On 24/02/2020 19:02, Jens Axboe wrote:
>> Usually doesn't work because of such possible "hackier assignments".
>> Ok, I have to go and experiment a bit. Anyway, it probably generates a lot of
>> useless stuff, e.g. for req->ctx
> 
> Tried this, and it generates the same code...

Maybe it wasn't able to optimise in the first place

E.g. for the following code any compiler generates 2 reads (thanks godbolt).

extern void foo(int);
int bar(const int *v)
{
    foo(*v);
    return *v;
}

> 
> diff --git a/fs/io_uring.c b/fs/io_uring.c
> index ba8d4e2d9f99..8de5863aa749 100644
> --- a/fs/io_uring.c
> +++ b/fs/io_uring.c
> @@ -598,7 +598,7 @@ struct io_kiocb {
>  
>  	struct io_async_ctx		*io;
>  	bool				needs_fixed_file;
> -	u8				opcode;
> +	const u8			opcode;
>  
>  	struct io_ring_ctx	*ctx;
>  	struct list_head	list;
> @@ -5427,6 +5427,8 @@ static bool io_get_sqring(struct io_ring_ctx *ctx, struct io_kiocb *req,
>  	 */
>  	head = READ_ONCE(sq_array[ctx->cached_sq_head & ctx->sq_mask]);
>  	if (likely(head < ctx->sq_entries)) {
> +		u8 *op;
> +
>  		/*
>  		 * All io need record the previous position, if LINK vs DARIN,
>  		 * it can be used to mark the position of the first IO in the
> @@ -5434,7 +5436,8 @@ static bool io_get_sqring(struct io_ring_ctx *ctx, struct io_kiocb *req,
>  		 */
>  		req->sequence = ctx->cached_sq_head;
>  		*sqe_ptr = &ctx->sq_sqes[head];
> -		req->opcode = READ_ONCE((*sqe_ptr)->opcode);
> +		op = (void *) req + offsetof(struct io_kiocb, opcode);
> +		*op = READ_ONCE((*sqe_ptr)->opcode);
>  		req->user_data = READ_ONCE((*sqe_ptr)->user_data);
>  		ctx->cached_sq_head++;
>  		return true;
> 

-- 
Pavel Begunkov