On 2018-11-05 22:48, Bart Van Assche wrote:
> On Mon, 2018-11-05 at 13:13 -0800, Andrew Morton wrote:
>> On Mon, 5 Nov 2018 12:40:00 -0800 Bart Van Assche <bvanassche@xxxxxxx> wrote:
>>
>>> This patch suppresses the following sparse warning:
>>>
>>> ./include/linux/slab.h:332:43: warning: dubious: x & !y
>>>
>>> ...
>>>
>>> --- a/include/linux/slab.h
>>> +++ b/include/linux/slab.h
>>> @@ -329,7 +329,7 @@ static __always_inline enum kmalloc_cache_type kmalloc_type(gfp_t flags)
>>> * If an allocation is both __GFP_DMA and __GFP_RECLAIMABLE, return
>>> * KMALLOC_DMA and effectively ignore __GFP_RECLAIMABLE
>>> */
>>> - return type_dma + (is_reclaimable & !is_dma) * KMALLOC_RECLAIM;
>>> + return type_dma + is_reclaimable * !is_dma * KMALLOC_RECLAIM;
>>> }
>>>
>>> /*
>>
>> I suppose so.
>>
>> That function seems too clever for its own good :(. I wonder if these
>> branch-avoiding tricks are really worthwhile.
>
> From what I have seen in gcc disassembly it seems to me like gcc uses the
> cmov instruction to implement e.g. the ternary operator (?:). So I think none
> of the cleverness in kmalloc_type() is really necessary to avoid conditional
> branches. I think this function would become much more readable when using a
> switch statement or when rewriting it as follows (untested):
>
> static __always_inline enum kmalloc_cache_type kmalloc_type(gfp_t flags)
> {
> - int is_dma = 0;
> - int type_dma = 0;
> - int is_reclaimable;
> -
> -#ifdef CONFIG_ZONE_DMA
> - is_dma = !!(flags & __GFP_DMA);
> - type_dma = is_dma * KMALLOC_DMA;
> -#endif
> -
> - is_reclaimable = !!(flags & __GFP_RECLAIMABLE);
> -
> /*
> * If an allocation is both __GFP_DMA and __GFP_RECLAIMABLE, return
> * KMALLOC_DMA and effectively ignore __GFP_RECLAIMABLE
> */
> - return type_dma + (is_reclaimable & !is_dma) * KMALLOC_RECLAIM;
> + static const enum kmalloc_cache_type flags_to_type[2][2] = {
> + { 0, KMALLOC_RECLAIM },
> + { KMALLOC_DMA, KMALLOC_DMA },
> + };
> +#ifdef CONFIG_ZONE_DMA
> + bool is_dma = !!(flags & __GFP_DMA);
> +#endif
> + bool is_reclaimable = !!(flags & __GFP_RECLAIMABLE);
> +
> + return flags_to_type[is_dma][is_reclaimable];
> }
>
Won't that pessimize the cases where gfp is a constant to actually do
the table lookup, and add 16 bytes to every translation unit?
Another option is to add a fake KMALLOC_DMA_RECLAIM so the
kmalloc_caches[] array has size 4, then assign the same dma
kmalloc_cache pointer to [2][i] and [3][i] (so that costs perhaps a
dozen pointers in .data), and then just compute kmalloc_type() as
((flags & __GFP_RECLAIMABLE) >> someshift) | ((flags & __GFP_DMA) >>
someothershift).
Perhaps one could even shuffle the GFP flags so the two shifts are the same.
Rasmus
