[PATCH] rust: add intrinsics to fix `-Os` builds

[PATCH] rust: add intrinsics to fix `-Os` builds

[Miguel Ojeda]

Alice reported [1] that an arm64 build failed with:

    ld.lld: error: undefined symbol: __extendsfdf2
    >>> referenced by core.a6f5fc5794e7b7b3-cgu.0
    >>>               rust/core.o:(<f32>::midpoint) in archive vmlinux.a
    >>> referenced by core.a6f5fc5794e7b7b3-cgu.0
    >>>               rust/core.o:(<f32>::midpoint) in archive vmlinux.a

    ld.lld: error: undefined symbol: __truncdfsf2
    >>> referenced by core.a6f5fc5794e7b7b3-cgu.0
    >>>               rust/core.o:(<f32>::midpoint) in archive vmlinux.a

Rust 1.80.0 or later together with `CONFIG_CC_OPTIMIZE_FOR_SIZE=y`
is what triggers it.

In addition, x86_64 builds also fail the same way.

Similarly, compiling with Rust 1.82.0 (currently in nightly) makes
another one appear, possibly due to the LLVM 19 upgrade there:

    ld.lld: error: undefined symbol: __eqdf2
    >>> referenced by core.20495ea57a9f069d-cgu.0
    >>>               rust/core.o:(<f64>::next_up) in archive vmlinux.a
    >>> referenced by core.20495ea57a9f069d-cgu.0
    >>>               rust/core.o:(<f64>::next_down) in archive vmlinux.a

Gary adds [1]:

> Usually the fix on rustc side is to mark those functions as `#[inline]`
>
> All of {midpoint,next_up,next_down} are indeed unstable functions not
> marked as inline...

Fix all those by adding those intrinsics to our usual workaround.

Cc: Gary Guo <gary@garyguo.net>
Reported-by: Alice Ryhl <aliceryhl@google.com>
Closes: https://rust-for-linux.zulipchat.com/#narrow/stream/288089-General/topic/v6.2E11-rc1.20doesn't.20build.20for.20arm64/near/455637364
Signed-off-by: Miguel Ojeda <ojeda@kernel.org>
---
 rust/Makefile             | 4 ++--
 rust/compiler_builtins.rs | 3 +++
 2 files changed, 5 insertions(+), 2 deletions(-)

diff --git a/rust/Makefile b/rust/Makefile
index 1f10f92737f2..77836388377d 100644
--- a/rust/Makefile
+++ b/rust/Makefile
@@ -354,8 +354,8 @@ rust-analyzer:
 		$(if $(KBUILD_EXTMOD),$(extmod_prefix),$(objtree))/rust-project.json
 
 redirect-intrinsics = \
-	__addsf3 __eqsf2 __gesf2 __lesf2 __ltsf2 __mulsf3 __nesf2 __unordsf2 \
-	__adddf3 __ledf2 __ltdf2 __muldf3 __unorddf2 \
+	__addsf3 __eqsf2 __extendsfdf2 __gesf2 __lesf2 __ltsf2 __mulsf3 __nesf2 __truncdfsf2 __unordsf2 \
+	__adddf3 __eqdf2 __ledf2 __ltdf2 __muldf3 __unorddf2 \
 	__muloti4 __multi3 \
 	__udivmodti4 __udivti3 __umodti3
 
diff --git a/rust/compiler_builtins.rs b/rust/compiler_builtins.rs
index bba2922c6ef7..f14b8d7caf89 100644
--- a/rust/compiler_builtins.rs
+++ b/rust/compiler_builtins.rs
@@ -40,16 +40,19 @@ pub extern "C" fn $ident() {
 define_panicking_intrinsics!("`f32` should not be used", {
     __addsf3,
     __eqsf2,
+    __extendsfdf2,
     __gesf2,
     __lesf2,
     __ltsf2,
     __mulsf3,
     __nesf2,
+    __truncdfsf2,
     __unordsf2,
 });
 
 define_panicking_intrinsics!("`f64` should not be used", {
     __adddf3,
+    __eqdf2,
     __ledf2,
     __ltdf2,
     __muldf3,

base-commit: de9c2c66ad8e787abec7c9d7eff4f8c3cdd28aed
-- 
2.46.0

Re: [PATCH] rust: add intrinsics to fix `-Os` builds

[Alice Ryhl]

On Tue, Aug 6, 2024 at 5:06 PM Miguel Ojeda <ojeda@kernel.org> wrote:
>
> Alice reported [1] that an arm64 build failed with:
>
>     ld.lld: error: undefined symbol: __extendsfdf2
>     >>> referenced by core.a6f5fc5794e7b7b3-cgu.0
>     >>>               rust/core.o:(<f32>::midpoint) in archive vmlinux.a
>     >>> referenced by core.a6f5fc5794e7b7b3-cgu.0
>     >>>               rust/core.o:(<f32>::midpoint) in archive vmlinux.a
>
>     ld.lld: error: undefined symbol: __truncdfsf2
>     >>> referenced by core.a6f5fc5794e7b7b3-cgu.0
>     >>>               rust/core.o:(<f32>::midpoint) in archive vmlinux.a
>
> Rust 1.80.0 or later together with `CONFIG_CC_OPTIMIZE_FOR_SIZE=y`
> is what triggers it.
>
> In addition, x86_64 builds also fail the same way.
>
> Similarly, compiling with Rust 1.82.0 (currently in nightly) makes
> another one appear, possibly due to the LLVM 19 upgrade there:
>
>     ld.lld: error: undefined symbol: __eqdf2
>     >>> referenced by core.20495ea57a9f069d-cgu.0
>     >>>               rust/core.o:(<f64>::next_up) in archive vmlinux.a
>     >>> referenced by core.20495ea57a9f069d-cgu.0
>     >>>               rust/core.o:(<f64>::next_down) in archive vmlinux.a
>
> Gary adds [1]:
>
> > Usually the fix on rustc side is to mark those functions as `#[inline]`
> >
> > All of {midpoint,next_up,next_down} are indeed unstable functions not
> > marked as inline...
>
> Fix all those by adding those intrinsics to our usual workaround.
>
> Cc: Gary Guo <gary@garyguo.net>
> Reported-by: Alice Ryhl <aliceryhl@google.com>
> Closes: https://rust-for-linux.zulipchat.com/#narrow/stream/288089-General/topic/v6.2E11-rc1.20doesn't.20build.20for.20arm64/near/455637364
> Signed-off-by: Miguel Ojeda <ojeda@kernel.org>

This fixes the issue I ran into.

Tested-by: Alice Ryhl <aliceryhl@google.com>

Re: [PATCH] rust: add intrinsics to fix `-Os` builds

[Boqun Feng]

On Tue, Aug 06, 2024 at 05:06:19PM +0200, Miguel Ojeda wrote:
> Alice reported [1] that an arm64 build failed with:
> 
>     ld.lld: error: undefined symbol: __extendsfdf2
>     >>> referenced by core.a6f5fc5794e7b7b3-cgu.0
>     >>>               rust/core.o:(<f32>::midpoint) in archive vmlinux.a
>     >>> referenced by core.a6f5fc5794e7b7b3-cgu.0
>     >>>               rust/core.o:(<f32>::midpoint) in archive vmlinux.a
> 
>     ld.lld: error: undefined symbol: __truncdfsf2
>     >>> referenced by core.a6f5fc5794e7b7b3-cgu.0
>     >>>               rust/core.o:(<f32>::midpoint) in archive vmlinux.a
> 
> Rust 1.80.0 or later together with `CONFIG_CC_OPTIMIZE_FOR_SIZE=y`
> is what triggers it.
> 
> In addition, x86_64 builds also fail the same way.
> 
> Similarly, compiling with Rust 1.82.0 (currently in nightly) makes
> another one appear, possibly due to the LLVM 19 upgrade there:
> 
>     ld.lld: error: undefined symbol: __eqdf2
>     >>> referenced by core.20495ea57a9f069d-cgu.0
>     >>>               rust/core.o:(<f64>::next_up) in archive vmlinux.a
>     >>> referenced by core.20495ea57a9f069d-cgu.0
>     >>>               rust/core.o:(<f64>::next_down) in archive vmlinux.a
> 
> Gary adds [1]:
> 
> > Usually the fix on rustc side is to mark those functions as `#[inline]`
> >
> > All of {midpoint,next_up,next_down} are indeed unstable functions not
> > marked as inline...
> 
> Fix all those by adding those intrinsics to our usual workaround.
> 
> Cc: Gary Guo <gary@garyguo.net>
> Reported-by: Alice Ryhl <aliceryhl@google.com>
> Closes: https://rust-for-linux.zulipchat.com/#narrow/stream/288089-General/topic/v6.2E11-rc1.20doesn't.20build.20for.20arm64/near/455637364
> Signed-off-by: Miguel Ojeda <ojeda@kernel.org>

Put this one into rust-dev and confirm this could fix the issue with
Rust 1.80. I also add a test with CONFIG_CC_OPTIMIZE_FOR_SIZE=y in my
rust-dev script, so we can catch the similar issues earlier.

Tested-by: Boqun Feng <boqun.feng@gmail.com>

Regards,
Boqun

> ---
>  rust/Makefile             | 4 ++--
>  rust/compiler_builtins.rs | 3 +++
>  2 files changed, 5 insertions(+), 2 deletions(-)
> 
> diff --git a/rust/Makefile b/rust/Makefile
> index 1f10f92737f2..77836388377d 100644
> --- a/rust/Makefile
> +++ b/rust/Makefile
> @@ -354,8 +354,8 @@ rust-analyzer:
>  		$(if $(KBUILD_EXTMOD),$(extmod_prefix),$(objtree))/rust-project.json
>  
>  redirect-intrinsics = \
> -	__addsf3 __eqsf2 __gesf2 __lesf2 __ltsf2 __mulsf3 __nesf2 __unordsf2 \
> -	__adddf3 __ledf2 __ltdf2 __muldf3 __unorddf2 \
> +	__addsf3 __eqsf2 __extendsfdf2 __gesf2 __lesf2 __ltsf2 __mulsf3 __nesf2 __truncdfsf2 __unordsf2 \
> +	__adddf3 __eqdf2 __ledf2 __ltdf2 __muldf3 __unorddf2 \
>  	__muloti4 __multi3 \
>  	__udivmodti4 __udivti3 __umodti3
>  
> diff --git a/rust/compiler_builtins.rs b/rust/compiler_builtins.rs
> index bba2922c6ef7..f14b8d7caf89 100644
> --- a/rust/compiler_builtins.rs
> +++ b/rust/compiler_builtins.rs
> @@ -40,16 +40,19 @@ pub extern "C" fn $ident() {
>  define_panicking_intrinsics!("`f32` should not be used", {
>      __addsf3,
>      __eqsf2,
> +    __extendsfdf2,
>      __gesf2,
>      __lesf2,
>      __ltsf2,
>      __mulsf3,
>      __nesf2,
> +    __truncdfsf2,
>      __unordsf2,
>  });
>  
>  define_panicking_intrinsics!("`f64` should not be used", {
>      __adddf3,
> +    __eqdf2,
>      __ledf2,
>      __ltdf2,
>      __muldf3,
> 
> base-commit: de9c2c66ad8e787abec7c9d7eff4f8c3cdd28aed
> -- 
> 2.46.0
> 
> 

Re: [PATCH] rust: add intrinsics to fix `-Os` builds

[Miguel Ojeda]

On Tue, Aug 6, 2024 at 7:28 PM Boqun Feng <boqun.feng@gmail.com> wrote:
>
> Put this one into rust-dev and confirm this could fix the issue with
> Rust 1.80. I also add a test with CONFIG_CC_OPTIMIZE_FOR_SIZE=y in my
> rust-dev script, so we can catch the similar issues earlier.

+1, same on my side.

Cheers,
Miguel

Re: [PATCH] rust: add intrinsics to fix `-Os` builds

[Trevor Gross]

On Tue, Aug 6, 2024 at 11:06 AM Miguel Ojeda <ojeda@kernel.org> wrote:
>
> Alice reported [1] that an arm64 build failed with:
>
>     ld.lld: error: undefined symbol: __extendsfdf2
>     >>> referenced by core.a6f5fc5794e7b7b3-cgu.0
>     >>>               rust/core.o:(<f32>::midpoint) in archive vmlinux.a
>     >>> referenced by core.a6f5fc5794e7b7b3-cgu.0
>     >>>               rust/core.o:(<f32>::midpoint) in archive vmlinux.a
>
>     ld.lld: error: undefined symbol: __truncdfsf2
>     >>> referenced by core.a6f5fc5794e7b7b3-cgu.0
>     >>>               rust/core.o:(<f32>::midpoint) in archive vmlinux.a
>
> Rust 1.80.0 or later together with `CONFIG_CC_OPTIMIZE_FOR_SIZE=y`
> is what triggers it.
>
> In addition, x86_64 builds also fail the same way.
>
> Similarly, compiling with Rust 1.82.0 (currently in nightly) makes
> another one appear, possibly due to the LLVM 19 upgrade there:
>
>     ld.lld: error: undefined symbol: __eqdf2
>     >>> referenced by core.20495ea57a9f069d-cgu.0
>     >>>               rust/core.o:(<f64>::next_up) in archive vmlinux.a
>     >>> referenced by core.20495ea57a9f069d-cgu.0
>     >>>               rust/core.o:(<f64>::next_down) in archive vmlinux.a
>
> Gary adds [1]:
>
> > Usually the fix on rustc side is to mark those functions as `#[inline]`
> >
> > All of {midpoint,next_up,next_down} are indeed unstable functions not
> > marked as inline...
>
> Fix all those by adding those intrinsics to our usual workaround.
>
> Cc: Gary Guo <gary@garyguo.net>
> Reported-by: Alice Ryhl <aliceryhl@google.com>
> Closes: https://rust-for-linux.zulipchat.com/#narrow/stream/288089-General/topic/v6.2E11-rc1.20doesn't.20build.20for.20arm64/near/455637364

Should this link have the `[1]`?

> Signed-off-by: Miguel Ojeda <ojeda@kernel.org>
> ---
>  rust/Makefile             | 4 ++--
>  rust/compiler_builtins.rs | 3 +++
>  2 files changed, 5 insertions(+), 2 deletions(-)

Reviewed-by: Trevor Gross <tmgross@umich.edu>

Also submitted a fix upstream https://github.com/rust-lang/rust/pull/128749.

- Trevor

Re: [PATCH] rust: add intrinsics to fix `-Os` builds

[Miguel Ojeda]

On Tue, Aug 6, 2024 at 8:30 PM Trevor Gross <tmgross@umich.edu> wrote:
>
> Should this link have the `[1]`?

Yeah, good eyes :)

> Reviewed-by: Trevor Gross <tmgross@umich.edu>
>
> Also submitted a fix upstream https://github.com/rust-lang/rust/pull/128749.

Thanks a lot! Linked in the `core` sublist in issue #2.

Cheers,
Miguel

Re: [PATCH] rust: add intrinsics to fix `-Os` builds

[Gary Guo]

On Tue,  6 Aug 2024 17:06:19 +0200
Miguel Ojeda <ojeda@kernel.org> wrote:

> Alice reported [1] that an arm64 build failed with:
> 
>     ld.lld: error: undefined symbol: __extendsfdf2
>     >>> referenced by core.a6f5fc5794e7b7b3-cgu.0
>     >>>               rust/core.o:(<f32>::midpoint) in archive vmlinux.a
>     >>> referenced by core.a6f5fc5794e7b7b3-cgu.0
>     >>>               rust/core.o:(<f32>::midpoint) in archive vmlinux.a  
> 
>     ld.lld: error: undefined symbol: __truncdfsf2
>     >>> referenced by core.a6f5fc5794e7b7b3-cgu.0
>     >>>               rust/core.o:(<f32>::midpoint) in archive vmlinux.a  
> 
> Rust 1.80.0 or later together with `CONFIG_CC_OPTIMIZE_FOR_SIZE=y`
> is what triggers it.
> 
> In addition, x86_64 builds also fail the same way.
> 
> Similarly, compiling with Rust 1.82.0 (currently in nightly) makes
> another one appear, possibly due to the LLVM 19 upgrade there:
> 
>     ld.lld: error: undefined symbol: __eqdf2
>     >>> referenced by core.20495ea57a9f069d-cgu.0
>     >>>               rust/core.o:(<f64>::next_up) in archive vmlinux.a
>     >>> referenced by core.20495ea57a9f069d-cgu.0
>     >>>               rust/core.o:(<f64>::next_down) in archive vmlinux.a  
> 
> Gary adds [1]:
> 
> > Usually the fix on rustc side is to mark those functions as `#[inline]`
> >
> > All of {midpoint,next_up,next_down} are indeed unstable functions not
> > marked as inline...  
> 
> Fix all those by adding those intrinsics to our usual workaround.
> 
> Cc: Gary Guo <gary@garyguo.net>
> Reported-by: Alice Ryhl <aliceryhl@google.com>
> Closes: https://rust-for-linux.zulipchat.com/#narrow/stream/288089-General/topic/v6.2E11-rc1.20doesn't.20build.20for.20arm64/near/455637364
> Signed-off-by: Miguel Ojeda <ojeda@kernel.org>

Reviewed-by: Gary Guo <gary@garyguo.net>

> ---
>  rust/Makefile             | 4 ++--
>  rust/compiler_builtins.rs | 3 +++
>  2 files changed, 5 insertions(+), 2 deletions(-)

Re: [PATCH] rust: add intrinsics to fix `-Os` builds

[Miguel Ojeda]

On Tue, Aug 6, 2024 at 5:06 PM Miguel Ojeda <ojeda@kernel.org> wrote:
>
> Alice reported [1] that an arm64 build failed with:
>
>     ld.lld: error: undefined symbol: __extendsfdf2
>     >>> referenced by core.a6f5fc5794e7b7b3-cgu.0
>     >>>               rust/core.o:(<f32>::midpoint) in archive vmlinux.a
>     >>> referenced by core.a6f5fc5794e7b7b3-cgu.0
>     >>>               rust/core.o:(<f32>::midpoint) in archive vmlinux.a
>
>     ld.lld: error: undefined symbol: __truncdfsf2
>     >>> referenced by core.a6f5fc5794e7b7b3-cgu.0
>     >>>               rust/core.o:(<f32>::midpoint) in archive vmlinux.a
>
> Rust 1.80.0 or later together with `CONFIG_CC_OPTIMIZE_FOR_SIZE=y`
> is what triggers it.
>
> In addition, x86_64 builds also fail the same way.
>
> Similarly, compiling with Rust 1.82.0 (currently in nightly) makes
> another one appear, possibly due to the LLVM 19 upgrade there:
>
>     ld.lld: error: undefined symbol: __eqdf2
>     >>> referenced by core.20495ea57a9f069d-cgu.0
>     >>>               rust/core.o:(<f64>::next_up) in archive vmlinux.a
>     >>> referenced by core.20495ea57a9f069d-cgu.0
>     >>>               rust/core.o:(<f64>::next_down) in archive vmlinux.a
>
> Gary adds [1]:
>
> > Usually the fix on rustc side is to mark those functions as `#[inline]`
> >
> > All of {midpoint,next_up,next_down} are indeed unstable functions not
> > marked as inline...
>
> Fix all those by adding those intrinsics to our usual workaround.
>
> Cc: Gary Guo <gary@garyguo.net>
> Reported-by: Alice Ryhl <aliceryhl@google.com>
> Closes: https://rust-for-linux.zulipchat.com/#narrow/stream/288089-General/topic/v6.2E11-rc1.20doesn't.20build.20for.20arm64/near/455637364
> Signed-off-by: Miguel Ojeda <ojeda@kernel.org>

Applied to `rust-fixes` -- thanks everyone!

    [ Trevor quickly submitted a fix to upstream Rust [2] that has already
      been merged, to be released in Rust 1.82.0 (2024-10-17). - Miguel ]

    [ Shortened Zulip link. - Miguel ]

Cheers,
Miguel