[Flutter GPU] FlutterGpuCommandBufferTest aborts in debug builds: CommandBuffer is created on the stack #194149

Problem

The three tests in lib/gpu/command_buffer_unittests.cc (FlutterGpuCommandBufferTest.*) abort in an unoptimized debug build:

[ RUN      ] FlutterGpuCommandBufferTest.RejectsCompletionCallbacksAfterSubmit
[FATAL:flutter/fml/memory/ref_counted_internal.h(97)] Check failed: !adoption_required_.
[ERROR:flutter/fml/backtrace.cc(108)] Caught signal SIGABRT during program execution.
...
Frame 4: fml::internal::RefCountedThreadSafeBase::~RefCountedThreadSafeBase()

Each test builds the flutter::gpu::CommandBuffer on the stack:

CommandBuffer command_buffer(context, impeller_command_buffer);

CommandBuffer is ref-counted, and an fml ref-counted object has to be adopted (created through fml::MakeRefCounted) before it is destroyed. A stack instance is never adopted, so its destructor trips the debug check. The check is debug-only, which I assume is why CI doesn't see it.

Steps to reproduce

macOS arm64, an existing host_debug_unopt_arm64 build. The test file is the same on current main.

cd engine/src
ninja -C out/host_debug_unopt_arm64 flutter/lib/gpu:gpu_unittests
out/host_debug_unopt_arm64/gpu_unittests --gtest_filter='FlutterGpuCommandBufferTest.*'

All three tests abort as above. The other gpu_unittests suites pass.

Proposed fix

Create the command buffer with fml::MakeRefCounted<CommandBuffer>(...) in the three tests and use it through the returned fml::RefPtr. No engine code changes.

I ran into this while testing a few Flutter GPU fixes locally; until it's fixed I leave FlutterGpuCommandBufferTest.* out of the gpu_unittests filter in those PRs.