Enable Clone Address Arithmetic pass on retry & RPE cutoff

Original pass: 131312e
Clone address arithmetic pass is enabled on retry. RPE is used as a
cutoff to prevent unnecessary rematerialization on kernels that no
longer have spills.

Author: esukhov

IGC/Compiler/CISACodeGen/IGCLivenessAnalysis.cpp

@@ -38,6 +38,9 @@ IGC_INITIALIZE_PASS_END(IGCLivenessAnalysis, PASS_FLAG2, PASS_DESCRIPTION2,
                         PASS_CFG_ONLY2, PASS_ANALYSIS2)
 #define PRINT(str) llvm::errs() << str
 
+
+IGCLivenessAnalysis::IGCLivenessAnalysis() : FunctionPass(ID) { UseWIAnalysis = true; };
+
 IGCLivenessAnalysis::IGCLivenessAnalysis(bool UseWiAnalysisVar /* = false */,
                                          bool DumpToFileVar /* = false */,
                                          const std::string &DumpFileNameVar)

IGC/Compiler/CISACodeGen/IGCLivenessAnalysis.h

@@ -55,9 +55,11 @@ class IGCLivenessAnalysis : public llvm::FunctionPass {
     llvm::StringRef getPassName() const override {
         return "IGCLivenessAnalysis";
     }
+
+    IGCLivenessAnalysis();
     IGCLivenessAnalysis(
-        bool UseWIAnalysis = false, bool DumpToFile = false,
-        const std::string &DumpFileName = "RegPressureEstimate.ll");
+        bool UseWIAnalysis, bool DumpToFile,
+        const std::string &DumpFileName);
 
     virtual ~IGCLivenessAnalysis() {}
 

IGC/Compiler/CISACodeGen/RematAddressArithmetic.cpp

@@ -14,6 +14,7 @@ SPDX-License-Identifier: MIT
 #include "llvm/Analysis/PostDominators.h"
 #include "llvm/ADT/BreadthFirstIterator.h"
 #include "common/LLVMWarningsPop.hpp"
+#include "Compiler/CISACodeGen/IGCLivenessAnalysis.h"
 
 using namespace llvm;
 using namespace IGC;
@@ -60,6 +61,7 @@ class CloneAddressArithmetic : public FunctionPass {
 
     virtual void getAnalysisUsage(llvm::AnalysisUsage& AU) const override {
         AU.setPreservesCFG();
+        AU.addRequired<IGCLivenessAnalysis>();
     }
 
     bool runOnFunction(Function&) override;
@@ -88,6 +90,7 @@ char CloneAddressArithmetic::ID = 0;
 #define PASS_ANALYSIS_2 false
 namespace IGC {
 IGC_INITIALIZE_PASS_BEGIN(CloneAddressArithmetic, PASS_FLAG_2, PASS_DESC_2, PASS_CFG_ONLY_2, PASS_ANALYSIS_2)
+IGC_INITIALIZE_PASS_DEPENDENCY(IGCLivenessAnalysis)
 IGC_INITIALIZE_PASS_END(CloneAddressArithmetic, PASS_FLAG_2, PASS_DESC_2, PASS_CFG_ONLY_2, PASS_ANALYSIS_2)
 }
 
@@ -184,6 +187,12 @@ bool CloneAddressArithmetic::greedyRemat(Function &F) {
 
   bool Result = false;
 
+  auto RPE = &getAnalysis<IGCLivenessAnalysis>();
+  unsigned int SIMD = numLanes(RPE->bestGuessSIMDSize());
+  unsigned int PressureLimit = IGC_GET_FLAG_VALUE(RematRPELimit);
+  if(RPE->getMaxRegCountForFunction(F, SIMD) < PressureLimit)
+      return Result;
+
   for (BasicBlock &BB : F) {
     for (auto &I : BB) { Uses[&I] = I.getNumUses(); }
   }

IGC/Compiler/CISACodeGen/ShaderCodeGen.cpp

@@ -893,7 +893,8 @@ void AddLegalizationPasses(CodeGenContext& ctx, IGCPassManager& mpm, PSSignature
     // Run address remat after GVN as it may hoist address calculations and
     // create PHI nodes with addresses.
 
-    if (IGC_IS_FLAG_ENABLED(EnableRemat)) {
+    if (IGC_IS_FLAG_ENABLED(EnableRemat) || (ctx.m_retryManager.AllowCloneAddressArithmetic() && ctx.type == ShaderType::OPENCL_SHADER)) {
+
 
       // TODO: This is a workaround that helps to reduce amount of instructions for clone address arithmetic
       // it helps with chain of instructions like this

IGC/Compiler/tests/RematAddressArithmetic/cloneAddressArithmetic.ll

@@ -6,7 +6,8 @@
 ;
 ;============================ end_copyright_notice =============================
 
-; RUN: igc_opt %s -S -o - -igc-clone-address-arithmetic --regkey=RematChainLimit=10 --dce | FileCheck %s
+; UNSUPPORTED: system-windows
+; RUN: igc_opt %s -S -o - -igc-clone-address-arithmetic --regkey=RematChainLimit=10 --regkey=RematRPELimit=0 --dce | FileCheck %s
 
 define spir_kernel void @main(double addrspace(1)* %base, i64 %offset, i64 %I, i64 %J) {
 ; CHECK: define spir_kernel void @main(double addrspace(1)* [[BASE:%.*]], i64 [[OFFSET:%.*]], i64 %I, i64 %J)

IGC/common/igc_flags.h

@@ -736,6 +736,7 @@ DECLARE_IGC_REGKEY(DWORD, MemOptWindowSize,   150,  "Size of the window in unit
 DECLARE_IGC_REGKEY(DWORD, RematBlockSize,   10,  "Represents a threshold for a basic block size which determines whether this block will be processed for rematerialization or not", false)
 DECLARE_IGC_REGKEY(DWORD, RematUsesThreshold,   5,  "Amount of uses after which operand is not rematerialized", false)
 DECLARE_IGC_REGKEY(DWORD, RematChainLimit,   10,  "If number of instructions we've collected is more than this value, we bail on it", false)
+DECLARE_IGC_REGKEY(DWORD, RematRPELimit,   100,  "Cutoff value for register estimator, lower than that, kernel won't be rematted", false)
 DECLARE_IGC_REGKEY(bool, RematSameBBScope, true, "Confine rematerialization only to variables within the same BB, we won't pull down values from predeccors", false)
 DECLARE_IGC_REGKEY(bool, EnableRemat, false, "Enable clone adress arithmetic pass not only on retry", false)
 DECLARE_IGC_REGKEY(bool, RematReassocBefore, false, "Enable short sequence of passes before clone address arithmetic pass to potentially decrese amount of operations that will be rematerialized", false)