From bbdf3192331e911fcdfbf6b60644a5b983872187 Mon Sep 17 00:00:00 2001
From: Jan Wassenberg <janwas@google.com>
Date: Wed, 10 Apr 2024 22:47:43 -0700
Subject: [PATCH] Re-enable (#3) native support of dynamic dispatch on Clang.
 Refs #838

PiperOrigin-RevId: 623714267
---
 hwy/contrib/math/math-inl.h   |  3 ++-
 hwy/contrib/math/math_test.cc |  8 ++++----
 hwy/detect_targets.h          | 13 +++++--------
 hwy/ops/arm_sve-inl.h         |  8 +++++---
 hwy/ops/set_macros-inl.h      |  2 +-
 5 files changed, 17 insertions(+), 17 deletions(-)

diff --git a/hwy/contrib/math/math-inl.h b/hwy/contrib/math/math-inl.h
index 10d7c176bf..5687f669a0 100644
--- a/hwy/contrib/math/math-inl.h
+++ b/hwy/contrib/math/math-inl.h
@@ -331,7 +331,8 @@ HWY_NOINLINE V CallTanh(const D d, VecArg<V> x) {
 template <class D, class V>
 HWY_INLINE void SinCos(D d, V x, V& s, V& c);
 template <class D, class V>
-HWY_NOINLINE V CallSinCos(const D d, VecArg<V> x, VecArg<V>& s, VecArg<V>& c) {
+HWY_NOINLINE void CallSinCos(const D d, VecArg<V> x, VecArg<V>& s,
+                             VecArg<V>& c) {
   SinCos(d, x, s, c);
 }
 
diff --git a/hwy/contrib/math/math_test.cc b/hwy/contrib/math/math_test.cc
index 02eb4abe0f..f2d2d18198 100644
--- a/hwy/contrib/math/math_test.cc
+++ b/hwy/contrib/math/math_test.cc
@@ -119,7 +119,7 @@ HWY_NOINLINE void TestMath(const char* name, T (*fx1)(T),
       const auto ulp = hwy::detail::ComputeUlpDelta(actual, expected);
       max_ulp = HWY_MAX(max_ulp, ulp);
       if (ulp > max_error_ulp) {
-        fprintf(stderr, "%s: %s(%f) expected %f actual %f ulp %g max ulp %u\n",
+        fprintf(stderr, "%s: %s(%f) expected %E actual %E ulp %g max ulp %u\n",
                 hwy::TypeName(T(), Lanes(d)).c_str(), name, value, expected,
                 actual, static_cast<double>(ulp),
                 static_cast<uint32_t>(max_error_ulp));
@@ -181,14 +181,14 @@ constexpr uint64_t ACosh32ULP() {
 template <class D>
 static Vec<D> SinCosSin(const D d, VecArg<Vec<D>> x) {
   Vec<D> s, c;
-  SinCos(d, x, s, c);
+  CallSinCos(d, x, s, c);
   return s;
 }
 
 template <class D>
 static Vec<D> SinCosCos(const D d, VecArg<Vec<D>> x) {
   Vec<D> s, c;
-  SinCos(d, x, s, c);
+  CallSinCos(d, x, s, c);
   return c;
 }
 
@@ -390,7 +390,7 @@ struct TestAtan2 {
       if (!AllTrue(d, ok)) {
         const size_t mismatch =
             static_cast<size_t>(FindKnownFirstTrue(d, Not(ok)));
-        fprintf(stderr, "Mismatch for i=%d expected %f actual %f\n",
+        fprintf(stderr, "Mismatch for i=%d expected %E actual %E\n",
                 static_cast<int>(i + mismatch), expected[i + mismatch],
                 ExtractLane(actual, mismatch));
         HWY_ASSERT(0);
diff --git a/hwy/detect_targets.h b/hwy/detect_targets.h
index f3571bcf7a..b75c0eb85f 100644
--- a/hwy/detect_targets.h
+++ b/hwy/detect_targets.h
@@ -554,14 +554,11 @@
 // Clang, GCC and MSVC allow runtime dispatch on x86.
 #if HWY_ARCH_X86
 #define HWY_HAVE_RUNTIME_DISPATCH 1
-// On Arm, PPC, S390X, and RISC-V: GCC and Clang 16+ do, and we require Linux
-// to detect CPU capabilities. Currently require opt-in for Clang on Arm
-// because it is experimental.
-#elif (HWY_ARCH_ARM || HWY_ARCH_PPC || HWY_ARCH_S390X || HWY_ARCH_RISCV) && \
-    (HWY_COMPILER_GCC_ACTUAL ||                                           \
-     (HWY_COMPILER_CLANG >= 1600 &&                                       \
-      (!HWY_ARCH_ARM || defined(HWY_ENABLE_CLANG_ARM_DISPATCH)))) &&      \
-    HWY_OS_LINUX && !defined(TOOLCHAIN_MISS_SYS_AUXV_H)
+// On Arm, PPC, S390X, and RISC-V: GCC and Clang 17+ do, and we require Linux
+// to detect CPU capabilities.
+#elif (HWY_ARCH_ARM || HWY_ARCH_PPC || HWY_ARCH_S390X || HWY_ARCH_RISCV) &&    \
+    (HWY_COMPILER_GCC_ACTUAL || HWY_COMPILER_CLANG >= 1700) && HWY_OS_LINUX && \
+    !defined(TOOLCHAIN_MISS_SYS_AUXV_H)
 #define HWY_HAVE_RUNTIME_DISPATCH 1
 #else
 #define HWY_HAVE_RUNTIME_DISPATCH 0
diff --git a/hwy/ops/arm_sve-inl.h b/hwy/ops/arm_sve-inl.h
index 211c3653c9..64d70c0213 100644
--- a/hwy/ops/arm_sve-inl.h
+++ b/hwy/ops/arm_sve-inl.h
@@ -3257,13 +3257,15 @@ HWY_API TFromV<V> ExtractLane(V v, size_t i) {
 }
 
 // ------------------------------ InsertLane (IfThenElse)
-template <class V>
-HWY_API V InsertLane(const V v, size_t i, TFromV<V> t) {
+template <class V, typename T>
+HWY_API V InsertLane(const V v, size_t i, T t) {
+  static_assert(sizeof(TFromV<V>) == sizeof(T), "Lane size mismatch");
   const DFromV<V> d;
   const RebindToSigned<decltype(d)> di;
   using TI = TFromD<decltype(di)>;
   const svbool_t is_i = detail::EqN(Iota(di, 0), static_cast<TI>(i));
-  return IfThenElse(RebindMask(d, is_i), Set(d, t), v);
+  return IfThenElse(RebindMask(d, is_i),
+                    Set(d, hwy::ConvertScalarTo<TFromV<V>>(t)), v);
 }
 
 // ------------------------------ DupEven
diff --git a/hwy/ops/set_macros-inl.h b/hwy/ops/set_macros-inl.h
index ef172e7a28..104c2ea1f4 100644
--- a/hwy/ops/set_macros-inl.h
+++ b/hwy/ops/set_macros-inl.h
@@ -416,7 +416,7 @@
 #if HWY_TARGET == HWY_NEON_WITHOUT_AES
 // Do not define HWY_TARGET_STR (no pragma).
 #else
-#define HWY_TARGET_STR "+crypto"
+#define HWY_TARGET_STR "+aes"
 #endif  // HWY_TARGET == HWY_NEON_WITHOUT_AES
 
 #endif  // HWY_ARCH_ARM_V7