162306a36Sopenharmony_ci/* SPDX-License-Identifier: GPL-2.0 */ 262306a36Sopenharmony_ci/* 362306a36Sopenharmony_ci * Copyright (C) 1999-2002 Hewlett-Packard Co 462306a36Sopenharmony_ci * Stephane Eranian <eranian@hpl.hp.com> 562306a36Sopenharmony_ci * David Mosberger-Tang <davidm@hpl.hp.com> 662306a36Sopenharmony_ci * Copyright (C) 2002 Ken Chen <kenneth.w.chen@intel.com> 762306a36Sopenharmony_ci * 862306a36Sopenharmony_ci * 1/06/01 davidm Tuned for Itanium. 962306a36Sopenharmony_ci * 2/12/02 kchen Tuned for both Itanium and McKinley 1062306a36Sopenharmony_ci * 3/08/02 davidm Some more tweaking 1162306a36Sopenharmony_ci */ 1262306a36Sopenharmony_ci 1362306a36Sopenharmony_ci#include <linux/export.h> 1462306a36Sopenharmony_ci#include <asm/asmmacro.h> 1562306a36Sopenharmony_ci#include <asm/page.h> 1662306a36Sopenharmony_ci 1762306a36Sopenharmony_ci#ifdef CONFIG_ITANIUM 1862306a36Sopenharmony_ci# define L3_LINE_SIZE 64 // Itanium L3 line size 1962306a36Sopenharmony_ci# define PREFETCH_LINES 9 // magic number 2062306a36Sopenharmony_ci#else 2162306a36Sopenharmony_ci# define L3_LINE_SIZE 128 // McKinley L3 line size 2262306a36Sopenharmony_ci# define PREFETCH_LINES 12 // magic number 2362306a36Sopenharmony_ci#endif 2462306a36Sopenharmony_ci 2562306a36Sopenharmony_ci#define saved_lc r2 2662306a36Sopenharmony_ci#define dst_fetch r3 2762306a36Sopenharmony_ci#define dst1 r8 2862306a36Sopenharmony_ci#define dst2 r9 2962306a36Sopenharmony_ci#define dst3 r10 3062306a36Sopenharmony_ci#define dst4 r11 3162306a36Sopenharmony_ci 3262306a36Sopenharmony_ci#define dst_last r31 3362306a36Sopenharmony_ci 3462306a36Sopenharmony_ciGLOBAL_ENTRY(clear_page) 3562306a36Sopenharmony_ci .prologue 3662306a36Sopenharmony_ci .regstk 1,0,0,0 3762306a36Sopenharmony_ci mov r16 = PAGE_SIZE/L3_LINE_SIZE-1 // main loop count, -1=repeat/until 3862306a36Sopenharmony_ci .save ar.lc, saved_lc 3962306a36Sopenharmony_ci mov saved_lc = ar.lc 4062306a36Sopenharmony_ci 4162306a36Sopenharmony_ci .body 4262306a36Sopenharmony_ci mov ar.lc = (PREFETCH_LINES - 1) 4362306a36Sopenharmony_ci mov dst_fetch = in0 4462306a36Sopenharmony_ci adds dst1 = 16, in0 4562306a36Sopenharmony_ci adds dst2 = 32, in0 4662306a36Sopenharmony_ci ;; 4762306a36Sopenharmony_ci.fetch: stf.spill.nta [dst_fetch] = f0, L3_LINE_SIZE 4862306a36Sopenharmony_ci adds dst3 = 48, in0 // executing this multiple times is harmless 4962306a36Sopenharmony_ci br.cloop.sptk.few .fetch 5062306a36Sopenharmony_ci ;; 5162306a36Sopenharmony_ci addl dst_last = (PAGE_SIZE - PREFETCH_LINES*L3_LINE_SIZE), dst_fetch 5262306a36Sopenharmony_ci mov ar.lc = r16 // one L3 line per iteration 5362306a36Sopenharmony_ci adds dst4 = 64, in0 5462306a36Sopenharmony_ci ;; 5562306a36Sopenharmony_ci#ifdef CONFIG_ITANIUM 5662306a36Sopenharmony_ci // Optimized for Itanium 5762306a36Sopenharmony_ci1: stf.spill.nta [dst1] = f0, 64 5862306a36Sopenharmony_ci stf.spill.nta [dst2] = f0, 64 5962306a36Sopenharmony_ci cmp.lt p8,p0=dst_fetch, dst_last 6062306a36Sopenharmony_ci ;; 6162306a36Sopenharmony_ci#else 6262306a36Sopenharmony_ci // Optimized for McKinley 6362306a36Sopenharmony_ci1: stf.spill.nta [dst1] = f0, 64 6462306a36Sopenharmony_ci stf.spill.nta [dst2] = f0, 64 6562306a36Sopenharmony_ci stf.spill.nta [dst3] = f0, 64 6662306a36Sopenharmony_ci stf.spill.nta [dst4] = f0, 128 6762306a36Sopenharmony_ci cmp.lt p8,p0=dst_fetch, dst_last 6862306a36Sopenharmony_ci ;; 6962306a36Sopenharmony_ci stf.spill.nta [dst1] = f0, 64 7062306a36Sopenharmony_ci stf.spill.nta [dst2] = f0, 64 7162306a36Sopenharmony_ci#endif 7262306a36Sopenharmony_ci stf.spill.nta [dst3] = f0, 64 7362306a36Sopenharmony_ci(p8) stf.spill.nta [dst_fetch] = f0, L3_LINE_SIZE 7462306a36Sopenharmony_ci br.cloop.sptk.few 1b 7562306a36Sopenharmony_ci ;; 7662306a36Sopenharmony_ci mov ar.lc = saved_lc // restore lc 7762306a36Sopenharmony_ci br.ret.sptk.many rp 7862306a36Sopenharmony_ciEND(clear_page) 7962306a36Sopenharmony_ciEXPORT_SYMBOL(clear_page) 80