From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1755329AbcE0PHo (ORCPT ); Fri, 27 May 2016 11:07:44 -0400 Received: from science.sciencehorizons.net ([71.41.210.147]:61184 "HELO ns.sciencehorizons.net" rhost-flags-OK-FAIL-OK-OK) by vger.kernel.org with SMTP id S1751206AbcE0PHn (ORCPT ); Fri, 27 May 2016 11:07:43 -0400 Date: 26 May 2016 13:21:01 -0400 Message-ID: <20160526172101.24571.qmail@ns.sciencehorizons.net> From: "George Spelvin" To: alistair.francis@xilinx.com, linux-kernel@vger.kernel.org, linux@sciencehorizons.net, michal.simek@xilinx.com Subject: [PATCH v2 09/10] microblaze: Add Cc: geert@linux-m68k.org, gerg@linux-m68k.org, linux-m68k@vger.kernel.org, phdm@macq.eu, schwab@linux-m68k.org, uclinux-h8-devel@lists.sourceforge.jp, ysato@users.sourceforge.jp In-Reply-To: <20160525073705.5792.qmail@ns.sciencehorizons.net> Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org Microblaze is an FPGA soft core that can be configured various ways. If it is configured without a multiplier, the standard __hash_32() will require a call to __mulsi3, which is a slow software loop. Instead, use a shift-and-add sequence for the constant multiply. GCC knows how to do this, but it's not as clever as some. Signed-off-by: George Spelvin Cc: Alistair Francis Cc: Michal Simek --- arch/microblaze/Kconfig | 1 + arch/microblaze/include/asm/hash.h | 81 ++++++++++++++++++++++++++++++++++++++ 2 files changed, 82 insertions(+) create mode 100644 arch/microblaze/include/asm/hash.h diff --git a/arch/microblaze/Kconfig b/arch/microblaze/Kconfig index 3d793b55..ce3e5125 100644 --- a/arch/microblaze/Kconfig +++ b/arch/microblaze/Kconfig @@ -16,6 +16,7 @@ config MICROBLAZE select GENERIC_IRQ_SHOW select GENERIC_PCI_IOMAP select GENERIC_SCHED_CLOCK + select HAVE_ARCH_HASH select HAVE_ARCH_KGDB select HAVE_DEBUG_KMEMLEAK select HAVE_DMA_API_DEBUG diff --git a/arch/microblaze/include/asm/hash.h b/arch/microblaze/include/asm/hash.h new file mode 100644 index 00000000..2f0c2a69 --- /dev/null +++ b/arch/microblaze/include/asm/hash.h @@ -0,0 +1,81 @@ +#ifndef _ASM_HASH_H +#define _ASM_HASH_H + +/* + * Fortunately, most people who want to run Linux on Microblaze enable + * both multiplier and barrel shifter, but omitting them is technically + * a supported configuration. + * + * With just a barrel shifter, we can implement an efficient constant + * multiply using shifts and adds. GCC can find a 9-step solution, but + * this 6-step solution was found by Yevgen Voronenko's implementation + * of the Hcub algorithm at http://spiral.ece.cmu.edu/mcm/gen.html. + * + * That software is really not designed for a single multiplier this large, + * but if you run it enough times with different seeds, it'll find several + * 6-shift, 6-add sequences for computing x * 0x61C88647. They are all + * c = (x << 19) + x; + * a = (x << 9) + c; + * b = (x << 23) + a; + * return (a<<11) + (b<<6) + (c<<3) - b; + * with variations on the order of the final add. + * + * Without even a shifter, it's hopless; any hash function will suck. + */ + +#if CONFIG_XILINX_MICROBLAZE0_USE_HW_MUL == 0 + +#define HAVE_ARCH__HASH_32 1 + +/* Multiply by GOLDEN_RATIO_32 = 0x61C88647 */ +static inline u32 __attribute_const__ __hash_32(u32 a) +{ +#if CONFIG_XILINX_MICROBLAZE0_USE_BARREL + unsigned b, c; + + /* Phase 1: Compute three intermediate values */ + b = a << 23; + c = (a << 19) + a; + a = (a << 9) + c; + b += a; + + /* Phase 2: Compute (a << 11) + (b << 6) + (c << 3) - b */ + a <<= 5; + a += b; /* (a << 5) + b */ + a <<= 3; + a += c; /* (a << 8) + (b << 3) + c */ + a <<= 3; + return a - b; /* (a << 11) + (b << 6) + (c << 3) - b */ +#else + /* + * "This is really going to hurt." + * + * Without a barrel shifter, left shifts are implemented as + * repeated additions, and the best we can do is an optimal + * addition-subtraction chain. This one is not known to be + * optimal, but at 37 steps, it's decent for a 31-bit multiplier. + * + * Question: given its size (37*4 = 148 bytes per instance), + * and slowness, is this worth having inline? + */ + unsigned b, c, d; + + b = a << 4; /* 4 */ + c = b << 1; /* 1 5 */ + b += a; /* 1 6 */ + c += b; /* 1 7 */ + c <<= 3; /* 3 10 */ + c -= a; /* 1 11 */ + d = c << 7; /* 7 18 */ + d += b; /* 1 19 */ + d <<= 8; /* 8 27 */ + d += a; /* 1 28 */ + d <<= 1; /* 1 29 */ + d += b; /* 1 30 */ + d <<= 6; /* 6 36 */ + return d + c; /* 1 37 total instructions*/ +#endif +} + +#endif /* !CONFIG_XILINX_MICROBLAZE0_USE_HW_MUL */ +#endif /* _ASM_HASH_H */ -- 2.8.1