diff --git a/internal/cabi/_testdata/arch/amd64/composite.ll b/internal/cabi/_testdata/arch/amd64/composite.ll
new file mode 100644
index 00000000..696afdfd
--- /dev/null
+++ b/internal/cabi/_testdata/arch/amd64/composite.ll
@@ -0,0 +1,175 @@
+; ModuleID = '../../wrap/composite.c'
+source_filename = "../../wrap/composite.c"
+target datalayout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "amd64-unknown-linux-gnu"
+
+%struct.array10 = type { %struct.basearray1 }
+%struct.basearray1 = type { [1 x i32] }
+%struct.array11 = type { %struct.basearray1, i32 }
+%struct.point10 = type { %struct.basepoint1 }
+%struct.basepoint1 = type { i32 }
+%struct.point11 = type { %struct.basepoint1, i32 }
+%struct.array20 = type { %struct.basearray2 }
+%struct.basearray2 = type { [2 x i32] }
+%struct.array21 = type { %struct.basearray2, i32 }
+%struct.point20 = type { %struct.basepoint2 }
+%struct.basepoint2 = type { i32, i32 }
+%struct.point21 = type { %struct.basepoint2, i32 }
+
+; Function Attrs: noinline nounwind optnone uwtable
+define dso_local i32 @demo_array10(i32 %0) #0 {
+  %2 = alloca %struct.array10, align 4
+  %3 = alloca %struct.array10, align 4
+  %4 = getelementptr inbounds %struct.array10, %struct.array10* %3, i32 0, i32 0
+  %5 = getelementptr inbounds %struct.basearray1, %struct.basearray1* %4, i32 0, i32 0
+  %6 = bitcast [1 x i32]* %5 to i32*
+  store i32 %0, i32* %6, align 4
+  %7 = bitcast %struct.array10* %2 to i8*
+  %8 = bitcast %struct.array10* %3 to i8*
+  call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 4 %7, i8* align 4 %8, i64 4, i1 false)
+  %9 = getelementptr inbounds %struct.array10, %struct.array10* %2, i32 0, i32 0
+  %10 = getelementptr inbounds %struct.basearray1, %struct.basearray1* %9, i32 0, i32 0
+  %11 = bitcast [1 x i32]* %10 to i32*
+  %12 = load i32, i32* %11, align 4
+  ret i32 %12
+}
+
+; Function Attrs: argmemonly nofree nounwind willreturn
+declare void @llvm.memcpy.p0i8.p0i8.i64(i8* noalias nocapture writeonly, i8* noalias nocapture readonly, i64, i1 immarg) #1
+
+; Function Attrs: noinline nounwind optnone uwtable
+define dso_local i64 @demo_array11(i64 %0) #0 {
+  %2 = alloca %struct.array11, align 4
+  %3 = alloca %struct.array11, align 4
+  %4 = bitcast %struct.array11* %3 to i64*
+  store i64 %0, i64* %4, align 4
+  %5 = bitcast %struct.array11* %2 to i8*
+  %6 = bitcast %struct.array11* %3 to i8*
+  call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 4 %5, i8* align 4 %6, i64 8, i1 false)
+  %7 = bitcast %struct.array11* %2 to i64*
+  %8 = load i64, i64* %7, align 4
+  ret i64 %8
+}
+
+; Function Attrs: noinline nounwind optnone uwtable
+define dso_local i32 @demo_point10(i32 %0) #0 {
+  %2 = alloca %struct.point10, align 4
+  %3 = alloca %struct.point10, align 4
+  %4 = getelementptr inbounds %struct.point10, %struct.point10* %3, i32 0, i32 0
+  %5 = getelementptr inbounds %struct.basepoint1, %struct.basepoint1* %4, i32 0, i32 0
+  store i32 %0, i32* %5, align 4
+  %6 = bitcast %struct.point10* %2 to i8*
+  %7 = bitcast %struct.point10* %3 to i8*
+  call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 4 %6, i8* align 4 %7, i64 4, i1 false)
+  %8 = getelementptr inbounds %struct.point10, %struct.point10* %2, i32 0, i32 0
+  %9 = getelementptr inbounds %struct.basepoint1, %struct.basepoint1* %8, i32 0, i32 0
+  %10 = load i32, i32* %9, align 4
+  ret i32 %10
+}
+
+; Function Attrs: noinline nounwind optnone uwtable
+define dso_local i64 @demo_point11(i64 %0) #0 {
+  %2 = alloca %struct.point11, align 4
+  %3 = alloca %struct.point11, align 4
+  %4 = bitcast %struct.point11* %3 to i64*
+  store i64 %0, i64* %4, align 4
+  %5 = bitcast %struct.point11* %2 to i8*
+  %6 = bitcast %struct.point11* %3 to i8*
+  call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 4 %5, i8* align 4 %6, i64 8, i1 false)
+  %7 = bitcast %struct.point11* %2 to i64*
+  %8 = load i64, i64* %7, align 4
+  ret i64 %8
+}
+
+; Function Attrs: noinline nounwind optnone uwtable
+define dso_local i64 @demo_array20(i64 %0) #0 {
+  %2 = alloca %struct.array20, align 4
+  %3 = alloca %struct.array20, align 4
+  %4 = getelementptr inbounds %struct.array20, %struct.array20* %3, i32 0, i32 0
+  %5 = getelementptr inbounds %struct.basearray2, %struct.basearray2* %4, i32 0, i32 0
+  %6 = bitcast [2 x i32]* %5 to i64*
+  store i64 %0, i64* %6, align 4
+  %7 = bitcast %struct.array20* %2 to i8*
+  %8 = bitcast %struct.array20* %3 to i8*
+  call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 4 %7, i8* align 4 %8, i64 8, i1 false)
+  %9 = getelementptr inbounds %struct.array20, %struct.array20* %2, i32 0, i32 0
+  %10 = getelementptr inbounds %struct.basearray2, %struct.basearray2* %9, i32 0, i32 0
+  %11 = bitcast [2 x i32]* %10 to i64*
+  %12 = load i64, i64* %11, align 4
+  ret i64 %12
+}
+
+; Function Attrs: noinline nounwind optnone uwtable
+define dso_local { i64, i32 } @demo_array21(i64 %0, i32 %1) #0 {
+  %3 = alloca %struct.array21, align 4
+  %4 = alloca %struct.array21, align 4
+  %5 = alloca { i64, i32 }, align 4
+  %6 = alloca { i64, i32 }, align 8
+  %7 = getelementptr inbounds { i64, i32 }, { i64, i32 }* %5, i32 0, i32 0
+  store i64 %0, i64* %7, align 4
+  %8 = getelementptr inbounds { i64, i32 }, { i64, i32 }* %5, i32 0, i32 1
+  store i32 %1, i32* %8, align 4
+  %9 = bitcast %struct.array21* %4 to i8*
+  %10 = bitcast { i64, i32 }* %5 to i8*
+  call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 4 %9, i8* align 4 %10, i64 12, i1 false)
+  %11 = bitcast %struct.array21* %3 to i8*
+  %12 = bitcast %struct.array21* %4 to i8*
+  call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 4 %11, i8* align 4 %12, i64 12, i1 false)
+  %13 = bitcast { i64, i32 }* %6 to i8*
+  %14 = bitcast %struct.array21* %3 to i8*
+  call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 %13, i8* align 4 %14, i64 12, i1 false)
+  %15 = load { i64, i32 }, { i64, i32 }* %6, align 8
+  ret { i64, i32 } %15
+}
+
+; Function Attrs: noinline nounwind optnone uwtable
+define dso_local i64 @demo_point20(i64 %0) #0 {
+  %2 = alloca %struct.point20, align 4
+  %3 = alloca %struct.point20, align 4
+  %4 = getelementptr inbounds %struct.point20, %struct.point20* %3, i32 0, i32 0
+  %5 = bitcast %struct.basepoint2* %4 to i64*
+  store i64 %0, i64* %5, align 4
+  %6 = bitcast %struct.point20* %2 to i8*
+  %7 = bitcast %struct.point20* %3 to i8*
+  call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 4 %6, i8* align 4 %7, i64 8, i1 false)
+  %8 = getelementptr inbounds %struct.point20, %struct.point20* %2, i32 0, i32 0
+  %9 = bitcast %struct.basepoint2* %8 to i64*
+  %10 = load i64, i64* %9, align 4
+  ret i64 %10
+}
+
+; Function Attrs: noinline nounwind optnone uwtable
+define dso_local { i64, i32 } @demo_point21(i64 %0, i32 %1) #0 {
+  %3 = alloca %struct.point21, align 4
+  %4 = alloca %struct.point21, align 4
+  %5 = alloca { i64, i32 }, align 4
+  %6 = alloca { i64, i32 }, align 8
+  %7 = getelementptr inbounds { i64, i32 }, { i64, i32 }* %5, i32 0, i32 0
+  store i64 %0, i64* %7, align 4
+  %8 = getelementptr inbounds { i64, i32 }, { i64, i32 }* %5, i32 0, i32 1
+  store i32 %1, i32* %8, align 4
+  %9 = bitcast %struct.point21* %4 to i8*
+  %10 = bitcast { i64, i32 }* %5 to i8*
+  call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 4 %9, i8* align 4 %10, i64 12, i1 false)
+  %11 = bitcast %struct.point21* %3 to i8*
+  %12 = bitcast %struct.point21* %4 to i8*
+  call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 4 %11, i8* align 4 %12, i64 12, i1 false)
+  %13 = bitcast { i64, i32 }* %6 to i8*
+  %14 = bitcast %struct.point21* %3 to i8*
+  call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 %13, i8* align 4 %14, i64 12, i1 false)
+  %15 = load { i64, i32 }, { i64, i32 }* %6, align 8
+  ret { i64, i32 } %15
+}
+
+attributes #0 = { noinline nounwind optnone uwtable "frame-pointer"="all" "min-legal-vector-width"="0" "no-trapping-math"="true" "stack-protector-buffer-size"="8" "target-cpu"="x86-64" "target-features"="+cx8,+fxsr,+mmx,+sse,+sse2,+x87" "tune-cpu"="generic" }
+attributes #1 = { argmemonly nofree nounwind willreturn }
+
+!llvm.module.flags = !{!0, !1, !2, !3, !4}
+!llvm.ident = !{!5}
+
+!0 = !{i32 1, !"wchar_size", i32 4}
+!1 = !{i32 7, !"PIC Level", i32 2}
+!2 = !{i32 7, !"PIE Level", i32 2}
+!3 = !{i32 7, !"uwtable", i32 2}
+!4 = !{i32 7, !"frame-pointer", i32 2}
+!5 = !{!"Apple clang version 14.0.3 (clang-1403.0.22.14.1)"}
diff --git a/internal/cabi/_testdata/arch/arm64/composite.ll b/internal/cabi/_testdata/arch/arm64/composite.ll
new file mode 100644
index 00000000..fb98e086
--- /dev/null
+++ b/internal/cabi/_testdata/arch/arm64/composite.ll
@@ -0,0 +1,177 @@
+; ModuleID = '../../wrap/composite.c'
+source_filename = "../../wrap/composite.c"
+target datalayout = "e-m:e-i8:8:32-i16:16:32-i64:64-i128:128-n32:64-S128"
+target triple = "aarch64-unknown-linux-gnu"
+
+%struct.array10 = type { %struct.basearray1 }
+%struct.basearray1 = type { [1 x i32] }
+%struct.array11 = type { %struct.basearray1, i32 }
+%struct.point10 = type { %struct.basepoint1 }
+%struct.basepoint1 = type { i32 }
+%struct.point11 = type { %struct.basepoint1, i32 }
+%struct.array20 = type { %struct.basearray2 }
+%struct.basearray2 = type { [2 x i32] }
+%struct.array21 = type { %struct.basearray2, i32 }
+%struct.point20 = type { %struct.basepoint2 }
+%struct.basepoint2 = type { i32, i32 }
+%struct.point21 = type { %struct.basepoint2, i32 }
+
+; Function Attrs: noinline nounwind optnone uwtable
+define dso_local i32 @demo_array10(i64 %0) #0 {
+  %2 = alloca %struct.array10, align 4
+  %3 = alloca %struct.array10, align 4
+  %4 = alloca i64, align 8
+  %5 = getelementptr inbounds %struct.array10, %struct.array10* %3, i32 0, i32 0
+  %6 = getelementptr inbounds %struct.basearray1, %struct.basearray1* %5, i32 0, i32 0
+  store i64 %0, i64* %4, align 8
+  %7 = bitcast [1 x i32]* %6 to i8*
+  %8 = bitcast i64* %4 to i8*
+  call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 4 %7, i8* align 8 %8, i64 4, i1 false)
+  %9 = bitcast %struct.array10* %2 to i8*
+  %10 = bitcast %struct.array10* %3 to i8*
+  call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 4 %9, i8* align 4 %10, i64 4, i1 false)
+  %11 = getelementptr inbounds %struct.array10, %struct.array10* %2, i32 0, i32 0
+  %12 = getelementptr inbounds %struct.basearray1, %struct.basearray1* %11, i32 0, i32 0
+  %13 = bitcast [1 x i32]* %12 to i32*
+  %14 = load i32, i32* %13, align 4
+  ret i32 %14
+}
+
+; Function Attrs: argmemonly nofree nounwind willreturn
+declare void @llvm.memcpy.p0i8.p0i8.i64(i8* noalias nocapture writeonly, i8* noalias nocapture readonly, i64, i1 immarg) #1
+
+; Function Attrs: noinline nounwind optnone uwtable
+define dso_local i64 @demo_array11(i64 %0) #0 {
+  %2 = alloca %struct.array11, align 4
+  %3 = alloca %struct.array11, align 4
+  %4 = bitcast %struct.array11* %3 to i64*
+  store i64 %0, i64* %4, align 4
+  %5 = bitcast %struct.array11* %2 to i8*
+  %6 = bitcast %struct.array11* %3 to i8*
+  call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 4 %5, i8* align 4 %6, i64 8, i1 false)
+  %7 = bitcast %struct.array11* %2 to i64*
+  %8 = load i64, i64* %7, align 4
+  ret i64 %8
+}
+
+; Function Attrs: noinline nounwind optnone uwtable
+define dso_local i32 @demo_point10(i64 %0) #0 {
+  %2 = alloca %struct.point10, align 4
+  %3 = alloca %struct.point10, align 4
+  %4 = getelementptr inbounds %struct.point10, %struct.point10* %3, i32 0, i32 0
+  %5 = getelementptr inbounds %struct.basepoint1, %struct.basepoint1* %4, i32 0, i32 0
+  %6 = trunc i64 %0 to i32
+  store i32 %6, i32* %5, align 4
+  %7 = bitcast %struct.point10* %2 to i8*
+  %8 = bitcast %struct.point10* %3 to i8*
+  call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 4 %7, i8* align 4 %8, i64 4, i1 false)
+  %9 = getelementptr inbounds %struct.point10, %struct.point10* %2, i32 0, i32 0
+  %10 = getelementptr inbounds %struct.basepoint1, %struct.basepoint1* %9, i32 0, i32 0
+  %11 = load i32, i32* %10, align 4
+  ret i32 %11
+}
+
+; Function Attrs: noinline nounwind optnone uwtable
+define dso_local i64 @demo_point11(i64 %0) #0 {
+  %2 = alloca %struct.point11, align 4
+  %3 = alloca %struct.point11, align 4
+  %4 = bitcast %struct.point11* %3 to i64*
+  store i64 %0, i64* %4, align 4
+  %5 = bitcast %struct.point11* %2 to i8*
+  %6 = bitcast %struct.point11* %3 to i8*
+  call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 4 %5, i8* align 4 %6, i64 8, i1 false)
+  %7 = bitcast %struct.point11* %2 to i64*
+  %8 = load i64, i64* %7, align 4
+  ret i64 %8
+}
+
+; Function Attrs: noinline nounwind optnone uwtable
+define dso_local i64 @demo_array20(i64 %0) #0 {
+  %2 = alloca %struct.array20, align 4
+  %3 = alloca %struct.array20, align 4
+  %4 = getelementptr inbounds %struct.array20, %struct.array20* %3, i32 0, i32 0
+  %5 = getelementptr inbounds %struct.basearray2, %struct.basearray2* %4, i32 0, i32 0
+  %6 = bitcast [2 x i32]* %5 to i64*
+  store i64 %0, i64* %6, align 4
+  %7 = bitcast %struct.array20* %2 to i8*
+  %8 = bitcast %struct.array20* %3 to i8*
+  call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 4 %7, i8* align 4 %8, i64 8, i1 false)
+  %9 = getelementptr inbounds %struct.array20, %struct.array20* %2, i32 0, i32 0
+  %10 = getelementptr inbounds %struct.basearray2, %struct.basearray2* %9, i32 0, i32 0
+  %11 = bitcast [2 x i32]* %10 to i64*
+  %12 = load i64, i64* %11, align 4
+  ret i64 %12
+}
+
+; Function Attrs: noinline nounwind optnone uwtable
+define dso_local [2 x i64] @demo_array21([2 x i64] %0) #0 {
+  %2 = alloca %struct.array21, align 4
+  %3 = alloca %struct.array21, align 4
+  %4 = alloca [2 x i64], align 8
+  %5 = alloca [2 x i64], align 8
+  store [2 x i64] %0, [2 x i64]* %4, align 8
+  %6 = bitcast %struct.array21* %3 to i8*
+  %7 = bitcast [2 x i64]* %4 to i8*
+  call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 4 %6, i8* align 8 %7, i64 12, i1 false)
+  %8 = bitcast %struct.array21* %2 to i8*
+  %9 = bitcast %struct.array21* %3 to i8*
+  call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 4 %8, i8* align 4 %9, i64 12, i1 false)
+  %10 = bitcast [2 x i64]* %5 to i8*
+  %11 = bitcast %struct.array21* %2 to i8*
+  call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 %10, i8* align 4 %11, i64 12, i1 false)
+  %12 = load [2 x i64], [2 x i64]* %5, align 8
+  ret [2 x i64] %12
+}
+
+; Function Attrs: noinline nounwind optnone uwtable
+define dso_local i64 @demo_point20(i64 %0) #0 {
+  %2 = alloca %struct.point20, align 4
+  %3 = alloca %struct.point20, align 4
+  %4 = getelementptr inbounds %struct.point20, %struct.point20* %3, i32 0, i32 0
+  %5 = bitcast %struct.basepoint2* %4 to i64*
+  store i64 %0, i64* %5, align 4
+  %6 = bitcast %struct.point20* %2 to i8*
+  %7 = bitcast %struct.point20* %3 to i8*
+  call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 4 %6, i8* align 4 %7, i64 8, i1 false)
+  %8 = getelementptr inbounds %struct.point20, %struct.point20* %2, i32 0, i32 0
+  %9 = bitcast %struct.basepoint2* %8 to i64*
+  %10 = load i64, i64* %9, align 4
+  ret i64 %10
+}
+
+; Function Attrs: noinline nounwind optnone uwtable
+define dso_local [2 x i64] @demo_point21([2 x i64] %0) #0 {
+  %2 = alloca %struct.point21, align 4
+  %3 = alloca %struct.point21, align 4
+  %4 = alloca [2 x i64], align 8
+  %5 = alloca [2 x i64], align 8
+  store [2 x i64] %0, [2 x i64]* %4, align 8
+  %6 = bitcast %struct.point21* %3 to i8*
+  %7 = bitcast [2 x i64]* %4 to i8*
+  call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 4 %6, i8* align 8 %7, i64 12, i1 false)
+  %8 = bitcast %struct.point21* %2 to i8*
+  %9 = bitcast %struct.point21* %3 to i8*
+  call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 4 %8, i8* align 4 %9, i64 12, i1 false)
+  %10 = bitcast [2 x i64]* %5 to i8*
+  %11 = bitcast %struct.point21* %2 to i8*
+  call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 %10, i8* align 4 %11, i64 12, i1 false)
+  %12 = load [2 x i64], [2 x i64]* %5, align 8
+  ret [2 x i64] %12
+}
+
+attributes #0 = { noinline nounwind optnone uwtable "frame-pointer"="non-leaf" "min-legal-vector-width"="0" "no-trapping-math"="true" "stack-protector-buffer-size"="8" "target-cpu"="generic" "target-features"="+neon,+v8a" }
+attributes #1 = { argmemonly nofree nounwind willreturn }
+
+!llvm.module.flags = !{!0, !1, !2, !3, !4, !5, !6, !7, !8}
+!llvm.ident = !{!9}
+
+!0 = !{i32 1, !"wchar_size", i32 4}
+!1 = !{i32 8, !"branch-target-enforcement", i32 0}
+!2 = !{i32 8, !"sign-return-address", i32 0}
+!3 = !{i32 8, !"sign-return-address-all", i32 0}
+!4 = !{i32 8, !"sign-return-address-with-bkey", i32 0}
+!5 = !{i32 7, !"PIC Level", i32 2}
+!6 = !{i32 7, !"PIE Level", i32 2}
+!7 = !{i32 7, !"uwtable", i32 2}
+!8 = !{i32 7, !"frame-pointer", i32 1}
+!9 = !{!"Apple clang version 14.0.3 (clang-1403.0.22.14.1)"}
diff --git a/internal/cabi/_testdata/arch/armv6/composite.ll b/internal/cabi/_testdata/arch/armv6/composite.ll
new file mode 100644
index 00000000..d677b6f8
--- /dev/null
+++ b/internal/cabi/_testdata/arch/armv6/composite.ll
@@ -0,0 +1,139 @@
+; ModuleID = '../../wrap/composite.c'
+source_filename = "../../wrap/composite.c"
+target datalayout = "e-m:e-p:32:32-Fi8-i64:64-v128:64:128-a:0:32-n32-S64"
+target triple = "armv6kz-unknown-linux-gnueabihf"
+
+%struct.array10 = type { %struct.basearray1 }
+%struct.basearray1 = type { [1 x i32] }
+%struct.array11 = type { %struct.basearray1, i32 }
+%struct.point10 = type { %struct.basepoint1 }
+%struct.basepoint1 = type { i32 }
+%struct.point11 = type { %struct.basepoint1, i32 }
+%struct.array20 = type { %struct.basearray2 }
+%struct.basearray2 = type { [2 x i32] }
+%struct.array21 = type { %struct.basearray2, i32 }
+%struct.point20 = type { %struct.basepoint2 }
+%struct.basepoint2 = type { i32, i32 }
+%struct.point21 = type { %struct.basepoint2, i32 }
+
+; Function Attrs: noinline nounwind optnone
+define dso_local i32 @demo_array10([1 x i32] %0) #0 {
+  %2 = alloca %struct.array10, align 4
+  %3 = alloca %struct.array10, align 4
+  %4 = getelementptr inbounds %struct.array10, %struct.array10* %3, i32 0, i32 0
+  %5 = getelementptr inbounds %struct.basearray1, %struct.basearray1* %4, i32 0, i32 0
+  store [1 x i32] %0, [1 x i32]* %5, align 4
+  %6 = bitcast %struct.array10* %2 to i8*
+  %7 = bitcast %struct.array10* %3 to i8*
+  call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 %6, i8* align 4 %7, i32 4, i1 false)
+  %8 = getelementptr inbounds %struct.array10, %struct.array10* %2, i32 0, i32 0
+  %9 = getelementptr inbounds %struct.basearray1, %struct.basearray1* %8, i32 0, i32 0
+  %10 = bitcast [1 x i32]* %9 to i32*
+  %11 = load i32, i32* %10, align 4
+  ret i32 %11
+}
+
+; Function Attrs: argmemonly nofree nounwind willreturn
+declare void @llvm.memcpy.p0i8.p0i8.i32(i8* noalias nocapture writeonly, i8* noalias nocapture readonly, i32, i1 immarg) #1
+
+; Function Attrs: noinline nounwind optnone
+define dso_local void @demo_array11(%struct.array11* noalias sret(%struct.array11) align 4 %0, [2 x i32] %1) #0 {
+  %3 = alloca %struct.array11, align 4
+  %4 = bitcast %struct.array11* %3 to [2 x i32]*
+  store [2 x i32] %1, [2 x i32]* %4, align 4
+  %5 = bitcast %struct.array11* %0 to i8*
+  %6 = bitcast %struct.array11* %3 to i8*
+  call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 %5, i8* align 4 %6, i32 8, i1 false)
+  ret void
+}
+
+; Function Attrs: noinline nounwind optnone
+define dso_local i32 @demo_point10([1 x i32] %0) #0 {
+  %2 = alloca %struct.point10, align 4
+  %3 = alloca %struct.point10, align 4
+  %4 = getelementptr inbounds %struct.point10, %struct.point10* %3, i32 0, i32 0
+  %5 = getelementptr inbounds %struct.basepoint1, %struct.basepoint1* %4, i32 0, i32 0
+  %6 = bitcast i32* %5 to [1 x i32]*
+  store [1 x i32] %0, [1 x i32]* %6, align 4
+  %7 = bitcast %struct.point10* %2 to i8*
+  %8 = bitcast %struct.point10* %3 to i8*
+  call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 %7, i8* align 4 %8, i32 4, i1 false)
+  %9 = getelementptr inbounds %struct.point10, %struct.point10* %2, i32 0, i32 0
+  %10 = getelementptr inbounds %struct.basepoint1, %struct.basepoint1* %9, i32 0, i32 0
+  %11 = load i32, i32* %10, align 4
+  ret i32 %11
+}
+
+; Function Attrs: noinline nounwind optnone
+define dso_local void @demo_point11(%struct.point11* noalias sret(%struct.point11) align 4 %0, [2 x i32] %1) #0 {
+  %3 = alloca %struct.point11, align 4
+  %4 = bitcast %struct.point11* %3 to [2 x i32]*
+  store [2 x i32] %1, [2 x i32]* %4, align 4
+  %5 = bitcast %struct.point11* %0 to i8*
+  %6 = bitcast %struct.point11* %3 to i8*
+  call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 %5, i8* align 4 %6, i32 8, i1 false)
+  ret void
+}
+
+; Function Attrs: noinline nounwind optnone
+define dso_local void @demo_array20(%struct.array20* noalias sret(%struct.array20) align 4 %0, [2 x i32] %1) #0 {
+  %3 = alloca %struct.array20, align 4
+  %4 = getelementptr inbounds %struct.array20, %struct.array20* %3, i32 0, i32 0
+  %5 = getelementptr inbounds %struct.basearray2, %struct.basearray2* %4, i32 0, i32 0
+  store [2 x i32] %1, [2 x i32]* %5, align 4
+  %6 = bitcast %struct.array20* %0 to i8*
+  %7 = bitcast %struct.array20* %3 to i8*
+  call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 %6, i8* align 4 %7, i32 8, i1 false)
+  ret void
+}
+
+; Function Attrs: noinline nounwind optnone
+define dso_local void @demo_array21(%struct.array21* noalias sret(%struct.array21) align 4 %0, [3 x i32] %1) #0 {
+  %3 = alloca %struct.array21, align 4
+  %4 = bitcast %struct.array21* %3 to [3 x i32]*
+  store [3 x i32] %1, [3 x i32]* %4, align 4
+  %5 = bitcast %struct.array21* %0 to i8*
+  %6 = bitcast %struct.array21* %3 to i8*
+  call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 %5, i8* align 4 %6, i32 12, i1 false)
+  ret void
+}
+
+; Function Attrs: noinline nounwind optnone
+define dso_local void @demo_point20(%struct.point20* noalias sret(%struct.point20) align 4 %0, [2 x i32] %1) #0 {
+  %3 = alloca %struct.point20, align 4
+  %4 = getelementptr inbounds %struct.point20, %struct.point20* %3, i32 0, i32 0
+  %5 = bitcast %struct.basepoint2* %4 to [2 x i32]*
+  store [2 x i32] %1, [2 x i32]* %5, align 4
+  %6 = bitcast %struct.point20* %0 to i8*
+  %7 = bitcast %struct.point20* %3 to i8*
+  call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 %6, i8* align 4 %7, i32 8, i1 false)
+  ret void
+}
+
+; Function Attrs: noinline nounwind optnone
+define dso_local void @demo_point21(%struct.point21* noalias sret(%struct.point21) align 4 %0, [3 x i32] %1) #0 {
+  %3 = alloca %struct.point21, align 4
+  %4 = bitcast %struct.point21* %3 to [3 x i32]*
+  store [3 x i32] %1, [3 x i32]* %4, align 4
+  %5 = bitcast %struct.point21* %0 to i8*
+  %6 = bitcast %struct.point21* %3 to i8*
+  call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 %5, i8* align 4 %6, i32 12, i1 false)
+  ret void
+}
+
+attributes #0 = { noinline nounwind optnone "frame-pointer"="all" "min-legal-vector-width"="0" "no-trapping-math"="true" "stack-protector-buffer-size"="8" "target-cpu"="arm1176jzf-s" "target-features"="+armv6kz,+dsp,+fp64,+strict-align,+vfp2,+vfp2sp,-aes,-d32,-fp-armv8,-fp-armv8d16,-fp-armv8d16sp,-fp-armv8sp,-fp16,-fp16fml,-fullfp16,-neon,-sha2,-thumb-mode,-vfp3,-vfp3d16,-vfp3d16sp,-vfp3sp,-vfp4,-vfp4d16,-vfp4d16sp,-vfp4sp" }
+attributes #1 = { argmemonly nofree nounwind willreturn }
+
+!llvm.module.flags = !{!0, !1, !2, !3, !4, !5, !6, !7, !8}
+!llvm.ident = !{!9}
+
+!0 = !{i32 1, !"wchar_size", i32 4}
+!1 = !{i32 1, !"min_enum_size", i32 4}
+!2 = !{i32 8, !"branch-target-enforcement", i32 0}
+!3 = !{i32 8, !"sign-return-address", i32 0}
+!4 = !{i32 8, !"sign-return-address-all", i32 0}
+!5 = !{i32 8, !"sign-return-address-with-bkey", i32 0}
+!6 = !{i32 7, !"PIC Level", i32 2}
+!7 = !{i32 7, !"PIE Level", i32 2}
+!8 = !{i32 7, !"frame-pointer", i32 2}
+!9 = !{!"Apple clang version 14.0.3 (clang-1403.0.22.14.1)"}
diff --git a/internal/cabi/_testdata/arch/i386/composite.ll b/internal/cabi/_testdata/arch/i386/composite.ll
new file mode 100644
index 00000000..1a8a6c30
--- /dev/null
+++ b/internal/cabi/_testdata/arch/i386/composite.ll
@@ -0,0 +1,122 @@
+; ModuleID = '../../wrap/composite.c'
+source_filename = "../../wrap/composite.c"
+target datalayout = "e-m:e-p:32:32-p270:32:32-p271:32:32-p272:64:64-f64:32:64-f80:32-n8:16:32-S128"
+target triple = "i386-unknown-linux-gnu"
+
+%struct.array10 = type { %struct.basearray1 }
+%struct.basearray1 = type { [1 x i32] }
+%struct.array11 = type { %struct.basearray1, i32 }
+%struct.point10 = type { %struct.basepoint1 }
+%struct.basepoint1 = type { i32 }
+%struct.point11 = type { %struct.basepoint1, i32 }
+%struct.array20 = type { %struct.basearray2 }
+%struct.basearray2 = type { [2 x i32] }
+%struct.array21 = type { %struct.basearray2, i32 }
+%struct.point20 = type { %struct.basepoint2 }
+%struct.basepoint2 = type { i32, i32 }
+%struct.point21 = type { %struct.basepoint2, i32 }
+
+; Function Attrs: noinline nounwind optnone uwtable
+define dso_local void @demo_array10(%struct.array10* noalias sret(%struct.array10) align 4 %0, %struct.array10* noundef byval(%struct.array10) align 4 %1) #0 {
+  %3 = alloca i8*, align 4
+  %4 = bitcast %struct.array10* %0 to i8*
+  store i8* %4, i8** %3, align 4
+  %5 = bitcast %struct.array10* %0 to i8*
+  %6 = bitcast %struct.array10* %1 to i8*
+  call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 %5, i8* align 4 %6, i32 4, i1 false)
+  ret void
+}
+
+; Function Attrs: argmemonly nofree nounwind willreturn
+declare void @llvm.memcpy.p0i8.p0i8.i32(i8* noalias nocapture writeonly, i8* noalias nocapture readonly, i32, i1 immarg) #1
+
+; Function Attrs: noinline nounwind optnone uwtable
+define dso_local void @demo_array11(%struct.array11* noalias sret(%struct.array11) align 4 %0, %struct.array11* noundef byval(%struct.array11) align 4 %1) #0 {
+  %3 = alloca i8*, align 4
+  %4 = bitcast %struct.array11* %0 to i8*
+  store i8* %4, i8** %3, align 4
+  %5 = bitcast %struct.array11* %0 to i8*
+  %6 = bitcast %struct.array11* %1 to i8*
+  call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 %5, i8* align 4 %6, i32 8, i1 false)
+  ret void
+}
+
+; Function Attrs: noinline nounwind optnone uwtable
+define dso_local void @demo_point10(%struct.point10* noalias sret(%struct.point10) align 4 %0, %struct.point10* noundef byval(%struct.point10) align 4 %1) #0 {
+  %3 = alloca i8*, align 4
+  %4 = bitcast %struct.point10* %0 to i8*
+  store i8* %4, i8** %3, align 4
+  %5 = bitcast %struct.point10* %0 to i8*
+  %6 = bitcast %struct.point10* %1 to i8*
+  call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 %5, i8* align 4 %6, i32 4, i1 false)
+  ret void
+}
+
+; Function Attrs: noinline nounwind optnone uwtable
+define dso_local void @demo_point11(%struct.point11* noalias sret(%struct.point11) align 4 %0, %struct.point11* noundef byval(%struct.point11) align 4 %1) #0 {
+  %3 = alloca i8*, align 4
+  %4 = bitcast %struct.point11* %0 to i8*
+  store i8* %4, i8** %3, align 4
+  %5 = bitcast %struct.point11* %0 to i8*
+  %6 = bitcast %struct.point11* %1 to i8*
+  call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 %5, i8* align 4 %6, i32 8, i1 false)
+  ret void
+}
+
+; Function Attrs: noinline nounwind optnone uwtable
+define dso_local void @demo_array20(%struct.array20* noalias sret(%struct.array20) align 4 %0, %struct.array20* noundef byval(%struct.array20) align 4 %1) #0 {
+  %3 = alloca i8*, align 4
+  %4 = bitcast %struct.array20* %0 to i8*
+  store i8* %4, i8** %3, align 4
+  %5 = bitcast %struct.array20* %0 to i8*
+  %6 = bitcast %struct.array20* %1 to i8*
+  call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 %5, i8* align 4 %6, i32 8, i1 false)
+  ret void
+}
+
+; Function Attrs: noinline nounwind optnone uwtable
+define dso_local void @demo_array21(%struct.array21* noalias sret(%struct.array21) align 4 %0, %struct.array21* noundef byval(%struct.array21) align 4 %1) #0 {
+  %3 = alloca i8*, align 4
+  %4 = bitcast %struct.array21* %0 to i8*
+  store i8* %4, i8** %3, align 4
+  %5 = bitcast %struct.array21* %0 to i8*
+  %6 = bitcast %struct.array21* %1 to i8*
+  call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 %5, i8* align 4 %6, i32 12, i1 false)
+  ret void
+}
+
+; Function Attrs: noinline nounwind optnone uwtable
+define dso_local void @demo_point20(%struct.point20* noalias sret(%struct.point20) align 4 %0, %struct.point20* noundef byval(%struct.point20) align 4 %1) #0 {
+  %3 = alloca i8*, align 4
+  %4 = bitcast %struct.point20* %0 to i8*
+  store i8* %4, i8** %3, align 4
+  %5 = bitcast %struct.point20* %0 to i8*
+  %6 = bitcast %struct.point20* %1 to i8*
+  call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 %5, i8* align 4 %6, i32 8, i1 false)
+  ret void
+}
+
+; Function Attrs: noinline nounwind optnone uwtable
+define dso_local void @demo_point21(%struct.point21* noalias sret(%struct.point21) align 4 %0, %struct.point21* noundef byval(%struct.point21) align 4 %1) #0 {
+  %3 = alloca i8*, align 4
+  %4 = bitcast %struct.point21* %0 to i8*
+  store i8* %4, i8** %3, align 4
+  %5 = bitcast %struct.point21* %0 to i8*
+  %6 = bitcast %struct.point21* %1 to i8*
+  call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 %5, i8* align 4 %6, i32 12, i1 false)
+  ret void
+}
+
+attributes #0 = { noinline nounwind optnone uwtable "frame-pointer"="all" "min-legal-vector-width"="0" "no-trapping-math"="true" "stack-protector-buffer-size"="8" "target-cpu"="pentium4" "target-features"="+cx8,+fxsr,+mmx,+sse,+sse2,+x87" "tune-cpu"="generic" }
+attributes #1 = { argmemonly nofree nounwind willreturn }
+
+!llvm.module.flags = !{!0, !1, !2, !3, !4, !5}
+!llvm.ident = !{!6}
+
+!0 = !{i32 1, !"NumRegisterParameters", i32 0}
+!1 = !{i32 1, !"wchar_size", i32 4}
+!2 = !{i32 7, !"PIC Level", i32 2}
+!3 = !{i32 7, !"PIE Level", i32 2}
+!4 = !{i32 7, !"uwtable", i32 2}
+!5 = !{i32 7, !"frame-pointer", i32 2}
+!6 = !{!"Apple clang version 14.0.3 (clang-1403.0.22.14.1)"}
diff --git a/internal/cabi/_testdata/arch/riscv64/composite.ll b/internal/cabi/_testdata/arch/riscv64/composite.ll
new file mode 100644
index 00000000..1c7040d8
--- /dev/null
+++ b/internal/cabi/_testdata/arch/riscv64/composite.ll
@@ -0,0 +1,176 @@
+; ModuleID = '../../wrap/composite.c'
+source_filename = "../../wrap/composite.c"
+target datalayout = "e-m:e-p:64:64-i64:64-i128:128-n64-S128"
+target triple = "riscv64-unknown-unknown-elf"
+
+%struct.array10 = type { %struct.basearray1 }
+%struct.basearray1 = type { [1 x i32] }
+%struct.array11 = type { %struct.basearray1, i32 }
+%struct.point10 = type { %struct.basepoint1 }
+%struct.basepoint1 = type { i32 }
+%struct.point11 = type { %struct.basepoint1, i32 }
+%struct.array20 = type { %struct.basearray2 }
+%struct.basearray2 = type { [2 x i32] }
+%struct.array21 = type { %struct.basearray2, i32 }
+%struct.point20 = type { %struct.basepoint2 }
+%struct.basepoint2 = type { i32, i32 }
+%struct.point21 = type { %struct.basepoint2, i32 }
+
+; Function Attrs: noinline nounwind optnone
+define dso_local i64 @demo_array10(i64 %0) #0 {
+  %2 = alloca %struct.array10, align 4
+  %3 = alloca %struct.array10, align 4
+  %4 = alloca i64, align 8
+  %5 = alloca i64, align 8
+  %6 = getelementptr inbounds %struct.array10, %struct.array10* %3, i32 0, i32 0
+  %7 = getelementptr inbounds %struct.basearray1, %struct.basearray1* %6, i32 0, i32 0
+  store i64 %0, i64* %4, align 8
+  %8 = bitcast [1 x i32]* %7 to i8*
+  %9 = bitcast i64* %4 to i8*
+  call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 4 %8, i8* align 8 %9, i64 4, i1 false)
+  %10 = bitcast %struct.array10* %2 to i8*
+  %11 = bitcast %struct.array10* %3 to i8*
+  call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 4 %10, i8* align 4 %11, i64 4, i1 false)
+  %12 = getelementptr inbounds %struct.array10, %struct.array10* %2, i32 0, i32 0
+  %13 = getelementptr inbounds %struct.basearray1, %struct.basearray1* %12, i32 0, i32 0
+  %14 = bitcast i64* %5 to i8*
+  %15 = bitcast [1 x i32]* %13 to i8*
+  call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 %14, i8* align 4 %15, i64 4, i1 false)
+  %16 = load i64, i64* %5, align 8
+  ret i64 %16
+}
+
+; Function Attrs: argmemonly nofree nounwind willreturn
+declare void @llvm.memcpy.p0i8.p0i8.i64(i8* noalias nocapture writeonly, i8* noalias nocapture readonly, i64, i1 immarg) #1
+
+; Function Attrs: noinline nounwind optnone
+define dso_local i64 @demo_array11(i64 %0) #0 {
+  %2 = alloca %struct.array11, align 4
+  %3 = alloca %struct.array11, align 4
+  %4 = bitcast %struct.array11* %3 to i64*
+  store i64 %0, i64* %4, align 4
+  %5 = bitcast %struct.array11* %2 to i8*
+  %6 = bitcast %struct.array11* %3 to i8*
+  call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 4 %5, i8* align 4 %6, i64 8, i1 false)
+  %7 = bitcast %struct.array11* %2 to i64*
+  %8 = load i64, i64* %7, align 4
+  ret i64 %8
+}
+
+; Function Attrs: noinline nounwind optnone
+define dso_local i64 @demo_point10(i64 %0) #0 {
+  %2 = alloca %struct.point10, align 4
+  %3 = alloca %struct.point10, align 4
+  %4 = getelementptr inbounds %struct.point10, %struct.point10* %3, i32 0, i32 0
+  %5 = getelementptr inbounds %struct.basepoint1, %struct.basepoint1* %4, i32 0, i32 0
+  %6 = trunc i64 %0 to i32
+  store i32 %6, i32* %5, align 4
+  %7 = bitcast %struct.point10* %2 to i8*
+  %8 = bitcast %struct.point10* %3 to i8*
+  call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 4 %7, i8* align 4 %8, i64 4, i1 false)
+  %9 = getelementptr inbounds %struct.point10, %struct.point10* %2, i32 0, i32 0
+  %10 = getelementptr inbounds %struct.basepoint1, %struct.basepoint1* %9, i32 0, i32 0
+  %11 = load i32, i32* %10, align 4
+  %12 = zext i32 %11 to i64
+  ret i64 %12
+}
+
+; Function Attrs: noinline nounwind optnone
+define dso_local i64 @demo_point11(i64 %0) #0 {
+  %2 = alloca %struct.point11, align 4
+  %3 = alloca %struct.point11, align 4
+  %4 = bitcast %struct.point11* %3 to i64*
+  store i64 %0, i64* %4, align 4
+  %5 = bitcast %struct.point11* %2 to i8*
+  %6 = bitcast %struct.point11* %3 to i8*
+  call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 4 %5, i8* align 4 %6, i64 8, i1 false)
+  %7 = bitcast %struct.point11* %2 to i64*
+  %8 = load i64, i64* %7, align 4
+  ret i64 %8
+}
+
+; Function Attrs: noinline nounwind optnone
+define dso_local i64 @demo_array20(i64 %0) #0 {
+  %2 = alloca %struct.array20, align 4
+  %3 = alloca %struct.array20, align 4
+  %4 = getelementptr inbounds %struct.array20, %struct.array20* %3, i32 0, i32 0
+  %5 = getelementptr inbounds %struct.basearray2, %struct.basearray2* %4, i32 0, i32 0
+  %6 = bitcast [2 x i32]* %5 to i64*
+  store i64 %0, i64* %6, align 4
+  %7 = bitcast %struct.array20* %2 to i8*
+  %8 = bitcast %struct.array20* %3 to i8*
+  call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 4 %7, i8* align 4 %8, i64 8, i1 false)
+  %9 = getelementptr inbounds %struct.array20, %struct.array20* %2, i32 0, i32 0
+  %10 = getelementptr inbounds %struct.basearray2, %struct.basearray2* %9, i32 0, i32 0
+  %11 = bitcast [2 x i32]* %10 to i64*
+  %12 = load i64, i64* %11, align 4
+  ret i64 %12
+}
+
+; Function Attrs: noinline nounwind optnone
+define dso_local [2 x i64] @demo_array21([2 x i64] %0) #0 {
+  %2 = alloca %struct.array21, align 4
+  %3 = alloca %struct.array21, align 4
+  %4 = alloca [2 x i64], align 8
+  %5 = alloca [2 x i64], align 8
+  store [2 x i64] %0, [2 x i64]* %4, align 8
+  %6 = bitcast %struct.array21* %3 to i8*
+  %7 = bitcast [2 x i64]* %4 to i8*
+  call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 4 %6, i8* align 8 %7, i64 12, i1 false)
+  %8 = bitcast %struct.array21* %2 to i8*
+  %9 = bitcast %struct.array21* %3 to i8*
+  call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 4 %8, i8* align 4 %9, i64 12, i1 false)
+  %10 = bitcast [2 x i64]* %5 to i8*
+  %11 = bitcast %struct.array21* %2 to i8*
+  call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 %10, i8* align 4 %11, i64 12, i1 false)
+  %12 = load [2 x i64], [2 x i64]* %5, align 8
+  ret [2 x i64] %12
+}
+
+; Function Attrs: noinline nounwind optnone
+define dso_local i64 @demo_point20(i64 %0) #0 {
+  %2 = alloca %struct.point20, align 4
+  %3 = alloca %struct.point20, align 4
+  %4 = getelementptr inbounds %struct.point20, %struct.point20* %3, i32 0, i32 0
+  %5 = bitcast %struct.basepoint2* %4 to i64*
+  store i64 %0, i64* %5, align 4
+  %6 = bitcast %struct.point20* %2 to i8*
+  %7 = bitcast %struct.point20* %3 to i8*
+  call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 4 %6, i8* align 4 %7, i64 8, i1 false)
+  %8 = getelementptr inbounds %struct.point20, %struct.point20* %2, i32 0, i32 0
+  %9 = bitcast %struct.basepoint2* %8 to i64*
+  %10 = load i64, i64* %9, align 4
+  ret i64 %10
+}
+
+; Function Attrs: noinline nounwind optnone
+define dso_local [2 x i64] @demo_point21([2 x i64] %0) #0 {
+  %2 = alloca %struct.point21, align 4
+  %3 = alloca %struct.point21, align 4
+  %4 = alloca [2 x i64], align 8
+  %5 = alloca [2 x i64], align 8
+  store [2 x i64] %0, [2 x i64]* %4, align 8
+  %6 = bitcast %struct.point21* %3 to i8*
+  %7 = bitcast [2 x i64]* %4 to i8*
+  call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 4 %6, i8* align 8 %7, i64 12, i1 false)
+  %8 = bitcast %struct.point21* %2 to i8*
+  %9 = bitcast %struct.point21* %3 to i8*
+  call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 4 %8, i8* align 4 %9, i64 12, i1 false)
+  %10 = bitcast [2 x i64]* %5 to i8*
+  %11 = bitcast %struct.point21* %2 to i8*
+  call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 8 %10, i8* align 4 %11, i64 12, i1 false)
+  %12 = load [2 x i64], [2 x i64]* %5, align 8
+  ret [2 x i64] %12
+}
+
+attributes #0 = { noinline nounwind optnone "frame-pointer"="all" "min-legal-vector-width"="0" "no-trapping-math"="true" "stack-protector-buffer-size"="8" "target-features"="+64bit,+a,+c,+m,+relax,-save-restore" }
+attributes #1 = { argmemonly nofree nounwind willreturn }
+
+!llvm.module.flags = !{!0, !1, !2, !3}
+!llvm.ident = !{!4}
+
+!0 = !{i32 1, !"wchar_size", i32 4}
+!1 = !{i32 1, !"target-abi", !"lp64"}
+!2 = !{i32 7, !"frame-pointer", i32 2}
+!3 = !{i32 1, !"SmallDataLimit", i32 8}
+!4 = !{!"Apple clang version 14.0.3 (clang-1403.0.22.14.1)"}
diff --git a/internal/cabi/_testdata/arch/wasm32/composite.ll b/internal/cabi/_testdata/arch/wasm32/composite.ll
new file mode 100644
index 00000000..227eb95d
--- /dev/null
+++ b/internal/cabi/_testdata/arch/wasm32/composite.ll
@@ -0,0 +1,111 @@
+; ModuleID = '../../wrap/composite.c'
+source_filename = "../../wrap/composite.c"
+target datalayout = "e-m:e-p:32:32-p10:8:8-p20:8:8-i64:64-f128:64-n32:64-S128-ni:1:10:20"
+target triple = "wasm32-unknown-emscripten"
+
+%struct.array10 = type { %struct.basearray1 }
+%struct.basearray1 = type { [1 x i32] }
+%struct.array11 = type { %struct.basearray1, i32 }
+%struct.point10 = type { %struct.basepoint1 }
+%struct.basepoint1 = type { i32 }
+%struct.point11 = type { %struct.basepoint1, i32 }
+%struct.array20 = type { %struct.basearray2 }
+%struct.basearray2 = type { [2 x i32] }
+%struct.array21 = type { %struct.basearray2, i32 }
+%struct.point20 = type { %struct.basepoint2 }
+%struct.basepoint2 = type { i32, i32 }
+%struct.point21 = type { %struct.basepoint2, i32 }
+
+; Function Attrs: noinline nounwind optnone
+define hidden i32 @demo_array10(i32 %0) #0 {
+  %2 = alloca %struct.array10, align 4
+  %3 = alloca %struct.array10, align 4
+  %4 = getelementptr inbounds %struct.array10, %struct.array10* %3, i32 0, i32 0
+  %5 = getelementptr inbounds %struct.basearray1, %struct.basearray1* %4, i32 0, i32 0
+  %6 = bitcast [1 x i32]* %5 to i32*
+  store i32 %0, i32* %6, align 4
+  %7 = bitcast %struct.array10* %2 to i8*
+  %8 = bitcast %struct.array10* %3 to i8*
+  call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 %7, i8* align 4 %8, i32 4, i1 false)
+  %9 = getelementptr inbounds %struct.array10, %struct.array10* %2, i32 0, i32 0
+  %10 = getelementptr inbounds %struct.basearray1, %struct.basearray1* %9, i32 0, i32 0
+  %11 = bitcast [1 x i32]* %10 to i32*
+  %12 = load i32, i32* %11, align 4
+  ret i32 %12
+}
+
+; Function Attrs: argmemonly nofree nounwind willreturn
+declare void @llvm.memcpy.p0i8.p0i8.i32(i8* noalias nocapture writeonly, i8* noalias nocapture readonly, i32, i1 immarg) #1
+
+; Function Attrs: noinline nounwind optnone
+define hidden void @demo_array11(%struct.array11* noalias sret(%struct.array11) align 4 %0, %struct.array11* noundef byval(%struct.array11) align 4 %1) #0 {
+  %3 = bitcast %struct.array11* %0 to i8*
+  %4 = bitcast %struct.array11* %1 to i8*
+  call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 %3, i8* align 4 %4, i32 8, i1 false)
+  ret void
+}
+
+; Function Attrs: noinline nounwind optnone
+define hidden i32 @demo_point10(i32 %0) #0 {
+  %2 = alloca %struct.point10, align 4
+  %3 = alloca %struct.point10, align 4
+  %4 = getelementptr inbounds %struct.point10, %struct.point10* %3, i32 0, i32 0
+  %5 = getelementptr inbounds %struct.basepoint1, %struct.basepoint1* %4, i32 0, i32 0
+  store i32 %0, i32* %5, align 4
+  %6 = bitcast %struct.point10* %2 to i8*
+  %7 = bitcast %struct.point10* %3 to i8*
+  call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 %6, i8* align 4 %7, i32 4, i1 false)
+  %8 = getelementptr inbounds %struct.point10, %struct.point10* %2, i32 0, i32 0
+  %9 = getelementptr inbounds %struct.basepoint1, %struct.basepoint1* %8, i32 0, i32 0
+  %10 = load i32, i32* %9, align 4
+  ret i32 %10
+}
+
+; Function Attrs: noinline nounwind optnone
+define hidden void @demo_point11(%struct.point11* noalias sret(%struct.point11) align 4 %0, %struct.point11* noundef byval(%struct.point11) align 4 %1) #0 {
+  %3 = bitcast %struct.point11* %0 to i8*
+  %4 = bitcast %struct.point11* %1 to i8*
+  call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 %3, i8* align 4 %4, i32 8, i1 false)
+  ret void
+}
+
+; Function Attrs: noinline nounwind optnone
+define hidden void @demo_array20(%struct.array20* noalias sret(%struct.array20) align 4 %0, %struct.array20* noundef byval(%struct.array20) align 4 %1) #0 {
+  %3 = bitcast %struct.array20* %0 to i8*
+  %4 = bitcast %struct.array20* %1 to i8*
+  call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 %3, i8* align 4 %4, i32 8, i1 false)
+  ret void
+}
+
+; Function Attrs: noinline nounwind optnone
+define hidden void @demo_array21(%struct.array21* noalias sret(%struct.array21) align 4 %0, %struct.array21* noundef byval(%struct.array21) align 4 %1) #0 {
+  %3 = bitcast %struct.array21* %0 to i8*
+  %4 = bitcast %struct.array21* %1 to i8*
+  call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 %3, i8* align 4 %4, i32 12, i1 false)
+  ret void
+}
+
+; Function Attrs: noinline nounwind optnone
+define hidden void @demo_point20(%struct.point20* noalias sret(%struct.point20) align 4 %0, %struct.point20* noundef byval(%struct.point20) align 4 %1) #0 {
+  %3 = bitcast %struct.point20* %0 to i8*
+  %4 = bitcast %struct.point20* %1 to i8*
+  call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 %3, i8* align 4 %4, i32 8, i1 false)
+  ret void
+}
+
+; Function Attrs: noinline nounwind optnone
+define hidden void @demo_point21(%struct.point21* noalias sret(%struct.point21) align 4 %0, %struct.point21* noundef byval(%struct.point21) align 4 %1) #0 {
+  %3 = bitcast %struct.point21* %0 to i8*
+  %4 = bitcast %struct.point21* %1 to i8*
+  call void @llvm.memcpy.p0i8.p0i8.i32(i8* align 4 %3, i8* align 4 %4, i32 12, i1 false)
+  ret void
+}
+
+attributes #0 = { noinline nounwind optnone "frame-pointer"="none" "min-legal-vector-width"="0" "no-trapping-math"="true" "stack-protector-buffer-size"="8" "target-cpu"="generic" }
+attributes #1 = { argmemonly nofree nounwind willreturn }
+
+!llvm.module.flags = !{!0}
+!llvm.ident = !{!1}
+
+!0 = !{i32 1, !"wchar_size", i32 4}
+!1 = !{!"Apple clang version 14.0.3 (clang-1403.0.22.14.1)"}
diff --git a/internal/cabi/_testdata/demo/composite.go b/internal/cabi/_testdata/demo/composite.go
new file mode 100644
index 00000000..5ccb5bf2
--- /dev/null
+++ b/internal/cabi/_testdata/demo/composite.go
@@ -0,0 +1,176 @@
+package main
+
+import "unsafe"
+
+const (
+	LLGoFiles = "../wrap/composite.c"
+)
+
+//go:linkname printf C.printf
+func printf(format *byte, __llgo_va_list ...any) int32
+
+func assert(info string, b bool) {
+	if !b {
+		printf(unsafe.StringData("Assertion failed: %s\n\000"), unsafe.StringData(info))
+	}
+}
+
+func main() {
+}
+
+type basearray1 struct {
+	x [1]int32
+}
+
+type array10 struct {
+	base basearray1
+}
+
+//go:linkname cdemo_array10 C.demo_array10
+func cdemo_array10(a array10) array10
+
+func demo_array10(a array10) array10 {
+	return a
+}
+
+func init() {
+	v := array10{basearray1{[1]int32{1}}}
+	assert("cdemo_array10\000", cdemo_array10(v) == v)
+	assert("demo_array10\000", demo_array10(v) == v)
+}
+
+type array11 struct {
+	base basearray1
+	z    int32
+}
+
+//go:linkname cdemo_array11 C.demo_array11
+func cdemo_array11(a array11) array11
+
+func demo_array11(a array11) array11 {
+	return a
+}
+
+func init() {
+	v := array11{basearray1{[1]int32{1}}, 3}
+	assert("cdemo_array11\000", cdemo_array11(v) == v)
+	assert("demo_array11\000", demo_array11(v) == v)
+}
+
+type basepoint1 struct {
+	x int32
+}
+
+type point10 struct {
+	base basepoint1
+}
+
+//go:linkname cdemo_point10 C.demo_point10
+func cdemo_point10(a point10) point10
+
+func demo_point10(a point10) point10 {
+	return a
+}
+
+func init() {
+	v := point10{basepoint1{1}}
+	assert("cdemo_point10\000", cdemo_point10(v) == v)
+	assert("demo_point10\000", demo_point10(v) == v)
+}
+
+type point11 struct {
+	base basepoint1
+	z    int32
+}
+
+//go:linkname cdemo_point11 C.demo_point11
+func cdemo_point11(a point11) point11
+
+func demo_point11(a point11) point11 {
+	return a
+}
+
+func init() {
+	v := point11{basepoint1{1}, 3}
+	assert("cdemo_point11\000", cdemo_point11(v) == v)
+	assert("demo_point11\000", demo_point11(v) == v)
+}
+
+type basearray2 struct {
+	x [2]int32
+}
+
+type array20 struct {
+	base basearray2
+}
+
+//go:linkname cdemo_array20 C.demo_array20
+func cdemo_array20(a array20) array20
+
+func demo_array20(a array20) array20 {
+	return a
+}
+
+func init() {
+	v := array20{basearray2{[2]int32{1, 2}}}
+	assert("cdemo_array20\000", cdemo_array20(v) == v)
+	assert("demo_array20\000", demo_array20(v) == v)
+}
+
+type array21 struct {
+	base basearray2
+	z    int32
+}
+
+//go:linkname cdemo_array21 C.demo_array21
+func cdemo_array21(a array21) array21
+
+func demo_array21(a array21) array21 {
+	return a
+}
+
+func init() {
+	v := array21{basearray2{[2]int32{1, 2}}, 3}
+	assert("cdemo_array21\000", cdemo_array21(v) == v)
+	assert("demo_array21\000", demo_array21(v) == v)
+}
+
+type basepoint2 struct {
+	x int32
+	y int32
+}
+
+type point20 struct {
+	base basepoint2
+}
+
+//go:linkname cdemo_point20 C.demo_point20
+func cdemo_point20(a point20) point20
+
+func demo_point20(a point20) point20 {
+	return a
+}
+
+func init() {
+	v := point20{basepoint2{1, 2}}
+	assert("cdemo_point20\000", cdemo_point20(v) == v)
+	assert("demo_point20\000", demo_point20(v) == v)
+}
+
+type point21 struct {
+	base basepoint2
+	z    int32
+}
+
+//go:linkname cdemo_point21 C.demo_point21
+func cdemo_point21(a point21) point21
+
+func demo_point21(a point21) point21 {
+	return a
+}
+
+func init() {
+	v := point21{basepoint2{1, 2}, 3}
+	assert("cdemo_point21\000", cdemo_point21(v) == v)
+	assert("demo_point21\000", demo_point21(v) == v)
+}
diff --git a/internal/cabi/_testdata/wrap/composite.c b/internal/cabi/_testdata/wrap/composite.c
new file mode 100644
index 00000000..0de8d6bc
--- /dev/null
+++ b/internal/cabi/_testdata/wrap/composite.c
@@ -0,0 +1,84 @@
+struct basearray1 {
+    int x[1];
+};
+
+struct array10 {
+    struct basearray1 base;
+};
+
+struct array10 demo_array10(struct array10 a) {
+    return a;
+}
+
+struct array11 {
+    struct basearray1 base;
+    int z;
+};
+
+struct array11 demo_array11(struct array11 a) {
+    return a;
+}
+
+struct basepoint1 {
+    int x; 
+};
+
+struct point10 {
+    struct basepoint1 base;
+};
+
+struct point10 demo_point10(struct point10 a) {
+    return a;
+}
+
+struct point11 {
+    struct basepoint1 base;
+    int z;
+};
+
+struct point11 demo_point11(struct point11 a) {
+    return a;
+}
+
+struct basearray2 {
+    int x[2];
+};
+
+struct array20 {
+    struct basearray2 base;
+};
+
+struct array20 demo_array20(struct array20 a) {
+    return a;
+}
+
+struct array21 {
+    struct basearray2 base;
+    int z;
+};
+
+struct array21 demo_array21(struct array21 a) {
+    return a;
+}
+
+struct basepoint2 {
+    int x;
+    int y;  
+};
+
+struct point20 {
+    struct basepoint2 base;
+};
+
+struct point20 demo_point20(struct point20 a) {
+    return a;
+}
+
+struct point21 {
+    struct basepoint2 base;
+    int z;
+};
+
+struct point21 demo_point21(struct point21 a) {
+    return a;
+}
\ No newline at end of file