[clang] [clang][AMDGPU] Widen ballot for read_exec_lo/hi to wavefront size (PR #212813)

[Matt Arsenault via cfe-commits]

================
@@ -275,13 +275,22 @@ Value *EmitAMDGPUGridSize(CodeGenFunction &CGF, unsigned Index) {
 
 // Generates the IR for __builtin_read_exec_*.
 // Lowers the builtin to amdgcn_ballot intrinsic.
+//
+// The ballot must be taken at the wavefront width: a ballot narrower than the
+// wave size cannot represent one bit per lane and fails to select. Request the
+// mask at the wave width and narrow it afterwards for the _lo and _hi halves.
 static Value *EmitAMDGCNBallotForExec(CodeGenFunction &CGF, const CallExpr *E,
                                       llvm::Type *RegisterType,
                                       llvm::Type *ValueType, bool isExecHi) {
   CodeGen::CGBuilderTy &Builder = CGF.Builder;
   CodeGen::CodeGenModule &CGM = CGF.CGM;
 
-  Function *F = CGM.getIntrinsic(Intrinsic::amdgcn_ballot, {RegisterType});
+  unsigned WaveSize = CGF.getTarget().getGridValue().GV_Warp_Size;
+  llvm::Type *BallotType = RegisterType;
+  if (BallotType->getIntegerBitWidth() < WaveSize)
----------------
arsenm wrote:

This shouldn't need to figure out the type, it should directly take the type 

https://github.com/llvm/llvm-project/pull/212813