1 #include <asm/processor.h>
2 #include <asm/ppc_asm.h>
4 #include <asm/asm-offsets.h>
5 #include <asm/cputable.h>
6 #include <asm/thread_info.h>
8 #include <asm/ptrace.h>
11 * load_up_altivec(unused, unused, tsk)
12 * Disable VMX for the task which had it previously,
13 * and save its vector registers in its thread_struct.
14 * Enables the VMX for use in the kernel on return.
15 * On SMP we know the VMX is free, since we give it up every
16 * switch (ie, no lazy save of the vector registers).
18 _GLOBAL(load_up_altivec)
19 mfmsr r5 /* grab the current MSR */
21 MTMSRD(r5) /* enable use of AltiVec now */
25 * For SMP, we don't do lazy VMX switching because it just gets too
26 * horrendously complex, especially when a task switches from one CPU
27 * to another. Instead we call giveup_altvec in switch_to.
28 * VRSAVE isn't dealt with here, that is done in the normal context
29 * switch code. Note that we could rely on vrsave value to eventually
30 * avoid saving all of the VREGs here...
33 LOAD_REG_ADDRBASE(r3, last_task_used_altivec)
35 PPC_LL r4,ADDROFF(last_task_used_altivec)(r3)
39 /* Save VMX state to last_task_used_altivec's THREAD struct */
46 /* Disable VMX for last_task_used_altivec */
49 PPC_LL r4,_MSR-STACK_FRAME_OVERHEAD(r5)
52 PPC_STL r4,_MSR-STACK_FRAME_OVERHEAD(r5)
54 #endif /* CONFIG_SMP */
56 /* Hack: if we get an altivec unavailable trap with VRSAVE
57 * set to all zeros, we assume this is a broken application
58 * that fails to set it properly, and thus we switch it to
67 /* enable use of VMX after return */
69 mfspr r5,SPRN_SPRG_THREAD /* current task's THREAD (phys) */
72 ld r4,PACACURRENT(r13)
73 addi r5,r4,THREAD /* Get THREAD */
74 oris r12,r12,MSR_VEC@h
79 stw r4,THREAD_USED_VR(r5)
84 /* Update last_task_used_altivec to 'current' */
85 subi r4,r5,THREAD /* Back to 'current' */
87 PPC_STL r4,ADDROFF(last_task_used_altivec)(r3)
88 #endif /* CONFIG_SMP */
89 /* restore registers and return */
94 * Disable VMX for the task given as the argument,
95 * and save the vector registers in its thread_struct.
96 * Enables the VMX for use in the kernel on return.
98 _GLOBAL(giveup_altivec)
102 MTMSRD(r5) /* enable use of VMX now */
105 beqlr- /* if no previous owner, done */
106 addi r3,r3,THREAD /* want THREAD of task */
107 PPC_LL r5,PT_REGS(r3)
114 PPC_LL r4,_MSR-STACK_FRAME_OVERHEAD(r5)
117 lis r3,(MSR_VEC|MSR_VSX)@h
120 ALT_FTR_SECTION_END_IFSET(CPU_FTR_VSX)
124 andc r4,r4,r3 /* disable FP for previous task */
125 PPC_STL r4,_MSR-STACK_FRAME_OVERHEAD(r5)
129 LOAD_REG_ADDRBASE(r4,last_task_used_altivec)
130 PPC_STL r5,ADDROFF(last_task_used_altivec)(r4)
131 #endif /* CONFIG_SMP */
137 #error This asm code isn't ready for 32-bit kernels
141 * load_up_vsx(unused, unused, tsk)
142 * Disable VSX for the task which had it previously,
143 * and save its vector registers in its thread_struct.
144 * Reuse the fp and vsx saves, but first check to see if they have
145 * been saved already.
148 /* Load FP and VSX registers if they haven't been done yet */
150 beql+ load_up_fpu /* skip if already loaded */
151 andis. r5,r12,MSR_VEC@h
152 beql+ load_up_altivec /* skip if already loaded */
155 ld r3,last_task_used_vsx@got(r2)
159 /* Disable VSX for last_task_used_vsx */
162 ld r4,_MSR-STACK_FRAME_OVERHEAD(r5)
165 std r6,_MSR-STACK_FRAME_OVERHEAD(r5)
167 #endif /* CONFIG_SMP */
168 ld r4,PACACURRENT(r13)
169 addi r4,r4,THREAD /* Get THREAD */
171 stw r6,THREAD_USED_VSR(r4) /* ... also set thread used vsr */
172 /* enable use of VSX after return */
173 oris r12,r12,MSR_VSX@h
176 /* Update last_task_used_vsx to 'current' */
177 ld r4,PACACURRENT(r13)
179 #endif /* CONFIG_SMP */
180 b fast_exception_return
184 * Disable VSX for the task given as the argument.
185 * Does NOT save vsx registers.
186 * Enables the VSX for use in the kernel on return.
188 _GLOBAL(__giveup_vsx)
191 mtmsrd r5 /* enable use of VSX now */
195 beqlr- /* if no previous owner, done */
196 addi r3,r3,THREAD /* want THREAD of task */
200 ld r4,_MSR-STACK_FRAME_OVERHEAD(r5)
202 andc r4,r4,r3 /* disable VSX for previous task */
203 std r4,_MSR-STACK_FRAME_OVERHEAD(r5)
207 ld r4,last_task_used_vsx@got(r2)
209 #endif /* CONFIG_SMP */
212 #endif /* CONFIG_VSX */
216 * The routines below are in assembler so we can closely control the
217 * usage of floating-point registers. These routines must be called
218 * with preempt disabled.
225 .long 0x3f800000 /* 1.0 in single-precision FP */
227 .long 0x3f000000 /* 0.5 in single-precision FP */
229 #define LDCONST(fr, name) \
238 .tc FD_3ff00000_0[TC],0x3ff0000000000000 /* 1.0 */
240 .tc FD_3fe00000_0[TC],0x3fe0000000000000 /* 0.5 */
242 #define LDCONST(fr, name) \
248 * Internal routine to enable floating point and set FPSCR to 0.
249 * Don't call it from C; it doesn't use the normal calling convention.
281 * Vector add, floating point.
298 * Vector subtract, floating point.
315 * Vector multiply and add, floating point.
327 fmadds fr0,fr0,fr2,fr1
335 * Vector negative multiply and subtract, floating point.
347 fnmsubs fr0,fr0,fr2,fr1
355 * Vector reciprocal estimate. We just compute 1.0/x.
356 * r3 -> destination, r4 -> source.
373 * Vector reciprocal square-root estimate, floating point.
374 * We use the frsqrte instruction for the initial estimate followed
375 * by 2 iterations of Newton-Raphson to get sufficient accuracy.
376 * r3 -> destination, r4 -> source.
391 frsqrte fr1,fr0 /* r = frsqrte(s) */
392 fmuls fr3,fr1,fr0 /* r * s */
393 fmuls fr2,fr1,fr5 /* r * 0.5 */
394 fnmsubs fr3,fr1,fr3,fr4 /* 1 - s * r * r */
395 fmadds fr1,fr2,fr3,fr1 /* r = r + 0.5 * r * (1 - s * r * r) */
396 fmuls fr3,fr1,fr0 /* r * s */
397 fmuls fr2,fr1,fr5 /* r * 0.5 */
398 fnmsubs fr3,fr1,fr3,fr4 /* 1 - s * r * r */
399 fmadds fr1,fr2,fr3,fr1 /* r = r + 0.5 * r * (1 - s * r * r) */