Executable Bloat - How it happens and how we can ﬁght it

Executable Bloat
How it happens and how we can ﬁght it

Andreas Fredriksson <dep@dice.se>

Thursday, February 24, 2011

Background
• Why is ELF/EXE size a problem?
• Executable code takes up memory
• Consoles have precious little RAM
• More available memory enables richer
gameplay
• It is a lot of work to ﬁx after the fact

Software Bloat
Software bloat is a term used to describe the
tendency of newer computer programs to have a
larger installation footprint, or have many
unnecessary features that are not used by end
users, or just generally use more system resources
than necessary, while offering little or no beneﬁt
to its users.

• Wikipedia has this deﬁnition (emphasis mine)
• We are wasting memory we could be using
to make better games


C++ Bloat

• Let’s look at a few typical C++ techniques
• Some are even enforced in coding
standards
• Often recommended in literature
• Often accepted without question


Interface Class Hiding
• Hiding single class Foo struct IFoo {
virtual void func() = 0;
behind interface IFoo };

• Intent is to avoid class Foo : public IFoo {

};
virtual void func();

including heavy Foo
header everywhere
• Sometimes used for
PRX/DLL loading


struct IFoo { 0 IFoo vtbl
virtual void a() = 0; typeinfo_ptr
virtual void b() = 0; pure_vcall
}; pure_vcall
class Foo : public IFoo {
virtual void a();
virtual void b(); 0
Foo vtbl
}; typeinfo_ptr
Foo_a
Foo_b

• One hidden cost - virtual tables
• One for each class
• In PPU ABI cost is higher as Foo_a will be a
pointer to a 4+4 byte ABI struct


• Additional overhead
• Every call site needs vcall adjustment
• ctor/dtor needs vptr adjustment
• Total SNC overhead, 8 functions: 528 bytes
• Likely more, due to callsites
• These bytes have no value = bloat

Excessive Inlining

• Typically associated with templates
• Templates are almost always inline
• Smart pointers
• String classes


Excessive Inlining
bool eastl_string_find(eastl::string& str) {
! return str.find("foo") != eastl::string::npos;
}
bool c_string_find(const char* str) {
! return strstr(str, "foo") != 0;
}

• Compare two string searches
• c_string_ﬁnd version - 56 bytes (SNC)
• eastl_string_ﬁnd - 504 bytes (SNC)
• These bytes add zero value = bloat

Excessive Inlining
• Many other common inlining culprits
• operator+= string concatenation
• Call by value w/ inline ctor
• Hard to control via compiler options
• Sometimes you want inlining
• Better to avoid pointless cases

Static Initialization
static const eastl::string strings[] = { "foo0", "foo1" };
static const Vec3 vecs[] = { { 0.2, 0.3, 0.5 }, { ... } };

• Extreme hidden costs for constructors
• Generates initializer function
• Have seen arrays of this kind generate over 20
kb of initializer code in the wild (SNC)
• Array of 10 eastl::strings - 1292 bytes
• Array of 10 vec3 - 368 bytes


Static Initialization
static const char* strings[] = { "foo0", "foo1", ... };
static const float vecs[][3] = { { 0.2, 0.3, 0.5 }, ... };

• Just don’t do it - prefer POD types
• Make sure data ends up in .rodata segment
• Adjust code using array accordingly
• Alternatively make data load with level
• No space overhead when not used

operator<<
• A special case of excessive inlining
• Creeps up in formatted I/O
• Assert macros
• Prefer snprintf()-style APIs if you must
format text at runtime
• Usually less than half the overhead
• Ideally avoid text formatting altogether

Sorting
• STL sort is a bloat generator
• Specialized for each type - faster
compares..
• ..but usually whole merge/quicksort
duplicated per parameter type! - often
around 1-2kb code
• We have 140 sort calls in the code base - up
to 280 kb overhead..


Sorting
• Prefer qsort() for small/medium datasets
• Adds callback overhead on PPU..
• Rule of thumb - qsort < 32-64 elements
• Same applies to many other template
algorithms
• Use only when it really buys something

Part 2:
What you can do


Accept the Domain
• Console coding is a very speciﬁc problem space
• Think and verify before you apply general
desktop C++ advice or patterns
• Bloat is caused by humans, not compilers
• Example: "Virtual functions are essentially free"
• True on x86 architecture (most of the time)
• On PS3 PPU often two cache misses - ~1200
cycle penalty + ELF size bloat already covered


Day to day
• Think about full impact of your changes
• .text size impact
• .data/.rodata size impact
• Bring it up & discuss in code reviews
• Make sure your code is reasonably sized for
the problem it solves!


Avoid “reuse” bloat
• YAGNI - “You ain’t gonna need it”
• Just write simple code that can be extended if
needed
• We typically never extend systems without
altering their interfaces anyway
• Game code is disposable, a means to an end
• Make sure it works well NOW
• Avoid “single-feature frameworks”


Avoid repetition
• Can often move repeated code to data
• Higher information density but same end result
RegisterFunc("foo", func_foo);
RegisterFunc("bar", func_bar);
RegisterFunc("baz", func_baz);
// ...

static const struct {
const char *name; void (*func)(void);
} opdata[] = {
! { "foo", func_foo },
! { "bar", func_bar },
! { "baz", func_baz },
// ...
};

for (int i=0; i < sizeof_array(opdata); ++i)
RegisterFunc(opdata[i].name, opdata[i].func);


Compiler Output
• Look at the generated code!
• That’s what you’re checking in, not C++
• Don’t assume code is improved by the
compiler
• No magic going on, compilers are stupid
• Develop an intuition for what to expect
• Verify assumptions as you code

Assembly
• Learn enough assembly to read compiler
output
• Function calls (calling convention)
• Memory loads and stores
• FP/Vector instructions
• It’s not very difﬁcult - just do it
• Also improves your debugging skills

Guidelines
• Avoid string classes, concatenation
• Excessive inlining
• Avoid template containers for simple
problems
• Inlining + instantiation cost
• Prefer C arrays for most jobs

Guidelines
• Avoid complex types in function signatures
and interfaces
• Requires caller to jump through hoops
• Often bloats all call sites
• Prefer raw POD types
• (T* ptr, int count) is better
than (std::vector<T>&)


Guidelines
• Avoid inheritance, interfaces and virtual
functions
• Hidden costs are subtle
• Prefer function pointers for callbacks
• Prefer free functions on predeclared
types for header stripping


Guidelines
• Avoid operator<< streaming
• Prefer printf() style APIs
• Easy to make your own formatter for often-
used types
• Avoid singletons
• They just add bloat around data that's just as
global anyway
• Prefer free functions around static data


Summary
• Make sure the code/data you’re adding is
reasonably sized for the problem it solves
• Use no more than necessary
• Pick up some assembly and look at the
compiler output
• Always measure, examine & question!


Questions?
Twitter: @deplinenoise
Email: <dep@dice.se>


Executable Bloat - How it happens and how we can ﬁght it

More Related Content

What's hot (20)

Viewers also liked (20)

Similar to Executable Bloat - How it happens and how we can ﬁght it (20)

More from Electronic Arts / DICE (20)

Executable Bloat - How it happens and how we can ﬁght it