ValueType Array Operations on Valhalla LWorld
Overview
I justed started getting involved with the Vahalla LWorld project, which is part of OpenJDK’s investigation of Value Types.
So this is what I just learned: how operations on an array of Values is much faster than an array of objects. Here’s a simple test case.
final __ByValue class V { // A Value
public final int v1;
public final int v2;
public final int v3;
V () {
v1 = 0;
v2 = 0;
v3 = 0;
}
public static V makeV(int v1) {
V p = __MakeDefault V();
p = __WithField(p.v1, v1);
return p;
}
}
class R { // Plain Old Java Object (POJO)
int r1, r2, r3;
}
public class Bench {
public static V[] varray = new V[10000];
public static R[] rarray = new R[10000];
static int testv() { // operate on an array of Value
V[] a = varray;
int c = 0;
for (int i=0; i<a.length; i++) {
c += a[i].v1;
}
return c;
}
static int testr() { // operate on an array of POJO
R[] a = rarray;
int c = 0;
for (int i=0; i<a.length; i++) {
c += a[i].r1;
}
return c;
}
public static void main(String args[]) {
testv();
testr();
for (int i=0; i<rarray.length; i++) {
rarray[i] = new R();
}
}
}
To test this, you need to build from the lworld branch of the
valhalla repo](http://hg.openjdk.java.net/valhalla/valhalla)
Bytecodes that Operate on an Array of Values.
Here’s the testv and testr methods in bytecodes. You can see that they are identical.
$ javap -c Bench.class
static int testv();
Code:
0: getstatic #2 // Field varray:[LV;
3: astore_0
4: iconst_0
5: istore_1
6: iconst_0
7: istore_2
8: iload_2
9: aload_0
10: arraylength
11: if_icmpge 29
14: iload_1
15: aload_0
16: iload_2
17: aaload
18: getfield #3 // Field V.v1:I
21: iadd
22: istore_1
23: iinc 2, 1
26: goto 8
29: iload_1
30: ireturn
static int testr();
Code:
0: getstatic #4 // Field rarray:[LR;
3: astore_0
4: iconst_0
5: istore_1
6: iconst_0
7: istore_2
8: iload_2
9: aload_0
10: arraylength
11: if_icmpge 29
14: iload_1
15: aload_0
16: iload_2
17: aaload
18: getfield #5 // Field R.r1:I
21: iadd
22: istore_1
23: iinc 2, 1
26: goto 8
29: iload_1
30: ireturn
The layout of varray is very different than rarray. On 64-bit platform with compressed oops, rarray looks like this
@0 object header
@8 array length
@16 rarray[0] - 4-byte compressed oop pointer
@20 rarray[1] - 4-byte compressed oop pointer
@24 ....
But varray contains “flattened” copies of V (think about how C++ lays out an array struct {int x, y, z;} varray[123];), so it looks like this
@0 object header
@8 array length
@16 rarray[0].v0, rarray[0].v1, rarray[0].v2, (filler) - 16 bytes
@32 rarray[1].v0, rarray[1].v1, rarray[1].v2, (filler) - 16 bytes
@48 ....
Comparison of Code Generated by JIT Compiler
You can see the code generated by the JIT compiler (currently only the server JIT compiler, aka C2, is enabled for Valhalla):
$ java -XX:+EnableValhalla -cp classes -XX:+UnlockDiagnosticVMOptions \
-XX:CompileCommand=quiet -XX:-BackgroundCompilation -XX:CICompilerCount=1 \
-XX:+ValueTypePassFieldsAsArgs -XX:+ValueTypeReturnedAsFields \
-XX:+ValueArrayFlatten -Xcomp -XX:+PrintAssembly -XX:+LogCompilation \
-XX:CompileCommand=compileonly,Bench::testr \
Bench
----------------------------------------------------------------------
Bench.testv()I [0x00007f661cc6a6a0, 0x00007f661cc6a7f8] 344 bytes
[Disassembling for mach='i386:x86-64']
[Entry Point]
[Verified Entry Point]
[Constants]
# {method} {0x00007f65eb0f53c0} 'testv' '()I' in 'Bench'
# [sp+0x20] (sp of caller)
0x00007f661cc6a6a0: mov %eax,-0x14000(%rsp)
0x00007f661cc6a6a7: push %rbp
0x00007f661cc6a6a8: sub $0x10,%rsp ;*synchronization entry
; - Bench::testv@-1 (line 10)
0x00007f661cc6a6ac: mov $0x4510965d8,%r10 ; {oop(a 'java/lang/Class'{0x00000004510965d8} = 'Bench')}
0x00007f661cc6a6b6: mov 0x70(%r10),%ebp ;*getstatic varray {reexecute=0 rethrow=0 return_oop=0 return_vt=0}
; - Bench::testv@0 (line 10)
0x00007f661cc6a6ba: mov 0xc(%r12,%rbp,8),%ecx ;*arraylength {reexecute=0 rethrow=0 return_oop=0 return_vt=0}
; - Bench::testv@10 (line 12)
; implicit exception: dispatches to 0x00007f661cc6a7ce
0x00007f661cc6a6bf: test %ecx,%ecx
0x00007f661cc6a6c1: jbe 0x00007f661cc6a7ae ;*if_icmpge {reexecute=0 rethrow=0 return_oop=0 return_vt=0}
; - Bench::testv@11 (line 12)
0x00007f661cc6a6c7: mov %ecx,%r11d
0x00007f661cc6a6ca: dec %r11d
0x00007f661cc6a6cd: cmp %ecx,%r11d
0x00007f661cc6a6d0: jae 0x00007f661cc6a7c0
0x00007f661cc6a6d6: mov 0x10(%r12,%rbp,8),%eax ;*aaload {reexecute=0 rethrow=0 return_oop=0 return_vt=0}
; - Bench::testv@17 (line 13)
0x00007f661cc6a6db: lea (%r12,%rbp,8),%r10 ;*getstatic varray {reexecute=0 rethrow=0 return_oop=0 return_vt=0}
; - Bench::testv@0 (line 10)
0x00007f661cc6a6df: mov %ecx,%r8d
0x00007f661cc6a6e2: add $0xfffffffffffffff1,%r8d
0x00007f661cc6a6e6: mov $0x80000000,%r9d
0x00007f661cc6a6ec: cmp %r8d,%r11d
0x00007f661cc6a6ef: cmovl %r9d,%r8d
0x00007f661cc6a6f3: mov $0x1,%r11d
0x00007f661cc6a6f9: cmp $0x1,%r8d
0x00007f661cc6a6fd: jle 0x00007f661cc6a791 ;*iload_1 {reexecute=0 rethrow=0 return_oop=0 return_vt=0}
; - Bench::testv@14 (line 13)
0x00007f661cc6a703: movslq %r11d,%r9
0x00007f661cc6a706: shl $0x4,%r9 ;*aaload {reexecute=0 rethrow=0 return_oop=0 return_vt=0}
; - Bench::testv@17 (line 13)
0x00007f661cc6a70a: add 0x10(%r10,%r9,1),%eax
0x00007f661cc6a70f: add 0x20(%r10,%r9,1),%eax
0x00007f661cc6a714: add 0x30(%r10,%r9,1),%eax
0x00007f661cc6a719: add 0x40(%r10,%r9,1),%eax
0x00007f661cc6a71e: add 0x50(%r10,%r9,1),%eax
0x00007f661cc6a723: add 0x60(%r10,%r9,1),%eax
0x00007f661cc6a728: add 0x70(%r10,%r9,1),%eax
0x00007f661cc6a72d: add 0x80(%r10,%r9,1),%eax
0x00007f661cc6a735: add 0x90(%r10,%r9,1),%eax
0x00007f661cc6a73d: add 0xa0(%r10,%r9,1),%eax
0x00007f661cc6a745: add 0xb0(%r10,%r9,1),%eax
0x00007f661cc6a74d: add 0xc0(%r10,%r9,1),%eax
0x00007f661cc6a755: add 0xd0(%r10,%r9,1),%eax
0x00007f661cc6a75d: add 0xe0(%r10,%r9,1),%eax
0x00007f661cc6a765: add 0xf0(%r10,%r9,1),%eax
0x00007f661cc6a76d: add 0x100(%r10,%r9,1),%eax ;*iadd {reexecute=0 rethrow=0 return_oop=0 return_vt=0}
; - Bench::testv@21 (line 13)
0x00007f661cc6a775: add $0x10,%r11d ;*iinc {reexecute=0 rethrow=0 return_oop=0 return_vt=0}
; - Bench::testv@23 (line 12)
0x00007f661cc6a779: cmp %r8d,%r11d
0x00007f661cc6a77c: jl 0x00007f661cc6a703 ;*if_icmpge {reexecute=0 rethrow=0 return_oop=0 return_vt=0}
; - Bench::testv@11 (line 12)
0x00007f661cc6a77e: mov 0xe8(%r15),%r9 ; ImmutableOopMap{r10=Oop }
;*goto {reexecute=1 rethrow=0 return_oop=0 return_vt=0}
; - Bench::testv@26 (line 12)
0x00007f661cc6a785: test %eax,(%r9) ;*goto {reexecute=0 rethrow=0 return_oop=0 return_vt=0}
; - Bench::testv@26 (line 12)
; {poll}
0x00007f661cc6a788: cmp %r8d,%r11d
0x00007f661cc6a78b: jl 0x00007f661cc6a703
0x00007f661cc6a791: cmp %ecx,%r11d
0x00007f661cc6a794: jge 0x00007f661cc6a7b0
0x00007f661cc6a796: xchg %ax,%ax ;*iload_1 {reexecute=0 rethrow=0 return_oop=0 return_vt=0}
; - Bench::testv@14 (line 13)
0x00007f661cc6a798: movslq %r11d,%r8
0x00007f661cc6a79b: shl $0x4,%r8
0x00007f661cc6a79f: add 0x10(%r10,%r8,1),%eax ;*iadd {reexecute=0 rethrow=0 return_oop=0 return_vt=0}
; - Bench::testv@21 (line 13)
0x00007f661cc6a7a4: inc %r11d ;*iinc {reexecute=0 rethrow=0 return_oop=0 return_vt=0}
; - Bench::testv@23 (line 12)
0x00007f661cc6a7a7: cmp %ecx,%r11d
0x00007f661cc6a7aa: jl 0x00007f661cc6a798
0x00007f661cc6a7ac: jmp 0x00007f661cc6a7b0
0x00007f661cc6a7ae: xor %eax,%eax ;*if_icmpge {reexecute=0 rethrow=0 return_oop=0 return_vt=0}
; - Bench::testv@11 (line 12)
0x00007f661cc6a7b0: add $0x10,%rsp
0x00007f661cc6a7b4: pop %rbp
0x00007f661cc6a7b5: mov 0xe8(%r15),%r10
0x00007f661cc6a7bc: test %eax,(%r10) ; {poll_return}
0x00007f661cc6a7bf: retq
0x00007f661cc6a7c0: mov $0xffffff86,%esi
0x00007f661cc6a7c5: xchg %ax,%ax
0x00007f661cc6a7c7: callq 0x00007f661cc2bd80 ; ImmutableOopMap{rbp=NarrowOop }
;*iload_1 {reexecute=0 rethrow=0 return_oop=0 return_vt=0}
; - Bench::testv@14 (line 13)
; {runtime_call UncommonTrapBlob}
0x00007f661cc6a7cc: ud2a ;*iload_1 {reexecute=0 rethrow=0 return_oop=0 return_vt=0}
; - Bench::testv@14 (line 13)
0x00007f661cc6a7ce: mov $0xfffffff6,%esi
0x00007f661cc6a7d3: callq 0x00007f661cc2bd80 ; ImmutableOopMap{}
;*arraylength {reexecute=0 rethrow=0 return_oop=0 return_vt=0}
; - Bench::testv@10 (line 12)
; {runtime_call UncommonTrapBlob}
0x00007f661cc6a7d8: ud2a ;*iload_1 {reexecute=0 rethrow=0 return_oop=0 return_vt=0}
; - Bench::testv@14 (line 13)
0x00007f661cc6a7da: hlt
[snip]
Note that starting from 0x00007f661cc6a70a, we have an unrolled loop of adding up
16 elements together for the c += a[i].v1 expression. Note that the v1 elements are spaced
out bt 16 bytes each.
In contrast, the POJO version looks like this. Although the loop body is also unrolled, the access of each element requires two loads, so it’s much slower:
Bench.testr()I [0x00007f2130c6a920, 0x00007f2130c6ab38] 536 bytes
[Disassembling for mach='i386:x86-64']
[Entry Point]
[Verified Entry Point]
[Constants]
# {method} {0x00007f20f7c004a0} 'testr' '()I' in 'Bench'
# [sp+0x40] (sp of caller)
0x00007f2130c6a920: mov %eax,-0x14000(%rsp)
0x00007f2130c6a927: push %rbp
0x00007f2130c6a928: sub $0x30,%rsp ;*synchronization entry
; - Bench::testr@-1 (line 19)
0x00007f2130c6a92c: mov $0x4510965d8,%r10 ; {oop(a 'java/lang/Class'{0x00000004510965d8} = 'Bench')}
0x00007f2130c6a936: mov 0x74(%r10),%ebp ;*getstatic rarray {reexecute=0 rethrow=0 return_oop=0 return_vt=0}
; - Bench::testr@0 (line 19)
0x00007f2130c6a93a: mov 0xc(%r12,%rbp,8),%esi ;*arraylength {reexecute=0 rethrow=0 return_oop=0 return_vt=0}
; - Bench::testr@10 (line 21)
; implicit exception: dispatches to 0x00007f2130c6aafa
0x00007f2130c6a93f: test %esi,%esi
0x00007f2130c6a941: jbe 0x00007f2130c6a983 ;*if_icmpge {reexecute=0 rethrow=0 return_oop=0 return_vt=0}
; - Bench::testr@11 (line 21)
0x00007f2130c6a943: mov %esi,%r8d
0x00007f2130c6a946: dec %r8d
0x00007f2130c6a949: cmp %esi,%r8d
0x00007f2130c6a94c: jae 0x00007f2130c6aadf
0x00007f2130c6a952: mov 0x10(%r12,%rbp,8),%r10d ;*aaload {reexecute=0 rethrow=0 return_oop=0 return_vt=0}
; - Bench::testr@17 (line 22)
0x00007f2130c6a957: mov 0xc(%r12,%r10,8),%eax ;*getfield r1 {reexecute=0 rethrow=0 return_oop=0 return_vt=0}
; - Bench::testr@18 (line 22)
; implicit exception: dispatches to 0x00007f2130c6aaee
0x00007f2130c6a95c: lea (%r12,%rbp,8),%r13 ;*getstatic rarray {reexecute=0 rethrow=0 return_oop=0 return_vt=0}
; - Bench::testr@0 (line 19)
0x00007f2130c6a960: mov %esi,%edi
0x00007f2130c6a962: add $0xfffffffffffffff1,%edi
0x00007f2130c6a965: mov $0x1,%r10d
0x00007f2130c6a96b: mov $0x80000000,%r9d
0x00007f2130c6a971: cmp %edi,%r8d
0x00007f2130c6a974: cmovl %r9d,%edi
0x00007f2130c6a978: cmp $0x1,%edi
0x00007f2130c6a97b: jle 0x00007f2130c6aabd ;*goto {reexecute=0 rethrow=0 return_oop=0 return_vt=0}
; - Bench::testr@26 (line 21)
0x00007f2130c6a981: jmp 0x00007f2130c6a9a7
0x00007f2130c6a983: xor %eax,%eax ;*if_icmpge {reexecute=0 rethrow=0 return_oop=0 return_vt=0}
; - Bench::testr@11 (line 21)
0x00007f2130c6a985: add $0x30,%rsp
0x00007f2130c6a989: pop %rbp
0x00007f2130c6a98a: mov 0xe8(%r15),%r10
0x00007f2130c6a991: test %eax,(%r10) ; {poll_return}
0x00007f2130c6a994: retq
0x00007f2130c6a995: nopw 0x0(%rax,%rax,1)
0x00007f2130c6a9a0: vmovd %xmm0,%esi
0x00007f2130c6a9a4: mov (%rsp),%edi ;*iload_1 {reexecute=0 rethrow=0 return_oop=0 return_vt=0}
; - Bench::testr@14 (line 22)
0x00007f2130c6a9a7: mov 0x10(%r13,%r10,4),%r11d
0x00007f2130c6a9ac: add 0xc(%r12,%r11,8),%eax ; implicit exception: dispatches to 0x00007f2130c6aaee
0x00007f2130c6a9b1: mov 0x14(%r13,%r10,4),%r11d ;*aaload {reexecute=0 rethrow=0 return_oop=0 return_vt=0}
; - Bench::testr@17 (line 22)
0x00007f2130c6a9b6: mov 0xc(%r12,%r11,8),%r8d ;*getfield r1 {reexecute=0 rethrow=0 return_oop=0 return_vt=0}
; - Bench::testr@18 (line 22)
; implicit exception: dispatches to 0x00007f2130c6aaee
0x00007f2130c6a9bb: mov 0x18(%r13,%r10,4),%r9d ;*aaload {reexecute=0 rethrow=0 return_oop=0 return_vt=0}
; - Bench::testr@17 (line 22)
0x00007f2130c6a9c0: mov 0xc(%r12,%r9,8),%ebx ;*getfield r1 {reexecute=0 rethrow=0 return_oop=0 return_vt=0}
; - Bench::testr@18 (line 22)
; implicit exception: dispatches to 0x00007f2130c6aaee
0x00007f2130c6a9c5: mov 0x1c(%r13,%r10,4),%r11d ;*aaload {reexecute=0 rethrow=0 return_oop=0 return_vt=0}
; - Bench::testr@17 (line 22)
0x00007f2130c6a9ca: mov 0xc(%r12,%r11,8),%ecx ;*getfield r1 {reexecute=0 rethrow=0 return_oop=0 return_vt=0}
; - Bench::testr@18 (line 22)
; implicit exception: dispatches to 0x00007f2130c6aaee
0x00007f2130c6a9cf: mov 0x20(%r13,%r10,4),%r9d ;*aaload {reexecute=0 rethrow=0 return_oop=0 return_vt=0}
; - Bench::testr@17 (line 22)
0x00007f2130c6a9d4: mov 0xc(%r12,%r9,8),%edx ;*getfield r1 {reexecute=0 rethrow=0 return_oop=0 return_vt=0}
; - Bench::testr@18 (line 22)
; implicit exception: dispatches to 0x00007f2130c6aaee
0x00007f2130c6a9d9: mov 0x24(%r13,%r10,4),%r11d ;*aaload {reexecute=0 rethrow=0 return_oop=0 return_vt=0}
; - Bench::testr@17 (line 22)
0x00007f2130c6a9de: mov 0xc(%r12,%r11,8),%ebp ;*getfield r1 {reexecute=0 rethrow=0 return_oop=0 return_vt=0}
; - Bench::testr@18 (line 22)
; implicit exception: dispatches to 0x00007f2130c6aaee
0x00007f2130c6a9e3: mov 0x28(%r13,%r10,4),%r9d ;*aaload {reexecute=0 rethrow=0 return_oop=0 return_vt=0}
; - Bench::testr@17 (line 22)
0x00007f2130c6a9e8: mov 0xc(%r12,%r9,8),%r14d ;*getfield r1 {reexecute=0 rethrow=0 return_oop=0 return_vt=0}
; - Bench::testr@18 (line 22)
; implicit exception: dispatches to 0x00007f2130c6aaee
0x00007f2130c6a9ed: mov 0x2c(%r13,%r10,4),%r11d ;*aaload {reexecute=0 rethrow=0 return_oop=0 return_vt=0}
; - Bench::testr@17 (line 22)
0x00007f2130c6a9f2: mov 0xc(%r12,%r11,8),%r11d ;*getfield r1 {reexecute=0 rethrow=0 return_oop=0 return_vt=0}
; - Bench::testr@18 (line 22)
; implicit exception: dispatches to 0x00007f2130c6aaee
0x00007f2130c6a9f7: mov 0x30(%r13,%r10,4),%r9d ;*aaload {reexecute=0 rethrow=0 return_oop=0 return_vt=0}
; - Bench::testr@17 (line 22)
0x00007f2130c6a9fc: mov 0xc(%r12,%r9,8),%r9d ;*getfield r1 {reexecute=0 rethrow=0 return_oop=0 return_vt=0}
; - Bench::testr@18 (line 22)
; implicit exception: dispatches to 0x00007f2130c6aaee
0x00007f2130c6aa01: mov %ebp,0x14(%rsp)
0x00007f2130c6aa05: mov %edx,0x10(%rsp)
0x00007f2130c6aa09: mov %ecx,0xc(%rsp)
0x00007f2130c6aa0d: mov %ebx,0x8(%rsp)
0x00007f2130c6aa11: mov %r8d,0x4(%rsp)
0x00007f2130c6aa16: mov %edi,(%rsp)
0x00007f2130c6aa19: vmovd %esi,%xmm0
0x00007f2130c6aa1d: mov 0x34(%r13,%r10,4),%ecx ;*aaload {reexecute=0 rethrow=0 return_oop=0 return_vt=0}
; - Bench::testr@17 (line 22)
0x00007f2130c6aa22: mov 0xc(%r12,%rcx,8),%r8d ;*getfield r1 {reexecute=0 rethrow=0 return_oop=0 return_vt=0}
; - Bench::testr@18 (line 22)
; implicit exception: dispatches to 0x00007f2130c6aaee
0x00007f2130c6aa27: mov 0x38(%r13,%r10,4),%ebx ;*aaload {reexecute=0 rethrow=0 return_oop=0 return_vt=0}
; - Bench::testr@17 (line 22)
0x00007f2130c6aa2c: mov 0xc(%r12,%rbx,8),%ebx ;*getfield r1 {reexecute=0 rethrow=0 return_oop=0 return_vt=0}
; - Bench::testr@18 (line 22)
; implicit exception: dispatches to 0x00007f2130c6aaee
0x00007f2130c6aa31: mov 0x3c(%r13,%r10,4),%edi ;*aaload {reexecute=0 rethrow=0 return_oop=0 return_vt=0}
; - Bench::testr@17 (line 22)
0x00007f2130c6aa36: mov 0xc(%r12,%rdi,8),%ecx ;*getfield r1 {reexecute=0 rethrow=0 return_oop=0 return_vt=0}
; - Bench::testr@18 (line 22)
; implicit exception: dispatches to 0x00007f2130c6aaee
0x00007f2130c6aa3b: mov 0x40(%r13,%r10,4),%edx ;*aaload {reexecute=0 rethrow=0 return_oop=0 return_vt=0}
; - Bench::testr@17 (line 22)
0x00007f2130c6aa40: mov 0xc(%r12,%rdx,8),%edx ;*getfield r1 {reexecute=0 rethrow=0 return_oop=0 return_vt=0}
; - Bench::testr@18 (line 22)
; implicit exception: dispatches to 0x00007f2130c6aaee
0x00007f2130c6aa45: mov 0x44(%r13,%r10,4),%esi ;*aaload {reexecute=0 rethrow=0 return_oop=0 return_vt=0}
; - Bench::testr@17 (line 22)
0x00007f2130c6aa4a: mov 0xc(%r12,%rsi,8),%edi ;*getfield r1 {reexecute=0 rethrow=0 return_oop=0 return_vt=0}
; - Bench::testr@18 (line 22)
; implicit exception: dispatches to 0x00007f2130c6aaee
0x00007f2130c6aa4f: mov 0x48(%r13,%r10,4),%ebp ;*aaload {reexecute=0 rethrow=0 return_oop=0 return_vt=0}
; - Bench::testr@17 (line 22)
0x00007f2130c6aa54: mov 0xc(%r12,%rbp,8),%ebp ;*getfield r1 {reexecute=0 rethrow=0 return_oop=0 return_vt=0}
; - Bench::testr@18 (line 22)
; implicit exception: dispatches to 0x00007f2130c6aaee
0x00007f2130c6aa59: mov 0x4c(%r13,%r10,4),%esi ;*aaload {reexecute=0 rethrow=0 return_oop=0 return_vt=0}
; - Bench::testr@17 (line 22)
0x00007f2130c6aa5e: mov 0xc(%r12,%rsi,8),%esi ;*getfield r1 {reexecute=0 rethrow=0 return_oop=0 return_vt=0}
; - Bench::testr@18 (line 22)
; implicit exception: dispatches to 0x00007f2130c6aaee
0x00007f2130c6aa63: add 0x4(%rsp),%eax
0x00007f2130c6aa67: add 0x8(%rsp),%eax
0x00007f2130c6aa6b: add 0xc(%rsp),%eax
0x00007f2130c6aa6f: add 0x10(%rsp),%eax
0x00007f2130c6aa73: add 0x14(%rsp),%eax
0x00007f2130c6aa77: add %r14d,%eax
0x00007f2130c6aa7a: add %r11d,%eax
0x00007f2130c6aa7d: add %r9d,%eax
0x00007f2130c6aa80: add %r8d,%eax
0x00007f2130c6aa83: add %ebx,%eax
0x00007f2130c6aa85: add %ecx,%eax
0x00007f2130c6aa87: add %edx,%eax
0x00007f2130c6aa89: add %edi,%eax
0x00007f2130c6aa8b: add %ebp,%eax
0x00007f2130c6aa8d: add %esi,%eax ;*iadd {reexecute=0 rethrow=0 return_oop=0 return_vt=0}
; - Bench::testr@21 (line 22)
0x00007f2130c6aa8f: add $0x10,%r10d ;*iinc {reexecute=0 rethrow=0 return_oop=0 return_vt=0}
; - Bench::testr@23 (line 21)
0x00007f2130c6aa93: cmp (%rsp),%r10d
0x00007f2130c6aa97: jl 0x00007f2130c6a9a0 ;*if_icmpge {reexecute=0 rethrow=0 return_oop=0 return_vt=0}
; - Bench::testr@11 (line 21)
0x00007f2130c6aa9d: mov 0xe8(%r15),%r11 ; ImmutableOopMap{r13=Oop }
;*goto {reexecute=1 rethrow=0 return_oop=0 return_vt=0}
; - Bench::testr@26 (line 21)
0x00007f2130c6aaa4: test %eax,(%r11) ;*goto {reexecute=0 rethrow=0 return_oop=0 return_vt=0}
; - Bench::testr@26 (line 21)
; {poll}
0x00007f2130c6aaa7: cmp (%rsp),%r10d
0x00007f2130c6aaab: jge 0x00007f2130c6aab9
0x00007f2130c6aaad: vmovd %xmm0,%esi
0x00007f2130c6aab1: mov (%rsp),%edi
0x00007f2130c6aab4: jmpq 0x00007f2130c6a9a7
0x00007f2130c6aab9: vmovd %xmm0,%esi
0x00007f2130c6aabd: cmp %esi,%r10d
0x00007f2130c6aac0: jge 0x00007f2130c6a985
0x00007f2130c6aac6: xchg %ax,%ax ;*iload_1 {reexecute=0 rethrow=0 return_oop=0 return_vt=0}
; - Bench::testr@14 (line 22)
0x00007f2130c6aac8: mov 0x10(%r13,%r10,4),%r8d ;*aaload {reexecute=0 rethrow=0 return_oop=0 return_vt=0}
; - Bench::testr@17 (line 22)
0x00007f2130c6aacd: add 0xc(%r12,%r8,8),%eax ;*iadd {reexecute=0 rethrow=0 return_oop=0 return_vt=0}
; - Bench::testr@21 (line 22)
; implicit exception: dispatches to 0x00007f2130c6aaee
0x00007f2130c6aad2: inc %r10d ;*iinc {reexecute=0 rethrow=0 return_oop=0 return_vt=0}
; - Bench::testr@23 (line 21)
0x00007f2130c6aad5: cmp %esi,%r10d
0x00007f2130c6aad8: jl 0x00007f2130c6aac8 ;*if_icmpge {reexecute=0 rethrow=0 return_oop=0 return_vt=0}
; - Bench::testr@11 (line 21)
0x00007f2130c6aada: jmpq 0x00007f2130c6a985
0x00007f2130c6aadf: mov $0xffffff86,%esi
0x00007f2130c6aae4: xchg %ax,%ax
0x00007f2130c6aae7: callq 0x00007f2130c2bd80 ; ImmutableOopMap{rbp=NarrowOop }
;*iload_1 {reexecute=0 rethrow=0 return_oop=0 return_vt=0}
; - Bench::testr@14 (line 22)
; {runtime_call UncommonTrapBlob}
0x00007f2130c6aaec: ud2a ;*iload_1 {reexecute=0 rethrow=0 return_oop=0 return_vt=0}
; - Bench::testr@14 (line 22)
0x00007f2130c6aaee: mov $0xfffffff6,%esi
0x00007f2130c6aaf3: callq 0x00007f2130c2bd80 ; ImmutableOopMap{}
;*getfield r1 {reexecute=1 rethrow=0 return_oop=0 return_vt=0}
; - Bench::testr@18 (line 22)
; {runtime_call UncommonTrapBlob}
0x00007f2130c6aaf8: ud2a ;*getfield r1 {reexecute=0 rethrow=0 return_oop=0 return_vt=0}
; - Bench::testr@18 (line 22)
0x00007f2130c6aafa: mov $0xfffffff6,%esi
0x00007f2130c6aaff: callq 0x00007f2130c2bd80 ; ImmutableOopMap{}
;*arraylength {reexecute=0 rethrow=0 return_oop=0 return_vt=0}
; - Bench::testr@10 (line 21)
; {runtime_call UncommonTrapBlob}
0x00007f2130c6ab04: ud2a ;*getfield r1 {reexecute=0 rethrow=0 return_oop=0 return_vt=0}
; - Bench::testr@18 (line 22)
0x00007f2130c6ab06: hlt
[snip]
Operations Inside the Interpreter
You may wonder – if the testv and testr methods have the exact same bytecode, how does the interpreter handle the aaload
bytecode, which needs to operate on the two arrays that are laid out different?
You can see this in the TemplateTable::aaload() handler of the interpreter:
if (ValueArrayFlatten) {
Label is_flat_array, done;
__ test_flat_array_oop(array, rbx, is_flat_array);
do_oop_load(_masm,
Address(array, index,
UseCompressedOops ? Address::times_4 : Address::times_ptr,
arrayOopDesc::base_offset_in_bytes(T_OBJECT)),
rax,
IN_HEAP_ARRAY);
__ jmp(done);
__ bind(is_flat_array);
__ call_VM(rax, CAST_FROM_FN_PTR(address,
InterpreterRuntime::value_array_load), array, index);
__ bind(done);
If the array is laid out as a ‘flat’ array, we will call InterpreterRuntime::value_array_load(), which essentially allocates and returns a new wrapper object that contains all the fields of the V object at the given array index.