From fa151a79ed2d0efc8fe1d7378be264b301d8caf1 Mon Sep 17 00:00:00 2001
From: Nick Lewycky <nicholas@mxc.ca>
Date: Tue, 6 Sep 2011 05:05:14 +0000
Subject: [PATCH] The logic inside getMulExpr to simplify {a,+,b}*{c,+,d} was
 wrong, which was visible given a=b=c=d=1, on iteration #1 (the second
 iteration). Replace it with correct math. Fixes PR10383!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@139133 91177308-0d34-0410-b5e6-96231b3b80d8
---
 lib/Analysis/ScalarEvolution.cpp              | 33 ++++++++++-------
 .../ScalarEvolution/SolveQuadraticEquation.ll | 37 +++++++++++++++++--
 2 files changed, 54 insertions(+), 16 deletions(-)

diff --git a/lib/Analysis/ScalarEvolution.cpp b/lib/Analysis/ScalarEvolution.cpp
index 8859c3b22fb..6d28f6912fd 100644
--- a/lib/Analysis/ScalarEvolution.cpp
+++ b/lib/Analysis/ScalarEvolution.cpp
@@ -1735,7 +1735,7 @@ const SCEV *ScalarEvolution::getAddExpr(SmallVectorImpl<const SCEV *> &Ops,
       // If all of the other operands were loop invariant, we are done.
       if (Ops.size() == 1) return NewRec;
 
-      // Otherwise, add the folded AddRec by the non-liv parts.
+      // Otherwise, add the folded AddRec by the non-live parts.
       for (unsigned i = 0;; ++i)
         if (Ops[i] == AddRec) {
           Ops[i] = NewRec;
@@ -1960,7 +1960,7 @@ const SCEV *ScalarEvolution::getMulExpr(SmallVectorImpl<const SCEV *> &Ops,
       // If all of the other operands were loop invariant, we are done.
       if (Ops.size() == 1) return NewRec;
 
-      // Otherwise, multiply the folded AddRec by the non-liv parts.
+      // Otherwise, multiply the folded AddRec by the non-live parts.
       for (unsigned i = 0;; ++i)
         if (Ops[i] == AddRec) {
           Ops[i] = NewRec;
@@ -1976,22 +1976,29 @@ const SCEV *ScalarEvolution::getMulExpr(SmallVectorImpl<const SCEV *> &Ops,
          OtherIdx < Ops.size() && isa<SCEVAddRecExpr>(Ops[OtherIdx]);
          ++OtherIdx)
       if (AddRecLoop == cast<SCEVAddRecExpr>(Ops[OtherIdx])->getLoop()) {
-        // F * G, where F = {A,+,B}<L> and G = {C,+,D}<L>  -->
-        // {A*C,+,F*D + G*B + B*D}<L>
+        // {A,+,B}<L> * {C,+,D}<L>  -->  {A*C,+,A*D + B*C + B*D,+,2*B*D}<L>
         for (; OtherIdx != Ops.size() && isa<SCEVAddRecExpr>(Ops[OtherIdx]);
              ++OtherIdx)
           if (const SCEVAddRecExpr *OtherAddRec =
                 dyn_cast<SCEVAddRecExpr>(Ops[OtherIdx]))
             if (OtherAddRec->getLoop() == AddRecLoop) {
-              const SCEVAddRecExpr *F = AddRec, *G = OtherAddRec;
-              const SCEV *NewStart = getMulExpr(F->getStart(), G->getStart());
-              const SCEV *B = F->getStepRecurrence(*this);
-              const SCEV *D = G->getStepRecurrence(*this);
-              const SCEV *NewStep = getAddExpr(getMulExpr(F, D),
-                                               getMulExpr(G, B),
-                                               getMulExpr(B, D));
-              const SCEV *NewAddRec = getAddRecExpr(NewStart, NewStep,
-                                                    F->getLoop(),
+              const SCEV *A = AddRec->getStart();
+              const SCEV *B = AddRec->getStepRecurrence(*this);
+              const SCEV *C = OtherAddRec->getStart();
+              const SCEV *D = OtherAddRec->getStepRecurrence(*this);
+              const SCEV *NewStart = getMulExpr(A, C);
+              const SCEV *BD = getMulExpr(B, D);
+              const SCEV *NewStep = getAddExpr(getMulExpr(A, D),
+                                               getMulExpr(B, C), BD);
+              const SCEV *NewSecondOrderStep =
+                  getMulExpr(BD, getConstant(BD->getType(), 2));
+
+              SmallVector<const SCEV *, 3> AddRecOps;
+              AddRecOps.push_back(NewStart);
+              AddRecOps.push_back(NewStep);
+              AddRecOps.push_back(NewSecondOrderStep);
+              const SCEV *NewAddRec = getAddRecExpr(AddRecOps,
+                                                    AddRec->getLoop(),
                                                     SCEV::FlagAnyWrap);
               if (Ops.size() == 2) return NewAddRec;
               Ops[Idx] = AddRec = cast<SCEVAddRecExpr>(NewAddRec);
diff --git a/test/Analysis/ScalarEvolution/SolveQuadraticEquation.ll b/test/Analysis/ScalarEvolution/SolveQuadraticEquation.ll
index ec95141fbe9..2c4ef16d2e3 100644
--- a/test/Analysis/ScalarEvolution/SolveQuadraticEquation.ll
+++ b/test/Analysis/ScalarEvolution/SolveQuadraticEquation.ll
@@ -1,11 +1,11 @@
 ; RUN: opt < %s -analyze -scalar-evolution \
-; RUN:   -scalar-evolution-max-iterations=0 | grep {backedge-taken count is 100}
+; RUN:   -scalar-evolution-max-iterations=0 | FileCheck %s
+
 ; PR1101
 
 @A = weak global [1000 x i32] zeroinitializer, align 32         
 
-
-define void @test(i32 %N) {
+define void @test1(i32 %N) {
 entry:
         %"alloca point" = bitcast i32 0 to i32           ; <i32> [#uses=0]
         br label %bb3
@@ -30,3 +30,34 @@ bb5:            ; preds = %bb3
 return:         ; preds = %bb5
         ret void
 }
+; CHECK: Determining loop execution counts for: @test1
+; CHECK-NEXT: backedge-taken count is 100
+
+
+; PR10383
+; This used to crash.
+
+define void @test2(i1 %cmp, i64 %n) {
+entry:
+  br label %for.body1
+
+for.body1:
+  %a0.08 = phi i64 [ 0, %entry ], [ %inc512, %for.body1 ]
+  %inc512 = add i64 %a0.08, 1
+  br i1 %cmp, label %preheader, label %for.body1
+
+preheader:
+  br label %for.body2
+
+for.body2:
+  %indvar = phi i64 [ 0, %preheader ], [ %indvar.next, %for.body2 ]
+  %tmp111 = add i64 %n, %indvar
+  %tmp114 = mul i64 %a0.08, %indvar
+  %mul542 = mul i64 %tmp114, %tmp111
+  %indvar.next = add i64 %indvar, 1
+  br i1 undef, label %end, label %for.body2
+
+end:
+  ret void
+}
+; CHECK: Determining loop execution counts for: @test2