18c2ecf20Sopenharmony_ci/* SPDX-License-Identifier: GPL-2.0 */ 28c2ecf20Sopenharmony_ci/* 38c2ecf20Sopenharmony_ci * Copyright (C) 1999-2002 Hewlett-Packard Co 48c2ecf20Sopenharmony_ci * Stephane Eranian <eranian@hpl.hp.com> 58c2ecf20Sopenharmony_ci * David Mosberger-Tang <davidm@hpl.hp.com> 68c2ecf20Sopenharmony_ci * Copyright (C) 2002 Ken Chen <kenneth.w.chen@intel.com> 78c2ecf20Sopenharmony_ci * 88c2ecf20Sopenharmony_ci * 1/06/01 davidm Tuned for Itanium. 98c2ecf20Sopenharmony_ci * 2/12/02 kchen Tuned for both Itanium and McKinley 108c2ecf20Sopenharmony_ci * 3/08/02 davidm Some more tweaking 118c2ecf20Sopenharmony_ci */ 128c2ecf20Sopenharmony_ci 138c2ecf20Sopenharmony_ci#include <asm/asmmacro.h> 148c2ecf20Sopenharmony_ci#include <asm/page.h> 158c2ecf20Sopenharmony_ci#include <asm/export.h> 168c2ecf20Sopenharmony_ci 178c2ecf20Sopenharmony_ci#ifdef CONFIG_ITANIUM 188c2ecf20Sopenharmony_ci# define L3_LINE_SIZE 64 // Itanium L3 line size 198c2ecf20Sopenharmony_ci# define PREFETCH_LINES 9 // magic number 208c2ecf20Sopenharmony_ci#else 218c2ecf20Sopenharmony_ci# define L3_LINE_SIZE 128 // McKinley L3 line size 228c2ecf20Sopenharmony_ci# define PREFETCH_LINES 12 // magic number 238c2ecf20Sopenharmony_ci#endif 248c2ecf20Sopenharmony_ci 258c2ecf20Sopenharmony_ci#define saved_lc r2 268c2ecf20Sopenharmony_ci#define dst_fetch r3 278c2ecf20Sopenharmony_ci#define dst1 r8 288c2ecf20Sopenharmony_ci#define dst2 r9 298c2ecf20Sopenharmony_ci#define dst3 r10 308c2ecf20Sopenharmony_ci#define dst4 r11 318c2ecf20Sopenharmony_ci 328c2ecf20Sopenharmony_ci#define dst_last r31 338c2ecf20Sopenharmony_ci 348c2ecf20Sopenharmony_ciGLOBAL_ENTRY(clear_page) 358c2ecf20Sopenharmony_ci .prologue 368c2ecf20Sopenharmony_ci .regstk 1,0,0,0 378c2ecf20Sopenharmony_ci mov r16 = PAGE_SIZE/L3_LINE_SIZE-1 // main loop count, -1=repeat/until 388c2ecf20Sopenharmony_ci .save ar.lc, saved_lc 398c2ecf20Sopenharmony_ci mov saved_lc = ar.lc 408c2ecf20Sopenharmony_ci 418c2ecf20Sopenharmony_ci .body 428c2ecf20Sopenharmony_ci mov ar.lc = (PREFETCH_LINES - 1) 438c2ecf20Sopenharmony_ci mov dst_fetch = in0 448c2ecf20Sopenharmony_ci adds dst1 = 16, in0 458c2ecf20Sopenharmony_ci adds dst2 = 32, in0 468c2ecf20Sopenharmony_ci ;; 478c2ecf20Sopenharmony_ci.fetch: stf.spill.nta [dst_fetch] = f0, L3_LINE_SIZE 488c2ecf20Sopenharmony_ci adds dst3 = 48, in0 // executing this multiple times is harmless 498c2ecf20Sopenharmony_ci br.cloop.sptk.few .fetch 508c2ecf20Sopenharmony_ci ;; 518c2ecf20Sopenharmony_ci addl dst_last = (PAGE_SIZE - PREFETCH_LINES*L3_LINE_SIZE), dst_fetch 528c2ecf20Sopenharmony_ci mov ar.lc = r16 // one L3 line per iteration 538c2ecf20Sopenharmony_ci adds dst4 = 64, in0 548c2ecf20Sopenharmony_ci ;; 558c2ecf20Sopenharmony_ci#ifdef CONFIG_ITANIUM 568c2ecf20Sopenharmony_ci // Optimized for Itanium 578c2ecf20Sopenharmony_ci1: stf.spill.nta [dst1] = f0, 64 588c2ecf20Sopenharmony_ci stf.spill.nta [dst2] = f0, 64 598c2ecf20Sopenharmony_ci cmp.lt p8,p0=dst_fetch, dst_last 608c2ecf20Sopenharmony_ci ;; 618c2ecf20Sopenharmony_ci#else 628c2ecf20Sopenharmony_ci // Optimized for McKinley 638c2ecf20Sopenharmony_ci1: stf.spill.nta [dst1] = f0, 64 648c2ecf20Sopenharmony_ci stf.spill.nta [dst2] = f0, 64 658c2ecf20Sopenharmony_ci stf.spill.nta [dst3] = f0, 64 668c2ecf20Sopenharmony_ci stf.spill.nta [dst4] = f0, 128 678c2ecf20Sopenharmony_ci cmp.lt p8,p0=dst_fetch, dst_last 688c2ecf20Sopenharmony_ci ;; 698c2ecf20Sopenharmony_ci stf.spill.nta [dst1] = f0, 64 708c2ecf20Sopenharmony_ci stf.spill.nta [dst2] = f0, 64 718c2ecf20Sopenharmony_ci#endif 728c2ecf20Sopenharmony_ci stf.spill.nta [dst3] = f0, 64 738c2ecf20Sopenharmony_ci(p8) stf.spill.nta [dst_fetch] = f0, L3_LINE_SIZE 748c2ecf20Sopenharmony_ci br.cloop.sptk.few 1b 758c2ecf20Sopenharmony_ci ;; 768c2ecf20Sopenharmony_ci mov ar.lc = saved_lc // restore lc 778c2ecf20Sopenharmony_ci br.ret.sptk.many rp 788c2ecf20Sopenharmony_ciEND(clear_page) 798c2ecf20Sopenharmony_ciEXPORT_SYMBOL(clear_page) 80