[U-Boot] [RFC PATCH] SPL: replace ctype implementation with tiny version
The ctype implementation (isdigit() & friends) works with an array of 256 Bytes - one for each character. This is pretty big in SPL terms, so let's replace this "bloated" implementation with a tiny version using C statements. This only implements the functions that the SPL requires and confines this change only to an actual SPL build. Saves about 200 Bytes from the SPL code size.
Signed-off-by: Andre Przywara andre.przywara@arm.com --- Hi,
some people voiced concerns about running out of SPL code space when adding new features. In this particular case this was an issue when looking at the SPL FIT extension series[1]. This patch here on top of this series saves more space than the SPL FIT series consumed, so I trade this as a bait to people wrestling with this problem ;-)
Cheers, Andre.
[1] http://lists.denx.de/pipermail/u-boot/2017-January/278772.html
include/linux/tiny_ctype.h | 12 ++++++++++++ lib/Makefile | 2 ++ lib/strto.c | 4 ++++ 3 files changed, 18 insertions(+) create mode 100644 include/linux/tiny_ctype.h
diff --git a/include/linux/tiny_ctype.h b/include/linux/tiny_ctype.h new file mode 100644 index 0000000..4910412 --- /dev/null +++ b/include/linux/tiny_ctype.h @@ -0,0 +1,12 @@ +#ifndef _LINUX_CTYPE_H +#define _LINUX_CTYPE_H + +#define isdigit(c) (((c) >= '0') && ((c) <= '9')) +#define isxdigit(c) (isdigit(c) || \ + ((c) >= 'A' && (c) <= 'F') || \ + ((c) >= 'a' && (c) <= 'f')) +#define islower(c) ((((c) >= 'a') && ((c) <= 'z')) || ((c) >= 223)) + +#define toupper(c) ((((c) >= 'a') && ((c) <= 'z')) ? (c) - 32 : (c)) + +#endif diff --git a/lib/Makefile b/lib/Makefile index 23e9f1e..15385fd 100644 --- a/lib/Makefile +++ b/lib/Makefile @@ -66,7 +66,9 @@ obj-y += display_options.o CFLAGS_display_options.o := $(if $(BUILD_TAG),-DBUILD_TAG='"$(BUILD_TAG)"') obj-$(CONFIG_BCH) += bch.o obj-y += crc32.o +ifndef CONFIG_SPL_BUILD obj-y += ctype.o +endif obj-y += div64.o obj-y += hang.o obj-y += linux_compat.o diff --git a/lib/strto.c b/lib/strto.c index e93a4f5..9e9ba75 100644 --- a/lib/strto.c +++ b/lib/strto.c @@ -11,7 +11,11 @@
#include <common.h> #include <errno.h> +#ifdef CONFIG_SPL_BUILD +#include <linux/tiny_ctype.h> +#else #include <linux/ctype.h> +#endif
unsigned long simple_strtoul(const char *cp, char **endp, unsigned int base)
On Fri, Jan 20, 2017 at 10:33:28PM +0000, Andre Przywara wrote:
The ctype implementation (isdigit() & friends) works with an array of 256 Bytes - one for each character. This is pretty big in SPL terms, so let's replace this "bloated" implementation with a tiny version using C statements. This only implements the functions that the SPL requires and confines this change only to an actual SPL build. Saves about 200 Bytes from the SPL code size.
Signed-off-by: Andre Przywara andre.przywara@arm.com
Hi,
some people voiced concerns about running out of SPL code space when adding new features. In this particular case this was an issue when looking at the SPL FIT extension series[1]. This patch here on top of this series saves more space than the SPL FIT series consumed, so I trade this as a bait to people wrestling with this problem ;-)
So.. where might this fail? Also, this doesn't look like it comes from the kernel exactly, so "linux/tiny_ctype.h" doesn't seem right (and, is missing the license boilerplate). It does bear some resemblance to the kernel ctype header so it would of course be correct to attribute that in the boilerplate.
On 22/01/17 16:09, Tom Rini wrote:
On Fri, Jan 20, 2017 at 10:33:28PM +0000, Andre Przywara wrote:
The ctype implementation (isdigit() & friends) works with an array of 256 Bytes - one for each character. This is pretty big in SPL terms, so let's replace this "bloated" implementation with a tiny version using C statements. This only implements the functions that the SPL requires and confines this change only to an actual SPL build. Saves about 200 Bytes from the SPL code size.
Signed-off-by: Andre Przywara andre.przywara@arm.com
Hi,
some people voiced concerns about running out of SPL code space when adding new features. In this particular case this was an issue when looking at the SPL FIT extension series[1]. This patch here on top of this series saves more space than the SPL FIT series consumed, so I trade this as a bait to people wrestling with this problem ;-)
So.. where might this fail?
What do you mean by that?
Also, this doesn't look like it comes from the kernel exactly, so "linux/tiny_ctype.h" doesn't seem right (and, is missing the license boilerplate). It does bear some resemblance to the kernel ctype header so it would of course be correct to attribute that in the boilerplate.
Actually I quickly hacked that up while looking at the existing map, trying to convert it into statements. It was mostly an experiment to see if that gives me something (hence the RFC tag).
If the general idea is acceptable, I can surely make this pretty.
Cheers, Andre.
On Sun, Jan 22, 2017 at 11:30:04PM +0000, André Przywara wrote:
On 22/01/17 16:09, Tom Rini wrote:
On Fri, Jan 20, 2017 at 10:33:28PM +0000, Andre Przywara wrote:
The ctype implementation (isdigit() & friends) works with an array of 256 Bytes - one for each character. This is pretty big in SPL terms, so let's replace this "bloated" implementation with a tiny version using C statements. This only implements the functions that the SPL requires and confines this change only to an actual SPL build. Saves about 200 Bytes from the SPL code size.
Signed-off-by: Andre Przywara andre.przywara@arm.com
Hi,
some people voiced concerns about running out of SPL code space when adding new features. In this particular case this was an issue when looking at the SPL FIT extension series[1]. This patch here on top of this series saves more space than the SPL FIT series consumed, so I trade this as a bait to people wrestling with this problem ;-)
So.. where might this fail?
What do you mean by that?
Also, this doesn't look like it comes from the kernel exactly, so "linux/tiny_ctype.h" doesn't seem right (and, is missing the license boilerplate). It does bear some resemblance to the kernel ctype header so it would of course be correct to attribute that in the boilerplate.
Actually I quickly hacked that up while looking at the existing map, trying to convert it into statements. It was mostly an experiment to see if that gives me something (hence the RFC tag).
If the general idea is acceptable, I can surely make this pretty.
Let me answer both parts at once. If this is as functional in all cases, why not just always do this instead?
On 01/20/2017 04:33 PM, Andre Przywara wrote:
The ctype implementation (isdigit() & friends) works with an array of 256 Bytes - one for each character. This is pretty big in SPL terms, so let's replace this "bloated" implementation with a tiny version using C statements. This only implements the functions that the SPL requires and confines this change only to an actual SPL build. Saves about 200 Bytes from the SPL code size.
Saves us a very needed 256 bytes on AM335x!
Acked-by: Andrew F. Davis afd@ti.com
Thanks, Andrew
Signed-off-by: Andre Przywara andre.przywara@arm.com
Hi,
some people voiced concerns about running out of SPL code space when adding new features. In this particular case this was an issue when looking at the SPL FIT extension series[1]. This patch here on top of this series saves more space than the SPL FIT series consumed, so I trade this as a bait to people wrestling with this problem ;-)
Cheers, Andre.
[1] http://lists.denx.de/pipermail/u-boot/2017-January/278772.html
include/linux/tiny_ctype.h | 12 ++++++++++++ lib/Makefile | 2 ++ lib/strto.c | 4 ++++ 3 files changed, 18 insertions(+) create mode 100644 include/linux/tiny_ctype.h
diff --git a/include/linux/tiny_ctype.h b/include/linux/tiny_ctype.h new file mode 100644 index 0000000..4910412 --- /dev/null +++ b/include/linux/tiny_ctype.h @@ -0,0 +1,12 @@ +#ifndef _LINUX_CTYPE_H +#define _LINUX_CTYPE_H
+#define isdigit(c) (((c) >= '0') && ((c) <= '9')) +#define isxdigit(c) (isdigit(c) || \
((c) >= 'A' && (c) <= 'F') || \((c) >= 'a' && (c) <= 'f'))+#define islower(c) ((((c) >= 'a') && ((c) <= 'z')) || ((c) >= 223))
+#define toupper(c) ((((c) >= 'a') && ((c) <= 'z')) ? (c) - 32 : (c))
+#endif diff --git a/lib/Makefile b/lib/Makefile index 23e9f1e..15385fd 100644 --- a/lib/Makefile +++ b/lib/Makefile @@ -66,7 +66,9 @@ obj-y += display_options.o CFLAGS_display_options.o := $(if $(BUILD_TAG),-DBUILD_TAG='"$(BUILD_TAG)"') obj-$(CONFIG_BCH) += bch.o obj-y += crc32.o +ifndef CONFIG_SPL_BUILD obj-y += ctype.o +endif obj-y += div64.o obj-y += hang.o obj-y += linux_compat.o diff --git a/lib/strto.c b/lib/strto.c index e93a4f5..9e9ba75 100644 --- a/lib/strto.c +++ b/lib/strto.c @@ -11,7 +11,11 @@
#include <common.h> #include <errno.h> +#ifdef CONFIG_SPL_BUILD +#include <linux/tiny_ctype.h> +#else #include <linux/ctype.h> +#endif
unsigned long simple_strtoul(const char *cp, char **endp, unsigned int base)
On 01/22/2017 03:58 PM, Andrew F. Davis wrote:
On 01/20/2017 04:33 PM, Andre Przywara wrote:
The ctype implementation (isdigit() & friends) works with an array of 256 Bytes - one for each character. This is pretty big in SPL terms, so let's replace this "bloated" implementation with a tiny version using C statements. This only implements the functions that the SPL requires and confines this change only to an actual SPL build. Saves about 200 Bytes from the SPL code size.
Saves us a very needed 256 bytes on AM335x!
Acked-by: Andrew F. Davis afd@ti.com
Thanks, Andrew
Signed-off-by: Andre Przywara andre.przywara@arm.com
Hi,
some people voiced concerns about running out of SPL code space when adding new features. In this particular case this was an issue when looking at the SPL FIT extension series[1]. This patch here on top of this series saves more space than the SPL FIT series consumed, so I trade this as a bait to people wrestling with this problem ;-)
Cheers, Andre.
[1] http://lists.denx.de/pipermail/u-boot/2017-January/278772.html
include/linux/tiny_ctype.h | 12 ++++++++++++ lib/Makefile | 2 ++ lib/strto.c | 4 ++++ 3 files changed, 18 insertions(+) create mode 100644 include/linux/tiny_ctype.h
diff --git a/include/linux/tiny_ctype.h b/include/linux/tiny_ctype.h new file mode 100644 index 0000000..4910412 --- /dev/null +++ b/include/linux/tiny_ctype.h @@ -0,0 +1,12 @@ +#ifndef _LINUX_CTYPE_H +#define _LINUX_CTYPE_H
+#define isdigit(c) (((c) >= '0') && ((c) <= '9')) +#define isxdigit(c) (isdigit(c) || \
((c) >= 'A' && (c) <= 'F') || \((c) >= 'a' && (c) <= 'f'))+#define islower(c) ((((c) >= 'a') && ((c) <= 'z')) || ((c) >= 223))
+#define toupper(c) ((((c) >= 'a') && ((c) <= 'z')) ? (c) - 32 : (c))
+#endif diff --git a/lib/Makefile b/lib/Makefile index 23e9f1e..15385fd 100644 --- a/lib/Makefile +++ b/lib/Makefile @@ -66,7 +66,9 @@ obj-y += display_options.o CFLAGS_display_options.o := $(if $(BUILD_TAG),-DBUILD_TAG='"$(BUILD_TAG)"') obj-$(CONFIG_BCH) += bch.o obj-y += crc32.o +ifndef CONFIG_SPL_BUILD obj-y += ctype.o +endif
One minor note, this causes build failures for code that uses the functions not defined here (or not including this header). Keeping the same space saving technique, we could avoid those changes by moving the SPL_BUILD check to inside ctype.h, then define a set of macros that don't use the _ctype table. The linker would drop the table if it is not needed without needing to have the Makefile change.
Andrew
obj-y += div64.o obj-y += hang.o obj-y += linux_compat.o diff --git a/lib/strto.c b/lib/strto.c index e93a4f5..9e9ba75 100644 --- a/lib/strto.c +++ b/lib/strto.c @@ -11,7 +11,11 @@
#include <common.h> #include <errno.h> +#ifdef CONFIG_SPL_BUILD +#include <linux/tiny_ctype.h> +#else #include <linux/ctype.h> +#endif
unsigned long simple_strtoul(const char *cp, char **endp, unsigned int base)
U-Boot mailing list U-Boot@lists.denx.de http://lists.denx.de/mailman/listinfo/u-boot
On Fri, Jan 20, 2017 at 10:33:28PM +0000, Andre Przywara wrote:
The ctype implementation (isdigit() & friends) works with an array of 256 Bytes - one for each character. This is pretty big in SPL terms, so let's replace this "bloated" implementation with a tiny version using C statements. This only implements the functions that the SPL requires and confines this change only to an actual SPL build. Saves about 200 Bytes from the SPL code size.
Signed-off-by: Andre Przywara andre.przywara@arm.com Acked-by: Andrew F. Davis afd@ti.com
Hi,
some people voiced concerns about running out of SPL code space when adding new features. In this particular case this was an issue when looking at the SPL FIT extension series[1]. This patch here on top of this series saves more space than the SPL FIT series consumed, so I trade this as a bait to people wrestling with this problem ;-)
Cheers, Andre.
[1] http://lists.denx.de/pipermail/u-boot/2017-January/278772.html
include/linux/tiny_ctype.h | 12 ++++++++++++ lib/Makefile | 2 ++ lib/strto.c | 4 ++++ 3 files changed, 18 insertions(+) create mode 100644 include/linux/tiny_ctype.h
So, cycling back. I think this is a good idea, but breaks on say socfpga_de0_nano_soc currently: arm: + socfpga_de0_nano_soc +(socfpga_de0_nano_soc) lib/built-in.o: In function `_vprintf': +(socfpga_de0_nano_soc) lib/tiny-printf.c:315: undefined reference to `_ctype' +(socfpga_de0_nano_soc) make[2]: *** [spl/u-boot-spl] Error 1 +(socfpga_de0_nano_soc) make[1]: *** [spl/u-boot-spl] Error 2 +(socfpga_de0_nano_soc) make: *** [sub-make] Error 2
Hi,
On 05/06/17 16:01, Tom Rini wrote:
On Fri, Jan 20, 2017 at 10:33:28PM +0000, Andre Przywara wrote:
The ctype implementation (isdigit() & friends) works with an array of 256 Bytes - one for each character. This is pretty big in SPL terms, so let's replace this "bloated" implementation with a tiny version using C statements. This only implements the functions that the SPL requires and confines this change only to an actual SPL build. Saves about 200 Bytes from the SPL code size.
Signed-off-by: Andre Przywara andre.przywara@arm.com Acked-by: Andrew F. Davis afd@ti.com
Hi,
some people voiced concerns about running out of SPL code space when adding new features. In this particular case this was an issue when looking at the SPL FIT extension series[1]. This patch here on top of this series saves more space than the SPL FIT series consumed, so I trade this as a bait to people wrestling with this problem ;-)
Cheers, Andre.
[1] http://lists.denx.de/pipermail/u-boot/2017-January/278772.html
include/linux/tiny_ctype.h | 12 ++++++++++++ lib/Makefile | 2 ++ lib/strto.c | 4 ++++ 3 files changed, 18 insertions(+) create mode 100644 include/linux/tiny_ctype.h
So, cycling back. I think this is a good idea, but breaks on say socfpga_de0_nano_soc currently: arm: + socfpga_de0_nano_soc +(socfpga_de0_nano_soc) lib/built-in.o: In function `_vprintf': +(socfpga_de0_nano_soc) lib/tiny-printf.c:315: undefined reference to `_ctype' +(socfpga_de0_nano_soc) make[2]: *** [spl/u-boot-spl] Error 1 +(socfpga_de0_nano_soc) make[1]: *** [spl/u-boot-spl] Error 2 +(socfpga_de0_nano_soc) make: *** [sub-make] Error 2
Yeah, I found this version broken too for some buildman targets. So I came up with some fixes, and IIRC at least this compiled cleanly. Unfortunately some boards *regressed* in terms of size, apparently because the ctype functions are now inlined and if a board calls them several times we end up with multiple copies, offsetting the code size savings, which were quite substantial for many boards, otherwise. I tried one more fix, which improved the situation, but still left boards with bigger code sizes. At this point I gave up, but I can post what I had so far for others to pick up.
Cheers, Andre.
participants (4)
-
Andre Przywara -
Andrew F. Davis -
André Przywara
-
Tom Rini