[Mlir-commits] [mlir] [MLIR] Modify lowering of gpu.alloc op to llvm (PR #69969)

[Guray Ozen]

================
@@ -836,7 +836,11 @@ LogicalResult ConvertAllocOpToGpuRuntimeCallPattern::matchAndRewrite(
   // Allocate the underlying buffer and store a pointer to it in the MemRef
   // descriptor.
   Type elementPtrType = this->getElementPtrType(memRefType);
-  auto stream = adaptor.getAsyncDependencies().front();
+
+  Value stream =
+      adaptor.getAsyncDependencies().empty()
----------------
grypp wrote:

> ok I agree. I made the change. 

Thanks for doing that! It looks better now. We just need a test for this PR as it’s regular llvm code submission policy. 

> I added the stream in the first place is because all the runtime signatures for mgpuMemAlloc expects a stream (albeit it doesn't use it now) ...but in future if they do require the stream passed for allocation ...but I agree for now we can pass nullptr 

Would you like to separate runtime APIs into mgpuMemAlloc and mgpuMemAllocAsync? The first one doesn't need a stream in the function declaration, while the second one works asynchronously and requires a stream. This is an additional work if you are interested. 

https://github.com/llvm/llvm-project/pull/69969