[llvm] [IR] Add CallBr intrinsics support (PR #133907)

[Matt Arsenault via llvm-commits]

================
@@ -3009,8 +3009,64 @@ bool IRTranslator::translateInvoke(const User &U,
 
 bool IRTranslator::translateCallBr(const User &U,
                                    MachineIRBuilder &MIRBuilder) {
-  // FIXME: Implement this.
-  return false;
+  const CallBrInst &I = cast<CallBrInst>(U);
+  MachineBasicBlock *CallBrMBB = &MIRBuilder.getMBB();
+
+  // TODO: operand bundles (see SelDAG implementation of callbr)?
+  assert(!I.hasOperandBundles() &&
+         "Cannot lower callbrs with operand bundles yet");
+
+  if (I.isInlineAsm()) {
+    // FIXME: inline asm not yet supported
+    if (!translateInlineAsm(I, MIRBuilder))
+      return false;
+  } else if (I.getIntrinsicID() != Intrinsic::not_intrinsic) {
+    switch (I.getIntrinsicID()) {
+    default:
+      report_fatal_error("Unsupported intrinsic for callbr");
+    case Intrinsic::amdgcn_kill:
+      if (I.getNumIndirectDests() != 1)
+        report_fatal_error(
+            "amdgcn.kill supportes exactly one indirect destination");
+      CallInst *CI =
+          CallInst::Create(I.getFunctionType(), I.getCalledFunction(),
+                           SmallVector<Value *, 1>(I.args()));
+      bool Success = translateCall(*CI, MIRBuilder);
+      CI->deleteValue();
----------------
arsenm wrote:

The IRTRanslator should not be hacking on the IR. You don't need or want to delete this 

https://github.com/llvm/llvm-project/pull/133907