[llvm] r325062 - [InstCombine] (lshr X, 31) * Y --> (ashr X, 31) & Y

[Sanjay Patel via llvm-commits]

Author: spatel
Date: Tue Feb 13 14:24:37 2018
New Revision: 325062

URL: http://llvm.org/viewvc/llvm-project?rev=325062&view=rev
Log:
[InstCombine] (lshr X, 31) * Y --> (ashr X, 31) & Y

This replaces the bit-tracking based fold that did the same thing,
but it only worked for scalars and not directly. 

There is no evidence in existing regression tests that the greater 
power of bit-tracking was needed here, but we should be aware of 
this potential loss of optimization.


Modified:
    llvm/trunk/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp
    llvm/trunk/test/Transforms/InstCombine/mul.ll

Modified: llvm/trunk/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp?rev=325062&r1=325061&r2=325062&view=diff
==============================================================================
--- llvm/trunk/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp (original)
+++ llvm/trunk/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp Tue Feb 13 14:24:37 2018
@@ -340,37 +340,25 @@ Instruction *InstCombiner::visitMul(Bina
     }
   }
 
-  Value *X;
   // (bool X) * Y --> X ? Y : 0
+  // Y * (bool X) --> X ? Y : 0
+  Value *X;
   if (match(Op0, m_ZExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1))
     return SelectInst::Create(X, Op1, ConstantInt::get(I.getType(), 0));
-
-  // Y * (bool X) --> X ? Y : 0
   if (match(Op1, m_ZExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1))
     return SelectInst::Create(X, Op0, ConstantInt::get(I.getType(), 0));
 
-  // If one of the operands of the multiply is a cast from a boolean value, then
-  // we know the bool is either zero or one, so this is a 'masking' multiply.
-  //   X * Y (where Y is 0 or 1) -> X & (0-Y)
-  if (!I.getType()->isVectorTy()) {
-    // -2 is "-1 << 1" so it is all bits set except the low one.
-    APInt Negative2(I.getType()->getPrimitiveSizeInBits(), (uint64_t)-2, true);
-
-    Value *BoolCast = nullptr, *OtherOp = nullptr;
-    if (MaskedValueIsZero(Op0, Negative2, 0, &I)) {
-      BoolCast = Op0;
-      OtherOp = Op1;
-    } else if (MaskedValueIsZero(Op1, Negative2, 0, &I)) {
-      BoolCast = Op1;
-      OtherOp = Op0;
-    }
-
-    if (BoolCast) {
-      Value *V = Builder.CreateSub(Constant::getNullValue(I.getType()),
-                                    BoolCast);
-      return BinaryOperator::CreateAnd(V, OtherOp);
-    }
-  }
+  // (lshr X, 31) * Y --> (ashr X, 31) & Y
+  // Y * (lshr X, 31) --> (ashr X, 31) & Y
+  // TODO: We are not checking one-use because the elimination of the multiply
+  //       is better for analysis?
+  // TODO: Should we canonicalize to '(X < 0) ? Y : 0' instead? That would be
+  //       more similar to what we're doing above.
+  const APInt *C;
+  if (match(Op0, m_LShr(m_Value(X), m_APInt(C))) && *C == C->getBitWidth() - 1)
+    return BinaryOperator::CreateAnd(Builder.CreateAShr(X, *C), Op1);
+  if (match(Op1, m_LShr(m_Value(X), m_APInt(C))) && *C == C->getBitWidth() - 1)
+    return BinaryOperator::CreateAnd(Builder.CreateAShr(X, *C), Op0);
 
   // Check for (mul (sext x), y), see if we can merge this into an
   // integer mul followed by a sext.

Modified: llvm/trunk/test/Transforms/InstCombine/mul.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/Transforms/InstCombine/mul.ll?rev=325062&r1=325061&r2=325062&view=diff
==============================================================================
--- llvm/trunk/test/Transforms/InstCombine/mul.ll (original)
+++ llvm/trunk/test/Transforms/InstCombine/mul.ll Tue Feb 13 14:24:37 2018
@@ -158,8 +158,8 @@ define i32 @signbit_mul_commute_extra_us
 
 define <2 x i32> @signbit_mul_vec(<2 x i32> %a, <2 x i32> %b) {
 ; CHECK-LABEL: @signbit_mul_vec(
-; CHECK-NEXT:    [[D:%.*]] = lshr <2 x i32> [[A:%.*]], <i32 31, i32 31>
-; CHECK-NEXT:    [[E:%.*]] = mul nuw <2 x i32> [[D]], [[B:%.*]]
+; CHECK-NEXT:    [[TMP1:%.*]] = ashr <2 x i32> [[A:%.*]], <i32 31, i32 31>
+; CHECK-NEXT:    [[E:%.*]] = and <2 x i32> [[TMP1]], [[B:%.*]]
 ; CHECK-NEXT:    ret <2 x i32> [[E]]
 ;
   %d = lshr <2 x i32> %a, <i32 31, i32 31>
@@ -169,8 +169,8 @@ define <2 x i32> @signbit_mul_vec(<2 x i
 
 define <2 x i32> @signbit_mul_vec_commute(<2 x i32> %a, <2 x i32> %b) {
 ; CHECK-LABEL: @signbit_mul_vec_commute(
-; CHECK-NEXT:    [[D:%.*]] = lshr <2 x i32> [[A:%.*]], <i32 31, i32 31>
-; CHECK-NEXT:    [[E:%.*]] = mul nuw <2 x i32> [[D]], [[B:%.*]]
+; CHECK-NEXT:    [[TMP1:%.*]] = ashr <2 x i32> [[A:%.*]], <i32 31, i32 31>
+; CHECK-NEXT:    [[E:%.*]] = and <2 x i32> [[TMP1]], [[B:%.*]]
 ; CHECK-NEXT:    ret <2 x i32> [[E]]
 ;
   %d = lshr <2 x i32> %a, <i32 31, i32 31>