llvm dev - Dec 2012 - [LLVMdev] Proposal: Adding aligned instruction bundle support to MC

> How will these bundles interact with ARM codegen? The constant island pass
in particular.
>
Hi Jim,

This is a great question.
>From the compiler's point of view, these bundles indeed pose a problem
for the constant island pass. At this point for the NaCl platform we
disable this pass, generating instructions like movt instead (NaCl
requires at least v7 for ARM). However, as a future enhancement we'd
like to follow gcc's path and adjust the pass to coexist with
potential bundling by computing worst-case estimates for the sizes of
bundle-locked instruction sequences.
>From the assembler's point of view, assembly code that refers to data
using labels should generate errors in the event that these labels
become too too far to reach, similarly to the current situation in
gas.

Eli

On Dec 6, 2012, at 2:45 PM, Eli Bendersky <eliben at google.com> wrote:
>> How will these bundles interact with ARM codegen? The constant island
pass in particular.
>> 
> 
> Hi Jim,
> 
> This is a great question.
> From the compiler's point of view, these bundles indeed pose a problem
> for the constant island pass. At this point for the NaCl platform we
> disable this pass, generating instructions like movt instead (NaCl
> requires at least v7 for ARM). However, as a future enhancement we'd
> like to follow gcc's path and adjust the pass to coexist with
> potential bundling by computing worst-case estimates for the sizes of
> bundle-locked instruction sequences.
That's going to be challenging. Not impossible, and you can probably
leverage the existing basic-block alignment code, but very tricky.

I'm curious. Have you done any code size comparisons using movw/movt instead
of constant pools? What was the impact?

-Jim
> 
> From the assembler's point of view, assembly code that refers to data
> using labels should generate errors in the event that these labels
> become too too far to reach, similarly to the current situation in
> gas.
> 
> Eli

>>
>> This is a great question.
>> From the compiler's point of view, these bundles indeed pose a
problem
>> for the constant island pass. At this point for the NaCl platform we
>> disable this pass, generating instructions like movt instead (NaCl
>> requires at least v7 for ARM). However, as a future enhancement
we'd
>> like to follow gcc's path and adjust the pass to coexist with
>> potential bundling by computing worst-case estimates for the sizes of
>> bundle-locked instruction sequences.
>
> That's going to be challenging. Not impossible, and you can probably
leverage the existing basic-block alignment code, but very tricky.
>
> I'm curious. Have you done any code size comparisons using movw/movt
instead of constant pools? What was the impact?
Even if we did, it was probably a while ago and they have to be done
again. We'll look into it when upstreaming the relevant parts.

Eli