[llvm-dev] RFC: Harvard architectures and default address spaces

[Dylan McKay via llvm-dev]

> Very nice if something that solves this issue for real could make it into
tree.

I'll keep you in the loop, thanks for the attached patch by the way.

Cheers,
Dylan

On Wed, Jul 12, 2017 at 1:39 AM, Mikael Holmén <mikael.holmen at ericsson.com>
wrote:

> Hi,
>
> We have an out-of-tree target that is also Harvard architecture, so we're
> interested in this as well.
>
> So far, we've "solved" the issue by extending the datalayout so we can
> specify our "function pointer address space" there and then add/use it when
> necessary. Our current patch for this is small but hacky (too hacky to be
> upstreamed), but it has done the job for us for quite some time.
>
> Very nice if something that solves this issue for real could make it into
> tree.
>
> Regards,
> Mikael
>
> On 07/11/2017 07:54 AM, Dylan McKay via llvm-dev wrote:
>
>> Hello all, I’m looking into solving an AVR-specific issue and would love
>> to hear peoples thoughts on how to best fix it.
>>
>>
>>   Background
>>
>> As you may or may not know, I maintain the in-tree AVR backend, which
>> also happens to be (to the best of my knowledge) the first in-tree backend
>> for a Harvard architecture.
>>
>> In this architecture, code lives inside the ‘program memory’ space
>> (numbered 1), whereas data lives inside RAM “data space”, which corresponds
>> to the default address space 0. This is important because loads/stores use
>> different instruction/pointer formats depending on the address space used,
>> and so we need correct address space information available to the backend
>> itself.
>>
>> Due to the fact that address spaces in LLVM default to 0, this means that
>> all global or constant variables default to living inside data space. This
>> causes a few issues, including the fact that the SimplifyCFG pass creates
>> switch lookup tables, which default to data space, causing us to emit
>> broken table lookups and also wasting precious RAM.
>>
>>
>>   The problem - emitting pointers as operands
>>
>> *NOTE*: Feel free to skip to tl;dr of this section if you don’t care too
>> much about the details
>>
>> There are different instructions which require different fixups to be
>> applied depending on whether pointers are located in data space or program
>> space.
>>
>> Take the |ICALL| instruction - it performs an indirect call to the
>> pointer stored in the |Z| register.
>>
>> We must first load the pointer into |Z| via the ‘ldi’ instruction. If the
>> pointer is actually a pointer to a symbol, we need to emit a
>> |AVR_LO8_LDI_GS| relocation, otherwise we emit a |AVR_LO8_LDI| relocation.
>> There are a few other cases, but they’re irrelevant for this discussion.
>>
>> We can quite easily look at the |GlobalValue*| that corresponds to the
>> pointer if it is a symbol and select the fixup based on that, but that
>> assumes that the address spaces are always correct.
>>
>> Now, imagine that the pointer is actually a function pointer. LLVM does
>> not expose any way to set address space in the IR for functions, but
>> because it derived from GlobalValue, it does have an address space, and
>> that address space defaults to zero. Because of this, every single function
>> pointer in the AVR backend that gets loaded by the |ldi| will be associated
>> with data space, and not program space, which it actually belongs to.
>>
>> *tl;dr* functions default to address space zero, even though they are in
>> a different space on Harvard architectures, which causes silent codegen
>> bugs when we rely on the address space of a global value
>>
>>
>>   Proposed solution
>>
>> It would be impossible to set the address space correctly on creation of
>> |llvm::Function| objects because at that point in the pipeline, we do not
>> know the target architecture.
>>
>> Because of this, I’d like to extend |TargetTransformInfo| with hooks that
>> like |getSwitchTableAddressSpace()|, |getFunctionAddressSpace()|. I have
>> already got a WIP patch for this here <https://reviews.llvm.org/D34983>.
>>
>> Once we have that information available to TargetTransformInfo, I propose
>> we add a pass (very early in the codegen pipeline) that sets the address
>> space of all functions to whatever value is specified in the hooks.
>>
>> This works well because we don’t let frontends specify address space on
>> functions, nor do we even mention that functions have address spaces in the
>> language reference.
>>
>> The downside of it it is that you wouldn’t normally expect something like
>> an address space to change midway through the compilation process. To
>> counter that however, I doubt the pre-codegen code cares much about the
>> value of function address spaces, if at all.
>>
>> On top of this, at the current point in time,
>> |Pointer<Function>::getAddressSpace| is downright incorrect on any
>> Harvard architecture, and for other architectures, the address space for
>> functions will still stay the default of zero and will not change at all.
>>
>> Does anybody know anything I haven’t thought of? Any reasons why this
>> solution is suboptimal?
>>
>> ​
>>
>>
>> _______________________________________________
>> LLVM Developers mailing list
>> llvm-dev at lists.llvm.org
>> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
>>
>>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20170712/64d64e64/attachment-0001.html>