The "m" memory constraint in inline assembly is broken (generates incorrect code or triggers LLVM asserts) and should not be used. Instead, indirect memory operands should be used with "\*m", "=\*m" and "+\*m".
Clang does this transparently by transforming "m" constraints into "\*m" indirect constraints, but for now just being able to use "\*m" directly is enough since asm! isn't stable.
While "\*m" works fine as an input operand, "=\*m" and "+\*m" need to be specified as input operands because they take a pointer value as an input. This PR relaxes the constraint checker to allow constraints starting with "=" or "+" if the constraint string contains a "\*", which indicates an indirect operand.
This (indirectly) fixes these issues: #29382, #16383 and #13366. The code will need to be changed to use "\*m" instead of "m".
Function arguments that are small aggregates get passed as integer types
instead. To correctly handle that, we need to use store_ty instead of
plain Store.
This branch implements a variant of trans that is based on MIR. It is very incomplete (intentionally), and had only the goal of laying out enough work to enable more incremental follow-on patches. Currently, only fns tagged with `#[rustc_mir]` use the new trans code. I plan to build up a meta-issue as well that tracks the various "not-yet-implemented" points. The only fn that has been tested so far is this amazingly complex "spike" fn:
```rust
#[rustc_mir]
fn sum(x: i32, y: i32) -> i32 {
x + y
}
```
In general, the most interesting commit is the last one. There are some points on which I would like feedback from @rust-lang/compiler:
- I did not use `Datum`. Originally, I thought that maybe just a `ValueRef` would be enough but I wound up with two very simple structures, `LvalueRef` and `OperandRef`, that just package up a `ValueRef` and a type. Because of MIR's structure, you don't wind up mixing by-ref and by-value so much, and I tend to think that a thinner abstraction layer is better here, but I'm not sure.
- Related to the above, I expect that sooner or later we will analyze temps (and maybe variables too) to find those whose address is never taken and which are word-sized and which perhaps meet a few other criteria. For those, we'll probably want to avoid the alloca, just because it means prettier code.
- I generally tried to re-use data structures from elsewhere in trans, though I'm sure we can trim these down.
- I didn't do any debuginfo primarily because it seems to want node-ids and we have only spans. I haven't really read into that code so I don't know what's going on there.
r? @nrc
`resolve_identifier` used to mark a variable as an upvar when used within a closure. However, the function is also used for the "unnecessary qualification" lint, which would mark paths whose last component had the same name as a local as upvars.
Fixes#29522
r? @eddyb
I think this should fix the test failures in debug mode from #29492
The assertion was written incorrectly, and I don't like the way the new assertion is written, but I _think_ it does the right thing now.
A line may be indented with either spaces or tabs, but not a mix of both. If there is a mix of tabs and spaces, only the kind that occurs first is counted.
This addresses issue #29268.
