[SCEV] Commute zero extends through <nuw> additions

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251052 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Analysis/ScalarEvolution.cpp b/lib/Analysis/ScalarEvolution.cpp
index d784eb9ace4914aa259d1ddbd9c5e9f8abdf7e2a..42e38ca8049bf0abf4edd0c0d52d0c970565a9de 100644 (file)
--- a/lib/Analysis/ScalarEvolution.cpp
+++ b/lib/Analysis/ScalarEvolution.cpp
@@ -1559,6 +1559,18 @@ const SCEV *ScalarEvolution::getZeroExtendExpr(const SCEV *Op,
       }
     }
 
+  if (auto *SA = dyn_cast<SCEVAddExpr>(Op)) {
+    // zext((A + B + ...)<nuw>) --> (zext(A) + zext(B) + ...)<nuw>
+    if (SA->getNoWrapFlags(SCEV::FlagNUW)) {
+      // If the addition does not unsign overflow then we can, by definition,
+      // commute the zero extension with the addition operation.
+      SmallVector<const SCEV *, 4> Ops;
+      for (const auto *Op : SA->operands())
+        Ops.push_back(getZeroExtendExpr(Op, Ty));
+      return getAddExpr(Ops, SCEV::FlagNUW);
+    }
+  }
+
   // The cast wasn't folded; create an explicit cast node.
   // Recompute the insert position, as it may have been invalidated.
   if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S;

diff --git a/test/Analysis/ScalarEvolution/no-wrap-add-exprs.ll b/test/Analysis/ScalarEvolution/no-wrap-add-exprs.ll
index fc5c35f8788ebd02870623f8a8758d215eb13a56..c24d7173e4ee507ce376757a3eb68c34f5486c92 100644 (file)
--- a/test/Analysis/ScalarEvolution/no-wrap-add-exprs.ll
+++ b/test/Analysis/ScalarEvolution/no-wrap-add-exprs.ll
@@ -83,3 +83,40 @@ define void @f1(i8* %len_addr) {
 
   ret void
 }
+
+define void @f2(i8* %len_addr) {
+; CHECK-LABEL: Classifying expressions for: @f2
+ entry:
+  %len = load i8, i8* %len_addr, !range !0
+  %len_norange = load i8, i8* %len_addr
+; CHECK:  %len = load i8, i8* %len_addr, !range !0
+; CHECK-NEXT:  -->  %len U: [0,127) S: [0,127)
+; CHECK:  %len_norange = load i8, i8* %len_addr
+; CHECK-NEXT:  -->  %len_norange U: full-set S: full-set
+
+  %t0 = add i8 %len, 1
+  %t1 = add i8 %len, 2
+; CHECK:  %t0 = add i8 %len, 1
+; CHECK-NEXT:  -->  (1 + %len)<nuw><nsw>
+; CHECK:  %t1 = add i8 %len, 2
+; CHECK-NEXT:  -->  (2 + %len)<nuw>
+
+  %t0.zext = zext i8 %t0 to i16
+  %t1.zext = zext i8 %t1 to i16
+; CHECK:  %t0.zext = zext i8 %t0 to i16
+; CHECK-NEXT: -->  (1 + (zext i8 %len to i16))<nuw><nsw> U: [1,128) S: [1,128)
+; CHECK:  %t1.zext = zext i8 %t1 to i16
+; CHECK-NEXT:  -->  (2 + (zext i8 %len to i16))<nuw><nsw> U: [2,129) S: [2,129)
+
+  %q0 = add i8 %len_norange, 1
+  %q1 = add i8 %len_norange, 2
+  %q0.zext = zext i8 %q0 to i16
+  %q1.zext = zext i8 %q1 to i16
+
+; CHECK:  %q0.zext = zext i8 %q0 to i16
+; CHECK-NEXT:  -->  (zext i8 (1 + %len_norange) to i16) U: [0,256) S: [0,256)
+; CHECK:  %q1.zext = zext i8 %q1 to i16
+; CHECK-NEXT:  -->  (zext i8 (2 + %len_norange) to i16) U: [0,256) S: [0,256)
+
+  ret void
+}