WIA without LCSSA

This is the solution when we do WIAnalysis without requiring LCSSA
form. I have added detailed comment to explain what we do in
this situation

Author: Gang Y Chen

IGC/Compiler/CISACodeGen/ConstantCoalescing.hpp

@@ -19,6 +19,7 @@ SPDX-License-Identifier: MIT
 #include <llvm/IR/PassManager.h>
 #include <llvm/IR/IRBuilder.h>
 #include <llvm/ADT/SmallBitVector.h>
+#include <llvmWrapper/Transforms/Utils.h>
 #include "common/LLVMWarningsPop.hpp"
 #include "common/IGCIRBuilder.h"
 

IGC/Compiler/CISACodeGen/ShaderCodeGen.cpp

@@ -266,6 +266,7 @@ static void AddAnalysisPasses(CodeGenContext& ctx, IGCPassManager& mpm)
     mpm.add(new Legalizer::PeepholeTypeLegalizer());
 
     // need this before WIAnalysis:
+    // insert phi to prevent changing of WIAnalysis result by later code-motion
     mpm.add(llvm::createLCSSAPass());
     // Fixup extract value pairs.
     mpm.add(createExtractValuePairFixupPass());
@@ -308,8 +309,6 @@ static void AddAnalysisPasses(CodeGenContext& ctx, IGCPassManager& mpm)
         mpm.add(createPromoteMemoryToRegisterPass());
         // Resolving private memory allocas
         mpm.add(CreatePrivateMemoryResolution());
-        // need this before WIAnalysis:
-        mpm.add(llvm::createLCSSAPass());
     }
 
     // This is for dumping register pressure info
@@ -535,8 +534,6 @@ static void AddLegalizationPasses(CodeGenContext& ctx, IGCPassManager& mpm, PSSi
               IGC_IS_FLAG_ENABLED(LateInlineUnmaskedFunc)))
         {
             mpm.add(createBreakCriticalEdgesPass());
-            // need this before WIAnalysis:
-            mpm.add(llvm::createLCSSAPass());
             mpm.add(createAnnotateUniformAllocasPass());
 
             if (IGC_IS_FLAG_DISABLED(DisablePromotePrivMem) &&
@@ -604,8 +601,7 @@ static void AddLegalizationPasses(CodeGenContext& ctx, IGCPassManager& mpm, PSSi
     if (!isOptDisabled &&
         ctx.m_instrTypes.hasLoadStore && IGC_IS_FLAG_DISABLED(DisableMemOpt)) {
         // run AdvMemOpt and MemOPt back-to-back so that we only
-        // need to run WIAnalysis once, also need LCSSA before WIAnalysis:
-        mpm.add(llvm::createLCSSAPass());
+        // need to run WIAnalysis once
         if (IGC_IS_FLAG_ENABLED(EnableAdvMemOpt))
             mpm.add(createAdvMemOptPass());
         bool AllowNegativeSymPtrsForLoad =
@@ -772,8 +768,6 @@ static void AddLegalizationPasses(CodeGenContext& ctx, IGCPassManager& mpm, PSSi
     if (IGC_IS_FLAG_DISABLED(DisableConstantCoalescing))
     {
         mpm.add(createBreakCriticalEdgesPass());
-        // need this before WIAnalysis:
-        mpm.add(llvm::createLCSSAPass());
         mpm.add(new ConstantCoalescing());
     }
 

IGC/Compiler/CISACodeGen/WIAnalysis.cpp

@@ -946,18 +946,23 @@ void WIAnalysisRunner::update_cf_dep(const IGCLLVM::TerminatorInst* inst)
                     m_pChangedNew->push_back(it->second);
             }
 
-            // Experiment to see when the code after this test is needed
-            if (IGC_IS_FLAG_ENABLED(WIAExperimental))
-                continue;
-
+            // This is an optimization that tries to detect instruction
+            // not really affected by control-flow divergency because
+            // all the sources are outside the region.
+            // However this is only as good as we can get because we
+            // only search limited depth
             if (isRegionInvariant(defi, &br_info, 0))
             {
                 continue;
             }
-            // look at the uses
+            // We need to look at where the use is in order to decide
+            // we should make def to be "random" when loop is not in
+            // LCSSA form because we do not have LCSSA phi-nodes.
+            // 1) if use is in the full-join
+            // 2) if use is even outside the full-join
+            // 3) if use is in partial-join but def is not in partial-join
             Value::use_iterator use_it = defi->use_begin();
             Value::use_iterator use_e = defi->use_end();
-
             for (; use_it != use_e; ++use_it)
             {
                 Instruction* user = dyn_cast<Instruction>((*use_it).getUser());
@@ -966,32 +971,20 @@ void WIAnalysisRunner::update_cf_dep(const IGCLLVM::TerminatorInst* inst)
                 PHINode* phi = dyn_cast<PHINode>(user);
                 if (phi)
                 {
-                    // another place we assume all critical edges have been split and
-                    // phi-move will be placed on the blocks created on those
+                    // another place we assume all critical edges have been
+                    // split and phi-move will be placed on those splitters
                     user_blk = phi->getIncomingBlock(*use_it);
                 }
-                auto canUserBeHoistedToDefBlock = [](Instruction* UI, Instruction* DI)->bool {
-                    for (auto& uo : UI->operands())
-                    {
-                        Value* V = uo.get();
-                        if (V == DI)
-                            continue;
-                        else if (dyn_cast<Constant>(V))
-                            continue;
-                        else
-                            return false;
-                    }
-                    return true;
-                };
-                if ((user_blk == def_blk) ||
-                    (canUserBeHoistedToDefBlock(user, defi) && !phi))
+                if (user_blk == def_blk)
                 {
                     // local def-use, not related to control-dependence
                     continue; // skip
                 }
                 if (user_blk == br_info.full_join ||
-                    br_info.partial_joins.count(user_blk) ||
-                    !br_info.influence_region.count(user_blk))
+                    !br_info.influence_region.count(user_blk) ||
+                    (br_info.partial_joins.count(user_blk) &&
+                     !br_info.partial_joins.count(def_blk))
+                   )
                 {
                     updateDepMap(defi, instDep);
                     // break out of the use loop

IGC/Compiler/CISACodeGen/WIAnalysis.hpp

@@ -28,7 +28,6 @@ SPDX-License-Identifier: MIT
 #include <llvm/IR/InstIterator.h>
 #include <llvm/IR/Dominators.h>
 #include <llvm/Analysis/PostDominators.h>
-#include <llvmWrapper/Transforms/Utils.h>
 #include "common/LLVMWarningsPop.hpp"
 
 
@@ -327,8 +326,6 @@ namespace IGC
             AU.addRequired<MetaDataUtilsWrapper>();
             AU.addRequired<CodeGenContextWrapper>();
             AU.addRequired<TranslationTable>();
-            AU.addRequiredID(llvm::LCSSAID);
-            AU.addPreservedID(llvm::LCSSAID);
         }
 
         /// @brief LLVM llvm::Function pass entry

IGC/common/igc_flags.h

@@ -241,8 +241,6 @@ DECLARE_IGC_REGKEY(bool, EnablerReadSuppressionWA,    true, "Enable read suppres
 DECLARE_IGC_REGKEY(bool, DPASReadSuppressionWA,    true, "Enable read suppression WA for the send and indirect access", false)
 DECLARE_IGC_REGKEY(DWORD, RSWARegNum,    0, "dummy register used for EnablerReadSuppressionWA", false)
 DECLARE_IGC_REGKEY(bool, EnableDivergentBarrierCheck,   false,  "Uses WIAnalysis to find barriers in divergent flow control. May have false positives.", false)
-DECLARE_IGC_REGKEY(bool, WIAExperimental, false,  "Experimental flag to find certain check in WIA is really needed or not", false)
-
 DECLARE_IGC_GROUP("Shader dumping")
 DECLARE_IGC_REGKEY(bool, EnableCosDump, false, "Enable cos dump", true)
 DECLARE_IGC_REGKEY(bool, EnableCisDump, false, "Enable cis dump", true)