[Mlir-commits] [mlir] [mlir] AMDGPUToROCDL: lower `amdgpu.swizzle_bitmode` (PR #136223)

[Ivan Butygin]

================
@@ -1377,6 +1377,38 @@ struct AMDGPUDPPLowering : public ConvertOpToLLVMPattern<DPPOp> {
   }
 };
 
+struct AMDGPUSwizzleBitModeLowering
+    : public ConvertOpToLLVMPattern<SwizzleBitModeOp> {
----------------
Hardcode84 wrote:

QDMode and BitMode variants have a different offsets format, my idea was to have them as separate ops in `AMDGPU`, so user won't need to bother with offset bitpacking even if they are both lowered to `ds_swizzle` eventually.

https://github.com/llvm/llvm-project/pull/136223