This pass is after ProgramScopeConstantAnalysis, which propagates llvm.used GV. Because of that, there is one additional Constant user (not GEPs), and pass would stop early, so let's ignore this specific one. Also added flag to ignore loop requirement. (disabled by default) With this and with modifying ConstantPromotionSize, we can promote bigger tables to registers for experiments.

Change-Id: I9f4dee238aefeb8faf65cbb3ae2b39ebea2017f0

Author: bgajdaINTC

IGC/Compiler/CISACodeGen/SimplifyConstant.cpp

@@ -260,9 +260,29 @@ FunctionPass* IGC::createPromoteConstantPass() { return new PromoteConstant(); }
 //
 // We only promote when there is load of GV inside loops.
 //
-static bool checkUses(GlobalVariable* GV, const Function* F, LoopInfo& LI) {
+static bool checkUses(GlobalVariable* GV, const Function* F, LoopInfo& LI, const ConstantArray* llvm_used) {
     bool LoadInLoop = false;
     for (auto U : GV->users()) {
+        // Because of ProgramScopeConstantAnalysis there might be user in llvm.used
+        // let's ignore this one, as it doesn't prevent opt.
+        // TODO: is there any other case of Constant to worry about. (?)
+        if (auto Const = dyn_cast<Constant>(U))
+        {
+            bool foundInLLVMUsed = false;
+            unsigned numOperands = llvm_used ? llvm_used->getNumOperands() : 0;
+            for (unsigned i = 0; i != numOperands; ++i)
+            {
+                Value* gvi = llvm_used->getOperand(i)->stripPointerCastsNoFollowAliases();
+                if (gvi == Const->stripPointerCastsNoFollowAliases())
+                {
+                    foundInLLVMUsed = true;
+                    break;
+                }
+            }
+            if (foundInLLVMUsed)
+                continue;
+        }
+
         auto Inst = dyn_cast<Instruction>(U);
         if (!Inst)
             // TODO: this may be a const expr.
@@ -284,6 +304,9 @@ static bool checkUses(GlobalVariable* GV, const Function* F, LoopInfo& LI) {
         }
     }
 
+    if (IGC_IS_FLAG_ENABLED(AllowNonLoopConstantPromotion))
+        return true;
+
     return LoadInLoop;
 }
 
@@ -342,6 +365,11 @@ static bool checkSize(GlobalVariable* GV, VectorType*& DataType,
         N *= VectorSize;
     }
     unsigned VS = getLegalVectorSize(N);
+
+    // Allow experimental overriding for bigger than 32 elem. tables
+    if (IGC_IS_FLAG_ENABLED(AllowNonLoopConstantPromotion))
+        VS = 1U << Log2_32_Ceil(N);
+
     if (VS == 0)
         return false;
 
@@ -409,7 +437,7 @@ static bool checkType(GlobalVariable* GV) {
     if (auto CDA = dyn_cast<ConstantDataArray>(Init))
         return !CDA->isString();
 
-    // Not simply data. Need to check if all emements have the same type in a
+    // Not simply data. Need to check if all elements have the same type in a
     // constant array.
     if (auto CA = dyn_cast<ConstantArray>(Init)) {
         auto EltTy = CA->getType()->getElementType();
@@ -499,8 +527,8 @@ static void promote(GlobalVariable* GV, VectorType* AllocaType, bool IsSigned,
     // Transform all uses
     for (auto UI = GV->user_begin(); UI != GV->user_end(); /*empty*/) {
         auto GEP = dyn_cast<GetElementPtrInst>(*UI++);
-        assert(GEP && "nullptr");
-        if (GEP->getParent()->getParent() != F)
+        // might be Constant user in llvm.used
+        if (!GEP || GEP->getParent()->getParent() != F)
             continue;
         assert(GEP->getNumIndices() == 2);
         // This is the index to address the array, and the first index is to address
@@ -569,8 +597,8 @@ static bool rewriteAsCmpSel(GlobalVariable* GV, Function& F) {
 
     for (auto UI = GV->user_begin(); UI != GV->user_end(); /*empty*/) {
         auto GEP = dyn_cast<GetElementPtrInst>(*UI++);
-        assert(GEP && "nullptr");
-        if (GEP->getParent()->getParent() != &F)
+        // might be Constant user in llvm.used
+        if (!GEP || GEP->getParent()->getParent() != &F)
             continue;
 
         // This is the index to address the array, and the first index is to
@@ -634,14 +662,17 @@ bool PromoteConstant::runOnFunction(Function& F) {
     Module* M = F.getParent();
     bool Changed = false;
 
+    auto gv = M->getGlobalVariable("llvm.used");
+    ConstantArray* llvm_used = gv ? dyn_cast_or_null<ConstantArray>(gv->getInitializer()) : nullptr;
+
     for (auto I = M->global_begin(); I != M->global_end(); /*empty*/) {
         GlobalVariable* GV = &(*I++);
         if (GV->user_empty() || !GV->isConstant() || !GV->hasInitializer() ||
             GV->getType()->getAddressSpace() != ADDRESS_SPACE_CONSTANT)
             continue;
         if (!checkType(GV))
             continue;
-        if (!checkUses(GV, &F, LI))
+        if (!checkUses(GV, &F, LI, llvm_used))
             continue;
 
         // Rewrite as cmp+sel sequence if profitable.

IGC/common/igc_flags.def

@@ -287,6 +287,7 @@ DECLARE_IGC_REGKEY(bool, EnableThreadCombiningWithNoSLM, false, "Enable thread c
 DECLARE_IGC_REGKEY(DWORD, SubroutineThreshold,          110000, "Minimal kernel size to enable subroutines")
 DECLARE_IGC_REGKEY(DWORD, SubroutineInlinerThreshold,   3000, "Subroutine inliner threshold")
 DECLARE_IGC_REGKEY(bool, EnableConstantPromotion,       true, "Enable global constant data to register promotion")
+DECLARE_IGC_REGKEY(bool, AllowNonLoopConstantPromotion, false, "Allows promotion for constants not in loop (e.g. used once)")
 DECLARE_IGC_REGKEY(DWORD, ConstantPromotionSize,        2, "Threshold in number of GRFs")
 DECLARE_IGC_REGKEY(DWORD, ConstantPromotionCmpSelSize,  4, "Array size threshold for cmp-sel transform")
 DECLARE_IGC_REGKEY(bool, EnableVariableReuse,           true, "Enable local variable reuse")