Fix ldexp implementations.

The previous implementations produced garbage values if the exponent did
not fit within the exponent bits.  See #2131 for a complete discussion,
and !375 for other possible implementations.

Here we implement the 4-factor version. See `pldexp_impl` in
`GenericPacketMathFunctions.h` for a full description.

The SSE `pcmp*` methods were moved down since `pcmp_le<Packet4i>`
requires `por`.

Left as a "TODO" is to delegate to a faster version if we know the
exponent does fit within the exponent bits.

Fixes #2131.

1 files changed, 31 insertions, 22 deletions

diff --git a/Eigen/src/Core/arch/SSE/PacketMath.h b/Eigen/src/Core/arch/SSE/PacketMath.h
index 05d9f8edd..422b0fce7 100755
--- a/Eigen/src/Core/arch/SSE/PacketMath.h
+++ b/Eigen/src/Core/arch/SSE/PacketMath.h
@@ -399,21 +399,6 @@ template<> EIGEN_DEVICE_FUNC inline Packet16b pselect(const Packet16b& mask, con
 }
 #endif
 
-template<> EIGEN_STRONG_INLINE Packet4f pcmp_le(const Packet4f& a, const Packet4f& b) { return _mm_cmple_ps(a,b); }
-template<> EIGEN_STRONG_INLINE Packet4f pcmp_lt(const Packet4f& a, const Packet4f& b) { return _mm_cmplt_ps(a,b); }
-template<> EIGEN_STRONG_INLINE Packet4f pcmp_lt_or_nan(const Packet4f& a, const Packet4f& b) { return _mm_cmpnge_ps(a,b); }
-template<> EIGEN_STRONG_INLINE Packet4f pcmp_eq(const Packet4f& a, const Packet4f& b) { return _mm_cmpeq_ps(a,b); }
-
-template<> EIGEN_STRONG_INLINE Packet2d pcmp_le(const Packet2d& a, const Packet2d& b) { return _mm_cmple_pd(a,b); }
-template<> EIGEN_STRONG_INLINE Packet2d pcmp_lt(const Packet2d& a, const Packet2d& b) { return _mm_cmplt_pd(a,b); }
-template<> EIGEN_STRONG_INLINE Packet2d pcmp_lt_or_nan(const Packet2d& a, const Packet2d& b) { return _mm_cmpnge_pd(a,b); }
-template<> EIGEN_STRONG_INLINE Packet2d pcmp_eq(const Packet2d& a, const Packet2d& b) { return _mm_cmpeq_pd(a,b); }
-
-template<> EIGEN_STRONG_INLINE Packet4i pcmp_lt(const Packet4i& a, const Packet4i& b) { return _mm_cmplt_epi32(a,b); }
-template<> EIGEN_STRONG_INLINE Packet4i pcmp_eq(const Packet4i& a, const Packet4i& b) { return _mm_cmpeq_epi32(a,b); }
-template<> EIGEN_STRONG_INLINE Packet16b pcmp_eq(const Packet16b& a, const Packet16b& b) { return _mm_cmpeq_epi8(a,b); }
-
-
 template<> EIGEN_STRONG_INLINE Packet4i ptrue<Packet4i>(const Packet4i& a) { return _mm_cmpeq_epi32(a, a); }
 template<> EIGEN_STRONG_INLINE Packet16b ptrue<Packet16b>(const Packet16b& a) { return _mm_cmpeq_epi8(a, a); }
 template<> EIGEN_STRONG_INLINE Packet4f
@@ -447,6 +432,21 @@ template<> EIGEN_STRONG_INLINE Packet4f pandnot<Packet4f>(const Packet4f& a, con
 template<> EIGEN_STRONG_INLINE Packet2d pandnot<Packet2d>(const Packet2d& a, const Packet2d& b) { return _mm_andnot_pd(b,a); }
 template<> EIGEN_STRONG_INLINE Packet4i pandnot<Packet4i>(const Packet4i& a, const Packet4i& b) { return _mm_andnot_si128(b,a); }
 
+template<> EIGEN_STRONG_INLINE Packet4f pcmp_le(const Packet4f& a, const Packet4f& b) { return _mm_cmple_ps(a,b); }
+template<> EIGEN_STRONG_INLINE Packet4f pcmp_lt(const Packet4f& a, const Packet4f& b) { return _mm_cmplt_ps(a,b); }
+template<> EIGEN_STRONG_INLINE Packet4f pcmp_lt_or_nan(const Packet4f& a, const Packet4f& b) { return _mm_cmpnge_ps(a,b); }
+template<> EIGEN_STRONG_INLINE Packet4f pcmp_eq(const Packet4f& a, const Packet4f& b) { return _mm_cmpeq_ps(a,b); }
+
+template<> EIGEN_STRONG_INLINE Packet2d pcmp_le(const Packet2d& a, const Packet2d& b) { return _mm_cmple_pd(a,b); }
+template<> EIGEN_STRONG_INLINE Packet2d pcmp_lt(const Packet2d& a, const Packet2d& b) { return _mm_cmplt_pd(a,b); }
+template<> EIGEN_STRONG_INLINE Packet2d pcmp_lt_or_nan(const Packet2d& a, const Packet2d& b) { return _mm_cmpnge_pd(a,b); }
+template<> EIGEN_STRONG_INLINE Packet2d pcmp_eq(const Packet2d& a, const Packet2d& b) { return _mm_cmpeq_pd(a,b); }
+
+template<> EIGEN_STRONG_INLINE Packet4i pcmp_lt(const Packet4i& a, const Packet4i& b) { return _mm_cmplt_epi32(a,b); }
+template<> EIGEN_STRONG_INLINE Packet4i pcmp_eq(const Packet4i& a, const Packet4i& b) { return _mm_cmpeq_epi32(a,b); }
+template<> EIGEN_STRONG_INLINE Packet16b pcmp_eq(const Packet16b& a, const Packet16b& b) { return _mm_cmpeq_epi8(a,b); }
+template<> EIGEN_STRONG_INLINE Packet4i pcmp_le(const Packet4i& a, const Packet4i& b) { return por(pcmp_lt(a,b), pcmp_eq(a,b)); }
+
 template<> EIGEN_STRONG_INLINE Packet4f pmin<Packet4f>(const Packet4f& a, const Packet4f& b) {
 #if EIGEN_COMP_GNUC && EIGEN_COMP_GNUC < 63
   // There appears to be a bug in GCC, by which the optimizer may
@@ -907,13 +907,22 @@ template<> EIGEN_STRONG_INLINE Packet4f pldexp<Packet4f>(const Packet4f& a, cons
 // We specialize pldexp here, since the generic implementation uses Packet2l, which is not well
 // supported by SSE, and has more range than is needed for exponents.
 template<> EIGEN_STRONG_INLINE Packet2d pldexp<Packet2d>(const Packet2d& a, const Packet2d& exponent) {
-  const Packet2d cst_1023 = pset1<Packet2d>(1023.0);
-  // Add exponent offset.
-  Packet4i ei = _mm_cvtpd_epi32(padd(exponent, cst_1023));
-  // Convert to exponents to int64 and swizzle to the low-order 32 bits.
-  ei = vec4i_swizzle1(ei, 0, 3, 1, 3);
-  // return a * 2^exponent
-  return pmul(a, _mm_castsi128_pd(_mm_slli_epi64(ei, 52)));
+  // Clamp exponent to [-2099, 2099]
+  const Packet2d max_exponent = pset1<Packet2d>(2099.0);
+  const Packet2d e = pmin(pmax(exponent, pnegate(max_exponent)), max_exponent);
+  
+  // Convert e to integer and swizzle to low-order bits.
+  const Packet4i ei = vec4i_swizzle1(_mm_cvtpd_epi32(e), 0, 3, 1, 3);
+  
+  // Split 2^e into four factors and multiply:
+  const Packet4i bias = _mm_set_epi32(0, 1023, 0, 1023);
+  Packet4i b = parithmetic_shift_right<2>(ei);  // floor(e/4)
+  Packet2d c = _mm_castsi128_pd(_mm_slli_epi64(padd(b, bias), 52));  // 2^b
+  Packet2d out = pmul(pmul(pmul(a, c), c), c); // a * 2^(3b)
+  b = psub(psub(psub(ei, b), b), b);  // e - 3b
+  c = _mm_castsi128_pd(_mm_slli_epi64(padd(b, bias), 52));  // 2^(e - 3b)
+  out = pmul(out, c);  // a * 2^e
+  return out;
 }
 
 // with AVX, the default implementations based on pload1 are faster