[Mlir-commits] [mlir] [uArch][XeGPU] Add XeGPU uArch definition. (PR #153706)

[Adam Siemieniuk]

================
@@ -35,6 +36,14 @@ void XeGPUDialect::initialize() {
 #define GET_ATTRDEF_LIST
 #include <mlir/Dialect/XeGPU/IR/XeGPUAttrs.cpp.inc>
       >();
+
+  // Populate the uArchMap with the supported target devices
----------------
adam-smnk wrote:

Right now, uArch info is not strictly necessary for the dialect i.e., you can still do a lot like transforming ops and running generic transforms without pulling in the hardware info. It makes me "pay the price" for sth I don't need.
I can see that moving forward this might change as a lot of passes might want to use it. But that's a bit premature decision.

Also, AFAIK the whole runtime initialization is the reason why you had to introduce mutexes into table builders. It adds extra complexity both design and runtime without a clear payoff.

https://github.com/llvm/llvm-project/pull/153706