ARM: 8341/1: io: Unpessimize relaxed io accessors
commit 195bbcac2e5c12f7fb ("ARM: 7500/1: io: avoid writeback addressing
modes for __raw_ accessors") disables writeback addressing modes for
raw i/o. However, the "+Q" output constraint forces the compiler to
disable load hoist optimizations (because the output constraint informs the
compiler of memory stores which the compiler assumes may alias other memory).
Since the relaxed accessors only guarantee ordering wrt i/o accesses to the
same device and not to main memory, there's never a possibility of an accessor
invalidating a hoisted load (because only non-i/o loads would have been hoisted).
The effect is especially noticable with complex address inputs in loops.
For example, the following code:
#include <linux/kernel.h>
#include <linux/io.h>
static const int *remap;
void wr_loop(void __iomem *base, int c, int val)
{
int i;
for (i = 0; i < c; i++)
writew_relaxed(val, base + remap[c >> 2]);
}
generates
current master | this patch
0: e3510000 cmp r1, #0 | 0: e3510000 cmp r1, #0
4: d12fff1e bxle lr | 4: d12fff1e bxle lr
8: e3003000 movw r3, #0 | 8: e3c1c003 bic ip, r1, #3
c: e3403000 movt r3, #0 | c: e6ff2072 uxth r2, r2
10: e92d4010 push {r4, lr} | 10: e3a03000 mov r3, #0
14: e6ff2072 uxth r2, r2 | 14: e59cc000 ldr ip, [ip]
18: e3c14003 bic r4, r1, #3 | 18: e080000c add r0, r0, ip
1c: e593e000 ldr lr, [r3] |
20: e3a03000 mov r3, #0 | 1c: e1c020b0 strh r2, [r0]
| 20: e2833001 add r3, r3, #1
24: e79ec004 ldr ip, [lr, r4] | 24: e1530001 cmp r3, r1
28: e080c00c add ip, r0, ip | 28: 1afffffb bne 1c
2c: e1cc20b0 strh r2, [ip] | 2c: e12fff1e bx lr
30: e2833001 add r3, r3, #1 |
34: e1530001 cmp r3, r1 |
38: 1afffff9 bne 24 |
|
3c: e8bd8010 pop {r4, pc} |
Acked-by: Will Deacon <will.deacon@arm.com>
Signed-off-by: Peter Hurley <peter@hurleysoftware.com>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
1 file changed