Skip to contentSkip to conversation

Clang compatibility improvements for RISC-V

PATCH In review Yao Zi 4 participants 12 messages Apr 27, 2025 – May 12, 2025

Subsystems (2)
lore
Share & tools
lore

Yao Zi

author
#0 lore raw

cover Clang compatibility improvements for RISC-V

Originally reported by Nathaniel Hourt[1], U-Boot fails to boot on JH7110 platforms (starfive_visionfive2_defconfig) when building with Clang.

This series fixes configuration generation with Clang's preprocessor and add an alternative codepath for RISC-V to access gd to workaround misoptimization of Clang, which is sufficient for building images for JH7110 boards.

Tested with

- starfive_visionfive2_defconfig
- qemu-riscv64_smode_defconfig
- sifive_unleashed_defconfig

built with either GCC 14 or LLVM 19.

Note that images built with LLVM lld are still broken and binutils ld has to be used for now. To test the changes on JH7110 platforms, this patch[2] must be applied for SPL to function.

[1]: https://lore.kernel.org/u-boot/[email protected]/
[2]: https://lore.kernel.org/all/[email protected]/

Yao Zi (2):
Makefile: Strip leading spaces when preprocessing generated_defconfig
riscv: Access gd with inline assembly when building with LTO or Clang

 arch/riscv/cpu/cpu.c                 |  6 ++++++
 arch/riscv/include/asm/global_data.h | 19 +++++++++++++++++++
 common/board_r.c                     |  4 +++-
 common/init/board_init.c             |  7 +++++--
 scripts/kconfig/Makefile             |  1 +
 5 files changed, 34 insertions(+), 3 deletions(-)
Signature · 1 line
2.49.0
author
#1 lore raw

PATCH 1/2 Makefile: Strip leading spaces when preprocessing generated_defconfig

Clang's preprocessor may emit extra spaces for lines starting with '#'. Lines with these extra characters cannot be handled by Kconfig and will be ignored with warnings like,

unexpected data: # CONFIG_OF_BOARD_FIXUP is not set

Those options that is expected to be assigned explicitly with N will be set to the default value, messing up board configurations.

Let's sed these spaces away to ensure board configurations could be correctly generated with Clang.

Fixes 2027e99e61a ("Makefile: Run defconfig files through the C preprocessor")
---
 scripts/kconfig/Makefile | 1 +
 1 file changed, 1 insertion(+)
scripts/kconfig/Makefile +1 -0
index 079add4d5da..ba30652f01a 100644
@@ -94,6 +94,7 @@ endif
 
 %_defconfig: $(obj)/conf
 	$(Q)$(CPP) -nostdinc -P -I $(srctree) -undef -x assembler-with-cpp $(srctree)/arch/$(SRCARCH)/configs/$@ -o generated_defconfig
+	$(Q)sed -i -e 's/^[[:space:]]//' generated_defconfig
 	$(Q)$< $(silent) --defconfig=generated_defconfig $(Kconfig)
 
 # Added for U-Boot (backward compatibility)
Signature · 1 line
2.49.0
#2 lore raw

Re: [PATCH 1/2] Makefile: Strip leading spaces when preprocessing generated_defconfig

Am 27. April 2025 16:50:10 MESZ schrieb Yao Zi <[email protected]>:

Quoted 4 lines

Clang's preprocessor may emit extra spaces for lines starting with '#'. Lines with these extra characters cannot be handled by Kconfig and will be ignored with warnings like,

Do you have an example for reprocing the issue? Is there an understanding why Clang behaves in this way?

Best regards

Heinrich

Quoted 27 lines

unexpected data: # CONFIG_OF_BOARD_FIXUP is not set

Those options that is expected to be assigned explicitly with N will be set to the default value, messing up board configurations.

Let's sed these spaces away to ensure board configurations could be correctly generated with Clang.

Fixes 2027e99e61a ("Makefile: Run defconfig files through the C preprocessor")

---
scripts/kconfig/Makefile | 1 +

1 file changed, 1 insertion(+)
scripts/kconfig/Makefile +0 -0
index 079add4d5da..ba30652f01a 100644
@@ -94,6 +94,7 @@ endif
 

%_defconfig: $(obj)/conf

	$(Q)$(CPP) -nostdinc -P -I $(srctree) -undef -x assembler-with-cpp $(srctree)/arch/$(SRCARCH)/configs/$@ -o generated_defconfig
scripts/kconfig/Makefile +1 -0
+	$(Q)sed -i -e 's/^[[:space:]]//' generated_defconfig
	$(Q)$< $(silent) --defconfig=generated_defconfig $(Kconfig)

# Added for U-Boot (backward compatibility)

author
#3 lore raw

Re: [PATCH 1/2] Makefile: Strip leading spaces when preprocessing generated_defconfig

Quoted 8 lines · Heinrich Schuchardt

Am 27. April 2025 16:50:10 MESZ schrieb Yao Zi <[email protected]>:

Quoted 4 lines

Clang's preprocessor may emit extra spaces for lines starting with '#'. Lines with these extra characters cannot be handled by Kconfig and will be ignored with warnings like,

Do you have an example for reprocing the issue?

Sure,

	clang-19 -E -nostdinc -P -I . -undef -x assembler-with-cpp \
		configs/starfive_visionfive2_defconfig

or a smaller example for demonstrating the behaviour,

cat << EOF | clang -E -P -x assembler-with-cpp - # comment line
normal line
EOF

and you could see the strange indentation. For reproducing the exact Kconfig warnings,

	make ARCH=riscv \
		CC='clang-19 --target=riscv64-unknown-linux-musl' \
		starfive_visionfive2_defconfig

(Clang is called clang-19 on my machine)

Quoted 1 line

Is there an understanding why Clang behaves in this way?

Sadly I have no idea. I guess it may serve for improving human-readability of the preprocessed output.

This piece of Makefile (processing defconfigs with C preprocessor) is U-Boot specific and doesn't exist in Linux kernel, thus it's unlikely to have someone noted the bahaviour difference of preprocessors before.

Honestly saying sedding the output isn't a clean solution, but I cannot come up with a better way. It'll be nice to have a flag to control the format of output, but I didn't find one.

Am looking forward to further suggestions and documentation on this.

Quoted 3 lines

Best regards

Heinrich

Thanks,
Yao Zi

Quoted 28 lines
Quoted 27 lines

unexpected data: # CONFIG_OF_BOARD_FIXUP is not set

Those options that is expected to be assigned explicitly with N will be set to the default value, messing up board configurations.

Let's sed these spaces away to ensure board configurations could be correctly generated with Clang.

Fixes 2027e99e61a ("Makefile: Run defconfig files through the C preprocessor")

---
scripts/kconfig/Makefile | 1 +

1 file changed, 1 insertion(+)
scripts/kconfig/Makefile +0 -0
index 079add4d5da..ba30652f01a 100644
@@ -94,6 +94,7 @@ endif
 

%_defconfig: $(obj)/conf

	$(Q)$(CPP) -nostdinc -P -I $(srctree) -undef -x assembler-with-cpp $(srctree)/arch/$(SRCARCH)/configs/$@ -o generated_defconfig
scripts/kconfig/Makefile +1 -0
+	$(Q)sed -i -e 's/^[[:space:]]//' generated_defconfig
	$(Q)$< $(silent) --defconfig=generated_defconfig $(Kconfig)

# Added for U-Boot (backward compatibility)

#4 lore raw

Re: [PATCH 1/2] Makefile: Strip leading spaces when preprocessing generated_defconfig

Quoted 29 lines · Yao Zi
Quoted 8 lines · Heinrich Schuchardt

Am 27. April 2025 16:50:10 MESZ schrieb Yao Zi <[email protected]>:

Quoted 4 lines

Clang's preprocessor may emit extra spaces for lines starting with '#'. Lines with these extra characters cannot be handled by Kconfig and will be ignored with warnings like,

Do you have an example for reprocing the issue?

Sure,

	clang-19 -E -nostdinc -P -I . -undef -x assembler-with-cpp \
		configs/starfive_visionfive2_defconfig

or a smaller example for demonstrating the behaviour,

cat << EOF | clang -E -P -x assembler-with-cpp - # comment line
normal line
EOF

and you could see the strange indentation. For reproducing the exact Kconfig warnings,

	make ARCH=riscv \
		CC='clang-19 --target=riscv64-unknown-linux-musl' \
		starfive_visionfive2_defconfig
Quoted 6 lines

(Clang is called clang-19 on my machine)

Quoted 1 line

Is there an understanding why Clang behaves in this way?

Sadly I have no idea. I guess it may serve for improving human-readability of the preprocessed output.

This is https://github.com/llvm/llvm-project/issues/78778

Signature · 1 line
Tom
author
#5 lore raw

Re: [PATCH 1/2] Makefile: Strip leading spaces when preprocessing generated_defconfig

Quoted 30 lines · Tom Rini
Quoted 29 lines · Yao Zi
Quoted 8 lines · Heinrich Schuchardt

Am 27. April 2025 16:50:10 MESZ schrieb Yao Zi <[email protected]>:

Quoted 4 lines

Clang's preprocessor may emit extra spaces for lines starting with '#'. Lines with these extra characters cannot be handled by Kconfig and will be ignored with warnings like,

Do you have an example for reprocing the issue?

Sure,

	clang-19 -E -nostdinc -P -I . -undef -x assembler-with-cpp \
		configs/starfive_visionfive2_defconfig

or a smaller example for demonstrating the behaviour,

cat << EOF | clang -E -P -x assembler-with-cpp - # comment line
normal line
EOF

and you could see the strange indentation. For reproducing the exact Kconfig warnings,

	make ARCH=riscv \
		CC='clang-19 --target=riscv64-unknown-linux-musl' \
		starfive_visionfive2_defconfig
Quoted 8 lines
Quoted 6 lines

(Clang is called clang-19 on my machine)

Quoted 1 line

Is there an understanding why Clang behaves in this way?

Sadly I have no idea. I guess it may serve for improving human-readability of the preprocessed output.

This is https://github.com/llvm/llvm-project/issues/78778

Thanks for the reference! I did a quick search among LLVM's issues before sending the patch but didn't find anything useful.

Do you think such workaround is acceptable? It seems the link should be included in commit message as well.

Quoted 2 lines
Signature · 1 line
Tom

Thanks,
Yao Zi

#6 lore raw

Re: [PATCH 1/2] Makefile: Strip leading spaces when preprocessing generated_defconfig

Quoted 38 lines · Yao Zi
Quoted 37 lines · Tom Rini
Quoted 35 lines · Yao Zi
Quoted 8 lines · Heinrich Schuchardt

Am 27. April 2025 16:50:10 MESZ schrieb Yao Zi <[email protected]>:

Quoted 4 lines

Clang's preprocessor may emit extra spaces for lines starting with '#'. Lines with these extra characters cannot be handled by Kconfig and will be ignored with warnings like,

Do you have an example for reprocing the issue?

Sure,

	clang-19 -E -nostdinc -P -I . -undef -x assembler-with-cpp \
		configs/starfive_visionfive2_defconfig

or a smaller example for demonstrating the behaviour,

cat << EOF | clang -E -P -x assembler-with-cpp - # comment line
normal line
EOF

and you could see the strange indentation. For reproducing the exact Kconfig warnings,

	make ARCH=riscv \
		CC='clang-19 --target=riscv64-unknown-linux-musl' \
		starfive_visionfive2_defconfig

(Clang is called clang-19 on my machine)

Quoted 1 line

Is there an understanding why Clang behaves in this way?

Sadly I have no idea. I guess it may serve for improving human-readability of the preprocessed output.

Quoted 7 lines
Quoted 1 line

This is https://github.com/llvm/llvm-project/issues/78778

Thanks for the reference! I did a quick search among LLVM's issues before sending the patch but didn't find anything useful.

Do you think such workaround is acceptable? It seems the link should be included in commit message as well.

I'd like to see if we can get llvm to fix this moving forward first. Can you please comment on the issue as you're hitting this now too?

Signature · 1 line
Tom
author
#7 lore raw

Re: [PATCH 1/2] Makefile: Strip leading spaces when preprocessing generated_defconfig

Quoted 41 lines · Tom Rini
Quoted 40 lines · Yao Zi
Quoted 38 lines · Tom Rini
Quoted 29 lines · Yao Zi
Quoted 8 lines · Heinrich Schuchardt

Am 27. April 2025 16:50:10 MESZ schrieb Yao Zi <[email protected]>:

Quoted 4 lines

Clang's preprocessor may emit extra spaces for lines starting with '#'. Lines with these extra characters cannot be handled by Kconfig and will be ignored with warnings like,

Do you have an example for reprocing the issue?

Sure,

	clang-19 -E -nostdinc -P -I . -undef -x assembler-with-cpp \
		configs/starfive_visionfive2_defconfig

or a smaller example for demonstrating the behaviour,

cat << EOF | clang -E -P -x assembler-with-cpp - # comment line
normal line
EOF

and you could see the strange indentation. For reproducing the exact Kconfig warnings,

	make ARCH=riscv \
		CC='clang-19 --target=riscv64-unknown-linux-musl' \
		starfive_visionfive2_defconfig
Quoted 6 lines

(Clang is called clang-19 on my machine)

Quoted 1 line

Is there an understanding why Clang behaves in this way?

Sadly I have no idea. I guess it may serve for improving human-readability of the preprocessed output.

This is https://github.com/llvm/llvm-project/issues/78778

Quoted 8 lines
Quoted 5 lines

Thanks for the reference! I did a quick search among LLVM's issues before sending the patch but didn't find anything useful.

Do you think such workaround is acceptable? It seems the link should be included in commit message as well.

I'd like to see if we can get llvm to fix this moving forward first. Can you please comment on the issue as you're hitting this now too?

Sure it should be fixed in LLVM and I've left a comment in the issue.

But even though it's fixed in upstream, it will be nice to have such compatibility workarounds for some time to keep compatibile with older Clang toolchains, about which my concern is most.

Quoted 2 lines
Signature · 1 line
Tom

Best regards,
Yao Zi

#8 lore raw

Re: [PATCH 1/2] Makefile: Strip leading spaces when preprocessing generated_defconfig

Quoted 48 lines · Yao Zi
Quoted 47 lines · Tom Rini
Quoted 45 lines · Yao Zi
Quoted 30 lines · Tom Rini
Quoted 29 lines · Yao Zi
Quoted 8 lines · Heinrich Schuchardt

Am 27. April 2025 16:50:10 MESZ schrieb Yao Zi <[email protected]>:

Quoted 4 lines

Clang's preprocessor may emit extra spaces for lines starting with '#'. Lines with these extra characters cannot be handled by Kconfig and will be ignored with warnings like,

Do you have an example for reprocing the issue?

Sure,

	clang-19 -E -nostdinc -P -I . -undef -x assembler-with-cpp \
		configs/starfive_visionfive2_defconfig

or a smaller example for demonstrating the behaviour,

cat << EOF | clang -E -P -x assembler-with-cpp - # comment line
normal line
EOF

and you could see the strange indentation. For reproducing the exact Kconfig warnings,

	make ARCH=riscv \
		CC='clang-19 --target=riscv64-unknown-linux-musl' \
		starfive_visionfive2_defconfig
Quoted 8 lines
Quoted 6 lines

(Clang is called clang-19 on my machine)

Quoted 1 line

Is there an understanding why Clang behaves in this way?

Sadly I have no idea. I guess it may serve for improving human-readability of the preprocessed output.

This is https://github.com/llvm/llvm-project/issues/78778

Thanks for the reference! I did a quick search among LLVM's issues before sending the patch but didn't find anything useful.

Do you think such workaround is acceptable? It seems the link should be included in commit message as well.

Quoted 8 lines
Quoted 2 lines

I'd like to see if we can get llvm to fix this moving forward first. Can you please comment on the issue as you're hitting this now too?

Sure it should be fixed in LLVM and I've left a comment in the issue.

But even though it's fixed in upstream, it will be nice to have such compatibility workarounds for some time to keep compatibile with older Clang toolchains, about which my concern is most.

Depending on what the LLVM people say, we can say what / how long of a workaround to keep in.

Signature · 1 line
Tom
author
#9 lore raw

PATCH 2/2 riscv: Access gd with inline assembly when building with LTO or Clang

In reply to Yao Zi

Similar to AArch64's case, Clang may wrongly fold accesses to gd pointer which is defined with register qualifier into constants, breaking various components.

This patch defines gd as a macro when building with Clang or LTO, which expands to get_gd() that accesses gp pointer in assembly, making RISC-V ports function properly and preparing for introduction of LTO in the future. Board initialization code is also adapted for non-assignable gd.

---
 arch/riscv/cpu/cpu.c                 |  6 ++++++
 arch/riscv/include/asm/global_data.h | 19 +++++++++++++++++++
 common/board_r.c                     |  4 +++-
 common/init/board_init.c             |  7 +++++--
 4 files changed, 33 insertions(+), 3 deletions(-)
arch/riscv/cpu/cpu.c +6 -0
index 5b31da64cbd..15c4e14599d 100644
@@ -18,6 +18,7 @@
 #include <asm/hwcap.h>
 #include <asm/cpufeature.h>
 #include <asm/cache.h>
+#include <asm/global_data.h>
 #include <dm/uclass-internal.h>
 #include <linux/bitops.h>
 #include <linux/log2.h>
@@ -746,3 +747,8 @@ __weak int cleanup_before_linux(void)
 
 	return 0;
 }
+
+void arch_setup_gd(gd_t *new_gd)
+{
+	set_gd(new_gd);
+}
arch/riscv/include/asm/global_data.h +19 -0
index d356752a56a..47b5e2cfc8f 100644
@@ -14,6 +14,7 @@
 #include <asm/smp.h>
 #include <asm/u-boot.h>
 #include <compiler.h>
+#include <config.h>
 
 /* Architecture-specific global data */
 struct arch_global_data {
@@ -47,8 +48,26 @@ struct arch_global_data {
 
 #include <asm-generic/global_data.h>
 
+#if defined(__clang__) || CONFIG_IS_ENABLED(LTO)
+
+#define DECLARE_GLOBAL_DATA_PTR
+#define gd			get_gd()
+
+static inline gd_t *get_gd(void)
+{
+	gd_t *gd_ptr;
+
+	__asm__ volatile ("mv %0, gp\n" : "=r" (gd_ptr));
+
+	return gd_ptr;
+}
+
+#else
+
 #define DECLARE_GLOBAL_DATA_PTR register gd_t *gd asm ("gp")
 
+#endif
+
 static inline void set_gd(volatile gd_t *gd_ptr)
 {
 #ifdef CONFIG_64BIT
common/board_r.c +3 -1
index bc6fd6448c2..4fa092ab596 100644
@@ -815,7 +815,9 @@ void board_init_r(gd_t *new_gd, ulong dest_addr)
 	if (CONFIG_IS_ENABLED(X86_64) && !IS_ENABLED(CONFIG_EFI_APP))
 		arch_setup_gd(new_gd);
 
-#if !defined(CONFIG_X86) && !defined(CONFIG_ARM) && !defined(CONFIG_ARM64)
+#if defined(CONFIG_RISCV)
+	set_gd(new_gd);
+#elif !defined(CONFIG_X86) && !defined(CONFIG_ARM) && !defined(CONFIG_ARM64)
 	gd = new_gd;
 #endif
 	gd->flags &= ~GD_FLG_LOG_READY;
common/init/board_init.c +5 -2
index a06ec1caa2c..2a6f39f51ad 100644
@@ -13,8 +13,11 @@
 
 DECLARE_GLOBAL_DATA_PTR;
 
-/* Unfortunately x86 or ARM can't compile this code as gd cannot be assigned */
-#if !defined(CONFIG_X86) && !defined(CONFIG_ARM)
+/*
+ * Unfortunately x86, ARM and RISC-V can't compile this code as gd is defined
+ * as macro and cannot be assigned.
+ */
+#if !defined(CONFIG_X86) && !defined(CONFIG_ARM) && !defined(CONFIG_RISCV)
 __weak void arch_setup_gd(struct global_data *gd_ptr)
 {
 	gd = gd_ptr;
Signature · 1 line
2.49.0
Reviewed-by on v1 2/2
#10 lore raw

Re: [PATCH 2/2] riscv: Access gd with inline assembly when building with LTO or Clang

Quoted 9 lines · Yao Zi

Similar to AArch64's case, Clang may wrongly fold accesses to gd pointer which is defined with register qualifier into constants, breaking various components.

This patch defines gd as a macro when building with Clang or LTO, which expands to get_gd() that accesses gp pointer in assembly, making RISC-V ports function properly and preparing for introduction of LTO in the future. Board initialization code is also adapted for non-assignable gd.

Quoted 8 lines
---
 arch/riscv/cpu/cpu.c                 |  6 ++++++
 arch/riscv/include/asm/global_data.h | 19 +++++++++++++++++++
 common/board_r.c                     |  4 +++-
 common/init/board_init.c             |  7 +++++--
 4 files changed, 33 insertions(+), 3 deletions(-)
#11 lore raw

Re: (subset) [PATCH 0/2] Clang compatibility improvements for RISC-V

In reply to Yao Zi
Quoted 10 lines · Yao Zi

Originally reported by Nathaniel Hourt[1], U-Boot fails to boot on JH7110 platforms (starfive_visionfive2_defconfig) when building with Clang.

This series fixes configuration generation with Clang's preprocessor and add an alternative codepath for RISC-V to access gd to workaround misoptimization of Clang, which is sufficient for building images for JH7110 boards.

[...]

Applied to u-boot/master, thanks!

[1/2] Makefile: Strip leading spaces when preprocessing generated_defconfig

      commit: 252cd205300c1b9945079c0ae1fe5980ef9dfa56
Signature · 1 line
Tom