Joseph Ferano c2d378957e The x86 backend and dyn finally meet, which is where the dev loop is

x86 is what flan dev takes by default and dyn is the iteration feature, so
a backend that refused dyn meant the two halves of the dev loop could not
be in the same program. The refusal was one arm of is_agg, and it said the
true thing: it was never the representation that was missing. A dyn is
uint64_t, a scalar in both calling conventions, classified by every rule
this file already had; every operation on one is a Tast.Rt primitive and
call_rt has always known how to make one of those. What the lane actually
cost was the collector's root discipline.

Which is emit.ml's, reused rather than rewritten: Emit.dyn_roots counts the
roots for both backends now, so the pushes and the pops balance because one
counter decides both ends, and the two backends root the same nodes because
there is one counter and not two. A zeroed frame slot per dyn slot and per
dyn-producing call, minted beside the channel and outside every scoped --
the bump allocator reclaims at the end of a statement and a slot minted in
the body would be handed out again while the collector still held its
address. Pushed from the body buffer, not the prologue's, because a call
clobbers the registers the prologue is still spilling from. And one pop in
the epilogue, which is the whole of why this backend needed no landing-pad
work for it: there is exactly one epilogue, and the return, the fall-through
and the transfer exit all arrive at it. emit.ml needs the same pop at five
separate rets.

The ABI point the dyn handoff left open for the integrator is settled by
reading the other side rather than by agreeing: flan_dyn.c's mark follows a
value only when the quiet-NaN prefix is set, and the zero word does not have
it, so a zeroed root decodes as the double 0.0 and is never an address
anything dereferences. Zero is safe for a reason. The header says so now.

And one line in dev.ml that was never x86's: the merged dev host resets the
condition stacks and the frame chain between runs, because main is
re-entered by longjmp and pops no frame -- and it never reset the root
stack, so every root a finished run pushed still named stack the next run
was about to write over. That gap was an LLVM dev build's too.

Verification, and one of the numbers is new. @x86: MATCH 129 -> 135, DIFFER
0, REFUSED 0 -- the five dyn programs off survey.sh's llvmonly list, which
is gone rather than empty, plus p13. dune test --force green, with --x86
acceptance rows beside the LLVM ones for all five dyn programs, dyn-boundary
asserted on the same exit 134 and the same sentence on both.

p13-dyn-collect.flan is the one that is not a formality. Nothing else in
this repository allocates past flan_dyn.c's one-megabyte floor, so nothing
else collects even once, so a program whose roots are entirely wrong passes
every output test there is -- the handoff wrote that about the stub and it
outlived the stub. p13 allocates several megabytes of garbage while holding
live values across it: at forty times the corpus size it peaks at 4MB of
RSS, which is the collector running many times over, and both backends
still print the same four lines.

2026-09-19 14:22:55 +07:00

8.9 KiB

Raw Blame History

dyn, milestone 1 — what was decided and what is left

The compiler half of dynamic-by-default. The runtime half is a sibling's, built in parallel against runtime/flan_dyn.h, which is the fixed ABI and the thing the two copies are diffed against.

Two decisions that differ from the brief

The return slot stays mandatory; dyn is written out in it. The brief expected ret = None to grow a third state meaning "unannotated", and listed mechanical fallout in load.ml, shim.ml and cimport.ml. That fallout does not exist, because the change was not made. The reason is in parse.ml beside the defn case: an optional return slot has no syntactic resolution, since a capitalised head in a list is both a type application and a struct literal — (defn f [] (Rune {.code 65}) (bar)) is the misparse that removed the old optional slot, and it would come straight back. A parameter vector has no such case, because every slot in it is a name or a type and never an expression. So ret = None still means Unit and only declare and the shim produce it. One token in the return position buys a decision the file paid for twice in one day.

The parameter rule is resolved in Check, not in parse.ml. The brief asked for "a known type name is a type, anything else is another dyn param", written where parse.ml argues its other misparse-closing decisions. That lookup is exactly the one parse.ml:904 records being removed for being wrong twice in one day, and at parse time the set of type names is incomplete by construction — macros generate definitions, packages are loaded later, C headers are imported later. cimport.ml decides it: named env n = tname n passes C type names through verbatim, so POSIX's stat and timespec are lowercase Flan type names writable in parameter position, and no syntactic rule ("capitalised is a type") can be made sound.

So the vector is carried undecided as Ast.pitems and paired in Check.pair_params, after every file is loaded, every macro expanded and every header imported. The argument is written at parse.ml's defn case as asked.

The residual the parent owns

The set of type names is complete at a point in time and not across time. (defn f [x y] ...) is two dyn parameters until somebody writes (defstruct y ...) — or imports a header that declares one — and then it is one parameter of type y, with no edit to f. The signature changes underneath it, and arity changes with it.

Session.compatible is where that is felt: it compares with Types.equal over parameters and return, so a redefinition that changes dyn-ness is refused like any other signature change (this falls out; it is pinned in test_session.ml). But the first definition after such an edit is the one that changes, and nothing warns.

What the feature costs, and what was taken back

A parameter slot with no type used to be a syntax error. It is now a dyn parameter, so a mistyped type silently becomes an extra parameter — the arity changes with no diagnostic, which is the failure class parse.ml calls the worst available. Two rules take most of it back, in dyn_param_or_typo:

a name within one edit of a type's name gets the resolver's own "did you mean", and
an unknown capitalised name is reported as an unknown type. Not one parameter in the corpus is capitalised, while Form, Cursor and Vector2 appear in these vectors constantly.

What is left uncovered is a lowercase name resembling no type: (defn f [x widget] ()) is two dyn parameters and nothing in the text says otherwise. That is the feature working as specified.

Sharp edge of the near-miss rule. near_miss treats any two single-char names as one edit apart, and it compares against every struct name in scope. So a (defstruct D ...) anywhere in the program makes (defn f [a d] ...) a refusal rather than two dyn parameters. The message is actionable — write the type, or rename — but it is a refusal a user will meet without having done anything wrong.

Open ABI point for the integrator

A rooted slot holding 0 is not a value, and the collector must skip it. This is written into runtime/flan_dyn.h beside the root functions, and it is the one thing in that header decided by one side alone. Roots are pushed in the function's entry block, before the code that fills them has run and possibly for a branch that never runs, so the compiler zeroes every root slot and must mean something by it — and 0 is the only pattern it can write without knowing the encoding.

If the real runtime NaN-boxes and integer zero is the zero word, this is wrong and the two sides need a different sentinel. Do not fix it on one side.

Settled, by reading the other side. The real runtime does NaN-box, and the zero word is not the zero integer: an integer is boxed, and boxed means the quiet-NaN prefix is set. mark_value in runtime/flan_dyn.c follows a value only when dyn_boxed holds, which tests (v & 0xFFF8000000000000) == 0xFFF8000000000000, and the zero word fails it. So a rooted slot holding 0 decodes as the double 0.0 — an ordinary value rather than a marker, and crucially never an address the collector dereferences. Zero is safe, and the header's sentence is true for a reason both sides can check rather than by the two of them having guessed alike. No sentinel is needed and neither side changes.

Not in milestone 1, each refused by name with a location

a typed container boxing into dyn ((Vec i64) → dyn): "not yet"; the heterogeneous container is the runtime's own from (vec-new dyn)
a dyn in a condition's payload, or in a field of one: milestone 2 — a payload crosses a handler boundary and must stay rooted across the transfer
a dyn crossing to C through declare/declare-c: it is one word and would have passed as an integer with nothing on the other side able to ask what it means. This one was not in the brief and is the dangerous one, because the general "cannot cross to C" arm would have caught it with advice (pass (Ptr T)) that is wrong for dyn.
integer widths other than i64 and floats other than f64 unboxing from dyn: the ABI carries one of each, and a need_i64 plus a truncation would put an implicit narrowing at the one boundary where the value's type was already uncertain
~~the x86 dev backend, and~~ the JS dialect, refuse dyn entirely. The x86 backend does not any more: a dyn is one machine word in both calling conventions and every operation on one is an ordinary Tast.Rt call, so what the lane cost was the root discipline and not the arithmetic — a zeroed frame slot per dyn local and per dyn-producing call, pushed in the body buffer at entry, and one flan_dyn_root_pop in the epilogue that every return and every transfer out of the frame already went through. Emit.dyn_roots is called by both backends, which is what makes the counts agree rather than merely both being written down

Roots: what is and is not verified

Every dyn slot and every dyn-producing runtime call is rooted, pushed in the entry block and popped at every ret — which is the funnel all five exits pass through, the transfer landing block included. Pushes and pops balance by construction: dyn_roots counts before emission, the slots are minted from that count, and dyn_tmp only hands them out.

The stub verifies none of this. flan_dyn_stub.c mallocs and never frees, so a program with entirely wrong root discipline passes every test that runs against it. What is checked instead is the IR, and that check earned its keep — it found a real hole. A defer appears twice in the typed IR, spliced into body for the normal path and again in fdefers for the path a transfer leaves through, so a dyn temporary inside one is emitted twice; dyn_roots counted only the body's, and the second copy went into slots the collector had never been told about.

Nothing failed, which is the point. dyn_tmp falls back to a plain unrooted slot rather than unbalancing the stack, so the pushes and the pops still matched, the program ran and printed the right answer, and four dyn values were simply invisible. Under a stub that never collects there is no symptom at all.

The assertion that caught it is in test_acceptance.ml: a rooted slot is spelled %dr and the fallback %dx, and no dyn program in the corpus may emit the latter. When the real collector lands, that is the check to extend rather than replace — it is the only one that can see a missing root before there is a collector to lose one by.

Cost: a rooted alloca has its address escape through flan_dyn_root_push, so mem2reg cannot promote it. Every dyn local and every dyn temporary is a real stack slot with a real store, at every optimisation level. That is inherent to a precise collector with an address-registration ABI rather than stack maps.

A function with no dyn emits nothing — no push, no pop, not a pop(0) — which is what makes --no-gc byte-identity hold.

8.9 KiB Raw Blame History