Use bit_cast to create -0.0 for floating point types to avoid compiler optimization changing sign with --ffast-math enabled.

6 files changed, 16 insertions, 12 deletions

diff --git a/Eigen/src/Core/arch/AltiVec/PacketMath.h b/Eigen/src/Core/arch/AltiVec/PacketMath.h
index 7c70c07b1..d4aee3e21 100755
--- a/Eigen/src/Core/arch/AltiVec/PacketMath.h
+++ b/Eigen/src/Core/arch/AltiVec/PacketMath.h
@@ -2260,7 +2260,8 @@ static Packet2ul p2ul_SIGN = { 0x8000000000000000ull, 0x8000000000000000ull };
 static Packet2ul p2ul_PREV0DOT5 = { 0x3FDFFFFFFFFFFFFFull, 0x3FDFFFFFFFFFFFFFull };
 static Packet2d  p2d_ONE  = { 1.0, 1.0 };
 static Packet2d  p2d_ZERO = reinterpret_cast<Packet2d>(p4f_ZERO);
-static Packet2d  p2d_MZERO = { -0.0, -0.0 };
+static Packet2d  p2d_MZERO = { numext::bit_cast<double>(0x8000000000000000ull),
+                               numext::bit_cast<double>(0x8000000000000000ull) };
 
 #ifdef _BIG_ENDIAN
 static Packet2d p2d_COUNTDOWN = reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4f>(p2d_ZERO), reinterpret_cast<Packet4f>(p2d_ONE), 8));
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h
index 87e8c2703..8f1c1a874 100644
--- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h
+++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h
@@ -839,7 +839,8 @@ Packet psqrt_complex(const Packet& a) {
 
   // Step 4. Compute solution for inputs with negative real part:
   //         [|eta0|, sign(y0)*rho0, |eta1|, sign(y1)*rho1]
-  const RealPacket cst_imag_sign_mask = pset1<Packet>(Scalar(RealScalar(0.0), RealScalar(-0.0))).v;
+  const RealScalar neg_zero = RealScalar(numext::bit_cast<float>(0x80000000u));
+  const RealPacket cst_imag_sign_mask = pset1<Packet>(Scalar(RealScalar(0.0), neg_zero)).v;
   RealPacket imag_signs = pand(a.v, cst_imag_sign_mask);
   Packet negative_real_result;
   // Notice that rho is positive, so taking it's absolute value is a noop.
diff --git a/Eigen/src/Core/arch/NEON/PacketMath.h b/Eigen/src/Core/arch/NEON/PacketMath.h
index 73a35c570..e1efe9bcb 100644
--- a/Eigen/src/Core/arch/NEON/PacketMath.h
+++ b/Eigen/src/Core/arch/NEON/PacketMath.h
@@ -866,12 +866,12 @@ template<> EIGEN_STRONG_INLINE Packet2ul psub<Packet2ul>(const Packet2ul& a, con
 
 template<> EIGEN_STRONG_INLINE Packet2f pxor<Packet2f>(const Packet2f& a, const Packet2f& b);
 template<> EIGEN_STRONG_INLINE Packet2f paddsub<Packet2f>(const Packet2f& a, const Packet2f & b) {
-  Packet2f mask = {-0.0f, 0.0f};
+  Packet2f mask = {numext::bit_cast<float>(0x80000000u), 0.0f};
   return padd(a, pxor(mask, b));
 }
 template<> EIGEN_STRONG_INLINE Packet4f pxor<Packet4f>(const Packet4f& a, const Packet4f& b);
 template<> EIGEN_STRONG_INLINE Packet4f paddsub<Packet4f>(const Packet4f& a, const Packet4f& b) {
-  Packet4f mask = {-0.0f, 0.0f, -0.0f, 0.0f};
+  Packet4f mask = {numext::bit_cast<float>(0x80000000u), 0.0f, numext::bit_cast<float>(0x80000000u), 0.0f};
   return padd(a, pxor(mask, b));
 }
 
@@ -3751,7 +3751,7 @@ template<> EIGEN_STRONG_INLINE Packet2d psub<Packet2d>(const Packet2d& a, const
 
 template<> EIGEN_STRONG_INLINE Packet2d pxor<Packet2d>(const Packet2d& , const Packet2d& );
 template<> EIGEN_STRONG_INLINE Packet2d paddsub<Packet2d>(const Packet2d& a, const Packet2d& b){
-  const Packet2d mask = {-0.0,0.0};
+  const Packet2d mask = {numext::bit_cast<double>(0x8000000000000000ull),0.0};
   return padd(a, pxor(mask, b));
 }
 
diff --git a/Eigen/src/Core/arch/ZVector/PacketMath.h b/Eigen/src/Core/arch/ZVector/PacketMath.h
index b10c1f6c7..2246439cc 100755
--- a/Eigen/src/Core/arch/ZVector/PacketMath.h
+++ b/Eigen/src/Core/arch/ZVector/PacketMath.h
@@ -94,8 +94,9 @@ static _EIGEN_DECLARE_CONST_FAST_Packet2d(ZERO, 0);
 static _EIGEN_DECLARE_CONST_FAST_Packet2l(ZERO, 0);
 static _EIGEN_DECLARE_CONST_FAST_Packet2l(ONE, 1);
 
-static Packet2d p2d_ONE = { 1.0, 1.0 }; 
-static Packet2d p2d_ZERO_ = { -0.0, -0.0 };
+static Packet2d p2d_ONE = { 1.0, 1.0 };
+static Packet2d p2d_ZERO_ = { numext::bit_cast<double>0x8000000000000000ull),
+                              numext::bit_cast<double>0x8000000000000000ull) };
 
 #if !defined(__ARCH__) || (defined(__ARCH__) && __ARCH__ >= 12)
 #define _EIGEN_DECLARE_CONST_FAST_Packet4f(NAME,X) \
diff --git a/Eigen/src/Geometry/arch/Geometry_SIMD.h b/Eigen/src/Geometry/arch/Geometry_SIMD.h
index 9c15bfb98..89ac92062 100644
--- a/Eigen/src/Geometry/arch/Geometry_SIMD.h
+++ b/Eigen/src/Geometry/arch/Geometry_SIMD.h
@@ -146,8 +146,9 @@ struct quat_conj<Architecture::Target, Derived, double>
   {
     evaluator<typename Derived::Coefficients> qe(q.coeffs());
     Quaternion<double> res;
-    double arr1[2] = {-0.0, -0.0};
-    double arr2[2] = {-0.0,  0.0};
+    const double neg_zero = numext::bit_cast<double>(0x8000000000000000ull);
+    double arr1[2] = {neg_zero, neg_zero};
+    double arr2[2] = {neg_zero,  0.0};
     const Packet2d mask0 = pset<Packet2d>(arr1);
     const Packet2d mask2 = pset<Packet2d>(arr2);
     pstoret<double,Packet2d,ResAlignment>(&res.x(), pxor(mask0, qe.template packet<traits<Derived>::Alignment,Packet2d>(0)));
diff --git a/Eigen/src/LU/arch/InverseSize4.h b/Eigen/src/LU/arch/InverseSize4.h
index ee5548aed..106224bbc 100644
--- a/Eigen/src/LU/arch/InverseSize4.h
+++ b/Eigen/src/LU/arch/InverseSize4.h
@@ -143,7 +143,7 @@ struct compute_inverse_size4<Architecture::Target, float, MatrixType, ResultType
     iC = psub(iC, pmul(vec4f_swizzle2(A, A, 1, 0, 3, 2), vec4f_swizzle2(DC, DC, 2, 1, 2, 1)));
     iC = psub(pmul(B, vec4f_duplane(dC, 0)), iC);
 
-    const float sign_mask[4] = {0.0f, -0.0f, -0.0f, 0.0f};
+    const float sign_mask[4] = {0.0f, numext::bit_cast<float>(0x80000000u), numext::bit_cast<float>(0x80000000u), 0.0f};
     const Packet4f p4f_sign_PNNP = pset<Packet4f>(sign_mask);
     rd = pxor(rd, p4f_sign_PNNP);
     iA = pmul(iA, rd);
@@ -326,8 +326,8 @@ struct compute_inverse_size4<Architecture::Target, double, MatrixType, ResultTyp
     iC1 = psub(pmul(B1, dC), iC1);
     iC2 = psub(pmul(B2, dC), iC2);
 
-    const double sign_mask1[2] = {0.0, -0.0};
-    const double sign_mask2[2] = {-0.0, 0.0};
+    const double sign_mask1[2] = {0.0, numext::bit_cast<double>(0x8000000000000000ull)};
+    const double sign_mask2[2] = {numext::bit_cast<double>(0x8000000000000000ull), 0.0};
     const Packet2d sign_PN = pset<Packet2d>(sign_mask1);
     const Packet2d sign_NP = pset<Packet2d>(sign_mask2);
     d1 = pxor(rd, sign_PN);