warning: 'resolved_axis' may be used uninitialized in this function (reduce.cc)

[tcal-x]

This was seen when we attempted to bump the tflite-micro submodule in our project up to current main, from about a month ago (google/CFU-Playground#230 (comment)).

Since we compile TFL-micro with -Werror (using riscv gcc v8.3), this warning becomes an error for us:

In file included from /home/tim/tcal-x/CFU-Playground/proj/proj_template_v/build/src/tensorflow/lite/kernels/internal/common.h:29,
                 from /home/tim/tcal-x/CFU-Playground/proj/proj_template_v/build/src/tensorflow/lite/kernels/internal/reference/reduce.h:19,
                 from /home/tim/tcal-x/CFU-Playground/proj/proj_template_v/build/src/tensorflow/lite/micro/kernels/reduce.cc:16:
/home/tim/tcal-x/CFU-Playground/proj/proj_template_v/build/src/tensorflow/lite/kernels/internal/types.h: In function 'TfLiteStatus tflite::ops::micro::reduce::EvalMean(TfLiteContext*, TfLiteNode*)':
/home/tim/tcal-x/CFU-Playground/proj/proj_template_v/build/src/tensorflow/lite/kernels/internal/types.h:185:33: warning: 'resolved_axis' may be used uninitialized in this function [-Wmaybe-uninitialized]
         if (idx == axis[axis_idx]) {
                    ~~~~~~~~~~~~~^
/home/tim/tcal-x/CFU-Playground/proj/proj_template_v/build/src/tensorflow/lite/kernels/internal/types.h:185:33: warning: 'resolved_axis' may be used uninitialized in this function [-Wmaybe-uninitialized]
         if (idx == axis[axis_idx]) {
                    ~~~~~~~~~~~~~^
/home/tim/tcal-x/CFU-Playground/proj/proj_template_v/build/src/tensorflow/lite/kernels/internal/types.h:185:33: warning: 'resolved_axis' may be used uninitialized in this function [-Wmaybe-uninitialized]
         if (idx == axis[axis_idx]) {
                    ~~~~~~~~~~~~~^

I think it's referencing this: https://github.com/tensorflow/tflite-micro/blob/main/tensorflow/lite/micro/kernels/reduce.cc#L149 , and I can't see where/how it gets initialized.

If you don't see these warnings, then there must be something wrong in our usage of tfl-micro.

[tcal-x]

When I compiled without -Werror, there were no additional warnings and compilation finished successfully.

[advaitjain]

As an aside, if you'd like to avoid pulling in a whole lot of unrelated code into your repo, you could use TFLM's new support for external integration:

• https://github.com/tensorflow/tflite-micro/blob/main/tensorflow/lite/micro/docs/new_platform_support.md
• Move hardware / platform support to repositories that are external to tflite-micro #407

To address this specific issue, while the code in question is certainly complex, AFAICT this is a false positive.

The variable is passed as a pre-allocated buffer and gives the impression of being an input parameter. However, if we trace the call stack it is initialized.

For example,

tflite-micro/tensorflow/lite/micro/kernels/reduce.cc

Lines 171 to 177 in a9f2e03

	reference_ops::Mean(
	tflite::micro::GetTensorData<float>(input), input->dims->data,
	input->dims->size, tflite::micro::GetTensorData<float>(output),
	output->dims->data, output->dims->size,
	tflite::micro::GetTensorData<int>(axis), num_axis,
	params->keep_dims, temp_index, resolved_axis,
	tflite::micro::GetTensorData<float>(output)));

calls into

tflite-micro/tensorflow/lite/kernels/internal/reference/reduce.h

Lines 215 to 240 in a9f2e03

	inline bool Mean(const T* input_data, const int* input_dims,
	const int input_num_dims, T* output_data,
	const int* output_dims, const int output_num_dims,
	const int* axis, const int num_axis_dimensions, bool keep_dims,
	int* temp_index, int* resolved_axis, U* temp_sum) {
	ruy::profiler::ScopeLabel label("Mean");
	// Reset output data.
	size_t num_outputs = 1;
	for (int idx = 0; idx < output_num_dims; ++idx) {
	size_t current = static_cast<size_t>(output_dims[idx]);
	// Overflow prevention.
	if (num_outputs > std::numeric_limits<size_t>::max() / current) {
	return false;
	}
	num_outputs *= current;
	}
	for (size_t idx = 0; idx < num_outputs; ++idx) {
	output_data[idx] = T();
	temp_sum[idx] = U();
	}
	// Resolve axis.
	int num_resolved_axis = 0;
	if (!ResolveAxis(input_num_dims, axis, num_axis_dimensions, resolved_axis,
	&num_resolved_axis)) {
	return false;

which then initializes the variable in

tflite-micro/tensorflow/lite/kernels/internal/reference/reduce.h

Lines 107 to 109 in a9f2e03

	inline bool ResolveAxis(const int num_dims, const int* axis,
	const int64_t num_axis, int* out_axis,
	int* out_num_axis) {

It may be worth adding -Wno-maybe-uninitialized to the flags for TFLM in your repo.

For completeness, here are the flags that the TFLM repo uses:

gcc on x86:

tflite-micro/tensorflow/lite/micro/tools/make/Makefile

Lines 130 to 145 in a9f2e03

	CC_WARNINGS := \
	-Wsign-compare \
	-Wdouble-promotion \
	-Wshadow \
	-Wunused-variable \
	-Wmissing-field-initializers \
	-Wunused-function \
	-Wswitch \
	-Wvla \
	-Wall \
	-Wextra \
	-Wstrict-aliasing \
	-Wno-unused-parameter
	COMMON_FLAGS := \
	-Werror \

clang on x86:

tflite-micro/tensorflow/lite/micro/build_def.bzl

Lines 1 to 6 in a9f2e03

	def micro_copts():
	return [
	"-Wall",
	"-Werror",
	"-DFLATBUFFERS_LOCALE_INDEPENDENT=0",
	]

[tcal-x]

Thank you for the detailed answer, @advaitjain . We are currently manually copying the directories and files that we need from TFLM; we should switch over to the new support.

FYI @alanvgreen