cranelift: force little-endian on inline bulk copy chunks

Pulley's `VLoad`/`VStore` for `v128` only encode the little-endian
variant, so `MemFlagsData::trusted()` (which inherits the target's
native endianness) tripped an emitter assertion when targeting
`pulley32be`/`pulley64be`. Pinning the chunk load/store endianness to
`Little` is safe because each loaded value feeds only its paired store
and is never observed as a number, so the destination's byte sequence
is independent of the chosen endianness as long as load and store
agree.

Author: claude

crates/cranelift/src/func_environ.rs

@@ -4461,7 +4461,16 @@ impl FuncEnvironment<'_> {
         // operand that requires alignment, so the backend selects unaligned
         // moves regardless of the `aligned` flag. The range was already
         // bounds-checked, so `notrap` is fine too.
-        let flags = ir::MemFlagsData::trusted();
+        //
+        // Endianness is pinned to `Little` rather than inheriting the target's
+        // native endianness for two reasons: (1) Pulley's `v128` load/store only
+        // encode the little-endian variant and would assert otherwise, and (2)
+        // the byte sequence written to the destination is independent of the
+        // chosen endianness as long as load and store agree, since each loaded
+        // chunk is fed straight back to its paired store without being observed
+        // as a number. On big-endian native targets this costs two byteswaps
+        // per chunk that cancel each other out, which we accept.
+        let flags = ir::MemFlagsData::trusted().with_endianness(Endianness::Little);
         const WIDTHS: &[(u64, ir::Type)] = &[
             (16, ir::types::I8X16),
             (8, ir::types::I64),

tests/disas/array-copy-inline.wat

@@ -67,12 +67,12 @@
 ;; @002a                               v56 = uadd_overflow_trap v44, v98, user2  ; v98 = 28
 ;; @002a                               v57 = icmp ugt v56, v50
 ;; @002a                               trapnz v57, user2
-;; @002a                               v58 = load.i8x16 notrap aligned v44
-;; @002a                               v59 = load.i64 notrap aligned v44+16
-;; @002a                               v60 = load.i32 notrap aligned v44+24
-;; @002a                               store notrap aligned v58, v25
-;; @002a                               store notrap aligned v59, v25+16
-;; @002a                               store notrap aligned v60, v25+24
+;; @002a                               v58 = load.i8x16 notrap aligned little v44
+;; @002a                               v59 = load.i64 notrap aligned little v44+16
+;; @002a                               v60 = load.i32 notrap aligned little v44+24
+;; @002a                               store notrap aligned little v58, v25
+;; @002a                               store notrap aligned little v59, v25+16
+;; @002a                               store notrap aligned little v60, v25+24
 ;; @002e                               jump block1
 ;;
 ;;                                 block1:

tests/disas/memory-copy-inline.wat

@@ -33,9 +33,9 @@
 ;; @0024                               trapnz v22, heap_oob
 ;; @0024                               v12 = load.i64 notrap aligned readonly can_move v0+56
 ;; @0024                               v26 = iadd v12, v18
-;; @0024                               v28 = load.i8x16 notrap aligned v26
+;; @0024                               v28 = load.i8x16 notrap aligned little v26
 ;; @0024                               v15 = iadd v12, v7
-;; @0024                               store notrap aligned v28, v15
+;; @0024                               store notrap aligned little v28, v15
 ;; @0028                               jump block1
 ;;
 ;;                                 block1: