Walter's talk on D backend

[user1234]

On Monday, 23 September 2024 at 11:49:27 UTC, claptrap wrote:
> On Monday, 23 September 2024 at 11:03:45 UTC, user1234 wrote:
>> On Monday, 23 September 2024 at 09:43:50 UTC, claptrap wrote:
>
>> I cant reply for DMD but as I detect a confusion, let me add 
>> something about that... SSA is not about variables it's about 
>> nodes assigned uniquely once
>>
>> So if X is a variable then each use of it is either a load or 
>> a store. (there can be bit cast and ptr arithmetic too, but 
>> let keep thing simple.)
>>
>> 
>> (pseudo LLVM IR, to illustrate, with X, a global uint)
>>
>> ```
>> %1 = i32  load  ptr @X
>> %3 = i32  add   i32 %1, i32 %1
>>      i32  store ptr @X, i32 %3
>>     (LLVM does not make store reuseable, that'd be be a 
>> non-sense)
>> 
>>
>> You see variables are not accessed as-is, either you have a 
>> load or a store.
>> An SSA node is what starts with "%"
>> ```
>> 
>
> That's counter to everything I've read about it. It is defined 
> as "variables are only assigned once", so each time you assign 
> a new value to X, it gets a suffix, so each assignment to X 
> becomes a new variable X0,X1... and so on.
>
> It makes it explicit what definition reaches what uses of X. 
> Even through all the control flow. That's actually what makes 
> SSA useful.
>
> A quick google says LLVM IR is SSA for registers but not memory 
> locations. So I don't think your example shows anything.
>
> What would show SSA is if you cant assign to a register 
> multiple times. Eg..
>
> ```
> %1 = i32  load  ptr @X
> %2 = i32  load  ptr @Y
> %2 = i32  add   i32 %1, i32 %2
> ```
>
> That would be invalid in SSA

Yes. That's impossible. Also the reasoning that each SSA 
instruction is a state of a register is not totally exact (that's 
not yours tho). At the IR stage the notion of register does not 
exist *yet*. However it's true that it's often the case.

Otherwise, my example showed that reasoning at the level of a 
variable is not relevant, for DCE at least. Actually an unused 
load should never be generated.
What *can* happen is something like

```
X + 1; // %1 = load ...
        // %2 = add
```

then DCE will realize that %2 is never used and drop %2 and %1.

IIRC D detects such cases at the front-end level and emits an 
error: "expression <...> has not effect".