From 1fdfae05b0a4356d1ed0633bf3d6cdc6eba2e173 Mon Sep 17 00:00:00 2001
From: Benjamin Kramer <benny.kra@googlemail.com>
Date: Sat, 24 Dec 2011 17:31:38 +0000
Subject: [PATCH] InstCombine: Canonicalize (2^n)-1 - x into (2^n)-1 ^ x iff x
 is known to be smaller than 2^n.

This has the obvious advantage of being commutable and is always a win on x86 because
const - x wastes a register there. On less weird architectures this may lead to
a regression because other arithmetic doesn't fuse with it anymore. I'll address that
problem in a followup.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@147254 91177308-0d34-0410-b5e6-96231b3b80d8
---
 lib/Transforms/InstCombine/InstCombineAddSub.cpp     |  3 +++
 .../InstCombine/InstCombineSimplifyDemanded.cpp      | 11 +++++++++++
 test/Transforms/InstCombine/sub-xor.ll               | 12 ++++++++++++
 3 files changed, 26 insertions(+)
 create mode 100644 test/Transforms/InstCombine/sub-xor.ll

diff --git a/lib/Transforms/InstCombine/InstCombineAddSub.cpp b/lib/Transforms/InstCombine/InstCombineAddSub.cpp
index d10046c10ba..604d9c8bb44 100644
--- a/lib/Transforms/InstCombine/InstCombineAddSub.cpp
+++ b/lib/Transforms/InstCombine/InstCombineAddSub.cpp
@@ -587,6 +587,9 @@ Instruction *InstCombiner::visitSub(BinaryOperator &I) {
     ConstantInt *C2;
     if (match(Op1, m_Add(m_Value(X), m_ConstantInt(C2))))
       return BinaryOperator::CreateSub(ConstantExpr::getSub(C, C2), X);
+
+    if (SimplifyDemandedInstructionBits(I))
+      return &I;
   }
 
   
diff --git a/lib/Transforms/InstCombine/InstCombineSimplifyDemanded.cpp b/lib/Transforms/InstCombine/InstCombineSimplifyDemanded.cpp
index 5cd9a4b7954..4c720203a28 100644
--- a/lib/Transforms/InstCombine/InstCombineSimplifyDemanded.cpp
+++ b/lib/Transforms/InstCombine/InstCombineSimplifyDemanded.cpp
@@ -567,9 +567,20 @@ Value *InstCombiner::SimplifyDemandedUseBits(Value *V, APInt DemandedMask,
                                LHSKnownZero, LHSKnownOne, Depth+1))
         return I;
     }
+
     // Otherwise just hand the sub off to ComputeMaskedBits to fill in
     // the known zeros and ones.
     ComputeMaskedBits(V, DemandedMask, KnownZero, KnownOne, Depth);
+
+    // Turn this into a xor if LHS is 2^n-1 and the remaining bits are known
+    // zero.
+    if (ConstantInt *C0 = dyn_cast<ConstantInt>(I->getOperand(0))) {
+      APInt I0 = C0->getValue();
+      if ((I0 + 1).isPowerOf2() && (I0 | KnownZero).isAllOnesValue()) {
+        Instruction *Xor = BinaryOperator::CreateXor(I->getOperand(1), C0);
+        return InsertNewInstWith(Xor, *I);
+      }
+    }
     break;
   case Instruction::Shl:
     if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
diff --git a/test/Transforms/InstCombine/sub-xor.ll b/test/Transforms/InstCombine/sub-xor.ll
new file mode 100644
index 00000000000..41e01fbdae8
--- /dev/null
+++ b/test/Transforms/InstCombine/sub-xor.ll
@@ -0,0 +1,12 @@
+; RUN: opt -instcombine -S < %s | FileCheck %s
+
+define i32 @test1(i32 %x) nounwind {
+  %and = and i32 %x, 31
+  %sub = sub i32 63, %and
+  ret i32 %sub
+
+; CHECK: @test1
+; CHECK-NEXT: and i32 %x, 31
+; CHECK-NEXT: xor i32 %and, 63
+; CHECK-NEXT: ret
+}