================
@@ -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)
----------------
aobolensk wrote:
Simplified for trunc operation
https://github.com/llvm/llvm-project/pull/212813
_______________________________________________
cfe-commits mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
