN 1024
-static VM_TLS void *linkstk[LINKN];
+static VM_TLS void * (*linkstk_tls)[LINKN];
+#define linkstk TLS_BIG(linkstk_tls)
+
 static VM_TLS Cell jit_depth;
 #endif
 static void be_drop(int j) { tab[ix_blk[j].a] = 0; cdec(cov, ix_blk[j].a, ix_blk[j].b); if (trace) fprintf(stderr, "jit: dropped %d..%d\n", ix_blk[j].a, ix_blk[j].b - 1); }
@@ -3963,7 +4037,7 @@ static int spec_enh(int s, Cell *ks, int nk, Enh *te)
 {
 	Enh *e = &enh[s];
 	int nin = e->nin - nk, i, j, out[8], na;
-	static VM_TLS int mp[MAXUOP * 4 + 16];
+	static VM_TLS int (*mp_tls)[MAXUOP * 4 + 16]; int *mp = TLS_BIG(mp_tls);
 	if (e->nv > MAXUOP * 4) return -1;
 	un = 0; unv = nin; uerr = 0;
 	for (j = 0; j < nin; j++) mp[j] = j;
@@ -5237,9 +5311,13 @@ static Blk lcompile(int p)
 #if !defined(ZYLLVM) && !defined(ZYJIT_X64)
 static VM_TLS unsigned char *rcov;	/* 1 on the words of a lifted region */
 #endif
-static VM_TLS Cell rvals[MAXUOP * 4 + 16];
+static VM_TLS Cell (*rvals_tls)[MAXUOP * 4 + 16];
+#define rvals TLS_BIG(rvals_tls)
+
 #define RPOOL 40
-static VM_TLS Cell rpool[RPOOL][MAXUOP * 4 + 16];
+static VM_TLS Cell (*rpool_tls)[RPOOL][MAXUOP * 4 + 16];
+#define rpool TLS_BIG(rpool_tls)
+
 static VM_TLS int rpool_n;
 static int rdrop(int a)		/* a store landed on region code: drop those regions */
 {
@@ -6090,7 +6168,11 @@ static void float_op(int i, int k)
 static const int xreg[] = { r_si, r_di, r_8, r_9, r_10, r_11, r_bp };
 #define XNREG 7
 #define XMAXOP 400
-static VM_TLS int xkc[MAXUOP * 4 + 16]; static VM_TLS Cell xkv[MAXUOP * 4 + 16];
+static VM_TLS int (*xkc_tls)[MAXUOP * 4 + 16];
+#define xkc TLS_BIG(xkc_tls)
+static VM_TLS Cell (*xkv_tls)[MAXUOP * 4 + 16];
+#define xkv TLS_BIG(xkv_tls)
+
 static VM_TLS signed char *xr;
 static int x64_op_ok(int op)
 {
@@ -6110,7 +6192,7 @@ static int x64_spec_enh(int s, Cell *ks, int nk, Enh *te)
 {
 	Enh *e = &enh[s];
 	int nin = e->nin - nk, i, j, out[8], na;
-	static VM_TLS int mp[MAXUOP * 4 + 16];
+	static VM_TLS int (*mp_tls)[MAXUOP * 4 + 16]; int *mp = TLS_BIG(mp_tls);
 	if (e->nv > MAXUOP * 4 || e->n > XMAXOP) return -1;
 	for (i = 0; i < e->n; i++) if (!x64_op_ok(e->u[i].op)) return -1;
 	un = 0; unv = nin; uerr = 0;
@@ -6131,7 +6213,9 @@ static int x64_spec_enh(int s, Cell *ks, int nk, Enh *te)
 	return 0;
 }
 /* the register value v lives in, or scratch with the constant v loaded into it */
-static VM_TLS int xvh[MAXUOP * 4 + 16];	/* regions: 1 + the var a value reads at block entry, 0 if none */
+static VM_TLS int (*xvh_tls)[MAXUOP * 4 + 16];
+#define xvh TLS_BIG(xvh_tls)
+/* regions: 1 + the var a value reads at block entry, 0 if none */
 static VM_TLS Rg *xR;
 static VM_TLS Cell xsyn[RMAXV];	/* the home of a var the lifter made (a hoisted value), which has no frame slot */
 static int xv(int v, int scratch)
@@ -6250,7 +6334,9 @@ static void x64_enh(Enh *e)
    under ZYO3INTERP=1. */
 static VM_TLS int x64_o3interp;
 static void perfmap_note(JW *start, JW *end, const char *kind, int a, int b);
-static VM_TLS signed char rxr[MAXUOP * 4 + 16];
+static VM_TLS signed char (*rxr_tls)[MAXUOP * 4 + 16];
+#define rxr TLS_BIG(rxr_tls)
+
 #define RXNREG 6
 static void rsave(void) { int j; for (j = 0; j < RXNREG; j++) push_r(xreg[j]); }
 static void rrest(void) { int j; for (j = RXNREG - 1; j >= 0; j--) pop_r(xreg[j]); }
@@ -6258,7 +6344,7 @@ static Cell xev(long op, Cell a, Cell b, Cell c) { return ev((int)op, a, b, c);
 /* a value's last use, registers on the six, -1 if the block doesn't fit or can't be compiled */
 static int x64_ralloc(RB *B)
 {
-	static VM_TLS int last[MAXUOP * 4 + 16];
+	static VM_TLS int (*last_tls)[MAXUOP * 4 + 16]; int *last = TLS_BIG(last_tls);
 	int i, j, k, na, fr = (1 << RXNREG) - 1, n = B->n;
 	if (B->nv > MAXUOP * 4) return -1;
 	for (i = 0; i < B->nv; i++) { last[i] = -1; xkc[i] = 0; xvh[i] = 0; rxr[i] = -1; }
@@ -7363,6 +7449,85 @@ static int code_range_live(Cell start, Cell count)
 
 /* each worker owns a host thread, vm memory and translation state. */
 #if !defined(_WIN32) && !defined(__wasm__)
+#include <sys/mman.h>
+#ifdef __linux__
+#include <sys/syscall.h>
+#endif
+/* vm memory for a worker is copy-on-write against a snapshot of its parent.
+   the first spawn writes the parent's live words [0, memory_words) to an anonymous
+   shared file and maps the parent's memory privately over it at the same address,
+   so the parent reads the same pages and its later writes stay its own. a child maps
+   the same file privately, so code, genome and everything the parent has not
+   touched since are one set of physical pages for all of them, and only the pages
+   the parent changed since the snapshot are copied in. the file never changes, so a
+   child never sees a later write of its parent. a parent whose memory has drifted
+   far from its snapshot takes a fresh one. without a file (or on failure) a spawn
+   falls back to a private copy. */
+typedef struct { int fd; Cell *view; size_t words; } MemSnap;
+static VM_TLS MemSnap mem_snap = { -1, 0, 0 };
+static VM_TLS int memory_mapped;	/* m came from vm_memory_map */
+#define SNAP_PAGE 512	/* words per 4 KB page */
+static Cell *vm_memory_map(void)
+{
+	void *p = mmap(0, (size_t)MW * sizeof(Cell), PROT_READ | PROT_WRITE, MAP_PRIVATE | MAP_ANON | MAP_NORESERVE, -1, 0);
+	return p == MAP_FAILED ? 0 : p;
+}
+static void vm_memory_unmap(Cell *p) { if (p) munmap(p, (size_t)MW * sizeof(Cell)); }
+static int snap_file(size_t bytes)
+{
+	int fd = -1;
+#ifdef __linux__
+#ifdef SYS_memfd_create
+	fd = (int)syscall(SYS_memfd_create, "zy-snapshot", 1u /* MFD_CLOEXEC */);
+#endif
+#else
+	char name[64]; snprintf(name, sizeof name, "/zy-%d-%p", (int)getpid(), (void *)&name);
+	fd = shm_open(name, O_RDWR | O_CREAT | O_EXCL, 0600); if (fd >= 0) shm_unlink(name);
+#endif
+	if (fd < 0) return -1;
+	if (ftruncate(fd, (off_t)bytes) < 0) { close(fd); return -1; }
+	return fd;
+}
+static void snap_drop(void)
+{
+	if (mem_snap.view) munmap(mem_snap.view, mem_snap.words * sizeof(Cell));
+	if (mem_snap.fd >= 0) close(mem_snap.fd);
+	mem_snap.fd = -1; mem_snap.view = 0; mem_snap.words = 0;
+}
+/* pages of [0, words) where m differs from the snapshot */
+static size_t snap_dirty(size_t words)
+{
+	size_t p, n = 0, lim = words < mem_snap.words ? words : mem_snap.words;
+	for (p = 0; p < lim; p += SNAP_PAGE) if (memcmp(m + p, mem_snap.view + p, SNAP_PAGE * sizeof(Cell))) n++;
+	return n + (words > mem_snap.words ? (words - mem_snap.words) / SNAP_PAGE : 0);
+}
+static int snap_take(void)
+{
+	size_t words = ((size_t)memory_words + SNAP_PAGE - 1) / SNAP_PAGE * SNAP_PAGE, bytes = words * sizeof(Cell), done = 0;
+	int fd; void *view, *over;
+	if (words > (size_t)MW) return -1;
+	fd = snap_file(bytes); if (fd < 0) return -1;
+	while (done < bytes) { ssize_t k = pwrite(fd, (char *)m + done, bytes - done, (off_t)done); if (k <= 0) { close(fd); return -1; } done += (size_t)k; }
+	view = mmap(0, bytes, PROT_READ, MAP_SHARED, fd, 0);
+	if (view == MAP_FAILED) { close(fd); return -1; }
+	over = mmap(m, bytes, PROT_READ | PROT_WRITE, MAP_PRIVATE | MAP_FIXED | MAP_NORESERVE, fd, 0);
+	if (over == MAP_FAILED) { munmap(view, bytes); close(fd); return -1; }	/* m is untouched when the map fails */
+	snap_drop(); mem_snap.fd = fd; mem_snap.view = view; mem_snap.words = words; return 0;
+}
+/* a child's memory: the snapshot, plus the pages the parent changed since */
+static Cell *snap_clone(void)
+{
+	size_t p, words = (size_t)memory_words;
+	Cell *c;
+	if (mem_snap.fd < 0 || snap_dirty(words) * 4 > mem_snap.words / SNAP_PAGE) { if (snap_take() < 0) return 0; }
+	c = vm_memory_map(); if (!c) return 0;
+	if (mmap(c, mem_snap.words * sizeof(Cell), PROT_READ | PROT_WRITE, MAP_PRIVATE | MAP_FIXED | MAP_NORESERVE, mem_snap.fd, 0) == MAP_FAILED) { vm_memory_unmap(c); return 0; }
+	for (p = 0; p < words; p += SNAP_PAGE) {
+		size_t k = words - p < SNAP_PAGE ? words - p : SNAP_PAGE;
+		if (p >= mem_snap.words || memcmp(m + p, mem_snap.view + p, k * sizeof(Cell))) memcpy(c + p, m + p, k * sizeof(Cell));
+	}
+	return c;
+}
 typedef struct ContextChannel { Cell handle; int fd; struct ContextChannel *next; } ContextChannel;
 static VM_TLS ContextChannel *context_channels;
 static atomic_llong context_id_sequence;
@@ -7374,6 +7539,7 @@ struct VmContext {
  Cell token, entry, sp, rp, stack, words, limit, growth, code_size;
  Cell frame_floor, frame_top, frame_guard;
  Cell *memory;
+ int memory_mapped;
  int exit_code, requested_exit;
  void *input_snapshot;
  int argc, envc;
@@ -7387,6 +7553,7 @@ struct VmContext {
  int region_count;
  unsigned short *entries;
  unsigned char *entry_indices;
+ int *entry_at, entry_count;
 };
 static int context_descriptor(Cell handle) {
  ContextChannel *channel;
@@ -7405,7 +7572,7 @@ static void context_snapshot_free(VmContext *context) {
  int i;
  for (i = 0; i < NENH; i++) { free(context->enhancements[i].u); free(context->enhancements[i].reg); }
  for (i = 0; i < context->region_count; i++) free(context->regions[i]);
- free(context->input_snapshot); free(context->entries); free(context->entry_indices); if (context->memory) { free(context->memory); atomic_fetch_sub(&context_memory_live, 1); } context->memory = 0;
+ free(context->input_snapshot); free(context->entries); free(context->entry_indices); free(context->entry_at); if (context->memory) { if (context->memory_mapped) vm_memory_unmap(context->memory); else free(context->memory); atomic_fetch_sub(&context_memory_live, 1); } context->memory = 0;
 }
 static int context_cancelled(void) { return context_current && atomic_load_explicit(&context_current->stop, memory_order_acquire); }
 static void context_exit(int value) { context_current->exit_code = value; context_current->requested_exit = 1; atomic_store_explicit(&context_current->stop, 1, memory_order_release); }
@@ -7425,7 +7592,7 @@ static void context_engine_free(void) {
  free(lcov); free(lt); free(hot);
 #endif
 #endif
- free(m); m = 0; atomic_fetch_sub(&context_memory_live, 1);
+ if (memory_mapped) vm_memory_unmap(m); else free(m); m = 0; snap_drop(); tls_big_free(); io_context_release(); atomic_fetch_sub(&context_memory_live, 1);
 }
 static void *context_worker(void *opaque) {
  VmContext *context = opaque;
@@ -7434,7 +7601,7 @@ static void *context_worker(void *opaque) {
  optimizer_storage();
  context_current = context; io_context_restore(context->input_snapshot); free(context->input_snapshot); context->input_snapshot = 0; trace = getenv("ZYTRACE") != 0;
  context->worker_channel.next = context_channels; context_channels = &context->worker_channel;
- m = context->memory; context->memory = 0;
+ m = context->memory; context->memory = 0; memory_mapped = context->memory_mapped;
  memory_limit = (int)context->limit; memory_words = (int)context->words; memory_growth_limit = (int)context->growth;
  stack_control = context->stack; frame_floor_word = context->frame_floor; frame_top_word = context->frame_top; frame_guard_word = context->frame_guard;
  zy_argc = context->argc; zy_argv = context->argv; zy_envc = context->envc; zy_envp = context->envp;
@@ -7463,8 +7630,8 @@ static void *context_worker(void *opaque) {
  memcpy(enh, context->enhancements, sizeof enh); memset(context->enhancements, 0, sizeof context->enhancements);
  for (i = 0; i < context->region_count; i++) { rgs[i] = context->regions[i]; context->regions[i] = 0; }
  nrg = context->region_count; context->region_count = 0;
- memcpy(o3ent, context->entries, (size_t)code_words * sizeof *o3ent); memcpy(o3ei, context->entry_indices, (size_t)code_words);
- free(context->entries); context->entries = 0; free(context->entry_indices); context->entry_indices = 0;
+ for (i = 0; i < context->entry_count; i++) { o3ent[context->entry_at[i]] = context->entries[i]; o3ei[context->entry_at[i]] = context->entry_indices[i]; }
+ free(context->entries); context->entries = 0; free(context->entry_indices); context->entry_indices = 0; free(context->entry_at); context->entry_at = 0;
 #ifdef ZYJIT_X64
  if (x64_boot(sp, rp, &context->stop)) { context->exit_code = -1; context_channel_close(context->worker_channel.handle); context_engine_free(); return 0; }
  atomic_fetch_add(&context_code_live, 1);
@@ -7489,7 +7656,7 @@ static void *context_worker(void *opaque) {
  if (!context->requested_exit) context->exit_code = context_cancelled() ? 0 : stopped(back) ? -1 : inmem(sp) ? (int)m[sp] : 0;
  context_channel_close(context->worker_channel.handle);
  while (context_children) context_join(context_children->token, 1);
- context_engine_free(); context_current = 0; return 0;
+ hold_release_all(); context_engine_free(); context_current = 0; return 0;
 }
 static Cell context_start(Cell entry, Cell channel) {
  VmContext *context; int fd[2], i; pthread_attr_t attributes; Cell saved;
@@ -7506,12 +7673,17 @@ static Cell context_start(Cell entry, Cell channel) {
  context->parent_channel.fd = fd[0]; context->worker_channel.fd = fd[1]; atomic_init(&context->stop, 0);
  context->input_snapshot = malloc(io_context_size());
  if (!context->input_snapshot) goto failed; io_context_snapshot(context->input_snapshot);
- context->memory = calloc((size_t)memory_limit, sizeof(Cell));
+ context->memory = snap_clone(); context->memory_mapped = context->memory != 0;
+ if (!context->memory) context->memory = calloc((size_t)memory_limit, sizeof(Cell));
  if (context->memory) atomic_fetch_add(&context_memory_live, 1);
- context->entries = malloc((size_t)code_words * sizeof *o3ent); context->entry_indices = malloc((size_t)code_words);
- if (!context->memory || !context->entries || !context->entry_indices) goto failed;
- saved = m[channel]; m[channel] = context->worker_channel.handle; memcpy(context->memory, m, (size_t)memory_words * sizeof(Cell)); m[channel] = saved;
- memcpy(context->entries, o3ent, (size_t)code_words * sizeof *o3ent); memcpy(context->entry_indices, o3ei, (size_t)code_words);
+ /* o3 entry points are few and the tables mostly zero: hand over only the set ones */
+ { Cell k; context->entry_count = 0; for (k = 0; k < code_words; k++) if (o3ent[k]) context->entry_count++; }
+ context->entry_at = malloc((size_t)(context->entry_count + 1) * sizeof *context->entry_at);
+ context->entries = malloc((size_t)(context->entry_count + 1) * sizeof *o3ent); context->entry_indices = malloc((size_t)(context->entry_count + 1));
+ if (!context->memory || !context->entries || !context->entry_indices || !context->entry_at) goto failed;
+ if (context->memory_mapped) context->memory[channel] = context->worker_channel.handle;
+ else { saved = m[channel]; m[channel] = context->worker_channel.handle; memcpy(context->memory, m, (size_t)memory_words * sizeof(Cell)); m[channel] = saved; }
+ { Cell k; int j = 0; for (k = 0; k < code_words; k++) if (o3ent[k]) { context->entry_at[j] = (int)k; context->entries[j] = o3ent[k]; context->entry_indices[j] = o3ei[k]; j++; } }
  for (i = 0; i < NENH; i++) {
   context->enhancements[i] = enh[i]; context->enhancements[i].u = 0; context->enhancements[i].reg = 0;
   if (enh[i].u) { context->enhancements[i].u = malloc((size_t)(enh[i].n + 1) * sizeof(Uop)); if (!context->enhancements[i].u) goto failed; memcpy(context->enhancements[i].u, enh[i].u, (size_t)(enh[i].n + 1) * sizeof(Uop)); }
@@ -7526,7 +7698,7 @@ static Cell context_start(Cell entry, Cell channel) {
  memcpy(context->callback_snapshot, ffi_cb, sizeof ffi_cb); memcpy(context->symbol_snapshot, ffi_sym, sizeof ffi_sym); memcpy(context->library_snapshot, ffi_lib, sizeof ffi_lib);
  context->ffi_symbols = ffi_nsym; context->ffi_libraries = ffi_nlib; context->ffi_libm = ffi_tried_libm;
 #endif
- pthread_attr_init(&attributes); pthread_attr_setstacksize(&attributes, 16u << 20);
+ pthread_attr_init(&attributes); pthread_attr_setstacksize(&attributes, 4u << 20);	/* the vm keeps its own stacks in vm memory; 4 MB of address space covers the deepest native recursion (jit compile, o3) */
  i = pthread_create(&context->thread, &attributes, context_worker, context); pthread_attr_destroy(&attributes);
  if (i) goto failed;
  context->parent_channel.next = context_channels; context_channels = &context->parent_channel;
@@ -7603,6 +7775,10 @@ int main(int argc, char **argv)
 	memory_words = memory_option("TINYVM_MEMORY_WORDS", HEAP_HI > PW ? HEAP_HI : PW, HEAP_HI > PW ? HEAP_HI : PW);
 	if (memory_words < 0) return 1;
 	if (memory_words > memory_limit) { fprintf(stderr, "zy: initial memory exceeds its limit\n"); return 1; }
+#if !defined(_WIN32) && !defined(__wasm__)
+	m = vm_memory_map(); memory_mapped = m != 0;
+	if (!m)
+#endif
 	m = calloc((size_t)MW, sizeof *m);
 #ifdef ZYNATIVE
 	cov = calloc((size_t)MW, 1);
diff --git a/src/core/hold.h b/src/core/hold.h
new file mode 100644
index 0000000..d4ab1f0
--- /dev/null
+++ b/src/core/hold.h
@@ -0,0 +1,116 @@
+/* hold blocks (spec 4.16): the host side of a monitor. a function of a hold block
+   stores op 57 (enter) at its start and op 58 (leave) at its exit, with the block's
+   rank and the address of its descriptor word. the descriptor points at a table in vm
+   memory: n, then n pairs (address, words) of the block's shared variables.
+
+   the shared values live here, outside every vm, so all actors (each with its own
+   memory) see one copy. entering a block from outside takes its lock and copies the
+   values into the vm's variables; leaving copies them back and unlocks. inside, they
+   are plain globals. the lock is per block and reentrant per thread (a nested call
+   only counts), and blocks are taken in rank (declaration) order: entering a block
+   while holding a later one is an error, which rules out lock-order deadlocks between
+   blocks. a worker that ends while holding blocks releases them without writing back.
+
+   included by src/core/core.c and src/micro/micro.c; HOLD_THREADS selects pthreads,
+   otherwise a single machine needs no locks (wasm, bare metal). no malloc. */
+#ifndef HOLD_H
+#define HOLD_H
+#define HOLD_BLOCKS 256
+#define HOLD_CELLS 65536
+#ifdef HOLD_THREADS
+#include <pthread.h>
+static pthread_mutex_t hold_lock[HOLD_BLOCKS], hold_meta = PTHREAD_MUTEX_INITIALIZER;
+static pthread_once_t hold_once = PTHREAD_ONCE_INIT;
+static void hold_init(void) { int i; for (i = 0; i < HOLD_BLOCKS; i++) pthread_mutex_init(&hold_lock[i], 0); }
+#define HOLD_TLS _Thread_local
+#else
+#define HOLD_TLS
+#endif
+static int64_t hold_cells[HOLD_CELLS];
+static int hold_base[HOLD_BLOCKS], hold_size[HOLD_BLOCKS], hold_ready[HOLD_BLOCKS], hold_next;
+static HOLD_TLS int hold_depth[HOLD_BLOCKS], hold_top;	/* this thread: nesting per block, highest rank held */
+static HOLD_TLS unsigned char hold_order[HOLD_BLOCKS];	/* ranks held, in the order taken */
+static HOLD_TLS int hold_n;
+
+/* the table's entries, checked against the machine. -1 when it is malformed */
+static int64_t hold_table(int64_t *mem, uint64_t words, int64_t desc, int64_t *n)
+{
+	int64_t t, i;
+	if (desc <= 0 || (uint64_t)desc >= words) return -1;
+	t = mem[desc];
+	if (t <= 0 || (uint64_t)t >= words) return -1;
+	*n = mem[t];
+	if (*n < 0 || *n > 64 || (uint64_t)(t + 1 + 2 * *n) > words) return -1;
+	for (i = 0; i < *n; i++) {
+		int64_t a = mem[t + 1 + 2 * i], w = mem[t + 2 + 2 * i];
+		if (a <= 0 || w <= 0 || (uint64_t)(a + w) > words) return -1;
+	}
+	return t;
+}
+static void hold_copy(int64_t *mem, int64_t t, int64_t n, int rank, int in)
+{
+	int64_t i, k, c = hold_base[rank];
+	for (i = 0; i < n; i++) {
+		int64_t a = mem[t + 1 + 2 * i], w = mem[t + 2 + 2 * i];
+		for (k = 0; k < w; k++, c++) { if (in) mem[a + k] = hold_cells[c]; else hold_cells[c] = mem[a + k]; }
+	}
+}
+/* op 57: 0, or -1 (bad block), or -2 (taken out of rank order; the caller stops) */
+static int64_t hold_enter(int64_t *mem, uint64_t words, int64_t rank, int64_t desc)
+{
+	int64_t t, n, i, total = 0;
+	if (rank <= 0 || rank >= HOLD_BLOCKS) return -1;
+	if (hold_depth[rank]) { hold_depth[rank]++; return 0; }
+	if (hold_top > rank) return -2;
+	t = hold_table(mem, words, desc, &n); if (t < 0) return -1;
+	for (i = 0; i < n; i++) total += mem[t + 2 + 2 * i];
+#ifdef HOLD_THREADS
+	pthread_once(&hold_once, hold_init);
+	pthread_mutex_lock(&hold_lock[rank]);
+	pthread_mutex_lock(&hold_meta);
+#endif
+	if (!hold_ready[rank]) {	/* the first entry anywhere: the image's initial values become the shared ones */
+		if (hold_next + total > HOLD_CELLS) {
+#ifdef HOLD_THREADS
+			pthread_mutex_unlock(&hold_meta); pthread_mutex_unlock(&hold_lock[rank]);
+#endif
+			return -1;
+		}
+		hold_base[rank] = hold_next; hold_size[rank] = (int)total; hold_next += (int)total; hold_ready[rank] = 1;
+		hold_copy(mem, t, n, (int)rank, 0);
+	}
+#ifdef HOLD_THREADS
+	pthread_mutex_unlock(&hold_meta);
+#endif
+	hold_copy(mem, t, n, (int)rank, 1);
+	hold_depth[rank] = 1; hold_order[hold_n++] = (unsigned char)rank; hold_top = (int)rank;
+	return 0;
+}
+/* op 58 */
+static int64_t hold_leave(int64_t *mem, uint64_t words, int64_t rank, int64_t desc)
+{
+	int64_t t, n;
+	if (rank <= 0 || rank >= HOLD_BLOCKS || !hold_depth[rank]) return -1;
+	if (--hold_depth[rank]) return 0;
+	t = hold_table(mem, words, desc, &n);
+	if (t >= 0) hold_copy(mem, t, n, (int)rank, 0);
+	if (hold_n > 0 && hold_order[hold_n - 1] == rank) hold_n--;
+	hold_top = hold_n ? hold_order[hold_n - 1] : 0;
+#ifdef HOLD_THREADS
+	pthread_mutex_unlock(&hold_lock[rank]);
+#endif
+	return 0;
+}
+/* a machine that ends while inside blocks gives them up, without writing back */
+static void hold_release_all(void)
+{
+	while (hold_n > 0) {
+		int r = hold_order[--hold_n];
+		hold_depth[r] = 0;
+#ifdef HOLD_THREADS
+		pthread_mutex_unlock(&hold_lock[r]);
+#endif
+	}
+	hold_top = 0;
+}
+#endif
diff --git a/src/core/io_linux.c b/src/core/io_linux.c
index aaf6a18..c3e05dd 100644
--- a/src/core/io_linux.c
+++ b/src/core/io_linux.c
@@ -1,3 +1,4 @@
+#include <stdlib.h>
 #include <stdio.h>
 #include <string.h>
 #ifdef _WIN32
@@ -25,7 +26,11 @@ char Q(void){int c=getchar();J=pk();return c;}
    one would wait for a message that comes only after the answer. so J is 1 only
    when the end is known, and 0 when more may come. the output is flushed before
    any read that can wait, for the same reason. */
-static VM_TLS unsigned char ib[65536];
+#define IBN 65536
+/* each vm thread's buffer is allocated when it first reads, so an actor that never
+   reads standard input carries none */
+static VM_TLS unsigned char *ibp;
+#define ib (ibp ? ibp : (ibp = malloc(IBN)) ? ibp : (fputs("zy: out of memory\n", stderr), exit(2), ibp))
 static VM_TLS int ip, in, ieof;
 /* make a byte available. wait is 0 to look without waiting. returns 1 when one is ready */
 static int fill(int wait)
@@ -37,7 +42,7 @@ static int fill(int wait)
 		struct pollfd f; f.fd = 0; f.events = POLLIN; f.revents = 0;
 		if (poll(&f, 1, 0) <= 0) return 0;	/* nothing yet: not the end, not a byte */
 	} else fflush(stdout);
-	n = (int)read(0, ib, sizeof ib);
+	n = (int)read(0, ib, IBN);
 	if (n <= 0) { ieof = 1; return 0; }
 	ip = 0; in = n;
 	return 1;
@@ -62,16 +67,20 @@ int input_pending(void)
 #endif
 }
 #ifndef _WIN32
-struct IoSnapshot { unsigned char bytes[65536]; int position, count, ended, flag; };
-size_t io_context_size(void) { return sizeof(struct IoSnapshot); }
+/* what a new worker inherits: the bytes read ahead and not consumed yet, and the flags */
+struct IoSnapshot { int position, count, ended, flag; unsigned char bytes[]; };
+size_t io_context_size(void) { return sizeof(struct IoSnapshot) + (size_t)(ip < in ? in - ip : 0); }
 void io_context_snapshot(void *destination) {
  struct IoSnapshot *snapshot = destination;
- memcpy(snapshot->bytes, ib, sizeof ib); snapshot->position = ip; snapshot->count = in; snapshot->ended = ieof; snapshot->flag = J;
+ if (ip < in) memcpy(snapshot->bytes, ib + ip, (size_t)(in - ip));
+ snapshot->position = ip; snapshot->count = in; snapshot->ended = ieof; snapshot->flag = J;
 }
 void io_context_restore(const void *source) {
  const struct IoSnapshot *snapshot = source;
- memcpy(ib, snapshot->bytes, sizeof ib); ip = snapshot->position; in = snapshot->count; ieof = snapshot->ended; J = snapshot->flag;
+ ip = snapshot->position; in = snapshot->count; ieof = snapshot->ended; J = snapshot->flag;
+ if (ip < in) memcpy(ib + ip, snapshot->bytes, (size_t)(in - ip));
 }
+void io_context_release(void) { free(ibp); ibp = 0; }
 #endif
 void G(char c){putchar(c);}