JosephFerano/flan
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity
flan/test/test_reload.ml

516 lines
26 KiB
OCaml
Raw Blame History

(* The reload primitive, measured (NEXT.md, dev loop step 1).
One function is recompiled into its own object and loaded into a process
that is already running. Everything after this — indirection cells, the
agent in the game, the daemon — assumes this works and is fast; nothing in
the codebase had ever done it, and plan.org's 16ms was measured with clang
in isolation somewhere else.
The parts, all of them new here:
Emit.program ~dev a cell per function; every call goes through one
Emit.redefinition a form list defined, everything else [external],
plus [flan_reload_install] to publish it into its cell
flan_dev.c the by-name registry a run-time-new name needs
Build.shared that IR text through llc + ld -shared, timed
reload_host.c dlopen, install, call — twice, in one process
The host is C rather than OCaml because that is where it has to end up: the
agent of step 3 lives in the game process, next to flan_rt.c, and there is
no OCaml runtime there. *)
open Flan
(* The watchdog first: a hang is the one failure mode that reports
nothing at all. See watchdog.ml. *)
let () = Watchdog.arm ~seconds:600 "test_reload"
let failures = ref 0
let fail fmt = Printf.ksprintf (fun s -> incr failures; print_endline ("FAIL " ^ s)) fmt
let scratch = Filename.get_temp_dir_name ()
let tmp name = Filename.concat scratch ("flan-reload-" ^ name)
let checked path =
Check.program
(Load.program ~file:path (Reader.read_file path)).Load.decls
let ms f =
let t0 = Unix.gettimeofday () in
let x = f () in
(x, (Unix.gettimeofday () -. t0) *. 1000.)
let () =
match Sys.command "command -v clang > /dev/null 2>&1 && command -v llc > /dev/null 2>&1" with
| 0 ->
let p1 = checked "programs/reload.flan" in
let p2 = checked "programs/reload-v2.flan" in
let p3 = checked "programs/reload-v3.flan" in
let p4 = checked "programs/reload-v4.flan" in
(* What the running process was built with. Everything else — v3's [extra]
and [added] — has no symbol to bind to and goes through the registry.
A session would keep this set and grow it; the test states it. *)
let host_names =
List.map (fun (f : Tast.fn) -> f.Tast.name) p1.Tast.fns
@ List.map (fun (g : Tast.global) -> g.Tast.gname) p1.Tast.globals
in
let known n = List.exists (String.equal n) host_names in
(* [dev] is the two halves of a reloadable build together: cells, so a
call site can be made to follow a redefinition, and [-rdynamic], so the
cells and globals are visible to a dlopen'd object at all. [-ldl] is the
host's own, for its dlopen. *)
let dev = { Build.default with Build.dev = true } in
let host = tmp "host" in
ignore
(Build.executable ~opts:dev ~csrcs:[ "reload_host.c" ]
~lflags:[ "-ldl" ] p1 ~out:host);
(* Two paths, not one rewritten in place: dlopen caches by path and would
hand back the first handle, so the swap would silently not happen. *)
let module_of p fns name =
let out = tmp name in
let ir, emit_ms = ms (fun () -> Emit.redefinition ~dev:true ~known p ~fns) in
let t = Build.shared ~opts:dev ~ir ~out () in
(out, ir, emit_ms, t)
in
let so1, _ir1, emit_ms, t1 = module_of p1 [ "bump" ] "v1.so" in
let so2, ir2, emit2_ms, t2 = module_of p2 [ "bump" ] "v2.so" in
(* One module, two forms: the var and the function that uses it have to
arrive together or the intermediate state refers to storage that does
not exist. This is the C-c C-k unit. *)
let so3, ir3, _, _ = module_of p3 [ "bump"; "added" ] "v3.so" in
let so4, ir4, _, _ = module_of p4 [ "added" ] "v4.so" in
(* v5 retypes [extra], which v3 introduced at run time. It is built here
and loaded in a process of its own below: what it does is abort. *)
let p5 = checked "programs/reload-v5.flan" in
let so5, _, _, _ = module_of p5 [ "added" ] "v5.so" in
(* A redefinition module must not define what the host already owns:
defining [counter] would give the loaded object a private copy and the
state would reset on every reload, and defining [helper] would freeze a
stale copy of it into the module. *)
let has hay needle =
let n = String.length needle and h = String.length hay in
let rec go i = i + n <= h && (String.sub hay i n = needle || go (i + 1)) in
go 0
in
(* No Str, for the same reason the reader is hand-written. [`First] and
[`Last] are which occurrence; the pair shape keeps the match below
readable. *)
let find hay needle which =
let n = String.length needle and h = String.length hay in
let rec go i acc =
if i + n > h then acc
else if String.sub hay i n = needle then
match which with `First -> Some i | `Last -> go (i + 1) (Some i)
else go (i + 1) acc
in
go 0 None
in
if not (has ir2 "@\"flan.counter\" = external global i64") then
fail "redefinition defines the global instead of declaring it";
(* In a dev module a sibling is reached only through its cell, so there is
nothing to declare and a [define] would be a private copy. *)
if has ir2 "declare i64 @\"flan.helper\"" then
fail "dev redefinition declares a sibling it should reach by cell";
if has ir2 "define i64 @\"flan.helper\"" then
fail "redefinition emitted a second body for a function it does not own";
if has ir2 "define i32 @main" then fail "redefinition emitted an entry point";
(* [hidden], or the module's own [@"flan.bump"] is interposed by the host's
and the installer publishes the very function it is replacing. *)
if not (has ir2 "define hidden i64 @\"flan.bump\"") then
fail "redefinition's own body is interposable";
if not (has ir2 "@\"flan.cell.helper\" = external global ptr") then
fail "redefinition defines a cell instead of using the host's";
(* A name the host has is a symbol; a name it lacks is a registry lookup
cached in a module-local slot. Getting this backwards either fails to
link or silently gives each module its own copy. *)
if not (has ir3 "@\"flan.cellp.added\" = internal global ptr null") then
fail "a run-time-new function did not get a slot";
if not (has ir3 "@\"flan.gp.extra\" = internal global ptr null") then
fail "a run-time-new global did not get a slot";
if has ir3 "@\"flan.extra\" = " then
fail "a run-time-new global was given storage in the module";
(* Every lookup is resolved before any body is published: publishing first
exposes a function whose module-local slots are still null to anything
that calls it. Not race-testable, so it is asserted on the text. *)
(* v4 redefines a name that exists only in the registry, so it publishes
through the cell it looked up rather than into a symbol — there is no
[@"flan.cell.added"] anywhere to store into. *)
if has ir4 "@\"flan.cell.added\"" then
fail "a run-time-new function was published into a symbol";
if not (has ir4 "call ptr @flan_dev_cell") then
fail "v4 did not look its target up by name";
(match find ir3 "store ptr @\"flan." `First, find ir3 "@flan_dev_(" `Last with
| Some publish, Some resolve when resolve > publish ->
fail "flan_reload_install publishes a body before resolving a lookup"
| None, _ -> fail "flan_reload_install publishes nothing"
| _ -> ());
(* A dev host's calls are indirect; a release host's are not. That is the
only difference between the two, and the whole of C-c C-c rests on it. *)
let host_ir = Emit.program ~dev:true p1 in
if not (has host_ir "@\"flan.cell.bump\" = global ptr @\"flan.bump\"") then
fail "dev build emitted no cell";
if has (Emit.program p1) "flan.cell." then
fail "release build emitted a cell";
let out = tmp "out" in
let cmd =
(* stderr kept apart from stdout: the host times its own dlopen there,
and stdout is what the expected transcript is compared against. *)
Printf.sprintf "%s %s %s %s %s > %s 2> %s" (Filename.quote host)
(Filename.quote so1) (Filename.quote so2) (Filename.quote so3)
(Filename.quote so4) (Filename.quote out)
(Filename.quote (tmp "err"))
in
let (code, dlopen_ms) = ms (fun () -> Sys.command cmd) in
let text = In_channel.with_open_bin out In_channel.input_all in
let timings = In_channel.with_open_bin (tmp "err") In_channel.input_all in
(* Every call is [outer], compiled once into the host and never rebuilt, so
a changed answer can only mean its call site followed the redefinition.
The arithmetic, in order:
host counter 0 -> 1, helper 1 = 2
v1 counter 1 -> 2, helper 2 = 4 (a rebuild of the same)
v2 +10 and +1000, recursing through its own cell until the
counter passes 100: 2 -> 12 -> ... -> 102, ten "v2" lines,
helper 102 = 204, so 1204. An interposed self-call would reach
the host's v1 body, print "v1", and land nowhere near it.
v3 extra 0 -> 7, counter 102 -> 109, helper 109 = 218. [extra]
and [added] are new names, so both came from the registry.
v4 redefines [added] only. v3's [bump] is still the installed one
and is not rebuilt here, so it reaches v4 only through a cell
the two modules found by the same name: extra 7 -> 107,
counter 109 -> 216, helper 216 = 432. Had v3 cached the
function's address rather than its cell's, this would be 246.
The "v1"/"v2"/"v3" lines come from inside each [bump] and are what
exercise a redefinition module's own string constants. 1204 rather than
1236 is [helper]: v2's text for it multiplies by three, and the module
declares it rather than defining it, so the host's copy is the one that
ran. *)
let v2s = String.concat "" (List.init 10 (fun _ -> "v2\n")) in
let want =
"v1\nhost 2\nv1\nafter1 4\n" ^ v2s ^ "after2 1204\n\
v3\nafter3 218\nv3\nafter4 432\ncounter 216\n"
in
if code <> 0 || text <> want then
fail "reload\n got: %S (exit %d)\n wanted: %S" text code want;
(* The same thing again, compiled by the dev backend end to end (x86.ml's
header, HANDOFF-x86-rt.md item 1). Both halves, host and module, because
the two backends' conventions agree on every scalar and disagree on
every aggregate: an LLVM-built module dlopened into an --x86 host would
be correct until the first redefined function took or returned a struct.
So an --x86 host gets --x86 modules and the two never meet.
Only v1 and v2, which is the whole of what X86.redefinition compiles:
every name they touch is one the host was built with. v3 and v4
introduce a function and a global at run time, and the flan_dev_cell /
flan_dev_global lookups that needs are refused there by name.
Read by running, not by reading. A disassembly reads correctly beside a
wrong answer often enough (DISCUSS.md item 15) that only the printed
transcript settles it: [outer] is compiled once into the host and never
rebuilt, so "after2 1204" can only mean its call site followed a body
that this backend emitted, published through a cell it reached via the
GOT. *)
let x86 = { dev with Build.x86 = true } in
let xhost = tmp "xhost" in
ignore
(Build.executable ~opts:x86 ~csrcs:[ "reload_host.c" ]
~lflags:[ "-ldl" ] p1 ~out:xhost);
let xmodule q fns name =
let o = tmp name in
let asm = X86.redefinition ~checks:true ~dev:true ~known q ~fns in
ignore (Build.shared_x86 ~opts:x86 ~asm ~out:o ());
o
in
let xso1 = xmodule p1 [ "bump" ] "xv1.so" in
let xso2 = xmodule p2 [ "bump" ] "xv2.so" in
let xout = tmp "xout" in
let xcode =
Sys.command
(Printf.sprintf "%s %s %s > %s 2> %s" (Filename.quote xhost)
(Filename.quote xso1) (Filename.quote xso2) (Filename.quote xout)
(Filename.quote (tmp "xerr")))
in
let xtext = In_channel.with_open_bin xout In_channel.input_all in
let xwant =
"v1\nhost 2\nv1\nafter1 4\n" ^ v2s ^ "after2 1204\ncounter 102\n"
in
if xcode <> 0 || xtext <> xwant then
fail "x86 reload\n got: %S (exit %d)\n wanted: %S" xtext xcode
xwant;
(* A name the host was never built with has no symbol to bind to, and the
registry path is not built here. It has to refuse rather than emit
something that links and then stores through a null. *)
(match X86.redefinition ~checks:true ~dev:true ~known p3 ~fns:[ "added" ] with
| _ -> fail "x86 redefinition accepted a name the host does not have"
| exception X86.Unsupported _ -> ());
(* The aggregate case, which is the whole reason X86.redefinition exists
rather than an --x86 host dlopening what Emit.redefinition made.
Everything above this point is scalar, and scalars are the half of the
calling convention the two backends cannot disagree about. They disagree
on every aggregate: x86.ml passes each one by pointer and returns it
through a hidden sret, LLVM classifies per eightbyte. So a redefined
function taking or returning a struct is the case that would expose a
mismatch, and until now the claim that an --x86 host plus --x86 modules
is same-convention-by-construction was an argument rather than a
measurement.
programs/reload-agg.flan crosses the boundary in four shapes at once —
two integer eightbytes, thirty-two bytes of MEMORY, two SSE eightbytes,
and one of each — because SysV treats those four differently and this
backend treats them identically, so a single shape would measure a
quarter of the disagreement and read like all of it. Each `step' takes
an aggregate and returns one, so a single call crosses in both
directions, and each calls a `weigh' the module does not define, which
hands an aggregate the other way.
Both backends run the same fixture and are compared against the same
transcript. The LLVM row is not decoration: a wrong expected number
would otherwise be indistinguishable from a backend that is right, and
two independently-built agreements on one string are what rule that
out. *)
let a1 = checked "programs/reload-agg.flan" in
let a2 = checked "programs/reload-agg-v2.flan" in
let agg_known =
let names =
List.map (fun (f : Tast.fn) -> f.Tast.name) a1.Tast.fns
@ List.map (fun (g : Tast.global) -> g.Tast.gname) a1.Tast.globals
in
fun n -> List.exists (String.equal n) names
in
let agg_fns = [ "step-pair"; "step-quad"; "step-duo"; "step-mix" ] in
(* The arithmetic, derived rather than observed, because a number read off
a run is a record of what happened and not a statement of what should:
v1 Pair {1,2} -> weigh 1+3*2 = 7, so {2, 2+7} and 2 + 100*9 = 902
Quad {1,2,3,4} -> weigh (1+6)+(15+28) = 50, so {2,4,6,54} and
2 + 400 + 60000 + 54000000 = 54060402
Duo {1,2} -> weigh 7, so {2.0, 9.0} and 902
Mix {1,2} -> weigh 7, so {2, 9.0} and 902
total 54063108
v2 Pair -> weigh is still the *host's* 7, so {11, 2+14} and 1611
Quad -> weigh still 50, so {11,22,33,104} and
11 + 2200 + 330000 + 104000000 = 104332211
Duo -> {11.0, 16.0} and 1611
Mix -> {11, 16.0} and 1611
total 104337044
1611 rather than 12411 in the first term is the tripwire: v2's text for
`weigh-pair' multiplies by thirty, and a module that grew its own copy
of a sibling rather than reaching the host's through a cell would say
so here. `counter' is stepped from inside the redefined body, by one in
v1 and by ten in v2, so 1 + 1 + 10 = 12 is the host's global being
written by three different bodies in turn. *)
let agg_want =
"a1\nhost 54063108\na1\nafter1 54063108\na2\nafter2 104337044\n\
counter 12\n"
in
let agg_run label opts mkmod =
let h = tmp ("agg-host-" ^ label) in
ignore
(Build.executable ~opts ~csrcs:[ "reload_host.c" ] ~lflags:[ "-ldl" ]
a1 ~out:h);
let m1 = mkmod a1 ("agg-" ^ label ^ "-1.so") in
let m2 = mkmod a2 ("agg-" ^ label ^ "-2.so") in
let o = tmp ("agg-out-" ^ label) and e = tmp ("agg-err-" ^ label) in
let code =
Sys.command
(Printf.sprintf "%s %s %s > %s 2> %s" (Filename.quote h)
(Filename.quote m1) (Filename.quote m2) (Filename.quote o)
(Filename.quote e))
in
let text = In_channel.with_open_bin o In_channel.input_all in
if code <> 0 || text <> agg_want then
fail "%s aggregate reload\n got: %S (exit %d)\n wanted: %S"
label text code agg_want;
List.iter (fun p -> try Sys.remove p with Sys_error _ -> ())
[ h; m1; m2; o; e ]
in
let agg_llvm_mod q name =
let o = tmp name in
let ir = Emit.redefinition ~dev:true ~known:agg_known q ~fns:agg_fns in
ignore (Build.shared ~opts:dev ~ir ~out:o ());
o
in
let agg_x86_mod q name =
let o = tmp name in
let asm =
X86.redefinition ~checks:true ~dev:true ~known:agg_known q ~fns:agg_fns
in
ignore (Build.shared_x86 ~opts:x86 ~asm ~out:o ());
o
in
agg_run "llvm" dev agg_llvm_mod;
agg_run "x86" x86 agg_x86_mod;
(* The same measurement run crossed, which is what the marker symbol is
for. Before it, an --x86 host given LLVM-built modules loaded them and
then died with SIGSEGV on the first call into a redefined aggregate
body:
got: "a1\nhost 54063108\na1\n" (exit 139)
That could not be asserted. It was undefined behaviour and what it
printed was a property of whichever LLVM happened to be installed; a
test pinning it would have been pinning the shape of a crash.
It is deterministic now, which is why it is here. A dev build defines a
marker naming the backend that built it — [flan.abi.x86] or
[flan.abi.llvm] — and a redefinition module holds a pointer to the one
it was itself built for. That pointer is a relocation the loader has to
resolve while it maps the object, so a crossed pair fails the [dlopen]
outright, before a single instruction of the new body runs. Both
directions, because a marker only one of the two backends emitted would
refuse in one direction and say nothing in the other.
Asserted on the message as well as the exit status, the way the
retyped-global and registry-overflow cases below are: a nonzero exit is
not by itself this refusal, and the point of the exercise is that what
reaches a user names the reason rather than repeating the loader's
"undefined symbol". *)
let agg_cross label opts mkmod wants =
let h = tmp ("agg-xhost-" ^ label) in
ignore
(Build.executable ~opts ~csrcs:[ "reload_host.c" ] ~lflags:[ "-ldl" ]
a1 ~out:h);
let m1 = mkmod a1 ("agg-cross-" ^ label ^ "-1.so") in
let o = tmp ("agg-xout-" ^ label) and e = tmp ("agg-xerr-" ^ label) in
let code =
Sys.command
(Printf.sprintf "%s %s > %s 2> %s" (Filename.quote h)
(Filename.quote m1) (Filename.quote o) (Filename.quote e))
in
let said = In_channel.with_open_bin e In_channel.input_all in
if code = 0 then
fail "%s: a crossed pair loaded and ran (exit 0)" label;
if not (has said "built by different backends") then
fail "%s: a crossed pair was refused without naming the reason: %S"
label said;
(* Which marker is missing is which backend built the module, so this is
also what says the refusal fired for the right direction rather than
for the other one. *)
if not (has said wants) then
fail "%s: the refusal named the wrong marker (wanted %s): %S" label
wants said;
List.iter (fun p -> try Sys.remove p with Sys_error _ -> ())
[ h; m1; o; e ]
in
(* An --x86 host handed an LLVM module: the pair the CLI can build today,
since [flan reload] has no --x86 spelling. *)
agg_cross "x86-host-llvm-module" x86 agg_llvm_mod "flan.abi.llvm";
(* And the reverse, which no command spells but [X86.redefinition] does. *)
agg_cross "llvm-host-x86-module" dev agg_x86_mod "flan.abi.x86";
(* The option-record guard, which is the older and narrower half of the
same answer: [Build.opts] is where the backend choice lives, so a
builder handed the *other* backend's option record refuses by name. It
catches a caller holding one option record and reaching for the wrong
builder. It cannot catch a caller holding two — the crossed runs above
pass both of these refusals — which is what the marker is for. See
HANDOFF-x86-aggregates.md and HANDOFF-x86-abi-marker.md. *)
(match Build.shared ~opts:x86 ~ir:"" ~out:(tmp "never.so") () with
| _ -> fail "Build.shared accepted an --x86 option record"
| exception Failure m when has m "--x86" -> ()
| exception Failure m -> fail "Build.shared refused for the wrong reason: %s" m);
(match Build.shared_x86 ~opts:dev ~asm:"" ~out:(tmp "never.so") () with
| _ -> fail "Build.shared_x86 accepted an LLVM option record"
| exception Failure m when has m "--x86" -> ()
| exception Failure m ->
fail "Build.shared_x86 refused for the wrong reason: %s" m);
(* The layout-drift guard, which needs a process of its own because what it
does is abort one. [extra] does not exist in the host: v3 introduced it
at run time, so flan_dev.c allocated its storage and recorded its size,
and every later module asking for that name is handed the same
allocation back. v5 asks for it as an i32. Handing back eight bytes for
a four-byte type is not an error anything downstream can detect — the
new body simply reads fields at offsets the allocation was never laid
out for — so the registry compares sizes and dies at the first chance
it has.
Asserted on the message as well as on the exit status: a process that
died for some other reason is not this guard firing, and the exit code
alone cannot tell the two apart. *)
let out5 = tmp "out5" and err5 = tmp "err5" in
let code5 =
Sys.command
(Printf.sprintf "%s %s %s %s > %s 2> %s" (Filename.quote host)
(Filename.quote so1) (Filename.quote so3) (Filename.quote so5)
(Filename.quote out5) (Filename.quote err5))
in
let said = In_channel.with_open_bin err5 In_channel.input_all in
if code5 = 0 then
fail "a global retyped across a reload was accepted (exit 0)";
if not (has said "size changed") then
fail "a retyped global did not stop on the size guard: %S" said;
(* The two fixed-size limits in flan_dev.c, which nothing had ever
reached: the 4K result buffer a renderer emits into, and the 4096-name
registry. Both are driven from dev_limits.c rather than from Flan,
because neither has a Flan spelling and a program that reached either
one by accident would be a program nobody wants in the corpus.
One process per mode. The name table never shrinks, so the two cases
would contaminate each other, and the overflow case ends in abort. *)
let limits = tmp "limits" in
ignore
(Build.executable ~opts:dev ~csrcs:[ "dev_limits.c" ] p1 ~out:limits);
let mode m =
let o = tmp ("limits-" ^ m ^ ".out") and e = tmp ("limits-" ^ m ^ ".err") in
let code =
Sys.command
(Printf.sprintf "%s %s > %s 2> %s" (Filename.quote limits) m
(Filename.quote o) (Filename.quote e))
in
let out = In_channel.with_open_bin o In_channel.input_all in
let err = In_channel.with_open_bin e In_channel.input_all in
List.iter (fun p -> try Sys.remove p with Sys_error _ -> ()) [ o; e ];
(code, out, err)
in
(* 6000 bytes emitted into 4096. The length is the cap itself, the three
dots are what says the value was cut rather than being that short, the
middle byte says the content before the cut is the content that was
emitted, and the generation moved exactly once — a reader waits on that
counter and a value published twice would be read half-formed. The last
line is the flag being cleared: a short value after a truncated one must
not inherit its ellipsis. *)
let code, out, _ = mode "cap" in
let want_cap = "len 4096\ntail ...\nmid b\nhead a\ngen 1\nagain 12\n" in
if code <> 0 || out <> want_cap then
fail "the 4K result cap\n got: %S (exit %d)\n wanted: %S"
out code want_cap;
(* 4096 distinct names fit; the next one stops the process. The table is
fixed and never moves, because a loaded module holds the address of a
cell in it, so growing is not available and overrunning is the only
other thing it could do. *)
let code, out, err = mode "names" in
if code = 0 then fail "the registry accepted a 4097th name (exit 0)";
if out <> "interned 4096\n" then
fail "the registry did not take 4096 names first: %S" out;
if not (has err "out of dev name slots") then
fail "the registry overflowed without saying so: %S" err;
Printf.printf
"reload: emit %.1fms llc %.1fms ld %.1fms (v2: emit %.1fms llc %.1fms ld %.1fms) host run %.1fms\n"
emit_ms t1.Build.llc_ms t1.Build.link_ms emit2_ms t2.Build.llc_ms
t2.Build.link_ms dlopen_ms;
print_string timings;
List.iter (fun p -> try Sys.remove p with Sys_error _ -> ())
[ host; limits; so1; so2; so3; so4; so5; out; out5; err5; tmp "err" ];
if !failures = 0 then print_endline "reload: all tests passed"
else begin
Printf.printf "\n%d failure(s)\n" !failures;
exit 1
end
| _ -> print_endline "reload: skipped (no clang or llc on PATH)"
Reference in New Issue View Git Blame Copy Permalink