Python low-level AOP using AST rewriting – part II

In the previous post we started modifying functions body for AOP purposes. So far we’ve transformed a function to an AST node with extended code. In this post we will transform it to bytecode and replace the original method code with the new one.

As to this moment, we have these two:

  1. original_function – the decorated function, without any modification.
  2. modified_node – the modified node, with extended code and fixed locations.

Compiling the node into a python module

Given the node, we can use Python built-in compile. Compile takes either string or AST node. The node is transformed into a module which we can later execute.

def compile_node(modified_node):
    compiled_method = compile(modified_node, '<string>', 'exec')
    return compiled_method

The two arguments we send to compile are the source file name to which the method will be linked. In our case, we’re not going to keep the method compiled method, and only use its bytecode. Therefore, we can leave a default value ‘<string>’ which is commonly used. The second argument is ‘exec’, which signals the compiler that multiple statements may appear in the input code and that we don’t care about the returned value. This is a simplified explanation, but sufficient for this context.

Getting replacement bytecode

Python code object does not allow setting the bytecode. Therefore we need to create a new code object, which is identical to the original code, other than the bytecode and stacksize:

def create_replacement_code(original_function, compiled_function):
    # these are the code object field names, ordered by the expected order of code object constructor
    code_arg_names = ['co_argcount', 'co_nlocals', 'co_stacksize',
                      'co_flags', 'co_code', 'co_consts', 'co_names',
                      'co_varnames', 'co_filename', 'co_name',
                      'co_firstlineno', 'co_lnotab']
    # fill all the args based on the original code
    code_args = [getattr(original_function.func_code, key, None)
                 for key in code_arg_names]
    # replace the bytecode and stacksize with the compiled bytecode and stacksize from the modified function
    code = extract_code(compiled_function)
    # replace bytecode and stacksize with the compiled method values
    code_args[code_arg_names.index('co_code')] = code.co_code
    code_args[code_arg_names.index('co_stacksize')] = code.co_stacksize
    return CodeType(*code_args)

The instantiation of the code object is undocumented. Yet, some examples are available online. Please note that this signature fits python 2.7 and was changed in 3.0. Now we finished transforming a python code (which is the original function) into a new code object, containing the modified bytecode.

The last two methods calls can be extracted to:

def create_func_code(original_function, modified_node):
    compiled_method = compile_node(modified_node)
    return create_replacement_code(original_function, compiled_method)

Replacing the code

The last step required is to switch the function’s code into the modified one. Luckily, this step is very simple:

def replace_code(original_function, code):
    original_function.func_code = code

Now, when the decorated method is executed, it’ll contain the modified bytecode, which checks if variables used in comparisons are not None.

The full code

from _ast import BoolOp, And, Name, Compare, IsNot, Load, Num
from ast import NodeTransformer, fix_missing_locations
import ast
import inspect
from types import FunctionType, CodeType


class ComparisonTransformer(NodeTransformer):
    def visit_Compare(self, node):
        parts = [node.left] + node.comparators

        # check if any constant number is involved in the comparison
        if not any(isinstance(part, Num) for part in parts):
            return node

        # get all the "variables" involved in the comparison
        names = [element for element in parts if isinstance(element, Name)]
        if len(names) == 0:
            return node

        # create a reference to None
        none = Name(id='None', ctx=Load())
        # create for each variable a node that represents 'var is not None'
        node_verifiers = [Compare(left=name, ops=[IsNot()], comparators=[none]) for name in names]
        # combine the None checks with the original comparison
        # e.g. 'a &lt; b &lt; 1' --&gt; 'a is not None and b is not None and a &lt; b &lt; 1
        return BoolOp(op=And(), values=node_verifiers + [node])


def rewrite_comparisons(original_function):
    assert isinstance(original_function, FunctionType)

    node = parse_method(original_function)
    rewrite_method(node)
    code = create_func_code(original_function, node)
    replace_code(original_function, code)
    return original_function


def replace_code(original_function, code):
    original_function.func_code = code


def parse_method(original_function):
    return ast.parse(inspect.getsource(original_function))


def rewrite_method(node):
    # assuming the method has single decorator (which is the rewriter) - remove it
    node.body[0].decorator_list.pop()
    # we rename the method to ensure separation from the original one.
    # this step has no real meaning and not really required.
    node.body[0].name = 'internal_method'
    # transform Compare nodes to fit the 'is not None' requirement
    ComparisonTransformer().visit(node)
    # let python try and fill code locations for the new elements
    fix_missing_locations(node)


def create_func_code(original_function, modified_node):
    compiled_method = compile_node(modified_node)
    return create_replacement_code(original_function, compiled_method)


def compile_node(modified_node):
    compiled_method = compile(modified_node, '&lt;string&gt;', 'exec')
    return compiled_method


def extract_code(compiled_method):
    exec compiled_method
    generated_func = locals()['internal_method']
    return generated_func.func_code


def create_replacement_code(original_function, compiled_function):
    # these are the code object field names, ordered by the expected order of code object constructor
    code_arg_names = ['co_argcount', 'co_nlocals', 'co_stacksize',
                      'co_flags', 'co_code', 'co_consts', 'co_names',
                      'co_varnames', 'co_filename', 'co_name',
                      'co_firstlineno', 'co_lnotab']
    # fill all the args based on the original code
    code_args = [getattr(original_function.func_code, key, None)
                 for key in code_arg_names]
    # replace the bytecode and stacksize with the compiled bytecode and stacksize from the modified function
    code = extract_code(compiled_function)
    # replace bytecode and stacksize with the compiled method values
    code_args
[code_arg_names.index('co_code')] = code.co_code
    code_args[code_arg_names.index('co_stacksize')] = code.co_stacksize
    return CodeType(*code_args)


@rewrite_comparisons
def foo(x):
    return x < 1

print foo(None)

This code, will output the following, as expected:

False

To prove that this code is works in other context, the following code:

@rewrite_comparisons
def bar(x, y, z, w):
    if x < 1 or y < 2 or z < 3:
        return 'Default behavior'
    if w < 1:
        return 'Expected behavior'
    return 'Failure'


print bar(None, None, None, 0)

Prints the following result:

Expected behavior

Summary

Python is amazingly flexible and using ~50 lines of code we can create a micro-framework to manipulate methods behavior. In addition, this flexibility allows code changes in many levels, including instructions level changes.

The vast majority of use cases are on the method level and not the instructions. Yet, this facilities can be useful in some cases an we should be able to take advantage of them.

Python low-level AOP using AST rewriting – part I

This post and the next post will address AOP in Python. In general AOP in Python is very simple thanks to Python’s decorators. The aspects which we would like to apply in this post are low-level, meaning they’ll be applied on in-body instructions and not just on method level. The way in which we’re going to implement it will be using code weaving and rewriting.
I previously blogged about similar concept in .Net using Mono Cecil, where we tracked IL instructions.

The topic will be covered by two posts, where the first one will address rewriting code and the second one will deal with replacing the original code.

Background

Motivation

The general motivation for AOP is to separate the business logic from other functional logic, like logging, security or error handling. Most of the common examples fit the pattern of wrapping the function with new one. Then, perform logic before/after the method is executed. This is very useful, yet, limits our ability to change behavior of specific instructions inside the method which are relevant to the aspect.

Example

During the post we will use a concrete simple example. Let us observe the following example (Python 2.7):

def foo(x):
    return x < 1

print foo(None)

As you probably know, this will print:

True

This is a common Python (2.7) behavior but might not be intuitive. In general, assuming we had many variables and many comparisons, we’d like to change all to the pattern: VAR is not None and VAR < CONST

The goal of our process will be to transform the method to:

def foo(x):
    return x is not None and x &lt; 1

Where the aspect we’re applying is Update Comparison of None and Constants.

The required steps

The steps required by this solution are the following:

  1. Decorate the method – create an entry point for the mechanism which’ll apply the aspect.
  2. Create an AST from the method –  prepare a modifiable syntax tree from the original method.
  3. Rewrite the AST – find the instructions influenced by the aspect and modify them.
  4. Create bytecode – create identical code to the original one other than newly generated bytecode.
  5. Update the method – replace the original method code with the new one.

Decorating the method

Like the common approach, we will use a decorator to modify the function. We will start from this simple decorator and build over it:

def rewrite_comparisons(original_function):
    assert isinstance(original_function, FunctionType)
    return original_function

@rewrite_comparisons
def foo(x):
    return x &lt; 1

This decorator does nothing, so far.

Getting code from function

The first challenge is getting the method code from a function and make it modifiable. Since Python provides bytecode by default for a method, we will use built-in inspect to extract the original source code:

function_source_code = inspect.getsource(original_function)

Inspect uses the code locations linked to the function and read them from the source file. The return value is a string with function. This is different from disassembling code from the method bytecode.

We can assume that for our functions the source code is available. Otherwise, this first step will fail, and the processing will need to be in bytecode level (which might be covered in other post). In addition, this constrains us to ensure decorator is called before any other decorator. Otherwise, previous decorators might be ignore since their effect is not reflected in the original source code.

Building an AST (abstract syntax tree)

After the previous line of code extracted the source, we can parse it to an AST. The motivation for building an abstract syntax tree, is that it’s modifiable and we can compile is back to bytecode.

function_source_code = inspect.getsource(original_function)
node = ast.parse(function_source_code)

The node we get is the root one of the parsed code. It links to all the elements in the hierarchy and represents a simplified module code.

Taking for example the foo function, the tree is:

Module
  # the method declaration (foo)
  FunctionDef
    # the arguments list (x)
    arguments
      Name
        Param
    # return instruction
    Return
      # comparison of two elements
      Compare
        # load variable (x)
        Name
          Load
        # comparison operator (<)
        Lt
        # load constant (1)
        Num

The AST represents the function, while the decorator is omitted for simplicity. As can easily be seen, the tree represents all the content of the method, including declaration, other methods in context if there are and more. Given the AST, we’d like to modify it a fit the need that our aspect requires.

Transforming the AST

AST visitors

We will use the AST visitors as an introduction to syntax tree traversal. The node visitors follow a convention where callback names are of pattern visit_NODETYPE(self, node), where node type can be any these. For example, if we want a callback on method calls, we can define one for the Call node and name it visit_Call(self, node).

In our example, we can visit the compare nodes, and print all the operands:

from ast import NodeVisitor


class ComparisonVisitor(NodeVisitor):
    def visit_Compare(self, node):
        operands = [node.left] + node.comparators
        print '; '.join(type(operand).__name__ for operand in operands)

For every callback, we are assured the type of the node fits the Compare node type. Given the type, we can investigate it’s members. Comparison in Python is composed of operators (one or more) and operands (two or more). In the case of Compare node, the first operand is called left, and the rest are called comparators. One of the reason for the complicated structure is to support expressions like:

0 &lt; x &lt; 100

Using the visitor we can query the nodes, but not modify them. If we visit the the original foo function:

&lt;/pre&gt;
node = ast.parse(inspect.getsource(foo))
ComparisonVisitor().visit(node)

The result we expect is:

Name; Num

Since comparison is x < 1, where x is Name load in the context and 1 is a Constant Number in the context.

AST transformers

Python provides transformers, which are a special type of AST visitors. The transformers, in contrast to nodes visitors,  modify the nodes they visit. In our example, we’ll look for nodes that represent comparison between variables and numbers, and then extend them to comply with the aspect.

from ast import NodeTransformer
from _ast import BoolOp, And, Name, Compare, IsNot, Load, Num


class ComparisonTransformer(NodeTransformer):
    def visit_Compare(self, node):
        parts = [node.left] + node.comparators

        # check if any constant number is involved in the comparison
        if not any(isinstance(part, Num) for part in parts):
            return node

        # get all the "variables" involved in the comparison
        names = [element for element in parts if isinstance(element, Name)]
        if len(names) == 0:
            return node

        # create a reference to None
        none = Name(id='None', ctx=Load())
        # create for each variable a node that represents 'var is not None'
        node_verifiers = [Compare(left=name, ops=[IsNot()], comparators=[none]) for name in names]
        # combine the None checks with the original comparison
        # e.g. 'a &lt; b &lt; 1' --&gt; 'a is not None and b is not None and a &lt; b &lt; 1
        return BoolOp(op=And(), values=node_verifiers + [node])

This chunk of code is a simplified (relaxed type input checks no attempts to code location fixes) version of a transformer that visits all nodes of type Compare. The transformer methods names use the same convention as the visitors.

According to the original behavior a new node is being built. This node is a new Boolean expression, which requires all the variables[2] in use to be not None and to satisfy the original comparison.

If we’d look at the output, the the AST will be modified and verify variables are not None before they’re compared to None. The out tree for the modified foo is:

Module
  # the method declaration (foo)
  FunctionDef
    # the arguments list (x)
    arguments
      Name
        Param
    # return instruction
    Return
      # the bool expression that combines with And:
      # 1. the original comparison
      # 2. the new check 'VAR is not None'
      BoolOp
        And
        # the 'x is not None' comparison
        Compare
          Name
            Load
          IsNot
          Name
            Load
        # the original comparison 'x < 1'
        Compare
          Name
            Load
          Lt
          Num

Prepare the node forrecompilation

In the next phase, we’re going to import the new code as temporary module, which will case the declaration of the new method to be executed again. In order to do so, we’d like to remove the rewriter decorator, since we don’t want it to process the modified function. In addition, we rename the function for safety to avoid collisions between the declared function and other locals.  Lastly, we ask python to fix code locations for the new nodes so they can be compiled later on. This is done using fix_missing_locations.

from ast fix_missing_locations


def rewrite_method(node):
    # assuming the method has single decorator (which is the rewriter) - remove it
    node.body[0].decorator_list.pop()
    # we rename the method to ensure separation from the original one.
    # this step has no real meaning and not really required.
    node.body[0].name = 'internal_method'
    # transform Compare nodes to fit the 'is not None' requirement
    ComparisonTransformer().visit(node)
    # let python try and fill code locations for the new elements
    fix_missing_locations(node)

Summary

During the first phase we got as an input a function (through a decorator), then modified it’s body by visiting it’s body using a syntactic level. Lastly, we modified it’s declaration and source locations so it can be safely imported as a new function.

As you probably notice, the only part in this code which is concerned by the aspect is the transformer. Meaning, if we’d like to apply a different aspect the only part which’ll change is the transformer. In our example the ComparisonTransformer is hard-coded for simplicity, but in real solution we’d provide it as an argument to the decorator.

Next phase

In the next phase we’ll use the modified function to generate replacement bytecode.

QuickSelect in CoffeeScript

QuickSelect is a known and simple algorithm for finding the kth smallest element in an array. The advantages of this algorithm are its linear average performance and the constant memory it requires, in addition to simple implementation.

The simplest solution is of course sorting the array using a built-in method of any language and then accessing the kth element in the sorted array. But, it may become an issue when working with large sets – O(n*log n) vs. O(n).

Implementation

class QuickSelect
constructor: (@arr) ->

kth: (k) ->
@_select(0, @arr.length - 1, k)

median: ->
@kth Math.floor(@arr.length / 2)

_swap: (i, j) ->
tmp = @arr[i]
@arr[i] = @arr[j]
@arr[j] = tmp

_partition: (left, right, pivotIndex) ->
pivotValue = @arr[pivotIndex]
@_swap pivotIndex, right
storeIndex = left
for i in [left..right]
if @arr[i] < pivotValue
@_swap storeIndex, i
storeIndex++
@_swap right, storeIndex
return storeIndex

_choose_random_pivot: (left, right) ->
left + Math.floor(Math.random() * (right - left + 1))

_select: (left, right, k) ->
if left == right
return @arr[left]

while true
pivotIndex = @_choose_random_pivot left, right
pivotIndex = @_partition(left, right, pivotIndex)
if k == pivotIndex
return @arr[k]
else if k < pivotIndex
right = pivotIndex - 1
else
left = pivotIndex + 1


Usage examples

describe "QuickSelect", ->
describe "1 item", ->
selector = new QuickSelect [1]
it "Should return the item", -> selector.kth(0).should.equal(1)

describe "2 items", ->
selector = new QuickSelect [2,1]
it "Min should be 1", -> selector.kth(0).should.equal(1)
it "Max should be 2", -> selector.kth(1).should.equal(2)

describe "3 items", ->
selector = new QuickSelect [2,1,3]
it "Min should be 1", -> selector.kth(0).should.equal(1)
it "Second should be 2", -> selector.kth(1).should.equal(2)
it "Max should be 3", -> selector.kth(2).should.equal(3)

These simple tests explains how the access to kth element is done. The tests are written also in CoffeeScript.

Performance comparison

Measuring time for 10,000 arrays, each accessed once at random location:
Array size 5,000 10,000 20,000
QuickSelect 1758ms 3393ms 6912ms
Sort 14302ms 32136ms 70566ms
As can be easily seen, the access to the kth smallest element is much faster. In cases where large arrays are used and repeated accesses are performed, this could be a bottleneck which can be easily solved.

Improving performance using page guards

The problems we’re facing today is, a little bit, unique. Given:

  • n contiguous arrays
  • Each array has m cells
  • Each cell is a Boolean flag

We receive a stream of signals, each signal is an absolute offset from the first array. For each signal we need to set the correct flag AND the first flag of the array. The motivation for setting the first flag is to enable quick filtering of arrays having some flags set.
For example, we have a usage tracking system for n websites and m users. If user i visited website j we’d like to signal that by setting the ith flag in the jth array. After some time, we’d like to query which sites had any visit and who visited them.

The intuitive solution

Assuming you don’t care too much for the performance the solution is straight forward. Whenever setting a flag in an array set also the array in offset 0. If the input is index, then the array index is index / m and the item index is index % m. Pretty simple. For simplicity the source of indexes will be an array named items and the address of the first array will be baseAddress:

for (int i = 0; i < numOfItems; ++i)
{
char* hitAddress = baseAddress + items[i];
*
hitAddress = 1;
char* blockStartAddress = hitAddress - (hitAddress - baseAddress) % dwPageSize;
*
blockStartAddress = 1;
}

It is clear that the first action, *hitAddress = 1, is impossible to avoid. But, what about the set of the signal at index 0? We can replace it with a condition but it is clear it won’t affect much the performance. So, how can we improve that part?

Enabling page guards

Windows provides several memory protections, one of them is the page guard. When allocating a new memory scope we can declare it as protected. Defining it as protected means that each page (page is an arbitrary partition of the memory based on OS page size) will throw an exception on the first access to it. After throwing the exception the protection is removed. We would like to use this mechanism to avoid re-setting the flag at index 0.

In order to define such a scope, we will use the VirtualAlloc method:

VirtualAlloc(NULL, TOTAL_SIZE,
MEM_RESERVE | MEM_COMMIT
PAGE_READWRITE | PAGE_GUARD)

It returns a pointer to the memory scope with size of TOTAL_SIZE in bytes. If a page size P then the new scope has TOTAL_SIZE / P pages.

Tracking page hit

As mentioned, at the first time the memory inside a page is accessed an exception is being thrown. We would like to catch it. In order to do so in the fastest way, we will use windows SetUnhandledExceptionFilter API. The filter is a simple method receiving the exception information and deciding how to treat it. Treating it has three options:

  1. Handling it
  2. Handling it and continue the code execution
  3. Pass the decision to other handler

As a simple filter we can request the runtime to ignore all page guards exceptions:

LONG WINAPI SmartFilter(_EXCEPTION_POINTERS *ep)
{
if (ep->ExceptionRecord->ExceptionCode != STATUS_GUARD_PAGE_VIOLATION)
{
return EXCEPTION_CONTINUE_SEARCH;
}

return EXCEPTION_CONTINUE_EXECUTION;
}

So after setting it as the filter all page guards exceptions will be ignored:

SetUnhandledExceptionFilter(&SmartFilter);

Extending the exceptions filter logic

Let’s assume that all arrays are smaller than the opration system page size and we’ll assume that we don’t care about reserving extra space to pad each array. We’ll denote the page size with dwPageSize.
Now, we can make our SmartFilter really smart. We will add to it the logic for setting the first flag on each array. Assuming baseAddress is some global variable:

LONG WINAPI SmartFilter(_EXCEPTION_POINTERS *ep)
{
if (ep->ExceptionRecord->ExceptionCode != STATUS_GUARD_PAGE_VIOLATION)
{
return EXCEPTION_CONTINUE_SEARCH;
}

char* hitAddress = (char*)ep->ExceptionRecord->ExceptionInformation[1];
char* blockStartAddress = hitAddress - (hitAddress - baseAddress) % dwPageSize;

*
blockStartAddress = 1;

return EXCEPTION_CONTINUE_EXECUTION;
}

We extract the exact address being touched by accessing ep->ExceptionRecord->ExceptionInformation[1]. Through it it’s easy to get the start address of the page. When having this filter method registered we can be sure that whenever we set a flag in the array the first flag will be set too.
Now, we can alter the original code which was in charge of setting the first signal whenever a flag was set:

for (int i = 0; i < numOfItems; ++i)
{
char* hitAddress = baseAddress + items[i];
*
hitAddress = 1;
}

Comparing the results

In order to make our comparison interesting let’s assume that we have 10000 arrays (websites in the tracking system) and each array has 25000 flags (users for example). In order to make it intense we’ll assume that during a short period 10% of the arrays were visited, for exmaple having 250000000 signals sent through the stream (repeating actions in a website by same users are allowed). On average, the time it took to run:

Seconds
Straight forward 1.744
Page guards 0.78

As can easily seen, the page guards solution saves ~50% of the runtime.

Conclusion

The operating system provides a few very fast facilities which can be exploited. Even though most of those facilities are designed for different purposes they can still be useful in different cases, like the one here requiring single time signal for a scope. Since Windows puts a lot of focus in being backward compatible, those exploitations are not too risky. As usual – if it doesn’t require performance optimization, don’t do it. The price of maintaining the code might not worth it.

Conditional attribute and arguments evaluation

What is the conditional attribute?

The conditional attribute enables including/omitting methods calls during compilation depending on compilation symbols. For example, we can condition that a specific log method calls will be included only when we compile in debug. The compiler in this case will omit the calls to the method. Looking at the next code:

public class Logger
{
[
Conditional("DEBUG")]
public void LogDebugMessage(string str)
{

}
}

And the code calling it:

class MyClass
{
private readonly Logger logger = new Logger();

public void Foo()
{
logger
.LogDebugMessage("Foo");
}
}

We expect the compiler to omit the body of Foo(). As we can see with a disassembler this is exactly what happens:

.method public hidebysig instance void  Foo() cil managed
{
// Code size 1 (0x1)
.maxstack 8
IL_0000: ret
} // end of method MyClass::Foo

How method arguments are treated?

Temp variable assignment optimization

Regardless the conditional attribute, in release mode the compiler performs many optimization which one of them is skipping local variable assignment. You’re most likely to notice it when you assign a value into a local variable and pass it to a method as an argument (while this is the only variable usage). For example:

public void Foo()
{
var foo = "Foo";
logger
.LogDebugMessage(foo);
}

Translates into:

0527C028  mov         edx,dword ptr ds:[2EE78B0h]  	// Load the string address
0527C02E mov ecx,dword ptr [ecx+4] // Load the logger instance
0527C031 cmp dword ptr [ecx],ecx // Null check
0527C033 call dword ptr ds:[50C5BD8h] // Call LogDebugMessage
0527C039 ret

While:

public void Foo()
{
logger
.LogDebugMessage("Foo");
}

Translates into:

0552C028  mov         ecx,dword ptr [ecx+4]  		// Load the logger instance
0552C02B mov edx,dword ptr ds:[31478B0h] // Load the string address
0552C031 cmp dword ptr [ecx],ecx // Null check
0552C033 call dword ptr ds:[5375C30h] // Call LogDebugMessage
0552C039 ret

Which are basically the same. So in case we’re not using the conditional attribute we shouldn’t care about local assignment. We can expect to have no difference in runtime.

Temp variable sent to omitted call optimization?

So an interesting question is what happens to an argument we’re about to send to a conditional method? If call to LogDebugMessage are omitted, what should we expect in this case:

public void Foo()
{
var method = MethodBase.GetCurrentMethod().Name;
logger
.LogDebugMessage(method);
}

And in this case:

public void Foo()
{
logger
.LogDebugMessage(MethodBase.GetCurrentMethod().Name);
}

The answer can be easily found by looking at the methods IL. The first version with temp assignment to a variable compiles into:

.method public hidebysig instance void  Foo() cil managed
{
// Code size 12 (0xc)
.maxstack 8
IL_0000: call class [mscorlib]System.Reflection.MethodBase [mscorlib]System.Reflection.MethodBase::GetCurrentMethod()
IL_0005: callvirt instance string [mscorlib]System.Reflection.MemberInfo::get_Name()
IL_000a: pop
IL_000b: ret
} // end of method MyClass::Foo

While the second version compiles into:

.method public hidebysig instance void  Foo() cil managed
{
// Code size 1 (0x1)
.maxstack 8
IL_0000: ret
} // end of method MyClass::Foo

As we can see, in this case the argument was not even evaluated and the whole statement was omitted from the IL. Meaning that in this case, inlining the variable would have influence on the performance. It didn’t happen by chance, this is the defined behavior of the compiler as stated in the Conditional attribute documentation:

“If the symbol is defined, the call is included; otherwise, the call (including evaluation of the parameters of the call) is omitted.”

Conclusion

The most common scenario in which the conditional attribute is involved is logging. Since the main advantage of omitting the logs is usually to avoid performance hit in production it is important to take into consideration the price of evaluating the arguments values. The simplest solution is to inline the variable. This can be done easily when the argument is string.Format() or similar. In case it is more complicated or unreadable it can always be solved by preprocessor directive such as #if.