diff mbox

[U-Boot,v2] arm: support Thumb-1 with CONFIG_SYS_THUMB_BUILD

Message ID 1424786016-12976-1-git-send-email-albert.u.boot@aribaud.net
State Changes Requested
Delegated to: Albert ARIBAUD
Headers show

Commit Message

Albert ARIBAUD Feb. 24, 2015, 1:53 p.m. UTC
When building a THumb-1-only target with CONFIG_SYS_THUMB_BUILD,
some files fail to build, most of the time because they include
mcr instructions, which only exist for Thumb-2.

Thos patch introduces a Kconfig option CONFIG_THUMB2 and uses
it to select between Thumb-2 and ARM mode for the aforementioned
files.

Signed-off-by: Albert ARIBAUD <albert.u.boot@aribaud.net>
---
This patch has been build-tested and run-tested on ED Mini V2,
above the "edmini: switch to SPL" patch, and found to reduce
U-Boot size by 25% and SPL size by 14%... and to run fine. :)

This patch has also been tested against side effects on the
non-Thumb wireless_space target. The binaries produced with
and without ths patch were found to differ only by their
version string.

Changes in v2:
- fixed a typo in the commit message
- added file arch/arm/thumb1/include/asm/proc-armv/system.h,
  which overrides arch/arm/include/asm/proc-armv/system.h
  when building for Thumb-1 and provides non-functional but
  Thumb-compilable IRQ and FIQ related macros and functions.

 Makefile                                       |  2 +
 arch/arm/Kconfig                               |  4 ++
 arch/arm/cpu/arm926ejs/Makefile                | 11 ++++
 arch/arm/lib/Makefile                          | 11 ++++
 arch/arm/mach-orion5x/Makefile                 | 10 ++++
 arch/arm/thumb1/include/asm/proc-armv/system.h | 69 ++++++++++++++++++++++++++
 examples/standalone/Makefile                   | 10 ++++
 include/configs/edminiv2.h                     |  4 +-
 8 files changed, 120 insertions(+), 1 deletion(-)
 create mode 100644 arch/arm/thumb1/include/asm/proc-armv/system.h

Comments

Albert ARIBAUD Feb. 24, 2015, 2 p.m. UTC | #1
On Tue, 24 Feb 2015 14:53:36 +0100, Albert ARIBAUD
<albert.u.boot@aribaud.net> wrote:

> When building a THumb-1-only target with CONFIG_SYS_THUMB_BUILD,

> - fixed a typo in the commit message

Drat. :/

If that version gets accepted, whoever commits it please s/THumb/Thumb/.

Amicalement,
Albert ARIBAUD Feb. 25, 2015, 10:09 p.m. UTC | #2
Hello Albert,

On Tue, 24 Feb 2015 14:53:36 +0100, Albert ARIBAUD
<albert.u.boot@aribaud.net> wrote:
> When building a THumb-1-only target with CONFIG_SYS_THUMB_BUILD,
> some files fail to build, most of the time because they include
> mcr instructions, which only exist for Thumb-2.
> 
> Thos patch introduces a Kconfig option CONFIG_THUMB2 and uses
> it to select between Thumb-2 and ARM mode for the aforementioned
> files.
> 
> Signed-off-by: Albert ARIBAUD <albert.u.boot@aribaud.net>
> ---
> This patch has been build-tested and run-tested on ED Mini V2,
> above the "edmini: switch to SPL" patch, and found to reduce
> U-Boot size by 25% and SPL size by 14%... and to run fine. :)
> 
> This patch has also been tested against side effects on the
> non-Thumb wireless_space target. The binaries produced with
> and without ths patch were found to differ only by their
> version string.
> 
> Changes in v2:
> - fixed a typo in the commit message
> - added file arch/arm/thumb1/include/asm/proc-armv/system.h,
>   which overrides arch/arm/include/asm/proc-armv/system.h
>   when building for Thumb-1 and provides non-functional but
>   Thumb-compilable IRQ and FIQ related macros and functions.

Ok, this does not fare as good as I have hoped, because there are also
thumb1-unfrendly macros in arch/arm/include/asm/cache.h, this time with
mcr and mrc instructions.

/me hates macros used as inline functions. :(

I cannot just replace the macros with empty inline functions as I did
with arch/arm/include/asm/proc-armv/system.h, since this would cause
cache to not work. :(

This leaves only one solution: when buildding for thumb1, replace the
macros in cache.h with plain function prototypes, and provide simple
ARM-state implementations for these. We'll lose a bit of performance as
instead of a single mcr or mrc instruction, we'll have a call switching
from Thumb to ARM state, the mcr/mrc instruction, and a return to Thumb
state. Hopefully that won't hurt performance too much.

V3 in the works.

Amicalement,
Stefan Roese April 9, 2015, 9:22 a.m. UTC | #3
Hi Albert,

On 25.02.2015 23:09, Albert ARIBAUD wrote:
> On Tue, 24 Feb 2015 14:53:36 +0100, Albert ARIBAUD
> <albert.u.boot@aribaud.net> wrote:
>> When building a THumb-1-only target with CONFIG_SYS_THUMB_BUILD,
>> some files fail to build, most of the time because they include
>> mcr instructions, which only exist for Thumb-2.
>>
>> Thos patch introduces a Kconfig option CONFIG_THUMB2 and uses
>> it to select between Thumb-2 and ARM mode for the aforementioned
>> files.
>>
>> Signed-off-by: Albert ARIBAUD <albert.u.boot@aribaud.net>
>> ---
>> This patch has been build-tested and run-tested on ED Mini V2,
>> above the "edmini: switch to SPL" patch, and found to reduce
>> U-Boot size by 25% and SPL size by 14%... and to run fine. :)
>>
>> This patch has also been tested against side effects on the
>> non-Thumb wireless_space target. The binaries produced with
>> and without ths patch were found to differ only by their
>> version string.
>>
>> Changes in v2:
>> - fixed a typo in the commit message
>> - added file arch/arm/thumb1/include/asm/proc-armv/system.h,
>>    which overrides arch/arm/include/asm/proc-armv/system.h
>>    when building for Thumb-1 and provides non-functional but
>>    Thumb-compilable IRQ and FIQ related macros and functions.
>
> Ok, this does not fare as good as I have hoped, because there are also
> thumb1-unfrendly macros in arch/arm/include/asm/cache.h, this time with
> mcr and mrc instructions.
>
> /me hates macros used as inline functions. :(
>
> I cannot just replace the macros with empty inline functions as I did
> with arch/arm/include/asm/proc-armv/system.h, since this would cause
> cache to not work. :(
>
> This leaves only one solution: when buildding for thumb1, replace the
> macros in cache.h with plain function prototypes, and provide simple
> ARM-state implementations for these. We'll lose a bit of performance as
> instead of a single mcr or mrc instruction, we'll have a call switching
> from Thumb to ARM state, the mcr/mrc instruction, and a return to Thumb
> state. Hopefully that won't hurt performance too much.
>
> V3 in the works.

Any update on this?

Thanks,
Stefan
Albert ARIBAUD April 13, 2015, 5:47 p.m. UTC | #4
Hello Stefan,

On Thu, 09 Apr 2015 11:22:27 +0200, Stefan Roese <sr@denx.de> wrote:
> Hi Albert,
> 
> On 25.02.2015 23:09, Albert ARIBAUD wrote:
> > On Tue, 24 Feb 2015 14:53:36 +0100, Albert ARIBAUD
> > <albert.u.boot@aribaud.net> wrote:
> >> When building a THumb-1-only target with CONFIG_SYS_THUMB_BUILD,
> >> some files fail to build, most of the time because they include
> >> mcr instructions, which only exist for Thumb-2.
> >>
> >> Thos patch introduces a Kconfig option CONFIG_THUMB2 and uses
> >> it to select between Thumb-2 and ARM mode for the aforementioned
> >> files.
> >>
> >> Signed-off-by: Albert ARIBAUD <albert.u.boot@aribaud.net>
> >> ---
> >> This patch has been build-tested and run-tested on ED Mini V2,
> >> above the "edmini: switch to SPL" patch, and found to reduce
> >> U-Boot size by 25% and SPL size by 14%... and to run fine. :)
> >>
> >> This patch has also been tested against side effects on the
> >> non-Thumb wireless_space target. The binaries produced with
> >> and without ths patch were found to differ only by their
> >> version string.
> >>
> >> Changes in v2:
> >> - fixed a typo in the commit message
> >> - added file arch/arm/thumb1/include/asm/proc-armv/system.h,
> >>    which overrides arch/arm/include/asm/proc-armv/system.h
> >>    when building for Thumb-1 and provides non-functional but
> >>    Thumb-compilable IRQ and FIQ related macros and functions.
> >
> > Ok, this does not fare as good as I have hoped, because there are also
> > thumb1-unfrendly macros in arch/arm/include/asm/cache.h, this time with
> > mcr and mrc instructions.
> >
> > /me hates macros used as inline functions. :(
> >
> > I cannot just replace the macros with empty inline functions as I did
> > with arch/arm/include/asm/proc-armv/system.h, since this would cause
> > cache to not work. :(
> >
> > This leaves only one solution: when buildding for thumb1, replace the
> > macros in cache.h with plain function prototypes, and provide simple
> > ARM-state implementations for these. We'll lose a bit of performance as
> > instead of a single mcr or mrc instruction, we'll have a call switching
> > from Thumb to ARM state, the mcr/mrc instruction, and a return to Thumb
> > state. Hopefully that won't hurt performance too much.
> >
> > V3 in the works.
> 
> Any update on this?

I am just restarting on this right now.

> Thanks,
> Stefan

Amicalement,
diff mbox

Patch

diff --git a/Makefile b/Makefile
index bd4abab..7dc9dc6 100644
--- a/Makefile
+++ b/Makefile
@@ -593,6 +593,8 @@  KBUILD_CFLAGS += $(KCFLAGS)
 UBOOTINCLUDE    := \
 		-Iinclude \
 		$(if $(KBUILD_SRC), -I$(srctree)/include) \
+		$(if $(CONFIG_SYS_THUMB_BUILD), $(if $(CONFIG_HAS_THUMB2),, \
+			-I$(srctree)/arch/$(ARCH)/thumb1/include),) \
 		-I$(srctree)/arch/$(ARCH)/include \
 		-include $(srctree)/include/linux/kconfig.h
 
diff --git a/arch/arm/Kconfig b/arch/arm/Kconfig
index 820ba1c..35420ca 100644
--- a/arch/arm/Kconfig
+++ b/arch/arm/Kconfig
@@ -10,6 +10,9 @@  config ARM64
 config HAS_VBAR
         bool
 
+config HAS_THUMB2
+        bool
+
 config CPU_ARM720T
         bool
 
@@ -32,6 +35,7 @@  config CPU_ARM1176
 config CPU_V7
         bool
         select HAS_VBAR
+        select HAS_THUMB2
 
 config CPU_PXA
         bool
diff --git a/arch/arm/cpu/arm926ejs/Makefile b/arch/arm/cpu/arm926ejs/Makefile
index f5944cc..090d64b 100644
--- a/arch/arm/cpu/arm926ejs/Makefile
+++ b/arch/arm/cpu/arm926ejs/Makefile
@@ -22,3 +22,14 @@  obj-$(CONFIG_MX27) += mx27/
 obj-$(if $(filter mxs,$(SOC)),y) += mxs/
 obj-$(CONFIG_PANTHEON) += pantheon/
 obj-$(if $(filter spear,$(SOC)),y) += spear/
+
+# some files can only build in ARM or THUMB2, not THUMB1
+
+ifdef CONFIG_SYS_THUMB_BUILD
+ifndef CONFIG_HAS_THUMB2
+
+CFLAGS_cpu.o := -marm
+CFLAGS_cache.o := -marm
+
+endif
+endif
diff --git a/arch/arm/lib/Makefile b/arch/arm/lib/Makefile
index da8ed72..a753bca 100644
--- a/arch/arm/lib/Makefile
+++ b/arch/arm/lib/Makefile
@@ -55,3 +55,14 @@  obj-$(CONFIG_DEBUG_LL)	+= debug.o
 ifneq (,$(findstring -mabi=aapcs-linux,$(PLATFORM_CPPFLAGS)))
 extra-y	+= eabi_compat.o
 endif
+
+# some files can only build in ARM or THUMB2, not THUMB1
+
+ifdef CONFIG_SYS_THUMB_BUILD
+ifndef CONFIG_HAS_THUMB2
+
+CFLAGS_cache.o := -marm
+CFLAGS_cache-cp15.o := -marm
+
+endif
+endif
diff --git a/arch/arm/mach-orion5x/Makefile b/arch/arm/mach-orion5x/Makefile
index 546ebcb..33dcad4 100644
--- a/arch/arm/mach-orion5x/Makefile
+++ b/arch/arm/mach-orion5x/Makefile
@@ -16,3 +16,13 @@  obj-y	+= timer.o
 ifndef CONFIG_SKIP_LOWLEVEL_INIT
 obj-y	+= lowlevel_init.o
 endif
+
+# some files can only build in ARM or THUMB2, not THUMB1
+
+ifdef CONFIG_SYS_THUMB_BUILD
+ifndef CONFIG_HAS_THUMB2
+
+CFLAGS_cpu.o := -marm
+
+endif
+endif
diff --git a/arch/arm/thumb1/include/asm/proc-armv/system.h b/arch/arm/thumb1/include/asm/proc-armv/system.h
new file mode 100644
index 0000000..7dfbf3d
--- /dev/null
+++ b/arch/arm/thumb1/include/asm/proc-armv/system.h
@@ -0,0 +1,69 @@ 
+/*
+ *  Thumb-1 drop-in for the linux/include/asm-arm/proc-armv/system.h
+ *
+ *  (C) Copyright 2015
+ *  Albert ARIBAUD <albert.u.boot@aribaud.net>
+ *
+ * The original file does not build in Thumb mode. However, in U-Boot
+ * we don't use interrupt context, so we can redefine these as empty
+ * memory barriers, which makes Thumb-1 compiler happy.
+ *
+ *  SPDX-License-Identifier:	GPL-2.0+
+ */
+
+/*
+ * Use the same macro name as linux/include/asm-arm/proc-armv/system.h
+ * here, so that if the original ever gets included after us, it won't
+ * try to re-redefine anything.
+ */
+
+#ifndef __ASM_PROC_SYSTEM_H
+#define __ASM_PROC_SYSTEM_H
+
+/*
+ * Redefine all original macros with static inline functions containing
+ * a simple memory barrier, so that they produce the same instruction
+ * ordering constraints as their original counterparts.
+ * We use static inline functions rather than macros so that we can tell
+ * the compiler to not complain about unused arguments.
+ */
+
+static inline void local_irq_save(
+	unsigned long flags __attribute__((unused)))
+{
+	__asm__ __volatile__ ("" : : : "memory");
+}
+
+static inline void local_irq_enable(void)
+{
+	__asm__ __volatile__ ("" : : : "memory");
+}
+
+static inline void local_irq_disable(void)
+{
+	__asm__ __volatile__ ("" : : : "memory");
+}
+
+static inline void __stf(void)
+{
+	__asm__ __volatile__ ("" : : : "memory");
+}
+
+static inline void __clf(void)
+{
+	__asm__ __volatile__ ("" : : : "memory");
+}
+
+static inline void local_save_flags(
+	unsigned long flags __attribute__((unused)))
+{
+	__asm__ __volatile__ ("" : : : "memory");
+}
+
+static inline void local_irq_restore(
+	unsigned long flags __attribute__((unused)))
+{
+	__asm__ __volatile__ ("" : : : "memory");
+}
+
+#endif /*  __ASM_PROC_SYSTEM_H */
diff --git a/examples/standalone/Makefile b/examples/standalone/Makefile
index 0863a8c..5a6ae00 100644
--- a/examples/standalone/Makefile
+++ b/examples/standalone/Makefile
@@ -73,3 +73,13 @@  $(obj)/%.srec: $(obj)/% FORCE
 $(obj)/%.bin: OBJCOPYFLAGS := -O binary
 $(obj)/%.bin: $(obj)/% FORCE
 	$(call if_changed,objcopy)
+
+# some files can only build in ARM or THUMB2, not THUMB1
+
+ifdef CONFIG_SYS_THUMB_BUILD
+ifndef CONFIG_HAS_THUMB2
+
+CFLAGS_stubs.o := -marm
+
+endif
+endif
diff --git a/include/configs/edminiv2.h b/include/configs/edminiv2.h
index 70a698a..112583b 100644
--- a/include/configs/edminiv2.h
+++ b/include/configs/edminiv2.h
@@ -12,6 +12,8 @@ 
 #ifndef _CONFIG_EDMINIV2_H
 #define _CONFIG_EDMINIV2_H
 
+#define CONFIG_SYS_THUMB_BUILD
+
 /*
  * Version number information
  */
@@ -203,7 +205,7 @@ 
 /*
  * Size of malloc() pool
  */
-#define CONFIG_SYS_MALLOC_LEN	(1024 * 256) /* 256kB for malloc() */
+#define CONFIG_SYS_MALLOC_LEN		(1024 * 1024) /* 1MB for malloc() */
 
 /*
  * Other required minimal configurations