Unions: Remove feature-gate for stable rustc #415

flukejones · 2017-07-21T22:55:14Z

Removes all feature-gating of union, and now requires rust 1.19
Updates readme to reflect this.

EPashkin · 2017-07-22T03:57:28Z

Travis build failure is legit

EPashkin · 2017-07-22T04:03:10Z

src/codegen/sys/statics.rs

    let v = vec![
-        u,
+        "#![feature(untagged_unions)]",


This line unneeded just "" instead

Ah, yes. I was using nightly rust still.

EPashkin · 2017-07-22T04:34:00Z

Only last problem with error: unions with non-Copy fields are unstable (see issue #32836)
I thought that we can just add #[derive(Clone, Copy)] to child structs, but in case of GDoubleIEEE754_mpn it won't works because _truncated_record_marker: c_void.
Maybe we need change its type to u8.
Also from gdk's "GdkEvent" seems we need derive Copy for many child structs and it inners (GdkRectangle) etc.,
so IMHO simpler derive Copy for all structs and possibly for unions.
@GuillaumeGomez you agree with this?
@sdroege As I remember you was for derive Copy from start?

flukejones · 2017-07-22T05:13:17Z

Not sure how to go about #[derive(Clone, Copy)]... To be honest, I don't think we should be adding this manually as this implies that we are checking each case to be sure all fields can be Copy - which we are not doing unless it is at the parser/output level.

Changing _truncated_record_marker: c_void to u8 would be a good idea I think. I'll work on removing the #![feature(untagged_unions)] and implementing #[derive(Copy, Clone)] for those particular types. Do you have any suggestions for tracking which ones need it?

EPashkin · 2017-07-22T05:26:14Z

@Luke-Nukem Lets wait for other say their options. For GdkEvent need or implement recursive check and update for all children or just derive for all structs.

flukejones · 2017-07-22T05:40:07Z

@EPashkin no problem. I do know we absolutely shouldn't (and can't) implement #[derive(Copy, Clone)] for all structs, it would introduce inefficiencies such as copying large data structures automatically instead of using a move. Looking at the structure of what we use to output the bindings, there doesn't seem to be anything easily used for it - unless perhaps we insert some flag at the parse stage?

EPashkin · 2017-07-22T05:50:55Z

Hm, I don't though that some structures can be large and Copy for these unreasonable.
Maybe we can add some config for Copy and for expanding enum.
Or really check\add Copy flag all unions children and hope that its will not cause problem.

EPashkin · 2017-07-22T06:25:35Z

last commit produce strange diff:

// Records
 #[repr(C)]
-pub struct GArray {
+GArraypub struct GArray {
     pub data: *mut c_char,
     pub len: c_uint,
 }

flukejones · 2017-07-22T06:27:01Z

Oh dear, looks like I'm making a mess of git. Hang on, I'll try and fix this up.

flukejones · 2017-07-22T06:34:53Z

The sys crates I've compiled so far seem to be fine now.

EPashkin · 2017-07-22T06:34:59Z

src/codegen/sys/lib_.rs

@@ -434,7 +434,7 @@ fn generate_records(w: &mut Write, env: &Env, records: &[&Record]) -> Result<()>
            if record.derive_copy {
                try!(writeln!(w, "{}#[derive(Copy,Clone)]", comment));
            }
-            try!(writeln!(w, "{0}pub struct {} {{", record.c_type));
+            try!(writeln!(w, "pub struct {} {{", record.c_type));


This not correct too, you lost comment text

🤦‍ 🤦‍ 🤦‍
Added back, squashed last series of commits. Should be good now.

EPashkin · 2017-07-22T06:52:37Z

Only last questionable problems with GdkEvent and GtkTextAppearance_u1, for this better wait opinions from others.

EPashkin · 2017-07-22T06:58:48Z

With GtkTextAppearance_u1 is problem that it used gdk::GdkRGBA so we can't add Copy for it automaticly.

EPashkin · 2017-07-22T07:06:45Z

Seems GtkTextAppearance_u1 actually unused as GtkTextAppearance is struct with bit-sized fields and truncated before union.

flukejones · 2017-07-22T07:33:39Z

Seems GtkTextAppearance_u1 actually unused as GtkTextAppearance is struct with bit-sized fields and truncated before union.

I guess that's just a side effect of how things are parsed then read out. Not sure what we can do about it other than just leave it. Or add in a special case for it? Do you think that is the only problem area?

sdroege · 2017-07-22T07:56:45Z

Just short comments

u8 or array thereof for bit fields makes sense. Just count the bits
I think deriving copy for all ffi structs is probably safe. As in, except for very few exceptions you only get them as pointers anyway. Using a configuration seems cleaner though, but OTOH we need to do that automatically for nested structs anyway. Copy for structs with pointers seems dangerous, but this is all unsafe code anyway so maybe OK. And in C, all structs are "Copy" too. Performance-wide I see no problem here in any case: if your code moved before, it will still do that. You might just copy without noticing, but that's only a problem if you have the whole struct and not just a pointer to it.

So in summary, I don't know. Making all Copy seems fine to me, a configuration also. Both are ugly in their own ways and in practice it probably makes no difference

flukejones · 2017-07-22T08:53:18Z

@sdroege the one problem with making all Copy though, is that there will be some things that shouldn't be or are inherently non-copy. I think c_void is one of those so if any structs have that as a field type then we'll have issues. Having said that I know c_void is u8 sized, and don't quite understand the purpose of c_void.

Copy on pointers... iirc raw pointers are fine and safe for copy, it's only when we deref them that it's unsafe. (?)

sdroege · 2017-07-22T16:59:48Z

For pointers, correct. You just need to be careful to free all copies later, I.e. do double-frees. But that's already with plain pointers

EPashkin · 2017-07-28T18:15:05Z

@GuillaumeGomez what you think about implementing Copy for almost all structs in sys crates to make unions work on stable?

GuillaumeGomez · 2017-07-28T18:20:27Z

That's something I proposed not that long ago so I'm totally for it!

EPashkin · 2017-07-28T18:39:31Z

@Luke-Nukem sorry for delay, I thought that I already asked.

flukejones · 2017-07-29T02:08:02Z

Sorry about the delay here, I'll push through an update soon. Was a bit tied up with study this week.

flukejones · 2017-07-29T02:32:26Z

I'm not sure about pub struct Volatile<T>(T); being copy,clone... Thoughts?

Appveyor is pretty far behind in rust versions isn't it...

flukejones · 2017-07-29T02:47:53Z

error[E0277]: the trait bound `[f64; 128]: std::clone::Clone` is not satisfied
    --> /home/travis/build/gtk-rs/gir/tests/sys/gdk-sys/src/lib.rs:3743:5
     |
3743 |     pub axes: [c_double; 128],
     |     ^^^^^^^^^^^^^^^^^^^^^^^^^ the trait `std::clone::Clone` is not implemented for `[f64; 128]`
     |
     = help: the following implementations were found:
               <[T; 28] as std::clone::Clone>
               <[T; 14] as std::clone::Clone>
               <[T; 7] as std::clone::Clone>
               <[T; 8] as std::clone::Clone>
             and 29 others
     = note: required by `std::clone::AssertParamIsClone`

Suggestions?

EPashkin · 2017-07-29T05:38:53Z

About old appveyour version: added fix #420

EPashkin · 2017-07-29T06:13:40Z

As gtk has single problem too, IMHO better add config parameter for object like "dont_copy=true"

error[E0277]: the trait bound `unsafe extern "C" fn(*mut GtkStyle, *mut cairo::cairo_t, GtkStateType, GtkShadowType, *mut GtkWidget, *const i8, i32, i32, i32, i32, GtkPositionType, i32, i32): std::clone::Clone` is not satisfied
    --> d:/eap/rust/0/sys/gtk-sys\src\lib.rs:5295:5
     |
5295 |     pub draw_shadow_gap: Option<unsafe extern "C" fn(*mut GtkStyle, *mut cairo::cairo_t, GtkStateType, GtkShadowType, *mut GtkWidget, *const c_char, c_int, c_int, c_int, c_int, GtkPositionType, c_int, c_int)>,
     |     ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ the trait `std::clone::Clone` is not implemented for `unsafe extern "C" fn(*mut GtkStyle, *mut cairo::cairo_t, GtkStateType, GtkShadowType, *mut GtkWidget, *const i8, i32, i32, i32, i32, GtkPositionType, i32, i32)`
     |
     = note: required because of the requirements on the impl of `std::clone::Clone` for `std::option::Option<unsafe extern "C" fn(*mut GtkStyle, *mut cairo::cairo_t, GtkStateType, GtkShadowType, *mut GtkWidget, *const i8, i32, i32, i32, i32, GtkPositionType, i32, i32)>`
     = note: required by `std::clone::AssertParamIsClone`

EPashkin · 2017-07-29T06:14:47Z

src/codegen/sys/lib_.rs

@@ -574,7 +516,7 @@ fn generate_fields(env: &Env, struct_name: &str, fields: &[Field]) -> (Vec<Strin
            if c_type.is_err() {
                commented = true;
            }
-            if !cfg!(feature = "use_unions") && is_gvalue && field.name == "data" {
+            if is_gvalue && field.name == "data" {


Why this don't removed?

EPashkin · 2017-07-29T06:14:55Z

src/codegen/sys/lib_.rs

@@ -565,7 +507,7 @@ fn generate_fields(env: &Env, struct_name: &str, fields: &[Field]) -> (Vec<Strin
            continue 'fields;
        }

-        if is_gweakref && !cfg!(feature = "use_unions") {
+        if is_gweakref {


Why this don't removed?

EPashkin · 2017-07-29T06:29:23Z

src/codegen/sys/lib_.rs

@@ -444,7 +417,7 @@ fn generate_records(w: &mut Write, env: &Env, records: &[&Record]) -> Result<()>
        if lines.is_empty() {
            try!(writeln!(
                w,
-                "{}#[repr(C)]\n{0}pub struct {}(c_void);\n",
+                "{0}#[repr(C)]\n{0}#[derive(Copy,Clone)]\n{0}pub struct {1}(u8);\n",


This better changed back (it will produce only 1 extra "dont_copy")

..because it just don't right, copy record with unknown size

I'm not sure I understand, pub struct GIConv(u8); should not be copy?

Yes, with GIOChannel

Not sure I understand either. What is the broken code generated by this?

this case produce warning: found non-foreign-function-safe member in struct marked #[repr(C)]: found Rust tuple type in foreign module; consider using a struct instead
just struct produce warning: found zero-size struct in foreign module, consider adding a member to this struct

I see... a case for the new #[repr(transparent)] I guess

Sorry, now I don't understand how it related to struct without fields.

You would want to represent an opaque C type. It's not really a struct without fields, it's just that the struct definition is unknown from the headers. So you can only ever use it as a pointer.

But that's actually extern type: rust-lang/rfcs#1861

extern type is that we need for this case but it not realized and may cause build error when inserted in other struct directly (as with GIConv)
#[repr(transparent)] can be used for minimize alignment errors for known fields, so also not for this case, so last repr(C).
I prefer c_void over u8 as it don't implement Copy while has same size and alignment.

EPashkin · 2017-07-29T06:36:01Z

About Volatile: as it applied only to simple types it can be copy.

EPashkin · 2017-07-29T06:54:50Z

src/codegen/sys/lib_.rs

+            if lines.is_empty() {
+                try!(writeln!(
+                    w,
+                    "{0}#[repr(C)]\n{0}#[derive(Copy,Clone)]\n{0}pub union {1}(c_void);\n",


Copy better removed as for records
PS. this code actually unused.

flukejones · 2017-07-29T06:56:38Z

config parameter for object like "dont_copy=true"

@EPashkin can you advise where I would look for examples of how to do this?

EPashkin · 2017-07-29T07:17:31Z

Add field to GObject same as https://github.com/gtk-rs/gir/blob/master/src/config/gobjects.rs#L151
then get it in generate_records

let full_name = format!("{}.{}", env.namespaces.main().name, record.name);
let config = env.config.objects.get(&full_name);

EPashkin · 2017-07-29T07:33:03Z

.. As config can be None for ffi, use if-let or config.map(|o| o.dont_copy).unwrap_or(false) to get value.

flukejones · 2017-07-30T04:41:03Z

I need to put this on the backlog for a week or so while I do a few other things of importance. I may work on it off/on through the week however, we'll see.

sdroege · 2017-09-18T08:03:37Z

What's the status here?

flukejones · 2017-09-18T08:08:47Z

I've not had much free time. I'll try and get a look at it again this week, I should be able to juggle my schedule a bit.

…

On 18/09/2017 8:03 PM, "Sebastian Dröge" ***@***.***> wrote: What's the status here? — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#415 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AB7Cfsc3v_nuFfAeRZFZYm3CzjgQcSsuks5sjiPagaJpZM4Of-Go> .

flukejones · 2017-10-25T08:41:54Z

@sdroege @EPashkin I've done a rebase of this on to master and run a couple of small gir tests - seems okay.

Looking back at previous comments it looks like I still have a few things to do such as "dont_copy=true". Is this still a blocker?

sdroege · 2017-10-25T08:59:13Z

IMHO just having copy on all FFI things is fine, configuration can be added later once needed :)

flukejones · 2017-10-25T09:07:12Z

Okidoki, then this may be good to go if edge cases are okay.

…

On 25 October 2017 at 21:59, Sebastian Dröge ***@***.***> wrote: IMHO just having copy on all FFI things is fine, configuration can be added later once needed :) — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#415 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AB7CfrQG3LWkO6JcR6tH2eXYFTh7M9SFks5svvhigaJpZM4Of-Go> .

flukejones · 2017-11-13T04:14:54Z

Tests will pass with this commit to cairo.

I get the impression that this is going to require a more in-depth solution rather than just implement Copy,Clone for all. Perhaps by marking unions in codegen and everything branching from it, and only derive/implement Copy,Clone for those, along with setting up the correct sized type to replace c_void for each architecture (c_void does not allow Copy).

EPashkin reviewed Jul 22, 2017

View reviewed changes

flukejones force-pushed the master branch 3 times, most recently from 83d94a7 to e9a7910 Compare July 22, 2017 06:24

EPashkin reviewed Jul 22, 2017

View reviewed changes

flukejones force-pushed the master branch from 4b5b88e to a703885 Compare July 22, 2017 06:39

EPashkin reviewed Jul 29, 2017

View reviewed changes

flukejones force-pushed the master branch from b86b995 to c199ba4 Compare July 29, 2017 06:59

flukejones force-pushed the master branch from c199ba4 to ac06389 Compare October 25, 2017 08:38

Luke Jones added 3 commits November 13, 2017 10:19

Unions: Remove feature-gate for stable rustc

21a7e4f

Unions: make all structs and unions copy,clone

a5e2bbb

Unions: tweaks to copy,clone, and remove unneeded if/else

aeefc4d

flukejones force-pushed the master branch from ac06389 to 612089f Compare November 13, 2017 00:37

unions: rebase & fix copy,clone for unions

8560db7

flukejones force-pushed the master branch from 612089f to 8560db7 Compare November 13, 2017 00:39

flukejones mentioned this pull request Nov 13, 2017

unions: cairo_pattern_t to u8 gtk-rs/cairo#166

Merged

tests: update some crate versions, eg, bitflags

0e6665f

ghost mentioned this pull request Jan 29, 2018

Union, Clone and Copy #527

Merged

EPashkin merged commit 0e6665f into gtk-rs:master Feb 1, 2018

ghost mentioned this pull request Mar 10, 2018

Add support for unions #334

Closed

sdroege mentioned this pull request Feb 4, 2021

Don't derive Copy for FFI structs/unions that contain pointers #1048

Open

Unions: Remove feature-gate for stable rustc #415

Unions: Remove feature-gate for stable rustc #415

Conversation

flukejones commented Jul 21, 2017

EPashkin commented Jul 22, 2017

Choose a reason for hiding this comment

flukejones Jul 22, 2017 • edited Loading

Choose a reason for hiding this comment

EPashkin commented Jul 22, 2017

flukejones commented Jul 22, 2017

EPashkin commented Jul 22, 2017

flukejones commented Jul 22, 2017

EPashkin commented Jul 22, 2017

EPashkin commented Jul 22, 2017

flukejones commented Jul 22, 2017

flukejones commented Jul 22, 2017

Choose a reason for hiding this comment

Choose a reason for hiding this comment

EPashkin commented Jul 22, 2017

EPashkin commented Jul 22, 2017

EPashkin commented Jul 22, 2017

flukejones commented Jul 22, 2017

sdroege commented Jul 22, 2017

flukejones commented Jul 22, 2017

sdroege commented Jul 22, 2017

EPashkin commented Jul 28, 2017

GuillaumeGomez commented Jul 28, 2017

EPashkin commented Jul 28, 2017

flukejones commented Jul 29, 2017

flukejones commented Jul 29, 2017 • edited Loading

flukejones commented Jul 29, 2017

EPashkin commented Jul 29, 2017

EPashkin commented Jul 29, 2017

Choose a reason for hiding this comment

Choose a reason for hiding this comment

EPashkin Jul 29, 2017 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

EPashkin commented Jul 29, 2017

Choose a reason for hiding this comment

flukejones commented Jul 29, 2017

EPashkin commented Jul 29, 2017 • edited Loading

EPashkin commented Jul 29, 2017 • edited Loading

flukejones commented Jul 30, 2017

sdroege commented Sep 18, 2017

flukejones commented Sep 18, 2017 via email

flukejones commented Oct 25, 2017

sdroege commented Oct 25, 2017

flukejones commented Oct 25, 2017 via email

flukejones commented Nov 13, 2017 • edited Loading

flukejones Jul 22, 2017 •

edited

Loading

flukejones commented Jul 29, 2017 •

edited

Loading

EPashkin Jul 29, 2017 •

edited

Loading

EPashkin commented Jul 29, 2017 •

edited

Loading

EPashkin commented Jul 29, 2017 •

edited

Loading

flukejones commented Nov 13, 2017 •

edited

Loading