LSR: Reuse the post-inc expansion of expressions.

This avoids unnecessary expansion of expressions and allows the SCEV
expander to work on expression DAGs, not just trees.
Fixes PR11090.


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@141870 91177308-0d34-0410-b5e6-96231b3b80d8

2 files changed, 125 insertions, 2 deletions

diff --git a/lib/Analysis/ScalarEvolutionExpander.cpp b/lib/Analysis/ScalarEvolutionExpander.cpp
index 81f4db1c7d..e76b6b10cf 100644
--- a/lib/Analysis/ScalarEvolutionExpander.cpp
+++ b/lib/Analysis/ScalarEvolutionExpander.cpp
@@ -1120,6 +1120,14 @@ Value *SCEVExpander::expandAddRecExprLiterally(const SCEVAddRecExpr *S) {
     BasicBlock *LatchBlock = L->getLoopLatch();
     assert(LatchBlock && "PostInc mode requires a unique loop latch!");
     Result = PN->getIncomingValueForBlock(LatchBlock);
+
+    // For an expansion to use the postinc form, the client must call
+    // expandCodeFor with an InsertPoint that is either outside the PostIncLoop
+    // or dominated by IVIncInsertPos.
+    assert((!isa<Instruction>(Result) ||
+            SE.DT->dominates(cast<Instruction>(Result),
+                             Builder.GetInsertPoint())) &&
+           "postinc expansion does not dominate use");
   }
 
   // Re-apply any non-loop-dominating scale.
@@ -1410,8 +1418,12 @@ Value *SCEVExpander::expand(const SCEV *S) {
   Value *V = visit(S);
 
   // Remember the expanded value for this SCEV at this location.
-  if (PostIncLoops.empty())
-    InsertedExpressions[std::make_pair(S, InsertPt)] = V;
+  //
+  // This is independent of PostIncLoops. The mapped value simply materializes
+  // the expression at this insertion point. If the mapped value happened to be
+  // a postinc expansion, it could be reused by a non postinc user, but only if
+  // its insertion point was already at the head of the loop.
+  InsertedExpressions[std::make_pair(S, InsertPt)] = V;
 
   restoreInsertPoint(SaveInsertBB, SaveInsertPt);
   return V;
diff --git a/test/Transforms/LoopStrengthReduce/2011-10-13-SCEVChain.ll b/test/Transforms/LoopStrengthReduce/2011-10-13-SCEVChain.ll
new file mode 100644
index 0000000000..4718529bfd
--- /dev/null
+++ b/test/Transforms/LoopStrengthReduce/2011-10-13-SCEVChain.ll
@@ -0,0 +1,111 @@
+; RUN: opt -loop-reduce -S < %s | FileCheck %s
+;
+; Test TransformForPostIncUse and LSR's expansion of expressions in
+; post-inc form to ensure the implementation can handle expressions
+; DAGs, not just trees.
+
+target triple = "x86_64-apple-darwin"
+
+; Verify that -loop-reduce runs without "hanging" and reuses post-inc
+; expansions.
+; CHECK: @test
+; CHECK: icmp
+; CHECK: icmp
+; CHECK: icmp
+; CHECK: icmp
+; CHECK: icmp
+; CHECK: icmp
+; CHECK: icmp
+; CHECK: icmp
+; CHECK: icmp
+; CHECK: icmp
+; CHECK: icmp
+; CHECK: icmp
+; CHECK: icmp
+; CHECK: icmp
+; CHECK: icmp
+; CHECK: icmp
+; CHECK: icmp
+; CHECK-NOT: icmp
+define void @test(i8* %base, i32 %a0) nounwind {
+entry:
+  br label %bb1
+bb1:
+  %n0 = sub i32 0, %a0
+  %t0 = icmp ugt i32 %n0, -4
+  %m0 = select i1 %t0, i32 %n0, i32 -4
+  %a1 = add i32 %m0, %a0
+  %n1 = sub i32 0, %a1
+  %t1 = icmp ugt i32 %n1, -4
+  %m1 = select i1 %t1, i32 %n1, i32 -4
+  %a2 = add i32 %m1, %a1
+  %n2 = sub i32 0, %a2
+  %t2 = icmp ugt i32 %n2, -4
+  %m2 = select i1 %t2, i32 %n2, i32 -4
+  %a3 = add i32 %m2, %a2
+  %n3 = sub i32 0, %a3
+  %t3 = icmp ugt i32 %n3, -4
+  %m3 = select i1 %t3, i32 %n3, i32 -4
+  %a4 = add i32 %m3, %a3
+  %n4 = sub i32 0, %a4
+  %t4 = icmp ugt i32 %n4, -4
+  %m4 = select i1 %t4, i32 %n4, i32 -4
+  %a5 = add i32 %m4, %a4
+  %n5 = sub i32 0, %a5
+  %t5 = icmp ugt i32 %n5, -4
+  %m5 = select i1 %t5, i32 %n5, i32 -4
+  %a6 = add i32 %m5, %a5
+  %n6 = sub i32 0, %a6
+  %t6 = icmp ugt i32 %n6, -4
+  %m6 = select i1 %t6, i32 %n6, i32 -4
+  %a7 = add i32 %m6, %a6
+  %n7 = sub i32 0, %a7
+  %t7 = icmp ugt i32 %n7, -4
+  %m7 = select i1 %t7, i32 %n7, i32 -4
+  %a8 = add i32 %m7, %a7
+  %n8 = sub i32 0, %a8
+  %t8 = icmp ugt i32 %n8, -4
+  %m8 = select i1 %t8, i32 %n8, i32 -4
+  %a9 = add i32 %m8, %a8
+  %n9 = sub i32 0, %a9
+  %t9 = icmp ugt i32 %n9, -4
+  %m9 = select i1 %t9, i32 %n9, i32 -4
+  %a10 = add i32 %m9, %a9
+  %n10 = sub i32 0, %a10
+  %t10 = icmp ugt i32 %n10, -4
+  %m10 = select i1 %t10, i32 %n10, i32 -4
+  %a11 = add i32 %m10, %a10
+  %n11 = sub i32 0, %a11
+  %t11 = icmp ugt i32 %n11, -4
+  %m11 = select i1 %t11, i32 %n11, i32 -4
+  %a12 = add i32 %m11, %a11
+  %n12 = sub i32 0, %a12
+  %t12 = icmp ugt i32 %n12, -4
+  %m12 = select i1 %t12, i32 %n12, i32 -4
+  %a13 = add i32 %m12, %a12
+  %n13 = sub i32 0, %a13
+  %t13 = icmp ugt i32 %n13, -4
+  %m13 = select i1 %t13, i32 %n13, i32 -4
+  %a14 = add i32 %m13, %a13
+  %n14 = sub i32 0, %a14
+  %t14 = icmp ugt i32 %n14, -4
+  %m14 = select i1 %t14, i32 %n14, i32 -4
+  %a15 = add i32 %m14, %a14
+  %n15 = sub i32 0, %a15
+  %t15 = icmp ugt i32 %n15, -4
+  %m15 = select i1 %t15, i32 %n15, i32 -4
+  %a16 = add i32 %m15, %a15
+  %gep = getelementptr i8* %base, i32 %a16
+  %ofs = add i32 %a16, 4
+  %limit = getelementptr i8* %base, i32 %ofs
+  br label %loop
+
+loop:
+  %iv = phi i8* [ %gep, %bb1 ], [ %inc, %loop ]
+  %inc = getelementptr inbounds i8* %iv, i64 1
+  %exitcond = icmp eq i8* %inc, %limit
+  br i1 %exitcond, label %loop, label %exit
+
+exit:
+  ret void
+}