From f606b2f403448ea2c3a8f866e999a0a9f46a136d Mon Sep 17 00:00:00 2001
From: Sanjoy Das <sanjoy@playingwithpointers.com>
Date: Thu, 22 Oct 2015 19:57:25 +0000
Subject: [PATCH] [SCEV] Commute sign extends through nsw additions

Summary: Depends on D13613.

Reviewers: atrick, hfinkel, reames, nlewycky

Subscribers: llvm-commits

Differential Revision: http://reviews.llvm.org/D13685

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251049 91177308-0d34-0410-b5e6-96231b3b80d8
---
 lib/Analysis/ScalarEvolution.cpp              | 10 +++++
 .../ScalarEvolution/no-wrap-add-exprs.ll      | 41 +++++++++++++++++++
 2 files changed, 51 insertions(+)

diff --git a/lib/Analysis/ScalarEvolution.cpp b/lib/Analysis/ScalarEvolution.cpp
index 8a257b46e7e..c5ca839e474 100644
--- a/lib/Analysis/ScalarEvolution.cpp
+++ b/lib/Analysis/ScalarEvolution.cpp
@@ -1631,6 +1631,16 @@ const SCEV *ScalarEvolution::getSignExtendExpr(const SCEV *Op,
         }
       }
     }
+
+    // sext((A + B + ...)<nsw>) --> (sext(A) + sext(B) + ...)<nsw>
+    if (SA->getNoWrapFlags(SCEV::FlagNSW)) {
+      // If the addition does not sign overflow then we can, by definition,
+      // commute the sign extension with the addition operation.
+      SmallVector<const SCEV *, 4> Ops;
+      for (const auto *Op : SA->operands())
+        Ops.push_back(getSignExtendExpr(Op, Ty));
+      return getAddExpr(Ops, SCEV::FlagNSW);
+    }
   }
   // If the input value is a chrec scev, and we can prove that the value
   // did not overflow the old, smaller, value, we can sign extend all of the
diff --git a/test/Analysis/ScalarEvolution/no-wrap-add-exprs.ll b/test/Analysis/ScalarEvolution/no-wrap-add-exprs.ll
index 49a4f4f1668..fc5c35f8788 100644
--- a/test/Analysis/ScalarEvolution/no-wrap-add-exprs.ll
+++ b/test/Analysis/ScalarEvolution/no-wrap-add-exprs.ll
@@ -42,3 +42,44 @@ define void @f0(i8* %len_addr) {
 
   ret void
 }
+
+define void @f1(i8* %len_addr) {
+; CHECK-LABEL: Classifying expressions for: @f1
+ 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)<nsw> U: [-1,126) S: [-1,126)
+; CHECK:  %t1 = add i8 %len, -2
+; CHECK-NEXT:  -->  (-2 + %len)<nsw> U: [-2,125) S: [-2,125)
+
+  %t0.sext = sext i8 %t0 to i16
+  %t1.sext = sext i8 %t1 to i16
+; CHECK:  %t0.sext = sext i8 %t0 to i16
+; CHECK-NEXT:  -->  (-1 + (zext i8 %len to i16))<nsw> U: [-1,126) S: [-1,126)
+; CHECK:  %t1.sext = sext i8 %t1 to i16
+; CHECK-NEXT:  -->  (-2 + (zext i8 %len to i16))<nsw> U: [-2,125) S: [-2,125)
+
+  %q0 = add i8 %len_norange, 1
+  %q1 = add i8 %len_norange, 2
+; CHECK:  %q0 = add i8 %len_norange, 1
+; CHECK-NEXT:  -->  (1 + %len_norange) U: full-set S: full-set
+; CHECK:  %q1 = add i8 %len_norange, 2
+; CHECK-NEXT:  -->  (2 + %len_norange) U: full-set S: full-set
+
+  %q0.sext = sext i8 %q0 to i16
+  %q1.sext = sext i8 %q1 to i16
+; CHECK:  %q0.sext = sext i8 %q0 to i16
+; CHECK-NEXT:  -->  (sext i8 (1 + %len_norange) to i16) U: [-128,128) S: [-128,128)
+; CHECK:  %q1.sext = sext i8 %q1 to i16
+; CHECK-NEXT:  -->  (sext i8 (2 + %len_norange) to i16) U: [-128,128) S: [-128,128)
+
+  ret void
+}