[RFC PATCH v4 10/11] lib: vdso: Allow arches to override the ns shift operation

[Christophe Leroy]

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
arches.

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 724b45c3e8ac..9ba92058cfd7 100644
--- a/lib/vdso/gettimeofday.c
+++ b/lib/vdso/gettimeofday.c
@@ -39,6 +39,13 @@ 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
+
 #ifndef __arch_vdso_hres_capable
 static inline bool __arch_vdso_hres_capable(void)
 {
@@ -148,7 +155,7 @@ static __always_inline int do_hres(const struct vdso_data *vd, clockid_t clk,
 		ns = vdso_ts->nsec;
 		last = vd->cycle_last;
 		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)));
 
-- 
2.13.3