[runtime] Make sure the movaps instruction only accesses 16-byte aligned memory addresses.

runtime/trampolines-i386-objc_msgSend-copyframe.inc

@@ -23,7 +23,7 @@
 	sall	$4,%eax
 	subl	%eax,%esp
 	# store the number somewhere so we can restore the stack pointer later
-	movl	%eax,-80(%ebp)
+	movl	%eax,-16(%ebp)
 	
 	# copy arguments from old location in the stack to new location in the stack
 	# %ecx will hold the amount of bytes left to copy
@@ -34,7 +34,7 @@
 	# (arguments smaller than 4 bytes are extended to 4 bytes according to
 	# http://developer.apple.com/library/mac/#documentation/DeveloperTools/Conceptual/LowLevelABI/130-IA-32_Function_Calling_Conventions/IA32.html#//apple_ref/doc/uid/TP40002492-SW4)
 	
-	movl	-80(%ebp),%ecx # ecx = frame_length
+	movl	-16(%ebp),%ecx # ecx = frame_length
 	leal	8(%ebp),%esi   # esi = address of first argument we got (buffer)
 	movl	%esp,%edi      # edi = address of the bottom of the stack
 
@@ -47,7 +47,7 @@ L_start:
 	jmp L_start               # }
 L_end:
 
-	movaps	-16(%ebp), %xmm0
-	movaps	-32(%ebp), %xmm1
-	movaps	-48(%ebp), %xmm2
-	movaps	-64(%ebp), %xmm3
+	movaps	-24(%ebp), %xmm0
+	movaps	-40(%ebp), %xmm1
+	movaps	-56(%ebp), %xmm2
+	movaps	-72(%ebp), %xmm3

runtime/trampolines-i386-objc_msgSend-post.inc

@@ -1,6 +1,6 @@
 Lafterinvoke:
-	addl	-80(%ebp), %esp
-	addl	$68, %esp
+	addl	-16(%ebp), %esp
+	addl	$92, %esp
 	popl	%ebx
 	popl	%edi
 	popl	%esi

runtime/trampolines-i386-objc_msgSend-pre.inc

@@ -17,11 +17,11 @@ Lfunc_begin0:
 	pushl	%esi		# %ebp-4
 	pushl	%edi		# %ebp-8
 	pushl	%ebx		# %ebp-12
-	# we need 64 bytes to store xmm0-3
-	# we need 4 bytes to store the result for get_frame_length, which we store in %ebp-80
+	# we need 64 bytes to store xmm0-3, and those 64 bytes need to be 16-byte aligned.
+	# we need 4 bytes to store the result for get_frame_length, which we store in %ebp-16
 	# then 4 more bytes for stack space for the call to get_frame_length (which takes 2 arguments)
-	subl	$76, %esp # to store xmm0-3
-	movaps	%xmm0, -16(%ebp)
-	movaps	%xmm1, -32(%ebp)
-	movaps	%xmm2, -48(%ebp)
-	movaps	%xmm3, -64(%ebp)
+	subl	$92, %esp # to store xmm0-3
+	movaps	%xmm0, -24(%ebp)
+	movaps	%xmm1, -40(%ebp)
+	movaps	%xmm2, -56(%ebp)
+	movaps	%xmm3, -72(%ebp)