As part of the Rust support for UML, we disable SSE (and similar flags)
to match the normal x86 builds. This both makes sense (we ideally want a
similar configuration to x86), and works around a crash bug with SSE
generation under Rust with LLVM.
However, this breaks compiling stdlib.h under gcc < 11, as the x86_64
ABI requires floating-point return values be stored in an SSE register.
gcc 11 fixes this by only doing register allocation when a function is
actually used, and since we never use atof(), it shouldn't be a problem:
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=99652
Nevertheless, only disable SSE on clang setups, as that's a simple way
of working around everyone's bugs.
Fixes: 884981867947 ("rust: arch/um: Disable FP/SIMD instruction to match x86")
Reported-by: Roberto Sassu <[email protected]>
Link: https://lore.kernel.org/linux-um/[email protected]/
Tested-by: Roberto Sassu <[email protected]>
Tested-by: SeongJae Park <[email protected]>
Signed-off-by: David Gow <[email protected]>
---
arch/x86/Makefile.um | 5 +++++
1 file changed, 5 insertions(+)
diff --git a/arch/x86/Makefile.um b/arch/x86/Makefile.um
index b70559b821df..2106a2bd152b 100644
--- a/arch/x86/Makefile.um
+++ b/arch/x86/Makefile.um
@@ -3,9 +3,14 @@ core-y += arch/x86/crypto/
#
# Disable SSE and other FP/SIMD instructions to match normal x86
+# This is required to work around issues in older LLVM versions, but breaks
+# GCC versions < 11. See:
+# https://gcc.gnu.org/bugzilla/show_bug.cgi?id=99652
#
+ifeq ($(CONFIG_CC_IS_CLANG),y)
KBUILD_CFLAGS += -mno-sse -mno-mmx -mno-sse2 -mno-3dnow -mno-avx
KBUILD_RUSTFLAGS += -Ctarget-feature=-sse,-sse2,-sse3,-ssse3,-sse4.1,-sse4.2,-avx,-avx2
+endif
ifeq ($(CONFIG_X86_32),y)
START := 0x8048000
--
2.40.0.rc1.284.g88254d51c5-goog
> As part of the Rust support for UML, we disable SSE (and similar flags)
> to match the normal x86 builds. This both makes sense (we ideally want a
> similar configuration to x86), and works around a crash bug with SSE
> generation under Rust with LLVM.
>
> However, this breaks compiling stdlib.h under gcc < 11, as the x86_64
> ABI requires floating-point return values be stored in an SSE register.
> gcc 11 fixes this by only doing register allocation when a function is
> actually used, and since we never use atof(), it shouldn't be a problem:
> https://gcc.gnu.org/bugzilla/show_bug.cgi?id=99652
>
> Nevertheless, only disable SSE on clang setups, as that's a simple way
> of working around everyone's bugs.
>
> Fixes: 884981867947 ("rust: arch/um: Disable FP/SIMD instruction to match x86")
> Reported-by: Roberto Sassu <[email protected]>
> Link: https://lore.kernel.org/linux-um/[email protected]/
> Tested-by: Roberto Sassu <[email protected]>
> Tested-by: SeongJae Park <[email protected]>
> Signed-off-by: David Gow <[email protected]>
> ---
> arch/x86/Makefile.um | 5 +++++
> 1 file changed, 5 insertions(+)
>
> diff --git a/arch/x86/Makefile.um b/arch/x86/Makefile.um
> index b70559b821df..2106a2bd152b 100644
> --- a/arch/x86/Makefile.um
> +++ b/arch/x86/Makefile.um
> @@ -3,9 +3,14 @@ core-y += arch/x86/crypto/
>
> #
> # Disable SSE and other FP/SIMD instructions to match normal x86
> +# This is required to work around issues in older LLVM versions, but breaks
> +# GCC versions < 11. See:
> +# https://gcc.gnu.org/bugzilla/show_bug.cgi?id=99652
> #
> +ifeq ($(CONFIG_CC_IS_CLANG),y)
> KBUILD_CFLAGS += -mno-sse -mno-mmx -mno-sse2 -mno-3dnow -mno-avx
> KBUILD_RUSTFLAGS += -Ctarget-feature=-sse,-sse2,-sse3,-ssse3,-sse4.1,-sse4.2,-avx,-avx2
> +endif
>
> ifeq ($(CONFIG_X86_32),y)
> START := 0x8048000
> --
> 2.40.0.rc1.284.g88254d51c5-goog
Reviewed-by: Vincenzo Palazzo <[email protected]>
On 18/03/23 01:15, 'David Gow' via KUnit Development wrote:
> As part of the Rust support for UML, we disable SSE (and similar flags)
> to match the normal x86 builds. This both makes sense (we ideally want a
> similar configuration to x86), and works around a crash bug with SSE
> generation under Rust with LLVM.
>
> However, this breaks compiling stdlib.h under gcc < 11, as the x86_64
> ABI requires floating-point return values be stored in an SSE register.
> gcc 11 fixes this by only doing register allocation when a function is
> actually used, and since we never use atof(), it shouldn't be a problem:
> https://gcc.gnu.org/bugzilla/show_bug.cgi?id=99652
>
> Nevertheless, only disable SSE on clang setups, as that's a simple way
> of working around everyone's bugs.
>
> Fixes: 884981867947 ("rust: arch/um: Disable FP/SIMD instruction to match x86")
> Reported-by: Roberto Sassu <[email protected]>
> Link: https://lore.kernel.org/linux-um/[email protected]/
> Tested-by: Roberto Sassu <[email protected]>
> Tested-by: SeongJae Park <[email protected]>
> Signed-off-by: David Gow <[email protected]>
> ---
> arch/x86/Makefile.um | 5 +++++
> 1 file changed, 5 insertions(+)
>
> diff --git a/arch/x86/Makefile.um b/arch/x86/Makefile.um
> index b70559b821df..2106a2bd152b 100644
> --- a/arch/x86/Makefile.um
> +++ b/arch/x86/Makefile.um
> @@ -3,9 +3,14 @@ core-y += arch/x86/crypto/
>
> #
> # Disable SSE and other FP/SIMD instructions to match normal x86
> +# This is required to work around issues in older LLVM versions, but breaks
> +# GCC versions < 11. See:
> +# https://gcc.gnu.org/bugzilla/show_bug.cgi?id=99652
> #
> +ifeq ($(CONFIG_CC_IS_CLANG),y)
> KBUILD_CFLAGS += -mno-sse -mno-mmx -mno-sse2 -mno-3dnow -mno-avx
> KBUILD_RUSTFLAGS += -Ctarget-feature=-sse,-sse2,-sse3,-ssse3,-sse4.1,-sse4.2,-avx,-avx2
> +endif
>
> ifeq ($(CONFIG_X86_32),y)
> START := 0x8048000
Tested-by: Arthur Grillo <[email protected]>
On Sat, 18 Mar 2023 at 12:16, David Gow <[email protected]> wrote:
>
> As part of the Rust support for UML, we disable SSE (and similar flags)
> to match the normal x86 builds. This both makes sense (we ideally want a
> similar configuration to x86), and works around a crash bug with SSE
> generation under Rust with LLVM.
>
> However, this breaks compiling stdlib.h under gcc < 11, as the x86_64
> ABI requires floating-point return values be stored in an SSE register.
> gcc 11 fixes this by only doing register allocation when a function is
> actually used, and since we never use atof(), it shouldn't be a problem:
> https://gcc.gnu.org/bugzilla/show_bug.cgi?id=99652
>
> Nevertheless, only disable SSE on clang setups, as that's a simple way
> of working around everyone's bugs.
>
> Fixes: 884981867947 ("rust: arch/um: Disable FP/SIMD instruction to match x86")
> Reported-by: Roberto Sassu <[email protected]>
> Link: https://lore.kernel.org/linux-um/[email protected]/
> Tested-by: Roberto Sassu <[email protected]>
> Tested-by: SeongJae Park <[email protected]>
> Signed-off-by: David Gow <[email protected]>
> ---
Any chance we could get this picked up as a fix for 6.3? The UML build
is still broken with older gcc versions.
Cheers,
-- David
----- Ursprüngliche Mail -----
> Von: "davidgow" <[email protected]>
> An: "richard" <[email protected]>, "anton ivanov" <[email protected]>, "Johannes Berg"
> Any chance we could get this picked up as a fix for 6.3? The UML build
> is still broken with older gcc versions.
Sure. Just got flooded with other stuff. :-S
Thanks,
//richard
On Sat, Mar 18, 2023 at 5:16 AM David Gow <[email protected]> wrote:
>
> Nevertheless, only disable SSE on clang setups, as that's a simple way
> of working around everyone's bugs.
Sounds good to me. If GCC >= 11 has fixed it, should we instead test
for GCC < 11, i.e. it could help catching future new regressions like
the linked GCC bug, if I understand correctly, right? Of course, that
could be done for the next cycle, since this one minimizes the
original change w.r.t. v6.2.
By the way, missing space on the title (`arch:um`).
Cheers,
Miguel
----- Ursprüngliche Mail -----
>> Von: "davidgow" <[email protected]>
>> An: "richard" <[email protected]>, "anton ivanov"
>> <[email protected]>, "Johannes Berg"
>> Any chance we could get this picked up as a fix for 6.3? The UML build
>> is still broken with older gcc versions.
>
> Sure. Just got flooded with other stuff. :-S
Patch applied to fixes branch. PR will happen soon.
Thanks everyone and sorry for the delay on my side!
--
Thanks,
//richard