Bug 4397 – const/CTFE does not work

In the following example, foo() will create an array literal on each call, instead of returning a statically shared / compile time value. On D2, the same may happen when using enum instead of const (I didn't test it). In many cases, this codegen bug may go unnoticed, which makes it just worse. $ cat rtz.d struct X { int[] a; } const cX = X([1,2]); X foo() { return cX; } $ dmd -c rtz.d $ objdump -d -Mintel rtz.o -r ... 00000000 <_D3rtz3fooFZS3rtz1X>: 0: 55 push ebp 1: 8b ec mov ebp,esp 3: 83 ec 08 sub esp,0x8 6: 53 push ebx 7: 6a 02 push 0x2 9: 6a 01 push 0x1 b: 6a 02 push 0x2 d: b8 00 00 00 00 mov eax,0x0 e: R_386_32 _D11TypeInfo_Ai6__initZ 12: 50 push eax 13: e8 fc ff ff ff call 14 <_D3rtz3fooFZS3rtz1X+0x14> 14: R_386_PC32 _d_arrayliteralT 18: 89 c1 mov ecx,eax 1a: bb 02 00 00 00 mov ebx,0x2 1f: 89 5d f8 mov DWORD PTR [ebp-0x8],ebx 22: 89 4d fc mov DWORD PTR [ebp-0x4],ecx 25: 8b 55 fc mov edx,DWORD PTR [ebp-0x4] 28: 8b 45 f8 mov eax,DWORD PTR [ebp-0x8] 2b: 83 c4 10 add esp,0x10 2e: 5b pop ebx 2f: c9 leave 30: c3 ret

Does anyone know a workaround for this? It's starting to get annoying, and Walter probably died, so it will take a while until this is fixed.

(Wrong code on v2.046 too, when switching const with enum.)

This generated code is identical on 1.020: this is not a regression. Downgrading to major. I don't see any evidence that it's wrong-code, either: it seems to be a performance issue. Finally, it looks like a duplicate of bug 2356.

Shouldn't literals be immutable and the code - invalid?

(In reply to comment #4) > Shouldn't literals be immutable and the code - invalid? I think they *should*. I argued strongly for immutable array literals. But I lost. So the code is valid, but likely to be very slow for a very long time.

It's not valid. This is a systems programming language, and the compiler can't just randomly insert memory allocations. What if you wrote a kernel in D? I insist on the wrong-code keyword.

>and the compiler can't just randomly insert memory allocations. Add "that are not supposed to be there".

Well... there can be a problem with immutable literals because immutability is transitive... hmm... Even if literals aren't immutable, compiler can still catch assignment of literal to mutable array and report error.

(In reply to comment #8) > Well... there can be a problem with immutable literals because immutability is > transitive... hmm... > Even if literals aren't immutable, compiler can still catch assignment of > literal to mutable array and report error. But it is EXPLICITLY LEGAL to assign a literal to a mutable array. There is an invisible dup by design. I don't like this, I argued strongly against it, but it's in there. This isn't wrong code. (Similarly, you can write const C x = new C; The C will be allocated on the heap, even though it will never change afterwards).

The main problem from my point of view is that this fails: struct X { int[] a; } const cX = X([1,2]); void main() { assert(cX.a.ptr is cX.a.ptr); } Which is an issue very similar to bug 2526 .

Btw. this is a CTFE problem, and the first example is in D1. I don't know what the hell is with D2 and immutable and implicit dups, but to get the same behavior on D2, replace const with enum. The array should be on the *datsegment*, not somehow constructed on the fly. Not sure about the exact testcase in bug 2356, but it's definitely different from writing "int[3] x; x[] = [1,2,3];".

I just noticed that the wrong-code keyword is gone. Sorry that's just bullshit.

*** This issue has been marked as a duplicate of issue 2526 ***

This isn't the same as bug 2526, which is an accepts-invalid bug. This has nothing to do with templates. The glue layer needs to detect when a constant is being initialized with an array literal, and when that happens, change it into a value on the static data segment (ie, it needs to create an XXXinit initializer). Doing this would also help with bug 2356, but it isn't the same. (Bug 2356 will often need to create values at runtime, so it will need to copy data from the initializer).

It passes if you specify the type in the const/enum. The reason is, that in declaration.c, VarDeclaration::semantic(), we see this code: else if (storage_class & (STCconst | STCimmutable | STCmanifest) || type->isConst() || type->isImmutable()) { /* Because we may need the results of a const declaration in a * subsequent type, such as an array dimension, before semantic2() * gets ordinarily run, try to run semantic2() now. * Ignore failure. */ if (!global.errors && !inferred) .... If we're inferring type inference, then semantic2 doesn't get run on the variable declaration (cX in the original bug report). So when it runs semantic on "return cx;", DsymbolExp::semantic() for cx, returns cx->init, which hasn't had semantic run on it yet. So it behaves like a copy-and-paste of the initializer.

There is no compiler bug here, it is working as designed, at least for D2. However, the design doesn't make sense. I have opened bug 9953 for the design change. It would be possible to give the desired behaviour in D1, though it's hardly worthwhile. In D2 I think it is a logical impossibility. IMHO this should just fail to compile. You should write "static const" instead of "enum" when you want to declare a constant of a reference type.

This works fine in D2: `enum` gives you a new value every time, but `[static] immutable` does what you expect.