AMDGPU/GlobalISel: add RegBankLegalize rules for extends and trunc

Uniform S1:
Truncs to uniform S1 and AnyExts from S1 are left as is as they are meant
to be combined away. Uniform S1 ZExt and SExt are lowered using select.
Divergent S1:
Trunc of VGPR to VCC is lowered as compare.
Extends of VCC are lowered using select.

For remaining types:
S32 to S64 ZExt and SExt are lowered using merge values, AnyExt and Trunc
are again left as is to be combined away.
Notably uniform S16 for SExt and Zext is not lowered to S32 and left as is
for instruction select to deal with them. This is because there are patterns
that check for S16 type.

Author: petar-avramovic

llvm/lib/Target/AMDGPU/AMDGPURegBankLegalize.cpp

@@ -214,6 +214,13 @@ class AMDGPURegBankLegalizeCombiner {
       return;
     }
 
+    if (DstTy == S64 && TruncSrcTy == S32) {
+      B.buildMergeLikeInstr(MI.getOperand(0).getReg(),
+                            {TruncSrc, B.buildUndef({SgprRB, S32})});
+      cleanUpAfterCombine(MI, Trunc);
+      return;
+    }
+
     if (DstTy == S32 && TruncSrcTy == S16) {
       B.buildAnyExt(Dst, TruncSrc);
       cleanUpAfterCombine(MI, Trunc);

llvm/lib/Target/AMDGPU/AMDGPURegBankLegalizeHelper.cpp

@@ -131,6 +131,40 @@ void RegBankLegalizeHelper::widenLoad(MachineInstr &MI, LLT WideTy,
   MI.eraseFromParent();
 }
 
+void RegBankLegalizeHelper::lowerVccExtToSel(MachineInstr &MI) {
+  Register Dst = MI.getOperand(0).getReg();
+  LLT Ty = MRI.getType(Dst);
+  Register Src = MI.getOperand(1).getReg();
+  unsigned Opc = MI.getOpcode();
+  if (Ty == S32 || Ty == S16) {
+    auto True = B.buildConstant({VgprRB, Ty}, Opc == G_SEXT ? -1 : 1);
+    auto False = B.buildConstant({VgprRB, Ty}, 0);
+    B.buildSelect(Dst, Src, True, False);
+  }
+  if (Ty == S64) {
+    auto True = B.buildConstant({VgprRB, S32}, Opc == G_SEXT ? -1 : 1);
+    auto False = B.buildConstant({VgprRB, S32}, 0);
+    auto Lo = B.buildSelect({VgprRB, S32}, Src, True, False);
+    MachineInstrBuilder Hi;
+    switch (Opc) {
+    case G_SEXT:
+      Hi = Lo;
+      break;
+    case G_ZEXT:
+      Hi = False;
+      break;
+    case G_ANYEXT:
+      Hi = B.buildUndef({VgprRB_S32});
+      break;
+    default:
+      llvm_unreachable("Opcode not supported");
+    }
+
+    B.buildMergeValues(Dst, {Lo.getReg(0), Hi.getReg(0)});
+  }
+  MI.eraseFromParent();
+}
+
 bool isSignedBFE(MachineInstr &MI) {
   if (isa<GIntrinsic>(MI)) {
     switch (MI.getOperand(1).getIntrinsicID()) {
@@ -267,26 +301,8 @@ void RegBankLegalizeHelper::lower(MachineInstr &MI,
   switch (Mapping.LoweringMethod) {
   case DoNotLower:
     return;
-  case VccExtToSel: {
-    LLT Ty = MRI.getType(MI.getOperand(0).getReg());
-    Register Src = MI.getOperand(1).getReg();
-    unsigned Opc = MI.getOpcode();
-    if (Ty == S32 || Ty == S16) {
-      auto True = B.buildConstant({VgprRB, Ty}, Opc == G_SEXT ? -1 : 1);
-      auto False = B.buildConstant({VgprRB, Ty}, 0);
-      B.buildSelect(MI.getOperand(0).getReg(), Src, True, False);
-    }
-    if (Ty == S64) {
-      auto True = B.buildConstant({VgprRB, S32}, Opc == G_SEXT ? -1 : 1);
-      auto False = B.buildConstant({VgprRB, S32}, 0);
-      auto Sel = B.buildSelect({VgprRB, S32}, Src, True, False);
-      B.buildMergeValues(
-          MI.getOperand(0).getReg(),
-          {Sel.getReg(0), Opc == G_SEXT ? Sel.getReg(0) : False.getReg(0)});
-    }
-    MI.eraseFromParent();
-    return;
-  }
+  case VccExtToSel:
+    return lowerVccExtToSel(MI);
   case UniExtToSel: {
     LLT Ty = MRI.getType(MI.getOperand(0).getReg());
     auto True = B.buildConstant({SgprRB, Ty},
@@ -303,13 +319,23 @@ void RegBankLegalizeHelper::lower(MachineInstr &MI,
   case Ext32To64: {
     const RegisterBank *RB = MRI.getRegBank(MI.getOperand(0).getReg());
     MachineInstrBuilder Hi;
-
-    if (MI.getOpcode() == AMDGPU::G_ZEXT) {
+    switch (MI.getOpcode()) {
+    case AMDGPU::G_ZEXT: {
       Hi = B.buildConstant({RB, S32}, 0);
-    } else {
+      break;
+    }
+    case AMDGPU::G_SEXT: {
       // Replicate sign bit from 32-bit extended part.
       auto ShiftAmt = B.buildConstant({RB, S32}, 31);
       Hi = B.buildAShr({RB, S32}, MI.getOperand(1).getReg(), ShiftAmt);
+      break;
+    }
+    case AMDGPU::G_ANYEXT: {
+      Hi = B.buildUndef({RB, S32});
+      break;
+    }
+    default:
+      llvm_unreachable("Unsuported Opcode in Ext32To64");
     }
 
     B.buildMergeLikeInstr(MI.getOperand(0).getReg(),
@@ -332,7 +358,7 @@ void RegBankLegalizeHelper::lower(MachineInstr &MI,
     // compares all bits in register.
     Register BoolSrc = MRI.createVirtualRegister({VgprRB, Ty});
     if (Ty == S64) {
-      auto Src64 = B.buildUnmerge({VgprRB, Ty}, Src);
+      auto Src64 = B.buildUnmerge(VgprRB_S32, Src);
       auto One = B.buildConstant(VgprRB_S32, 1);
       auto AndLo = B.buildAnd(VgprRB_S32, Src64.getReg(0), One);
       auto Zero = B.buildConstant(VgprRB_S32, 0);
@@ -420,8 +446,11 @@ LLT RegBankLegalizeHelper::getTyFromID(RegBankLLTMappingApplyID ID) {
   case Sgpr32AExt:
   case Sgpr32AExtBoolInReg:
   case Sgpr32SExt:
+  case Sgpr32ZExt:
   case UniInVgprS32:
   case Vgpr32:
+  case Vgpr32SExt:
+  case Vgpr32ZExt:
     return LLT::scalar(32);
   case Sgpr64:
   case Vgpr64:
@@ -532,6 +561,7 @@ RegBankLegalizeHelper::getRegBankFromID(RegBankLLTMappingApplyID ID) {
   case Sgpr32AExt:
   case Sgpr32AExtBoolInReg:
   case Sgpr32SExt:
+  case Sgpr32ZExt:
     return SgprRB;
   case Vgpr16:
   case Vgpr32:
@@ -548,6 +578,8 @@ RegBankLegalizeHelper::getRegBankFromID(RegBankLLTMappingApplyID ID) {
   case VgprB128:
   case VgprB256:
   case VgprB512:
+  case Vgpr32SExt:
+  case Vgpr32ZExt:
     return VgprRB;
   default:
     return nullptr;
@@ -753,8 +785,8 @@ void RegBankLegalizeHelper::applyMappingSrc(
       assert(Ty.getSizeInBits() == 1);
       assert(RB == SgprRB);
       auto Aext = B.buildAnyExt(SgprRB_S32, Reg);
-      // Zext SgprS1 is not legal, this instruction is most of times meant to be
-      // combined away in RB combiner, so do not make AND with 1.
+      // Zext SgprS1 is not legal, make AND with 1 instead. This instruction is
+      // most of times meant to be combined away in AMDGPURegBankCombiner.
       auto Cst1 = B.buildConstant(SgprRB_S32, 1);
       auto BoolInReg = B.buildAnd(SgprRB_S32, Aext, Cst1);
       Op.setReg(BoolInReg.getReg(0));
@@ -767,6 +799,29 @@ void RegBankLegalizeHelper::applyMappingSrc(
       Op.setReg(Sext.getReg(0));
       break;
     }
+    case Sgpr32ZExt: {
+      assert(1 < Ty.getSizeInBits() && Ty.getSizeInBits() < 32);
+      assert(RB == SgprRB);
+      auto Zext = B.buildZExt({SgprRB, S32}, Reg);
+      Op.setReg(Zext.getReg(0));
+      break;
+    }
+    case Vgpr32SExt: {
+      // Note this ext allows S1, and it is meant to be combined away.
+      assert(Ty.getSizeInBits() < 32);
+      assert(RB == VgprRB);
+      auto Sext = B.buildSExt({VgprRB, S32}, Reg);
+      Op.setReg(Sext.getReg(0));
+      break;
+    }
+    case Vgpr32ZExt: {
+      // Note this ext allows S1, and it is meant to be combined away.
+      assert(Ty.getSizeInBits() < 32);
+      assert(RB == VgprRB);
+      auto Zext = B.buildZExt({VgprRB, S32}, Reg);
+      Op.setReg(Zext.getReg(0));
+      break;
+    }
     default:
       llvm_unreachable("ID not supported");
     }

llvm/lib/Target/AMDGPU/AMDGPURegBankLegalizeHelper.h

@@ -110,6 +110,7 @@ class RegBankLegalizeHelper {
   void lower(MachineInstr &MI, const RegBankLLTMapping &Mapping,
              SmallSet<Register, 4> &SgprWaterfallOperandRegs);
 
+  void lowerVccExtToSel(MachineInstr &MI);
   void lowerDiv_BFE(MachineInstr &MI);
   void lowerUni_BFE(MachineInstr &MI);
   void lowerSplitTo32(MachineInstr &MI);

llvm/lib/Target/AMDGPU/AMDGPURegBankLegalizeRules.cpp

@@ -489,22 +489,61 @@ RegBankLegalizeRules::RegBankLegalizeRules(const GCNSubtarget &_ST,
       .Uni(B32, {{SgprB32}, {Sgpr32AExtBoolInReg, SgprB32, SgprB32}});
 
   addRulesForGOpcs({G_ANYEXT})
+      .Any({{UniS16, S1}, {{None}, {None}}}) // should be combined away
       .Any({{UniS32, S1}, {{None}, {None}}}) // should be combined away
-      .Any({{UniS32, S16}, {{Sgpr32}, {Sgpr16}}});
+      .Any({{UniS64, S1}, {{None}, {None}}}) // should be combined away
+      .Any({{DivS16, S1}, {{Vgpr16}, {Vcc}, VccExtToSel}})
+      .Any({{DivS32, S1}, {{Vgpr32}, {Vcc}, VccExtToSel}})
+      .Any({{DivS64, S1}, {{Vgpr64}, {Vcc}, VccExtToSel}})
+      .Any({{UniS64, S32}, {{Sgpr64}, {Sgpr32}, Ext32To64}})
+      .Any({{DivS64, S32}, {{Vgpr64}, {Vgpr32}, Ext32To64}})
+      .Any({{UniS32, S16}, {{Sgpr32}, {Sgpr16}}})
+      .Any({{DivS32, S16}, {{Vgpr32}, {Vgpr16}}});
 
   // In global-isel G_TRUNC in-reg is treated as no-op, inst selected into COPY.
   // It is up to user to deal with truncated bits.
   addRulesForGOpcs({G_TRUNC})
+      .Any({{UniS1, UniS16}, {{None}, {None}}}) // should be combined away
       .Any({{UniS1, UniS32}, {{None}, {None}}}) // should be combined away
+      .Any({{UniS1, UniS64}, {{None}, {None}}}) // should be combined away
       .Any({{UniS16, S32}, {{Sgpr16}, {Sgpr32}}})
+      .Any({{DivS16, S32}, {{Vgpr16}, {Vgpr32}}})
+      .Any({{UniS32, S64}, {{Sgpr32}, {Sgpr64}}})
+      .Any({{DivS32, S64}, {{Vgpr32}, {Vgpr64}}})
       // This is non-trivial. VgprToVccCopy is done using compare instruction.
-      .Any({{DivS1, DivS32}, {{Vcc}, {Vgpr32}, VgprToVccCopy}});
+      .Any({{DivS1, DivS16}, {{Vcc}, {Vgpr16}, VgprToVccCopy}})
+      .Any({{DivS1, DivS32}, {{Vcc}, {Vgpr32}, VgprToVccCopy}})
+      .Any({{DivS1, DivS64}, {{Vcc}, {Vgpr64}, VgprToVccCopy}});
 
-  addRulesForGOpcs({G_ZEXT, G_SEXT})
+  addRulesForGOpcs({G_ZEXT})
+      .Any({{UniS16, S1}, {{Sgpr32Trunc}, {Sgpr32AExtBoolInReg}, UniExtToSel}})
+      .Any({{UniS32, S1}, {{Sgpr32}, {Sgpr32AExtBoolInReg}, UniExtToSel}})
+      .Any({{UniS64, S1}, {{Sgpr64}, {Sgpr32AExtBoolInReg}, UniExtToSel}})
+      .Any({{DivS16, S1}, {{Vgpr16}, {Vcc}, VccExtToSel}})
+      .Any({{DivS32, S1}, {{Vgpr32}, {Vcc}, VccExtToSel}})
+      .Any({{DivS64, S1}, {{Vgpr64}, {Vcc}, VccExtToSel}})
+      .Any({{UniS64, S32}, {{Sgpr64}, {Sgpr32}, Ext32To64}})
+      .Any({{DivS64, S32}, {{Vgpr64}, {Vgpr32}, Ext32To64}})
+      // not extending S16 to S32 is questionable.
+      .Any({{UniS64, S16}, {{Sgpr64}, {Sgpr32ZExt}, Ext32To64}})
+      .Any({{DivS64, S16}, {{Vgpr64}, {Vgpr32ZExt}, Ext32To64}})
+      .Any({{UniS32, S16}, {{Sgpr32}, {Sgpr16}}})
+      .Any({{DivS32, S16}, {{Vgpr32}, {Vgpr16}}});
+
+  addRulesForGOpcs({G_SEXT})
+      .Any({{UniS16, S1}, {{Sgpr32Trunc}, {Sgpr32AExtBoolInReg}, UniExtToSel}})
       .Any({{UniS32, S1}, {{Sgpr32}, {Sgpr32AExtBoolInReg}, UniExtToSel}})
+      .Any({{UniS64, S1}, {{Sgpr64}, {Sgpr32AExtBoolInReg}, UniExtToSel}})
+      .Any({{DivS16, S1}, {{Vgpr16}, {Vcc}, VccExtToSel}})
       .Any({{DivS32, S1}, {{Vgpr32}, {Vcc}, VccExtToSel}})
+      .Any({{DivS64, S1}, {{Vgpr64}, {Vcc}, VccExtToSel}})
       .Any({{UniS64, S32}, {{Sgpr64}, {Sgpr32}, Ext32To64}})
-      .Any({{DivS64, S32}, {{Vgpr64}, {Vgpr32}, Ext32To64}});
+      .Any({{DivS64, S32}, {{Vgpr64}, {Vgpr32}, Ext32To64}})
+      // not extending S16 to S32 is questionable.
+      .Any({{UniS64, S16}, {{Sgpr64}, {Sgpr32SExt}, Ext32To64}})
+      .Any({{DivS64, S16}, {{Vgpr64}, {Vgpr32SExt}, Ext32To64}})
+      .Any({{UniS32, S16}, {{Sgpr32}, {Sgpr16}}})
+      .Any({{DivS32, S16}, {{Vgpr32}, {Vgpr16}}});
 
   bool hasUnalignedLoads = ST->getGeneration() >= AMDGPUSubtarget::GFX12;
   bool hasSMRDSmall = ST->hasScalarSubwordLoads();

llvm/lib/Target/AMDGPU/AMDGPURegBankLegalizeRules.h

@@ -159,6 +159,9 @@ enum RegBankLLTMappingApplyID {
   Sgpr32AExt,
   Sgpr32AExtBoolInReg,
   Sgpr32SExt,
+  Sgpr32ZExt,
+  Vgpr32SExt,
+  Vgpr32ZExt,
 };
 
 // Instruction needs to be replaced with sequence of instructions. Lowering was