On 29-11-2018 14:42, Jose Abreu wrote:
> On 29-11-2018 14:38, David Laight wrote:
>> From: Jose Abreu
>>> Sent: 29 November 2018 14:29
>>>
>>> Some ARC CPU's do not support unaligned loads/stores. Currently, generic
>>> implementation of reads{b/w/l}()/writes{b/w/l}() is being used with ARC.
>>> This can lead to misfunction of some drivers as generic functions do a
>>> plain dereference of a pointer that can be unaligned.
>>>
>>> Let's use {get/put}_unaligned() helper instead of plain dereference of
>>> pointer in order to fix this.
>> ...
>>> +#define __raw_readsx(t,f) \
>>> +static inline void __raw_reads##f(const volatile void __iomem *addr, \
>>> + void *buffer, unsigned int count) \
>>> +{ \
>>> + if (count) { \
>>> + const unsigned long bptr = (unsigned long)buffer; \
>>> + u##t *buf = buffer; \
>>> +\
>>> + do { \
>>> + u##t x = __raw_read##f(addr); \
>>> +\
>>> + /* Some ARC CPU's don't support unaligned accesses */ \
>>> + if (bptr % ((t) / 8)) { \
>>> + put_unaligned(x, buf++); \
>>> + } else { \
>>> + *buf++ = x; \
>>> + } \
>>> + } while (--count); \
>>> + } \
>>> +}
>> Does the compiler move the alignment test outside the loop?
>> You really want two copies of the loop body.
> Hmm, I would expect so because the if condition takes two const
> args ... I will try check that.
And it did optimize :)
Sample C Source:
--->8--
static noinline void test_readsl(char *buf, int len)
{
readsl(0xdeadbeef, buf, len);
}
--->8---
And the disassembly:
--->8---
00000e88 <test_readsl>:
e88: breq.dr1,0,eac <0xeac> /* if (count) */
e8c: and r2,r0,3
e90: mov_s lp_count,r1 /* r1 = count */
e92: brne r2,0,eb0 <0xeb0> /* if (bptr % ((t) / 8)) */
e96: sub r0,r0,4
e9a: nop_s
e9c: lp eac <0xeac> /* first loop */
ea0: ld r2,[0xdeadbeef]
ea8: st.a r2,[r0,4]
eac: j_s [blink]
eae: nop_s
eb0: lp ed6 <0xed6> /* second loop */
eb4: ld r2,[0xdeadbeef]
ebc: lsr r5,r2,8
ec0: lsr r4,r2,16
ec4: lsr r3,r2,24
ec8: stb_s r2,[r0,0]
eca: stb r5,[r0,1]
ece: stb r4,[r0,2]
ed2: stb_s r3,[r0,3]
ed4: add_s r0,r0,4
ed6: j_s [blink]
--->8---
See how the if condition added in this version is checked in
<test_readsl+0xe92> and then it takes two different loops.
Thanks and Best Regards,
Jose Miguel Abreu
>
> Thanks and Best Regards,
> Jose Miguel Abreu
>
>> David
>>
>> -
>> Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK
>> Registration No: 1397386 (Wales)
>>
_______________________________________________
linux-snps-arc mailing list
linux-snps-arc@xxxxxxxxxxxxxxxxxxx
http://lists.infradead.org/mailman/listinfo/linux-snps-arc
