Re: Using C++ Classes From D: dmd cannot link, while ldc segfault
If I use array: ``` extern(C++) { void getInts(core.stdcpp.array.array!(int, 10) vec) { foreach (int i; 0 .. 10) { vec.at(i) = i; } } } ``` ``` #include using namespace std; void getInts(array* vector); ``` Both DMD and LDC has link error: base.cpp:42: undefined reference to `getInts(std::array10ul>*)'
Re: Using C++ Classes From D: dmd cannot link, while ldc segfault
On 19/06/2023 5:54 PM, mw wrote: Ha, I saw vector.d there, So I can use this vector.d as the D side of C++'s std::vector? Probably, I just don't know how well tested it is. But worth a go!
Re: Using C++ Classes From D: dmd cannot link, while ldc segfault
On Monday, 19 June 2023 at 05:56:54 UTC, Richard (Rikki) Andrew Cattermole wrote: On 19/06/2023 5:54 PM, mw wrote: Ha, I saw vector.d there, So I can use this vector.d as the D side of C++'s std::vector? Probably, I just don't know how well tested it is. But worth a go! ``` import core.stdcpp.vector; extern(C++) { void getInts(core.stdcpp.vector.vector!(int) vec) { foreach (int i; 0 .. 10) { vec.push_back(i); } } } ``` dmd v2.104.0 failed: ``` /usr/include/dmd/druntime/import/core/stdcpp/vector.d(58): Error: undefined identifier `size`, did you mean alias `size_t`? /usr/include/dmd/druntime/import/core/stdcpp/vector.d(33): Error: template instance `core.stdcpp.vector.vector!(int, allocator!int)` error instantiating main.d(32):instantiated from here: `vector!int` ``` LDC - the LLVM D compiler (1.32.2): ``` main.d(32): Error: undefined identifier `vector` in module `core.stdcpp.vector`, did you mean enum member `MIctor`? ``` So what's wrong the LDC? how do I write `core.stdcpp.vector.vector` then?
Re: How does D’s ‘import’ work?
On Sunday, 18 June 2023 at 20:24:10 UTC, Cecil Ward wrote: I wasn’t intending to use DMD, rather ldc if possible or GDC because of their excellent optimisation, in which DMD seems lacking, is that fair? (Have only briefly looked at dmd+x86 and haven’t given DMD’s back end a fair trial.) Other than the execution runtime, one other very important problem with DMD that wasn't refered is that is only support x86 and x86_64. LDC and GDC support LLVM's and GCC's architectures respectively.
Re: How does D’s ‘import’ work?
On Sunday, 18 June 2023 at 20:17:50 UTC, Cecil Ward wrote: target.obj: target.c include1.h include2.h cc.exe cc target.c and you either have to pray that you have kept the list of .h files that are mentioned inside target.c and other .h files referenced recursively from within these .h files etc. I listed the compiler as a dependency too, and I should really have a pseudo-target somehow that depends on the nature of the command line because changing the command line affects the generated code. If you have an automaking compiler that will generate the list of .h files then that’s so, so much safer. First of all, If we are talking about C files, D can import and compile them so you don't even need a Makefile! Now, if you need to compile C++ files and then either link or create a library (and link with it from the D project), then you can just run Dub in the end of the job in your make file! You can then have a variable called `DUB_FLAGS` in your Makefile and this is where the arguments that will be passed for the Dub will be. Will this be good enough for you?
Re: How does D’s ‘import’ work?
On Monday, 19 June 2023 at 08:46:31 UTC, rempas wrote: On Sunday, 18 June 2023 at 20:17:50 UTC, Cecil Ward wrote: target.obj: target.c include1.h include2.h cc.exe cc target.c and you either have to pray that you have kept the list of .h files that are mentioned inside target.c and other .h files referenced recursively from within these .h files etc. I listed the compiler as a dependency too, and I should really have a pseudo-target somehow that depends on the nature of the command line because changing the command line affects the generated code. If you have an automaking compiler that will generate the list of .h files then that’s so, so much safer. First of all, If we are talking about C files, D can import and compile them so you don't even need a Makefile! Now, if you need to compile C++ files and then either link or create a library (and link with it from the D project), then you can just run Dub in the end of the job in your make file! You can then have a variable called `DUB_FLAGS` in your Makefile and this is where the arguments that will be passed for the Dub will be. Will this be good enough for you? If I have sources to all the library routines, not libraries or .obj files. I am simply completely ignorant about the D tools including DUB, so off to do some reading. I’ve just been seeing how good LDC and GDC are, and the answer is extremely, especially LDC, which perhaps has a slight edge in code generation quality. I haven’t looked at AAarch64 code yet, only AMD64. Very impressed with all the work!
Re: SIMD c = a op b
On Sunday, 18 June 2023 at 05:01:16 UTC, Cecil Ward wrote: On Sunday, 18 June 2023 at 04:54:08 UTC, Cecil Ward wrote: Is it true that this doesn’t always work (in either branch)? float4 a,b; static if (__traits(compiles, a/b)) c = a / b; else c[] = a[] / b[]; It's because SIMD stuff doesn't always works that intel-intrinsics was created. It insulates you from the compiler underneath. import inteli.emmintrin; void main() { float4 a, b, c; c = a / b;// _always_ works c = _mm_div_ps(a, b); // _always_ works } Sure in some case it may emulate those vectors, but for vector of float it's only in DMD -m32. It relies on excellent __vector work made a long time ago, and supplements it. For 32-byte vectors such as __vector(float[8]), you will have trouble on GDC when -mavx isn't there, or with DMD. Do you think the builtin __vector support the same operations across the compilers? The answer is "it's getting there", in the meanwhile using intel-intrinsics will lower your exposure to the compiler woes. If you want to use DMD and -O -inline, you should also expect much more problems unless working extra in order to have SIMD.
Re: How does D’s ‘import’ work?
On Monday, 19 June 2023 at 12:48:26 UTC, Cecil Ward wrote: If I have sources to all the library routines, not libraries or .obj files. I am simply completely ignorant about the D tools including DUB, so off to do some reading. I’ve just been seeing how good LDC and GDC are, and the answer is extremely, especially LDC, which perhaps has a slight edge in code generation quality. I haven’t looked at AAarch64 code yet, only AMD64. Very impressed with all the work! Of course, DMD uses it's own custom backend so it's only fair to not expect for it to have the same runtime performance and optimizations as the other two compilers than use LLVM and GCC. If you only use x86_64, DMD will be amazing for your debug cycles!
GC doesn't collect where expected
I'm doing some experiments with ldc2 GC, by instrumenting it and printing basic information (what is allocated and freed) My first tests are made on this sample : ``` cat test2.d import core.memory; class Bar { int bar; } class Foo { this() { this.bar = new Bar; } Bar bar; } void func() { Foo f2 = new Foo; } int main() { Foo f = new Foo; func(); GC.collect(); return 0; } ``` When trying to run the instrumented druntime, I get a strange behavior : the first collection (done with GC.collect) doesn't sweep anything (in particular, it doesn't sweep memory allocated in _func()_). The whole sweeping is done when program finish, at cleanup. I don't understand why : memory allocated in _func()_ shouldn't be accessible from any root at first collection, right ? ``` ╰─> /instrumented-ldc2 -g -O0 test2.d --disable-gc2stack --disable-d-passes --of test2 && ./test2 "--DRT-gcopt=cleanup:collect fork:0 parallel:0 verbose:2" [test2.d:26] new 'test2.Foo' (24 bytes) => p = 0x7f3a0454d000 [test2.d:10] new 'test2.Bar' (20 bytes) => p = 0x7f3a0454d020 [test2.d:21] new 'test2.Foo' (24 bytes) => p = 0x7f3a0454d040 [test2.d:10] new 'test2.Bar' (20 bytes) => p = 0x7f3a0454d060 COLLECTION = = MARKING == marking range: [0x7fff22337a60..0x7fff22339000] (0x15a0) range: [0x7f3a0454d000..0x7f3a0454d020] (0x20) range: [0x7f3a0454d040..0x7f3a0454d060] (0x20) marking range: [0x7f3a0464d720..0x7f3a0464d8b9] (0x199) marking range: [0x46c610..0x47b3b8] (0xeda8) = SWEEPING == = COLLECTION = = MARKING == marking range: [0x46c610..0x47b3b8] (0xeda8) = SWEEPING == Freeing test2.Foo (test2.d:26; 24 bytes) (0x7f3a0454d000). AGE : 1/2 Freeing test2.Bar (test2.d:10; 20 bytes) (0x7f3a0454d020). AGE : 1/2 Freeing test2.Foo (test2.d:21; 24 bytes) (0x7f3a0454d040). AGE : 1/2 Freeing test2.Bar (test2.d:10; 20 bytes) (0x7f3a0454d060). AGE : 1/2 = ```
Re: GC doesn't collect where expected
On 6/19/23 12:13 PM, axricard wrote: I'm doing some experiments with ldc2 GC, by instrumenting it and printing basic information (what is allocated and freed) My first tests are made on this sample : ``` cat test2.d import core.memory; class Bar { int bar; } class Foo { this() { this.bar = new Bar; } Bar bar; } void func() { Foo f2 = new Foo; } int main() { Foo f = new Foo; func(); GC.collect(); return 0; } ``` When trying to run the instrumented druntime, I get a strange behavior : the first collection (done with GC.collect) doesn't sweep anything (in particular, it doesn't sweep memory allocated in _func()_). The whole sweeping is done when program finish, at cleanup. I don't understand why : memory allocated in _func()_ shouldn't be accessible from any root at first collection, right ? ``` ╰─> /instrumented-ldc2 -g -O0 test2.d --disable-gc2stack --disable-d-passes --of test2 && ./test2 "--DRT-gcopt=cleanup:collect fork:0 parallel:0 verbose:2" [test2.d:26] new 'test2.Foo' (24 bytes) => p = 0x7f3a0454d000 [test2.d:10] new 'test2.Bar' (20 bytes) => p = 0x7f3a0454d020 [test2.d:21] new 'test2.Foo' (24 bytes) => p = 0x7f3a0454d040 [test2.d:10] new 'test2.Bar' (20 bytes) => p = 0x7f3a0454d060 COLLECTION = = MARKING == marking range: [0x7fff22337a60..0x7fff22339000] (0x15a0) range: [0x7f3a0454d000..0x7f3a0454d020] (0x20) range: [0x7f3a0454d040..0x7f3a0454d060] (0x20) marking range: [0x7f3a0464d720..0x7f3a0464d8b9] (0x199) marking range: [0x46c610..0x47b3b8] (0xeda8) = SWEEPING == = COLLECTION = = MARKING == marking range: [0x46c610..0x47b3b8] (0xeda8) = SWEEPING == Freeing test2.Foo (test2.d:26; 24 bytes) (0x7f3a0454d000). AGE : 1/2 Freeing test2.Bar (test2.d:10; 20 bytes) (0x7f3a0454d020). AGE : 1/2 Freeing test2.Foo (test2.d:21; 24 bytes) (0x7f3a0454d040). AGE : 1/2 Freeing test2.Bar (test2.d:10; 20 bytes) (0x7f3a0454d060). AGE : 1/2 = ``` In general, the language does not guarantee when the GC will collect your item. In this specific case, most likely it's a stale register or stack reference. One way I usually use to ensure such things is to call a function that destroys the existing stack: ```d void clobber() { int[2048] x; } ``` Calling this function will clear out 2048x4 bytes of data to 0 on the stack. -Steve
Re: GC doesn't collect where expected
On 6/19/23 12:51 PM, Anonymouse wrote: On Monday, 19 June 2023 at 16:43:30 UTC, Steven Schveighoffer wrote: In this specific case, most likely it's a stale register or stack reference. One way I usually use to ensure such things is to call a function that destroys the existing stack: ```d void clobber() { int[2048] x; } ``` Calling this function will clear out 2048x4 bytes of data to 0 on the stack. Could you elaborate on how you use this? When do you call it? Just, ever so often, or is there thought behind it? Just before forcing a collect. The stack is *always* scanned conservatively, and even though really the stack data should be blown away by the next function call (probably GC.collect), it doesn't always work out that way. Indeed, even just declaring `x` might not do it if the compiler decides it doesn't actually have to. But I've found that seems to help. -Steve
Re: GC doesn't collect where expected
On Monday, 19 June 2023 at 16:43:30 UTC, Steven Schveighoffer wrote: In general, the language does not guarantee when the GC will collect your item. In this specific case, most likely it's a stale register or stack reference. One way I usually use to ensure such things is to call a function that destroys the existing stack: ```d void clobber() { int[2048] x; } ``` Calling this function will clear out 2048x4 bytes of data to 0 on the stack. -Steve All clear, thank you !
Re: GC doesn't collect where expected
On Monday, 19 June 2023 at 16:43:30 UTC, Steven Schveighoffer wrote: In this specific case, most likely it's a stale register or stack reference. One way I usually use to ensure such things is to call a function that destroys the existing stack: ```d void clobber() { int[2048] x; } ``` Calling this function will clear out 2048x4 bytes of data to 0 on the stack. -Steve Could you elaborate on how you use this? When do you call it? Just, ever so often, or is there thought behind it?
Re: GC doesn't collect where expected
On Monday, 19 June 2023 at 16:43:30 UTC, Steven Schveighoffer wrote: In general, the language does not guarantee when the GC will collect your item. In this specific case, most likely it's a stale register or stack reference. One way I usually use to ensure such things is to call a function that destroys the existing stack: ```d void clobber() { int[2048] x; } ``` Calling this function will clear out 2048x4 bytes of data to 0 on the stack. -Steve Does it mean that if my function _func()_ is as following (say I don't use clobber), I could keep a lot of memory for a very long time (until the stack is fully erased by other function calls) ? ``` void func() { Foo[2048] x; foreach(i; 0 .. 2048) x[i] = new Foo; } ```
is ref inout redundant in: ref inout(T) opIndex(size_t index)
Hi, I just saw this line: https://github.com/dlang/dmd/blob/master/druntime/src/core/stdcpp/vector.d#LL66C5-L66C39 ``` 66:ref inout(T) opIndex(size_t index) inout pure nothrow @safe @nogc { return as_array[index]; } ``` I'm wondering if the `ref` and `inout` redundant here? They both mean the same thing? in C++ terms both return the reference of the i-th element? so only one of them should be enough? If not, can someone help to explain the difference? the following 4 return types: 1) `ref T` alone 2) `inout T` alone 3) `ref inout(T)` 4) `inout ref(T)` BTW, what does the second `inout` before `pure` do? it's also redundant? Thanks.
Re: GC doesn't collect where expected
On 6/19/23 2:01 PM, axricard wrote: Does it mean that if my function _func()_ is as following (say I don't use clobber), I could keep a lot of memory for a very long time (until the stack is fully erased by other function calls) ? ``` void func() { Foo[2048] x; foreach(i; 0 .. 2048) x[i] = new Foo; } ``` When the GC stops all threads, each of them registers their *current* stack as the target to scan, so most likely not. However, the compiler/optimizer is not trying to zero out stack unnecessarily, and likely this leads in some cases to false pointers. Like I said, even the "clobber" function might not actually zero out any stack because the compiler decides writing zeros to the stack that will never be read is a "dead store" and just omit that. This question comes up somewhat frequently "why isn't the GC collecting the garbage I gave it!", and the answer is mostly "don't worry about it". There is no real good way to guarantee an interaction between the compiler, the optimizer, and the runtime to make sure something happens one way or another. The only thing you really should care about is if you have a reference to an item and it's prematurely collected. Then there is a bug. Other than that, just don't worry about it. -Steve
Re: is ref inout redundant in: ref inout(T) opIndex(size_t index)
On 6/19/23 2:19 PM, mw wrote: Hi, I just saw this line: https://github.com/dlang/dmd/blob/master/druntime/src/core/stdcpp/vector.d#LL66C5-L66C39 ``` 66: ref inout(T) opIndex(size_t index) inout pure nothrow @safe @nogc { return as_array[index]; } ``` I'm wondering if the `ref` and `inout` redundant here? They both mean the same thing? in C++ terms both return the reference of the i-th element? so only one of them should be enough? No, they do not both mean the same thing. inout is a form of mutability that is unique to D. It does *not* mean the same as `ref` like other languages (or even D1). What `inout` does is forward the mutability of the parameter to the return type. If not, can someone help to explain the difference? the following 4 return types: 1) `ref T` alone a reference to a T. 2) `inout T` alone An inout T passed by value. Sorry for the recursive definition, but inout is kinda unique with D. 3) `ref inout(T)` A reference to an inout T. 4) `inout ref(T)` I'm not sure that's valid. `ref` is a storage class, not a type modifier. BTW, what does the second `inout` before `pure` do? it's also redundant? This is the qualifier put onto the `this` parameter (i.e. the `vector` in this case). Because of this, you get the mutability of the parameter forwarded to the return type. ```d const vector!int c; immutable vector!int i; vector!int m; static assert(is(typeof(c[0]) == const(int))); static assert(is(typeof(i[0]) == immutable(int))); static assert(is(typeof(m[0]) == int)); ``` I gave a presentation on const/inout, which you might find helpful. https://dconf.org/2016/talks/schveighoffer.html -Steve