Use out pointer more eagerly (return value optimisation)

[huonw]

The following pattern occurs reasonably often in the standard lib:

fn new() -> SomeStruct {
   let mut foo = SomeStruct { empty... };
   // ... initialise foo ...
   foo
}

which currently uses a memcpy at the end to copy the result into the out pointer, rather than just directly placing foo into it to begin with.

An example:

pub fn foo() -> [uint, .. 8] {
    [0, .. 8]
}
pub fn bar() -> [uint, .. 8] {
    let a = [0, .. 8];
    a
}

Compiling with -O -S --emit-llvm:

define void @_ZN3foo15_5274bbbd7427d03_00E([8 x i64]* nocapture, { i64, %tydesc*, i8*, i8*, i8 } addrspace(1)* nocapture) #1 {
static_allocas:
  %2 = bitcast [8 x i64]* %0 to i8*
  call void @llvm.memset.p0i8.i64(i8* %2, i8 0, i64 64, i32 8, i1 false)
  ret void
}

define void @_ZN3bar15_5274bbbd7427d03_00E([8 x i64]* nocapture, { i64, %tydesc*, i8*, i8*, i8 } addrspace(1)* nocapture) #1 {
static_allocas:
  %2 = alloca [8 x i64], align 8
  %3 = bitcast [8 x i64]* %2 to i8*
  call void @llvm.memset.p0i8.i64(i8* %3, i8 0, i64 64, i32 8, i1 false)
  %4 = bitcast [8 x i64]* %0 to i8*
  call void @llvm.memcpy.p0i8.p0i8.i64(i8* %4, i8* %3, i64 64, i32 8, i1 false)
  ret void
}

[mstewartgallus]

So I thought that this might be due to being only at optimization level 2 but even at optimization level 3 the code is still the same. By the way what does optimization level 3 actually enable over 2?

define void @_ZN3foo17_87e6ad83abde57f23_00E([8 x i64]* nocapture, { i64, %tydesc*, i8*, i8*, i8 } addrspace(1)* nocapture) #1 {
static_allocas:
   %2 = bitcast [8 x i64]* %0 to i8*
   call void @llvm.memset.p0i8.i64(i8* %2, i8 0, i64 64, i32 8, i1 false)
   ret void
}

define void @_ZN3bar17_87e6ad83abde57f23_00E([8 x i64]* nocapture, { i64, %tydesc*, i8*, i8*, i8 } addrspace(1)* nocapture) #1 {
static_allocas:
   %2 = alloca [8 x i64], align 8
   3 = bitcast [8 x i64]* %2 to i8*
   call void @llvm.memset.p0i8.i64(i8* %3, i8 0, i64 64, i32 8, i1 false)
   %4 = bitcast [8 x i64]* %0 to i8*
   call void @llvm.memcpy.p0i8.p0i8.i64(i8* %4, i8* %3, i64 64, i32 8, i1 false)
   ret void
}

[huonw]

If I'm reading back::passes correctly, --opt-level=3 adds argpromotion and loop-vectorize.

[thestinger]

This has regressed in LLVM after the issue was originally closed. Despite having both sret and dereferenceable, this does not optimize out.

[luqmana]

So it turn out that it's the lifetime markers that are stopping LLVM from optimizing this into just one memset. I have patch for LLVM upstream: http://reviews.llvm.org/D5020