cgen: function-tier units for i386 ELF objects

The load bias of a listed object (GEX_ELF_UNITS, ?elfunits=) was read
from ELFCLASS64 headers only, so an i386 library never got its unit.
Read ELFCLASS32 headers and program headers at their own offsets and
widths, natively (pe_units) and in the browser (libprov).

tests/elf-units-remap/run32.sh: run.sh's two libraries at one address,
built with gcc -m32; four placements and each load's own output, t1
and fjit.

Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_8b3490cccb4848b495dde1c5cdcacba0

shims/browser/src/cgen/libprov.rs

@@ -991,13 +991,28 @@ fn elf_bias(bytes: &[u8], start: u64, offset: u64) -> Option<u64> {
         v[..n].copy_from_slice(b);
         Some(u64::from_le_bytes(v))
     };
-    if bytes.get(0..4)? != b"\x7fELF" || *bytes.get(4)? != 2 {
+    if bytes.get(0..4)? != b"\x7fELF" {
         return None;
     }
-    let (phoff, phentsize, phnum) = (rd(0x20, 8)? as usize, rd(0x36, 2)? as usize, rd(0x38, 2)? as usize);
+    // ELFCLASS64, or ELFCLASS32 (an i386 object): the header and program
+    // header fields at their own offsets and widths.
+    let wide = match *bytes.get(4)? {
+        2 => true,
+        1 => false,
+        _ => return None,
+    };
+    let (phoff, phentsize, phnum) = if wide {
+        (rd(0x20, 8)? as usize, rd(0x36, 2)? as usize, rd(0x38, 2)? as usize)
+    } else {
+        (rd(0x1c, 4)? as usize, rd(0x2a, 2)? as usize, rd(0x2c, 2)? as usize)
+    };
     for i in 0..phnum {
         let ph = phoff + i * phentsize;
-        let (ty, flags, off, vaddr) = (rd(ph, 4)?, rd(ph + 4, 4)?, rd(ph + 8, 8)?, rd(ph + 16, 8)?);
+        let (ty, flags, off, vaddr) = if wide {
+            (rd(ph, 4)?, rd(ph + 4, 4)?, rd(ph + 8, 8)?, rd(ph + 16, 8)?)
+        } else {
+            (rd(ph, 4)?, rd(ph + 24, 4)?, rd(ph + 4, 4)?, rd(ph + 8, 4)?)
+        };
         if ty == 1 && flags & 1 != 0 && off & !0xfff == offset {
             return Some(start.wrapping_sub(vaddr & !0xfff));
         }

shims/native/src/pe_units.rs

@@ -1807,13 +1807,28 @@ fn elf_bias(bytes: &[u8], start: u64, offset: u64) -> Option<u64> {
         v[..n].copy_from_slice(b);
         Some(u64::from_le_bytes(v))
     };
-    if bytes.get(0..4)? != b"\x7fELF" || *bytes.get(4)? != 2 {
+    if bytes.get(0..4)? != b"\x7fELF" {
         return None;
     }
-    let (phoff, phentsize, phnum) = (rd(0x20, 8)? as usize, rd(0x36, 2)? as usize, rd(0x38, 2)? as usize);
+    // ELFCLASS64, or ELFCLASS32 (an i386 object): the header and program
+    // header fields at their own offsets and widths.
+    let wide = match *bytes.get(4)? {
+        2 => true,
+        1 => false,
+        _ => return None,
+    };
+    let (phoff, phentsize, phnum) = if wide {
+        (rd(0x20, 8)? as usize, rd(0x36, 2)? as usize, rd(0x38, 2)? as usize)
+    } else {
+        (rd(0x1c, 4)? as usize, rd(0x2a, 2)? as usize, rd(0x2c, 2)? as usize)
+    };
     for i in 0..phnum {
         let ph = phoff + i * phentsize;
-        let (ty, flags, off, vaddr) = (rd(ph, 4)?, rd(ph + 4, 4)?, rd(ph + 8, 8)?, rd(ph + 16, 8)?);
+        let (ty, flags, off, vaddr) = if wide {
+            (rd(ph, 4)?, rd(ph + 4, 4)?, rd(ph + 8, 8)?, rd(ph + 16, 8)?)
+        } else {
+            (rd(ph, 4)?, rd(ph + 24, 4)?, rd(ph + 4, 4)?, rd(ph + 8, 4)?)
+        };
         if ty == 1 && flags & 1 != 0 && off & !0xfff == offset {
             return Some(start.wrapping_sub(vaddr & !0xfff));
         }

tests/elf-units-remap/run32.sh

@@ -0,0 +1,39 @@
+#!/bin/sh
+# run.sh's two libraries at one address, built for i386: the load bias of an
+# ELFCLASS32 object is found from its own program headers, so each i386
+# library gets its function-tier unit (four placements), and each load runs
+# its own code.
+#
+#   tests/elf-units-remap/run32.sh [gexc]    (needs gcc -m32 and an i386 libc)
+set -u
+export GEX_LOG=${GEX_LOG:-1}
+HERE=$(cd "$(dirname "$0")" && pwd)
+GEXC=${1:-target/x86/release/gexc}
+W=${WORK:-$(mktemp -d)}
+mkdir -p "$W"
+if ! echo 'int main(void){return 0;}' | gcc -m32 -x c -o "$W/probe" - 2>/dev/null; then
+	echo "skip elf-units-remap i386 (no gcc -m32)"
+	exit 0
+fi
+gcc -m32 -O1 -shared -fPIC -DK=3 -DC=1 -o "$W/liba.so" "$HERE/lib.c" || exit 1
+gcc -m32 -O1 -shared -fPIC -DK=7 -DC=-2 -o "$W/libb.so" "$HERE/lib.c" || exit 1
+gcc -m32 -O1 -o "$W/main" "$HERE/main.c" -ldl || exit 1
+BASE=0x30000000
+printf '%s 0x10000 liba.so\n%s 0x10000 libb.so\n' $BASE $BASE > "$W/pins.txt"
+export GEX_CGEN_CACHE="$W/cgen" GEX_ELF_UNITS=liba.so,libb.so GEX_ELF_UNITS_BUILD=0
+for l in liba libb; do
+	"$GEXC" cgen-image --elf "$W/$l.so" --base $BASE --jobs 2 > "$W/$l.cgen.log" 2>&1 || { cat "$W/$l.cgen.log"; exit 1; }
+done
+native=$("$W/main" "$W")
+fail=0
+for flags in "" "--fjit"; do
+	gex=$(timeout 60 "$GEXC" run $flags --pin "$W/pins.txt" "$W/main" "$W" 2> "$W/err.txt")
+	placed=$(grep -c 'function tier for .*/lib[ab]\.so' "$W/err.txt")
+	if [ "$native" = "$gex" ] && [ "$placed" = 4 ]; then
+		echo "ok   elf-units-remap i386 ${flags:-t1}"
+	else
+		echo "FAIL elf-units-remap i386 ${flags:-t1} (library units placed $placed times, want 4)"
+		printf 'native:\n%s\ngex:\n%s\n' "$native" "$gex"; tail -20 "$W/err.txt"; fail=1
+	fi
+done
+exit $fail