Index: .travis.yml
==================================================================
--- .travis.yml
+++ .travis.yml
@@ -223,23 +223,23 @@
       env:
         - config=amigaos
 
     # Nintendo 3DS
     - os: linux
-      dist: trusty
+      dist: bionic
       env:
         - config=nintendo_3ds
 
     # Nintendo DS
     - os: linux
-      dist: trusty
+      dist: bionic
       env:
         - config=nintendo_ds
 
     # Nintendo Wii
     - os: linux
-      dist: trusty
+      dist: bionic
       env:
         - config=wii
 
 before_install:
   - if [ "$TRAVIS_OS_NAME" = "linux" -a -z "$config" ]; then
@@ -268,12 +268,14 @@
             fi;
     fi
 
   - if [ "$config" = "nintendo_3ds" -o "$config" = "nintendo_ds"
         -o "$config" = "wii" ]; then
-            wget https://github.com/devkitPro/pacman/releases/download/devkitpro-pacman-1.0.1/devkitpro-pacman.deb;
-            sudo dpkg -i devkitpro-pacman.deb;
+            deb=devkitpro-pacman.amd64.deb;
+            wget https://github.com/devkitPro/pacman/releases/download/v1.0.2/$deb;
+            sudo apt install gdebi;
+            sudo gdebi -n $deb;
     fi
 
   - if [ "$config" = "nintendo_3ds" ]; then
             sudo dkp-pacman --noconfirm -Syu 3ds-dev;
     fi

Index: src/forwarding/apple-forwarding-i386.S
==================================================================
--- src/forwarding/apple-forwarding-i386.S
+++ src/forwarding/apple-forwarding-i386.S
@@ -15,123 +15,133 @@
  * file.
  */
 
 #include "config.h"
 
+.intel_syntax noprefix
+
 .globl _of_forward
 .globl _of_forward_stret
 
 .section __TEXT, __cstring, cstring_literals
-str_forwardingTargetForSelector_:
+Lstr_forwardingTargetForSelector_:
 	.asciz "forwardingTargetForSelector:"
 
 .section __OBJC, __message_refs, literal_pointers, no_dead_strip
-sel_forwardingTargetForSelector_:
-	.long str_forwardingTargetForSelector_
+Lsel_forwardingTargetForSelector_:
+	.long Lstr_forwardingTargetForSelector_
 
 .section __OBJC, __image_info
 	.long 0, 0
 
 .section __TEXT, __text, regular, pure_instructions
 _of_forward:
-	pushl	%ebp
-	movl	%esp, %ebp
+	push	ebp
+	mov	ebp, esp
 
-	pushl	%ebx
-	subl	$20, %esp
+	push	ebx
+	sub	esp, 20
 
 	call	get_eip
 0:
 
-	movl	8(%ebp), %eax
-	movl	%eax, (%esp)
+	mov	eax, [ebp+8]
+	mov	[esp], eax
 	call	_object_getClass
 
-	movl	%eax, (%esp)
-	movl	sel_forwardingTargetForSelector_-0b(%ebx), %eax
-	movl	%eax, 4(%esp)
+	mov	[esp], eax
+	.att_syntax	/* Next line is broken in Intel syntax */
+	movl	Lsel_forwardingTargetForSelector_-0b(%ebx), %eax
+	.intel_syntax noprefix
+	mov	[esp+4], eax
 	call	_class_respondsToSelector
 
-	testl	%eax, %eax
+	test	eax, eax
 	jz	0f
 
-	movl	8(%ebp), %eax
-	movl	%eax, (%esp)
-	movl	sel_forwardingTargetForSelector_-0b(%ebx), %eax
-	movl	%eax, 4(%esp)
-	movl	12(%ebp), %eax
-	movl	%eax, 8(%esp)
+	mov	eax, [ebp+8]
+	mov	[esp], eax
+	.att_syntax	/* Next line is broken in Intel syntax */
+	movl	Lsel_forwardingTargetForSelector_-0b(%ebx), %eax
+	.intel_syntax noprefix
+	mov	[esp+4], eax
+	mov	eax, [ebp+12]
+	mov	[esp+8], eax
 	call	_objc_msgSend
 
-	testl	%eax, %eax
+	test	eax, eax
 	jz	0f
-	cmpl	8(%ebp), %eax
+	cmp	eax, [ebp+8]
 	je	0f
 
-	movl	%eax, 8(%ebp)
+	mov	[ebp+8], eax
 
-	addl	$20, %esp
-	popl	%ebx
-	popl	%ebp
+	add	esp, 20
+	pop	ebx
+	pop	ebp
 
 	jmp	_objc_msgSend
 
 0:
-	addl	$20, %esp
-	popl	%ebx
-	popl	%ebp
+	add	esp, 20
+	pop	ebx
+	pop	ebp
 
 	jmp	_of_method_not_found
 
 _of_forward_stret:
-	pushl	%ebp
-	movl	%esp, %ebp
+	push	ebp
+	mov	ebp, esp
 
-	pushl	%ebx
-	subl	$20, %esp
+	push	ebx
+	sub	esp, 20
 
 	call	get_eip
 0:
 
-	movl	12(%ebp), %eax
-	movl	%eax, (%esp)
+	mov	eax, [ebp+12]
+	mov	[esp], eax
 	call	_object_getClass
 
-	movl	%eax, (%esp)
-	movl	sel_forwardingTargetForSelector_-0b(%ebx), %eax
-	movl	%eax, 4(%esp)
+	mov	[esp], eax
+	.att_syntax	/* Next line is broken in Intel syntax */
+	movl	Lsel_forwardingTargetForSelector_-0b(%ebx), %eax
+	.intel_syntax noprefix
+	mov	[esp+4], eax
 	call	_class_respondsToSelector
 
-	testl	%eax, %eax
+	test	eax, eax
 	jz	0f
 
-	movl	12(%ebp), %eax
-	movl	%eax, (%esp)
-	movl	sel_forwardingTargetForSelector_-0b(%ebx), %eax
-	movl	%eax, 4(%esp)
-	movl	16(%ebp), %eax
-	movl	%eax, 8(%esp)
+	mov	eax, [ebp+12]
+	mov	[esp], eax
+	.att_syntax	/* Next line is broken in Intel syntax */
+	movl	Lsel_forwardingTargetForSelector_-0b(%ebx), %eax
+	.intel_syntax noprefix
+	mov	[esp+4], eax
+	mov	eax, [ebp+16]
+	mov	[esp+8], eax
 	call	_objc_msgSend
 
-	testl	%eax, %eax
+	test	eax, eax
 	jz	0f
-	cmpl	12(%ebp), %eax
+	cmp	eax, [ebp+12]
 	je	0f
 
-	movl	%eax, 12(%ebp)
+	mov	[ebp+12], eax
 
-	addl	$20, %esp
-	popl	%ebx
-	popl	%ebp
+	add	esp, 20
+	pop	ebx
+	pop	ebp
 
 	jmp	_objc_msgSend_stret
 
 0:
-	addl	$20, %esp
-	popl	%ebx
-	popl	%ebp
+	add	esp, 20
+	pop	ebx
+	pop	ebp
 
 	jmp	_of_method_not_found_stret
 
 get_eip:
-	movl	(%esp), %ebx
+	mov	ebx, [esp]
 	ret

Index: src/forwarding/apple-forwarding-x86_64.S
==================================================================
--- src/forwarding/apple-forwarding-x86_64.S
+++ src/forwarding/apple-forwarding-x86_64.S
@@ -15,10 +15,12 @@
  * file.
  */
 
 #include "config.h"
 
+.intel_syntax noprefix
+
 .globl _of_forward
 .globl _of_forward_stret
 
 .section __TEXT, __objc_methname, cstring_literals
 str_forwardingTargetForSelector_:
@@ -31,150 +33,150 @@
 .section __DATA, __objc_imageinfo, regular, no_dead_strip
 	.long 0, 0
 
 .section __TEXT, __text, regular, pure_instructions
 _of_forward:
-	pushq	%rbp
-	movq	%rsp, %rbp
+	push	rbp
+	mov	rbp, rsp
 
 	/* 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)
-	movaps	%xmm0, -0x50(%rbp)
-	movaps	%xmm1, -0x60(%rbp)
-	movaps	%xmm2, -0x70(%rbp)
-	movaps	%xmm3, -0x80(%rbp)
-	movaps	%xmm4, -0x90(%rbp)
-	movaps	%xmm5, -0xA0(%rbp)
-	movaps	%xmm6, -0xB0(%rbp)
-	movaps	%xmm7, -0xC0(%rbp)
+	sub	rsp, 0xC0	/* 16-byte alignment */
+	mov	[rbp-0x08], rax
+	mov	[rbp-0x10], rdi
+	mov	[rbp-0x18], rsi
+	mov	[rbp-0x20], rdx
+	mov	[rbp-0x28], rcx
+	mov	[rbp-0x30], r8
+	mov	[rbp-0x38], r9
+	movaps	[rbp-0x50], xmm0
+	movaps	[rbp-0x60], xmm1
+	movaps	[rbp-0x70], xmm2
+	movaps	[rbp-0x80], xmm3
+	movaps	[rbp-0x90], xmm4
+	movaps	[rbp-0xA0], xmm5
+	movaps	[rbp-0xB0], xmm6
+	movaps	[rbp-0xC0], xmm7
 
 	call	_object_getClass
 
-	movq	%rax, %rdi
-	movq	sel_forwardingTargetForSelector_(%rip), %rsi
+	mov	rdi, rax
+	mov	rsi, [rip+sel_forwardingTargetForSelector_]
 	call	_class_respondsToSelector
 
-	testq	%rax, %rax
+	test	rax, rax
 	jz	0f
 
-	movq	-0x10(%rbp), %rdi
-	movq	sel_forwardingTargetForSelector_(%rip), %rsi
-	movq	-0x18(%rbp), %rdx
+	mov	rdi, [rbp-0x10]
+	mov	rsi, [rip+sel_forwardingTargetForSelector_]
+	mov	rdx, [rbp-0x18]
 	call	_objc_msgSend
 
-	testq	%rax, %rax
+	test	rax, rax
 	jz	0f
-	cmpq	-0x10(%rbp), %rax
+	cmp	rax, [rbp-0x10]
 	je	0f
 
-	movq	%rax, %rdi
+	mov	rdi, rax
 
 	/* Restore all arguments, except %rdi */
-	movaps	-0xC0(%rbp), %xmm7
-	movaps	-0xB0(%rbp), %xmm6
-	movaps	-0xA0(%rbp), %xmm5
-	movaps	-0x90(%rbp), %xmm4
-	movaps	-0x80(%rbp), %xmm3
-	movaps	-0x70(%rbp), %xmm2
-	movaps	-0x60(%rbp), %xmm1
-	movaps	-0x50(%rbp), %xmm0
-	movq	-0x38(%rbp), %r9
-	movq	-0x30(%rbp), %r8
-	movq	-0x28(%rbp), %rcx
-	movq	-0x20(%rbp), %rdx
-	movq	-0x18(%rbp), %rsi
-	movq	-0x8(%rbp), %rax
-
-	movq	%rbp, %rsp
-	popq	%rbp
+	movaps	xmm7, [rbp-0xC0]
+	movaps	xmm6, [rbp-0xB0]
+	movaps	xmm5, [rbp-0xA0]
+	movaps	xmm4, [rbp-0x90]
+	movaps	xmm3, [rbp-0x80]
+	movaps	xmm2, [rbp-0x70]
+	movaps	xmm1, [rbp-0x60]
+	movaps	xmm0, [rbp-0x50]
+	mov	r9,   [rbp-0x38]
+	mov	r8,   [rbp-0x30]
+	mov	rcx,  [rbp-0x28]
+	mov	rdx,  [rbp-0x20]
+	mov	rsi,  [rbp-0x18]
+	mov	rax,  [rbp-0x08]
+
+	mov	rsp, rbp
+	pop	rbp
 
 	jmp	_objc_msgSend
 
 0:
-	movq	-0x10(%rbp), %rdi
-	movq	-0x18(%rbp), %rsi
+	mov	rdi, [rbp-0x10]
+	mov	rsi, [rbp-0x18]
 
-	movq	%rbp, %rsp
-	popq	%rbp
+	mov	rsp, rbp
+	pop	rbp
 
 	jmp	_of_method_not_found
 
 _of_forward_stret:
-	pushq	%rbp
-	movq	%rsp, %rbp
+	push	rbp
+	mov	rbp, rsp
 
 	/* 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)
-	movaps	%xmm0, -0x50(%rbp)
-	movaps	%xmm1, -0x60(%rbp)
-	movaps	%xmm2, -0x70(%rbp)
-	movaps	%xmm3, -0x80(%rbp)
-	movaps	%xmm4, -0x90(%rbp)
-	movaps	%xmm5, -0xA0(%rbp)
-	movaps	%xmm6, -0xB0(%rbp)
-	movaps	%xmm7, -0xC0(%rbp)
-
-	movq	%rsi, %rdi
+	sub	rsp, 0xC0	/* 16-byte alignment */
+	mov	[rbp-0x08], rax
+	mov	[rbp-0x10], rdi
+	mov	[rbp-0x18], rsi
+	mov	[rbp-0x20], rdx
+	mov	[rbp-0x28], rcx
+	mov	[rbp-0x30], r8
+	mov	[rbp-0x38], r9
+	movaps	[rbp-0x50], xmm0
+	movaps	[rbp-0x60], xmm1
+	movaps	[rbp-0x70], xmm2
+	movaps	[rbp-0x80], xmm3
+	movaps	[rbp-0x90], xmm4
+	movaps	[rbp-0xA0], xmm5
+	movaps	[rbp-0xB0], xmm6
+	movaps	[rbp-0xC0], xmm7
+
+	mov	rdi, rsi
 	call	_object_getClass
 
-	movq	%rax, %rdi
-	movq	sel_forwardingTargetForSelector_(%rip), %rsi
+	mov	rdi, rax
+	mov	rsi, [rip+sel_forwardingTargetForSelector_]
 	call	_class_respondsToSelector
-	testq	%rax, %rax
+	test	rax, rax
 	jz	0f
 
-	movq	-0x18(%rbp), %rdi
-	movq	sel_forwardingTargetForSelector_(%rip), %rsi
-	movq	-0x20(%rbp), %rdx
+	mov	rdi, [rbp-0x18]
+	mov	rsi, [rip+sel_forwardingTargetForSelector_]
+	mov	rdx, [rbp-0x20]
 	call	_objc_msgSend
 
-	testq	%rax, %rax
+	test	rax, rax
 	jz	0f
-	cmpq	-0x18(%rbp), %rax
+	cmp	rax, [rbp-0x18]
 	je	0f
 
-	movq	%rax, %rsi
+	mov	rsi, rax
 
 	/* Restore all arguments, except %rsi */
-	movaps	-0xC0(%rbp), %xmm7
-	movaps	-0xB0(%rbp), %xmm6
-	movaps	-0xA0(%rbp), %xmm5
-	movaps	-0x90(%rbp), %xmm4
-	movaps	-0x80(%rbp), %xmm3
-	movaps	-0x70(%rbp), %xmm2
-	movaps	-0x60(%rbp), %xmm1
-	movaps	-0x50(%rbp), %xmm0
-	movq	-0x38(%rbp), %r9
-	movq	-0x30(%rbp), %r8
-	movq	-0x28(%rbp), %rcx
-	movq	-0x20(%rbp), %rdx
-	movq	-0x10(%rbp), %rdi
-	movq	-0x8(%rbp), %rax
-
-	movq	%rbp, %rsp
-	popq	%rbp
+	movaps	xmm7, [rbp-0xC0]
+	movaps	xmm6, [rbp-0xB0]
+	movaps	xmm5, [rbp-0xA0]
+	movaps	xmm4, [rbp-0x90]
+	movaps	xmm3, [rbp-0x80]
+	movaps	xmm2, [rbp-0x70]
+	movaps	xmm1, [rbp-0x60]
+	movaps	xmm0, [rbp-0x50]
+	mov	r9,   [rbp-0x38]
+	mov	r8,   [rbp-0x30]
+	mov	rcx,  [rbp-0x28]
+	mov	rdx,  [rbp-0x20]
+	mov	rdi,  [rbp-0x10]
+	mov	rax,  [rbp-0x08]
+
+	mov	rsp, rbp
+	pop	rbp
 
 	jmp	_objc_msgSend_stret
 
 0:
-	movq	-0x10(%rbp), %rdi
-	movq	-0x18(%rbp), %rsi
-	movq	-0x20(%rbp), %rdx
+	mov	rdi, [rbp-0x10]
+	mov	rsi, [rbp-0x18]
+	mov	rdx, [rbp-0x20]
 
-	movq	%rbp, %rsp
-	popq	%rbp
+	mov	rsp, rbp
+	pop	rbp
 
 	jmp	_of_method_not_found_stret

Index: src/forwarding/forwarding-x86_64-macho.S
==================================================================
--- src/forwarding/forwarding-x86_64-macho.S
+++ src/forwarding/forwarding-x86_64-macho.S
@@ -17,184 +17,186 @@
 
 #include "config.h"
 
 #include "platform.h"
 
+.intel_syntax noprefix
+
 .globl _of_forward
 .globl _of_forward_stret
 
 .section __TEXT, __text, regular, pure_instructions
 _of_forward:
-	pushq	%rbp
-	movq	%rsp, %rbp
+	push	rbp
+	mov	rbp, rsp
 
 	/* 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)
-	movaps	%xmm0, -0x50(%rbp)
-	movaps	%xmm1, -0x60(%rbp)
-	movaps	%xmm2, -0x70(%rbp)
-	movaps	%xmm3, -0x80(%rbp)
-	movaps	%xmm4, -0x90(%rbp)
-	movaps	%xmm5, -0xA0(%rbp)
-	movaps	%xmm6, -0xB0(%rbp)
-	movaps	%xmm7, -0xC0(%rbp)
+	sub	rsp, 0xC0	/* 16-byte alignment */
+	mov	[rbp-0x08], rax
+	mov	[rbp-0x10], rdi
+	mov	[rbp-0x18], rsi
+	mov	[rbp-0x20], rdx
+	mov	[rbp-0x28], rcx
+	mov	[rbp-0x30], r8
+	mov	[rbp-0x38], r9
+	movaps	[rbp-0x50], xmm0
+	movaps	[rbp-0x60], xmm1
+	movaps	[rbp-0x70], xmm2
+	movaps	[rbp-0x80], xmm3
+	movaps	[rbp-0x90], xmm4
+	movaps	[rbp-0xA0], xmm5
+	movaps	[rbp-0xB0], xmm6
+	movaps	[rbp-0xC0], xmm7
 
 	call	_object_getClass
 
-	movq	%rax, %rdi
-	leaq	sel_forwardingTargetForSelector_(%rip), %rsi
+	mov	rdi, rax
+	lea	rsi, [rip+sel_forwardingTargetForSelector_]
 	call	_class_respondsToSelector
 
-	testq	%rax, %rax
+	test	rax, rax
 	jz	0f
 
-	movq	-0x10(%rbp), %rdi
-	leaq	sel_forwardingTargetForSelector_(%rip), %rsi
+	mov	rdi, [rbp-0x10]
+	lea	rsi, [rip+sel_forwardingTargetForSelector_]
 	call	_objc_msg_lookup
 
-	movq	-0x10(%rbp), %rdi
-	leaq	sel_forwardingTargetForSelector_(%rip), %rsi
-	movq	-0x18(%rbp), %rdx
-	call	*%rax
+	mov	rdi, [rbp-0x10]
+	lea	rsi, [rip+sel_forwardingTargetForSelector_]
+	mov	rdx, [rbp-0x18]
+	call	rax
 
-	testq	%rax, %rax
+	test	rax, rax
 	jz	0f
-	cmpq	-0x10(%rbp), %rax
+	cmp	rax, [rbp-0x10]
 	je	0f
 
-	movq	%rax, -0x10(%rbp)
+	mov	[rbp-0x10], rax
 
-	movq	%rax, %rdi
-	movq	-0x18(%rbp), %rsi
+	mov	rdi, rax
+	mov	rsi, [rbp-0x18]
 	call	_objc_msg_lookup
-	movq	%rax, %r11
+	mov	r11, rax
 
 	/* Restore all arguments */
-	movaps	-0xC0(%rbp), %xmm7
-	movaps	-0xB0(%rbp), %xmm6
-	movaps	-0xA0(%rbp), %xmm5
-	movaps	-0x90(%rbp), %xmm4
-	movaps	-0x80(%rbp), %xmm3
-	movaps	-0x70(%rbp), %xmm2
-	movaps	-0x60(%rbp), %xmm1
-	movaps	-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
+	movaps	xmm7, [rbp-0xC0]
+	movaps	xmm6, [rbp-0xB0]
+	movaps	xmm5, [rbp-0xA0]
+	movaps	xmm4, [rbp-0x90]
+	movaps	xmm3, [rbp-0x80]
+	movaps	xmm2, [rbp-0x70]
+	movaps	xmm1, [rbp-0x60]
+	movaps	xmm0, [rbp-0x50]
+	mov	r9,   [rbp-0x38]
+	mov	r8,   [rbp-0x30]
+	mov	rcx,  [rbp-0x28]
+	mov	rdx,  [rbp-0x20]
+	mov	rsi,  [rbp-0x18]
+	mov	rdi,  [rbp-0x10]
+	mov	rax,  [rbp-0x08]
+
+	mov	rsp, rbp
+	pop	rbp
+
+	jmp	r11
 
 0:
-	movq	-0x10(%rbp), %rdi
-	movq	-0x18(%rbp), %rsi
+	mov	rdi, [rbp-0x10]
+	mov	rsi, [rbp-0x18]
 
-	movq	%rbp, %rsp
-	popq	%rbp
+	mov	rsp, rbp
+	pop	rbp
 
 	jmp	_of_method_not_found
 
 _of_forward_stret:
-	pushq	%rbp
-	movq	%rsp, %rbp
+	push	rbp
+	mov	rbp, rsp
 
 	/* 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)
-	movaps	%xmm0, -0x50(%rbp)
-	movaps	%xmm1, -0x60(%rbp)
-	movaps	%xmm2, -0x70(%rbp)
-	movaps	%xmm3, -0x80(%rbp)
-	movaps	%xmm4, -0x90(%rbp)
-	movaps	%xmm5, -0xA0(%rbp)
-	movaps	%xmm6, -0xB0(%rbp)
-	movaps	%xmm7, -0xC0(%rbp)
-
-	movq	%rsi, %rdi
+	sub	rsp, 0xC0	/* 16-byte alignment */
+	mov	[rbp-0x08], rax
+	mov	[rbp-0x10], rdi
+	mov	[rbp-0x18], rsi
+	mov	[rbp-0x20], rdx
+	mov	[rbp-0x28], rcx
+	mov	[rbp-0x30], r8
+	mov	[rbp-0x38], r9
+	movaps	[rbp-0x50], xmm0
+	movaps	[rbp-0x60], xmm1
+	movaps	[rbp-0x70], xmm2
+	movaps	[rbp-0x80], xmm3
+	movaps	[rbp-0x90], xmm4
+	movaps	[rbp-0xA0], xmm5
+	movaps	[rbp-0xB0], xmm6
+	movaps	[rbp-0xC0], xmm7
+
+	mov	rdi, rsi
 	call	_object_getClass
 
-	movq	%rax, %rdi
-	leaq	sel_forwardingTargetForSelector_(%rip), %rsi
+	mov	rdi, rax
+	lea	rsi, [rip+sel_forwardingTargetForSelector_]
 	call	_class_respondsToSelector
 
-	testq	%rax, %rax
+	test	rax, rax
 	jz	0f
 
-	movq	-0x18(%rbp), %rdi
-	leaq	sel_forwardingTargetForSelector_(%rip), %rsi
+	mov	rdi, [rbp-0x18]
+	lea	rsi, [rip+sel_forwardingTargetForSelector_]
 	call	_objc_msg_lookup
 
-	movq	-0x18(%rbp), %rdi
-	leaq	sel_forwardingTargetForSelector_(%rip), %rsi
-	movq	-0x20(%rbp), %rdx
-	call	*%rax
+	mov	rdi, [rbp-0x18]
+	lea	rsi, [rip+sel_forwardingTargetForSelector_]
+	mov	rdx, [rbp-0x20]
+	call	rax
 
-	testq	%rax, %rax
+	test	rax, rax
 	jz	0f
-	cmpq	-0x18(%rbp), %rax
+	cmp	rax, [rbp-0x18]
 	je	0f
 
-	movq	%rax, -0x18(%rbp)
+	mov	[rbp-0x18], rax
 
-	movq	%rax, %rdi
-	movq	-0x20(%rbp), %rsi
+	mov	rdi, rax
+	mov	rsi, [rbp-0x20]
 	call	_objc_msg_lookup_stret
-	movq	%rax, %r11
+	mov	r11, rax
 
 	/* Restore all arguments */
-	movaps	-0xC0(%rbp), %xmm7
-	movaps	-0xB0(%rbp), %xmm6
-	movaps	-0xA0(%rbp), %xmm5
-	movaps	-0x90(%rbp), %xmm4
-	movaps	-0x80(%rbp), %xmm3
-	movaps	-0x70(%rbp), %xmm2
-	movaps	-0x60(%rbp), %xmm1
-	movaps	-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
+	movaps	xmm7, [rbp-0xC0]
+	movaps	xmm6, [rbp-0xB0]
+	movaps	xmm5, [rbp-0xA0]
+	movaps	xmm4, [rbp-0x90]
+	movaps	xmm3, [rbp-0x80]
+	movaps	xmm2, [rbp-0x70]
+	movaps	xmm1, [rbp-0x60]
+	movaps	xmm0, [rbp-0x50]
+	mov	r9,   [rbp-0x38]
+	mov	r8,   [rbp-0x30]
+	mov	rcx,  [rbp-0x28]
+	mov	rdx,  [rbp-0x20]
+	mov	rsi,  [rbp-0x18]
+	mov	rdi,  [rbp-0x10]
+	mov	rax,  [rbp-0x08]
+
+	mov	rsp, rbp
+	pop	rbp
+
+	jmp	r11
 
 0:
-	movq	-0x10(%rbp), %rdi
-	movq	-0x18(%rbp), %rsi
-	movq	-0x20(%rbp), %rdx
+	mov	rdi, [rbp-0x10]
+	mov	rsi, [rbp-0x18]
+	mov	rdx, [rbp-0x20]
 
-	movq	%rbp, %rsp
-	popq	%rbp
+	mov	rsp, rbp
+	pop	rbp
 
 	jmp	_of_method_not_found_stret
 
 init:
-	leaq	module(%rip), %rdi
+	lea	rdi, [rip+module]
 	jmp	___objc_exec_class
 
 .section __DATA, __mod_init_func, mod_init_funcs
 	.quad init
 

Index: src/invocation/apple-call-x86_64.S
==================================================================
--- src/invocation/apple-call-x86_64.S
+++ src/invocation/apple-call-x86_64.S
@@ -16,103 +16,105 @@
  */
 
 #include "config.h"
 
 #include "invoke-x86_64.h"
+
+.intel_syntax noprefix
 
 .globl _of_invocation_call
 
 .section __TEXT, __text, regular, pure_instructions
 _of_invocation_call:
-	pushq	%rbp
-	movq	%rsp, %rbp
-
-	subq	$16, %rsp
-	andq	$-16, %rsp
-	movq	%rdi, -8(%rbp)
-
-	leaq	OFFSET_STACK(%rdi), %rdx
-	movq	OFFSET_STACK_SIZE(%rdi), %rcx
-
-	testq	$1, %rcx
+	push	rbp
+	mov	rbp, rsp
+
+	sub	rsp, 16
+	and	rsp, -16
+	mov	[rbp-8], rdi
+
+	lea	rdx, [rdi+OFFSET_STACK]
+	mov	rcx, [rdi+OFFSET_STACK_SIZE]
+
+	test	rcx, 1
 	jnz	Lfix_align
 
 Lfill_stack:
-	testq	%rcx, %rcx
+	test	rcx, rcx
 	jz	Lstack_filled
 
-	decq	%rcx
-	movq	(%rdx,%rcx,8), %r11
-	pushq	%r11
+	dec	rcx
+	mov	r11, [rdx+rcx*8]
+	push	r11
 
 	jmp	Lfill_stack
 
 Lstack_filled:
-	movb	OFFSET_NUM_SSE_USED(%rdi), %al
-
-	movaps	OFFSET_SSE_INOUT+112(%rdi), %xmm7
-	movaps	OFFSET_SSE_INOUT+96(%rdi), %xmm6
-	movaps	OFFSET_SSE_INOUT+80(%rdi), %xmm5
-	movaps	OFFSET_SSE_INOUT+64(%rdi), %xmm4
-	movaps	OFFSET_SSE_INOUT+48(%rdi), %xmm3
-	movaps	OFFSET_SSE_INOUT+32(%rdi), %xmm2
-	movaps	OFFSET_SSE_INOUT+16(%rdi), %xmm1
-	movaps	OFFSET_SSE_INOUT(%rdi), %xmm0
-
-	movq	OFFSET_GPR_IN+40(%rdi), %r9
-	movq	OFFSET_GPR_IN+32(%rdi), %r8
-	movq	OFFSET_GPR_IN+24(%rdi), %rcx
-	movq	OFFSET_GPR_IN+16(%rdi), %rdx
-	movq	OFFSET_GPR_IN+8(%rdi), %rsi
-
-	movb	OFFSET_RETURN_TYPE(%rdi), %r11b
-	movq	OFFSET_GPR_IN(%rdi), %rdi
-
-	cmpb	$RETURN_TYPE_STRET, %r11b
+	mov	al, [rdi+OFFSET_NUM_SSE_USED]
+
+	movaps	xmm7, [rdi+OFFSET_SSE_INOUT+112]
+	movaps	xmm6, [rdi+OFFSET_SSE_INOUT+96]
+	movaps	xmm5, [rdi+OFFSET_SSE_INOUT+80]
+	movaps	xmm4, [rdi+OFFSET_SSE_INOUT+64]
+	movaps	xmm3, [rdi+OFFSET_SSE_INOUT+48]
+	movaps	xmm2, [rdi+OFFSET_SSE_INOUT+32]
+	movaps	xmm1, [rdi+OFFSET_SSE_INOUT+16]
+	movaps	xmm0, [rdi+OFFSET_SSE_INOUT]
+
+	mov	r9,  [rdi+OFFSET_GPR_IN+40]
+	mov	r8,  [rdi+OFFSET_GPR_IN+32]
+	mov	rcx, [rdi+OFFSET_GPR_IN+24]
+	mov	rdx, [rdi+OFFSET_GPR_IN+16]
+	mov	rsi, [rdi+OFFSET_GPR_IN+8]
+
+	mov	r11b, [rdi+OFFSET_RETURN_TYPE]
+	mov	rdi,  [rdi+OFFSET_GPR_IN]
+
+	cmp	r11b, RETURN_TYPE_STRET
 	je	Lcall_send_stret
 
-	cmpb	$RETURN_TYPE_JMP, %r11b
+	cmp	r11b, RETURN_TYPE_JMP
 	je	_objc_msgSend
 
-	cmpb	$RETURN_TYPE_JMP_STRET, %r11b
+	cmp	r11b, RETURN_TYPE_JMP_STRET
 	je	_objc_msgSend_stret
 
 	call	_objc_msgSend
 
 Lafter_send:
-	movq	-8(%rbp), %rdi
-	movq	%rax, OFFSET_GPR_OUT(%rdi)
-	movq	%rdx, OFFSET_GPR_OUT+8(%rdi)
-	movaps	%xmm0, OFFSET_SSE_INOUT(%rdi)
-	movaps	%xmm1, OFFSET_SSE_INOUT+16(%rdi)
-
-	movb	OFFSET_RETURN_TYPE(%rdi), %r11b
-
-	cmpb	$RETURN_TYPE_X87, %r11b
+	mov	rdi, [rbp-8]
+	mov	[rdi+OFFSET_GPR_OUT], rax
+	mov	[rdi+OFFSET_GPR_OUT+8], rdx
+	movaps	[rdi+OFFSET_SSE_INOUT], xmm0
+	movaps	[rdi+OFFSET_SSE_INOUT+16], xmm1
+
+	mov	r11b, [rdi+OFFSET_RETURN_TYPE]
+
+	cmp	r11b, RETURN_TYPE_X87
 	je	Lpop_long_double
 
-	cmpb	$RETURN_TYPE_COMPLEX_X87, %r11b
+	cmp	r11b, RETURN_TYPE_COMPLEX_X87
 	je	Lpop_complex_long_double
 
 Lreturn:
-	movq	%rbp, %rsp
-	popq	%rbp
+	mov	rsp, rbp
+	pop	rbp
 
 	ret
 
 Lfix_align:
-	xorq	%r11, %r11
-	pushq	%r11
+	xor	r11, r11
+	push	r11
 	jmp	Lfill_stack
 
 Lcall_send_stret:
 	call	_objc_msgSend_stret
 	jmp	Lafter_send
 
 Lpop_long_double:
-	fstpt	OFFSET_X87_OUT(%rdi)
+	fstp	tbyte ptr [rdi+OFFSET_X87_OUT]
 	jmp	Lreturn
 
 Lpop_complex_long_double:
-	fstpt	OFFSET_X87_OUT(%rdi)
-	fstpt	OFFSET_X87_OUT+16(%rdi)
+	fstp	tbyte ptr [rdi+OFFSET_X87_OUT]
+	fstp	tbyte ptr [rdi+OFFSET_X87_OUT+16]
 	jmp	Lreturn

Index: src/invocation/call-x86_64-elf.S
==================================================================
--- src/invocation/call-x86_64-elf.S
+++ src/invocation/call-x86_64-elf.S
@@ -16,125 +16,127 @@
  */
 
 #include "config.h"
 
 #include "invoke-x86_64.h"
+
+.intel_syntax noprefix
 
 .globl of_invocation_call
 
 .section .text
 of_invocation_call:
-	pushq	%rbp
-	movq	%rsp, %rbp
-
-	subq	$16, %rsp
-	andq	$-16, %rsp
-	movq	%rdi, -8(%rbp)
-
-	movb	OFFSET_RETURN_TYPE(%rdi), %r11b
-	cmpb	$RETURN_TYPE_STRET, %r11b
-	je	.lookup_stret
-	cmpb	$RETURN_TYPE_JMP_STRET, %r11b
-	je	.lookup_stret
-
-	movq	OFFSET_GPR_IN+8(%rdi), %rsi
-	movq	OFFSET_GPR_IN+0(%rdi), %rdi
+	pushq	rbp
+	mov	rbp, rsp
+
+	sub	rsp, 16
+	and	rsp, -16
+	mov	[rbp-8], rdi
+
+	mov	r11b, [rdi+OFFSET_RETURN_TYPE]
+	cmp	r11b, RETURN_TYPE_STRET
+	je	short .Llookup_stret
+	cmp	r11b, RETURN_TYPE_JMP_STRET
+	je	short .Llookup_stret
+
+	mov	rsi, [rdi+OFFSET_GPR_IN+8]
+	mov	rdi, [rdi+OFFSET_GPR_IN]
 	call	objc_msg_lookup@PLT
 
-.after_lookup:
-	movq	%rax, -16(%rbp)
-	movq	-8(%rbp), %rdi
-
-	leaq	OFFSET_STACK(%rdi), %rdx
-	movq	OFFSET_STACK_SIZE(%rdi), %rcx
-
-	testq	$1, %rcx
-	jnz	.fix_align
-
-.fill_stack:
-	testq	%rcx, %rcx
-	jz	.stack_filled
-
-	decq	%rcx
-	movq	(%rdx,%rcx,8), %r11
-	pushq	%r11
-
-	jmp	.fill_stack
-
-.stack_filled:
-	movb	OFFSET_NUM_SSE_USED(%rdi), %al
-
-	movaps	OFFSET_SSE_INOUT+112(%rdi), %xmm7
-	movaps	OFFSET_SSE_INOUT+96(%rdi), %xmm6
-	movaps	OFFSET_SSE_INOUT+80(%rdi), %xmm5
-	movaps	OFFSET_SSE_INOUT+64(%rdi), %xmm4
-	movaps	OFFSET_SSE_INOUT+48(%rdi), %xmm3
-	movaps	OFFSET_SSE_INOUT+32(%rdi), %xmm2
-	movaps	OFFSET_SSE_INOUT+16(%rdi), %xmm1
-	movaps	OFFSET_SSE_INOUT(%rdi), %xmm0
-
-	movq	OFFSET_GPR_IN+40(%rdi), %r9
-	movq	OFFSET_GPR_IN+32(%rdi), %r8
-	movq	OFFSET_GPR_IN+24(%rdi), %rcx
-	movq	OFFSET_GPR_IN+16(%rdi), %rdx
-	movq	OFFSET_GPR_IN+8(%rdi), %rsi
-
-	movb	OFFSET_RETURN_TYPE(%rdi), %r11b
-	movq	OFFSET_GPR_IN(%rdi), %rdi
-
-	cmpb	$RETURN_TYPE_JMP, %r11b
-	je	.jmp_into_method
-	cmpb	$RETURN_TYPE_JMP_STRET, %r11b
-	je	.jmp_into_method
-
-	movq	-16(%rbp), %r11
-	call	*%r11
-
-.after_send:
-	movq	-8(%rbp), %rdi
-	movq	%rax, OFFSET_GPR_OUT(%rdi)
-	movq	%rdx, OFFSET_GPR_OUT+8(%rdi)
-	movaps	%xmm0, OFFSET_SSE_INOUT(%rdi)
-	movaps	%xmm1, OFFSET_SSE_INOUT+16(%rdi)
-
-	movb	OFFSET_RETURN_TYPE(%rdi), %r11b
-
-	cmpb	$RETURN_TYPE_X87, %r11b
-	je	.pop_long_double
-
-	cmpb	$RETURN_TYPE_COMPLEX_X87, %r11b
-	je	.pop_complex_long_double
-
-.return:
-	movq	%rbp, %rsp
-	popq	%rbp
+.Lafter_lookup:
+	mov	[rbp-16], rax
+	mov	rdi, [rbp-8]
+
+	lea	rdx, [rdi+OFFSET_STACK]
+	mov	rcx, [rdi+OFFSET_STACK_SIZE]
+
+	test	rcx, 1
+	jnz	short .Lfix_align
+
+.Lfill_stack:
+	test	rcx, rcx
+	jz	short .Lstack_filled
+
+	dec	rcx
+	mov	r11, [rdx+rcx*8]
+	push	r11
+
+	jmp	short .Lfill_stack
+
+.Lstack_filled:
+	mov	al, [rdi+OFFSET_NUM_SSE_USED]
+
+	movaps	xmm7, [rdi+OFFSET_SSE_INOUT+112]
+	movaps	xmm6, [rdi+OFFSET_SSE_INOUT+96]
+	movaps	xmm5, [rdi+OFFSET_SSE_INOUT+80]
+	movaps	xmm4, [rdi+OFFSET_SSE_INOUT+64]
+	movaps	xmm3, [rdi+OFFSET_SSE_INOUT+48]
+	movaps	xmm2, [rdi+OFFSET_SSE_INOUT+32]
+	movaps	xmm1, [rdi+OFFSET_SSE_INOUT+16]
+	movaps	xmm0, [rdi+OFFSET_SSE_INOUT]
+
+	mov	r9,  [rdi+OFFSET_GPR_IN+40]
+	mov	r8,  [rdi+OFFSET_GPR_IN+32]
+	mov	rcx, [rdi+OFFSET_GPR_IN+24]
+	mov	rdx, [rdi+OFFSET_GPR_IN+16]
+	mov	rsi, [rdi+OFFSET_GPR_IN+8]
+
+	mov	r11b, [rdi+OFFSET_RETURN_TYPE]
+	mov	rdi,  [rdi+OFFSET_GPR_IN]
+
+	cmp	r11b, RETURN_TYPE_JMP
+	je	short .Ljmp_into_method
+	cmp	r11b, RETURN_TYPE_JMP_STRET
+	je	short .Ljmp_into_method
+
+	mov	r11, [rbp-16]
+	call	r11
+
+.Lafter_send:
+	mov	rdi, [rbp-8]
+	mov	[rdi+OFFSET_GPR_OUT], rax
+	mov	[rdi+OFFSET_GPR_OUT+8], rdx
+	movaps	[rdi+OFFSET_SSE_INOUT], xmm0
+	movaps	[rdi+OFFSET_SSE_INOUT+16], xmm1
+
+	mov	r11b, [rdi+OFFSET_RETURN_TYPE]
+
+	cmp	r11b, RETURN_TYPE_X87
+	je	short .Lpop_long_double
+
+	cmp	r11b, RETURN_TYPE_COMPLEX_X87
+	je	short .Lpop_complex_long_double
+
+.Lreturn:
+	mov	rsp, rbp
+	pop	rbp
 
 	ret
 
-.fix_align:
-	xorq	%r11, %r11
-	pushq	%r11
-	jmp	.fill_stack
-
-.lookup_stret:
-	movq	OFFSET_GPR_IN+16(%rdi), %rsi
-	movq	OFFSET_GPR_IN+8(%rdi), %rdi
+.Lfix_align:
+	xor	r11, r11
+	push	r11
+	jmp	short .Lfill_stack
+
+.Llookup_stret:
+	mov	rsi, [rdi+OFFSET_GPR_IN+16]
+	mov	rdi, [rdi+OFFSET_GPR_IN+8]
 	call	objc_msg_lookup_stret@PLT
 
-	jmp	.after_lookup
-
-.jmp_into_method:
-	movq	-16(%rbp), %r11
-	jmp	*%r11
-
-.pop_long_double:
-	fstpt	OFFSET_X87_OUT(%rdi)
-	jmp	.return
-
-.pop_complex_long_double:
-	fstpt	OFFSET_X87_OUT(%rdi)
-	fstpt	OFFSET_X87_OUT+16(%rdi)
-	jmp	.return
+	jmp	short .Lafter_lookup
+
+.Ljmp_into_method:
+	mov	r11, [rbp-16]
+	jmp	r11
+
+.Lpop_long_double:
+	fstp	tbyte ptr [rdi+OFFSET_X87_OUT]
+	jmp	short .Lreturn
+
+.Lpop_complex_long_double:
+	fstp	tbyte ptr [rdi+OFFSET_X87_OUT]
+	fstp	tbyte ptr [rdi+OFFSET_X87_OUT+16]
+	jmp	short .Lreturn
 
 #ifdef OF_LINUX
 .section .note.GNU-stack, "", %progbits
 #endif

Index: src/runtime/lookup-asm/lookup-asm-x86_64-macho.S
==================================================================
--- src/runtime/lookup-asm/lookup-asm-x86_64-macho.S
+++ src/runtime/lookup-asm/lookup-asm-x86_64-macho.S
@@ -14,62 +14,64 @@
  * LICENSE.GPLv2 or LICENSE.GPLv3 respectively included in the packaging of this
  * file.
  */
 
 #include "config.h"
+
+.intel_syntax noprefix
 
 .globl _objc_msg_lookup
 .globl _objc_msg_lookup_stret
 .globl _objc_msg_lookup_super
 .globl _objc_msg_lookup_super_stret
 
 .section __TEXT, __text, regular, pure_instructions
 .macro generate_lookup
 $0:
-	testq	%rdi, %rdi
+	test	rdi, rdi
 	jz	ret_nil
 
-	movq	(%rdi), %r8
-	movq	64(%r8), %r8
+	mov	r8, [rdi]
+	mov	r8, [r8+64]
 
 Lmain_$0:
-	movq	(%rsi), %rax
-	movzbl	%ah, %ecx
-	movzbl	%al, %edx
+	mov	rax, [rsi]
+	movzx	ecx, ah
+	movzx	edx, al
 #ifdef OF_SELUID24
-	shrl	$$16, %eax
+	shr	eax, 16
 
-	movq	(%r8,%rax,8), %r8
+	mov	r8, [r8+rax*8]
 #endif
-	movq	(%r8,%rcx,8), %r8
-	movq	(%r8,%rdx,8), %rax
+	mov	r8, [r8+rcx*8]
+	mov	rax, [r8+rdx*8]
 
-	testq	%rax, %rax
+	test	rax, rax
 	jz	$1
 
 	ret
 .endmacro
 
 .macro generate_lookup_super
 $0:
-	movq	%rdi, %r8
-	movq	(%rdi), %rdi
-	testq	%rdi, %rdi
+	mov	r8, rdi
+	mov	rdi, [rdi]
+	test	rdi, rdi
 	jz	ret_nil
 
-	movq	8(%r8), %r8
-	movq	64(%r8), %r8
+	mov	r8, [r8+8]
+	mov	r8, [r8+64]
 	jmp	Lmain_$1
 .endmacro
 
 generate_lookup _objc_msg_lookup, _objc_method_not_found
 generate_lookup _objc_msg_lookup_stret, _objc_method_not_found_stret
 generate_lookup_super _objc_msg_lookup_super, _objc_msg_lookup
 generate_lookup_super _objc_msg_lookup_super_stret, _objc_msg_lookup_stret
 
 ret_nil:
-	leaq	nil_method(%rip), %rax
+	lea	rax, [rip+nil_method]
 	ret
 
 nil_method:
-	movq	%rdi, %rax
+	mov	rax, rdi
 	ret