diff --git a/src/abi_ppc64le.cpp b/src/abi_ppc64le.cpp
index 12c4782547202..40c75247e33b3 100644
--- a/src/abi_ppc64le.cpp
+++ b/src/abi_ppc64le.cpp
@@ -42,46 +42,67 @@
 typedef bool AbiState;
 AbiState default_abi_state = 0;
 
-static bool isHFA(jl_datatype_t *ty)
+// count the homogeneous floating agregate size (saturating at max count of 8)
+static unsigned isHFA(jl_datatype_t *ty, jl_datatype_t **ty0)
 {
     size_t i, l = ty->nfields;
-    jl_value_t *fld0 = jl_field_type(ty, 0);
-    if (l > 8)
-        return false;
-    if (fld0 != (jl_value_t*)jl_float64_type && fld0 != (jl_value_t*)jl_float32_type)
-        return false;
-    for (i = 1; i < l; i++) {
+    if (l == 0) {
+        if (*ty0 == NULL) {
+            if (ty == jl_float64_type || ty == jl_float32_type)
+                *ty0 = ty;
+        }
+        return ty == *ty0 ? 1 : 9;
+    }
+    int n = 0;
+    for (i = 0; i < l; i++) {
         jl_value_t *fld = jl_field_type(ty, i);
-        if (fld != fld0)
-            return false;
+        if (!jl_is_datatype(fld))
+            return 9;
+        n += isHFA((jl_datatype_t*)fld, ty0);
+        if (n > 8)
+            return 9;
     }
-    return true;
+    return n;
 }
 
 bool use_sret(AbiState *state, jl_value_t *ty)
 {
     // Assume jl_is_datatype(ty) && !jl_is_abstracttype(ty)
     jl_datatype_t *dt = (jl_datatype_t*)ty;
-    if (dt->size > 16 && !isHFA(dt))
+    jl_datatype_t *ty0 = NULL;
+    if (dt->size > 16 && isHFA(dt, &ty0) > 8)
         return true;
     return false;
 }
 
 void needPassByRef(AbiState *state, jl_value_t *ty, bool *byRef, bool *inReg)
 {
-    *byRef = false;
+    if (!jl_is_datatype(ty) || jl_is_abstracttype(ty) || jl_is_cpointer_type(ty) || jl_is_array_type(ty))
+        return;
+    size_t size = jl_datatype_size(ty);
+    if (size > 64)
+        *byRef = true;
 }
 
 Type *preferred_llvm_type(jl_value_t *ty, bool isret)
 {
     // Arguments are either scalar or passed by value
-    if (!jl_is_datatype(ty) || jl_is_abstracttype(ty))
+    if (!jl_is_datatype(ty) || jl_is_abstracttype(ty) || jl_is_cpointer_type(ty) || jl_is_array_type(ty))
         return NULL;
     jl_datatype_t *dt = (jl_datatype_t*)ty;
-    // rewrite integer sized (non-HFA) struct to the corresponding integer
-    if (!dt->nfields || (isret ? use_sret(NULL, ty) : false) || isHFA(dt))
+    size_t size = dt->size;
+    // don't need to change bitstypes
+    if (!dt->nfields)
         return NULL;
-    return Type::getIntNTy(jl_LLVMContext, dt->size * 8);
+    // legalize this into [n x f32/f64]
+    jl_datatype_t *ty0 = NULL;
+    int hfa = isHFA(dt, &ty0);
+    if (hfa <= 8)
+        return ArrayType::get(ty0 == jl_float32_type ? T_float32 : T_float64, hfa);
+    // rewrite integer-sized (non-HFA) struct to an array of i64
+    if (size > 8)
+        return ArrayType::get(T_int64, (size + 7) / 8);
+    return Type::getIntNTy(jl_LLVMContext, size * 8);
 }
 
 bool need_private_copy(jl_value_t *ty, bool byRef)
diff --git a/src/abi_win32.cpp b/src/abi_win32.cpp
index 1209f72520c57..c9cfd10f2b545 100644
--- a/src/abi_win32.cpp
+++ b/src/abi_win32.cpp
@@ -67,7 +67,7 @@ Type *preferred_llvm_type(jl_value_t *ty, bool isret)
         return NULL;
     jl_datatype_t *dt = (jl_datatype_t*)ty;
     // rewrite integer sized (non-sret) struct to the corresponding integer
-    if (!dt->nfields || use_sret(NULL, ty))
+    if (!dt->nfields)
         return NULL;
     return Type::getIntNTy(jl_LLVMContext, dt->size * 8);
 }
diff --git a/src/ccall.cpp b/src/ccall.cpp
index 7d9b2f7302d13..9682a269fe132 100644
--- a/src/ccall.cpp
+++ b/src/ccall.cpp
@@ -876,7 +876,6 @@ static std::string generate_func_sig(
     size_t nargt = jl_svec_len(tt);
     assert(rt && !jl_is_abstract_ref_type(rt));
 
-    AttrBuilder retattrs;
     std::vector<AttrBuilder> paramattrs;
     AbiState abi = default_abi_state;
     sret = 0;
@@ -885,7 +884,7 @@ static std::string generate_func_sig(
         *prt = *lrt = T_void;
     }
     else {
-        *prt = preferred_llvm_type(rt, true);
+        *prt = sret ? NULL : preferred_llvm_type(rt, true);
         if (*prt == NULL)
             *prt = *lrt;
 
@@ -895,6 +894,7 @@ static std::string generate_func_sig(
 #if !defined(_OS_WINDOWS_) || defined(LLVM35) // llvm used to use the old mingw ABI, skipping this marking works around that difference
             paramattrs[0].addAttribute(Attribute::StructRet);
 #endif
+            paramattrs[0].addAttribute(Attribute::NoAlias);
             fargt_sig.push_back(PointerType::get(*prt, 0));
             sret = 1;
         }
@@ -953,14 +953,8 @@ static std::string generate_func_sig(
             needPassByRef(&abi, tti, &byRef, &inReg);
         }
 
-        Type *pat = preferred_llvm_type(tti, false);
-        if (pat != NULL) {
-            assert(!byRef); // it is an error for an ABI to specify a preferred type for a pointer arg
-        }
-        else if (byRef) {
-            pat = PointerType::get(t, 0);
-        }
-        else {
+        Type *pat = byRef ? PointerType::get(t, 0) : preferred_llvm_type(tti, false);
+        if (pat == NULL) {
             pat = t;
         }
 
@@ -994,10 +988,6 @@ static std::string generate_func_sig(
         } while (current_isVa && i < nargs); // if is this is the vararg, loop to the end
     }
 
-    if (retattrs.hasAttributes()) {
-        attributes = AttributeSet::get(jl_LLVMContext, AttributeSet::ReturnIndex, retattrs);
-    }
-
     for (i = 0; i < nargs + sret; ++i) {
         if (paramattrs[i].hasAttributes()) {
             attributes = attributes.addAttributes(jl_LLVMContext, i + 1,
diff --git a/src/codegen.cpp b/src/codegen.cpp
index 9df623d2970ac..d7d2005b3cdd5 100644
--- a/src/codegen.cpp
+++ b/src/codegen.cpp
@@ -4126,8 +4126,10 @@ static std::unique_ptr<Module> emit_function(jl_lambda_info_t *lam, jl_llvm_func
         f = Function::Create(FunctionType::get(rt, fsig, false),
                              GlobalVariable::ExternalLinkage,
                              funcName.str(), M);
-        if (ctx.sret)
+        if (ctx.sret) {
             f->addAttribute(1, Attribute::StructRet);
+            f->addAttribute(1, Attribute::NoAlias);
+        }
         addComdat(f);
 #ifdef LLVM37
         f->addFnAttr("no-frame-pointer-elim", "true");