x86: Improve String.hashCode for medium/long strings

This commit improves String.hashCode() performance
by calling vectorizedHashCodeLoopHelper(...) to
process longer string in an unrolled vectorized loop
with larger vector lengths. Minor tweaks were made,
including to change how multiplication vectors are
loaded.

Signed-off-by: Bradley Wood <bradley.wood@ibm.com>

Author: Bradley Wood

runtime/compiler/x/codegen/J9TreeEvaluator.cpp

@@ -9372,7 +9372,10 @@ TR::Register* J9::X86::TreeEvaluator::inlineMathFma(TR::Node* node, TR::CodeGene
    return result;
    }
 
-// Convert serial String.hashCode computation into vectorization copy and implement with SSE instruction
+// Convert serial String.hashCode computation into vectorization copy and implement with vector instructions.
+// This algorithm processes 4-characters at a time in a vectorized loop and prepends zeros to the input characters
+// to handle residual elements. To improve performance for large strings, this code calls vectorizedHashCodeLoopHelper
+// and makes use of loop unrolling and larger vector lengths.
 //
 // Conversion process example:
 //
@@ -9464,6 +9467,9 @@ static TR::Register* inlineStringHashCode(TR::Node* node, bool isCompressed, TR:
    {
    TR_ASSERT(node->getChild(1)->getOpCodeValue() == TR::iconst && node->getChild(1)->getInt() == 0, "String hashcode offset can only be const zero.");
 
+   TR::VectorLength vl = cg->getMaxPreferredVectorLength();
+   static int32_t vectorSizes[3] = { 4, 8, 16 };
+
    const int size = 4;
    auto shift = isCompressed ? 0 : 1;
 
@@ -9472,29 +9478,48 @@ static TR::Register* inlineStringHashCode(TR::Node* node, bool isCompressed, TR:
    auto index = cg->allocateRegister();
    auto hash = cg->allocateRegister();
    auto tmp = cg->allocateRegister();
+   auto loopLimit = cg->allocateRegister();
    auto hashXMM = cg->allocateRegister(TR_VRF);
    auto tmpXMM = cg->allocateRegister(TR_VRF);
    auto multiplierXMM = cg->allocateRegister(TR_VRF);
 
    auto begLabel = generateLabelSymbol(cg);
    auto endLabel = generateLabelSymbol(cg);
    auto loopLabel = generateLabelSymbol(cg);
+   auto bigLoopLabel = generateLabelSymbol(cg);
+   auto doneLabel = generateLabelSymbol(cg);
    begLabel->setStartInternalControlFlow();
    endLabel->setEndInternalControlFlow();
    auto deps = generateRegisterDependencyConditions((uint8_t)6, (uint8_t)6, cg);
    deps->addPreCondition(address, TR::RealRegister::NoReg, cg);
    deps->addPreCondition(index, TR::RealRegister::NoReg, cg);
-   deps->addPreCondition(length, TR::RealRegister::NoReg, cg);
+   deps->addPreCondition(loopLimit, TR::RealRegister::NoReg, cg);
    deps->addPreCondition(multiplierXMM, TR::RealRegister::NoReg, cg);
    deps->addPreCondition(tmpXMM, TR::RealRegister::NoReg, cg);
    deps->addPreCondition(hashXMM, TR::RealRegister::NoReg, cg);
    deps->addPostCondition(address, TR::RealRegister::NoReg, cg);
    deps->addPostCondition(index, TR::RealRegister::NoReg, cg);
-   deps->addPostCondition(length, TR::RealRegister::NoReg, cg);
+   deps->addPostCondition(loopLimit, TR::RealRegister::NoReg, cg);
    deps->addPostCondition(multiplierXMM, TR::RealRegister::NoReg, cg);
    deps->addPostCondition(tmpXMM, TR::RealRegister::NoReg, cg);
    deps->addPostCondition(hashXMM, TR::RealRegister::NoReg, cg);
 
+   // Generate Main Loop; 4x Unrolled seems to yield the best performance for large arrays
+#ifdef TR_TARGET_64BIT
+   static char *unrollVar = feGetEnv("TR_setInlineStringHashCodeUnrollCount");
+   int32_t unrollCount = unrollVar ? atoi(unrollVar) : 4;
+#else
+   int32_t unrollCount = 1;
+#endif
+
+   int32_t charsPerMainLoopIteration = unrollCount * vectorSizes[vl - TR::VectorLength128];
+
+   generateRegRegInstruction(TR::InstOpCode::XORRegReg(), node, index, index, cg);
+   generateRegRegInstruction(TR::InstOpCode::XORRegReg(), node, hash, hash, cg);
+   generateRegRegInstruction(TR::InstOpCode::MOV4RegReg, node, loopLimit, length, cg);
+   generateRegImmInstruction(TR::InstOpCode::AND4RegImm4, node, loopLimit, charsPerMainLoopIteration - 1, cg);
+   generateLabelInstruction(TR::InstOpCode::JE4, node, bigLoopLabel, cg);
+
    generateRegRegInstruction(TR::InstOpCode::MOV4RegReg, node, index, length, cg);
    generateRegImmInstruction(TR::InstOpCode::AND4RegImms, node, index, size-1, cg); // mod size
    generateRegMemInstruction(TR::InstOpCode::CMOVE4RegMem, node, index, generateX86MemoryReference(cg->findOrCreate4ByteConstant(node, size), cg), cg);
@@ -9523,17 +9548,21 @@ static TR::Register* inlineStringHashCode(TR::Node* node, bool isCompressed, TR:
 
    // Reduction Loop
    {
-   static uint32_t multiplier[] = { 31*31*31*31, 31*31*31*31, 31*31*31*31, 31*31*31*31 };
    generateLabelInstruction(TR::InstOpCode::label, node, begLabel, cg);
-   generateRegRegInstruction(TR::InstOpCode::CMP4RegReg, node, index, length, cg);
+   generateRegRegInstruction(TR::InstOpCode::CMP4RegReg, node, index, loopLimit, cg);
    generateLabelInstruction(TR::InstOpCode::JGE4, node, endLabel, cg);
-   generateRegMemInstruction(TR::InstOpCode::MOVDQURegMem, node, multiplierXMM, generateX86MemoryReference(cg->findOrCreate16ByteConstant(node, multiplier), cg), cg);
+
+   TR::Register *broadcastReg = TR::TreeEvaluator::loadConstant(node, 31*31*31*31, TR_RematerializableInt, cg);
+   generateRegRegInstruction(TR::InstOpCode::MOVDRegReg4, node, multiplierXMM, broadcastReg, cg);
+   TR::TreeEvaluator::broadcastHelper(node, multiplierXMM, vl, TR::Int32, cg);
+   cg->stopUsingRegister(broadcastReg);
+
    generateLabelInstruction(TR::InstOpCode::label, node, loopLabel, cg);
    generateRegRegInstruction(TR::InstOpCode::PMULLDRegReg, node, hashXMM, multiplierXMM, cg);
    generateRegMemInstruction(isCompressed ? TR::InstOpCode::PMOVZXBDRegMem : TR::InstOpCode::PMOVZXWDRegMem, node, tmpXMM, generateX86MemoryReference(address, index, shift, TR::Compiler->om.contiguousArrayHeaderSizeInBytes(), cg), cg);
    generateRegImmInstruction(TR::InstOpCode::ADD4RegImms, node, index, 4, cg);
    generateRegRegInstruction(TR::InstOpCode::PADDDRegReg, node, hashXMM, tmpXMM, cg);
-   generateRegRegInstruction(TR::InstOpCode::CMP4RegReg, node, index, length, cg);
+   generateRegRegInstruction(TR::InstOpCode::CMP4RegReg, node, index, loopLimit, cg);
    generateLabelInstruction(TR::InstOpCode::JL4, node, loopLabel, cg);
    generateLabelInstruction(TR::InstOpCode::label, node, endLabel, deps, cg);
    }
@@ -9550,6 +9579,17 @@ static TR::Register* inlineStringHashCode(TR::Node* node, bool isCompressed, TR:
 
    generateRegRegInstruction(TR::InstOpCode::MOVDReg4Reg, node, hash, hashXMM, cg);
 
+   // Skip secondary loop for small arrays
+   generateRegImmInstruction(TR::InstOpCode::CMPRegImm4(), node, length, charsPerMainLoopIteration, cg);
+   generateLabelInstruction(TR::InstOpCode::JL4, node, doneLabel, cg);
+
+   generateLabelInstruction(TR::InstOpCode::label, node, bigLoopLabel, cg);
+
+   // Secondary unrolled vectorized loop with larger vector lengths
+   TR::TreeEvaluator::vectorizedHashCodeLoopHelper(node, isCompressed ? TR::Int8 : TR::Int16, vl, false, hash, hash, index, length, address, unrollCount, cg);
+
+   generateLabelInstruction(TR::InstOpCode::label, node, doneLabel, cg);
+
    cg->stopUsingRegister(index);
    cg->stopUsingRegister(tmp);
    cg->stopUsingRegister(hashXMM);
@@ -9608,7 +9648,7 @@ J9::X86::TreeEvaluator::vectorizedHashCodeReductionHelper(TR::Node* node, TR::Re
    // then proceed to do horizontal reduction
    for (int32_t i = 1; i < numVectors; i++)
       {
-      OMR::X86::Encoding opcodeEncoding = opcode.getSIMDEncoding(&cg->comp()->target().cpu, vl);
+      OMR::X86::Encoding opcodeEncoding = opcode.getSIMDEncoding(&cg->comp()->target().cpu, vl, vl == TR::VectorLength512);
       generateRegRegInstruction(TR::InstOpCode::PADDDRegReg, node, vectorRegVRF, vectorRegisters[i], cg, opcodeEncoding);
       }
 
@@ -9625,7 +9665,7 @@ J9::X86::TreeEvaluator::vectorizedHashCodeReductionHelper(TR::Node* node, TR::Re
       case TR::VectorLength256:
           // extract 128 bits from ymm and store in xmm, then perform vertical operation
          generateRegRegImmInstruction(TR::InstOpCode::VEXTRACTF128RegRegImm1, node, tmpVectorRegVRF, vectorRegVRF, 0xFF, cg);
-         generateRegRegInstruction(opcode.getMnemonic(), node, vectorRegVRF, tmpVectorRegVRF, cg, opcode.getSIMDEncoding(&cg->comp()->target().cpu, TR::VectorLength128));
+         generateRegRegInstruction(opcode.getMnemonic(), node, vectorRegVRF, tmpVectorRegVRF, cg, OMR::X86::VEX_L128);
          // Fallthrough to treat remaining result as 128-bit vector
       case TR::VectorLength128:
          generateRegRegImmInstruction(TR::InstOpCode::PSHUFDRegRegImm1, node, tmpVectorRegVRF, vectorRegVRF, 0x0e, cg);
@@ -9732,7 +9772,11 @@ J9::X86::TreeEvaluator::vectorizedHashCodeLoopHelper(TR::Node *node,
    begLabel->setStartInternalControlFlow();
    endLabel->setEndInternalControlFlow();
 
-   generateRegRegInstruction(TR::InstOpCode::MOVRegReg(), node, result, initialHash, cg);
+   if (result != initialHash)
+      {
+      generateRegRegInstruction(TR::InstOpCode::MOVRegReg(), node, result, initialHash, cg);
+      }
+
    generateLabelInstruction(TR::InstOpCode::label, node, begLabel, cg);
    generateRegRegInstruction(TR::InstOpCode::MOVRegReg(), node, tmp, length, cg);
    generateRegImmInstruction(TR::InstOpCode::AND4RegImm4, node, tmp, ~(numElements - 1), cg);
@@ -9750,7 +9794,11 @@ J9::X86::TreeEvaluator::vectorizedHashCodeLoopHelper(TR::Node *node,
    int32_t multiplier31PowNData[16];
    // Fill multiplier array with 31^numElements
    std::fill_n(multiplier31PowNData, 16, powersOf31[64 - numElements]);
-   generateRegMemInstruction(TR::InstOpCode::MOVDQURegMem, node, multiplierVRF, generateX86MemoryReference(cg->findOrCreateConstantDataSnippet(node, multiplier31PowNData, vectorSizeElements * sizeof(int32_t)), cg), cg, vectorEncoding);
+
+   TR::Register *broadcastReg = TR::TreeEvaluator::loadConstant(node, powersOf31[64 - numElements], TR_RematerializableInt, cg);
+   generateRegRegInstruction(TR::InstOpCode::MOVDRegReg4, node, multiplierVRF, broadcastReg, cg);
+   TR::TreeEvaluator::broadcastHelper(node, multiplierVRF, vl, TR::Int32, cg);
+   cg->stopUsingRegister(broadcastReg);
 
    for (int32_t i = 0; i < unrollCount; i++)
       {