Start to vectorize SkTileGrid.

This adds Sk4x.h to help.

BUG=skia:

Review URL: https://codereview.chromium.org/634543004

9 files changed, 625 insertions, 56 deletions

diff --git a/bench/GeometryBench.cpp b/bench/GeometryBench.cpp
index a5c4643f9e..65398fa34a 100644
--- a/bench/GeometryBench.cpp
+++ b/bench/GeometryBench.cpp
@@ -6,6 +6,7 @@
  */
 
 #include "Benchmark.h"
+#include "Sk4x.h"
 #include "SkGeometry.h"
 #include "SkRandom.h"
 #include "SkRect.h"
@@ -44,6 +45,9 @@ public:
     GeoRectBench(const char suffix[]) : GeometryBench(suffix) {}
 
 protected:
+    // void* vptr;
+    size_t align_fRects_to_16Bytes[sizeof(void*) == 8 ? 1 : 3];
+
     SkRect fRects[2048];
 
     virtual void onPreDraw() {
@@ -97,7 +101,7 @@ protected:
 class GeoRectBench_Intersects : public GeoRectBench {
 public:
     GeoRectBench_Intersects() : GeoRectBench("rect_Intersects") {}
-    
+
 protected:
     virtual void onDraw(const int loops, SkCanvas* canvas) SK_OVERRIDE {
         for (int outer = 0; outer < loops; ++outer) {
@@ -113,7 +117,7 @@ protected:
 class GeoRectBench_sort : public GeoRectBench {
 public:
     GeoRectBench_sort() : GeoRectBench("rect_sort") {}
-    
+
 protected:
     virtual void onDraw(const int loops, SkCanvas* canvas) SK_OVERRIDE {
         for (int outer = 0; outer < loops; ++outer) {
@@ -129,3 +133,59 @@ DEF_BENCH( return new GeoRectBench_intersect_rect; )
 DEF_BENCH( return new GeoRectBench_Intersects; )
 
 DEF_BENCH( return new GeoRectBench_sort; )
+
+class GeoRectBench_sort_4f : public GeoRectBench {
+public:
+    GeoRectBench_sort_4f() : GeoRectBench("rect_sort_4f") { }
+
+protected:
+    static SkRect Sort(const SkRect& rect) {
+        // To sort:
+        //   left, right = minmax(left, right)
+        //   top, bottom = minmax(top, bottom)
+        Sk4f ltrb(&rect.fLeft),
+             rblt = ltrb.zwxy(),
+             ltlt = Sk4f::Min(ltrb, rblt),  // Holds (2 copies of) new left and top.
+             rbrb = Sk4f::Max(ltrb, rblt),  // Holds (2 copies of) new right and bottom.
+             sort = Sk4f::XYAB(ltlt, rbrb);
+
+        SkRect sorted;
+        sort.store(&sorted.fLeft);
+        return sorted;
+    }
+
+    virtual void onDraw(const int loops, SkCanvas* canvas) SK_OVERRIDE {
+        for (int outer = 0; outer < loops; ++outer) {
+            for (size_t i = 0; i < SK_ARRAY_COUNT(fRects); ++i) {
+                fRects[i] = Sort(fRects[i]);
+            }
+        }
+    }
+};
+DEF_BENCH( return new GeoRectBench_sort_4f; )
+
+class GeoRectBench_Intersects_4f : public GeoRectBench {
+public:
+    GeoRectBench_Intersects_4f() : GeoRectBench("rect_Intersects_4f") {}
+
+protected:
+    static bool Intersects(const SkRect& a, const SkRect& b) {
+        Sk4f r1(&a.fLeft),
+             r2(&b.fLeft),
+             lt = Sk4f::XYAB(r1, r2),  // a.L a.T b.L b.T <
+             rb = Sk4f::ZWCD(r2, r1);  // b.R b.B a.R a.B ?
+        return lt.lessThan(rb).allTrue();
+    }
+
+    virtual void onDraw(const int loops, SkCanvas* canvas) SK_OVERRIDE {
+        for (int outer = 0; outer < loops; ++outer) {
+            int count = 0;
+            for (size_t i = 0; i < SK_ARRAY_COUNT(fRects); ++i) {
+                count += Intersects(fRects[0], fRects[i]);
+            }
+            this->virtualCallToFoilOptimizers(count);
+        }
+    }
+};
+DEF_BENCH( return new GeoRectBench_Intersects_4f; )
+
diff --git a/gyp/common_conditions.gypi b/gyp/common_conditions.gypi
index d9c048771e..d213e17e4f 100644
--- a/gyp/common_conditions.gypi
+++ b/gyp/common_conditions.gypi
@@ -46,6 +46,7 @@
             4275,  # An exported class was derived from a class that was not exported
             4345,  # This is an FYI about a behavior change from long ago. Chrome stifles it too.
             4355,  # 'this' used in base member initializer list. Off by default in newer compilers.
+            4800,  # forcing value to bool 'true' or 'false'
         ],
         'msvs_cygwin_shell': 0,
         'msvs_settings': {
diff --git a/include/core/SkTypes.h b/include/core/SkTypes.h
index 0e9e230349..2625b7343d 100644
--- a/include/core/SkTypes.h
+++ b/include/core/SkTypes.h
@@ -300,6 +300,9 @@ static inline bool SkIsU16(long x) {
 #define SkAlign8(x)     (((x) + 7) >> 3 << 3)
 #define SkIsAlign8(x)   (0 == ((x) & 7))
 
+#define SkAlign16(x)    (((x) + 15) >> 4 << 4)
+#define SkIsAlign16(x)  (0 == ((x) & 15))
+
 #define SkAlignPtr(x)   (sizeof(void*) == 8 ?   SkAlign8(x) :   SkAlign4(x))
 #define SkIsAlignPtr(x) (sizeof(void*) == 8 ? SkIsAlign8(x) : SkIsAlign4(x))
 
diff --git a/src/core/Sk4x.h b/src/core/Sk4x.h
new file mode 100644
index 0000000000..9af12c4e91
--- /dev/null
+++ b/src/core/Sk4x.h
@@ -0,0 +1,98 @@
+#ifndef Sk4x_DEFINED
+#define Sk4x_DEFINED
+
+#include "SkTypes.h"
+
+// First we'll let Clang or GCC try their best with whatever instructions are available.
+// Otherwise fall back on portable code.  This really should be a last resort.
+
+#define SK4X_PREAMBLE 1
+    #if defined(__clang__)
+        #include "Sk4x_clang.h"
+    #elif defined(__GNUC__)
+        #include "Sk4x_gcc.h"
+    #else
+        #include "Sk4x_portable.h"
+    #endif
+#undef SK4X_PREAMBLE
+
+template <typename T> class Sk4x;
+typedef Sk4x<int>   Sk4i;
+typedef Sk4x<float> Sk4f;
+
+template <typename T> class Sk4x {
+public:
+    Sk4x();  // Uninitialized; use Sk4x(0,0,0,0) for zero.
+    Sk4x(T, T, T, T);
+    explicit Sk4x(const T[4]);
+
+    Sk4x(const Sk4x&);
+    Sk4x& operator=(const Sk4x&);
+
+    void set(T, T, T, T);
+
+    void store(T[4]) const;
+
+    template <typename Dst> Dst reinterpret() const;
+    template <typename Dst> Dst cast() const;
+
+    bool allTrue() const;
+    bool anyTrue() const;
+
+    Sk4x bitNot() const;
+    Sk4x bitAnd(const Sk4x&) const;
+    Sk4x bitOr (const Sk4x&) const;
+
+    Sk4i            equal(const Sk4x&) const;
+    Sk4i         notEqual(const Sk4x&) const;
+    Sk4i         lessThan(const Sk4x&) const;
+    Sk4i      greaterThan(const Sk4x&) const;
+    Sk4i    lessThanEqual(const Sk4x&) const;
+    Sk4i greaterThanEqual(const Sk4x&) const;
+
+    Sk4x      add(const Sk4x&) const;
+    Sk4x subtract(const Sk4x&) const;
+    Sk4x multiply(const Sk4x&) const;
+    Sk4x   divide(const Sk4x&) const;
+
+    static Sk4x Min(const Sk4x& a, const Sk4x& b);
+    static Sk4x Max(const Sk4x& a, const Sk4x& b);
+
+    // Swizzles follow OpenCL xyzw convention.
+    Sk4x zwxy() const;
+
+    // When there's a second argument, it's abcd.
+    static Sk4x XYAB(const Sk4x& xyzw, const Sk4x& abcd);
+    static Sk4x ZWCD(const Sk4x& xyzw, const Sk4x& abcd);
+
+private:
+    // It's handy to have Sk4f and Sk4i be mutual friends.
+    template <typename S> friend class Sk4x;
+
+#define SK4X_PRIVATE 1
+    #if defined(__clang__)
+        #include "Sk4x_clang.h"
+    #elif defined(__GNUC__)
+        #include "Sk4x_gcc.h"
+    #else
+        #include "Sk4x_portable.h"
+    #endif
+#undef SK4X_PRIVATE
+};
+
+#if defined(__clang__)
+    #include "Sk4x_clang.h"
+#elif defined(__GNUC__)
+    #include "Sk4x_gcc.h"
+#else
+    #include "Sk4x_portable.h"
+#endif
+
+// TODO ideas for enterprising coders:
+//   1) Code generated for Max() isn't as good in Sk4x_gcc.h as it is in _clang.  Why?
+//   2) Sk4x_sse.h would be good for Windows, and could possibly beat _clang / _gcc
+//      (e.g. they can't generate _mm_movemask_ps for allTrue/anyTrue).
+//   3) Sk4x_neon.h might be a good idea if _clang / _gcc aren't good enough on ARM.
+
+
+#endif//Sk4x_DEFINED
diff --git a/src/core/Sk4x_clang.h b/src/core/Sk4x_clang.h
new file mode 100644
index 0000000000..ae976ba3da
--- /dev/null
+++ b/src/core/Sk4x_clang.h
@@ -0,0 +1,125 @@
+// It is important _not_ to put header guards here.
+// This file will be intentionally included three times.
+
+// Useful reading:
+//   http://clang.llvm.org/docs/LanguageExtensions.html#vectors-and-extended-vectors
+
+#if defined(SK4X_PREAMBLE)
+
+#elif defined(SK4X_PRIVATE)
+    typedef T Vector __attribute__((ext_vector_type(4)));
+
+    /*implicit*/ Sk4x(Vector vec) : fVec(vec) {}
+
+    template <int m, int a, int s, int k>
+    static Sk4x Shuffle(const Sk4x&, const Sk4x&);
+
+    Vector fVec;
+
+#else  // defined(SK4X_PRIVATE)
+
+template <typename T>
+Sk4x<T>::Sk4x() { }
+
+template <typename T>
+Sk4x<T>::Sk4x(T a, T b, T c, T d) { this->set(a,b,c,d); }
+
+template <typename T>
+Sk4x<T>::Sk4x(const T vals[4]) { this->set(vals[0], vals[1], vals[2], vals[3]); }
+
+template <typename T>
+Sk4x<T>::Sk4x(const Sk4x<T>& other) { *this = other; }
+
+template <typename T>
+Sk4x<T>& Sk4x<T>::operator=(const Sk4x<T>& other) { fVec = other.fVec; return *this; }
+
+template <typename T>
+void Sk4x<T>::set(T a, T b, T c, T d) {
+    Vector v = { a, b, c, d };
+    fVec = v;
+}
+
+template <typename T>
+void Sk4x<T>::store(T vals[4]) const {
+    SkASSERT(SkIsAlign16((uintptr_t)vals));
+    *reinterpret_cast<Vector*>(vals) = fVec;
+}
+
+template <typename T>
+template <typename Dst> Dst Sk4x<T>::reinterpret() const {
+    return Dst((typename Dst::Vector)fVec);
+}
+
+template <typename T>
+template <typename Dst> Dst Sk4x<T>::cast() const {
+    #if __has_builtin(__builtin_convertvector)
+        return Dst(__builtin_convertvector(fVec, typename Dst::Vector));
+    #else
+        return Dst(fVec[0], fVec[1], fVec[2], fVec[3]);
+    #endif
+}
+
+template <typename T>
+bool Sk4x<T>::allTrue() const { return fVec[0] & fVec[1] & fVec[2] & fVec[3]; }
+template <typename T>
+bool Sk4x<T>::anyTrue() const { return fVec[0] | fVec[1] | fVec[2] | fVec[3]; }
+
+template <typename T> Sk4x<T> Sk4x<T>::bitNot() const { return ~fVec; }
+
+template <typename T> Sk4x<T> Sk4x<T>::bitAnd(const Sk4x& other) const { return fVec & other.fVec; }
+template <typename T> Sk4x<T> Sk4x<T>::bitOr (const Sk4x& other) const { return fVec | other.fVec; }
+
+template <typename T>
+Sk4i Sk4x<T>::           equal(const Sk4x<T>& other) const { return fVec == other.fVec; }
+template <typename T>
+Sk4i Sk4x<T>::        notEqual(const Sk4x<T>& other) const { return fVec != other.fVec; }
+template <typename T>
+Sk4i Sk4x<T>::        lessThan(const Sk4x<T>& other) const { return fVec  < other.fVec; }
+template <typename T>
+Sk4i Sk4x<T>::     greaterThan(const Sk4x<T>& other) const { return fVec  > other.fVec; }
+template <typename T>
+Sk4i Sk4x<T>::   lessThanEqual(const Sk4x<T>& other) const { return fVec <= other.fVec; }
+template <typename T>
+Sk4i Sk4x<T>::greaterThanEqual(const Sk4x<T>& other) const { return fVec >= other.fVec; }
+
+template <typename T>
+Sk4x<T> Sk4x<T>::     add(const Sk4x<T>& other) const { return fVec + other.fVec; }
+template <typename T>
+Sk4x<T> Sk4x<T>::subtract(const Sk4x<T>& other) const { return fVec - other.fVec; }
+template <typename T>
+Sk4x<T> Sk4x<T>::multiply(const Sk4x<T>& other) const { return fVec * other.fVec; }
+template <typename T>
+Sk4x<T> Sk4x<T>::  divide(const Sk4x<T>& other) const { return fVec / other.fVec; }
+
+template <typename T>
+Sk4x<T> Sk4x<T>::Min(const Sk4x<T>& a, const Sk4x<T>& b) {
+    Sk4i less(a.lessThan(b));
+    Sk4i val = a.reinterpret<Sk4i>().bitAnd(less).bitOr(
+               b.reinterpret<Sk4i>().bitAnd(less.bitNot()));
+    return val.reinterpret<Sk4x>();
+}
+
+template <typename T>
+Sk4x<T> Sk4x<T>::Max(const Sk4x<T>& a, const Sk4x<T>& b) {
+    Sk4i less(a.lessThan(b));
+    Sk4i val = b.reinterpret<Sk4i>().bitAnd(less).bitOr(
+               a.reinterpret<Sk4i>().bitAnd(less.bitNot()));
+    return val.reinterpret<Sk4x>();
+}
+
+template <typename T>
+template <int m, int a, int s, int k>
+Sk4x<T> Sk4x<T>::Shuffle(const Sk4x<T>& x, const Sk4x<T>& y) {
+    return __builtin_shufflevector(x.fVec, y.fVec, m,a,s,k);
+}
+
+template <typename T>
+Sk4x<T> Sk4x<T>::zwxy() const { return fVec.zwxy; }
+
+template <typename T>
+Sk4x<T> Sk4x<T>::XYAB(const Sk4x& xyzw, const Sk4x& abcd) { return Shuffle<0,1,4,5>(xyzw, abcd); }
+
+template <typename T>
+Sk4x<T> Sk4x<T>::ZWCD(const Sk4x& xyzw, const Sk4x& abcd) { return Shuffle<2,3,6,7>(xyzw, abcd); }
+
+#endif // defined(SK4X_PRIVATE)
diff --git a/src/core/Sk4x_gcc.h b/src/core/Sk4x_gcc.h
new file mode 100644
index 0000000000..dcef4c2978
--- /dev/null
+++ b/src/core/Sk4x_gcc.h
@@ -0,0 +1,135 @@
+// It is important _not_ to put header guards here.
+// This file will be intentionally included three times.
+
+// Useful reading:
+//   https://gcc.gnu.org/onlinedocs/gcc/Vector-Extensions.html
+
+#if defined(SK4X_PREAMBLE)
+
+#elif defined(SK4X_PRIVATE)
+    typedef T Vector __attribute__((vector_size(16)));
+
+    /*implicit*/ Sk4x(Vector vec) : fVec(vec) {}
+    static inline Vector ShuffleImpl(Vector a, Vector b, int __attribute__((vector_size(16))) mask);
+    template <int m, int a, int s, int k>
+    static Sk4x Shuffle(const Sk4x&, const Sk4x&);
+
+    Vector fVec;
+
+#else  // defined(SK4X_PRIVATE)
+
+template <typename T>
+Sk4x<T>::Sk4x() { }
+
+template <typename T>
+Sk4x<T>::Sk4x(T a, T b, T c, T d) { this->set(a,b,c,d); }
+
+template <typename T>
+Sk4x<T>::Sk4x(const T vals[4]) {
+    fVec = *reinterpret_cast<const Vector*>(vals);  // Should compile to moveaps or moveups.
+}
+
+template <typename T>
+Sk4x<T>::Sk4x(const Sk4x<T>& other) { *this = other; }
+
+template <typename T>
+Sk4x<T>& Sk4x<T>::operator=(const Sk4x<T>& other) { fVec = other.fVec; return *this; }
+
+template <typename T>
+void Sk4x<T>::set(T a, T b, T c, T d) {
+    Vector v = { a, b, c, d };
+    fVec = v;
+}
+
+template <typename T>
+void Sk4x<T>::store(T vals[4]) const {
+    SkASSERT(SkIsAlign16((uintptr_t)vals));
+    *reinterpret_cast<Vector*>(vals) = fVec;
+}
+
+template <typename T>
+template <typename Dst> Dst Sk4x<T>::reinterpret() const {
+    return Dst((typename Dst::Vector)fVec);
+}
+
+template <typename T>
+template <typename Dst> Dst Sk4x<T>::cast() const {
+    return Dst(fVec[0], fVec[1], fVec[2], fVec[3]);
+}
+
+template <typename T>
+bool Sk4x<T>::allTrue() const { return fVec[0] & fVec[1] & fVec[2] & fVec[3]; }
+template <typename T>
+bool Sk4x<T>::anyTrue() const { return fVec[0] | fVec[1] | fVec[2] | fVec[3]; }
+
+template <typename T> Sk4x<T> Sk4x<T>::bitNot() const { return Sk4i(~fVec); }
+
+template <typename T> Sk4x<T> Sk4x<T>::bitAnd(const Sk4x& other) const { return fVec & other.fVec; }
+template <typename T> Sk4x<T> Sk4x<T>::bitOr (const Sk4x& other) const { return fVec | other.fVec; }
+
+template <typename T>
+Sk4i Sk4x<T>::           equal(const Sk4x<T>& other) const { return fVec == other.fVec; }
+template <typename T>
+Sk4i Sk4x<T>::        notEqual(const Sk4x<T>& other) const { return fVec != other.fVec; }
+template <typename T>
+Sk4i Sk4x<T>::        lessThan(const Sk4x<T>& other) const { return fVec  < other.fVec; }
+template <typename T>
+Sk4i Sk4x<T>::     greaterThan(const Sk4x<T>& other) const { return fVec  > other.fVec; }
+template <typename T>
+Sk4i Sk4x<T>::   lessThanEqual(const Sk4x<T>& other) const { return fVec <= other.fVec; }
+template <typename T>
+Sk4i Sk4x<T>::greaterThanEqual(const Sk4x<T>& other) const { return fVec >= other.fVec; }
+
+template <typename T>
+Sk4x<T> Sk4x<T>::     add(const Sk4x<T>& other) const { return fVec + other.fVec; }
+template <typename T>
+Sk4x<T> Sk4x<T>::subtract(const Sk4x<T>& other) const { return fVec - other.fVec; }
+template <typename T>
+Sk4x<T> Sk4x<T>::multiply(const Sk4x<T>& other) const { return fVec * other.fVec; }
+template <typename T>
+Sk4x<T> Sk4x<T>::  divide(const Sk4x<T>& other) const { return fVec / other.fVec; }
+
+template <typename T>
+Sk4x<T> Sk4x<T>::Min(const Sk4x<T>& a, const Sk4x<T>& b) {
+    return a.fVec < b.fVec ? a.fVec : b.fVec;  // This makes great SSE code (1 minps op)...
+}
+
+template <typename T>
+Sk4x<T> Sk4x<T>::Max(const Sk4x<T>& a, const Sk4x<T>& b) {
+    return a.fVec < b.fVec ? b.fVec : a.fVec;  // ...but this doesn't look so good (7 ops?).
+}
+
+// GCC 4.8 has a bug that leads it to segfault when presented with the obvious code for Shuffle:
+//   Sk4i::Vector mask = { m,a,s,k };
+//   return __builtin_shuffle(x.fVec, y.fVec, mask);
+//
+// This roundabout implementation via ShuffleImpl works around that bug,
+//   https://gcc.gnu.org/bugzilla/show_bug.cgi?id=57509
+
+template <>
+inline Sk4i::Vector Sk4i::ShuffleImpl(Sk4i::Vector x, Sk4i::Vector y, Sk4i::Vector mask) {
+    return __builtin_shuffle(x,y, mask);
+}
+
+template <>
+inline Sk4f::Vector Sk4f::ShuffleImpl(Sk4f::Vector x, Sk4f::Vector y, Sk4i::Vector mask) {
+    return __builtin_shuffle(x,y, mask);
+}
+
+template <typename T>
+template <int m, int a, int s, int k>
+Sk4x<T> Sk4x<T>::Shuffle(const Sk4x<T>& x, const Sk4x<T>& y) {
+    Sk4i::Vector mask = { m,a,s,k };
+    return ShuffleImpl(x.fVec, y.fVec, mask);
+}
+
+template <typename T>
+Sk4x<T> Sk4x<T>::zwxy() const { return Shuffle<2,3,0,1>(*this, *this); }
+
+template <typename T>
+Sk4x<T> Sk4x<T>::XYAB(const Sk4x& xyzw, const Sk4x& abcd) { return Shuffle<0,1,4,5>(xyzw, abcd); }
+
+template <typename T>
+Sk4x<T> Sk4x<T>::ZWCD(const Sk4x& xyzw, const Sk4x& abcd) { return Shuffle<2,3,6,7>(xyzw, abcd); }
+
+#endif // defined(SK4X_PRIVATE)
diff --git a/src/core/Sk4x_portable.h b/src/core/Sk4x_portable.h
new file mode 100644
index 0000000000..0515a9bd40
--- /dev/null
+++ b/src/core/Sk4x_portable.h
@@ -0,0 +1,134 @@
+// It is important _not_ to put header guards here.
+// This file will be intentionally included three times.
+
+#if defined(SK4X_PREAMBLE)
+
+#elif defined(SK4X_PRIVATE)
+    typedef T Vector[4];
+
+    Vector fVec;
+
+    template <int m, int a, int s, int k>
+    static Sk4x Shuffle(const Sk4x&, const Sk4x&);
+
+#else  // defined(SK4X_PRIVATE)
+
+template <typename T>
+Sk4x<T>::Sk4x() { }
+
+template <typename T>
+Sk4x<T>::Sk4x(T a, T b, T c, T d) { this->set(a,b,c,d); }
+
+template <typename T>
+Sk4x<T>::Sk4x(const T vals[4]) { this->set(vals[0], vals[1], vals[2], vals[3]); }
+
+template <typename T>
+Sk4x<T>::Sk4x(const Sk4x<T>& other) { *this = other; }
+
+template <typename T>
+Sk4x<T>& Sk4x<T>::operator=(const Sk4x<T>& other) {
+    this->set(other.fVec[0], other.fVec[1], other.fVec[2], other.fVec[3]);
+    return *this;
+}
+
+template <typename T>
+void Sk4x<T>::set(T a, T b, T c, T d) {
+    fVec[0] = a;
+    fVec[1] = b;
+    fVec[2] = c;
+    fVec[3] = d;
+}
+
+template <typename T>
+void Sk4x<T>::store(T vals[4]) const {
+    vals[0] = fVec[0];
+    vals[1] = fVec[1];
+    vals[2] = fVec[2];
+    vals[3] = fVec[3];
+}
+
+template <typename T>
+template <typename Dst> Dst Sk4x<T>::reinterpret() const {
+    return Dst(reinterpret_cast<const typename Dst::Vector*>(fVec));
+}
+
+template <typename T>
+template <typename Dst> Dst Sk4x<T>::cast() const {
+    return Dst(fVec[0], fVec[1], fVec[2], fVec[3]);
+}
+
+template <typename T>
+bool Sk4x<T>::allTrue() const { return fVec[0] & fVec[1] & fVec[2] & fVec[3]; }
+template <typename T>
+bool Sk4x<T>::anyTrue() const { return fVec[0] | fVec[1] | fVec[2] | fVec[3]; }
+
+template <typename T>
+Sk4x<T> Sk4x<T>::bitNot() const { return Sk4x(~fVec[0], ~fVec[1], ~fVec[2], ~fVec[3]); }
+
+#define BINOP(op) fVec[0] op other.fVec[0], \
+                  fVec[1] op other.fVec[1], \
+                  fVec[2] op other.fVec[2], \
+                  fVec[3] op other.fVec[3]
+
+template <typename T> Sk4x<T> Sk4x<T>::bitAnd(const Sk4x& other) const { return Sk4x(BINOP(&)); }
+template <typename T> Sk4x<T> Sk4x<T>::bitOr (const Sk4x& other) const { return Sk4x(BINOP(|)); }
+
+template <typename T>
+Sk4i Sk4x<T>::           equal(const Sk4x<T>& other) const { return Sk4i(BINOP(==)); }
+template <typename T>
+Sk4i Sk4x<T>::        notEqual(const Sk4x<T>& other) const { return Sk4i(BINOP(!=)); }
+template <typename T>
+Sk4i Sk4x<T>::        lessThan(const Sk4x<T>& other) const { return Sk4i(BINOP( <)); }
+template <typename T>
+Sk4i Sk4x<T>::     greaterThan(const Sk4x<T>& other) const { return Sk4i(BINOP( >)); }
+template <typename T>
+Sk4i Sk4x<T>::   lessThanEqual(const Sk4x<T>& other) const { return Sk4i(BINOP(<=)); }
+template <typename T>
+Sk4i Sk4x<T>::greaterThanEqual(const Sk4x<T>& other) const { return Sk4i(BINOP(>=)); }
+
+template <typename T>
+Sk4x<T> Sk4x<T>::     add(const Sk4x<T>& other) const { return Sk4x(BINOP(+)); }
+template <typename T>
+Sk4x<T> Sk4x<T>::subtract(const Sk4x<T>& other) const { return Sk4x(BINOP(-)); }
+template <typename T>
+Sk4x<T> Sk4x<T>::multiply(const Sk4x<T>& other) const { return Sk4x(BINOP(*)); }
+template <typename T>
+Sk4x<T> Sk4x<T>::  divide(const Sk4x<T>& other) const { return Sk4x(BINOP(/)); }
+
+#undef BINOP
+
+template <typename T>
+Sk4x<T> Sk4x<T>::Min(const Sk4x<T>& a, const Sk4x<T>& b) {
+    return Sk4x(SkTMin(a.fVec[0], b.fVec[0]),
+                SkTMin(a.fVec[1], b.fVec[1]),
+                SkTMin(a.fVec[2], b.fVec[2]),
+                SkTMin(a.fVec[3], b.fVec[3]));
+}
+
+template <typename T>
+Sk4x<T> Sk4x<T>::Max(const Sk4x<T>& a, const Sk4x<T>& b) {
+    return Sk4x(SkTMax(a.fVec[0], b.fVec[0]),
+                SkTMax(a.fVec[1], b.fVec[1]),
+                SkTMax(a.fVec[2], b.fVec[2]),
+                SkTMax(a.fVec[3], b.fVec[3]));
+}
+
+template <typename T>
+template <int m, int a, int s, int k>
+Sk4x<T> Sk4x<T>::Shuffle(const Sk4x<T>& x, const Sk4x<T>& y) {
+    return Sk4x(m < 4 ? x.fVec[m] : y.fVec[m-4],
+                a < 4 ? x.fVec[a] : y.fVec[a-4],
+                s < 4 ? x.fVec[s] : y.fVec[s-4],
+                k < 4 ? x.fVec[k] : y.fVec[k-4]);
+}
+
+template <typename T>
+Sk4x<T> Sk4x<T>::zwxy() const { return Shuffle<2,3,0,1>(*this, *this); }
+
+template <typename T>
+Sk4x<T> Sk4x<T>::XYAB(const Sk4x& xyzw, const Sk4x& abcd) { return Shuffle<0,1,4,5>(xyzw, abcd); }
+
+template <typename T>
+Sk4x<T> Sk4x<T>::ZWCD(const Sk4x& xyzw, const Sk4x& abcd) { return Shuffle<2,3,6,7>(xyzw, abcd); }
+
+#endif // defined(SK4X_PRIVATE)
diff --git a/src/core/SkTileGrid.cpp b/src/core/SkTileGrid.cpp
index 10782c4d6d..03b30f22ee 100644
--- a/src/core/SkTileGrid.cpp
+++ b/src/core/SkTileGrid.cpp
@@ -6,25 +6,34 @@
  */
 
 #include "SkTileGrid.h"
+#include "Sk4x.h"
 
 SkTileGrid::SkTileGrid(int xTiles, int yTiles, const SkTileGridFactory::TileGridInfo& info)
     : fXTiles(xTiles)
-    , fYTiles(yTiles)
-    , fInvWidth( SkScalarInvert(info.fTileInterval.width()))
-    , fInvHeight(SkScalarInvert(info.fTileInterval.height()))
-    , fMarginWidth (info.fMargin.fWidth +1)  // Margin is offset by 1 as a provision for AA and
-    , fMarginHeight(info.fMargin.fHeight+1)  // to cancel the outset applied by getClipDeviceBounds.
-    , fOffset(SkPoint::Make(info.fOffset.fX, info.fOffset.fY))
+    , fNumTiles(xTiles * yTiles)
     , fGridBounds(SkRect::MakeWH(xTiles * info.fTileInterval.width(),
                                  yTiles * info.fTileInterval.height()))
-    , fTiles(SkNEW_ARRAY(SkTDArray<unsigned>, xTiles * yTiles)) {}
+    , fMargin(-info.fMargin.fWidth  - 1,  // Outset margin by 1 as a provision for AA and to
+              -info.fMargin.fHeight - 1,  // cancel the outset applied by getClipDeviceBounds().
+              +info.fMargin.fWidth  + 1,
+              +info.fMargin.fHeight + 1)
+    , fOffset(info.fOffset.fX,
+              info.fOffset.fY,
+              info.fOffset.fX - SK_ScalarNearlyZero,  // We scrunch user-provided bounds in a little
+              info.fOffset.fY - SK_ScalarNearlyZero)  // to make right and bottom edges exclusive.
+    , fUserToGrid(SkScalarInvert(info.fTileInterval.width()),
+                  SkScalarInvert(info.fTileInterval.height()),
+                  SkScalarInvert(info.fTileInterval.width()),
+                  SkScalarInvert(info.fTileInterval.height()))
+    , fGridHigh(fXTiles - 1, yTiles - 1, fXTiles - 1, yTiles - 1)
+    , fTiles(SkNEW_ARRAY(SkTDArray<unsigned>, fNumTiles)) {}
 
 SkTileGrid::~SkTileGrid() {
     SkDELETE_ARRAY(fTiles);
 }
 
 void SkTileGrid::reserve(unsigned opCount) {
-    if (fXTiles * fYTiles == 0) {
+    if (fNumTiles == 0) {
         return;  // A tileless tile grid is nonsensical, but happens in at least cc_unittests.
     }
 
@@ -34,9 +43,9 @@ void SkTileGrid::reserve(unsigned opCount) {
 
     // If we take those observations and further assume the ops are distributed evenly
     // across the picture, we get this guess for number of ops per tile:
-    const int opsPerTileGuess = (2 * opCount) / (fXTiles * fYTiles);
+    const int opsPerTileGuess = (2 * opCount) / fNumTiles;
 
-    for (SkTDArray<unsigned>* tile = fTiles; tile != fTiles + (fXTiles * fYTiles); tile++) {
+    for (SkTDArray<unsigned>* tile = fTiles; tile != fTiles + fNumTiles; tile++) {
         tile->setReserve(opsPerTileGuess);
     }
 
@@ -45,39 +54,51 @@ void SkTileGrid::reserve(unsigned opCount) {
 }
 
 void SkTileGrid::flushDeferredInserts() {
-    for (SkTDArray<unsigned>* tile = fTiles; tile != fTiles + (fXTiles * fYTiles); tile++) {
+    for (SkTDArray<unsigned>* tile = fTiles; tile != fTiles + fNumTiles; tile++) {
         tile->shrinkToFit();
     }
 }
 
-// Adjustments to user-provided bounds common to both insert() and search().
-// Call this after making insert- or search- specific adjustments.
-void SkTileGrid::commonAdjust(SkRect* rect) const {
-    // Apply our offset.
-    rect->offset(fOffset);
+// Convert user-space bounds to grid tiles they cover (LT+RB both inclusive).
+// Out of bounds queries are clamped to the single nearest tile.
+void SkTileGrid::userToGrid(const Sk4f& user, SkIRect* out) const {
+    // Map from user coordinates to grid tile coordinates.
+    Sk4f grid = user.multiply(fUserToGrid);
 
-    // Scrunch the bounds in just a little to make the right and bottom edges
-    // exclusive.  We want bounds of exactly one tile to hit exactly one tile.
-    rect->fRight  -= SK_ScalarNearlyZero;
-    rect->fBottom -= SK_ScalarNearlyZero;
+    // Now that we're in grid coordinates, clamp to the grid bounds.
+    grid = Sk4f::Max(grid, Sk4f(0,0,0,0));
+    grid = Sk4f::Min(grid, fGridHigh);
+
+    // Truncate to integers.
+    grid.cast<Sk4i>().store(&out->fLeft);
+}
+
+// If the rect is inverted, sort it.
+static Sk4f sorted(const Sk4f& ltrb) {
+    // To sort:
+    //   left, right = minmax(left, right)
+    //   top, bottom = minmax(top, bottom)
+    Sk4f rblt = ltrb.zwxy(),
+         ltlt = Sk4f::Min(ltrb, rblt),  // Holds (2 copies of) new left and top.
+         rbrb = Sk4f::Max(ltrb, rblt),  // Holds (2 copies of) new right and bottom.
+         sort = Sk4f::XYAB(ltlt, rbrb);
+    return sort;
 }
 
-// Convert user-space bounds to grid tiles they cover (LT and RB both inclusive).
-void SkTileGrid::userToGrid(const SkRect& user, SkIRect* grid) const {
-    grid->fLeft   = SkPin32(user.left()   * fInvWidth , 0, fXTiles - 1);
-    grid->fTop    = SkPin32(user.top()    * fInvHeight, 0, fYTiles - 1);
-    grid->fRight  = SkPin32(user.right()  * fInvWidth , 0, fXTiles - 1);
-    grid->fBottom = SkPin32(user.bottom() * fInvHeight, 0, fYTiles - 1);
+// Does this rect intersect the grid?
+bool SkTileGrid::intersectsGrid(const Sk4f& ltrb) const {
+    SkRect bounds;
+    ltrb.store(&bounds.fLeft);
+    return SkRect::Intersects(bounds, fGridBounds);
+    // TODO: If we can get it fast enough, write intersect using Sk4f.
 }
 
 void SkTileGrid::insert(unsigned opIndex, const SkRect& originalBounds, bool) {
-    SkRect bounds = originalBounds;
-    bounds.outset(fMarginWidth, fMarginHeight);
-    this->commonAdjust(&bounds);
+    Sk4f bounds = Sk4f(&originalBounds.fLeft).add(fMargin).add(fOffset);
+    SkASSERT(sorted(bounds).equal(bounds).allTrue());
 
-    // TODO(mtklein): can we assert this instead to save an intersection in Release mode,
-    // or just allow out-of-bound insertions to insert anyway (clamped to nearest tile)?
-    if (!SkRect::Intersects(bounds, fGridBounds)) {
+    // TODO(mtklein): skip this check and just let out-of-bounds rects insert into nearest tile?
+    if (!this->intersectsGrid(bounds)) {
         return;
     }
 
@@ -103,20 +124,11 @@ void SkTileGrid::insert(unsigned opIndex, const SkRect& originalBounds, bool) {
 static const int kStackAllocationTileCount = 1024;
 
 void SkTileGrid::search(const SkRect& originalQuery, SkTDArray<unsigned>* results) const {
-    // The inset counteracts the outset that applied in 'insert', which optimizes
-    // for lookups of size 'tileInterval + 2 * margin' (aligned with the tile grid).
-    SkRect query = originalQuery;
-    query.inset(fMarginWidth, fMarginHeight);
-    this->commonAdjust(&query);
-
-    // The inset may have inverted the rectangle, so sort().
-    // TODO(mtklein): It looks like we only end up with inverted bounds in unit tests
-    // that make explicitly inverted queries, not from insetting.  If we can drop support for
-    // unsorted bounds (i.e. we don't see them outside unit tests), I think we can drop this.
-    query.sort();
-
-    // No intersection check.  We optimize for queries that are in bounds.
-    // We're safe anyway: userToGrid() will clamp out-of-bounds queries to nearest tile.
+    // The .subtract(fMargin) counteracts the .add(fMargin) applied in insert(),
+    // which optimizes for lookups of size tileInterval + 2 * margin (aligned with the tile grid).
+    // That .subtract(fMargin) may have inverted the rect, so we sort it.
+    Sk4f query = sorted(Sk4f(&originalQuery.fLeft).subtract(fMargin).add(fOffset));
+
     SkIRect grid;
     this->userToGrid(query, &grid);
 
diff --git a/src/core/SkTileGrid.h b/src/core/SkTileGrid.h
index fd7584fd9c..d556f8074b 100644
--- a/src/core/SkTileGrid.h
+++ b/src/core/SkTileGrid.h
@@ -8,6 +8,7 @@
 #ifndef SkTileGrid_DEFINED
 #define SkTileGrid_DEFINED
 
+#include "Sk4x.h"
 #include "SkBBHFactory.h"
 #include "SkBBoxHierarchy.h"
 
@@ -43,16 +44,16 @@ public:
     virtual void flushDeferredInserts() SK_OVERRIDE;
 
 private:
-    void commonAdjust(SkRect*) const;
-    void userToGrid(const SkRect&, SkIRect* grid) const;
+    void userToGrid(const Sk4f&, SkIRect*) const;
+    bool intersectsGrid(const Sk4f&) const;
 
-    const int fXTiles, fYTiles;
-    const SkScalar fInvWidth, fInvHeight;
-    const SkScalar fMarginWidth, fMarginHeight;
-    const SkPoint fOffset;
-    const SkRect  fGridBounds;
+    const int fXTiles,    // Number of tiles in a single row.
+              fNumTiles;  // Total number of tiles.
 
-    // (fXTiles * fYTiles) SkTDArrays, each listing ops overlapping that tile in order.
+    const SkRect fGridBounds;  // Only used for intersectsGrid().  Remove if that's removed.
+    const Sk4f fMargin, fOffset, fUserToGrid, fGridHigh;
+
+    // fNumTiles SkTDArrays, each listing ops overlapping that tile in order.
     SkTDArray<unsigned>* fTiles;
 
     typedef SkBBoxHierarchy INHERITED;