Re-apply r101000, with a fix: Don't eliminate an icmp which is part of

the loop exit test. This usually doesn't come up for a variety of
reasons, but it isn't impossible, so make IndVarSimplify handle it
conservatively.


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@101008 91177308-0d34-0410-b5e6-96231b3b80d8

lib/Transforms/Scalar/IndVarSimplify.cpp

@@ -97,6 +97,7 @@ namespace {
 
   private:
 
+    void EliminateIVComparisons();
     void RewriteNonIntegerIVs(Loop *L);
 
     ICmpInst *LinearFunctionTestReplace(Loop *L, const SCEV *BackedgeTakenCount,
@@ -336,6 +337,40 @@ void IndVarSimplify::RewriteNonIntegerIVs(Loop *L) {
     SE->forgetLoop(L);
 }
 
+void IndVarSimplify::EliminateIVComparisons() {
+  // Look for ICmp users.
+  for (IVUsers::iterator I = IU->begin(), E = IU->end(); I != E;) {
+    IVStrideUse &UI = *I++;
+    ICmpInst *ICmp = dyn_cast<ICmpInst>(UI.getUser());
+    if (!ICmp) continue;
+
+    bool Swapped = UI.getOperandValToReplace() == ICmp->getOperand(1);
+    ICmpInst::Predicate Pred = ICmp->getPredicate();
+    if (Swapped) Pred = ICmpInst::getSwappedPredicate(Pred);
+
+    // Get the SCEVs for the ICmp operands.
+    const SCEV *S = IU->getReplacementExpr(UI);
+    const SCEV *X = SE->getSCEV(ICmp->getOperand(!Swapped));
+
+    // Simplify unnecessary loops away.
+    const Loop *ICmpLoop = LI->getLoopFor(ICmp->getParent());
+    S = SE->getSCEVAtScope(S, ICmpLoop);
+    X = SE->getSCEVAtScope(X, ICmpLoop);
+
+    // If the condition is always true or always false, replace it with
+    // a constant value.
+    if (SE->isKnownPredicate(Pred, S, X))
+      ICmp->replaceAllUsesWith(ConstantInt::getTrue(ICmp->getContext()));
+    else if (SE->isKnownPredicate(ICmpInst::getInversePredicate(Pred), S, X))
+      ICmp->replaceAllUsesWith(ConstantInt::getFalse(ICmp->getContext()));
+    else
+      continue;
+
+    DEBUG(dbgs() << "INDVARS: Eliminated comparison: " << *ICmp << '\n');
+    ICmp->eraseFromParent();
+  }
+}
+
 bool IndVarSimplify::runOnLoop(Loop *L, LPPassManager &LPM) {
   IU = &getAnalysis<IVUsers>();
   LI = &getAnalysis<LoopInfo>();
@@ -427,11 +462,20 @@ bool IndVarSimplify::runOnLoop(Loop *L, LPPassManager &LPM) {
       ExitingBlock) {
     assert(NeedCannIV &&
            "LinearFunctionTestReplace requires a canonical induction variable");
+
     // Can't rewrite non-branch yet.
-    if (BranchInst *BI = dyn_cast<BranchInst>(ExitingBlock->getTerminator()))
+    if (BranchInst *BI = dyn_cast<BranchInst>(ExitingBlock->getTerminator())) {
+      // Eliminate comparisons which are always true or always false, due to
+      // the known backedge-taken count. This may include comparisons which
+      // are currently controlling (part of) the loop exit, so we can only do
+      // it when we know we're going to insert our own loop exit code.
+      EliminateIVComparisons();
+
+      // Insert new loop exit code.
       NewICmp = LinearFunctionTestReplace(L, BackedgeTakenCount, IndVar,
                                           ExitingBlock, BI, Rewriter);
     }
+  }
 
   // Rewrite IV-derived expressions. Clears the rewriter cache.
   RewriteIVExpressions(L, Rewriter);

test/Transforms/IndVarSimplify/eliminate-comparison.ll

@@ -0,0 +1,84 @@
+; RUN: opt -indvars -S < %s | FileCheck %s
+
+target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16:32:64"
+
+@X = external global [0 x double]
+
+; Indvars should be able to simplify simple comparisons involving
+; induction variables.
+
+; CHECK: @foo
+; CHECK: %cond = and i1 %tobool.not, true
+
+define void @foo(i64 %n, i32* nocapture %p) nounwind {
+entry:
+  %cmp9 = icmp sgt i64 %n, 0
+  br i1 %cmp9, label %pre, label %return
+
+pre:
+  %t3 = load i32* %p
+  %tobool.not = icmp ne i32 %t3, 0
+  br label %loop
+
+loop:
+  %i = phi i64 [ 0, %pre ], [ %inc, %for.inc ]
+  %cmp6 = icmp slt i64 %i, %n
+  %cond = and i1 %tobool.not, %cmp6
+  br i1 %cond, label %if.then, label %for.inc
+
+if.then:
+  %arrayidx = getelementptr [0 x double]* @X, i64 0, i64 %i
+  store double 3.200000e+00, double* %arrayidx
+  br label %for.inc
+
+for.inc:
+  %inc = add nsw i64 %i, 1
+  %exitcond = icmp sge i64 %inc, %n
+  br i1 %exitcond, label %return, label %loop
+
+return:
+  ret void
+}
+
+; Don't eliminate an icmp that's contributing to the loop exit test though.
+
+; CHECK: @_ZNK4llvm5APInt3ultERKS0_
+; CHECK: %tmp99 = icmp sgt i32 %i, -1
+
+define i32 @_ZNK4llvm5APInt3ultERKS0_(i32 %tmp2.i1, i64** %tmp65, i64** %tmp73, i64** %tmp82, i64** %tmp90) {
+entry:
+  br label %bb18
+
+bb13:
+  %tmp66 = load i64** %tmp65, align 4
+  %tmp68 = getelementptr inbounds i64* %tmp66, i32 %i
+  %tmp69 = load i64* %tmp68, align 4
+  %tmp74 = load i64** %tmp73, align 4
+  %tmp76 = getelementptr inbounds i64* %tmp74, i32 %i
+  %tmp77 = load i64* %tmp76, align 4
+  %tmp78 = icmp ugt i64 %tmp69, %tmp77
+  br i1 %tmp78, label %bb20.loopexit, label %bb15
+
+bb15:
+  %tmp83 = load i64** %tmp82, align 4
+  %tmp85 = getelementptr inbounds i64* %tmp83, i32 %i
+  %tmp86 = load i64* %tmp85, align 4
+  %tmp91 = load i64** %tmp90, align 4
+  %tmp93 = getelementptr inbounds i64* %tmp91, i32 %i
+  %tmp94 = load i64* %tmp93, align 4
+  %tmp95 = icmp ult i64 %tmp86, %tmp94
+  br i1 %tmp95, label %bb20.loopexit, label %bb17
+
+bb17:
+  %tmp97 = add nsw i32 %i, -1
+  br label %bb18
+
+bb18:
+  %i = phi i32 [ %tmp2.i1, %entry ], [ %tmp97, %bb17 ]
+  %tmp99 = icmp sgt i32 %i, -1
+  br i1 %tmp99, label %bb13, label %bb20.loopexit
+
+bb20.loopexit:
+  %tmp.0.ph = phi i32 [ 0, %bb18 ], [ 1, %bb15 ], [ 0, %bb13 ]
+  ret i32 %tmp.0.ph
+}