[MIPS GlobalISel] Lower fptoui

Add lower for G_FPTOUI. Algorithm is similar to the SDAG version
in TargetLowering::expandFP_TO_UINT.
Lower G_FPTOUI for MIPS32.

Differential Revision: https://reviews.llvm.org/D66929

llvm-svn: 370431

1 files changed, 44 insertions, 0 deletions

diff --git a/llvm/lib/CodeGen/GlobalISel/LegalizerHelper.cpp b/llvm/lib/CodeGen/GlobalISel/LegalizerHelper.cpp
index 79297d35b0c..c37b115fc91 100644
--- a/llvm/lib/CodeGen/GlobalISel/LegalizerHelper.cpp
+++ b/llvm/lib/CodeGen/GlobalISel/LegalizerHelper.cpp
@@ -2125,6 +2125,8 @@ LegalizerHelper::lower(MachineInstr &MI, unsigned TypeIdx, LLT Ty) {
     return lowerUITOFP(MI, TypeIdx, Ty);
   case G_SITOFP:
     return lowerSITOFP(MI, TypeIdx, Ty);
+  case G_FPTOUI:
+    return lowerFPTOUI(MI, TypeIdx, Ty);
   case G_SMIN:
   case G_SMAX:
   case G_UMIN:
@@ -3715,6 +3717,48 @@ LegalizerHelper::lowerSITOFP(MachineInstr &MI, unsigned TypeIdx, LLT Ty) {
   return UnableToLegalize;
 }
 
+LegalizerHelper::LegalizeResult
+LegalizerHelper::lowerFPTOUI(MachineInstr &MI, unsigned TypeIdx, LLT Ty) {
+  Register Dst = MI.getOperand(0).getReg();
+  Register Src = MI.getOperand(1).getReg();
+  LLT DstTy = MRI.getType(Dst);
+  LLT SrcTy = MRI.getType(Src);
+  const LLT S64 = LLT::scalar(64);
+  const LLT S32 = LLT::scalar(32);
+
+  if (SrcTy != S64 && SrcTy != S32)
+    return UnableToLegalize;
+  if (DstTy != S32 && DstTy != S64)
+    return UnableToLegalize;
+
+  // FPTOSI gives same result as FPTOUI for positive signed integers.
+  // FPTOUI needs to deal with fp values that convert to unsigned integers
+  // greater or equal to 2^31 for float or 2^63 for double. For brevity 2^Exp.
+
+  APInt TwoPExpInt = APInt::getSignMask(DstTy.getSizeInBits());
+  APFloat TwoPExpFP(SrcTy.getSizeInBits() == 32 ? APFloat::IEEEsingle()
+                                                : APFloat::IEEEdouble(),
+                    APInt::getNullValue(SrcTy.getSizeInBits()));
+  TwoPExpFP.convertFromAPInt(TwoPExpInt, false, APFloat::rmNearestTiesToEven);
+
+  MachineInstrBuilder FPTOSI = MIRBuilder.buildFPTOSI(DstTy, Src);
+
+  MachineInstrBuilder Threshold = MIRBuilder.buildFConstant(SrcTy, TwoPExpFP);
+  // For fp Value greater or equal to Threshold(2^Exp), we use FPTOSI on
+  // (Value - 2^Exp) and add 2^Exp by setting highest bit in result to 1.
+  MachineInstrBuilder FSub = MIRBuilder.buildFSub(SrcTy, Src, Threshold);
+  MachineInstrBuilder ResLowBits = MIRBuilder.buildFPTOSI(DstTy, FSub);
+  MachineInstrBuilder ResHighBit = MIRBuilder.buildConstant(DstTy, TwoPExpInt);
+  MachineInstrBuilder Res = MIRBuilder.buildXor(DstTy, ResLowBits, ResHighBit);
+
+  MachineInstrBuilder FCMP =
+      MIRBuilder.buildFCmp(CmpInst::FCMP_ULT, DstTy, Src, Threshold);
+  MIRBuilder.buildSelect(Dst, FCMP, FPTOSI, Res);
+
+  MI.eraseFromParent();
+  return Legalized;
+}
+
 static CmpInst::Predicate minMaxToCompare(unsigned Opc) {
   switch (Opc) {
   case TargetOpcode::G_SMIN: