Alignment of section does not reflect the maximum alignment of its contents

[dlangBugzillaToGithub]

Yuxuan Shui (@yshui) reported this on 2018-08-08T00:46:05Z

Transferred from https://issues.dlang.org/show_bug.cgi?id=19148

Description

This cause ldc compiled with dmd to crash, because of an unaligned SIMD instruction.

[dlangBugzillaToGithub]

yshuiv7 commented on 2018-08-08T01:23:04Z

Whops, not the problem. But ldc still crashes, looking into it...

[dlangBugzillaToGithub]

yshuiv7 commented on 2018-08-09T10:14:28Z

Looks like dmd mis-aligned a __gshared variable. .alignof reports an alignment of 16 bytes, while the variable is aligned only to 8 bytes.

[dlangBugzillaToGithub]

yshuiv7 commented on 2018-08-10T00:39:25Z

I found the actual bug in codegen. In elfobj.c data_start(), when a symbol doesn't have a explicit alignment (Salignment <= 0), the symbol itself is aligned, but its section's alignment is not updated.

To reproduce:

__gshared real x;
pragma(msg, x.alignof);

Compiling it prints: 16LU, but the .data section of the resulting object file is only aligned to 8 bytes.

[dlangBugzillaToGithub]

dbugz commented on 2018-08-25T15:25:27Z

Which ldc and DMD versions? Does the just-released LDC 1.11 have this problem?

[dlangBugzillaToGithub]

yshuiv7 commented on 2018-08-25T19:21:19Z

(In reply to Joakim from comment #4)
> Which ldc and DMD versions? Does the just-released LDC 1.11 have this
> problem?

Most recent ldc release fix the build problem. But this is a dmd bug nonetheless.

This bug still exists in dmd master

[dlangBugzillaToGithub]

dbugz commented on 2018-08-26T04:13:40Z

There was an issue with a ldc header mismatch that kinke found when merging the 2.081 frontend into ldc: are you sure that's not what you're seeing instead?

https://github.com/ldc-developers/ldc/pull/2752#issuecomment-398897813
https://github.com/ldc-developers/ldc/pull/2752#issuecomment-399691107

It's a known bug that DMD 2.081 won't build a working ldc 1.10, have to use 1.11.

[dlangBugzillaToGithub]

yshuiv7 commented on 2018-08-26T04:18:15Z

(In reply to Joakim from comment #6)
> There was an issue with a ldc header mismatch that kinke found when merging
> the 2.081 frontend into ldc: are you sure that's not what you're seeing
> instead?
> 
> https://github.com/ldc-developers/ldc/pull/2752#issuecomment-398897813
> https://github.com/ldc-developers/ldc/pull/2752#issuecomment-399691107
> 
> It's a known bug that DMD 2.081 won't build a working ldc 1.10, have to use
> 1.11.

I have inspected the generated binary, and found problem in DMD's code. This bug is real. Please read comment #3

[dlangBugzillaToGithub]

dbugz commented on 2018-08-26T04:26:27Z

OK, if you're so sure it's a problem in dmd's backend, a pull request is welcome.

[dlangBugzillaToGithub]

yshuiv7 commented on 2018-08-26T18:55:51Z

(In reply to Joakim from comment #8)
> OK, if you're so sure it's a problem in dmd's backend, a pull request is
> welcome.

While I know where the problem is, I don't know what is the proper fix. So I would need some help.