Bug 24440 – [REG 2.107] multiple multidimensional array appends broke in 2.107

Reduced example: ``` immutable int n = 500; void main() { auto a = new long [] [] [] (n, n); foreach (i; 0..n * 4) foreach (j; 0..n) a[j][0] ~= 0; assert(false); } ``` Compile string (Windows): `dmd -m32 -g reg107.d` or `dmd -m64 -g reg107.d` Expected behavior (works in 2.106.0): The program correctly goes to `assert(false)` and halts there. This way is used instead of debug output to avoid libraries. Issue (in 2.107.0, 2.107.1, and 2.108.0-rc1): The Win32 build breaks with: ``` object.Error@(0): Win32 Exception ---------------- 0x740CF41C in wcslen 0x00401101 in D main at C:\programs\stuff\dmd\reg107.d(6) 0x004042CF in void rt.dmain2._d_run_main2(char[][], uint, extern (C) int function(char[][])*).runAll().__lambda2() ... ``` The Win64 build just breaks silently. Issue was reported on Linux too, but I can't check there at the moment.

Created attachment 1911 reduced example

With Linux, it's just "segmentation fault" with 2.107. Correctly goes to assert(false) with 2.106.

Originally got hit by it in my toy language project which I use to teach parallel computing. It maintains a three-dimensional array to store queues of messages between different simulated processes. For now, circumvented the issue by transforming it to a two-dimensional array of structs containing arrays. https://github.com/GassaFM/interpr/commit/b959bb7870 Still, multidimensional arrays are a pretty core language feature. Can I expect it would work again at some point?.. Ivan Kazmenko.

Sorry if this is unrelated, but I'm trying to add an example with similar behavior. The following code, given `1000 1000` on the first line and 1000 tokens in each of the subsequence 1000 lines from standard input, fails to run successfully and it is just "Segmentation fault" (DMD>=2.107.0, WSL2). ``` import std.conv, std.stdio, std.string; void main() { string[] tokens = readln.chomp.split; const M = tokens[0].to!int; const N = tokens[1].to!int; auto A = new string[][](M, N); foreach (x; 0 .. M) { tokens = readln.chomp.split; foreach (y; 0 .. N) { A[x][y] = tokens[y]; } } } ``` For further info, I found this at when solving a problem at https://yukicoder.me/ - my submission link: https://yukicoder.me/submissions/974786 where I see that when the input is smaller the segmentation fault does not happen (RE (Runtime Error) vs WA (Wrong Answer)).

This looks like the same problem as in issue 24498. The expression `new long [][][](n, n)` creates memory, which is not scanned by the GC, but contains pointers. The GC will free the inner arrays and the program crashes, when they are accessed.

*** This issue has been marked as a duplicate of issue 24498 ***