moving int to byte[] in single step
Ulf Zibis
Ulf.Zibis at gmx.de
Wed Dec 23 16:02:13 PST 2009
In my code I have following snippet:
byte[] da = ...;
int dp = ...;
int qb = ...;
da[dp+3] = (byte)qb;
da[dp+2] = (byte)(qb >> 8);
da[dp+1] = (byte)(qb >> 16);
da[dp+0] = (byte)(qb >> 24);
This could be compiled to something like:
mov 0x24(%esp),%edi
..
mov %eax,%ebx
bswap %ebx ; only little-endian architectures
mov 0x20(%esp),%esi
mov 0x8(%esi),%edx ; implicit exception: dispatches to 0x00ba49e4
cmp %edx,%edi
jae 0x00ba49a5
mov 0x24(%esp),%ebp
add $0x3,%ebp
cmp %edx,%ebp
jae 0x00ba49c1
mov %ebx,0xc(%esi,%edi,4) ;*bastore
...
... but disassembly output results:
(copies each of the 4 bytes in single steps)
0x00ba48d6: mov 0x24(%esp),%edi
...
0x00ba492a: mov %eax,%ebx
0x00ba492c: sar $0x18,%ebx ;*ishr
; -
sun.nio.cs.ext.EUC_TW_C_d_b_c1_f3_shortMap4$Encoder::encode at 93 (line 491)
0x00ba492f: mov 0x20(%esp),%esi
0x00ba4933: mov 0x8(%esi),%edx ; implicit exception: dispatches
to 0x00ba49e4
0x00ba4936: cmp %edx,%edi
0x00ba4938: jae 0x00ba49a5
0x00ba493a: mov %bl,0xc(%esi,%edi,1) ;*bastore
; -
sun.nio.cs.ext.EUC_TW_C_d_b_c1_f3_shortMap4$Encoder::encode at 95 (line 491)
0x00ba493e: mov 0x24(%esp),%ebp
0x00ba4942: add $0x3,%ebp
0x00ba4945: mov %eax,%ecx
0x00ba4947: sar $0x10,%ecx ;*ishr
; -
sun.nio.cs.ext.EUC_TW_C_d_b_c1_f3_shortMap4$Encoder::encode at 104 (line 492)
0x00ba494a: cmp %edx,%ebp
0x00ba494c: jae 0x00ba49c1
0x00ba494e: mov %cl,0xd(%esi,%edi,1) ;*bastore
; -
sun.nio.cs.ext.EUC_TW_C_d_b_c1_f3_shortMap4$Encoder::encode at 106 (line 492)
0x00ba4952: mov %eax,%ebx
0x00ba4954: sar $0x8,%ebx
0x00ba4957: mov %bl,0xe(%esi,%edi,1) ;*bastore
; -
sun.nio.cs.ext.EUC_TW_C_d_b_c1_f3_shortMap4$Encoder::encode at 117 (line 493)
0x00ba495b: mov %al,0xf(%esi,%edi,1) ;*bastore
; -
sun.nio.cs.ext.EUC_TW_C_d_b_c1_f3_shortMap4$Encoder::encode at 125 (line 494)
...
Complete output here (line 2497):
https://java-nio-charset-enhanced.dev.java.net/source/browse/java-nio-charset-enhanced/branches/j7_EUC_TW/log/C_d_b_c1_f3_shortMap4_PA_2.xml?rev=888&view=markup
As you can see, copying the 4 bytes could be done at once.
I guess this would be faster here.
-Ulf
More information about the hotspot-compiler-dev
mailing list