[AMDGPU] Adapt new lowering sequence for fdiv16

The current lowering of fdiv16 can generate incorrectly rounded result in some
cases.

Fixes SWDEV-47760.

Author: shiltian

llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp

@@ -4900,16 +4900,33 @@ bool AMDGPULegalizerInfo::legalizeFDIV16(MachineInstr &MI,
   LLT S16 = LLT::scalar(16);
   LLT S32 = LLT::scalar(32);
 
+  // a32.u = opx(V_CVT_F32_F16, a.u);
+  // b32.u = opx(V_CVT_F32_F16, b.u);
+  // r32.u = opx(V_RCP_F32, b32.u);
+  // q32.u = opx(V_MUL_F32, a32.u, r32.u);
+  // e32.u = opx(V_MAD_F32, (b32.u^_neg32), q32.u, a32.u);
+  // q32.u = opx(V_MAD_F32, e32.u, r32.u, q32.u);
+  // e32.u = opx(V_MAD_F32, (b32.u^_neg32), q32.u, a32.u);
+  // tmp.u = opx(V_MUL_F32, e32.u, r32.u);
+  // tmp.u = opx(V_AND_B32, tmp.u, 0xff800000);
+  // q32.u = opx(V_ADD_F32, tmp.u, q32.u);
+  // q16.u = opx(V_CVT_F16_F32, q32.u);
+  // q16.u = opx(V_DIV_FIXUP_F16, q16.u, b.u, a.u);
+
   auto LHSExt = B.buildFPExt(S32, LHS, Flags);
   auto RHSExt = B.buildFPExt(S32, RHS, Flags);
-
-  auto RCP = B.buildIntrinsic(Intrinsic::amdgcn_rcp, {S32})
+  auto NegRHSExt = B.buildFNeg(S32, RHSExt);
+  auto Rcp = B.buildIntrinsic(Intrinsic::amdgcn_rcp, {S32})
                  .addUse(RHSExt.getReg(0))
                  .setMIFlags(Flags);
-
-  auto QUOT = B.buildFMul(S32, LHSExt, RCP, Flags);
-  auto RDst = B.buildFPTrunc(S16, QUOT, Flags);
-
+  auto Quot = B.buildFMul(S32, LHSExt, Rcp);
+  auto Err = B.buildFMA(S32, NegRHSExt, Quot, LHSExt);
+  Quot = B.buildFMA(S32, Err, Rcp, Quot);
+  Err = B.buildFMA(S32, NegRHSExt, Quot, LHSExt);
+  auto Tmp = B.buildFMul(S32, Err, Rcp);
+  Tmp = B.buildAnd(S32, Tmp, B.buildConstant(S32, 0xff800000));
+  Quot = B.buildFAdd(S32, Tmp, Quot);
+  auto RDst = B.buildFPTrunc(S16, Quot, Flags);
   B.buildIntrinsic(Intrinsic::amdgcn_div_fixup, Res)
       .addUse(RDst.getReg(0))
       .addUse(RHS)

llvm/lib/Target/AMDGPU/SIISelLowering.cpp

@@ -10606,19 +10606,40 @@ SDValue SITargetLowering::LowerFDIV16(SDValue Op, SelectionDAG &DAG) const {
     return FastLowered;
 
   SDLoc SL(Op);
-  SDValue Src0 = Op.getOperand(0);
-  SDValue Src1 = Op.getOperand(1);
-
-  SDValue CvtSrc0 = DAG.getNode(ISD::FP_EXTEND, SL, MVT::f32, Src0);
-  SDValue CvtSrc1 = DAG.getNode(ISD::FP_EXTEND, SL, MVT::f32, Src1);
-
-  SDValue RcpSrc1 = DAG.getNode(AMDGPUISD::RCP, SL, MVT::f32, CvtSrc1);
-  SDValue Quot = DAG.getNode(ISD::FMUL, SL, MVT::f32, CvtSrc0, RcpSrc1);
-
-  SDValue FPRoundFlag = DAG.getTargetConstant(0, SL, MVT::i32);
-  SDValue BestQuot = DAG.getNode(ISD::FP_ROUND, SL, MVT::f16, Quot, FPRoundFlag);
+  SDValue LHS = Op.getOperand(0);
+  SDValue RHS = Op.getOperand(1);
 
-  return DAG.getNode(AMDGPUISD::DIV_FIXUP, SL, MVT::f16, BestQuot, Src1, Src0);
+  // a32.u = opx(V_CVT_F32_F16, a.u);
+  // b32.u = opx(V_CVT_F32_F16, b.u);
+  // r32.u = opx(V_RCP_F32, b32.u);
+  // q32.u = opx(V_MUL_F32, a32.u, r32.u);
+  // e32.u = opx(V_MAD_F32, (b32.u^_neg32), q32.u, a32.u);
+  // q32.u = opx(V_MAD_F32, e32.u, r32.u, q32.u);
+  // e32.u = opx(V_MAD_F32, (b32.u^_neg32), q32.u, a32.u);
+  // tmp.u = opx(V_MUL_F32, e32.u, r32.u);
+  // tmp.u = opx(V_AND_B32, tmp.u, 0xff800000)
+  // tmp.u = opx(V_FREXP_MANT_F32, tmp.u);
+  // q32.u = opx(V_ADD_F32, tmp.u, q32.u);
+  // q16.u = opx(V_CVT_F16_F32, q32.u);
+  // q16.u = opx(V_DIV_FIXUP_F16, q16.u, b.u, a.u);
+
+  SDValue LHSExt = DAG.getNode(ISD::FP_EXTEND, SL, MVT::f32, LHS);
+  SDValue RHSExt = DAG.getNode(ISD::FP_EXTEND, SL, MVT::f32, RHS);
+  SDValue NegRHSExt = DAG.getNode(ISD::FNEG, SL, MVT::f32, RHSExt);
+  SDValue Rcp = DAG.getNode(AMDGPUISD::RCP, SL, MVT::f32, RHSExt);
+  SDValue Quot = DAG.getNode(ISD::FMUL, SL, MVT::f32, LHSExt, Rcp);
+  SDValue Err = DAG.getNode(ISD::FMA, SL, MVT::f32, NegRHSExt, Quot, LHSExt);
+  Quot = DAG.getNode(ISD::FMA, SL, MVT::f32, Err, Rcp, Quot);
+  Err = DAG.getNode(ISD::FMA, SL, MVT::f32, NegRHSExt, Quot, LHSExt);
+  SDValue Tmp = DAG.getNode(ISD::FMUL, SL, MVT::f32, Err, Rcp);
+  SDValue TmpCast = DAG.getNode(ISD::BITCAST, SL, MVT::i32, Tmp);
+  TmpCast = DAG.getNode(ISD::AND, SL, MVT::i32, TmpCast,
+                        DAG.getConstant(0xff800000, SL, MVT::i32));
+  Tmp = DAG.getNode(ISD::BITCAST, SL, MVT::f32, TmpCast);
+  Quot = DAG.getNode(ISD::FADD, SL, MVT::f32, Tmp, Quot);
+  SDValue FPRoundFlag = DAG.getConstant(0, SL, MVT::i32);
+  SDValue RDst = DAG.getNode(ISD::FP_ROUND, SL, MVT::f16, Quot, FPRoundFlag);
+  return DAG.getNode(AMDGPUISD::DIV_FIXUP, SL, MVT::f16, RDst, RHS, LHS);
 }
 
 // Faster 2.5 ULP division that does not support denormals.