summaryrefslogtreecommitdiff
path: root/indra/llmath
diff options
context:
space:
mode:
Diffstat (limited to 'indra/llmath')
-rw-r--r--indra/llmath/CMakeLists.txt1
-rw-r--r--indra/llmath/llcamera.h12
-rw-r--r--indra/llmath/llmatrix3a.h2
-rw-r--r--indra/llmath/llmatrix4a.h2
-rw-r--r--indra/llmath/lloctree.h4
-rw-r--r--indra/llmath/llplane.h4
-rw-r--r--indra/llmath/llsimdmath.h3
-rw-r--r--indra/llmath/llsimdtypes.inl2
-rw-r--r--indra/llmath/llvector4a.cpp8
-rw-r--r--indra/llmath/llvector4a.h6
-rw-r--r--indra/llmath/llvector4a.inl1
-rw-r--r--indra/llmath/llvector4logical.h2
-rw-r--r--indra/llmath/llvolume.cpp45
-rw-r--r--indra/llmath/llvolumeoctree.h26
14 files changed, 74 insertions, 44 deletions
diff --git a/indra/llmath/CMakeLists.txt b/indra/llmath/CMakeLists.txt
index b5e59c1ca3..5865ae030c 100644
--- a/indra/llmath/CMakeLists.txt
+++ b/indra/llmath/CMakeLists.txt
@@ -117,6 +117,7 @@ if (LL_TESTS)
# INTEGRATION TESTS
set(test_libs llmath llcommon ${LLCOMMON_LIBRARIES} ${WINDOWS_LIBRARIES})
# TODO: Some of these need refactoring to be proper Unit tests rather than Integration tests.
+ LL_ADD_INTEGRATION_TEST(alignment "" "${test_libs}")
LL_ADD_INTEGRATION_TEST(llbbox llbbox.cpp "${test_libs}")
LL_ADD_INTEGRATION_TEST(llquaternion llquaternion.cpp "${test_libs}")
LL_ADD_INTEGRATION_TEST(mathmisc "" "${test_libs}")
diff --git a/indra/llmath/llcamera.h b/indra/llmath/llcamera.h
index ec67b91d05..0b591be622 100644
--- a/indra/llmath/llcamera.h
+++ b/indra/llmath/llcamera.h
@@ -60,7 +60,7 @@ static const F32 MAX_FIELD_OF_VIEW = 175.f * DEG_TO_RAD;
// roll(), pitch(), yaw()
// etc...
-
+LL_ALIGN_PREFIX(16)
class LLCamera
: public LLCoordFrame
{
@@ -108,7 +108,7 @@ public:
};
private:
- LLPlane mAgentPlanes[7]; //frustum planes in agent space a la gluUnproject (I'm a bastard, I know) - DaveP
+ LL_ALIGN_16(LLPlane mAgentPlanes[7]); //frustum planes in agent space a la gluUnproject (I'm a bastard, I know) - DaveP
U8 mPlaneMask[8]; // 8 for alignment
F32 mView; // angle between top and bottom frustum planes in radians.
@@ -116,13 +116,13 @@ private:
S32 mViewHeightInPixels; // for ViewHeightInPixels() only
F32 mNearPlane;
F32 mFarPlane;
- LLPlane mLocalPlanes[4];
+ LL_ALIGN_16(LLPlane mLocalPlanes[4]);
F32 mFixedDistance; // Always return this distance, unless < 0
LLVector3 mFrustCenter; // center of frustum and radius squared for ultra-quick exclusion test
F32 mFrustRadiusSquared;
- LLPlane mWorldPlanes[PLANE_NUM];
- LLPlane mHorizPlanes[HORIZ_PLANE_NUM];
+ LL_ALIGN_16(LLPlane mWorldPlanes[PLANE_NUM]);
+ LL_ALIGN_16(LLPlane mHorizPlanes[HORIZ_PLANE_NUM]);
U32 mPlaneCount; //defaults to 6, if setUserClipPlane is called, uses user supplied clip plane in
@@ -208,7 +208,7 @@ protected:
void calculateFrustumPlanes(F32 left, F32 right, F32 top, F32 bottom);
void calculateFrustumPlanesFromWindow(F32 x1, F32 y1, F32 x2, F32 y2);
void calculateWorldFrustumPlanes();
-};
+} LL_ALIGN_POSTFIX(16);
#endif
diff --git a/indra/llmath/llmatrix3a.h b/indra/llmath/llmatrix3a.h
index adb7e3389d..9916cfd2da 100644
--- a/indra/llmath/llmatrix3a.h
+++ b/indra/llmath/llmatrix3a.h
@@ -111,7 +111,7 @@ public:
protected:
- LLVector4a mColumns[3];
+ LL_ALIGN_16(LLVector4a mColumns[3]);
};
diff --git a/indra/llmath/llmatrix4a.h b/indra/llmath/llmatrix4a.h
index 27cf5b79f6..c4cefdb4fa 100644
--- a/indra/llmath/llmatrix4a.h
+++ b/indra/llmath/llmatrix4a.h
@@ -34,7 +34,7 @@
class LLMatrix4a
{
public:
- LLVector4a mMatrix[4];
+ LL_ALIGN_16(LLVector4a mMatrix[4]);
inline void clear()
{
diff --git a/indra/llmath/lloctree.h b/indra/llmath/lloctree.h
index 1b11e83b4a..6c7744cdf1 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/llplane.h b/indra/llmath/llplane.h
index a611894721..3c32441b11 100644
--- a/indra/llmath/llplane.h
+++ b/indra/llmath/llplane.h
@@ -36,6 +36,8 @@
// The plane normal = [A, B, C]
// The closest approach = D / sqrt(A*A + B*B + C*C)
+
+LL_ALIGN_PREFIX(16)
class LLPlane
{
public:
@@ -94,7 +96,7 @@ public:
private:
LLVector4a mV;
-};
+} LL_ALIGN_POSTFIX(16);
diff --git a/indra/llmath/llsimdmath.h b/indra/llmath/llsimdmath.h
index c7cdf7b32c..01458521ec 100644
--- a/indra/llmath/llsimdmath.h
+++ b/indra/llmath/llsimdmath.h
@@ -67,11 +67,10 @@ template <typename T> T* LL_NEXT_ALIGNED_ADDRESS_64(T* address)
#define LL_ALIGN_16(var) LL_ALIGN_PREFIX(16) var LL_ALIGN_POSTFIX(16)
-
-
#include <xmmintrin.h>
#include <emmintrin.h>
+#include "llmemory.h"
#include "llsimdtypes.h"
#include "llsimdtypes.inl"
diff --git a/indra/llmath/llsimdtypes.inl b/indra/llmath/llsimdtypes.inl
index 712239e425..e905c84954 100644
--- a/indra/llmath/llsimdtypes.inl
+++ b/indra/llmath/llsimdtypes.inl
@@ -62,6 +62,7 @@ inline LLSimdScalar operator/(const LLSimdScalar& a, const LLSimdScalar& b)
inline LLSimdScalar operator-(const LLSimdScalar& a)
{
static LL_ALIGN_16(const U32 signMask[4]) = {0x80000000, 0x80000000, 0x80000000, 0x80000000 };
+ ll_assert_aligned(signMask,16);
return _mm_xor_ps(*reinterpret_cast<const LLQuad*>(signMask), a);
}
@@ -146,6 +147,7 @@ inline LLSimdScalar& LLSimdScalar::operator/=(const LLSimdScalar& rhs)
inline LLSimdScalar LLSimdScalar::getAbs() const
{
static const LL_ALIGN_16(U32 F_ABS_MASK_4A[4]) = { 0x7FFFFFFF, 0x7FFFFFFF, 0x7FFFFFFF, 0x7FFFFFFF };
+ ll_assert_aligned(F_ABS_MASK_4A,16);
return _mm_and_ps( mQ, *reinterpret_cast<const LLQuad*>(F_ABS_MASK_4A));
}
diff --git a/indra/llmath/llvector4a.cpp b/indra/llmath/llvector4a.cpp
index b66b7a7076..6edeb0fefe 100644
--- a/indra/llmath/llvector4a.cpp
+++ b/indra/llmath/llvector4a.cpp
@@ -24,6 +24,7 @@
* $/LicenseInfo$
*/
+#include "llmemory.h"
#include "llmath.h"
#include "llquantize.h"
@@ -44,7 +45,10 @@ extern const LLVector4a LL_V4A_EPSILON = reinterpret_cast<const LLVector4a&> ( F
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)
@@ -189,6 +193,8 @@ void LLVector4a::quantize16( const LLVector4a& low, const LLVector4a& high )
LLVector4a oneOverDelta;
{
static LL_ALIGN_16( const F32 F_TWO_4A[4] ) = { 2.f, 2.f, 2.f, 2.f };
+ ll_assert_aligned(F_TWO_4A,16);
+
LLVector4a two; two.load4a( F_TWO_4A );
// Here we use _mm_rcp_ps plus one round of newton-raphson
diff --git a/indra/llmath/llvector4a.h b/indra/llmath/llvector4a.h
index 596082509d..0526793d3a 100644
--- a/indra/llmath/llvector4a.h
+++ b/indra/llmath/llvector4a.h
@@ -32,6 +32,7 @@ class LLRotation;
#include <assert.h>
#include "llpreprocessor.h"
+#include "llmemory.h"
///////////////////////////////////
// FIRST TIME USERS PLEASE READ
@@ -46,6 +47,7 @@ class LLRotation;
// LLVector3/LLVector4.
/////////////////////////////////
+LL_ALIGN_PREFIX(16)
class LLVector4a
{
public:
@@ -82,6 +84,7 @@ public:
}
// Copy words 16-byte blocks from src to dst. Source and destination must not overlap.
+ // Source and dest must be 16-byte aligned and size must be multiple of 16.
static void memcpyNonAliased16(F32* __restrict dst, const F32* __restrict src, size_t bytes);
////////////////////////////////////
@@ -90,6 +93,7 @@ public:
LLVector4a()
{ //DO NOT INITIALIZE -- The overhead is completely unnecessary
+ ll_assert_aligned(this,16);
}
LLVector4a(F32 x, F32 y, F32 z, F32 w = 0.f)
@@ -313,7 +317,7 @@ public:
private:
LLQuad mQ;
-};
+} LL_ALIGN_POSTFIX(16);
inline void update_min_max(LLVector4a& min, LLVector4a& max, const LLVector4a& p)
{
diff --git a/indra/llmath/llvector4a.inl b/indra/llmath/llvector4a.inl
index 7ad22a5631..7c52ffef21 100644
--- a/indra/llmath/llvector4a.inl
+++ b/indra/llmath/llvector4a.inl
@@ -475,6 +475,7 @@ inline void LLVector4a::setLerp(const LLVector4a& lhs, const LLVector4a& rhs, F3
inline LLBool32 LLVector4a::isFinite3() const
{
static LL_ALIGN_16(const U32 nanOrInfMask[4]) = { 0x7f800000, 0x7f800000, 0x7f800000, 0x7f800000 };
+ ll_assert_aligned(nanOrInfMask,16);
const __m128i nanOrInfMaskV = *reinterpret_cast<const __m128i*> (nanOrInfMask);
const __m128i maskResult = _mm_and_si128( _mm_castps_si128(mQ), nanOrInfMaskV );
const LLVector4Logical equalityCheck = _mm_castsi128_ps(_mm_cmpeq_epi32( maskResult, nanOrInfMaskV ));
diff --git a/indra/llmath/llvector4logical.h b/indra/llmath/llvector4logical.h
index dd66b09d43..c5698f7cea 100644
--- a/indra/llmath/llvector4logical.h
+++ b/indra/llmath/llvector4logical.h
@@ -27,6 +27,7 @@
#ifndef LL_VECTOR4LOGICAL_H
#define LL_VECTOR4LOGICAL_H
+#include "llmemory.h"
////////////////////////////
// LLVector4Logical
@@ -77,6 +78,7 @@ public:
inline LLVector4Logical& invert()
{
static const LL_ALIGN_16(U32 allOnes[4]) = { 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF };
+ ll_assert_aligned(allOnes,16);
mQ = _mm_andnot_ps( mQ, *(LLQuad*)(allOnes) );
return *this;
}
diff --git a/indra/llmath/llvolume.cpp b/indra/llmath/llvolume.cpp
index cc9744756f..11fa7080ce 100644
--- a/indra/llmath/llvolume.cpp
+++ b/indra/llmath/llvolume.cpp
@@ -95,17 +95,6 @@ const S32 SCULPT_MIN_AREA_DETAIL = 1;
extern BOOL gDebugGL;
-void assert_aligned(void* ptr, uintptr_t alignment)
-{
-#if 0
- uintptr_t t = (uintptr_t) ptr;
- if (t%alignment != 0)
- {
- llerrs << "Alignment check failed." << llendl;
- }
-#endif
-}
-
BOOL check_same_clock_dir( const LLVector3& pt1, const LLVector3& pt2, const LLVector3& pt3, const LLVector3& norm)
{
LLVector3 test = (pt2-pt1)%(pt3-pt2);
@@ -6962,14 +6951,14 @@ void LLVolumeFace::resizeVertices(S32 num_verts)
if (num_verts)
{
mPositions = (LLVector4a*) ll_aligned_malloc_16(sizeof(LLVector4a)*num_verts);
- assert_aligned(mPositions, 16);
+ ll_assert_aligned(mPositions, 16);
mNormals = (LLVector4a*) ll_aligned_malloc_16(sizeof(LLVector4a)*num_verts);
- assert_aligned(mNormals, 16);
+ ll_assert_aligned(mNormals, 16);
//pad texture coordinate block end to allow for QWORD reads
S32 size = ((num_verts*sizeof(LLVector2)) + 0xF) & ~0xF;
mTexCoords = (LLVector2*) ll_aligned_malloc_16(size);
- assert_aligned(mTexCoords, 16);
+ ll_assert_aligned(mTexCoords, 16);
}
else
{
@@ -6993,14 +6982,17 @@ void LLVolumeFace::pushVertex(const LLVector4a& pos, const LLVector4a& norm, con
// S32 old_size = mNumVertices*16;
//positions
- mPositions = (LLVector4a*) realloc(mPositions, new_size);
+ mPositions = (LLVector4a*) ll_aligned_realloc_16(mPositions, new_size);
+ ll_assert_aligned(mPositions,16);
//normals
- mNormals = (LLVector4a*) realloc(mNormals, new_size);
-
+ mNormals = (LLVector4a*) ll_aligned_realloc_16(mNormals, new_size);
+ ll_assert_aligned(mNormals,16);
+
//tex coords
new_size = ((new_verts*8)+0xF) & ~0xF;
- mTexCoords = (LLVector2*) realloc(mTexCoords, new_size);
+ mTexCoords = (LLVector2*) ll_aligned_realloc_16(mTexCoords, new_size);
+ ll_assert_aligned(mTexCoords,16);
//just clear binormals
@@ -7053,7 +7045,8 @@ void LLVolumeFace::pushIndex(const U16& idx)
S32 old_size = ((mNumIndices*2)+0xF) & ~0xF;
if (new_size != old_size)
{
- mIndices = (U16*) realloc(mIndices, new_size);
+ mIndices = (U16*) ll_aligned_realloc_16(mIndices, new_size);
+ ll_assert_aligned(mIndices,16);
}
mIndices[mNumIndices++] = idx;
@@ -7094,12 +7087,12 @@ void LLVolumeFace::appendFace(const LLVolumeFace& face, LLMatrix4& mat_in, LLMat
}
//allocate new buffer space
- mPositions = (LLVector4a*) realloc(mPositions, new_count*sizeof(LLVector4a));
- assert_aligned(mPositions, 16);
- mNormals = (LLVector4a*) realloc(mNormals, new_count*sizeof(LLVector4a));
- assert_aligned(mNormals, 16);
- mTexCoords = (LLVector2*) realloc(mTexCoords, (new_count*sizeof(LLVector2)+0xF) & ~0xF);
- assert_aligned(mTexCoords, 16);
+ mPositions = (LLVector4a*) ll_aligned_realloc_16(mPositions, new_count*sizeof(LLVector4a));
+ ll_assert_aligned(mPositions, 16);
+ mNormals = (LLVector4a*) ll_aligned_realloc_16(mNormals, new_count*sizeof(LLVector4a));
+ ll_assert_aligned(mNormals, 16);
+ mTexCoords = (LLVector2*) ll_aligned_realloc_16(mTexCoords, (new_count*sizeof(LLVector2)+0xF) & ~0xF);
+ ll_assert_aligned(mTexCoords, 16);
mNumVertices = new_count;
@@ -7145,7 +7138,7 @@ void LLVolumeFace::appendFace(const LLVolumeFace& face, LLMatrix4& mat_in, LLMat
new_count = mNumIndices + face.mNumIndices;
//allocate new index buffer
- mIndices = (U16*) realloc(mIndices, (new_count*sizeof(U16)+0xF) & ~0xF);
+ mIndices = (U16*) ll_aligned_realloc_16(mIndices, (new_count*sizeof(U16)+0xF) & ~0xF);
//get destination address into new index buffer
U16* dst_idx = mIndices+mNumIndices;
diff --git a/indra/llmath/llvolumeoctree.h b/indra/llmath/llvolumeoctree.h
index 688d91dc40..dac97b14b5 100644
--- a/indra/llmath/llvolumeoctree.h
+++ b/indra/llmath/llvolumeoctree.h
@@ -37,6 +37,16 @@
class LLVolumeTriangle : public LLRefCount
{
public:
+ void* operator new(size_t size)
+ {
+ return ll_aligned_malloc_16(size);
+ }
+
+ void operator delete(void* ptr)
+ {
+ ll_aligned_free_16(ptr);
+ }
+
LLVolumeTriangle()
{
@@ -58,7 +68,7 @@ public:
}
- LLVector4a mPositionGroup;
+ LL_ALIGN_16(LLVector4a mPositionGroup);
const LLVector4a* mV[3];
U16 mIndex[3];
@@ -73,6 +83,16 @@ 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 +119,8 @@ 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>
generated by cgit v1.2.3 (git 2.25.1) at 2025-10-31 10:35:55 +0000