Sk4x_sse.h

tests/Sk4xTest.cpp

 #include "Test.h"
 #include "Sk4x.h"
 
 #define ASSERT_EQ(a, b) REPORTER_ASSERT(r, a.equal(b).allTrue())
 #define ASSERT_NE(a, b) REPORTER_ASSERT(r, a.notEqual(b).allTrue())
 
 DEF_TEST(Sk4x_Construction, r) {
     Sk4f uninitialized;
     Sk4f zero(0,0,0,0);
     Sk4f foo(1,2,3,4),
          bar(foo),
          baz = bar;
     ASSERT_EQ(foo, bar);
     ASSERT_EQ(bar, baz);
     ASSERT_EQ(baz, foo);
 }
 
+struct AlignedFloats {
+    Sk4f forces16ByteAlignment;   // On 64-bit machines, the stack starts 128-bit aligned,
+    float fs[5];                  // but not necessarily so on 32-bit.  Adding an Sk4f forces it.
+};
+
 DEF_TEST(Sk4x_LoadStore, r) {
+    AlignedFloats aligned;
     // fs will be 16-byte aligned, fs+1 not.
-    float fs[] = { 5,6,7,8,9 };
+    float* fs = aligned.fs;
+    for (int i = 0; i < 5; i++) {  // set to 5,6,7,8,9
+        fs[i] = float(i+5);
+    }
 
     Sk4f foo = Sk4f::Load(fs);
     Sk4f bar = Sk4f::LoadAligned(fs);
     ASSERT_EQ(foo, bar);
 
     foo = Sk4f::Load(fs+1);
     ASSERT_NE(foo, bar);
 
     foo.storeAligned(fs);
     bar.store(fs+1);
@@ skipping 73 matching lines @@
 }
 
 DEF_TEST(Sk4x_Swizzle, r) {
     ASSERT_EQ(Sk4f(3,4,1,2), Sk4f(1,2,3,4).zwxy());
     ASSERT_EQ(Sk4f(1,2,5,6), Sk4f::XYAB(Sk4f(1,2,3,4), Sk4f(5,6,7,8)));
     ASSERT_EQ(Sk4f(3,4,7,8), Sk4f::ZWCD(Sk4f(1,2,3,4), Sk4f(5,6,7,8)));
     ASSERT_EQ(Sk4i(3,4,1,2), Sk4i(1,2,3,4).zwxy());
     ASSERT_EQ(Sk4i(1,2,5,6), Sk4i::XYAB(Sk4i(1,2,3,4), Sk4i(5,6,7,8)));
     ASSERT_EQ(Sk4i(3,4,7,8), Sk4i::ZWCD(Sk4i(1,2,3,4), Sk4i(5,6,7,8)));
 }