From 394aabb0a3976d95a5c6f286d49e43bb49558cc2 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Antonio=20S=C3=A1nchez?= <cantonios@google.com>
Date: Fri, 10 Mar 2023 22:36:57 +0000
Subject: [PATCH] Fix failing MSVC tests due to compiler bugs.

---
 Eigen/src/Core/arch/AVX/PacketMath.h    | 13 ++++++++++---
 Eigen/src/Core/arch/AVX512/PacketMath.h | 11 +++++++++--
 test/packet_ostream.h                   | 22 ++++++++++++++++++++++
 test/packetmath.cpp                     |  1 +
 4 files changed, 42 insertions(+), 5 deletions(-)
 create mode 100644 test/packet_ostream.h

diff --git a/Eigen/src/Core/arch/AVX/PacketMath.h b/Eigen/src/Core/arch/AVX/PacketMath.h
index 89963eb7a..3eb439434 100644
--- a/Eigen/src/Core/arch/AVX/PacketMath.h
+++ b/Eigen/src/Core/arch/AVX/PacketMath.h
@@ -984,13 +984,20 @@ template<> EIGEN_STRONG_INLINE void pstoreu<int>(int*       to, const Packet8i&
 template<> EIGEN_STRONG_INLINE void pstoreu<float>(float*   to, const Packet8f& from, uint8_t umask) {
 #ifdef EIGEN_VECTORIZE_AVX512
   __mmask16 mask = static_cast<__mmask16>(umask & 0x00FF);
-  EIGEN_DEBUG_UNALIGNED_STORE return _mm512_mask_storeu_ps(to, mask, _mm512_castps256_ps512(from));
+  EIGEN_DEBUG_UNALIGNED_STORE _mm512_mask_storeu_ps(to, mask, _mm512_castps256_ps512(from));
 #else
   Packet8i mask = _mm256_set1_epi8(static_cast<char>(umask));
-  const Packet8i bit_mask = _mm256_set_epi32(0xffffff7f, 0xffffffbf, 0xffffffdf, 0xffffffef, 0xfffffff7, 0xfffffffb, 0xfffffffd, 0xfffffffe);
+  const Packet8i bit_mask = _mm256_set_epi32(0x7f7f7f7f, 0xbfbfbfbf, 0xdfdfdfdf, 0xefefefef, 0xf7f7f7f7, 0xfbfbfbfb, 0xfdfdfdfd, 0xfefefefe);
   mask = por<Packet8i>(mask, bit_mask);
   mask = pcmp_eq<Packet8i>(mask, _mm256_set1_epi32(0xffffffff));
-  EIGEN_DEBUG_UNALIGNED_STORE return _mm256_maskstore_ps(to, mask, from);
+#if EIGEN_COMP_MSVC
+  // MSVC sometimes seems to use a bogus mask with maskstore.
+  const __m256i ifrom = _mm256_castps_si256(from);
+  EIGEN_DEBUG_UNALIGNED_STORE _mm_maskmoveu_si128(_mm256_extractf128_si256(ifrom, 0), _mm256_extractf128_si256(mask, 0), reinterpret_cast<char*>(to));
+  EIGEN_DEBUG_UNALIGNED_STORE _mm_maskmoveu_si128(_mm256_extractf128_si256(ifrom, 1), _mm256_extractf128_si256(mask, 1), reinterpret_cast<char*>(to + 4));
+#else
+  EIGEN_DEBUG_UNALIGNED_STORE _mm256_maskstore_ps(to, mask, from);
+#endif
 #endif
 }
 
diff --git a/Eigen/src/Core/arch/AVX512/PacketMath.h b/Eigen/src/Core/arch/AVX512/PacketMath.h
index 262211705..98b55ea06 100644
--- a/Eigen/src/Core/arch/AVX512/PacketMath.h
+++ b/Eigen/src/Core/arch/AVX512/PacketMath.h
@@ -336,12 +336,19 @@ EIGEN_STRONG_INLINE Packet16i psub<Packet16i>(const Packet16i& a,
 
 template <>
 EIGEN_STRONG_INLINE Packet16f pnegate(const Packet16f& a) {
-  const __m512i mask = _mm512_set1_epi32(0x80000000);
+  // NOTE: MSVC seems to struggle with _mm512_set1_epi32, leading to random results.
+  //       The intel docs give it a relatively high latency as well, so we're probably
+  //       better off with using _mm512_set_epi32 directly anyways.
+  const __m512i mask = _mm512_set_epi32(0x80000000,0x80000000,0x80000000,0x80000000,
+                                        0x80000000,0x80000000,0x80000000,0x80000000,
+                                        0x80000000,0x80000000,0x80000000,0x80000000,
+                                        0x80000000,0x80000000,0x80000000,0x80000000);
   return _mm512_castsi512_ps(_mm512_xor_epi32(_mm512_castps_si512(a), mask));
 }
 template <>
 EIGEN_STRONG_INLINE Packet8d pnegate(const Packet8d& a) {
-  const __m512i mask = _mm512_set1_epi64(0x8000000000000000ULL);
+  const __m512i mask = _mm512_set_epi64(0x8000000000000000ULL, 0x8000000000000000ULL, 0x8000000000000000ULL, 0x8000000000000000ULL,
+                                        0x8000000000000000ULL, 0x8000000000000000ULL, 0x8000000000000000ULL, 0x8000000000000000ULL);
   return _mm512_castsi512_pd(_mm512_xor_epi64(_mm512_castpd_si512(a), mask));
 }
 template <>
diff --git a/test/packet_ostream.h b/test/packet_ostream.h
new file mode 100644
index 000000000..ece499575
--- /dev/null
+++ b/test/packet_ostream.h
@@ -0,0 +1,22 @@
+#ifndef TEST_PACKET_OSTREAM
+#define TEST_PACKET_OSTREAM
+
+#include <type_traits>
+#include <ostream>
+
+// Include this header to be able to print Packets while debugging.
+
+template<typename Packet, typename EnableIf = std::enable_if_t<Eigen::internal::unpacket_traits<Packet>::vectorizable> >
+std::ostream& operator<<(std::ostream& os, const Packet& packet) {
+  using Scalar = typename Eigen::internal::unpacket_traits<Packet>::type;
+  Scalar v[Eigen::internal::unpacket_traits<Packet>::size];
+  Eigen::internal::pstoreu(v, packet);
+  os << "{" << v[0];
+  for (int i=1; i<Eigen::internal::unpacket_traits<Packet>::size; ++i) {
+    os << "," << v[i];
+  }
+  os << "}";
+  return os;
+}
+
+#endif  // TEST_PACKET_OSTREAM
\ No newline at end of file
diff --git a/test/packetmath.cpp b/test/packetmath.cpp
index 76332d7f7..a98a014df 100644
--- a/test/packetmath.cpp
+++ b/test/packetmath.cpp
@@ -10,6 +10,7 @@
 
 #include "packetmath_test_shared.h"
 #include "random_without_cast_overflow.h"
+#include "packet_ostream.h"
 
 template <typename T>
 inline T REF_ADD(const T& a, const T& b) {