Converting RUN_TEST() from a macro to a function significantly reduces the size
of the compiled binary. On amd64, the largest test runner in the test suite
(testsample_DefaultsThroughCommandLine_runner.o) was reduced from 3.4 kB to 2.4
kB (stripped).
- Running time macros have been made more portable, previously it was not
possible to override all macros
- Running time macros will be executed by default test runner, and auto test
runners
- Adds a default execution time implementation for unix. (Previous default
implementation only worked on Windows)
- For embedded platforms there is a simple method of getting a default
implementation by defining a single macro UNITY_CLOCK_MS()
- Removed need for UNITY_EXEC_TIME_RESET. This was not being used for the default
implementations, if anything ever did need reset-like functionality it could
simply be wrapped up with the start or stop macros for that platform
The existing implementation was not very good:
- It printed all very small values as "0.000000..."
- It did not distinguish positive and negative zero
- In some cases it printed extra garbage digits for single-precision values
(e.g. 3.9e+30 was printed as 3.90000013+30)
Tests have been updated to check that we now match printf("%.6g") for
1,000,000 randomly chosen values, except for rounding of the 6th digit.
Using this option changes the control flow of Unity, but is useful on
constrained embedded systems. You can't fully simulate the power of
'longjmp' with just 'return', but Unity still works well, all tests pass.
Originated from commit c6dc96f3 but are rarely used and not in math.h
Removes float to double promotion warning and simplifies the code
'gcc -std=gnu99 -Wdouble-promotion' spits out these warnings
Takes about 10 minutes to run all floats, so split into 3 tests for parallel
running later. This was useful during development for finding hard corner cases
and getting the routine to high quality. Off by default.
Note that all floats 16.0 and up can be represented uniquely (in this format) and will
round-trip back to the exact same float with sscanf(). This property is true for
UnityPrintFloat, despite a few rounding error cases, it will produce output identity.
Better comments and refactor on round ties to even
Add upper threshold value on round to even feature since numerical precision issues
start to give approximations when dividing one large double by another. When tested
on float values, using 1e22 gave the fewest rounding errors.
Fix warnings from gcc. Some float constants do not behave well in existing tests.
Add casts where conversions could be imprecise.
Remove UNITY_FLOAT_VERBOSE entirely, add option UNITY_EXCLUDE_FLOAT_PRINT
Remove some questionable float casts from doubles
Default to Round Ties to Even behavior, add option to Round Ties Away from Zero
John Lindgren