Odd Loop x86 CodeGen Behavior
Vladimir Kozlov
Vladimir.Kozlov at Sun.COM
Tue Dec 9 13:29:55 PST 2008
The stack spills you see are generated by Register Allocator.
And I agree they are not needed.
Do you have a small test to show this problem?
Thanks,
Vladimir
Edward Lee wrote:
> Some reason codegen is adding unnecessary stores that in turn cause
> unnecessary loads to be added. Below is a snippet of the output of
> opto assembly after some changes I've done to the nodes such as
> removing the safepoint for this particular backedge. Any tips on where
> I should look would be useful.
>
> The high level code is doing HashMap.get where the object has a custom
> equals method inlined that checks a long field. Kinda like..
>
> if (entry.hash == hash_ECX && entry.key.longField == long_ESP_64)
>
> The odd behavior is in B8 and B10, where long_ESP_64 gets loaded into
> EAX,EDX in B8 and written back to the stack in B10 even though both
> pair of load/stores are unnecessary. It seems that code gen is
> confused by the fact that B3 loads long_ESP_64 into EAX,EDX for the
> long compare and the branch from B3->B9 has a higher frequency
> (0.029548) than the branch B2->B8 (0.023747).
>
> But the really odd thing is why does B10 even exist in the first
> place? There's no point in writing EAX,EDX back to the stack, and this
> seems to be purely an artifact of code gen as the ideal graph doesn't
> have any store nodes to that location.
>
> 080 B2: # B8 B3 <- B1 B10 Loop: B2-B10 inner Freq: 1.01271
> 080 MOV EDI,[EBX + #8] ! Field java/util/HashMap$Entry.hash
> 083 CMP EDI,ECX
> 085 Jne,s B8 P=0.023747 C=6443.000000
> 085
> 087 B3: # B9 B4 <- B2 Freq: 0.98866
> 087 MOV ESI,[EBX + #12] ! Field java/util/HashMap$Entry.key
> 08a MOV EBP.lo,[ESI + #8]
> MOV EBP.hi,[ESI + #8]+4 ! Field customEquals.longField
> 090 MOV EAX,[ESP + #64]
> MOV EDX,[ESP + #68]
> 098 CMP EAX.lo,EBP.lo ! Long compare; set flags for low bits
> JNE,s skip
> CMP EAX.hi,EBP.hi
> skip:
> 09e Jne,s B9 P=0.029548 C=6701.000000
> 09e
> 0a0 B4: # B5 <- B3 B9 Freq: 0.959448
> 0a0 MOV EBX,[EBX + #16] ! Field java/util/HashMap$Entry.value
> ...
> 0d8 RET
>
> 0e0 B8: # B9 <- B2 Freq: 0.0240485
> 0e0 MOV EAX,[ESP + #64]
> MOV EDX,[ESP + #68]
> 0e0
> 0e8 B9: # B7 B10 <- B3 B8 Freq: 0.0532613
> 0e8 MOV EBX,[EBX + #20] ! Field java/util/HashMap$Entry.next
> 0eb TEST EBX,EBX
> 0ed Je,s B4 P=0.038502 C=718.000000
> 0ed
> 0ef B10: # B2 <- B9 Freq: 0.0512106
> 0ef MOV [ESP + #64],EAX
> MOV [ESP + #68],EDX
> 0f7 JMP B2
>
> Ed
More information about the hotspot-dev
mailing list