6502/llvm-6502
1
0
Fork 0
mirror of https://github.com/c64scene-ar/llvm-6502.git synced 2024-07-17 20:29:31 +00:00
Code Issues Projects Releases Wiki Activity

Adding 6-, 12-, and 18-bit limited-precision floating-point support for exp2

Browse Source 
function.


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@56025 91177308-0d34-0410-b5e6-96231b3b80d8
This commit is contained in:
Bill Wendling 2008-09-09 22:39:21 +00:00
parent d2e51af035
commit e10c814c52
1 changed files with 111 additions and 26 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

137
lib/CodeGen/SelectionDAG/SelectionDAGBuild.cpp
View File

@ -3276,42 +3276,127 @@ SelectionDAGLowering::visitLog10(CallInst &I) {
setValue(&I, result);
}
/// visitExp2 - lower an exp2 intrinsic. Handles the special sequences
/// for limited-precision mode.
/// visitExp2 - Lower an exp2 intrinsic. Handles the special sequences for
/// limited-precision mode.
void
SelectionDAGLowering::visitExp2(CallInst &I) {
SDValue result;
if (getValue(I.getOperand(1)).getValueType() == MVT::f32 &&
LimitFloatPrecision>0 && LimitFloatPrecision<=12) {
SDValue operand = getValue(I.getOperand(1));
SDValue t0 = DAG.getNode(ISD::FP_TO_SINT, MVT::i32, operand);
SDValue t1 = DAG.getNode(ISD::SINT_TO_FP, MVT::f32, t0);
SDValue t2 = DAG.getNode(ISD::FSUB, MVT::f32, operand, t1);
SDValue t3 = DAG.getNode(ISD::SHL, MVT::i32, t0,
DAG.getConstant(23, MVT::i32));
SDValue t4 = DAG.getNode(ISD::FMUL, MVT::f32, t2,
DAG.getConstantFP(APFloat(
APInt(32, 0x3da235e3)), MVT::f32));
SDValue t5 = DAG.getNode(ISD::FADD, MVT::f32, t4,
DAG.getConstantFP(APFloat(
APInt(32, 0x3e65b8f3)), MVT::f32));
SDValue t6 = DAG.getNode(ISD::FMUL, MVT::f32, t5, t2);
SDValue t7 = DAG.getNode(ISD::FADD, MVT::f32, t6,
DAG.getConstantFP(APFloat(
APInt(32, 0x3f324b07)), MVT::f32));
SDValue t8 = DAG.getNode(ISD::FMUL, MVT::f32, t7, t2);
SDValue t9 = DAG.getNode(ISD::FADD, MVT::f32, t8,
DAG.getConstantFP(APFloat(
APInt(32, 0x3f7ff8fd)), MVT::f32));
SDValue t10 = DAG.getNode(ISD::BIT_CONVERT, MVT::i32, t9);
SDValue t11 = DAG.getNode(ISD::ADD, MVT::i32, t10, t3);
result = DAG.getNode(ISD::BIT_CONVERT, MVT::f32, t11);
LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
SDValue Op = getValue(I.getOperand(1));
SDValue IntegerPartOfX = DAG.getNode(ISD::FP_TO_SINT, MVT::i32, Op);
// FractionalPartOfX = x - (float)IntegerPartOfX;
SDValue t1 = DAG.getNode(ISD::SINT_TO_FP, MVT::f32, IntegerPartOfX);
SDValue X = DAG.getNode(ISD::FSUB, MVT::f32, Op, 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 t6 = DAG.getNode(ISD::BIT_CONVERT, MVT::i32, t5);
SDValue TwoToFractionalPartOfX =
DAG.getNode(ISD::ADD, MVT::i32, t6, IntegerPartOfX);
result = DAG.getNode(ISD::BIT_CONVERT, MVT::f32, TwoToFractionalPartOfX);
} 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;
//
// error 0.000107046256, 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 t8 = DAG.getNode(ISD::BIT_CONVERT, MVT::i32, t7);
SDValue TwoToFractionalPartOfX =
DAG.getNode(ISD::ADD, MVT::i32, t8, IntegerPartOfX);
result = DAG.getNode(ISD::BIT_CONVERT, MVT::f32, TwoToFractionalPartOfX);
} 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 t14 = DAG.getNode(ISD::BIT_CONVERT, MVT::i32, t13);
SDValue TwoToFractionalPartOfX =
DAG.getNode(ISD::ADD, MVT::i32, t14, IntegerPartOfX);
result = DAG.getNode(ISD::BIT_CONVERT, MVT::f32, TwoToFractionalPartOfX);
}
} else {
// No special expansion.
result = DAG.getNode(ISD::FEXP2,
getValue(I.getOperand(1)).getValueType(),
getValue(I.getOperand(1)));
}
setValue(&I, result);
}