|
Description
|
while looking at code generated for other things I noticed some bad handling of constant values. It seems that when constants are commoned up in a compilation if the register pressure is high enough you end up with multiple copies of the same constant being produced repeatedly within a block. You also seem to end up with instruction variants which expect the values to be in registers instead of versions which could handle it as a constant. The problem is particularly bad on customer where there are few registers and in general when long constants are used. Here's a program which shows the problem. It's not interesting by itself, just for showing the problem.
public class zero {
int a;
int b;
long c;
long d;
public static void main(String[] args) {
zero zero = new zero();
System.out.println(zero);
for (int i = 0; i < 1000000; i++) {
}
zero.a = 0;
zero.b = 1;
zero.c = 0xffffffffffL;
zero.d = 0xfffffffff0L;
System.out.println(zero);
if (args.length == 0) {
zero.a = 0;
zero.b = 0;
zero.c = 0xffffffffffL;
zero.d = 0xfffffffff0L;
} else if (args.length == 2) {
zero.a = 1;
zero.b = 1;
zero.c = 0xffffffffffL;
zero.d = 0xffffffffffL;
} else if (args.length == 3) {
zero.a = 0;
zero.b = 0;
zero.c = 0xfffffffff0L;
zero.d = 0xfffffffff0L;
} else if (args.length == 4) {
zero.a = 1;
zero.b = 1;
zero.c = 0xffffffffffL;
zero.d = 0xfffffffff0L;
} else if (args.length == 5) {
zero.a = 0;
zero.b = 1;
zero.c = 0xffffffffffL;
zero.d = 0xfffffffff0L;
}
}
}
running it using java_g -Xbatch -Xcomp -XX:CompileOnly=zero -XX:+PrintOptoAssembly zero you get the following output.
<?xml version='1.0' encoding='UTF-8'?>
<hotspot_log version='142 1' process='22460' time_ms='1082065721374'>
<vm_version>
<name>
Java HotSpot(TM) Server VM
</name>
<release>
1.5.0-never-intrinsic-jvmg2-debug
</release>
<info>
Java HotSpot(TM) Server VM (1.5.0-never-intrinsic-jvmg2) for solaris-x86, built on Apr 14 2004 10:15:40 by unknown with unknown Workshop:0x550
</info>
</vm_version>
<tty>
Java HotSpot(TM) Server VM warning: Attempt to use MPSS failed.
<writer thread='12'/>
<opto_assembly compile_id='1'>
{method}
- klass: {other class}
- method holder: 'zero'
- constants: 0xdd3eb024{constant pool}
- access: 0xc1000009 public static
- name: 'main'
- signature: '([Ljava/lang/String;)V'
- max stack: 3
- max locals: 3
- size of params: 1
- method size: 18
- vtable index: -1
- code size: 222
- code start: 0x92d68ae8
- code end (excl): 0x92d68bc6
- method data: 0x92d68e70
- checked ex length: 0
- linenumber start: 0x92d68bc6
- localvar length: 0
#
# void ( java/lang/String:exact *[int+]:exact* )
#
#r000 ECX : parm 0: java/lang/String:exact *[int+]:exact*
# -- Old ESP -- Framesize: 16 --
#r059 ESP+12: return address
#r058 ESP+ 8: pad2, in_preserve
#r061 ESP+ 4: spill
#r060 ESP+ 0: spill
#
abababab N1: # B1 <- B7 B10 Freq: 0.000666667
abababab
000 B1: # B5 B2 <- BLOCK HEAD IS JUNK Freq: 0.000666667
000 # stack bang
SUB ESP,12 # Create frame
00d MOV [ESP + #0],ECX
010 MOV EAX, Thread::current()
02c MOV ECX,EAX
02e MOV EAX,[EAX + #76]
031 LEA EBX,[EAX + #32]
034 CMPu EBX,[ECX + #80]
037 Jge,us B5 P=0.000100 C=-1.000000
037
039 B2: # B3 <- B1 Freq: 0.0006666
039 MOV [ECX + #76],EBX
03c MOV [EAX],0x00000001
042 MOV [EAX + #4],precise klass zero: 0x08176aa8:Constant:exact *
049 MOV EBX,#0
04e MOV [EAX + #8],EBX
051 MOV [EAX + #12],EBX
054 MOV [EAX + #16],EBX
057 MOV [EAX + #20],EBX
05a MOV [EAX + #24],EBX
05d MOV [EAX + #28],EBX
05d
060 B3: # B8 B4 <- B2 B6 Freq: 0.000666667
060 MOV EBX,EAX
062 #checkcastPP of EBX
062 MOV [ESP + #4],EBX
066 MOV ECX,EBX
068 NOP # Pad for loops and calls
069 NOP # Pad for loops and calls
06a NOP # Pad for loops and calls
06b CALL,static zero::<init>
# zero::main @ bci:4 L0=ESP + #0 L1=_ L2=_ STK0=ESP + #4
# EBP=Callers_EBP EDI=Callers_EDI ESI=Callers_ESI [0]=Oop [4]=Oop
070
070 B4: # B7 <- B3 Freq: 0
070 JMP,s B7
070
072 B5: # B9 B6 <- B1 Freq: 6.66777e-08
072 MOV ECX,precise klass zero: 0x08176aa8:Constant:exact *
077 CALL,static wrapper for: _new_Java
# zero::main @ bci:0 L0=ESP + #0 L1=_ L2=_
# EBP=Callers_EBP EDI=Callers_EDI ESI=Callers_ESI [0]=Oop
07c
07c B6: # B3 <- B5 Freq: 0
07c JMP,s B3
07c
07e B7: # N1 <- B4 Freq: 1e-06
07e MOV ECX,#32
083 CALL,static wrapper for: uncommon_trap
# zero::main @ bci:8 L0=ESP + #0 L1=ESP + #4 L2=_
# EBP=Callers_EBP EDI=Callers_EDI ESI=Callers_ESI [0]=Oop [4]=Oop
088 INT3 ; ShouldNotReachHere
088
08d B8: # B10 <- B3 Freq: 6.66667e-09
08d # exception oop is in EAX; no code emitted
08d MOV ECX,EAX
08f JMP,s B10
08f
091 B9: # B10 <- B5 Freq: 6.66777e-13
091 # exception oop is in EAX; no code emitted
091 MOV ECX,EAX
091
093 B10: # N1 <- B8 B9 Freq: 6.66733e-09
093 ADD ESP,12 # Destroy frame
096 JMP rethrow_stub
096
</opto_assembly>
<opto_assembly compile_id='2'>
{method}
- klass: {other class}
- method holder: 'zero'
- constants: 0xdd3eb024{constant pool}
- access: 0x81000001 public
- name: '<init>'
- signature: '()V'
- max stack: 1
- max locals: 1
- size of params: 1
- method size: 18
- vtable index: -1
- code size: 5
- code start: 0x92d68a68
- code end (excl): 0x92d68a6d
- method data: 0x92d68f88
- checked ex length: 0
- linenumber start: 0x92d68a6d
- localvar length: 0
#
# void ( zero:NotNull * )
#
#r000 ECX : parm 0: zero:NotNull *
# -- Old ESP -- Framesize: 8 --
#r059 ESP+ 4: return address
#r058 ESP+ 0: pad2, in_preserve
#
000 N30: # B1 <- BLOCK HEAD IS JUNK Freq: 0.000666667
000 CMP EAX,[ECX+4] # Inline cache check
JNE OptoRuntime::handle_ic_miss_stub
NOP
NOP
NOP
000
00c B1: # B3 B2 <- BLOCK HEAD IS JUNK Freq: 0.000666667
00c # stack bang
SUB ESP,4 # Create frame
019 NOP # Pad for loops and calls
01a NOP # Pad for loops and calls
01b CALL,static java.lang.Object::<init>
# zero::<init> @ bci:1 L0=_
# EBP=Callers_EBP EDI=Callers_EDI ESI=Callers_ESI
020
020 B2: # N30 <- B1 Freq: 0.000666653
020 ADD ESP,4 # Destroy frame
TEST PollPage,EAX ! Poll Safepoint
029 RET
029
02a B3: # N30 <- B1 Freq: 6.66667e-09
02a # exception oop is in EAX; no code emitted
02a MOV ECX,EAX
02c ADD ESP,4 # Destroy frame
02f JMP rethrow_stub
02f
</opto_assembly>
<opto_assembly compile_id='1' compile_kind='osr'>
{method}
- klass: {other class}
- method holder: 'zero'
- constants: 0xdd3eb024{constant pool}
- access: 0xc1000009 public static
- name: 'main'
- signature: '([Ljava/lang/String;)V'
- max stack: 3
- max locals: 3
- size of params: 1
- method size: 18
- vtable index: -1
- code size: 222
- code start: 0x92d68ae8
- code end (excl): 0x92d68bc6
- method data: 0x92d68e70
- checked ex length: 0
- linenumber start: 0x92d68bc6
- localvar length: 0
#
# void ( rawptr:BotPTR, rawptr:BotPTR )
#
#r004 EAX : parm 0: rawptr:BotPTR
#r001 EBX : parm 1: rawptr:BotPTR
# -- Old ESP -- Framesize: 32 --
#r059 ESP+28: return address
#r058 ESP+24: pad2, in_preserve
#r065 ESP+20: spill
#r064 ESP+16: spill
#r063 ESP+12: spill
#r062 ESP+ 8: spill
#r061 ESP+ 4: spill
#r060 ESP+ 0: spill
#
000 N253: # B1 <- BLOCK HEAD IS JUNK Freq: 0.000666667
000 INT3
NOP # Pad for loops and calls
NOP # Pad for loops and calls
NOP # Pad for loops and calls
008 B1: # B22 B2 <- BLOCK HEAD IS JUNK Freq: 0.000666667
008 # stack bang
MOV EBX,EDI # Move locals ptr to interpreter_arg_ptr_reg
SUB ESP,28 # Create frame
017 MOV [ESP + #8],EBP
MOV [ESP + #12],EDI
01f MOV EBP,[EBX + #-4]
022 MOV EDI,[EBX]
024 TEST EDI,EDI
026 Jeq B22 P=0.001000 C=-1.000000
026
02c B2: # B23 B3 <- B1 Freq: 0.000666
02c MOV [ESP + #16],ESI
030 MOV EAX,[EDI + #4]
033 CMPu EAX,precise klass [Ljava/lang/String;: 0x08166b10:Constant:exact *
039 Jne,u B23 P=0.100000 C=-1.000000
039
03f B3: # B4 <- B2 Freq: 0.0005994
03f MOV ESI,EDI
041 #checkcastPP of ESI
041
041 B4: # B24 B5 <- B22 B3 Freq: 0.000600067
041 TEST EBP,EBP
043 Jeq B24 P=0.001000 C=-1.000000
043
049 B5: # B23 B6 <- B4 Freq: 0.000599467
049 MOV EAX,[EBP + #4]
04c CMPu EAX,precise klass zero: 0x08166a48:Constant:exact *
052 Jne,u B23 P=0.100000 C=-1.000000
052
058 B6: # B7 <- B5 Freq: 0.00053952
058 #checkcastPP of EBP
058
058 B7: # B28 B8 <- B24 B6 Freq: 0.00054012
058 TEST EBP,EBP
05a Jeq B28 P=0.000001 C=-1.000000
05a
060 B8: # B27 B9 <- B7 Freq: 0.000540119
060 MOV [EBP + #28],#1
067 MOV EBX,#300
06c MOV EBX,[EBX + precise klass java/lang/System: 0x08166ba0:Constant:exact *]
072 MOV [ESP + #4],EBX
076 MOV ECX.lo,#1099511627760.lo
MOV ECX.hi,#1099511627760.hi
080 MOV [EBP + #16],ECX.lo
MOV [EBP + #16]+4,ECX.hi
086 MOV [ESP + #0],EBP
089 MOV EBX,#0
08e MOV [EBP + #24],EBX
091 MOV EBP.lo,#1099511627775.lo
MOV EBP.hi,#1099511627775.hi
09b MOV ECX,[ESP + #0]
09e MOV [ECX + #8],EBP.lo
MOV [ECX + #8]+4,EBP.hi
0a4 MOV EBX,[ESP + #4]
0a8 TEST EBX,EBX
0aa Jeq B27 P=0.000001 C=-1.000000
0aa
0b0 B9: # B26 B10 <- B8 Freq: 0.000540119
0b0 MOV ECX,EBX
0b2 MOV EDX,[ESP + #0]
0b5 NOP # Pad for loops and calls
0b6 NOP # Pad for loops and calls
0b7 CALL,static java.io.PrintStream::println
# zero::main @ bci:57 L0=ESI L1=ESP + #0 L2=_
# ESI=Oop [0]=Oop [8]=Callers_EBP [12]=Callers_EDI [16]=Callers_ESI
0bc
0bc B10: # B25 B11 <- B9 Freq: 0.000540108
0bc MOV ECX,[ESI + #8]
0bf NullCheck ESI
0bf
0bf B11: # B20 B12 <- B10 Freq: 0.000540108
0bf TEST ECX,ECX
0c1 Jeq B20 P=0.100000 C=-1.000000
0c1
0c7 B12: # B19 B13 <- B11 Freq: 0.000486097
0c7 CMP ECX,#2
0ca Jeq B19 P=0.100000 C=-1.000000
0ca
0d0 B13: # B18 B14 <- B12 Freq: 0.000437487
0d0 CMP ECX,#3
0d3 Jeq,s B18 P=0.100000 C=-1.000000
0d3
0d5 B14: # B17 B15 <- B13 Freq: 0.000393738
0d5 CMP ECX,#4
0d8 Jeq,s B17 P=0.100000 C=-1.000000
0d8
0da B15: # B21 B16 <- B14 Freq: 0.000354365
0da CMP ECX,#5
0dd Jne B21 P=0.900000 C=-1.000000
0dd
0e3 B16: # B21 <- B15 Freq: 3.54365e-05
0e3 MOV EAX,[ESP + #0]
0e6 MOV ECX.lo,#1099511627760.lo
MOV ECX.hi,#1099511627760.hi
0f0 MOV [EAX + #16],ECX.lo
MOV [EAX + #16]+4,ECX.hi
0f6 MOV ECX,EAX
0f8 MOV EBX,#0
0fd MOV [ECX + #24],EBX
100 MOV [ECX + #8],EBP.lo
MOV [ECX + #8]+4,EBP.hi
106 MOV [ECX + #28],#1
10d JMP B21
10d
112 B17: # B21 <- B14 Freq: 3.93738e-05
112 MOV EDX,[ESP + #0]
115 MOV ECX.lo,#1099511627760.lo
MOV ECX.hi,#1099511627760.hi
11f MOV [EDX + #16],ECX.lo
MOV [EDX + #16]+4,ECX.hi
125 MOV EBX,EDX
127 MOV [EBX + #8],EBP.lo
MOV [EBX + #8]+4,EBP.hi
12d MOV ECX,EBX
12f MOV [ECX + #24],#1
136 MOV [ECX + #28],#1
13d JMP B21
13d
142 B18: # B21 <- B13 Freq: 4.37487e-05
142 MOV ECX,[ESP + #0]
145 MOV EBP.lo,#1099511627760.lo
MOV EBP.hi,#1099511627760.hi
14f MOV [ECX + #16],EBP.lo
MOV [ECX + #16]+4,EBP.hi
155 MOV EBP,#0
15a MOV [ECX + #24],EBP
15d MOV EBX,#0
162 MOV [ECX + #28],EBX
165 MOV EBP.lo,#1099511627760.lo
MOV EBP.hi,#1099511627760.hi
16f MOV [ECX + #8],EBP.lo
MOV [ECX + #8]+4,EBP.hi
175 JMP,s B21
175
177 B19: # B21 <- B12 Freq: 4.86097e-05
177 MOV ECX,[ESP + #0]
17a MOV [ECX + #16],EBP.lo
MOV [ECX + #16]+4,EBP.hi
180 MOV [ECX + #8],EBP.lo
MOV [ECX + #8]+4,EBP.hi
186 MOV [ECX + #24],#1
18d MOV [ECX + #28],#1
194 JMP,s B21
194
196 B20: # B21 <- B11 Freq: 5.40108e-05
196 MOV EDX,[ESP + #0]
199 MOV ECX.lo,#1099511627760.lo
MOV ECX.hi,#1099511627760.hi
1a3 MOV [EDX + #16],ECX.lo
MOV [EDX + #16]+4,ECX.hi
1a9 MOV ECX,EDX
1ab MOV EBX,#0
1b0 MOV [ECX + #24],EBX
1b3 MOV EBX,#0
1b8 MOV [ECX + #28],EBX
1bb MOV [ECX + #8],EBP.lo
MOV [ECX + #8]+4,EBP.hi
1bb
1c1 B21: # N253 <- B16 B20 B17 B18 B19 B15 Freq: 0.000540108
1c1 MOV EBP,[ESP + #8]
MOV EDI,[ESP + #12]
1c9 MOV ESI,[ESP + #16]
1cd ADD ESP,28 # Destroy frame
TEST PollPage,EAX ! Poll Safepoint
1d6 RET
1d6
1d7 B22: # B4 <- B1 Freq: 6.66658e-07
1d7 MOV [ESP + #16],ESI
1db MOV ESI,NULL
1e0 JMP B4
1e0
1e5 B23: # N253 <- B2 B5 Freq: 1e-06
1e5 MOV ESI,[EBX + #-8]
1e8 MOV ECX,#-75
1ed NOP # Pad for loops and calls
1ee NOP # Pad for loops and calls
1ef CALL,static wrapper for: uncommon_trap
# zero::main @ bci:17 L0=EDI L1=EBP L2=ESI
# EBP=Oop EDI=Oop [8]=Callers_EBP [12]=Callers_EDI [16]=Callers_ESI
1f4 INT3 ; ShouldNotReachHere
1f4
1f9 B24: # B7 <- B4 Freq: 6.00059e-07
1f9 MOV EBP,NULL
1fe JMP B7
1fe
203 B25: # N253 <- B10 Freq: 5.4728e-10
203 MOV ECX,#-10
208 NOP # Pad for loops and calls
209 NOP # Pad for loops and calls
20a NOP # Pad for loops and calls
20b CALL,static wrapper for: uncommon_trap
# zero::main @ bci:61 L0=_ L1=_ L2=_ STK0=_
# [8]=Callers_EBP [12]=Callers_EDI [16]=Callers_ESI
210 INT3 ; ShouldNotReachHere
210
215 B26: # N253 <- B9 Freq: 5.40119e-09
215 # exception oop is in EAX; no code emitted
215 MOV ECX,EAX
217 MOV EBP,[ESP + #8]
MOV EDI,[ESP + #12]
21f MOV ESI,[ESP + #16]
223 ADD ESP,28 # Destroy frame
226 JMP rethrow_stub
226
22b B27: # N253 <- B8 Freq: 5.47292e-10
22b MOV ECX,#-10
230 NOP # Pad for loops and calls
231 NOP # Pad for loops and calls
232 NOP # Pad for loops and calls
233 CALL,static wrapper for: uncommon_trap
# zero::main @ bci:57 L0=_ L1=_ L2=_ STK0=ESP + #4 STK1=ESP + #0
# [0]=Oop [4]=Oop [8]=Callers_EBP [12]=Callers_EDI [16]=Callers_ESI
238 INT3 ; ShouldNotReachHere
238
23d B28: # N253 <- B7
Freq: 5.47292e-10
23d MOV ECX,#-10
242 NOP # Pad for loops and calls
243 CALL,static wrapper for: uncommon_trap
# zero::main @ bci:31 L0=_ L1=_ L2=_ STK0=EBP STK1=#0
# EBP=Oop [8]=Callers_EBP [12]=Callers_EDI [16]=Callers_ESI
248 INT3 ; ShouldNotReachHere
248
</opto_assembly>
<writer thread='1'/>
--- Compiler Statistics ---
Methods seen: 3 Methods parsed: 3 Nodes created: 3832
32 original NULL checks - 29 elided (90%); optimizer leaves 27,
1 made implicit ( 3%)
CCP: 3 constants found: 0
Total frameslots = 234, Max frameslots = 12
Inserted 0 spill loads, 0 spill stores, 0 mem-mem moves and 0 copies.
Total load cost= 0, store cost = 0, mem-mem cost = 0.00, copy cost = 0.
Adjusted spill cost = 0.
Conservatively coalesced 21 copies, 0 pairs, 197 post alloc.
Average allocation trips 0.103448
High Pressure Blocks = 0, Low Pressure Blocks = 0
Nops added 0 bytes to total of 5781 bytes, for 0.00%
Deoptimization traps recorded:
1 (100.0%) total
unloaded/reinterpret/getstatic: 1 (100.0%)
0 inline cache miss in compiled
0 wrong method
0 unresolved static call site
0 unresolved virtual call site
3 unresolved opt virtual call site
1 removed paired I2C/C2I adapters during deopt
Peephole: peephole rules applied: 0
PhaseIdealLoop=4, sum _unique=732
<tty_done stamp='7.355'/>
</tty>
<hotspot_log_done stamp='7.356'/>
</hotspot_log>
Blocks 16, 17, 18 and 8 all show the problem I'm talking about. I've seen the same problem on sparc, in particular when using tiered compilation which produces a lot of references to constants representing methodDataOops. I can show you how to reproduce those as well if you are interested.
Part of the problem with longs appears to be the misnamed value short_spill_offset_limit which from it's name you'd assume has something to do with spilling but as far as I can tell it's only purpose it to determine which long constants get forced into registers. On intel this can greatly increase register pressure.
xxxxx@xxxxx 2004-04-15
Posted Date : 2005-07-22 03:26:19.0
|
