6502/llvm-6502
1
0
Fork 0
mirror of https://github.com/c64scene-ar/llvm-6502.git synced 2024-07-22 09:29:31 +00:00
Code Issues Projects Releases Wiki Activity

Fixed sdiv and udiv for <4 x i16>. The test from r125402 still applies for this change.

Browse Source 
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@131630 91177308-0d34-0410-b5e6-96231b3b80d8
This commit is contained in:
Mon P Wang 2011-05-19 04:15:07 +00:00
parent 91d9a1c0f7
commit 28e2b1d29a
1 changed files with 7 additions and 7 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

14
lib/Target/ARM/ARMISelLowering.cpp
View File

@ -4663,10 +4663,10 @@ LowerSDIV_v4i16(SDValue N0, SDValue N1, DebugLoc dl, SelectionDAG &DAG) {
// Because short has a smaller range than ushort, we can actually get away
// with only a single newton step. This requires that we use a weird bias
// of 89, however (again, this has been exhaustively tested).
// float4 result = as_float4(as_int4(xf*recip) + 89);
// float4 result = as_float4(as_int4(xf*recip) + 0x89);
N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
N1 = DAG.getConstant(89, MVT::i32);
N1 = DAG.getConstant(0x89, MVT::i32);
N1 = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, N1, N1, N1, N1);
N0 = DAG.getNode(ISD::ADD, dl, MVT::v4i32, N0, N1);
N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, N0);
@ -4753,26 +4753,26 @@ static SDValue LowerUDIV(SDValue Op, SelectionDAG &DAG) {
N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N0);
N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v4i32, N1);
N0 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N0);
N1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
SDValue BN1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::v4f32, N1);
// Use reciprocal estimate and two refinement steps.
// float4 recip = vrecpeq_f32(yf);
// recip *= vrecpsq_f32(yf, recip);
// recip *= vrecpsq_f32(yf, recip);
N2 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), N1);
DAG.getConstant(Intrinsic::arm_neon_vrecpe, MVT::i32), BN1);
N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
N1, N2);
BN1, N2);
N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
N1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::v4f32,
DAG.getConstant(Intrinsic::arm_neon_vrecps, MVT::i32),
N1, N2);
BN1, N2);
N2 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N1, N2);
// Simply multiplying by the reciprocal estimate can leave us a few ulps
// too low, so we add 2 ulps (exhaustive testing shows that this is enough,
// and that it will never cause us to return an answer too large).
// float4 result = as_float4(as_int4(xf*recip) + 89);
// float4 result = as_float4(as_int4(xf*recip) + 2);
N0 = DAG.getNode(ISD::FMUL, dl, MVT::v4f32, N0, N2);
N0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, N0);
N1 = DAG.getConstant(2, MVT::i32);