Request for reviews (XS): 7043552: regression after 7042327
tom.rodriguez at oracle.com
Tue May 10 12:46:12 PDT 2011
Looks good then.
On May 10, 2011, at 12:37 PM, Vladimir Kozlov wrote:
> Yes, one after an other for each unroll:
> 0e0 movl R11, #-2147483648 # int
> 0e6 cmpl R10, [RSP + #8 (32-bit)]
> 0eb movl R8, [rsp + #8] # spill
> 0f0 cmovll R8, R11 # signed, int
> 0f4 movl RCX, R8 # spill
> 0f7 addl RCX, #-2 # int
> 0fa cmpl R8, RCX
> 0fd cmovll RCX, R11 # signed, int
> 101 movl R8, RCX # spill
> 104 addl R8, #-4 # int
> 108 cmpl RCX, R8
> 10b cmovll R8, R11 # signed, int
> 10f movl RCX, R8 # spill
> 112 addl RCX, #-8 # int
> 115 cmpl R8, RCX
> 118 cmovll RCX, R11 # signed, int
> Tom Rodriguez wrote:
>> So the old code ended up stacking them up?
>> On May 10, 2011, at 12:06 PM, Vladimir Kozlov wrote:
>>> Fixed 7043552: regression after 7042327
>>> I screw up the fix for 7042327 and Opaque2 node is generated for each unroll preventing folding CMoveI nodes. Generate Opaque2 node only during first unroll.
>>> Tested with failed case from 7042327.
More information about the hotspot-compiler-dev