No-op module ctors cause circular dependencies.

[dlangBugzillaToGithub]

Marco Leise (@mleise) reported this on 2014-10-18T10:09:06Z

Transferred from https://issues.dlang.org/show_bug.cgi?id=13631

Description

Based on commit 15b68997d0ae2332ed653658e5d7643fcf0d354f

In some cases dmd now seems to emit module ctors/dtors for empty modules just depending on what they import. This in turn causes new circular dependency conflicts that weren't there previously.

I didn't reduce my code base, but it probably rings a bell already when I say that this file now gets an auto-generated(?) module ctor:

------------------
module Lib.Audio.Sample;

import Lib.Audio.AudioLoader;
------------------

(AudioLoader itself also imports Lib.Audio.Sample.)

The generated file is ~810 KiB large and the output of nm for `Lib.Audio.Sample.o` contains as well as a lot of symbols from Phobos like std.parallelism or std.container:

0000000000000000 T _D3Lib5Audio6Sample9__modctorFZv
0000000000000000 T _D3Lib5Audio6Sample9__moddtorFZv

objdump further reveals that DMD now emulates -ffunction-sections and that the code it generates for the module ctor/dtor is a glorified no-op.

Disassembly of section .text._D3Lib5Audio6Sample9__modctorFZv:

0000000000000000 <_D3Lib5Audio6Sample9__modctorFZv>:
   0:   push   rbp
   1:   mov    rbp,rsp
   4:   push   rbx
   5:   mov    rax,QWORD PTR fs:0x0
   e:   mov    rcx,QWORD PTR [rip+0x0]
  15:   inc    DWORD PTR [rax+rcx*1]
  18:   mov    rdx,QWORD PTR fs:0x0
  21:   mov    rbx,QWORD PTR [rip+0x0]
  28:   inc    DWORD PTR [rdx+rbx*1]
  2b:   pop    rbx
  2c:   pop    rbp
  2d:   ret    
  2e:   xchg   ax,ax

Disassembly of section .text._D3Lib5Audio6Sample9__moddtorFZv:

0000000000000000 <_D3Lib5Audio6Sample9__moddtorFZv>:
   0:   push   rbp
   1:   mov    rbp,rsp
   4:   call   9 <_D3Lib5Audio6Sample9__moddtorFZv+0x9>
   9:   call   e <_D3Lib5Audio6Sample9__moddtorFZv+0xe>
   e:   pop    rbp
   f:   ret

[dlangBugzillaToGithub]

Marco.Leise commented on 2014-10-18T10:47:15Z

Actually it happens with with 2.066.1 release, too.

[dlangBugzillaToGithub]

Marco.Leise commented on 2014-10-18T13:08:54Z

The issue is actually present in 2.066.0, so not a new regression. Maybe no regression at all. It just happens in -debug builds and maybe would have happened in earlier version of dmd, too. But I can't easily test this with the code already depending on 2.066.1 fixes.