Hi Michael, On Thu, May 24, 2018 at 05:44:33PM +1000, Michael Ellerman wrote: > Hi Simon, > > wei.guo.si...@gmail.com writes: > > From: Simon Guo <wei.guo.si...@gmail.com> > > > > This patch add VMX primitives to do memcmp() in case the compare size > > exceeds 4K bytes. KSM feature can benefit from this. > > You say "exceeds 4K" here. > it should be >= 4k. I will correct the message.
> > diff --git a/arch/powerpc/lib/memcmp_64.S b/arch/powerpc/lib/memcmp_64.S > > index f20e883..6303bbf 100644 > > --- a/arch/powerpc/lib/memcmp_64.S > > +++ b/arch/powerpc/lib/memcmp_64.S > > @@ -27,12 +27,73 @@ > > #define LH lhbrx > > #define LW lwbrx > > #define LD ldbrx > > +#define LVS lvsr > > +#define VPERM(_VRT,_VRA,_VRB,_VRC) \ > > + vperm _VRT,_VRB,_VRA,_VRC > > #else > > #define LH lhzx > > #define LW lwzx > > #define LD ldx > > +#define LVS lvsl > > +#define VPERM(_VRT,_VRA,_VRB,_VRC) \ > > + vperm _VRT,_VRA,_VRB,_VRC > > #endif > > > > +#define VMX_OPS_THRES 4096 > > THRES == 4096 > > BTW, can we call it VMX_THRESH ? > Sure. I will update it. > > +#define ENTER_VMX_OPS \ > > + mflr r0; \ > > + std r3,-STACKFRAMESIZE+STK_REG(R31)(r1); \ > > + std r4,-STACKFRAMESIZE+STK_REG(R30)(r1); \ > > + std r5,-STACKFRAMESIZE+STK_REG(R29)(r1); \ > > + std r0,16(r1); \ > > + stdu r1,-STACKFRAMESIZE(r1); \ > > + bl enter_vmx_ops; \ > > + cmpwi cr1,r3,0; \ > > + ld r0,STACKFRAMESIZE+16(r1); \ > > + ld r3,STK_REG(R31)(r1); \ > > + ld r4,STK_REG(R30)(r1); \ > > + ld r5,STK_REG(R29)(r1); \ > > + addi r1,r1,STACKFRAMESIZE; \ > > + mtlr r0 > > + > > +#define EXIT_VMX_OPS \ > > + mflr r0; \ > > + std r3,-STACKFRAMESIZE+STK_REG(R31)(r1); \ > > + std r4,-STACKFRAMESIZE+STK_REG(R30)(r1); \ > > + std r5,-STACKFRAMESIZE+STK_REG(R29)(r1); \ > > + std r0,16(r1); \ > > + stdu r1,-STACKFRAMESIZE(r1); \ > > + bl exit_vmx_ops; \ > > + ld r0,STACKFRAMESIZE+16(r1); \ > > + ld r3,STK_REG(R31)(r1); \ > > + ld r4,STK_REG(R30)(r1); \ > > + ld r5,STK_REG(R29)(r1); \ > > + addi r1,r1,STACKFRAMESIZE; \ > > + mtlr r0 > > + > > +/* > > + * LD_VSR_CROSS16B load the 2nd 16 bytes for _vaddr which is unaligned with > > + * 16 bytes boundary and permute the result with the 1st 16 bytes. > > + > > + * | y y y y y y y y y y y y y 0 1 2 | 3 4 5 6 7 8 9 a b c d e f z z z > > | > > + * ^ ^ > > ^ > > + * 0xbbbb10 0xbbbb20 > > 0xbbb30 > > + * ^ > > + * _vaddr > > + * > > + * > > + * _vmask is the mask generated by LVS > > + * _v1st_qw is the 1st aligned QW of current addr which is already loaded. > > + * for example: 0xyyyyyyyyyyyyy012 for big endian > > + * _v2nd_qw is the 2nd aligned QW of cur _vaddr to be loaded. > > + * for example: 0x3456789abcdefzzz for big endian > > + * The permute result is saved in _v_res. > > + * for example: 0x0123456789abcdef for big endian. > > + */ > > +#define LD_VSR_CROSS16B(_vaddr,_vmask,_v1st_qw,_v2nd_qw,_v_res) \ > > + lvx _v2nd_qw,_vaddr,off16; \ > > + VPERM(_v_res,_v1st_qw,_v2nd_qw,_vmask) > > + > > /* > > * There are 2 categories for memcmp: > > * 1) src/dst has the same offset to the 8 bytes boundary. The handlers > > @@ -174,6 +235,13 @@ _GLOBAL(memcmp) > > blr > > > > .Llong: > > +#ifdef CONFIG_ALTIVEC > > + /* Try to use vmx loop if length is larger than 4K */ > > + cmpldi cr6,r5,VMX_OPS_THRES > > + bge cr6,.Lsameoffset_vmx_cmp > > Here we compare the length to 4K and if it's greater *or equal* then we > go to the VMX case. Or am I reading it backward? > > So we should say "if the size is 4K or more we do VMX" shouldn't we? Yes. Again I need reword the comment to "equal or greater than 4K" here. Thanks, - Simon
