linux/arch/powerpc/kernel/fpu.S

153 lines
3.8 KiB
ArmAsm
Raw Normal View History

/*
* FPU support code, moved here from head.S so that it can be used
* by chips which use other head-whatever.S files.
*
* Copyright (C) 1995-1996 Gary Thomas (gdt@linuxppc.org)
* Copyright (C) 1996 Cort Dougan <cort@cs.nmt.edu>
* Copyright (C) 1996 Paul Mackerras.
* Copyright (C) 1997 Dan Malek (dmalek@jlc.net).
*
* This program is free software; you can redistribute it and/or
* modify it under the terms of the GNU General Public License
* as published by the Free Software Foundation; either version
* 2 of the License, or (at your option) any later version.
*
*/
#include <asm/reg.h>
#include <asm/page.h>
#include <asm/mmu.h>
#include <asm/pgtable.h>
#include <asm/cputable.h>
#include <asm/cache.h>
#include <asm/thread_info.h>
#include <asm/ppc_asm.h>
#include <asm/asm-offsets.h>
#include <asm/ptrace.h>
#include <asm/export.h>
#ifdef CONFIG_VSX
#define __REST_32FPVSRS(n,c,base) \
BEGIN_FTR_SECTION \
b 2f; \
END_FTR_SECTION_IFSET(CPU_FTR_VSX); \
REST_32FPRS(n,base); \
b 3f; \
2: REST_32VSRS(n,c,base); \
3:
#define __SAVE_32FPVSRS(n,c,base) \
BEGIN_FTR_SECTION \
b 2f; \
END_FTR_SECTION_IFSET(CPU_FTR_VSX); \
SAVE_32FPRS(n,base); \
b 3f; \
2: SAVE_32VSRS(n,c,base); \
3:
#else
#define __REST_32FPVSRS(n,b,base) REST_32FPRS(n, base)
#define __SAVE_32FPVSRS(n,b,base) SAVE_32FPRS(n, base)
#endif
#define REST_32FPVSRS(n,c,base) __REST_32FPVSRS(n,__REG_##c,__REG_##base)
#define SAVE_32FPVSRS(n,c,base) __SAVE_32FPVSRS(n,__REG_##c,__REG_##base)
/*
* Load state from memory into FP registers including FPSCR.
* Assumes the caller has enabled FP in the MSR.
*/
_GLOBAL(load_fp_state)
lfd fr0,FPSTATE_FPSCR(r3)
MTFSF_L(fr0)
REST_32FPVSRS(0, R4, R3)
blr
EXPORT_SYMBOL(load_fp_state)
/*
* Store FP state into memory, including FPSCR
* Assumes the caller has enabled FP in the MSR.
*/
_GLOBAL(store_fp_state)
SAVE_32FPVSRS(0, R4, R3)
mffs fr0
stfd fr0,FPSTATE_FPSCR(r3)
blr
EXPORT_SYMBOL(store_fp_state)
/*
* This task wants to use the FPU now.
* On UP, disable FP for the task which had the FPU previously,
* and save its floating-point registers in its thread_struct.
* Load up this task's FP registers from its thread_struct,
* enable the FPU for the current task and return to the task.
* Note that on 32-bit this can only use registers that will be
* restored by fast_exception_return, i.e. r3 - r6, r10 and r11.
*/
_GLOBAL(load_up_fpu)
mfmsr r5
ori r5,r5,MSR_FP
#ifdef CONFIG_VSX
BEGIN_FTR_SECTION
oris r5,r5,MSR_VSX@h
END_FTR_SECTION_IFSET(CPU_FTR_VSX)
#endif
SYNC
MTMSRD(r5) /* enable use of fpu now */
isync
/* enable use of FP after return */
#ifdef CONFIG_PPC32
mfspr r5,SPRN_SPRG_THREAD /* current task's THREAD (phys) */
lwz r4,THREAD_FPEXC_MODE(r5)
ori r9,r9,MSR_FP /* enable FP for current */
or r9,r9,r4
#else
ld r4,PACACURRENT(r13)
addi r5,r4,THREAD /* Get THREAD */
lwz r4,THREAD_FPEXC_MODE(r5)
ori r12,r12,MSR_FP
or r12,r12,r4
std r12,_MSR(r1)
#endif
powerpc: Restore FPU/VEC/VSX if previously used Currently the FPU, VEC and VSX facilities are lazily loaded. This is not a problem unless a process is using these facilities. Modern versions of GCC are very good at automatically vectorising code, new and modernised workloads make use of floating point and vector facilities, even the kernel makes use of vectorised memcpy. All this combined greatly increases the cost of a syscall since the kernel uses the facilities sometimes even in syscall fast-path making it increasingly common for a thread to take an *_unavailable exception soon after a syscall, not to mention potentially taking all three. The obvious overcompensation to this problem is to simply always load all the facilities on every exit to userspace. Loading up all FPU, VEC and VSX registers every time can be expensive and if a workload does avoid using them, it should not be forced to incur this penalty. An 8bit counter is used to detect if the registers have been used in the past and the registers are always loaded until the value wraps to back to zero. Several versions of the assembly in entry_64.S were tested: 1. Always calling C. 2. Performing a common case check and then calling C. 3. A complex check in asm. After some benchmarking it was determined that avoiding C in the common case is a performance benefit (option 2). The full check in asm (option 3) greatly complicated that codepath for a negligible performance gain and the trade-off was deemed not worth it. Signed-off-by: Cyril Bur <cyrilbur@gmail.com> [mpe: Move load_vec in the struct to fill an existing hole, reword change log] Signed-off-by: Michael Ellerman <mpe@ellerman.id.au> fixup
2016-02-29 06:53:47 +00:00
/* Don't care if r4 overflows, this is desired behaviour */
lbz r4,THREAD_LOAD_FP(r5)
addi r4,r4,1
stb r4,THREAD_LOAD_FP(r5)
addi r10,r5,THREAD_FPSTATE
lfd fr0,FPSTATE_FPSCR(r10)
MTFSF_L(fr0)
REST_32FPVSRS(0, R4, R10)
/* restore registers and return */
/* we haven't used ctr or xer or lr */
blr
/*
* save_fpu(tsk)
* Save the floating-point registers in its thread_struct.
* Enables the FPU for use in the kernel on return.
*/
_GLOBAL(save_fpu)
addi r3,r3,THREAD /* want THREAD of task */
PPC_LL r6,THREAD_FPSAVEAREA(r3)
PPC_LL r5,PT_REGS(r3)
PPC_LCMPI 0,r6,0
bne 2f
addi r6,r3,THREAD_FPSTATE
2: SAVE_32FPVSRS(0, R4, R6)
mffs fr0
stfd fr0,FPSTATE_FPSCR(r6)
blr
[PATCH] powerpc: Fix handling of fpscr on 64-bit The recent merge of fpu.S broken the handling of fpscr for ARCH=powerpc and CONFIG_PPC64=y. FP registers could be corrupted, leading to strange random application crashes. The confusion arises, because the thread_struct has (and requires) a 64-bit area to save the fpscr, because we use load/store double instructions to get it in to/out of the FPU. However, only the low 32-bits are actually used, so we want to treat it as a 32-bit quantity when manipulating its bits to avoid extra load/stores on 32-bit. This patch replaces the current definition with a structure of two 32-bit quantities (pad and val), to clarify things as much as is possible. The 'val' field is used when manipulating bits, the structure itself is used when obtaining the address for loading/unloading the value from the FPU. While we're at it, consolidate the 4 (!) almost identical versions of cvt_fd() and cvt_df() (arch/ppc/kernel/misc.S, arch/ppc64/kernel/misc.S, arch/powerpc/kernel/misc_32.S, arch/powerpc/kernel/misc_64.S) into a single version in fpu.S. The new version takes a pointer to thread_struct and applies the correct offset itself, rather than a pointer to the fpscr field itself, again to avoid confusion as to which is the correct field to use. Finally, this patch makes ARCH=ppc64 also use the consolidated fpu.S code, which it previously did not. Built for G5 (ARCH=ppc64 and ARCH=powerpc), 32-bit powermac (ARCH=ppc and ARCH=powerpc) and Walnut (ARCH=ppc, CONFIG_MATH_EMULATION=y). Booted on G5 (ARCH=powerpc) and things which previously fell over no longer do. Signed-off-by: David Gibson <dwg@au1.ibm.com> Signed-off-by: Paul Mackerras <paulus@samba.org>
2005-10-27 06:27:25 +00:00
/*
* These are used in the alignment trap handler when emulating
* single-precision loads and stores.
*/
_GLOBAL(cvt_fd)
lfs 0,0(r3)
stfd 0,0(r4)
blr
_GLOBAL(cvt_df)
lfd 0,0(r3)
stfs 0,0(r4)
blr