SH-2789 WIP - various fixes to force 16-byte alignment

4 files changed, 52 insertions, 47 deletions

diff --git a/indra/llmath/lloctree.h b/indra/llmath/lloctree.h
index 3c1ae45d68..374858be51 100644
--- a/indra/llmath/lloctree.h
+++ b/indra/llmath/lloctree.h
@@ -88,7 +88,7 @@ public:
 	typedef LLOctreeNode<T>		oct_node;
 	typedef LLOctreeListener<T>	oct_listener;
 
-	/*void* operator new(size_t size)
+	void* operator new(size_t size)
 	{
 		return ll_aligned_malloc_16(size);
 	}
@@ -96,7 +96,7 @@ public:
 	void operator delete(void* ptr)
 	{
 		ll_aligned_free_16(ptr);
-	}*/
+	}
 
 	LLOctreeNode(	const LLVector4a& center, 
 					const LLVector4a& size, 
diff --git a/indra/llmath/llvector4a.cpp b/indra/llmath/llvector4a.cpp
index 49b8754cd0..7602ef0cb2 100644..100755
--- a/indra/llmath/llvector4a.cpp
+++ b/indra/llmath/llvector4a.cpp
@@ -41,11 +41,15 @@ extern const LLVector4a LL_V4A_EPSILON = reinterpret_cast<const LLVector4a&> ( F
 
 /*static */void LLVector4a::memcpyNonAliased16(F32* __restrict dst, const F32* __restrict src, size_t bytes)
 {
+	memcpy((void*)dst,(const void*)src,bytes);
+#if 0
 	assert(src != NULL);
 	assert(dst != NULL);
 	assert(bytes > 0);
 	assert((bytes % sizeof(F32))== 0); 
-	
+	ll_assert_aligned(src,16);
+	ll_assert_aligned(dst,16);
+
 	F32* end = dst + (bytes / sizeof(F32) );
 
 	if (bytes > 64)
@@ -87,6 +91,7 @@ extern const LLVector4a LL_V4A_EPSILON = reinterpret_cast<const LLVector4a&> ( F
 		dst += 4;
 		src += 4;
 	}
+#endif
 }
 
 void LLVector4a::setRotated( const LLRotation& rot, const LLVector4a& vec )
diff --git a/indra/llmath/llvolumeoctree.h b/indra/llmath/llvolumeoctree.h
index 688d91dc40..6de7b223be 100644..100755
--- a/indra/llmath/llvolumeoctree.h
+++ b/indra/llmath/llvolumeoctree.h
@@ -73,6 +73,16 @@ class LLVolumeOctreeListener : public LLOctreeListener<LLVolumeTriangle>
 {
 public:
 	
+	void* operator new(size_t size)
+	{
+		return ll_aligned_malloc_16(size);
+	}
+
+	void operator delete(void* ptr)
+	{
+		ll_aligned_free_16(ptr);
+	}
+
 	LLVolumeOctreeListener(LLOctreeNode<LLVolumeTriangle>* node);
 	~LLVolumeOctreeListener();
 	
@@ -99,8 +109,8 @@ public:
 	
 
 public:
-	LLVector4a mBounds[2]; // bounding box (center, size) of this node and all its children (tight fit to objects)
-	LLVector4a mExtents[2]; // extents (min, max) of this node and all its children
+	LL_ALIGN_16(LLVector4a mBounds[2]); // bounding box (center, size) of this node and all its children (tight fit to objects)
+	LL_ALIGN_16(LLVector4a mExtents[2]); // extents (min, max) of this node and all its children
 };
 
 class LLOctreeTriangleRayIntersect : public LLOctreeTraveler<LLVolumeTriangle>
diff --git a/indra/llmath/tests/alignment_test.cpp b/indra/llmath/tests/alignment_test.cpp
index 6b4a454fd2..51a7051e69 100755
--- a/indra/llmath/tests/alignment_test.cpp
+++ b/indra/llmath/tests/alignment_test.cpp
@@ -34,10 +34,21 @@
 #include "../llsimdmath.h"

 #include "../llvector4a.h"

 

+void* operator new(size_t size)

+{

+	return ll_aligned_malloc_16(size);

+}

+

+void operator delete(void *p)

+{

+	ll_aligned_free_16(p);

+}

+

 namespace tut

 {

 

 #define is_aligned(ptr,alignment) ((reinterpret_cast<uintptr_t>(ptr))%(alignment)==0)

+#define is_aligned_relative(ptr,base_ptr,alignment) ((reinterpret_cast<uintptr_t>(ptr)-reinterpret_cast<uintptr_t>(base_ptr))%(alignment)==0)

 

 struct alignment_test {};

 

@@ -51,38 +62,40 @@ class MyVector4a
 	LLQuad mQ;

 } LL_ALIGN_POSTFIX(16);

 

-LL_ALIGN_PREFIX(64)

-class MyBigBlob

+// Verify that aligned allocators perform as advertised.

+template<> template<>

+void alignment_test_object_t::test<1>()

 {

-public:

-	~MyBigBlob() {}

-private:

-	LLQuad mQ[4];

-} LL_ALIGN_POSTFIX(64);

+	const int num_tests = 7;

+	void *align_ptr;

+	for (int i=0; i<num_tests; i++)

+	{

+		align_ptr = ll_aligned_malloc_16(sizeof(MyVector4a));

+		ensure("ll_aligned_malloc_16 failed", is_aligned(align_ptr,16));

+		ll_aligned_free_16(align_ptr);

+

+		align_ptr = ll_aligned_malloc_32(sizeof(MyVector4a));

+		ensure("ll_aligned_malloc_32 failed", is_aligned(align_ptr,32));

+		ll_aligned_free_32(align_ptr);

+	}

+}

 

 // In-place allocation of objects and arrays.

 template<> template<>

-void alignment_test_object_t::test<1>()

+void alignment_test_object_t::test<2>()

 {

 	ensure("LLAlignment reality is broken: ", (1==1));

 

 	MyVector4a vec1;

 	ensure("LLAlignment vec1 unaligned", is_aligned(&vec1,16));

 	

-	MyBigBlob bb1;

-	ensure("LLAlignment bb1 unaligned", is_aligned(&bb1,64));

-		   

-

 	MyVector4a veca[12];

 	ensure("LLAlignment veca unaligned", is_aligned(veca,16));

-

-	MyBigBlob bba[12];

-	ensure("LLAlignment bba unaligned", is_aligned(bba,64));

 }

 

 // Heap allocation of objects and arrays.

 template<> template<>

-void alignment_test_object_t::test<2>()

+void alignment_test_object_t::test<3>()

 {

 	const int ARR_SIZE = 7;

 	for(int i=0; i<ARR_SIZE; i++)

@@ -93,35 +106,12 @@ void alignment_test_object_t::test<2>()
 	}

 

 	MyVector4a *veca = new MyVector4a[ARR_SIZE];

+	ensure("LLAligment veca base", is_aligned(veca,16));

 	for(int i=0; i<ARR_SIZE; i++)

 	{

-		ensure("LLAlignment veca unaligned", is_aligned(&veca[i],16));

-	}

-

-	for(int i=0; i<ARR_SIZE; i++)

-	{

-		void *aligned_addr = _aligned_malloc(sizeof(MyBigBlob),64);

-		MyBigBlob *bbp = new(aligned_addr) MyBigBlob;

-		ensure("LLAlignment bbp unaligned", is_aligned(bbp,64));

-		bbp->~MyBigBlob();

-		_aligned_free(aligned_addr);

-	}

-

-	ensure("LLAlignment big blob size",sizeof(MyBigBlob)==64);

-	void *aligned_addr = _aligned_malloc(ARR_SIZE*sizeof(MyBigBlob),64);

-	MyBigBlob *bba = new(aligned_addr) MyBigBlob[ARR_SIZE];

-	std::cout << "aligned_addr " << aligned_addr << std::endl;

-	std::cout << "bba " << bba << std::endl;

-	for(int i=0; i<ARR_SIZE; i++)

-	{

-		std::cout << "bba test " << i << std::endl;

-		ensure("LLAlignment bba unaligned", is_aligned(&bba[i],64));

-	}

-	for(int i=0; i<ARR_SIZE; i++)

-	{

-		bba[i].~MyBigBlob();

+		std::cout << "veca[" << i << "]" << std::endl;

+		ensure("LLAlignment veca member unaligned", is_aligned(&veca[i],16));

 	}

-	_aligned_free(aligned_addr);

 }

 

 }