From: David Majnemer Date: Tue, 30 Jul 2013 21:01:36 +0000 (+0000) Subject: isKnownToBeAPowerOfTwo: Strengthen isKnownToBeAPowerOfTwo's analysis on add instructions X-Git-Url: http://review.tizen.org/git/?a=commitdiff_plain;h=b7d5409ad292f54def13989d868f2536713d6ece;p=platform%2Fupstream%2Fllvm.git isKnownToBeAPowerOfTwo: Strengthen isKnownToBeAPowerOfTwo's analysis on add instructions Call into ComputeMaskedBits to figure out which bits are set on both add operands and determine if the value is a power-of-two-or-zero or not. llvm-svn: 187445 --- diff --git a/llvm/lib/Analysis/ValueTracking.cpp b/llvm/lib/Analysis/ValueTracking.cpp index 84a6346..4591af8 100644 --- a/llvm/lib/Analysis/ValueTracking.cpp +++ b/llvm/lib/Analysis/ValueTracking.cpp @@ -855,22 +855,36 @@ bool llvm::isKnownToBeAPowerOfTwo(Value *V, bool OrZero, unsigned Depth) { return false; } - if (match(V, m_Add(m_Value(X), m_Value(Y)))) - if (OverflowingBinaryOperator *VOBO = cast(V)) - if (OrZero || VOBO->hasNoUnsignedWrap() || VOBO->hasNoSignedWrap()) { - // Adding a power of two to the same power of two is a power of two or - // zero. - if (BinaryOperator *XBO = dyn_cast(X)) - if (XBO->getOpcode() == Instruction::And) - if (XBO->getOperand(0) == Y || XBO->getOperand(1) == Y) - if (isKnownToBeAPowerOfTwo(Y, OrZero, Depth)) - return true; - if (BinaryOperator *YBO = dyn_cast(Y)) - if (YBO->getOpcode() == Instruction::And) - if (YBO->getOperand(0) == X || YBO->getOperand(1) == X) - if (isKnownToBeAPowerOfTwo(X, OrZero, Depth)) - return true; - } + // Adding a power-of-two or zero to the same power-of-two or zero yields + // either the original power-of-two, a larger power-of-two or zero. + if (match(V, m_Add(m_Value(X), m_Value(Y)))) { + OverflowingBinaryOperator *VOBO = cast(V); + if (OrZero || VOBO->hasNoUnsignedWrap() || VOBO->hasNoSignedWrap()) { + if (match(X, m_And(m_Specific(Y), m_Value())) || + match(X, m_And(m_Value(), m_Specific(Y)))) + if (isKnownToBeAPowerOfTwo(Y, OrZero, Depth)) + return true; + if (match(Y, m_And(m_Specific(X), m_Value())) || + match(Y, m_And(m_Value(), m_Specific(X)))) + if (isKnownToBeAPowerOfTwo(X, OrZero, Depth)) + return true; + + unsigned BitWidth = V->getType()->getScalarSizeInBits(); + APInt LHSZeroBits(BitWidth, 0), LHSOneBits(BitWidth, 0); + ComputeMaskedBits(X, LHSZeroBits, LHSOneBits, 0, Depth); + + APInt RHSZeroBits(BitWidth, 0), RHSOneBits(BitWidth, 0); + ComputeMaskedBits(Y, RHSZeroBits, RHSOneBits, 0, Depth); + // If i8 V is a power of two or zero: + // ZeroBits: 1 1 1 0 1 1 1 1 + // ~ZeroBits: 0 0 0 1 0 0 0 0 + if ((~(LHSZeroBits & RHSZeroBits)).isPowerOf2()) + // If OrZero isn't set, we cannot give back a zero result. + // Make sure either the LHS or RHS has a bit set. + if (OrZero || RHSOneBits.getBoolValue() || LHSOneBits.getBoolValue()) + return true; + } + } // An exact divide or right shift can only shift off zero bits, so the result // is a power of two only if the first operand is a power of two and not diff --git a/llvm/test/Transforms/InstCombine/rem.ll b/llvm/test/Transforms/InstCombine/rem.ll index 9272f83..22fd90b 100644 --- a/llvm/test/Transforms/InstCombine/rem.ll +++ b/llvm/test/Transforms/InstCombine/rem.ll @@ -172,3 +172,35 @@ define i32 @test17(i32 %X) { %A = urem i32 1, %X ret i32 %A } + +define i32 @test18(i16 %x, i32 %y) { +; CHECK: @test18 +; CHECK-NEXT: [[AND:%.*]] = and i16 %x, 4 +; CHECK-NEXT: [[EXT:%.*]] = zext i16 [[AND]] to i32 +; CHECK-NEXT: [[SHL:%.*]] = shl nuw nsw i32 [[EXT]], 3 +; CHECK-NEXT: [[XOR:%.*]] = xor i32 [[SHL]], 63 +; CHECK-NEXT: [[REM:%.*]] = and i32 [[XOR]], %y +; CHECK-NEXT: ret i32 [[REM]] + %1 = and i16 %x, 4 + %2 = icmp ne i16 %1, 0 + %3 = select i1 %2, i32 32, i32 64 + %4 = urem i32 %y, %3 + ret i32 %4 +} + +define i32 @test19(i32 %x, i32 %y) { +; CHECK: @test19 +; CHECK-NEXT: [[SHL1:%.*]] = shl i32 1, %x +; CHECK-NEXT: [[SHL2:%.*]] = shl i32 1, %y +; CHECK-NEXT: [[AND:%.*]] = and i32 [[SHL1]], [[SHL2]] +; CHECK-NEXT: [[ADD:%.*]] = add i32 [[AND]], [[SHL1]] +; CHECK-NEXT: [[SUB:%.*]] = add i32 [[ADD]], -1 +; CHECK-NEXT: [[REM:%.*]] = and i32 [[SUB]], %y +; CHECK-NEXT: ret i32 [[REM]] + %A = shl i32 1, %x + %B = shl i32 1, %y + %C = and i32 %A, %B + %D = add i32 %C, %A + %E = urem i32 %y, %D + ret i32 %E +}