From 97d969a338c1e4f973eb817ba7701aff51a02ccb Mon Sep 17 00:00:00 2001 From: Oz Linden Date: Wed, 12 Sep 2012 14:36:37 -0400 Subject: initial attempt to restore changes that make removing tcmalloc possible; not tested --- indra/llcommon/llallocator.cpp | 2 +- indra/llcommon/llmemory.cpp | 12 +++++++++++ indra/llcommon/llmemory.h | 45 +++++++++++++++++++++++++++++------------- 3 files changed, 44 insertions(+), 15 deletions(-) (limited to 'indra/llcommon') diff --git a/indra/llcommon/llallocator.cpp b/indra/llcommon/llallocator.cpp index 6f6abefc67..87654b5b97 100644 --- a/indra/llcommon/llallocator.cpp +++ b/indra/llcommon/llallocator.cpp @@ -27,7 +27,7 @@ #include "linden_common.h" #include "llallocator.h" -#if LL_USE_TCMALLOC +#if (LL_USE_TCMALLOC && LL_USE_HEAP_PROFILER) #include "google/heap-profiler.h" #include "google/commandlineflags_public.h" diff --git a/indra/llcommon/llmemory.cpp b/indra/llcommon/llmemory.cpp index 3b9758f996..afaf366668 100644 --- a/indra/llcommon/llmemory.cpp +++ b/indra/llcommon/llmemory.cpp @@ -61,6 +61,18 @@ BOOL LLMemory::sEnableMemoryFailurePrevention = FALSE; LLPrivateMemoryPoolManager::mem_allocation_info_t LLPrivateMemoryPoolManager::sMemAllocationTracker; #endif +void ll_assert_aligned_func(uintptr_t ptr,U32 alignment) +{ +#ifdef SHOW_ASSERT + // Redundant, place to set breakpoints. + if (ptr%alignment!=0) + { + llwarns << "alignment check failed" << llendl; + } + llassert(ptr%alignment==0); +#endif +} + //static void LLMemory::initClass() { diff --git a/indra/llcommon/llmemory.h b/indra/llcommon/llmemory.h index bbbdaa6497..9dd776ff57 100644 --- a/indra/llcommon/llmemory.h +++ b/indra/llcommon/llmemory.h @@ -27,7 +27,6 @@ #define LLMEMORY_H #include "llmemtype.h" -#if LL_DEBUG inline void* ll_aligned_malloc( size_t size, int align ) { void* mem = malloc( size + (align - 1) + sizeof(void*) ); @@ -43,10 +42,11 @@ inline void ll_aligned_free( void* ptr ) free( ((void**)ptr)[-1] ); } +#if !LL_USE_TCMALLOC 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); + return _aligned_malloc(size, 16); #elif defined(LL_DARWIN) return malloc(size); // default osx malloc is 16 byte aligned. #else @@ -58,21 +58,38 @@ inline void* ll_aligned_malloc_16(size_t size) // returned hunk MUST be freed wi #endif } +inline void* ll_aligned_realloc_16(void* ptr, size_t size) // returned hunk MUST be freed with ll_aligned_free_16(). +{ +#if defined(LL_WINDOWS) + return _aligned_realloc(ptr, size, 16); +#elif defined(LL_DARWIN) + return realloc(ptr,size); // default osx malloc is 16 byte aligned. +#else + return realloc(ptr,size); // FIXME not guaranteed to be aligned. +#endif +} + inline void ll_aligned_free_16(void *p) { #if defined(LL_WINDOWS) - _mm_free(p); + _aligned_free(p); #elif defined(LL_DARWIN) return free(p); #else free(p); // posix_memalign() is compatible with heap deallocator #endif } +#else // USE_TCMALLOC +// ll_aligned_foo_16 are not needed with tcmalloc +#define ll_aligned_malloc_16 malloc +#define ll_aligned_realloc_16 realloc +#define ll_aligned_free_16 free +#endif // USE_TCMALLOC 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); + return _aligned_malloc(size, 32); #elif defined(LL_DARWIN) return ll_aligned_malloc( size, 32 ); #else @@ -87,22 +104,13 @@ inline void* ll_aligned_malloc_32(size_t size) // returned hunk MUST be freed wi inline void ll_aligned_free_32(void *p) { #if defined(LL_WINDOWS) - _mm_free(p); + _aligned_free(p); #elif defined(LL_DARWIN) ll_aligned_free( p ); #else free(p); // posix_memalign() is compatible with heap deallocator #endif } -#else // LL_DEBUG -// ll_aligned_foo are noops now that we use tcmalloc everywhere (tcmalloc aligns automatically at appropriate intervals) -#define ll_aligned_malloc( size, align ) malloc(size) -#define ll_aligned_free( ptr ) free(ptr) -#define ll_aligned_malloc_16 malloc -#define ll_aligned_free_16 free -#define ll_aligned_malloc_32 malloc -#define ll_aligned_free_32 free -#endif // LL_DEBUG #ifndef __DEBUG_PRIVATE_MEM__ #define __DEBUG_PRIVATE_MEM__ 0 @@ -512,4 +520,13 @@ void LLPrivateMemoryPoolTester::operator delete[](void* addr) // LLSingleton moved to llsingleton.h +LL_COMMON_API void ll_assert_aligned_func(uintptr_t ptr,U32 alignment); + +#ifdef SHOW_ASSERT +#define ll_assert_aligned(ptr,alignment) ll_assert_aligned_func(reinterpret_cast(ptr),((U32)alignment)) +#else +#define ll_assert_aligned(ptr,alignment) +#endif + + #endif -- cgit v1.3 From 9ec263d2f1f51b334949281393ef2446796f1404 Mon Sep 17 00:00:00 2001 From: Dave Parks Date: Wed, 12 Sep 2012 15:44:36 -0500 Subject: MAINT-1503 Minimal set of changes needed to safely disable tcmalloc --- indra/cmake/GooglePerfTools.cmake | 2 +- indra/llcommon/llmemory.h | 24 ++++++++++++++++-------- indra/llmath/llvolume.cpp | 19 ++++++++++--------- indra/llmath/tests/alignment_test.cpp | 2 +- 4 files changed, 28 insertions(+), 19 deletions(-) (limited to 'indra/llcommon') diff --git a/indra/cmake/GooglePerfTools.cmake b/indra/cmake/GooglePerfTools.cmake index 09501e0406..73b3642ae6 100644 --- a/indra/cmake/GooglePerfTools.cmake +++ b/indra/cmake/GooglePerfTools.cmake @@ -3,7 +3,7 @@ include(Prebuilt) # If you want to enable or disable TCMALLOC in viewer builds, this is the place. # set ON or OFF as desired. -set (USE_TCMALLOC ON) +set (USE_TCMALLOC OFF) if (STANDALONE) include(FindGooglePerfTools) diff --git a/indra/llcommon/llmemory.h b/indra/llcommon/llmemory.h index 9dd776ff57..f37e8459ff 100644 --- a/indra/llcommon/llmemory.h +++ b/indra/llcommon/llmemory.h @@ -58,27 +58,35 @@ inline void* ll_aligned_malloc_16(size_t size) // returned hunk MUST be freed wi #endif } -inline void* ll_aligned_realloc_16(void* ptr, size_t size) // returned hunk MUST be freed with ll_aligned_free_16(). +inline void ll_aligned_free_16(void *p) { #if defined(LL_WINDOWS) - return _aligned_realloc(ptr, size, 16); + _aligned_free(p); #elif defined(LL_DARWIN) - return realloc(ptr,size); // default osx malloc is 16 byte aligned. + return free(p); #else - return realloc(ptr,size); // FIXME not guaranteed to be aligned. + free(p); // posix_memalign() is compatible with heap deallocator #endif } -inline void ll_aligned_free_16(void *p) +inline void* ll_aligned_realloc_16(void* ptr, size_t size, size_t old_size) // returned hunk MUST be freed with ll_aligned_free_16(). { #if defined(LL_WINDOWS) - _aligned_free(p); + return _aligned_realloc(ptr, size, 16); #elif defined(LL_DARWIN) - return free(p); + return realloc(ptr,size); // default osx malloc is 16 byte aligned. #else - free(p); // posix_memalign() is compatible with heap deallocator + //FIXME: memcpy is SLOW + void* ret = ll_aligned_malloc_16(size); + if (ptr) + { + memcpy(ret, ptr, old_size); + ll_aligned_free_16(ptr); + } + return ret; #endif } + #else // USE_TCMALLOC // ll_aligned_foo_16 are not needed with tcmalloc #define ll_aligned_malloc_16 malloc diff --git a/indra/llmath/llvolume.cpp b/indra/llmath/llvolume.cpp index 11fa7080ce..ea09a3afe7 100644 --- a/indra/llmath/llvolume.cpp +++ b/indra/llmath/llvolume.cpp @@ -6979,19 +6979,20 @@ void LLVolumeFace::pushVertex(const LLVector4a& pos, const LLVector4a& norm, con { S32 new_verts = mNumVertices+1; S32 new_size = new_verts*16; -// S32 old_size = mNumVertices*16; + S32 old_size = mNumVertices*16; //positions - mPositions = (LLVector4a*) ll_aligned_realloc_16(mPositions, new_size); + mPositions = (LLVector4a*) ll_aligned_realloc_16(mPositions, new_size, old_size); ll_assert_aligned(mPositions,16); //normals - mNormals = (LLVector4a*) ll_aligned_realloc_16(mNormals, new_size); + mNormals = (LLVector4a*) ll_aligned_realloc_16(mNormals, new_size, old_size); ll_assert_aligned(mNormals,16); //tex coords new_size = ((new_verts*8)+0xF) & ~0xF; - mTexCoords = (LLVector2*) ll_aligned_realloc_16(mTexCoords, new_size); + old_size = ((mNumVertices*8)+0xF) & ~0xF; + mTexCoords = (LLVector2*) ll_aligned_realloc_16(mTexCoords, new_size, old_size); ll_assert_aligned(mTexCoords,16); @@ -7045,7 +7046,7 @@ void LLVolumeFace::pushIndex(const U16& idx) S32 old_size = ((mNumIndices*2)+0xF) & ~0xF; if (new_size != old_size) { - mIndices = (U16*) ll_aligned_realloc_16(mIndices, new_size); + mIndices = (U16*) ll_aligned_realloc_16(mIndices, new_size, old_size); ll_assert_aligned(mIndices,16); } @@ -7087,11 +7088,11 @@ void LLVolumeFace::appendFace(const LLVolumeFace& face, LLMatrix4& mat_in, LLMat } //allocate new buffer space - mPositions = (LLVector4a*) ll_aligned_realloc_16(mPositions, new_count*sizeof(LLVector4a)); + mPositions = (LLVector4a*) ll_aligned_realloc_16(mPositions, new_count*sizeof(LLVector4a), mNumVertices*sizeof(LLVector4a)); ll_assert_aligned(mPositions, 16); - mNormals = (LLVector4a*) ll_aligned_realloc_16(mNormals, new_count*sizeof(LLVector4a)); + mNormals = (LLVector4a*) ll_aligned_realloc_16(mNormals, new_count*sizeof(LLVector4a), mNumVertices*sizeof(LLVector4a)); ll_assert_aligned(mNormals, 16); - mTexCoords = (LLVector2*) ll_aligned_realloc_16(mTexCoords, (new_count*sizeof(LLVector2)+0xF) & ~0xF); + mTexCoords = (LLVector2*) ll_aligned_realloc_16(mTexCoords, (new_count*sizeof(LLVector2)+0xF) & ~0xF, (mNumVertices*sizeof(LLVector2)+0xF) & ~0xF); ll_assert_aligned(mTexCoords, 16); mNumVertices = new_count; @@ -7138,7 +7139,7 @@ void LLVolumeFace::appendFace(const LLVolumeFace& face, LLMatrix4& mat_in, LLMat new_count = mNumIndices + face.mNumIndices; //allocate new index buffer - mIndices = (U16*) ll_aligned_realloc_16(mIndices, (new_count*sizeof(U16)+0xF) & ~0xF); + mIndices = (U16*) ll_aligned_realloc_16(mIndices, (new_count*sizeof(U16)+0xF) & ~0xF, (mNumIndices*sizeof(U16)+0xF) & ~0xF); //get destination address into new index buffer U16* dst_idx = mIndices+mNumIndices; diff --git a/indra/llmath/tests/alignment_test.cpp b/indra/llmath/tests/alignment_test.cpp index ac0c45ae6f..5c426c57c3 100644 --- a/indra/llmath/tests/alignment_test.cpp +++ b/indra/llmath/tests/alignment_test.cpp @@ -78,7 +78,7 @@ void alignment_test_object_t::test<1>() align_ptr = ll_aligned_malloc_16(sizeof(MyVector4a)); ensure("ll_aligned_malloc_16 failed", is_aligned(align_ptr,16)); - align_ptr = ll_aligned_realloc_16(align_ptr,2*sizeof(MyVector4a)); + align_ptr = ll_aligned_realloc_16(align_ptr,2*sizeof(MyVector4a), sizeof(MyVector4a)); ensure("ll_aligned_realloc_16 failed", is_aligned(align_ptr,16)); ll_aligned_free_16(align_ptr); -- cgit v1.3