Fix SCEV overly optimistic back edge taken count for multi-exit loops.

Fixes PR11375: Different results for 'clang++ huh.cpp'...

llvm-svn: 144746

lib/Analysis/ScalarEvolution.cpp

@@ -4153,13 +4153,19 @@ void ScalarEvolution::forgetValue(Value *V) {
 }
 
 /// getExact - Get the exact loop backedge taken count considering all loop
-/// exits. If all exits are computable, this is the minimum computed count.
+/// exits. A computable result can only be return for loops with a single exit.
+/// Returning the minimum taken count among all exits is incorrect because one
+/// of the loop's exit limit's may have been skipped. HowFarToZero assumes that
+/// the limit of each loop test is never skipped. This is a valid assumption as
+/// long as the loop exits via that test. For precise results, it is the
+/// caller's responsibility to specify the relevant loop exit using
+/// getExact(ExitingBlock, SE).
 const SCEV *
 ScalarEvolution::BackedgeTakenInfo::getExact(ScalarEvolution *SE) const {
   // If any exits were not computable, the loop is not computable.
   if (!ExitNotTaken.isCompleteList()) return SE->getCouldNotCompute();
 
-  // We need at least one computable exit.
+  // We need exactly one computable exit.
   if (!ExitNotTaken.ExitingBlock) return SE->getCouldNotCompute();
   assert(ExitNotTaken.ExactNotTaken && "uninitialized not-taken info");
 
@@ -4171,8 +4177,8 @@ ScalarEvolution::BackedgeTakenInfo::getExact(ScalarEvolution *SE) const {
 
     if (!BECount)
       BECount = ENT->ExactNotTaken;
-    else
-      BECount = SE->getUMinFromMismatchedTypes(BECount, ENT->ExactNotTaken);
+    else if (BECount != ENT->ExactNotTaken)
+      return SE->getCouldNotCompute();
   }
   assert(BECount && "Invalid not taken count for loop exit");
   return BECount;
@@ -4253,8 +4259,15 @@ ScalarEvolution::ComputeBackedgeTakenCount(const Loop *L) {
 
     if (MaxBECount == getCouldNotCompute())
       MaxBECount = EL.Max;
-    else if (EL.Max != getCouldNotCompute())
-      MaxBECount = getUMinFromMismatchedTypes(MaxBECount, EL.Max);
+    else if (EL.Max != getCouldNotCompute()) {
+      // We cannot take the "min" MaxBECount, because non-unit stride loops may
+      // skip some loop tests. Taking the max over the exits is sufficiently
+      // conservative.  TODO: We could do better taking into consideration
+      // that (1) the loop has unit stride (2) the last loop test is
+      // less-than/greater-than (3) any loop test is less-than/greater-than AND
+      // falls-through some constant times less then the other tests.
+      MaxBECount = getUMaxFromMismatchedTypes(MaxBECount, EL.Max);
+    }
   }
 
   return BackedgeTakenInfo(ExitCounts, CouldComputeBECount, MaxBECount);
@@ -4920,7 +4933,7 @@ const SCEV *ScalarEvolution::ComputeExitCountExhaustively(const Loop *L,
   // the loop symbolically to determine when the condition gets a value of
   // "ExitWhen".
 
-  unsigned MaxIterations = MaxBruteForceIterations;   // Limit analysis.  
+  unsigned MaxIterations = MaxBruteForceIterations;   // Limit analysis.
   for (unsigned IterationNum = 0; IterationNum != MaxIterations;++IterationNum){
     ConstantInt *CondVal =
       dyn_cast_or_null<ConstantInt>(EvaluateExpression(Cond, L,
@@ -5507,10 +5520,10 @@ ScalarEvolution::HowFarToZero(const SCEV *V, const Loop *L) {
   // behavior. Loops must exhibit defined behavior until a wrapped value is
   // actually used. So the trip count computed by udiv could be smaller than the
   // number of well-defined iterations.
-  if (AddRec->getNoWrapFlags(SCEV::FlagNW))
+  if (AddRec->getNoWrapFlags(SCEV::FlagNW)) {
     // FIXME: We really want an "isexact" bit for udiv.
     return getUDivExpr(Distance, CountDown ? getNegativeSCEV(Step) : Step);
-
+  }
   // Then, try to solve the above equation provided that Start is constant.
   if (const SCEVConstant *StartC = dyn_cast<SCEVConstant>(Start))
     return SolveLinEquationWithOverflow(StepC->getValue()->getValue(),

test/Transforms/IndVarSimplify/2011-11-15-multiexit.ll

@@ -0,0 +1,40 @@
+; RUN: opt < %s -indvars -S | FileCheck %s
+;
+; Prior to the fix for PR11375, indvars would replace %firstIV with a
+; loop-invariant gep computed in the preheader. This was incorrect
+; because it was based on the minimum "ExitNotTaken" count. If the
+; final loop test is skipped (odd number of elements) then the early
+; exit would be taken and the loop invariant value would be incorrect.
+
+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-S128"
+target triple = "x86_64-apple-darwin"
+
+; CHECK: if.end:
+; CHECK: phi i32* [ %first.lcssa, %early.exit ]
+define i32 @test(i32* %first, i32* %last) uwtable ssp {
+entry:
+  br i1 undef, label %if.end, label %if.then
+
+if.then:                                          ; preds = %entry
+  br i1 undef, label %if.end, label %do.body
+
+do.body:                                          ; preds = %if.else, %if.then
+  %firstIV = phi i32* [ %incdec.ptr2, %if.else ], [ %first, %if.then ]
+  %incdec.ptr1 = getelementptr inbounds i32* %firstIV, i64 1
+  %cmp1 = icmp eq i32* %incdec.ptr1, %last
+  br i1 %cmp1, label %early.exit, label %if.else
+
+if.else:                                        ; preds = %do.body
+  %incdec.ptr2 = getelementptr inbounds i32* %firstIV, i64 2
+  %cmp2 = icmp eq i32* %incdec.ptr2, %last
+  br i1 %cmp2, label %if.end, label %do.body
+
+early.exit:
+  %first.lcssa = phi i32* [ %firstIV, %do.body ]
+  br label %if.end
+
+if.end:
+  %tmp = phi i32* [ %first.lcssa, %early.exit ], [ %first, %if.then ], [ %first, %entry ], [ undef, %if.else ]
+  %val = load i32* %tmp
+  ret i32 %val
+}

test/Transforms/IndVarSimplify/loop_evaluate10.ll

@@ -1,8 +1,14 @@
 ; RUN: opt < %s -indvars -S \
 ; RUN:   | grep {%b.1 = phi i32 \\\[ 2, %bb \\\], \\\[ 1, %bb2 \\\]}
-
+;
 ; This loop has multiple exits, and the value of %b1 depends on which
 ; exit is taken. Indvars should correctly compute the exit values.
+;
+; XFAIL: *
+; Indvars does not currently replace loop invariant values unless all
+; loop exits have the same exit value. We could handle some cases,
+; such as this, by making getSCEVAtScope() sensitive to a particular
+; loop exit.  See PR11388.
 
 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"
 target triple = "x86_64-pc-linux-gnu"

test/Transforms/IndVarSimplify/loop_evaluate9.ll

@@ -2,8 +2,13 @@
 ; RUN: grep {\[%\]tmp7 = icmp eq i8 -28, -28} %t
 ; RUN: grep {\[%\]tmp8 = icmp eq i8 63, 63} %t
 ; PR4477
-
 ; Indvars should compute the exit values in loop.
+;
+; XFAIL: *
+; Indvars does not currently replace loop invariant values unless all
+; loop exits have the same exit value. We could handle some cases,
+; such as this, by making getSCEVAtScope() sensitive to a particular
+; loop exit.  See PR11388.
 
 target datalayout = "e-p:32:32:32-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:32:64-f32:32:32-f64:32:64-v64:64:64-v128:128:128-a0:0:64-f80:32:32"
 target triple = "i386-pc-linux-gnu"