[SystemZ] Support conditional indirect sibling calls via BCR.

[• koriakin]

Summary

This adds a conditional variant of CallBR instruction, CallBCR. Also,
it can be fused with integer comparisons, resulting in one of the new
C*BCall instructions.

In addition to CallBRCL limitations, this has another one: it won't trigger
if the function to call isn't already in %r1 - see f22 in the test for
an example (it's also why the loads in tests are volatile).

[uweigand]

It would be nice if the optimization applied in more cases. (On the other hand, we don't want to move too many operations out of the conditional block into the unconditional block either, in case the condition holds true only rarely ...)

In any case, it's good to apply the patch, even if it triggers only in a couple of cases.

Patch LGTM, I'll check it in.

lib/Target/SystemZ/SystemZAsmPrinter.cpp

LoweredMI = MCInstBuilder(SystemZ::BRCL)
       .addImm(MI->getOperand(1).getImm())
       .addExpr(Lower.getExpr(MI->getOperand(2), MCSymbolRefExpr::VK_PLT));
     break;
 
   case SystemZ::CallBR:
     LoweredMI = MCInstBuilder(SystemZ::BR).addReg(SystemZ::R1D);
     break;
 
+  case SystemZ::CallBCR:
+    LoweredMI = MCInstBuilder(SystemZ::BCR)
+      .addImm(MI->getOperand(0).getImm())
+      .addImm(MI->getOperand(1).getImm())
+      .addReg(SystemZ::R1D);
+    break;
+
+  case SystemZ::CRBCall:
+    LoweredMI = MCInstBuilder(SystemZ::CRB)
+      .addReg(MI->getOperand(0).getReg())
+      .addReg(MI->getOperand(1).getReg())
+      .addImm(MI->getOperand(2).getImm())
+      .addReg(SystemZ::R1D)
+      .addImm(0);
+    break;
+
+  case SystemZ::CGRBCall:
+    LoweredMI = MCInstBuilder(SystemZ::CGRB)
+      .addReg(MI->getOperand(0).getReg())
+      .addReg(MI->getOperand(1).getReg())
+      .addImm(MI->getOperand(2).getImm())
+      .addReg(SystemZ::R1D)
+      .addImm(0);
+    break;
+
+  case SystemZ::CIBCall:
+    LoweredMI = MCInstBuilder(SystemZ::CIB)
+      .addReg(MI->getOperand(0).getReg())
+      .addImm(MI->getOperand(1).getImm())
+      .addImm(MI->getOperand(2).getImm())
+      .addReg(SystemZ::R1D)
+      .addImm(0);
+    break;
+
+  case SystemZ::CGIBCall:
+    LoweredMI = MCInstBuilder(SystemZ::CGIB)
+      .addReg(MI->getOperand(0).getReg())
+      .addImm(MI->getOperand(1).getImm())
+      .addImm(MI->getOperand(2).getImm())
+      .addReg(SystemZ::R1D)
+      .addImm(0);
+    break;
+
+  case SystemZ::CLRBCall:
+    LoweredMI = MCInstBuilder(SystemZ::CLRB)
+      .addReg(MI->getOperand(0).getReg())
+      .addReg(MI->getOperand(1).getReg())
+      .addImm(MI->getOperand(2).getImm())
+      .addReg(SystemZ::R1D)
+      .addImm(0);
+    break;
+
+  case SystemZ::CLGRBCall:
+    LoweredMI = MCInstBuilder(SystemZ::CLGRB)
+      .addReg(MI->getOperand(0).getReg())
+      .addReg(MI->getOperand(1).getReg())
+      .addImm(MI->getOperand(2).getImm())
+      .addReg(SystemZ::R1D)
+      .addImm(0);
+    break;
+
+  case SystemZ::CLIBCall:
+    LoweredMI = MCInstBuilder(SystemZ::CLIB)
+      .addReg(MI->getOperand(0).getReg())
+      .addImm(MI->getOperand(1).getImm())
+      .addImm(MI->getOperand(2).getImm())
+      .addReg(SystemZ::R1D)
+      .addImm(0);
+    break;
+
+  case SystemZ::CLGIBCall:
+    LoweredMI = MCInstBuilder(SystemZ::CLGIB)
+      .addReg(MI->getOperand(0).getReg())
+      .addImm(MI->getOperand(1).getImm())
+      .addImm(MI->getOperand(2).getImm())
+      .addReg(SystemZ::R1D)
+      .addImm(0);
+    break;
+
   case SystemZ::TLS_GDCALL:
     LoweredMI = MCInstBuilder(SystemZ::BRASL)
       .addReg(SystemZ::R14D)
       .addExpr(getTLSGetOffset(MF->getContext()))
       .addExpr(Lower.getExpr(MI->getOperand(0), MCSymbolRefExpr::VK_TLSGD));
     break;
 
   case SystemZ::TLS_LDCALL:

lib/Target/SystemZ/SystemZElimCompare.cpp

fuseCompareAndBranch(MachineInstr *Compare,
   SystemZII::CompareAndBranchType Type;
   switch (Branch->getOpcode()) {
   case SystemZ::BRC:
     Type = SystemZII::CompareAndBranch;
     break;
   case SystemZ::CondReturn:
     Type = SystemZII::CompareAndReturn;
     break;
+  case SystemZ::CallBCR:
+    Type = SystemZII::CompareAndSibcall;
+    break;
   default:
     return false;
   }
 
   // See whether we have a comparison that can be fused.
   unsigned FusedOpcode = TII->getCompareAndBranch(Compare->getOpcode(),
                                                   Type, Compare);
   if (!FusedOpcode)
     return false;
 
   // Make sure that the operands are available at the branch.
   unsigned SrcReg = Compare->getOperand(0).getReg();
   unsigned SrcReg2 = (Compare->getOperand(1).isReg() ?
                       Compare->getOperand(1).getReg() : 0);
   MachineBasicBlock::iterator MBBI = Compare, MBBE = Branch;
   for (++MBBI; MBBI != MBBE; ++MBBI)
     if (MBBI->modifiesRegister(SrcReg, TRI) ||
         (SrcReg2 && MBBI->modifiesRegister(SrcReg2, TRI)))
       return false;
 
-  // Read the branch mask and target (if applicable).
+  // Read the branch mask, target (if applicable), regmask (if applicable).
   MachineOperand CCMask(MBBI->getOperand(1));
   assert((CCMask.getImm() & ~SystemZ::CCMASK_ICMP) == 0 &&
          "Invalid condition-code mask for integer comparison");
   // This is only valid for CompareAndBranch.
   MachineOperand Target(MBBI->getOperand(
     Type == SystemZII::CompareAndBranch ? 2 : 0));
+  const uint32_t *RegMask;
+  if (Type == SystemZII::CompareAndSibcall)
+    RegMask = MBBI->getOperand(2).getRegMask();
 
   // Clear out all current operands.
   int CCUse = MBBI->findRegisterUseOperandIdx(SystemZ::CC, false, TRI);
   assert(CCUse >= 0 && "BRC/BCR must use CC");
   Branch->RemoveOperand(CCUse);
-  if (Type == SystemZII::CompareAndBranch)
+  // Remove target (branch) or regmask (sibcall).
+  if (Type == SystemZII::CompareAndBranch ||
+      Type == SystemZII::CompareAndSibcall)
     Branch->RemoveOperand(2);
   Branch->RemoveOperand(1);
   Branch->RemoveOperand(0);
 
   // Rebuild Branch as a fused compare and branch.
   Branch->setDesc(TII->get(FusedOpcode));
   MachineInstrBuilder MIB(*Branch->getParent()->getParent(), Branch);
   MIB.addOperand(Compare->getOperand(0))
      .addOperand(Compare->getOperand(1))
      .addOperand(CCMask);
 
   if (Type == SystemZII::CompareAndBranch) {
     // Only conditional branches define CC, as they may be converted back
     // to a non-fused branch because of a long displacement.  Conditional
     // returns don't have that problem.
     MIB.addOperand(Target)
        .addReg(SystemZ::CC, RegState::ImplicitDefine);
   }
 
+  if (Type == SystemZII::CompareAndSibcall)
+    MIB.addRegMask(RegMask);
+
   // Clear any intervening kills of SrcReg and SrcReg2.
   MBBI = Compare;
   for (++MBBI; MBBI != MBBE; ++MBBI) {
     MBBI->clearRegisterKills(SrcReg, TRI);
     if (SrcReg2)
       MBBI->clearRegisterKills(SrcReg2, TRI);
   }
   FusedComparisons += 1;

lib/Target/SystemZ/SystemZInstrInfo.h

 // Kinds of branch in compare-and-branch instructions.  Together with type
 // of the converted compare, this identifies the compare-and-branch
 // instruction.
 enum CompareAndBranchType {
   // Relative branch - CRJ etc.
   CompareAndBranch,
 
   // Indirect branch, used for return - CRBReturn etc.
-  CompareAndReturn
+  CompareAndReturn,
+
+  // Indirect branch, used for sibcall - CRBCall etc.
+  CompareAndSibcall
 };
 } // end namespace SystemZII
 
 class SystemZSubtarget;
 class SystemZInstrInfo : public SystemZGenInstrInfo {
   const SystemZRegisterInfo RI;
   SystemZSubtarget &STI;
 

lib/Target/SystemZ/SystemZInstrInfo.cpp

static unsigned getConditionalMove(unsigned Opcode) {
   }
 }
 
 bool SystemZInstrInfo::isPredicable(MachineInstr &MI) const {
   unsigned Opcode = MI.getOpcode();
   if (STI.hasLoadStoreOnCond() && getConditionalMove(Opcode))
     return true;
   if (Opcode == SystemZ::Return ||
-      Opcode == SystemZ::CallJG)
+      Opcode == SystemZ::CallJG ||
+      Opcode == SystemZ::CallBR)
     return true;
   return false;
 }
 
 bool SystemZInstrInfo::
 isProfitableToIfCvt(MachineBasicBlock &MBB,
                     unsigned NumCycles, unsigned ExtraPredCycles,
                     BranchProbability Probability) const {
if (Opcode == SystemZ::CallJG) {
     MI.setDesc(get(SystemZ::CallBRCL));
     MachineInstrBuilder(*MI.getParent()->getParent(), MI)
       .addImm(CCValid).addImm(CCMask)
       .addGlobalAddress(Global)
       .addRegMask(RegMask)
       .addReg(SystemZ::CC, RegState::Implicit);
     return true;
   }
+  if (Opcode == SystemZ::CallBR) {
+    const uint32_t *RegMask = MI.getOperand(0).getRegMask();
+    MI.RemoveOperand(0);
+    MI.setDesc(get(SystemZ::CallBCR));
+    MachineInstrBuilder(*MI.getParent()->getParent(), MI)
+      .addImm(CCValid).addImm(CCMask)
+      .addRegMask(RegMask)
+      .addReg(SystemZ::CC, RegState::Implicit);
+    return true;
+  }
   return false;
 }
 
 void SystemZInstrInfo::copyPhysReg(MachineBasicBlock &MBB,
                                    MachineBasicBlock::iterator MBBI,
                                    DebugLoc DL, unsigned DestReg,
                                    unsigned SrcReg, bool KillSrc) const {
   // Split 128-bit GPR moves into two 64-bit moves.  This handles ADDR128 too.
case SystemZ::CLGR:
       return SystemZ::CLGRBReturn;
     case SystemZ::CLFI:
       return SystemZ::CLIBReturn;
     case SystemZ::CLGFI:
       return SystemZ::CLGIBReturn;
     default:
       return 0;
     }
+  case SystemZII::CompareAndSibcall:
+    switch (Opcode) {
+    case SystemZ::CR:
+      return SystemZ::CRBCall;
+    case SystemZ::CGR:
+      return SystemZ::CGRBCall;
+    case SystemZ::CHI:
+      return SystemZ::CIBCall;
+    case SystemZ::CGHI:
+      return SystemZ::CGIBCall;
+    case SystemZ::CLR:
+      return SystemZ::CLRBCall;
+    case SystemZ::CLGR:
+      return SystemZ::CLGRBCall;
+    case SystemZ::CLFI:
+      return SystemZ::CLIBCall;
+    case SystemZ::CLGFI:
+      return SystemZ::CLGIBCall;
+    default:
+      return 0;
+    }
   }
   return 0;
 }
 
 void SystemZInstrInfo::loadImmediate(MachineBasicBlock &MBB,
                                      MachineBasicBlock::iterator MBBI,
                                      unsigned Reg, uint64_t Value) const {
   DebugLoc DL = MBBI != MBB.end() ? MBBI->getDebugLoc() : DebugLoc();

lib/Target/SystemZ/SystemZInstrInfo.td

 // are argument registers and since branching to R0 is a no-op.
 let isCall = 1, isTerminator = 1, isReturn = 1, isBarrier = 1 in {
   def CallJG : Alias<6, (outs), (ins pcrel32:$I2),
                      [(z_sibcall pcrel32:$I2)]>;
   let Uses = [R1D] in
     def CallBR : Alias<2, (outs), (ins), [(z_sibcall R1D)]>;
 }
 
-let CCMaskFirst = 1, isCall = 1, isReturn = 1 in
+let CCMaskFirst = 1, isCall = 1, isTerminator = 1, isReturn = 1 in {
   def CallBRCL : Alias<6, (outs), (ins cond4:$valid, cond4:$R1,
                                    pcrel32:$I2), []>;
 
+  let Uses = [R1D] in
+    def CallBCR : Alias<2, (outs), (ins cond4:$valid, cond4:$R1), []>;
+}
+
+// Fused compare and conditional sibling calls.
+let isCall = 1, isTerminator = 1, isReturn = 1, Uses = [R1D] in {
+  def CRBCall : Alias<6, (outs), (ins GR32:$R1, GR32:$R2, cond4:$M3), []>;
+  def CGRBCall : Alias<6, (outs), (ins GR64:$R1, GR64:$R2, cond4:$M3), []>;
+  def CIBCall : Alias<6, (outs), (ins GR32:$R1, imm32sx8:$I2, cond4:$M3), []>;
+  def CGIBCall : Alias<6, (outs), (ins GR64:$R1, imm64sx8:$I2, cond4:$M3), []>;
+  def CLRBCall : Alias<6, (outs), (ins GR32:$R1, GR32:$R2, cond4:$M3), []>;
+  def CLGRBCall : Alias<6, (outs), (ins GR64:$R1, GR64:$R2, cond4:$M3), []>;
+  def CLIBCall : Alias<6, (outs), (ins GR32:$R1, imm32zx8:$I2, cond4:$M3), []>;
+  def CLGIBCall : Alias<6, (outs), (ins GR64:$R1, imm64zx8:$I2, cond4:$M3), []>;
+}
+
 // TLS calls.  These will be lowered into a call to __tls_get_offset,
 // with an extra relocation specifying the TLS symbol.
 let isCall = 1, Defs = [R14D, CC] in {
   def TLS_GDCALL : Alias<6, (outs), (ins tlssym:$I2, variable_ops),
                          [(z_tls_gdcall tglobaltlsaddr:$I2)]>;
   def TLS_LDCALL : Alias<6, (outs), (ins tlssym:$I2, variable_ops),
                          [(z_tls_ldcall tglobaltlsaddr:$I2)]>;
 }

test/CodeGen/SystemZ/call-05.ll

+; Test conditional sibling calls.
+;
+; RUN: llc < %s -mtriple=s390x-linux-gnu | FileCheck %s
+
+
+@var = global i32 1;
+@fun_a = global void()* null;
+@fun_b = global void()* null;
+@fun_c = global void(i32)* null;
+
+; Check a conditional sibling call.
+define void @f1(i32 %val1, i32 %val2) {
+; CHECK-LABEL: f1:
+; CHECK: crbl %r2, %r3, 0(%r1)
+; CHECK: br %r14
+  %fun_a = load volatile void() *, void()** @fun_a;
+  %cond = icmp slt i32 %val1, %val2;
+  br i1 %cond, label %a, label %b;
+
+a:
+  tail call void %fun_a()
+  ret void
+
+b:
+  store i32 1, i32 *@var;
+  ret void
+}
+
+; Check a conditional sibling call when there are two possibilities.
+define void @f2(i32 %val1, i32 %val2) {
+; CHECK-LABEL: f2:
+; CHECK: crbl %r2, %r3, 0(%r1)
+; CHECK: br %r1
+  %fun_a = load volatile void() *, void()** @fun_a;
+  %fun_b = load volatile void() *, void()** @fun_b;
+  %cond = icmp slt i32 %val1, %val2;
+  br i1 %cond, label %a, label %b;
+
+a:
+  tail call void %fun_a()
+  ret void
+
+b:
+  tail call void %fun_b()
+  ret void
+}
+
+; Check a conditional sibling call with an argument - not supported.
+define void @f3(i32 %val1, i32 %val2) {
+; CHECK-LABEL: f3:
+; CHECK: crjhe %r2, %r3
+; CHECK: br %r1
+; CHECK: br %r14
+  %fun_c = load volatile void(i32) *, void(i32)** @fun_c;
+  %cond = icmp slt i32 %val1, %val2;
+  br i1 %cond, label %a, label %b;
+
+a:
+  tail call void %fun_c(i32 1)
+  ret void
+
+b:
+  store i32 1, i32 *@var;
+  ret void
+}
+
+; Check a conditional sibling call - unsigned compare.
+define void @f4(i32 %val1, i32 %val2) {
+; CHECK-LABEL: f4:
+; CHECK: clrbl %r2, %r3, 0(%r1)
+; CHECK: br %r14
+  %fun_a = load volatile void() *, void()** @fun_a;
+  %cond = icmp ult i32 %val1, %val2;
+  br i1 %cond, label %a, label %b;
+
+a:
+  tail call void %fun_a()
+  ret void
+
+b:
+  store i32 1, i32 *@var;
+  ret void
+}
+
+; Check a conditional sibling call - 64-bit compare.
+define void @f5(i64 %val1, i64 %val2) {
+; CHECK-LABEL: f5:
+; CHECK: cgrbl %r2, %r3, 0(%r1)
+; CHECK: br %r14
+  %fun_a = load volatile void() *, void()** @fun_a;
+  %cond = icmp slt i64 %val1, %val2;
+  br i1 %cond, label %a, label %b;
+
+a:
+  tail call void %fun_a()
+  ret void
+
+b:
+  store i32 1, i32 *@var;
+  ret void
+}
+
+; Check a conditional sibling call - unsigned 64-bit compare.
+define void @f6(i64 %val1, i64 %val2) {
+; CHECK-LABEL: f6:
+; CHECK: clgrbl %r2, %r3, 0(%r1)
+; CHECK: br %r14
+  %fun_a = load volatile void() *, void()** @fun_a;
+  %cond = icmp ult i64 %val1, %val2;
+  br i1 %cond, label %a, label %b;
+
+a:
+  tail call void %fun_a()
+  ret void
+
+b:
+  store i32 1, i32 *@var;
+  ret void
+}
+
+; Check a conditional sibling call - less-equal compare.
+define void @f7(i32 %val1, i32 %val2) {
+; CHECK-LABEL: f7:
+; CHECK: crble %r2, %r3, 0(%r1)
+; CHECK: br %r14
+  %fun_a = load volatile void() *, void()** @fun_a;
+  %cond = icmp sle i32 %val1, %val2;
+  br i1 %cond, label %a, label %b;
+
+a:
+  tail call void %fun_a()
+  ret void
+
+b:
+  store i32 1, i32 *@var;
+  ret void
+}
+
+; Check a conditional sibling call - high compare.
+define void @f8(i32 %val1, i32 %val2) {
+; CHECK-LABEL: f8:
+; CHECK: crbh %r2, %r3, 0(%r1)
+; CHECK: br %r14
+  %fun_a = load volatile void() *, void()** @fun_a;
+  %cond = icmp sgt i32 %val1, %val2;
+  br i1 %cond, label %a, label %b;
+
+a:
+  tail call void %fun_a()
+  ret void
+
+b:
+  store i32 1, i32 *@var;
+  ret void
+}
+
+; Check a conditional sibling call - high-equal compare.
+define void @f9(i32 %val1, i32 %val2) {
+; CHECK-LABEL: f9:
+; CHECK: crbhe %r2, %r3, 0(%r1)
+; CHECK: br %r14
+  %fun_a = load volatile void() *, void()** @fun_a;
+  %cond = icmp sge i32 %val1, %val2;
+  br i1 %cond, label %a, label %b;
+
+a:
+  tail call void %fun_a()
+  ret void
+
+b:
+  store i32 1, i32 *@var;
+  ret void
+}
+
+; Check a conditional sibling call - equal compare.
+define void @f10(i32 %val1, i32 %val2) {
+; CHECK-LABEL: f10:
+; CHECK: crbe %r2, %r3, 0(%r1)
+; CHECK: br %r14
+  %fun_a = load volatile void() *, void()** @fun_a;
+  %cond = icmp eq i32 %val1, %val2;
+  br i1 %cond, label %a, label %b;
+
+a:
+  tail call void %fun_a()
+  ret void
+
+b:
+  store i32 1, i32 *@var;
+  ret void
+}
+
+; Check a conditional sibling call - unequal compare.
+define void @f11(i32 %val1, i32 %val2) {
+; CHECK-LABEL: f11:
+; CHECK: crblh %r2, %r3, 0(%r1)
+; CHECK: br %r14
+  %fun_a = load volatile void() *, void()** @fun_a;
+  %cond = icmp ne i32 %val1, %val2;
+  br i1 %cond, label %a, label %b;
+
+a:
+  tail call void %fun_a()
+  ret void
+
+b:
+  store i32 1, i32 *@var;
+  ret void
+}
+
+; Check a conditional sibling call - immediate slt.
+define void @f12(i32 %val1) {
+; CHECK-LABEL: f12:
+; CHECK: cible %r2, 4, 0(%r1)
+; CHECK: br %r14
+  %fun_a = load volatile void() *, void()** @fun_a;
+  %cond = icmp slt i32 %val1, 5;
+  br i1 %cond, label %a, label %b;
+
+a:
+  tail call void %fun_a()
+  ret void
+
+b:
+  store i32 1, i32 *@var;
+  ret void
+}
+
+; Check a conditional sibling call - immediate sle.
+define void @f13(i32 %val1) {
+; CHECK-LABEL: f13:
+; CHECK: cible %r2, 5, 0(%r1)
+; CHECK: br %r14
+  %fun_a = load volatile void() *, void()** @fun_a;
+  %cond = icmp sle i32 %val1, 5;
+  br i1 %cond, label %a, label %b;
+
+a:
+  tail call void %fun_a()
+  ret void
+
+b:
+  store i32 1, i32 *@var;
+  ret void
+}
+
+; Check a conditional sibling call - immediate sgt.
+define void @f14(i32 %val1) {
+; CHECK-LABEL: f14:
+; CHECK: cibhe %r2, 6, 0(%r1)
+; CHECK: br %r14
+  %fun_a = load volatile void() *, void()** @fun_a;
+  %cond = icmp sgt i32 %val1, 5;
+  br i1 %cond, label %a, label %b;
+
+a:
+  tail call void %fun_a()
+  ret void
+
+b:
+  store i32 1, i32 *@var;
+  ret void
+}
+
+; Check a conditional sibling call - immediate sge.
+define void @f15(i32 %val1) {
+; CHECK-LABEL: f15:
+; CHECK: cibhe %r2, 5, 0(%r1)
+; CHECK: br %r14
+  %fun_a = load volatile void() *, void()** @fun_a;
+  %cond = icmp sge i32 %val1, 5;
+  br i1 %cond, label %a, label %b;
+
+a:
+  tail call void %fun_a()
+  ret void
+
+b:
+  store i32 1, i32 *@var;
+  ret void
+}
+
+; Check a conditional sibling call - immediate eq.
+define void @f16(i32 %val1) {
+; CHECK-LABEL: f16:
+; CHECK: cibe %r2, 5, 0(%r1)
+; CHECK: br %r14
+  %fun_a = load volatile void() *, void()** @fun_a;
+  %cond = icmp eq i32 %val1, 5;
+  br i1 %cond, label %a, label %b;
+
+a:
+  tail call void %fun_a()
+  ret void
+
+b:
+  store i32 1, i32 *@var;
+  ret void
+}
+
+; Check a conditional sibling call - immediate ne.
+define void @f17(i32 %val1) {
+; CHECK-LABEL: f17:
+; CHECK: ciblh %r2, 5, 0(%r1)
+; CHECK: br %r14
+  %fun_a = load volatile void() *, void()** @fun_a;
+  %cond = icmp ne i32 %val1, 5;
+  br i1 %cond, label %a, label %b;
+
+a:
+  tail call void %fun_a()
+  ret void
+
+b:
+  store i32 1, i32 *@var;
+  ret void
+}
+
+; Check a conditional sibling call - immediate ult.
+define void @f18(i32 %val1) {
+; CHECK-LABEL: f18:
+; CHECK: clible %r2, 4, 0(%r1)
+; CHECK: br %r14
+  %fun_a = load volatile void() *, void()** @fun_a;
+  %cond = icmp ult i32 %val1, 5;
+  br i1 %cond, label %a, label %b;
+
+a:
+  tail call void %fun_a()
+  ret void
+
+b:
+  store i32 1, i32 *@var;
+  ret void
+}
+
+; Check a conditional sibling call - immediate 64-bit slt.
+define void @f19(i64 %val1) {
+; CHECK-LABEL: f19:
+; CHECK: cgible %r2, 4, 0(%r1)
+; CHECK: br %r14
+  %fun_a = load volatile void() *, void()** @fun_a;
+  %cond = icmp slt i64 %val1, 5;
+  br i1 %cond, label %a, label %b;
+
+a:
+  tail call void %fun_a()
+  ret void
+
+b:
+  store i32 1, i32 *@var;
+  ret void
+}
+
+; Check a conditional sibling call - immediate 64-bit ult.
+define void @f20(i64 %val1) {
+; CHECK-LABEL: f20:
+; CHECK: clgible %r2, 4, 0(%r1)
+; CHECK: br %r14
+  %fun_a = load volatile void() *, void()** @fun_a;
+  %cond = icmp ult i64 %val1, 5;
+  br i1 %cond, label %a, label %b;
+
+a:
+  tail call void %fun_a()
+  ret void
+
+b:
+  store i32 1, i32 *@var;
+  ret void
+}
+
+; Check a conditional sibling call to an argument - will fail due to
+; intervening lgr.
+define void @f21(i32 %val1, i32 %val2, void()* %fun) {
+; CHECK-LABEL: f21:
+; CHECK: crjhe %r2, %r3
+; CHECK: lgr %r1, %r4
+; CHECK: br %r1
+; CHECK: br %r14
+  %cond = icmp slt i32 %val1, %val2;
+  br i1 %cond, label %a, label %b;
+
+a:
+  tail call void %fun()
+  ret void
+
+b:
+  store i32 1, i32 *@var;
+  ret void
+}
+
+; Check a conditional sibling call - float olt compare.
+define void @f22(float %val1, float %val2) {
+; CHECK-LABEL: f22:
+; CHECK: cebr %f0, %f2
+; CHECK: blr %r1
+; CHECK: br %r14
+  %fun_a = load volatile void() *, void()** @fun_a;
+  %cond = fcmp olt float %val1, %val2;
+  br i1 %cond, label %a, label %b;
+
+a:
+  tail call void %fun_a()
+  ret void
+
+b:
+  store i32 1, i32 *@var;
+  ret void
+}
+
+; Check a conditional sibling call - float ult compare.
+define void @f23(float %val1, float %val2) {
+; CHECK-LABEL: f23:
+; CHECK: cebr %f0, %f2
+; CHECK: bnher %r1
+; CHECK: br %r14
+  %fun_a = load volatile void() *, void()** @fun_a;
+  %cond = fcmp ult float %val1, %val2;
+  br i1 %cond, label %a, label %b;
+
+a:
+  tail call void %fun_a()
+  ret void
+
+b:
+  store i32 1, i32 *@var;
+  ret void
+}
+
+; Check a conditional sibling call - float ord compare.
+define void @f24(float %val1, float %val2) {
+; CHECK-LABEL: f24:
+; CHECK: cebr %f0, %f2
+; CHECK: bnor %r1
+; CHECK: br %r14
+  %fun_a = load volatile void() *, void()** @fun_a;
+  %cond = fcmp ord float %val1, %val2;
+  br i1 %cond, label %a, label %b;
+
+a:
+  tail call void %fun_a()
+  ret void
+
+b:
+  store i32 1, i32 *@var;
+  ret void
+}
+
+; Check a conditional sibling call - float uno compare.
+define void @f25(float %val1, float %val2) {
+; CHECK-LABEL: f25:
+; CHECK: cebr %f0, %f2
+; CHECK: bor %r1
+; CHECK: br %r14
+  %fun_a = load volatile void() *, void()** @fun_a;
+  %cond = fcmp uno float %val1, %val2;
+  br i1 %cond, label %a, label %b;
+
+a:
+  tail call void %fun_a()
+  ret void
+
+b:
+  store i32 1, i32 *@var;
+  ret void
+}