Move fast memcpy to llcommon and use it in llalignedarray pushback on all platforms. Code Review: DaveP

4 files changed, 8 insertions, 85 deletions

diff --git a/indra/llmath/llsimdmath.h b/indra/llmath/llsimdmath.h
index 01458521ec..cebd2ace7d 100644
--- a/indra/llmath/llsimdmath.h
+++ b/indra/llmath/llsimdmath.h
@@ -39,34 +39,6 @@
 #include <stdint.h>
 #endif
 
-template <typename T> T* LL_NEXT_ALIGNED_ADDRESS(T* address) 
-{ 
-	return reinterpret_cast<T*>(
-		(reinterpret_cast<uintptr_t>(address) + 0xF) & ~0xF);
-}
-
-template <typename T> T* LL_NEXT_ALIGNED_ADDRESS_64(T* address) 
-{ 
-	return reinterpret_cast<T*>(
-		(reinterpret_cast<uintptr_t>(address) + 0x3F) & ~0x3F);
-}
-
-#if LL_LINUX || LL_DARWIN
-
-#define			LL_ALIGN_PREFIX(x)
-#define			LL_ALIGN_POSTFIX(x)		__attribute__((aligned(x)))
-
-#elif LL_WINDOWS
-
-#define			LL_ALIGN_PREFIX(x)		__declspec(align(x))
-#define			LL_ALIGN_POSTFIX(x)
-
-#else
-#error "LL_ALIGN_PREFIX and LL_ALIGN_POSTFIX undefined"
-#endif
-
-#define LL_ALIGN_16(var) LL_ALIGN_PREFIX(16) var LL_ALIGN_POSTFIX(16)
-
 #include <xmmintrin.h>
 #include <emmintrin.h>
 
diff --git a/indra/llmath/llvector4a.cpp b/indra/llmath/llvector4a.cpp
index 6edeb0fefe..570fa41a43 100644
--- a/indra/llmath/llvector4a.cpp
+++ b/indra/llmath/llvector4a.cpp
@@ -41,55 +41,7 @@ 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)
 {
-	assert(src != NULL);
-	assert(dst != NULL);
-	assert(bytes > 0);
-	assert((bytes % sizeof(F32))== 0); 
-	ll_assert_aligned(src,16);
-	ll_assert_aligned(dst,16);
-	assert(bytes%16==0);
-
-	F32* end = dst + (bytes / sizeof(F32) );
-
-	if (bytes > 64)
-	{
-		F32* begin_64 = LL_NEXT_ALIGNED_ADDRESS_64(dst);
-		
-		//at least 64 (16*4) bytes before the end of the destination, switch to 16 byte copies
-		F32* end_64 = end-16;
-		
-		_mm_prefetch((char*)begin_64, _MM_HINT_NTA);
-		_mm_prefetch((char*)begin_64 + 64, _MM_HINT_NTA);
-		_mm_prefetch((char*)begin_64 + 128, _MM_HINT_NTA);
-		_mm_prefetch((char*)begin_64 + 192, _MM_HINT_NTA);
-		
-		while (dst < begin_64)
-		{
-			copy4a(dst, src);
-			dst += 4;
-			src += 4;
-		}
-		
-		while (dst < end_64)
-		{
-			_mm_prefetch((char*)src + 512, _MM_HINT_NTA);
-			_mm_prefetch((char*)dst + 512, _MM_HINT_NTA);
-			copy4a(dst, src);
-			copy4a(dst+4, src+4);
-			copy4a(dst+8, src+8);
-			copy4a(dst+12, src+12);
-			
-			dst += 16;
-			src += 16;
-		}
-	}
-
-	while (dst < end)
-	{
-		copy4a(dst, src);
-		dst += 4;
-		src += 4;
-	}
+        ll_memcpy_nonaliased_aligned_16((char*)dst, (char*)src, bytes);
 }
 
 void LLVector4a::setRotated( const LLRotation& rot, const LLVector4a& vec )
diff --git a/indra/llmath/llvector4a.h b/indra/llmath/llvector4a.h
index 1a478bc8de..0526793d3a 100644
--- a/indra/llmath/llvector4a.h
+++ b/indra/llmath/llvector4a.h
@@ -93,11 +93,7 @@ public:
 	
 	LLVector4a()
 	{ //DO NOT INITIALIZE -- The overhead is completely unnecessary
-// This assert is causing spurious referenced before set warnings on GCC 4.3.4
-//
-#if !LL_LINUX
 		ll_assert_aligned(this,16);
-#endif
 	}
 	
 	LLVector4a(F32 x, F32 y, F32 z, F32 w = 0.f)
diff --git a/indra/llmath/llvolume.cpp b/indra/llmath/llvolume.cpp
index 4f3e753276..e7a0bc7df7 100644
--- a/indra/llmath/llvolume.cpp
+++ b/indra/llmath/llvolume.cpp
@@ -4729,10 +4729,13 @@ void LLVolumeFace::optimize(F32 angle_cutoff)
 		}
 	}
 
-	llassert(new_face.mNumIndices == mNumIndices);
-	llassert(new_face.mNumVertices <= mNumVertices);
-
-	swapData(new_face);
+	// Only swap data if we've actually optimized the mesh
+	//
+	if (new_face.mNumVertices < mNumVertices)
+	{
+		llassert(new_face.mNumIndices == mNumIndices);
+		swapData(new_face);
+	}
 }
 
 class LLVCacheTriangleData;