[PATCH v5 5/6] lib/vdso: Allow architectures to override the ns shift operation
Christophe Leroy
christophe.leroy at c-s.fr
Fri Jan 31 03:08:39 AEDT 2020
On powerpc/32, GCC (8.1) generates pretty bad code for the
ns >>= vd->shift operation taking into account that the
shift is always <= 32 and the upper part of the result is
likely to be nul. GCC makes reversed assumptions considering
the shift to be likely >= 32 and the upper part to be like not nul.
unsigned long long shift(unsigned long long x, unsigned char s)
{
return x >> s;
}
results in:
00000018 <shift>:
18: 35 25 ff e0 addic. r9,r5,-32
1c: 41 80 00 10 blt 2c <shift+0x14>
20: 7c 64 4c 30 srw r4,r3,r9
24: 38 60 00 00 li r3,0
28: 4e 80 00 20 blr
2c: 54 69 08 3c rlwinm r9,r3,1,0,30
30: 21 45 00 1f subfic r10,r5,31
34: 7c 84 2c 30 srw r4,r4,r5
38: 7d 29 50 30 slw r9,r9,r10
3c: 7c 63 2c 30 srw r3,r3,r5
40: 7d 24 23 78 or r4,r9,r4
44: 4e 80 00 20 blr
Even when forcing the shift with an &= 31, it still considers
the shift as likely >= 32.
Define a vdso_shift_ns() macro that can be overriden by
architectures.
Signed-off-by: Christophe Leroy <christophe.leroy at c-s.fr>
---
lib/vdso/gettimeofday.c | 9 ++++++++-
1 file changed, 8 insertions(+), 1 deletion(-)
diff --git a/lib/vdso/gettimeofday.c b/lib/vdso/gettimeofday.c
index 0b4b71880f22..a5ee53d1ecee 100644
--- a/lib/vdso/gettimeofday.c
+++ b/lib/vdso/gettimeofday.c
@@ -38,6 +38,13 @@ u64 vdso_calc_delta(u64 cycles, u64 last, u64 mask, u32 mult)
}
#endif
+#ifndef vdso_shift_ns
+static __always_inline u64 vdso_shift_ns(u64 ns, unsigned long shift)
+{
+ return ns >> shift;
+}
+#endif
+
#ifdef CONFIG_TIME_NS
static int do_hres_timens(const struct vdso_data *vdns, clockid_t clk,
struct __kernel_timespec *ts)
@@ -134,7 +141,7 @@ static __always_inline int do_hres(const struct vdso_data *vd, clockid_t clk,
#endif
ns += vdso_calc_delta(cycles, last, vd->mask, vd->mult);
- ns >>= vd->shift;
+ ns = vdso_shift_ns(ns, vd->shift);
sec = vdso_ts->sec;
} while (unlikely(vdso_read_retry(vd, seq)));
--
2.25.0
More information about the Linuxppc-dev
mailing list