[SeparateConstOffsetFromGEP] sext(a)+sext(b) => sext(a+b) when a+b can't sign-overflow.

[jingyue]

Summary

This patch implements my promised optimization to reunites certain sexts from
operands after we extract the constant offset. See the header comment of
reuniteExts for its motivation.

One key building block that enables this optimization is Bjarke's poison value
analysis (D11212). That helps to prove "a +nsw b" can't overflow.

[broune]

lib/Transforms/Scalar/SeparateConstOffsetFromGEP.cpp

1103

a -> an

1106

I think that you also need to check that I has a NSW or NUW flag.

I wonder if it would be faster to insert all the add/sub dominators with such a flag and then check isKnownNotFullPoison(I) only when an instruction is returned from findClosestMatchingDominator().

[jingyue]

lib/Transforms/Scalar/SeparateConstOffsetFromGEP.cpp

1106

I did check that I has a NSW flag by using m_NSWAdd and m_NSWSub.

It's hard to tell. In the current way, the hash table is smaller and findClosestMatchingDominator more likely returns nullptr. Though I agree it calls isKnownNotFullPoison more if matched SCEVs are rare.

[broune]

LGTM.

[jingyue]

fixed a typo and add more comments on the test

[jingyue]

rebase against master

lib/Transforms/Scalar/SeparateConstOffsetFromGEP.cpp

 // parts and variant parts, LICM can hoist/sink those invariant parts.
 // CGP (CodeGen Prepare) tries to sink address calculations that match the
 // target's addressing modes. A GEP with multiple indices may not match and will
 // not be sunk. If we lower such GEP into smaller parts, CGP may sink some of
 // them. So we end up with a better addressing mode.
 //
 //===----------------------------------------------------------------------===//
 
+#include "llvm/Analysis/ScalarEvolution.h"
 #include "llvm/Analysis/TargetTransformInfo.h"
 #include "llvm/Analysis/ValueTracking.h"
 #include "llvm/IR/Constants.h"
 #include "llvm/IR/DataLayout.h"
 #include "llvm/IR/Dominators.h"
 #include "llvm/IR/Instructions.h"
 #include "llvm/IR/LLVMContext.h"
 #include "llvm/IR/Module.h"
+#include "llvm/IR/PatternMatch.h"
 #include "llvm/IR/Operator.h"
 #include "llvm/Support/CommandLine.h"
 #include "llvm/Support/raw_ostream.h"
 #include "llvm/Transforms/Scalar.h"
 #include "llvm/Transforms/Utils/Local.h"
 #include "llvm/Target/TargetMachine.h"
 #include "llvm/Target/TargetSubtargetInfo.h"
 #include "llvm/IR/IRBuilder.h"
 
 using namespace llvm;
+using namespace llvm::PatternMatch;
 
 static cl::opt<bool> DisableSeparateConstOffsetFromGEP(
     "disable-separate-const-offset-from-gep", cl::init(false),
     cl::desc("Do not separate the constant offset from a GEP instruction"),
     cl::Hidden);
 // Setting this flag may emit false positives when the input module already
 // contains dead instructions. Therefore, we set it only in unit tests that are
 // free of dead code.
public:
   SeparateConstOffsetFromGEP(const TargetMachine *TM = nullptr,
                              bool LowerGEP = false)
       : FunctionPass(ID), DL(nullptr), DT(nullptr), TM(TM), LowerGEP(LowerGEP) {
     initializeSeparateConstOffsetFromGEPPass(*PassRegistry::getPassRegistry());
   }
 
   void getAnalysisUsage(AnalysisUsage &AU) const override {
     AU.addRequired<DominatorTreeWrapperPass>();
+    AU.addRequired<ScalarEvolution>();
     AU.addRequired<TargetTransformInfoWrapperPass>();
     AU.setPreservesCFG();
   }
 
   bool doInitialization(Module &M) override {
     DL = &M.getDataLayout();
     return false;
   }
private:
   ///
   /// This canonicalization is very likely already done in clang and
   /// instcombine. Therefore, the program will probably remain the same.
   ///
   /// Returns true if the module changes.
   ///
   /// Verified in @i32_add in split-gep.ll
   bool canonicalizeArrayIndicesToPointerSize(GetElementPtrInst *GEP);
+  /// Optimize sext(a)+sext(b) to sext(a+b) when a+b can't sign overflow.
+  /// SeparateConstOffsetFromGEP distributes a sext to leaves before extracting
+  /// the constant offset. After extraction, it becomes desirable to reunion the
+  /// distributed sexts. For example,
+  ///
+  ///                              &a[sext(i +nsw (j +nsw 5)]
+  ///   => distribute              &a[sext(i) +nsw (sext(j) +nsw 5)]
+  ///   => constant extraction     &a[sext(i) + sext(j)] + 5
+  ///   => reunion                 &a[sext(i +nsw j)] + 5
+  bool reuniteExts(Function &F);
+  /// A helper that reunites sexts in an instruction.
+  bool reuniteExts(Instruction *I);
+  /// Find the closest dominator of <Dominatee> that is equivalent to <Key>.
+  Instruction *findClosestMatchingDominator(const SCEV *Key,
+                                            Instruction *Dominatee);
   /// Verify F is free of dead code.
   void verifyNoDeadCode(Function &F);
 
   const DataLayout *DL;
-  const DominatorTree *DT;
+  DominatorTree *DT;
+  ScalarEvolution *SE;
   const TargetMachine *TM;
   /// Whether to lower a GEP with multiple indices into arithmetic operations or
   /// multiple GEPs with a single index.
   bool LowerGEP;
+  DenseMap<const SCEV *, SmallVector<Instruction *, 2>> DominatingExprs;
 };
 }  // anonymous namespace
 
 char SeparateConstOffsetFromGEP::ID = 0;
 INITIALIZE_PASS_BEGIN(
     SeparateConstOffsetFromGEP, "separate-const-offset-from-gep",
     "Split GEPs to a variadic base and a constant offset for better CSE", false,
     false)
 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
+INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
 INITIALIZE_PASS_END(
     SeparateConstOffsetFromGEP, "separate-const-offset-from-gep",
     "Split GEPs to a variadic base and a constant offset for better CSE", false,
     false)
 
 FunctionPass *
 llvm::createSeparateConstOffsetFromGEPPass(const TargetMachine *TM,
if (isa<SequentialType>(*GTI)) {
         RecursivelyDeleteTriviallyDeadInstructions(OldIdx);
       }
     }
   }
 
   // Clear the inbounds attribute because the new index may be off-bound.
   // e.g.,
   //
-  // b     = add i64 a, 5
-  // addr  = gep inbounds float, float* p, i64 b
+  //   b     = add i64 a, 5
+  //   addr  = gep inbounds float, float* p, i64 b
   //
   // is transformed to:
   //
-  // addr2 = gep float, float* p, i64 a ; inbounds removed
-  // addr  = gep inbounds float, float* addr2, i64 5
+  //   addr2 = gep float, float* p, i64 a ; inbounds removed
+  //   addr  = gep inbounds float, float* addr2, i64 5
   //
   // If a is -4, although the old index b is in bounds, the new index a is
   // off-bound. http://llvm.org/docs/LangRef.html#id181 says "if the
   // inbounds keyword is not present, the offsets are added to the base
   // address with silently-wrapping two's complement arithmetic".
   // Therefore, the final code will be a semantically equivalent.
   //
   // TODO(jingyue): do some range analysis to keep as many inbounds as

 bool SeparateConstOffsetFromGEP::runOnFunction(Function &F) {
   if (skipOptnoneFunction(F))
     return false;
 
   if (DisableSeparateConstOffsetFromGEP)
     return false;
 
   DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
+  SE = &getAnalysis<ScalarEvolution>();
 
   bool Changed = false;
   for (Function::iterator B = F.begin(), BE = F.end(); B != BE; ++B) {
     for (BasicBlock::iterator I = B->begin(), IE = B->end(); I != IE; ) {
       if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I++)) {
         Changed |= splitGEP(GEP);
       }
       // No need to split GEP ConstantExprs because all its indices are constant
       // already.
     }
   }
 
+  Changed |= reuniteExts(F);
+
   if (VerifyNoDeadCode)
     verifyNoDeadCode(F);
 
   return Changed;
 }
 
+Instruction *SeparateConstOffsetFromGEP::findClosestMatchingDominator(
+    const SCEV *Key, Instruction *Dominatee) {
+  auto Pos = DominatingExprs.find(Key);
+  if (Pos == DominatingExprs.end())
+    return nullptr;
+
+  auto &Candidates = Pos->second;
+  // Because we process the basic blocks in pre-order of the dominator tree, a
+  // candidate that doesn't dominate the current instruction won't dominate any
+  // future instruction either. Therefore, we pop it out of the stack. This
+  // optimization makes the algorithm O(n).
+  while (!Candidates.empty()) {
+    Instruction *Candidate = Candidates.back();
+    if (DT->dominates(Candidate, Dominatee))
+      return Candidate;
+    Candidates.pop_back();
+  }
+  return nullptr;
+}
+
+bool SeparateConstOffsetFromGEP::reuniteExts(Instruction *I) {
+  if (!SE->isSCEVable(I->getType()))
+    return false;
+
+  //   Dom: LHS+RHS
+  //   I: sext(LHS)+sext(RHS)
+  // If Dom can't sign overflow and Dom dominates I, optimize I to sext(Dom).
+  // TODO: handle zext
+  Value *LHS = nullptr, *RHS = nullptr;
+  if (match(I, m_Add(m_SExt(m_Value(LHS)), m_SExt(m_Value(RHS)))) ||
+      match(I, m_Sub(m_SExt(m_Value(LHS)), m_SExt(m_Value(RHS))))) {
+    if (LHS->getType() == RHS->getType()) {
+      const SCEV *Key =
+          SE->getAddExpr(SE->getUnknown(LHS), SE->getUnknown(RHS));
+      if (auto *Dom = findClosestMatchingDominator(Key, I)) {
+        Instruction *NewSExt = new SExtInst(Dom, I->getType(), "", I);
+        NewSExt->takeName(I);
+        I->replaceAllUsesWith(NewSExt);
+        RecursivelyDeleteTriviallyDeadInstructions(I);
+        return true;
+      }
+    }
+  }
+
+  // Add I to DominatingExprs if it's an add/sub that can't sign overflow.
+  if (match(I, m_NSWAdd(m_Value(LHS), m_Value(RHS))) ||
+      match(I, m_NSWSub(m_Value(LHS), m_Value(RHS)))) {
+    if (isKnownNotFullPoison(I)) {
+      const SCEV *Key =
+          SE->getAddExpr(SE->getUnknown(LHS), SE->getUnknown(RHS));
+      DominatingExprs[Key].push_back(I);
+    }
+  }
+  return false;
+}
+
+bool SeparateConstOffsetFromGEP::reuniteExts(Function &F) {
+  bool Changed = false;
+  DominatingExprs.clear();
+  for (auto Node = GraphTraits<DominatorTree *>::nodes_begin(DT);
+       Node != GraphTraits<DominatorTree *>::nodes_end(DT); ++Node) {
+    BasicBlock *BB = Node->getBlock();
+    for (auto I = BB->begin(); I != BB->end(); ) {
+      Instruction *Cur = I++;
+      Changed |= reuniteExts(Cur);
+    }
+  }
+  return Changed;
+}
+
 void SeparateConstOffsetFromGEP::verifyNoDeadCode(Function &F) {
   for (auto &B : F) {
     for (auto &I : B) {
       if (isInstructionTriviallyDead(&I)) {
         std::string ErrMessage;
         raw_string_ostream RSO(ErrMessage);
         RSO << "Dead instruction detected!\n" << I << "\n";
         llvm_unreachable(RSO.str().c_str());
       }
     }
   }
 }

test/Transforms/SeparateConstOffsetFromGEP/NVPTX/split-gep-and-gvn.ll

 ; PTX: ld.shared.f32 {{%f[0-9]+}}, {{\[}}[[BASE_REG]]+128{{\]}}
 ; PTX: ld.shared.f32 {{%f[0-9]+}}, {{\[}}[[BASE_REG]]+132{{\]}}
 
 ; IR-LABEL: @sum_of_array4(
 ; IR: [[BASE_PTR:%[a-zA-Z0-9]+]] = getelementptr inbounds [32 x [32 x float]], [32 x [32 x float]] addrspace(3)* @array, i64 0, i64 %{{[a-zA-Z0-9]+}}, i64 %{{[a-zA-Z0-9]+}}
 ; IR: getelementptr inbounds float, float addrspace(3)* [[BASE_PTR]], i64 1
 ; IR: getelementptr inbounds float, float addrspace(3)* [[BASE_PTR]], i64 32
 ; IR: getelementptr inbounds float, float addrspace(3)* [[BASE_PTR]], i64 33
+
+
+; The source code is:
+;   p0 = &input[sext(x + y)];
+;   p1 = &input[sext(x + (y + 5))];
+;
+; Without reuniting extensions, SeparateConstOffsetFromGEP would emit
+;   p0 = &input[sext(x + y)];
+;   t1 = &input[sext(x) + sext(y)];
+;   p1 = &t1[5];
+;
+; With reuniting extensions, it merges p0 and t1 and thus emits
+;   p0 = &input[sext(x + y)];
+;   p1 = &p0[5];
+define void @reunion(i32 %x, i32 %y, float* %input) {
+; IR-LABEL: @reunion(
+; PTX-LABEL: reunion(
+entry:
+  %xy = add nsw i32 %x, %y
+  %0 = sext i32 %xy to i64
+  %p0 = getelementptr inbounds float, float* %input, i64 %0
+  %v0 = load float, float* %p0, align 4
+; PTX: ld.f32 %f{{[0-9]+}}, {{\[}}[[p0:%rd[0-9]+]]{{\]}}
+  call void @use(float %v0)
+
+  %y5 = add nsw i32 %y, 5
+  %xy5 = add nsw i32 %x, %y5
+  %1 = sext i32 %xy5 to i64
+  %p1 = getelementptr inbounds float, float* %input, i64 %1
+; IR: getelementptr inbounds float, float* %p0, i64 5
+  %v1 = load float, float* %p1, align 4
+; PTX: ld.f32 %f{{[0-9]+}}, {{\[}}[[p0]]+20{{\]}}
+  call void @use(float %v1)
+
+  ret void
+}
+
+declare void @use(float)