[RISCV] Add short forward branch support for min, max, maxu and minu (llvm#164394)

Author: hchandel

llvm/lib/Target/RISCV/RISCVExpandPseudoInsts.cpp

@@ -127,6 +127,10 @@ bool RISCVExpandPseudo::expandMI(MachineBasicBlock &MBB,
   case RISCV::PseudoCCAND:
   case RISCV::PseudoCCOR:
   case RISCV::PseudoCCXOR:
+  case RISCV::PseudoCCMAX:
+  case RISCV::PseudoCCMAXU:
+  case RISCV::PseudoCCMIN:
+  case RISCV::PseudoCCMINU:
   case RISCV::PseudoCCADDW:
   case RISCV::PseudoCCSUBW:
   case RISCV::PseudoCCSLL:
@@ -217,6 +221,7 @@ bool RISCVExpandPseudo::expandCCOp(MachineBasicBlock &MBB,
         .addImm(0);
   } else {
     unsigned NewOpc;
+    // clang-format off
     switch (MI.getOpcode()) {
     default:
       llvm_unreachable("Unexpected opcode!");
@@ -228,6 +233,10 @@ bool RISCVExpandPseudo::expandCCOp(MachineBasicBlock &MBB,
     case RISCV::PseudoCCAND:   NewOpc = RISCV::AND;   break;
     case RISCV::PseudoCCOR:    NewOpc = RISCV::OR;    break;
     case RISCV::PseudoCCXOR:   NewOpc = RISCV::XOR;   break;
+    case RISCV::PseudoCCMAX:   NewOpc = RISCV::MAX;   break;
+    case RISCV::PseudoCCMIN:   NewOpc = RISCV::MIN;   break;
+    case RISCV::PseudoCCMAXU:  NewOpc = RISCV::MAXU;  break;
+    case RISCV::PseudoCCMINU:  NewOpc = RISCV::MINU;  break;
     case RISCV::PseudoCCADDI:  NewOpc = RISCV::ADDI;  break;
     case RISCV::PseudoCCSLLI:  NewOpc = RISCV::SLLI;  break;
     case RISCV::PseudoCCSRLI:  NewOpc = RISCV::SRLI;  break;
@@ -250,6 +259,7 @@ bool RISCVExpandPseudo::expandCCOp(MachineBasicBlock &MBB,
     case RISCV::PseudoCCNDS_BFOS: NewOpc = RISCV::NDS_BFOS; break;
     case RISCV::PseudoCCNDS_BFOZ: NewOpc = RISCV::NDS_BFOZ; break;
     }
+    // clang-format on
 
     if (NewOpc == RISCV::NDS_BFOZ || NewOpc == RISCV::NDS_BFOS) {
       BuildMI(TrueBB, DL, TII->get(NewOpc), DestReg)

llvm/lib/Target/RISCV/RISCVFeatures.td

@@ -1851,6 +1851,11 @@ def TuneShortForwardBranchOpt
 def HasShortForwardBranchOpt : Predicate<"Subtarget->hasShortForwardBranchOpt()">;
 def NoShortForwardBranchOpt : Predicate<"!Subtarget->hasShortForwardBranchOpt()">;
 
+def TuneShortForwardBranchIMinMax
+    : SubtargetFeature<"short-forward-branch-i-minmax", "HasShortForwardBranchIMinMax",
+                       "true", "Enable short forward branch optimization for min,max instructions in Zbb",
+                       [TuneShortForwardBranchOpt]>;
+
 // Some subtargets require a S2V transfer buffer to move scalars into vectors.
 // FIXME: Forming .vx/.vf/.wx/.wf can reduce register pressure.
 def TuneNoSinkSplatOperands

llvm/lib/Target/RISCV/RISCVInstrInfo.cpp

@@ -1699,6 +1699,10 @@ unsigned getPredicatedOpcode(unsigned Opcode) {
   case RISCV::AND:   return RISCV::PseudoCCAND;
   case RISCV::OR:    return RISCV::PseudoCCOR;
   case RISCV::XOR:   return RISCV::PseudoCCXOR;
+  case RISCV::MAX:   return RISCV::PseudoCCMAX;
+  case RISCV::MAXU:  return RISCV::PseudoCCMAXU;
+  case RISCV::MIN:   return RISCV::PseudoCCMIN;
+  case RISCV::MINU:  return RISCV::PseudoCCMINU;
 
   case RISCV::ADDI:  return RISCV::PseudoCCADDI;
   case RISCV::SLLI:  return RISCV::PseudoCCSLLI;
@@ -1735,14 +1739,21 @@ unsigned getPredicatedOpcode(unsigned Opcode) {
 /// return the defining instruction.
 static MachineInstr *canFoldAsPredicatedOp(Register Reg,
                                            const MachineRegisterInfo &MRI,
-                                           const TargetInstrInfo *TII) {
+                                           const TargetInstrInfo *TII,
+                                           const RISCVSubtarget &STI) {
   if (!Reg.isVirtual())
     return nullptr;
   if (!MRI.hasOneNonDBGUse(Reg))
     return nullptr;
   MachineInstr *MI = MRI.getVRegDef(Reg);
   if (!MI)
     return nullptr;
+
+  if (!STI.hasShortForwardBranchIMinMax() &&
+      (MI->getOpcode() == RISCV::MAX || MI->getOpcode() == RISCV::MIN ||
+       MI->getOpcode() == RISCV::MINU || MI->getOpcode() == RISCV::MAXU))
+    return nullptr;
+
   // Check if MI can be predicated and folded into the CCMOV.
   if (getPredicatedOpcode(MI->getOpcode()) == RISCV::INSTRUCTION_LIST_END)
     return nullptr;
@@ -1806,10 +1817,10 @@ RISCVInstrInfo::optimizeSelect(MachineInstr &MI,
 
   MachineRegisterInfo &MRI = MI.getParent()->getParent()->getRegInfo();
   MachineInstr *DefMI =
-      canFoldAsPredicatedOp(MI.getOperand(5).getReg(), MRI, this);
+      canFoldAsPredicatedOp(MI.getOperand(5).getReg(), MRI, this, STI);
   bool Invert = !DefMI;
   if (!DefMI)
-    DefMI = canFoldAsPredicatedOp(MI.getOperand(4).getReg(), MRI, this);
+    DefMI = canFoldAsPredicatedOp(MI.getOperand(4).getReg(), MRI, this, STI);
   if (!DefMI)
     return nullptr;
 

llvm/lib/Target/RISCV/RISCVInstrInfoSFB.td

@@ -106,6 +106,10 @@ def PseudoCCSRA : SFBALU_rr;
 def PseudoCCAND : SFBALU_rr;
 def PseudoCCOR  : SFBALU_rr;
 def PseudoCCXOR : SFBALU_rr;
+def PseudoCCMAX : SFBALU_rr;
+def PseudoCCMIN : SFBALU_rr;
+def PseudoCCMAXU : SFBALU_rr;
+def PseudoCCMINU : SFBALU_rr;
 
 def PseudoCCADDI : SFBALU_ri;
 def PseudoCCANDI : SFBALU_ri;

llvm/test/CodeGen/RISCV/features-info.ll

@@ -136,6 +136,7 @@
 ; CHECK-NEXT:   shgatpa                          - 'Shgatpa' (SvNNx4 mode supported for all modes supported by satp, as well as Bare).
 ; CHECK-NEXT:   shifted-zextw-fusion             - Enable SLLI+SRLI to be fused when computing (shifted) word zero extension.
 ; CHECK-NEXT:   shlcofideleg                     - 'Shlcofideleg' (Delegating LCOFI Interrupts to VS-mode).
+; CHECK-NEXT:   short-forward-branch-i-minmax    - Enable short forward branch optimization for min,max instructions in Zbb.
 ; CHECK-NEXT:   short-forward-branch-opt         - Enable short forward branch optimization.
 ; CHECK-NEXT:   shtvala                          - 'Shtvala' (htval provides all needed values).
 ; CHECK-NEXT:   shvsatpa                         - 'Shvsatpa' (vsatp supports all modes supported by satp).