Declare by-value on-stack parameters to be noalias

[dotdash]

Function parameters that are to be passed by value but don't fit into a
single register are currently passed by creating a copy on the stack and
passing a pointer to that copy to the callee. Since the copy is made
just for the function call, there are no aliases.

For example, this sometimes allows LLVM to eliminate unnecessary calls
to drop glue. Given

struct Foo {
    a: int,
    b: Option<~str>,
}

extern {
    fn eat(eat: Option<~str>);
}

pub fn foo(v: Foo) {
    match v {
        Foo { a: _, b } => unsafe { eat(b) }
    }
}

LLVM currently can't eliminate the drop call for the string, because it
only sees a pointer to Foo, for which it has to expect an alias. So we
get:

; Function Attrs: uwtable
define void @_ZN3foo20h9f32c90ae7201edbxaa4v0.0E(%struct.Foo* nocapture) unnamed_addr #0 {
"_ZN34std..option..Option$LT$$UP$str$GT$9glue_drop17hc39b3015f3b9c69dE.exit":
  %1 = getelementptr inbounds %struct.Foo* %0, i64 0, i32 1, i32 0
  %2 = load { i64, i64, [0 x i8] }** %1, align 8
  store { i64, i64, [0 x i8] }* null, { i64, i64, [0 x i8] }** %1, align 8
  %3 = ptrtoint { i64, i64, [0 x i8] }* %2 to i64
  %.fca.0.insert = insertvalue { i64 } undef, i64 %3, 0
  tail call void @eat({ i64 } %.fca.0.insert)
  %4 = load { i64, i64, [0 x i8] }** %1, align 8
  %5 = icmp eq { i64, i64, [0 x i8] }* %4, null
  br i1 %5, label %_ZN3Foo9glue_drop17hf611996539d3036fE.exit, label %"_ZN8_$UP$str9glue_drop17h15dbdbe2b8897a98E.exit.i.i"

"_ZN8_$UP$str9glue_drop17h15dbdbe2b8897a98E.exit.i.i": ; preds = %"_ZN34std..option..Option$LT$$UP$str$GT$9glue_drop17hc39b3015f3b9c69dE.exit"
  %6 = bitcast { i64, i64, [0 x i8] }* %4 to i8*
  tail call void @free(i8* %6) #1
  br label %_ZN3Foo9glue_drop17hf611996539d3036fE.exit

_ZN3Foo9glue_drop17hf611996539d3036fE.exit:       ; preds = %"_ZN34std..option..Option$LT$$UP$str$GT$9glue_drop17hc39b3015f3b9c69dE.exit", %"_ZN8_$UP$str9glue_drop17h15dbdbe2b8897a98E.exit.i.i"
  ret void
}

But with the noalias attribute, it can safely optimize that to:

define void @_ZN3foo20hd28431f929f0d6c4xaa4v0.0E(%struct.Foo* noalias nocapture) unnamed_addr #0 {
_ZN3Foo9glue_drop17he9afbc09d4e9c851E.exit:
  %1 = getelementptr inbounds %struct.Foo* %0, i64 0, i32 1, i32 0
  %2 = load { i64, i64, [0 x i8] }** %1, align 8
  store { i64, i64, [0 x i8] }* null, { i64, i64, [0 x i8] }** %1, align 8
  %3 = ptrtoint { i64, i64, [0 x i8] }* %2 to i64
  %.fca.0.insert = insertvalue { i64 } undef, i64 %3, 0
  tail call void @eat({ i64 } %.fca.0.insert)
  ret void
}

[flaper87]

I think we need to have more codegen tests for this kind of changes. Getting to 1.0 and introducing backwards compatibility also means having a good way to avoid introducing regressions for this kind of optimization.

I'd like to see one for this fix too.