Move Sk4px Xfermode code to a header so we can use it twice.

  - Once in SkXfermode as usual to pick up compile-time SSE and NEON
  - Once in SkXfermode_arm_neon to pick up run-time NEON

This allows us to start cleaning up SkXfermode_arm_neon as we've done
for SkXfermode_SSE2.  I'm saving this catharsis for a day when I need it.

The Sk4px xfermodes are generally faster than the existing NEON procs,
so this should also have the side effect of a perf win there.

This means our new Plus-AA code works for runtime NEON too.
BUG=skia:3852

Review URL: https://codereview.chromium.org/1150313003

3 files changed, 167 insertions, 146 deletions

diff --git a/src/core/Sk4pxXfermode.h b/src/core/Sk4pxXfermode.h
new file mode 100644
index 0000000000..ff26436134
--- /dev/null
+++ b/src/core/Sk4pxXfermode.h
@@ -0,0 +1,156 @@
+/*
+ * Copyright 2015 Google Inc.
+ *
+ * Use of this source code is governed by a BSD-style license that can be
+ * found in the LICENSE file.
+ */
+
+#ifndef Sk4pxXfermode_DEFINED
+#define Sk4pxXfermode_DEFINED
+
+#include "Sk4px.h"
+
+// This file is possibly included into multiple .cpp files.
+// Each gets its own independent instantiation by wrapping in an anonymous namespace.
+namespace {
+
+#define XFERMODE(Name)                                                    \
+    struct Name {                                                         \
+        static Sk4px Xfer(const Sk4px&, const Sk4px&);                    \
+        static const SkXfermode::Mode kMode = SkXfermode::k##Name##_Mode; \
+    };                                                                    \
+    inline Sk4px Name::Xfer(const Sk4px& s, const Sk4px& d)
+
+XFERMODE(Clear) { return Sk4px((SkPMColor)0); }
+XFERMODE(Src)   { return s; }
+XFERMODE(Dst)   { return d; }
+XFERMODE(SrcIn)   { return     s.fastMulDiv255Round(d.alphas()      ); }
+XFERMODE(SrcOut)  { return     s.fastMulDiv255Round(d.alphas().inv()); }
+XFERMODE(SrcOver) { return s + d.fastMulDiv255Round(s.alphas().inv()); }
+XFERMODE(DstIn)   { return SrcIn  ::Xfer(d,s); }
+XFERMODE(DstOut)  { return SrcOut ::Xfer(d,s); }
+XFERMODE(DstOver) { return SrcOver::Xfer(d,s); }
+
+// [ S * Da + (1 - Sa) * D]
+XFERMODE(SrcATop) {
+    return Sk4px::Wide(s.mulWiden(d.alphas()) + d.mulWiden(s.alphas().inv()))
+        .div255RoundNarrow();
+}
+XFERMODE(DstATop) { return SrcATop::Xfer(d,s); }
+//[ S * (1 - Da) + (1 - Sa) * D ]
+XFERMODE(Xor) {
+    return Sk4px::Wide(s.mulWiden(d.alphas().inv()) + d.mulWiden(s.alphas().inv()))
+        .div255RoundNarrow();
+}
+// [S + D ]
+XFERMODE(Plus) { return s.saturatedAdd(d); }
+// [S * D ]
+XFERMODE(Modulate) { return s.fastMulDiv255Round(d); }
+// [S + D - S * D]
+XFERMODE(Screen) {
+    // Doing the math as S + (1-S)*D or S + (D - S*D) means the add and subtract can be done
+    // in 8-bit space without overflow.  S + (1-S)*D is a touch faster because inv() is cheap.
+    return s + d.fastMulDiv255Round(s.inv());
+}
+XFERMODE(Multiply) {
+    return Sk4px::Wide(s.mulWiden(d.alphas().inv()) +
+            d.mulWiden(s.alphas().inv()) +
+            s.mulWiden(d))
+        .div255RoundNarrow();
+}
+// [ Sa + Da - Sa*Da, Sc + Dc - 2*min(Sc*Da, Dc*Sa) ]  (And notice Sa*Da == min(Sa*Da, Da*Sa).)
+XFERMODE(Difference) {
+    auto m = Sk4px::Wide(Sk16h::Min(s.mulWiden(d.alphas()), d.mulWiden(s.alphas())))
+        .div255RoundNarrow();
+    // There's no chance of underflow, and if we subtract m before adding s+d, no overflow.
+    return (s - m) + (d - m.zeroAlphas());
+}
+// [ Sa + Da - Sa*Da, Sc + Dc - 2*Sc*Dc ]
+XFERMODE(Exclusion) {
+    auto p = s.fastMulDiv255Round(d);
+    // There's no chance of underflow, and if we subtract p before adding src+dst, no overflow.
+    return (s - p) + (d - p.zeroAlphas());
+}
+
+#undef XFERMODE
+
+// A reasonable fallback mode for doing AA is to simply apply the transfermode first,
+// then linearly interpolate the AA.
+template <typename Mode>
+static Sk4px xfer_aa(const Sk4px& s, const Sk4px& d, const Sk16b& aa) {
+    Sk4px noAA = Mode::Xfer(s, d);
+    return Sk4px::Wide(noAA.mulWiden(aa) + d.mulWiden(Sk4px(aa).inv()))
+        .div255RoundNarrow();
+}
+
+// For some transfermodes we specialize AA, either for correctness or performance.
+#ifndef SK_NO_SPECIALIZED_AA_XFERMODES
+    #define XFERMODE_AA(Name) \
+        template <> Sk4px xfer_aa<Name>(const Sk4px& s, const Sk4px& d, const Sk16b& aa)
+
+    // Plus' clamp needs to happen after AA.  skia:3852
+    XFERMODE_AA(Plus) {  // [ clamp(D + AA*S) ]
+        // We implement this as D + Min(S*AA, (1-D)) to fit the arguments to Min in 16 bits.
+        return d +
+            Sk4px::Wide(Sk16h::Min(s.mulWiden(aa), d.inv().mul255Widen())).div255RoundNarrow();
+    }
+
+    #undef XFERMODE_AA
+#endif
+
+template <typename ProcType>
+class SkT4pxXfermode : public SkProcCoeffXfermode {
+public:
+    static SkProcCoeffXfermode* Create(const ProcCoeff& rec) {
+        return SkNEW_ARGS(SkT4pxXfermode, (rec));
+    }
+
+    void xfer32(SkPMColor dst[], const SkPMColor src[], int n, const SkAlpha aa[]) const override {
+        if (NULL == aa) {
+            Sk4px::MapDstSrc(n, dst, src, [&](const Sk4px& dst4, const Sk4px& src4) {
+                return ProcType::Xfer(src4, dst4);
+            });
+        } else {
+            Sk4px::MapDstSrcAlpha(n, dst, src, aa,
+                    [&](const Sk4px& dst4, const Sk4px& src4, const Sk16b& alpha) {
+                return xfer_aa<ProcType>(src4, dst4, alpha);
+            });
+        }
+    }
+
+private:
+    SkT4pxXfermode(const ProcCoeff& rec) : SkProcCoeffXfermode(rec, ProcType::kMode) {}
+
+    typedef SkProcCoeffXfermode INHERITED;
+};
+
+static SkProcCoeffXfermode* SkCreate4pxXfermode(const ProcCoeff& rec, SkXfermode::Mode mode) {
+#if !defined(SK_CPU_ARM32) || defined(SK_ARM_HAS_NEON)
+    switch (mode) {
+        case SkXfermode::kClear_Mode:      return SkT4pxXfermode<Clear>::Create(rec);
+        case SkXfermode::kSrc_Mode:        return SkT4pxXfermode<Src>::Create(rec);
+        case SkXfermode::kDst_Mode:        return SkT4pxXfermode<Dst>::Create(rec);
+        case SkXfermode::kSrcOver_Mode:    return SkT4pxXfermode<SrcOver>::Create(rec);
+        case SkXfermode::kDstOver_Mode:    return SkT4pxXfermode<DstOver>::Create(rec);
+        case SkXfermode::kSrcIn_Mode:      return SkT4pxXfermode<SrcIn>::Create(rec);
+        case SkXfermode::kDstIn_Mode:      return SkT4pxXfermode<DstIn>::Create(rec);
+        case SkXfermode::kSrcOut_Mode:     return SkT4pxXfermode<SrcOut>::Create(rec);
+        case SkXfermode::kDstOut_Mode:     return SkT4pxXfermode<DstOut>::Create(rec);
+        case SkXfermode::kSrcATop_Mode:    return SkT4pxXfermode<SrcATop>::Create(rec);
+        case SkXfermode::kDstATop_Mode:    return SkT4pxXfermode<DstATop>::Create(rec);
+        case SkXfermode::kXor_Mode:        return SkT4pxXfermode<Xor>::Create(rec);
+        case SkXfermode::kPlus_Mode:       return SkT4pxXfermode<Plus>::Create(rec);
+        case SkXfermode::kModulate_Mode:   return SkT4pxXfermode<Modulate>::Create(rec);
+        case SkXfermode::kScreen_Mode:     return SkT4pxXfermode<Screen>::Create(rec);
+        case SkXfermode::kMultiply_Mode:   return SkT4pxXfermode<Multiply>::Create(rec);
+        case SkXfermode::kDifference_Mode: return SkT4pxXfermode<Difference>::Create(rec);
+        case SkXfermode::kExclusion_Mode:  return SkT4pxXfermode<Exclusion>::Create(rec);
+        default: break;
+    }
+#endif
+    return nullptr;
+}
+
+} // namespace
+
+#endif//Sk4pxXfermode_DEFINED
diff --git a/src/core/SkXfermode.cpp b/src/core/SkXfermode.cpp
index 99f83f9f3d..d8ca8f1e6e 100644
--- a/src/core/SkXfermode.cpp
+++ b/src/core/SkXfermode.cpp
@@ -9,7 +9,7 @@
 #include "SkXfermode.h"
 #include "SkXfermode_opts_SSE2.h"
 #include "SkXfermode_proccoeff.h"
-#include "Sk4px.h"
+#include "Sk4pxXfermode.h"
 #include "SkColorPriv.h"
 #include "SkLazyPtr.h"
 #include "SkMathPriv.h"
@@ -19,14 +19,6 @@
 #include "SkUtilsArm.h"
 #include "SkWriteBuffer.h"
 
-#if SK_CPU_X86 && SK_CPU_SSE_LEVEL < SK_CPU_SSE_LEVEL_SSE2
-    #warning "SkXfermode will be much faster if you compile with support for SSE2."
-#endif
-
-#if SK_CPU_X86 || defined(SK_ARM_HAS_NEON)
-    #define SK_USE_4PX_XFERMODES
-#endif
-
 #if !SK_ARM_NEON_IS_NONE
     #include "SkXfermode_opts_arm_neon.h"
 #endif
@@ -1181,116 +1173,6 @@ void SkDstInXfermode::toString(SkString* str) const {
 
 ///////////////////////////////////////////////////////////////////////////////
 
-#define XFERMODE(Name)                                                    \
-    struct Name {                                                         \
-        static Sk4px Xfer(const Sk4px&, const Sk4px&);                    \
-        static const SkXfermode::Mode kMode = SkXfermode::k##Name##_Mode; \
-    };                                                                    \
-    inline Sk4px Name::Xfer(const Sk4px& s, const Sk4px& d)
-
-XFERMODE(Clear) { return Sk4px((SkPMColor)0); }
-XFERMODE(Src)   { return s; }
-XFERMODE(Dst)   { return d; }
-XFERMODE(SrcIn)   { return     s.fastMulDiv255Round(d.alphas()      ); }
-XFERMODE(SrcOut)  { return     s.fastMulDiv255Round(d.alphas().inv()); }
-XFERMODE(SrcOver) { return s + d.fastMulDiv255Round(s.alphas().inv()); }
-XFERMODE(DstIn)   { return SrcIn  ::Xfer(d,s); }
-XFERMODE(DstOut)  { return SrcOut ::Xfer(d,s); }
-XFERMODE(DstOver) { return SrcOver::Xfer(d,s); }
-
-// [ S * Da + (1 - Sa) * D]
-XFERMODE(SrcATop) {
-    return Sk4px::Wide(s.mulWiden(d.alphas()) + d.mulWiden(s.alphas().inv()))
-        .div255RoundNarrow();
-}
-XFERMODE(DstATop) { return SrcATop::Xfer(d,s); }
-//[ S * (1 - Da) + (1 - Sa) * D ]
-XFERMODE(Xor) {
-    return Sk4px::Wide(s.mulWiden(d.alphas().inv()) + d.mulWiden(s.alphas().inv()))
-        .div255RoundNarrow();
-}
-// [S + D ]
-XFERMODE(Plus) { return s.saturatedAdd(d); }
-// [S * D ]
-XFERMODE(Modulate) { return s.fastMulDiv255Round(d); }
-// [S + D - S * D]
-XFERMODE(Screen) {
-    // Doing the math as S + (1-S)*D or S + (D - S*D) means the add and subtract can be done
-    // in 8-bit space without overflow.  S + (1-S)*D is a touch faster because inv() is cheap.
-    return s + d.fastMulDiv255Round(s.inv());
-}
-XFERMODE(Multiply) {
-    return Sk4px::Wide(s.mulWiden(d.alphas().inv()) +
-            d.mulWiden(s.alphas().inv()) +
-            s.mulWiden(d))
-        .div255RoundNarrow();
-}
-// [ Sa + Da - Sa*Da, Sc + Dc - 2*min(Sc*Da, Dc*Sa) ]  (And notice Sa*Da == min(Sa*Da, Da*Sa).)
-XFERMODE(Difference) {
-    auto m = Sk4px::Wide(Sk16h::Min(s.mulWiden(d.alphas()), d.mulWiden(s.alphas())))
-        .div255RoundNarrow();
-    // There's no chance of underflow, and if we subtract m before adding s+d, no overflow.
-    return (s - m) + (d - m.zeroAlphas());
-}
-// [ Sa + Da - Sa*Da, Sc + Dc - 2*Sc*Dc ]
-XFERMODE(Exclusion) {
-    auto p = s.fastMulDiv255Round(d);
-    // There's no chance of underflow, and if we subtract p before adding src+dst, no overflow.
-    return (s - p) + (d - p.zeroAlphas());
-}
-
-#undef XFERMODE
-
-// A reasonable fallback mode for doing AA is to simply apply the transfermode first,
-// then linearly interpolate the AA.
-template <typename Mode>
-static Sk4px xfer_aa(const Sk4px& s, const Sk4px& d, const Sk16b& aa) {
-    Sk4px noAA = Mode::Xfer(s, d);
-    return Sk4px::Wide(noAA.mulWiden(aa) + d.mulWiden(Sk4px(aa).inv()))
-        .div255RoundNarrow();
-}
-
-// For some transfermodes we specialize AA, either for correctness or performance.
-#ifndef SK_NO_SPECIALIZED_AA_XFERMODES
-    #define XFERMODE_AA(Name) \
-        template <> Sk4px xfer_aa<Name>(const Sk4px& s, const Sk4px& d, const Sk16b& aa)
-
-    // Plus' clamp needs to happen after AA.  skia:3852
-    XFERMODE_AA(Plus) {  // [ clamp(D + AA*S) ]
-        // We implement this as D + Min(S*AA, (1-D)) to fit the arguments to Min in 16 bits.
-        return d +
-            Sk4px::Wide(Sk16h::Min(s.mulWiden(aa), d.inv().mul255Widen())).div255RoundNarrow();
-    }
-
-    #undef XFERMODE_AA
-#endif
-
-template <typename ProcType>
-class SkT4pxXfermode : public SkProcCoeffXfermode {
-public:
-    static SkXfermode* Create(const ProcCoeff& rec) {
-        return SkNEW_ARGS(SkT4pxXfermode, (rec));
-    }
-
-    void xfer32(SkPMColor dst[], const SkPMColor src[], int n, const SkAlpha aa[]) const override {
-        if (NULL == aa) {
-            Sk4px::MapDstSrc(n, dst, src, [&](const Sk4px& dst4, const Sk4px& src4) {
-                return ProcType::Xfer(src4, dst4);
-            });
-        } else {
-            Sk4px::MapDstSrcAlpha(n, dst, src, aa,
-                    [&](const Sk4px& dst4, const Sk4px& src4, const Sk16b& alpha) {
-                return xfer_aa<ProcType>(src4, dst4, alpha);
-            });
-        }
-    }
-
-private:
-    SkT4pxXfermode(const ProcCoeff& rec) : SkProcCoeffXfermode(rec, ProcType::kMode) {}
-
-    typedef SkProcCoeffXfermode INHERITED;
-};
-
 ///////////////////////////////////////////////////////////////////////////////
 
 class SkDstOutXfermode : public SkProcCoeffXfermode {
@@ -1351,29 +1233,9 @@ SkXfermode* create_mode(int iMode) {
         rec.fProc = pp;
     }
 
-#if defined(SK_USE_4PX_XFERMODES)
-    switch (mode) {
-        case SkXfermode::kClear_Mode:      return SkT4pxXfermode<Clear>::Create(rec);
-        case SkXfermode::kSrc_Mode:        return SkT4pxXfermode<Src>::Create(rec);
-        case SkXfermode::kDst_Mode:        return SkT4pxXfermode<Dst>::Create(rec);
-        case SkXfermode::kSrcOver_Mode:    return SkT4pxXfermode<SrcOver>::Create(rec);
-        case SkXfermode::kDstOver_Mode:    return SkT4pxXfermode<DstOver>::Create(rec);
-        case SkXfermode::kSrcIn_Mode:      return SkT4pxXfermode<SrcIn>::Create(rec);
-        case SkXfermode::kDstIn_Mode:      return SkT4pxXfermode<DstIn>::Create(rec);
-        case SkXfermode::kSrcOut_Mode:     return SkT4pxXfermode<SrcOut>::Create(rec);
-        case SkXfermode::kDstOut_Mode:     return SkT4pxXfermode<DstOut>::Create(rec);
-        case SkXfermode::kSrcATop_Mode:    return SkT4pxXfermode<SrcATop>::Create(rec);
-        case SkXfermode::kDstATop_Mode:    return SkT4pxXfermode<DstATop>::Create(rec);
-        case SkXfermode::kXor_Mode:        return SkT4pxXfermode<Xor>::Create(rec);
-        case SkXfermode::kPlus_Mode:       return SkT4pxXfermode<Plus>::Create(rec);
-        case SkXfermode::kModulate_Mode:   return SkT4pxXfermode<Modulate>::Create(rec);
-        case SkXfermode::kScreen_Mode:     return SkT4pxXfermode<Screen>::Create(rec);
-        case SkXfermode::kMultiply_Mode:   return SkT4pxXfermode<Multiply>::Create(rec);
-        case SkXfermode::kDifference_Mode: return SkT4pxXfermode<Difference>::Create(rec);
-        case SkXfermode::kExclusion_Mode:  return SkT4pxXfermode<Exclusion>::Create(rec);
-        default: break;
+    if (auto xfermode = SkCreate4pxXfermode(rec, mode)) {
+        return xfermode;
     }
-#endif
 
     SkXfermode* xfer = NULL;
 
diff --git a/src/opts/SkXfermode_opts_arm_neon.cpp b/src/opts/SkXfermode_opts_arm_neon.cpp
index 12c2f57d4f..1759429c57 100644
--- a/src/opts/SkXfermode_opts_arm_neon.cpp
+++ b/src/opts/SkXfermode_opts_arm_neon.cpp
@@ -1,3 +1,5 @@
+// Copyright 2013 unknown
+
 #include "SkXfermode.h"
 #include "SkXfermode_proccoeff.h"
 #include "SkColorPriv.h"
@@ -5,6 +7,7 @@
 #include <arm_neon.h>
 #include "SkColor_opts_neon.h"
 #include "SkXfermode_opts_arm_neon.h"
+#include "Sk4pxXfermode.h"
 
 #define SkAlphaMulAlpha(a, b)   SkMulDiv255Round(a, b)
 
@@ -1010,11 +1013,11 @@ SK_COMPILE_ASSERT(
 
 SkProcCoeffXfermode* SkPlatformXfermodeFactory_impl_neon(const ProcCoeff& rec,
                                                          SkXfermode::Mode mode) {
-
-    void* procSIMD = reinterpret_cast<void*>(gNEONXfermodeProcs[mode]);
-
-    if (procSIMD != NULL) {
-        return SkNEW_ARGS(SkNEONProcCoeffXfermode, (rec, mode, procSIMD));
+    if (auto xfermode = SkCreate4pxXfermode(rec, mode)) {
+        return xfermode;
+    }
+    if (auto proc = gNEONXfermodeProcs[mode]) {
+        return SkNEW_ARGS(SkNEONProcCoeffXfermode, (rec, mode, (void*)proc));
     }
     return NULL;
 }