kammce / exception_research

Exists to test exceptions and result types and compare their impact on ROM and RAM as well as Happy/Sad path performance

exception_v_result

Warning

These instructions are old and much of this project has been refactored but is not currently working.

Exists to test exceptions and result types and compare their impact on ROM and RAM as well as Happy/Sad path performance.

How to run Performance benchmarks

To generate the code & compile it:

python3 generate.py > info.csv
conan build . -pr baremetal.profile

The info.csv has information about each group ID and what its call depth is and its percentage of functions with clean up routines.

Open a new terminal and run the following, replace the target with your own:

pyocd gdbserver --target=lpc4088 --persist

You'll need a JTAG or SWD debug probe connected to your host machine as well as have that debug probe connected to your target device.

In the same terminal you compiled the code run:

source build/MinSizeRel/generators/conanbuild.sh

This will add the tools downloaded by conan to the path of your terminal, which will give you access to arm-none-eabi-gdb.

Run this command to start GDB and begin debugging:

arm-none-eabi-gdb build/MinSizeRel/except.elf -ex "target remote :3333" -ex "set mem inaccessible-by-default off"

From here run the following gdb commands:

load
b main
continue

• load: loads the elf file into the device
• b main: sets a breakpoint at main
• continue: runs the code until it reaches main (start and run are not used)

Put a break point at the line number where the while(true) { continue; } is and enter continue. When the program stops inspect the cycles_map function.

print cycles_map

The cycles map has the number of cycles required to throw an exception and catch it. Save this to a file and replace all , (comma with a space) with a new line character. Import the CSV into a spreadsheet program and copy and paste the vertical list into the last column of the spreadsheet. And now you have performance data.

How to run Size benchmarks

Haven't written this yet.

Assumptions about software in general

1. Error handlers are rare in code.

Most of the code written in an application is dedicated to performing work. It is not dedicated to handling errors.

It would be very rare to find a code base that is composed of mostly (above 50%) error handling code over productive code.

2. Most errors require context to be handled

In general, the only way to properly handle an error is to understand exactly how and why the error was generated and also understands that there is a way to recover from that error in software, and this requires context.

Extreme example: You have a function called int sum(int a, int b). It can detect overflows. This function could emit the information in some way, exceptions or result return types, or it can terminate the application. Termination seems like it should ALWAYS be the wrong choice because:

1. This function is making an assumption that the application can tolerate terminate being called as a result of this function.
2. Raising the risk of calling this function with invalid inputs, requiring the developer to add additional checks before calling the function, or wrapping the function to detect potential overflows and emitting an error, before calling this function. This defeats the purpose of the original function.

Typically, what the developer wants is to be allow to choose to ignore or handle this error rather than letting the library make critical decision about the application.

3. Error information tends to be more meaningful at higher stack frame.

The higher the call frame, the more context about the program is known. The highest call frame of an application is main(). Developers writing their main() function have full understanding of what they set out to do in their application, thus main has the MOST context regarding what to do when errors occur and how to handle them.

main() could be too high up the stack in cases where main runs the real application function or it simply spawns a bunch of threads. In which case, the application function or the threads would be the functions with the most context about the execution of the function and would be the best places to handle errors.

It may not always be the highest stack, but some starting stack frame has the most context regarding handling errors for its call stack.

4. Errors are exceptional situations

An 64-bit integer overflow, when dealing with real numbers, is a very exceptional case. Out of bounds access are exceptional, if not simply a bug in the program. Speaking to a sensor over I2C should almost always work after the first time communication is established.

Costs

All tested on a Cortex M4 MCU (lpc4078), using ARM GNU GCC 12.2.1.

💾 Storage Costs

• Exceptions:
  • Enabling exceptions: 5005 bytes
  • catch block: TBD
  • catching a new type: TBD
  • throwing a type: TBD
  • throwing a new type: TBD
• Result<T,E>:
  • Enabling/using std::expected: TBD
  • Handling an expected value: TBD (should be 0 bytes)
  • Returning std::unexpected: TDB
  • Checking & returning an unexpected value up a level: TBD
  • Returning a new error object

⏱️ Time Costs

All tests are conducted with call depths of 1, 2, 4, 8, 10, 16, 32, 50, 64, and 100. Results will be in the number of CPU cycles, NOT in time as time can be relative to the CPU operating frequency. With cycles, the amount of time between an error being emitted and handled by dividing the number of cycles by the CPU operating frequency.

NOTE: the size of an exception object does not affect the speed in which an exception is thrown beyond constructing the object. The size of the object does matter to std::expected. The inheritance hierarchy of an error object returned using std::unexpected does not affect the speed of returning the object, but this can greatly affect the speed in which a catch statement can be found.

• Exceptions:
  • Throwing T, catching T, where T is a trivially destructible object
  • Throwing T, catching T*, where T is a trivially destructible object
  • Throwing V, catching I, where I is an interface and V implements I
  • Throwing W, catching E, where E is some base type of W but W has more than 2 bases (should be the slowest).
• Result:
  • std::unexpected(T) where sizeof(T) == 4 bytes
  • std::unexpected(U) where sizeof(U) > 65 bytes (calls memcpy)
  • std::unexpected(W) where sizeof(W) == 8 bytes
  • std::unexpected(V) where V is a std::unique_ptr<I>, where I is an interface V implements I.

🧠 Memory Costs (RAM)

• Exceptions:
  • Amount of allocated memory (statically or dynamically) when throwing an exception.
  • Amount of .data/.bss memory taken up
  • Stack required to run libunwind routines after throw (use stack watermark).
• Result:
  • Amount of additional stack memory used to support result types

⚖️ Factors that contribute to code size

Error transport mechanism

✅ Exception

Exceptions utilize an algorithm along with some slight code instrumentation in order to manage the task of bringing an error object up to a catch statement, and running all necessary destructors on its way up to the catch. The cost for the unwinding mechanism is a fixed cost. As a project gets larger, the percentage of the mechanism relative to the overall project gets smaller.

We've been capable of getting the total cost of the error transport mechanism, in Arm Gnu GCC 12.2.1, down to 5005 bytes.

❌ Result<T,E>

Result return types require each call site to be responsible for either handling the error or helping to transport it up a level, deferring it to a higher function with more context to deal with it. This error handling mechanism has no initial up front cost and thus looks cheap to use. Note that, in most cases most functions do not have the context to determine how an error should be handled so they typically hoist it up rather than handle it directly. As an example, if a fixed point math operation overflows and an error is generated, it is unlikely (or not acceptable) that the library developer of the math library will decide to terminate or handle this error for you. That would give less control to the application developer. Thus, the cost of each call to a function that returns a result type requires the following:

1. Cost of the stack memory to hold the result
2. An additional instruction to pass the address of that stack memory location to the function.
3. The instructions to check if the result type is an error not
4. The instructions to copy/construct the error object
5. The instructions to return that error object up one function call

This will cost very little initial, but the more you make this pervasive throughout the code base, the more the costs add up, until you finally reach beyond 5005 bytes worth of memory. In which case exceptions become the cheaper alternative.

❓ Caveats

If the device you are working with has flash below 32kB, then result return types can be a decent alternative. Although, in such situations where error handling will be sparse, it may also be a good idea to simply save error information to a region of RAM that will persist and terminate/reset the controller.

Number of types thrown/unexpected

❌ Exceptions

Each thrown type gets its own type info object created. Along with the <typeinfo name for X>, where X is the name of the string literal mangled name of the type. This string information is inaccessible to programs using -fno-rtti. Each type info object occupy's a place in the .rodata region of the binary. The address of that type info object is compared to the address of the type info object in the catch statement. Note that the name of the type is not used for comparisons as that would be too slow. The longer the type's name, the larger that type info is. This cost is only paid once, so the same object can be thrown multiple times without generating additional type info objects.

How to fix this for exceptions? Patch GCC to add support for -fstrip-typeinfo-name, which would strip out type info type names when -fno-rtti and -fexceptions are both being used. This way, we save a bit more memory when using exceptions and no longer have to be concerned with the length and number of exceptions through. IMHO a 32-bit word for each thrown value is a very cheap price.

🟡 Result<T,E>

The number of error types does not directly affect the binary size. But there is a limitation where, if the error type is NOT the same across all functions in a call stack, then when a function of return type Result<T, E1> calls a function of return type Result<T, E2>, then the first function either needs to handle the error within its scope, which it may not have enough context to do so, or it must transform E2 to E1 which may result in information loss. An extreme example is return std::expected<int, my_error_t> to a function that return std::optional<int>. The info in my_error_t will be destroyed.

Number of functions using classes with non-trivial destructors

Both systems have to pay a cost for handling non-trivial destructors. There is no way around this.

✅ Exception

Exceptions utilize a region of the function where the non-trivial destructors can be called in reverse order of their creation. It will look something like this:

6e2e: e008       b.n 6e42 <funct_group13_18()+0x56>
6e30: a801       add r0, sp, #4
6e32: f7f9 fa45  bl 2c0 <class_1::~class_1()>
6e36: 4668       mov r0, sp
6e38: f7f9 fa1a  bl 270 <class_0::~class_0()>
6e3c: f000 fbce  bl 75dc <__cxa_end_cleanup>
6e40: e7f9       b.n 6e36 <funct_group13_18()+0x4a>
6e42: a801       add r0, sp, #4

In address 6e2e you can see the instruction b.n 6e42 which means, unconditional branch to location 6e42. Its jumping over the region that destructs the classes with non-trivial destructors. When the exception unwind mechanism is running, it will find this location in the list of functions with cleanup work and will either place the program counter at 6e30 which will destruct both classes or will place it at 6e36 to destruct only class_0. The function __cxa_end_cleanup simply runs the unwind_resume() function from libunwind to continue up the stack until a catch is reached.

❌ Result<T,E>

Result types on the other hand, require additional branching to support running the destructors and returning the error value. Below we see a factory function that returns tl::expected<class_0, my_error_t>. If the result of the factory function is not a error, then it branches over the else handler and continues running. If the result was an error, then the else branch is taken and it has to manage constructing the result object and destructing the original object.

ad2e: a802       add r0, sp, #8
ad30: f7f5 fa40  bl 1b4 <class_0::make(long)>
ad34: f89d 300c  ldrb.w r3, [sp, #12]
ad38: b94b       cbnz r3, ad4e <funct_group13_18()+0x32>
ad3a: 9a02       ldr r2, [sp, #8]
ad3c: 9206       str r2, [sp, #24]
ad3e: 6022       str r2, [r4, #0]
ad40: 7123       strb r3, [r4, #4]
ad42: a802       add r0, sp, #8
ad44: f7f5 fa82  bl 24c <tl::detail::expected_storage_base<class_0, my_error_t, false, true>::~expected_storage_base()>
ad48: 4620       mov r0, r4
ad4a: b009       add sp, #36 ; 0x24
ad4c: bd30       pop {r4, r5, pc}
ad4e: 9902       ldr r1, [sp, #8]

⚖️ Comparing the two:

Exceptions are not burdened with the task of transporting errors to their destination. This responsibility is placed on the exception runtime algorithms that can be reused for every possible function in the entire application. With result types, each function that participates in error transport must also do the work of transporting errors.

Cost of error handling routine

❓ Exceptions

The cost of handling an error with a try/catch is:

TBD (need to investigate this...)

✅ Result<T,E>

Since the machinery for handling errors is always local, the moment a handler gets a response from a function, it can choose to handle it or transport it up. This means that the error can be handled at any point in the stack. The overall cost of handling an exception is no greater than the cost of an if/else statement.

Because we assume that handlers are rare in software, the overall cost for handling a routine with result types is smaller.

🏎️ Factors that contribute to sad path performance

Reaching the error handler

[INCOMPLETE]

✅ Exceptions when they work well

Additional information: exceptions work in two phases in GCC, the search phase and unwind phase. The search phase traverses up the call stack quickly to find the appropriate catch statement. If it does not find one, it executes terminate. This allows the program to terminate quickly when no viable handlers are available. If it does find a suitable handler, it begins phase 2. Phase 2 is where the exception mechanism follows the stack link registers like a linked list, figures out which functions have destructors to execute, executes them if they are required, and resumes stack traversal until it reaches the designated catch blocks. The list of functions with non-trivial destructors with possible throwing behavior is organized in a sorted list of functions based on their addresses. Given the current

❌ Exceptions when they work poorly

🟡 Result<T,E>

TBD

🏎️ Factors that contribute to happy path performance

Limitations

TBD

Benefits of

TBD

Tricks of the trade

Weakening symbols so they can be overridden

arm-none-eabi-objcopy \
/Users/kammce/.conan2/p/b/arm-g788a41ed25167/p/bin/bin/../lib/gcc/arm-none-eabi/12.2.1/thumb/v7e-m+fp/hard/libgcc.a \
--globalize-symbol=search_EIT_table --weaken-symbol=search_EIT_table \
/Users/kammce/.conan2/p/b/arm-g788a41ed25167/p/bin/bin/../lib/gcc/arm-none-eabi/12.2.1/thumb/v7e-m+fp/hard/libgcc.a

Full:

arm-none-eabi-objcopy \
<insert_path_to_arm-gnu-toolchain>p/bin/bin/../lib/gcc/arm-none-eabi/12.2.1/thumb/v7e-m+fp/hard/libgcc.a \
--globalize-symbol=search_EIT_table --weaken-symbol=search_EIT_table \
<insert_path_to_arm-gnu-toolchain>p/bin/bin/../lib/gcc/arm-none-eabi/12.2.1/thumb/v7e-m+fp/hard/libgcc.a

arm-none-eabi-objcopy \
<insert_path_to_arm-gnu-toolchain>p/bin/bin/../lib/gcc/arm-none-eabi/12.2.1/thumb/v7e-m+fp/hard/libgcc.a \
--globalize-symbol=_Unwind_VRS_Pop --weaken-symbol=_Unwind_VRS_Pop \
<insert_path_to_arm-gnu-toolchain>p/bin/bin/../lib/gcc/arm-none-eabi/12.2.1/thumb/v7e-m+fp/hard/libgcc.a

arm-none-eabi-objcopy \
<insert_path_to_arm-gnu-toolchain>p/bin/bin/../lib/gcc/arm-none-eabi/12.2.1/thumb/v7e-m+fp/hard/libgcc.a \
--globalize-symbol=__gnu_unwind_execute --weaken-symbol=__gnu_unwind_execute \
<insert_path_to_arm-gnu-toolchain>p/bin/bin/../lib/gcc/arm-none-eabi/12.2.1/thumb/v7e-m+fp/hard/libgcc.a

License

Apache 2.0; see LICENSE for details.