blind micro optimisation
-
on a code review someone commented I was enumerating twice a really big list, could I have only one loop instead, they said initial code, where
sourceis a large listlist.AddRange(source.OfType<A>());
DoThings(source.OfType<B>());void DoThings(IEnumerable<B> items)
{
foreach (var b in item)
{
// do something
}
}so to please the crowd I refactored new code
foreach (var item in source)
{
if (item is A a)
list.Add(a);
if (item is B b)
DoThing(b);
}
void DoThings(IEnumerable<B> items) // still in use through an other codepath
{
foreach (var b in item)
DoThing(b);
}
void DoThing(B b)
{
// do something
}now... without doing any measurement (hence the blind adjective) I am suspecting version 2 is in fact slower, since enumerating is cheap, but there a lot more method calls. What says you?
A new .NET Serializer All in one Menu-Ribbon Bar Taking over the world since 1371!
I would measure them to be sure. But I agree that calling
DoThingsonce is likely to be better than callingDoThinga great many times. The particular micro-optimization I'm working on now will likely lead to a post in Weird and Wonderful. -
on a code review someone commented I was enumerating twice a really big list, could I have only one loop instead, they said initial code, where
sourceis a large listlist.AddRange(source.OfType<A>());
DoThings(source.OfType<B>());void DoThings(IEnumerable<B> items)
{
foreach (var b in item)
{
// do something
}
}so to please the crowd I refactored new code
foreach (var item in source)
{
if (item is A a)
list.Add(a);
if (item is B b)
DoThing(b);
}
void DoThings(IEnumerable<B> items) // still in use through an other codepath
{
foreach (var b in item)
DoThing(b);
}
void DoThing(B b)
{
// do something
}now... without doing any measurement (hence the blind adjective) I am suspecting version 2 is in fact slower, since enumerating is cheap, but there a lot more method calls. What says you?
A new .NET Serializer All in one Menu-Ribbon Bar Taking over the world since 1371!
I'm guessing the refactor is faster since it only has to enum once.
To err is human. Fortune favors the monsters.
-
on a code review someone commented I was enumerating twice a really big list, could I have only one loop instead, they said initial code, where
sourceis a large listlist.AddRange(source.OfType<A>());
DoThings(source.OfType<B>());void DoThings(IEnumerable<B> items)
{
foreach (var b in item)
{
// do something
}
}so to please the crowd I refactored new code
foreach (var item in source)
{
if (item is A a)
list.Add(a);
if (item is B b)
DoThing(b);
}
void DoThings(IEnumerable<B> items) // still in use through an other codepath
{
foreach (var b in item)
DoThing(b);
}
void DoThing(B b)
{
// do something
}now... without doing any measurement (hence the blind adjective) I am suspecting version 2 is in fact slower, since enumerating is cheap, but there a lot more method calls. What says you?
A new .NET Serializer All in one Menu-Ribbon Bar Taking over the world since 1371!
I'd argue that you really need two lists, one of Type[A] and one of Type[B] (sorry, couldn't figure out how to avoid font changes with the <>). This appears to really be an issue of delaying the decision until too deep in the code's decision trees. The new code is easier to read and it really points out that your list should be two lists.
-
on a code review someone commented I was enumerating twice a really big list, could I have only one loop instead, they said initial code, where
sourceis a large listlist.AddRange(source.OfType<A>());
DoThings(source.OfType<B>());void DoThings(IEnumerable<B> items)
{
foreach (var b in item)
{
// do something
}
}so to please the crowd I refactored new code
foreach (var item in source)
{
if (item is A a)
list.Add(a);
if (item is B b)
DoThing(b);
}
void DoThings(IEnumerable<B> items) // still in use through an other codepath
{
foreach (var b in item)
DoThing(b);
}
void DoThing(B b)
{
// do something
}now... without doing any measurement (hence the blind adjective) I am suspecting version 2 is in fact slower, since enumerating is cheap, but there a lot more method calls. What says you?
A new .NET Serializer All in one Menu-Ribbon Bar Taking over the world since 1371!
To me, this is a clear case of premature optimization and (almost) every optimization is premature if it sacrifices clarity when the need for optimization has not be demonstrated. There are certain cases involving function calls that I know for a fact can be done better so I deal with those up front rather than later since the delta time spent is zero. Here's an exmple :
x*xis considerably faster thanpow(x,2)so I deal with that up front instead of returning to it later. Especially since calling thepowfunction does not enhance clarity."They have a consciousness, they have a life, they have a soul! Damn you! Let the rabbits wear glasses! Save our brothers! Can I get an amen?"
-
on a code review someone commented I was enumerating twice a really big list, could I have only one loop instead, they said initial code, where
sourceis a large listlist.AddRange(source.OfType<A>());
DoThings(source.OfType<B>());void DoThings(IEnumerable<B> items)
{
foreach (var b in item)
{
// do something
}
}so to please the crowd I refactored new code
foreach (var item in source)
{
if (item is A a)
list.Add(a);
if (item is B b)
DoThing(b);
}
void DoThings(IEnumerable<B> items) // still in use through an other codepath
{
foreach (var b in item)
DoThing(b);
}
void DoThing(B b)
{
// do something
}now... without doing any measurement (hence the blind adjective) I am suspecting version 2 is in fact slower, since enumerating is cheap, but there a lot more method calls. What says you?
A new .NET Serializer All in one Menu-Ribbon Bar Taking over the world since 1371!
-
I'm guessing the refactor is faster since it only has to enum once.
To err is human. Fortune favors the monsters.
but it call a method for every item..... I was unsure, or maybe thought it was going to be a bit in fact slower (extra method call, where enumeration is cheap)... Anyway, I was so annoyed I made a simple test program at home and.. turns out new method is faster indeed (by 20% on my simple test thing) After that, well, I beautify the new code a bit and rolled with it! :)
A new .NET Serializer All in one Menu-Ribbon Bar Taking over the world since 1371!
-
I would measure them to be sure. But I agree that calling
DoThingsonce is likely to be better than callingDoThinga great many times. The particular micro-optimization I'm working on now will likely lead to a post in Weird and Wonderful.well.. I did think that too.... then I did a simple test home and, well, to my stunned surprise the optimisation was 20% faster!
A new .NET Serializer All in one Menu-Ribbon Bar Taking over the world since 1371!
-
but it call a method for every item..... I was unsure, or maybe thought it was going to be a bit in fact slower (extra method call, where enumeration is cheap)... Anyway, I was so annoyed I made a simple test program at home and.. turns out new method is faster indeed (by 20% on my simple test thing) After that, well, I beautify the new code a bit and rolled with it! :)
A new .NET Serializer All in one Menu-Ribbon Bar Taking over the world since 1371!
Looking at the code I would expect the second to be faster. The reason is that enumeration is slow relative to function calls. The original code enumerates three times over source and creates temporary objects. The rewritten code enumerates once over source and creates no temporary objects. Super Lloyd actually tested and found the second code is 20% faster. [The Lounge](https://www.codeproject.com/Lounge.aspx?msg=5880319#xx5880319xx)
-
on a code review someone commented I was enumerating twice a really big list, could I have only one loop instead, they said initial code, where
sourceis a large listlist.AddRange(source.OfType<A>());
DoThings(source.OfType<B>());void DoThings(IEnumerable<B> items)
{
foreach (var b in item)
{
// do something
}
}so to please the crowd I refactored new code
foreach (var item in source)
{
if (item is A a)
list.Add(a);
if (item is B b)
DoThing(b);
}
void DoThings(IEnumerable<B> items) // still in use through an other codepath
{
foreach (var b in item)
DoThing(b);
}
void DoThing(B b)
{
// do something
}now... without doing any measurement (hence the blind adjective) I am suspecting version 2 is in fact slower, since enumerating is cheap, but there a lot more method calls. What says you?
A new .NET Serializer All in one Menu-Ribbon Bar Taking over the world since 1371!
Performance is not the main reason to avoid multiple enumeration. Depending on the origin of the source, the collection may be changed between any two enumerations, so the result will be a bug that is extremely hard to reproduce.
-
but it call a method for every item..... I was unsure, or maybe thought it was going to be a bit in fact slower (extra method call, where enumeration is cheap)... Anyway, I was so annoyed I made a simple test program at home and.. turns out new method is faster indeed (by 20% on my simple test thing) After that, well, I beautify the new code a bit and rolled with it! :)
A new .NET Serializer All in one Menu-Ribbon Bar Taking over the world since 1371!
In my experience with codegen method calls don't require much overhead. I haven't run any hard tests directly, but I have written a codegen system that did method calls to separate the non-terminals and one that did not. The one that did was significantly faster, I suspect due to keeping repeatedly used code cached at the CPU level whereas with longer methods that's not possible. That's what leads me to lean toward what I leaned toward. I suspect JIT magic to reduce overhead, but there's not much you can do to with JIT to avoid allocation overhead. With .NET allocation is cheap in theory, but people forget about the overhead of the constructors.
To err is human. Fortune favors the monsters.
-
In my experience with codegen method calls don't require much overhead. I haven't run any hard tests directly, but I have written a codegen system that did method calls to separate the non-terminals and one that did not. The one that did was significantly faster, I suspect due to keeping repeatedly used code cached at the CPU level whereas with longer methods that's not possible. That's what leads me to lean toward what I leaned toward. I suspect JIT magic to reduce overhead, but there's not much you can do to with JIT to avoid allocation overhead. With .NET allocation is cheap in theory, but people forget about the overhead of the constructors.
To err is human. Fortune favors the monsters.
nice hindsight! :)
A new .NET Serializer All in one Menu-Ribbon Bar Taking over the world since 1371!
-
nice hindsight! :)
A new .NET Serializer All in one Menu-Ribbon Bar Taking over the world since 1371!
You know what you might try, if the container is indexable (your code is way up top of the thread and I don't recall it offhand) is do the loops without the enumerator.
for(int i = 0; i < container.Count; ++i) { ...
Like that. My bet is your results will be more similar at that point, but if i had to guess which one would be faster, I still suspect the method call one will be faster due to the overhead of running loops.
To err is human. Fortune favors the monsters.
