On Tue, Dec 31, 2019 at 12:41 AM Tom Rini trini@konsulko.com wrote:
On Tue, Dec 17, 2019 at 08:52:01PM +0100, Marek Vasut wrote:
On 12/17/19 8:32 PM, Heinrich Schuchardt wrote:
On 12/17/19 1:19 PM, Marek Vasut wrote:
On 12/17/19 12:59 PM, Heinrich Schuchardt wrote:
On 12/17/19 12:19 PM, Marek Vasut wrote:
On 12/17/19 12:14 PM, Heinrich Schuchardt wrote: > On 12/17/19 11:00 AM, Marek Vasut wrote: >> On 12/17/19 10:27 AM, Heinrich Schuchardt wrote: >>> With GCC 9.2.1 errors of type -Werror=address-of-packed-member occur >>> when >>> passing a member of a packed structure to le16_to_cpus() on a big >>> endian >>> system (e.g. P2041RDB_defconfig). >>> >>> Replace le16_to_cpus() by get_unaligned_le16(). Check >>> defined(__BIG_ENDIAN) >>> to avoid the introduction of unnecessary instructions on little >>> endian >>> systems as seen on aarch64. >> >> I would expect the compiler would optimize such stuff out ? >> Can we do without the ifdef ? > > When compiling qemu_arm64_defconfig without the #ifdef the GCC 9.2.1 > adds assembler instructions:
Why ?
I am not a GCC developer. I simply observed that GCC currently cannot optimize this away on its own. That is why I added the #ifdef.
Are we now adding workarounds instead of solving issues properly?
Identifying that bit operations like << 8 in __get_unaligned_le16() have zero effect is not an easy task when developing a compiler. You would have to follow the flow of every bit.
Maybe the fix is then to somehow optimize the get_unaligned_le16() to help the compiler ?
Inside get_unaligned_le16() it is not known that we will be reassigning to the same memory location. So I cannot imagine what to improve here.
You could invent new functions for in place byte swapping. But that would only move the #ifdef to a different place.
Isn't there already such a function in Linux ?
Also, why would you need the ifdef if the compiler would now know that the operation is noop ?
This particular patch looks to me exactly why we want to follow the Linux Kernel and disable this particular warning for GCC like we already do for LLVM.
Agree. I think we can follow Linux commit 6f303d60534c46aa1a239f29c321f95c83dda748