From a1444219b271cab6fbfe340c1328b0ab10d8f7b6 Mon Sep 17 00:00:00 2001 From: Shuxin Yang Date: Tue, 15 Jan 2013 21:09:32 +0000 Subject: 1. Hoist minus sign as high as possible in an attempt to reveal some optimization opportunities (in the enclosing supper-expressions). rule 1. (-0.0 - X ) * Y => -0.0 - (X * Y) if expression "-0.0 - X" has only one reference. rule 2. (0.0 - X ) * Y => -0.0 - (X * Y) if expression "0.0 - X" has only one reference, and the instruction is marked "noSignedZero". 2. Eliminate negation (The compiler was already able to handle these opt if the 0.0s are replaced with -0.0.) rule 3: (0.0 - X) * (0.0 - Y) => X * Y rule 4: (0.0 - X) * C => X * -C if the expr is flagged "noSignedZero". 3. Rule 5: (X*Y) * X => (X*X) * Y if X!=Y and the expression is flagged with "UnsafeAlgebra". The purpose of this transformation is two-fold: a) to form a power expression (of X). b) potentially shorten the critical path: After transformation, the latency of the instruction Y is amortized by the expression of X*X, and therefore Y is in a "less critical" position compared to what it was before the transformation. 4. Remove the InstCombine code about simplifiying "X * select". The reasons are following: a) The "select" is somewhat architecture-dependent, therefore the higher level optimizers are not able to precisely predict if the simplification really yields any performance improvement or not. b) The "select" operator is bit complicate, and tends to obscure optimization opportunities. It is btter to keep it as low as possible in expr tree, and let CodeGen to tackle the optimization. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@172551 91177308-0d34-0410-b5e6-96231b3b80d8 --- .../InstCombine/InstCombineMulDivRem.cpp | 91 ++++++++++++++-------- 1 file changed, 60 insertions(+), 31 deletions(-) (limited to 'lib/Transforms') diff --git a/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp b/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp index 29846c1..8e4267f 100644 --- a/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp +++ b/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp @@ -377,6 +377,8 @@ Instruction *InstCombiner::visitFMul(BinaryOperator &I) { if (Value *V = SimplifyFMulInst(Op0, Op1, I.getFastMathFlags(), TD)) return ReplaceInstUsesWith(I, V); + bool AllowReassociate = I.hasUnsafeAlgebra(); + // Simplify mul instructions with a constant RHS. if (isa(Op1)) { // Try to fold constant mul into select arguments. @@ -389,7 +391,7 @@ Instruction *InstCombiner::visitFMul(BinaryOperator &I) { return NV; ConstantFP *C = dyn_cast(Op1); - if (C && I.hasUnsafeAlgebra() && C->getValueAPF().isNormal()) { + if (C && AllowReassociate && C->getValueAPF().isNormal()) { // Let MDC denote an expression in one of these forms: // X * C, C/X, X/C, where C is a constant. // @@ -430,7 +432,7 @@ Instruction *InstCombiner::visitFMul(BinaryOperator &I) { BinaryOperator::CreateFAdd(M0, M1) : BinaryOperator::CreateFSub(M0, M1); Instruction *RI = cast(R); - RI->setHasUnsafeAlgebra(true); + RI->copyFastMathFlags(&I); return RI; } } @@ -438,9 +440,6 @@ Instruction *InstCombiner::visitFMul(BinaryOperator &I) { } } - if (Value *Op0v = dyn_castFNegVal(Op0)) // -X * -Y = X*Y - if (Value *Op1v = dyn_castFNegVal(Op1)) - return BinaryOperator::CreateFMul(Op0v, Op1v); // Under unsafe algebra do: // X * log2(0.5*Y) = X*log2(Y) - X @@ -469,36 +468,66 @@ Instruction *InstCombiner::visitFMul(BinaryOperator &I) { } } - // X * cond ? 1.0 : 0.0 => cond ? X : 0.0 - if (I.hasNoNaNs() && I.hasNoSignedZeros()) { - Value *V0 = I.getOperand(0); - Value *V1 = I.getOperand(1); - Value *Cond, *SLHS, *SRHS; - bool Match = false; - - if (match(V0, m_Select(m_Value(Cond), m_Value(SLHS), m_Value(SRHS)))) { - Match = true; - } else if (match(V1, m_Select(m_Value(Cond), m_Value(SLHS), - m_Value(SRHS)))) { - Match = true; - std::swap(V0, V1); + // Handle symmetric situation in a 2-iteration loop + Value *Opnd0 = Op0; + Value *Opnd1 = Op1; + for (int i = 0; i < 2; i++) { + bool IgnoreZeroSign = I.hasNoSignedZeros(); + if (BinaryOperator::isFNeg(Opnd0, IgnoreZeroSign)) { + Value *N0 = dyn_castFNegVal(Opnd0, IgnoreZeroSign); + Value *N1 = dyn_castFNegVal(Opnd1, IgnoreZeroSign); + + // -X * -Y => X*Y + if (N1) + return BinaryOperator::CreateFMul(N0, N1); + + if (Opnd0->hasOneUse()) { + // -X * Y => -(X*Y) (Promote negation as high as possible) + Value *T = Builder->CreateFMul(N0, Opnd1); + cast(T)->setDebugLoc(I.getDebugLoc()); + Instruction *Neg = BinaryOperator::CreateFNeg(T); + if (I.getFastMathFlags().any()) { + cast(T)->copyFastMathFlags(&I); + Neg->copyFastMathFlags(&I); + } + return Neg; + } } - if (Match) { - ConstantFP *C0 = dyn_cast(SLHS); - ConstantFP *C1 = dyn_cast(SRHS); - - if (C0 && C1 && - ((C0->isZero() && C1->isExactlyValue(1.0)) || - (C1->isZero() && C0->isExactlyValue(1.0)))) { - Value *T; - if (C0->isZero()) - T = Builder->CreateSelect(Cond, SLHS, V1); - else - T = Builder->CreateSelect(Cond, V1, SRHS); - return ReplaceInstUsesWith(I, T); + // (X*Y) * X => (X*X) * Y where Y != X + // The purpose is two-fold: + // 1) to form a power expression (of X). + // 2) potentially shorten the critical path: After transformation, the + // latency of the instruction Y is amortized by the expression of X*X, + // and therefore Y is in a "less critical" position compared to what it + // was before the transformation. + // + if (AllowReassociate) { + Value *Opnd0_0, *Opnd0_1; + if (Opnd0->hasOneUse() && + match(Opnd0, m_FMul(m_Value(Opnd0_0), m_Value(Opnd0_1)))) { + Value *Y = 0; + if (Opnd0_0 == Opnd1 && Opnd0_1 != Opnd1) + Y = Opnd0_1; + else if (Opnd0_1 == Opnd1 && Opnd0_0 != Opnd1) + Y = Opnd0_0; + + if (Y) { + Instruction *T = cast(Builder->CreateFMul(Opnd1, Opnd1)); + T->copyFastMathFlags(&I); + T->setDebugLoc(I.getDebugLoc()); + + Instruction *R = BinaryOperator::CreateFMul(T, Y); + R->copyFastMathFlags(&I); + return R; + } } } + + if (!isa(Op1)) + std::swap(Opnd0, Opnd1); + else + break; } return Changed ? &I : 0; -- cgit v1.1