llvm dev - Jul 2015 - [LLVMdev] How to not zeroinitialize array

Hi,

I'm trying to do a proof of concept of compiling some code down to a very
simple runtime that doesn't provide support for zeroinitialize.  If I run
the below code through llc:

; ModuleID = '.\t.bc'
target datalayout = "e-m:w-p:32:32-i64:64-f80:32-n8:16:32-S32"
target triple = "i686-pc-windows-gnu"

@_ZN3Foo11ZeroAndZeroE = global [2 x i32] [i32 0, i32 0], align 4

!llvm.ident = !{!0}

!0 = !{!"clang version 3.6.0 (tags/RELEASE_360/final)"}

It turns into this:

; ModuleID = '.\t.ll'
target datalayout = "e-m:w-p:32:32-i64:64-f80:32-n8:16:32-S32"
target triple = "i686-pc-windows-gnu"

@_ZN3Foo11ZeroAndZeroE = global [2 x i32] zeroinitializer, align 4

!llvm.ident = !{!0}

!0 = !{!"clang version 3.6.0 (tags/RELEASE_360/final)"}

My question is where is this transform happening - it looks like it's a
byproduct of parsing the ll file which was super surprising to me.  And second
is there a know what to circumvent this?  For my app I just want the zero'd
space in the section so I can copy it into memory.

Thanks,

-R
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20150721/b55d4235/attachment.html>

On Tue, Jul 21, 2015 at 1:31 PM, Russell Hadley <rhadley at microsoft.com>
wrote:
>  Hi,
>
>
>
> I’m trying to do a proof of concept of compiling some code down to a very
> simple runtime that doesn’t provide support for zeroinitialize.  If I run
> the below code through llc:
>
>
>
> ; ModuleID = '.\t.bc'
>
> target datalayout = "e-m:w-p:32:32-i64:64-f80:32-n8:16:32-S32"
>
> target triple = "i686-pc-windows-gnu"
>
>
>
> @_ZN3Foo11ZeroAndZeroE = global [2 x i32] [i32 0, i32 0], align 4
>
>
>
> !llvm.ident = !{!0}
>
>
>
> !0 = !{!"clang version 3.6.0 (tags/RELEASE_360/final)"}
>
>
>
> It turns into this:
>
>
>
> ; ModuleID = '.\t.ll'
>
> target datalayout = "e-m:w-p:32:32-i64:64-f80:32-n8:16:32-S32"
>
> target triple = "i686-pc-windows-gnu"
>
>
>
> @_ZN3Foo11ZeroAndZeroE = global [2 x i32] zeroinitializer, align 4
>
>
>
> !llvm.ident = !{!0}
>
>
>
> !0 = !{!"clang version 3.6.0 (tags/RELEASE_360/final)"}
>
>
>
> My question is where is this transform happening – it looks like it’s a
> byproduct of parsing the ll file which was super surprising to me.
>
This is a natural result of the IR constant folder.  It isn't an
optimization per-se, it is part of the mechanism which canonicalizes
constants.

>   And second is there a know what to circumvent this?
>
It is impossible to disable it.

>   For my app I just want the zero’d space in the section so I can copy it
> into memory.
>
I don't quite follow.  @_ZN3Foo11ZeroAndZeroE will be in the .bss section
section and will end up in memory.  Do you need it in some other section?

>
>
> Thanks,
>
>
>
> -R
>
> _______________________________________________
> LLVM Developers mailing list
> LLVMdev at cs.uiuc.edu         http://llvm.cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20150721/34dcefbb/attachment.html>

Hi Russel,

On 21 July 2015 at 13:31, Russell Hadley <rhadley at microsoft.com>
wrote:
> My question is where is this transform happening – it looks like it’s a
> byproduct of parsing the ll file which was super surprising to me.
It's not even a transformation. It's just a shorthand used for
printing out (and reading in) a constant 0. What you want to do (well,
what you think you want to do) is prevent LLVM from putting any 0 data
in a special section when compiling.
> And second is there a know what to circumvent this?  For my app I just want
the
> zero’d space in the section so I can copy it into memory.
This is probably better handled at the linker level, but for hacking
purposes there does seem to be an llc option to override the choice to
use automatic zero-fill: -nozero-initialized-in-bss. I've never used
it, but it may help for some quick tests.

Cheers.

Tim.

Thanks the .bss trick was part of it.  I had another wire crossed but I've
got it straightened now.

-----Original Message-----
From: Tim Northover [mailto:t.p.northover at gmail.com] 
Sent: Tuesday, July 21, 2015 1:46 PM
To: Russell Hadley
Cc: llvmdev at cs.uiuc.edu
Subject: Re: [LLVMdev] How to not zeroinitialize array

Hi Russel,

On 21 July 2015 at 13:31, Russell Hadley <rhadley at microsoft.com>
wrote:
> My question is where is this transform happening – it looks like it’s 
> a byproduct of parsing the ll file which was super surprising to me.
It's not even a transformation. It's just a shorthand used for printing
out (and reading in) a constant 0. What you want to do (well, what you think you
want to do) is prevent LLVM from putting any 0 data in a special section when
compiling.
> And second is there a know what to circumvent this?  For my app I just 
> want the zero’d space in the section so I can copy it into memory.
This is probably better handled at the linker level, but for hacking purposes
there does seem to be an llc option to override the choice to use automatic
zero-fill: -nozero-initialized-in-bss. I've never used it, but it may help
for some quick tests.

Cheers.

Tim.