xref: /openbmc/linux/arch/arc/kernel/fpu.c (revision bc5aa3a0)
1 /*
2  * fpu.c - save/restore of Floating Point Unit Registers on task switch
3  *
4  * Copyright (C) 2004, 2007-2010, 2011-2012 Synopsys, Inc. (www.synopsys.com)
5  *
6  * This program is free software; you can redistribute it and/or modify
7  * it under the terms of the GNU General Public License version 2 as
8  * published by the Free Software Foundation.
9  */
10 
11 #include <linux/sched.h>
12 #include <asm/switch_to.h>
13 
14 /*
15  * To save/restore FPU regs, simplest scheme would use LR/SR insns.
16  * However since SR serializes the pipeline, an alternate "hack" can be used
17  * which uses the FPU Exchange insn (DEXCL) to r/w FPU regs.
18  *
19  * Store to 64bit dpfp1 reg from a pair of core regs:
20  *   dexcl1 0, r1, r0  ; where r1:r0 is the 64 bit val
21  *
22  * Read from dpfp1 into pair of core regs (w/o clobbering dpfp1)
23  *   mov_s    r3, 0
24  *   daddh11  r1, r3, r3   ; get "hi" into r1 (dpfp1 unchanged)
25  *   dexcl1   r0, r1, r3   ; get "low" into r0 (dpfp1 low clobbered)
26  *   dexcl1    0, r1, r0   ; restore dpfp1 to orig value
27  *
28  * However we can tweak the read, so that read-out of outgoing task's FPU regs
29  * and write of incoming task's regs happen in one shot. So all the work is
30  * done before context switch
31  */
32 
33 void fpu_save_restore(struct task_struct *prev, struct task_struct *next)
34 {
35 	unsigned int *saveto = &prev->thread.fpu.aux_dpfp[0].l;
36 	unsigned int *readfrom = &next->thread.fpu.aux_dpfp[0].l;
37 
38 	const unsigned int zero = 0;
39 
40 	__asm__ __volatile__(
41 		"daddh11  %0, %2, %2\n"
42 		"dexcl1   %1, %3, %4\n"
43 		: "=&r" (*(saveto + 1)), /* early clobber must here */
44 		  "=&r" (*(saveto))
45 		: "r" (zero), "r" (*(readfrom + 1)), "r" (*(readfrom))
46 	);
47 
48 	__asm__ __volatile__(
49 		"daddh22  %0, %2, %2\n"
50 		"dexcl2   %1, %3, %4\n"
51 		: "=&r"(*(saveto + 3)),	/* early clobber must here */
52 		  "=&r"(*(saveto + 2))
53 		: "r" (zero), "r" (*(readfrom + 3)), "r" (*(readfrom + 2))
54 	);
55 }
56