Z: Accelerate inlining Thread.onSpinWait() #20714
Conversation
| TR::Instruction* cursor = new (cg->trHeapMemory()) TR::S390NOPInstruction(TR::InstOpCode::NOP, 2, node, cg); | ||
|
|
||
| TR::Compilation *comp = cg->comp(); | ||
| static const bool printIt = feGetEnv("TR_showPauseOnSpinWait") != NULL; |
There was a problem hiding this comment.
TR_showPauseOnSpinWait would be a little misleading since "PAUSE" instruction is x86 specific but I have to use the same env variable because it was used in shared code and x86!
There was a problem hiding this comment.
Do we even need condition for the traceMsg? It should be done in the entry unconditionally,
|
@r30shah I renamed the branch and it caused the older PR to close. I opened this one with a different branch name. Please take a look. |
ehsankianifar
force-pushed
the
Z_Accelerate-onSpinWait-inlining
branch
from
72e0d9d to
a5374a1
Compare
| TR::Instruction* cursor = new (cg->trHeapMemory()) TR::S390NOPInstruction(TR::InstOpCode::NOP, 2, node, cg); | ||
|
|
||
| TR::Compilation *comp = cg->comp(); | ||
| static const bool printIt = feGetEnv("TR_showPauseOnSpinWait") != NULL; |
There was a problem hiding this comment.
Do we even need condition for the traceMsg? It should be done in the entry unconditionally,
| static const bool printIt = feGetEnv("TR_showPauseOnSpinWait") != NULL; | ||
| if (printIt && comp->getOption(TR_TraceCG)) | ||
| { | ||
| traceMsg(comp, "Insert NOP for onSpinWait : node=%p, %s\n", node, comp->signature()); |
There was a problem hiding this comment.
I think printing the method name in the JIT compilation log is redundant. We would see this message in the compilation log of the method we want to log. I think, a simple message saying inlining call to Thread.onSpinWait to NOP is sufficient.
| @@ -14665,3 +14665,18 @@ J9::Z::TreeEvaluator::inlineIntegerStringSize(TR::Node* node, TR::CodeGenerator* | |||
|
|
|||
| return node->setRegister(lengthReg); | |||
| } | |||
|
|
|||
| TR::Register* | |||
| J9::Z::TreeEvaluator::inlineOnSpinWait(TR::Node *node, TR::CodeGenerator *cg) | |||
There was a problem hiding this comment.
Please add reference / Comment to the JNI call to OnSpinWait that emits NOP for reference to explain why we are generating NOP.
ehsankianifar
force-pushed
the
Z_Accelerate-onSpinWait-inlining
branch
2 times, most recently
from
e70d501 to
aba70ba
Compare
|
@r30shah thank you for your comments. When you have time, could you please check if the recent changes look good to you? thanks. |
| TR::Instruction* cursor = new (cg->trHeapMemory()) TR::S390NOPInstruction(TR::InstOpCode::NOP, 2, node, cg); | ||
|
|
||
| TR::Compilation *comp = cg->comp(); | ||
| if (comp->getOption(TR_TraceCG)) |
There was a problem hiding this comment.
Can you move this at the entry of the function - It just seems bit odd reading the code - Cosmetic change.
Thread.onSpinWait() is a simple `nop` instruction on z. Added instructions to recognize that method and inline it in jit code. Enabled by default. Disable by setting the `TR_noPauseOnSpinWait` environment variable. Signed-off-by: Ehsan Kiani Far <[email protected]>
ehsankianifar
force-pushed
the
Z_Accelerate-onSpinWait-inlining
branch
from
aba70ba to
4e89f97
Compare
|
Jenkins test sanity zlinux jdk11,jdk21 |
|
Based on the tests done by @ehsankianifar internally and on PR, merging this change. |
Thread.onSpinWait() is a simple
nopinstruction on z. Enabled by default. Disable by setting theTR_noPauseOnSpinWaitenvironment variable.