linux_dsm_epyc7002/arch/powerpc/kernel/vector.S
Anton Blanchard 35000870fc powerpc: Optimise enable_kernel_altivec
Add two optimisations to enable_kernel_altivec:

- enable_kernel_altivec has already determined if we need to
save the previous task's state but we call giveup_altivec
in both cases, requiring an extra branch in giveup_altivec. Create
giveup_altivec_notask which only turns on the VMX bit in the
MSR.

- We write the VMX MSR bit each time we call enable_kernel_altivec
even it was already set. Check the bit and branch out if we have
already set it. The classic case for this is vectored IO
where we have to copy multiple buffers to or from userspace.

The following testcase was used to confirm this patch improves
performance:

http://ozlabs.org/~anton/junkcode/copy_to_user.c

Since the current breakpoint for using VMX in copy_tofrom_user is
4096 bytes, I'm using buffers of 4096 + 1 cacheline (4224) bytes.
A benchmark of 16 entry readvs (-s 16):

time copy_to_user -l 4224 -s 16 -i 1000000

completes 5.2% faster on a POWER7 PS700.

Signed-off-by: Anton Blanchard <anton@samba.org>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
2012-04-30 15:37:17 +10:00

418 lines
8.7 KiB
ArmAsm

#include <asm/processor.h>
#include <asm/ppc_asm.h>
#include <asm/reg.h>
#include <asm/asm-offsets.h>
#include <asm/cputable.h>
#include <asm/thread_info.h>
#include <asm/page.h>
#include <asm/ptrace.h>
/*
* load_up_altivec(unused, unused, tsk)
* Disable VMX for the task which had it previously,
* and save its vector registers in its thread_struct.
* Enables the VMX for use in the kernel on return.
* On SMP we know the VMX is free, since we give it up every
* switch (ie, no lazy save of the vector registers).
*/
_GLOBAL(load_up_altivec)
mfmsr r5 /* grab the current MSR */
oris r5,r5,MSR_VEC@h
MTMSRD(r5) /* enable use of AltiVec now */
isync
/*
* For SMP, we don't do lazy VMX switching because it just gets too
* horrendously complex, especially when a task switches from one CPU
* to another. Instead we call giveup_altvec in switch_to.
* VRSAVE isn't dealt with here, that is done in the normal context
* switch code. Note that we could rely on vrsave value to eventually
* avoid saving all of the VREGs here...
*/
#ifndef CONFIG_SMP
LOAD_REG_ADDRBASE(r3, last_task_used_altivec)
toreal(r3)
PPC_LL r4,ADDROFF(last_task_used_altivec)(r3)
PPC_LCMPI 0,r4,0
beq 1f
/* Save VMX state to last_task_used_altivec's THREAD struct */
toreal(r4)
addi r4,r4,THREAD
SAVE_32VRS(0,r5,r4)
mfvscr vr0
li r10,THREAD_VSCR
stvx vr0,r10,r4
/* Disable VMX for last_task_used_altivec */
PPC_LL r5,PT_REGS(r4)
toreal(r5)
PPC_LL r4,_MSR-STACK_FRAME_OVERHEAD(r5)
lis r10,MSR_VEC@h
andc r4,r4,r10
PPC_STL r4,_MSR-STACK_FRAME_OVERHEAD(r5)
1:
#endif /* CONFIG_SMP */
/* Hack: if we get an altivec unavailable trap with VRSAVE
* set to all zeros, we assume this is a broken application
* that fails to set it properly, and thus we switch it to
* all 1's
*/
mfspr r4,SPRN_VRSAVE
cmpwi 0,r4,0
bne+ 1f
li r4,-1
mtspr SPRN_VRSAVE,r4
1:
/* enable use of VMX after return */
#ifdef CONFIG_PPC32
mfspr r5,SPRN_SPRG_THREAD /* current task's THREAD (phys) */
oris r9,r9,MSR_VEC@h
#else
ld r4,PACACURRENT(r13)
addi r5,r4,THREAD /* Get THREAD */
oris r12,r12,MSR_VEC@h
std r12,_MSR(r1)
#endif
li r4,1
li r10,THREAD_VSCR
stw r4,THREAD_USED_VR(r5)
lvx vr0,r10,r5
mtvscr vr0
REST_32VRS(0,r4,r5)
#ifndef CONFIG_SMP
/* Update last_task_used_altivec to 'current' */
subi r4,r5,THREAD /* Back to 'current' */
fromreal(r4)
PPC_STL r4,ADDROFF(last_task_used_altivec)(r3)
#endif /* CONFIG_SMP */
/* restore registers and return */
blr
_GLOBAL(giveup_altivec_notask)
mfmsr r3
andis. r4,r3,MSR_VEC@h
bnelr /* Already enabled? */
oris r3,r3,MSR_VEC@h
SYNC
MTMSRD(r3) /* enable use of VMX now */
isync
blr
/*
* giveup_altivec(tsk)
* Disable VMX for the task given as the argument,
* and save the vector registers in its thread_struct.
* Enables the VMX for use in the kernel on return.
*/
_GLOBAL(giveup_altivec)
mfmsr r5
oris r5,r5,MSR_VEC@h
SYNC
MTMSRD(r5) /* enable use of VMX now */
isync
PPC_LCMPI 0,r3,0
beqlr /* if no previous owner, done */
addi r3,r3,THREAD /* want THREAD of task */
PPC_LL r5,PT_REGS(r3)
PPC_LCMPI 0,r5,0
SAVE_32VRS(0,r4,r3)
mfvscr vr0
li r4,THREAD_VSCR
stvx vr0,r4,r3
beq 1f
PPC_LL r4,_MSR-STACK_FRAME_OVERHEAD(r5)
#ifdef CONFIG_VSX
BEGIN_FTR_SECTION
lis r3,(MSR_VEC|MSR_VSX)@h
FTR_SECTION_ELSE
lis r3,MSR_VEC@h
ALT_FTR_SECTION_END_IFSET(CPU_FTR_VSX)
#else
lis r3,MSR_VEC@h
#endif
andc r4,r4,r3 /* disable FP for previous task */
PPC_STL r4,_MSR-STACK_FRAME_OVERHEAD(r5)
1:
#ifndef CONFIG_SMP
li r5,0
LOAD_REG_ADDRBASE(r4,last_task_used_altivec)
PPC_STL r5,ADDROFF(last_task_used_altivec)(r4)
#endif /* CONFIG_SMP */
blr
#ifdef CONFIG_VSX
#ifdef CONFIG_PPC32
#error This asm code isn't ready for 32-bit kernels
#endif
/*
* load_up_vsx(unused, unused, tsk)
* Disable VSX for the task which had it previously,
* and save its vector registers in its thread_struct.
* Reuse the fp and vsx saves, but first check to see if they have
* been saved already.
*/
_GLOBAL(load_up_vsx)
/* Load FP and VSX registers if they haven't been done yet */
andi. r5,r12,MSR_FP
beql+ load_up_fpu /* skip if already loaded */
andis. r5,r12,MSR_VEC@h
beql+ load_up_altivec /* skip if already loaded */
#ifndef CONFIG_SMP
ld r3,last_task_used_vsx@got(r2)
ld r4,0(r3)
cmpdi 0,r4,0
beq 1f
/* Disable VSX for last_task_used_vsx */
addi r4,r4,THREAD
ld r5,PT_REGS(r4)
ld r4,_MSR-STACK_FRAME_OVERHEAD(r5)
lis r6,MSR_VSX@h
andc r6,r4,r6
std r6,_MSR-STACK_FRAME_OVERHEAD(r5)
1:
#endif /* CONFIG_SMP */
ld r4,PACACURRENT(r13)
addi r4,r4,THREAD /* Get THREAD */
li r6,1
stw r6,THREAD_USED_VSR(r4) /* ... also set thread used vsr */
/* enable use of VSX after return */
oris r12,r12,MSR_VSX@h
std r12,_MSR(r1)
#ifndef CONFIG_SMP
/* Update last_task_used_vsx to 'current' */
ld r4,PACACURRENT(r13)
std r4,0(r3)
#endif /* CONFIG_SMP */
b fast_exception_return
/*
* __giveup_vsx(tsk)
* Disable VSX for the task given as the argument.
* Does NOT save vsx registers.
* Enables the VSX for use in the kernel on return.
*/
_GLOBAL(__giveup_vsx)
mfmsr r5
oris r5,r5,MSR_VSX@h
mtmsrd r5 /* enable use of VSX now */
isync
cmpdi 0,r3,0
beqlr- /* if no previous owner, done */
addi r3,r3,THREAD /* want THREAD of task */
ld r5,PT_REGS(r3)
cmpdi 0,r5,0
beq 1f
ld r4,_MSR-STACK_FRAME_OVERHEAD(r5)
lis r3,MSR_VSX@h
andc r4,r4,r3 /* disable VSX for previous task */
std r4,_MSR-STACK_FRAME_OVERHEAD(r5)
1:
#ifndef CONFIG_SMP
li r5,0
ld r4,last_task_used_vsx@got(r2)
std r5,0(r4)
#endif /* CONFIG_SMP */
blr
#endif /* CONFIG_VSX */
/*
* The routines below are in assembler so we can closely control the
* usage of floating-point registers. These routines must be called
* with preempt disabled.
*/
#ifdef CONFIG_PPC32
.data
fpzero:
.long 0
fpone:
.long 0x3f800000 /* 1.0 in single-precision FP */
fphalf:
.long 0x3f000000 /* 0.5 in single-precision FP */
#define LDCONST(fr, name) \
lis r11,name@ha; \
lfs fr,name@l(r11)
#else
.section ".toc","aw"
fpzero:
.tc FD_0_0[TC],0
fpone:
.tc FD_3ff00000_0[TC],0x3ff0000000000000 /* 1.0 */
fphalf:
.tc FD_3fe00000_0[TC],0x3fe0000000000000 /* 0.5 */
#define LDCONST(fr, name) \
lfd fr,name@toc(r2)
#endif
.text
/*
* Internal routine to enable floating point and set FPSCR to 0.
* Don't call it from C; it doesn't use the normal calling convention.
*/
fpenable:
#ifdef CONFIG_PPC32
stwu r1,-64(r1)
#else
stdu r1,-64(r1)
#endif
mfmsr r10
ori r11,r10,MSR_FP
mtmsr r11
isync
stfd fr0,24(r1)
stfd fr1,16(r1)
stfd fr31,8(r1)
LDCONST(fr1, fpzero)
mffs fr31
MTFSF_L(fr1)
blr
fpdisable:
mtlr r12
MTFSF_L(fr31)
lfd fr31,8(r1)
lfd fr1,16(r1)
lfd fr0,24(r1)
mtmsr r10
isync
addi r1,r1,64
blr
/*
* Vector add, floating point.
*/
_GLOBAL(vaddfp)
mflr r12
bl fpenable
li r0,4
mtctr r0
li r6,0
1: lfsx fr0,r4,r6
lfsx fr1,r5,r6
fadds fr0,fr0,fr1
stfsx fr0,r3,r6
addi r6,r6,4
bdnz 1b
b fpdisable
/*
* Vector subtract, floating point.
*/
_GLOBAL(vsubfp)
mflr r12
bl fpenable
li r0,4
mtctr r0
li r6,0
1: lfsx fr0,r4,r6
lfsx fr1,r5,r6
fsubs fr0,fr0,fr1
stfsx fr0,r3,r6
addi r6,r6,4
bdnz 1b
b fpdisable
/*
* Vector multiply and add, floating point.
*/
_GLOBAL(vmaddfp)
mflr r12
bl fpenable
stfd fr2,32(r1)
li r0,4
mtctr r0
li r7,0
1: lfsx fr0,r4,r7
lfsx fr1,r5,r7
lfsx fr2,r6,r7
fmadds fr0,fr0,fr2,fr1
stfsx fr0,r3,r7
addi r7,r7,4
bdnz 1b
lfd fr2,32(r1)
b fpdisable
/*
* Vector negative multiply and subtract, floating point.
*/
_GLOBAL(vnmsubfp)
mflr r12
bl fpenable
stfd fr2,32(r1)
li r0,4
mtctr r0
li r7,0
1: lfsx fr0,r4,r7
lfsx fr1,r5,r7
lfsx fr2,r6,r7
fnmsubs fr0,fr0,fr2,fr1
stfsx fr0,r3,r7
addi r7,r7,4
bdnz 1b
lfd fr2,32(r1)
b fpdisable
/*
* Vector reciprocal estimate. We just compute 1.0/x.
* r3 -> destination, r4 -> source.
*/
_GLOBAL(vrefp)
mflr r12
bl fpenable
li r0,4
LDCONST(fr1, fpone)
mtctr r0
li r6,0
1: lfsx fr0,r4,r6
fdivs fr0,fr1,fr0
stfsx fr0,r3,r6
addi r6,r6,4
bdnz 1b
b fpdisable
/*
* Vector reciprocal square-root estimate, floating point.
* We use the frsqrte instruction for the initial estimate followed
* by 2 iterations of Newton-Raphson to get sufficient accuracy.
* r3 -> destination, r4 -> source.
*/
_GLOBAL(vrsqrtefp)
mflr r12
bl fpenable
stfd fr2,32(r1)
stfd fr3,40(r1)
stfd fr4,48(r1)
stfd fr5,56(r1)
li r0,4
LDCONST(fr4, fpone)
LDCONST(fr5, fphalf)
mtctr r0
li r6,0
1: lfsx fr0,r4,r6
frsqrte fr1,fr0 /* r = frsqrte(s) */
fmuls fr3,fr1,fr0 /* r * s */
fmuls fr2,fr1,fr5 /* r * 0.5 */
fnmsubs fr3,fr1,fr3,fr4 /* 1 - s * r * r */
fmadds fr1,fr2,fr3,fr1 /* r = r + 0.5 * r * (1 - s * r * r) */
fmuls fr3,fr1,fr0 /* r * s */
fmuls fr2,fr1,fr5 /* r * 0.5 */
fnmsubs fr3,fr1,fr3,fr4 /* 1 - s * r * r */
fmadds fr1,fr2,fr3,fr1 /* r = r + 0.5 * r * (1 - s * r * r) */
stfsx fr1,r3,r6
addi r6,r6,4
bdnz 1b
lfd fr5,56(r1)
lfd fr4,48(r1)
lfd fr3,40(r1)
lfd fr2,32(r1)
b fpdisable