Bug 8854 – incomprehensible bug on windows with import side effect

Created attachment 1150 zip containing all 4 files to reproduce bug I've attached a set of 4 files I've simplified as much as I could. This gives rise to a very weird bug (only on windows 32 bits). Here's the contents of the readme: steps to reproduce bug: ---- cd to directory containing this readme rdmd --force -I.\tests main ---- note: dmd -I.\tests main main_aux1 tests/main_testfun and then running main still has the bug note: any of the following changes will remove the bug (ie the assert will pass): rename tests/main_testfun to tests/main_aux2 (or probably other stuff) and reflecting this in main_testfun.d and the import statement in main.d) rename directory tests to test123 (or something else) and reflecting this in the -I flag remove the (useless) import main_aux1; in main_testfun.d remove the (useless) import std.stdio in main_aux1.d replace assert([0].map!(a=>b.length)[0]==1); by assert(b.length==1); in main_testfun.d ├── main.d ├── main_aux1.d ├── readme.txt └── tests └── main_testfun.d -------------------- contents of each file: cat main.d -------------------- import main_testfun; void main(){ testfun; } -------------------- cat main_aux1.d -------------------- module main_aux1; import std.stdio;//works wo this -------------------- cat tests/main_testfun.d -------------------- module main_testfun; import main_aux1; import std.algorithm:map; void testfun(){ auto b=[1]; assert([0].map!(a=>b.length)[0]==1); } --------------------

This happens in linux too with both options -m32 and -m64 (dmd 2.060). Also if dummy import of std.stdio is changed to other module (eg. bigint) the bug still happens.

Forgot to mention - I cannot reproduce bug related to naming files and directories. Comparing disassembles of testfunc shows that they are same, except __lamda2. Buggy version (presence of dummy import) boils down to following asm: 0x0000000000419b84 <+0>: push %rbp 0x0000000000419b85 <+1>: mov %rsp,%rbp 0x0000000000419b88 <+4>: sub $0x10,%rsp 0x0000000000419b8c <+8>: mov (%rdi),%rax 0x0000000000419b8f <+11>: leaveq 0x0000000000419b90 <+12>: retq However, in correct version of program without dummy import asm is: 0x0000000000419c68 <+0>: push %rbp 0x0000000000419c69 <+1>: mov %rsp,%rbp 0x0000000000419c6c <+4>: sub $0x10,%rsp 0x0000000000419c70 <+8>: mov -0x30(%rdi),%rax 0x0000000000419c74 <+12>: leaveq 0x0000000000419c75 <+13>: retq It seems that importing some code changes size of type. In everything else assemblies look similar, except in changes of relatives addresses (binaries are of different sizes.) May be worth mentioning that running two version shows that differences started at calling opIndex of MapStruct (https://github.com/D-Programming-Language/phobos/blob/master/std/algorithm.d#L438). In buggy version index=42_998_32, in correct version index=42_982_32. Both function return same value.

Fixed in issue 8774