[LLVMdev] implicit register usage

[Jim Grosbach]

On Feb 20, 2013, at 1:46 PM, Jakob Stoklund Olesen <stoklund at 2pi.dk> wrote:

> 
> On Feb 20, 2013, at 1:27 PM, Jim Grosbach <grosbach at apple.com> wrote:
> 
>> 
>> On Feb 20, 2013, at 1:22 PM, Jakob Stoklund Olesen <stoklund at 2pi.dk> wrote:
>> 
>>> 
>>> On Feb 20, 2013, at 11:44 AM, Jim Grosbach <grosbach at apple.com> wrote:
>>> 
>>>> I'm not sure. It certainly seems reasonable to me that the building could/should set default Define flags on operands based on the MCInstrDesc. Dunno why it doesn't.
>>> 
>>> Once you move beyond the fixed operands, it can't be determined automatically if the added register should be a use or a def.
>> 
>> Yes, very true. It seems the builder should use the information for the fixed operands, though? Or at least perhaps the machine verifier should check that they match what's actually on the operands (does it currently?). Right now it's pretty easy to construct an MI that's inconsistent with its MCInstrDesc information.
> 
> The verifier checks uses and defs when it can (i.e., the fixed ops).

OK, cool. That alleviates the vast majority of my concern.

> I am not a big fan of automagic features that only work some of the time. You tend to rely on them and not notice when they fail. They should at least fail loudly then.

That's an entirely fair argument.

> Imagine if Tablegen would infer the attributes of instructions most of the time, but just revert to guessing when the inference fails. We would get all kinds of weird attributes on our instructions without even noticing. Crazy! ;)
> 
Touche, sir!

> Are there cases where code calls BuildMI().addReg() without knowing if it is adding a use or a def? That seems like it would be rare.
> 

Yeah, it's pretty rare. I'm unconvinced there would realistically be cases that both don't know and can't be taught to know.


> We could add assertions inside MachineInstr::addOperand() to check the uses and defs against the MCInstrDesc. We do have code, though, that modifies an instruction in place before swizzling the descriptor. Such code could be fixed, I guess.

Yeah, switching the description is a touchy one. We take advantage of that a fair bit. On the other hand, I don't know that we ever do it in a situation that would change the Def flags.

I like the idea of adding assert()s to validate.

-Jim