fix(dwarf): remap DW_AT_ranges lists to the fused layout (#319 inc 2) (#321)

* fix(dwarf): remap DW_AT_ranges lists to the fused layout (#319 inc 2)

The residual "DIE address ranges are not contained in its parent's ranges"
errors after inc 1 came from base-relative .debug_ranges offsets.

Root cause: DWARF v4 .debug_ranges entries are offset pairs relative to a
base (the CU's DW_AT_low_pc, which is nonzero for wasi-libc CUs). gimli
parses them as AddressOrOffsetPair and routes the *offsets* through
convert_address, which treats a small offset (e.g. 0x141) as an absolute
code address and remaps it to an unrelated output location — so the resolved
range escapes its parent (the .debug_ranges analogue of the inc-1 high_pc
bug, one level down). The CU base (an address form) is remapped correctly,
so consumer = out_base + garbage_offset → e.g. [0x1c439,..) outside the
parent [0x133b6,0x1353d).

Fix: correct_die_ranges — a post-convert pass that, walking read↔write DIEs
in lockstep pre-order, resolves each DIE's absolute input ranges via
gimli read `die_ranges` (base already applied), translates both endpoints to
the fused layout, and re-emits them as OffsetPair relative to the output CU
base. Sub-ranges whose endpoints can't be mapped (dropped/tombstoned code)
are dropped; an emptied list deletes DW_AT_ranges — never a wrong or
escaping range (LS-D-1).

Effect: the "not contained" class is eliminated — records-only 4->0,
records+variants 11->0 (cumulative with inc 1: 1049->35 multi, ~510->8
single). The remaining 35/8 are Cause #2 (.debug_loc tombstone vs
base-address-selection escape) = inc 3.

Test: inc2_die_ranges_stay_within_enclosing_subprogram — fuses records.wasm,
re-parses the fused DWARF, asserts every DW_AT_ranges entry sits inside its
enclosing subprogram (the invariant that failed). 2 witness + 24 dwarf unit
tests pass.

Refs: #319.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* harden(dwarf): drop zero-length ranges in correct_die_ranges (#319 inc 2)

Mythos delta-pass robustness: a degenerate begin==end OffsetPair is rejected
by gimli's range writer (InvalidRange), which would abort the whole DWARF
write and fall back to strip. Require strict oe > ob; a zero-length range
carries no info, so dropping it is lossless. (No mis-attribution — this is a
write-abort/reliability guard.)

Refs: #319.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Author: avrabe

meld-core/src/dwarf.rs

@@ -575,6 +575,9 @@ fn rewrite_debug_sections(
     // `llvm-dwarfdump --verify` flags as out-of-parent / overlapping. Fix
     // the lengths to the fused layout now that `low_pc` is remapped.
     correct_high_pc_lengths(&mut write_dwarf, remap);
+    // #319 inc 2: fix DW_AT_ranges lists (base-relative offsets that gimli
+    // mis-routes through convert_address) from the read side.
+    correct_die_ranges(&mut write_dwarf, &read_dwarf, remap);
     // #144 inc 3: the synthetic `<meld-adapter>` unit rides the SAME
     // write as the remapped original units, so every cross-section
     // offset is computed in one shared offset space (appending
@@ -644,6 +647,138 @@ fn correct_high_pc_lengths(write_dwarf: &mut gimli::write::Dwarf, remap: &Addres
     }
 }
 
+/// Collect a write unit's DIE ids in first-child-first pre-order — the same
+/// order `gimli::read`'s `next_dfs` yields, so the two can be walked in
+/// lockstep. (Distinct from [`correct_high_pc_lengths`]'s order-agnostic
+/// stack walk, which doesn't need to align with the read side.)
+fn write_dies_preorder(
+    unit: &gimli::write::Unit,
+    id: gimli::write::UnitEntryId,
+    out: &mut Vec<gimli::write::UnitEntryId>,
+) {
+    out.push(id);
+    let children: Vec<_> = unit.get(id).children().copied().collect();
+    for c in children {
+        write_dies_preorder(unit, c, out);
+    }
+}
+
+/// Rewrite every `DW_AT_ranges` list to the fused layout (#319 inc 2).
+///
+/// DWARF v4 `.debug_ranges` entries are base-relative offset pairs. gimli
+/// parses them as `AddressOrOffsetPair` and routes the *offsets* through
+/// `convert_address` — which treats a small offset as an absolute code
+/// address and remaps it to an unrelated output location, so the resolved
+/// range escapes its parent (the `.debug_ranges` analogue of the inc-1
+/// `high_pc`-length bug). Fix it from the read side, where the base is still
+/// applied: resolve each DIE's absolute input ranges via `die_ranges`,
+/// `translate` both endpoints to the fused layout, and re-emit them as
+/// offsets relative to the output CU base. Drop sub-ranges (or the whole
+/// list) that can't be mapped — never emit a wrong or escaping range
+/// (LS-D-1).
+fn correct_die_ranges<R: gimli::read::Reader<Offset = usize>>(
+    write_dwarf: &mut gimli::write::Dwarf,
+    read_dwarf: &gimli::read::Dwarf<R>,
+    remap: &AddressRemap,
+) {
+    use gimli::constants::{DW_AT_low_pc, DW_AT_ranges};
+    use gimli::write::{Address, AttributeValue, Range, RangeList};
+
+    let mut headers = read_dwarf.units();
+    let mut unit_idx = 0usize;
+    while let Ok(Some(header)) = headers.next() {
+        if unit_idx >= write_dwarf.units.count() {
+            break;
+        }
+        let uid = write_dwarf.units.id(unit_idx);
+        unit_idx += 1;
+        let unit = match read_dwarf.unit(header) {
+            Ok(u) => u,
+            Err(_) => continue,
+        };
+
+        // Output CU base = the write root's already-remapped low_pc; v4
+        // range entries are emitted relative to it. No usable base → skip.
+        let out_cu_base = {
+            let root = write_dwarf.units.get(uid).root();
+            match write_dwarf.units.get(uid).get(root).get(DW_AT_low_pc) {
+                Some(AttributeValue::Address(Address::Constant(a))) => *a as u32,
+                _ => continue,
+            }
+        };
+
+        // Write DIE ids in the same pre-order the read cursor will walk.
+        let write_ids = {
+            let wunit = write_dwarf.units.get(uid);
+            let mut ids = Vec::new();
+            write_dies_preorder(wunit, wunit.root(), &mut ids);
+            ids
+        };
+
+        let mut entries = unit.entries();
+        let mut wi = 0usize;
+        while let Ok(Some((_, entry))) = entries.next_dfs() {
+            if wi >= write_ids.len() {
+                break;
+            }
+            let die_id = write_ids[wi];
+            wi += 1;
+
+            // Only DIEs that carry a range LIST (not low/high — those are the
+            // inc-1 path). `die_ranges` would otherwise synthesize a single
+            // range from low/high and we'd wrongly convert a subprogram.
+            if entry.attr_value(DW_AT_ranges).ok().flatten().is_none() {
+                continue;
+            }
+
+            let mut ranges = match read_dwarf.die_ranges(&unit, entry) {
+                Ok(r) => r,
+                Err(_) => continue,
+            };
+            let mut new_ranges = Vec::new();
+            while let Ok(Some(r)) = ranges.next() {
+                // `die_ranges` yields absolute input addresses (base applied).
+                let (Some(ob), Some(oe)) = (
+                    remap.translate(r.begin as u32),
+                    remap.translate(r.end as u32),
+                ) else {
+                    continue; // unmappable (dropped/tombstoned) sub-range → drop
+                };
+                // Require a strictly-positive extent: a degenerate `begin ==
+                // end` OffsetPair is rejected by gimli's range writer
+                // (`InvalidRange`), which would abort the whole DWARF write
+                // and strip. A zero-length range carries no information, so
+                // dropping it is lossless.
+                if ob >= out_cu_base && oe > ob {
+                    new_ranges.push(Range::OffsetPair {
+                        begin: (ob - out_cu_base) as u64,
+                        end: (oe - out_cu_base) as u64,
+                    });
+                }
+            }
+
+            if new_ranges.is_empty() {
+                write_dwarf
+                    .units
+                    .get_mut(uid)
+                    .get_mut(die_id)
+                    .delete(DW_AT_ranges);
+            } else {
+                let rid = write_dwarf
+                    .units
+                    .get_mut(uid)
+                    .ranges
+                    .add(RangeList(new_ranges));
+                write_dwarf
+                    .units
+                    .get_mut(uid)
+                    .get_mut(die_id)
+                    .set(DW_AT_ranges, AttributeValue::RangeListRef(rid));
+            }
+        }
+    }
+}
+
 /// Top-level entry point for [`crate::DwarfHandling::Remap`].
 ///
 /// Inspects the input components for `.debug_*` sections and, when

meld-core/tests/dwarf_remap_witness.rs

@@ -190,3 +190,99 @@ fn remapped_subprogram_low_pcs_match_fused_body_starts() {
          fused body starts, {tombstoned} dead-code subprograms tombstoned"
     );
 }
+
+/// A multi-inner-module fixture whose `.debug_ranges` carry inlined
+/// subroutines (base-relative offset pairs) — the #319 inc-2 surface.
+const RANGES_FIXTURE: &str = "../tests/wit_bindgen/fixtures/records.wasm";
+
+/// #319 inc 2: no `DW_AT_ranges` DIE may escape its enclosing subprogram in
+/// the remapped output. `records.wasm` carries inlined subroutines whose
+/// `.debug_ranges` are base-relative offset pairs; before the fix gimli
+/// mis-routed those *offsets* through `convert_address` (treating a small
+/// offset as an absolute address), sending the ranges to unrelated output
+/// locations — `llvm-dwarfdump --verify`: "DIE address ranges are not
+/// contained in its parent's ranges". This re-parses the fused DWARF and
+/// asserts every resolved range sits inside the enclosing subprogram.
+#[test]
+fn inc2_die_ranges_stay_within_enclosing_subprogram() {
+    if !std::path::Path::new(RANGES_FIXTURE).is_file() {
+        eprintln!("skipping: fixture not found at {RANGES_FIXTURE}");
+        return;
+    }
+    let input = std::fs::read(RANGES_FIXTURE).expect("read fixture");
+    let fused = fuse_remap(&input);
+    let sections = debug_sections(&fused);
+    assert!(
+        sections.contains_key(".debug_info"),
+        "Remap must emit remapped .debug_info for the ranges fixture"
+    );
+
+    let endian = gimli::LittleEndian;
+    let load = |id: gimli::SectionId| -> Result<gimli::EndianSlice<'_, gimli::LittleEndian>, gimli::Error> {
+        let data = sections.get(id.name()).map(|v| v.as_slice()).unwrap_or(&[]);
+        Ok(gimli::EndianSlice::new(data, endian))
+    };
+    let dwarf = gimli::Dwarf::load(load).expect("load output dwarf");
+
+    let mut checked = 0usize;
+    let mut units = dwarf.units();
+    while let Some(header) = units.next().expect("unit header") {
+        let unit = dwarf.unit(header).expect("parse unit");
+        let mut entries = unit.entries();
+        let mut depth = 0isize;
+        // Stack of enclosing subprogram ranges: (die_depth, low, end).
+        let mut encl: Vec<(isize, u64, u64)> = Vec::new();
+        while let Some((delta, entry)) = entries.next_dfs().expect("dfs walk") {
+            depth += delta;
+            while matches!(encl.last(), Some(&(d, _, _)) if d >= depth) {
+                encl.pop();
+            }
+            if entry.tag() == gimli::constants::DW_TAG_subprogram
+                && let Some(gimli::AttributeValue::Addr(low)) = entry
+                    .attr_value(gimli::constants::DW_AT_low_pc)
+                    .expect("low_pc")
+                && low != TOMBSTONE
+                && let Some(hp) = entry
+                    .attr_value(gimli::constants::DW_AT_high_pc)
+                    .expect("high_pc")
+            {
+                let end = match hp {
+                    gimli::AttributeValue::Udata(len) => low + len,
+                    gimli::AttributeValue::Addr(a) => a,
+                    _ => low,
+                };
+                encl.push((depth, low, end));
+            }
+            // A DIE carrying a range LIST, checked against its enclosing
+            // subprogram (the class that escaped in #319).
+            if entry
+                .attr_value(gimli::constants::DW_AT_ranges)
+                .expect("ranges attr")
+                .is_some()
+                && let Some(&(_, plow, pend)) = encl.last()
+            {
+                let mut ranges = dwarf.die_ranges(&unit, entry).expect("die_ranges");
+                while let Some(r) = ranges.next().expect("range") {
+                    // Skip tombstone / dead-code sentinel entries.
+                    if r.begin >= 0xffff_fffe {
+                        continue;
+                    }
+                    assert!(
+                        r.begin >= plow && r.end <= pend,
+                        "inc2: DW_AT_ranges [{:#x},{:#x}) escapes enclosing \
+                         subprogram [{plow:#x},{pend:#x}) — base-relative range \
+                         offset was remapped as an absolute address (#319)",
+                        r.begin,
+                        r.end
+                    );
+                    checked += 1;
+                }
+            }
+        }
+    }
+    assert!(
+        checked > 0,
+        "expected at least one DW_AT_ranges DIE inside a subprogram to check"
+    );
+    eprintln!("inc2: {checked} DW_AT_ranges entries all contained in their subprogram");
+}