x86: Implement vblend - #8010
Conversation
|
jenkins build all |
| cg->decReferenceCount(rhsNode); | ||
| cg->decReferenceCount(maskNode); | ||
|
|
||
| if (node->getNumChildren() == 4) { |
There was a problem hiding this comment.
According to [1], vblend only has 3 children. What is this check for?
[1]
omr/compiler/il/VectorOperations.enum
Line 939 in 0c59477
There was a problem hiding this comment.
When I implemented this, there was a bug in the optimizer where an additional erroneous child argument was present. I'm not sure if that issue was fixed, I will have to check.
There was a problem hiding this comment.
n4306n ( 0) vstoreiVector128Int32 <array-shadow>[#241 Shadow] [flags 0x80000613 0x0 ] ()
n6457n ( 0) aladd (X>=0 internalPtr )
n35n ( 1) ==>newarray (in &GPR_0127) (highWordZero Unsigned X!=0 allocationCanBeRemoved )
n1617n ( 0) lconst 8 (highWordZero X!=0 X>=0 cannotOverflow )
n3768n ( 0) vblendVector128Int32 (in VRF_0137) ()
n2295n ( 0) vloadiVector128Int32 <array-shadow>[#241 Shadow] [flags 0x80000613 0x0 ] (in VR
n6413n ( 0) aladd (X>=0 internalPtr )
n6535n ( 0) ==>aRegLoad (in &GPR_0017) (X!=0 SeenRealReference )
n1617n ( 0) ==>lconst 8 (highWordZero X!=0 X>=0 cannotOverflow )
n3265n ( 0) vloadiVector128Int32 <array-shadow>[#241 Shadow] [flags 0x80000613 0x0 ] (in VR
n6428n ( 0) aladd (X>=0 internalPtr )
n6534n ( 0) ==>aRegLoad (in &GPR_0016) (X!=0 SeenRealReference )
n1617n ( 0) ==>lconst 8 (highWordZero X!=0 X>=0 cannotOverflow )
n1134n ( 0) i2mVector128Int32 (in VRF_0136) ()
n6408n ( 0) iloadi <generic int shadow>[#1226 Shadow] [flags 0x603 0x0 ] (in GPR_0134)
n6397n ( 0) aladd (X>=0 internalPtr )
n6536n ( 0) ==>aRegLoad (in &GPR_0018) (X!=0 SeenRealReference )
n1617n ( 0) ==>lconst 8 (highWordZero X!=0 X>=0 cannotOverflow )
n4096n ( 1) l2a (node obj11)
n6642n ( 1) lshl (compressionSequence )
n6641n ( 1) ==>iu2l (in GPR_0131)
n6639n ( 1) iconst 3 (X!=0 X>=0 )
-----------------------------
There was a problem hiding this comment.
Your tree suggests the bug is still there. How pervasive is the problem and what's the status on resolving it? I am dubious about checking in a workaround to counter the case of an invalid vblend.
There was a problem hiding this comment.
I think the extra child is always present if vblend is generated by OpenJ9. I asked @gita-omr to look into it. I don't have a timeline for that.
There was a problem hiding this comment.
A fix has been proposed so I have removed this check. This PR should not be merged until the corresponding OpenJ9 fix is in.
There was a problem hiding this comment.
I think the extra child is always present if vblend is generated by OpenJ9. I asked @gita-omr to look into it. I don't have a timeline for that.
It's 10 characters change. I will do it now.
There was a problem hiding this comment.
Opened eclipse-openj9/openj9#22876. Hopefully this is not holding this PR any more.
There was a problem hiding this comment.
Pull request eclipse-openj9/openj9#22876 has been merged, so merging this pull request shouldn't harm that downstream project.
Signed-off-by: Bradley Wood <bradley.wood@ibm.com>
Signed-off-by: Bradley Wood <bradley.wood@ibm.com>
Signed-off-by: Bradley Wood <bradley.wood@ibm.com>
4 participants
No description provided.