AMDGPU/GlobalISel: Legalize G_FRINT

[arsenm]

Summary

Copied as directly as possible from the current lowering

[tstellar]

LGTM.

[arsenm]

r361026

lib/Target/AMDGPU/AMDGPULegalizerInfo.h

bool legalizeCustom(MachineInstr &MI, MachineRegisterInfo &MRI,
                       GISelChangeObserver &Observer) const override;
 
   unsigned getSegmentAperture(unsigned AddrSpace,
                               MachineRegisterInfo &MRI,
                               MachineIRBuilder &MIRBuilder) const;
 
   bool legalizeAddrSpaceCast(MachineInstr &MI, MachineRegisterInfo &MRI,
                              MachineIRBuilder &MIRBuilder) const;
+  bool legalizeFrint(MachineInstr &MI, MachineRegisterInfo &MRI,
+                     MachineIRBuilder &MIRBuilder) const;
+
 };
 } // End llvm namespace.
 #endif

lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp

AMDGPULegalizerInfo::AMDGPULegalizerInfo(const GCNSubtarget &ST,
   getActionDefinitionsBuilder({G_FPTOSI, G_FPTOUI})
     .legalFor({{S32, S32}, {S32, S64}})
     .scalarize(0);
 
   getActionDefinitionsBuilder({G_INTRINSIC_TRUNC, G_INTRINSIC_ROUND})
     .legalFor({S32, S64})
     .scalarize(0);
 
+  if (ST.getGeneration() >= AMDGPUSubtarget::SEA_ISLANDS) {
+    getActionDefinitionsBuilder(G_FRINT)
+      .legalFor({S32, S64})
+      .clampScalar(0, S32, S64)
+      .scalarize(0);
+  } else {
+    getActionDefinitionsBuilder(G_FRINT)
+      .legalFor({S32})
+      .customFor({S64})
+      .clampScalar(0, S32, S64)
+      .scalarize(0);
+  }
 
   getActionDefinitionsBuilder(G_GEP)
     .legalForCartesianProduct(AddrSpaces64, {S64})
     .legalForCartesianProduct(AddrSpaces32, {S32})
     .scalarize(0);
 
   setAction({G_BLOCK_ADDR, CodePtr}, Legal);
 

 
 bool AMDGPULegalizerInfo::legalizeCustom(MachineInstr &MI,
                                          MachineRegisterInfo &MRI,
                                          MachineIRBuilder &MIRBuilder,
                                          GISelChangeObserver &Observer) const {
   switch (MI.getOpcode()) {
   case TargetOpcode::G_ADDRSPACE_CAST:
     return legalizeAddrSpaceCast(MI, MRI, MIRBuilder);
+  case TargetOpcode::G_FRINT:
+    return legalizeFrint(MI, MRI, MIRBuilder);
   default:
     return false;
   }
 
   llvm_unreachable("expected switch to return");
 }
 
 unsigned AMDGPULegalizerInfo::getSegmentAperture(
bool AMDGPULegalizerInfo::legalizeAddrSpaceCast(
   // TODO: Should we allow mismatched types but matching sizes in merges to
   // avoid the ptrtoint?
   MIRBuilder.buildMerge(BuildPtr, {SrcAsInt, ApertureReg});
   MIRBuilder.buildSelect(Dst, CmpRes, BuildPtr, FlatNull.getReg(0));
 
   MI.eraseFromParent();
   return true;
 }
+
+bool AMDGPULegalizerInfo::legalizeFrint(
+  MachineInstr &MI, MachineRegisterInfo &MRI,
+  MachineIRBuilder &MIRBuilder) const {
+  MIRBuilder.setInstr(MI);
+
+  unsigned Src = MI.getOperand(1).getReg();
+  LLT Ty = MRI.getType(Src);
+  assert(Ty.isScalar() && Ty.getSizeInBits() == 64);
+
+  APFloat C1Val(APFloat::IEEEdouble(), "0x1.0p+52");
+  APFloat C2Val(APFloat::IEEEdouble(), "0x1.fffffffffffffp+51");
+
+  auto C1 = MIRBuilder.buildFConstant(Ty, C1Val);
+  auto CopySign = MIRBuilder.buildFCopysign(Ty, C1, Src);
+
+  // TODO: Should this propagate fast-math-flags?
+  auto Tmp1 = MIRBuilder.buildFAdd(Ty, Src, CopySign);
+  auto Tmp2 = MIRBuilder.buildFSub(Ty, Tmp1, CopySign);
+
+  auto C2 = MIRBuilder.buildFConstant(Ty, C2Val);
+  auto Fabs = MIRBuilder.buildFAbs(Ty, Src);
+
+  auto Cond = MIRBuilder.buildFCmp(CmpInst::FCMP_OGT, LLT::scalar(1), Fabs, C2);
+  MIRBuilder.buildSelect(MI.getOperand(0).getReg(), Cond, Src, Tmp2);
+  return true;
+}

test/CodeGen/AMDGPU/GlobalISel/legalize-frint.mir

+# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
+# RUN: llc -mtriple=amdgcn-mesa-mesa3d -mcpu=tahiti -run-pass=legalizer -o - %s | FileCheck -check-prefix=SI %s
+# RUN: llc -mtriple=amdgcn-mesa-mesa3d -mcpu=hawaii -run-pass=legalizer -o - %s | FileCheck -check-prefix=CI %s
+
+---
+name: test_frint_s16
+body: |
+  bb.0:
+    liveins: $vgpr0
+
+    ; SI-LABEL: name: test_frint_s16
+    ; SI: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0
+    ; SI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[COPY]](s32)
+    ; SI: [[FPEXT:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC]](s16)
+    ; SI: [[FRINT:%[0-9]+]]:_(s32) = G_FRINT [[FPEXT]]
+    ; SI: [[FPTRUNC:%[0-9]+]]:_(s16) = G_FPTRUNC [[FRINT]](s32)
+    ; SI: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[FPTRUNC]](s16)
+    ; SI: $vgpr0 = COPY [[ANYEXT]](s32)
+    ; CI-LABEL: name: test_frint_s16
+    ; CI: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0
+    ; CI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[COPY]](s32)
+    ; CI: [[FPEXT:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC]](s16)
+    ; CI: [[FRINT:%[0-9]+]]:_(s32) = G_FRINT [[FPEXT]]
+    ; CI: [[FPTRUNC:%[0-9]+]]:_(s16) = G_FPTRUNC [[FRINT]](s32)
+    ; CI: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[FPTRUNC]](s16)
+    ; CI: $vgpr0 = COPY [[ANYEXT]](s32)
+    %0:_(s32) = COPY $vgpr0
+    %1:_(s16) = G_TRUNC %0
+    %2:_(s16) = G_FRINT %1
+    %3:_(s32) = G_ANYEXT %2
+    $vgpr0 = COPY %3
+...
+
+---
+name: test_frint_s32
+body: |
+  bb.0:
+    liveins: $vgpr0
+
+    ; SI-LABEL: name: test_frint_s32
+    ; SI: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0
+    ; SI: $vgpr0 = COPY [[COPY]](s32)
+    ; CI-LABEL: name: test_frint_s32
+    ; CI: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0
+    ; CI: $vgpr0 = COPY [[COPY]](s32)
+    %0:_(s32) = COPY $vgpr0
+    %1:_(s32) = G_FRINT %0
+    $vgpr0 = COPY %0
+...
+
+---
+name: test_frint_s64
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_frint_s64
+    ; SI: [[COPY:%[0-9]+]]:_(s64) = COPY $vgpr0_vgpr1
+    ; SI: [[C:%[0-9]+]]:_(s64) = G_FCONSTANT double 0x4330000000000000
+    ; SI: [[FCOPYSIGN:%[0-9]+]]:_(s64) = G_FCOPYSIGN [[C]], [[COPY]](s64)
+    ; SI: [[FADD:%[0-9]+]]:_(s64) = G_FADD [[COPY]], [[FCOPYSIGN]]
+    ; SI: [[FNEG:%[0-9]+]]:_(s64) = G_FNEG [[FCOPYSIGN]]
+    ; SI: [[FADD1:%[0-9]+]]:_(s64) = G_FADD [[FADD]], [[FNEG]]
+    ; SI: [[C1:%[0-9]+]]:_(s64) = G_FCONSTANT double 0x432FFFFFFFFFFFFF
+    ; SI: [[FABS:%[0-9]+]]:_(s64) = G_FABS [[COPY]]
+    ; SI: [[FCMP:%[0-9]+]]:_(s1) = G_FCMP floatpred(ogt), [[FABS]](s64), [[C1]]
+    ; SI: [[SELECT:%[0-9]+]]:_(s64) = G_SELECT [[FCMP]](s1), [[COPY]], [[FADD1]]
+    ; SI: [[FRINT:%[0-9]+]]:_(s64) = G_FRINT [[COPY]]
+    ; SI: $vgpr0_vgpr1 = COPY [[FRINT]](s64)
+    ; CI-LABEL: name: test_frint_s64
+    ; CI: [[COPY:%[0-9]+]]:_(s64) = COPY $vgpr0_vgpr1
+    ; CI: [[FRINT:%[0-9]+]]:_(s64) = G_FRINT [[COPY]]
+    ; CI: $vgpr0_vgpr1 = COPY [[FRINT]](s64)
+    %0:_(s64) = COPY $vgpr0_vgpr1
+    %1:_(s64) = G_FRINT %0
+    $vgpr0_vgpr1 = COPY %1
+...
+
+---
+name: test_frint_v2s16
+body: |
+  bb.0:
+    liveins: $vgpr0
+
+    ; SI-LABEL: name: test_frint_v2s16
+    ; SI: [[COPY:%[0-9]+]]:_(<2 x s16>) = COPY $vgpr0
+    ; SI: $vgpr0 = COPY [[COPY]](<2 x s16>)
+    ; CI-LABEL: name: test_frint_v2s16
+    ; CI: [[COPY:%[0-9]+]]:_(<2 x s16>) = COPY $vgpr0
+    ; CI: $vgpr0 = COPY [[COPY]](<2 x s16>)
+    %0:_(<2 x s16>) = COPY $vgpr0
+    %1:_(<2 x s16>) = G_FRINT %0
+    $vgpr0 = COPY %0
+...
+
+---
+name: test_frint_v2s32
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_frint_v2s32
+    ; SI: [[COPY:%[0-9]+]]:_(<2 x s32>) = COPY $vgpr0_vgpr1
+    ; SI: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[COPY]](<2 x s32>)
+    ; SI: [[FRINT:%[0-9]+]]:_(s32) = G_FRINT [[UV]]
+    ; SI: [[FRINT1:%[0-9]+]]:_(s32) = G_FRINT [[UV1]]
+    ; SI: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[FRINT]](s32), [[FRINT1]](s32)
+    ; SI: $vgpr0_vgpr1 = COPY [[BUILD_VECTOR]](<2 x s32>)
+    ; CI-LABEL: name: test_frint_v2s32
+    ; CI: [[COPY:%[0-9]+]]:_(<2 x s32>) = COPY $vgpr0_vgpr1
+    ; CI: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[COPY]](<2 x s32>)
+    ; CI: [[FRINT:%[0-9]+]]:_(s32) = G_FRINT [[UV]]
+    ; CI: [[FRINT1:%[0-9]+]]:_(s32) = G_FRINT [[UV1]]
+    ; CI: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[FRINT]](s32), [[FRINT1]](s32)
+    ; CI: $vgpr0_vgpr1 = COPY [[BUILD_VECTOR]](<2 x s32>)
+    %0:_(<2 x s32>) = COPY $vgpr0_vgpr1
+    %1:_(<2 x s32>) = G_FRINT %0
+    $vgpr0_vgpr1 = COPY %1
+...
+
+---
+name: test_frint_v2s64
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1_vgpr2_vgpr3
+
+    ; SI-LABEL: name: test_frint_v2s64
+    ; SI: [[COPY:%[0-9]+]]:_(<2 x s64>) = COPY $vgpr0_vgpr1_vgpr2_vgpr3
+    ; SI: [[UV:%[0-9]+]]:_(s64), [[UV1:%[0-9]+]]:_(s64) = G_UNMERGE_VALUES [[COPY]](<2 x s64>)
+    ; SI: [[C:%[0-9]+]]:_(s64) = G_FCONSTANT double 0x4330000000000000
+    ; SI: [[FCOPYSIGN:%[0-9]+]]:_(s64) = G_FCOPYSIGN [[C]], [[UV]](s64)
+    ; SI: [[FADD:%[0-9]+]]:_(s64) = G_FADD [[UV]], [[FCOPYSIGN]]
+    ; SI: [[FNEG:%[0-9]+]]:_(s64) = G_FNEG [[FCOPYSIGN]]
+    ; SI: [[FADD1:%[0-9]+]]:_(s64) = G_FADD [[FADD]], [[FNEG]]
+    ; SI: [[C1:%[0-9]+]]:_(s64) = G_FCONSTANT double 0x432FFFFFFFFFFFFF
+    ; SI: [[FABS:%[0-9]+]]:_(s64) = G_FABS [[UV]]
+    ; SI: [[FCMP:%[0-9]+]]:_(s1) = G_FCMP floatpred(ogt), [[FABS]](s64), [[C1]]
+    ; SI: [[SELECT:%[0-9]+]]:_(s64) = G_SELECT [[FCMP]](s1), [[UV]], [[FADD1]]
+    ; SI: [[FRINT:%[0-9]+]]:_(s64) = G_FRINT [[UV]]
+    ; SI: [[FCOPYSIGN1:%[0-9]+]]:_(s64) = G_FCOPYSIGN [[C]], [[UV1]](s64)
+    ; SI: [[FADD2:%[0-9]+]]:_(s64) = G_FADD [[UV1]], [[FCOPYSIGN1]]
+    ; SI: [[FNEG1:%[0-9]+]]:_(s64) = G_FNEG [[FCOPYSIGN1]]
+    ; SI: [[FADD3:%[0-9]+]]:_(s64) = G_FADD [[FADD2]], [[FNEG1]]
+    ; SI: [[FABS1:%[0-9]+]]:_(s64) = G_FABS [[UV1]]
+    ; SI: [[FCMP1:%[0-9]+]]:_(s1) = G_FCMP floatpred(ogt), [[FABS1]](s64), [[C1]]
+    ; SI: [[SELECT:%[0-9]+]]:_(s64) = G_SELECT [[FCMP1]](s1), [[UV1]], [[FADD3]]
+    ; SI: [[FRINT1:%[0-9]+]]:_(s64) = G_FRINT [[UV1]]
+    ; SI: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s64>) = G_BUILD_VECTOR [[FRINT]](s64), [[FRINT1]](s64)
+    ; SI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BUILD_VECTOR]](<2 x s64>)
+    ; CI-LABEL: name: test_frint_v2s64
+    ; CI: [[COPY:%[0-9]+]]:_(<2 x s64>) = COPY $vgpr0_vgpr1_vgpr2_vgpr3
+    ; CI: [[UV:%[0-9]+]]:_(s64), [[UV1:%[0-9]+]]:_(s64) = G_UNMERGE_VALUES [[COPY]](<2 x s64>)
+    ; CI: [[FRINT:%[0-9]+]]:_(s64) = G_FRINT [[UV]]
+    ; CI: [[FRINT1:%[0-9]+]]:_(s64) = G_FRINT [[UV1]]
+    ; CI: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s64>) = G_BUILD_VECTOR [[FRINT]](s64), [[FRINT1]](s64)
+    ; CI: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BUILD_VECTOR]](<2 x s64>)
+    %0:_(<2 x s64>) = COPY $vgpr0_vgpr1_vgpr2_vgpr3
+    %1:_(<2 x s64>) = G_FRINT %0
+    $vgpr0_vgpr1_vgpr2_vgpr3 = COPY %1
+...