librustc: Don't create extra alloca slot for by value bindings in match.

[luqmana]

structWith{x:int,f:NoCopy}#[no_mangle]fnbar(){let mine = With{x:3,f:NoCopy};match mine {
c => {foo(c);}}}#[no_mangle]fnfoo(_:With){}

Before:

defineinternalvoid@bar() unnamed_addr #1 {
entry-block:
%mine = alloca %"struct.With<[]>"%__llmatch = alloca %"struct.With<[]>"*
%c = alloca %"struct.With<[]>"%0 = getelementptrinbounds %"struct.With<[]>"* %mine, i320, i320storei643, i64*%0%1 = getelementptrinbounds %"struct.With<[]>"* %mine, i320, i321store %"struct.With<[]>"* %mine, %"struct.With<[]>"** %__llmatchbrlabel%case_body
case_body: ; preds = %entry-block%2 = load %"struct.With<[]>"** %__llmatch%3 = bitcast %"struct.With<[]>"* %2toi8*%4 = bitcast %"struct.With<[]>"* %ctoi8*callvoid@llvm.memcpy.p0i8.p0i8.i64(i8*%4, i8*%3, i648, i328, i1false)
%5 = load %"struct.With<[]>"* %ccallvoid@foo(%"struct.With<[]>"%5)
brlabel%join
join: ; preds = %case_bodyretvoid
}

After:

defineinternalvoid@bar() unnamed_addr #1 {
entry-block:
%mine = alloca %"struct.With<[]>"%c = alloca %"struct.With<[]>"*
%0 = getelementptrinbounds %"struct.With<[]>"* %mine, i320, i320storei643, i64*%0%1 = getelementptrinbounds %"struct.With<[]>"* %mine, i320, i321store %"struct.With<[]>"* %mine, %"struct.With<[]>"** %cbrlabel%case_body
case_body: ; preds = %entry-block%2 = load %"struct.With<[]>"** %c%3 = load %"struct.With<[]>"* %2callvoid@foo(%"struct.With<[]>"%3)
brlabel%join
join: ; preds = %case_bodyretvoid
}

r? @pcwalton

[pcwalton]

Does this correctly handle:

let mut x = 3;
match x {
mut y => { y = 5; }
}

In that case you must actually create an extra slot.

[pcwalton]

(Actually I don't think you need to create an extra slot per binding, but you should create one for the matched value.)

[pcwalton]

@luqmana I think what you need to do is to make a temporary copy of the value in the head of the match if and only if the matched value is Copy AND (there are mutable by-value bindings OR the value is not Share). That's because Cells can mess things up too. For a first cut it may be simpler to just unconditionally make a temporary copy if the value is Copy.

This is great work BTW, thanks!

[pcwalton]

I think this may be wrong in some cases: what about

match box Foo { x: String::new("hi") } {
box Foo { x } => ... use x ...
}

This might result in use-after-free. I think we may have to restrict the optimization to cases in which it's a local variable or an rvalue guaranteed to be on the stack that you're moving out of.

[pcwalton]

@bors: retry

[huonw]

This apparently caused a 1 GB memory regression!

[huonw]

(IRC informs me that #15442 is the fix; nice catch @dotdash, nice fix @luqmana.)

[lilyball]

I think this is the cause for the double-drop seen in #16151.