Merge 183060: Prevent loop-unroll from making assumptions about undefined behavior.

Fixes rdar:14036816, PR16130.

    There is an opportunity to compute precise trip counts for 'or'
    expressions and multi-exit loops.
    rdar:14038809: Optimize trip count computation for multi-exit loops.

    To do this we need to record the fact that ExitLimit assumes NSW. When
    it does not we can safely assume that the loop trip count is the
    minimum ExitLimt across all subexpressions and loop exits.

    git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@183060 91177308-0d34-0410-b5e6-96231b3b80d8


git-svn-id: https://llvm.org/svn/llvm-project/llvm/branches/release_33@183066 91177308-0d34-0410-b5e6-96231b3b80d8

lib/Analysis/ScalarEvolution.cpp

@@ -3937,10 +3937,19 @@ const SCEV *ScalarEvolution::createSCEV(Value *V) {
 /// before taking the branch. For loops with multiple exits, it may not be the
 /// number times that the loop header executes because the loop may exit
 /// prematurely via another branch.
+///
+/// FIXME: We conservatively call getBackedgeTakenCount(L) instead of
+/// getExitCount(L, ExitingBlock) to compute a safe trip count considering all
+/// loop exits. getExitCount() may return an exact count for this branch
+/// assuming no-signed-wrap. The number of well-defined iterations may actually
+/// be higher than this trip count if this exit test is skipped and the loop
+/// exits via a different branch. Ideally, getExitCount() would know whether it
+/// depends on a NSW assumption, and we would only fall back to a conservative
+/// trip count in that case.
 unsigned ScalarEvolution::
-getSmallConstantTripCount(Loop *L, BasicBlock *ExitingBlock) {
+getSmallConstantTripCount(Loop *L, BasicBlock */*ExitingBlock*/) {
   const SCEVConstant *ExitCount =
-    dyn_cast<SCEVConstant>(getExitCount(L, ExitingBlock));
+    dyn_cast<SCEVConstant>(getBackedgeTakenCount(L));
   if (!ExitCount)
     return 0;
 
@@ -3967,8 +3976,8 @@ getSmallConstantTripCount(Loop *L, BasicBlock *ExitingBlock) {
 /// As explained in the comments for getSmallConstantTripCount, this assumes
 /// that control exits the loop via ExitingBlock.
 unsigned ScalarEvolution::
-getSmallConstantTripMultiple(Loop *L, BasicBlock *ExitingBlock) {
-  const SCEV *ExitCount = getExitCount(L, ExitingBlock);
+getSmallConstantTripMultiple(Loop *L, BasicBlock */*ExitingBlock*/) {
+  const SCEV *ExitCount = getBackedgeTakenCount(L);
   if (ExitCount == getCouldNotCompute())
     return 1;
 
@@ -3997,7 +4006,7 @@ getSmallConstantTripMultiple(Loop *L, BasicBlock *ExitingBlock) {
 }
 
 // getExitCount - Get the expression for the number of loop iterations for which
-// this loop is guaranteed not to exit via ExitintBlock. Otherwise return
+// this loop is guaranteed not to exit via ExitingBlock. Otherwise return
 // SCEVCouldNotCompute.
 const SCEV *ScalarEvolution::getExitCount(Loop *L, BasicBlock *ExitingBlock) {
   return getBackedgeTakenInfo(L).getExact(ExitingBlock, this);

test/Transforms/LoopUnroll/scevunroll.ll

@@ -66,13 +66,16 @@ exit2:
 
 ; SCEV properly unrolls multi-exit loops.
 ;
+; SCEV cannot currently unroll this loop.
+; It should ideally detect a trip count of 5.
+; rdar:14038809 [SCEV]: Optimize trip count computation for multi-exit loops.
 ; CHECK: @multiExit
-; CHECK: getelementptr i32* %base, i32 10
-; CHECK-NEXT: load i32*
-; CHECK: br i1 false, label %l2.10, label %exit1
-; CHECK: l2.10:
-; CHECK-NOT: br
-; CHECK: ret i32
+; CHECKFIXME: getelementptr i32* %base, i32 10
+; CHECKFIXME-NEXT: load i32*
+; CHECKFIXME: br i1 false, label %l2.10, label %exit1
+; CHECKFIXME: l2.10:
+; CHECKFIXME-NOT: br
+; CHECKFIXME: ret i32
 define i32 @multiExit(i32* %base) nounwind {
 entry:
   br label %l1
@@ -170,3 +173,38 @@ for.body87:
   br label %for.body87
 }
 
+; PR16130: clang produces incorrect code with loop/expression at -O2
+; rdar:14036816 loop-unroll makes assumptions about undefined behavior
+;
+; The loop latch is assumed to exit after the first iteration because
+; of the induction variable's NSW flag. However, the loop latch's
+; equality test is skipped and the loop exits after the second
+; iteration via the early exit. So loop unrolling cannot assume that
+; the loop latch's exit count of zero is an upper bound on the number
+; of iterations.
+;
+; CHECK: @nsw_latch
+; CHECK: for.body:
+; CHECK: %b.03 = phi i32 [ 0, %entry ], [ %add, %for.cond ]
+; CHECK: return:
+; CHECK: %b.03.lcssa = phi i32 [ %b.03, %for.body ], [ %b.03, %for.cond ]
+define void @nsw_latch(i32* %a) nounwind {
+entry:
+  br label %for.body
+
+for.body:                                         ; preds = %for.cond, %entry
+  %b.03 = phi i32 [ 0, %entry ], [ %add, %for.cond ]
+  %tobool = icmp eq i32 %b.03, 0
+  %add = add nsw i32 %b.03, 8
+  br i1 %tobool, label %for.cond, label %return
+
+for.cond:                                         ; preds = %for.body
+  %cmp = icmp eq i32 %add, 13
+  br i1 %cmp, label %return, label %for.body
+
+return:                                           ; preds = %for.body, %for.cond
+  %b.03.lcssa = phi i32 [ %b.03, %for.body ], [ %b.03, %for.cond ]
+  %retval.0 = phi i32 [ 1, %for.body ], [ 0, %for.cond ]
+  store i32 %b.03.lcssa, i32* %a, align 4
+  ret void
+}

test/Transforms/LoopUnroll/unloop.ll

@@ -21,8 +21,8 @@ outer:
 inner:
   %iv = phi i32 [ 0, %outer ], [ %inc, %tail ]
   %inc = add i32 %iv, 1
-  %wbucond = call zeroext i1 @check()
-  br i1 %wbucond, label %outer.backedge, label %tail
+  call zeroext i1 @check()
+  br i1 true, label %outer.backedge, label %tail
 
 tail:
   br i1 false, label %inner, label %exit
@@ -126,25 +126,27 @@ return:
 ; Ensure that only the middle loop is removed and rely on verify-loopinfo to
 ; check soundness.
 ;
-; CHECK: @unloopDeepNested
+; This test must be disabled until trip count computation can be optimized...
+; rdar:14038809 [SCEV]: Optimize trip count computation for multi-exit loops.
+; CHECKFIXME: @unloopDeepNested
 ; Inner-inner loop control.
-; CHECK: while.cond.us.i:
-; CHECK: br i1 %cmp.us.i, label %next_data.exit, label %while.body.us.i
-; CHECK: if.then.us.i:
-; CHECK: br label %while.cond.us.i
+; CHECKFIXME: while.cond.us.i:
+; CHECKFIXME: br i1 %cmp.us.i, label %next_data.exit, label %while.body.us.i
+; CHECKFIXME: if.then.us.i:
+; CHECKFIXME: br label %while.cond.us.i
 ; Inner loop tail.
-; CHECK: if.else.i:
-; CHECK: br label %while.cond.outer.i
+; CHECKFIXME: if.else.i:
+; CHECKFIXME: br label %while.cond.outer.i
 ; Middle loop control (removed).
-; CHECK: valid_data.exit:
-; CHECK-NOT: br
-; CHECK: %cmp = call zeroext i1 @check()
+; CHECKFIXME: valid_data.exit:
+; CHECKFIXME-NOT: br
+; CHECKFIXME: %cmp = call zeroext i1 @check()
 ; Outer loop control.
-; CHECK: copy_data.exit:
-; CHECK: br i1 %cmp38, label %if.then39, label %while.cond.outer
+; CHECKFIXME: copy_data.exit:
+; CHECKFIXME: br i1 %cmp38, label %if.then39, label %while.cond.outer
 ; Outer-outer loop tail.
-; CHECK: while.cond.outer.outer.backedge:
-; CHECK: br label %while.cond.outer.outer
+; CHECKFIXME: while.cond.outer.outer.backedge:
+; CHECKFIXME: br label %while.cond.outer.outer
 define void @unloopDeepNested() nounwind {
 for.cond8.preheader.i:
   %cmp113.i = call zeroext i1 @check()