From 222619278d287d178b741d9916e6f5ec017080d6 Mon Sep 17 00:00:00 2001
From: Sanjay Patel <spatel@rotateright.com>
Date: Mon, 30 Jan 2017 23:01:05 +0000
Subject: [PATCH] [InstCombine] enable more lshr(shl X, C1), C2 folds for
 vectors with splat constants

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@293562 91177308-0d34-0410-b5e6-96231b3b80d8
---
 .../InstCombine/InstCombineShifts.cpp         | 40 ++++++++-----------
 test/Transforms/InstCombine/shift.ll          | 16 ++++++--
 2 files changed, 29 insertions(+), 27 deletions(-)

diff --git a/lib/Transforms/InstCombine/InstCombineShifts.cpp b/lib/Transforms/InstCombine/InstCombineShifts.cpp
index edd51baf2a8..61716ca4d7f 100644
--- a/lib/Transforms/InstCombine/InstCombineShifts.cpp
+++ b/lib/Transforms/InstCombine/InstCombineShifts.cpp
@@ -373,24 +373,6 @@ foldShiftByConstOfShiftByConst(BinaryOperator &I, const APInt *COp1,
   if (ShiftAmt2 < ShiftAmt1) {
     uint32_t ShiftDiff = ShiftAmt1 - ShiftAmt2;
 
-    // (X << C1) >>u C2  --> X << (C1-C2) & (-1 >> C2)
-    if (I.getOpcode() == Instruction::LShr &&
-        ShiftOp->getOpcode() == Instruction::Shl) {
-      ConstantInt *ShiftDiffCst = ConstantInt::get(Ty, ShiftDiff);
-      if (ShiftOp->hasNoUnsignedWrap()) {
-        // (X <<nuw C1) >>u C2 --> X <<nuw (C1-C2)
-        BinaryOperator *NewShl =
-            BinaryOperator::Create(Instruction::Shl, X, ShiftDiffCst);
-        NewShl->setHasNoUnsignedWrap(true);
-        return NewShl;
-      }
-      Value *Shift = Builder->CreateShl(X, ShiftDiffCst);
-
-      APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
-      return BinaryOperator::CreateAnd(Shift,
-                                       ConstantInt::get(I.getContext(), Mask));
-    }
-
     // We can't handle (X << C1) >>s C2, it shifts arbitrary bits in. However,
     // we can handle (X <<nsw C1) >>s C2 since it only shifts in sign bits.
     if (I.getOpcode() == Instruction::AShr &&
@@ -740,11 +722,6 @@ Instruction *InstCombiner::visitLShr(BinaryOperator &I) {
     const APInt *ShlAmtAPInt;
     if (match(Op0, m_Shl(m_Value(X), m_APInt(ShlAmtAPInt)))) {
       unsigned ShlAmt = ShlAmtAPInt->getZExtValue();
-      if (ShlAmt == ShAmt) {
-        // (X << C) >>u C --> X & (-1 >>u C)
-        APInt Mask(APInt::getLowBitsSet(BitWidth, BitWidth - ShAmt));
-        return BinaryOperator::CreateAnd(X, ConstantInt::get(Ty, Mask));
-      }
       if (ShlAmt < ShAmt) {
         Constant *ShiftDiff = ConstantInt::get(Ty, ShAmt - ShlAmt);
         if (cast<BinaryOperator>(Op0)->hasNoUnsignedWrap()) {
@@ -758,6 +735,23 @@ Instruction *InstCombiner::visitLShr(BinaryOperator &I) {
         APInt Mask(APInt::getLowBitsSet(BitWidth, BitWidth - ShAmt));
         return BinaryOperator::CreateAnd(NewLShr, ConstantInt::get(Ty, Mask));
       }
+      if (ShlAmt > ShAmt) {
+        Constant *ShiftDiff = ConstantInt::get(Ty, ShlAmt - ShAmt);
+        if (cast<BinaryOperator>(Op0)->hasNoUnsignedWrap()) {
+          // (X <<nuw C1) >>u C2 --> X <<nuw (C1 - C2)
+          auto *NewShl = BinaryOperator::CreateShl(X, ShiftDiff);
+          NewShl->setHasNoUnsignedWrap(true);
+          return NewShl;
+        }
+        // (X << C1) >>u C2  --> X << (C1 - C2) & (-1 >> C2)
+        Value *NewShl = Builder->CreateShl(X, ShiftDiff);
+        APInt Mask(APInt::getLowBitsSet(BitWidth, BitWidth - ShAmt));
+        return BinaryOperator::CreateAnd(NewShl, ConstantInt::get(Ty, Mask));
+      }
+      assert(ShlAmt == ShAmt);
+      // (X << C) >>u C --> X & (-1 >>u C)
+      APInt Mask(APInt::getLowBitsSet(BitWidth, BitWidth - ShAmt));
+      return BinaryOperator::CreateAnd(X, ConstantInt::get(Ty, Mask));
     }
 
     // If the shifted-out value is known-zero, then this is an exact shift.
diff --git a/test/Transforms/InstCombine/shift.ll b/test/Transforms/InstCombine/shift.ll
index 573cb8dfe1d..b105c58201e 100644
--- a/test/Transforms/InstCombine/shift.ll
+++ b/test/Transforms/InstCombine/shift.ll
@@ -997,6 +997,8 @@ define i32 @test52(i32 %x) {
   ret i32 %B
 }
 
+; (X <<nuw C1) >>u C2 --> X <<nuw (C1 - C2)
+
 define i32 @test53(i32 %x) {
 ; CHECK-LABEL: @test53(
 ; CHECK-NEXT:    [[B:%.*]] = shl nuw i32 %x, 2
@@ -1007,10 +1009,11 @@ define i32 @test53(i32 %x) {
   ret i32 %B
 }
 
+; (X <<nuw C1) >>u C2 --> X <<nuw (C1 - C2)
+
 define <2 x i32> @test53_splat_vec(<2 x i32> %x) {
 ; CHECK-LABEL: @test53_splat_vec(
-; CHECK-NEXT:    [[A:%.*]] = shl nuw <2 x i32> %x, <i32 3, i32 3>
-; CHECK-NEXT:    [[B:%.*]] = lshr exact <2 x i32> [[A]], <i32 1, i32 1>
+; CHECK-NEXT:    [[B:%.*]] = shl nuw <2 x i32> %x, <i32 2, i32 2>
 ; CHECK-NEXT:    ret <2 x i32> [[B]]
 ;
   %A = shl nuw <2 x i32> %x, <i32 3, i32 3>
@@ -1018,6 +1021,8 @@ define <2 x i32> @test53_splat_vec(<2 x i32> %x) {
   ret <2 x i32> %B
 }
 
+; (X << C1) >>u C2  --> X << (C1 - C2) & (-1 >> C2)
+
 define i8 @test53_no_nuw(i8 %x) {
 ; CHECK-LABEL: @test53_no_nuw(
 ; CHECK-NEXT:    [[TMP1:%.*]] = shl i8 %x, 2
@@ -1029,10 +1034,13 @@ define i8 @test53_no_nuw(i8 %x) {
   ret i8 %B
 }
 
+; (X << C1) >>u C2  --> X << (C1 - C2) & (-1 >> C2)
+; FIXME: Demanded bits should change the mask constant as it does for the scalar case.
+
 define <2 x i8> @test53_no_nuw_splat_vec(<2 x i8> %x) {
 ; CHECK-LABEL: @test53_no_nuw_splat_vec(
-; CHECK-NEXT:    [[A:%.*]] = shl <2 x i8> %x, <i8 3, i8 3>
-; CHECK-NEXT:    [[B:%.*]] = lshr exact <2 x i8> [[A]], <i8 1, i8 1>
+; CHECK-NEXT:    [[TMP1:%.*]] = shl <2 x i8> %x, <i8 2, i8 2>
+; CHECK-NEXT:    [[B:%.*]] = and <2 x i8> [[TMP1]], <i8 127, i8 127>
 ; CHECK-NEXT:    ret <2 x i8> [[B]]
 ;
   %A = shl <2 x i8> %x, <i8 3, i8 3>
-- 
2.40.0