From d022964a25572218c81364d24dfb2aa5a2dedfab Mon Sep 17 00:00:00 2001
From: Rye Mutt <rye@alchemyviewer.org>
Date: Mon, 24 Aug 2020 01:12:52 -0400
Subject: [PATCH] Use _mm_setzero and _mm_set1 instead of static global
variables for the SIMD epsilon and zero vectors as this generates better code
under modern compilers
---
indra/llappearance/llpolymorph.cpp | 7 ++++---
indra/llmath/llsimdtypes.h | 5 ++---
indra/llmath/llvector4a.cpp | 15 ++-------------
indra/llmath/llvector4a.h | 10 ++++------
4 files changed, 12 insertions(+), 25 deletions(-)
diff --git a/indra/llappearance/llpolymorph.cpp b/indra/llappearance/llpolymorph.cpp
index 3f0423fc498..072efeeee34 100644
--- a/indra/llappearance/llpolymorph.cpp
+++ b/indra/llappearance/llpolymorph.cpp
@@ -445,7 +445,7 @@ LLVector4a LLPolyMorphTarget::getVertexDistortion(S32 requested_index, LLPolyMes
//-----------------------------------------------------------------------------
const LLVector4a *LLPolyMorphTarget::getFirstDistortion(U32 *index, LLPolyMesh **poly_mesh)
{
- if (!mMorphData) return &LLVector4a::getZero();
+ if (!mMorphData) return nullptr;
LLVector4a* resultVec;
mMorphData->mCurrentIndex = 0;
@@ -471,7 +471,7 @@ const LLVector4a *LLPolyMorphTarget::getFirstDistortion(U32 *index, LLPolyMesh *
//-----------------------------------------------------------------------------
const LLVector4a *LLPolyMorphTarget::getNextDistortion(U32 *index, LLPolyMesh **poly_mesh)
{
- if (!mMorphData) return &LLVector4a::getZero();
+ if (!mMorphData) return nullptr;
LLVector4a* resultVec;
mMorphData->mCurrentIndex++;
@@ -517,7 +517,8 @@ const LLVector4a& LLPolyMorphTarget::getAvgDistortion()
}
else
{
- return LLVector4a::getZero();
+ static LLVector4a zero = LLVector4a::getZero();
+ return zero;
}
}
diff --git a/indra/llmath/llsimdtypes.h b/indra/llmath/llsimdtypes.h
index 5290630ce58..805da7ce121 100644
--- a/indra/llmath/llsimdtypes.h
+++ b/indra/llmath/llsimdtypes.h
@@ -64,10 +64,9 @@ class LLSimdScalar
mQ = _mm_set_ss(f);
}
- static inline const LLSimdScalar& getZero()
+ static inline LLSimdScalar getZero()
{
- extern const LLQuad F_ZERO_4A;
- return reinterpret_cast<const LLSimdScalar&>(F_ZERO_4A);
+ return _mm_setzero_ps();
}
inline F32 getF32() const;
diff --git a/indra/llmath/llvector4a.cpp b/indra/llmath/llvector4a.cpp
index 570fa41a439..25df427633d 100644
--- a/indra/llmath/llvector4a.cpp
+++ b/indra/llmath/llvector4a.cpp
@@ -28,17 +28,6 @@
#include "llmath.h"
#include "llquantize.h"
-extern const LLQuad F_ZERO_4A = { 0, 0, 0, 0 };
-extern const LLQuad F_APPROXIMATELY_ZERO_4A = {
- F_APPROXIMATELY_ZERO,
- F_APPROXIMATELY_ZERO,
- F_APPROXIMATELY_ZERO,
- F_APPROXIMATELY_ZERO
-};
-
-extern const LLVector4a LL_V4A_ZERO = reinterpret_cast<const LLVector4a&> ( F_ZERO_4A );
-extern const LLVector4a LL_V4A_EPSILON = reinterpret_cast<const LLVector4a&> ( F_APPROXIMATELY_ZERO_4A );
-
/*static */void LLVector4a::memcpyNonAliased16(F32* __restrict dst, const F32* __restrict src, size_t bytes)
{
ll_memcpy_nonaliased_aligned_16((char*)dst, (char*)src, bytes);
@@ -128,7 +117,7 @@ void LLVector4a::quantize8( const LLVector4a& low, const LLVector4a& high )
{
LLVector4a maxError; maxError.setMul(delta, *reinterpret_cast<const LLVector4a*>(F_OOU8MAX_4A));
LLVector4a absVal; absVal.setAbs( val );
- setSelectWithMask( absVal.lessThan( maxError ), F_ZERO_4A, val );
+ setSelectWithMask( absVal.lessThan( maxError ), getZero(), val );
}
}
@@ -175,6 +164,6 @@ void LLVector4a::quantize16( const LLVector4a& low, const LLVector4a& high )
{
LLVector4a maxError; maxError.setMul(delta, *reinterpret_cast<const LLVector4a*>(F_OOU16MAX_4A));
LLVector4a absVal; absVal.setAbs( val );
- setSelectWithMask( absVal.lessThan( maxError ), F_ZERO_4A, val );
+ setSelectWithMask( absVal.lessThan( maxError ), getZero(), val );
}
}
diff --git a/indra/llmath/llvector4a.h b/indra/llmath/llvector4a.h
index 043a59a750a..19f29d59b3a 100644
--- a/indra/llmath/llvector4a.h
+++ b/indra/llmath/llvector4a.h
@@ -70,17 +70,15 @@ class LLVector4a
}
// Return a vector of all zeros
- static inline const LLVector4a& getZero()
+ static inline LLVector4a getZero()
{
- extern const LLVector4a LL_V4A_ZERO;
- return LL_V4A_ZERO;
+ return _mm_setzero_ps();
}
// Return a vector of all epsilon, where epsilon is a small float suitable for approximate equality checks
- static inline const LLVector4a& getEpsilon()
+ static inline LLVector4a getEpsilon()
{
- extern const LLVector4a LL_V4A_EPSILON;
- return LL_V4A_EPSILON;
+ return _mm_set1_ps(F_APPROXIMATELY_ZERO);
}
// Copy 16 bytes from src to dst. Source and destination must be 16-byte aligned
--
GitLab