Fix LSR to tolerate cases where ScalarEvolution initially
misses an opportunity to fold add operands, but folds them
after LSR has separated them out. This fixes rdar://7886751.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@102157 91177308-0d34-0410-b5e6-96231b3b80d8
diff --git a/lib/Analysis/README.txt b/lib/Analysis/README.txt
index c401090..88ea9f1 100644
--- a/lib/Analysis/README.txt
+++ b/lib/Analysis/README.txt
@@ -16,3 +16,15 @@
which is very inefficient when expanded into code.
//===---------------------------------------------------------------------===//
+
+In test/CodeGen/X86/lsr-delayed-fold.ll,
+
+ScalarEvolution is forming this expression:
+
+((trunc i64 (-1 * %arg5) to i32) + (trunc i64 %arg5 to i32) + (-1 * (trunc i64 undef to i32)))
+
+This could be folded to
+
+(-1 * (trunc i64 undef to i32))
+
+//===---------------------------------------------------------------------===//
diff --git a/lib/Transforms/Scalar/LoopStrengthReduce.cpp b/lib/Transforms/Scalar/LoopStrengthReduce.cpp
index a09bca8..a09b3dc 100644
--- a/lib/Transforms/Scalar/LoopStrengthReduce.cpp
+++ b/lib/Transforms/Scalar/LoopStrengthReduce.cpp
@@ -2060,8 +2060,11 @@
LU.Kind, LU.AccessTy, TLI, SE))
continue;
+ const SCEV *InnerSum = SE.getAddExpr(InnerAddOps);
+ if (InnerSum->isZero())
+ continue;
Formula F = Base;
- F.BaseRegs[i] = SE.getAddExpr(InnerAddOps);
+ F.BaseRegs[i] = InnerSum;
F.BaseRegs.push_back(*J);
if (InsertFormula(LU, LUIdx, F))
// If that formula hadn't been seen before, recurse to find more like