llvm dev - Dec 2011 - [LLVMdev] A Question about LLVM structures - alignment of data members

Hi All,

Today, LLVM IR doesn't include a description for alignment for specific data
members inside structures.
The implies that LLVM IR Generators need to do the padding themselves in order
to achieve the requested alignment.
Since the information is not represented explicitly by the LLVM IR, optimization
passes might unintentionally break the required alignment and lead to unexpected
results when executing the generated program.

Here is an example "C" code (and OpenCL as well)
struct st1_t {
    int  i;
    float __attribute__((align(64))) f;
};
struct st1_t ist1;

Generated IR by CLANG:
%struct.st1_t = type { i32, [60 x i8], float, [60 x i8] }
@ist1 = common global %struct.st1_t zeroinitializer, align 64

What is the reason for disallowing alignment of data members inside structs in
LLVM IR? What were the design considerations?

Thanks,
Boaz
---------------------------------------------------------------------
Intel Israel (74) Limited

This e-mail and any attachments may contain confidential material for
the sole use of the intended recipient(s). Any review or distribution
by others is strictly prohibited. If you are not the intended
recipient, please contact the sender and delete all copies.
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20111214/6a0bc7a3/attachment.html>

On Wed, Dec 14, 2011 at 12:44 PM, Ouriel, Boaz <boaz.ouriel at intel.com>
wrote:
> Hi All,
>
>
>
> Today, LLVM IR doesn’t include a description for alignment for specific
data
> members inside structures.
>
> The implies that LLVM IR Generators need to do the padding themselves in
> order to achieve the requested alignment.
If the alignment required is higher than native alignment, yes.
> Since the information is not represented explicitly by the LLVM IR,
> optimization passes might unintentionally break the required alignment and
> lead to unexpected results when executing the generated program.
I don't see how this follows: you can explicitly specify alignment for
every operation where it matters. If some operation requires a
specific alignment specified in the IR, the optimizers won't break it.
> What is the reason for disallowing alignment of data members inside structs
> in LLVM IR? What were the design considerations?
We have a perfectly usable design for handling explicitly specifying
alignment.  Allowing explicit alignment on struct types would make the
IR more complicated, and making the IR more complicated requires a
strong justification.

-Eli