[Spice-devel] [PATCH spice-common] codegen: Make the compiler work about better way to write unaligned memory

[Christophe Fergeau]

On Mon, Sep 11, 2017 at 11:13:04AM +0100, Frediano Ziglio wrote:
> Instead of assuming that the system can safely do unaligned access
> to memory use packed structures to allow the compiler generate
> best code possible.

It will generate the best code possible, but assuming all accesses are
unaligned I believe? Ie a perfectly fine 32 bit aligned field may still
be read using byte-sized accesses if the arch/cpu does not support
unaligned 32 bit reads?

The big missing part in this commit log is what this commit is doing in
human readable form...
This changes the current read_intX methods from:

#ifdef WORDS_BIGENDIAN
#define read_uint32(ptr) ((uint32_t)SPICE_BYTESWAP32(*((uint32_t *)(ptr))))
#define write_uint32(ptr, val) *(uint32_t *)(ptr) = SPICE_BYTESWAP32((uint32_t)val)
#else
#define read_uint32(ptr) (((uint32_unaligned_p)(ptr))->v)
#define read_uint32(ptr) (*((uint32_t *)(ptr)))
#endif

to:
#include <spice/start-packed.h>
typedef struct {
    uint32_t v;
} *uint32_unaligned_p;
include <spice/end-packed.h>

#ifdef WORDS_BIGENDIAN
#define read_uint32(ptr) ((uint32_t)SPICE_BYTESWAP32(((uint32_unaligned_p)(ptr))->v))
#define write_uint32(ptr, val) ((uint32_unaligned_p)(ptr))->v = SPICE_BYTESWAP32((uint32_t)val)
#else
#define read_uint32(ptr) (((uint32_unaligned_p)(ptr))->v)
#define write_uint32(ptr, val) (((uint32_unaligned_p)(ptr))->v) = val
#endif

The struct is missing SPICE_ATTR_PACKED, and I think I prefer typedef
struct {} uint32_unaligned_t;
and then that the casts become ((uint32_unaligned_t *)(ptr))->v

Christophe