[llvm] r293208 - [InstCombine] use m_APInt to allow (X << C) >>u C --> X & (-1 >>u C) with splat vectors

[Sanjay Patel via llvm-commits]

Author: spatel
Date: Thu Jan 26 14:52:27 2017
New Revision: 293208

URL: http://llvm.org/viewvc/llvm-project?rev=293208&view=rev
Log:
[InstCombine] use m_APInt to allow (X << C) >>u C --> X & (-1 >>u C) with splat vectors

Modified:
    llvm/trunk/lib/Transforms/InstCombine/InstCombineShifts.cpp
    llvm/trunk/test/Transforms/InstCombine/apint-shift.ll

Modified: llvm/trunk/lib/Transforms/InstCombine/InstCombineShifts.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Transforms/InstCombine/InstCombineShifts.cpp?rev=293208&r1=293207&r2=293208&view=diff
==============================================================================
--- llvm/trunk/lib/Transforms/InstCombine/InstCombineShifts.cpp (original)
+++ llvm/trunk/lib/Transforms/InstCombine/InstCombineShifts.cpp Thu Jan 26 14:52:27 2017
@@ -353,29 +353,37 @@ foldShiftByConstOfShiftByConst(BinaryOpe
   // Combinations of right and left shifts will still be optimized in
   // DAGCombine where scalar evolution no longer applies.
 
+  Value *X = ShiftOp->getOperand(0);
+  unsigned ShiftAmt1 = ShAmt1->getLimitedValue();
+  unsigned ShiftAmt2 = COp1->getLimitedValue();
+  assert(ShiftAmt2 != 0 && "Should have been simplified earlier");
+  if (ShiftAmt1 == 0)
+    return nullptr; // Will be simplified in the future.
+
+  if (ShiftAmt1 == ShiftAmt2) {
+    // FIXME: This repeats a fold that exists in foldShiftedShift(), but we're
+    // not handling the related fold here:
+    // (X >>u C) << C --> X & (-1 << C).
+    // foldShiftedShift() is always called before this, but it is restricted to
+    // only handle cases where the ShiftOp has one use. We don't have that
+    // restriction here.
+    if (I.getOpcode() != Instruction::LShr ||
+        ShiftOp->getOpcode() != Instruction::Shl)
+      return nullptr;
+
+    // (X << C) >>u C --> X & (-1 >>u C).
+    APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt1));
+    return BinaryOperator::CreateAnd(X, ConstantInt::get(I.getType(), Mask));
+  }
+
   // FIXME: Everything under here should be extended to work with vector types.
 
   auto *ShiftAmt1C = dyn_cast<ConstantInt>(ShiftOp->getOperand(1));
   if (!ShiftAmt1C)
     return nullptr;
 
-  uint32_t ShiftAmt1 = ShiftAmt1C->getLimitedValue(TypeBits);
-  uint32_t ShiftAmt2 = COp1->getLimitedValue(TypeBits);
-  assert(ShiftAmt2 != 0 && "Should have been simplified earlier");
-  if (ShiftAmt1 == 0)
-    return nullptr; // Will be simplified in the future.
-
-  Value *X = ShiftOp->getOperand(0);
   IntegerType *Ty = cast<IntegerType>(I.getType());
-  if (ShiftAmt1 == ShiftAmt2) {
-    // If we have ((X << C) >>u C), turn this into X & (-1 >>u C).
-    if (I.getOpcode() == Instruction::LShr &&
-        ShiftOp->getOpcode() == Instruction::Shl) {
-      APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt1));
-      return BinaryOperator::CreateAnd(X,
-                                       ConstantInt::get(I.getContext(), Mask));
-    }
-  } else if (ShiftAmt1 < ShiftAmt2) {
+  if (ShiftAmt1 < ShiftAmt2) {
     uint32_t ShiftDiff = ShiftAmt2 - ShiftAmt1;
 
     // (X >>?,exact C1) << C2 --> X << (C2-C1)

Modified: llvm/trunk/test/Transforms/InstCombine/apint-shift.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/Transforms/InstCombine/apint-shift.ll?rev=293208&r1=293207&r2=293208&view=diff
==============================================================================
--- llvm/trunk/test/Transforms/InstCombine/apint-shift.ll (original)
+++ llvm/trunk/test/Transforms/InstCombine/apint-shift.ll Thu Jan 26 14:52:27 2017
@@ -444,12 +444,12 @@ define i44 @shl_lshr_eq_amt_multi_use(i4
   ret i44 %D
 }
 
-; FIXME: Fold vector lshr (shl X, C), C -> and X, C' regardless of the number of uses of the shl.
+; Fold vector lshr (shl X, C), C -> and X, C' regardless of the number of uses of the shl.
 
 define <2 x i44> @shl_lshr_eq_amt_multi_use_splat_vec(<2 x i44> %A) {
 ; CHECK-LABEL: @shl_lshr_eq_amt_multi_use_splat_vec(
 ; CHECK-NEXT:    [[B:%.*]] = shl <2 x i44> %A, <i44 33, i44 33>
-; CHECK-NEXT:    [[C:%.*]] = lshr exact <2 x i44> [[B]], <i44 33, i44 33>
+; CHECK-NEXT:    [[C:%.*]] = and <2 x i44> %A, <i44 2047, i44 2047>
 ; CHECK-NEXT:    [[D:%.*]] = or <2 x i44> [[B]], [[C]]
 ; CHECK-NEXT:    ret <2 x i44> [[D]]
 ;