[PATCH 0/2] alpha: enable building with clang

[PATCH 0/2] alpha: enable building with clang

[Matt Turner]

Two small patches to let the alpha kernel build with clang.

The first registers the clang target triple and stops passing -Wa,-mev6
when the compiler is not gcc.  That flag exists to keep gas from emulating
instructions it believes the target lacks; it is a gas-only option and
clang's integrated assembler does not emulate instructions, so it is not
needed there.

The second fixes two uses of local register-asm variables that clang does
not honor.  clang treats `register unsigned long x __asm__("$N")` as the
named register only where the variable appears as an inline-asm operand,
so reading one to get the live $gp or $sp yields an undefined value.
trap_init() passed that to PAL_wrkgp and load_PCB() stored it into the
PCB for swpctx, either of which wedges an early boot.

Note that the alpha backend is not in upstream LLVM.  It lives in

  https://github.com/alphalinux-org/llvm-project

and is a work in progress, so the scripts/Makefile.clang entry has no
effect with an upstream clang today.  I am sending this now because the
second patch is a real bug in its own right -- the register-asm reads are
only guaranteed to work by gcc's implementation, not by anything either
compiler documents -- but I understand if the kbuild side would rather
wait for the backend to land upstream.

---
Matt Turner (2):
      alpha: enable building with clang
      alpha: read $gp and $sp explicitly for clang

 arch/alpha/Makefile       | 8 +++++++-
 arch/alpha/kernel/traps.c | 4 +++-
 arch/alpha/mm/init.c      | 3 +--
 scripts/Makefile.clang    | 1 +
 4 files changed, 12 insertions(+), 4 deletions(-)
---
base-commit: 0e6be1d34ae92e2b0dbc1b7410d422b22464389a
change-id: 20260803-alpha-clang-6acd144eb86c

Best regards,
-- 
Matt Turner <mattst88@gmail.com>

[PATCH 1/2] alpha: enable building with clang

[Matt Turner]

Register the clang target triple for alpha, and pass -Wa,-mev6 only when
building with gcc.  That flag exists to stop gas emulating instructions
the assembler thinks the target lacks; it is a gas-only option, and
clang's integrated assembler does not emulate instructions in the first
place, so nothing is needed there.

With this, LLVM=1 (and LLVM_IAS=1) builds work.  Note that the alpha
backend is not in upstream LLVM yet; it lives in

  https://github.com/alphalinux-org/llvm-project

and is a work in progress.

Signed-off-by: Matt Turner <mattst88@gmail.com>
---
 arch/alpha/Makefile    | 8 +++++++-
 scripts/Makefile.clang | 1 +
 2 files changed, 8 insertions(+), 1 deletion(-)

diff --git a/arch/alpha/Makefile b/arch/alpha/Makefile
index 35445ff2e489..54283705911f 100644
--- a/arch/alpha/Makefile
+++ b/arch/alpha/Makefile
@@ -27,10 +27,16 @@ cpuflags-$(CONFIG_ALPHA_GENERIC)	:= -mcpu=ev56 -mtune=ev6
 cflags-y				+= $(cpuflags-y)
 
 
+KBUILD_CFLAGS += $(cflags-y)
+
 # For TSUNAMI, we must have the assembler not emulate our instructions.
 # The same is true for IRONGATE, POLARIS, PYXIS.
 # BWX is most important, but we don't really want any emulation ever.
-KBUILD_CFLAGS += $(cflags-y) -Wa,-mev6
+# Only gas emulates instructions the target does not implement, and only gas
+# understands -mev6; clang's integrated assembler never emulates.
+ifdef CONFIG_CC_IS_GCC
+KBUILD_CFLAGS += -Wa,-mev6
+endif
 
 libs-y				+= arch/alpha/lib/
 
diff --git a/scripts/Makefile.clang b/scripts/Makefile.clang
index b67636b28c35..2434cf90ad7e 100644
--- a/scripts/Makefile.clang
+++ b/scripts/Makefile.clang
@@ -1,6 +1,7 @@
 # Individual arch/{arch}/Makefiles should use -EL/-EB to set intended
 # endianness and -m32/-m64 to set word size based on Kconfigs instead of
 # relying on the target triple.
+CLANG_TARGET_FLAGS_alpha	:= alpha-linux-gnu
 CLANG_TARGET_FLAGS_arm		:= arm-linux-gnueabi
 CLANG_TARGET_FLAGS_arm64	:= aarch64-linux-gnu
 CLANG_TARGET_FLAGS_hexagon	:= hexagon-linux-musl

-- 
2.54.0

[PATCH 2/2] alpha: read $gp and $sp explicitly for clang

[Matt Turner]

clang honors a local `register unsigned long x __asm__("$N")` variable
only where it appears as an inline-asm operand; merely reading it does
not produce the contents of that register.  So trap_init() passed an
undefined global pointer to PAL_wrkgp, and load_PCB() stored an undefined
stack pointer into the PCB that swpctx then loaded.  Either one wedges an
early boot.

Read the registers explicitly instead: an inline mov for $gp in
trap_init(), and the file-scope current_stack_pointer for $sp in
load_PCB().  A file-scope register-asm variable is the form clang does
support.

Signed-off-by: Matt Turner <mattst88@gmail.com>
---
 arch/alpha/kernel/traps.c | 4 +++-
 arch/alpha/mm/init.c      | 3 +--
 2 files changed, 4 insertions(+), 3 deletions(-)

diff --git a/arch/alpha/kernel/traps.c b/arch/alpha/kernel/traps.c
index 7631129ac914..5b4f1ae2b74b 100644
--- a/arch/alpha/kernel/traps.c
+++ b/arch/alpha/kernel/traps.c
@@ -925,7 +925,9 @@ void
 trap_init(void)
 {
 	/* Tell PAL-code what global pointer we want in the kernel.  */
-	register unsigned long gptr __asm__("$29");
+	unsigned long gptr;
+
+	__asm__ volatile("mov $29, %0" : "=r"(gptr));
 	wrkgp(gptr);
 
 	wrent(entArith, 1);
diff --git a/arch/alpha/mm/init.c b/arch/alpha/mm/init.c
index 9531cbc761c0..f4d65a60c869 100644
--- a/arch/alpha/mm/init.c
+++ b/arch/alpha/mm/init.c
@@ -63,8 +63,7 @@ pgd_alloc(struct mm_struct *mm)
 static inline unsigned long
 load_PCB(struct pcb_struct *pcb)
 {
-	register unsigned long sp __asm__("$30");
-	pcb->ksp = sp;
+	pcb->ksp = (unsigned long)current_stack_pointer;
 	return __reload_thread(pcb);
 }
 

-- 
2.54.0

Re: [PATCH 0/2] alpha: enable building with clang

[Nathan Chancellor]

Hi Matt,

On Mon, Aug 03, 2026 at 01:08:40PM -0400, Matt Turner wrote:
> Two small patches to let the alpha kernel build with clang.

Nice!

> The first registers the clang target triple and stops passing -Wa,-mev6
> when the compiler is not gcc.  That flag exists to keep gas from emulating
> instructions it believes the target lacks; it is a gas-only option and
> clang's integrated assembler does not emulate instructions, so it is not
> needed there.
> 
> The second fixes two uses of local register-asm variables that clang does
> not honor.  clang treats `register unsigned long x __asm__("$N")` as the
> named register only where the variable appears as an inline-asm operand,
> so reading one to get the live $gp or $sp yields an undefined value.
> trap_init() passed that to PAL_wrkgp and load_PCB() stored it into the
> PCB for swpctx, either of which wedges an early boot.
> 
> Note that the alpha backend is not in upstream LLVM.  It lives in
> 
>   https://github.com/alphalinux-org/llvm-project
> 
> and is a work in progress, so the scripts/Makefile.clang entry has no
> effect with an upstream clang today.  I am sending this now because the
> second patch is a real bug in its own right -- the register-asm reads are
> only guaranteed to work by gcc's implementation, not by anything either
> compiler documents -- but I understand if the kbuild side would rather
> wait for the backend to land upstream.

Yeah, I am not sure how I feel taking the target triple part of the
first patch. On the one hand, I want it to be easy for you to test
against upstream Linux but on the other, I do not want people to read
this Makefile and assume that ARCH=alpha will work with an upstream
clang.

We could add a comment that the backend is currently out of tree but
that would go stale once it is actually upstream and it will be floating
around for forever. Maybe a better compromise is taking arch/alpha
changes now then landing the scripts/Makefile.clang change when you
actually start upstreaming the backend, as being able to use an upstream
version of clang should be relatively imminent at that point.

-- 
Cheers,
Nathan

Re: [PATCH 1/2] alpha: enable building with clang

[Magnus Lindholm]

On Mon, Aug 3, 2026 at 7:08 PM Matt Turner <mattst88@gmail.com> wrote:
>
> Register the clang target triple for alpha, and pass -Wa,-mev6 only when
> building with gcc.  That flag exists to stop gas emulating instructions
> the assembler thinks the target lacks; it is a gas-only option, and
> clang's integrated assembler does not emulate instructions in the first
> place, so nothing is needed there.
>
> With this, LLVM=1 (and LLVM_IAS=1) builds work.  Note that the alpha
> backend is not in upstream LLVM yet; it lives in
>
>   https://github.com/alphalinux-org/llvm-project
>
> and is a work in progress.
>
> Signed-off-by: Matt Turner <mattst88@gmail.com>
> ---
>  arch/alpha/Makefile    | 8 +++++++-
>  scripts/Makefile.clang | 1 +
>  2 files changed, 8 insertions(+), 1 deletion(-)
>
> diff --git a/arch/alpha/Makefile b/arch/alpha/Makefile
> index 35445ff2e489..54283705911f 100644
> --- a/arch/alpha/Makefile
> +++ b/arch/alpha/Makefile
> @@ -27,10 +27,16 @@ cpuflags-$(CONFIG_ALPHA_GENERIC)    := -mcpu=ev56 -mtune=ev6
>  cflags-y                               += $(cpuflags-y)
>
>
> +KBUILD_CFLAGS += $(cflags-y)
> +
>  # For TSUNAMI, we must have the assembler not emulate our instructions.
>  # The same is true for IRONGATE, POLARIS, PYXIS.
>  # BWX is most important, but we don't really want any emulation ever.
> -KBUILD_CFLAGS += $(cflags-y) -Wa,-mev6
> +# Only gas emulates instructions the target does not implement, and only gas
> +# understands -mev6; clang's integrated assembler never emulates.
> +ifdef CONFIG_CC_IS_GCC
> +KBUILD_CFLAGS += -Wa,-mev6
> +endif
>
>  libs-y                         += arch/alpha/lib/
>

Hi Matt,

I agree with Nathan's suggestion to leave the
CLANG_TARGET_FLAGS_alpha entry out for now and add it separately when
the Alpha backend is being upstreamed to LLVM.

Could you send a v2 containing just the arch/alpha/Makefile change?

Since -Wa,-mev6 is an assembler-specific option, I think maybe the condition
should also test the assembler rather than the compiler?

ifdef CONFIG_AS_IS_GNU
KBUILD_CFLAGS += -Wa,-mev6
endif

That preserves the option for both GCC and clang builds using GNU as,
including clang with LLVM_IAS=0, while omitting it for LLVM's integrated
assembler.

The revised Alpha Makefile patch can then go through the Alpha tree,
while the scripts/Makefile.clang change can wait until the backend is
available upstream.

Thanks,
Magnus

Re: [PATCH 2/2] alpha: read $gp and $sp explicitly for clang

[Magnus Lindholm]

Hi Matt,

On Mon, Aug 3, 2026 at 7:08 PM Matt Turner <mattst88@gmail.com> wrote:
>
> clang honors a local `register unsigned long x __asm__("$N")` variable
> only where it appears as an inline-asm operand; merely reading it does
> not produce the contents of that register.  So trap_init() passed an
> undefined global pointer to PAL_wrkgp, and load_PCB() stored an undefined
> stack pointer into the PCB that swpctx then loaded.  Either one wedges an
> early boot.
>
> Read the registers explicitly instead: an inline mov for $gp in
> trap_init(), and the file-scope current_stack_pointer for $sp in
> load_PCB().  A file-scope register-asm variable is the form clang does
> support.
>
> Signed-off-by: Matt Turner <mattst88@gmail.com>
> ---
>  arch/alpha/kernel/traps.c | 4 +++-
>  arch/alpha/mm/init.c      | 3 +--
>  2 files changed, 4 insertions(+), 3 deletions(-)
>
> diff --git a/arch/alpha/kernel/traps.c b/arch/alpha/kernel/traps.c
> index 7631129ac914..5b4f1ae2b74b 100644
> --- a/arch/alpha/kernel/traps.c
> +++ b/arch/alpha/kernel/traps.c
> @@ -925,7 +925,9 @@ void
>  trap_init(void)
>  {
>         /* Tell PAL-code what global pointer we want in the kernel.  */
> -       register unsigned long gptr __asm__("$29");
> +       unsigned long gptr;
> +
> +       __asm__ volatile("mov $29, %0" : "=r"(gptr));
>         wrkgp(gptr);
>
>         wrent(entArith, 1);
> diff --git a/arch/alpha/mm/init.c b/arch/alpha/mm/init.c
> index 9531cbc761c0..f4d65a60c869 100644
> --- a/arch/alpha/mm/init.c
> +++ b/arch/alpha/mm/init.c
> @@ -63,8 +63,7 @@ pgd_alloc(struct mm_struct *mm)
>  static inline unsigned long
>  load_PCB(struct pcb_struct *pcb)
>  {
> -       register unsigned long sp __asm__("$30");
> -       pcb->ksp = sp;
> +       pcb->ksp = (unsigned long)current_stack_pointer;
>         return __reload_thread(pcb);
>  }
>
>
> --
> 2.54.0
>

This looks good to me. The explicit $gp read and use of
current_stack_pointer avoid relying on the unsupported local
register-asm behavior.

I also built and booted the patched kernel successfully with GCC on an
AlphaStation DS10.

Reviewed-by: Magnus Lindholm linmag7@gmail.com
Tested-by: Magnus Lindholm linmag7@gmail.com

I am happy to take this through the Alpha tree.

Thanks,
Magnus