Merge pull request eclipse-omr#7553 from Spencer-Comin/popcnt

X: Implement popcnt
knn-k · Nov 29, 2024 · 37e9b35 · 37e9b35
2 parents 7a568a0 + 35faffb
commit 37e9b35
Show file tree

Hide file tree

Showing 8 changed files with 51 additions and 16 deletions.
diff --git a/compiler/x/amd64/codegen/OMRTreeEvaluator.cpp b/compiler/x/amd64/codegen/OMRTreeEvaluator.cpp
@@ -2195,12 +2195,6 @@ OMR::X86::AMD64::TreeEvaluator::inotzEvaluator(TR::Node *node, TR::CodeGenerator
    return TR::TreeEvaluator::badILOpEvaluator(node, cg);
    }
 
-TR::Register*
-OMR::X86::AMD64::TreeEvaluator::ipopcntEvaluator(TR::Node *node, TR::CodeGenerator *cg)
-   {
-   return TR::TreeEvaluator::badILOpEvaluator(node, cg);
-   }
-
 TR::Register*
 OMR::X86::AMD64::TreeEvaluator::lhbitEvaluator(TR::Node *node, TR::CodeGenerator *cg)
    {
@@ -2228,7 +2222,14 @@ OMR::X86::AMD64::TreeEvaluator::lnotzEvaluator(TR::Node *node, TR::CodeGenerator
 TR::Register*
 OMR::X86::AMD64::TreeEvaluator::lpopcntEvaluator(TR::Node *node, TR::CodeGenerator *cg)
    {
-   return TR::TreeEvaluator::badILOpEvaluator(node, cg);
+   TR::Node *child = node->getFirstChild();
+   TR::Register *inputReg = cg->longClobberEvaluate(child);
+
+   generateRegRegInstruction(TR::InstOpCode::POPCNT8RegReg, node, inputReg, inputReg, cg);
+
+   node->setRegister(inputReg);
+   cg->decReferenceCount(child);
+   return inputReg;
    }
 
 TR::Register*

diff --git a/compiler/x/amd64/codegen/OMRTreeEvaluator.hpp b/compiler/x/amd64/codegen/OMRTreeEvaluator.hpp
@@ -411,7 +411,6 @@ class OMR_EXTENSIBLE TreeEvaluator: public OMR::X86::TreeEvaluator
    static TR::Register *ilbitEvaluator(TR::Node *node, TR::CodeGenerator *cg);
    static TR::Register *inolzEvaluator(TR::Node *node, TR::CodeGenerator *cg);
    static TR::Register *inotzEvaluator(TR::Node *node, TR::CodeGenerator *cg);
-   static TR::Register *ipopcntEvaluator(TR::Node *node, TR::CodeGenerator *cg);
    static TR::Register *lhbitEvaluator(TR::Node *node, TR::CodeGenerator *cg);
    static TR::Register *llbitEvaluator(TR::Node *node, TR::CodeGenerator *cg);
    static TR::Register *lnolzEvaluator(TR::Node *node, TR::CodeGenerator *cg);

diff --git a/compiler/x/codegen/OMRTreeEvaluator.hpp b/compiler/x/codegen/OMRTreeEvaluator.hpp
@@ -149,6 +149,7 @@ class OMR_EXTENSIBLE TreeEvaluator: public OMR::TreeEvaluator
    static TR::Register *s2iEvaluator(TR::Node *node, TR::CodeGenerator *cg);
    static TR::Register *su2iEvaluator(TR::Node *node, TR::CodeGenerator *cg);
    static TR::Register *c2iEvaluator(TR::Node *node, TR::CodeGenerator *cg);
+   static TR::Register *ipopcntEvaluator(TR::Node *node, TR::CodeGenerator *cg);
    static TR::Register *ibits2fEvaluator(TR::Node *node, TR::CodeGenerator *cg);
    static TR::Register *fbits2iEvaluator(TR::Node *node, TR::CodeGenerator *cg);
    static TR::Register *compareFloatAndBranchEvaluator(TR::Node *node, TR::CodeGenerator *cg);

diff --git a/compiler/x/codegen/UnaryEvaluator.cpp b/compiler/x/codegen/UnaryEvaluator.cpp
@@ -455,3 +455,15 @@ TR::Register *OMR::X86::TreeEvaluator::c2iEvaluator(TR::Node *node, TR::CodeGene
    {
    return TR::TreeEvaluator::conversionAnalyser(node, TR::InstOpCode::MOVZXReg4Mem2, TR::InstOpCode::MOVZXReg4Reg2, cg);
    }
+
+TR::Register *OMR::X86::TreeEvaluator::ipopcntEvaluator(TR::Node *node, TR::CodeGenerator *cg)
+   {
+   TR::Node *child = node->getFirstChild();
+   TR::Register *inputReg = cg->intClobberEvaluate(child);
+
+   generateRegRegInstruction(TR::InstOpCode::POPCNT4RegReg, node, inputReg, inputReg, cg);
+
+   node->setRegister(inputReg);
+   cg->decReferenceCount(child);
+   return inputReg;
+   }
diff --git a/compiler/x/env/OMRCPU.cpp b/compiler/x/env/OMRCPU.cpp
@@ -158,6 +158,15 @@ OMR::X86::CPU::getSupportsHardwareSQRT()
    return true;
    }
 
+bool
+OMR::X86::CPU::hasPopulationCountInstruction()
+   {
+   if ((self()->getX86ProcessorFeatureFlags2() & TR_POPCNT) != 0x00000000)
+      return true;
+   else
+     return false;
+   }
+
 bool
 OMR::X86::CPU::supportsTransactionalMemoryInstructions()
    {

diff --git a/compiler/x/env/OMRCPU.hpp b/compiler/x/env/OMRCPU.hpp
@@ -71,6 +71,12 @@ class OMR_EXTENSIBLE CPU : public OMR::CPU
 
    bool getSupportsHardwareSQRT();
 
+   /** @brief Determines whether the popcnt instruction is available on the current processor.
+    *
+    *  @return true if popcnt is available, false otherwise.
+    */
+   bool hasPopulationCountInstruction();
+
    /** @brief Determines whether the Transactional Memory (TM) facility is available on the current processor.
     *
     *  @return true if TM is available, false otherwise.

diff --git a/compiler/x/i386/codegen/OMRTreeEvaluator.cpp b/compiler/x/i386/codegen/OMRTreeEvaluator.cpp
@@ -2274,12 +2274,6 @@ OMR::X86::I386::TreeEvaluator::inotzEvaluator(TR::Node *node, TR::CodeGenerator
    return TR::TreeEvaluator::badILOpEvaluator(node, cg);
    }
 
-TR::Register*
-OMR::X86::I386::TreeEvaluator::ipopcntEvaluator(TR::Node *node, TR::CodeGenerator *cg)
-   {
-   return TR::TreeEvaluator::badILOpEvaluator(node, cg);
-   }
-
 TR::Register*
 OMR::X86::I386::TreeEvaluator::lhbitEvaluator(TR::Node *node, TR::CodeGenerator *cg)
    {
@@ -2307,7 +2301,21 @@ OMR::X86::I386::TreeEvaluator::lnotzEvaluator(TR::Node *node, TR::CodeGenerator
 TR::Register*
 OMR::X86::I386::TreeEvaluator::lpopcntEvaluator(TR::Node *node, TR::CodeGenerator *cg)
    {
-   return TR::TreeEvaluator::badILOpEvaluator(node, cg);
+   TR::Node *child = node->getFirstChild();
+   TR::Register *inputReg = cg->longClobberEvaluate(child);
+   TR::Register *inputHigh = inputReg->getHighOrder();
+   TR::Register *inputLow = inputReg->getLowOrder();
+   TR::Register *resultReg = inputLow;
+
+   //add low result and high result together
+   generateRegRegInstruction(TR::InstOpCode::POPCNT4RegReg, node, inputLow, inputLow, cg);
+   generateRegRegInstruction(TR::InstOpCode::POPCNT4RegReg, node, inputHigh, inputHigh, cg);
+   generateRegRegInstruction(TR::InstOpCode::ADD4RegReg, node, inputLow, inputHigh, cg);
+
+   cg->stopUsingRegister(inputHigh);
+   node->setRegister(resultReg);
+   cg->decReferenceCount(child);
+   return resultReg;
    }
 
 TR::Register*

diff --git a/compiler/x/i386/codegen/OMRTreeEvaluator.hpp b/compiler/x/i386/codegen/OMRTreeEvaluator.hpp
@@ -392,7 +392,6 @@ class OMR_EXTENSIBLE TreeEvaluator: public OMR::X86::TreeEvaluator
    static TR::Register *ilbitEvaluator(TR::Node *node, TR::CodeGenerator *cg);
    static TR::Register *inolzEvaluator(TR::Node *node, TR::CodeGenerator *cg);
    static TR::Register *inotzEvaluator(TR::Node *node, TR::CodeGenerator *cg);
-   static TR::Register *ipopcntEvaluator(TR::Node *node, TR::CodeGenerator *cg);
    static TR::Register *lhbitEvaluator(TR::Node *node, TR::CodeGenerator *cg);
    static TR::Register *llbitEvaluator(TR::Node *node, TR::CodeGenerator *cg);
    static TR::Register *lnolzEvaluator(TR::Node *node, TR::CodeGenerator *cg);