CtrlK
BlogDocsLog inGet started
Tessl Logo

testland/googletest-embedded-arm

Author and run GoogleTest 1.17+ for embedded C++ on ARM targets - TEST() / TEST_F() / TEST_P() / TYPED_TEST(), EXPECT_* vs ASSERT_* assertions, fixtures with SetUp() / TearDown(), value-parameterised tests, GoogleMock when paired, cross-compile with arm-none-eabi-g++, run on host or under QEMU via the qemu-system-test-runner skill, --gtest_filter / --gtest_output=xml:results.xml / --gtest_shuffle / --gtest_repeat command-line flags, and XML / JSON output parsing for CI. Use when the unit-under-test is C++ (modern C++17+) and the team wants the de-facto C++ test framework instead of the C-only Unity. For C use unity-test-framework-c; for pure C mocks see the CMock reference in ceedling-build-runner.

72

Quality

90%

Does it follow best practices?

Run evals on this skill

Adds up to 20 points to the overall score

View guide
SecuritybySnyk

Passed

No findings from the security scan

Overview
Quality
Evals
Security
Files
name:
googletest-embedded-arm
description:
Author and run GoogleTest 1.17+ for embedded C++ on ARM targets - TEST() / TEST_F() / TEST_P() / TYPED_TEST(), EXPECT_* vs ASSERT_* assertions, fixtures with SetUp() / TearDown(), value-parameterised tests, GoogleMock when paired, cross-compile with arm-none-eabi-g++, run on host or under QEMU via the qemu-system-test-runner skill, --gtest_filter / --gtest_output=xml:results.xml / --gtest_shuffle / --gtest_repeat command-line flags, and XML / JSON output parsing for CI. Use when the unit-under-test is C++ (modern C++17+) and the team wants the de-facto C++ test framework instead of the C-only Unity. For C use unity-test-framework-c; for pure C mocks see the CMock reference in ceedling-build-runner.
metadata:
{"keywords":"googletest, gtest, gmock, embedded, arm, c++, cortex-m, cortex-a","version":"1.3.8"}

googletest-embedded-arm

Overview

GoogleTest is, per the README at github.com/google/googletest, "Google's C++ test framework" - an xUnit-style framework merged with GoogleMock so the two ship together. The 1.17.x branch "requires at least C++17". For embedded ARM use, the framework runs anywhere a hosted C++ standard library is available - on the host, under QEMU, and on Cortex-A Linux targets; for Cortex-M without an OS, a minimal _write / _exit semihosting stub is the bridge.

This skill wraps GoogleTest for embedded C++ work. Composes with:

  • embedded-coverage-strategy-reference for gcov / llvm-cov instrumentation.
  • qemu-system-test-runner for running the binary on a virtual Cortex-M / Cortex-A.
  • For pure-C suites, prefer unity-test-framework-c; for mock-heavy suites prefer the Ceedling stack - ceedling-build-runner (CMock semantics in its references/cmock.md).

When to use

  • Unit-under-test is C++ (not C). For C, use unity-test-framework-c.
  • Target is Cortex-A running Linux or Cortex-M with semihosting - anything with enough RAM for the gtest binary (~300 KB stripped).
  • Team wants the de-facto C++ framework with rich matchers (GoogleMock, value-parameterised tests, typed tests).
  • The build pipeline can run a host build for speed and a cross-compile under QEMU for architecture sanity.

How to use

  1. Confirm the unit-under-test is C++17+ (for C, use unity-test-framework-c).
  2. Fetch GoogleTest and build the suite on the host for the fast inner loop (gtest_main supplies main()); author TEST() / TEST_F() cases - see Authoring.
  3. Cross-compile the same suite for the ARM target with arm-none-eabi-g++ and run it under QEMU for architecture sanity - full toolchain, CMake, and build invariants in references/cross-compile-and-ci.md.
  4. Emit --gtest_output=xml:results.xml; gate the host job on the JUnit XML and the QEMU job on the semihosting exit code.
  5. Reach for value-parameterised (TEST_P), typed (TYPED_TEST), and death tests as coverage grows - see references/authoring-variants.md.

The full single-test path (author to QEMU) is in Worked example below.

Authoring

Minimal test

Per google.github.io/googletest/primer.html:

#include <gtest/gtest.h>
#include "ringbuffer.h"

TEST(RingbufferTest, EmptyOnInit) {
    Ringbuffer<int, 8> rb;
    EXPECT_TRUE(rb.empty());
    EXPECT_EQ(0u, rb.size());
}

TEST(RingbufferTest, PushIncrementsSize) {
    Ringbuffer<int, 8> rb;
    rb.push(42);
    EXPECT_EQ(1u, rb.size());
    EXPECT_FALSE(rb.empty());
}

TEST(SuiteName, TestName) registers an independent test; the first argument is "the name of the test suite, and the second argument is the test's name within the test suite" (Primer doc).

Fixtures (TEST_F)

When several tests share setup, derive from ::testing::Test:

class RingbufferTest : public ::testing::Test {
protected:
    void SetUp() override { rb.clear(); }
    void TearDown() override { /* nothing */ }
    Ringbuffer<int, 8> rb;
};

TEST_F(RingbufferTest, PopReturnsPushedValue) {
    rb.push(7);
    int v = 0;
    EXPECT_TRUE(rb.pop(&v));
    EXPECT_EQ(7, v);
}

Per the Primer: "GoogleTest does not reuse the same test fixture for multiple tests" - each TEST_F gets a fresh instance.

Fatal vs non-fatal

FamilyOn failureUse when
EXPECT_* (e.g. EXPECT_EQ)Records failure; continuesEach assertion is independent; collect all failures
ASSERT_* (e.g. ASSERT_EQ)Records failure; aborts current functionSubsequent assertions would crash (e.g. null pointer deref)

Per the Primer: "Usually EXPECT_* are preferred, as they allow more than one failure to be reported in a test." Use ASSERT_* only when the next line would dereference a possibly-null pointer returned from the previous check.

Value-parameterised (TEST_P), typed (TYPED_TEST), and death tests live in references/authoring-variants.md.

Worked example

One test from host to QEMU, end to end. It uses the RingbufferTest cases from Authoring above.

  1. Put those cases in test/ringbuffer_test.cpp.

  2. Host build + run (fast inner loop) with the CMake config from references/cross-compile-and-ci.md:

cmake -S . -B build-host && cmake --build build-host
./build-host/ringbuffer_test
# [==========] Running 2 tests from 1 test suite.
# [  PASSED  ] 2 tests.
  1. Cross-compile the same source for Cortex-M4 and run it under QEMU:
arm-none-eabi-g++ -mcpu=cortex-m4 -mthumb -O0 -g -std=c++17 \
    -DGTEST_HAS_PTHREAD=0 --specs=rdimon.specs \
    -I ext/googletest/googletest/include \
    test/ringbuffer_test.cpp ext/googletest/googletest/src/gtest-all.cc \
    -o ringbuffer_test.elf -lrdimon

qemu-system-arm -M mps2-an385 -cpu cortex-m4 -nographic \
    -semihosting-config enable=on,target=native \
    -kernel ringbuffer_test.elf
echo $?   # gtest RUN_ALL_TESTS() return code, propagated via semihosting exit
  1. Two flags make the target run work: -DGTEST_HAS_PTHREAD=0 is mandatory on bare-metal M-profile (no pthreads, link fails without it), and --specs=rdimon.specs pulls in the librdimon semihosting library so stdout and exit() reach QEMU. The -kernel flag loads the ELF.

  2. For a machine-readable gate, add --gtest_output=xml:results.xml to the host run and count failing testcases - the XML schema and parsing pattern are in references/cross-compile-and-ci.md.

Command-line flags

The flags every embedded run reaches for (per the Advanced Guide):

FlagEffect
--gtest_filter=Pattern:-separated wildcard list; supports *, ?, negative -
--gtest_output=xml:results.xml / json:results.jsonMachine-readable report for CI
--gtest_shuffleRandom order each run - reveals inter-test dependencies
--gtest_repeat=NRepeat N times (-1 = infinite, for flake hunting)

The full flag set (--gtest_break_on_failure, --gtest_catch_exceptions=0, --gtest_brief, --gtest_list_tests, and more), the XML / JSON output schema, and the CI-gate parsing pattern are in references/cross-compile-and-ci.md.

Anti-patterns

Anti-patternWhy it failsFix
ASSERT_EQ then ignore the abort and continueSubsequent assertions undefined; can crashUse EXPECT_EQ unless next line dereferences the value
GTEST_HAS_PTHREAD=1 on bare-metal Cortex-MLink fails: pthread symbols missingAlways -DGTEST_HAS_PTHREAD=0 on M-profile cross-builds
Death tests on bare-metalNo fork(); test hangs or aborts oddlySkip death tests on M-profile; or use --gtest_filter=-*DeathTest*
Linking gtest without gtest_main then no own main"undefined reference to main"Link gtest_main or write int main(int argc, char **argv) { testing::InitGoogleTest(&argc, argv); return RUN_ALL_TESTS(); }
Test order dependence--gtest_shuffle reveals; CI breaks intermittentlyEach TEST_F must work in any order; reset state in SetUp
Optimised coverage build-O2 collapses branches gcov can't see-O0 -g for the coverage build per coverage-reference skill
TEST_P without INSTANTIATE_TEST_SUITE_PCompiles but never runsAlways pair
Mocking everythingTests measure mocks not behaviourMock at the I/O boundary only - see ceedling-build-runner's references/cmock.md

Limitations

  • C++ only. For pure C suites use unity-test-framework-c.
  • Heap required. GoogleTest allocates internally; ultra-low- RAM Cortex-M0 (8 - 16 KB) may not have headroom. Use Unity for those targets.
  • No native parameterised test = different test names. Each TEST_P instance is Boundaries/WrapTest.WrapAtCapacity/4 - the /4 is the param index, not the value. Use INSTANTIATE_TEST_SUITE_P(..., ::testing::PrintToStringParamName()) to encode the value into the name.
  • Death tests require fork() - not available on bare-metal or under most RTOSes. Per the Advanced Guide they run "in separate child processes".
  • No determinism guarantee with --gtest_shuffle - the seed is logged but rerunning with the same seed and a code change produces a different schedule.
  • GoogleMock pulls in std::function-based call recording - on cross-compiles with -fno-exceptions you may need -DGTEST_HAS_EXCEPTIONS=0 to compile cleanly.

References

Cited inline. Foundational documents:

Workspace
testland
Visibility
Public
Created
Last updated
Publish Source
GitHub
Badge
testland/googletest-embedded-arm badge