Bug ID: 6547163 6498658 changes introduced regression in arraycopy on sparc

6547163 : 6498658 changes introduced regression in arraycopy on sparc

Details

Type:	Bug	Submit Date:	2007-04-18
Status:	Resolved	Updated Date:	2010-04-03
Project Name:	JDK	Resolved Date:	2007-05-24
Component:	hotspot	OS:	solaris_9
Sub-Component:	compiler	CPU:	sparc
Priority:	P2
Resolution:	Fixed
Affected Versions:	7
Fixed Versions:	hs10

Related Reports

Backport:	JDK-2148621 - 6498658 changes introduced regression in arraycopy on sparc
Backport:	JDK-2176927 - 6498658 changes introduced regression in arraycopy on sparc
Relates:	JDK-6498658 - System.arraycopy performance lags
Relates:	JDK-6563987 - api/org_ietf/jgss/GSSContext/index.html#SetGetTests: VM Crash JDK 6u2 since b03 with -server -Xcomp

Sub Tasks

Description

6498658 changes introduced regression in arraycopy on sparc (32-bits VM)
when the copy offsets statically known constants which have heapWordSize mod
values (4 bytes in 32-bits VM) and offset's difference is not 8 bytes mod:

20070214101449.sgoldman.6524971
jaberwocky% gamma -Xbatch ArrayCopyTest5

Disjoint +4 bytes offset small arraycopy:
Time to copy 128 bytes (128 bytes): 122
Time to copy 64  chars (128 bytes): 134
Time to copy 32   ints (128 bytes): 125
Time to copy 32   oops (128 bytes): 134

Disjoint +8 bytes offset small arraycopy:
Time to copy 128 bytes (128 bytes): 124
Time to copy 64  chars (128 bytes): 98
Time to copy 32   ints (128 bytes): 103
Time to copy 32   oops (128 bytes): 110
Time to copy 16  longs (128 bytes): 87

Disjoint +4 bytes offset small arraycopy:
Time to copy 128 bytes (128 bytes): 117
Time to copy 64  chars (128 bytes): 103
Time to copy 32   ints (128 bytes): 100
Time to copy 32   oops (128 bytes): 109

Disjoint +8 bytes offset small arraycopy:
Time to copy 128 bytes (128 bytes): 103
Time to copy 64  chars (128 bytes): 72
Time to copy 32   ints (128 bytes): 77
Time to copy 32   oops (128 bytes): 86
Time to copy 16  longs (128 bytes): 62

Disjoint +4 bytes offset small arraycopy:
Time to copy 128 bytes (128 bytes): 89
Time to copy 64  chars (128 bytes): 99
Time to copy 32   ints (128 bytes): 108
Time to copy 32   oops (128 bytes): 96

Disjoint +8 bytes offset small arraycopy:
Time to copy 128 bytes (128 bytes): 90
Time to copy 64  chars (128 bytes): 65
Time to copy 32   ints (128 bytes): 65
Time to copy 32   oops (128 bytes): 78
Time to copy 16  longs (128 bytes): 55

20070214151808.kvn.6498658
jaberwocky% gamma -Xbatch ArrayCopyTest5

Disjoint +4 bytes offset small arraycopy:
Time to copy 128 bytes (128 bytes): 18451
Time to copy 64  chars (128 bytes): 18401
Time to copy 32   ints (128 bytes): 17089
Time to copy 32   oops (128 bytes): 17070

Disjoint +8 bytes offset small arraycopy:
Time to copy 128 bytes (128 bytes): 81
Time to copy 64  chars (128 bytes): 80
Time to copy 32   ints (128 bytes): 81
Time to copy 32   oops (128 bytes): 80
Time to copy 16  longs (128 bytes): 68

Disjoint +4 bytes offset small arraycopy:
Time to copy 128 bytes (128 bytes): 18694
Time to copy 64  chars (128 bytes): 18658
Time to copy 32   ints (128 bytes): 17317
Time to copy 32   oops (128 bytes): 17259

Disjoint +8 bytes offset small arraycopy:
Time to copy 128 bytes (128 bytes): 57
Time to copy 64  chars (128 bytes): 54
Time to copy 32   ints (128 bytes): 54
Time to copy 32   oops (128 bytes): 60
Time to copy 16  longs (128 bytes): 43

Disjoint +4 bytes offset small arraycopy:
Time to copy 128 bytes (128 bytes): 18660
Time to copy 64  chars (128 bytes): 18573
Time to copy 32   ints (128 bytes): 17239
Time to copy 32   oops (128 bytes): 17221

Disjoint +8 bytes offset small arraycopy:
Time to copy 128 bytes (128 bytes): 53
Time to copy 64  chars (128 bytes): 47
Time to copy 32   ints (128 bytes): 45
Time to copy 32   oops (128 bytes): 46
Time to copy 16  longs (128 bytes): 35

Fixed VM
jaberwocky% gamma -Xbatch ArrayCopyTest5

Disjoint +4 bytes offset small arraycopy:
Time to copy 128 bytes (128 bytes): 98
Time to copy 64  chars (128 bytes): 87
Time to copy 32   ints (128 bytes): 89
Time to copy 32   oops (128 bytes): 93

Disjoint +8 bytes offset small arraycopy:
Time to copy 128 bytes (128 bytes): 80
Time to copy 64  chars (128 bytes): 73
Time to copy 32   ints (128 bytes): 79
Time to copy 32   oops (128 bytes): 78
Time to copy 16  longs (128 bytes): 67

Disjoint +4 bytes offset small arraycopy:
Time to copy 128 bytes (128 bytes): 84
Time to copy 64  chars (128 bytes): 67
Time to copy 32   ints (128 bytes): 65
Time to copy 32   oops (128 bytes): 68

Disjoint +8 bytes offset small arraycopy:
Time to copy 128 bytes (128 bytes): 55
Time to copy 64  chars (128 bytes): 49
Time to copy 32   ints (128 bytes): 54
Time to copy 32   oops (128 bytes): 52
Time to copy 16  longs (128 bytes): 41

Disjoint +4 bytes offset small arraycopy:
Time to copy 128 bytes (128 bytes): 59
Time to copy 64  chars (128 bytes): 55
Time to copy 32   ints (128 bytes): 55
Time to copy 32   oops (128 bytes): 62

Disjoint +8 bytes offset small arraycopy:
Time to copy 128 bytes (128 bytes): 50
Time to copy 64  chars (128 bytes): 49
Time to copy 32   ints (128 bytes): 42
Time to copy 32   oops (128 bytes): 51
Time to copy 16  longs (128 bytes): 35


64-bits VM is fine since heapWordSize is 8 bytes:

20070214101449.sgoldman.6524971
jaberwocky% gamma -Xbatch ArrayCopyTest5

Disjoint +4 bytes offset small arraycopy:
Time to copy 128 bytes (128 bytes): 135
Time to copy 64  chars (128 bytes): 142
Time to copy 32   ints (128 bytes): 133
Time to copy 32   oops (128 bytes): 145

Disjoint +8 bytes offset small arraycopy:
Time to copy 128 bytes (128 bytes): 135
Time to copy 64  chars (128 bytes): 110
Time to copy 32   ints (128 bytes): 111
Time to copy 32   oops (128 bytes): 145
Time to copy 16  longs (128 bytes): 91

Disjoint +4 bytes offset small arraycopy:
Time to copy 128 bytes (128 bytes): 126
Time to copy 64  chars (128 bytes): 115
Time to copy 32   ints (128 bytes): 110
Time to copy 32   oops (128 bytes): 118

Disjoint +8 bytes offset small arraycopy:
Time to copy 128 bytes (128 bytes): 112
Time to copy 64  chars (128 bytes): 82
Time to copy 32   ints (128 bytes): 85
Time to copy 32   oops (128 bytes): 118
Time to copy 16  longs (128 bytes): 68

Disjoint +4 bytes offset small arraycopy:
Time to copy 128 bytes (128 bytes): 101
Time to copy 64  chars (128 bytes): 108
Time to copy 32   ints (128 bytes): 112
Time to copy 32   oops (128 bytes): 113

Disjoint +8 bytes offset small arraycopy:
Time to copy 128 bytes (128 bytes): 103
Time to copy 64  chars (128 bytes): 77
Time to copy 32   ints (128 bytes): 72
Time to copy 32   oops (128 bytes): 106
Time to copy 16  longs (128 bytes): 60

20070214151808.kvn.6498658
jaberwocky% gamma -Xbatch ArrayCopyTest5

Disjoint +4 bytes offset small arraycopy:
Time to copy 128 bytes (128 bytes): 99
Time to copy 64  chars (128 bytes): 107
Time to copy 32   ints (128 bytes): 98
Time to copy 32   oops (128 bytes): 134

Disjoint +8 bytes offset small arraycopy:
Time to copy 128 bytes (128 bytes): 78
Time to copy 64  chars (128 bytes): 84
Time to copy 32   ints (128 bytes): 84
Time to copy 32   oops (128 bytes): 134
Time to copy 16  longs (128 bytes): 76

Disjoint +4 bytes offset small arraycopy:
Time to copy 128 bytes (128 bytes): 91
Time to copy 64  chars (128 bytes): 81
Time to copy 32   ints (128 bytes): 72
Time to copy 32   oops (128 bytes): 106

Disjoint +8 bytes offset small arraycopy:
Time to copy 128 bytes (128 bytes): 53
Time to copy 64  chars (128 bytes): 60
Time to copy 32   ints (128 bytes): 59
Time to copy 32   oops (128 bytes): 112
Time to copy 16  longs (128 bytes): 50

Disjoint +4 bytes offset small arraycopy:
Time to copy 128 bytes (128 bytes): 71
Time to copy 64  chars (128 bytes): 70
Time to copy 32   ints (128 bytes): 66
Time to copy 32   oops (128 bytes): 96

Disjoint +8 bytes offset small arraycopy:
Time to copy 128 bytes (128 bytes): 46
Time to copy 64  chars (128 bytes): 52
Time to copy 32   ints (128 bytes): 46
Time to copy 32   oops (128 bytes): 101
Time to copy 16  longs (128 bytes): 43

Comments

SUGGESTED FIX

Webrev:                 http://prt-web.sfbay.sun.com/net/prt-archiver.sfbay/data/archived_workspaces/main/c2_baseline/2007/20070418210619.kvn.6547163/workspace/webrevs/webrev-2007.04.18/index.html

Solution:
Check source address in aligned stubs also and use copy with shift
if it is not aligned to 8 bytes.
Added regression test.

2007-04-19

EVALUATION

The aligned arraycopy stubs are called when the copy offsets statically known constants 
which have heapWordSize mod values (4 bytes in 32-bits VM).
In aligned arraycopy stubs only destination address was checked and aligned to 8 bytes
causing the execution of ldx/stx instructions on address not aligned to 8 bytes.
Which causes HW traps on each instruction.

2007-04-18

SELECT A COUNTRY/REGION
Africa Operation Argentina Australia Austria Bahrain Bangladesh Belgium & Luxembourg Belize Bhutan Bolivia Bosnia & Herzegovina Brasil Brunei Bulgaria Cambodia Canada - English Canada - French Chile China Colombia	Costa Rica Croatia Cyprus Czech Republic Denmark Ecuador Egypt Estonia Finland France Germany Greece Guatemala Honduras Hong Kong Hungary India Indonesia Iraq Ireland	Israel Italy Japan Jordan Kazakhstan Korea Kuwait Laos Latvia Lebanon Lithuania Malaysia Maldives Malta Mexico Moldova Nepal Netherlands New Zealand Nicaragua	Norway Oman Pakistan Panama Paraguay Peru Philippines Poland Portugal Puerto Rico Qatar Romania Russia Saudi Arabia Serbia & Montenegro Singapore Slovakia Slovenia South Africa Spain	Sri Lanka Sweden Switzerland -- French Switzerland -- German Taiwan Thailand Turkey Ukraine United Arab Emirates United Kingdom United States Uruguay Venezuela Vietnam Yemen

Hardware and Software, Engineered to Work Together