Switch TPM to SVM entirely

zuban32 · igcbot · commit 89ca3bdb6a2c · 2021-03-15T19:26:09.000+01:00
diff --git a/IGC/VectorCompiler/include/vc/GenXCodeGen/GenXInternalMetadata.h b/IGC/VectorCompiler/include/vc/GenXCodeGen/GenXInternalMetadata.h
@@ -38,6 +38,14 @@ namespace FunctionMD {
 inline constexpr const char GenXKernelInternal[] = "genx.kernel.internal";
 }
 
+namespace InstMD {
+inline constexpr const char SVMBlockType[] = "SVMBlockType";
+}
+
+namespace ModuleMD {
+inline constexpr const char UseSVMStack[] = "genx.useGlobalMem";
+}
+
 namespace internal {
 
 namespace KernelMDOp {
diff --git a/IGC/VectorCompiler/lib/GenXCodeGen/GenXCisaBuilder.cpp b/IGC/VectorCompiler/lib/GenXCodeGen/GenXCisaBuilder.cpp
@@ -3633,7 +3633,10 @@ void GenXKernelBuilder::buildIntrinsic(CallInst *CI, unsigned IntrinID,
     Value *V = CI;
     if (!AI.isRet())
       V = CI->getArgOperand(AI.getArgIdx());
-    unsigned ElBytes = getResultedTypeSize(V->getType()->getScalarType(), DL);
+    auto *EltType = V->getType()->getScalarType();
+    if (auto *MDType = CI->getMetadata(InstMD::SVMBlockType))
+      EltType = cast<ValueAsMetadata>(MDType->getOperand(0).get())->getType();
+    unsigned ElBytes = getResultedTypeSize(EltType, DL);
     switch (ElBytes) {
       // For N = 2 byte data type, use block size 1 and block count 2.
       // Otherwise, use block size N and block count 1.
diff --git a/IGC/VectorCompiler/lib/GenXCodeGen/GenXThreadPrivateMemory.cpp b/IGC/VectorCompiler/lib/GenXCodeGen/GenXThreadPrivateMemory.cpp
@@ -35,6 +35,7 @@ IN THE SOFTWARE.
 #include "GenXTargetMachine.h"
 #include "GenXUtil.h"
 #include "GenXVisa.h"
+#include "vc/GenXCodeGen/GenXInternalMetadata.h"
 
 #include "Probe/Assertion.h"
 #include "llvmWrapper/IR/DerivedTypes.h"
@@ -208,6 +209,7 @@ std::pair<Value *, unsigned>
 GenXThreadPrivateMemory::NormalizeVector(Value *From, Type *To,
                                          Instruction *Inst) {
   Type *I32Ty = Type::getInt32Ty(Inst->getContext());
+  Type *I64Ty = Type::getInt64Ty(Inst->getContext());
   Value *Res = From;
   Type *FromTy = From->getType();
   IGC_ASSERT(isa<VectorType>(FromTy));
@@ -234,22 +236,22 @@ GenXThreadPrivateMemory::NormalizeVector(Value *From, Type *To,
     To = IGCLLVM::FixedVectorType::get(I32Ty, NumElts);
     EltSz = I32Ty->getPrimitiveSizeInBits() / genx::ByteBits;
     Res = CastInst::Create(Instruction::BitCast, Res, To, "", Inst);
-  } else if (cast<VectorType>(To)->getElementType()->getPrimitiveSizeInBits() <
-                 genx::DWordBits
-             // this is required for correct generation of svm.gather/scatter
-             // of data of type which size is < i32 because these intrinsics
-             // infer their block size from the type of the data they handle
-             && !m_useGlobalMem) {
+  } else if (m_DL->getTypeSizeInBits(cast<VectorType>(To)->getElementType()) <
+             genx::DWordBits) {
     To = IGCLLVM::FixedVectorType::get(I32Ty, NumElts);
-
-    Res = CastInst::Create(Instruction::ZExt, From, To, "", Inst);
-  } else if (cast<VectorType>(To)->getElementType()->getPrimitiveSizeInBits() ==
-             genx::QWordBits) {
+    Res = CastInst::CreateZExtOrBitCast(From, To, "", Inst);
+  } else if (!m_useGlobalMem &&
+             m_DL->getTypeSizeInBits(cast<VectorType>(To)->getElementType()) ==
+                 genx::QWordBits) {
+    if (From->getType()->getScalarType()->isPointerTy()) {
+      auto *NewType = IGCLLVM::FixedVectorType::get(I64Ty, NumElts);
+      From = CastInst::Create(CastInst::PtrToInt, From, NewType, "", Inst);
+    }
     NumElts *= 2;
     EltSz = I32Ty->getPrimitiveSizeInBits() / genx::ByteBits;
     To = IGCLLVM::FixedVectorType::get(I32Ty, NumElts);
 
-    Res = CastInst::Create(Instruction::BitCast, From, To, "", Inst);
+    Res = CastInst::CreateBitOrPointerCast(From, To, "", Inst);
   }
 
   return std::make_pair(Res, EltSz);
@@ -258,6 +260,8 @@ GenXThreadPrivateMemory::NormalizeVector(Value *From, Type *To,
 Instruction *
 GenXThreadPrivateMemory::RestoreVectorAfterNormalization(Instruction *From,
                                                          Type *To) {
+  if (From->getType() == To)
+    return From;
   Instruction *Restored = From;
   unsigned EltSz = m_DL->getTypeSizeInBits(To->getScalarType());
   IGC_ASSERT(EltSz > 0);
@@ -519,35 +523,19 @@ bool GenXThreadPrivateMemory::replaceLoad(LoadInst *LdI) {
     LdTy = IGCLLVM::FixedVectorType::get(LdTy, 1);
 
   unsigned NumEltsToLoad = cast<VectorType>(LdTy)->getNumElements();
-  unsigned LdEltTySz = m_DL->getTypeSizeInBits(LdEltTy);
-  if (!(m_useGlobalMem && LdEltTy->isIntegerTy(64)) &&
-      LdEltTySz == genx::QWordBits)
-    NumEltsToLoad *= 2;
+  unsigned ValueEltSz = m_DL->getTypeSizeInBits(LdEltTy) / genx::ByteBits;
 
   Value *PredVal = ConstantInt::get(Type::getInt1Ty(*m_ctx), 1);
   Value *Pred = Builder.CreateVectorSplat(NumEltsToLoad, PredVal);
 
   Type *I32Ty = Type::getInt32Ty(*m_ctx);
   Type *I64Ty = Type::getInt64Ty(*m_ctx);
-  Type *TyToLoad = (m_useGlobalMem && LdEltTy->isIntegerTy(64)) ? I64Ty : I32Ty;
-  if (LdEltTy->isFloatTy())
-    TyToLoad = LdEltTy;
-  Type *RealTyToLoad = LdEltTy;
-  if (!(m_useGlobalMem && LdEltTy->isIntegerTy(64)) &&
-      m_DL->getTypeSizeInBits(RealTyToLoad) == genx::QWordBits)
-    RealTyToLoad = I32Ty;
-  unsigned RealTyToLoadSz =
-      m_DL->getTypeSizeInBits(RealTyToLoad) / genx::ByteBits;
-  // we don't want to use improper block sizes for loads of i8/i16
-  // to make sure we comply with alignment rules for gathers
-  bool NoExtToDword =
-      m_useGlobalMem &&
-      !(LdI->getType()->isAggregateType() || LdI->getType()->isVectorTy()) &&
-      m_DL->getTypeSizeInBits(LdI->getType()) < genx::DWordBits;
-  if (NoExtToDword)
-    TyToLoad = LdI->getType();
   Value *OldValOfTheDataRead =
-      Builder.CreateVectorSplat(NumEltsToLoad, UndefValue::get(TyToLoad));
+      Builder.CreateVectorSplat(NumEltsToLoad, UndefValue::get(LdEltTy));
+  std::tie(OldValOfTheDataRead, ValueEltSz) =
+      NormalizeVector(OldValOfTheDataRead, LdTy, LdI);
+  NumEltsToLoad =
+      cast<VectorType>(OldValOfTheDataRead->getType())->getNumElements();
 
   Value *PointerOp = LdI->getPointerOperand();
   Value *Offset = lookForPtrReplacement(PointerOp);
@@ -557,10 +545,13 @@ bool GenXThreadPrivateMemory::replaceLoad(LoadInst *LdI) {
                  ? llvm::GenXIntrinsic::genx_svm_gather
                  : llvm::GenXIntrinsic::genx_gather_scaled;
 
-  Value *EltsOffset = FormEltsOffsetVector(NumEltsToLoad, RealTyToLoadSz, LdI);
+  Value *EltsOffset = FormEltsOffsetVector(NumEltsToLoad, ValueEltSz, LdI);
 
-  unsigned SrcSize = genx::log2(RealTyToLoadSz);
-  Value *logNumBlocks = ConstantInt::get(I32Ty, m_useGlobalMem ? 0 : SrcSize);
+  unsigned NumBlocks = m_DL->getTypeSizeInBits(LdEltTy) / genx::ByteBits;
+  // This logic is aligned with the on in CisaBuilder and GenXLowering
+  // The reason behind check for == 2 is that svm intrinsics don't support
+  // BlockSize of 2, so for ops with i16s we have to use BlockSize == 1 and NumBlocks == 2
+  Value *logNumBlocks = ConstantInt::get(I32Ty, genx::log2(NumBlocks == 2 ? NumBlocks : 1));
   Value *Scale = ConstantInt::get(Type::getInt16Ty(*m_ctx), 0);
   Value *Surface = ConstantInt::get(I32Ty,
                                     visa::getReservedSurfaceIndex(m_stack));
@@ -601,6 +592,10 @@ bool GenXThreadPrivateMemory::replaceLoad(LoadInst *LdI) {
     ProperGather = LdVal;
   }
 
+  Gather->setMetadata(InstMD::SVMBlockType,
+                      MDNode::get(*m_ctx, llvm::ValueAsMetadata::get(
+                                              UndefValue::get(LdEltTy))));
+
   LLVM_DEBUG(dbgs() << *Gather << "\n");
   LdI->replaceAllUsesWith(ProperGather);
   LdI->eraseFromParent();
@@ -647,7 +642,9 @@ bool GenXThreadPrivateMemory::replaceStore(StoreInst *StI) {
       {Pred->getType(),
        (m_useGlobalMem ? Offset : EltsOffset)->getType(),
        ValueOp->getType()});
-  Value *logNumBlocks = ConstantInt::get(I32Ty, m_useGlobalMem ? 0 : genx::log2(ValueEltSz));
+  unsigned NumBlocks = m_DL->getTypeSizeInBits(ValueOpTy->getScalarType()) / genx::ByteBits;
+  // see the comment in replaceLoad above
+  Value *logNumBlocks = ConstantInt::get(I32Ty, genx::log2(NumBlocks == 2 ? NumBlocks : 1));
   Value *Scale = ConstantInt::get(Type::getInt16Ty(*m_ctx), 0);
   Value *Surface = ConstantInt::get(I32Ty,
                                     visa::getReservedSurfaceIndex(m_stack));
@@ -662,6 +659,11 @@ bool GenXThreadPrivateMemory::replaceStore(StoreInst *StI) {
   Scatter->insertAfter(StI);
   StI->eraseFromParent();
 
+  Scatter->setMetadata(
+      InstMD::SVMBlockType,
+      MDNode::get(*m_ctx, llvm::ValueAsMetadata::get(
+                              UndefValue::get(ValueOpTy->getScalarType()))));
+
   LLVM_DEBUG(dbgs() << *Scatter << "\n");
   m_scatter.push_back(Scatter);
 
@@ -1094,6 +1096,12 @@ void SplitScatter(CallInst *CI) {
   }
   IGC_ASSERT(FirstScatter && SecondScatter);
 
+  auto *MD = CI->getMetadata(InstMD::SVMBlockType);
+  if (MD) {
+    FirstScatter->setMetadata(InstMD::SVMBlockType, MD);
+    SecondScatter->setMetadata(InstMD::SVMBlockType, MD);
+  }
+
   FirstScatter->insertAfter(CI);
   SecondScatter->insertAfter(FirstScatter);
 
@@ -1163,6 +1171,12 @@ void SplitGather(CallInst *CI) {
   }
   IGC_ASSERT(FirstGather && SecondGather);
 
+  auto *MD = CI->getMetadata(InstMD::SVMBlockType);
+  if (MD) {
+    FirstGather->setMetadata(InstMD::SVMBlockType, MD);
+    SecondGather->setMetadata(InstMD::SVMBlockType, MD);
+  }
+
   FirstGather->insertAfter(CI);
   SecondGather->insertAfter(FirstGather);
 
@@ -1280,14 +1294,16 @@ bool GenXThreadPrivateMemory::runOnModule(Module &M) {
   m_ST = &getAnalysis<TargetPassConfig>()
               .getTM<GenXTargetMachine>()
               .getGenXSubtarget();
+  if (!m_ST->isOCLRuntime())
+    m_useGlobalMem = false;
   for (auto &F : M)
     visit(F);
-  if (!m_useGlobalMem &&
-      std::find_if(m_alloca.begin(), m_alloca.end(), SVMChecker()) !=
-          m_alloca.end()) {
+  if (m_useGlobalMem ||
+      (m_ST->isOCLRuntime() && std::find_if(m_alloca.begin(), m_alloca.end(),
+                                            SVMChecker()) != m_alloca.end())) {
     LLVM_DEBUG(dbgs() << "Switching TPM to SVM\n");
     // TODO: move the name string to vc-intrinsics *MD::useGlobalMem
-    M.addModuleFlag(Module::ModFlagBehavior::Error, "genx.useGlobalMem", 1);
+    M.addModuleFlag(Module::ModFlagBehavior::Error, ModuleMD::UseSVMStack, 1);
     m_useGlobalMem = true;
   }
   bool Result = false;
