From 16bfdc8ee68b04c832c04bcd39e29ce9dbb502c1 Mon Sep 17 00:00:00 2001
From: theraven <theraven@72102866-910b-0410-8b05-ffd578937521>
Date: Mon, 8 Jul 2013 11:02:52 +0000
Subject: [PATCH] Align ivars to 16 byte boundaries when they are larger than a
 pointer.  This is currently overly defensive.

Add a test case for this realignment.
---
 Test/CMakeLists.txt |  1 +
 Test/alignTest.m    | 42 ++++++++++++++++++++++++++++++++++++++++++
 ivar.c              | 31 +++++++++++++++++++++++++++++++
 3 files changed, 74 insertions(+)
 create mode 100644 Test/alignTest.m

diff --git a/Test/CMakeLists.txt b/Test/CMakeLists.txt
index b94979f..2b6b28a 100644
--- a/Test/CMakeLists.txt
+++ b/Test/CMakeLists.txt
@@ -5,6 +5,7 @@
 
 # List of single-file tests.
 set(TESTS
+	alignTest.m
 	AllocatePair.m
 	BlockImpTest.m
 	BlockTest_arc.m
diff --git a/Test/alignTest.m b/Test/alignTest.m
new file mode 100644
index 0000000..92e2164
--- /dev/null
+++ b/Test/alignTest.m
@@ -0,0 +1,42 @@
+#include "stdio.h"
+#include "Test.h"
+
+// This is a large vector type, which the compiler will lower to some sequence
+// of vector ops on the target, or scalar ops if there is no vector FPU.
+typedef double __attribute__((vector_size(32))) v4d;
+
+@interface X : Test
+{
+	id f;
+	id g;
+}
+@end
+@implementation X @end
+
+@interface Vector : X
+{
+	v4d x;
+}
+@end
+@implementation Vector
++ (Vector*)alloc
+{
+	Vector *v = class_createInstance(self, 0);
+	// The initialisation might be done with memset, but will probably be a
+	// vector load / store and so will likely fail if x is incorrectly aligned.
+	v->x = (v4d){1,2,3,4};
+	return v;
+}
+- (void)permute
+{
+	// This will become a sequence of one or more vector operations.  We must
+	// have the correct alignment for x, even after the instance variable
+	// munging, or this will break.
+	x *= (v4d){2,3,4,5};
+}
+@end
+
+int main(void)
+{
+	[[Vector alloc] permute];
+}
diff --git a/ivar.c b/ivar.c
index 3c639aa..471042d 100644
--- a/ivar.c
+++ b/ivar.c
@@ -4,6 +4,7 @@
 #include "class.h"
 #include "ivar.h"
 #include "visibility.h"
+#include "gc_ops.h"
 
 ptrdiff_t objc_alignof_type(const char *);
 ptrdiff_t objc_sizeof_type(const char *);
@@ -28,6 +29,7 @@ PRIVATE void objc_compute_ivar_offsets(Class class)
 			}
 			ivar_start = super->instance_size;
 		}
+		long class_size = 0 - class->instance_size;
 		class->instance_size = ivar_start - class->instance_size;
 		/* For each instance variable, we add the offset if required (it will be zero
 		* if this class is compiled with a static ivar layout).  We then set the
@@ -45,6 +47,35 @@ PRIVATE void objc_compute_ivar_offsets(Class class)
 			for (i = 0 ; i < class->ivars->count ; i++)
 			{
 				struct objc_ivar *ivar = &class->ivars->ivar_list[i];
+				// We are going to be allocating an extra word for the reference count
+				// in front of the object.  This doesn't matter for aligment most of
+				// the time, but if we have an instance variable that is a vector type
+				// then we will need to ensure that we are properly aligned again.
+				long ivar_size = (i+1 == class->ivars->count)
+					? (class_size - ivar->offset)
+					: ivar->offset - class->ivars->ivar_list[i+1].offset;
+				// We only need to do the realignment for things that are
+				// bigger than a pointer, and we don't need to do it in GC mode
+				// where we don't add any extra padding.
+				if (!isGCEnabled && (ivar_size > sizeof(void*)))
+				{
+					long fudge = (ivar_start +ivar->offset + sizeof(void*)) % 16;
+					if (fudge != 0)
+					{
+						// If this is the first ivar in the class, then
+						// we can eat some of the padding that the compiler
+						// added...
+						if ((i == 0) && (ivar->offset > 0) && ((ivar_start + sizeof(void*) %16) == 0))
+						{
+							ivar->offset = 0;
+						}
+						else
+						{
+							ivar_start += fudge;
+							class->instance_size += fudge;
+						}
+					}
+				}
 				ivar->offset += ivar_start;
 				/* If we're using the new ABI then we also set up the faster ivar
 				* offset variables.