Index: src/OFObject.h
==================================================================
--- src/OFObject.h
+++ src/OFObject.h
@@ -97,12 +97,17 @@
 #  define OF_HAVE_FORWARDING_TARGET_FOR_SELECTOR
 #  define OF_HAVE_FORWARDING_TARGET_FOR_SELECTOR_STRET
 # endif
 #else
 # ifdef __ELF__
-#  if defined(__amd64__) || defined(__x86_64__) || defined(__i386__) || \
-    defined(__arm__) || defined(__ARM__)
+#  if defined(__amd64__) || defined(__x86_64__)
+#   define OF_HAVE_FORWARDING_TARGET_FOR_SELECTOR
+#   if __has_feature(objc_msg_lookup_stret)
+#    define OF_HAVE_FORWARDING_TARGET_FOR_SELECTOR_STRET
+#   endif
+#  endif
+#  if  defined(__i386__) || defined(__arm__) || defined(__ARM__)
 #   define OF_HAVE_FORWARDING_TARGET_FOR_SELECTOR
 #  endif
 #  if (defined(_MIPS_SIM) && _MIPS_SIM == _ABIO32) || \
     (defined(__mips_eabi) && _MIPS_SZPTR == 32)
 #   define OF_HAVE_FORWARDING_TARGET_FOR_SELECTOR

Index: src/forwarding-amd64-elf.S
==================================================================
--- src/forwarding-amd64-elf.S
+++ src/forwarding-amd64-elf.S
@@ -17,11 +17,10 @@
 .globl of_forward
 .globl of_forward_stret
 
 .section .text
 of_forward:
-of_forward_stret: /* FIXME */
 	pushq	%rbp
 	movq	%rsp, %rbp
 
 	/* Save all arguments */
 	subq	$0xC0, %rsp	/* 16-byte alignment */
@@ -65,10 +64,11 @@
 	movdqa	-0x50(%rbp), %xmm0
 	movq	-0x38(%rbp), %r9
 	movq	-0x30(%rbp), %r8
 	movq	-0x28(%rbp), %rcx
 	movq	-0x20(%rbp), %rdx
+	movq	-0x18(%rbp), %rsi
 	movq	-0x10(%rbp), %rdi
 	movq	-0x8(%rbp), %rax
 
 	movq	%rbp, %rsp
 	popq	%rbp
@@ -75,10 +75,70 @@
 
 	jmpq     *%r11
 .type of_forward, %function
 .size of_forward, .-of_forward
 
+of_forward_stret:
+	pushq	%rbp
+	movq	%rsp, %rbp
+
+	/* Save all arguments */
+	subq	$0xC0, %rsp	/* 16-byte alignment */
+	movq	%rax, -0x8(%rbp)
+	movq	%rdi, -0x10(%rbp)
+	movq	%rsi, -0x18(%rbp)
+	movq	%rdx, -0x20(%rbp)
+	movq	%rcx, -0x28(%rbp)
+	movq	%r8, -0x30(%rbp)
+	movq	%r9, -0x38(%rbp)
+	movdqa	%xmm0, -0x50(%rbp)
+	movdqa	%xmm1, -0x60(%rbp)
+	movdqa	%xmm2, -0x70(%rbp)
+	movdqa	%xmm3, -0x80(%rbp)
+	movdqa	%xmm4, -0x90(%rbp)
+	movdqa	%xmm5, -0xA0(%rbp)
+	movdqa	%xmm6, -0xB0(%rbp)
+	movdqa	%xmm7, -0xC0(%rbp)
+
+	movq	%rsi, %rdi
+	leaq	sel_forwardingTargetForSelector_(%rip), %rsi
+	call	objc_msg_lookup@PLT
+	movq	-0x18(%rbp), %rdi
+	leaq	sel_forwardingTargetForSelector_(%rip), %rsi
+	movq	-0x20(%rbp), %rdx
+	call	*%rax
+	movq	%rax, -0x18(%rbp)
+
+	movq	%rax, %rdi
+	movq	-0x20(%rbp), %rsi
+	call	objc_msg_lookup_stret@PLT
+	movq	%rax, %r11
+
+	/* Restore all arguments */
+	movdqa	-0xC0(%rbp), %xmm7
+	movdqa	-0xB0(%rbp), %xmm6
+	movdqa	-0xA0(%rbp), %xmm5
+	movdqa	-0x90(%rbp), %xmm4
+	movdqa	-0x80(%rbp), %xmm3
+	movdqa	-0x70(%rbp), %xmm2
+	movdqa	-0x60(%rbp), %xmm1
+	movdqa	-0x50(%rbp), %xmm0
+	movq	-0x38(%rbp), %r9
+	movq	-0x30(%rbp), %r8
+	movq	-0x28(%rbp), %rcx
+	movq	-0x20(%rbp), %rdx
+	movq	-0x18(%rbp), %rsi
+	movq	-0x10(%rbp), %rdi
+	movq	-0x8(%rbp), %rax
+
+	movq	%rbp, %rsp
+	popq	%rbp
+
+	jmpq     *%r11
+.type of_forward_stret, %function
+.size of_forward_stret, .-of_forward_stret
+
 init:
 	leaq	module(%rip), %rdi
 	jmp	__objc_exec_class@PLT
 
 .section .ctors, "a", %progbits