Skip to content

Commit 531d463

Browse files
zx2c4zx2c4
committed
go-patches: read nanotime in 1 instruction on 64bit
Signed-off-by: Jason A. Donenfeld <[email protected]>
1 parent 1d87c38 commit 531d463

3 files changed

+200
-6
lines changed

‎go-patches/0001-runtime-allow-builtin-write-function-to-be-redirecte.patch

Lines changed: 0 additions & 3 deletions
Original file line numberDiff line numberDiff line change
@@ -54,6 +54,3 @@ index 5a4ceaf43d..68c01805a5 100644
5454
+ }
5555
return write1(fd, p, n)
5656
}
57-
--
58-
2.33.0
59-

‎go-patches/0002-release-branch.go1.17-cmd-compile-speed-up-unsafe.Sl.patch

Lines changed: 0 additions & 3 deletions
Original file line numberDiff line numberDiff line change
@@ -326,6 +326,3 @@ index 4c940aa855..d613088827 100644
326326
- }()
327327
- f()
328328
-}
329-
--
330-
2.33.0
331-

‎go-patches/0003-runtime-on-windows-read-nanotime-with-one-instructio.patch

Lines changed: 200 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,200 @@
1+
From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
2+
From: "Jason A. Donenfeld" <[email protected]>
3+
Date: Wed, 3 Nov 2021 18:19:04 +0100
4+
Subject: [PATCH] runtime: on windows, read nanotime with one instruction or
5+
issue barrier
6+
7+
On 64-bit, this is more efficient, and on ARM64, this prevents the time
8+
from moving backwards due to the weaker memory model. On ARM32 due to
9+
the weaker memory model, we issue a memory barrier.
10+
11+
Updates #48072.
12+
13+
Change-Id: If4695716c3039d8af14e14808af217f5c99fc93a
14+
Reviewed-on: https://go-review.googlesource.com/c/go/+/361057
15+
Trust: Jason A. Donenfeld <[email protected]>
16+
Run-TryBot: Jason A. Donenfeld <[email protected]>
17+
TryBot-Result: Go Bot <[email protected]>
18+
Reviewed-by: Austin Clements <[email protected]>
19+
---
20+
src/runtime/sys_windows_amd64.s | 13 +++----------
21+
src/runtime/sys_windows_arm.s | 2 ++
22+
src/runtime/sys_windows_arm64.s | 10 +---------
23+
src/runtime/time_windows.h | 1 +
24+
src/runtime/time_windows_amd64.s | 19 +++----------------
25+
src/runtime/time_windows_arm.s | 4 ++++
26+
src/runtime/time_windows_arm64.s | 22 +++-------------------
27+
7 files changed, 17 insertions(+), 54 deletions(-)
28+
29+
diff --git a/src/runtime/sys_windows_amd64.s b/src/runtime/sys_windows_amd64.s
30+
index e7782846b2..64fa6791f4 100644
31+
--- a/src/runtime/sys_windows_amd64.s
32+
+++ b/src/runtime/sys_windows_amd64.s
33+
@@ -348,16 +348,9 @@ TEXT runtime·nanotime1(SB),NOSPLIT,$0-8
34+
CMPB runtime·useQPCTime(SB), $0
35+
JNE useQPC
36+
MOVQ $_INTERRUPT_TIME, DI
37+
-loop:
38+
- MOVL time_hi1(DI), AX
39+
- MOVL time_lo(DI), BX
40+
- MOVL time_hi2(DI), CX
41+
- CMPL AX, CX
42+
- JNE loop
43+
- SHLQ $32, CX
44+
- ORQ BX, CX
45+
- IMULQ $100, CX
46+
- MOVQ CX, ret+0(FP)
47+
+ MOVQ time_lo(DI), AX
48+
+ IMULQ $100, AX
49+
+ MOVQ AX, ret+0(FP)
50+
RET
51+
useQPC:
52+
JMP runtime·nanotimeQPC(SB)
53+
diff --git a/src/runtime/sys_windows_arm.s b/src/runtime/sys_windows_arm.s
54+
index 48f8c7dedf..d7ad244161 100644
55+
--- a/src/runtime/sys_windows_arm.s
56+
+++ b/src/runtime/sys_windows_arm.s
57+
@@ -350,7 +350,9 @@ TEXT runtime·nanotime1(SB),NOSPLIT|NOFRAME,$0-8
58+
MOVW $_INTERRUPT_TIME, R3
59+
loop:
60+
MOVW time_hi1(R3), R1
61+
+ DMB MB_ISH
62+
MOVW time_lo(R3), R0
63+
+ DMB MB_ISH
64+
MOVW time_hi2(R3), R2
65+
CMP R1, R2
66+
BNE loop
67+
diff --git a/src/runtime/sys_windows_arm64.s b/src/runtime/sys_windows_arm64.s
68+
index 7a2e11f5ae..183128dd09 100644
69+
--- a/src/runtime/sys_windows_arm64.s
70+
+++ b/src/runtime/sys_windows_arm64.s
71+
@@ -415,15 +415,7 @@ TEXT runtime·nanotime1(SB),NOSPLIT|NOFRAME,$0-8
72+
CMP $0, R0
73+
BNE useQPC
74+
MOVD $_INTERRUPT_TIME, R3
75+
-loop:
76+
- MOVWU time_hi1(R3), R1
77+
- MOVWU time_lo(R3), R0
78+
- MOVWU time_hi2(R3), R2
79+
- CMP R1, R2
80+
- BNE loop
81+
-
82+
- // wintime = R1:R0, multiply by 100
83+
- ORR R1<<32, R0
84+
+ MOVD time_lo(R3), R0
85+
MOVD $100, R1
86+
MUL R1, R0
87+
MOVD R0, ret+0(FP)
88+
diff --git a/src/runtime/time_windows.h b/src/runtime/time_windows.h
89+
index cd16fd163b..7c2e65c328 100644
90+
--- a/src/runtime/time_windows.h
91+
+++ b/src/runtime/time_windows.h
92+
@@ -9,6 +9,7 @@
93+
// http://web.archive.org/web/20210411000829/https://wrkhpi.wordpress.com/2007/08/09/getting-os-information-the-kuser_shared_data-structure/
94+
95+
// Must read hi1, then lo, then hi2. The snapshot is valid if hi1 == hi2.
96+
+// Or, on 64-bit, just read lo:hi1 all at once atomically.
97+
#define _INTERRUPT_TIME 0x7ffe0008
98+
#define _SYSTEM_TIME 0x7ffe0014
99+
#define time_lo 0
100+
diff --git a/src/runtime/time_windows_amd64.s b/src/runtime/time_windows_amd64.s
101+
index 93ab960b06..045f64eb46 100644
102+
--- a/src/runtime/time_windows_amd64.s
103+
+++ b/src/runtime/time_windows_amd64.s
104+
@@ -12,27 +12,14 @@
105+
TEXT time·now(SB),NOSPLIT,$0-24
106+
CMPB runtime·useQPCTime(SB), $0
107+
JNE useQPC
108+
+
109+
MOVQ $_INTERRUPT_TIME, DI
110+
-loop:
111+
- MOVL time_hi1(DI), AX
112+
- MOVL time_lo(DI), BX
113+
- MOVL time_hi2(DI), CX
114+
- CMPL AX, CX
115+
- JNE loop
116+
- SHLQ $32, AX
117+
- ORQ BX, AX
118+
+ MOVQ time_lo(DI), AX
119+
IMULQ $100, AX
120+
MOVQ AX, mono+16(FP)
121+
122+
MOVQ $_SYSTEM_TIME, DI
123+
-wall:
124+
- MOVL time_hi1(DI), AX
125+
- MOVL time_lo(DI), BX
126+
- MOVL time_hi2(DI), CX
127+
- CMPL AX, CX
128+
- JNE wall
129+
- SHLQ $32, AX
130+
- ORQ BX, AX
131+
+ MOVQ time_lo(DI), AX
132+
MOVQ $116444736000000000, DI
133+
SUBQ DI, AX
134+
IMULQ $100, AX
135+
diff --git a/src/runtime/time_windows_arm.s b/src/runtime/time_windows_arm.s
136+
index 7c763b66ed..6552d75ff1 100644
137+
--- a/src/runtime/time_windows_arm.s
138+
+++ b/src/runtime/time_windows_arm.s
139+
@@ -17,7 +17,9 @@ TEXT time·now(SB),NOSPLIT|NOFRAME,$0-20
140+
MOVW $_INTERRUPT_TIME, R3
141+
loop:
142+
MOVW time_hi1(R3), R1
143+
+ DMB MB_ISH
144+
MOVW time_lo(R3), R0
145+
+ DMB MB_ISH
146+
MOVW time_hi2(R3), R2
147+
CMP R1, R2
148+
BNE loop
149+
@@ -34,7 +36,9 @@ loop:
150+
MOVW $_SYSTEM_TIME, R3
151+
wall:
152+
MOVW time_hi1(R3), R1
153+
+ DMB MB_ISH
154+
MOVW time_lo(R3), R0
155+
+ DMB MB_ISH
156+
MOVW time_hi2(R3), R2
157+
CMP R1, R2
158+
BNE wall
159+
diff --git a/src/runtime/time_windows_arm64.s b/src/runtime/time_windows_arm64.s
160+
index ef52ce4c99..e8a0eb2f93 100644
161+
--- a/src/runtime/time_windows_arm64.s
162+
+++ b/src/runtime/time_windows_arm64.s
163+
@@ -13,34 +13,18 @@ TEXT time·now(SB),NOSPLIT|NOFRAME,$0-24
164+
MOVB runtime·useQPCTime(SB), R0
165+
CMP $0, R0
166+
BNE useQPC
167+
- MOVD $_INTERRUPT_TIME, R3
168+
-loop:
169+
- MOVWU time_hi1(R3), R1
170+
- MOVWU time_lo(R3), R0
171+
- MOVWU time_hi2(R3), R2
172+
- CMP R1, R2
173+
- BNE loop
174+
175+
- // wintime = R1:R0, multiply by 100
176+
- ORR R1<<32, R0
177+
+ MOVD $_INTERRUPT_TIME, R3
178+
+ MOVD time_lo(R3), R0
179+
MOVD $100, R1
180+
MUL R1, R0
181+
MOVD R0, mono+16(FP)
182+
183+
MOVD $_SYSTEM_TIME, R3
184+
-wall:
185+
- MOVWU time_hi1(R3), R1
186+
- MOVWU time_lo(R3), R0
187+
- MOVWU time_hi2(R3), R2
188+
- CMP R1, R2
189+
- BNE wall
190+
-
191+
- // w = R1:R0 in 100ns units
192+
+ MOVD time_lo(R3), R0
193+
// convert to Unix epoch (but still 100ns units)
194+
#define delta 116444736000000000
195+
- ORR R1<<32, R0
196+
SUB $delta, R0
197+
-
198+
// Convert to nSec
199+
MOVD $100, R1
200+
MUL R1, R0

0 commit comments

Comments
 (0)