summaryrefslogtreecommitdiff
diff options
context:
space:
mode:
Diffstat
-rw-r--r--indra/llcommon/llmemory.h32
-rw-r--r--indra/llmath/lloctree.h6
-rw-r--r--indra/llmath/llvolume.cpp89
-rw-r--r--indra/llrender/llrender.cpp8
-rw-r--r--indra/llrender/llrender.h1
-rw-r--r--indra/llrender/llvertexbuffer.cpp17
-rw-r--r--indra/newview/lldrawable.cpp4
-rw-r--r--indra/newview/llface.cpp8
-rw-r--r--indra/newview/llspatialpartition.cpp14
9 files changed, 102 insertions, 77 deletions
diff --git a/indra/llcommon/llmemory.h b/indra/llcommon/llmemory.h
index 117268cfe7..1c8c91f57e 100644
--- a/indra/llcommon/llmemory.h
+++ b/indra/llcommon/llmemory.h
@@ -34,7 +34,7 @@
#include <stdlib.h>
-inline void* ll_aligned_malloc_16(size_t size) // returned hunk MUST be freed with ll_aligned_free().
+inline void* ll_aligned_malloc_16(size_t size) // returned hunk MUST be freed with ll_aligned_free_16().
{
#if defined(LL_WINDOWS)
return _mm_malloc(size, 16);
@@ -43,13 +43,9 @@ inline void* ll_aligned_malloc_16(size_t size) // returned hunk MUST be freed wi
#else
void *rtn;
if (LL_LIKELY(0 == posix_memalign(&rtn, 16, size)))
- {
return rtn;
- }
else // bad alignment requested, or out of memory
- {
return NULL;
- }
#endif
}
@@ -64,6 +60,32 @@ inline void ll_aligned_free_16(void *p)
#endif
}
+inline void* ll_aligned_malloc_32(size_t size) // returned hunk MUST be freed with ll_aligned_free_32().
+{
+#if defined(LL_WINDOWS)
+ return _mm_malloc(size, 32);
+#elif defined(LL_DARWIN)
+# error implement me.
+#else
+ void *rtn;
+ if (LL_LIKELY(0 == posix_memalign(&rtn, 32, size)))
+ return rtn;
+ else // bad alignment requested, or out of memory
+ return NULL;
+#endif
+}
+
+inline void ll_aligned_free_32(void *p)
+{
+#if defined(LL_WINDOWS)
+ _mm_free(p);
+#elif defined(LL_DARWIN)
+# error implement me.
+#else
+ free(p); // posix_memalign() is compatible with heap deallocator
+#endif
+}
+
class LL_COMMON_API LLMemory
{
public:
diff --git a/indra/llmath/lloctree.h b/indra/llmath/lloctree.h
index f61ce6ce05..59828ae565 100644
--- a/indra/llmath/lloctree.h
+++ b/indra/llmath/lloctree.h
@@ -102,7 +102,7 @@ public:
: mParent((oct_node*)parent),
mOctant(octant)
{
- mD = (LLVector4a*) _mm_malloc(sizeof(LLVector4a)*4, 16);
+ mD = (LLVector4a*) ll_aligned_malloc_16(sizeof(LLVector4a)*4);
mD[CENTER] = center;
mD[SIZE] = size;
@@ -125,7 +125,7 @@ public:
delete getChild(i);
}
- _mm_free(mD);
+ ll_aligned_free_16(mD);
}
inline const BaseType* getParent() const { return mParent; }
@@ -640,7 +640,7 @@ public:
const LLVector4a& v = data->getPositionGroup();
LLVector4a val;
- val.setSub(v, mD[CENTER]);
+ val.setSub(v, BaseType::mD[BaseType::CENTER]);
val.setAbs(val);
S32 lt = val.lessThan4(MAX_MAG).getComparisonMask() & 0x7;
diff --git a/indra/llmath/llvolume.cpp b/indra/llmath/llvolume.cpp
index 8cb9475994..09ab47b890 100644
--- a/indra/llmath/llvolume.cpp
+++ b/indra/llmath/llvolume.cpp
@@ -31,6 +31,7 @@
*/
#include "linden_common.h"
+#include "llmemory.h"
#include "llmath.h"
#include <set>
@@ -1987,7 +1988,7 @@ BOOL LLVolume::generate()
void LLVolumeFace::VertexData::init()
{
- mData = (LLVector4a*) _mm_malloc(32, 16);
+ mData = (LLVector4a*) ll_aligned_malloc_16(32);
}
LLVolumeFace::VertexData::VertexData()
@@ -2004,7 +2005,7 @@ LLVolumeFace::VertexData::VertexData(const VertexData& rhs)
LLVolumeFace::VertexData::~VertexData()
{
- _mm_free(mData);
+ ll_aligned_free_16(mData);
}
LLVector4a& LLVolumeFace::VertexData::getPosition()
@@ -4406,7 +4407,7 @@ S32 LLVolume::lineSegmentIntersect(const LLVector4a& start, const LLVector4a& en
face.createOctree();
}
- LLVector4a* p = (LLVector4a*) face.mPositions;
+ //LLVector4a* p = (LLVector4a*) face.mPositions;
LLOctreeTriangleRayIntersect intersect(start, dir, &face, &closest_t, intersection, tex_coord, normal, bi_normal);
intersect.traverse(face.mOctree);
@@ -5196,7 +5197,7 @@ LLVolumeFace::LLVolumeFace() :
mWeights(NULL),
mOctree(NULL)
{
- mExtents = (LLVector4a*) _mm_malloc(48, 16);
+ mExtents = (LLVector4a*) ll_aligned_malloc_16(48);
mCenter = mExtents+2;
}
@@ -5217,7 +5218,7 @@ LLVolumeFace::LLVolumeFace(const LLVolumeFace& src)
mWeights(NULL),
mOctree(NULL)
{
- mExtents = (LLVector4a*) _mm_malloc(48, 16);
+ mExtents = (LLVector4a*) ll_aligned_malloc_16(48);
mCenter = mExtents+2;
*this = src;
}
@@ -5264,7 +5265,7 @@ LLVolumeFace& LLVolumeFace::operator=(const LLVolumeFace& src)
}
else
{
- _mm_free(mBinormals);
+ ll_aligned_free_16(mBinormals);
mBinormals = NULL;
}
@@ -5275,7 +5276,7 @@ LLVolumeFace& LLVolumeFace::operator=(const LLVolumeFace& src)
}
else
{
- _mm_free(mWeights);
+ ll_aligned_free_16(mWeights);
mWeights = NULL;
}
}
@@ -5295,7 +5296,7 @@ LLVolumeFace& LLVolumeFace::operator=(const LLVolumeFace& src)
LLVolumeFace::~LLVolumeFace()
{
- _mm_free(mExtents);
+ ll_aligned_free_16(mExtents);
mExtents = NULL;
freeData();
@@ -5303,17 +5304,17 @@ LLVolumeFace::~LLVolumeFace()
void LLVolumeFace::freeData()
{
- _mm_free(mPositions);
+ ll_aligned_free_16(mPositions);
mPositions = NULL;
- _mm_free(mNormals);
+ ll_aligned_free_16(mNormals);
mNormals = NULL;
- _mm_free(mTexCoords);
+ ll_aligned_free_16(mTexCoords);
mTexCoords = NULL;
- _mm_free(mIndices);
+ ll_aligned_free_16(mIndices);
mIndices = NULL;
- _mm_free(mBinormals);
+ ll_aligned_free_16(mBinormals);
mBinormals = NULL;
- _mm_free(mWeights);
+ ll_aligned_free_16(mWeights);
mWeights = NULL;
delete mOctree;
@@ -6084,21 +6085,21 @@ void LLVolumeFace::createBinormals()
void LLVolumeFace::resizeVertices(S32 num_verts)
{
- _mm_free(mPositions);
- _mm_free(mNormals);
- _mm_free(mBinormals);
- _mm_free(mTexCoords);
+ ll_aligned_free_16(mPositions);
+ ll_aligned_free_16(mNormals);
+ ll_aligned_free_16(mBinormals);
+ ll_aligned_free_16(mTexCoords);
mBinormals = NULL;
if (num_verts)
{
- mPositions = (LLVector4a*) _mm_malloc(num_verts*16, 16);
- mNormals = (LLVector4a*) _mm_malloc(num_verts*16, 16);
+ mPositions = (LLVector4a*) ll_aligned_malloc_16(num_verts*16);
+ mNormals = (LLVector4a*) ll_aligned_malloc_16(num_verts*16);
//pad texture coordinate block end to allow for QWORD reads
S32 size = ((num_verts*8) + 0xF) & ~0xF;
- mTexCoords = (LLVector2*) _mm_malloc(size, 16);
+ mTexCoords = (LLVector2*) ll_aligned_malloc_16(size);
}
else
{
@@ -6121,20 +6122,20 @@ void LLVolumeFace::pushVertex(const LLVector4a& pos, const LLVector4a& norm, con
S32 new_size = new_verts*16;
//positions
- LLVector4a* dst = (LLVector4a*) _mm_malloc(new_size, 16);
+ LLVector4a* dst = (LLVector4a*) ll_aligned_malloc_16(new_size);
if (mPositions)
{
LLVector4a::memcpyNonAliased16((F32*) dst, (F32*) mPositions, new_size/4);
- _mm_free(mPositions);
+ ll_aligned_free_16(mPositions);
}
mPositions = dst;
//normals
- dst = (LLVector4a*) _mm_malloc(new_size, 16);
+ dst = (LLVector4a*) ll_aligned_malloc_16(new_size);
if (mNormals)
{
LLVector4a::memcpyNonAliased16((F32*) dst, (F32*) mNormals, new_size/4);
- _mm_free(mNormals);
+ ll_aligned_free_16(mNormals);
}
mNormals = dst;
@@ -6142,16 +6143,16 @@ void LLVolumeFace::pushVertex(const LLVector4a& pos, const LLVector4a& norm, con
new_size = ((new_verts*8)+0xF) & ~0xF;
{
- LLVector2* dst = (LLVector2*) _mm_malloc(new_size, 16);
+ LLVector2* dst = (LLVector2*) ll_aligned_malloc_16(new_size);
if (mTexCoords)
{
LLVector4a::memcpyNonAliased16((F32*) dst, (F32*) mTexCoords, new_size/4);
- _mm_free(mTexCoords);
+ ll_aligned_free_16(mTexCoords);
}
}
//just clear binormals
- _mm_free(mBinormals);
+ ll_aligned_free_16(mBinormals);
mBinormals = NULL;
mPositions[mNumVertices] = pos;
@@ -6163,26 +6164,26 @@ void LLVolumeFace::pushVertex(const LLVector4a& pos, const LLVector4a& norm, con
void LLVolumeFace::allocateBinormals(S32 num_verts)
{
- _mm_free(mBinormals);
- mBinormals = (LLVector4a*) _mm_malloc(num_verts*16, 16);
+ ll_aligned_free_16(mBinormals);
+ mBinormals = (LLVector4a*) ll_aligned_malloc_16(num_verts*16);
}
void LLVolumeFace::allocateWeights(S32 num_verts)
{
- _mm_free(mWeights);
- mWeights = (LLVector4a*) _mm_malloc(num_verts*16, 16);
+ ll_aligned_free_16(mWeights);
+ mWeights = (LLVector4a*) ll_aligned_malloc_16(num_verts*16);
}
void LLVolumeFace::resizeIndices(S32 num_indices)
{
- _mm_free(mIndices);
+ ll_aligned_free_16(mIndices);
if (num_indices)
{
//pad index block end to allow for QWORD reads
S32 size = ((num_indices*2) + 0xF) & ~0xF;
- mIndices = (U16*) _mm_malloc(size,16);
+ mIndices = (U16*) ll_aligned_malloc_16(size);
}
else
{
@@ -6200,9 +6201,9 @@ void LLVolumeFace::pushIndex(const U16& idx)
S32 old_size = (mNumIndices+0xF) & ~0xF;
if (new_size != old_size)
{
- U16* dst = (U16*) _mm_malloc(new_size, 16);
+ U16* dst = (U16*) ll_aligned_malloc_16(new_size);
LLVector4a::memcpyNonAliased16((F32*) dst, (F32*) mIndices, new_size/4);
- _mm_free(mIndices);
+ ll_aligned_free_16(mIndices);
mIndices = dst;
}
@@ -6239,17 +6240,17 @@ void LLVolumeFace::appendFace(const LLVolumeFace& face, LLMatrix4& mat_in, LLMat
}
- LLVector4a* new_pos = (LLVector4a*) _mm_malloc(new_count*16, 16);
- LLVector4a* new_norm = (LLVector4a*) _mm_malloc(new_count*16, 16);
- LLVector2* new_tc = (LLVector2*) _mm_malloc((new_count*8+0xF) & ~0xF, 16);
+ LLVector4a* new_pos = (LLVector4a*) ll_aligned_malloc_16(new_count*16);
+ LLVector4a* new_norm = (LLVector4a*) ll_aligned_malloc_16(new_count*16);
+ LLVector2* new_tc = (LLVector2*) ll_aligned_malloc_16((new_count*8+0xF) & ~0xF);
LLVector4a::memcpyNonAliased16((F32*) new_pos, (F32*) mPositions, new_count*4);
LLVector4a::memcpyNonAliased16((F32*) new_norm, (F32*) mNormals, new_count*4);
LLVector4a::memcpyNonAliased16((F32*) new_tc, (F32*) mTexCoords, new_count*2);
- _mm_free(mPositions);
- _mm_free(mNormals);
- _mm_free(mTexCoords);
+ ll_aligned_free_16(mPositions);
+ ll_aligned_free_16(mNormals);
+ ll_aligned_free_16(mTexCoords);
mPositions = new_pos;
mNormals = new_norm;
@@ -6289,9 +6290,9 @@ void LLVolumeFace::appendFace(const LLVolumeFace& face, LLMatrix4& mat_in, LLMat
new_count = mNumIndices + face.mNumIndices;
- U16* new_indices = (U16*) _mm_malloc((new_count*2+0xF) & ~0xF, 16);
+ U16* new_indices = (U16*) ll_aligned_malloc_16((new_count*2+0xF) & ~0xF);
LLVector4a::memcpyNonAliased16((F32*) new_indices, (F32*) mIndices, new_count/2);
- _mm_free(mIndices);
+ ll_aligned_free_16(mIndices);
mIndices = new_indices;
mNumIndices = new_count;
diff --git a/indra/llrender/llrender.cpp b/indra/llrender/llrender.cpp
index 3f70ccacb1..70601663e6 100644
--- a/indra/llrender/llrender.cpp
+++ b/indra/llrender/llrender.cpp
@@ -777,8 +777,8 @@ LLRender::LLRender()
mTexcoordsp = tc.get();
mColorsp = color.get();
- mUIOffset = (LLVector4a*) _mm_malloc(LL_MAX_UI_STACK_DEPTH*sizeof(LLVector4a), 16);
- mUIScale = (LLVector4a*) _mm_malloc(LL_MAX_UI_STACK_DEPTH*sizeof(LLVector4a), 16);
+ mUIOffset = (LLVector4a*) ll_aligned_malloc_16(LL_MAX_UI_STACK_DEPTH*sizeof(LLVector4a));
+ mUIScale = (LLVector4a*) ll_aligned_malloc_16(LL_MAX_UI_STACK_DEPTH*sizeof(LLVector4a));
mTexUnits.reserve(LL_NUM_TEXTURE_LAYERS);
for (U32 i = 0; i < LL_NUM_TEXTURE_LAYERS; i++)
@@ -816,9 +816,9 @@ void LLRender::shutdown()
delete mDummyTexUnit;
mDummyTexUnit = NULL;
- _mm_free(mUIOffset);
+ ll_aligned_free_16(mUIOffset);
mUIOffset = NULL;
- _mm_free(mUIScale);
+ ll_aligned_free_16(mUIScale);
mUIScale = NULL;
}
diff --git a/indra/llrender/llrender.h b/indra/llrender/llrender.h
index 2bacf16dc6..11cd95646f 100644
--- a/indra/llrender/llrender.h
+++ b/indra/llrender/llrender.h
@@ -45,6 +45,7 @@
#include "v4coloru.h"
#include "llstrider.h"
#include "llpointer.h"
+#include "llmemory.h"
#include "llglheaders.h"
class LLVertexBuffer;
diff --git a/indra/llrender/llvertexbuffer.cpp b/indra/llrender/llvertexbuffer.cpp
index 4f2dae0cdf..48c20b09a8 100644
--- a/indra/llrender/llvertexbuffer.cpp
+++ b/indra/llrender/llvertexbuffer.cpp
@@ -31,6 +31,7 @@
*/
#include "linden_common.h"
+#include "llmemory.h"
#include <boost/static_assert.hpp>
@@ -627,7 +628,7 @@ void LLVertexBuffer::createGLBuffer()
{
static int gl_buffer_idx = 0;
mGLBuffer = ++gl_buffer_idx;
- mMappedData = (U8*) _mm_malloc(size, 16);
+ mMappedData = (U8*) ll_aligned_malloc_16(size);
memset(mMappedData, 0, size);
}
}
@@ -662,7 +663,7 @@ void LLVertexBuffer::createGLIndices()
}
else
{
- mMappedIndexData = (U8*) _mm_malloc(size, 16);
+ mMappedIndexData = (U8*) ll_aligned_malloc_16(size);
static int gl_buffer_idx = 0;
mGLIndices = ++gl_buffer_idx;
}
@@ -683,7 +684,7 @@ void LLVertexBuffer::destroyGLBuffer()
}
else
{
- _mm_free(mMappedData);
+ ll_aligned_free_16(mMappedData);
mMappedData = NULL;
mEmpty = TRUE;
}
@@ -710,7 +711,7 @@ void LLVertexBuffer::destroyGLIndices()
}
else
{
- _mm_free(mMappedIndexData);
+ ll_aligned_free_16(mMappedIndexData);
mMappedIndexData = NULL;
mEmpty = TRUE;
}
@@ -846,8 +847,8 @@ void LLVertexBuffer::resizeBuffer(S32 newnverts, S32 newnindices)
{
if (!useVBOs())
{
- _mm_free(mMappedData);
- mMappedData = (U8*) _mm_malloc(newsize, 16);
+ ll_aligned_free_16(mMappedData);
+ mMappedData = (U8*) ll_aligned_malloc_16(newsize);
}
mResized = TRUE;
}
@@ -867,8 +868,8 @@ void LLVertexBuffer::resizeBuffer(S32 newnverts, S32 newnindices)
{
if (!useVBOs())
{
- _mm_free(mMappedIndexData);
- mMappedIndexData = (U8*) _mm_malloc(new_index_size, 16);
+ ll_aligned_free_16(mMappedIndexData);
+ mMappedIndexData = (U8*) ll_aligned_malloc_16(new_index_size);
}
mResized = TRUE;
}
diff --git a/indra/newview/lldrawable.cpp b/indra/newview/lldrawable.cpp
index 04e433dcfd..ca408a309e 100644
--- a/indra/newview/lldrawable.cpp
+++ b/indra/newview/lldrawable.cpp
@@ -95,7 +95,7 @@ void LLDrawable::incrementVisible()
void LLDrawable::init()
{
- mExtents = (LLVector4a*) _mm_malloc(sizeof(LLVector4a)*3, 32);
+ mExtents = (LLVector4a*) ll_aligned_malloc_32(sizeof(LLVector4a)*3);
mPositionGroup = mExtents + 2;
// mXform
@@ -150,7 +150,7 @@ void LLDrawable::destroy()
llinfos << "- Zombie drawables: " << sNumZombieDrawables << llendl;
}*/
- _mm_free(mExtents);
+ ll_aligned_free_32(mExtents);
mExtents = mPositionGroup = NULL;
}
diff --git a/indra/newview/llface.cpp b/indra/newview/llface.cpp
index 0b5cf78261..dea160ae64 100644
--- a/indra/newview/llface.cpp
+++ b/indra/newview/llface.cpp
@@ -152,7 +152,7 @@ void cylindricalProjection(LLVector2 &tc, const LLVector4a& normal, const LLVect
void LLFace::init(LLDrawable* drawablep, LLViewerObject* objp)
{
- mExtents = (LLVector4a*) _mm_malloc(sizeof(LLVector4a)*2, 16);
+ mExtents = (LLVector4a*) ll_aligned_malloc_16(sizeof(LLVector4a)*2);
mLastUpdateTime = gFrameTimeSeconds;
mLastMoveTime = 0.f;
@@ -269,7 +269,7 @@ void LLFace::destroy()
mVObjp = NULL;
}
- _mm_free(mExtents);
+ ll_aligned_free_16(mExtents);
mExtents = NULL;
}
@@ -1328,8 +1328,8 @@ BOOL LLFace::getGeometryVolume(const LLVolume& volume,
for (S32 i = 0; i < num_vertices; i++)
{
LLVector2 tc(vf.mTexCoords[i]);
- LLVector4a& norm = vf.mNormals[i];
- LLVector4a& center = *(vf.mCenter);
+ //LLVector4a& norm = vf.mNormals[i];
+ //LLVector4a& center = *(vf.mCenter);
LLVector3 tmp(tc.mV[0], tc.mV[1], 0.f);
tmp = tmp * *mTextureMatrix;
diff --git a/indra/newview/llspatialpartition.cpp b/indra/newview/llspatialpartition.cpp
index 77d36b1c2e..d0e3a1428f 100644
--- a/indra/newview/llspatialpartition.cpp
+++ b/indra/newview/llspatialpartition.cpp
@@ -239,7 +239,7 @@ void LLSpatialGroup::buildOcclusion()
{
if (!mOcclusionVerts)
{
- mOcclusionVerts = (LLVector4a*) _mm_malloc(sizeof(LLVector4a)*8, 16);
+ mOcclusionVerts = (LLVector4a*) ll_aligned_malloc_16(sizeof(LLVector4a)*8);
}
LLVector4a fudge;
@@ -342,13 +342,13 @@ LLSpatialGroup::~LLSpatialGroup()
sQueryPool.release(mOcclusionQuery[LLViewerCamera::sCurCameraID]);
}
- _mm_free(mOcclusionVerts);
+ ll_aligned_free_16(mOcclusionVerts);
LLMemType mt(LLMemType::MTYPE_SPACE_PARTITION);
clearDrawMap();
clearAtlasList() ;
- _mm_free(mBounds);
+ ll_aligned_free_16(mBounds);
}
BOOL LLSpatialGroup::hasAtlas(LLTextureAtlas* atlasp)
@@ -1163,7 +1163,7 @@ LLSpatialGroup::LLSpatialGroup(OctreeNode* node, LLSpatialPartition* part) :
sNodeCount++;
LLMemType mt(LLMemType::MTYPE_SPACE_PARTITION);
- mBounds = (LLVector4a*) _mm_malloc(sizeof(LLVector4a) * V4_COUNT, 16);
+ mBounds = (LLVector4a*) ll_aligned_malloc_16(sizeof(LLVector4a) * V4_COUNT);
mExtents = mBounds + EXTENTS;
mObjectBounds = mBounds + OBJECT_BOUNDS;
mObjectExtents = mBounds + OBJECT_EXTENTS;
@@ -1434,7 +1434,7 @@ void LLSpatialGroup::destroyGL()
}
}
- _mm_free(mOcclusionVerts);
+ ll_aligned_free_16(mOcclusionVerts);
mOcclusionVerts = NULL;
for (LLSpatialGroup::element_iter i = getData().begin(); i != getData().end(); ++i)
@@ -3557,7 +3557,7 @@ LLDrawInfo::LLDrawInfo(U16 start, U16 end, U32 count, U32 offset,
mDrawMode(LLRender::TRIANGLES)
{
mVertexBuffer->validateRange(mStart, mEnd, mCount, mOffset);
- mExtents = (LLVector4a*) _mm_malloc(sizeof(LLVector4a)*2, 16);
+ mExtents = (LLVector4a*) ll_aligned_malloc_16(sizeof(LLVector4a)*2);
mDebugColor = (rand() << 16) + rand();
}
@@ -3579,7 +3579,7 @@ LLDrawInfo::~LLDrawInfo()
gPipeline.checkReferences(this);
}
- _mm_free(mExtents);
+ ll_aligned_free_16(mExtents);
}
void LLDrawInfo::validate()
generated by cgit v1.2.3 (git 2.25.1) at 2025-11-04 09:06:56 +0000