summaryrefslogtreecommitdiffstats
diff options
context:
space:
mode:
authorGeorge Spelvin <linux@sciencehorizons.net>2016-05-25 14:19:49 -0400
committerGeorge Spelvin <linux@sciencehorizons.net>2016-05-28 15:48:58 -0400
commit4684fe95300c071983f77653e354c040fe80a265 (patch)
treefec7e5e3b1034b346a723d5f09684b362d9261fc
parent7b13277b682972c2ff8f6419e86c333d81936023 (diff)
downloadlinux-4684fe95300c071983f77653e354c040fe80a265.tar.gz
linux-4684fe95300c071983f77653e354c040fe80a265.tar.bz2
linux-4684fe95300c071983f77653e354c040fe80a265.zip
h8300: Add <asm/hash.h>
This will improve the performance of hash_32() and hash_64(), but due to complete lack of multi-bit shift instructions on H8, performance will still be bad in surrounding code. Designing H8-specific hash algorithms to work around that is a separate project. (But if the maintainers would like to get in touch...) Signed-off-by: George Spelvin <linux@sciencehorizons.net> Cc: Yoshinori Sato <ysato@users.sourceforge.jp> Cc: uclinux-h8-devel@lists.sourceforge.jp
-rw-r--r--arch/h8300/Kconfig1
-rw-r--r--arch/h8300/include/asm/hash.h53
2 files changed, 54 insertions, 0 deletions
diff --git a/arch/h8300/Kconfig b/arch/h8300/Kconfig
index 986ea84caaed..6c583dbbc119 100644
--- a/arch/h8300/Kconfig
+++ b/arch/h8300/Kconfig
@@ -20,6 +20,7 @@ config H8300
select HAVE_KERNEL_GZIP
select HAVE_KERNEL_LZO
select HAVE_ARCH_KGDB
+ select HAVE_ARCH_HASH
config RWSEM_GENERIC_SPINLOCK
def_bool y
diff --git a/arch/h8300/include/asm/hash.h b/arch/h8300/include/asm/hash.h
new file mode 100644
index 000000000000..04cfbd2bd850
--- /dev/null
+++ b/arch/h8300/include/asm/hash.h
@@ -0,0 +1,53 @@
+#ifndef _ASM_HASH_H
+#define _ASM_HASH_H
+
+/*
+ * The later H8SX models have a 32x32-bit multiply, but the H8/300H
+ * and H8S have only 16x16->32. Since it's tolerably compact, this is
+ * basically an inlined version of the __mulsi3 code. Since the inputs
+ * are not expected to be small, it's also simplfied by skipping the
+ * early-out checks.
+ *
+ * (Since neither CPU has any multi-bit shift instructions, a
+ * shift-and-add version is a non-starter.)
+ *
+ * TODO: come up with an arch-specific version of the hashing in fs/namei.c,
+ * since that is heavily dependent on rotates. Which, as mentioned, suck
+ * horribly on H8.
+ */
+
+#if defined(CONFIG_CPU_H300H) || defined(CONFIG_CPU_H8S)
+
+#define HAVE_ARCH__HASH_32 1
+
+/*
+ * Multiply by k = 0x61C88647. Fitting this into three registers requires
+ * one extra instruction, but reducing register pressure will probably
+ * make that back and then some.
+ *
+ * GCC asm note: %e1 is the high half of operand %1, while %f1 is the
+ * low half. So if %1 is er4, then %e1 is e4 and %f1 is r4.
+ *
+ * This has been designed to modify x in place, since that's the most
+ * common usage, but preserve k, since hash_64() makes two calls in
+ * quick succession.
+ */
+static inline u32 __attribute_const__ __hash_32(u32 x)
+{
+ u32 temp;
+
+ asm( "mov.w %e1,%f0"
+ "\n mulxu.w %f2,%0" /* klow * xhigh */
+ "\n mov.w %f0,%e1" /* The extra instruction */
+ "\n mov.w %f1,%f0"
+ "\n mulxu.w %e2,%0" /* khigh * xlow */
+ "\n add.w %e1,%f0"
+ "\n mulxu.w %f2,%1" /* klow * xlow */
+ "\n add.w %f0,%e1"
+ : "=&r" (temp), "=r" (x)
+ : "%r" (GOLDEN_RATIO_32), "1" (x));
+ return x;
+}
+
+#endif
+#endif /* _ASM_HASH_H */