Enable Loop unrolling protmotion for Alloc for some

platforms

Enable Loop unrolling protmotion for Alloc for some platforms

Author: lioujheyu

IGC/Compiler/GenTTI.cpp

@@ -42,6 +42,24 @@ bool GenIntrinsicsTTIImpl::isLoweredToCall(const Function *F) {
 // instructions. Set this to false unless IGC legalization can fix them.
 bool GenIntrinsicsTTIImpl::shouldBuildLookupTables() { return false; }
 
+bool GenIntrinsicsTTIImpl::enablePromoteLoopUnrollwithAlloca() {
+  const IGC::TriboolFlag RK_PromoteLoopUnrollwithAlloca =
+      static_cast<TriboolFlag>(IGC_GET_FLAG_VALUE(ForcePromoteLoopUnrollwithAlloca));
+  switch (RK_PromoteLoopUnrollwithAlloca) {
+  case TriboolFlag::Enabled:
+    return true;
+  case TriboolFlag::Disabled:
+    return false;
+  default:
+    if (ctx->type == ShaderType::OPENCL_SHADER)
+      return false;
+    if (!ctx->platform.isCoreChildOf(IGFX_XE2_HPG_CORE))
+      return false;
+
+    return true;
+  }
+}
+
 void *GenIntrinsicsTTIImpl::getAdjustedAnalysisPointer(const void *ID) {
   if (ID == &TargetTransformInfoWrapperPass::ID)
     return (TargetTransformInfo *)this;
@@ -318,7 +336,7 @@ void GenIntrinsicsTTIImpl::getUnrollingPreferences(Loop *L, ScalarEvolution &SE,
   //       It can potentially do some global cost estimations.
   const unsigned UnrollMaxCountForAlloca = IGC_GET_FLAG_VALUE(PromoteLoopUnrollwithAllocaCountThreshold);
   bool AllocaFound = false;
-  if (MaxTripCount && MaxTripCount <= UnrollMaxCountForAlloca ) {
+  if (MaxTripCount && MaxTripCount <= UnrollMaxCountForAlloca) {
     unsigned int ThresholdBoost = 0;
     for (auto BB : L->blocks()) {
       for (auto &I : *BB) {
@@ -357,13 +375,12 @@ void GenIntrinsicsTTIImpl::getUnrollingPreferences(Loop *L, ScalarEvolution &SE,
       }
     }
     if (AllocaFound) {
-      // LLVM default only to 10, boost to UnrollMaxCountForAlloca
-      UP.MaxIterationsCountToAnalyze = UnrollMaxCountForAlloca;
       UP.UpperBound = true;
       UP.Force = UnrollLoopForCodeSizeOnly ? false : true;
 
-      if (IGC_IS_FLAG_ENABLED(EnablePromoteLoopUnrollwithAlloca) &&
-          ctx->type != ShaderType::OPENCL_SHADER) {
+      if (enablePromoteLoopUnrollwithAlloca()){
+        // LLVM default only to 10, boost to UnrollMaxCountForAlloca
+        UP.MaxIterationsCountToAnalyze = UnrollMaxCountForAlloca;
         UP.Threshold += ThresholdBoost;
         LLVM_DEBUG(dbgs() << "Increasing L:" << L->getName() << " threshold to " << UP.Threshold
                           << " due to Alloca accessed by:");
@@ -648,7 +665,7 @@ llvm::InstructionCost GenIntrinsicsTTIImpl::internalCalculateCost(const User *U,
     }
   }
 
-  if (IGC_IS_FLAG_ENABLED(EnablePromoteLoopUnrollwithAlloca)) {
+  if (enablePromoteLoopUnrollwithAlloca()) {
     const GetElementPtrInst *GEP = nullptr;
     if (Operator::getOpcode(U) == Instruction::Load)
       GEP = dyn_cast<GetElementPtrInst>(cast<LoadInst>(U)->getPointerOperand());

IGC/Compiler/GenTTI.h

@@ -32,6 +32,7 @@ class GenIntrinsicsTTIImpl : public IGCLLVM::TTIImplCRTPBase<GenIntrinsicsTTIImp
   DenseMap<Value *, bool> isGEPLoopInduction;
 
   bool shouldBuildLookupTables();
+  bool enablePromoteLoopUnrollwithAlloca();
 
   bool isLoweredToCall(const Function *F);
 

IGC/common/igc_flags.h

@@ -437,7 +437,7 @@ DECLARE_IGC_REGKEY(
     true)
 DECLARE_IGC_REGKEY(bool, UnrollLoopForCodeSizeOnly, false,
                    "Only unroll the loop if it can reduce program size/register pressure. Ignore all other threshold "
-                   "setting but still enable EnablePromoteLoopUnrollwithAlloca due to high likelyhood to reduce size.",
+                   "setting but still enable PromoteLoopUnrollwithAlloca due to high likelyhood to reduce size.",
                    true)
 DECLARE_IGC_REGKEY(DWORD, SetLoopUnrollThreshold, 0,
                    "Set the loop unroll threshold. Value 0 will use the default threshold.", false)
@@ -448,12 +448,14 @@ DECLARE_IGC_REGKEY(DWORD, SetLoopUnrollMaxPercentThresholdBoostForHighRegPressur
                    "Set the loop unroll max allowed threshold boost in percentage for shaders with high reg pressure. "
                    "The LLVM internal value is 400.",
                    false)
-DECLARE_IGC_REGKEY(
-    bool, EnablePromoteLoopUnrollwithAlloca, false,
+DECLARE_IGC_REGKEY_ENUM(ForcePromoteLoopUnrollwithAlloca, -1,
     "Loop cost estimation assumes Load/Store who accesses Alloca with index deductible to loop count having 0 cost. "
     "Disable this flag makes them always cost something as well as disables dynamic threshold increase based on the "
-    "size of alloca and number of GEP to the alloca in the loop, leading to the loop less likely to be unrolled.",
-    false)
+    "size of alloca and number of GEP to the alloca in the loop, leading to the loop less likely to be unrolled."
+    "-1 - default behavior, decided by platforms"
+    " 0 - force disabled"
+    " 1 - force enabled",
+    TRIBOOL_OPTIONS, false)
 DECLARE_IGC_REGKEY(DWORD, PromoteLoopUnrollwithAllocaCountThreshold, 256,
                    "The loop trip count OR number of alloca elements cutoff to stop regkey "
                    "EnablePromoteLoopUnrollwithAlloca (Check regkey description).",