[Mlir-commits] [mlir] [mlir][spirv] Add GpuToLLVM cconv suited to Vulkan, migrate last tests (PR #123384)

[Andrea Faulds]

================
@@ -970,33 +971,55 @@ LogicalResult LegalizeLaunchFuncOpPattern::matchAndRewrite(
   else if (launchOp.getAsyncToken())
     stream = streamCreateCallBuilder.create(loc, rewriter, {}).getResult();
 
-  if (typeCheckKernelArgs) {
-    // The current non-bare-pointer ABI is a bad fit for `mgpuLaunchKernel`,
-    // which takes an untyped list of arguments. The type check here prevents
-    // accidentally violating the assumption made in vulkan-runtime-wrappers.cpp
-    // and creating a unchecked runtime ABI mismatch.
-    // TODO(https://github.com/llvm/llvm-project/issues/73457): Change the ABI
-    // here to remove the need for this type check.
-    for (Value arg : launchOp.getKernelOperands()) {
-      if (auto memrefTy = dyn_cast<MemRefType>(arg.getType())) {
-        if (memrefTy.getRank() != 1 ||
-            memrefTy.getElementTypeBitWidth() != 32) {
-          return rewriter.notifyMatchFailure(
-              launchOp, "Operand to launch op is not a rank-1 memref with "
-                        "32-bit element type.");
-        }
-      } else {
+  // Lower the kernel operands to match kernel parameters.
+  // Note: If `useBarePtrCallConv` is set in the type converter's options,
+  // the value of `kernelBarePtrCallConv` will be ignored.
+  OperandRange origArguments = launchOp.getKernelOperands();
+  SmallVector<Value, 8> llvmArguments = getTypeConverter()->promoteOperands(
+      loc, origArguments, adaptor.getKernelOperands(), rewriter,
+      /*useBarePtrCallConv=*/kernelBarePtrCallConv);
+  SmallVector<Value, 8> llvmArgumentsWithSizes;
----------------
andfau-amd wrote:

I prefer the explicit size in this case because `promoteOperands` returns a `SmallVector<Value, 4>`, which isn't going to be large enough once the sizes are added (three operands becomes six). But I guess it doesn't matter too much...

https://github.com/llvm/llvm-project/pull/123384