[NVPTX] Add builtin for 'exit' handling

Summary:
The PTX ISA has always supported the 'exit' instruction to terminate
individual threads. This patch adds a builtin to handle it. See the PTX
documentation for further details.
https://docs.nvidia.com/cuda/parallel-thread-execution/index.html#control-flow-instructions-exit

Author: jhuber6

clang/include/clang/Basic/BuiltinsNVPTX.def

@@ -155,6 +155,7 @@ BUILTIN(__nvvm_read_ptx_sreg_pm3, "i", "n")
 // MISC
 
 BUILTIN(__nvvm_prmt, "UiUiUiUi", "")
+BUILTIN(__nvvm_exit, "v", "r")
 
 // Min Max
 

clang/test/CodeGen/builtins-nvptx.c

@@ -165,6 +165,14 @@ __device__ void sync() {
 
 }
 
+__device__ void exit() {
+
+// CHECK: call void @llvm.nvvm.exit()
+
+  __nvvm_exit();
+
+}
+
 
 // NVVM intrinsics
 

llvm/include/llvm/IR/IntrinsicsNVVM.td

@@ -4801,4 +4801,8 @@ def int_nvvm_setmaxnreg_dec_sync_aligned_u32
               [IntrConvergent, IntrNoMem, IntrHasSideEffects, ImmArg<ArgIndex<0>>],
               "llvm.nvvm.setmaxnreg.dec.sync.aligned.u32">;
 
+// Exit
+def int_nvvm_exit : ClangBuiltin<"__nvvm_exit">,
+    Intrinsic<[], [], [IntrConvergent, IntrInaccessibleMemOnly, IntrNoReturn]>;
+
 } // let TargetPrefix = "nvvm"

llvm/lib/Target/NVPTX/NVPTXIntrinsics.td

@@ -6832,4 +6832,7 @@ multiclass SET_MAXNREG<string Action, Intrinsic Intr> {
 
 defm INT_SET_MAXNREG_INC : SET_MAXNREG<"inc", int_nvvm_setmaxnreg_inc_sync_aligned_u32>;
 defm INT_SET_MAXNREG_DEC : SET_MAXNREG<"dec", int_nvvm_setmaxnreg_dec_sync_aligned_u32>;
+
 } // isConvergent
+
+def INT_EXIT : NVPTXInst<(outs), (ins), "exit;", [(int_nvvm_exit)]>;

llvm/test/CodeGen/NVPTX/intrinsics.ll

@@ -133,6 +133,13 @@ define i64 @test_clock64() {
   ret i64 %ret
 }
 
+; CHECK-LABEL: test_exit
+define void @test_exit() {
+; CHECK: exit;
+  call void @llvm.nvvm.exit()
+  ret void
+}
+
 declare float @llvm.fabs.f32(float)
 declare double @llvm.fabs.f64(double)
 declare float @llvm.nvvm.sqrt.f(float)
@@ -146,3 +153,4 @@ declare i64 @llvm.ctpop.i64(i64)
 declare i32 @llvm.nvvm.read.ptx.sreg.tid.x()
 declare i32 @llvm.nvvm.read.ptx.sreg.clock()
 declare i64 @llvm.nvvm.read.ptx.sreg.clock64()
+declare void @llvm.nvvm.exit()