librustc: Implement a Pod kind for types that can be memcpy'd.

[pcwalton]

This will be used for the new Cell.

r? @nikomatsakis

[nikomatsakis]

Can you add a run-pass test like the following:

use std::kinds::Pod;

// Test that type parameters with the Pod bound are implicitly copyable.

fn can_copy_POD<T:Pod>(v: T) {
    let _a = v;
    let _b = v;
}

fn main() { }

[nikomatsakis]

Also, should we put Pod in the prelude, like Send etc?

[pcwalton]

Pod is already in the prelude.

[pcwalton]

re-r? @nikomatsakis

[nikomatsakis]

Given that POD is implicitly copyable, I think you can simplify Cell like so:

diff --git a/src/libstd/cell.rs b/src/libstd/cell.rs
index 6fb4ce4..582e989 100644
--- a/src/libstd/cell.rs
+++ b/src/libstd/cell.rs
@@ -12,6 +12,7 @@

 use prelude::*;
 use cast;
+#[cfg(stage0)]
 use unstable::intrinsics;
 use util::NonCopyable;

@@ -65,18 +66,15 @@ impl<T: ::kinds::Pod> Cell<T> {
     /// Returns a copy of the contained value.
     #[inline]
     pub fn get(&self) -> T {
-        unsafe {
-            let mut result = intrinsics::uninit();
-            intrinsics::copy_nonoverlapping_memory(&mut result, &self.value, 1);
-            result
-        }
+        self.value // because this is POD, we can copy it
     }

     /// Sets the contained value.
     #[inline]
     pub fn set(&self, value: T) {
         unsafe {
-            intrinsics::copy_nonoverlapping_memory(cast::transmute_mut(&self.value), &value, 1)
+            let p: &mut T = cast::transmute_mut(&self.value);
+            *p = value;
         }
     }
 }

[nikomatsakis]

Also, I think that InteriorNonpod is not needed. It is only set when you own something owned anyway. So this patch should be equivalent I believe:

diff --git a/src/librustc/middle/ty.rs b/src/librustc/middle/ty.rs
index 43932c5..11cbf3a 100644
--- a/src/librustc/middle/ty.rs
+++ b/src/librustc/middle/ty.rs
@@ -1766,7 +1766,6 @@ def_type_content_sets!(

         // Things that are interior to the value (first nibble):
         InteriorUnsized                     = 0b0000__00000000__0001,
-        InteriorNonpod                      = 0b0000__00000000__0010,
         // InteriorAll                         = 0b0000__00000000__1111,

         // Things that are owned by the value (second and third nibbles):
@@ -1808,7 +1807,7 @@ def_type_content_sets!(
         Nonsized                            = 0b0000__00000000__0001,

         // Things that make values considered not POD
-        Nonpod                              = 0b0000__00001011__0010,
+        Nonpod                              = 0b0000__00001011__0000,

         // Bits to set when a managed value is encountered
         //
@@ -1882,8 +1881,7 @@ impl TypeContents {
          * Includes only those bits that still apply
          * when indirected through a `~` pointer
          */
-        TC::OwnsOwned |
-        TC::InteriorNonpod | (
+        TC::OwnsOwned | (
             *self & (TC::OwnsAll | TC::ReachesAll))
     }

@@ -2014,7 +2012,7 @@ pub fn type_contents(cx: ctxt, ty: t) -> TypeContents {
             }

             ty_estr(vstore_uniq) => {
-                TC::OwnsOwned | TC::InteriorNonpod
+                TC::OwnsOwned
             }

             ty_closure(ref c) => {
@@ -2133,7 +2131,7 @@ pub fn type_contents(cx: ctxt, ty: t) -> TypeContents {
                 match sigil {
                     ast::BorrowedSigil => TC::ReachesBorrowed,
                     ast::ManagedSigil => TC::Managed,
-                    ast::OwnedSigil => TC::OwnsOwned | TC::InteriorNonpod,
+                    ast::OwnedSigil => TC::OwnsOwned,
                 }
             }

[nikomatsakis]

Regarding the prelude, I wasn't seeing it there, but that's probably because I only built rustc-stage1 or something.

[chris-morgan]

At least a single mention of what POD stands for ("Plain Old Data", I am told) would be nice to have in the Pod doc comment.

[emberian]

Manual needs to be updated, and ideally the error for non-pod would say why. But, rules in the manual are a fine substitute.

[pcwalton]

Updated.

r? @nikomatsakis

[nikomatsakis]

r+ if we resolve the matter of whether @T should be considered pod (I'm inclined to say no until GC is tracing, but I'd prefer to add a FIXME and link to #2997)

[thestinger]

@nikomatsakis: even if it's tracing, it's not really POD as you're going to need some kind of special annotations + rooting API calls AFAICT

[nikomatsakis]

On Mon, Dec 16, 2013 at 02:40:26AM -0800, Daniel Micay wrote:

@nikomatsakis: even if it's tracing, it's not really POD as you're going to need some kind of special annotations + rooting API calls AFAICT

I guess that will depend on precisely what annotations are required
and precisely what we mean by POD. I think we can end up with a scheme
where managed data is considered POD, but since we don't have one now
it shouldn't be included.