From b995b501ef5cf113534c5aead6c85baea6cc423c Mon Sep 17 00:00:00 2001
From: Behdad Esfahbod <behdad@behdad.org>
Date: Tue, 22 May 2018 21:06:22 -0700
Subject: [PATCH] Try enabling vectorization smartly

We'll see if this sticks to the bots.
---
 src/hb-private.hh     | 25 +++++++++++++++++--------
 src/hb-set-private.hh |  7 -------
 2 files changed, 17 insertions(+), 15 deletions(-)

diff --git a/src/hb-private.hh b/src/hb-private.hh
index 0dd9539..0b7d902 100644
--- a/src/hb-private.hh
+++ b/src/hb-private.hh
@@ -978,6 +978,14 @@ struct HbOpXor
   template <typename T> static void process (T &o, const T &a, const T &b) { o = a ^ b; }
 };
 
+
+/* Compiler-assisted vectorization. */
+
+/* The `vector_size' attribute was introduced in gcc 3.1. */
+#if defined( __GNUC__ ) && ( __GNUC__ >= 4 )
+#define HAVE_VECTOR_SIZE 1
+#endif
+
 /* Type behaving similar to vectorized vars defined using __attribute__((vector_size(...))). */
 template <typename elt_t, unsigned int byte_size>
 struct hb_vector_size_t
@@ -989,8 +997,11 @@ struct hb_vector_size_t
   inline hb_vector_size_t process (const hb_vector_size_t &o) const
   {
     hb_vector_size_t r;
-    for (unsigned int i = 0; i < ARRAY_LENGTH (v); i++)
-      Op::process (r.v[i], v[i], o.v[i]);
+    if (HAVE_VECTOR_SIZE+0)
+      Op::process (r.vec, vec, o.vec);
+    else
+      for (unsigned int i = 0; i < ARRAY_LENGTH (v); i++)
+	Op::process (r.v[i], v[i], o.v[i]);
     return r;
   }
   inline hb_vector_size_t operator | (const hb_vector_size_t &o) const
@@ -1009,14 +1020,12 @@ struct hb_vector_size_t
 
   private:
   static_assert (byte_size / sizeof (elt_t) * sizeof (elt_t) == byte_size, "");
-  elt_t v[byte_size / sizeof (elt_t)];
+  union {
+    elt_t v[byte_size / sizeof (elt_t)];
+    elt_t vec __attribute__((vector_size (byte_size))); /* Only usable if HAVE_VECTOR_SIZE */
+  };
 };
 
-/* The `vector_size' attribute was introduced in gcc 3.1. */
-#if defined( __GNUC__ ) && ( __GNUC__ >= 4 )
-#define HAVE_VECTOR_SIZE 1
-#endif
-
 
 /* Global runtime options. */
 
diff --git a/src/hb-set-private.hh b/src/hb-set-private.hh
index 9a9f37f..7a604e1 100644
--- a/src/hb-set-private.hh
+++ b/src/hb-set-private.hh
@@ -164,14 +164,7 @@ struct hb_set_t
     static inline unsigned int elt_get_min (const elt_t &elt) { return _hb_ctz (elt); }
     static inline unsigned int elt_get_max (const elt_t &elt) { return _hb_bit_storage (elt) - 1; }
 
-#if 0 && HAVE_VECTOR_SIZE
-    /* The vectorized version does not work with clang as non-const
-     * elt() errs "non-const reference cannot bind to vector element". */
-    typedef elt_t vector_t __attribute__((vector_size (PAGE_BITS / 8)));
-#else
     typedef hb_vector_size_t<elt_t, PAGE_BITS / 8> vector_t;
-#endif
-
     vector_t v;
 
     static const unsigned int ELT_BITS = sizeof (elt_t) * 8;
-- 
2.7.4