Legalize load.i128 and store.i128

cranelift-codegen/meta/src/shared/legalize.rs

@@ -148,6 +148,7 @@ pub fn define(insts: &InstructionGroup, immediates: &OperandKinds) -> TransformG
     let ieee32 = immediates.by_name("ieee32");
     let ieee64 = immediates.by_name("ieee64");
     let intcc = immediates.by_name("intcc");
+    let offset32 = immediates.by_name("offset32");
 
     // List of variables to reuse in patterns.
     let x = var("x");
@@ -286,6 +287,30 @@ pub fn define(insts: &InstructionGroup, immediates: &OperandKinds) -> TransformG
         ],
     );
 
+    // FIXME generalize to any offset once offset+8 can be represented
+    narrow.legalize(
+        def!(a = load.I128(flags, ptr, Literal::constant(offset32, 0))),
+        vec![
+            def!(al = load.I64(flags, ptr, Literal::constant(offset32, 0))),
+            def!(ah = load.I64(flags, ptr, Literal::constant(offset32, 8))),
+            // `iconcat` expects the same byte order as stored in memory,
+            // so no need to swap depending on endianness.
+            def!(a = iconcat(al, ah)),
+        ],
+    );
+
+    // FIXME generalize to any offset once offset+8 can be represented
+    narrow.legalize(
+        def!(store.I128(flags, a, ptr, Literal::constant(offset32, 0))),
+        vec![
+            // `isplit` gives the same byte order as stored in memory,
+            // so no need to swap depending on endianness.
+            def!((al, ah) = isplit(a)),
+            def!(store.I64(flags, al, ptr, Literal::constant(offset32, 0))),
+            def!(store.I64(flags, ah, ptr, Literal::constant(offset32, 8))),
+        ],
+    );
+
     // Widen instructions with one input operand.
     for &op in &[bnot, popcnt] {
         for &int_ty in &[I8, I16] {

filetests/isa/x86/i128.clif

@@ -26,3 +26,21 @@ ebb0(v0: i128):
     ; check: v6 = x86_pop.i64
     ; check: return v3, v4, v6
 }
+
+function u0:1(i64, i128) fast {
+; check: ebb0(v0: i64 [%rdi], v2: i64 [%rsi], v3: i64 [%rdx], v4: i64 [%rbp]):
+ebb0(v0: i64, v1: i128):
+    ; check: store v2, v0
+    ; check: store v3, v0+8
+    store v1, v0
+    return
+}
+
+function u0:1(i64) -> i128 fast {
+ebb0(v0: i64):
+    ; check: v2 = load.i64 v0
+    ; check: v3 = load.i64 v0+8
+    v1 = load.i128 v0
+    ; check: return v2, v3, v5
+    return v1
+}