fixup! [codegen] legalize imul.I64 for 32-bit ISAs

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

@@ -275,20 +275,22 @@ pub(crate) fn define(insts: &InstructionGroup, imm: &Immediates) -> TransformGro
 
     // TODO(ryzokuken): explore the perf diff w/ x86_umulx and consider have a
     // separate legalization for x86.
-    narrow.legalize(
-        def!(a = imul.I64(x, y)),
-        vec![
-            def!((xl, xh) = isplit(x)),
-            def!((yl, yh) = isplit(y)),
-            def!(al = imul(xl, yl)),
-            def!(a1 = imul(xh, yl)),
-            def!(a2 = imul(xl, yh)),
-            def!(a3 = umulhi(xl, yl)),
-            def!(a4 = iadd(a1, a2)),
-            def!(ah = iadd(a4, a3)),
-            def!(a = iconcat(al, ah)),
-        ],
-    );
+    for &(ty, ty_half) in &[(I64, I32), (I128, I64)] {
+        narrow.legalize(
+            def!(a = imul.ty(x, y)),
+            vec![
+                def!((xl, xh) = isplit.ty(x)),
+                def!((yl, yh) = isplit.ty(y)),
+                def!(al = imul.ty_half(xl, yl)),
+                def!(a1 = imul.ty_half(xh, yl)),
+                def!(a2 = imul.ty_half(xl, yh)),
+                def!(a3 = umulhi.ty_half(xl, yl)),
+                def!(a4 = iadd.ty_half(a1, a2)),
+                def!(ah = iadd.ty_half(a4, a3)),
+                def!(a = iconcat.ty_half(al, ah)),
+            ],
+        );
+    }
 
     // Widen instructions with one input operand.
     for &op in &[bnot, popcnt] {