Fix ARM and AArch64 ABI

Also add related ccall test and test both mutable and immutable types.

src/ccall.cpp

@@ -409,10 +409,26 @@ Value *llvm_type_rewrite(Value *v, Type *from_type, Type *target_type,
     // one or both of from_type and target_type is a VectorType or AggregateType
     // LLVM doesn't allow us to cast these values directly, so
     // we need to use this alloca copy trick instead
-    // NOTE: it is assumed that the ABI has ensured that sizeof(from_type) == sizeof(target_type)
-    Value *mem = emit_static_alloca(target_type, ctx);
-    builder.CreateStore(v, builder.CreatePointerCast(mem, from_type->getPointerTo()));
-    return builder.CreateLoad(mem);
+    // On ARM and AArch64, the ABI requires casting through memory to different
+    // sizes.
+    Value *from;
+    Value *to;
+    const DataLayout &DL =
+#if JL_LLVM_VERSION >= 30600
+        jl_ExecutionEngine->getDataLayout();
+#else
+    *jl_ExecutionEngine->getDataLayout();
+#endif
+    if (DL.getTypeAllocSize(target_type) >= DL.getTypeAllocSize(from_type)) {
+        to = emit_static_alloca(target_type, ctx);
+        from = builder.CreatePointerCast(to, from_type->getPointerTo());
+    }
+    else {
+        from = emit_static_alloca(from_type, ctx);
+        to = builder.CreatePointerCast(from, target_type->getPointerTo());
+    }
+    builder.CreateStore(v, from);
+    return builder.CreateLoad(to);
 }
 
 // --- argument passing and scratch space utilities ---
@@ -1839,8 +1855,24 @@ static jl_cgval_t emit_ccall(jl_value_t **args, size_t nargs, jl_codectx_t *ctx)
             jl_cgval_t newst = emit_new_struct(rt, 1, NULL, ctx); // emit a new, empty struct
             assert(newst.typ != NULL && "Type was not concrete");
             assert(newst.isboxed);
+            size_t rtsz = jl_datatype_size(rt);
+            assert(rtsz > 0);
+            int boxalign = jl_gc_alignment(rtsz);
+#ifndef NDEBUG
+            const DataLayout &DL =
+#if JL_LLVM_VERSION >= 30600
+                jl_ExecutionEngine->getDataLayout();
+#else
+            *jl_ExecutionEngine->getDataLayout();
+#endif
+            // ARM and AArch64 can use a LLVM type larger than the julia
+            // type. However, the LLVM type size should be no larger than
+            // the GC allocation size. (multiple of `sizeof(void*)`)
+            assert(DL.getTypeStoreSize(lrt) <= LLT_ALIGN(jl_datatype_size(rt),
+                                                         boxalign));
+#endif
             // copy the data from the return value to the new struct
-            tbaa_decorate(newst.tbaa, builder.CreateAlignedStore(result, emit_bitcast(newst.V, prt->getPointerTo()), 16)); // julia gc is aligned 16
+            tbaa_decorate(newst.tbaa, builder.CreateAlignedStore(result, emit_bitcast(newst.V, prt->getPointerTo()), boxalign));
             return newst;
         }
         else if (jlrt != prt) {

src/ccalltest.c

@@ -191,6 +191,17 @@ typedef struct {
     double a,b,c;
 } struct16;
 
+typedef struct {
+    int8_t a;
+    int16_t b;
+} struct17;
+
+typedef struct {
+    int8_t a;
+    int8_t b;
+    int8_t c;
+} struct18;
+
 typedef struct {
     jint x;
     jint y;
@@ -467,6 +478,24 @@ JL_DLLEXPORT struct16 test_16(struct16 a, float b) {
     return a;
 }
 
+JL_DLLEXPORT struct17 test_17(struct17 a, int8_t b) {
+    //Unpack a struct with non-obvious packing requirements
+    if (verbose) fprintf(stderr,"%d %d & %d\n", (int)a.a, (int)a.b, (int)b);
+    a.a += b*1;
+    a.b -= b*2;
+    return a;
+}
+
+JL_DLLEXPORT struct18 test_18(struct18 a, int8_t b) {
+    //Unpack a struct with non-obvious packing requirements
+    if (verbose) fprintf(stderr,"%d %d %d & %d\n",
+                         (int)a.a, (int)a.b, (int)a.c, (int)b);
+    a.a += b*1;
+    a.b -= b*2;
+    a.c += b*3;
+    return a;
+}
+
 // Note for AArch64:
 // `i128` is a native type on aarch64 so the type here is wrong.
 // However, it happens to have the same calling convention with `[2 x i64]`

src/julia_internal.h

@@ -92,6 +92,27 @@ static const int jl_gc_sizeclasses[JL_GC_N_POOLS] = {
 //    64,   32,  160,   64,   16,   64,  112,  128, bytes lost
 };
 
+STATIC_INLINE int jl_gc_alignment(size_t sz)
+{
+    if (sz == 0)
+        return sizeof(void*);
+#ifdef _P64
+    (void)sz;
+    return 16;
+#elif defined(_CPU_ARM_) || defined(_CPU_PPC_)
+    return rtsz < 4 : 8 : 16;
+#else
+    // szclass 8
+    if (sz < 4)
+        return 8;
+    // szclass 12
+    if (sz < 8)
+        return 4;
+    // szclass 16+
+    return 16;
+#endif
+}
+
 STATIC_INLINE int JL_CONST_FUNC jl_gc_szclass(size_t sz)
 {
 #ifdef _P64