amd64: align on 16 bytes instead of 8

Some instructions emitted by LLVM (like movaps) expect 16-byte
alignment, while the allocator assumed that 8-byte alignment is good
enough.

TODO: this issue came to light with LLVM optimizing a complex128 store
to the movaps instruction, which must be aligned. It looks like this
means that the <2 x double> IR type is actually 16-byte aligned instead
of 8-byte like a double. If this is the case, the alignment of complex
numbers needs to be updated in the whole compiler.

src/runtime/arch_amd64.go

@@ -5,7 +5,9 @@ const GOARCH = "amd64"
 // The bitness of the CPU (e.g. 8, 32, 64).
 const TargetBits = 64
 
-// Align on word boundary.
+// Align a pointer.
+// Note that some amd64 instructions (like movaps) expect 16-byte aligned
+// memory, thus the result must be 16-byte aligned.
 func align(ptr uintptr) uintptr {
-	return (ptr + 7) &^ 7
+	return (ptr + 15) &^ 15
 }