x86: Optimize arraycopy with constant copy size #7459
Conversation
a7ehuo
changed the title
a7ehuo
force-pushed
the
system-arraycopy-perf-20-optimize-constant-copy-size
branch
from
2432b6f
to
25ae04f
Compare
|
@0xdaryl May I ask you to review this change? Thank you! |
a7ehuo
force-pushed
the
system-arraycopy-perf-20-optimize-constant-copy-size
branch
from
25ae04f
to
d2c22d0
Compare
| { | ||
| regSize = (copySize < 4) ? 2 : 4; | ||
| } | ||
| else if (copySize < 16) // 8-15 |
There was a problem hiding this comment.
Is it true that copySize is always the full length of the destination array? That is, copySize is not a subsection of the array. If the answer is no, then I don't think this copy works in all cases because you may overwrite more bytes in the destination than you were allowed to copy. For example, let's say copySize is 13. This will generate two 8-byte loads and stores from source to destination. But I think that only works if you know that 13 is also the full length of the destination array (where you can copy some extra bytes into the slush between objects).
There was a problem hiding this comment.
Is it true that copySize is always the full length of the destination array?
No, the copy size isn't always the full length of the detonation array. It doesn't need to be
If copySize is 13, the generated load/store instructions from generateMemoryCopyInstructions as below:
[0x7f485307c070] mov rdx, qword ptr [rax+1*rdi-0x8] # L8RegMem, SymRef [#4758 -8]
[0x7f485307c190] mov rax, qword ptr [rax] # L8RegMem
[0x7f485307c2b0] mov qword ptr [rcx+1*rdi-0x8], rdx # S8MemReg, SymRef [#4760 -8]
[0x7f485307c3d0] mov qword ptr [rcx], rax # S8MemReg
src: |------- copysize=13 ------|
..... 1 2 3 4 5 6 7 8 9 10 11 12 13 .....
|------------------|
first load 8 bytes at src + (13 - 8) = src + 5
|--------------|
second load 8 bytes at src
dst: |------- copysize=13 ------|
..... 1 2 3 4 5 6 7 8 9 10 11 12 13 .....
|------------------|
first store 8 bytes at dst + (13 - 8) = dst + 5
|--------------|
second store 8 bytes at dst
Optimize arraycopy sequence if copy size is constant
and less than or equal to 64 bytes. This feature can
be disabled by env variable
TR_DisableArrayCopyInlineSmallSizeConstantCopySize.
Create generateMemoryCopyInstructions that loads and stores
memory based on specified register size. This method is used
by multiple arraycopy inline methods to reduce the repetition
of the same code.
Update generateArrayElement{Load|Store} to include
vector registers and instructions.
Signed-off-by: Annabelle Huo <[email protected]>
a7ehuo
force-pushed
the
system-arraycopy-perf-20-optimize-constant-copy-size
branch
from
d2c22d0
to
24c7d3f
Compare
|
@0xdaryl All comments are addressed. Ready for another review. Thank you! |
Optimize arraycopy sequence if copy size is constant and less than or equal to 64 bytes. This feature can be disabled by env variable
TR_DisableArrayCopyInlineSmallSizeConstantCopySize.Create
generateMemoryCopyInstructionsthat loads and stores memory based on specified register size. This method is used by multiple arraycopy inline methods to reduce the repetition of the same code.Update
generateArrayElement{Load|Store}to include vector registers and instructions.