From 57ed0948b87206beadce58c406f95dda345fe62c Mon Sep 17 00:00:00 2001
From: Nick Lewycky <nicholas@mxc.ca>
Date: Sat, 31 Dec 2011 21:30:22 +0000
Subject: [PATCH] Make use of the exact bit when optimizing '(X >>exact 3) <<
 1' to eliminate the 'and' that would zero out the trailing bits, and to
 produce an exact shift ourselves.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@147391 91177308-0d34-0410-b5e6-96231b3b80d8
---
 .../InstCombine/InstCombineShifts.cpp          | 16 +++++++++++-----
 test/Transforms/InstCombine/shift.ll           | 18 ++++++++++++++++++
 2 files changed, 29 insertions(+), 5 deletions(-)

diff --git a/lib/Transforms/InstCombine/InstCombineShifts.cpp b/lib/Transforms/InstCombine/InstCombineShifts.cpp
index 702e0f2ee8e..6779e4076cb 100644
--- a/lib/Transforms/InstCombine/InstCombineShifts.cpp
+++ b/lib/Transforms/InstCombine/InstCombineShifts.cpp
@@ -536,12 +536,11 @@ Instruction *InstCombiner::FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
     if (ShiftAmt1 == 0) return 0;  // Will be simplified in the future.
     Value *X = ShiftOp->getOperand(0);
     
-    uint32_t AmtSum = ShiftAmt1+ShiftAmt2;   // Fold into one big shift.
-    
     IntegerType *Ty = cast<IntegerType>(I.getType());
     
     // Check for (X << c1) << c2  and  (X >> c1) >> c2
     if (I.getOpcode() == ShiftOp->getOpcode()) {
+      uint32_t AmtSum = ShiftAmt1+ShiftAmt2;   // Fold into one big shift.
       // If this is oversized composite shift, then unsigned shifts get 0, ashr
       // saturates.
       if (AmtSum >= TypeBits) {
@@ -603,9 +602,16 @@ Instruction *InstCombiner::FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
       // (X >>? C1) << C2 --> X >>? (C1-C2) & (-1 << C2)
       if (I.getOpcode() == Instruction::Shl &&
           ShiftOp->getOpcode() != Instruction::Shl) {
-        Value *Shift = Builder->CreateBinOp(ShiftOp->getOpcode(), X,
-                                            ConstantInt::get(Ty, ShiftDiff));
-        
+        ConstantInt *ShiftDiffCst = ConstantInt::get(Ty, ShiftDiff);
+        if (ShiftOp->isExact()) {
+          // (X >>?exact C1) << C2 --> X >>?exact (C1-C2)
+          BinaryOperator *NewShr = BinaryOperator::Create(ShiftOp->getOpcode(),
+                                                          X, ShiftDiffCst);
+          NewShr->setIsExact(true);
+          return NewShr;
+        }
+        Value *Shift = Builder->CreateBinOp(ShiftOp->getOpcode(),
+                                            X, ShiftDiffCst);
         APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
         return BinaryOperator::CreateAnd(Shift,
                                          ConstantInt::get(I.getContext(),Mask));
diff --git a/test/Transforms/InstCombine/shift.ll b/test/Transforms/InstCombine/shift.ll
index 132d51a660b..0dd969b8c00 100644
--- a/test/Transforms/InstCombine/shift.ll
+++ b/test/Transforms/InstCombine/shift.ll
@@ -542,3 +542,21 @@ define i32 @test45(i32 %a) nounwind {
 ; CHECK-NEXT: %y = lshr i32 %a, 5
 ; CHECK-NEXT: ret i32 %y
 }
+
+define i32 @test46(i32 %a) {
+  %y = ashr exact i32 %a, 3
+  %z = shl i32 %y, 1
+  ret i32 %z
+; CHECK: @test46
+; CHECK-NEXT: %z = ashr exact i32 %a, 2
+; CHECK-NEXT: ret i32 %z
+}
+
+define i32 @test47(i32 %a) {
+  %y = lshr exact i32 %a, 3
+  %z = shl i32 %y, 1
+  ret i32 %z
+; CHECK: @test47
+; CHECK-NEXT: %z = lshr exact i32 %a, 2
+; CHECK-NEXT: ret i32 %z
+}
-- 
2.34.1