[llvm] r214423 - Fix ScalarEvolutionExpander when creating a PHI in a block with duplicate predecessors

[Hal Finkel]

Author: hfinkel
Date: Thu Jul 31 14:13:38 2014
New Revision: 214423

URL: http://llvm.org/viewvc/llvm-project?rev=214423&view=rev
Log:
Fix ScalarEvolutionExpander when creating a PHI in a block with duplicate predecessors

It seems that when I fixed this, almost exactly a year ago, I did not quite do
it correctly. When we have duplicate block predecessors, we can indeed not have
different incoming values for the same block, but we *must* have duplicate
entries. So, instead of skipping the duplicates, we explicitly add the
duplicate incoming values.

Fixes PR20442.

Added:
    llvm/trunk/test/CodeGen/PowerPC/pr20442.ll
Modified:
    llvm/trunk/lib/Analysis/ScalarEvolutionExpander.cpp

Modified: llvm/trunk/lib/Analysis/ScalarEvolutionExpander.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Analysis/ScalarEvolutionExpander.cpp?rev=214423&r1=214422&r2=214423&view=diff
==============================================================================
--- llvm/trunk/lib/Analysis/ScalarEvolutionExpander.cpp (original)
+++ llvm/trunk/lib/Analysis/ScalarEvolutionExpander.cpp Thu Jul 31 14:13:38 2014
@@ -1443,8 +1443,12 @@ Value *SCEVExpander::visitAddRecExpr(con
     Constant *One = ConstantInt::get(Ty, 1);
     for (pred_iterator HPI = HPB; HPI != HPE; ++HPI) {
       BasicBlock *HP = *HPI;
-      if (!PredSeen.insert(HP))
+      if (!PredSeen.insert(HP)) {
+        // There must be an incoming value for each predecessor, even the
+        // duplicates!
+        CanonicalIV->addIncoming(CanonicalIV->getIncomingValueForBlock(HP), HP);
         continue;
+      }
 
       if (L->contains(HP)) {
         // Insert a unit add instruction right before the terminator

Added: llvm/trunk/test/CodeGen/PowerPC/pr20442.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/PowerPC/pr20442.ll?rev=214423&view=auto
==============================================================================
--- llvm/trunk/test/CodeGen/PowerPC/pr20442.ll (added)
+++ llvm/trunk/test/CodeGen/PowerPC/pr20442.ll Thu Jul 31 14:13:38 2014
@@ -0,0 +1,79 @@
+; RUN: llc -mcpu=pwr7 < %s | FileCheck %s
+target datalayout = "E-m:e-p:32:32-i64:64-n32"
+target triple = "powerpc-unknown-linux-gnu"
+
+; This code would cause code generation like this after PPCCTRLoops ran:
+;  %indvar = phi i32 [ 0, %for.body ], [ %indvar.next, %if.then6 ]
+;  %j.1.ph13 = phi i32 [ %j.110, %if.then6 ], [ 0, %for.body ], [ 0, %for.body ]
+;  %c.0.ph12 = phi i32 [ %dec, %if.then6 ], [ %2, %for.body ], [ %2, %for.body ]
+; which would fail verification because the created induction variable does not
+; have as many predecessor entries as the other PHIs.
+; CHECK-LABEL: @fn1
+; CHECK: mtctr
+
+%struct.anon = type { i32 }
+%struct.anon.0 = type { i32 }
+
+ at b = common global %struct.anon* null, align 4
+ at a = common global %struct.anon.0* null, align 4
+
+; Function Attrs: nounwind readonly uwtable
+define i32 @fn1() #0 {
+entry:
+  %0 = load %struct.anon** @b, align 4
+  %1 = ptrtoint %struct.anon* %0 to i32
+  %cmp = icmp sgt %struct.anon* %0, null
+  %2 = load %struct.anon.0** @a, align 4
+  br i1 %cmp, label %for.bodythread-pre-split, label %if.end8
+
+for.bodythread-pre-split:                         ; preds = %entry
+  %aclass = getelementptr inbounds %struct.anon.0* %2, i32 0, i32 0
+  %.pr = load i32* %aclass, align 4
+  br label %for.body
+
+for.body:                                         ; preds = %for.bodythread-pre-split, %for.body
+  switch i32 %.pr, label %for.body [
+    i32 0, label %while.body.lr.ph.preheader
+    i32 2, label %while.body.lr.ph.preheader
+  ]
+
+while.body.lr.ph.preheader:                       ; preds = %for.body, %for.body
+  br label %while.body.lr.ph
+
+while.body.lr.ph:                                 ; preds = %while.body.lr.ph.preheader, %if.then6
+  %j.1.ph13 = phi i32 [ %j.110.lcssa, %if.then6 ], [ 0, %while.body.lr.ph.preheader ]
+  %c.0.ph12 = phi i32 [ %dec, %if.then6 ], [ %1, %while.body.lr.ph.preheader ]
+  br label %while.body
+
+while.cond:                                       ; preds = %while.body
+  %cmp2 = icmp slt i32 %inc7, %c.0.ph12
+  br i1 %cmp2, label %while.body, label %if.end8.loopexit
+
+while.body:                                       ; preds = %while.body.lr.ph, %while.cond
+  %j.110 = phi i32 [ %j.1.ph13, %while.body.lr.ph ], [ %inc7, %while.cond ]
+  %aclass_index = getelementptr inbounds %struct.anon* %0, i32 %j.110, i32 0
+  %3 = load i32* %aclass_index, align 4
+  %aclass5 = getelementptr inbounds %struct.anon.0* %2, i32 %3, i32 0
+  %4 = load i32* %aclass5, align 4
+  %tobool = icmp eq i32 %4, 0
+  %inc7 = add nsw i32 %j.110, 1
+  br i1 %tobool, label %while.cond, label %if.then6
+
+if.then6:                                         ; preds = %while.body
+  %j.110.lcssa = phi i32 [ %j.110, %while.body ]
+  %dec = add nsw i32 %c.0.ph12, -1
+  %cmp29 = icmp slt i32 %j.110.lcssa, %dec
+  br i1 %cmp29, label %while.body.lr.ph, label %if.end8.loopexit17
+
+if.end8.loopexit:                                 ; preds = %while.cond
+  br label %if.end8
+
+if.end8.loopexit17:                               ; preds = %if.then6
+  br label %if.end8
+
+if.end8:                                          ; preds = %if.end8.loopexit17, %if.end8.loopexit, %entry
+  ret i32 undef
+}
+
+attributes #0 = { nounwind readonly uwtable }
+