Why is there alignment in the LLVM datalayout?

[gergoerdi]

I've tracked down a bug to bad alignment here: #63 (comment) After a lot of pain, I've found out that there is actually a hardcoded datalayout in the AVR backend, in AVRTargetMachine.cpp, taking precedence over whatever the user puts in the LLVM IR file. This datalayout is currently as follows:

static const char *AVRDataLayout = "e-p:16:16:16-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-n8";

Expanded, the above states:

• 2-byte words are 2-byte-aligned, 4-byte words are 4-byte-aligned, 8-byte words are 8-byte-aligned
• floats are 4-byte-aligned, doubles are 8-byte-aligned
• Pointers take up 2 bytes and are 2-byte-aligned
• Aggregates are 8-byte-aligned

To my untrained eye, this looks completely bogus (except, yeah, pointers are 2 bytes:)), so I did a quick git blame in the hope that I'd find a good reason. There is a commit which gives an explanation for the 2-byte alignment:

commit 553362942dbc6baa23225d7e39020631a9c15b91
Author: Dylan McKay <...>
Date:   Wed Sep 28 13:29:10 2016 +0000

    [AVR] Update the data layout
    
    The previous data layout caused issues when dealing with atomics.
    
    Foe example, it is illegal to load a 16-bit value with less than 16-bits
    of alignment.
    
    This changes the data layout so that all types are aligned by at least
    their own width.
    
    Interestingly, this also _slightly_ decreased register pressure in some
    cases.
    
    git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@282587 91177308-0d34-0410-b5e6-96231b3b80d8

but it still doesn't explain what's going on with the 4- and even 8-byte alignments (which were also introduced in the same commit). The original code is already missing an a specifier, so that's probably just an oversight/typo.

If we need 2-byte alignment for >=2 byte words as the above commit states, what's the reason for going above that for wider types, that are going to be soft-implemented anyway?

Would this work instead? It would certainly fix #63

static const char *AVRDataLayout = "e-p:16:16-i8:8-i16:16-i32:16-i64:16-f32:16-f64:16-n8-a:16";

[gergoerdi]

I'm trying to get more information on AVR alignment on SO.

[gergoerdi]

On StackOverflow, user tofro writes in his answer:

This comment doesn't make any sense to me. AVR does not natively know 16-bit-types, and specifically not atomic access to 16-bit-types, regardless of any alignment.

On a classic AVR CPU, alignment of 16 bit data is not required as 16-bit memory access always evaluates to two 8-bit fetches to two registers.

There is a kind of "alignment restriction", though, when using the movw instruction available on some AVRs that moves data from one register pair to another - Here the register number of the lower registers must be even. This has nothing to do with memory alignment, though.

I'd like to try compiling LLVM/rustc with 8-bit alignment through and through and see how it goes. @dylanmckay, do you have a ticket # or a repro description or anything on the problems you've encountered that triggered the above commit?

[gergoerdi]

As a test, I've now recompiled libcore-mini + chip8-engine + chip8-avr with 8-bit alignment, and everything seems to work completely fine.

[dylanmckay]

Most of the alignment stuff has been untouched since I originally imported the old SVN repository from SourceForge. I haven't dealt with it much and so my knowledge is pretty poor.

Safest to assume that if something looks intentional, it probably isn't 😜

@dylanmckay, do you have a ticket # or a repro description or anything on the problems you've encountered that triggered the above commit?

From memory, I think any atomic-related code would trigger it. I think the related issue is avr-llvm/llvm#214.

[gergoerdi]

I have now pushed the changes on both the LLVM and the Rust side that turn off all alignment.

[dylanmckay]

My check-everything-upstreamed.rb script found that I never ended up upstreaming the fix for this to LLVM proper, doing now

[dylanmckay]

Committed in r314179