passes: implement control-flow pass (4 schemes + rotation)

FlowPass rebuilds one method per scheme picked from FlowConfig:
state-spread (dense tableswitch dispatcher), computed-switch
(sparse lookupswitch over a counter local), butterfly (branch
inversion), exception-goto (dispatcher under try with a dead
catch-all), plus opaque goto-splices inside every method and
StackMapTable recomputation afterwards. Now in standard_pipeline
(between constants and noise); --no-flow is an honest off-flag.

Acceptance is real JVM, not shapes: per-scheme verify, rotation
spread, calc-prints-45 execution, paranoid smoke on lorep-1.1.0
(131 methods rebuilt, no new failures over standard).

Two verifier rules learned the hard way are now construction
invariants: no inserted mark shares a bytecode offset with a
foreign one (nop pads), and no frameless gap follows a terminal
(dead blocks are opaque-entered and throw).

Also fixes a latent frame-emitter bug the pass exposed: wide
stack values were written as slot pairs ([Double, Top]) while
javac writes one entry ([Double]); HotSpot rejects the pair.
Stacks now collapse like locals (emit::stack_vars), with a
regression test in classfile/tests/frames.rs.
This commit is contained in:
Loki 2026-10-09 11:11:55 +02:00
parent 39178420e4
commit 2172206e83
Signed by: boba
GPG key ID: 253067914055423B
15 changed files with 1360 additions and 103 deletions

View file

@ -35,7 +35,7 @@ pub(super) fn emit_frames(
Some(previous) => target
.checked_sub(previous)
.and_then(|delta| delta.checked_sub(1))
.ok_or_else(|| merger.invalid("кадры не по порядку"))?,
.ok_or_else(|| merger.invalid("frames out of order"))?,
};
out.push(select(delta, &state.locals, &previous_locals, &state.stack));
previous_locals = state.locals.clone();
@ -54,13 +54,17 @@ fn select(
// otherwise halves would produce `Append{[Top, Long]}`.
let current = vars(locals);
let base = vars(previous);
// The operand stack needs the same collapsing: a `long`/`double` is one
// entry (`javac` writes `[Double, Integer]`, not `[Double, Top, Integer]`);
// an explicit half without its value is rejected by HotSpot.
let stack = stack_vars(stack);
if current == base {
if stack.is_empty() {
return StackMapFrame::Same {
offset_delta: delta,
};
}
if let [local] = stack {
if let [local] = stack.as_slice() {
return StackMapFrame::SameLocals {
offset_delta: delta,
local: local.clone(),
@ -94,6 +98,30 @@ fn select(
}
}
/// Stack entries → values: `[Double, Top]` — one entry (like `javac`).
///
/// Unlike [`vars`], no trailing-`Top` trimming: every stack entry is live by
/// construction, and dropping one would corrupt the depth.
fn stack_vars(stack: &[VerificationType]) -> Vec<VerificationType> {
let mut out = Vec::with_capacity(stack.len());
let mut index = 0;
while index < stack.len() {
match &stack[index] {
VerificationType::Long | VerificationType::Double
if stack.get(index + 1) == Some(&VerificationType::Top) =>
{
out.push(stack[index].clone());
index += 2;
}
other => {
out.push(other.clone());
index += 1;
}
}
}
out
}
/// Slots → variables: `[Long, Top]` — one entry; trailing `Top`s — gone.
///
/// Single `Top`s inside (holes) are kept: in `Full` they are a legal type, and we

View file

@ -1,9 +1,9 @@
//! Приёмка `compute_frames` на фикстурах (`tests/data`).
//! Acceptance for `compute_frames`: fixtures plus emitter edge cases.
//!
//! Кадры сносятся и считаются заново. Судьи: детерминизм + round-trip IR +
//! настоящий `java -Xverify:all` для классов без внешних зависимостей.
//! (`verify_class` backend'а для приёмки не годится: он ждёт кадры в байтах,
//! а `from_bytes` отдаёт в индексах — см. `todo.md`.)
//! Frames are dropped and recomputed. Judges: determinism + IR round-trip +
//! real `java -Xverify:all` for classes without external dependencies.
//! (The backend's `verify_class` is unfit for acceptance: it expects frames
//! in bytes while `from_bytes` hands out indices — see `todo.md`.)
#![allow(clippy::expect_used, clippy::unwrap_used)]
@ -11,9 +11,12 @@ use std::collections::BTreeMap;
use std::path::{Path, PathBuf};
use std::process::Command;
use noxis_classfile::{Hierarchy, compute_frames, format};
use noxis_classfile::{
Arith, BranchCond, ClassBuilder, Code, Cond, Const, Insn, Prim, SlotKind, Version,
{Hierarchy, compute_frames, format},
};
/// Карта `класс → суперкласс` для LUB.
/// A `class → superclass` map for LUB.
#[derive(Default)]
struct MapHierarchy {
supers: BTreeMap<String, String>,
@ -25,13 +28,13 @@ impl Hierarchy for MapHierarchy {
}
}
/// Классы только на JDK — их `java` проверит полностью.
/// JDK-only classes — `java` checks them fully.
const JDK_ONLY: [&str; 2] = [
"dev/loki/lorep/util/TimeUtil",
"dev/loki/lorep/util/PaginationUtil",
];
/// Фикстуры: снести кадры, посчитать заново, проверить выход.
/// Fixtures: drop frames, recompute, check the output.
#[test]
fn fixtures_recompute_and_verify() {
let dir = PathBuf::from(env!("CARGO_MANIFEST_DIR")).join("tests/data");
@ -41,7 +44,7 @@ fn fixtures_recompute_and_verify() {
.filter(|path| path.extension().is_some_and(|ext| ext == "class"))
.collect();
names.sort();
assert!(!names.is_empty(), "нет фикстур в tests/data");
assert!(!names.is_empty(), "no fixtures in tests/data");
let mut supers = BTreeMap::new();
let mut parsed = Vec::new();
@ -50,7 +53,7 @@ fn fixtures_recompute_and_verify() {
supers.insert(class.this_class.clone(), class.super_class.clone());
parsed.push(class);
}
// Без суперкласса — только `java/lang/Object`.
// Only `java/lang/Object` has no superclass.
supers.entry("java/lang/Object".to_string()).or_insert(None);
let hierarchy = MapHierarchy {
supers: supers
@ -71,11 +74,11 @@ fn fixtures_recompute_and_verify() {
}
let first = compute_frames(&class.this_class, method, &hierarchy).expect("recompute");
method.code.as_mut().expect("code").frames = first.clone();
// Детерминизм: повторный прогон даёт те же кадры.
// Determinism: a second run yields the same frames.
let second = compute_frames(&class.this_class, method, &hierarchy).expect("recompute");
assert_eq!(first, second, "недетерминированные кадры");
assert_eq!(first, second, "nondeterministic frames");
}
// Round-trip: записанное читается в тот же IR (включая кадры).
// Round-trip: written bytes read back into the same IR (frames included).
let written = format::write(class).expect("write");
assert_eq!(
format::parse(&written).expect("re-parse"),
@ -89,7 +92,85 @@ fn fixtures_recompute_and_verify() {
let _ = std::fs::remove_dir_all(&scratch);
}
/// Есть ли `java` для проверки настоящим верификатором.
/// A join with a `double` below on the stack must emit it as ONE entry.
///
/// Regression: the emitter used to write the slot pair (`[Double, Top]`),
/// like `javac` does for locals — but on the operand stack `javac` writes a
/// bare `[Double]`, and HotSpot rejects the pair with
/// `Inconsistent stackmap frames`.
#[test]
fn wide_stack_join_emits_single_entries() {
// iload_0, i2d, iload_1, <opaque join>, i2d, ddiv, d2i, ireturn.
// At the join the stack holds a double plus an int.
let (dead, into) = (100, 101);
let mut insns = vec![
Insn::Load {
slot: 0,
kind: SlotKind::Int,
},
Insn::Convert(noxis_classfile::Convert::I2D),
Insn::Load {
slot: 1,
kind: SlotKind::Int,
},
];
insns.extend([
Insn::Const(Const::Int(7)),
Insn::Dup,
Insn::Binary {
prim: Prim::Int,
op: Arith::Mul,
},
Insn::Const(Const::Int(7)),
Insn::Binary {
prim: Prim::Int,
op: Arith::Add,
},
Insn::Const(Const::Int(1)),
Insn::Binary {
prim: Prim::Int,
op: Arith::And,
},
Insn::Branch(BranchCond::Zero(Cond::Ne), dead),
Insn::Goto(into),
Insn::Mark(dead),
Insn::Nop,
Insn::Mark(into),
Insn::Convert(noxis_classfile::Convert::I2D),
Insn::Binary {
prim: Prim::Double,
op: Arith::Div,
},
Insn::Convert(noxis_classfile::Convert::D2I),
Insn::Return(Some(Prim::Int)),
]);
let code = Code {
max_stack: 6,
max_locals: 2,
insns,
exception_table: Vec::new(),
frames: Vec::new(),
attributes: Vec::new(),
};
let class = ClassBuilder::new("demo/Wide", Some("java/lang/Object"))
.version(Version::JAVA_21)
.method_with_code(0x0009, "calc", "(II)I", code)
.build();
let hierarchy = MapHierarchy::default();
let frames = compute_frames(&class.this_class, &class.methods[0], &hierarchy).expect("frames");
let mut class = class;
class.methods[0].code.as_mut().expect("code").frames = frames;
if !java_available() {
eprintln!("SKIP: no java");
return;
}
let scratch = std::env::temp_dir().join(format!("noxis-wide-{}", std::process::id()));
let bytes = format::write(&class).expect("write");
java_verify(&scratch, &class, &bytes).expect("java -Xverify:all");
let _ = std::fs::remove_dir_all(&scratch);
}
/// Whether `java` is available for checks with the real verifier.
fn java_available() -> bool {
Command::new("java")
.arg("-version")
@ -97,8 +178,8 @@ fn java_available() -> bool {
.is_ok_and(|out| out.status.success())
}
/// Записать класс в дерево пакетов и прогнать `java -Xverify:all`.
/// Успех — загрузка без ошибок верификации (`main` при этом отсутствует).
/// Write a class into a package tree and run `java -Xverify:all`.
/// Success means loading without verification errors (`main` may be missing).
fn java_verify(
scratch: &Path,
class: &noxis_classfile::ClassFile,
@ -118,7 +199,7 @@ fn java_verify(
.arg(scratch)
.arg(&dotted)
.output()
.map_err(|e| format!("запуск java: {e}"))?;
.map_err(|e| format!("running java: {e}"))?;
let stderr = String::from_utf8_lossy(&out.stderr);
for bad in [
"VerifyError",