[clang] [CIR][CUDA] Support built-in CUDA texture type (PR #214781)

via cfe-commits cfe-commits at lists.llvm.org
Sun Aug 16 21:58:34 PDT 2026


================
@@ -77,6 +77,13 @@ class NVPTXTargetCIRGenInfo : public TargetCIRGenInfo {
     return cir::IntType::get(&getABIInfo().cgt.getMLIRContext(), 64,
                              /*isSigned=*/true);
   }
+
+  mlir::Type getCUDADeviceBuiltinTextureDeviceType() const override {
+    // On the device side, texture reference is represented as an object handle
+    // in 64-bit integer.
+    return cir::IntType::get(&getABIInfo().cgt.getMLIRContext(), 64,
----------------
Siya-05 wrote:

Thanks! I've updated this to preserve the texture semantics at the CIR level with a dedicated `!cir.cuda_texture` type, and lower it to i64 during CIR-to-LLVM lowering.

https://github.com/llvm/llvm-project/pull/214781


More information about the cfe-commits mailing list