From b4ec2830499b8c3c5e0de56a2620fb2b21c88b9e Mon Sep 17 00:00:00 2001 From: Bill Wendling Date: Tue, 9 Sep 2008 22:13:54 +0000 Subject: [PATCH] Add support for 6-, 12-, and 18-bit limited precision calculations of exp for floating-point numbers. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56023 91177308-0d34-0410-b5e6-96231b3b80d8 --- .../SelectionDAG/SelectionDAGBuild.cpp | 141 +++++++++++++++++- 1 file changed, 134 insertions(+), 7 deletions(-) diff --git a/lib/CodeGen/SelectionDAG/SelectionDAGBuild.cpp b/lib/CodeGen/SelectionDAG/SelectionDAGBuild.cpp index 649604c0e5b..94ca366f36e 100644 --- a/lib/CodeGen/SelectionDAG/SelectionDAGBuild.cpp +++ b/lib/CodeGen/SelectionDAG/SelectionDAGBuild.cpp @@ -2761,7 +2761,7 @@ GetSignificand(SelectionDAG &DAG, SDValue Op) { static SDValue GetExponent(SelectionDAG &DAG, SDValue Op) { SDValue t1 = DAG.getNode(ISD::AND, MVT::i32, Op, - DAG.getConstant(0x7f800000, MVT::i32)); + DAG.getConstant(0x7f800000, MVT::i32)); SDValue t2 = DAG.getNode(ISD::SRL, MVT::i32, t1, DAG.getConstant(23, MVT::i32)); SDValue t3 = DAG.getNode(ISD::SUB, MVT::i32, t2, @@ -2784,15 +2784,142 @@ SelectionDAGLowering::implVisitBinaryAtomic(CallInst& I, ISD::NodeType Op) { return 0; } -/// visitExp - lower an exp intrinsic. Handles the special sequences -/// for limited-precision mode. +/// visitExp - Lower an exp intrinsic. Handles the special sequences for +/// limited-precision mode. void SelectionDAGLowering::visitExp(CallInst &I) { SDValue result; - // No special expansion. - result = DAG.getNode(ISD::FEXP, - getValue(I.getOperand(1)).getValueType(), - getValue(I.getOperand(1))); + + if (getValue(I.getOperand(1)).getValueType() == MVT::f32 && + LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) { + SDValue Op = getValue(I.getOperand(1)); + + // Put the exponent in the right bit position for later addition to the + // final result: + // + // #define LOG2OFe 1.4426950f + // IntegerPartOfX = ((int32_t)(X * LOG2OFe)); + SDValue t0 = DAG.getNode(ISD::FMUL, MVT::f32, Op, + DAG.getConstantFP(APFloat( + APInt(32, 0x3fb8aa3b)), MVT::f32)); + SDValue IntegerPartOfX = DAG.getNode(ISD::FP_TO_SINT, MVT::i32, t0); + + // FractionalPartOfX = (X * LOG2OFe) - (float)IntegerPartOfX; + SDValue t1 = DAG.getNode(ISD::SINT_TO_FP, MVT::f32, IntegerPartOfX); + SDValue X = DAG.getNode(ISD::FSUB, MVT::f32, t0, t1); + + // IntegerPartOfX <<= 23; + IntegerPartOfX = DAG.getNode(ISD::SHL, MVT::i32, IntegerPartOfX, + DAG.getConstant(23, MVT::i32)); + + if (LimitFloatPrecision <= 6) { + // For floating-point precision of 6: + // + // TwoToFractionalPartOfX = + // 0.997535578f + + // (0.735607626f + 0.252464424f * x) * x; + // + // error 0.0144103317, which is 6 bits + SDValue t2 = DAG.getNode(ISD::FMUL, MVT::f32, X, + DAG.getConstantFP(APFloat( + APInt(32, 0x3e814304)), MVT::f32)); + SDValue t3 = DAG.getNode(ISD::FADD, MVT::f32, t2, + DAG.getConstantFP(APFloat( + APInt(32, 0x3f3c50c8)), MVT::f32)); + SDValue t4 = DAG.getNode(ISD::FMUL, MVT::f32, t3, X); + SDValue t5 = DAG.getNode(ISD::FADD, MVT::f32, t4, + DAG.getConstantFP(APFloat( + APInt(32, 0x3f7f5e7e)), MVT::f32)); + SDValue TwoToFracPartOfX = DAG.getNode(ISD::BIT_CONVERT, MVT::i32, t5); + + // Add the exponent into the result in integer domain. + SDValue t6 = DAG.getNode(ISD::ADD, MVT::i32, + TwoToFracPartOfX, IntegerPartOfX); + + result = DAG.getNode(ISD::BIT_CONVERT, MVT::f32, t6); + } else if (LimitFloatPrecision > 6 && LimitFloatPrecision <= 12) { + // For floating-point precision of 12: + // + // TwoToFractionalPartOfX = + // 0.999892986f + + // (0.696457318f + + // (0.224338339f + 0.792043434e-1f * x) * x) * x; + // + // 0.000107046256 error, which is 13 to 14 bits + SDValue t2 = DAG.getNode(ISD::FMUL, MVT::f32, X, + DAG.getConstantFP(APFloat( + APInt(32, 0x3da235e3)), MVT::f32)); + SDValue t3 = DAG.getNode(ISD::FADD, MVT::f32, t2, + DAG.getConstantFP(APFloat( + APInt(32, 0x3e65b8f3)), MVT::f32)); + SDValue t4 = DAG.getNode(ISD::FMUL, MVT::f32, t3, X); + SDValue t5 = DAG.getNode(ISD::FADD, MVT::f32, t4, + DAG.getConstantFP(APFloat( + APInt(32, 0x3f324b07)), MVT::f32)); + SDValue t6 = DAG.getNode(ISD::FMUL, MVT::f32, t5, X); + SDValue t7 = DAG.getNode(ISD::FADD, MVT::f32, t6, + DAG.getConstantFP(APFloat( + APInt(32, 0x3f7ff8fd)), MVT::f32)); + SDValue TwoToFracPartOfX = DAG.getNode(ISD::BIT_CONVERT, MVT::i32, t7); + + // Add the exponent into the result in integer domain. + SDValue t8 = DAG.getNode(ISD::ADD, MVT::i32, + TwoToFracPartOfX, IntegerPartOfX); + + result = DAG.getNode(ISD::BIT_CONVERT, MVT::f32, t8); + } else { // LimitFloatPrecision > 12 && LimitFloatPrecision <= 18 + // For floating-point precision of 18: + // + // TwoToFractionalPartOfX = + // 0.999999982f + + // (0.693148872f + + // (0.240227044f + + // (0.554906021e-1f + + // (0.961591928e-2f + + // (0.136028312e-2f + 0.157059148e-3f *x)*x)*x)*x)*x)*x; + // + // error 2.47208000*10^(-7), which is better than 18 bits + SDValue t2 = DAG.getNode(ISD::FMUL, MVT::f32, X, + DAG.getConstantFP(APFloat( + APInt(32, 0x3924b03e)), MVT::f32)); + SDValue t3 = DAG.getNode(ISD::FADD, MVT::f32, t2, + DAG.getConstantFP(APFloat( + APInt(32, 0x3ab24b87)), MVT::f32)); + SDValue t4 = DAG.getNode(ISD::FMUL, MVT::f32, t3, X); + SDValue t5 = DAG.getNode(ISD::FADD, MVT::f32, t4, + DAG.getConstantFP(APFloat( + APInt(32, 0x3c1d8c17)), MVT::f32)); + SDValue t6 = DAG.getNode(ISD::FMUL, MVT::f32, t5, X); + SDValue t7 = DAG.getNode(ISD::FADD, MVT::f32, t6, + DAG.getConstantFP(APFloat( + APInt(32, 0x3d634a1d)), MVT::f32)); + SDValue t8 = DAG.getNode(ISD::FMUL, MVT::f32, t7, X); + SDValue t9 = DAG.getNode(ISD::FADD, MVT::f32, t8, + DAG.getConstantFP(APFloat( + APInt(32, 0x3e75fe14)), MVT::f32)); + SDValue t10 = DAG.getNode(ISD::FMUL, MVT::f32, t9, X); + SDValue t11 = DAG.getNode(ISD::FADD, MVT::f32, t10, + DAG.getConstantFP(APFloat( + APInt(32, 0x3f317234)), MVT::f32)); + SDValue t12 = DAG.getNode(ISD::FMUL, MVT::f32, t11, X); + SDValue t13 = DAG.getNode(ISD::FADD, MVT::f32, t12, + DAG.getConstantFP(APFloat( + APInt(32, 0x3f800000)), MVT::f32)); + SDValue TwoToFracPartOfX = DAG.getNode(ISD::BIT_CONVERT, MVT::i32, t13); + + // Add the exponent into the result in integer domain. + SDValue t14 = DAG.getNode(ISD::ADD, MVT::i32, + TwoToFracPartOfX, IntegerPartOfX); + + result = DAG.getNode(ISD::BIT_CONVERT, MVT::f32, t14); + } + } else { + // No special expansion. + result = DAG.getNode(ISD::FEXP, + getValue(I.getOperand(1)).getValueType(), + getValue(I.getOperand(1))); + } + setValue(&I, result); } -- 2.34.1