rt: exact-size heap bounds; n, the words new asked for, lives in bits 32..62 of p[-4]

the bnd recipe and paylen read n instead of the power-of-two block:
  bnd:    r6 = sub p 3; r6 = ld r6; r7 = xor p 0x2B7E1516; r6 = eq r6 r7;
          r7 = sub p 4; r7 = ld r7; r7 = shr r7 32; r7 = mul r7 q;
          r8 = mul i m; r8 = ltu r8 r7; r8 = xor r8 1; r0 = and r6 r8
  paylen: return p[-4] >> 32
two loads instead of three; osize, bchk, release, recycle and the memdebug
leak report all read n (rt.new(4) has length 4, rt.new(0) has 0).

writers: new_raw and new_k set n; stack.h line 87 and tinyvm.ts (BigInt) set
it on stack segments. release clears n when a block goes into quarantine
(len 0, indexing traps; under memdebug bchk reports a use after free); every
free-list path (fpush, mkblock, bdrop, recycle_leaf, balloc hold) rewrites
the word as k << 16, and an array that grows gets a fresh block. a request over
2^31-1-HDR words stops in new_raw (oom), and the compiler only takes the
new_k fast path for constant sizes that fit, so n is never truncated.

bit users checked (all mask or preserve bits 32..62; none compares or copies
the whole word): gc mark and sweep; release, free, recycle, recycle_leaf;
share and own; isobj, quar, checkptr, obj_base; bclass and the bfree buddy
merge; balloc, mkblock, fpush, funlink, heap_grow; frame_grow; stack_allocate
in src/core/stack.h and src/js/tinyvm.ts; micro.c and core.c (check word
only); the compiler's two p[-4] | F_INTERNAL writes (base.zc, builtins.zc).
32-bit ports: none (baremetal, wasm and micro use int64 Cell).

tests: tests/programs/exactsize.zc (n-1/n for non-power-of-two sizes, new(0),
packed bytes, recycle, grow, stack growth); 8-runtime: exact_size_bounds,
pointer_index_n_minus_1, pointer_index_n_traps, pointer_index_n_traps_memdebug,
new_zero_index_traps, new_too_big_traps, stale_index_after_free_traps,
stale_index_after_free_memdebug. expectations changed: leak report shows n
(tests/test.sh line 169, examples/learn/13-memdebug.out, docs/LEARN.md,
4-statements/memdebug_report_line.zc). programs adjusted to keep test.sh
expectations: quarantine.zc (unbounded), bounds_heap.zc (rt.new(11)).
spec 3.7.4 and 4.11 updated. fixpoint 402204 words (main 402197).

1 files changed, 4 insertions, 3 deletions

diff --git a/tests/programs/bounds_heap.zc b/tests/programs/bounds_heap.zc
index d0bd318..719757e 100644
--- a/tests/programs/bounds_heap.zc
+++ b/tests/programs/bounds_heap.zc
@@ -1,12 +1,13 @@
 namespace;
 include rt;
 
-// a pointer to the start of a live heap block is checked against the block's
-// payload, which a buddy block rounds up: rt.new(4) holds 11 words. a pointer
+// a pointer to the start of a live heap block is checked against exactly the
+// words new asked for: rt.new(11) holds 11 words (its block holds 11 too, but
+// rt.new(4) would hold 4, not the 11 its 16-word block has room for). a pointer
 // into the middle of a block, or to memory that is not a heap block, is not
 // checked. a char is a word, so a char* counts words.
 int main() {
-	int* p = rt.new(4); int* q; char* c = rt.new(2); int i = 10;
+	int* p = rt.new(11); int* q; char* c = rt.new(3); int i = 10;
 	p[i] = 3; rt.put_n(p[i]); rt.put_c(10);
 	q = p + 1; i = 10; rt.put_n(q[i - 1] - 3); rt.put_c(10);
 	i = 2; c[i] = 'x'; rt.put_c(c[i]); rt.put_c(10);