a couple trait method call optimizations

[eugals]

This is my first PR to the project.
It is intended to optimize/beautify the code generated in a few trivial trait operations.
Let's take the following code as an example:

trait Stuff {
    fn bar(&self);
}

fn callBar(s: &Stuff) {
    s.bar();
}

struct Foo;

impl Stuff for Foo {
    fn bar(&self) {
    }
}

pub fn main() {
    let o = Foo;
    callBar(&o as &Stuff);
}

At present it is translated into something like:

define void @_ZN7callBar_UUID.0E({ i32, %tydesc*, i8*, i8*, i8 }*, { %tydesc*, i8* }*) #4 {
"function top level":
  %__trait_callee = alloca { %tydesc*, i8* }
  %__auto_borrow_obj = alloca { %tydesc*, i8* }
  %2 = getelementptr inbounds { %tydesc*, i8* }* %1, i32 0, i32 0
  %3 = load %tydesc** %2
  %4 = getelementptr inbounds { %tydesc*, i8* }* %__auto_borrow_obj, i32 0, i32 0
  store %tydesc* %3, %tydesc** %4
  %5 = getelementptr inbounds { %tydesc*, i8* }* %1, i32 0, i32 1
  %6 = load i8** %5
  %7 = getelementptr inbounds { %tydesc*, i8* }* %__auto_borrow_obj, i32 0, i32 1
  store i8* %6, i8** %7
  %8 = bitcast { %tydesc*, i8* }* %__auto_borrow_obj to i8*
  %9 = bitcast { %tydesc*, i8* }* %__trait_callee to i8*
  call void @llvm.memcpy.p0i8.p0i8.i32(i8* %9, i8* %8, i32 8, i32 4, i1 false)
  %10 = getelementptr inbounds { %tydesc*, i8* }* %__trait_callee, i32 0, i32 1
  %11 = load i8** %10
  %12 = bitcast i8* %11 to { i32, %tydesc*, i8*, i8*, i8 }*
  %13 = getelementptr inbounds { %tydesc*, i8* }* %__trait_callee, i32 0, i32 0
  %14 = bitcast %tydesc** %13 to [1 x i8*]**
  %15 = load [1 x i8*]** %14
  %16 = getelementptr inbounds [1 x i8*]* %15, i32 0, i32 1
  %17 = load i8** %16
  %18 = bitcast i8* %17 to void ({ i32, %tydesc*, i8*, i8*, i8 }*)*
  call void %18({ i32, %tydesc*, i8*, i8*, i8 }* %12)
  ret void
}

...

define void @_ZN4main_UUID.0E({ i32, %tydesc*, i8*, i8*, i8 }*) #4 {
"function top level":
  %o = alloca %struct.Foo
  %1 = alloca { %tydesc*, i8* }
  %__auto_borrow_obj = alloca { %tydesc*, i8* }
  %2 = getelementptr inbounds { %tydesc*, i8* }* %1, i32 0, i32 1
  %3 = bitcast i8** %2 to %struct.Foo**
  store %struct.Foo* %o, %struct.Foo** %3
  %4 = getelementptr inbounds { %tydesc*, i8* }* %1, i32 0, i32 0
  %5 = bitcast %tydesc** %4 to { %tydesc*, void ({ i32, %tydesc*, i8*, i8*, i8 }*)* }**
  store { %tydesc*, void ({ i32, %tydesc*, i8*, i8*, i8 }*)* }* @vtable1081, { %tydesc*, void ({ i32, %tydesc*, i8*, i8*, i8 }*)* }** %5
  %6 = getelementptr inbounds { %tydesc*, i8* }* %1, i32 0, i32 0
  %7 = load %tydesc** %6
  %8 = getelementptr inbounds { %tydesc*, i8* }* %__auto_borrow_obj, i32 0, i32 0
  store %tydesc* %7, %tydesc** %8
  %9 = getelementptr inbounds { %tydesc*, i8* }* %1, i32 0, i32 1
  %10 = load i8** %9
  %11 = getelementptr inbounds { %tydesc*, i8* }* %__auto_borrow_obj, i32 0, i32 1
  store i8* %10, i8** %11
  call void @_ZN7callBar_UUID.0E({ i32, %tydesc*, i8*, i8*, i8 }* undef, { %tydesc*, i8* }* %__auto_borrow_obj)
  ret void
}

If you apply my patch, it would become way shorter and cleaner:

define void @_ZN7callBar_UUID.0E({ i32, %tydesc*, i8*, i8*, i8 }*, { %tydesc*, i8* }*) #4 {
"function top level":
  %2 = getelementptr inbounds { %tydesc*, i8* }* %1, i32 0, i32 1
  %3 = load i8** %2
  %4 = bitcast i8* %3 to { i32, %tydesc*, i8*, i8*, i8 }*
  %5 = getelementptr inbounds { %tydesc*, i8* }* %1, i32 0, i32 0
  %6 = bitcast %tydesc** %5 to [1 x i8*]**
  %7 = load [1 x i8*]** %6
  %8 = getelementptr inbounds [1 x i8*]* %7, i32 0, i32 1
  %9 = load i8** %8
  %10 = bitcast i8* %9 to void ({ i32, %tydesc*, i8*, i8*, i8 }*)*
  call void %10({ i32, %tydesc*, i8*, i8*, i8 }* %4)
  ret void
}

...

define void @_ZN4main_UUID.0E({ i32, %tydesc*, i8*, i8*, i8 }*) #4 {
"function top level":
  %o = alloca %struct.Foo
  %1 = alloca { %tydesc*, i8* }
  %2 = getelementptr inbounds { %tydesc*, i8* }* %1, i32 0, i32 1
  %3 = bitcast i8** %2 to %struct.Foo**
  store %struct.Foo* %o, %struct.Foo** %3
  %4 = getelementptr inbounds { %tydesc*, i8* }* %1, i32 0, i32 0
  %5 = bitcast %tydesc** %4 to { %tydesc*, void ({ i32, %tydesc*, i8*, i8*, i8 }*)* }**
  store { %tydesc*, void ({ i32, %tydesc*, i8*, i8*, i8 }*)* }* @vtable1081, { %tydesc*, void ({ i32, %tydesc*, i8*, i8*, i8 }*)* }** %5
  call void @_ZN7callBar_UUID.0E({ i32, %tydesc*, i8*, i8*, i8 }* undef, { %tydesc*, i8* }* %1)
  ret void
}

Although this change doesn't increase the compilation speed much (I mentioned only about 1-2% boost on "rustc -O -Z time-passes syntax.rs"), but I still think it's a good thing to do as it greatly simplifies/clarifies LL generated in some cases which would definitely help in the future code generation investigations.

I don't provide any new test cases in this patch as it is merely an optimization.
I tried to do "make check" and everything passed well except "run-pass/core-run-destroy.rs", but as far as I can see it doesn't pass on the current branch/master too.

[eugals]

Ouch. Sorry I've merged it wrong