Re: D Language Foundation October 2023 Quarterly Meeting Summary
On Wednesday, 6 December 2023 at 16:28:08 UTC, Mike Parker wrote: ## The Next Meetings We had our October monthly meeting one week after this meeting. The next quarterly should happen on January 5, 2024. We had no regular planning sessions in October, but two workgroup meetings took place regarding DMD-as-a-library. The monthly meeting summary is coming next, then I'll publish an update about the workgroup meetings. https://monkyyyscience.substack.com/p/d-data-structures please add this to the agenda
Re: D Language Foundation October 2023 Quarterly Meeting Summary
On Monday, 11 December 2023 at 19:55:38 UTC, Timon Gehr wrote: There is the following trick. Not ideal since the length cannot be inferred, but this successfully injects alloca into the caller's scope. Wow, what a great hack - I'd have never came up with that!
Re: D Language Foundation October 2023 Quarterly Meeting Summary
On Monday, 11 December 2023 at 19:55:38 UTC, Timon Gehr wrote: ... this successfully injects alloca into the caller's scope. ```d import core.stdc.stdlib:alloca; import std.range:ElementType; import core.lifetime:moveEmplace; struct VLA(T,alias len){ T[] storage; this(R)(R initializer,return void[] storage=alloca(len*T.sizeof)[0..len*T.sizeof]){ this.storage=cast(T[])storage; foreach(ref element;this.storage){ assert(!initializer.empty); auto init=initializer.front; moveEmplace!T(init,element); initializer.popFront(); } } ref T opIndex(size_t i)return{ return storage[i]; } T[] opSlice()return{ return storage; } } auto vla(alias len,R)(R initializer,void[] storage=alloca(len*ElementType!R.sizeof)[0..len*ElementType!R.sizeof]){ return VLA!(ElementType!R,len)(initializer,storage); } void main(){ import std.stdio,std.string,std.conv,std.range; int x=readln.strip.to!int; writeln(vla!x(2.repeat(x))[]); } ``` You guys are great!
Re: D Language Foundation October 2023 Quarterly Meeting Summary
On Monday, 11 December 2023 at 08:24:55 UTC, Bastiaan Veelo wrote: On Sunday, 10 December 2023 at 22:59:06 UTC, Nicholas Wilson wrote: Or you could use grep with `--output-ll` as noted by Johan https://github.com/ldc-developers/ldc/issues/4265#issuecomment-1376424944 although this will be with that `workaroundIssue1356` applied. Thanks for highlighting this, as I must have forgotten. I should be able to create a CI job that checks this as part of the release. This will give us the confidence that we need. I should note that regex will need some updating for the most recent LLVMs that have opaque pointers enabled: `ptr byval\(%[a-zA-Z_][a-zA-Z0-9_\.]*\) align`
Re: D Language Foundation October 2023 Quarterly Meeting Summary
On Monday, 11 December 2023 at 22:04:34 UTC, Nicholas Wilson wrote: And please do get in touch with Bruce Carneal if you want some tips and insight with the practical and applied side of dcompute (also with auto-vectorisation) as he has used it a lot more than I have. dcompute needs some love: https://github.com/libmir/dcompute/pull/74 Cheers, I look forward to some large speed increase reports. it will be amazing to see such reports
Re: D Language Foundation October 2023 Quarterly Meeting Summary
On Monday, 11 December 2023 at 08:24:55 UTC, Bastiaan Veelo wrote: On Sunday, 10 December 2023 at 22:59:06 UTC, Nicholas Wilson wrote: Always happy to help if you're interested in looking into using dcompute. Thank you, I'll let you know! And please do get in touch with Bruce Carneal if you want some tips and insight with the practical and applied side of dcompute (also with auto-vectorisation) as he has used it a lot more than I have. Or you could use grep with `--output-ll` as noted by Johan https://github.com/ldc-developers/ldc/issues/4265#issuecomment-1376424944 although this will be with that `workaroundIssue1356` applied. Thanks for highlighting this, as I must have forgotten. I should be able to create a CI job that checks this as part of the release. This will give us the confidence that we need. -- Bastiaan. Cheers, I look forward to some large speed increase reports.
Re: D Language Foundation October 2023 Quarterly Meeting Summary
On 12/11/23 20:55, Timon Gehr wrote: There is the following trick. Not ideal since the length cannot be inferred, but this successfully injects alloca into the caller's scope. I see Nick already brought it up.
Re: D Language Foundation October 2023 Quarterly Meeting Summary
On 12/6/23 17:28, Mike Parker wrote: One way to do that in D is to use `alloca`, but that's an issue because the memory it allocates has to be used in the same function that calls the `alloca`. So you can't, e.g., use `alloca` to alloc memory in a constructor, and that prevents using it in a custom array implementation. He couldn't think of a way to translate it. There is the following trick. Not ideal since the length cannot be inferred, but this successfully injects alloca into the caller's scope. ```d import core.stdc.stdlib:alloca; import std.range:ElementType; import core.lifetime:moveEmplace; struct VLA(T,alias len){ T[] storage; this(R)(R initializer,return void[] storage=alloca(len*T.sizeof)[0..len*T.sizeof]){ this.storage=cast(T[])storage; foreach(ref element;this.storage){ assert(!initializer.empty); auto init=initializer.front; moveEmplace!T(init,element); initializer.popFront(); } } ref T opIndex(size_t i)return{ return storage[i]; } T[] opSlice()return{ return storage; } } auto vla(alias len,R)(R initializer,void[] storage=alloca(len*ElementType!R.sizeof)[0..len*ElementType!R.sizeof]){ return VLA!(ElementType!R,len)(initializer,storage); } void main(){ import std.stdio,std.string,std.conv,std.range; int x=readln.strip.to!int; writeln(vla!x(2.repeat(x))[]); } ```
Re: D Language Foundation October 2023 Quarterly Meeting Summary
On Sunday, 10 December 2023 at 15:08:05 UTC, Bastiaan Veelo wrote: We are looking forward to being able to safely use LDC, because tests show that it has the potential to at least double the performance. Yes, and that's before you its excellent SIMD capabilities :)
Re: D Language Foundation October 2023 Quarterly Meeting Summary
On Sunday, 10 December 2023 at 22:59:06 UTC, Nicholas Wilson wrote: Always happy to help if you're interested in looking into using dcompute. Thank you, I'll let you know! Or you could use grep with `--output-ll` as noted by Johan https://github.com/ldc-developers/ldc/issues/4265#issuecomment-1376424944 although this will be with that `workaroundIssue1356` applied. Thanks for highlighting this, as I must have forgotten. I should be able to create a CI job that checks this as part of the release. This will give us the confidence that we need. -- Bastiaan.
Re: D Language Foundation October 2023 Quarterly Meeting Summary
On Sunday, 10 December 2023 at 18:16:05 UTC, Nick Treleaven wrote: You can call `alloca` as a default argument to a function. The memory will be allocated on the caller's stack before calling the function: https://github.com/ntrel/stuff/blob/master/util.d#L113C1-L131C2 I've just tested and it seems it works as a constructor default argument too. Clever!
Re: D Language Foundation October 2023 Quarterly Meeting Summary
On Sunday, 10 December 2023 at 16:08:45 UTC, Bastiaan Veelo wrote: On Sunday, 10 December 2023 at 15:31:55 UTC, Richard (Rikki) Andrew Cattermole wrote: It will be interesting to hear how dcompute will fare in your situation, due to it being D code it should be an incremental improvement once you're ready to move to D fully. Yes, dcompute could mean another leap forward. There are so many great things to look forward to. -- Bastiaan. Always happy to help if you're interested in looking into using dcompute. I can't remember if we've talked about it before, but if you were wanting to use it you'd need OpenCL 2.x (explicitly the 2.x version series, or make sure the 3.x implementation supports SPIRV) running on that 20 logical core box (or if it has GPUs attached to it, CUDA (any version should do) for NVidia GPUs or OpenCL 2.x (as above) on any other GPUs). With regards to the stack corruption there is https://github.com/ldc-developers/ldc/blob/master/gen/abi/x86.cpp#L260 which has been there for some time. It would be fairly simple to issue a diagnostic there (although getting source location from there might be a bit tricky) for when there is both a `byval` and an alignment specified. Or you could use grep with `--output-ll` as noted by Johan https://github.com/ldc-developers/ldc/issues/4265#issuecomment-1376424944 although this will be with that `workaroundIssue1356` applied.
Re: D Language Foundation October 2023 Quarterly Meeting Summary
On Wednesday, 6 December 2023 at 16:28:08 UTC, Mike Parker wrote: One way to do that in D is to use `alloca`, but that's an issue because the memory it allocates has to be used in the same function that calls the `alloca`. So you can't, e.g., use `alloca` to alloc memory in a constructor, and that prevents using it in a custom array implementation. You can call `alloca` as a default argument to a function. The memory will be allocated on the caller's stack before calling the function: https://github.com/ntrel/stuff/blob/master/util.d#L113C1-L131C2 I've just tested and it seems it works as a constructor default argument too.
Re: D Language Foundation October 2023 Quarterly Meeting Summary
On Sunday, 10 December 2023 at 17:11:04 UTC, Siarhei Siamashka wrote: On Sunday, 10 December 2023 at 15:08:05 UTC, Bastiaan Veelo wrote: The compiler can check if `scope` delegates escape a function, but it only does this in `@safe` code --- and our code is long from being `@safe`. So it was a bit of a puzzle to find out which arguments needed to be `scope` and which arguments couldn't be `scope`. This reminded me of https://forum.dlang.org/thread/myiqlzkghnnyykbyk...@forum.dlang.org LDC has a special GC2Stack IR optimization pass, which is a lifesaver in many cases like this. Interesting. Are there some known blocker bugs, which prevent a safe usage of LDC in production? This one: https://github.com/ldc-developers/ldc/issues/4265 Mike has summarized it: LDC unfortunately had an issue that caused stack corruption on 32-bit Windows. They'd hit it in one case and were able to work around it, but he couldn't be sure they wouldn't hit it somewhere else. He wasn't willing to risk unreliable computations. He said that LDC could do the right thing, but his understanding from talking to Martin was that implementing it would have a large time cost. Since Win32 is going to eventually go away, he wasn't very keen on paying that cost. They'd spoken at DConf about the possibility of LDC raising compilation errors when stack corruption could occur so that they could then work around those cases, but he hadn't followed up with Martin about it. -- Bastiaan.
Re: D Language Foundation October 2023 Quarterly Meeting Summary
On Sunday, 10 December 2023 at 15:08:05 UTC, Bastiaan Veelo wrote: 1) Missing `scope` storage class specifiers on `delegate` function arguments. This can be chalked down as a beginner error, but also one that is easy to miss. If you didn't know: without `scope` the compiler cannot be sure that the delegate is not stored in some variable that has a longer lifetime than the stack frame of the (nested) function pointed to by the delegate. Therefore, a dynamic closure is created, which means that the stack is copied to new GC-allocated memory. In the majority of our cases, delegate arguments are simple callbacks that are only stored on the stack, but a select number of delegates in the GUI are stored for longer. The compiler can check if `scope` delegates escape a function, but it only does this in `@safe` code --- and our code is long from being `@safe`. So it was a bit of a puzzle to find out which arguments needed to be `scope` and which arguments couldn't be `scope`. This reminded me of https://forum.dlang.org/thread/myiqlzkghnnyykbyk...@forum.dlang.org LDC has a special GC2Stack IR optimization pass, which is a lifesaver in many cases like this. So now all cores are finally under full load, which is a magnificent sight! Speed of DMD `release-nobounds` is on par with our Pascal version, if not slightly faster. We are looking forward to being able to safely use LDC, because tests show that it has the potential to at least double the performance. Are there some known blocker bugs, which prevent a safe usage of LDC in production?
Re: D Language Foundation October 2023 Quarterly Meeting Summary
On Sunday, 10 December 2023 at 15:31:55 UTC, Richard (Rikki) Andrew Cattermole wrote: It will be interesting to hear how dcompute will fare in your situation, due to it being D code it should be an incremental improvement once you're ready to move to D fully. Yes, dcompute could mean another leap forward. There are so many great things to look forward to. -- Bastiaan.
Re: D Language Foundation October 2023 Quarterly Meeting Summary
That is awesome to hear! If the move towards ldc has the potential to half your run time, that is quite a significant improvement for your customers. It will be interesting to hear how dcompute will fare in your situation, due to it being D code it should be an incremental improvement once you're ready to move to D fully. Based upon the estimates here already, it seems like acquiring an LDC developer in house might be well worth it.
Re: D Language Foundation October 2023 Quarterly Meeting Summary
On Wednesday, 6 December 2023 at 16:28:08 UTC, Mike Parker wrote: Bastiaan reported that SARC had been testing their D codebase (transpiled from Pascal---[see Bastiaan's DConf 2019 talk](https://youtu.be/HvunD0ZJqiA)). They'd found the multithreaded performance worse than the Pascal version. He said that execution time increased with more threads and that it didn't matter how many threads you throw at it. It's the latter problem he was focused on at the moment. I have an update on this issue. But first let me clarify how grave this situation is (was!) for us. There are certain tasks that we, and our customers, need to perform that involves a 20 logical core computer to crunch numbers for a week. This is painful, but it also means that a doubling of that time is completely unacceptable, let alone a 20-fold increase. It is the difference between in business and out of business. Aside from the allocation issue, there are several other properties that our array implementation needs to replicate from Extended Pascal: being able to have non-0 starting indices, having value semantics, having array limits that can be compile-time and run-time, and function arguments that must work on arrays of any limits, also for multi-dimensional arrays. So while trying to solve one aspect, care had to be taken not to break any of the other aspects. It turned out that thread contention had more than one causes, which made this an extra frustrating problem because just as we thought to have found the culprit, it did not have the effect that we expected. These were the three major reasons we were seeing large thread contention, in no particular order: 1) Missing `scope` storage class specifiers on `delegate` function arguments. This can be chalked down as a beginner error, but also one that is easy to miss. If you didn't know: without `scope` the compiler cannot be sure that the delegate is not stored in some variable that has a longer lifetime than the stack frame of the (nested) function pointed to by the delegate. Therefore, a dynamic closure is created, which means that the stack is copied to new GC-allocated memory. In the majority of our cases, delegate arguments are simple callbacks that are only stored on the stack, but a select number of delegates in the GUI are stored for longer. The compiler can check if `scope` delegates escape a function, but it only does this in `@safe` code --- and our code is long from being `@safe`. So it was a bit of a puzzle to find out which arguments needed to be `scope` and which arguments couldn't be `scope`. 2) Allocating heap memory in the array implementation, as discussed in the meeting. We followed Walter's advice and now use `alloca`. Not directly, but using string mixin's and static member functions that generate the appropriate code. 3) Stale calls to `GC.addRange` and `GC.removeRange`. These were left over from an experiment where we tried to circumvent the garbage collector. Without knowing these were still in there, we were puzzled because we even saw contention in code that was marked `@nogc`. It makes sense now, because even though `addRange` doesn't allocate, it does need the global GC lock to register the range safely. Because the stack is already scanned by default, these calls were now superfluous and could be removed. So now all cores are finally under full load, which is a magnificent sight! Speed of DMD `release-nobounds` is on par with our Pascal version, if not slightly faster. We are looking forward to being able to safely use LDC, because tests show that it has the potential to at least double the performance. A big sigh of relief from us as we have solved the biggest hurdle (hopefully!) on our way to full adoption of D. -- Bastiaan.
Re: D Language Foundation October 2023 Quarterly Meeting Summary
On Wednesday, 6 December 2023 at 16:28:08 UTC, Mike Parker wrote: ### Bastiaan They'd found the multithreaded performance worse than the Pascal version. He said that execution time increased with more threads and that it didn't matter how many threads you throw at it. It's the latter problem he was focused on at the moment. At first, they'd suspected the GC, but it turned out to be contention resulting from heap allocation. In Pascal, they'd heavily used variable-length arrays. For those, the length is determined at run time, but it's fixed. Since they can't grow, they're put on the stack. This makes them quite fast and avoids the global lock of the heap. I am kindly invite Bastiaan and his team to participate in this competition :) https://github.com/jinyus/related_post_gen fixed-sized arrays will suit perfectly for the task, and it also has multithreading comparison! Pascal should be good over there
Re: D Language Foundation October 2023 Quarterly Meeting Summary
This needs to be taken out of DRuntime because DRuntime is distributed pre-compiled, and that ties it to a specific compiler API, which isn't good. Instead, we should distribute it as a package. It's something he'd brought up before. Why not directly distribute DRuntime as a source? or rather, simplify how it can be used as a source ``dmd -i`` does the magic already, it'll be able to pick what ever module on the fly That's how i use my custom runtime, as source, makes things much smoother to use, however, in the case of druntime, it might highlight some compilation speed issues What was the rational behind distributing the runtime as a compiled library?
D Language Foundation October 2023 Quarterly Meeting Summary
The D Language Foundation's quarterly meeting for October 2023 took place on Friday the 6th at 15:00 UTC. This was quite a short one as far as quarterlies go, clocking in at around 35 minutes. ## The Attendees The following people attended the meeting: * Mathis Beer (Funkwerk) * Walter Bright (DLF) * Dennis Korpel (DLF) * Mario Kröplin (Funkwerk) * Mathias Lang (DLF/Symmetry) * Átila Neves (DLF/Symmetry) * Mike Parker (DLF) * Igor Pikovets (Ahrefs) * Carsten Rasmussen (Decard) * Robert Schadek (DLF/Symmetry) * Bastiaan Veelo (SARC) ## The Summary ### Bastiaan Bastiaan reported that SARC had been testing their D codebase (transpiled from Pascal---[see Bastiaan's DConf 2019 talk](https://youtu.be/HvunD0ZJqiA)). They'd found the multithreaded performance worse than the Pascal version. He said that execution time increased with more threads and that it didn't matter how many threads you throw at it. It's the latter problem he was focused on at the moment. At first, they'd suspected the GC, but it turned out to be contention resulting from heap allocation. In Pascal, they'd heavily used variable-length arrays. For those, the length is determined at run time, but it's fixed. Since they can't grow, they're put on the stack. This makes them quite fast and avoids the global lock of the heap. One way to do that in D is to use `alloca`, but that's an issue because the memory it allocates has to be used in the same function that calls the `alloca`. So you can't, e.g., use `alloca` to alloc memory in a constructor, and that prevents using it in a custom array implementation. He couldn't think of a way to translate it. He was able to work around it by using allocators in the array implementation with a thread-local free list. He found that promising. His current problem was that it took a lot of time to understand the experimental allocators package. Once he got this sorted, he would have to see if it helped solve the problem they were seeing with more threads resulting in worse performance. There was also a problem with DMD underperforming Pascal. DMD's output was about five times slower than Pascal's. His tests with LDC showed it was two times faster than Pascal. Unfortunately, they are currently limited to 32-bit Windows, and it will be a few years before they can migrate to 64-bit. LDC unfortunately had an issue that [caused stack corruption on 32-bit Windows](https://github.com/ldc-developers/ldc/issues/4265). They'd hit it in one case and were able to work around it, but he couldn't be sure they wouldn't hit it somewhere else. He wasn't willing to risk unreliable computations. He said that LDC could do the right thing, but his understanding from talking to Martin was that implementing it would have a large time cost. Since Win32 is going to eventually go away, he wasn't very keen on paying that cost. They'd spoken at DConf about the possibility of LDC raising compilation errors when stack corruption could occur so that they could then work around those cases, but he hadn't followed up with Martin about it. They'd spent seven years getting the transcompilation complete, so this was a critical issue they needed to resolve. He was hopeful that the experimental allocator package would help solve it. Robert asked if he'd looked into doing something like the small string optimization, where you set a default size that you use for static arrays and then only resort to heap allocation when you need something larger. Had they analyzed their code to determine the array sizes they were using? Bastiaan said yes, a consequence of this issue was that they were linking with a rather large stack size. Walter suggested he just use `alloca`. Just have the transcompiler emit calls to `alloca` in the first lines of the function body for any VLAs and they should be okay. Bastiaan said they'd thought of allocating large chunks of memory up front and just picking off chunks of that for a custom allocator. That works very close to a free list, then he discovered the std allocator package has a free list. His experiments with that worked, but it had been challenging to implement it more generally. He said he would have to take another look at `alloca`. Walter said `alloca` wasn't used very much in D, but it's there. If he were to implement C VLAs, that's what he'd use to do it. Robert stressed they should analyze their code to see what a magic maximum number of elements is and just use that for static arrays, allocating on the heap when they need more. Static arrays and `alloca` were comparable to some degree. Maybe they could get away with that. It should result in cleaner code. Robert also suggested that since this project has been going on for so long and was a good showcase for D in general, Bastiaan should come back and ask for help even on more than a quarterly basis. We then had a bit of discussion about what it would take to fix the LDC