[clang] [CIR] Fix cir.call_llvm_intrinsic lowering for 0-result ops (PR #199516)

[Andy Kaylor via cfe-commits]

================
@@ -0,0 +1,27 @@
+// RUN: cir-opt %s -cir-to-llvm -o - | FileCheck %s
+
+// Contract test for CIRToLLVMLLVMIntrinsicCallOpLowering: the op def
+// declares Optional<CIR_AnyType>:$result, so the lowering must accept
+// 0-result (void) calls in addition to the single-result case.
+
+!s32i = !cir.int<s, 32>
+
+module {
+  // 0-result, 0-operand.
+  // CHECK-LABEL: llvm.func @void_no_operands
+  // CHECK:         llvm.call_intrinsic "llvm.donothing"() : () -> ()
+  // CHECK:         llvm.return
+  cir.func @void_no_operands() {
+    cir.call_llvm_intrinsic "donothing" : () -> ()
----------------
andykaylor wrote:

How did you get into this situation? We have existing builtin handlers that call LLVM intrinsics with a void return type, and they're implemented using an explicit void type. For example:

```
void test_lfence(void) {
  __builtin_ia32_lfence();
}
```
Produces this CIR:
```
  !void = !cir.void
  cir.func @test_lfence() {
    %0 = cir.call_llvm_intrinsic "x86.sse2.lfence"  : () -> !void
    cir.return loc(#loc2)
  }
```
Which lowers to this LLVM IR:
```
define dso_local void @test_lfence() {
  call void @llvm.x86.sse2.lfence()
  ret void
}
```

https://github.com/llvm/llvm-project/pull/199516