Wilco (Wilco.Dijkstra.delete@this.ntlworld.com) on November 21, 2012 7:44 am wrote:
> Install-time compilation would be far better than using a JIT indeed. But you still may not have
> enough memory/performance for link-time whole program optimization with profile feedback.

On desktops the time consuming compilation is done at background, in phones it can be done during recharge time, no?

Oh, and with the most up to date and relevant profile feedback.

> Rubbish. GC is memory inefficient by definition, so claiming it is better for locality is just wishful
> thinking. Compacting GC's typically need 2-3 times more memory than a non-compacting GC, so are worse on
> average.

Your claims don't hold true...

Every allocator use more memory than strictly necessary (later about C++), GC does the job of moving data toghether from time to time wich negates fragmentation.

> Also the much higher memory allocation rate and resulting collections are bad for locality.

Allocations are cheap, collecting short lived objects is cheap too.

> Concurrent GC has even larger overheads. It stops threads for shorter periods but stops them more often, so
> takes far longer overall. And then we haven't considered the far higher overheads on the generated code.

I don't know about what GC you are talking...

The concurrent GC I have in mind originated from the non-concurrent GC, the most recent generation collections still blocking wich the full GC running in background until it needs blocking, so there is the same numbers of full stops but the most time consuming stop is shorter.

Yes, it uses more resources, but on s single-threaded application such resources does exist, and wouldn't be used in another way.

> No, no space is wasted, unlike GC which requires descriptors for every object.

Descriptors aren't strictly necessary and can be encoded in the vtable pointer, the reason many VMs doesn't omit the descriptor is because it is not that expensive.

But C++ allocators waste space, for example, on Windows, before LFH deallocated objects left holes in the heap, some wich were left unused for lack of a suitable object, a problem called "fragmentation" and it is accumalative, the longer the app runs the worse it gets, see the wasted space? Locality isn't good either, objects go were there is space, not near related objects.

LFH allocates 64k for every object size rounded to lowest higher power of 2, wich, by definition waste the space between that power of 2 and the actual object size, and, if there is not enough objects of a given size to fill the segment, it also waste the segment unused space. Locality suffers as well since objects are placed by it's size and related objects may end very far from each other.

> Instructions that can trap are bad. That's why you see modern FPUs
> implement IEEE so you never need traps, not even for denormals.

Ok, instructions that may trap, loads, store, branches and integer division...

On a normal applications it is about half the instructions, the reason modern FPU don't need to trap is mostly because having to handle an exception in normal flow is a pain in the access, exceptions are supposed to be exceptions, not part of normal flow.

> Since when is a memory access cheap? Every unnecesary instruction has a cost.

Sure it have a cost, a non-blocking load will consume one decoder, one port for one cycle and one slot in a modern processor, wich will likely pass without being noticed, so it is a small cost, that's why I called it cheap.

> The barriers and other checks for concurrent GC or multithreaded access
> to fields are not exactly zero-cost and block many optimizations.

Maybe that's why the "volatile" keyword exists?

Anyway, MS doesn't implement the weak memory model that the C# or C++ specs allow, many applications wouldn't work if they did.

> By your logic, C++ would be much slower than C as C++ is quite new. Clearly that's not the case - and the
> simple reason is that much of what goes on inside is unrelated to the source language or the target.

No, that's not my logic, reread my statement.

> Not sure what your point is here. Yes VC++ does extremely well as it is optimized
> to do well on the millions of lines of code that people actually write (such as Windows
> and its apps) - as opposed to showing huge gains on a few small benchmarks.

.Net code generation isn't all that different from VC++.

Did you read when I wrote VC++ is conservative?

> Wrong. Exceptions can only happen explicitly with throw in C++.

Or when the program hits an undefined behaviour...

And there are quite a few undefined behaviours in C++ spec.

BTW, I am not limiting myself to C++ exceptions, any hardware exception is valid, VMs may handle exceptions by traping the hardware exceptions, if there is no try...catch block no need to care about order or barries, the program will crash anyway... Like it is done in C++.

> The problem is not just the ordering, but the fact that more operations can cause exceptions. That alone
> creates a lot of overhead as you need to model flows from every possible exception to all possible exception
> handlers. Local variable values need to be preserved for example, severely limiting optimizations.

If there is a registered handler...

And the implementation can, for example, let the heavy lifting of restoring state for the exception handler, giving more freedom for optimizations in the normal flow.

> The performance cost is high if you happen to use arrays a lot.

And if the optimizer fails to remove the redundant checks.