llvm dev - Jul 2013 - [LLVMdev] Another memory alignment issue with SSE operations

Unfortunately, I've ran into a second issue where addpd is being 
performed on memory that isn't 16 byte aligned. Again, this only happens 
if the createJIT OptLevel is set to Default (vs None).

According to
http://www.jaist.ac.jp/iscenter-new/mpc/altix/altixdata/opt/intel/vtune/doc/users_guide/mergedProjects/analyzer_ec/mergedProjects/reference_olh/mergedProjects/instructions/instruct32_hh/vc8a.htm
that will cause a GPF.

I've attached the LLVM IR and a copy of the Disassembly this results in. 
The crash occurs at 00370872

At the time of the crash, ESP is set to 0018EEF8 - this results in a 
value is not 16 byte aligned. I notice that the offset is aligned though.

The crash occurs on the first instance of addpd applied to the stack (as 
I understand ESP is used for).

There is also raises the question of would it be worth requiring 
alignment of the function stack to improve performance (assuming movapd 
is faster then movupd). I'm not expecting LLVM to recognize this 
(although it would be neat) but is this something worth setting 
ourselves, knowing we're going to be using mostly SSE instructions? And 
how would we do that?

--
Peter N
-------------- next part --------------
A non-text attachment was scrubbed...
Name: addpd-unaligned.zip
Type: application/octet-stream
Size: 42810 bytes
Desc: not available
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20130720/387f45bf/attachment.obj>

I discovered the alignstack(16) attribute, and am now using this in our 
code.

This avoids this crash, and as a bonus, causes the resulting code to use 
movapd instead of movupd - so is potentially faster code, too.

--
Peter N

On 20/07/2013 6:08 PM, Peter Newman wrote:
> Unfortunately, I've ran into a second issue where addpd is being 
> performed on memory that isn't 16 byte aligned. Again, this only 
> happens if the createJIT OptLevel is set to Default (vs None).
>
> According to
>
http://www.jaist.ac.jp/iscenter-new/mpc/altix/altixdata/opt/intel/vtune/doc/users_guide/mergedProjects/analyzer_ec/mergedProjects/reference_olh/mergedProjects/instructions/instruct32_hh/vc8a.htm
>
> that will cause a GPF.
>
> I've attached the LLVM IR and a copy of the Disassembly this results 
> in. The crash occurs at 00370872
>
> At the time of the crash, ESP is set to 0018EEF8 - this results in a 
> value is not 16 byte aligned. I notice that the offset is aligned though.
>
> The crash occurs on the first instance of addpd applied to the stack 
> (as I understand ESP is used for).
>
> There is also raises the question of would it be worth requiring 
> alignment of the function stack to improve performance (assuming 
> movapd is faster then movupd). I'm not expecting LLVM to recognize 
> this (although it would be neat) but is this something worth setting 
> ourselves, knowing we're going to be using mostly SSE instructions? 
> And how would we do that?
>
> -- 
> Peter N
>
>
> _______________________________________________
> 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/20130720/1062ceec/attachment.html>