make pm4f be RGBA always, not pmcolor order

src/core/SkPM4f.h

Index: src/core/SkPM4f.h
diff --git a/src/core/SkPM4f.h b/src/core/SkPM4f.h
index fb22783dcc44203a8c3931982f645e3de7dde956..e99ddccd8effac93898897df523042ac38a7ccba 100644
--- a/src/core/SkPM4f.h
+++ b/src/core/SkPM4f.h
@@ -9,29 +9,51 @@
 #define SkPM4f_DEFINED
 
 #include "SkColorPriv.h"
+#include "SkNx.h"
+
+static inline Sk4f swizzle_rb(const Sk4f& x) {
+    return SkNx_shuffle<2, 1, 0, 3>(x);
+}
+
+static inline Sk4f swizzle_rb_if_bgra(const Sk4f& x) {
+#ifdef SK_PMCOLOR_IS_BGRA
+    return swizzle_rb(x);
+#else
+    return x;
+#endif
+}
 
 /*
- *  The float values are 0...1 premultiplied
+ *  The float values are 0...1 premultiplied in RGBA order (regardless of SkPMColor order)
  */
 struct SkPM4f {
     enum {
-        A = SK_A32_SHIFT/8,
-        R = SK_R32_SHIFT/8,
-        G = SK_G32_SHIFT/8,
-        B = SK_B32_SHIFT/8,
+        R, G, B, A,
     };
     float fVec[4];
 
+    float r() const { return fVec[R]; }
+    float g() const { return fVec[G]; }
+    float b() const { return fVec[B]; }
     float a() const { return fVec[A]; }
 
-    SkColor4f unpremul() const;
-
+    static SkPM4f From4f(const Sk4f& x) {
+        SkPM4f pm;
+        x.store(pm.fVec);
+        return pm;
+    }
+    static SkPM4f FromF16(const uint16_t[4]);
     static SkPM4f FromPMColor(SkPMColor);
 
-    // half-float routines
+    Sk4f to4f() const { return Sk4f::Load(fVec); }
+    Sk4f to4f_rgba() const { return this->to4f(); }
+    Sk4f to4f_bgra() const { return swizzle_rb(this->to4f()); }
+    Sk4f to4f_pmorder() const { return swizzle_rb_if_bgra(this->to4f()); }
+
     void toF16(uint16_t[4]) const;
     uint64_t toF16() const; // 4 float16 values packed into uint64_t
-    static SkPM4f FromF16(const uint16_t[4]);
+
+    SkColor4f unpremul() const;
 
 #ifdef SK_DEBUG
     void assertIsUnit() const;
@@ -42,5 +64,4 @@ struct SkPM4f {
 
 typedef SkPM4f (*SkXfermodeProc4f)(const SkPM4f& src, const SkPM4f& dst);
 
-
 #endif